]>
Commit | Line | Data |
---|---|---|
83ffe9cd | 1 | /* Copyright (C) 1988-2023 Free Software Foundation, Inc. |
2bf6d935 ML |
2 | |
3 | This file is part of GCC. | |
4 | ||
5 | GCC is free software; you can redistribute it and/or modify | |
6 | it under the terms of the GNU General Public License as published by | |
7 | the Free Software Foundation; either version 3, or (at your option) | |
8 | any later version. | |
9 | ||
10 | GCC is distributed in the hope that it will be useful, | |
11 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
13 | GNU General Public License for more details. | |
14 | ||
15 | You should have received a copy of the GNU General Public License | |
16 | along with GCC; see the file COPYING3. If not see | |
17 | <http://www.gnu.org/licenses/>. */ | |
18 | ||
19 | #define IN_TARGET_CODE 1 | |
20 | ||
21 | #include "config.h" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
24 | #include "backend.h" | |
25 | #include "rtl.h" | |
26 | #include "tree.h" | |
27 | #include "memmodel.h" | |
28 | #include "gimple.h" | |
29 | #include "cfghooks.h" | |
30 | #include "cfgloop.h" | |
31 | #include "df.h" | |
32 | #include "tm_p.h" | |
33 | #include "stringpool.h" | |
34 | #include "expmed.h" | |
35 | #include "optabs.h" | |
36 | #include "regs.h" | |
37 | #include "emit-rtl.h" | |
38 | #include "recog.h" | |
39 | #include "cgraph.h" | |
40 | #include "diagnostic.h" | |
41 | #include "cfgbuild.h" | |
42 | #include "alias.h" | |
43 | #include "fold-const.h" | |
44 | #include "attribs.h" | |
45 | #include "calls.h" | |
46 | #include "stor-layout.h" | |
47 | #include "varasm.h" | |
48 | #include "output.h" | |
49 | #include "insn-attr.h" | |
50 | #include "flags.h" | |
51 | #include "except.h" | |
52 | #include "explow.h" | |
53 | #include "expr.h" | |
54 | #include "cfgrtl.h" | |
55 | #include "common/common-target.h" | |
56 | #include "langhooks.h" | |
57 | #include "reload.h" | |
58 | #include "gimplify.h" | |
59 | #include "dwarf2.h" | |
60 | #include "tm-constrs.h" | |
2bf6d935 ML |
61 | #include "cselib.h" |
62 | #include "sched-int.h" | |
63 | #include "opts.h" | |
64 | #include "tree-pass.h" | |
65 | #include "context.h" | |
66 | #include "pass_manager.h" | |
67 | #include "target-globals.h" | |
68 | #include "gimple-iterator.h" | |
69 | #include "tree-vectorizer.h" | |
70 | #include "shrink-wrap.h" | |
71 | #include "builtins.h" | |
72 | #include "rtl-iter.h" | |
73 | #include "tree-iterator.h" | |
74 | #include "dbgcnt.h" | |
75 | #include "case-cfn-macros.h" | |
76 | #include "dojump.h" | |
77 | #include "fold-const-call.h" | |
78 | #include "tree-vrp.h" | |
79 | #include "tree-ssanames.h" | |
80 | #include "selftest.h" | |
81 | #include "selftest-rtl.h" | |
82 | #include "print-rtl.h" | |
83 | #include "intl.h" | |
84 | #include "ifcvt.h" | |
85 | #include "symbol-summary.h" | |
86 | #include "ipa-prop.h" | |
87 | #include "ipa-fnsummary.h" | |
88 | #include "wide-int-bitmask.h" | |
89 | #include "tree-vector-builder.h" | |
90 | #include "debug.h" | |
91 | #include "dwarf2out.h" | |
92 | #include "i386-options.h" | |
93 | #include "i386-builtins.h" | |
94 | #include "i386-expand.h" | |
bb576017 | 95 | #include "asan.h" |
2bf6d935 ML |
96 | |
97 | /* Split one or more double-mode RTL references into pairs of half-mode | |
98 | references. The RTL can be REG, offsettable MEM, integer constant, or | |
99 | CONST_DOUBLE. "operands" is a pointer to an array of double-mode RTLs to | |
100 | split and "num" is its length. lo_half and hi_half are output arrays | |
101 | that parallel "operands". */ | |
102 | ||
103 | void | |
104 | split_double_mode (machine_mode mode, rtx operands[], | |
105 | int num, rtx lo_half[], rtx hi_half[]) | |
106 | { | |
107 | machine_mode half_mode; | |
108 | unsigned int byte; | |
deeedbad JJ |
109 | rtx mem_op = NULL_RTX; |
110 | int mem_num = 0; | |
2bf6d935 ML |
111 | |
112 | switch (mode) | |
113 | { | |
114 | case E_TImode: | |
115 | half_mode = DImode; | |
116 | break; | |
117 | case E_DImode: | |
118 | half_mode = SImode; | |
119 | break; | |
58d6eea0 | 120 | case E_P2HImode: |
121 | half_mode = HImode; | |
122 | break; | |
123 | case E_P2QImode: | |
124 | half_mode = QImode; | |
125 | break; | |
2bf6d935 ML |
126 | default: |
127 | gcc_unreachable (); | |
128 | } | |
129 | ||
130 | byte = GET_MODE_SIZE (half_mode); | |
131 | ||
132 | while (num--) | |
133 | { | |
134 | rtx op = operands[num]; | |
135 | ||
136 | /* simplify_subreg refuse to split volatile memory addresses, | |
137 | but we still have to handle it. */ | |
138 | if (MEM_P (op)) | |
139 | { | |
deeedbad JJ |
140 | if (mem_op && rtx_equal_p (op, mem_op)) |
141 | { | |
142 | lo_half[num] = lo_half[mem_num]; | |
143 | hi_half[num] = hi_half[mem_num]; | |
144 | } | |
145 | else | |
146 | { | |
147 | mem_op = op; | |
148 | mem_num = num; | |
149 | lo_half[num] = adjust_address (op, half_mode, 0); | |
150 | hi_half[num] = adjust_address (op, half_mode, byte); | |
151 | } | |
2bf6d935 ML |
152 | } |
153 | else | |
154 | { | |
155 | lo_half[num] = simplify_gen_subreg (half_mode, op, | |
156 | GET_MODE (op) == VOIDmode | |
157 | ? mode : GET_MODE (op), 0); | |
d39fbed7 UB |
158 | |
159 | rtx tmp = simplify_gen_subreg (half_mode, op, | |
160 | GET_MODE (op) == VOIDmode | |
161 | ? mode : GET_MODE (op), byte); | |
162 | /* simplify_gen_subreg will return NULL RTX for the | |
163 | high half of the paradoxical subreg. */ | |
164 | hi_half[num] = tmp ? tmp : gen_reg_rtx (half_mode); | |
2bf6d935 ML |
165 | } |
166 | } | |
167 | } | |
168 | ||
16aafa31 RS |
169 | /* Emit the double word assignment DST = { LO, HI }. */ |
170 | ||
171 | void | |
172 | split_double_concat (machine_mode mode, rtx dst, rtx lo, rtx hi) | |
173 | { | |
174 | rtx dlo, dhi; | |
175 | int deleted_move_count = 0; | |
176 | split_double_mode (mode, &dst, 1, &dlo, &dhi); | |
2c089640 JJ |
177 | /* Constraints ensure that if both lo and hi are MEMs, then |
178 | dst has early-clobber and thus addresses of MEMs don't use | |
179 | dlo/dhi registers. Otherwise if at least one of li and hi are MEMs, | |
180 | dlo/dhi are registers. */ | |
181 | if (MEM_P (lo) | |
182 | && rtx_equal_p (dlo, hi) | |
183 | && reg_overlap_mentioned_p (dhi, lo)) | |
184 | { | |
185 | /* If dlo is same as hi and lo's address uses dhi register, | |
186 | code below would first emit_move_insn (dhi, hi) | |
187 | and then emit_move_insn (dlo, lo). But the former | |
188 | would invalidate lo's address. Load into dhi first, | |
189 | then swap. */ | |
190 | emit_move_insn (dhi, lo); | |
191 | lo = dhi; | |
192 | } | |
193 | else if (MEM_P (hi) | |
194 | && !MEM_P (lo) | |
195 | && !rtx_equal_p (dlo, lo) | |
196 | && reg_overlap_mentioned_p (dlo, hi)) | |
197 | { | |
198 | /* In this case, code below would first emit_move_insn (dlo, lo) | |
199 | and then emit_move_insn (dhi, hi). But the former would | |
42630fad JJ |
200 | invalidate hi's address. */ |
201 | if (rtx_equal_p (dhi, lo)) | |
202 | { | |
203 | /* We can't load into dhi first, so load into dlo | |
204 | first and we'll swap. */ | |
205 | emit_move_insn (dlo, hi); | |
206 | hi = dlo; | |
207 | } | |
208 | else | |
209 | { | |
210 | /* Load into dhi first. */ | |
211 | emit_move_insn (dhi, hi); | |
212 | hi = dhi; | |
213 | } | |
2c089640 | 214 | } |
16aafa31 RS |
215 | if (!rtx_equal_p (dlo, hi)) |
216 | { | |
217 | if (!rtx_equal_p (dlo, lo)) | |
218 | emit_move_insn (dlo, lo); | |
219 | else | |
220 | deleted_move_count++; | |
221 | if (!rtx_equal_p (dhi, hi)) | |
222 | emit_move_insn (dhi, hi); | |
223 | else | |
224 | deleted_move_count++; | |
225 | } | |
226 | else if (!rtx_equal_p (lo, dhi)) | |
227 | { | |
228 | if (!rtx_equal_p (dhi, hi)) | |
229 | emit_move_insn (dhi, hi); | |
230 | else | |
231 | deleted_move_count++; | |
232 | if (!rtx_equal_p (dlo, lo)) | |
233 | emit_move_insn (dlo, lo); | |
234 | else | |
235 | deleted_move_count++; | |
236 | } | |
237 | else if (mode == TImode) | |
238 | emit_insn (gen_swapdi (dlo, dhi)); | |
239 | else | |
240 | emit_insn (gen_swapsi (dlo, dhi)); | |
241 | ||
242 | if (deleted_move_count == 2) | |
243 | emit_note (NOTE_INSN_DELETED); | |
244 | } | |
245 | ||
246 | ||
2bf6d935 ML |
247 | /* Generate either "mov $0, reg" or "xor reg, reg", as appropriate |
248 | for the target. */ | |
249 | ||
250 | void | |
251 | ix86_expand_clear (rtx dest) | |
252 | { | |
253 | rtx tmp; | |
254 | ||
255 | /* We play register width games, which are only valid after reload. */ | |
256 | gcc_assert (reload_completed); | |
257 | ||
258 | /* Avoid HImode and its attendant prefix byte. */ | |
259 | if (GET_MODE_SIZE (GET_MODE (dest)) < 4) | |
260 | dest = gen_rtx_REG (SImode, REGNO (dest)); | |
261 | tmp = gen_rtx_SET (dest, const0_rtx); | |
262 | ||
263 | if (!TARGET_USE_MOV0 || optimize_insn_for_size_p ()) | |
264 | { | |
265 | rtx clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG)); | |
266 | tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, tmp, clob)); | |
267 | } | |
268 | ||
269 | emit_insn (tmp); | |
270 | } | |
271 | ||
edafb35b L |
272 | /* Return true if V can be broadcasted from an integer of WIDTH bits |
273 | which is returned in VAL_BROADCAST. Otherwise, return false. */ | |
274 | ||
275 | static bool | |
276 | ix86_broadcast (HOST_WIDE_INT v, unsigned int width, | |
277 | HOST_WIDE_INT &val_broadcast) | |
278 | { | |
279 | wide_int val = wi::uhwi (v, HOST_BITS_PER_WIDE_INT); | |
280 | val_broadcast = wi::extract_uhwi (val, 0, width); | |
281 | for (unsigned int i = width; i < HOST_BITS_PER_WIDE_INT; i += width) | |
282 | { | |
283 | HOST_WIDE_INT each = wi::extract_uhwi (val, i, width); | |
284 | if (val_broadcast != each) | |
285 | return false; | |
286 | } | |
287 | val_broadcast = sext_hwi (val_broadcast, width); | |
288 | return true; | |
289 | } | |
290 | ||
291 | /* Convert the CONST_WIDE_INT operand OP to broadcast in MODE. */ | |
292 | ||
293 | static rtx | |
294 | ix86_convert_const_wide_int_to_broadcast (machine_mode mode, rtx op) | |
295 | { | |
296 | /* Don't use integer vector broadcast if we can't move from GPR to SSE | |
297 | register directly. */ | |
298 | if (!TARGET_INTER_UNIT_MOVES_TO_VEC) | |
299 | return nullptr; | |
300 | ||
301 | /* Convert CONST_WIDE_INT to a non-standard SSE constant integer | |
302 | broadcast only if vector broadcast is available. */ | |
303 | if (!TARGET_AVX | |
304 | || !CONST_WIDE_INT_P (op) | |
963315a9 JJ |
305 | || standard_sse_constant_p (op, mode) |
306 | || (CONST_WIDE_INT_NUNITS (op) * HOST_BITS_PER_WIDE_INT | |
307 | != GET_MODE_BITSIZE (mode))) | |
edafb35b L |
308 | return nullptr; |
309 | ||
310 | HOST_WIDE_INT val = CONST_WIDE_INT_ELT (op, 0); | |
311 | HOST_WIDE_INT val_broadcast; | |
312 | scalar_int_mode broadcast_mode; | |
313 | if (TARGET_AVX2 | |
314 | && ix86_broadcast (val, GET_MODE_BITSIZE (QImode), | |
315 | val_broadcast)) | |
316 | broadcast_mode = QImode; | |
317 | else if (TARGET_AVX2 | |
318 | && ix86_broadcast (val, GET_MODE_BITSIZE (HImode), | |
319 | val_broadcast)) | |
320 | broadcast_mode = HImode; | |
321 | else if (ix86_broadcast (val, GET_MODE_BITSIZE (SImode), | |
322 | val_broadcast)) | |
323 | broadcast_mode = SImode; | |
324 | else if (TARGET_64BIT | |
325 | && ix86_broadcast (val, GET_MODE_BITSIZE (DImode), | |
326 | val_broadcast)) | |
327 | broadcast_mode = DImode; | |
328 | else | |
329 | return nullptr; | |
330 | ||
331 | /* Check if OP can be broadcasted from VAL. */ | |
332 | for (int i = 1; i < CONST_WIDE_INT_NUNITS (op); i++) | |
333 | if (val != CONST_WIDE_INT_ELT (op, i)) | |
334 | return nullptr; | |
335 | ||
336 | unsigned int nunits = (GET_MODE_SIZE (mode) | |
337 | / GET_MODE_SIZE (broadcast_mode)); | |
338 | machine_mode vector_mode; | |
339 | if (!mode_for_vector (broadcast_mode, nunits).exists (&vector_mode)) | |
340 | gcc_unreachable (); | |
77127363 | 341 | rtx target = gen_reg_rtx (vector_mode); |
edafb35b L |
342 | bool ok = ix86_expand_vector_init_duplicate (false, vector_mode, |
343 | target, | |
344 | GEN_INT (val_broadcast)); | |
345 | gcc_assert (ok); | |
346 | target = lowpart_subreg (mode, target, vector_mode); | |
347 | return target; | |
348 | } | |
349 | ||
2bf6d935 ML |
350 | void |
351 | ix86_expand_move (machine_mode mode, rtx operands[]) | |
352 | { | |
353 | rtx op0, op1; | |
354 | rtx tmp, addend = NULL_RTX; | |
355 | enum tls_model model; | |
356 | ||
357 | op0 = operands[0]; | |
358 | op1 = operands[1]; | |
359 | ||
be39636d RS |
360 | /* Avoid complex sets of likely spilled hard registers before reload. */ |
361 | if (!ix86_hardreg_mov_ok (op0, op1)) | |
362 | { | |
363 | tmp = gen_reg_rtx (mode); | |
364 | operands[0] = tmp; | |
365 | ix86_expand_move (mode, operands); | |
366 | operands[0] = op0; | |
367 | operands[1] = tmp; | |
368 | op1 = tmp; | |
369 | } | |
370 | ||
2bf6d935 ML |
371 | switch (GET_CODE (op1)) |
372 | { | |
373 | case CONST: | |
374 | tmp = XEXP (op1, 0); | |
375 | ||
376 | if (GET_CODE (tmp) != PLUS | |
377 | || GET_CODE (XEXP (tmp, 0)) != SYMBOL_REF) | |
378 | break; | |
379 | ||
380 | op1 = XEXP (tmp, 0); | |
381 | addend = XEXP (tmp, 1); | |
382 | /* FALLTHRU */ | |
383 | ||
384 | case SYMBOL_REF: | |
385 | model = SYMBOL_REF_TLS_MODEL (op1); | |
386 | ||
387 | if (model) | |
388 | op1 = legitimize_tls_address (op1, model, true); | |
389 | else if (ix86_force_load_from_GOT_p (op1)) | |
390 | { | |
391 | /* Load the external function address via GOT slot to avoid PLT. */ | |
392 | op1 = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, op1), | |
393 | (TARGET_64BIT | |
394 | ? UNSPEC_GOTPCREL | |
395 | : UNSPEC_GOT)); | |
396 | op1 = gen_rtx_CONST (Pmode, op1); | |
397 | op1 = gen_const_mem (Pmode, op1); | |
398 | set_mem_alias_set (op1, ix86_GOT_alias_set ()); | |
399 | } | |
400 | else | |
401 | { | |
402 | tmp = legitimize_pe_coff_symbol (op1, addend != NULL_RTX); | |
403 | if (tmp) | |
404 | { | |
405 | op1 = tmp; | |
406 | if (!addend) | |
407 | break; | |
408 | } | |
409 | else | |
410 | { | |
411 | op1 = operands[1]; | |
412 | break; | |
413 | } | |
414 | } | |
415 | ||
416 | if (addend) | |
417 | { | |
418 | op1 = force_operand (op1, NULL_RTX); | |
419 | op1 = expand_simple_binop (Pmode, PLUS, op1, addend, | |
420 | op0, 1, OPTAB_DIRECT); | |
421 | } | |
422 | else | |
423 | op1 = force_operand (op1, op0); | |
424 | ||
425 | if (op1 == op0) | |
426 | return; | |
427 | ||
428 | op1 = convert_to_mode (mode, op1, 1); | |
429 | ||
430 | default: | |
431 | break; | |
432 | } | |
433 | ||
434 | if ((flag_pic || MACHOPIC_INDIRECT) | |
435 | && symbolic_operand (op1, mode)) | |
436 | { | |
437 | if (TARGET_MACHO && !TARGET_64BIT) | |
438 | { | |
439 | #if TARGET_MACHO | |
440 | /* dynamic-no-pic */ | |
441 | if (MACHOPIC_INDIRECT) | |
442 | { | |
443 | rtx temp = (op0 && REG_P (op0) && mode == Pmode) | |
444 | ? op0 : gen_reg_rtx (Pmode); | |
445 | op1 = machopic_indirect_data_reference (op1, temp); | |
446 | if (MACHOPIC_PURE) | |
447 | op1 = machopic_legitimize_pic_address (op1, mode, | |
448 | temp == op1 ? 0 : temp); | |
449 | } | |
450 | if (op0 != op1 && GET_CODE (op0) != MEM) | |
451 | { | |
452 | rtx insn = gen_rtx_SET (op0, op1); | |
453 | emit_insn (insn); | |
454 | return; | |
455 | } | |
456 | if (GET_CODE (op0) == MEM) | |
457 | op1 = force_reg (Pmode, op1); | |
458 | else | |
459 | { | |
460 | rtx temp = op0; | |
461 | if (GET_CODE (temp) != REG) | |
462 | temp = gen_reg_rtx (Pmode); | |
463 | temp = legitimize_pic_address (op1, temp); | |
464 | if (temp == op0) | |
465 | return; | |
466 | op1 = temp; | |
467 | } | |
468 | /* dynamic-no-pic */ | |
469 | #endif | |
470 | } | |
471 | else | |
472 | { | |
473 | if (MEM_P (op0)) | |
474 | op1 = force_reg (mode, op1); | |
475 | else if (!(TARGET_64BIT && x86_64_movabs_operand (op1, DImode))) | |
476 | { | |
477 | rtx reg = can_create_pseudo_p () ? NULL_RTX : op0; | |
478 | op1 = legitimize_pic_address (op1, reg); | |
479 | if (op0 == op1) | |
480 | return; | |
481 | op1 = convert_to_mode (mode, op1, 1); | |
482 | } | |
483 | } | |
484 | } | |
485 | else | |
486 | { | |
487 | if (MEM_P (op0) | |
488 | && (PUSH_ROUNDING (GET_MODE_SIZE (mode)) != GET_MODE_SIZE (mode) | |
489 | || !push_operand (op0, mode)) | |
490 | && MEM_P (op1)) | |
491 | op1 = force_reg (mode, op1); | |
492 | ||
493 | if (push_operand (op0, mode) | |
494 | && ! general_no_elim_operand (op1, mode)) | |
495 | op1 = copy_to_mode_reg (mode, op1); | |
496 | ||
497 | /* Force large constants in 64bit compilation into register | |
498 | to get them CSEed. */ | |
499 | if (can_create_pseudo_p () | |
500 | && (mode == DImode) && TARGET_64BIT | |
501 | && immediate_operand (op1, mode) | |
502 | && !x86_64_zext_immediate_operand (op1, VOIDmode) | |
503 | && !register_operand (op0, mode) | |
504 | && optimize) | |
505 | op1 = copy_to_mode_reg (mode, op1); | |
506 | ||
edafb35b | 507 | if (can_create_pseudo_p ()) |
2bf6d935 | 508 | { |
edafb35b | 509 | if (CONST_DOUBLE_P (op1)) |
2bf6d935 | 510 | { |
edafb35b L |
511 | /* If we are loading a floating point constant to a |
512 | register, force the value to memory now, since we'll | |
513 | get better code out the back end. */ | |
514 | ||
515 | op1 = validize_mem (force_const_mem (mode, op1)); | |
516 | if (!register_operand (op0, mode)) | |
517 | { | |
518 | rtx temp = gen_reg_rtx (mode); | |
519 | emit_insn (gen_rtx_SET (temp, op1)); | |
520 | emit_move_insn (op0, temp); | |
521 | return; | |
522 | } | |
523 | } | |
96c3539f RS |
524 | else if (CONST_WIDE_INT_P (op1) |
525 | && GET_MODE_SIZE (mode) >= 16) | |
edafb35b L |
526 | { |
527 | rtx tmp = ix86_convert_const_wide_int_to_broadcast | |
528 | (GET_MODE (op0), op1); | |
529 | if (tmp != nullptr) | |
530 | op1 = tmp; | |
2bf6d935 ML |
531 | } |
532 | } | |
533 | } | |
534 | ||
535 | emit_insn (gen_rtx_SET (op0, op1)); | |
536 | } | |
537 | ||
a6291d88 | 538 | /* OP is a memref of CONST_VECTOR, return scalar constant mem |
539 | if CONST_VECTOR is a vec_duplicate, else return NULL. */ | |
edafb35b | 540 | static rtx |
a6291d88 | 541 | ix86_broadcast_from_constant (machine_mode mode, rtx op) |
edafb35b L |
542 | { |
543 | int nunits = GET_MODE_NUNITS (mode); | |
544 | if (nunits < 2) | |
545 | return nullptr; | |
546 | ||
547 | /* Don't use integer vector broadcast if we can't move from GPR to SSE | |
548 | register directly. */ | |
a6291d88 | 549 | if (!TARGET_INTER_UNIT_MOVES_TO_VEC |
550 | && INTEGRAL_MODE_P (mode)) | |
edafb35b L |
551 | return nullptr; |
552 | ||
553 | /* Convert CONST_VECTOR to a non-standard SSE constant integer | |
554 | broadcast only if vector broadcast is available. */ | |
555 | if (!(TARGET_AVX2 | |
556 | || (TARGET_AVX | |
557 | && (GET_MODE_INNER (mode) == SImode | |
a6291d88 | 558 | || GET_MODE_INNER (mode) == DImode)) |
559 | || FLOAT_MODE_P (mode)) | |
edafb35b L |
560 | || standard_sse_constant_p (op, mode)) |
561 | return nullptr; | |
562 | ||
a6291d88 | 563 | /* Don't broadcast from a 64-bit integer constant in 32-bit mode. |
564 | We can still put 64-bit integer constant in memory when | |
565 | avx512 embed broadcast is available. */ | |
566 | if (GET_MODE_INNER (mode) == DImode && !TARGET_64BIT | |
567 | && (!TARGET_AVX512F | |
568 | || (GET_MODE_SIZE (mode) < 64 && !TARGET_AVX512VL))) | |
edafb35b L |
569 | return nullptr; |
570 | ||
f7cad1a0 JJ |
571 | if (GET_MODE_INNER (mode) == TImode) |
572 | return nullptr; | |
573 | ||
edafb35b L |
574 | rtx constant = get_pool_constant (XEXP (op, 0)); |
575 | if (GET_CODE (constant) != CONST_VECTOR) | |
576 | return nullptr; | |
577 | ||
578 | /* There could be some rtx like | |
579 | (mem/u/c:V16QI (symbol_ref/u:DI ("*.LC1"))) | |
580 | but with "*.LC1" refer to V2DI constant vector. */ | |
581 | if (GET_MODE (constant) != mode) | |
582 | { | |
583 | constant = simplify_subreg (mode, constant, GET_MODE (constant), | |
584 | 0); | |
585 | if (constant == nullptr || GET_CODE (constant) != CONST_VECTOR) | |
586 | return nullptr; | |
587 | } | |
588 | ||
589 | rtx first = XVECEXP (constant, 0, 0); | |
590 | ||
591 | for (int i = 1; i < nunits; ++i) | |
592 | { | |
593 | rtx tmp = XVECEXP (constant, 0, i); | |
594 | /* Vector duplicate value. */ | |
595 | if (!rtx_equal_p (tmp, first)) | |
596 | return nullptr; | |
597 | } | |
598 | ||
599 | return first; | |
600 | } | |
601 | ||
2bf6d935 ML |
602 | void |
603 | ix86_expand_vector_move (machine_mode mode, rtx operands[]) | |
604 | { | |
605 | rtx op0 = operands[0], op1 = operands[1]; | |
606 | /* Use GET_MODE_BITSIZE instead of GET_MODE_ALIGNMENT for IA MCU | |
607 | psABI since the biggest alignment is 4 byte for IA MCU psABI. */ | |
608 | unsigned int align = (TARGET_IAMCU | |
609 | ? GET_MODE_BITSIZE (mode) | |
610 | : GET_MODE_ALIGNMENT (mode)); | |
611 | ||
612 | if (push_operand (op0, VOIDmode)) | |
613 | op0 = emit_move_resolve_push (mode, op0); | |
614 | ||
615 | /* Force constants other than zero into memory. We do not know how | |
616 | the instructions used to build constants modify the upper 64 bits | |
617 | of the register, once we have that information we may be able | |
618 | to handle some of them more efficiently. */ | |
619 | if (can_create_pseudo_p () | |
620 | && (CONSTANT_P (op1) | |
621 | || (SUBREG_P (op1) | |
622 | && CONSTANT_P (SUBREG_REG (op1)))) | |
623 | && ((register_operand (op0, mode) | |
624 | && !standard_sse_constant_p (op1, mode)) | |
625 | /* ix86_expand_vector_move_misalign() does not like constants. */ | |
626 | || (SSE_REG_MODE_P (mode) | |
627 | && MEM_P (op0) | |
628 | && MEM_ALIGN (op0) < align))) | |
629 | { | |
630 | if (SUBREG_P (op1)) | |
631 | { | |
632 | machine_mode imode = GET_MODE (SUBREG_REG (op1)); | |
633 | rtx r = force_const_mem (imode, SUBREG_REG (op1)); | |
634 | if (r) | |
635 | r = validize_mem (r); | |
636 | else | |
637 | r = force_reg (imode, SUBREG_REG (op1)); | |
638 | op1 = simplify_gen_subreg (mode, r, imode, SUBREG_BYTE (op1)); | |
639 | } | |
640 | else | |
edafb35b L |
641 | { |
642 | machine_mode mode = GET_MODE (op0); | |
643 | rtx tmp = ix86_convert_const_wide_int_to_broadcast | |
644 | (mode, op1); | |
645 | if (tmp == nullptr) | |
646 | op1 = validize_mem (force_const_mem (mode, op1)); | |
647 | else | |
648 | op1 = tmp; | |
649 | } | |
650 | } | |
651 | ||
652 | if (can_create_pseudo_p () | |
653 | && GET_MODE_SIZE (mode) >= 16 | |
a6291d88 | 654 | && VECTOR_MODE_P (mode) |
edafb35b L |
655 | && (MEM_P (op1) |
656 | && SYMBOL_REF_P (XEXP (op1, 0)) | |
657 | && CONSTANT_POOL_ADDRESS_P (XEXP (op1, 0)))) | |
658 | { | |
a6291d88 | 659 | rtx first = ix86_broadcast_from_constant (mode, op1); |
edafb35b L |
660 | if (first != nullptr) |
661 | { | |
662 | /* Broadcast to XMM/YMM/ZMM register from an integer | |
a6291d88 | 663 | constant or scalar mem. */ |
6e5401e8 | 664 | op1 = gen_reg_rtx (mode); |
a6291d88 | 665 | if (FLOAT_MODE_P (mode) |
666 | || (!TARGET_64BIT && GET_MODE_INNER (mode) == DImode)) | |
6e5401e8 | 667 | first = force_const_mem (GET_MODE_INNER (mode), first); |
edafb35b L |
668 | bool ok = ix86_expand_vector_init_duplicate (false, mode, |
669 | op1, first); | |
670 | gcc_assert (ok); | |
671 | emit_move_insn (op0, op1); | |
672 | return; | |
673 | } | |
2bf6d935 ML |
674 | } |
675 | ||
676 | /* We need to check memory alignment for SSE mode since attribute | |
677 | can make operands unaligned. */ | |
678 | if (can_create_pseudo_p () | |
679 | && SSE_REG_MODE_P (mode) | |
680 | && ((MEM_P (op0) && (MEM_ALIGN (op0) < align)) | |
681 | || (MEM_P (op1) && (MEM_ALIGN (op1) < align)))) | |
682 | { | |
683 | rtx tmp[2]; | |
684 | ||
685 | /* ix86_expand_vector_move_misalign() does not like both | |
686 | arguments in memory. */ | |
687 | if (!register_operand (op0, mode) | |
688 | && !register_operand (op1, mode)) | |
09dba016 | 689 | { |
77127363 | 690 | rtx scratch = gen_reg_rtx (mode); |
09dba016 L |
691 | emit_move_insn (scratch, op1); |
692 | op1 = scratch; | |
693 | } | |
2bf6d935 ML |
694 | |
695 | tmp[0] = op0; tmp[1] = op1; | |
696 | ix86_expand_vector_move_misalign (mode, tmp); | |
697 | return; | |
698 | } | |
699 | ||
96c3539f RS |
700 | /* Special case TImode to 128-bit vector conversions via V2DI. */ |
701 | if (VECTOR_MODE_P (mode) | |
702 | && GET_MODE_SIZE (mode) == 16 | |
fad14a02 RS |
703 | && SUBREG_P (op1) |
704 | && GET_MODE (SUBREG_REG (op1)) == TImode | |
705 | && TARGET_64BIT && TARGET_SSE | |
706 | && can_create_pseudo_p ()) | |
707 | { | |
708 | rtx tmp = gen_reg_rtx (V2DImode); | |
709 | rtx lo = gen_reg_rtx (DImode); | |
710 | rtx hi = gen_reg_rtx (DImode); | |
711 | emit_move_insn (lo, gen_lowpart (DImode, SUBREG_REG (op1))); | |
712 | emit_move_insn (hi, gen_highpart (DImode, SUBREG_REG (op1))); | |
713 | emit_insn (gen_vec_concatv2di (tmp, lo, hi)); | |
96c3539f | 714 | emit_move_insn (op0, gen_lowpart (mode, tmp)); |
fad14a02 RS |
715 | return; |
716 | } | |
717 | ||
97c32001 RS |
718 | /* If operand0 is a hard register, make operand1 a pseudo. */ |
719 | if (can_create_pseudo_p () | |
720 | && !ix86_hardreg_mov_ok (op0, op1)) | |
721 | { | |
722 | rtx tmp = gen_reg_rtx (GET_MODE (op0)); | |
723 | emit_move_insn (tmp, op1); | |
724 | emit_move_insn (op0, tmp); | |
725 | return; | |
726 | } | |
727 | ||
2bf6d935 ML |
728 | /* Make operand1 a register if it isn't already. */ |
729 | if (can_create_pseudo_p () | |
730 | && !register_operand (op0, mode) | |
731 | && !register_operand (op1, mode)) | |
732 | { | |
77127363 | 733 | rtx tmp = gen_reg_rtx (GET_MODE (op0)); |
7f4c3943 L |
734 | emit_move_insn (tmp, op1); |
735 | emit_move_insn (op0, tmp); | |
2bf6d935 ML |
736 | return; |
737 | } | |
738 | ||
739 | emit_insn (gen_rtx_SET (op0, op1)); | |
740 | } | |
741 | ||
742 | /* Split 32-byte AVX unaligned load and store if needed. */ | |
743 | ||
744 | static void | |
745 | ix86_avx256_split_vector_move_misalign (rtx op0, rtx op1) | |
746 | { | |
747 | rtx m; | |
748 | rtx (*extract) (rtx, rtx, rtx); | |
749 | machine_mode mode; | |
750 | ||
751 | if ((MEM_P (op1) && !TARGET_AVX256_SPLIT_UNALIGNED_LOAD) | |
752 | || (MEM_P (op0) && !TARGET_AVX256_SPLIT_UNALIGNED_STORE)) | |
753 | { | |
754 | emit_insn (gen_rtx_SET (op0, op1)); | |
755 | return; | |
756 | } | |
757 | ||
758 | rtx orig_op0 = NULL_RTX; | |
759 | mode = GET_MODE (op0); | |
760 | switch (GET_MODE_CLASS (mode)) | |
761 | { | |
762 | case MODE_VECTOR_INT: | |
763 | case MODE_INT: | |
764 | if (mode != V32QImode) | |
765 | { | |
766 | if (!MEM_P (op0)) | |
767 | { | |
768 | orig_op0 = op0; | |
769 | op0 = gen_reg_rtx (V32QImode); | |
770 | } | |
771 | else | |
772 | op0 = gen_lowpart (V32QImode, op0); | |
773 | op1 = gen_lowpart (V32QImode, op1); | |
774 | mode = V32QImode; | |
775 | } | |
776 | break; | |
777 | case MODE_VECTOR_FLOAT: | |
778 | break; | |
779 | default: | |
780 | gcc_unreachable (); | |
781 | } | |
782 | ||
783 | switch (mode) | |
784 | { | |
785 | default: | |
786 | gcc_unreachable (); | |
787 | case E_V32QImode: | |
788 | extract = gen_avx_vextractf128v32qi; | |
789 | mode = V16QImode; | |
790 | break; | |
60d1d296 L |
791 | case E_V16BFmode: |
792 | extract = gen_avx_vextractf128v16bf; | |
793 | mode = V8BFmode; | |
794 | break; | |
d959312b L |
795 | case E_V16HFmode: |
796 | extract = gen_avx_vextractf128v16hf; | |
797 | mode = V8HFmode; | |
798 | break; | |
2bf6d935 ML |
799 | case E_V8SFmode: |
800 | extract = gen_avx_vextractf128v8sf; | |
801 | mode = V4SFmode; | |
802 | break; | |
803 | case E_V4DFmode: | |
804 | extract = gen_avx_vextractf128v4df; | |
805 | mode = V2DFmode; | |
806 | break; | |
807 | } | |
808 | ||
809 | if (MEM_P (op1)) | |
810 | { | |
811 | rtx r = gen_reg_rtx (mode); | |
812 | m = adjust_address (op1, mode, 0); | |
813 | emit_move_insn (r, m); | |
814 | m = adjust_address (op1, mode, 16); | |
815 | r = gen_rtx_VEC_CONCAT (GET_MODE (op0), r, m); | |
816 | emit_move_insn (op0, r); | |
817 | } | |
818 | else if (MEM_P (op0)) | |
819 | { | |
820 | m = adjust_address (op0, mode, 0); | |
821 | emit_insn (extract (m, op1, const0_rtx)); | |
822 | m = adjust_address (op0, mode, 16); | |
823 | emit_insn (extract (m, copy_rtx (op1), const1_rtx)); | |
824 | } | |
825 | else | |
826 | gcc_unreachable (); | |
827 | ||
828 | if (orig_op0) | |
829 | emit_move_insn (orig_op0, gen_lowpart (GET_MODE (orig_op0), op0)); | |
830 | } | |
831 | ||
832 | /* Implement the movmisalign patterns for SSE. Non-SSE modes go | |
833 | straight to ix86_expand_vector_move. */ | |
834 | /* Code generation for scalar reg-reg moves of single and double precision data: | |
835 | if (x86_sse_partial_reg_dependency == true | x86_sse_split_regs == true) | |
836 | movaps reg, reg | |
837 | else | |
838 | movss reg, reg | |
839 | if (x86_sse_partial_reg_dependency == true) | |
840 | movapd reg, reg | |
841 | else | |
842 | movsd reg, reg | |
843 | ||
844 | Code generation for scalar loads of double precision data: | |
845 | if (x86_sse_split_regs == true) | |
846 | movlpd mem, reg (gas syntax) | |
847 | else | |
848 | movsd mem, reg | |
849 | ||
850 | Code generation for unaligned packed loads of single precision data | |
851 | (x86_sse_unaligned_move_optimal overrides x86_sse_partial_reg_dependency): | |
852 | if (x86_sse_unaligned_move_optimal) | |
853 | movups mem, reg | |
854 | ||
855 | if (x86_sse_partial_reg_dependency == true) | |
856 | { | |
857 | xorps reg, reg | |
858 | movlps mem, reg | |
859 | movhps mem+8, reg | |
860 | } | |
861 | else | |
862 | { | |
863 | movlps mem, reg | |
864 | movhps mem+8, reg | |
865 | } | |
866 | ||
867 | Code generation for unaligned packed loads of double precision data | |
868 | (x86_sse_unaligned_move_optimal overrides x86_sse_split_regs): | |
869 | if (x86_sse_unaligned_move_optimal) | |
870 | movupd mem, reg | |
871 | ||
872 | if (x86_sse_split_regs == true) | |
873 | { | |
874 | movlpd mem, reg | |
875 | movhpd mem+8, reg | |
876 | } | |
877 | else | |
878 | { | |
879 | movsd mem, reg | |
880 | movhpd mem+8, reg | |
881 | } | |
882 | */ | |
883 | ||
884 | void | |
885 | ix86_expand_vector_move_misalign (machine_mode mode, rtx operands[]) | |
886 | { | |
887 | rtx op0, op1, m; | |
888 | ||
889 | op0 = operands[0]; | |
890 | op1 = operands[1]; | |
891 | ||
892 | /* Use unaligned load/store for AVX512 or when optimizing for size. */ | |
893 | if (GET_MODE_SIZE (mode) == 64 || optimize_insn_for_size_p ()) | |
894 | { | |
895 | emit_insn (gen_rtx_SET (op0, op1)); | |
896 | return; | |
897 | } | |
898 | ||
899 | if (TARGET_AVX) | |
900 | { | |
901 | if (GET_MODE_SIZE (mode) == 32) | |
902 | ix86_avx256_split_vector_move_misalign (op0, op1); | |
903 | else | |
904 | /* Always use 128-bit mov<mode>_internal pattern for AVX. */ | |
905 | emit_insn (gen_rtx_SET (op0, op1)); | |
906 | return; | |
907 | } | |
908 | ||
909 | if (TARGET_SSE_UNALIGNED_LOAD_OPTIMAL | |
910 | || TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL) | |
911 | { | |
912 | emit_insn (gen_rtx_SET (op0, op1)); | |
913 | return; | |
914 | } | |
915 | ||
916 | /* ??? If we have typed data, then it would appear that using | |
917 | movdqu is the only way to get unaligned data loaded with | |
918 | integer type. */ | |
919 | if (TARGET_SSE2 && GET_MODE_CLASS (mode) == MODE_VECTOR_INT) | |
920 | { | |
921 | emit_insn (gen_rtx_SET (op0, op1)); | |
922 | return; | |
923 | } | |
924 | ||
925 | if (MEM_P (op1)) | |
926 | { | |
927 | if (TARGET_SSE2 && mode == V2DFmode) | |
928 | { | |
929 | rtx zero; | |
930 | ||
931 | /* When SSE registers are split into halves, we can avoid | |
932 | writing to the top half twice. */ | |
933 | if (TARGET_SSE_SPLIT_REGS) | |
934 | { | |
935 | emit_clobber (op0); | |
936 | zero = op0; | |
937 | } | |
938 | else | |
939 | { | |
940 | /* ??? Not sure about the best option for the Intel chips. | |
941 | The following would seem to satisfy; the register is | |
942 | entirely cleared, breaking the dependency chain. We | |
943 | then store to the upper half, with a dependency depth | |
944 | of one. A rumor has it that Intel recommends two movsd | |
945 | followed by an unpacklpd, but this is unconfirmed. And | |
946 | given that the dependency depth of the unpacklpd would | |
947 | still be one, I'm not sure why this would be better. */ | |
948 | zero = CONST0_RTX (V2DFmode); | |
949 | } | |
950 | ||
951 | m = adjust_address (op1, DFmode, 0); | |
952 | emit_insn (gen_sse2_loadlpd (op0, zero, m)); | |
953 | m = adjust_address (op1, DFmode, 8); | |
954 | emit_insn (gen_sse2_loadhpd (op0, op0, m)); | |
955 | } | |
956 | else | |
957 | { | |
958 | rtx t; | |
959 | ||
960 | if (mode != V4SFmode) | |
961 | t = gen_reg_rtx (V4SFmode); | |
962 | else | |
963 | t = op0; | |
964 | ||
965 | if (TARGET_SSE_PARTIAL_REG_DEPENDENCY) | |
966 | emit_move_insn (t, CONST0_RTX (V4SFmode)); | |
967 | else | |
968 | emit_clobber (t); | |
969 | ||
970 | m = adjust_address (op1, V2SFmode, 0); | |
971 | emit_insn (gen_sse_loadlps (t, t, m)); | |
972 | m = adjust_address (op1, V2SFmode, 8); | |
973 | emit_insn (gen_sse_loadhps (t, t, m)); | |
974 | if (mode != V4SFmode) | |
975 | emit_move_insn (op0, gen_lowpart (mode, t)); | |
976 | } | |
977 | } | |
978 | else if (MEM_P (op0)) | |
979 | { | |
980 | if (TARGET_SSE2 && mode == V2DFmode) | |
981 | { | |
982 | m = adjust_address (op0, DFmode, 0); | |
983 | emit_insn (gen_sse2_storelpd (m, op1)); | |
984 | m = adjust_address (op0, DFmode, 8); | |
985 | emit_insn (gen_sse2_storehpd (m, op1)); | |
986 | } | |
987 | else | |
988 | { | |
989 | if (mode != V4SFmode) | |
990 | op1 = gen_lowpart (V4SFmode, op1); | |
991 | ||
992 | m = adjust_address (op0, V2SFmode, 0); | |
993 | emit_insn (gen_sse_storelps (m, op1)); | |
994 | m = adjust_address (op0, V2SFmode, 8); | |
995 | emit_insn (gen_sse_storehps (m, copy_rtx (op1))); | |
996 | } | |
997 | } | |
998 | else | |
999 | gcc_unreachable (); | |
1000 | } | |
1001 | ||
b74ebb2a L |
1002 | /* Move bits 64:95 to bits 32:63. */ |
1003 | ||
1004 | void | |
1005 | ix86_move_vector_high_sse_to_mmx (rtx op) | |
1006 | { | |
1007 | rtx mask = gen_rtx_PARALLEL (VOIDmode, | |
1008 | gen_rtvec (4, GEN_INT (0), GEN_INT (2), | |
1009 | GEN_INT (0), GEN_INT (0))); | |
1010 | rtx dest = lowpart_subreg (V4SImode, op, GET_MODE (op)); | |
1011 | op = gen_rtx_VEC_SELECT (V4SImode, dest, mask); | |
1012 | rtx insn = gen_rtx_SET (dest, op); | |
1013 | emit_insn (insn); | |
1014 | } | |
1015 | ||
1016 | /* Split MMX pack with signed/unsigned saturation with SSE/SSE2. */ | |
1017 | ||
1018 | void | |
1019 | ix86_split_mmx_pack (rtx operands[], enum rtx_code code) | |
1020 | { | |
1021 | rtx op0 = operands[0]; | |
1022 | rtx op1 = operands[1]; | |
1023 | rtx op2 = operands[2]; | |
58e61a3a | 1024 | rtx src; |
b74ebb2a L |
1025 | |
1026 | machine_mode dmode = GET_MODE (op0); | |
1027 | machine_mode smode = GET_MODE (op1); | |
1028 | machine_mode inner_dmode = GET_MODE_INNER (dmode); | |
1029 | machine_mode inner_smode = GET_MODE_INNER (smode); | |
1030 | ||
1031 | /* Get the corresponding SSE mode for destination. */ | |
1032 | int nunits = 16 / GET_MODE_SIZE (inner_dmode); | |
1033 | machine_mode sse_dmode = mode_for_vector (GET_MODE_INNER (dmode), | |
1034 | nunits).require (); | |
1035 | machine_mode sse_half_dmode = mode_for_vector (GET_MODE_INNER (dmode), | |
1036 | nunits / 2).require (); | |
1037 | ||
1038 | /* Get the corresponding SSE mode for source. */ | |
1039 | nunits = 16 / GET_MODE_SIZE (inner_smode); | |
1040 | machine_mode sse_smode = mode_for_vector (GET_MODE_INNER (smode), | |
1041 | nunits).require (); | |
1042 | ||
1043 | /* Generate SSE pack with signed/unsigned saturation. */ | |
1044 | rtx dest = lowpart_subreg (sse_dmode, op0, GET_MODE (op0)); | |
1045 | op1 = lowpart_subreg (sse_smode, op1, GET_MODE (op1)); | |
1046 | op2 = lowpart_subreg (sse_smode, op2, GET_MODE (op2)); | |
1047 | ||
58e61a3a | 1048 | /* paskusdw/packuswb does unsigned saturation of a signed source |
1049 | which is different from generic us_truncate RTX. */ | |
1050 | if (code == US_TRUNCATE) | |
1051 | src = gen_rtx_UNSPEC (sse_dmode, | |
1052 | gen_rtvec (2, op1, op2), | |
1053 | UNSPEC_US_TRUNCATE); | |
1054 | else | |
1055 | { | |
1056 | op1 = gen_rtx_fmt_e (code, sse_half_dmode, op1); | |
1057 | op2 = gen_rtx_fmt_e (code, sse_half_dmode, op2); | |
1058 | src = gen_rtx_VEC_CONCAT (sse_dmode, op1, op2); | |
1059 | } | |
1060 | ||
1061 | emit_move_insn (dest, src); | |
b74ebb2a L |
1062 | |
1063 | ix86_move_vector_high_sse_to_mmx (op0); | |
1064 | } | |
1065 | ||
6e9fffcf L |
1066 | /* Split MMX punpcklXX/punpckhXX with SSE punpcklXX. */ |
1067 | ||
1068 | void | |
1069 | ix86_split_mmx_punpck (rtx operands[], bool high_p) | |
1070 | { | |
1071 | rtx op0 = operands[0]; | |
1072 | rtx op1 = operands[1]; | |
1073 | rtx op2 = operands[2]; | |
1074 | machine_mode mode = GET_MODE (op0); | |
1075 | rtx mask; | |
1076 | /* The corresponding SSE mode. */ | |
1077 | machine_mode sse_mode, double_sse_mode; | |
1078 | ||
1079 | switch (mode) | |
1080 | { | |
be8749f9 | 1081 | case E_V4QImode: |
6e9fffcf L |
1082 | case E_V8QImode: |
1083 | sse_mode = V16QImode; | |
1084 | double_sse_mode = V32QImode; | |
1085 | mask = gen_rtx_PARALLEL (VOIDmode, | |
1086 | gen_rtvec (16, | |
1087 | GEN_INT (0), GEN_INT (16), | |
1088 | GEN_INT (1), GEN_INT (17), | |
1089 | GEN_INT (2), GEN_INT (18), | |
1090 | GEN_INT (3), GEN_INT (19), | |
1091 | GEN_INT (4), GEN_INT (20), | |
1092 | GEN_INT (5), GEN_INT (21), | |
1093 | GEN_INT (6), GEN_INT (22), | |
1094 | GEN_INT (7), GEN_INT (23))); | |
1095 | break; | |
1096 | ||
1097 | case E_V4HImode: | |
be8749f9 | 1098 | case E_V2HImode: |
6e9fffcf L |
1099 | sse_mode = V8HImode; |
1100 | double_sse_mode = V16HImode; | |
1101 | mask = gen_rtx_PARALLEL (VOIDmode, | |
1102 | gen_rtvec (8, | |
1103 | GEN_INT (0), GEN_INT (8), | |
1104 | GEN_INT (1), GEN_INT (9), | |
1105 | GEN_INT (2), GEN_INT (10), | |
1106 | GEN_INT (3), GEN_INT (11))); | |
1107 | break; | |
1108 | ||
1109 | case E_V2SImode: | |
1110 | sse_mode = V4SImode; | |
1111 | double_sse_mode = V8SImode; | |
1112 | mask = gen_rtx_PARALLEL (VOIDmode, | |
1113 | gen_rtvec (4, | |
1114 | GEN_INT (0), GEN_INT (4), | |
1115 | GEN_INT (1), GEN_INT (5))); | |
1116 | break; | |
1117 | ||
a325bdd1 PB |
1118 | case E_V2SFmode: |
1119 | sse_mode = V4SFmode; | |
1120 | double_sse_mode = V8SFmode; | |
1121 | mask = gen_rtx_PARALLEL (VOIDmode, | |
1122 | gen_rtvec (4, | |
1123 | GEN_INT (0), GEN_INT (4), | |
1124 | GEN_INT (1), GEN_INT (5))); | |
1125 | break; | |
1126 | ||
6e9fffcf L |
1127 | default: |
1128 | gcc_unreachable (); | |
1129 | } | |
1130 | ||
1131 | /* Generate SSE punpcklXX. */ | |
1132 | rtx dest = lowpart_subreg (sse_mode, op0, GET_MODE (op0)); | |
1133 | op1 = lowpart_subreg (sse_mode, op1, GET_MODE (op1)); | |
1134 | op2 = lowpart_subreg (sse_mode, op2, GET_MODE (op2)); | |
1135 | ||
1136 | op1 = gen_rtx_VEC_CONCAT (double_sse_mode, op1, op2); | |
1137 | op2 = gen_rtx_VEC_SELECT (sse_mode, op1, mask); | |
1138 | rtx insn = gen_rtx_SET (dest, op2); | |
1139 | emit_insn (insn); | |
1140 | ||
be8749f9 | 1141 | /* Move high bits to low bits. */ |
6e9fffcf L |
1142 | if (high_p) |
1143 | { | |
a325bdd1 PB |
1144 | if (sse_mode == V4SFmode) |
1145 | { | |
1146 | mask = gen_rtx_PARALLEL (VOIDmode, | |
1147 | gen_rtvec (4, GEN_INT (2), GEN_INT (3), | |
1148 | GEN_INT (4), GEN_INT (5))); | |
1149 | op2 = gen_rtx_VEC_CONCAT (V8SFmode, dest, dest); | |
1150 | op1 = gen_rtx_VEC_SELECT (V4SFmode, op2, mask); | |
1151 | } | |
1152 | else | |
1153 | { | |
be8749f9 UB |
1154 | int sz = GET_MODE_SIZE (mode); |
1155 | ||
1156 | if (sz == 4) | |
1157 | mask = gen_rtx_PARALLEL (VOIDmode, | |
1158 | gen_rtvec (4, GEN_INT (1), GEN_INT (0), | |
1159 | GEN_INT (0), GEN_INT (1))); | |
1160 | else if (sz == 8) | |
1161 | mask = gen_rtx_PARALLEL (VOIDmode, | |
1162 | gen_rtvec (4, GEN_INT (2), GEN_INT (3), | |
1163 | GEN_INT (0), GEN_INT (1))); | |
1164 | else | |
1165 | gcc_unreachable (); | |
1166 | ||
a325bdd1 PB |
1167 | dest = lowpart_subreg (V4SImode, dest, GET_MODE (dest)); |
1168 | op1 = gen_rtx_VEC_SELECT (V4SImode, dest, mask); | |
1169 | } | |
1170 | ||
6e9fffcf L |
1171 | insn = gen_rtx_SET (dest, op1); |
1172 | emit_insn (insn); | |
1173 | } | |
1174 | } | |
1175 | ||
2bf6d935 ML |
1176 | /* Helper function of ix86_fixup_binary_operands to canonicalize |
1177 | operand order. Returns true if the operands should be swapped. */ | |
1178 | ||
1179 | static bool | |
1180 | ix86_swap_binary_operands_p (enum rtx_code code, machine_mode mode, | |
1181 | rtx operands[]) | |
1182 | { | |
1183 | rtx dst = operands[0]; | |
1184 | rtx src1 = operands[1]; | |
1185 | rtx src2 = operands[2]; | |
1186 | ||
1187 | /* If the operation is not commutative, we can't do anything. */ | |
1188 | if (GET_RTX_CLASS (code) != RTX_COMM_ARITH | |
1189 | && GET_RTX_CLASS (code) != RTX_COMM_COMPARE) | |
1190 | return false; | |
1191 | ||
1192 | /* Highest priority is that src1 should match dst. */ | |
1193 | if (rtx_equal_p (dst, src1)) | |
1194 | return false; | |
1195 | if (rtx_equal_p (dst, src2)) | |
1196 | return true; | |
1197 | ||
1198 | /* Next highest priority is that immediate constants come second. */ | |
1199 | if (immediate_operand (src2, mode)) | |
1200 | return false; | |
1201 | if (immediate_operand (src1, mode)) | |
1202 | return true; | |
1203 | ||
1204 | /* Lowest priority is that memory references should come second. */ | |
1205 | if (MEM_P (src2)) | |
1206 | return false; | |
1207 | if (MEM_P (src1)) | |
1208 | return true; | |
1209 | ||
1210 | return false; | |
1211 | } | |
1212 | ||
1213 | ||
1214 | /* Fix up OPERANDS to satisfy ix86_binary_operator_ok. Return the | |
1215 | destination to use for the operation. If different from the true | |
1216 | destination in operands[0], a copy operation will be required. */ | |
1217 | ||
1218 | rtx | |
1219 | ix86_fixup_binary_operands (enum rtx_code code, machine_mode mode, | |
1220 | rtx operands[]) | |
1221 | { | |
1222 | rtx dst = operands[0]; | |
1223 | rtx src1 = operands[1]; | |
1224 | rtx src2 = operands[2]; | |
1225 | ||
1226 | /* Canonicalize operand order. */ | |
1227 | if (ix86_swap_binary_operands_p (code, mode, operands)) | |
1228 | { | |
1229 | /* It is invalid to swap operands of different modes. */ | |
1230 | gcc_assert (GET_MODE (src1) == GET_MODE (src2)); | |
1231 | ||
1232 | std::swap (src1, src2); | |
1233 | } | |
1234 | ||
1235 | /* Both source operands cannot be in memory. */ | |
1236 | if (MEM_P (src1) && MEM_P (src2)) | |
1237 | { | |
1238 | /* Optimization: Only read from memory once. */ | |
1239 | if (rtx_equal_p (src1, src2)) | |
1240 | { | |
1241 | src2 = force_reg (mode, src2); | |
1242 | src1 = src2; | |
1243 | } | |
1244 | else if (rtx_equal_p (dst, src1)) | |
1245 | src2 = force_reg (mode, src2); | |
1246 | else | |
1247 | src1 = force_reg (mode, src1); | |
1248 | } | |
1249 | ||
1250 | /* If the destination is memory, and we do not have matching source | |
1251 | operands, do things in registers. */ | |
1252 | if (MEM_P (dst) && !rtx_equal_p (dst, src1)) | |
1253 | dst = gen_reg_rtx (mode); | |
1254 | ||
1255 | /* Source 1 cannot be a constant. */ | |
1256 | if (CONSTANT_P (src1)) | |
1257 | src1 = force_reg (mode, src1); | |
1258 | ||
1259 | /* Source 1 cannot be a non-matching memory. */ | |
1260 | if (MEM_P (src1) && !rtx_equal_p (dst, src1)) | |
1261 | src1 = force_reg (mode, src1); | |
1262 | ||
1263 | /* Improve address combine. */ | |
1264 | if (code == PLUS | |
1265 | && GET_MODE_CLASS (mode) == MODE_INT | |
1266 | && MEM_P (src2)) | |
1267 | src2 = force_reg (mode, src2); | |
1268 | ||
1269 | operands[1] = src1; | |
1270 | operands[2] = src2; | |
1271 | return dst; | |
1272 | } | |
1273 | ||
1274 | /* Similarly, but assume that the destination has already been | |
1275 | set up properly. */ | |
1276 | ||
1277 | void | |
1278 | ix86_fixup_binary_operands_no_copy (enum rtx_code code, | |
1279 | machine_mode mode, rtx operands[]) | |
1280 | { | |
1281 | rtx dst = ix86_fixup_binary_operands (code, mode, operands); | |
1282 | gcc_assert (dst == operands[0]); | |
1283 | } | |
1284 | ||
1285 | /* Attempt to expand a binary operator. Make the expansion closer to the | |
1286 | actual machine, then just general_operand, which will allow 3 separate | |
1287 | memory references (one output, two input) in a single insn. */ | |
1288 | ||
1289 | void | |
1290 | ix86_expand_binary_operator (enum rtx_code code, machine_mode mode, | |
1291 | rtx operands[]) | |
1292 | { | |
1293 | rtx src1, src2, dst, op, clob; | |
1294 | ||
1295 | dst = ix86_fixup_binary_operands (code, mode, operands); | |
1296 | src1 = operands[1]; | |
1297 | src2 = operands[2]; | |
1298 | ||
1299 | /* Emit the instruction. */ | |
1300 | ||
1301 | op = gen_rtx_SET (dst, gen_rtx_fmt_ee (code, mode, src1, src2)); | |
1302 | ||
1303 | if (reload_completed | |
1304 | && code == PLUS | |
1305 | && !rtx_equal_p (dst, src1)) | |
1306 | { | |
1307 | /* This is going to be an LEA; avoid splitting it later. */ | |
1308 | emit_insn (op); | |
1309 | } | |
1310 | else | |
1311 | { | |
1312 | clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG)); | |
1313 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, op, clob))); | |
1314 | } | |
1315 | ||
1316 | /* Fix up the destination if needed. */ | |
1317 | if (dst != operands[0]) | |
1318 | emit_move_insn (operands[0], dst); | |
1319 | } | |
1320 | ||
1321 | /* Expand vector logical operation CODE (AND, IOR, XOR) in MODE with | |
1322 | the given OPERANDS. */ | |
1323 | ||
1324 | void | |
1325 | ix86_expand_vector_logical_operator (enum rtx_code code, machine_mode mode, | |
1326 | rtx operands[]) | |
1327 | { | |
1328 | rtx op1 = NULL_RTX, op2 = NULL_RTX; | |
1329 | if (SUBREG_P (operands[1])) | |
1330 | { | |
1331 | op1 = operands[1]; | |
1332 | op2 = operands[2]; | |
1333 | } | |
1334 | else if (SUBREG_P (operands[2])) | |
1335 | { | |
1336 | op1 = operands[2]; | |
1337 | op2 = operands[1]; | |
1338 | } | |
1339 | /* Optimize (__m128i) d | (__m128i) e and similar code | |
1340 | when d and e are float vectors into float vector logical | |
1341 | insn. In C/C++ without using intrinsics there is no other way | |
1342 | to express vector logical operation on float vectors than | |
1343 | to cast them temporarily to integer vectors. */ | |
1344 | if (op1 | |
1345 | && !TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL | |
1346 | && (SUBREG_P (op2) || GET_CODE (op2) == CONST_VECTOR) | |
1347 | && GET_MODE_CLASS (GET_MODE (SUBREG_REG (op1))) == MODE_VECTOR_FLOAT | |
1348 | && GET_MODE_SIZE (GET_MODE (SUBREG_REG (op1))) == GET_MODE_SIZE (mode) | |
1349 | && SUBREG_BYTE (op1) == 0 | |
1350 | && (GET_CODE (op2) == CONST_VECTOR | |
1351 | || (GET_MODE (SUBREG_REG (op1)) == GET_MODE (SUBREG_REG (op2)) | |
1352 | && SUBREG_BYTE (op2) == 0)) | |
1353 | && can_create_pseudo_p ()) | |
1354 | { | |
1355 | rtx dst; | |
1356 | switch (GET_MODE (SUBREG_REG (op1))) | |
1357 | { | |
1358 | case E_V4SFmode: | |
1359 | case E_V8SFmode: | |
1360 | case E_V16SFmode: | |
1361 | case E_V2DFmode: | |
1362 | case E_V4DFmode: | |
1363 | case E_V8DFmode: | |
1364 | dst = gen_reg_rtx (GET_MODE (SUBREG_REG (op1))); | |
1365 | if (GET_CODE (op2) == CONST_VECTOR) | |
1366 | { | |
1367 | op2 = gen_lowpart (GET_MODE (dst), op2); | |
1368 | op2 = force_reg (GET_MODE (dst), op2); | |
1369 | } | |
1370 | else | |
1371 | { | |
1372 | op1 = operands[1]; | |
1373 | op2 = SUBREG_REG (operands[2]); | |
1374 | if (!vector_operand (op2, GET_MODE (dst))) | |
1375 | op2 = force_reg (GET_MODE (dst), op2); | |
1376 | } | |
1377 | op1 = SUBREG_REG (op1); | |
1378 | if (!vector_operand (op1, GET_MODE (dst))) | |
1379 | op1 = force_reg (GET_MODE (dst), op1); | |
1380 | emit_insn (gen_rtx_SET (dst, | |
1381 | gen_rtx_fmt_ee (code, GET_MODE (dst), | |
1382 | op1, op2))); | |
1383 | emit_move_insn (operands[0], gen_lowpart (mode, dst)); | |
1384 | return; | |
1385 | default: | |
1386 | break; | |
1387 | } | |
1388 | } | |
1389 | if (!vector_operand (operands[1], mode)) | |
1390 | operands[1] = force_reg (mode, operands[1]); | |
1391 | if (!vector_operand (operands[2], mode)) | |
1392 | operands[2] = force_reg (mode, operands[2]); | |
1393 | ix86_fixup_binary_operands_no_copy (code, mode, operands); | |
1394 | emit_insn (gen_rtx_SET (operands[0], | |
1395 | gen_rtx_fmt_ee (code, mode, operands[1], | |
1396 | operands[2]))); | |
1397 | } | |
1398 | ||
1399 | /* Return TRUE or FALSE depending on whether the binary operator meets the | |
1400 | appropriate constraints. */ | |
1401 | ||
1402 | bool | |
1403 | ix86_binary_operator_ok (enum rtx_code code, machine_mode mode, | |
1404 | rtx operands[3]) | |
1405 | { | |
1406 | rtx dst = operands[0]; | |
1407 | rtx src1 = operands[1]; | |
1408 | rtx src2 = operands[2]; | |
1409 | ||
1410 | /* Both source operands cannot be in memory. */ | |
7026bb95 | 1411 | if ((MEM_P (src1) || bcst_mem_operand (src1, mode)) |
1412 | && (MEM_P (src2) || bcst_mem_operand (src2, mode))) | |
2bf6d935 ML |
1413 | return false; |
1414 | ||
1415 | /* Canonicalize operand order for commutative operators. */ | |
1416 | if (ix86_swap_binary_operands_p (code, mode, operands)) | |
1417 | std::swap (src1, src2); | |
1418 | ||
1419 | /* If the destination is memory, we must have a matching source operand. */ | |
1420 | if (MEM_P (dst) && !rtx_equal_p (dst, src1)) | |
1421 | return false; | |
1422 | ||
1423 | /* Source 1 cannot be a constant. */ | |
1424 | if (CONSTANT_P (src1)) | |
1425 | return false; | |
1426 | ||
1427 | /* Source 1 cannot be a non-matching memory. */ | |
1428 | if (MEM_P (src1) && !rtx_equal_p (dst, src1)) | |
1429 | /* Support "andhi/andsi/anddi" as a zero-extending move. */ | |
1430 | return (code == AND | |
1431 | && (mode == HImode | |
1432 | || mode == SImode | |
1433 | || (TARGET_64BIT && mode == DImode)) | |
1434 | && satisfies_constraint_L (src2)); | |
1435 | ||
1436 | return true; | |
1437 | } | |
1438 | ||
1439 | /* Attempt to expand a unary operator. Make the expansion closer to the | |
1440 | actual machine, then just general_operand, which will allow 2 separate | |
1441 | memory references (one output, one input) in a single insn. */ | |
1442 | ||
1443 | void | |
1444 | ix86_expand_unary_operator (enum rtx_code code, machine_mode mode, | |
1445 | rtx operands[]) | |
1446 | { | |
1447 | bool matching_memory = false; | |
1448 | rtx src, dst, op, clob; | |
1449 | ||
1450 | dst = operands[0]; | |
1451 | src = operands[1]; | |
1452 | ||
1453 | /* If the destination is memory, and we do not have matching source | |
1454 | operands, do things in registers. */ | |
1455 | if (MEM_P (dst)) | |
1456 | { | |
1457 | if (rtx_equal_p (dst, src)) | |
1458 | matching_memory = true; | |
1459 | else | |
1460 | dst = gen_reg_rtx (mode); | |
1461 | } | |
1462 | ||
1463 | /* When source operand is memory, destination must match. */ | |
1464 | if (MEM_P (src) && !matching_memory) | |
1465 | src = force_reg (mode, src); | |
1466 | ||
1467 | /* Emit the instruction. */ | |
1468 | ||
1469 | op = gen_rtx_SET (dst, gen_rtx_fmt_e (code, mode, src)); | |
1470 | ||
1471 | if (code == NOT) | |
1472 | emit_insn (op); | |
1473 | else | |
1474 | { | |
1475 | clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG)); | |
1476 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, op, clob))); | |
1477 | } | |
1478 | ||
1479 | /* Fix up the destination if needed. */ | |
1480 | if (dst != operands[0]) | |
1481 | emit_move_insn (operands[0], dst); | |
1482 | } | |
1483 | ||
1484 | /* Predict just emitted jump instruction to be taken with probability PROB. */ | |
1485 | ||
1486 | static void | |
1487 | predict_jump (int prob) | |
1488 | { | |
1489 | rtx_insn *insn = get_last_insn (); | |
1490 | gcc_assert (JUMP_P (insn)); | |
1491 | add_reg_br_prob_note (insn, profile_probability::from_reg_br_prob_base (prob)); | |
1492 | } | |
1493 | ||
1494 | /* Split 32bit/64bit divmod with 8bit unsigned divmod if dividend and | |
1495 | divisor are within the range [0-255]. */ | |
1496 | ||
1497 | void | |
1498 | ix86_split_idivmod (machine_mode mode, rtx operands[], | |
40c81f84 | 1499 | bool unsigned_p) |
2bf6d935 ML |
1500 | { |
1501 | rtx_code_label *end_label, *qimode_label; | |
1502 | rtx div, mod; | |
1503 | rtx_insn *insn; | |
1504 | rtx scratch, tmp0, tmp1, tmp2; | |
1505 | rtx (*gen_divmod4_1) (rtx, rtx, rtx, rtx); | |
2bf6d935 | 1506 | |
2b399dba UB |
1507 | operands[2] = force_reg (mode, operands[2]); |
1508 | operands[3] = force_reg (mode, operands[3]); | |
1509 | ||
2bf6d935 ML |
1510 | switch (mode) |
1511 | { | |
1512 | case E_SImode: | |
1513 | if (GET_MODE (operands[0]) == SImode) | |
1514 | { | |
1515 | if (GET_MODE (operands[1]) == SImode) | |
40c81f84 | 1516 | gen_divmod4_1 = unsigned_p ? gen_udivmodsi4_1 : gen_divmodsi4_1; |
2bf6d935 ML |
1517 | else |
1518 | gen_divmod4_1 | |
40c81f84 | 1519 | = unsigned_p ? gen_udivmodsi4_zext_2 : gen_divmodsi4_zext_2; |
2bf6d935 ML |
1520 | } |
1521 | else | |
ea298f7a UB |
1522 | gen_divmod4_1 |
1523 | = unsigned_p ? gen_udivmodsi4_zext_1 : gen_divmodsi4_zext_1; | |
2bf6d935 | 1524 | break; |
ea298f7a | 1525 | |
2bf6d935 | 1526 | case E_DImode: |
40c81f84 | 1527 | gen_divmod4_1 = unsigned_p ? gen_udivmoddi4_1 : gen_divmoddi4_1; |
2bf6d935 | 1528 | break; |
ea298f7a | 1529 | |
2bf6d935 ML |
1530 | default: |
1531 | gcc_unreachable (); | |
1532 | } | |
1533 | ||
1534 | end_label = gen_label_rtx (); | |
1535 | qimode_label = gen_label_rtx (); | |
1536 | ||
1537 | scratch = gen_reg_rtx (mode); | |
1538 | ||
1539 | /* Use 8bit unsigned divimod if dividend and divisor are within | |
1540 | the range [0-255]. */ | |
1541 | emit_move_insn (scratch, operands[2]); | |
1542 | scratch = expand_simple_binop (mode, IOR, scratch, operands[3], | |
1543 | scratch, 1, OPTAB_DIRECT); | |
ea298f7a | 1544 | emit_insn (gen_test_ccno_1 (mode, scratch, GEN_INT (-0x100))); |
2bf6d935 ML |
1545 | tmp0 = gen_rtx_REG (CCNOmode, FLAGS_REG); |
1546 | tmp0 = gen_rtx_EQ (VOIDmode, tmp0, const0_rtx); | |
1547 | tmp0 = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp0, | |
1548 | gen_rtx_LABEL_REF (VOIDmode, qimode_label), | |
1549 | pc_rtx); | |
1550 | insn = emit_jump_insn (gen_rtx_SET (pc_rtx, tmp0)); | |
1551 | predict_jump (REG_BR_PROB_BASE * 50 / 100); | |
1552 | JUMP_LABEL (insn) = qimode_label; | |
1553 | ||
1554 | /* Generate original signed/unsigned divimod. */ | |
e9539592 UB |
1555 | emit_insn (gen_divmod4_1 (operands[0], operands[1], |
1556 | operands[2], operands[3])); | |
2bf6d935 ML |
1557 | |
1558 | /* Branch to the end. */ | |
1559 | emit_jump_insn (gen_jump (end_label)); | |
1560 | emit_barrier (); | |
1561 | ||
1562 | /* Generate 8bit unsigned divide. */ | |
1563 | emit_label (qimode_label); | |
1564 | /* Don't use operands[0] for result of 8bit divide since not all | |
1565 | registers support QImode ZERO_EXTRACT. */ | |
1566 | tmp0 = lowpart_subreg (HImode, scratch, mode); | |
1567 | tmp1 = lowpart_subreg (HImode, operands[2], mode); | |
1568 | tmp2 = lowpart_subreg (QImode, operands[3], mode); | |
1569 | emit_insn (gen_udivmodhiqi3 (tmp0, tmp1, tmp2)); | |
1570 | ||
40c81f84 | 1571 | if (unsigned_p) |
2bf6d935 | 1572 | { |
40c81f84 UB |
1573 | div = gen_rtx_UDIV (mode, operands[2], operands[3]); |
1574 | mod = gen_rtx_UMOD (mode, operands[2], operands[3]); | |
2bf6d935 ML |
1575 | } |
1576 | else | |
1577 | { | |
40c81f84 UB |
1578 | div = gen_rtx_DIV (mode, operands[2], operands[3]); |
1579 | mod = gen_rtx_MOD (mode, operands[2], operands[3]); | |
2bf6d935 ML |
1580 | } |
1581 | if (mode == SImode) | |
1582 | { | |
1583 | if (GET_MODE (operands[0]) != SImode) | |
1584 | div = gen_rtx_ZERO_EXTEND (DImode, div); | |
1585 | if (GET_MODE (operands[1]) != SImode) | |
1586 | mod = gen_rtx_ZERO_EXTEND (DImode, mod); | |
1587 | } | |
1588 | ||
1589 | /* Extract remainder from AH. */ | |
e9539592 UB |
1590 | scratch = gen_lowpart (GET_MODE (operands[1]), scratch); |
1591 | tmp1 = gen_rtx_ZERO_EXTRACT (GET_MODE (operands[1]), scratch, | |
1592 | GEN_INT (8), GEN_INT (8)); | |
1593 | insn = emit_move_insn (operands[1], tmp1); | |
2bf6d935 ML |
1594 | set_unique_reg_note (insn, REG_EQUAL, mod); |
1595 | ||
1596 | /* Zero extend quotient from AL. */ | |
1597 | tmp1 = gen_lowpart (QImode, tmp0); | |
ea298f7a UB |
1598 | insn = emit_insn (gen_extend_insn |
1599 | (operands[0], tmp1, | |
1600 | GET_MODE (operands[0]), QImode, 1)); | |
2bf6d935 ML |
1601 | set_unique_reg_note (insn, REG_EQUAL, div); |
1602 | ||
1603 | emit_label (end_label); | |
1604 | } | |
1605 | ||
1606 | /* Emit x86 binary operand CODE in mode MODE, where the first operand | |
1607 | matches destination. RTX includes clobber of FLAGS_REG. */ | |
1608 | ||
1609 | void | |
1610 | ix86_emit_binop (enum rtx_code code, machine_mode mode, | |
1611 | rtx dst, rtx src) | |
1612 | { | |
1613 | rtx op, clob; | |
1614 | ||
1615 | op = gen_rtx_SET (dst, gen_rtx_fmt_ee (code, mode, dst, src)); | |
1616 | clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG)); | |
1617 | ||
1618 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, op, clob))); | |
1619 | } | |
1620 | ||
1621 | /* Return true if regno1 def is nearest to the insn. */ | |
1622 | ||
1623 | static bool | |
1624 | find_nearest_reg_def (rtx_insn *insn, int regno1, int regno2) | |
1625 | { | |
1626 | rtx_insn *prev = insn; | |
1627 | rtx_insn *start = BB_HEAD (BLOCK_FOR_INSN (insn)); | |
1628 | ||
1629 | if (insn == start) | |
1630 | return false; | |
1631 | while (prev && prev != start) | |
1632 | { | |
1633 | if (!INSN_P (prev) || !NONDEBUG_INSN_P (prev)) | |
1634 | { | |
1635 | prev = PREV_INSN (prev); | |
1636 | continue; | |
1637 | } | |
1638 | if (insn_defines_reg (regno1, INVALID_REGNUM, prev)) | |
1639 | return true; | |
1640 | else if (insn_defines_reg (regno2, INVALID_REGNUM, prev)) | |
1641 | return false; | |
1642 | prev = PREV_INSN (prev); | |
1643 | } | |
1644 | ||
1645 | /* None of the regs is defined in the bb. */ | |
1646 | return false; | |
1647 | } | |
1648 | ||
d58a66aa JJ |
1649 | /* INSN_UID of the last insn emitted by zero store peephole2s. */ |
1650 | int ix86_last_zero_store_uid; | |
1651 | ||
2bf6d935 ML |
1652 | /* Split lea instructions into a sequence of instructions |
1653 | which are executed on ALU to avoid AGU stalls. | |
1654 | It is assumed that it is allowed to clobber flags register | |
1655 | at lea position. */ | |
1656 | ||
1657 | void | |
1658 | ix86_split_lea_for_addr (rtx_insn *insn, rtx operands[], machine_mode mode) | |
1659 | { | |
1660 | unsigned int regno0, regno1, regno2; | |
1661 | struct ix86_address parts; | |
1662 | rtx target, tmp; | |
1663 | int ok, adds; | |
1664 | ||
1665 | ok = ix86_decompose_address (operands[1], &parts); | |
1666 | gcc_assert (ok); | |
1667 | ||
1668 | target = gen_lowpart (mode, operands[0]); | |
1669 | ||
1670 | regno0 = true_regnum (target); | |
1671 | regno1 = INVALID_REGNUM; | |
1672 | regno2 = INVALID_REGNUM; | |
1673 | ||
1674 | if (parts.base) | |
1675 | { | |
1676 | parts.base = gen_lowpart (mode, parts.base); | |
1677 | regno1 = true_regnum (parts.base); | |
1678 | } | |
1679 | ||
1680 | if (parts.index) | |
1681 | { | |
1682 | parts.index = gen_lowpart (mode, parts.index); | |
1683 | regno2 = true_regnum (parts.index); | |
1684 | } | |
1685 | ||
1686 | if (parts.disp) | |
1687 | parts.disp = gen_lowpart (mode, parts.disp); | |
1688 | ||
1689 | if (parts.scale > 1) | |
1690 | { | |
1691 | /* Case r1 = r1 + ... */ | |
1692 | if (regno1 == regno0) | |
1693 | { | |
1694 | /* If we have a case r1 = r1 + C * r2 then we | |
1695 | should use multiplication which is very | |
1696 | expensive. Assume cost model is wrong if we | |
1697 | have such case here. */ | |
1698 | gcc_assert (regno2 != regno0); | |
1699 | ||
1700 | for (adds = parts.scale; adds > 0; adds--) | |
1701 | ix86_emit_binop (PLUS, mode, target, parts.index); | |
1702 | } | |
1703 | else | |
1704 | { | |
1705 | /* r1 = r2 + r3 * C case. Need to move r3 into r1. */ | |
1706 | if (regno0 != regno2) | |
1707 | emit_insn (gen_rtx_SET (target, parts.index)); | |
1708 | ||
d55ce33a JJ |
1709 | /* Use shift for scaling, but emit it as MULT instead |
1710 | to avoid it being immediately peephole2 optimized back | |
1711 | into lea. */ | |
1712 | ix86_emit_binop (MULT, mode, target, GEN_INT (parts.scale)); | |
2bf6d935 ML |
1713 | |
1714 | if (parts.base) | |
1715 | ix86_emit_binop (PLUS, mode, target, parts.base); | |
1716 | ||
1717 | if (parts.disp && parts.disp != const0_rtx) | |
1718 | ix86_emit_binop (PLUS, mode, target, parts.disp); | |
1719 | } | |
1720 | } | |
1721 | else if (!parts.base && !parts.index) | |
1722 | { | |
1723 | gcc_assert(parts.disp); | |
1724 | emit_insn (gen_rtx_SET (target, parts.disp)); | |
1725 | } | |
1726 | else | |
1727 | { | |
1728 | if (!parts.base) | |
1729 | { | |
1730 | if (regno0 != regno2) | |
1731 | emit_insn (gen_rtx_SET (target, parts.index)); | |
1732 | } | |
1733 | else if (!parts.index) | |
1734 | { | |
1735 | if (regno0 != regno1) | |
1736 | emit_insn (gen_rtx_SET (target, parts.base)); | |
1737 | } | |
1738 | else | |
1739 | { | |
1740 | if (regno0 == regno1) | |
1741 | tmp = parts.index; | |
1742 | else if (regno0 == regno2) | |
1743 | tmp = parts.base; | |
1744 | else | |
1745 | { | |
1746 | rtx tmp1; | |
1747 | ||
1748 | /* Find better operand for SET instruction, depending | |
1749 | on which definition is farther from the insn. */ | |
1750 | if (find_nearest_reg_def (insn, regno1, regno2)) | |
1751 | tmp = parts.index, tmp1 = parts.base; | |
1752 | else | |
1753 | tmp = parts.base, tmp1 = parts.index; | |
1754 | ||
1755 | emit_insn (gen_rtx_SET (target, tmp)); | |
1756 | ||
1757 | if (parts.disp && parts.disp != const0_rtx) | |
1758 | ix86_emit_binop (PLUS, mode, target, parts.disp); | |
1759 | ||
1760 | ix86_emit_binop (PLUS, mode, target, tmp1); | |
1761 | return; | |
1762 | } | |
1763 | ||
1764 | ix86_emit_binop (PLUS, mode, target, tmp); | |
1765 | } | |
1766 | ||
1767 | if (parts.disp && parts.disp != const0_rtx) | |
1768 | ix86_emit_binop (PLUS, mode, target, parts.disp); | |
1769 | } | |
1770 | } | |
1771 | ||
1772 | /* Post-reload splitter for converting an SF or DFmode value in an | |
1773 | SSE register into an unsigned SImode. */ | |
1774 | ||
1775 | void | |
1776 | ix86_split_convert_uns_si_sse (rtx operands[]) | |
1777 | { | |
1778 | machine_mode vecmode; | |
1779 | rtx value, large, zero_or_two31, input, two31, x; | |
1780 | ||
1781 | large = operands[1]; | |
1782 | zero_or_two31 = operands[2]; | |
1783 | input = operands[3]; | |
1784 | two31 = operands[4]; | |
1785 | vecmode = GET_MODE (large); | |
1786 | value = gen_rtx_REG (vecmode, REGNO (operands[0])); | |
1787 | ||
1788 | /* Load up the value into the low element. We must ensure that the other | |
1789 | elements are valid floats -- zero is the easiest such value. */ | |
1790 | if (MEM_P (input)) | |
1791 | { | |
1792 | if (vecmode == V4SFmode) | |
1793 | emit_insn (gen_vec_setv4sf_0 (value, CONST0_RTX (V4SFmode), input)); | |
1794 | else | |
1795 | emit_insn (gen_sse2_loadlpd (value, CONST0_RTX (V2DFmode), input)); | |
1796 | } | |
1797 | else | |
1798 | { | |
1799 | input = gen_rtx_REG (vecmode, REGNO (input)); | |
1800 | emit_move_insn (value, CONST0_RTX (vecmode)); | |
1801 | if (vecmode == V4SFmode) | |
febb58d2 | 1802 | emit_insn (gen_sse_movss_v4sf (value, value, input)); |
2bf6d935 | 1803 | else |
febb58d2 | 1804 | emit_insn (gen_sse2_movsd_v2df (value, value, input)); |
2bf6d935 ML |
1805 | } |
1806 | ||
1807 | emit_move_insn (large, two31); | |
1808 | emit_move_insn (zero_or_two31, MEM_P (two31) ? large : two31); | |
1809 | ||
1810 | x = gen_rtx_fmt_ee (LE, vecmode, large, value); | |
1811 | emit_insn (gen_rtx_SET (large, x)); | |
1812 | ||
1813 | x = gen_rtx_AND (vecmode, zero_or_two31, large); | |
1814 | emit_insn (gen_rtx_SET (zero_or_two31, x)); | |
1815 | ||
1816 | x = gen_rtx_MINUS (vecmode, value, zero_or_two31); | |
1817 | emit_insn (gen_rtx_SET (value, x)); | |
1818 | ||
1819 | large = gen_rtx_REG (V4SImode, REGNO (large)); | |
1820 | emit_insn (gen_ashlv4si3 (large, large, GEN_INT (31))); | |
1821 | ||
1822 | x = gen_rtx_REG (V4SImode, REGNO (value)); | |
1823 | if (vecmode == V4SFmode) | |
1824 | emit_insn (gen_fix_truncv4sfv4si2 (x, value)); | |
1825 | else | |
1826 | emit_insn (gen_sse2_cvttpd2dq (x, value)); | |
1827 | value = x; | |
1828 | ||
1829 | emit_insn (gen_xorv4si3 (value, value, large)); | |
1830 | } | |
1831 | ||
1832 | static bool ix86_expand_vector_init_one_nonzero (bool mmx_ok, | |
1833 | machine_mode mode, rtx target, | |
1834 | rtx var, int one_var); | |
1835 | ||
1836 | /* Convert an unsigned DImode value into a DFmode, using only SSE. | |
1837 | Expects the 64-bit DImode to be supplied in a pair of integral | |
1838 | registers. Requires SSE2; will use SSE3 if available. For x86_32, | |
1839 | -mfpmath=sse, !optimize_size only. */ | |
1840 | ||
1841 | void | |
1842 | ix86_expand_convert_uns_didf_sse (rtx target, rtx input) | |
1843 | { | |
1844 | REAL_VALUE_TYPE bias_lo_rvt, bias_hi_rvt; | |
1845 | rtx int_xmm, fp_xmm; | |
1846 | rtx biases, exponents; | |
1847 | rtx x; | |
1848 | ||
1849 | int_xmm = gen_reg_rtx (V4SImode); | |
1850 | if (TARGET_INTER_UNIT_MOVES_TO_VEC) | |
1851 | emit_insn (gen_movdi_to_sse (int_xmm, input)); | |
1852 | else if (TARGET_SSE_SPLIT_REGS) | |
1853 | { | |
1854 | emit_clobber (int_xmm); | |
1855 | emit_move_insn (gen_lowpart (DImode, int_xmm), input); | |
1856 | } | |
1857 | else | |
1858 | { | |
1859 | x = gen_reg_rtx (V2DImode); | |
1860 | ix86_expand_vector_init_one_nonzero (false, V2DImode, x, input, 0); | |
1861 | emit_move_insn (int_xmm, gen_lowpart (V4SImode, x)); | |
1862 | } | |
1863 | ||
1864 | x = gen_rtx_CONST_VECTOR (V4SImode, | |
1865 | gen_rtvec (4, GEN_INT (0x43300000UL), | |
1866 | GEN_INT (0x45300000UL), | |
1867 | const0_rtx, const0_rtx)); | |
1868 | exponents = validize_mem (force_const_mem (V4SImode, x)); | |
1869 | ||
1870 | /* int_xmm = {0x45300000UL, fp_xmm/hi, 0x43300000, fp_xmm/lo } */ | |
1871 | emit_insn (gen_vec_interleave_lowv4si (int_xmm, int_xmm, exponents)); | |
1872 | ||
1873 | /* Concatenating (juxtaposing) (0x43300000UL ## fp_value_low_xmm) | |
1874 | yields a valid DF value equal to (0x1.0p52 + double(fp_value_lo_xmm)). | |
1875 | Similarly (0x45300000UL ## fp_value_hi_xmm) yields | |
1876 | (0x1.0p84 + double(fp_value_hi_xmm)). | |
1877 | Note these exponents differ by 32. */ | |
1878 | ||
1879 | fp_xmm = copy_to_mode_reg (V2DFmode, gen_lowpart (V2DFmode, int_xmm)); | |
1880 | ||
1881 | /* Subtract off those 0x1.0p52 and 0x1.0p84 biases, to produce values | |
1882 | in [0,2**32-1] and [0]+[2**32,2**64-1] respectively. */ | |
1883 | real_ldexp (&bias_lo_rvt, &dconst1, 52); | |
1884 | real_ldexp (&bias_hi_rvt, &dconst1, 84); | |
1885 | biases = const_double_from_real_value (bias_lo_rvt, DFmode); | |
1886 | x = const_double_from_real_value (bias_hi_rvt, DFmode); | |
1887 | biases = gen_rtx_CONST_VECTOR (V2DFmode, gen_rtvec (2, biases, x)); | |
1888 | biases = validize_mem (force_const_mem (V2DFmode, biases)); | |
1889 | emit_insn (gen_subv2df3 (fp_xmm, fp_xmm, biases)); | |
1890 | ||
1891 | /* Add the upper and lower DFmode values together. */ | |
1892 | if (TARGET_SSE3) | |
1893 | emit_insn (gen_sse3_haddv2df3 (fp_xmm, fp_xmm, fp_xmm)); | |
1894 | else | |
1895 | { | |
1896 | x = copy_to_mode_reg (V2DFmode, fp_xmm); | |
1897 | emit_insn (gen_vec_interleave_highv2df (fp_xmm, fp_xmm, fp_xmm)); | |
1898 | emit_insn (gen_addv2df3 (fp_xmm, fp_xmm, x)); | |
1899 | } | |
1900 | ||
1901 | ix86_expand_vector_extract (false, target, fp_xmm, 0); | |
1902 | } | |
1903 | ||
1904 | /* Not used, but eases macroization of patterns. */ | |
1905 | void | |
1906 | ix86_expand_convert_uns_sixf_sse (rtx, rtx) | |
1907 | { | |
1908 | gcc_unreachable (); | |
1909 | } | |
1910 | ||
0cda606d UB |
1911 | static rtx ix86_expand_sse_fabs (rtx op0, rtx *smask); |
1912 | ||
2bf6d935 ML |
1913 | /* Convert an unsigned SImode value into a DFmode. Only currently used |
1914 | for SSE, but applicable anywhere. */ | |
1915 | ||
1916 | void | |
1917 | ix86_expand_convert_uns_sidf_sse (rtx target, rtx input) | |
1918 | { | |
1919 | REAL_VALUE_TYPE TWO31r; | |
1920 | rtx x, fp; | |
1921 | ||
1922 | x = expand_simple_binop (SImode, PLUS, input, GEN_INT (-2147483647 - 1), | |
1923 | NULL, 1, OPTAB_DIRECT); | |
1924 | ||
1925 | fp = gen_reg_rtx (DFmode); | |
1926 | emit_insn (gen_floatsidf2 (fp, x)); | |
1927 | ||
1928 | real_ldexp (&TWO31r, &dconst1, 31); | |
1929 | x = const_double_from_real_value (TWO31r, DFmode); | |
1930 | ||
1931 | x = expand_simple_binop (DFmode, PLUS, fp, x, target, 0, OPTAB_DIRECT); | |
0cda606d UB |
1932 | |
1933 | /* Remove the sign with FE_DOWNWARD, where x - x = -0.0. */ | |
1934 | if (HONOR_SIGNED_ZEROS (DFmode) && flag_rounding_math) | |
1935 | x = ix86_expand_sse_fabs (x, NULL); | |
1936 | ||
2bf6d935 ML |
1937 | if (x != target) |
1938 | emit_move_insn (target, x); | |
1939 | } | |
1940 | ||
1941 | /* Convert a signed DImode value into a DFmode. Only used for SSE in | |
1942 | 32-bit mode; otherwise we have a direct convert instruction. */ | |
1943 | ||
1944 | void | |
1945 | ix86_expand_convert_sign_didf_sse (rtx target, rtx input) | |
1946 | { | |
1947 | REAL_VALUE_TYPE TWO32r; | |
1948 | rtx fp_lo, fp_hi, x; | |
1949 | ||
1950 | fp_lo = gen_reg_rtx (DFmode); | |
1951 | fp_hi = gen_reg_rtx (DFmode); | |
1952 | ||
1953 | emit_insn (gen_floatsidf2 (fp_hi, gen_highpart (SImode, input))); | |
1954 | ||
1955 | real_ldexp (&TWO32r, &dconst1, 32); | |
1956 | x = const_double_from_real_value (TWO32r, DFmode); | |
1957 | fp_hi = expand_simple_binop (DFmode, MULT, fp_hi, x, fp_hi, 0, OPTAB_DIRECT); | |
1958 | ||
1959 | ix86_expand_convert_uns_sidf_sse (fp_lo, gen_lowpart (SImode, input)); | |
1960 | ||
1961 | x = expand_simple_binop (DFmode, PLUS, fp_hi, fp_lo, target, | |
1962 | 0, OPTAB_DIRECT); | |
1963 | if (x != target) | |
1964 | emit_move_insn (target, x); | |
1965 | } | |
1966 | ||
1967 | /* Convert an unsigned SImode value into a SFmode, using only SSE. | |
1968 | For x86_32, -mfpmath=sse, !optimize_size only. */ | |
1969 | void | |
1970 | ix86_expand_convert_uns_sisf_sse (rtx target, rtx input) | |
1971 | { | |
1972 | REAL_VALUE_TYPE ONE16r; | |
1973 | rtx fp_hi, fp_lo, int_hi, int_lo, x; | |
1974 | ||
1975 | real_ldexp (&ONE16r, &dconst1, 16); | |
1976 | x = const_double_from_real_value (ONE16r, SFmode); | |
1977 | int_lo = expand_simple_binop (SImode, AND, input, GEN_INT(0xffff), | |
1978 | NULL, 0, OPTAB_DIRECT); | |
1979 | int_hi = expand_simple_binop (SImode, LSHIFTRT, input, GEN_INT(16), | |
1980 | NULL, 0, OPTAB_DIRECT); | |
1981 | fp_hi = gen_reg_rtx (SFmode); | |
1982 | fp_lo = gen_reg_rtx (SFmode); | |
1983 | emit_insn (gen_floatsisf2 (fp_hi, int_hi)); | |
1984 | emit_insn (gen_floatsisf2 (fp_lo, int_lo)); | |
ad9fcb96 L |
1985 | if (TARGET_FMA) |
1986 | { | |
1987 | x = validize_mem (force_const_mem (SFmode, x)); | |
1988 | fp_hi = gen_rtx_FMA (SFmode, fp_hi, x, fp_lo); | |
1989 | emit_move_insn (target, fp_hi); | |
1990 | } | |
1991 | else | |
1992 | { | |
1993 | fp_hi = expand_simple_binop (SFmode, MULT, fp_hi, x, fp_hi, | |
1994 | 0, OPTAB_DIRECT); | |
1995 | fp_hi = expand_simple_binop (SFmode, PLUS, fp_hi, fp_lo, target, | |
1996 | 0, OPTAB_DIRECT); | |
1997 | if (!rtx_equal_p (target, fp_hi)) | |
1998 | emit_move_insn (target, fp_hi); | |
1999 | } | |
2bf6d935 ML |
2000 | } |
2001 | ||
2002 | /* floatunsv{4,8}siv{4,8}sf2 expander. Expand code to convert | |
2003 | a vector of unsigned ints VAL to vector of floats TARGET. */ | |
2004 | ||
2005 | void | |
2006 | ix86_expand_vector_convert_uns_vsivsf (rtx target, rtx val) | |
2007 | { | |
2008 | rtx tmp[8]; | |
2009 | REAL_VALUE_TYPE TWO16r; | |
2010 | machine_mode intmode = GET_MODE (val); | |
2011 | machine_mode fltmode = GET_MODE (target); | |
2012 | rtx (*cvt) (rtx, rtx); | |
2013 | ||
2014 | if (intmode == V4SImode) | |
2015 | cvt = gen_floatv4siv4sf2; | |
2016 | else | |
2017 | cvt = gen_floatv8siv8sf2; | |
2018 | tmp[0] = ix86_build_const_vector (intmode, 1, GEN_INT (0xffff)); | |
2019 | tmp[0] = force_reg (intmode, tmp[0]); | |
2020 | tmp[1] = expand_simple_binop (intmode, AND, val, tmp[0], NULL_RTX, 1, | |
2021 | OPTAB_DIRECT); | |
2022 | tmp[2] = expand_simple_binop (intmode, LSHIFTRT, val, GEN_INT (16), | |
2023 | NULL_RTX, 1, OPTAB_DIRECT); | |
2024 | tmp[3] = gen_reg_rtx (fltmode); | |
2025 | emit_insn (cvt (tmp[3], tmp[1])); | |
2026 | tmp[4] = gen_reg_rtx (fltmode); | |
2027 | emit_insn (cvt (tmp[4], tmp[2])); | |
2028 | real_ldexp (&TWO16r, &dconst1, 16); | |
2029 | tmp[5] = const_double_from_real_value (TWO16r, SFmode); | |
2030 | tmp[5] = force_reg (fltmode, ix86_build_const_vector (fltmode, 1, tmp[5])); | |
ad9fcb96 L |
2031 | if (TARGET_FMA) |
2032 | { | |
2033 | tmp[6] = gen_rtx_FMA (fltmode, tmp[4], tmp[5], tmp[3]); | |
2034 | emit_move_insn (target, tmp[6]); | |
2035 | } | |
2036 | else | |
2037 | { | |
2038 | tmp[6] = expand_simple_binop (fltmode, MULT, tmp[4], tmp[5], | |
2039 | NULL_RTX, 1, OPTAB_DIRECT); | |
2040 | tmp[7] = expand_simple_binop (fltmode, PLUS, tmp[3], tmp[6], | |
2041 | target, 1, OPTAB_DIRECT); | |
2042 | if (tmp[7] != target) | |
2043 | emit_move_insn (target, tmp[7]); | |
2044 | } | |
2bf6d935 ML |
2045 | } |
2046 | ||
2047 | /* Adjust a V*SFmode/V*DFmode value VAL so that *sfix_trunc* resp. fix_trunc* | |
fe42e7fe | 2048 | pattern can be used on it instead of fixuns_trunc*. |
2bf6d935 ML |
2049 | This is done by doing just signed conversion if < 0x1p31, and otherwise by |
2050 | subtracting 0x1p31 first and xoring in 0x80000000 from *XORP afterwards. */ | |
2051 | ||
2052 | rtx | |
2053 | ix86_expand_adjust_ufix_to_sfix_si (rtx val, rtx *xorp) | |
2054 | { | |
2055 | REAL_VALUE_TYPE TWO31r; | |
2056 | rtx two31r, tmp[4]; | |
2057 | machine_mode mode = GET_MODE (val); | |
2058 | machine_mode scalarmode = GET_MODE_INNER (mode); | |
2059 | machine_mode intmode = GET_MODE_SIZE (mode) == 32 ? V8SImode : V4SImode; | |
2060 | rtx (*cmp) (rtx, rtx, rtx, rtx); | |
2061 | int i; | |
2062 | ||
2063 | for (i = 0; i < 3; i++) | |
2064 | tmp[i] = gen_reg_rtx (mode); | |
2065 | real_ldexp (&TWO31r, &dconst1, 31); | |
2066 | two31r = const_double_from_real_value (TWO31r, scalarmode); | |
2067 | two31r = ix86_build_const_vector (mode, 1, two31r); | |
2068 | two31r = force_reg (mode, two31r); | |
2069 | switch (mode) | |
2070 | { | |
2071 | case E_V8SFmode: cmp = gen_avx_maskcmpv8sf3; break; | |
2072 | case E_V4SFmode: cmp = gen_sse_maskcmpv4sf3; break; | |
2073 | case E_V4DFmode: cmp = gen_avx_maskcmpv4df3; break; | |
2074 | case E_V2DFmode: cmp = gen_sse2_maskcmpv2df3; break; | |
2075 | default: gcc_unreachable (); | |
2076 | } | |
2077 | tmp[3] = gen_rtx_LE (mode, two31r, val); | |
2078 | emit_insn (cmp (tmp[0], two31r, val, tmp[3])); | |
2079 | tmp[1] = expand_simple_binop (mode, AND, tmp[0], two31r, tmp[1], | |
2080 | 0, OPTAB_DIRECT); | |
2081 | if (intmode == V4SImode || TARGET_AVX2) | |
2082 | *xorp = expand_simple_binop (intmode, ASHIFT, | |
2083 | gen_lowpart (intmode, tmp[0]), | |
2084 | GEN_INT (31), NULL_RTX, 0, | |
2085 | OPTAB_DIRECT); | |
2086 | else | |
2087 | { | |
6a556ba4 | 2088 | rtx two31 = gen_int_mode (HOST_WIDE_INT_1U << 31, SImode); |
2bf6d935 ML |
2089 | two31 = ix86_build_const_vector (intmode, 1, two31); |
2090 | *xorp = expand_simple_binop (intmode, AND, | |
2091 | gen_lowpart (intmode, tmp[0]), | |
2092 | two31, NULL_RTX, 0, | |
2093 | OPTAB_DIRECT); | |
2094 | } | |
2095 | return expand_simple_binop (mode, MINUS, val, tmp[1], tmp[2], | |
2096 | 0, OPTAB_DIRECT); | |
2097 | } | |
2098 | ||
2099 | /* Generate code for floating point ABS or NEG. */ | |
2100 | ||
2101 | void | |
2102 | ix86_expand_fp_absneg_operator (enum rtx_code code, machine_mode mode, | |
2103 | rtx operands[]) | |
2104 | { | |
f359611b | 2105 | rtx set, dst, src; |
2bf6d935 ML |
2106 | bool use_sse = false; |
2107 | bool vector_mode = VECTOR_MODE_P (mode); | |
2108 | machine_mode vmode = mode; | |
f359611b | 2109 | rtvec par; |
2bf6d935 | 2110 | |
75a97b59 L |
2111 | if (vector_mode || mode == TFmode || mode == HFmode) |
2112 | { | |
2113 | use_sse = true; | |
2114 | if (mode == HFmode) | |
2115 | vmode = V8HFmode; | |
2116 | } | |
2bf6d935 ML |
2117 | else if (TARGET_SSE_MATH) |
2118 | { | |
2119 | use_sse = SSE_FLOAT_MODE_P (mode); | |
2120 | if (mode == SFmode) | |
2121 | vmode = V4SFmode; | |
2122 | else if (mode == DFmode) | |
2123 | vmode = V2DFmode; | |
2124 | } | |
2125 | ||
2bf6d935 ML |
2126 | dst = operands[0]; |
2127 | src = operands[1]; | |
2128 | ||
2129 | set = gen_rtx_fmt_e (code, mode, src); | |
2130 | set = gen_rtx_SET (dst, set); | |
2131 | ||
f359611b | 2132 | if (use_sse) |
2bf6d935 | 2133 | { |
f359611b | 2134 | rtx mask, use, clob; |
2bf6d935 | 2135 | |
f359611b UB |
2136 | /* NEG and ABS performed with SSE use bitwise mask operations. |
2137 | Create the appropriate mask now. */ | |
2138 | mask = ix86_build_signbit_mask (vmode, vector_mode, code == ABS); | |
2bf6d935 | 2139 | use = gen_rtx_USE (VOIDmode, mask); |
94f687bd | 2140 | if (vector_mode || mode == TFmode) |
2bf6d935 ML |
2141 | par = gen_rtvec (2, set, use); |
2142 | else | |
2143 | { | |
2144 | clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG)); | |
2145 | par = gen_rtvec (3, set, use, clob); | |
2146 | } | |
2bf6d935 ML |
2147 | } |
2148 | else | |
f359611b UB |
2149 | { |
2150 | rtx clob; | |
2151 | ||
2152 | /* Changing of sign for FP values is doable using integer unit too. */ | |
2153 | clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG)); | |
2154 | par = gen_rtvec (2, set, clob); | |
2155 | } | |
2156 | ||
2157 | emit_insn (gen_rtx_PARALLEL (VOIDmode, par)); | |
2158 | } | |
2159 | ||
2160 | /* Deconstruct a floating point ABS or NEG operation | |
2161 | with integer registers into integer operations. */ | |
2162 | ||
2163 | void | |
2164 | ix86_split_fp_absneg_operator (enum rtx_code code, machine_mode mode, | |
2165 | rtx operands[]) | |
2166 | { | |
2167 | enum rtx_code absneg_op; | |
2168 | rtx dst, set; | |
2169 | ||
2170 | gcc_assert (operands_match_p (operands[0], operands[1])); | |
2171 | ||
2172 | switch (mode) | |
2173 | { | |
2174 | case E_SFmode: | |
2175 | dst = gen_lowpart (SImode, operands[0]); | |
2176 | ||
2177 | if (code == ABS) | |
2178 | { | |
2179 | set = gen_int_mode (0x7fffffff, SImode); | |
2180 | absneg_op = AND; | |
2181 | } | |
2182 | else | |
2183 | { | |
2184 | set = gen_int_mode (0x80000000, SImode); | |
2185 | absneg_op = XOR; | |
2186 | } | |
2187 | set = gen_rtx_fmt_ee (absneg_op, SImode, dst, set); | |
2188 | break; | |
2189 | ||
2190 | case E_DFmode: | |
2191 | if (TARGET_64BIT) | |
2192 | { | |
2193 | dst = gen_lowpart (DImode, operands[0]); | |
2194 | dst = gen_rtx_ZERO_EXTRACT (DImode, dst, const1_rtx, GEN_INT (63)); | |
2195 | ||
2196 | if (code == ABS) | |
2197 | set = const0_rtx; | |
2198 | else | |
2199 | set = gen_rtx_NOT (DImode, dst); | |
2200 | } | |
2201 | else | |
2202 | { | |
2203 | dst = gen_highpart (SImode, operands[0]); | |
2204 | ||
2205 | if (code == ABS) | |
2206 | { | |
2207 | set = gen_int_mode (0x7fffffff, SImode); | |
2208 | absneg_op = AND; | |
2209 | } | |
2210 | else | |
2211 | { | |
2212 | set = gen_int_mode (0x80000000, SImode); | |
2213 | absneg_op = XOR; | |
2214 | } | |
2215 | set = gen_rtx_fmt_ee (absneg_op, SImode, dst, set); | |
2216 | } | |
2217 | break; | |
2218 | ||
2219 | case E_XFmode: | |
2220 | dst = gen_rtx_REG (SImode, | |
2221 | REGNO (operands[0]) + (TARGET_64BIT ? 1 : 2)); | |
2222 | if (code == ABS) | |
2223 | { | |
2224 | set = GEN_INT (0x7fff); | |
2225 | absneg_op = AND; | |
2226 | } | |
2227 | else | |
2228 | { | |
2229 | set = GEN_INT (0x8000); | |
2230 | absneg_op = XOR; | |
2231 | } | |
2232 | set = gen_rtx_fmt_ee (absneg_op, SImode, dst, set); | |
2233 | break; | |
2234 | ||
2235 | default: | |
2236 | gcc_unreachable (); | |
2237 | } | |
2238 | ||
2239 | set = gen_rtx_SET (dst, set); | |
2240 | ||
2241 | rtx clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG)); | |
2242 | rtvec par = gen_rtvec (2, set, clob); | |
2243 | ||
2244 | emit_insn (gen_rtx_PARALLEL (VOIDmode, par)); | |
2bf6d935 ML |
2245 | } |
2246 | ||
2247 | /* Expand a copysign operation. Special case operand 0 being a constant. */ | |
2248 | ||
2249 | void | |
2250 | ix86_expand_copysign (rtx operands[]) | |
2251 | { | |
2252 | machine_mode mode, vmode; | |
7e691189 | 2253 | rtx dest, vdest, op0, op1, mask, op2, op3; |
2bf6d935 | 2254 | |
60efb1fe | 2255 | mode = GET_MODE (operands[0]); |
2bf6d935 | 2256 | |
75a97b59 L |
2257 | if (mode == HFmode) |
2258 | vmode = V8HFmode; | |
2259 | else if (mode == SFmode) | |
2bf6d935 ML |
2260 | vmode = V4SFmode; |
2261 | else if (mode == DFmode) | |
2262 | vmode = V2DFmode; | |
987a3082 | 2263 | else if (mode == TFmode) |
2bf6d935 | 2264 | vmode = mode; |
987a3082 UB |
2265 | else |
2266 | gcc_unreachable (); | |
2267 | ||
60efb1fe | 2268 | if (rtx_equal_p (operands[1], operands[2])) |
2bf6d935 | 2269 | { |
60efb1fe | 2270 | emit_move_insn (operands[0], operands[1]); |
2bf6d935 ML |
2271 | return; |
2272 | } | |
2273 | ||
7e691189 JJ |
2274 | dest = operands[0]; |
2275 | vdest = lowpart_subreg (vmode, dest, mode); | |
2276 | if (vdest == NULL_RTX) | |
2277 | vdest = gen_reg_rtx (vmode); | |
2278 | else | |
2279 | dest = NULL_RTX; | |
2280 | op1 = lowpart_subreg (vmode, force_reg (mode, operands[2]), mode); | |
864c6471 | 2281 | mask = ix86_build_signbit_mask (vmode, TARGET_AVX512F && mode != HFmode, 0); |
2bf6d935 | 2282 | |
60efb1fe | 2283 | if (CONST_DOUBLE_P (operands[1])) |
2bf6d935 | 2284 | { |
60efb1fe | 2285 | op0 = simplify_unary_operation (ABS, mode, operands[1], mode); |
2286 | /* Optimize for 0, simplify b = copy_signf (0.0f, a) to b = mask & a. */ | |
2287 | if (op0 == CONST0_RTX (mode)) | |
2bf6d935 | 2288 | { |
7e691189 JJ |
2289 | emit_move_insn (vdest, gen_rtx_AND (vmode, mask, op1)); |
2290 | if (dest) | |
2291 | emit_move_insn (dest, lowpart_subreg (mode, vdest, vmode)); | |
60efb1fe | 2292 | return; |
2bf6d935 | 2293 | } |
2bf6d935 | 2294 | |
60efb1fe | 2295 | if (GET_MODE_SIZE (mode) < 16) |
2296 | op0 = ix86_build_const_vector (vmode, false, op0); | |
2297 | op0 = force_reg (vmode, op0); | |
2bf6d935 | 2298 | } |
60efb1fe | 2299 | else |
7e691189 | 2300 | op0 = lowpart_subreg (vmode, force_reg (mode, operands[1]), mode); |
60efb1fe | 2301 | |
2302 | op2 = gen_reg_rtx (vmode); | |
2303 | op3 = gen_reg_rtx (vmode); | |
2304 | emit_move_insn (op2, gen_rtx_AND (vmode, | |
2305 | gen_rtx_NOT (vmode, mask), | |
2306 | op0)); | |
2307 | emit_move_insn (op3, gen_rtx_AND (vmode, mask, op1)); | |
7e691189 JJ |
2308 | emit_move_insn (vdest, gen_rtx_IOR (vmode, op2, op3)); |
2309 | if (dest) | |
2310 | emit_move_insn (dest, lowpart_subreg (mode, vdest, vmode)); | |
2bf6d935 ML |
2311 | } |
2312 | ||
2313 | /* Expand an xorsign operation. */ | |
2314 | ||
2315 | void | |
2316 | ix86_expand_xorsign (rtx operands[]) | |
2317 | { | |
2bf6d935 | 2318 | machine_mode mode, vmode; |
7e691189 | 2319 | rtx dest, vdest, op0, op1, mask, x, temp; |
2bf6d935 ML |
2320 | |
2321 | dest = operands[0]; | |
2322 | op0 = operands[1]; | |
2323 | op1 = operands[2]; | |
2324 | ||
2325 | mode = GET_MODE (dest); | |
2326 | ||
75a97b59 L |
2327 | if (mode == HFmode) |
2328 | vmode = V8HFmode; | |
2329 | else if (mode == SFmode) | |
987a3082 | 2330 | vmode = V4SFmode; |
2bf6d935 | 2331 | else if (mode == DFmode) |
987a3082 | 2332 | vmode = V2DFmode; |
2bf6d935 ML |
2333 | else |
2334 | gcc_unreachable (); | |
2335 | ||
7485a525 | 2336 | temp = gen_reg_rtx (vmode); |
2bf6d935 ML |
2337 | mask = ix86_build_signbit_mask (vmode, 0, 0); |
2338 | ||
7e691189 | 2339 | op1 = lowpart_subreg (vmode, force_reg (mode, op1), mode); |
7485a525 JJ |
2340 | x = gen_rtx_AND (vmode, op1, mask); |
2341 | emit_insn (gen_rtx_SET (temp, x)); | |
2bf6d935 | 2342 | |
7e691189 | 2343 | op0 = lowpart_subreg (vmode, force_reg (mode, op0), mode); |
7485a525 | 2344 | x = gen_rtx_XOR (vmode, temp, op0); |
652bef70 | 2345 | |
7e691189 JJ |
2346 | vdest = lowpart_subreg (vmode, dest, mode); |
2347 | if (vdest == NULL_RTX) | |
2348 | vdest = gen_reg_rtx (vmode); | |
2349 | else | |
2350 | dest = NULL_RTX; | |
2351 | emit_insn (gen_rtx_SET (vdest, x)); | |
2352 | ||
2353 | if (dest) | |
2354 | emit_move_insn (dest, lowpart_subreg (mode, vdest, vmode)); | |
2bf6d935 ML |
2355 | } |
2356 | ||
2357 | static rtx ix86_expand_compare (enum rtx_code code, rtx op0, rtx op1); | |
2358 | ||
2359 | void | |
2360 | ix86_expand_branch (enum rtx_code code, rtx op0, rtx op1, rtx label) | |
2361 | { | |
2362 | machine_mode mode = GET_MODE (op0); | |
2363 | rtx tmp; | |
2364 | ||
2365 | /* Handle special case - vector comparsion with boolean result, transform | |
2366 | it using ptest instruction. */ | |
850a13d7 | 2367 | if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT |
2368 | || mode == OImode) | |
2bf6d935 ML |
2369 | { |
2370 | rtx flag = gen_rtx_REG (CCZmode, FLAGS_REG); | |
2371 | machine_mode p_mode = GET_MODE_SIZE (mode) == 32 ? V4DImode : V2DImode; | |
2372 | ||
2373 | gcc_assert (code == EQ || code == NE); | |
850a13d7 | 2374 | |
2375 | if (mode == OImode) | |
2376 | { | |
2377 | op0 = lowpart_subreg (p_mode, force_reg (mode, op0), mode); | |
2378 | op1 = lowpart_subreg (p_mode, force_reg (mode, op1), mode); | |
2379 | mode = p_mode; | |
2380 | } | |
2bf6d935 ML |
2381 | /* Generate XOR since we can't check that one operand is zero vector. */ |
2382 | tmp = gen_reg_rtx (mode); | |
2383 | emit_insn (gen_rtx_SET (tmp, gen_rtx_XOR (mode, op0, op1))); | |
2384 | tmp = gen_lowpart (p_mode, tmp); | |
3635e8c6 RS |
2385 | emit_insn (gen_rtx_SET (gen_rtx_REG (CCZmode, FLAGS_REG), |
2386 | gen_rtx_UNSPEC (CCZmode, | |
2bf6d935 ML |
2387 | gen_rtvec (2, tmp, tmp), |
2388 | UNSPEC_PTEST))); | |
2389 | tmp = gen_rtx_fmt_ee (code, VOIDmode, flag, const0_rtx); | |
2390 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp, | |
2391 | gen_rtx_LABEL_REF (VOIDmode, label), | |
2392 | pc_rtx); | |
2393 | emit_jump_insn (gen_rtx_SET (pc_rtx, tmp)); | |
2394 | return; | |
2395 | } | |
2396 | ||
2397 | switch (mode) | |
2398 | { | |
a6841211 | 2399 | case E_HFmode: |
2bf6d935 ML |
2400 | case E_SFmode: |
2401 | case E_DFmode: | |
2402 | case E_XFmode: | |
2403 | case E_QImode: | |
2404 | case E_HImode: | |
2405 | case E_SImode: | |
2406 | simple: | |
2407 | tmp = ix86_expand_compare (code, op0, op1); | |
2408 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp, | |
2409 | gen_rtx_LABEL_REF (VOIDmode, label), | |
2410 | pc_rtx); | |
2411 | emit_jump_insn (gen_rtx_SET (pc_rtx, tmp)); | |
2412 | return; | |
2413 | ||
2414 | case E_DImode: | |
2415 | if (TARGET_64BIT) | |
2416 | goto simple; | |
2bf6d935 ML |
2417 | /* FALLTHRU */ |
2418 | case E_TImode: | |
43201f2c RS |
2419 | /* DI and TI mode equality/inequality comparisons may be performed |
2420 | on SSE registers. Avoid splitting them, except when optimizing | |
2421 | for size. */ | |
2422 | if ((code == EQ || code == NE) | |
2423 | && !optimize_insn_for_size_p ()) | |
2424 | goto simple; | |
2425 | ||
2bf6d935 ML |
2426 | /* Expand DImode branch into multiple compare+branch. */ |
2427 | { | |
2428 | rtx lo[2], hi[2]; | |
2429 | rtx_code_label *label2; | |
2430 | enum rtx_code code1, code2, code3; | |
2431 | machine_mode submode; | |
2432 | ||
2433 | if (CONSTANT_P (op0) && !CONSTANT_P (op1)) | |
2434 | { | |
2435 | std::swap (op0, op1); | |
2436 | code = swap_condition (code); | |
2437 | } | |
2438 | ||
2439 | split_double_mode (mode, &op0, 1, lo+0, hi+0); | |
2440 | split_double_mode (mode, &op1, 1, lo+1, hi+1); | |
2441 | ||
2442 | submode = mode == DImode ? SImode : DImode; | |
2443 | ||
43201f2c | 2444 | /* If we are doing less-than or greater-or-equal-than, |
2bf6d935 ML |
2445 | op1 is a constant and the low word is zero, then we can just |
2446 | examine the high word. Similarly for low word -1 and | |
2447 | less-or-equal-than or greater-than. */ | |
2448 | ||
2449 | if (CONST_INT_P (hi[1])) | |
2450 | switch (code) | |
2451 | { | |
2452 | case LT: case LTU: case GE: case GEU: | |
2453 | if (lo[1] == const0_rtx) | |
2454 | { | |
2455 | ix86_expand_branch (code, hi[0], hi[1], label); | |
2456 | return; | |
2457 | } | |
2458 | break; | |
2459 | case LE: case LEU: case GT: case GTU: | |
2460 | if (lo[1] == constm1_rtx) | |
2461 | { | |
2462 | ix86_expand_branch (code, hi[0], hi[1], label); | |
2463 | return; | |
2464 | } | |
2465 | break; | |
2466 | default: | |
2467 | break; | |
2468 | } | |
2469 | ||
2470 | /* Emulate comparisons that do not depend on Zero flag with | |
2471 | double-word subtraction. Note that only Overflow, Sign | |
2472 | and Carry flags are valid, so swap arguments and condition | |
2473 | of comparisons that would otherwise test Zero flag. */ | |
2474 | ||
2475 | switch (code) | |
2476 | { | |
2477 | case LE: case LEU: case GT: case GTU: | |
2478 | std::swap (lo[0], lo[1]); | |
2479 | std::swap (hi[0], hi[1]); | |
2480 | code = swap_condition (code); | |
2481 | /* FALLTHRU */ | |
2482 | ||
2483 | case LT: case LTU: case GE: case GEU: | |
2484 | { | |
2bf6d935 | 2485 | bool uns = (code == LTU || code == GEU); |
987a3082 UB |
2486 | rtx (*sbb_insn) (machine_mode, rtx, rtx, rtx) |
2487 | = uns ? gen_sub3_carry_ccc : gen_sub3_carry_ccgz; | |
2bf6d935 ML |
2488 | |
2489 | if (!nonimmediate_operand (lo[0], submode)) | |
2490 | lo[0] = force_reg (submode, lo[0]); | |
2491 | if (!x86_64_general_operand (lo[1], submode)) | |
2492 | lo[1] = force_reg (submode, lo[1]); | |
2493 | ||
2494 | if (!register_operand (hi[0], submode)) | |
2495 | hi[0] = force_reg (submode, hi[0]); | |
2496 | if ((uns && !nonimmediate_operand (hi[1], submode)) | |
2497 | || (!uns && !x86_64_general_operand (hi[1], submode))) | |
2498 | hi[1] = force_reg (submode, hi[1]); | |
2499 | ||
987a3082 | 2500 | emit_insn (gen_cmp_1 (submode, lo[0], lo[1])); |
2bf6d935 | 2501 | |
987a3082 UB |
2502 | tmp = gen_rtx_SCRATCH (submode); |
2503 | emit_insn (sbb_insn (submode, tmp, hi[0], hi[1])); | |
2bf6d935 | 2504 | |
987a3082 | 2505 | tmp = gen_rtx_REG (uns ? CCCmode : CCGZmode, FLAGS_REG); |
2bf6d935 ML |
2506 | ix86_expand_branch (code, tmp, const0_rtx, label); |
2507 | return; | |
2508 | } | |
2509 | ||
2510 | default: | |
2511 | break; | |
2512 | } | |
2513 | ||
2514 | /* Otherwise, we need two or three jumps. */ | |
2515 | ||
2516 | label2 = gen_label_rtx (); | |
2517 | ||
2518 | code1 = code; | |
2519 | code2 = swap_condition (code); | |
2520 | code3 = unsigned_condition (code); | |
2521 | ||
2522 | switch (code) | |
2523 | { | |
2524 | case LT: case GT: case LTU: case GTU: | |
2525 | break; | |
2526 | ||
2527 | case LE: code1 = LT; code2 = GT; break; | |
2528 | case GE: code1 = GT; code2 = LT; break; | |
2529 | case LEU: code1 = LTU; code2 = GTU; break; | |
2530 | case GEU: code1 = GTU; code2 = LTU; break; | |
2531 | ||
2532 | case EQ: code1 = UNKNOWN; code2 = NE; break; | |
2533 | case NE: code2 = UNKNOWN; break; | |
2534 | ||
2535 | default: | |
2536 | gcc_unreachable (); | |
2537 | } | |
2538 | ||
2539 | /* | |
2540 | * a < b => | |
2541 | * if (hi(a) < hi(b)) goto true; | |
2542 | * if (hi(a) > hi(b)) goto false; | |
2543 | * if (lo(a) < lo(b)) goto true; | |
2544 | * false: | |
2545 | */ | |
2546 | ||
2547 | if (code1 != UNKNOWN) | |
2548 | ix86_expand_branch (code1, hi[0], hi[1], label); | |
2549 | if (code2 != UNKNOWN) | |
2550 | ix86_expand_branch (code2, hi[0], hi[1], label2); | |
2551 | ||
2552 | ix86_expand_branch (code3, lo[0], lo[1], label); | |
2553 | ||
2554 | if (code2 != UNKNOWN) | |
2555 | emit_label (label2); | |
2556 | return; | |
2557 | } | |
2558 | ||
2559 | default: | |
2560 | gcc_assert (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC); | |
2561 | goto simple; | |
2562 | } | |
2563 | } | |
2564 | ||
2565 | /* Figure out whether to use unordered fp comparisons. */ | |
2566 | ||
2567 | static bool | |
2568 | ix86_unordered_fp_compare (enum rtx_code code) | |
2569 | { | |
2570 | if (!TARGET_IEEE_FP) | |
2571 | return false; | |
2572 | ||
2573 | switch (code) | |
2574 | { | |
2bf6d935 ML |
2575 | case LT: |
2576 | case LE: | |
d6038777 UB |
2577 | case GT: |
2578 | case GE: | |
2579 | case LTGT: | |
2bf6d935 ML |
2580 | return false; |
2581 | ||
2582 | case EQ: | |
2583 | case NE: | |
2584 | ||
2bf6d935 ML |
2585 | case UNORDERED: |
2586 | case ORDERED: | |
2587 | case UNLT: | |
2588 | case UNLE: | |
2589 | case UNGT: | |
2590 | case UNGE: | |
2591 | case UNEQ: | |
2592 | return true; | |
2593 | ||
2594 | default: | |
2595 | gcc_unreachable (); | |
2596 | } | |
2597 | } | |
2598 | ||
2599 | /* Return a comparison we can do and that it is equivalent to | |
2600 | swap_condition (code) apart possibly from orderedness. | |
2601 | But, never change orderedness if TARGET_IEEE_FP, returning | |
2602 | UNKNOWN in that case if necessary. */ | |
2603 | ||
2604 | static enum rtx_code | |
2605 | ix86_fp_swap_condition (enum rtx_code code) | |
2606 | { | |
2607 | switch (code) | |
2608 | { | |
2609 | case GT: /* GTU - CF=0 & ZF=0 */ | |
2610 | return TARGET_IEEE_FP ? UNKNOWN : UNLT; | |
2611 | case GE: /* GEU - CF=0 */ | |
2612 | return TARGET_IEEE_FP ? UNKNOWN : UNLE; | |
2613 | case UNLT: /* LTU - CF=1 */ | |
2614 | return TARGET_IEEE_FP ? UNKNOWN : GT; | |
2615 | case UNLE: /* LEU - CF=1 | ZF=1 */ | |
2616 | return TARGET_IEEE_FP ? UNKNOWN : GE; | |
2617 | default: | |
2618 | return swap_condition (code); | |
2619 | } | |
2620 | } | |
2621 | ||
2622 | /* Return cost of comparison CODE using the best strategy for performance. | |
2623 | All following functions do use number of instructions as a cost metrics. | |
2624 | In future this should be tweaked to compute bytes for optimize_size and | |
2625 | take into account performance of various instructions on various CPUs. */ | |
2626 | ||
2627 | static int | |
2628 | ix86_fp_comparison_cost (enum rtx_code code) | |
2629 | { | |
2630 | int arith_cost; | |
2631 | ||
2632 | /* The cost of code using bit-twiddling on %ah. */ | |
2633 | switch (code) | |
2634 | { | |
2635 | case UNLE: | |
2636 | case UNLT: | |
2637 | case LTGT: | |
2638 | case GT: | |
2639 | case GE: | |
2640 | case UNORDERED: | |
2641 | case ORDERED: | |
2642 | case UNEQ: | |
2643 | arith_cost = 4; | |
2644 | break; | |
2645 | case LT: | |
2646 | case NE: | |
2647 | case EQ: | |
2648 | case UNGE: | |
2649 | arith_cost = TARGET_IEEE_FP ? 5 : 4; | |
2650 | break; | |
2651 | case LE: | |
2652 | case UNGT: | |
2653 | arith_cost = TARGET_IEEE_FP ? 6 : 4; | |
2654 | break; | |
2655 | default: | |
2656 | gcc_unreachable (); | |
2657 | } | |
2658 | ||
2659 | switch (ix86_fp_comparison_strategy (code)) | |
2660 | { | |
2661 | case IX86_FPCMP_COMI: | |
2662 | return arith_cost > 4 ? 3 : 2; | |
2663 | case IX86_FPCMP_SAHF: | |
2664 | return arith_cost > 4 ? 4 : 3; | |
2665 | default: | |
2666 | return arith_cost; | |
2667 | } | |
2668 | } | |
2669 | ||
2670 | /* Swap, force into registers, or otherwise massage the two operands | |
2671 | to a fp comparison. The operands are updated in place; the new | |
2672 | comparison code is returned. */ | |
2673 | ||
2674 | static enum rtx_code | |
2675 | ix86_prepare_fp_compare_args (enum rtx_code code, rtx *pop0, rtx *pop1) | |
2676 | { | |
2677 | bool unordered_compare = ix86_unordered_fp_compare (code); | |
2678 | rtx op0 = *pop0, op1 = *pop1; | |
2679 | machine_mode op_mode = GET_MODE (op0); | |
a6841211 | 2680 | bool is_sse = SSE_FLOAT_MODE_SSEMATH_OR_HF_P (op_mode); |
2bf6d935 | 2681 | |
5792208f JJ |
2682 | if (op_mode == BFmode) |
2683 | { | |
2684 | rtx op = gen_lowpart (HImode, op0); | |
2685 | if (CONST_INT_P (op)) | |
2686 | op = simplify_const_unary_operation (FLOAT_EXTEND, SFmode, | |
2687 | op0, BFmode); | |
2688 | else | |
2689 | { | |
2690 | rtx t1 = gen_reg_rtx (SImode); | |
2691 | emit_insn (gen_zero_extendhisi2 (t1, op)); | |
2692 | emit_insn (gen_ashlsi3 (t1, t1, GEN_INT (16))); | |
2693 | op = gen_lowpart (SFmode, t1); | |
2694 | } | |
2695 | *pop0 = op; | |
2696 | op = gen_lowpart (HImode, op1); | |
2697 | if (CONST_INT_P (op)) | |
2698 | op = simplify_const_unary_operation (FLOAT_EXTEND, SFmode, | |
2699 | op1, BFmode); | |
2700 | else | |
2701 | { | |
2702 | rtx t1 = gen_reg_rtx (SImode); | |
2703 | emit_insn (gen_zero_extendhisi2 (t1, op)); | |
2704 | emit_insn (gen_ashlsi3 (t1, t1, GEN_INT (16))); | |
2705 | op = gen_lowpart (SFmode, t1); | |
2706 | } | |
2707 | *pop1 = op; | |
2708 | return ix86_prepare_fp_compare_args (code, pop0, pop1); | |
2709 | } | |
2710 | ||
2bf6d935 ML |
2711 | /* All of the unordered compare instructions only work on registers. |
2712 | The same is true of the fcomi compare instructions. The XFmode | |
2713 | compare instructions require registers except when comparing | |
2714 | against zero or when converting operand 1 from fixed point to | |
2715 | floating point. */ | |
2716 | ||
2717 | if (!is_sse | |
2718 | && (unordered_compare | |
2719 | || (op_mode == XFmode | |
2720 | && ! (standard_80387_constant_p (op0) == 1 | |
2721 | || standard_80387_constant_p (op1) == 1) | |
2722 | && GET_CODE (op1) != FLOAT) | |
2723 | || ix86_fp_comparison_strategy (code) == IX86_FPCMP_COMI)) | |
2724 | { | |
2725 | op0 = force_reg (op_mode, op0); | |
2726 | op1 = force_reg (op_mode, op1); | |
2727 | } | |
2728 | else | |
2729 | { | |
2730 | /* %%% We only allow op1 in memory; op0 must be st(0). So swap | |
2731 | things around if they appear profitable, otherwise force op0 | |
2732 | into a register. */ | |
2733 | ||
2734 | if (standard_80387_constant_p (op0) == 0 | |
2735 | || (MEM_P (op0) | |
2736 | && ! (standard_80387_constant_p (op1) == 0 | |
2737 | || MEM_P (op1)))) | |
2738 | { | |
2739 | enum rtx_code new_code = ix86_fp_swap_condition (code); | |
2740 | if (new_code != UNKNOWN) | |
2741 | { | |
2742 | std::swap (op0, op1); | |
2743 | code = new_code; | |
2744 | } | |
2745 | } | |
2746 | ||
2747 | if (!REG_P (op0)) | |
2748 | op0 = force_reg (op_mode, op0); | |
2749 | ||
2750 | if (CONSTANT_P (op1)) | |
2751 | { | |
2752 | int tmp = standard_80387_constant_p (op1); | |
2753 | if (tmp == 0) | |
2754 | op1 = validize_mem (force_const_mem (op_mode, op1)); | |
2755 | else if (tmp == 1) | |
2756 | { | |
2757 | if (TARGET_CMOVE) | |
2758 | op1 = force_reg (op_mode, op1); | |
2759 | } | |
2760 | else | |
2761 | op1 = force_reg (op_mode, op1); | |
2762 | } | |
2763 | } | |
2764 | ||
2765 | /* Try to rearrange the comparison to make it cheaper. */ | |
2766 | if (ix86_fp_comparison_cost (code) | |
2767 | > ix86_fp_comparison_cost (swap_condition (code)) | |
2768 | && (REG_P (op1) || can_create_pseudo_p ())) | |
2769 | { | |
2770 | std::swap (op0, op1); | |
2771 | code = swap_condition (code); | |
2772 | if (!REG_P (op0)) | |
2773 | op0 = force_reg (op_mode, op0); | |
2774 | } | |
2775 | ||
2776 | *pop0 = op0; | |
2777 | *pop1 = op1; | |
2778 | return code; | |
2779 | } | |
2780 | ||
2781 | /* Generate insn patterns to do a floating point compare of OPERANDS. */ | |
2782 | ||
2783 | static rtx | |
2784 | ix86_expand_fp_compare (enum rtx_code code, rtx op0, rtx op1) | |
2785 | { | |
2786 | bool unordered_compare = ix86_unordered_fp_compare (code); | |
2787 | machine_mode cmp_mode; | |
2788 | rtx tmp, scratch; | |
2789 | ||
2790 | code = ix86_prepare_fp_compare_args (code, &op0, &op1); | |
2791 | ||
2792 | tmp = gen_rtx_COMPARE (CCFPmode, op0, op1); | |
2793 | if (unordered_compare) | |
2794 | tmp = gen_rtx_UNSPEC (CCFPmode, gen_rtvec (1, tmp), UNSPEC_NOTRAP); | |
2795 | ||
2796 | /* Do fcomi/sahf based test when profitable. */ | |
2797 | switch (ix86_fp_comparison_strategy (code)) | |
2798 | { | |
2799 | case IX86_FPCMP_COMI: | |
2800 | cmp_mode = CCFPmode; | |
2801 | emit_insn (gen_rtx_SET (gen_rtx_REG (CCFPmode, FLAGS_REG), tmp)); | |
2802 | break; | |
2803 | ||
2804 | case IX86_FPCMP_SAHF: | |
2805 | cmp_mode = CCFPmode; | |
2806 | tmp = gen_rtx_UNSPEC (HImode, gen_rtvec (1, tmp), UNSPEC_FNSTSW); | |
2807 | scratch = gen_reg_rtx (HImode); | |
2808 | emit_insn (gen_rtx_SET (scratch, tmp)); | |
2809 | emit_insn (gen_x86_sahf_1 (scratch)); | |
2810 | break; | |
2811 | ||
2812 | case IX86_FPCMP_ARITH: | |
2813 | cmp_mode = CCNOmode; | |
2814 | tmp = gen_rtx_UNSPEC (HImode, gen_rtvec (1, tmp), UNSPEC_FNSTSW); | |
2815 | scratch = gen_reg_rtx (HImode); | |
2816 | emit_insn (gen_rtx_SET (scratch, tmp)); | |
2817 | ||
2818 | /* In the unordered case, we have to check C2 for NaN's, which | |
2819 | doesn't happen to work out to anything nice combination-wise. | |
2820 | So do some bit twiddling on the value we've got in AH to come | |
2821 | up with an appropriate set of condition codes. */ | |
2822 | ||
2823 | switch (code) | |
2824 | { | |
2825 | case GT: | |
2826 | case UNGT: | |
2827 | if (code == GT || !TARGET_IEEE_FP) | |
2828 | { | |
2829 | emit_insn (gen_testqi_ext_1_ccno (scratch, GEN_INT (0x45))); | |
2830 | code = EQ; | |
2831 | } | |
2832 | else | |
2833 | { | |
2834 | emit_insn (gen_andqi_ext_1 (scratch, scratch, GEN_INT (0x45))); | |
2835 | emit_insn (gen_addqi_ext_1 (scratch, scratch, constm1_rtx)); | |
2836 | emit_insn (gen_cmpqi_ext_3 (scratch, GEN_INT (0x44))); | |
2837 | cmp_mode = CCmode; | |
2838 | code = GEU; | |
2839 | } | |
2840 | break; | |
2841 | case LT: | |
2842 | case UNLT: | |
2843 | if (code == LT && TARGET_IEEE_FP) | |
2844 | { | |
2845 | emit_insn (gen_andqi_ext_1 (scratch, scratch, GEN_INT (0x45))); | |
2846 | emit_insn (gen_cmpqi_ext_3 (scratch, const1_rtx)); | |
2847 | cmp_mode = CCmode; | |
2848 | code = EQ; | |
2849 | } | |
2850 | else | |
2851 | { | |
2852 | emit_insn (gen_testqi_ext_1_ccno (scratch, const1_rtx)); | |
2853 | code = NE; | |
2854 | } | |
2855 | break; | |
2856 | case GE: | |
2857 | case UNGE: | |
2858 | if (code == GE || !TARGET_IEEE_FP) | |
2859 | { | |
2860 | emit_insn (gen_testqi_ext_1_ccno (scratch, GEN_INT (0x05))); | |
2861 | code = EQ; | |
2862 | } | |
2863 | else | |
2864 | { | |
2865 | emit_insn (gen_andqi_ext_1 (scratch, scratch, GEN_INT (0x45))); | |
2866 | emit_insn (gen_xorqi_ext_1_cc (scratch, scratch, const1_rtx)); | |
2867 | code = NE; | |
2868 | } | |
2869 | break; | |
2870 | case LE: | |
2871 | case UNLE: | |
2872 | if (code == LE && TARGET_IEEE_FP) | |
2873 | { | |
2874 | emit_insn (gen_andqi_ext_1 (scratch, scratch, GEN_INT (0x45))); | |
2875 | emit_insn (gen_addqi_ext_1 (scratch, scratch, constm1_rtx)); | |
2876 | emit_insn (gen_cmpqi_ext_3 (scratch, GEN_INT (0x40))); | |
2877 | cmp_mode = CCmode; | |
2878 | code = LTU; | |
2879 | } | |
2880 | else | |
2881 | { | |
2882 | emit_insn (gen_testqi_ext_1_ccno (scratch, GEN_INT (0x45))); | |
2883 | code = NE; | |
2884 | } | |
2885 | break; | |
2886 | case EQ: | |
2887 | case UNEQ: | |
2888 | if (code == EQ && TARGET_IEEE_FP) | |
2889 | { | |
2890 | emit_insn (gen_andqi_ext_1 (scratch, scratch, GEN_INT (0x45))); | |
2891 | emit_insn (gen_cmpqi_ext_3 (scratch, GEN_INT (0x40))); | |
2892 | cmp_mode = CCmode; | |
2893 | code = EQ; | |
2894 | } | |
2895 | else | |
2896 | { | |
2897 | emit_insn (gen_testqi_ext_1_ccno (scratch, GEN_INT (0x40))); | |
2898 | code = NE; | |
2899 | } | |
2900 | break; | |
2901 | case NE: | |
2902 | case LTGT: | |
2903 | if (code == NE && TARGET_IEEE_FP) | |
2904 | { | |
2905 | emit_insn (gen_andqi_ext_1 (scratch, scratch, GEN_INT (0x45))); | |
2906 | emit_insn (gen_xorqi_ext_1_cc (scratch, scratch, | |
2907 | GEN_INT (0x40))); | |
2908 | code = NE; | |
2909 | } | |
2910 | else | |
2911 | { | |
2912 | emit_insn (gen_testqi_ext_1_ccno (scratch, GEN_INT (0x40))); | |
2913 | code = EQ; | |
2914 | } | |
2915 | break; | |
2916 | ||
2917 | case UNORDERED: | |
2918 | emit_insn (gen_testqi_ext_1_ccno (scratch, GEN_INT (0x04))); | |
2919 | code = NE; | |
2920 | break; | |
2921 | case ORDERED: | |
2922 | emit_insn (gen_testqi_ext_1_ccno (scratch, GEN_INT (0x04))); | |
2923 | code = EQ; | |
2924 | break; | |
2925 | ||
2926 | default: | |
2927 | gcc_unreachable (); | |
2928 | } | |
2929 | break; | |
2930 | ||
2931 | default: | |
2932 | gcc_unreachable(); | |
2933 | } | |
2934 | ||
2935 | /* Return the test that should be put into the flags user, i.e. | |
2936 | the bcc, scc, or cmov instruction. */ | |
2937 | return gen_rtx_fmt_ee (code, VOIDmode, | |
2938 | gen_rtx_REG (cmp_mode, FLAGS_REG), | |
2939 | const0_rtx); | |
2940 | } | |
2941 | ||
2942 | /* Generate insn patterns to do an integer compare of OPERANDS. */ | |
2943 | ||
2944 | static rtx | |
2945 | ix86_expand_int_compare (enum rtx_code code, rtx op0, rtx op1) | |
2946 | { | |
2947 | machine_mode cmpmode; | |
2948 | rtx tmp, flags; | |
2949 | ||
86403f4e UB |
2950 | /* Swap operands to emit carry flag comparison. */ |
2951 | if ((code == GTU || code == LEU) | |
2952 | && nonimmediate_operand (op1, VOIDmode)) | |
2953 | { | |
2954 | std::swap (op0, op1); | |
2955 | code = swap_condition (code); | |
2956 | } | |
2957 | ||
2bf6d935 ML |
2958 | cmpmode = SELECT_CC_MODE (code, op0, op1); |
2959 | flags = gen_rtx_REG (cmpmode, FLAGS_REG); | |
2960 | ||
2961 | /* This is very simple, but making the interface the same as in the | |
2962 | FP case makes the rest of the code easier. */ | |
2963 | tmp = gen_rtx_COMPARE (cmpmode, op0, op1); | |
2964 | emit_insn (gen_rtx_SET (flags, tmp)); | |
2965 | ||
2966 | /* Return the test that should be put into the flags user, i.e. | |
2967 | the bcc, scc, or cmov instruction. */ | |
2968 | return gen_rtx_fmt_ee (code, VOIDmode, flags, const0_rtx); | |
2969 | } | |
2970 | ||
2971 | static rtx | |
2972 | ix86_expand_compare (enum rtx_code code, rtx op0, rtx op1) | |
2973 | { | |
2974 | rtx ret; | |
2975 | ||
2976 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC) | |
2977 | ret = gen_rtx_fmt_ee (code, VOIDmode, op0, op1); | |
2978 | ||
2979 | else if (SCALAR_FLOAT_MODE_P (GET_MODE (op0))) | |
2980 | { | |
2981 | gcc_assert (!DECIMAL_FLOAT_MODE_P (GET_MODE (op0))); | |
2982 | ret = ix86_expand_fp_compare (code, op0, op1); | |
2983 | } | |
2984 | else | |
2985 | ret = ix86_expand_int_compare (code, op0, op1); | |
2986 | ||
2987 | return ret; | |
2988 | } | |
2989 | ||
2990 | void | |
2991 | ix86_expand_setcc (rtx dest, enum rtx_code code, rtx op0, rtx op1) | |
2992 | { | |
2993 | rtx ret; | |
2994 | ||
2995 | gcc_assert (GET_MODE (dest) == QImode); | |
2996 | ||
2997 | ret = ix86_expand_compare (code, op0, op1); | |
2998 | PUT_MODE (ret, QImode); | |
2999 | emit_insn (gen_rtx_SET (dest, ret)); | |
3000 | } | |
3001 | ||
463d9108 JJ |
3002 | /* Expand floating point op0 <=> op1, i.e. |
3003 | dest = op0 == op1 ? 0 : op0 < op1 ? -1 : op0 > op1 ? 1 : 2. */ | |
3004 | ||
3005 | void | |
3006 | ix86_expand_fp_spaceship (rtx dest, rtx op0, rtx op1) | |
3007 | { | |
3008 | gcc_checking_assert (ix86_fp_comparison_strategy (GT) != IX86_FPCMP_ARITH); | |
3009 | rtx gt = ix86_expand_fp_compare (GT, op0, op1); | |
3010 | rtx l0 = gen_label_rtx (); | |
3011 | rtx l1 = gen_label_rtx (); | |
3012 | rtx l2 = TARGET_IEEE_FP ? gen_label_rtx () : NULL_RTX; | |
3013 | rtx lend = gen_label_rtx (); | |
3014 | rtx tmp; | |
3015 | rtx_insn *jmp; | |
3016 | if (l2) | |
3017 | { | |
3018 | rtx un = gen_rtx_fmt_ee (UNORDERED, VOIDmode, | |
3019 | gen_rtx_REG (CCFPmode, FLAGS_REG), const0_rtx); | |
3020 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, un, | |
3021 | gen_rtx_LABEL_REF (VOIDmode, l2), pc_rtx); | |
3022 | jmp = emit_jump_insn (gen_rtx_SET (pc_rtx, tmp)); | |
3023 | add_reg_br_prob_note (jmp, profile_probability:: very_unlikely ()); | |
3024 | } | |
3025 | rtx eq = gen_rtx_fmt_ee (UNEQ, VOIDmode, | |
3026 | gen_rtx_REG (CCFPmode, FLAGS_REG), const0_rtx); | |
3027 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, eq, | |
3028 | gen_rtx_LABEL_REF (VOIDmode, l0), pc_rtx); | |
3029 | jmp = emit_jump_insn (gen_rtx_SET (pc_rtx, tmp)); | |
3030 | add_reg_br_prob_note (jmp, profile_probability::unlikely ()); | |
3031 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, gt, | |
3032 | gen_rtx_LABEL_REF (VOIDmode, l1), pc_rtx); | |
3033 | jmp = emit_jump_insn (gen_rtx_SET (pc_rtx, tmp)); | |
3034 | add_reg_br_prob_note (jmp, profile_probability::even ()); | |
3035 | emit_move_insn (dest, constm1_rtx); | |
3036 | emit_jump (lend); | |
3037 | emit_label (l0); | |
3038 | emit_move_insn (dest, const0_rtx); | |
3039 | emit_jump (lend); | |
3040 | emit_label (l1); | |
3041 | emit_move_insn (dest, const1_rtx); | |
3042 | emit_jump (lend); | |
3043 | if (l2) | |
3044 | { | |
3045 | emit_label (l2); | |
3046 | emit_move_insn (dest, const2_rtx); | |
3047 | } | |
3048 | emit_label (lend); | |
3049 | } | |
3050 | ||
2bf6d935 ML |
3051 | /* Expand comparison setting or clearing carry flag. Return true when |
3052 | successful and set pop for the operation. */ | |
3053 | static bool | |
3054 | ix86_expand_carry_flag_compare (enum rtx_code code, rtx op0, rtx op1, rtx *pop) | |
3055 | { | |
3056 | machine_mode mode | |
3057 | = GET_MODE (op0) != VOIDmode ? GET_MODE (op0) : GET_MODE (op1); | |
3058 | ||
3059 | /* Do not handle double-mode compares that go through special path. */ | |
3060 | if (mode == (TARGET_64BIT ? TImode : DImode)) | |
3061 | return false; | |
3062 | ||
3063 | if (SCALAR_FLOAT_MODE_P (mode)) | |
3064 | { | |
3065 | rtx compare_op; | |
3066 | rtx_insn *compare_seq; | |
3067 | ||
3068 | gcc_assert (!DECIMAL_FLOAT_MODE_P (mode)); | |
3069 | ||
3070 | /* Shortcut: following common codes never translate | |
3071 | into carry flag compares. */ | |
3072 | if (code == EQ || code == NE || code == UNEQ || code == LTGT | |
3073 | || code == ORDERED || code == UNORDERED) | |
3074 | return false; | |
3075 | ||
3076 | /* These comparisons require zero flag; swap operands so they won't. */ | |
3077 | if ((code == GT || code == UNLE || code == LE || code == UNGT) | |
3078 | && !TARGET_IEEE_FP) | |
3079 | { | |
3080 | std::swap (op0, op1); | |
3081 | code = swap_condition (code); | |
3082 | } | |
3083 | ||
3084 | /* Try to expand the comparison and verify that we end up with | |
3085 | carry flag based comparison. This fails to be true only when | |
3086 | we decide to expand comparison using arithmetic that is not | |
3087 | too common scenario. */ | |
3088 | start_sequence (); | |
3089 | compare_op = ix86_expand_fp_compare (code, op0, op1); | |
3090 | compare_seq = get_insns (); | |
3091 | end_sequence (); | |
3092 | ||
3093 | if (GET_MODE (XEXP (compare_op, 0)) == CCFPmode) | |
3094 | code = ix86_fp_compare_code_to_integer (GET_CODE (compare_op)); | |
3095 | else | |
3096 | code = GET_CODE (compare_op); | |
3097 | ||
3098 | if (code != LTU && code != GEU) | |
3099 | return false; | |
3100 | ||
3101 | emit_insn (compare_seq); | |
3102 | *pop = compare_op; | |
3103 | return true; | |
3104 | } | |
3105 | ||
3106 | if (!INTEGRAL_MODE_P (mode)) | |
3107 | return false; | |
3108 | ||
3109 | switch (code) | |
3110 | { | |
3111 | case LTU: | |
3112 | case GEU: | |
3113 | break; | |
3114 | ||
3115 | /* Convert a==0 into (unsigned)a<1. */ | |
3116 | case EQ: | |
3117 | case NE: | |
3118 | if (op1 != const0_rtx) | |
3119 | return false; | |
3120 | op1 = const1_rtx; | |
3121 | code = (code == EQ ? LTU : GEU); | |
3122 | break; | |
3123 | ||
3124 | /* Convert a>b into b<a or a>=b-1. */ | |
3125 | case GTU: | |
3126 | case LEU: | |
3127 | if (CONST_INT_P (op1)) | |
3128 | { | |
3129 | op1 = gen_int_mode (INTVAL (op1) + 1, GET_MODE (op0)); | |
3130 | /* Bail out on overflow. We still can swap operands but that | |
3131 | would force loading of the constant into register. */ | |
3132 | if (op1 == const0_rtx | |
3133 | || !x86_64_immediate_operand (op1, GET_MODE (op1))) | |
3134 | return false; | |
3135 | code = (code == GTU ? GEU : LTU); | |
3136 | } | |
3137 | else | |
3138 | { | |
3139 | std::swap (op0, op1); | |
3140 | code = (code == GTU ? LTU : GEU); | |
3141 | } | |
3142 | break; | |
3143 | ||
3144 | /* Convert a>=0 into (unsigned)a<0x80000000. */ | |
3145 | case LT: | |
3146 | case GE: | |
3147 | if (mode == DImode || op1 != const0_rtx) | |
3148 | return false; | |
3149 | op1 = gen_int_mode (1 << (GET_MODE_BITSIZE (mode) - 1), mode); | |
3150 | code = (code == LT ? GEU : LTU); | |
3151 | break; | |
3152 | case LE: | |
3153 | case GT: | |
3154 | if (mode == DImode || op1 != constm1_rtx) | |
3155 | return false; | |
3156 | op1 = gen_int_mode (1 << (GET_MODE_BITSIZE (mode) - 1), mode); | |
3157 | code = (code == LE ? GEU : LTU); | |
3158 | break; | |
3159 | ||
3160 | default: | |
3161 | return false; | |
3162 | } | |
3163 | /* Swapping operands may cause constant to appear as first operand. */ | |
3164 | if (!nonimmediate_operand (op0, VOIDmode)) | |
3165 | { | |
3166 | if (!can_create_pseudo_p ()) | |
3167 | return false; | |
3168 | op0 = force_reg (mode, op0); | |
3169 | } | |
3170 | *pop = ix86_expand_compare (code, op0, op1); | |
3171 | gcc_assert (GET_CODE (*pop) == LTU || GET_CODE (*pop) == GEU); | |
3172 | return true; | |
3173 | } | |
3174 | ||
3175 | /* Expand conditional increment or decrement using adb/sbb instructions. | |
3176 | The default case using setcc followed by the conditional move can be | |
3177 | done by generic code. */ | |
3178 | bool | |
3179 | ix86_expand_int_addcc (rtx operands[]) | |
3180 | { | |
3181 | enum rtx_code code = GET_CODE (operands[1]); | |
3182 | rtx flags; | |
987a3082 | 3183 | rtx (*insn) (machine_mode, rtx, rtx, rtx, rtx, rtx); |
2bf6d935 ML |
3184 | rtx compare_op; |
3185 | rtx val = const0_rtx; | |
3186 | bool fpcmp = false; | |
3187 | machine_mode mode; | |
3188 | rtx op0 = XEXP (operands[1], 0); | |
3189 | rtx op1 = XEXP (operands[1], 1); | |
3190 | ||
3191 | if (operands[3] != const1_rtx | |
3192 | && operands[3] != constm1_rtx) | |
3193 | return false; | |
3194 | if (!ix86_expand_carry_flag_compare (code, op0, op1, &compare_op)) | |
3195 | return false; | |
3196 | code = GET_CODE (compare_op); | |
3197 | ||
3198 | flags = XEXP (compare_op, 0); | |
3199 | ||
3200 | if (GET_MODE (flags) == CCFPmode) | |
3201 | { | |
3202 | fpcmp = true; | |
3203 | code = ix86_fp_compare_code_to_integer (code); | |
3204 | } | |
3205 | ||
3206 | if (code != LTU) | |
3207 | { | |
3208 | val = constm1_rtx; | |
3209 | if (fpcmp) | |
3210 | PUT_CODE (compare_op, | |
3211 | reverse_condition_maybe_unordered | |
3212 | (GET_CODE (compare_op))); | |
3213 | else | |
3214 | PUT_CODE (compare_op, reverse_condition (GET_CODE (compare_op))); | |
3215 | } | |
3216 | ||
3217 | mode = GET_MODE (operands[0]); | |
3218 | ||
3219 | /* Construct either adc or sbb insn. */ | |
3220 | if ((code == LTU) == (operands[3] == constm1_rtx)) | |
987a3082 | 3221 | insn = gen_sub3_carry; |
2bf6d935 | 3222 | else |
987a3082 UB |
3223 | insn = gen_add3_carry; |
3224 | ||
3225 | emit_insn (insn (mode, operands[0], operands[2], val, flags, compare_op)); | |
2bf6d935 ML |
3226 | |
3227 | return true; | |
3228 | } | |
3229 | ||
3230 | bool | |
3231 | ix86_expand_int_movcc (rtx operands[]) | |
3232 | { | |
3233 | enum rtx_code code = GET_CODE (operands[1]), compare_code; | |
3234 | rtx_insn *compare_seq; | |
3235 | rtx compare_op; | |
3236 | machine_mode mode = GET_MODE (operands[0]); | |
3237 | bool sign_bit_compare_p = false; | |
f1652e33 | 3238 | bool negate_cc_compare_p = false; |
2bf6d935 ML |
3239 | rtx op0 = XEXP (operands[1], 0); |
3240 | rtx op1 = XEXP (operands[1], 1); | |
1ceddd74 JJ |
3241 | rtx op2 = operands[2]; |
3242 | rtx op3 = operands[3]; | |
2bf6d935 ML |
3243 | |
3244 | if (GET_MODE (op0) == TImode | |
3245 | || (GET_MODE (op0) == DImode | |
3246 | && !TARGET_64BIT)) | |
3247 | return false; | |
3248 | ||
5792208f JJ |
3249 | if (GET_MODE (op0) == BFmode |
3250 | && !ix86_fp_comparison_operator (operands[1], VOIDmode)) | |
3251 | return false; | |
3252 | ||
2bf6d935 ML |
3253 | start_sequence (); |
3254 | compare_op = ix86_expand_compare (code, op0, op1); | |
3255 | compare_seq = get_insns (); | |
3256 | end_sequence (); | |
3257 | ||
3258 | compare_code = GET_CODE (compare_op); | |
3259 | ||
3260 | if ((op1 == const0_rtx && (code == GE || code == LT)) | |
3261 | || (op1 == constm1_rtx && (code == GT || code == LE))) | |
3262 | sign_bit_compare_p = true; | |
3263 | ||
1ceddd74 JJ |
3264 | /* op0 == op1 ? op0 : op3 is equivalent to op0 == op1 ? op1 : op3, |
3265 | but if op1 is a constant, the latter form allows more optimizations, | |
3266 | either through the last 2 ops being constant handling, or the one | |
3267 | constant and one variable cases. On the other side, for cmov the | |
3268 | former might be better as we don't need to load the constant into | |
3269 | another register. */ | |
3270 | if (code == EQ && CONST_INT_P (op1) && rtx_equal_p (op0, op2)) | |
3271 | op2 = op1; | |
3272 | /* Similarly for op0 != op1 ? op2 : op0 and op0 != op1 ? op2 : op1. */ | |
3273 | else if (code == NE && CONST_INT_P (op1) && rtx_equal_p (op0, op3)) | |
3274 | op3 = op1; | |
3275 | ||
2bf6d935 ML |
3276 | /* Don't attempt mode expansion here -- if we had to expand 5 or 6 |
3277 | HImode insns, we'd be swallowed in word prefix ops. */ | |
3278 | ||
3279 | if ((mode != HImode || TARGET_FAST_PREFIX) | |
3280 | && (mode != (TARGET_64BIT ? TImode : DImode)) | |
1ceddd74 JJ |
3281 | && CONST_INT_P (op2) |
3282 | && CONST_INT_P (op3)) | |
2bf6d935 ML |
3283 | { |
3284 | rtx out = operands[0]; | |
1ceddd74 JJ |
3285 | HOST_WIDE_INT ct = INTVAL (op2); |
3286 | HOST_WIDE_INT cf = INTVAL (op3); | |
2bf6d935 ML |
3287 | HOST_WIDE_INT diff; |
3288 | ||
f1652e33 RS |
3289 | if ((mode == SImode |
3290 | || (TARGET_64BIT && mode == DImode)) | |
3291 | && (GET_MODE (op0) == SImode | |
3292 | || (TARGET_64BIT && GET_MODE (op0) == DImode))) | |
3293 | { | |
3294 | /* Special case x != 0 ? -1 : y. */ | |
3295 | if (code == NE && op1 == const0_rtx && ct == -1) | |
3296 | { | |
3297 | negate_cc_compare_p = true; | |
3298 | std::swap (ct, cf); | |
3299 | code = EQ; | |
3300 | } | |
3301 | else if (code == EQ && op1 == const0_rtx && cf == -1) | |
3302 | negate_cc_compare_p = true; | |
3303 | } | |
3304 | ||
2bf6d935 ML |
3305 | diff = ct - cf; |
3306 | /* Sign bit compares are better done using shifts than we do by using | |
3307 | sbb. */ | |
3308 | if (sign_bit_compare_p | |
f1652e33 | 3309 | || negate_cc_compare_p |
2bf6d935 ML |
3310 | || ix86_expand_carry_flag_compare (code, op0, op1, &compare_op)) |
3311 | { | |
9e6ac747 RS |
3312 | /* Detect overlap between destination and compare sources. */ |
3313 | rtx tmp = out; | |
2bf6d935 | 3314 | |
f1652e33 RS |
3315 | if (negate_cc_compare_p) |
3316 | { | |
3317 | if (GET_MODE (op0) == DImode) | |
3318 | emit_insn (gen_x86_negdi_ccc (gen_reg_rtx (DImode), op0)); | |
3319 | else | |
3320 | emit_insn (gen_x86_negsi_ccc (gen_reg_rtx (SImode), | |
3321 | gen_lowpart (SImode, op0))); | |
3322 | ||
9e6ac747 | 3323 | tmp = gen_reg_rtx (mode); |
f1652e33 RS |
3324 | if (mode == DImode) |
3325 | emit_insn (gen_x86_movdicc_0_m1_neg (tmp)); | |
3326 | else | |
3327 | emit_insn (gen_x86_movsicc_0_m1_neg (gen_lowpart (SImode, | |
3328 | tmp))); | |
3329 | } | |
3330 | else if (!sign_bit_compare_p) | |
2bf6d935 ML |
3331 | { |
3332 | rtx flags; | |
3333 | bool fpcmp = false; | |
3334 | ||
3335 | compare_code = GET_CODE (compare_op); | |
3336 | ||
3337 | flags = XEXP (compare_op, 0); | |
3338 | ||
3339 | if (GET_MODE (flags) == CCFPmode) | |
3340 | { | |
3341 | fpcmp = true; | |
3342 | compare_code | |
3343 | = ix86_fp_compare_code_to_integer (compare_code); | |
3344 | } | |
3345 | ||
3346 | /* To simplify rest of code, restrict to the GEU case. */ | |
3347 | if (compare_code == LTU) | |
3348 | { | |
3349 | std::swap (ct, cf); | |
3350 | compare_code = reverse_condition (compare_code); | |
3351 | code = reverse_condition (code); | |
3352 | } | |
3353 | else | |
3354 | { | |
3355 | if (fpcmp) | |
3356 | PUT_CODE (compare_op, | |
3357 | reverse_condition_maybe_unordered | |
3358 | (GET_CODE (compare_op))); | |
3359 | else | |
3360 | PUT_CODE (compare_op, | |
3361 | reverse_condition (GET_CODE (compare_op))); | |
3362 | } | |
3363 | diff = ct - cf; | |
3364 | ||
9e6ac747 RS |
3365 | if (reg_overlap_mentioned_p (out, compare_op)) |
3366 | tmp = gen_reg_rtx (mode); | |
3367 | ||
2bf6d935 ML |
3368 | if (mode == DImode) |
3369 | emit_insn (gen_x86_movdicc_0_m1 (tmp, flags, compare_op)); | |
3370 | else | |
3371 | emit_insn (gen_x86_movsicc_0_m1 (gen_lowpart (SImode, tmp), | |
3372 | flags, compare_op)); | |
3373 | } | |
3374 | else | |
3375 | { | |
3376 | if (code == GT || code == GE) | |
3377 | code = reverse_condition (code); | |
3378 | else | |
3379 | { | |
3380 | std::swap (ct, cf); | |
3381 | diff = ct - cf; | |
3382 | } | |
3383 | tmp = emit_store_flag (tmp, code, op0, op1, VOIDmode, 0, -1); | |
3384 | } | |
3385 | ||
3386 | if (diff == 1) | |
3387 | { | |
3388 | /* | |
3389 | * cmpl op0,op1 | |
3390 | * sbbl dest,dest | |
3391 | * [addl dest, ct] | |
3392 | * | |
3393 | * Size 5 - 8. | |
3394 | */ | |
3395 | if (ct) | |
9e6ac747 RS |
3396 | tmp = expand_simple_binop (mode, PLUS, |
3397 | tmp, GEN_INT (ct), | |
3398 | copy_rtx (tmp), 1, OPTAB_DIRECT); | |
2bf6d935 ML |
3399 | } |
3400 | else if (cf == -1) | |
3401 | { | |
3402 | /* | |
3403 | * cmpl op0,op1 | |
3404 | * sbbl dest,dest | |
3405 | * orl $ct, dest | |
3406 | * | |
3407 | * Size 8. | |
3408 | */ | |
9e6ac747 RS |
3409 | tmp = expand_simple_binop (mode, IOR, |
3410 | tmp, GEN_INT (ct), | |
3411 | copy_rtx (tmp), 1, OPTAB_DIRECT); | |
2bf6d935 ML |
3412 | } |
3413 | else if (diff == -1 && ct) | |
3414 | { | |
3415 | /* | |
3416 | * cmpl op0,op1 | |
3417 | * sbbl dest,dest | |
3418 | * notl dest | |
3419 | * [addl dest, cf] | |
3420 | * | |
3421 | * Size 8 - 11. | |
3422 | */ | |
9e6ac747 | 3423 | tmp = expand_simple_unop (mode, NOT, tmp, copy_rtx (tmp), 1); |
2bf6d935 | 3424 | if (cf) |
9e6ac747 RS |
3425 | tmp = expand_simple_binop (mode, PLUS, |
3426 | copy_rtx (tmp), GEN_INT (cf), | |
3427 | copy_rtx (tmp), 1, OPTAB_DIRECT); | |
2bf6d935 ML |
3428 | } |
3429 | else | |
3430 | { | |
3431 | /* | |
3432 | * cmpl op0,op1 | |
3433 | * sbbl dest,dest | |
3434 | * [notl dest] | |
3435 | * andl cf - ct, dest | |
3436 | * [addl dest, ct] | |
3437 | * | |
3438 | * Size 8 - 11. | |
3439 | */ | |
3440 | ||
3441 | if (cf == 0) | |
3442 | { | |
3443 | cf = ct; | |
3444 | ct = 0; | |
9e6ac747 | 3445 | tmp = expand_simple_unop (mode, NOT, tmp, copy_rtx (tmp), 1); |
2bf6d935 ML |
3446 | } |
3447 | ||
9e6ac747 RS |
3448 | tmp = expand_simple_binop (mode, AND, |
3449 | copy_rtx (tmp), | |
2bf6d935 | 3450 | gen_int_mode (cf - ct, mode), |
9e6ac747 | 3451 | copy_rtx (tmp), 1, OPTAB_DIRECT); |
2bf6d935 | 3452 | if (ct) |
9e6ac747 RS |
3453 | tmp = expand_simple_binop (mode, PLUS, |
3454 | copy_rtx (tmp), GEN_INT (ct), | |
3455 | copy_rtx (tmp), 1, OPTAB_DIRECT); | |
2bf6d935 ML |
3456 | } |
3457 | ||
9e6ac747 RS |
3458 | if (!rtx_equal_p (tmp, out)) |
3459 | emit_move_insn (copy_rtx (out), copy_rtx (tmp)); | |
3460 | ||
2bf6d935 ML |
3461 | return true; |
3462 | } | |
3463 | ||
3464 | if (diff < 0) | |
3465 | { | |
3466 | machine_mode cmp_mode = GET_MODE (op0); | |
3467 | enum rtx_code new_code; | |
3468 | ||
3469 | if (SCALAR_FLOAT_MODE_P (cmp_mode)) | |
3470 | { | |
3471 | gcc_assert (!DECIMAL_FLOAT_MODE_P (cmp_mode)); | |
3472 | ||
8f17461b UB |
3473 | /* We may be reversing a non-trapping |
3474 | comparison to a trapping comparison. */ | |
3475 | if (HONOR_NANS (cmp_mode) && flag_trapping_math | |
3476 | && code != EQ && code != NE | |
3477 | && code != ORDERED && code != UNORDERED) | |
3478 | new_code = UNKNOWN; | |
3479 | else | |
3480 | new_code = reverse_condition_maybe_unordered (code); | |
2bf6d935 ML |
3481 | } |
3482 | else | |
3483 | new_code = ix86_reverse_condition (code, cmp_mode); | |
3484 | if (new_code != UNKNOWN) | |
3485 | { | |
3486 | std::swap (ct, cf); | |
3487 | diff = -diff; | |
3488 | code = new_code; | |
3489 | } | |
3490 | } | |
3491 | ||
3492 | compare_code = UNKNOWN; | |
3493 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT | |
3494 | && CONST_INT_P (op1)) | |
3495 | { | |
3496 | if (op1 == const0_rtx | |
3497 | && (code == LT || code == GE)) | |
3498 | compare_code = code; | |
3499 | else if (op1 == constm1_rtx) | |
3500 | { | |
3501 | if (code == LE) | |
3502 | compare_code = LT; | |
3503 | else if (code == GT) | |
3504 | compare_code = GE; | |
3505 | } | |
3506 | } | |
3507 | ||
3508 | /* Optimize dest = (op0 < 0) ? -1 : cf. */ | |
3509 | if (compare_code != UNKNOWN | |
3510 | && GET_MODE (op0) == GET_MODE (out) | |
3511 | && (cf == -1 || ct == -1)) | |
3512 | { | |
3513 | /* If lea code below could be used, only optimize | |
3514 | if it results in a 2 insn sequence. */ | |
3515 | ||
3516 | if (! (diff == 1 || diff == 2 || diff == 4 || diff == 8 | |
3517 | || diff == 3 || diff == 5 || diff == 9) | |
3518 | || (compare_code == LT && ct == -1) | |
3519 | || (compare_code == GE && cf == -1)) | |
3520 | { | |
3521 | /* | |
3522 | * notl op1 (if necessary) | |
3523 | * sarl $31, op1 | |
3524 | * orl cf, op1 | |
3525 | */ | |
3526 | if (ct != -1) | |
3527 | { | |
3528 | cf = ct; | |
3529 | ct = -1; | |
3530 | code = reverse_condition (code); | |
3531 | } | |
3532 | ||
3533 | out = emit_store_flag (out, code, op0, op1, VOIDmode, 0, -1); | |
3534 | ||
3535 | out = expand_simple_binop (mode, IOR, | |
3536 | out, GEN_INT (cf), | |
3537 | out, 1, OPTAB_DIRECT); | |
3538 | if (out != operands[0]) | |
3539 | emit_move_insn (operands[0], out); | |
3540 | ||
3541 | return true; | |
3542 | } | |
3543 | } | |
3544 | ||
3545 | ||
3546 | if ((diff == 1 || diff == 2 || diff == 4 || diff == 8 | |
3547 | || diff == 3 || diff == 5 || diff == 9) | |
3548 | && ((mode != QImode && mode != HImode) || !TARGET_PARTIAL_REG_STALL) | |
3549 | && (mode != DImode | |
3550 | || x86_64_immediate_operand (GEN_INT (cf), VOIDmode))) | |
3551 | { | |
3552 | /* | |
3553 | * xorl dest,dest | |
3554 | * cmpl op1,op2 | |
3555 | * setcc dest | |
3556 | * lea cf(dest*(ct-cf)),dest | |
3557 | * | |
3558 | * Size 14. | |
3559 | * | |
3560 | * This also catches the degenerate setcc-only case. | |
3561 | */ | |
3562 | ||
3563 | rtx tmp; | |
3564 | int nops; | |
3565 | ||
3566 | out = emit_store_flag (out, code, op0, op1, VOIDmode, 0, 1); | |
3567 | ||
3568 | nops = 0; | |
3569 | /* On x86_64 the lea instruction operates on Pmode, so we need | |
3570 | to get arithmetics done in proper mode to match. */ | |
3571 | if (diff == 1) | |
3572 | tmp = copy_rtx (out); | |
3573 | else | |
3574 | { | |
3575 | rtx out1; | |
3576 | out1 = copy_rtx (out); | |
3577 | tmp = gen_rtx_MULT (mode, out1, GEN_INT (diff & ~1)); | |
3578 | nops++; | |
3579 | if (diff & 1) | |
3580 | { | |
3581 | tmp = gen_rtx_PLUS (mode, tmp, out1); | |
3582 | nops++; | |
3583 | } | |
3584 | } | |
3585 | if (cf != 0) | |
3586 | { | |
c3185b64 | 3587 | tmp = plus_constant (mode, tmp, cf); |
2bf6d935 ML |
3588 | nops++; |
3589 | } | |
3590 | if (!rtx_equal_p (tmp, out)) | |
3591 | { | |
3592 | if (nops == 1) | |
3593 | out = force_operand (tmp, copy_rtx (out)); | |
3594 | else | |
3595 | emit_insn (gen_rtx_SET (copy_rtx (out), copy_rtx (tmp))); | |
3596 | } | |
3597 | if (!rtx_equal_p (out, operands[0])) | |
3598 | emit_move_insn (operands[0], copy_rtx (out)); | |
3599 | ||
3600 | return true; | |
3601 | } | |
3602 | ||
3603 | /* | |
3604 | * General case: Jumpful: | |
3605 | * xorl dest,dest cmpl op1, op2 | |
3606 | * cmpl op1, op2 movl ct, dest | |
3607 | * setcc dest jcc 1f | |
3608 | * decl dest movl cf, dest | |
3609 | * andl (cf-ct),dest 1: | |
3610 | * addl ct,dest | |
3611 | * | |
3612 | * Size 20. Size 14. | |
3613 | * | |
3614 | * This is reasonably steep, but branch mispredict costs are | |
3615 | * high on modern cpus, so consider failing only if optimizing | |
3616 | * for space. | |
3617 | */ | |
3618 | ||
3619 | if ((!TARGET_CMOVE || (mode == QImode && TARGET_PARTIAL_REG_STALL)) | |
3620 | && BRANCH_COST (optimize_insn_for_speed_p (), | |
3621 | false) >= 2) | |
3622 | { | |
3623 | if (cf == 0) | |
3624 | { | |
3625 | machine_mode cmp_mode = GET_MODE (op0); | |
3626 | enum rtx_code new_code; | |
3627 | ||
3628 | if (SCALAR_FLOAT_MODE_P (cmp_mode)) | |
3629 | { | |
3630 | gcc_assert (!DECIMAL_FLOAT_MODE_P (cmp_mode)); | |
3631 | ||
8f17461b UB |
3632 | /* We may be reversing a non-trapping |
3633 | comparison to a trapping comparison. */ | |
3634 | if (HONOR_NANS (cmp_mode) && flag_trapping_math | |
3635 | && code != EQ && code != NE | |
3636 | && code != ORDERED && code != UNORDERED) | |
3637 | new_code = UNKNOWN; | |
3638 | else | |
3639 | new_code = reverse_condition_maybe_unordered (code); | |
3640 | ||
2bf6d935 ML |
3641 | } |
3642 | else | |
3643 | { | |
3644 | new_code = ix86_reverse_condition (code, cmp_mode); | |
3645 | if (compare_code != UNKNOWN && new_code != UNKNOWN) | |
3646 | compare_code = reverse_condition (compare_code); | |
3647 | } | |
3648 | ||
3649 | if (new_code != UNKNOWN) | |
3650 | { | |
3651 | cf = ct; | |
3652 | ct = 0; | |
3653 | code = new_code; | |
3654 | } | |
3655 | } | |
3656 | ||
3657 | if (compare_code != UNKNOWN) | |
3658 | { | |
3659 | /* notl op1 (if needed) | |
3660 | sarl $31, op1 | |
3661 | andl (cf-ct), op1 | |
3662 | addl ct, op1 | |
3663 | ||
3664 | For x < 0 (resp. x <= -1) there will be no notl, | |
3665 | so if possible swap the constants to get rid of the | |
3666 | complement. | |
3667 | True/false will be -1/0 while code below (store flag | |
3668 | followed by decrement) is 0/-1, so the constants need | |
3669 | to be exchanged once more. */ | |
3670 | ||
3671 | if (compare_code == GE || !cf) | |
3672 | { | |
3673 | code = reverse_condition (code); | |
3674 | compare_code = LT; | |
3675 | } | |
3676 | else | |
3677 | std::swap (ct, cf); | |
3678 | ||
3679 | out = emit_store_flag (out, code, op0, op1, VOIDmode, 0, -1); | |
3680 | } | |
3681 | else | |
3682 | { | |
3683 | out = emit_store_flag (out, code, op0, op1, VOIDmode, 0, 1); | |
3684 | ||
3685 | out = expand_simple_binop (mode, PLUS, copy_rtx (out), | |
3686 | constm1_rtx, | |
3687 | copy_rtx (out), 1, OPTAB_DIRECT); | |
3688 | } | |
3689 | ||
3690 | out = expand_simple_binop (mode, AND, copy_rtx (out), | |
3691 | gen_int_mode (cf - ct, mode), | |
3692 | copy_rtx (out), 1, OPTAB_DIRECT); | |
3693 | if (ct) | |
3694 | out = expand_simple_binop (mode, PLUS, copy_rtx (out), GEN_INT (ct), | |
3695 | copy_rtx (out), 1, OPTAB_DIRECT); | |
3696 | if (!rtx_equal_p (out, operands[0])) | |
3697 | emit_move_insn (operands[0], copy_rtx (out)); | |
3698 | ||
3699 | return true; | |
3700 | } | |
3701 | } | |
3702 | ||
3703 | if (!TARGET_CMOVE || (mode == QImode && TARGET_PARTIAL_REG_STALL)) | |
3704 | { | |
3705 | /* Try a few things more with specific constants and a variable. */ | |
3706 | ||
3707 | optab op; | |
3708 | rtx var, orig_out, out, tmp; | |
3709 | ||
3710 | if (BRANCH_COST (optimize_insn_for_speed_p (), false) <= 2) | |
3711 | return false; | |
3712 | ||
1ceddd74 JJ |
3713 | operands[2] = op2; |
3714 | operands[3] = op3; | |
3715 | ||
2bf6d935 ML |
3716 | /* If one of the two operands is an interesting constant, load a |
3717 | constant with the above and mask it in with a logical operation. */ | |
3718 | ||
3719 | if (CONST_INT_P (operands[2])) | |
3720 | { | |
3721 | var = operands[3]; | |
3722 | if (INTVAL (operands[2]) == 0 && operands[3] != constm1_rtx) | |
3723 | operands[3] = constm1_rtx, op = and_optab; | |
3724 | else if (INTVAL (operands[2]) == -1 && operands[3] != const0_rtx) | |
3725 | operands[3] = const0_rtx, op = ior_optab; | |
3726 | else | |
3727 | return false; | |
3728 | } | |
3729 | else if (CONST_INT_P (operands[3])) | |
3730 | { | |
3731 | var = operands[2]; | |
3732 | if (INTVAL (operands[3]) == 0 && operands[2] != constm1_rtx) | |
e4ced0b6 RS |
3733 | { |
3734 | /* For smin (x, 0), expand as "x < 0 ? x : 0" instead of | |
3735 | "x <= 0 ? x : 0" to enable sign_bit_compare_p. */ | |
3736 | if (code == LE && op1 == const0_rtx && rtx_equal_p (op0, var)) | |
3737 | operands[1] = simplify_gen_relational (LT, VOIDmode, | |
3738 | GET_MODE (op0), | |
3739 | op0, const0_rtx); | |
3740 | ||
3741 | operands[2] = constm1_rtx; | |
3742 | op = and_optab; | |
3743 | } | |
2bf6d935 ML |
3744 | else if (INTVAL (operands[3]) == -1 && operands[3] != const0_rtx) |
3745 | operands[2] = const0_rtx, op = ior_optab; | |
3746 | else | |
3747 | return false; | |
3748 | } | |
3749 | else | |
3750 | return false; | |
3751 | ||
3752 | orig_out = operands[0]; | |
3753 | tmp = gen_reg_rtx (mode); | |
3754 | operands[0] = tmp; | |
3755 | ||
3756 | /* Recurse to get the constant loaded. */ | |
3757 | if (!ix86_expand_int_movcc (operands)) | |
3758 | return false; | |
3759 | ||
3760 | /* Mask in the interesting variable. */ | |
3761 | out = expand_binop (mode, op, var, tmp, orig_out, 0, | |
3762 | OPTAB_WIDEN); | |
3763 | if (!rtx_equal_p (out, orig_out)) | |
3764 | emit_move_insn (copy_rtx (orig_out), copy_rtx (out)); | |
3765 | ||
3766 | return true; | |
3767 | } | |
3768 | ||
3769 | /* | |
3770 | * For comparison with above, | |
3771 | * | |
3772 | * movl cf,dest | |
3773 | * movl ct,tmp | |
3774 | * cmpl op1,op2 | |
3775 | * cmovcc tmp,dest | |
3776 | * | |
3777 | * Size 15. | |
3778 | */ | |
3779 | ||
3780 | if (! nonimmediate_operand (operands[2], mode)) | |
3781 | operands[2] = force_reg (mode, operands[2]); | |
3782 | if (! nonimmediate_operand (operands[3], mode)) | |
3783 | operands[3] = force_reg (mode, operands[3]); | |
3784 | ||
3785 | if (! register_operand (operands[2], VOIDmode) | |
3786 | && (mode == QImode | |
3787 | || ! register_operand (operands[3], VOIDmode))) | |
3788 | operands[2] = force_reg (mode, operands[2]); | |
3789 | ||
3790 | if (mode == QImode | |
3791 | && ! register_operand (operands[3], VOIDmode)) | |
3792 | operands[3] = force_reg (mode, operands[3]); | |
3793 | ||
3794 | emit_insn (compare_seq); | |
3795 | emit_insn (gen_rtx_SET (operands[0], | |
3796 | gen_rtx_IF_THEN_ELSE (mode, | |
3797 | compare_op, operands[2], | |
3798 | operands[3]))); | |
3799 | return true; | |
3800 | } | |
3801 | ||
3802 | /* Detect conditional moves that exactly match min/max operational | |
3803 | semantics. Note that this is IEEE safe, as long as we don't | |
3804 | interchange the operands. | |
3805 | ||
3806 | Returns FALSE if this conditional move doesn't match a MIN/MAX, | |
3807 | and TRUE if the operation is successful and instructions are emitted. */ | |
3808 | ||
3809 | static bool | |
3810 | ix86_expand_sse_fp_minmax (rtx dest, enum rtx_code code, rtx cmp_op0, | |
3811 | rtx cmp_op1, rtx if_true, rtx if_false) | |
3812 | { | |
3813 | machine_mode mode; | |
3814 | bool is_min; | |
3815 | rtx tmp; | |
3816 | ||
3817 | if (code == LT) | |
3818 | ; | |
3819 | else if (code == UNGE) | |
3820 | std::swap (if_true, if_false); | |
3821 | else | |
3822 | return false; | |
3823 | ||
3824 | if (rtx_equal_p (cmp_op0, if_true) && rtx_equal_p (cmp_op1, if_false)) | |
3825 | is_min = true; | |
3826 | else if (rtx_equal_p (cmp_op1, if_true) && rtx_equal_p (cmp_op0, if_false)) | |
3827 | is_min = false; | |
3828 | else | |
3829 | return false; | |
3830 | ||
3831 | mode = GET_MODE (dest); | |
3832 | ||
3833 | /* We want to check HONOR_NANS and HONOR_SIGNED_ZEROS here, | |
3834 | but MODE may be a vector mode and thus not appropriate. */ | |
3835 | if (!flag_finite_math_only || flag_signed_zeros) | |
3836 | { | |
3837 | int u = is_min ? UNSPEC_IEEE_MIN : UNSPEC_IEEE_MAX; | |
3838 | rtvec v; | |
3839 | ||
3840 | if_true = force_reg (mode, if_true); | |
3841 | v = gen_rtvec (2, if_true, if_false); | |
3842 | tmp = gen_rtx_UNSPEC (mode, v, u); | |
3843 | } | |
3844 | else | |
3845 | { | |
3846 | code = is_min ? SMIN : SMAX; | |
3847 | if (MEM_P (if_true) && MEM_P (if_false)) | |
3848 | if_true = force_reg (mode, if_true); | |
3849 | tmp = gen_rtx_fmt_ee (code, mode, if_true, if_false); | |
3850 | } | |
3851 | ||
3852 | emit_insn (gen_rtx_SET (dest, tmp)); | |
3853 | return true; | |
3854 | } | |
3855 | ||
8b905e9b HL |
3856 | /* Return true if MODE is valid for vector compare to mask register, |
3857 | Same result for conditionl vector move with mask register. */ | |
3858 | static bool | |
3859 | ix86_valid_mask_cmp_mode (machine_mode mode) | |
3860 | { | |
3861 | /* XOP has its own vector conditional movement. */ | |
a8654147 | 3862 | if (TARGET_XOP && !TARGET_AVX512F) |
8b905e9b HL |
3863 | return false; |
3864 | ||
0d788c35 | 3865 | /* HFmode only supports vcmpsh whose dest is mask register. */ |
3866 | if (TARGET_AVX512FP16 && mode == HFmode) | |
3867 | return true; | |
3868 | ||
8b905e9b HL |
3869 | /* AVX512F is needed for mask operation. */ |
3870 | if (!(TARGET_AVX512F && VECTOR_MODE_P (mode))) | |
3871 | return false; | |
3872 | ||
3873 | /* AVX512BW is needed for vector QI/HImode, | |
3874 | AVX512VL is needed for 128/256-bit vector. */ | |
3875 | machine_mode inner_mode = GET_MODE_INNER (mode); | |
3876 | int vector_size = GET_MODE_SIZE (mode); | |
3877 | if ((inner_mode == QImode || inner_mode == HImode) && !TARGET_AVX512BW) | |
3878 | return false; | |
3879 | ||
3880 | return vector_size == 64 || TARGET_AVX512VL; | |
3881 | } | |
3882 | ||
8d0737d8 | 3883 | /* Return true if integer mask comparison should be used. */ |
3884 | static bool | |
3885 | ix86_use_mask_cmp_p (machine_mode mode, machine_mode cmp_mode, | |
3886 | rtx op_true, rtx op_false) | |
3887 | { | |
92f372f0 UB |
3888 | int vector_size = GET_MODE_SIZE (mode); |
3889 | ||
0d788c35 | 3890 | if (cmp_mode == HFmode) |
3891 | return true; | |
3892 | else if (vector_size < 16) | |
92f372f0 UB |
3893 | return false; |
3894 | else if (vector_size == 64) | |
8d0737d8 | 3895 | return true; |
9ce50028 HW |
3896 | else if (GET_MODE_INNER (cmp_mode) == HFmode) |
3897 | return true; | |
8d0737d8 | 3898 | |
3899 | /* When op_true is NULL, op_false must be NULL, or vice versa. */ | |
3900 | gcc_assert (!op_true == !op_false); | |
3901 | ||
3902 | /* When op_true/op_false is NULL or cmp_mode is not valid mask cmp mode, | |
3903 | vector dest is required. */ | |
3904 | if (!op_true || !ix86_valid_mask_cmp_mode (cmp_mode)) | |
3905 | return false; | |
3906 | ||
3907 | /* Exclude those that could be optimized in ix86_expand_sse_movcc. */ | |
3908 | if (op_false == CONST0_RTX (mode) | |
3909 | || op_true == CONST0_RTX (mode) | |
3910 | || (INTEGRAL_MODE_P (mode) | |
3911 | && (op_true == CONSTM1_RTX (mode) | |
3912 | || op_false == CONSTM1_RTX (mode)))) | |
3913 | return false; | |
3914 | ||
3915 | return true; | |
3916 | } | |
3917 | ||
2bf6d935 ML |
3918 | /* Expand an SSE comparison. Return the register with the result. */ |
3919 | ||
3920 | static rtx | |
3921 | ix86_expand_sse_cmp (rtx dest, enum rtx_code code, rtx cmp_op0, rtx cmp_op1, | |
3922 | rtx op_true, rtx op_false) | |
3923 | { | |
3924 | machine_mode mode = GET_MODE (dest); | |
3925 | machine_mode cmp_ops_mode = GET_MODE (cmp_op0); | |
3926 | ||
3927 | /* In general case result of comparison can differ from operands' type. */ | |
3928 | machine_mode cmp_mode; | |
3929 | ||
3930 | /* In AVX512F the result of comparison is an integer mask. */ | |
3931 | bool maskcmp = false; | |
3932 | rtx x; | |
3933 | ||
8d0737d8 | 3934 | if (ix86_use_mask_cmp_p (mode, cmp_ops_mode, op_true, op_false)) |
2bf6d935 ML |
3935 | { |
3936 | unsigned int nbits = GET_MODE_NUNITS (cmp_ops_mode); | |
2bf6d935 | 3937 | maskcmp = true; |
8b905e9b | 3938 | cmp_mode = nbits > 8 ? int_mode_for_size (nbits, 0).require () : E_QImode; |
2bf6d935 ML |
3939 | } |
3940 | else | |
3941 | cmp_mode = cmp_ops_mode; | |
3942 | ||
3943 | cmp_op0 = force_reg (cmp_ops_mode, cmp_op0); | |
3944 | ||
a86b3453 | 3945 | bool (*op1_predicate)(rtx, machine_mode) |
2bf6d935 ML |
3946 | = VECTOR_MODE_P (cmp_ops_mode) ? vector_operand : nonimmediate_operand; |
3947 | ||
3948 | if (!op1_predicate (cmp_op1, cmp_ops_mode)) | |
3949 | cmp_op1 = force_reg (cmp_ops_mode, cmp_op1); | |
3950 | ||
3951 | if (optimize | |
3952 | || (maskcmp && cmp_mode != mode) | |
3953 | || (op_true && reg_overlap_mentioned_p (dest, op_true)) | |
3954 | || (op_false && reg_overlap_mentioned_p (dest, op_false))) | |
3955 | dest = gen_reg_rtx (maskcmp ? cmp_mode : mode); | |
3956 | ||
99e4891e | 3957 | if (maskcmp) |
3958 | { | |
3959 | bool ok = ix86_expand_mask_vec_cmp (dest, code, cmp_op0, cmp_op1); | |
3960 | gcc_assert (ok); | |
3961 | return dest; | |
3962 | } | |
3963 | ||
2bf6d935 ML |
3964 | x = gen_rtx_fmt_ee (code, cmp_mode, cmp_op0, cmp_op1); |
3965 | ||
8d0737d8 | 3966 | if (cmp_mode != mode) |
2bf6d935 ML |
3967 | { |
3968 | x = force_reg (cmp_ops_mode, x); | |
3969 | convert_move (dest, x, false); | |
3970 | } | |
3971 | else | |
3972 | emit_insn (gen_rtx_SET (dest, x)); | |
3973 | ||
3974 | return dest; | |
3975 | } | |
3976 | ||
b5193e35 UB |
3977 | /* Emit x86 binary operand CODE in mode MODE for SSE vector |
3978 | instructions that can be performed using GP registers. */ | |
3979 | ||
3980 | static void | |
3981 | ix86_emit_vec_binop (enum rtx_code code, machine_mode mode, | |
3982 | rtx dst, rtx src1, rtx src2) | |
3983 | { | |
3984 | rtx tmp; | |
3985 | ||
3986 | tmp = gen_rtx_SET (dst, gen_rtx_fmt_ee (code, mode, src1, src2)); | |
3987 | ||
3988 | if (GET_MODE_SIZE (mode) <= GET_MODE_SIZE (SImode) | |
3989 | && GET_MODE_CLASS (mode) == MODE_VECTOR_INT) | |
3990 | { | |
3991 | rtx clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG)); | |
3992 | tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, tmp, clob)); | |
3993 | } | |
3994 | ||
3995 | emit_insn (tmp); | |
3996 | } | |
3997 | ||
2bf6d935 ML |
3998 | /* Expand DEST = CMP ? OP_TRUE : OP_FALSE into a sequence of logical |
3999 | operations. This is used for both scalar and vector conditional moves. */ | |
4000 | ||
4001 | void | |
4002 | ix86_expand_sse_movcc (rtx dest, rtx cmp, rtx op_true, rtx op_false) | |
4003 | { | |
4004 | machine_mode mode = GET_MODE (dest); | |
4005 | machine_mode cmpmode = GET_MODE (cmp); | |
f4a2cecd | 4006 | rtx x; |
2bf6d935 | 4007 | |
9b5d50b7 | 4008 | /* Simplify trivial VEC_COND_EXPR to avoid ICE in pr97506. */ |
4009 | if (rtx_equal_p (op_true, op_false)) | |
4010 | { | |
4011 | emit_move_insn (dest, op_true); | |
4012 | return; | |
4013 | } | |
4014 | ||
2bf6d935 ML |
4015 | /* If we have an integer mask and FP value then we need |
4016 | to cast mask to FP mode. */ | |
4017 | if (mode != cmpmode && VECTOR_MODE_P (cmpmode)) | |
4018 | { | |
4019 | cmp = force_reg (cmpmode, cmp); | |
4020 | cmp = gen_rtx_SUBREG (mode, cmp, 0); | |
4021 | } | |
4022 | ||
8d0737d8 | 4023 | /* In AVX512F the result of comparison is an integer mask. */ |
4024 | if (mode != cmpmode | |
4025 | && GET_MODE_CLASS (cmpmode) == MODE_INT) | |
2bf6d935 | 4026 | { |
8d0737d8 | 4027 | gcc_assert (ix86_valid_mask_cmp_mode (mode)); |
0d788c35 | 4028 | /* Using scalar/vector move with mask register. */ |
8b905e9b HL |
4029 | cmp = force_reg (cmpmode, cmp); |
4030 | /* Optimize for mask zero. */ | |
4031 | op_true = (op_true != CONST0_RTX (mode) | |
4032 | ? force_reg (mode, op_true) : op_true); | |
4033 | op_false = (op_false != CONST0_RTX (mode) | |
4034 | ? force_reg (mode, op_false) : op_false); | |
4035 | if (op_true == CONST0_RTX (mode)) | |
2bf6d935 | 4036 | { |
ee78c20e | 4037 | if (cmpmode == E_DImode && !TARGET_64BIT) |
f4a2cecd UB |
4038 | { |
4039 | x = gen_reg_rtx (cmpmode); | |
4040 | emit_insn (gen_knotdi (x, cmp)); | |
4041 | } | |
ee78c20e | 4042 | else |
f4a2cecd UB |
4043 | x = expand_simple_unop (cmpmode, NOT, cmp, NULL, 1); |
4044 | cmp = x; | |
8b905e9b HL |
4045 | /* Reverse op_true op_false. */ |
4046 | std::swap (op_true, op_false); | |
2bf6d935 | 4047 | } |
8b905e9b | 4048 | |
0d788c35 | 4049 | if (mode == HFmode) |
4050 | emit_insn (gen_movhf_mask (dest, op_true, op_false, cmp)); | |
4051 | else | |
f4a2cecd UB |
4052 | emit_insn (gen_rtx_SET (dest, |
4053 | gen_rtx_VEC_MERGE (mode, | |
4054 | op_true, op_false, cmp))); | |
8b905e9b | 4055 | return; |
2bf6d935 | 4056 | } |
f4a2cecd UB |
4057 | |
4058 | if (vector_all_ones_operand (op_true, mode) | |
4059 | && op_false == CONST0_RTX (mode)) | |
2bf6d935 | 4060 | { |
f4a2cecd | 4061 | emit_move_insn (dest, cmp); |
2bf6d935 ML |
4062 | return; |
4063 | } | |
4064 | else if (op_false == CONST0_RTX (mode)) | |
4065 | { | |
f4a2cecd UB |
4066 | x = expand_simple_binop (mode, AND, cmp, op_true, |
4067 | dest, 1, OPTAB_DIRECT); | |
4068 | if (x != dest) | |
4069 | emit_move_insn (dest, x); | |
2bf6d935 ML |
4070 | return; |
4071 | } | |
4072 | else if (op_true == CONST0_RTX (mode)) | |
4073 | { | |
4074 | op_false = force_reg (mode, op_false); | |
4075 | x = gen_rtx_NOT (mode, cmp); | |
b5193e35 | 4076 | ix86_emit_vec_binop (AND, mode, dest, x, op_false); |
2bf6d935 ML |
4077 | return; |
4078 | } | |
f4a2cecd | 4079 | else if (vector_all_ones_operand (op_true, mode)) |
2bf6d935 | 4080 | { |
f4a2cecd UB |
4081 | x = expand_simple_binop (mode, IOR, cmp, op_false, |
4082 | dest, 1, OPTAB_DIRECT); | |
4083 | if (x != dest) | |
4084 | emit_move_insn (dest, x); | |
2bf6d935 ML |
4085 | return; |
4086 | } | |
f4a2cecd UB |
4087 | |
4088 | if (TARGET_XOP) | |
2bf6d935 ML |
4089 | { |
4090 | op_true = force_reg (mode, op_true); | |
4091 | ||
f1693741 UB |
4092 | if (GET_MODE_SIZE (mode) < 16 |
4093 | || !nonimmediate_operand (op_false, mode)) | |
2bf6d935 ML |
4094 | op_false = force_reg (mode, op_false); |
4095 | ||
f4a2cecd UB |
4096 | emit_insn (gen_rtx_SET (dest, |
4097 | gen_rtx_IF_THEN_ELSE (mode, cmp, | |
4098 | op_true, op_false))); | |
2bf6d935 ML |
4099 | return; |
4100 | } | |
4101 | ||
4102 | rtx (*gen) (rtx, rtx, rtx, rtx) = NULL; | |
f4a2cecd | 4103 | machine_mode blend_mode = mode; |
2bf6d935 | 4104 | |
f4a2cecd UB |
4105 | if (GET_MODE_SIZE (mode) < 16 |
4106 | || !vector_operand (op_true, mode)) | |
2bf6d935 ML |
4107 | op_true = force_reg (mode, op_true); |
4108 | ||
4109 | op_false = force_reg (mode, op_false); | |
4110 | ||
4111 | switch (mode) | |
4112 | { | |
b1f7fd8a UB |
4113 | case E_V2SFmode: |
4114 | if (TARGET_SSE4_1) | |
f4a2cecd | 4115 | gen = gen_mmx_blendvps; |
b1f7fd8a | 4116 | break; |
2bf6d935 ML |
4117 | case E_V4SFmode: |
4118 | if (TARGET_SSE4_1) | |
4119 | gen = gen_sse4_1_blendvps; | |
4120 | break; | |
4121 | case E_V2DFmode: | |
4122 | if (TARGET_SSE4_1) | |
4123 | gen = gen_sse4_1_blendvpd; | |
4124 | break; | |
4125 | case E_SFmode: | |
4126 | if (TARGET_SSE4_1) | |
f4a2cecd | 4127 | gen = gen_sse4_1_blendvss; |
2bf6d935 ML |
4128 | break; |
4129 | case E_DFmode: | |
4130 | if (TARGET_SSE4_1) | |
f4a2cecd | 4131 | gen = gen_sse4_1_blendvsd; |
2bf6d935 | 4132 | break; |
5795ec0e UB |
4133 | case E_V8QImode: |
4134 | case E_V4HImode: | |
4135 | case E_V2SImode: | |
4136 | if (TARGET_SSE4_1) | |
4137 | { | |
820ac79e | 4138 | gen = gen_mmx_pblendvb_v8qi; |
f4a2cecd | 4139 | blend_mode = V8QImode; |
5795ec0e UB |
4140 | } |
4141 | break; | |
2df9d3c5 UB |
4142 | case E_V4QImode: |
4143 | case E_V2HImode: | |
4144 | if (TARGET_SSE4_1) | |
4145 | { | |
820ac79e | 4146 | gen = gen_mmx_pblendvb_v4qi; |
f4a2cecd | 4147 | blend_mode = V4QImode; |
2df9d3c5 UB |
4148 | } |
4149 | break; | |
820ac79e UB |
4150 | case E_V2QImode: |
4151 | if (TARGET_SSE4_1) | |
f4a2cecd | 4152 | gen = gen_mmx_pblendvb_v2qi; |
820ac79e | 4153 | break; |
2bf6d935 ML |
4154 | case E_V16QImode: |
4155 | case E_V8HImode: | |
9e2a82e1 | 4156 | case E_V8HFmode: |
6910cad5 | 4157 | case E_V8BFmode: |
2bf6d935 ML |
4158 | case E_V4SImode: |
4159 | case E_V2DImode: | |
793f847b | 4160 | case E_V1TImode: |
2bf6d935 ML |
4161 | if (TARGET_SSE4_1) |
4162 | { | |
4163 | gen = gen_sse4_1_pblendvb; | |
f4a2cecd | 4164 | blend_mode = V16QImode; |
2bf6d935 ML |
4165 | } |
4166 | break; | |
4167 | case E_V8SFmode: | |
4168 | if (TARGET_AVX) | |
4169 | gen = gen_avx_blendvps256; | |
4170 | break; | |
4171 | case E_V4DFmode: | |
4172 | if (TARGET_AVX) | |
4173 | gen = gen_avx_blendvpd256; | |
4174 | break; | |
4175 | case E_V32QImode: | |
4176 | case E_V16HImode: | |
9e2a82e1 | 4177 | case E_V16HFmode: |
6910cad5 | 4178 | case E_V16BFmode: |
2bf6d935 ML |
4179 | case E_V8SImode: |
4180 | case E_V4DImode: | |
4181 | if (TARGET_AVX2) | |
4182 | { | |
4183 | gen = gen_avx2_pblendvb; | |
f4a2cecd | 4184 | blend_mode = V32QImode; |
2bf6d935 ML |
4185 | } |
4186 | break; | |
4187 | ||
4188 | case E_V64QImode: | |
4189 | gen = gen_avx512bw_blendmv64qi; | |
4190 | break; | |
4191 | case E_V32HImode: | |
4192 | gen = gen_avx512bw_blendmv32hi; | |
4193 | break; | |
9e2a82e1 | 4194 | case E_V32HFmode: |
4195 | gen = gen_avx512bw_blendmv32hf; | |
4196 | break; | |
6910cad5 | 4197 | case E_V32BFmode: |
4198 | gen = gen_avx512bw_blendmv32bf; | |
4199 | break; | |
2bf6d935 ML |
4200 | case E_V16SImode: |
4201 | gen = gen_avx512f_blendmv16si; | |
4202 | break; | |
4203 | case E_V8DImode: | |
4204 | gen = gen_avx512f_blendmv8di; | |
4205 | break; | |
4206 | case E_V8DFmode: | |
4207 | gen = gen_avx512f_blendmv8df; | |
4208 | break; | |
4209 | case E_V16SFmode: | |
4210 | gen = gen_avx512f_blendmv16sf; | |
4211 | break; | |
4212 | ||
4213 | default: | |
4214 | break; | |
4215 | } | |
4216 | ||
4217 | if (gen != NULL) | |
4218 | { | |
f4a2cecd UB |
4219 | if (blend_mode == mode) |
4220 | x = dest; | |
4221 | else | |
4222 | { | |
4223 | x = gen_reg_rtx (blend_mode); | |
4224 | op_false = gen_lowpart (blend_mode, op_false); | |
4225 | op_true = gen_lowpart (blend_mode, op_true); | |
4226 | cmp = gen_lowpart (blend_mode, cmp); | |
4227 | } | |
4228 | ||
4229 | emit_insn (gen (x, op_false, op_true, cmp)); | |
4230 | ||
4231 | if (x != dest) | |
4232 | emit_move_insn (dest, gen_lowpart (mode, x)); | |
2bf6d935 ML |
4233 | } |
4234 | else | |
4235 | { | |
f4a2cecd | 4236 | rtx t2, t3; |
2bf6d935 | 4237 | |
f4a2cecd UB |
4238 | t2 = expand_simple_binop (mode, AND, op_true, cmp, |
4239 | NULL, 1, OPTAB_DIRECT); | |
2bf6d935 | 4240 | |
f4a2cecd | 4241 | t3 = gen_reg_rtx (mode); |
2bf6d935 | 4242 | x = gen_rtx_NOT (mode, cmp); |
b5193e35 | 4243 | ix86_emit_vec_binop (AND, mode, t3, x, op_false); |
2bf6d935 | 4244 | |
f4a2cecd UB |
4245 | x = expand_simple_binop (mode, IOR, t3, t2, |
4246 | dest, 1, OPTAB_DIRECT); | |
4247 | if (x != dest) | |
4248 | emit_move_insn (dest, x); | |
2bf6d935 ML |
4249 | } |
4250 | } | |
4251 | ||
4252 | /* Swap, force into registers, or otherwise massage the two operands | |
4253 | to an sse comparison with a mask result. Thus we differ a bit from | |
4254 | ix86_prepare_fp_compare_args which expects to produce a flags result. | |
4255 | ||
4256 | The DEST operand exists to help determine whether to commute commutative | |
4257 | operators. The POP0/POP1 operands are updated in place. The new | |
4258 | comparison code is returned, or UNKNOWN if not implementable. */ | |
4259 | ||
4260 | static enum rtx_code | |
4261 | ix86_prepare_sse_fp_compare_args (rtx dest, enum rtx_code code, | |
4262 | rtx *pop0, rtx *pop1) | |
4263 | { | |
4264 | switch (code) | |
4265 | { | |
4266 | case LTGT: | |
4267 | case UNEQ: | |
4268 | /* AVX supports all the needed comparisons. */ | |
4269 | if (TARGET_AVX) | |
4270 | break; | |
4271 | /* We have no LTGT as an operator. We could implement it with | |
4272 | NE & ORDERED, but this requires an extra temporary. It's | |
4273 | not clear that it's worth it. */ | |
4274 | return UNKNOWN; | |
4275 | ||
4276 | case LT: | |
4277 | case LE: | |
4278 | case UNGT: | |
4279 | case UNGE: | |
4280 | /* These are supported directly. */ | |
4281 | break; | |
4282 | ||
4283 | case EQ: | |
4284 | case NE: | |
4285 | case UNORDERED: | |
4286 | case ORDERED: | |
4287 | /* AVX has 3 operand comparisons, no need to swap anything. */ | |
4288 | if (TARGET_AVX) | |
4289 | break; | |
4290 | /* For commutative operators, try to canonicalize the destination | |
4291 | operand to be first in the comparison - this helps reload to | |
4292 | avoid extra moves. */ | |
4293 | if (!dest || !rtx_equal_p (dest, *pop1)) | |
4294 | break; | |
4295 | /* FALLTHRU */ | |
4296 | ||
4297 | case GE: | |
4298 | case GT: | |
4299 | case UNLE: | |
4300 | case UNLT: | |
4301 | /* These are not supported directly before AVX, and furthermore | |
4302 | ix86_expand_sse_fp_minmax only optimizes LT/UNGE. Swap the | |
4303 | comparison operands to transform into something that is | |
4304 | supported. */ | |
4305 | std::swap (*pop0, *pop1); | |
4306 | code = swap_condition (code); | |
4307 | break; | |
4308 | ||
4309 | default: | |
4310 | gcc_unreachable (); | |
4311 | } | |
4312 | ||
4313 | return code; | |
4314 | } | |
4315 | ||
4316 | /* Expand a floating-point conditional move. Return true if successful. */ | |
4317 | ||
4318 | bool | |
4319 | ix86_expand_fp_movcc (rtx operands[]) | |
4320 | { | |
4321 | machine_mode mode = GET_MODE (operands[0]); | |
4322 | enum rtx_code code = GET_CODE (operands[1]); | |
4323 | rtx tmp, compare_op; | |
4324 | rtx op0 = XEXP (operands[1], 0); | |
4325 | rtx op1 = XEXP (operands[1], 1); | |
4326 | ||
5792208f JJ |
4327 | if (GET_MODE (op0) == BFmode |
4328 | && !ix86_fp_comparison_operator (operands[1], VOIDmode)) | |
4329 | return false; | |
4330 | ||
a6841211 | 4331 | if (SSE_FLOAT_MODE_SSEMATH_OR_HF_P (mode)) |
2bf6d935 ML |
4332 | { |
4333 | machine_mode cmode; | |
4334 | ||
4335 | /* Since we've no cmove for sse registers, don't force bad register | |
4336 | allocation just to gain access to it. Deny movcc when the | |
4337 | comparison mode doesn't match the move mode. */ | |
4338 | cmode = GET_MODE (op0); | |
4339 | if (cmode == VOIDmode) | |
4340 | cmode = GET_MODE (op1); | |
4341 | if (cmode != mode) | |
4342 | return false; | |
4343 | ||
4344 | code = ix86_prepare_sse_fp_compare_args (operands[0], code, &op0, &op1); | |
4345 | if (code == UNKNOWN) | |
4346 | return false; | |
4347 | ||
4348 | if (ix86_expand_sse_fp_minmax (operands[0], code, op0, op1, | |
4349 | operands[2], operands[3])) | |
4350 | return true; | |
4351 | ||
4352 | tmp = ix86_expand_sse_cmp (operands[0], code, op0, op1, | |
4353 | operands[2], operands[3]); | |
4354 | ix86_expand_sse_movcc (operands[0], tmp, operands[2], operands[3]); | |
4355 | return true; | |
4356 | } | |
4357 | ||
4358 | if (GET_MODE (op0) == TImode | |
4359 | || (GET_MODE (op0) == DImode | |
4360 | && !TARGET_64BIT)) | |
4361 | return false; | |
4362 | ||
4363 | /* The floating point conditional move instructions don't directly | |
4364 | support conditions resulting from a signed integer comparison. */ | |
4365 | ||
4366 | compare_op = ix86_expand_compare (code, op0, op1); | |
4367 | if (!fcmov_comparison_operator (compare_op, VOIDmode)) | |
4368 | { | |
4369 | tmp = gen_reg_rtx (QImode); | |
4370 | ix86_expand_setcc (tmp, code, op0, op1); | |
4371 | ||
4372 | compare_op = ix86_expand_compare (NE, tmp, const0_rtx); | |
4373 | } | |
4374 | ||
4375 | emit_insn (gen_rtx_SET (operands[0], | |
4376 | gen_rtx_IF_THEN_ELSE (mode, compare_op, | |
4377 | operands[2], operands[3]))); | |
4378 | ||
4379 | return true; | |
4380 | } | |
4381 | ||
4382 | /* Helper for ix86_cmp_code_to_pcmp_immediate for int modes. */ | |
4383 | ||
4384 | static int | |
4385 | ix86_int_cmp_code_to_pcmp_immediate (enum rtx_code code) | |
4386 | { | |
4387 | switch (code) | |
4388 | { | |
4389 | case EQ: | |
4390 | return 0; | |
4391 | case LT: | |
4392 | case LTU: | |
4393 | return 1; | |
4394 | case LE: | |
4395 | case LEU: | |
4396 | return 2; | |
4397 | case NE: | |
4398 | return 4; | |
4399 | case GE: | |
4400 | case GEU: | |
4401 | return 5; | |
4402 | case GT: | |
4403 | case GTU: | |
4404 | return 6; | |
4405 | default: | |
4406 | gcc_unreachable (); | |
4407 | } | |
4408 | } | |
4409 | ||
4410 | /* Helper for ix86_cmp_code_to_pcmp_immediate for fp modes. */ | |
4411 | ||
4412 | static int | |
4413 | ix86_fp_cmp_code_to_pcmp_immediate (enum rtx_code code) | |
4414 | { | |
4415 | switch (code) | |
4416 | { | |
4417 | case EQ: | |
4418 | return 0x00; | |
4419 | case NE: | |
4420 | return 0x04; | |
4421 | case GT: | |
4422 | return 0x0e; | |
4423 | case LE: | |
4424 | return 0x02; | |
4425 | case GE: | |
4426 | return 0x0d; | |
4427 | case LT: | |
4428 | return 0x01; | |
4429 | case UNLE: | |
4430 | return 0x0a; | |
4431 | case UNLT: | |
4432 | return 0x09; | |
4433 | case UNGE: | |
4434 | return 0x05; | |
4435 | case UNGT: | |
4436 | return 0x06; | |
4437 | case UNEQ: | |
4438 | return 0x18; | |
4439 | case LTGT: | |
4440 | return 0x0c; | |
4441 | case ORDERED: | |
4442 | return 0x07; | |
4443 | case UNORDERED: | |
4444 | return 0x03; | |
4445 | default: | |
4446 | gcc_unreachable (); | |
4447 | } | |
4448 | } | |
4449 | ||
4450 | /* Return immediate value to be used in UNSPEC_PCMP | |
4451 | for comparison CODE in MODE. */ | |
4452 | ||
4453 | static int | |
4454 | ix86_cmp_code_to_pcmp_immediate (enum rtx_code code, machine_mode mode) | |
4455 | { | |
4456 | if (FLOAT_MODE_P (mode)) | |
4457 | return ix86_fp_cmp_code_to_pcmp_immediate (code); | |
4458 | return ix86_int_cmp_code_to_pcmp_immediate (code); | |
4459 | } | |
4460 | ||
4461 | /* Expand AVX-512 vector comparison. */ | |
4462 | ||
4463 | bool | |
99e4891e | 4464 | ix86_expand_mask_vec_cmp (rtx dest, enum rtx_code code, rtx cmp_op0, rtx cmp_op1) |
2bf6d935 | 4465 | { |
99e4891e | 4466 | machine_mode mask_mode = GET_MODE (dest); |
4467 | machine_mode cmp_mode = GET_MODE (cmp_op0); | |
2bf6d935 ML |
4468 | rtx imm = GEN_INT (ix86_cmp_code_to_pcmp_immediate (code, cmp_mode)); |
4469 | int unspec_code; | |
4470 | rtx unspec; | |
4471 | ||
4472 | switch (code) | |
4473 | { | |
4474 | case LEU: | |
4475 | case GTU: | |
4476 | case GEU: | |
4477 | case LTU: | |
4478 | unspec_code = UNSPEC_UNSIGNED_PCMP; | |
4479 | break; | |
4480 | ||
4481 | default: | |
4482 | unspec_code = UNSPEC_PCMP; | |
4483 | } | |
4484 | ||
99e4891e | 4485 | unspec = gen_rtx_UNSPEC (mask_mode, gen_rtvec (3, cmp_op0, cmp_op1, imm), |
2bf6d935 | 4486 | unspec_code); |
99e4891e | 4487 | emit_insn (gen_rtx_SET (dest, unspec)); |
2bf6d935 ML |
4488 | |
4489 | return true; | |
4490 | } | |
4491 | ||
4492 | /* Expand fp vector comparison. */ | |
4493 | ||
4494 | bool | |
4495 | ix86_expand_fp_vec_cmp (rtx operands[]) | |
4496 | { | |
4497 | enum rtx_code code = GET_CODE (operands[1]); | |
4498 | rtx cmp; | |
4499 | ||
4500 | code = ix86_prepare_sse_fp_compare_args (operands[0], code, | |
4501 | &operands[2], &operands[3]); | |
4502 | if (code == UNKNOWN) | |
4503 | { | |
4504 | rtx temp; | |
4505 | switch (GET_CODE (operands[1])) | |
4506 | { | |
4507 | case LTGT: | |
4508 | temp = ix86_expand_sse_cmp (operands[0], ORDERED, operands[2], | |
4509 | operands[3], NULL, NULL); | |
4510 | cmp = ix86_expand_sse_cmp (operands[0], NE, operands[2], | |
4511 | operands[3], NULL, NULL); | |
4512 | code = AND; | |
4513 | break; | |
4514 | case UNEQ: | |
4515 | temp = ix86_expand_sse_cmp (operands[0], UNORDERED, operands[2], | |
4516 | operands[3], NULL, NULL); | |
4517 | cmp = ix86_expand_sse_cmp (operands[0], EQ, operands[2], | |
4518 | operands[3], NULL, NULL); | |
4519 | code = IOR; | |
4520 | break; | |
4521 | default: | |
4522 | gcc_unreachable (); | |
4523 | } | |
4524 | cmp = expand_simple_binop (GET_MODE (cmp), code, temp, cmp, cmp, 1, | |
4525 | OPTAB_DIRECT); | |
4526 | } | |
4527 | else | |
4528 | cmp = ix86_expand_sse_cmp (operands[0], code, operands[2], operands[3], | |
8d0737d8 | 4529 | NULL, NULL); |
2bf6d935 ML |
4530 | |
4531 | if (operands[0] != cmp) | |
4532 | emit_move_insn (operands[0], cmp); | |
4533 | ||
4534 | return true; | |
4535 | } | |
4536 | ||
4537 | static rtx | |
4538 | ix86_expand_int_sse_cmp (rtx dest, enum rtx_code code, rtx cop0, rtx cop1, | |
4539 | rtx op_true, rtx op_false, bool *negate) | |
4540 | { | |
4541 | machine_mode data_mode = GET_MODE (dest); | |
4542 | machine_mode mode = GET_MODE (cop0); | |
4543 | rtx x; | |
4544 | ||
4545 | *negate = false; | |
4546 | ||
4547 | /* XOP supports all of the comparisons on all 128-bit vector int types. */ | |
4548 | if (TARGET_XOP | |
6c67afaf UB |
4549 | && GET_MODE_CLASS (mode) == MODE_VECTOR_INT |
4550 | && GET_MODE_SIZE (mode) <= 16) | |
2bf6d935 | 4551 | ; |
8b905e9b HL |
4552 | /* AVX512F supports all of the comparsions |
4553 | on all 128/256/512-bit vector int types. */ | |
8d0737d8 | 4554 | else if (ix86_use_mask_cmp_p (data_mode, mode, op_true, op_false)) |
8b905e9b | 4555 | ; |
2bf6d935 ML |
4556 | else |
4557 | { | |
4558 | /* Canonicalize the comparison to EQ, GT, GTU. */ | |
4559 | switch (code) | |
4560 | { | |
4561 | case EQ: | |
4562 | case GT: | |
4563 | case GTU: | |
4564 | break; | |
4565 | ||
2bf6d935 ML |
4566 | case LE: |
4567 | case LEU: | |
fa271afb JJ |
4568 | /* x <= cst can be handled as x < cst + 1 unless there is |
4569 | wrap around in cst + 1. */ | |
4570 | if (GET_CODE (cop1) == CONST_VECTOR | |
4571 | && GET_MODE_INNER (mode) != TImode) | |
4572 | { | |
4573 | unsigned int n_elts = GET_MODE_NUNITS (mode), i; | |
4574 | machine_mode eltmode = GET_MODE_INNER (mode); | |
4575 | for (i = 0; i < n_elts; ++i) | |
4576 | { | |
4577 | rtx elt = CONST_VECTOR_ELT (cop1, i); | |
4578 | if (!CONST_INT_P (elt)) | |
4579 | break; | |
4580 | if (code == GE) | |
4581 | { | |
4582 | /* For LE punt if some element is signed maximum. */ | |
4583 | if ((INTVAL (elt) & (GET_MODE_MASK (eltmode) >> 1)) | |
4584 | == (GET_MODE_MASK (eltmode) >> 1)) | |
4585 | break; | |
4586 | } | |
4587 | /* For LEU punt if some element is unsigned maximum. */ | |
4588 | else if (elt == constm1_rtx) | |
4589 | break; | |
4590 | } | |
4591 | if (i == n_elts) | |
4592 | { | |
4593 | rtvec v = rtvec_alloc (n_elts); | |
4594 | for (i = 0; i < n_elts; ++i) | |
4595 | RTVEC_ELT (v, i) | |
49de156a JJ |
4596 | = gen_int_mode (INTVAL (CONST_VECTOR_ELT (cop1, i)) + 1, |
4597 | eltmode); | |
fa271afb JJ |
4598 | cop1 = gen_rtx_CONST_VECTOR (mode, v); |
4599 | std::swap (cop0, cop1); | |
4600 | code = code == LE ? GT : GTU; | |
4601 | break; | |
4602 | } | |
4603 | } | |
4604 | /* FALLTHRU */ | |
4605 | case NE: | |
2bf6d935 ML |
4606 | code = reverse_condition (code); |
4607 | *negate = true; | |
4608 | break; | |
4609 | ||
4610 | case GE: | |
4611 | case GEU: | |
fa271afb JJ |
4612 | /* x >= cst can be handled as x > cst - 1 unless there is |
4613 | wrap around in cst - 1. */ | |
4614 | if (GET_CODE (cop1) == CONST_VECTOR | |
4615 | && GET_MODE_INNER (mode) != TImode) | |
4616 | { | |
4617 | unsigned int n_elts = GET_MODE_NUNITS (mode), i; | |
4618 | machine_mode eltmode = GET_MODE_INNER (mode); | |
4619 | for (i = 0; i < n_elts; ++i) | |
4620 | { | |
4621 | rtx elt = CONST_VECTOR_ELT (cop1, i); | |
4622 | if (!CONST_INT_P (elt)) | |
4623 | break; | |
4624 | if (code == GE) | |
4625 | { | |
4626 | /* For GE punt if some element is signed minimum. */ | |
4627 | if (INTVAL (elt) < 0 | |
4628 | && ((INTVAL (elt) & (GET_MODE_MASK (eltmode) >> 1)) | |
4629 | == 0)) | |
4630 | break; | |
4631 | } | |
4632 | /* For GEU punt if some element is zero. */ | |
4633 | else if (elt == const0_rtx) | |
4634 | break; | |
4635 | } | |
4636 | if (i == n_elts) | |
4637 | { | |
4638 | rtvec v = rtvec_alloc (n_elts); | |
4639 | for (i = 0; i < n_elts; ++i) | |
4640 | RTVEC_ELT (v, i) | |
49de156a JJ |
4641 | = gen_int_mode (INTVAL (CONST_VECTOR_ELT (cop1, i)) - 1, |
4642 | eltmode); | |
fa271afb JJ |
4643 | cop1 = gen_rtx_CONST_VECTOR (mode, v); |
4644 | code = code == GE ? GT : GTU; | |
4645 | break; | |
4646 | } | |
4647 | } | |
2bf6d935 ML |
4648 | code = reverse_condition (code); |
4649 | *negate = true; | |
4650 | /* FALLTHRU */ | |
4651 | ||
4652 | case LT: | |
4653 | case LTU: | |
4654 | std::swap (cop0, cop1); | |
4655 | code = swap_condition (code); | |
4656 | break; | |
4657 | ||
4658 | default: | |
4659 | gcc_unreachable (); | |
4660 | } | |
4661 | ||
4662 | /* Only SSE4.1/SSE4.2 supports V2DImode. */ | |
4663 | if (mode == V2DImode) | |
4664 | { | |
4665 | switch (code) | |
4666 | { | |
4667 | case EQ: | |
4668 | /* SSE4.1 supports EQ. */ | |
4669 | if (!TARGET_SSE4_1) | |
4670 | return NULL; | |
4671 | break; | |
4672 | ||
4673 | case GT: | |
4674 | case GTU: | |
4675 | /* SSE4.2 supports GT/GTU. */ | |
4676 | if (!TARGET_SSE4_2) | |
4677 | return NULL; | |
4678 | break; | |
4679 | ||
4680 | default: | |
4681 | gcc_unreachable (); | |
4682 | } | |
4683 | } | |
4684 | ||
fa271afb JJ |
4685 | if (GET_CODE (cop0) == CONST_VECTOR) |
4686 | cop0 = force_reg (mode, cop0); | |
4687 | else if (GET_CODE (cop1) == CONST_VECTOR) | |
4688 | cop1 = force_reg (mode, cop1); | |
4689 | ||
2bf6d935 ML |
4690 | rtx optrue = op_true ? op_true : CONSTM1_RTX (data_mode); |
4691 | rtx opfalse = op_false ? op_false : CONST0_RTX (data_mode); | |
4692 | if (*negate) | |
4693 | std::swap (optrue, opfalse); | |
4694 | ||
4695 | /* Transform x > y ? 0 : -1 (i.e. x <= y ? -1 : 0 or x <= y) when | |
4696 | not using integer masks into min (x, y) == x ? -1 : 0 (i.e. | |
4697 | min (x, y) == x). While we add one instruction (the minimum), | |
4698 | we remove the need for two instructions in the negation, as the | |
4699 | result is done this way. | |
4700 | When using masks, do it for SI/DImode element types, as it is shorter | |
4701 | than the two subtractions. */ | |
4702 | if ((code != EQ | |
4703 | && GET_MODE_SIZE (mode) != 64 | |
4704 | && vector_all_ones_operand (opfalse, data_mode) | |
4705 | && optrue == CONST0_RTX (data_mode)) | |
4706 | || (code == GTU | |
4707 | && GET_MODE_SIZE (GET_MODE_INNER (mode)) >= 4 | |
4708 | /* Don't do it if not using integer masks and we'd end up with | |
4709 | the right values in the registers though. */ | |
4710 | && (GET_MODE_SIZE (mode) == 64 | |
4711 | || !vector_all_ones_operand (optrue, data_mode) | |
4712 | || opfalse != CONST0_RTX (data_mode)))) | |
4713 | { | |
4714 | rtx (*gen) (rtx, rtx, rtx) = NULL; | |
4715 | ||
4716 | switch (mode) | |
4717 | { | |
4718 | case E_V16SImode: | |
4719 | gen = (code == GTU) ? gen_uminv16si3 : gen_sminv16si3; | |
4720 | break; | |
4721 | case E_V8DImode: | |
4722 | gen = (code == GTU) ? gen_uminv8di3 : gen_sminv8di3; | |
4723 | cop0 = force_reg (mode, cop0); | |
4724 | cop1 = force_reg (mode, cop1); | |
4725 | break; | |
4726 | case E_V32QImode: | |
4727 | if (TARGET_AVX2) | |
4728 | gen = (code == GTU) ? gen_uminv32qi3 : gen_sminv32qi3; | |
4729 | break; | |
4730 | case E_V16HImode: | |
4731 | if (TARGET_AVX2) | |
4732 | gen = (code == GTU) ? gen_uminv16hi3 : gen_sminv16hi3; | |
4733 | break; | |
4734 | case E_V8SImode: | |
4735 | if (TARGET_AVX2) | |
4736 | gen = (code == GTU) ? gen_uminv8si3 : gen_sminv8si3; | |
4737 | break; | |
4738 | case E_V4DImode: | |
4739 | if (TARGET_AVX512VL) | |
4740 | { | |
4741 | gen = (code == GTU) ? gen_uminv4di3 : gen_sminv4di3; | |
4742 | cop0 = force_reg (mode, cop0); | |
4743 | cop1 = force_reg (mode, cop1); | |
4744 | } | |
4745 | break; | |
4746 | case E_V16QImode: | |
4747 | if (code == GTU && TARGET_SSE2) | |
4748 | gen = gen_uminv16qi3; | |
4749 | else if (code == GT && TARGET_SSE4_1) | |
4750 | gen = gen_sminv16qi3; | |
4751 | break; | |
f3661f2d UB |
4752 | case E_V8QImode: |
4753 | if (code == GTU && TARGET_SSE2) | |
4754 | gen = gen_uminv8qi3; | |
4755 | else if (code == GT && TARGET_SSE4_1) | |
4756 | gen = gen_sminv8qi3; | |
4757 | break; | |
2df9d3c5 UB |
4758 | case E_V4QImode: |
4759 | if (code == GTU && TARGET_SSE2) | |
4760 | gen = gen_uminv4qi3; | |
4761 | else if (code == GT && TARGET_SSE4_1) | |
4762 | gen = gen_sminv4qi3; | |
4763 | break; | |
04a74555 UB |
4764 | case E_V2QImode: |
4765 | if (code == GTU && TARGET_SSE2) | |
4766 | gen = gen_uminv2qi3; | |
4767 | else if (code == GT && TARGET_SSE4_1) | |
4768 | gen = gen_sminv2qi3; | |
4769 | break; | |
2bf6d935 ML |
4770 | case E_V8HImode: |
4771 | if (code == GTU && TARGET_SSE4_1) | |
4772 | gen = gen_uminv8hi3; | |
4773 | else if (code == GT && TARGET_SSE2) | |
4774 | gen = gen_sminv8hi3; | |
4775 | break; | |
f3661f2d UB |
4776 | case E_V4HImode: |
4777 | if (code == GTU && TARGET_SSE4_1) | |
4778 | gen = gen_uminv4hi3; | |
4779 | else if (code == GT && TARGET_SSE2) | |
4780 | gen = gen_sminv4hi3; | |
4781 | break; | |
2df9d3c5 UB |
4782 | case E_V2HImode: |
4783 | if (code == GTU && TARGET_SSE4_1) | |
4784 | gen = gen_uminv2hi3; | |
4785 | else if (code == GT && TARGET_SSE2) | |
4786 | gen = gen_sminv2hi3; | |
4787 | break; | |
2bf6d935 ML |
4788 | case E_V4SImode: |
4789 | if (TARGET_SSE4_1) | |
4790 | gen = (code == GTU) ? gen_uminv4si3 : gen_sminv4si3; | |
4791 | break; | |
f3661f2d UB |
4792 | case E_V2SImode: |
4793 | if (TARGET_SSE4_1) | |
4794 | gen = (code == GTU) ? gen_uminv2si3 : gen_sminv2si3; | |
4795 | break; | |
2bf6d935 ML |
4796 | case E_V2DImode: |
4797 | if (TARGET_AVX512VL) | |
4798 | { | |
4799 | gen = (code == GTU) ? gen_uminv2di3 : gen_sminv2di3; | |
4800 | cop0 = force_reg (mode, cop0); | |
4801 | cop1 = force_reg (mode, cop1); | |
4802 | } | |
4803 | break; | |
4804 | default: | |
4805 | break; | |
4806 | } | |
4807 | ||
4808 | if (gen) | |
4809 | { | |
4810 | rtx tem = gen_reg_rtx (mode); | |
4811 | if (!vector_operand (cop0, mode)) | |
4812 | cop0 = force_reg (mode, cop0); | |
4813 | if (!vector_operand (cop1, mode)) | |
4814 | cop1 = force_reg (mode, cop1); | |
4815 | *negate = !*negate; | |
4816 | emit_insn (gen (tem, cop0, cop1)); | |
4817 | cop1 = tem; | |
4818 | code = EQ; | |
4819 | } | |
4820 | } | |
4821 | ||
4822 | /* Unsigned parallel compare is not supported by the hardware. | |
4823 | Play some tricks to turn this into a signed comparison | |
4824 | against 0. */ | |
4825 | if (code == GTU) | |
4826 | { | |
4827 | cop0 = force_reg (mode, cop0); | |
4828 | ||
4829 | switch (mode) | |
4830 | { | |
4831 | case E_V16SImode: | |
4832 | case E_V8DImode: | |
4833 | case E_V8SImode: | |
4834 | case E_V4DImode: | |
4835 | case E_V4SImode: | |
f3661f2d | 4836 | case E_V2SImode: |
2bf6d935 ML |
4837 | case E_V2DImode: |
4838 | { | |
4839 | rtx t1, t2, mask; | |
83bc5e44 | 4840 | |
2bf6d935 ML |
4841 | /* Subtract (-(INT MAX) - 1) from both operands to make |
4842 | them signed. */ | |
4843 | mask = ix86_build_signbit_mask (mode, true, false); | |
4844 | t1 = gen_reg_rtx (mode); | |
83bc5e44 | 4845 | emit_insn (gen_sub3_insn (t1, cop0, mask)); |
2bf6d935 ML |
4846 | |
4847 | t2 = gen_reg_rtx (mode); | |
83bc5e44 | 4848 | emit_insn (gen_sub3_insn (t2, cop1, mask)); |
2bf6d935 ML |
4849 | |
4850 | cop0 = t1; | |
4851 | cop1 = t2; | |
4852 | code = GT; | |
4853 | } | |
4854 | break; | |
4855 | ||
4856 | case E_V64QImode: | |
4857 | case E_V32HImode: | |
4858 | case E_V32QImode: | |
4859 | case E_V16HImode: | |
4860 | case E_V16QImode: | |
f3661f2d | 4861 | case E_V8QImode: |
2df9d3c5 | 4862 | case E_V4QImode: |
04a74555 | 4863 | case E_V2QImode: |
2bf6d935 | 4864 | case E_V8HImode: |
f3661f2d | 4865 | case E_V4HImode: |
2df9d3c5 | 4866 | case E_V2HImode: |
2bf6d935 ML |
4867 | /* Perform a parallel unsigned saturating subtraction. */ |
4868 | x = gen_reg_rtx (mode); | |
83bc5e44 UB |
4869 | emit_insn (gen_rtx_SET |
4870 | (x, gen_rtx_US_MINUS (mode, cop0, cop1))); | |
2bf6d935 ML |
4871 | cop0 = x; |
4872 | cop1 = CONST0_RTX (mode); | |
4873 | code = EQ; | |
4874 | *negate = !*negate; | |
4875 | break; | |
4876 | ||
4877 | default: | |
4878 | gcc_unreachable (); | |
4879 | } | |
4880 | } | |
4881 | } | |
4882 | ||
4883 | if (*negate) | |
4884 | std::swap (op_true, op_false); | |
4885 | ||
fa271afb JJ |
4886 | if (GET_CODE (cop1) == CONST_VECTOR) |
4887 | cop1 = force_reg (mode, cop1); | |
4888 | ||
2bf6d935 ML |
4889 | /* Allow the comparison to be done in one mode, but the movcc to |
4890 | happen in another mode. */ | |
4891 | if (data_mode == mode) | |
fa271afb | 4892 | x = ix86_expand_sse_cmp (dest, code, cop0, cop1, op_true, op_false); |
2bf6d935 ML |
4893 | else |
4894 | { | |
4895 | gcc_assert (GET_MODE_SIZE (data_mode) == GET_MODE_SIZE (mode)); | |
4896 | x = ix86_expand_sse_cmp (gen_reg_rtx (mode), code, cop0, cop1, | |
4897 | op_true, op_false); | |
4898 | if (GET_MODE (x) == mode) | |
4899 | x = gen_lowpart (data_mode, x); | |
4900 | } | |
4901 | ||
4902 | return x; | |
4903 | } | |
4904 | ||
4905 | /* Expand integer vector comparison. */ | |
4906 | ||
4907 | bool | |
4908 | ix86_expand_int_vec_cmp (rtx operands[]) | |
4909 | { | |
4910 | rtx_code code = GET_CODE (operands[1]); | |
4911 | bool negate = false; | |
4912 | rtx cmp = ix86_expand_int_sse_cmp (operands[0], code, operands[2], | |
4913 | operands[3], NULL, NULL, &negate); | |
4914 | ||
4915 | if (!cmp) | |
4916 | return false; | |
4917 | ||
4918 | if (negate) | |
4919 | cmp = ix86_expand_int_sse_cmp (operands[0], EQ, cmp, | |
4920 | CONST0_RTX (GET_MODE (cmp)), | |
4921 | NULL, NULL, &negate); | |
4922 | ||
4923 | gcc_assert (!negate); | |
4924 | ||
4925 | if (operands[0] != cmp) | |
4926 | emit_move_insn (operands[0], cmp); | |
4927 | ||
4928 | return true; | |
4929 | } | |
4930 | ||
4931 | /* Expand a floating-point vector conditional move; a vcond operation | |
4932 | rather than a movcc operation. */ | |
4933 | ||
4934 | bool | |
4935 | ix86_expand_fp_vcond (rtx operands[]) | |
4936 | { | |
4937 | enum rtx_code code = GET_CODE (operands[3]); | |
4938 | rtx cmp; | |
4939 | ||
4940 | code = ix86_prepare_sse_fp_compare_args (operands[0], code, | |
4941 | &operands[4], &operands[5]); | |
4942 | if (code == UNKNOWN) | |
4943 | { | |
4944 | rtx temp; | |
4945 | switch (GET_CODE (operands[3])) | |
4946 | { | |
4947 | case LTGT: | |
4948 | temp = ix86_expand_sse_cmp (operands[0], ORDERED, operands[4], | |
4949 | operands[5], operands[0], operands[0]); | |
4950 | cmp = ix86_expand_sse_cmp (operands[0], NE, operands[4], | |
4951 | operands[5], operands[1], operands[2]); | |
4952 | code = AND; | |
4953 | break; | |
4954 | case UNEQ: | |
4955 | temp = ix86_expand_sse_cmp (operands[0], UNORDERED, operands[4], | |
4956 | operands[5], operands[0], operands[0]); | |
4957 | cmp = ix86_expand_sse_cmp (operands[0], EQ, operands[4], | |
4958 | operands[5], operands[1], operands[2]); | |
4959 | code = IOR; | |
4960 | break; | |
4961 | default: | |
4962 | gcc_unreachable (); | |
4963 | } | |
4964 | cmp = expand_simple_binop (GET_MODE (cmp), code, temp, cmp, cmp, 1, | |
4965 | OPTAB_DIRECT); | |
4966 | ix86_expand_sse_movcc (operands[0], cmp, operands[1], operands[2]); | |
4967 | return true; | |
4968 | } | |
4969 | ||
4970 | if (ix86_expand_sse_fp_minmax (operands[0], code, operands[4], | |
4971 | operands[5], operands[1], operands[2])) | |
4972 | return true; | |
4973 | ||
4974 | cmp = ix86_expand_sse_cmp (operands[0], code, operands[4], operands[5], | |
4975 | operands[1], operands[2]); | |
4976 | ix86_expand_sse_movcc (operands[0], cmp, operands[1], operands[2]); | |
4977 | return true; | |
4978 | } | |
4979 | ||
4980 | /* Expand a signed/unsigned integral vector conditional move. */ | |
4981 | ||
4982 | bool | |
4983 | ix86_expand_int_vcond (rtx operands[]) | |
4984 | { | |
4985 | machine_mode data_mode = GET_MODE (operands[0]); | |
4986 | machine_mode mode = GET_MODE (operands[4]); | |
4987 | enum rtx_code code = GET_CODE (operands[3]); | |
4988 | bool negate = false; | |
4989 | rtx x, cop0, cop1; | |
4990 | ||
4991 | cop0 = operands[4]; | |
4992 | cop1 = operands[5]; | |
4993 | ||
4994 | /* Try to optimize x < 0 ? -1 : 0 into (signed) x >> 31 | |
4995 | and x < 0 ? 1 : 0 into (unsigned) x >> 31. */ | |
4996 | if ((code == LT || code == GE) | |
4997 | && data_mode == mode | |
4998 | && cop1 == CONST0_RTX (mode) | |
4999 | && operands[1 + (code == LT)] == CONST0_RTX (data_mode) | |
5000 | && GET_MODE_UNIT_SIZE (data_mode) > 1 | |
5001 | && GET_MODE_UNIT_SIZE (data_mode) <= 8 | |
5002 | && (GET_MODE_SIZE (data_mode) == 16 | |
5003 | || (TARGET_AVX2 && GET_MODE_SIZE (data_mode) == 32))) | |
5004 | { | |
5005 | rtx negop = operands[2 - (code == LT)]; | |
5006 | int shift = GET_MODE_UNIT_BITSIZE (data_mode) - 1; | |
5007 | if (negop == CONST1_RTX (data_mode)) | |
5008 | { | |
5009 | rtx res = expand_simple_binop (mode, LSHIFTRT, cop0, GEN_INT (shift), | |
5010 | operands[0], 1, OPTAB_DIRECT); | |
5011 | if (res != operands[0]) | |
5012 | emit_move_insn (operands[0], res); | |
5013 | return true; | |
5014 | } | |
5015 | else if (GET_MODE_INNER (data_mode) != DImode | |
5016 | && vector_all_ones_operand (negop, data_mode)) | |
5017 | { | |
5018 | rtx res = expand_simple_binop (mode, ASHIFTRT, cop0, GEN_INT (shift), | |
5019 | operands[0], 0, OPTAB_DIRECT); | |
5020 | if (res != operands[0]) | |
5021 | emit_move_insn (operands[0], res); | |
5022 | return true; | |
5023 | } | |
5024 | } | |
5025 | ||
5026 | if (!nonimmediate_operand (cop1, mode)) | |
5027 | cop1 = force_reg (mode, cop1); | |
5028 | if (!general_operand (operands[1], data_mode)) | |
5029 | operands[1] = force_reg (data_mode, operands[1]); | |
5030 | if (!general_operand (operands[2], data_mode)) | |
5031 | operands[2] = force_reg (data_mode, operands[2]); | |
5032 | ||
5033 | x = ix86_expand_int_sse_cmp (operands[0], code, cop0, cop1, | |
5034 | operands[1], operands[2], &negate); | |
5035 | ||
5036 | if (!x) | |
5037 | return false; | |
5038 | ||
5039 | ix86_expand_sse_movcc (operands[0], x, operands[1+negate], | |
5040 | operands[2-negate]); | |
5041 | return true; | |
5042 | } | |
5043 | ||
5044 | static bool | |
5045 | ix86_expand_vec_perm_vpermt2 (rtx target, rtx mask, rtx op0, rtx op1, | |
5046 | struct expand_vec_perm_d *d) | |
5047 | { | |
5048 | /* ix86_expand_vec_perm_vpermt2 is called from both const and non-const | |
5049 | expander, so args are either in d, or in op0, op1 etc. */ | |
5050 | machine_mode mode = GET_MODE (d ? d->op0 : op0); | |
5051 | machine_mode maskmode = mode; | |
5052 | rtx (*gen) (rtx, rtx, rtx, rtx) = NULL; | |
5053 | ||
5054 | switch (mode) | |
5055 | { | |
faf2b6bc | 5056 | case E_V16QImode: |
5057 | if (TARGET_AVX512VL && TARGET_AVX512VBMI) | |
5058 | gen = gen_avx512vl_vpermt2varv16qi3; | |
5059 | break; | |
5060 | case E_V32QImode: | |
5061 | if (TARGET_AVX512VL && TARGET_AVX512VBMI) | |
5062 | gen = gen_avx512vl_vpermt2varv32qi3; | |
5063 | break; | |
5064 | case E_V64QImode: | |
5065 | if (TARGET_AVX512VBMI) | |
5066 | gen = gen_avx512bw_vpermt2varv64qi3; | |
5067 | break; | |
2bf6d935 ML |
5068 | case E_V8HImode: |
5069 | if (TARGET_AVX512VL && TARGET_AVX512BW) | |
5070 | gen = gen_avx512vl_vpermt2varv8hi3; | |
5071 | break; | |
5072 | case E_V16HImode: | |
5073 | if (TARGET_AVX512VL && TARGET_AVX512BW) | |
5074 | gen = gen_avx512vl_vpermt2varv16hi3; | |
5075 | break; | |
2bf6d935 ML |
5076 | case E_V32HImode: |
5077 | if (TARGET_AVX512BW) | |
5078 | gen = gen_avx512bw_vpermt2varv32hi3; | |
5079 | break; | |
5080 | case E_V4SImode: | |
5081 | if (TARGET_AVX512VL) | |
5082 | gen = gen_avx512vl_vpermt2varv4si3; | |
5083 | break; | |
5084 | case E_V8SImode: | |
5085 | if (TARGET_AVX512VL) | |
5086 | gen = gen_avx512vl_vpermt2varv8si3; | |
5087 | break; | |
5088 | case E_V16SImode: | |
5089 | if (TARGET_AVX512F) | |
5090 | gen = gen_avx512f_vpermt2varv16si3; | |
5091 | break; | |
5092 | case E_V4SFmode: | |
5093 | if (TARGET_AVX512VL) | |
5094 | { | |
5095 | gen = gen_avx512vl_vpermt2varv4sf3; | |
5096 | maskmode = V4SImode; | |
5097 | } | |
5098 | break; | |
5099 | case E_V8SFmode: | |
5100 | if (TARGET_AVX512VL) | |
5101 | { | |
5102 | gen = gen_avx512vl_vpermt2varv8sf3; | |
5103 | maskmode = V8SImode; | |
5104 | } | |
5105 | break; | |
5106 | case E_V16SFmode: | |
5107 | if (TARGET_AVX512F) | |
5108 | { | |
5109 | gen = gen_avx512f_vpermt2varv16sf3; | |
5110 | maskmode = V16SImode; | |
5111 | } | |
5112 | break; | |
5113 | case E_V2DImode: | |
5114 | if (TARGET_AVX512VL) | |
5115 | gen = gen_avx512vl_vpermt2varv2di3; | |
5116 | break; | |
5117 | case E_V4DImode: | |
5118 | if (TARGET_AVX512VL) | |
5119 | gen = gen_avx512vl_vpermt2varv4di3; | |
5120 | break; | |
5121 | case E_V8DImode: | |
5122 | if (TARGET_AVX512F) | |
5123 | gen = gen_avx512f_vpermt2varv8di3; | |
5124 | break; | |
5125 | case E_V2DFmode: | |
5126 | if (TARGET_AVX512VL) | |
5127 | { | |
5128 | gen = gen_avx512vl_vpermt2varv2df3; | |
5129 | maskmode = V2DImode; | |
5130 | } | |
5131 | break; | |
5132 | case E_V4DFmode: | |
5133 | if (TARGET_AVX512VL) | |
5134 | { | |
5135 | gen = gen_avx512vl_vpermt2varv4df3; | |
5136 | maskmode = V4DImode; | |
5137 | } | |
5138 | break; | |
5139 | case E_V8DFmode: | |
5140 | if (TARGET_AVX512F) | |
5141 | { | |
5142 | gen = gen_avx512f_vpermt2varv8df3; | |
5143 | maskmode = V8DImode; | |
5144 | } | |
5145 | break; | |
5146 | default: | |
5147 | break; | |
5148 | } | |
5149 | ||
5150 | if (gen == NULL) | |
5151 | return false; | |
5152 | ||
5153 | /* ix86_expand_vec_perm_vpermt2 is called from both const and non-const | |
5154 | expander, so args are either in d, or in op0, op1 etc. */ | |
5155 | if (d) | |
5156 | { | |
5157 | rtx vec[64]; | |
5158 | target = d->target; | |
5159 | op0 = d->op0; | |
5160 | op1 = d->op1; | |
5161 | for (int i = 0; i < d->nelt; ++i) | |
5162 | vec[i] = GEN_INT (d->perm[i]); | |
5163 | mask = gen_rtx_CONST_VECTOR (maskmode, gen_rtvec_v (d->nelt, vec)); | |
5164 | } | |
5165 | ||
5166 | emit_insn (gen (target, force_reg (maskmode, mask), op0, op1)); | |
5167 | return true; | |
5168 | } | |
5169 | ||
5170 | /* Expand a variable vector permutation. */ | |
5171 | ||
5172 | void | |
5173 | ix86_expand_vec_perm (rtx operands[]) | |
5174 | { | |
5175 | rtx target = operands[0]; | |
5176 | rtx op0 = operands[1]; | |
5177 | rtx op1 = operands[2]; | |
5178 | rtx mask = operands[3]; | |
5179 | rtx t1, t2, t3, t4, t5, t6, t7, t8, vt, vt2, vec[32]; | |
5180 | machine_mode mode = GET_MODE (op0); | |
5181 | machine_mode maskmode = GET_MODE (mask); | |
5182 | int w, e, i; | |
5183 | bool one_operand_shuffle = rtx_equal_p (op0, op1); | |
5184 | ||
5185 | /* Number of elements in the vector. */ | |
5186 | w = GET_MODE_NUNITS (mode); | |
5187 | e = GET_MODE_UNIT_SIZE (mode); | |
5188 | gcc_assert (w <= 64); | |
5189 | ||
be072bfa HW |
5190 | /* For HF mode vector, convert it to HI using subreg. */ |
5191 | if (GET_MODE_INNER (mode) == HFmode) | |
5192 | { | |
5193 | machine_mode orig_mode = mode; | |
5194 | mode = mode_for_vector (HImode, w).require (); | |
5195 | target = lowpart_subreg (mode, target, orig_mode); | |
5196 | op0 = lowpart_subreg (mode, op0, orig_mode); | |
5197 | op1 = lowpart_subreg (mode, op1, orig_mode); | |
5198 | } | |
5199 | ||
2bf6d935 ML |
5200 | if (TARGET_AVX512F && one_operand_shuffle) |
5201 | { | |
5202 | rtx (*gen) (rtx, rtx, rtx) = NULL; | |
5203 | switch (mode) | |
5204 | { | |
5205 | case E_V16SImode: | |
5206 | gen =gen_avx512f_permvarv16si; | |
5207 | break; | |
5208 | case E_V16SFmode: | |
5209 | gen = gen_avx512f_permvarv16sf; | |
5210 | break; | |
5211 | case E_V8DImode: | |
5212 | gen = gen_avx512f_permvarv8di; | |
5213 | break; | |
5214 | case E_V8DFmode: | |
5215 | gen = gen_avx512f_permvarv8df; | |
5216 | break; | |
5217 | default: | |
5218 | break; | |
5219 | } | |
5220 | if (gen != NULL) | |
5221 | { | |
5222 | emit_insn (gen (target, op0, mask)); | |
5223 | return; | |
5224 | } | |
5225 | } | |
5226 | ||
5227 | if (ix86_expand_vec_perm_vpermt2 (target, mask, op0, op1, NULL)) | |
5228 | return; | |
5229 | ||
5230 | if (TARGET_AVX2) | |
5231 | { | |
5232 | if (mode == V4DImode || mode == V4DFmode || mode == V16HImode) | |
5233 | { | |
5234 | /* Unfortunately, the VPERMQ and VPERMPD instructions only support | |
5235 | an constant shuffle operand. With a tiny bit of effort we can | |
5236 | use VPERMD instead. A re-interpretation stall for V4DFmode is | |
5237 | unfortunate but there's no avoiding it. | |
5238 | Similarly for V16HImode we don't have instructions for variable | |
5239 | shuffling, while for V32QImode we can use after preparing suitable | |
5240 | masks vpshufb; vpshufb; vpermq; vpor. */ | |
5241 | ||
5242 | if (mode == V16HImode) | |
5243 | { | |
5244 | maskmode = mode = V32QImode; | |
5245 | w = 32; | |
5246 | e = 1; | |
5247 | } | |
5248 | else | |
5249 | { | |
5250 | maskmode = mode = V8SImode; | |
5251 | w = 8; | |
5252 | e = 4; | |
5253 | } | |
5254 | t1 = gen_reg_rtx (maskmode); | |
5255 | ||
5256 | /* Replicate the low bits of the V4DImode mask into V8SImode: | |
5257 | mask = { A B C D } | |
5258 | t1 = { A A B B C C D D }. */ | |
5259 | for (i = 0; i < w / 2; ++i) | |
5260 | vec[i*2 + 1] = vec[i*2] = GEN_INT (i * 2); | |
5261 | vt = gen_rtx_CONST_VECTOR (maskmode, gen_rtvec_v (w, vec)); | |
5262 | vt = force_reg (maskmode, vt); | |
5263 | mask = gen_lowpart (maskmode, mask); | |
5264 | if (maskmode == V8SImode) | |
5265 | emit_insn (gen_avx2_permvarv8si (t1, mask, vt)); | |
5266 | else | |
5267 | emit_insn (gen_avx2_pshufbv32qi3 (t1, mask, vt)); | |
5268 | ||
5269 | /* Multiply the shuffle indicies by two. */ | |
5270 | t1 = expand_simple_binop (maskmode, PLUS, t1, t1, t1, 1, | |
5271 | OPTAB_DIRECT); | |
5272 | ||
5273 | /* Add one to the odd shuffle indicies: | |
5274 | t1 = { A*2, A*2+1, B*2, B*2+1, ... }. */ | |
5275 | for (i = 0; i < w / 2; ++i) | |
5276 | { | |
5277 | vec[i * 2] = const0_rtx; | |
5278 | vec[i * 2 + 1] = const1_rtx; | |
5279 | } | |
5280 | vt = gen_rtx_CONST_VECTOR (maskmode, gen_rtvec_v (w, vec)); | |
5281 | vt = validize_mem (force_const_mem (maskmode, vt)); | |
5282 | t1 = expand_simple_binop (maskmode, PLUS, t1, vt, t1, 1, | |
5283 | OPTAB_DIRECT); | |
5284 | ||
5285 | /* Continue as if V8SImode (resp. V32QImode) was used initially. */ | |
5286 | operands[3] = mask = t1; | |
5287 | target = gen_reg_rtx (mode); | |
5288 | op0 = gen_lowpart (mode, op0); | |
5289 | op1 = gen_lowpart (mode, op1); | |
5290 | } | |
5291 | ||
5292 | switch (mode) | |
5293 | { | |
5294 | case E_V8SImode: | |
5295 | /* The VPERMD and VPERMPS instructions already properly ignore | |
5296 | the high bits of the shuffle elements. No need for us to | |
5297 | perform an AND ourselves. */ | |
5298 | if (one_operand_shuffle) | |
5299 | { | |
5300 | emit_insn (gen_avx2_permvarv8si (target, op0, mask)); | |
5301 | if (target != operands[0]) | |
5302 | emit_move_insn (operands[0], | |
5303 | gen_lowpart (GET_MODE (operands[0]), target)); | |
5304 | } | |
5305 | else | |
5306 | { | |
5307 | t1 = gen_reg_rtx (V8SImode); | |
5308 | t2 = gen_reg_rtx (V8SImode); | |
5309 | emit_insn (gen_avx2_permvarv8si (t1, op0, mask)); | |
5310 | emit_insn (gen_avx2_permvarv8si (t2, op1, mask)); | |
5311 | goto merge_two; | |
5312 | } | |
5313 | return; | |
5314 | ||
5315 | case E_V8SFmode: | |
5316 | mask = gen_lowpart (V8SImode, mask); | |
5317 | if (one_operand_shuffle) | |
5318 | emit_insn (gen_avx2_permvarv8sf (target, op0, mask)); | |
5319 | else | |
5320 | { | |
5321 | t1 = gen_reg_rtx (V8SFmode); | |
5322 | t2 = gen_reg_rtx (V8SFmode); | |
5323 | emit_insn (gen_avx2_permvarv8sf (t1, op0, mask)); | |
5324 | emit_insn (gen_avx2_permvarv8sf (t2, op1, mask)); | |
5325 | goto merge_two; | |
5326 | } | |
5327 | return; | |
5328 | ||
5329 | case E_V4SImode: | |
5330 | /* By combining the two 128-bit input vectors into one 256-bit | |
5331 | input vector, we can use VPERMD and VPERMPS for the full | |
5332 | two-operand shuffle. */ | |
5333 | t1 = gen_reg_rtx (V8SImode); | |
5334 | t2 = gen_reg_rtx (V8SImode); | |
5335 | emit_insn (gen_avx_vec_concatv8si (t1, op0, op1)); | |
5336 | emit_insn (gen_avx_vec_concatv8si (t2, mask, mask)); | |
5337 | emit_insn (gen_avx2_permvarv8si (t1, t1, t2)); | |
5338 | emit_insn (gen_avx_vextractf128v8si (target, t1, const0_rtx)); | |
5339 | return; | |
5340 | ||
5341 | case E_V4SFmode: | |
5342 | t1 = gen_reg_rtx (V8SFmode); | |
5343 | t2 = gen_reg_rtx (V8SImode); | |
5344 | mask = gen_lowpart (V4SImode, mask); | |
5345 | emit_insn (gen_avx_vec_concatv8sf (t1, op0, op1)); | |
5346 | emit_insn (gen_avx_vec_concatv8si (t2, mask, mask)); | |
5347 | emit_insn (gen_avx2_permvarv8sf (t1, t1, t2)); | |
5348 | emit_insn (gen_avx_vextractf128v8sf (target, t1, const0_rtx)); | |
5349 | return; | |
5350 | ||
5351 | case E_V32QImode: | |
5352 | t1 = gen_reg_rtx (V32QImode); | |
5353 | t2 = gen_reg_rtx (V32QImode); | |
5354 | t3 = gen_reg_rtx (V32QImode); | |
5355 | vt2 = GEN_INT (-128); | |
5356 | vt = gen_const_vec_duplicate (V32QImode, vt2); | |
5357 | vt = force_reg (V32QImode, vt); | |
5358 | for (i = 0; i < 32; i++) | |
5359 | vec[i] = i < 16 ? vt2 : const0_rtx; | |
5360 | vt2 = gen_rtx_CONST_VECTOR (V32QImode, gen_rtvec_v (32, vec)); | |
5361 | vt2 = force_reg (V32QImode, vt2); | |
5362 | /* From mask create two adjusted masks, which contain the same | |
5363 | bits as mask in the low 7 bits of each vector element. | |
5364 | The first mask will have the most significant bit clear | |
5365 | if it requests element from the same 128-bit lane | |
5366 | and MSB set if it requests element from the other 128-bit lane. | |
5367 | The second mask will have the opposite values of the MSB, | |
5368 | and additionally will have its 128-bit lanes swapped. | |
5369 | E.g. { 07 12 1e 09 ... | 17 19 05 1f ... } mask vector will have | |
5370 | t1 { 07 92 9e 09 ... | 17 19 85 1f ... } and | |
5371 | t3 { 97 99 05 9f ... | 87 12 1e 89 ... } where each ... | |
5372 | stands for other 12 bytes. */ | |
5373 | /* The bit whether element is from the same lane or the other | |
5374 | lane is bit 4, so shift it up by 3 to the MSB position. */ | |
5375 | t5 = gen_reg_rtx (V4DImode); | |
5376 | emit_insn (gen_ashlv4di3 (t5, gen_lowpart (V4DImode, mask), | |
5377 | GEN_INT (3))); | |
5378 | /* Clear MSB bits from the mask just in case it had them set. */ | |
5379 | emit_insn (gen_avx2_andnotv32qi3 (t2, vt, mask)); | |
5380 | /* After this t1 will have MSB set for elements from other lane. */ | |
5381 | emit_insn (gen_xorv32qi3 (t1, gen_lowpart (V32QImode, t5), vt2)); | |
5382 | /* Clear bits other than MSB. */ | |
5383 | emit_insn (gen_andv32qi3 (t1, t1, vt)); | |
5384 | /* Or in the lower bits from mask into t3. */ | |
5385 | emit_insn (gen_iorv32qi3 (t3, t1, t2)); | |
5386 | /* And invert MSB bits in t1, so MSB is set for elements from the same | |
5387 | lane. */ | |
5388 | emit_insn (gen_xorv32qi3 (t1, t1, vt)); | |
5389 | /* Swap 128-bit lanes in t3. */ | |
5390 | t6 = gen_reg_rtx (V4DImode); | |
5391 | emit_insn (gen_avx2_permv4di_1 (t6, gen_lowpart (V4DImode, t3), | |
5392 | const2_rtx, GEN_INT (3), | |
5393 | const0_rtx, const1_rtx)); | |
5394 | /* And or in the lower bits from mask into t1. */ | |
5395 | emit_insn (gen_iorv32qi3 (t1, t1, t2)); | |
5396 | if (one_operand_shuffle) | |
5397 | { | |
5398 | /* Each of these shuffles will put 0s in places where | |
5399 | element from the other 128-bit lane is needed, otherwise | |
5400 | will shuffle in the requested value. */ | |
5401 | emit_insn (gen_avx2_pshufbv32qi3 (t3, op0, | |
5402 | gen_lowpart (V32QImode, t6))); | |
5403 | emit_insn (gen_avx2_pshufbv32qi3 (t1, op0, t1)); | |
5404 | /* For t3 the 128-bit lanes are swapped again. */ | |
5405 | t7 = gen_reg_rtx (V4DImode); | |
5406 | emit_insn (gen_avx2_permv4di_1 (t7, gen_lowpart (V4DImode, t3), | |
5407 | const2_rtx, GEN_INT (3), | |
5408 | const0_rtx, const1_rtx)); | |
5409 | /* And oring both together leads to the result. */ | |
5410 | emit_insn (gen_iorv32qi3 (target, t1, | |
5411 | gen_lowpart (V32QImode, t7))); | |
5412 | if (target != operands[0]) | |
5413 | emit_move_insn (operands[0], | |
5414 | gen_lowpart (GET_MODE (operands[0]), target)); | |
5415 | return; | |
5416 | } | |
5417 | ||
5418 | t4 = gen_reg_rtx (V32QImode); | |
5419 | /* Similarly to the above one_operand_shuffle code, | |
5420 | just for repeated twice for each operand. merge_two: | |
5421 | code will merge the two results together. */ | |
5422 | emit_insn (gen_avx2_pshufbv32qi3 (t4, op0, | |
5423 | gen_lowpart (V32QImode, t6))); | |
5424 | emit_insn (gen_avx2_pshufbv32qi3 (t3, op1, | |
5425 | gen_lowpart (V32QImode, t6))); | |
5426 | emit_insn (gen_avx2_pshufbv32qi3 (t2, op0, t1)); | |
5427 | emit_insn (gen_avx2_pshufbv32qi3 (t1, op1, t1)); | |
5428 | t7 = gen_reg_rtx (V4DImode); | |
5429 | emit_insn (gen_avx2_permv4di_1 (t7, gen_lowpart (V4DImode, t4), | |
5430 | const2_rtx, GEN_INT (3), | |
5431 | const0_rtx, const1_rtx)); | |
5432 | t8 = gen_reg_rtx (V4DImode); | |
5433 | emit_insn (gen_avx2_permv4di_1 (t8, gen_lowpart (V4DImode, t3), | |
5434 | const2_rtx, GEN_INT (3), | |
5435 | const0_rtx, const1_rtx)); | |
5436 | emit_insn (gen_iorv32qi3 (t4, t2, gen_lowpart (V32QImode, t7))); | |
5437 | emit_insn (gen_iorv32qi3 (t3, t1, gen_lowpart (V32QImode, t8))); | |
5438 | t1 = t4; | |
5439 | t2 = t3; | |
5440 | goto merge_two; | |
5441 | ||
5442 | default: | |
5443 | gcc_assert (GET_MODE_SIZE (mode) <= 16); | |
5444 | break; | |
5445 | } | |
5446 | } | |
5447 | ||
5448 | if (TARGET_XOP) | |
5449 | { | |
5450 | /* The XOP VPPERM insn supports three inputs. By ignoring the | |
5451 | one_operand_shuffle special case, we avoid creating another | |
5452 | set of constant vectors in memory. */ | |
5453 | one_operand_shuffle = false; | |
5454 | ||
5455 | /* mask = mask & {2*w-1, ...} */ | |
5456 | vt = GEN_INT (2*w - 1); | |
5457 | } | |
5458 | else | |
5459 | { | |
5460 | /* mask = mask & {w-1, ...} */ | |
5461 | vt = GEN_INT (w - 1); | |
5462 | } | |
5463 | ||
5464 | vt = gen_const_vec_duplicate (maskmode, vt); | |
5465 | mask = expand_simple_binop (maskmode, AND, mask, vt, | |
5466 | NULL_RTX, 0, OPTAB_DIRECT); | |
5467 | ||
5468 | /* For non-QImode operations, convert the word permutation control | |
5469 | into a byte permutation control. */ | |
5470 | if (mode != V16QImode) | |
5471 | { | |
5472 | mask = expand_simple_binop (maskmode, ASHIFT, mask, | |
5473 | GEN_INT (exact_log2 (e)), | |
5474 | NULL_RTX, 0, OPTAB_DIRECT); | |
5475 | ||
5476 | /* Convert mask to vector of chars. */ | |
5477 | mask = force_reg (V16QImode, gen_lowpart (V16QImode, mask)); | |
5478 | ||
5479 | /* Replicate each of the input bytes into byte positions: | |
5480 | (v2di) --> {0,0,0,0,0,0,0,0, 8,8,8,8,8,8,8,8} | |
5481 | (v4si) --> {0,0,0,0, 4,4,4,4, 8,8,8,8, 12,12,12,12} | |
5482 | (v8hi) --> {0,0, 2,2, 4,4, 6,6, ...}. */ | |
5483 | for (i = 0; i < 16; ++i) | |
5484 | vec[i] = GEN_INT (i/e * e); | |
5485 | vt = gen_rtx_CONST_VECTOR (V16QImode, gen_rtvec_v (16, vec)); | |
5486 | vt = validize_mem (force_const_mem (V16QImode, vt)); | |
5487 | if (TARGET_XOP) | |
5488 | emit_insn (gen_xop_pperm (mask, mask, mask, vt)); | |
5489 | else | |
5490 | emit_insn (gen_ssse3_pshufbv16qi3 (mask, mask, vt)); | |
5491 | ||
5492 | /* Convert it into the byte positions by doing | |
5493 | mask = mask + {0,1,..,16/w, 0,1,..,16/w, ...} */ | |
5494 | for (i = 0; i < 16; ++i) | |
5495 | vec[i] = GEN_INT (i % e); | |
5496 | vt = gen_rtx_CONST_VECTOR (V16QImode, gen_rtvec_v (16, vec)); | |
5497 | vt = validize_mem (force_const_mem (V16QImode, vt)); | |
5498 | emit_insn (gen_addv16qi3 (mask, mask, vt)); | |
5499 | } | |
5500 | ||
5501 | /* The actual shuffle operations all operate on V16QImode. */ | |
5502 | op0 = gen_lowpart (V16QImode, op0); | |
5503 | op1 = gen_lowpart (V16QImode, op1); | |
5504 | ||
5505 | if (TARGET_XOP) | |
5506 | { | |
5507 | if (GET_MODE (target) != V16QImode) | |
5508 | target = gen_reg_rtx (V16QImode); | |
5509 | emit_insn (gen_xop_pperm (target, op0, op1, mask)); | |
5510 | if (target != operands[0]) | |
5511 | emit_move_insn (operands[0], | |
5512 | gen_lowpart (GET_MODE (operands[0]), target)); | |
5513 | } | |
5514 | else if (one_operand_shuffle) | |
5515 | { | |
5516 | if (GET_MODE (target) != V16QImode) | |
5517 | target = gen_reg_rtx (V16QImode); | |
5518 | emit_insn (gen_ssse3_pshufbv16qi3 (target, op0, mask)); | |
5519 | if (target != operands[0]) | |
5520 | emit_move_insn (operands[0], | |
5521 | gen_lowpart (GET_MODE (operands[0]), target)); | |
5522 | } | |
5523 | else | |
5524 | { | |
5525 | rtx xops[6]; | |
5526 | bool ok; | |
5527 | ||
5528 | /* Shuffle the two input vectors independently. */ | |
5529 | t1 = gen_reg_rtx (V16QImode); | |
5530 | t2 = gen_reg_rtx (V16QImode); | |
5531 | emit_insn (gen_ssse3_pshufbv16qi3 (t1, op0, mask)); | |
5532 | emit_insn (gen_ssse3_pshufbv16qi3 (t2, op1, mask)); | |
5533 | ||
5534 | merge_two: | |
5535 | /* Then merge them together. The key is whether any given control | |
5536 | element contained a bit set that indicates the second word. */ | |
5537 | mask = operands[3]; | |
5538 | vt = GEN_INT (w); | |
5539 | if (maskmode == V2DImode && !TARGET_SSE4_1) | |
5540 | { | |
5541 | /* Without SSE4.1, we don't have V2DImode EQ. Perform one | |
5542 | more shuffle to convert the V2DI input mask into a V4SI | |
5543 | input mask. At which point the masking that expand_int_vcond | |
5544 | will work as desired. */ | |
5545 | rtx t3 = gen_reg_rtx (V4SImode); | |
5546 | emit_insn (gen_sse2_pshufd_1 (t3, gen_lowpart (V4SImode, mask), | |
5547 | const0_rtx, const0_rtx, | |
5548 | const2_rtx, const2_rtx)); | |
5549 | mask = t3; | |
5550 | maskmode = V4SImode; | |
5551 | e = w = 4; | |
5552 | } | |
5553 | ||
5554 | vt = gen_const_vec_duplicate (maskmode, vt); | |
5555 | vt = force_reg (maskmode, vt); | |
5556 | mask = expand_simple_binop (maskmode, AND, mask, vt, | |
5557 | NULL_RTX, 0, OPTAB_DIRECT); | |
5558 | ||
5559 | if (GET_MODE (target) != mode) | |
5560 | target = gen_reg_rtx (mode); | |
5561 | xops[0] = target; | |
5562 | xops[1] = gen_lowpart (mode, t2); | |
5563 | xops[2] = gen_lowpart (mode, t1); | |
5564 | xops[3] = gen_rtx_EQ (maskmode, mask, vt); | |
5565 | xops[4] = mask; | |
5566 | xops[5] = vt; | |
5567 | ok = ix86_expand_int_vcond (xops); | |
5568 | gcc_assert (ok); | |
5569 | if (target != operands[0]) | |
5570 | emit_move_insn (operands[0], | |
5571 | gen_lowpart (GET_MODE (operands[0]), target)); | |
5572 | } | |
5573 | } | |
5574 | ||
5575 | /* Unpack OP[1] into the next wider integer vector type. UNSIGNED_P is | |
5576 | true if we should do zero extension, else sign extension. HIGH_P is | |
5577 | true if we want the N/2 high elements, else the low elements. */ | |
5578 | ||
5579 | void | |
5580 | ix86_expand_sse_unpack (rtx dest, rtx src, bool unsigned_p, bool high_p) | |
5581 | { | |
5582 | machine_mode imode = GET_MODE (src); | |
5583 | rtx tmp; | |
5584 | ||
5585 | if (TARGET_SSE4_1) | |
5586 | { | |
5587 | rtx (*unpack)(rtx, rtx); | |
5588 | rtx (*extract)(rtx, rtx) = NULL; | |
5589 | machine_mode halfmode = BLKmode; | |
5590 | ||
5591 | switch (imode) | |
5592 | { | |
5593 | case E_V64QImode: | |
5594 | if (unsigned_p) | |
5595 | unpack = gen_avx512bw_zero_extendv32qiv32hi2; | |
5596 | else | |
5597 | unpack = gen_avx512bw_sign_extendv32qiv32hi2; | |
5598 | halfmode = V32QImode; | |
5599 | extract | |
5600 | = high_p ? gen_vec_extract_hi_v64qi : gen_vec_extract_lo_v64qi; | |
5601 | break; | |
5602 | case E_V32QImode: | |
5603 | if (unsigned_p) | |
5604 | unpack = gen_avx2_zero_extendv16qiv16hi2; | |
5605 | else | |
5606 | unpack = gen_avx2_sign_extendv16qiv16hi2; | |
5607 | halfmode = V16QImode; | |
5608 | extract | |
5609 | = high_p ? gen_vec_extract_hi_v32qi : gen_vec_extract_lo_v32qi; | |
5610 | break; | |
5611 | case E_V32HImode: | |
5612 | if (unsigned_p) | |
5613 | unpack = gen_avx512f_zero_extendv16hiv16si2; | |
5614 | else | |
5615 | unpack = gen_avx512f_sign_extendv16hiv16si2; | |
5616 | halfmode = V16HImode; | |
5617 | extract | |
5618 | = high_p ? gen_vec_extract_hi_v32hi : gen_vec_extract_lo_v32hi; | |
5619 | break; | |
5620 | case E_V16HImode: | |
5621 | if (unsigned_p) | |
5622 | unpack = gen_avx2_zero_extendv8hiv8si2; | |
5623 | else | |
5624 | unpack = gen_avx2_sign_extendv8hiv8si2; | |
5625 | halfmode = V8HImode; | |
5626 | extract | |
5627 | = high_p ? gen_vec_extract_hi_v16hi : gen_vec_extract_lo_v16hi; | |
5628 | break; | |
5629 | case E_V16SImode: | |
5630 | if (unsigned_p) | |
5631 | unpack = gen_avx512f_zero_extendv8siv8di2; | |
5632 | else | |
5633 | unpack = gen_avx512f_sign_extendv8siv8di2; | |
5634 | halfmode = V8SImode; | |
5635 | extract | |
5636 | = high_p ? gen_vec_extract_hi_v16si : gen_vec_extract_lo_v16si; | |
5637 | break; | |
5638 | case E_V8SImode: | |
5639 | if (unsigned_p) | |
5640 | unpack = gen_avx2_zero_extendv4siv4di2; | |
5641 | else | |
5642 | unpack = gen_avx2_sign_extendv4siv4di2; | |
5643 | halfmode = V4SImode; | |
5644 | extract | |
5645 | = high_p ? gen_vec_extract_hi_v8si : gen_vec_extract_lo_v8si; | |
5646 | break; | |
5647 | case E_V16QImode: | |
5648 | if (unsigned_p) | |
5649 | unpack = gen_sse4_1_zero_extendv8qiv8hi2; | |
5650 | else | |
5651 | unpack = gen_sse4_1_sign_extendv8qiv8hi2; | |
5652 | break; | |
5653 | case E_V8HImode: | |
5654 | if (unsigned_p) | |
5655 | unpack = gen_sse4_1_zero_extendv4hiv4si2; | |
5656 | else | |
5657 | unpack = gen_sse4_1_sign_extendv4hiv4si2; | |
5658 | break; | |
5659 | case E_V4SImode: | |
5660 | if (unsigned_p) | |
5661 | unpack = gen_sse4_1_zero_extendv2siv2di2; | |
5662 | else | |
5663 | unpack = gen_sse4_1_sign_extendv2siv2di2; | |
5664 | break; | |
836328b2 UB |
5665 | case E_V8QImode: |
5666 | if (unsigned_p) | |
5667 | unpack = gen_sse4_1_zero_extendv4qiv4hi2; | |
5668 | else | |
5669 | unpack = gen_sse4_1_sign_extendv4qiv4hi2; | |
5670 | break; | |
5671 | case E_V4HImode: | |
5672 | if (unsigned_p) | |
5673 | unpack = gen_sse4_1_zero_extendv2hiv2si2; | |
5674 | else | |
5675 | unpack = gen_sse4_1_sign_extendv2hiv2si2; | |
5676 | break; | |
663a014e UB |
5677 | case E_V4QImode: |
5678 | if (unsigned_p) | |
5679 | unpack = gen_sse4_1_zero_extendv2qiv2hi2; | |
5680 | else | |
5681 | unpack = gen_sse4_1_sign_extendv2qiv2hi2; | |
5682 | break; | |
2bf6d935 ML |
5683 | default: |
5684 | gcc_unreachable (); | |
5685 | } | |
5686 | ||
5687 | if (GET_MODE_SIZE (imode) >= 32) | |
5688 | { | |
5689 | tmp = gen_reg_rtx (halfmode); | |
5690 | emit_insn (extract (tmp, src)); | |
5691 | } | |
5692 | else if (high_p) | |
5693 | { | |
836328b2 UB |
5694 | switch (GET_MODE_SIZE (imode)) |
5695 | { | |
5696 | case 16: | |
5697 | /* Shift higher 8 bytes to lower 8 bytes. */ | |
5698 | tmp = gen_reg_rtx (V1TImode); | |
5699 | emit_insn (gen_sse2_lshrv1ti3 (tmp, gen_lowpart (V1TImode, src), | |
5700 | GEN_INT (64))); | |
5701 | break; | |
5702 | case 8: | |
5703 | /* Shift higher 4 bytes to lower 4 bytes. */ | |
5704 | tmp = gen_reg_rtx (V1DImode); | |
5705 | emit_insn (gen_mmx_lshrv1di3 (tmp, gen_lowpart (V1DImode, src), | |
5706 | GEN_INT (32))); | |
5707 | break; | |
663a014e UB |
5708 | case 4: |
5709 | /* Shift higher 2 bytes to lower 2 bytes. */ | |
5710 | tmp = gen_reg_rtx (V1SImode); | |
5711 | emit_insn (gen_mmx_lshrv1si3 (tmp, gen_lowpart (V1SImode, src), | |
5712 | GEN_INT (16))); | |
5713 | break; | |
836328b2 UB |
5714 | default: |
5715 | gcc_unreachable (); | |
5716 | } | |
5717 | ||
2bf6d935 ML |
5718 | tmp = gen_lowpart (imode, tmp); |
5719 | } | |
5720 | else | |
5721 | tmp = src; | |
5722 | ||
5723 | emit_insn (unpack (dest, tmp)); | |
5724 | } | |
5725 | else | |
5726 | { | |
5727 | rtx (*unpack)(rtx, rtx, rtx); | |
5728 | ||
5729 | switch (imode) | |
5730 | { | |
5731 | case E_V16QImode: | |
5732 | if (high_p) | |
5733 | unpack = gen_vec_interleave_highv16qi; | |
5734 | else | |
5735 | unpack = gen_vec_interleave_lowv16qi; | |
5736 | break; | |
5737 | case E_V8HImode: | |
5738 | if (high_p) | |
5739 | unpack = gen_vec_interleave_highv8hi; | |
5740 | else | |
5741 | unpack = gen_vec_interleave_lowv8hi; | |
5742 | break; | |
5743 | case E_V4SImode: | |
5744 | if (high_p) | |
5745 | unpack = gen_vec_interleave_highv4si; | |
5746 | else | |
5747 | unpack = gen_vec_interleave_lowv4si; | |
5748 | break; | |
836328b2 UB |
5749 | case E_V8QImode: |
5750 | if (high_p) | |
5751 | unpack = gen_mmx_punpckhbw; | |
5752 | else | |
5753 | unpack = gen_mmx_punpcklbw; | |
5754 | break; | |
5755 | case E_V4HImode: | |
5756 | if (high_p) | |
5757 | unpack = gen_mmx_punpckhwd; | |
5758 | else | |
5759 | unpack = gen_mmx_punpcklwd; | |
5760 | break; | |
663a014e UB |
5761 | case E_V4QImode: |
5762 | if (high_p) | |
5763 | unpack = gen_mmx_punpckhbw_low; | |
5764 | else | |
5765 | unpack = gen_mmx_punpcklbw_low; | |
5766 | break; | |
2bf6d935 ML |
5767 | default: |
5768 | gcc_unreachable (); | |
5769 | } | |
5770 | ||
5771 | if (unsigned_p) | |
5772 | tmp = force_reg (imode, CONST0_RTX (imode)); | |
5773 | else | |
5774 | tmp = ix86_expand_sse_cmp (gen_reg_rtx (imode), GT, CONST0_RTX (imode), | |
5775 | src, pc_rtx, pc_rtx); | |
5776 | ||
5777 | rtx tmp2 = gen_reg_rtx (imode); | |
5778 | emit_insn (unpack (tmp2, src, tmp)); | |
5779 | emit_move_insn (dest, gen_lowpart (GET_MODE (dest), tmp2)); | |
5780 | } | |
5781 | } | |
5782 | ||
faf2b6bc | 5783 | /* Return true if mem is pool constant which contains a const_vector |
5784 | perm index, assign the index to PERM. */ | |
5785 | bool | |
5786 | ix86_extract_perm_from_pool_constant (int* perm, rtx mem) | |
5787 | { | |
5788 | machine_mode mode = GET_MODE (mem); | |
5789 | int nelt = GET_MODE_NUNITS (mode); | |
5790 | ||
5791 | if (!INTEGRAL_MODE_P (mode)) | |
5792 | return false; | |
5793 | ||
5794 | /* Needs to be constant pool. */ | |
5795 | if (!(MEM_P (mem)) | |
5796 | || !SYMBOL_REF_P (XEXP (mem, 0)) | |
5797 | || !CONSTANT_POOL_ADDRESS_P (XEXP (mem, 0))) | |
5798 | return false; | |
5799 | ||
5800 | rtx constant = get_pool_constant (XEXP (mem, 0)); | |
5801 | ||
5802 | if (GET_CODE (constant) != CONST_VECTOR) | |
5803 | return false; | |
5804 | ||
5805 | /* There could be some rtx like | |
5806 | (mem/u/c:V16QI (symbol_ref/u:DI ("*.LC1"))) | |
5807 | but with "*.LC1" refer to V2DI constant vector. */ | |
5808 | if (GET_MODE (constant) != mode) | |
5809 | { | |
5810 | constant = simplify_subreg (mode, constant, GET_MODE (constant), 0); | |
5811 | ||
5812 | if (constant == nullptr || GET_CODE (constant) != CONST_VECTOR) | |
5813 | return false; | |
5814 | } | |
5815 | ||
5816 | for (int i = 0; i != nelt; i++) | |
5817 | perm[i] = UINTVAL (XVECEXP (constant, 0, i)); | |
5818 | ||
5819 | return true; | |
5820 | } | |
5821 | ||
2bf6d935 ML |
5822 | /* Split operands 0 and 1 into half-mode parts. Similar to split_double_mode, |
5823 | but works for floating pointer parameters and nonoffsetable memories. | |
5824 | For pushes, it returns just stack offsets; the values will be saved | |
5825 | in the right order. Maximally three parts are generated. */ | |
5826 | ||
5827 | static int | |
5828 | ix86_split_to_parts (rtx operand, rtx *parts, machine_mode mode) | |
5829 | { | |
5830 | int size; | |
5831 | ||
5832 | if (!TARGET_64BIT) | |
5833 | size = mode==XFmode ? 3 : GET_MODE_SIZE (mode) / 4; | |
5834 | else | |
5835 | size = (GET_MODE_SIZE (mode) + 4) / 8; | |
5836 | ||
5837 | gcc_assert (!REG_P (operand) || !MMX_REGNO_P (REGNO (operand))); | |
5838 | gcc_assert (size >= 2 && size <= 4); | |
5839 | ||
5840 | /* Optimize constant pool reference to immediates. This is used by fp | |
5841 | moves, that force all constants to memory to allow combining. */ | |
5842 | if (MEM_P (operand) && MEM_READONLY_P (operand)) | |
5843 | operand = avoid_constant_pool_reference (operand); | |
5844 | ||
5845 | if (MEM_P (operand) && !offsettable_memref_p (operand)) | |
5846 | { | |
5847 | /* The only non-offsetable memories we handle are pushes. */ | |
5848 | int ok = push_operand (operand, VOIDmode); | |
5849 | ||
5850 | gcc_assert (ok); | |
5851 | ||
5852 | operand = copy_rtx (operand); | |
5853 | PUT_MODE (operand, word_mode); | |
5854 | parts[0] = parts[1] = parts[2] = parts[3] = operand; | |
5855 | return size; | |
5856 | } | |
5857 | ||
5858 | if (GET_CODE (operand) == CONST_VECTOR) | |
5859 | { | |
5860 | scalar_int_mode imode = int_mode_for_mode (mode).require (); | |
5861 | /* Caution: if we looked through a constant pool memory above, | |
5862 | the operand may actually have a different mode now. That's | |
5863 | ok, since we want to pun this all the way back to an integer. */ | |
5864 | operand = simplify_subreg (imode, operand, GET_MODE (operand), 0); | |
5865 | gcc_assert (operand != NULL); | |
5866 | mode = imode; | |
5867 | } | |
5868 | ||
5869 | if (!TARGET_64BIT) | |
5870 | { | |
5871 | if (mode == DImode) | |
5872 | split_double_mode (mode, &operand, 1, &parts[0], &parts[1]); | |
5873 | else | |
5874 | { | |
5875 | int i; | |
5876 | ||
5877 | if (REG_P (operand)) | |
5878 | { | |
5879 | gcc_assert (reload_completed); | |
5880 | for (i = 0; i < size; i++) | |
5881 | parts[i] = gen_rtx_REG (SImode, REGNO (operand) + i); | |
5882 | } | |
5883 | else if (offsettable_memref_p (operand)) | |
5884 | { | |
5885 | operand = adjust_address (operand, SImode, 0); | |
5886 | parts[0] = operand; | |
5887 | for (i = 1; i < size; i++) | |
5888 | parts[i] = adjust_address (operand, SImode, 4 * i); | |
5889 | } | |
5890 | else if (CONST_DOUBLE_P (operand)) | |
5891 | { | |
5892 | const REAL_VALUE_TYPE *r; | |
5893 | long l[4]; | |
5894 | ||
5895 | r = CONST_DOUBLE_REAL_VALUE (operand); | |
5896 | switch (mode) | |
5897 | { | |
5898 | case E_TFmode: | |
5899 | real_to_target (l, r, mode); | |
5900 | parts[3] = gen_int_mode (l[3], SImode); | |
5901 | parts[2] = gen_int_mode (l[2], SImode); | |
5902 | break; | |
5903 | case E_XFmode: | |
5904 | /* We can't use REAL_VALUE_TO_TARGET_LONG_DOUBLE since | |
5905 | long double may not be 80-bit. */ | |
5906 | real_to_target (l, r, mode); | |
5907 | parts[2] = gen_int_mode (l[2], SImode); | |
5908 | break; | |
5909 | case E_DFmode: | |
5910 | REAL_VALUE_TO_TARGET_DOUBLE (*r, l); | |
5911 | break; | |
5912 | default: | |
5913 | gcc_unreachable (); | |
5914 | } | |
5915 | parts[1] = gen_int_mode (l[1], SImode); | |
5916 | parts[0] = gen_int_mode (l[0], SImode); | |
5917 | } | |
5918 | else | |
5919 | gcc_unreachable (); | |
5920 | } | |
5921 | } | |
5922 | else | |
5923 | { | |
5924 | if (mode == TImode) | |
5925 | split_double_mode (mode, &operand, 1, &parts[0], &parts[1]); | |
5926 | if (mode == XFmode || mode == TFmode) | |
5927 | { | |
5928 | machine_mode upper_mode = mode==XFmode ? SImode : DImode; | |
5929 | if (REG_P (operand)) | |
5930 | { | |
5931 | gcc_assert (reload_completed); | |
5932 | parts[0] = gen_rtx_REG (DImode, REGNO (operand) + 0); | |
5933 | parts[1] = gen_rtx_REG (upper_mode, REGNO (operand) + 1); | |
5934 | } | |
5935 | else if (offsettable_memref_p (operand)) | |
5936 | { | |
5937 | operand = adjust_address (operand, DImode, 0); | |
5938 | parts[0] = operand; | |
5939 | parts[1] = adjust_address (operand, upper_mode, 8); | |
5940 | } | |
5941 | else if (CONST_DOUBLE_P (operand)) | |
5942 | { | |
5943 | long l[4]; | |
5944 | ||
5945 | real_to_target (l, CONST_DOUBLE_REAL_VALUE (operand), mode); | |
5946 | ||
5947 | /* real_to_target puts 32-bit pieces in each long. */ | |
5948 | parts[0] = gen_int_mode ((l[0] & HOST_WIDE_INT_C (0xffffffff)) | |
5949 | | ((l[1] & HOST_WIDE_INT_C (0xffffffff)) | |
5950 | << 32), DImode); | |
5951 | ||
5952 | if (upper_mode == SImode) | |
5953 | parts[1] = gen_int_mode (l[2], SImode); | |
5954 | else | |
5955 | parts[1] | |
5956 | = gen_int_mode ((l[2] & HOST_WIDE_INT_C (0xffffffff)) | |
5957 | | ((l[3] & HOST_WIDE_INT_C (0xffffffff)) | |
5958 | << 32), DImode); | |
5959 | } | |
5960 | else | |
5961 | gcc_unreachable (); | |
5962 | } | |
5963 | } | |
5964 | ||
5965 | return size; | |
5966 | } | |
5967 | ||
5968 | /* Emit insns to perform a move or push of DI, DF, XF, and TF values. | |
5969 | Return false when normal moves are needed; true when all required | |
5970 | insns have been emitted. Operands 2-4 contain the input values | |
5971 | int the correct order; operands 5-7 contain the output values. */ | |
5972 | ||
5973 | void | |
5974 | ix86_split_long_move (rtx operands[]) | |
5975 | { | |
5976 | rtx part[2][4]; | |
5977 | int nparts, i, j; | |
5978 | int push = 0; | |
5979 | int collisions = 0; | |
5980 | machine_mode mode = GET_MODE (operands[0]); | |
5981 | bool collisionparts[4]; | |
5982 | ||
5983 | /* The DFmode expanders may ask us to move double. | |
5984 | For 64bit target this is single move. By hiding the fact | |
5985 | here we simplify i386.md splitters. */ | |
5986 | if (TARGET_64BIT && GET_MODE_SIZE (GET_MODE (operands[0])) == 8) | |
5987 | { | |
5988 | /* Optimize constant pool reference to immediates. This is used by | |
5989 | fp moves, that force all constants to memory to allow combining. */ | |
5990 | ||
5991 | if (MEM_P (operands[1]) | |
5992 | && GET_CODE (XEXP (operands[1], 0)) == SYMBOL_REF | |
5993 | && CONSTANT_POOL_ADDRESS_P (XEXP (operands[1], 0))) | |
5994 | operands[1] = get_pool_constant (XEXP (operands[1], 0)); | |
5995 | if (push_operand (operands[0], VOIDmode)) | |
5996 | { | |
5997 | operands[0] = copy_rtx (operands[0]); | |
5998 | PUT_MODE (operands[0], word_mode); | |
5999 | } | |
6000 | else | |
6001 | operands[0] = gen_lowpart (DImode, operands[0]); | |
6002 | operands[1] = gen_lowpart (DImode, operands[1]); | |
6003 | emit_move_insn (operands[0], operands[1]); | |
6004 | return; | |
6005 | } | |
6006 | ||
6007 | /* The only non-offsettable memory we handle is push. */ | |
6008 | if (push_operand (operands[0], VOIDmode)) | |
6009 | push = 1; | |
6010 | else | |
6011 | gcc_assert (!MEM_P (operands[0]) | |
6012 | || offsettable_memref_p (operands[0])); | |
6013 | ||
6014 | nparts = ix86_split_to_parts (operands[1], part[1], GET_MODE (operands[0])); | |
6015 | ix86_split_to_parts (operands[0], part[0], GET_MODE (operands[0])); | |
6016 | ||
6017 | /* When emitting push, take care for source operands on the stack. */ | |
6018 | if (push && MEM_P (operands[1]) | |
6019 | && reg_overlap_mentioned_p (stack_pointer_rtx, operands[1])) | |
6020 | { | |
6021 | rtx src_base = XEXP (part[1][nparts - 1], 0); | |
6022 | ||
6023 | /* Compensate for the stack decrement by 4. */ | |
6024 | if (!TARGET_64BIT && nparts == 3 | |
6025 | && mode == XFmode && TARGET_128BIT_LONG_DOUBLE) | |
6026 | src_base = plus_constant (Pmode, src_base, 4); | |
6027 | ||
6028 | /* src_base refers to the stack pointer and is | |
6029 | automatically decreased by emitted push. */ | |
6030 | for (i = 0; i < nparts; i++) | |
6031 | part[1][i] = change_address (part[1][i], | |
6032 | GET_MODE (part[1][i]), src_base); | |
6033 | } | |
6034 | ||
6035 | /* We need to do copy in the right order in case an address register | |
6036 | of the source overlaps the destination. */ | |
6037 | if (REG_P (part[0][0]) && MEM_P (part[1][0])) | |
6038 | { | |
6039 | rtx tmp; | |
6040 | ||
6041 | for (i = 0; i < nparts; i++) | |
6042 | { | |
6043 | collisionparts[i] | |
6044 | = reg_overlap_mentioned_p (part[0][i], XEXP (part[1][0], 0)); | |
6045 | if (collisionparts[i]) | |
6046 | collisions++; | |
6047 | } | |
6048 | ||
6049 | /* Collision in the middle part can be handled by reordering. */ | |
6050 | if (collisions == 1 && nparts == 3 && collisionparts [1]) | |
6051 | { | |
6052 | std::swap (part[0][1], part[0][2]); | |
6053 | std::swap (part[1][1], part[1][2]); | |
6054 | } | |
6055 | else if (collisions == 1 | |
6056 | && nparts == 4 | |
6057 | && (collisionparts [1] || collisionparts [2])) | |
6058 | { | |
6059 | if (collisionparts [1]) | |
6060 | { | |
6061 | std::swap (part[0][1], part[0][2]); | |
6062 | std::swap (part[1][1], part[1][2]); | |
6063 | } | |
6064 | else | |
6065 | { | |
6066 | std::swap (part[0][2], part[0][3]); | |
6067 | std::swap (part[1][2], part[1][3]); | |
6068 | } | |
6069 | } | |
6070 | ||
6071 | /* If there are more collisions, we can't handle it by reordering. | |
6072 | Do an lea to the last part and use only one colliding move. */ | |
6073 | else if (collisions > 1) | |
6074 | { | |
6075 | rtx base, addr; | |
6076 | ||
6077 | collisions = 1; | |
6078 | ||
6079 | base = part[0][nparts - 1]; | |
6080 | ||
6081 | /* Handle the case when the last part isn't valid for lea. | |
6082 | Happens in 64-bit mode storing the 12-byte XFmode. */ | |
6083 | if (GET_MODE (base) != Pmode) | |
6084 | base = gen_rtx_REG (Pmode, REGNO (base)); | |
6085 | ||
6086 | addr = XEXP (part[1][0], 0); | |
6087 | if (TARGET_TLS_DIRECT_SEG_REFS) | |
6088 | { | |
6089 | struct ix86_address parts; | |
6090 | int ok = ix86_decompose_address (addr, &parts); | |
6091 | gcc_assert (ok); | |
6092 | /* It is not valid to use %gs: or %fs: in lea. */ | |
6093 | gcc_assert (parts.seg == ADDR_SPACE_GENERIC); | |
6094 | } | |
6095 | emit_insn (gen_rtx_SET (base, addr)); | |
6096 | part[1][0] = replace_equiv_address (part[1][0], base); | |
6097 | for (i = 1; i < nparts; i++) | |
6098 | { | |
6099 | tmp = plus_constant (Pmode, base, UNITS_PER_WORD * i); | |
6100 | part[1][i] = replace_equiv_address (part[1][i], tmp); | |
6101 | } | |
6102 | } | |
6103 | } | |
6104 | ||
6105 | if (push) | |
6106 | { | |
6107 | if (!TARGET_64BIT) | |
6108 | { | |
6109 | if (nparts == 3) | |
6110 | { | |
6111 | if (TARGET_128BIT_LONG_DOUBLE && mode == XFmode) | |
d9330fb5 | 6112 | emit_insn (gen_add2_insn (stack_pointer_rtx, GEN_INT (-4))); |
2bf6d935 ML |
6113 | emit_move_insn (part[0][2], part[1][2]); |
6114 | } | |
6115 | else if (nparts == 4) | |
6116 | { | |
6117 | emit_move_insn (part[0][3], part[1][3]); | |
6118 | emit_move_insn (part[0][2], part[1][2]); | |
6119 | } | |
6120 | } | |
6121 | else | |
6122 | { | |
6123 | /* In 64bit mode we don't have 32bit push available. In case this is | |
6124 | register, it is OK - we will just use larger counterpart. We also | |
6125 | retype memory - these comes from attempt to avoid REX prefix on | |
6126 | moving of second half of TFmode value. */ | |
6127 | if (GET_MODE (part[1][1]) == SImode) | |
6128 | { | |
6129 | switch (GET_CODE (part[1][1])) | |
6130 | { | |
6131 | case MEM: | |
6132 | part[1][1] = adjust_address (part[1][1], DImode, 0); | |
6133 | break; | |
6134 | ||
6135 | case REG: | |
6136 | part[1][1] = gen_rtx_REG (DImode, REGNO (part[1][1])); | |
6137 | break; | |
6138 | ||
6139 | default: | |
6140 | gcc_unreachable (); | |
6141 | } | |
6142 | ||
6143 | if (GET_MODE (part[1][0]) == SImode) | |
6144 | part[1][0] = part[1][1]; | |
6145 | } | |
6146 | } | |
6147 | emit_move_insn (part[0][1], part[1][1]); | |
6148 | emit_move_insn (part[0][0], part[1][0]); | |
6149 | return; | |
6150 | } | |
6151 | ||
6152 | /* Choose correct order to not overwrite the source before it is copied. */ | |
6153 | if ((REG_P (part[0][0]) | |
6154 | && REG_P (part[1][1]) | |
6155 | && (REGNO (part[0][0]) == REGNO (part[1][1]) | |
6156 | || (nparts == 3 | |
6157 | && REGNO (part[0][0]) == REGNO (part[1][2])) | |
6158 | || (nparts == 4 | |
6159 | && REGNO (part[0][0]) == REGNO (part[1][3])))) | |
6160 | || (collisions > 0 | |
6161 | && reg_overlap_mentioned_p (part[0][0], XEXP (part[1][0], 0)))) | |
6162 | { | |
6163 | for (i = 0, j = nparts - 1; i < nparts; i++, j--) | |
6164 | { | |
6165 | operands[2 + i] = part[0][j]; | |
6166 | operands[6 + i] = part[1][j]; | |
6167 | } | |
6168 | } | |
6169 | else | |
6170 | { | |
6171 | for (i = 0; i < nparts; i++) | |
6172 | { | |
6173 | operands[2 + i] = part[0][i]; | |
6174 | operands[6 + i] = part[1][i]; | |
6175 | } | |
6176 | } | |
6177 | ||
6178 | /* If optimizing for size, attempt to locally unCSE nonzero constants. */ | |
6179 | if (optimize_insn_for_size_p ()) | |
6180 | { | |
6181 | for (j = 0; j < nparts - 1; j++) | |
6182 | if (CONST_INT_P (operands[6 + j]) | |
6183 | && operands[6 + j] != const0_rtx | |
6184 | && REG_P (operands[2 + j])) | |
6185 | for (i = j; i < nparts - 1; i++) | |
6186 | if (CONST_INT_P (operands[7 + i]) | |
6187 | && INTVAL (operands[7 + i]) == INTVAL (operands[6 + j])) | |
6188 | operands[7 + i] = operands[2 + j]; | |
6189 | } | |
6190 | ||
6191 | for (i = 0; i < nparts; i++) | |
6192 | emit_move_insn (operands[2 + i], operands[6 + i]); | |
6193 | ||
6194 | return; | |
6195 | } | |
6196 | ||
6197 | /* Helper function of ix86_split_ashl used to generate an SImode/DImode | |
6198 | left shift by a constant, either using a single shift or | |
6199 | a sequence of add instructions. */ | |
6200 | ||
6201 | static void | |
6202 | ix86_expand_ashl_const (rtx operand, int count, machine_mode mode) | |
6203 | { | |
2bf6d935 ML |
6204 | if (count == 1 |
6205 | || (count * ix86_cost->add <= ix86_cost->shift_const | |
6206 | && !optimize_insn_for_size_p ())) | |
6207 | { | |
2bf6d935 | 6208 | while (count-- > 0) |
83bc5e44 | 6209 | emit_insn (gen_add2_insn (operand, operand)); |
2bf6d935 ML |
6210 | } |
6211 | else | |
6212 | { | |
83bc5e44 UB |
6213 | rtx (*insn)(rtx, rtx, rtx); |
6214 | ||
2bf6d935 ML |
6215 | insn = mode == DImode ? gen_ashlsi3 : gen_ashldi3; |
6216 | emit_insn (insn (operand, operand, GEN_INT (count))); | |
6217 | } | |
6218 | } | |
6219 | ||
6220 | void | |
6221 | ix86_split_ashl (rtx *operands, rtx scratch, machine_mode mode) | |
6222 | { | |
6223 | rtx (*gen_ashl3)(rtx, rtx, rtx); | |
6224 | rtx (*gen_shld)(rtx, rtx, rtx); | |
6225 | int half_width = GET_MODE_BITSIZE (mode) >> 1; | |
987a3082 | 6226 | machine_mode half_mode; |
2bf6d935 ML |
6227 | |
6228 | rtx low[2], high[2]; | |
6229 | int count; | |
6230 | ||
6231 | if (CONST_INT_P (operands[2])) | |
6232 | { | |
6233 | split_double_mode (mode, operands, 2, low, high); | |
6234 | count = INTVAL (operands[2]) & (GET_MODE_BITSIZE (mode) - 1); | |
6235 | ||
6236 | if (count >= half_width) | |
6237 | { | |
6238 | emit_move_insn (high[0], low[1]); | |
38b649ec | 6239 | ix86_expand_clear (low[0]); |
2bf6d935 ML |
6240 | |
6241 | if (count > half_width) | |
6242 | ix86_expand_ashl_const (high[0], count - half_width, mode); | |
6243 | } | |
6244 | else | |
6245 | { | |
6246 | gen_shld = mode == DImode ? gen_x86_shld : gen_x86_64_shld; | |
6247 | ||
6248 | if (!rtx_equal_p (operands[0], operands[1])) | |
6249 | emit_move_insn (operands[0], operands[1]); | |
6250 | ||
6251 | emit_insn (gen_shld (high[0], low[0], GEN_INT (count))); | |
6252 | ix86_expand_ashl_const (low[0], count, mode); | |
6253 | } | |
6254 | return; | |
6255 | } | |
6256 | ||
6257 | split_double_mode (mode, operands, 1, low, high); | |
987a3082 | 6258 | half_mode = mode == DImode ? SImode : DImode; |
2bf6d935 ML |
6259 | |
6260 | gen_ashl3 = mode == DImode ? gen_ashlsi3 : gen_ashldi3; | |
6261 | ||
6262 | if (operands[1] == const1_rtx) | |
6263 | { | |
6264 | /* Assuming we've chosen a QImode capable registers, then 1 << N | |
6265 | can be done with two 32/64-bit shifts, no branches, no cmoves. */ | |
6266 | if (ANY_QI_REG_P (low[0]) && ANY_QI_REG_P (high[0])) | |
6267 | { | |
6268 | rtx s, d, flags = gen_rtx_REG (CCZmode, FLAGS_REG); | |
6269 | ||
6270 | ix86_expand_clear (low[0]); | |
6271 | ix86_expand_clear (high[0]); | |
6272 | emit_insn (gen_testqi_ccz_1 (operands[2], GEN_INT (half_width))); | |
6273 | ||
6274 | d = gen_lowpart (QImode, low[0]); | |
6275 | d = gen_rtx_STRICT_LOW_PART (VOIDmode, d); | |
6276 | s = gen_rtx_EQ (QImode, flags, const0_rtx); | |
6277 | emit_insn (gen_rtx_SET (d, s)); | |
6278 | ||
6279 | d = gen_lowpart (QImode, high[0]); | |
6280 | d = gen_rtx_STRICT_LOW_PART (VOIDmode, d); | |
6281 | s = gen_rtx_NE (QImode, flags, const0_rtx); | |
6282 | emit_insn (gen_rtx_SET (d, s)); | |
6283 | } | |
6284 | ||
6285 | /* Otherwise, we can get the same results by manually performing | |
6286 | a bit extract operation on bit 5/6, and then performing the two | |
6287 | shifts. The two methods of getting 0/1 into low/high are exactly | |
6288 | the same size. Avoiding the shift in the bit extract case helps | |
6289 | pentium4 a bit; no one else seems to care much either way. */ | |
6290 | else | |
6291 | { | |
2bf6d935 ML |
6292 | rtx (*gen_lshr3)(rtx, rtx, rtx); |
6293 | rtx (*gen_and3)(rtx, rtx, rtx); | |
6294 | rtx (*gen_xor3)(rtx, rtx, rtx); | |
6295 | HOST_WIDE_INT bits; | |
6296 | rtx x; | |
6297 | ||
6298 | if (mode == DImode) | |
6299 | { | |
2bf6d935 ML |
6300 | gen_lshr3 = gen_lshrsi3; |
6301 | gen_and3 = gen_andsi3; | |
6302 | gen_xor3 = gen_xorsi3; | |
6303 | bits = 5; | |
6304 | } | |
6305 | else | |
6306 | { | |
2bf6d935 ML |
6307 | gen_lshr3 = gen_lshrdi3; |
6308 | gen_and3 = gen_anddi3; | |
6309 | gen_xor3 = gen_xordi3; | |
6310 | bits = 6; | |
6311 | } | |
6312 | ||
6313 | if (TARGET_PARTIAL_REG_STALL && !optimize_insn_for_size_p ()) | |
6314 | x = gen_rtx_ZERO_EXTEND (half_mode, operands[2]); | |
6315 | else | |
6316 | x = gen_lowpart (half_mode, operands[2]); | |
6317 | emit_insn (gen_rtx_SET (high[0], x)); | |
6318 | ||
6319 | emit_insn (gen_lshr3 (high[0], high[0], GEN_INT (bits))); | |
6320 | emit_insn (gen_and3 (high[0], high[0], const1_rtx)); | |
6321 | emit_move_insn (low[0], high[0]); | |
6322 | emit_insn (gen_xor3 (low[0], low[0], const1_rtx)); | |
6323 | } | |
6324 | ||
6325 | emit_insn (gen_ashl3 (low[0], low[0], operands[2])); | |
6326 | emit_insn (gen_ashl3 (high[0], high[0], operands[2])); | |
6327 | return; | |
6328 | } | |
6329 | ||
6330 | if (operands[1] == constm1_rtx) | |
6331 | { | |
6332 | /* For -1 << N, we can avoid the shld instruction, because we | |
6333 | know that we're shifting 0...31/63 ones into a -1. */ | |
6334 | emit_move_insn (low[0], constm1_rtx); | |
6335 | if (optimize_insn_for_size_p ()) | |
6336 | emit_move_insn (high[0], low[0]); | |
6337 | else | |
6338 | emit_move_insn (high[0], constm1_rtx); | |
6339 | } | |
6340 | else | |
6341 | { | |
6342 | gen_shld = mode == DImode ? gen_x86_shld : gen_x86_64_shld; | |
6343 | ||
6344 | if (!rtx_equal_p (operands[0], operands[1])) | |
6345 | emit_move_insn (operands[0], operands[1]); | |
6346 | ||
6347 | split_double_mode (mode, operands, 1, low, high); | |
6348 | emit_insn (gen_shld (high[0], low[0], operands[2])); | |
6349 | } | |
6350 | ||
6351 | emit_insn (gen_ashl3 (low[0], low[0], operands[2])); | |
6352 | ||
6353 | if (TARGET_CMOVE && scratch) | |
6354 | { | |
2bf6d935 | 6355 | ix86_expand_clear (scratch); |
987a3082 UB |
6356 | emit_insn (gen_x86_shift_adj_1 |
6357 | (half_mode, high[0], low[0], operands[2], scratch)); | |
2bf6d935 ML |
6358 | } |
6359 | else | |
987a3082 | 6360 | emit_insn (gen_x86_shift_adj_2 (half_mode, high[0], low[0], operands[2])); |
2bf6d935 ML |
6361 | } |
6362 | ||
6363 | void | |
6364 | ix86_split_ashr (rtx *operands, rtx scratch, machine_mode mode) | |
6365 | { | |
6366 | rtx (*gen_ashr3)(rtx, rtx, rtx) | |
6367 | = mode == DImode ? gen_ashrsi3 : gen_ashrdi3; | |
6368 | rtx (*gen_shrd)(rtx, rtx, rtx); | |
6369 | int half_width = GET_MODE_BITSIZE (mode) >> 1; | |
6370 | ||
6371 | rtx low[2], high[2]; | |
6372 | int count; | |
6373 | ||
6374 | if (CONST_INT_P (operands[2])) | |
6375 | { | |
6376 | split_double_mode (mode, operands, 2, low, high); | |
6377 | count = INTVAL (operands[2]) & (GET_MODE_BITSIZE (mode) - 1); | |
6378 | ||
6379 | if (count == GET_MODE_BITSIZE (mode) - 1) | |
6380 | { | |
6381 | emit_move_insn (high[0], high[1]); | |
6382 | emit_insn (gen_ashr3 (high[0], high[0], | |
6383 | GEN_INT (half_width - 1))); | |
6384 | emit_move_insn (low[0], high[0]); | |
6385 | ||
6386 | } | |
6387 | else if (count >= half_width) | |
6388 | { | |
6389 | emit_move_insn (low[0], high[1]); | |
6390 | emit_move_insn (high[0], low[0]); | |
6391 | emit_insn (gen_ashr3 (high[0], high[0], | |
6392 | GEN_INT (half_width - 1))); | |
6393 | ||
6394 | if (count > half_width) | |
6395 | emit_insn (gen_ashr3 (low[0], low[0], | |
6396 | GEN_INT (count - half_width))); | |
6397 | } | |
6398 | else | |
6399 | { | |
6400 | gen_shrd = mode == DImode ? gen_x86_shrd : gen_x86_64_shrd; | |
6401 | ||
6402 | if (!rtx_equal_p (operands[0], operands[1])) | |
6403 | emit_move_insn (operands[0], operands[1]); | |
6404 | ||
6405 | emit_insn (gen_shrd (low[0], high[0], GEN_INT (count))); | |
6406 | emit_insn (gen_ashr3 (high[0], high[0], GEN_INT (count))); | |
6407 | } | |
6408 | } | |
6409 | else | |
6410 | { | |
987a3082 UB |
6411 | machine_mode half_mode; |
6412 | ||
2bf6d935 ML |
6413 | gen_shrd = mode == DImode ? gen_x86_shrd : gen_x86_64_shrd; |
6414 | ||
6415 | if (!rtx_equal_p (operands[0], operands[1])) | |
6416 | emit_move_insn (operands[0], operands[1]); | |
6417 | ||
6418 | split_double_mode (mode, operands, 1, low, high); | |
987a3082 | 6419 | half_mode = mode == DImode ? SImode : DImode; |
2bf6d935 ML |
6420 | |
6421 | emit_insn (gen_shrd (low[0], high[0], operands[2])); | |
6422 | emit_insn (gen_ashr3 (high[0], high[0], operands[2])); | |
6423 | ||
6424 | if (TARGET_CMOVE && scratch) | |
6425 | { | |
2bf6d935 ML |
6426 | emit_move_insn (scratch, high[0]); |
6427 | emit_insn (gen_ashr3 (scratch, scratch, | |
6428 | GEN_INT (half_width - 1))); | |
987a3082 UB |
6429 | emit_insn (gen_x86_shift_adj_1 |
6430 | (half_mode, low[0], high[0], operands[2], scratch)); | |
2bf6d935 ML |
6431 | } |
6432 | else | |
987a3082 UB |
6433 | emit_insn (gen_x86_shift_adj_3 |
6434 | (half_mode, low[0], high[0], operands[2])); | |
2bf6d935 ML |
6435 | } |
6436 | } | |
6437 | ||
6438 | void | |
6439 | ix86_split_lshr (rtx *operands, rtx scratch, machine_mode mode) | |
6440 | { | |
6441 | rtx (*gen_lshr3)(rtx, rtx, rtx) | |
6442 | = mode == DImode ? gen_lshrsi3 : gen_lshrdi3; | |
6443 | rtx (*gen_shrd)(rtx, rtx, rtx); | |
6444 | int half_width = GET_MODE_BITSIZE (mode) >> 1; | |
6445 | ||
6446 | rtx low[2], high[2]; | |
6447 | int count; | |
6448 | ||
6449 | if (CONST_INT_P (operands[2])) | |
6450 | { | |
6451 | split_double_mode (mode, operands, 2, low, high); | |
6452 | count = INTVAL (operands[2]) & (GET_MODE_BITSIZE (mode) - 1); | |
6453 | ||
6454 | if (count >= half_width) | |
6455 | { | |
6456 | emit_move_insn (low[0], high[1]); | |
6457 | ix86_expand_clear (high[0]); | |
6458 | ||
6459 | if (count > half_width) | |
6460 | emit_insn (gen_lshr3 (low[0], low[0], | |
6461 | GEN_INT (count - half_width))); | |
6462 | } | |
6463 | else | |
6464 | { | |
6465 | gen_shrd = mode == DImode ? gen_x86_shrd : gen_x86_64_shrd; | |
6466 | ||
6467 | if (!rtx_equal_p (operands[0], operands[1])) | |
6468 | emit_move_insn (operands[0], operands[1]); | |
6469 | ||
6470 | emit_insn (gen_shrd (low[0], high[0], GEN_INT (count))); | |
6471 | emit_insn (gen_lshr3 (high[0], high[0], GEN_INT (count))); | |
6472 | } | |
6473 | } | |
6474 | else | |
6475 | { | |
987a3082 UB |
6476 | machine_mode half_mode; |
6477 | ||
2bf6d935 ML |
6478 | gen_shrd = mode == DImode ? gen_x86_shrd : gen_x86_64_shrd; |
6479 | ||
6480 | if (!rtx_equal_p (operands[0], operands[1])) | |
6481 | emit_move_insn (operands[0], operands[1]); | |
6482 | ||
6483 | split_double_mode (mode, operands, 1, low, high); | |
987a3082 | 6484 | half_mode = mode == DImode ? SImode : DImode; |
2bf6d935 ML |
6485 | |
6486 | emit_insn (gen_shrd (low[0], high[0], operands[2])); | |
6487 | emit_insn (gen_lshr3 (high[0], high[0], operands[2])); | |
6488 | ||
6489 | if (TARGET_CMOVE && scratch) | |
6490 | { | |
2bf6d935 | 6491 | ix86_expand_clear (scratch); |
987a3082 UB |
6492 | emit_insn (gen_x86_shift_adj_1 |
6493 | (half_mode, low[0], high[0], operands[2], scratch)); | |
2bf6d935 ML |
6494 | } |
6495 | else | |
987a3082 UB |
6496 | emit_insn (gen_x86_shift_adj_2 |
6497 | (half_mode, low[0], high[0], operands[2])); | |
2bf6d935 ML |
6498 | } |
6499 | } | |
6500 | ||
1188cf5f RS |
6501 | /* Expand move of V1TI mode register X to a new TI mode register. */ |
6502 | static rtx | |
6503 | ix86_expand_v1ti_to_ti (rtx x) | |
6504 | { | |
6505 | rtx result = gen_reg_rtx (TImode); | |
a5d269f0 RS |
6506 | if (TARGET_SSE2) |
6507 | { | |
51e9e8a2 | 6508 | rtx temp = force_reg (V2DImode, gen_lowpart (V2DImode, x)); |
a5d269f0 RS |
6509 | rtx lo = gen_lowpart (DImode, result); |
6510 | emit_insn (gen_vec_extractv2didi (lo, temp, const0_rtx)); | |
6511 | rtx hi = gen_highpart (DImode, result); | |
6512 | emit_insn (gen_vec_extractv2didi (hi, temp, const1_rtx)); | |
6513 | } | |
6514 | else | |
6515 | emit_move_insn (result, gen_lowpart (TImode, x)); | |
1188cf5f RS |
6516 | return result; |
6517 | } | |
6518 | ||
6519 | /* Expand move of TI mode register X to a new V1TI mode register. */ | |
6520 | static rtx | |
6521 | ix86_expand_ti_to_v1ti (rtx x) | |
6522 | { | |
1188cf5f RS |
6523 | if (TARGET_SSE2) |
6524 | { | |
6525 | rtx lo = gen_lowpart (DImode, x); | |
6526 | rtx hi = gen_highpart (DImode, x); | |
6527 | rtx tmp = gen_reg_rtx (V2DImode); | |
6528 | emit_insn (gen_vec_concatv2di (tmp, lo, hi)); | |
51e9e8a2 | 6529 | return force_reg (V1TImode, gen_lowpart (V1TImode, tmp)); |
1188cf5f | 6530 | } |
51e9e8a2 RS |
6531 | |
6532 | return force_reg (V1TImode, gen_lowpart (V1TImode, x)); | |
1188cf5f RS |
6533 | } |
6534 | ||
6b8b2557 | 6535 | /* Expand V1TI mode shift (of rtx_code CODE) by constant. */ |
1188cf5f RS |
6536 | void |
6537 | ix86_expand_v1ti_shift (enum rtx_code code, rtx operands[]) | |
6b8b2557 | 6538 | { |
6b8b2557 RS |
6539 | rtx op1 = force_reg (V1TImode, operands[1]); |
6540 | ||
1188cf5f RS |
6541 | if (!CONST_INT_P (operands[2])) |
6542 | { | |
6543 | rtx tmp1 = ix86_expand_v1ti_to_ti (op1); | |
6544 | rtx tmp2 = gen_reg_rtx (TImode); | |
6545 | rtx (*shift) (rtx, rtx, rtx) | |
6546 | = (code == ASHIFT) ? gen_ashlti3 : gen_lshrti3; | |
6547 | emit_insn (shift (tmp2, tmp1, operands[2])); | |
6548 | rtx tmp3 = ix86_expand_ti_to_v1ti (tmp2); | |
6549 | emit_move_insn (operands[0], tmp3); | |
6550 | return; | |
6551 | } | |
6552 | ||
6553 | HOST_WIDE_INT bits = INTVAL (operands[2]) & 127; | |
6554 | ||
6b8b2557 RS |
6555 | if (bits == 0) |
6556 | { | |
6557 | emit_move_insn (operands[0], op1); | |
6558 | return; | |
6559 | } | |
6560 | ||
6561 | if ((bits & 7) == 0) | |
6562 | { | |
6563 | rtx tmp = gen_reg_rtx (V1TImode); | |
6564 | if (code == ASHIFT) | |
1188cf5f | 6565 | emit_insn (gen_sse2_ashlv1ti3 (tmp, op1, GEN_INT (bits))); |
6b8b2557 RS |
6566 | else |
6567 | emit_insn (gen_sse2_lshrv1ti3 (tmp, op1, GEN_INT (bits))); | |
6568 | emit_move_insn (operands[0], tmp); | |
6569 | return; | |
6570 | } | |
6571 | ||
6572 | rtx tmp1 = gen_reg_rtx (V1TImode); | |
6573 | if (code == ASHIFT) | |
6574 | emit_insn (gen_sse2_ashlv1ti3 (tmp1, op1, GEN_INT (64))); | |
6575 | else | |
6576 | emit_insn (gen_sse2_lshrv1ti3 (tmp1, op1, GEN_INT (64))); | |
6577 | ||
6578 | /* tmp2 is operands[1] shifted by 64, in V2DImode. */ | |
51e9e8a2 | 6579 | rtx tmp2 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp1)); |
6b8b2557 RS |
6580 | |
6581 | /* tmp3 will be the V2DImode result. */ | |
6582 | rtx tmp3 = gen_reg_rtx (V2DImode); | |
6583 | ||
6584 | if (bits > 64) | |
6585 | { | |
6586 | if (code == ASHIFT) | |
6587 | emit_insn (gen_ashlv2di3 (tmp3, tmp2, GEN_INT (bits - 64))); | |
6588 | else | |
6589 | emit_insn (gen_lshrv2di3 (tmp3, tmp2, GEN_INT (bits - 64))); | |
6590 | } | |
6591 | else | |
6592 | { | |
6593 | /* tmp4 is operands[1], in V2DImode. */ | |
51e9e8a2 | 6594 | rtx tmp4 = force_reg (V2DImode, gen_lowpart (V2DImode, op1)); |
6b8b2557 RS |
6595 | |
6596 | rtx tmp5 = gen_reg_rtx (V2DImode); | |
6597 | if (code == ASHIFT) | |
6598 | emit_insn (gen_ashlv2di3 (tmp5, tmp4, GEN_INT (bits))); | |
6599 | else | |
6600 | emit_insn (gen_lshrv2di3 (tmp5, tmp4, GEN_INT (bits))); | |
6601 | ||
6602 | rtx tmp6 = gen_reg_rtx (V2DImode); | |
6603 | if (code == ASHIFT) | |
6604 | emit_insn (gen_lshrv2di3 (tmp6, tmp2, GEN_INT (64 - bits))); | |
6605 | else | |
6606 | emit_insn (gen_ashlv2di3 (tmp6, tmp2, GEN_INT (64 - bits))); | |
6607 | ||
6608 | emit_insn (gen_iorv2di3 (tmp3, tmp5, tmp6)); | |
6609 | } | |
6610 | ||
6611 | /* Convert the result back to V1TImode and store in operands[0]. */ | |
51e9e8a2 | 6612 | rtx tmp7 = force_reg (V1TImode, gen_lowpart (V1TImode, tmp3)); |
6b8b2557 RS |
6613 | emit_move_insn (operands[0], tmp7); |
6614 | } | |
6615 | ||
6616 | /* Expand V1TI mode rotate (of rtx_code CODE) by constant. */ | |
1188cf5f RS |
6617 | void |
6618 | ix86_expand_v1ti_rotate (enum rtx_code code, rtx operands[]) | |
6b8b2557 | 6619 | { |
6b8b2557 RS |
6620 | rtx op1 = force_reg (V1TImode, operands[1]); |
6621 | ||
1188cf5f RS |
6622 | if (!CONST_INT_P (operands[2])) |
6623 | { | |
6624 | rtx tmp1 = ix86_expand_v1ti_to_ti (op1); | |
6625 | rtx tmp2 = gen_reg_rtx (TImode); | |
6626 | rtx (*rotate) (rtx, rtx, rtx) | |
6627 | = (code == ROTATE) ? gen_rotlti3 : gen_rotrti3; | |
6628 | emit_insn (rotate (tmp2, tmp1, operands[2])); | |
6629 | rtx tmp3 = ix86_expand_ti_to_v1ti (tmp2); | |
6630 | emit_move_insn (operands[0], tmp3); | |
6631 | return; | |
6632 | } | |
6633 | ||
6634 | HOST_WIDE_INT bits = INTVAL (operands[2]) & 127; | |
6635 | ||
6b8b2557 RS |
6636 | if (bits == 0) |
6637 | { | |
6638 | emit_move_insn (operands[0], op1); | |
6639 | return; | |
6640 | } | |
6641 | ||
6642 | if (code == ROTATERT) | |
6643 | bits = 128 - bits; | |
6644 | ||
6645 | if ((bits & 31) == 0) | |
6646 | { | |
6b8b2557 | 6647 | rtx tmp2 = gen_reg_rtx (V4SImode); |
51e9e8a2 | 6648 | rtx tmp1 = force_reg (V4SImode, gen_lowpart (V4SImode, op1)); |
6b8b2557 RS |
6649 | if (bits == 32) |
6650 | emit_insn (gen_sse2_pshufd (tmp2, tmp1, GEN_INT (0x93))); | |
6651 | else if (bits == 64) | |
6652 | emit_insn (gen_sse2_pshufd (tmp2, tmp1, GEN_INT (0x4e))); | |
6653 | else | |
6654 | emit_insn (gen_sse2_pshufd (tmp2, tmp1, GEN_INT (0x39))); | |
51e9e8a2 | 6655 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp2)); |
6b8b2557 RS |
6656 | return; |
6657 | } | |
6658 | ||
6659 | if ((bits & 7) == 0) | |
6660 | { | |
6661 | rtx tmp1 = gen_reg_rtx (V1TImode); | |
6662 | rtx tmp2 = gen_reg_rtx (V1TImode); | |
6663 | rtx tmp3 = gen_reg_rtx (V1TImode); | |
6664 | ||
6665 | emit_insn (gen_sse2_ashlv1ti3 (tmp1, op1, GEN_INT (bits))); | |
6666 | emit_insn (gen_sse2_lshrv1ti3 (tmp2, op1, GEN_INT (128 - bits))); | |
6667 | emit_insn (gen_iorv1ti3 (tmp3, tmp1, tmp2)); | |
6668 | emit_move_insn (operands[0], tmp3); | |
6669 | return; | |
6670 | } | |
6671 | ||
51e9e8a2 | 6672 | rtx op1_v4si = force_reg (V4SImode, gen_lowpart (V4SImode, op1)); |
6b8b2557 RS |
6673 | |
6674 | rtx lobits; | |
6675 | rtx hibits; | |
6676 | ||
6677 | switch (bits >> 5) | |
6678 | { | |
6679 | case 0: | |
6680 | lobits = op1_v4si; | |
6681 | hibits = gen_reg_rtx (V4SImode); | |
6682 | emit_insn (gen_sse2_pshufd (hibits, op1_v4si, GEN_INT (0x93))); | |
6683 | break; | |
6684 | ||
6685 | case 1: | |
6686 | lobits = gen_reg_rtx (V4SImode); | |
6687 | hibits = gen_reg_rtx (V4SImode); | |
6688 | emit_insn (gen_sse2_pshufd (lobits, op1_v4si, GEN_INT (0x93))); | |
6689 | emit_insn (gen_sse2_pshufd (hibits, op1_v4si, GEN_INT (0x4e))); | |
6690 | break; | |
6691 | ||
6692 | case 2: | |
6693 | lobits = gen_reg_rtx (V4SImode); | |
6694 | hibits = gen_reg_rtx (V4SImode); | |
6695 | emit_insn (gen_sse2_pshufd (lobits, op1_v4si, GEN_INT (0x4e))); | |
6696 | emit_insn (gen_sse2_pshufd (hibits, op1_v4si, GEN_INT (0x39))); | |
6697 | break; | |
6698 | ||
6699 | default: | |
6700 | lobits = gen_reg_rtx (V4SImode); | |
6701 | emit_insn (gen_sse2_pshufd (lobits, op1_v4si, GEN_INT (0x39))); | |
6702 | hibits = op1_v4si; | |
6703 | break; | |
6704 | } | |
6705 | ||
6706 | rtx tmp1 = gen_reg_rtx (V4SImode); | |
6707 | rtx tmp2 = gen_reg_rtx (V4SImode); | |
6708 | rtx tmp3 = gen_reg_rtx (V4SImode); | |
6b8b2557 RS |
6709 | |
6710 | emit_insn (gen_ashlv4si3 (tmp1, lobits, GEN_INT (bits & 31))); | |
6711 | emit_insn (gen_lshrv4si3 (tmp2, hibits, GEN_INT (32 - (bits & 31)))); | |
6712 | emit_insn (gen_iorv4si3 (tmp3, tmp1, tmp2)); | |
51e9e8a2 RS |
6713 | |
6714 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp3)); | |
6b8b2557 RS |
6715 | } |
6716 | ||
1188cf5f RS |
6717 | /* Expand V1TI mode ashiftrt by constant. */ |
6718 | void | |
6719 | ix86_expand_v1ti_ashiftrt (rtx operands[]) | |
6720 | { | |
6721 | rtx op1 = force_reg (V1TImode, operands[1]); | |
6722 | ||
6723 | if (!CONST_INT_P (operands[2])) | |
6724 | { | |
6725 | rtx tmp1 = ix86_expand_v1ti_to_ti (op1); | |
6726 | rtx tmp2 = gen_reg_rtx (TImode); | |
6727 | emit_insn (gen_ashrti3 (tmp2, tmp1, operands[2])); | |
6728 | rtx tmp3 = ix86_expand_ti_to_v1ti (tmp2); | |
6729 | emit_move_insn (operands[0], tmp3); | |
6730 | return; | |
6731 | } | |
6732 | ||
6733 | HOST_WIDE_INT bits = INTVAL (operands[2]) & 127; | |
6734 | ||
6735 | if (bits == 0) | |
6736 | { | |
6737 | emit_move_insn (operands[0], op1); | |
6738 | return; | |
6739 | } | |
6740 | ||
6741 | if (bits == 127) | |
6742 | { | |
6743 | /* Two operations. */ | |
51e9e8a2 | 6744 | rtx tmp1 = force_reg(V4SImode, gen_lowpart (V4SImode, op1)); |
1188cf5f | 6745 | rtx tmp2 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6746 | emit_insn (gen_sse2_pshufd (tmp2, tmp1, GEN_INT (0xff))); |
6747 | ||
6748 | rtx tmp3 = gen_reg_rtx (V4SImode); | |
6749 | emit_insn (gen_ashrv4si3 (tmp3, tmp2, GEN_INT (31))); | |
6750 | ||
51e9e8a2 | 6751 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp3)); |
1188cf5f RS |
6752 | return; |
6753 | } | |
6754 | ||
6755 | if (bits == 64) | |
6756 | { | |
6757 | /* Three operations. */ | |
51e9e8a2 | 6758 | rtx tmp1 = force_reg(V4SImode, gen_lowpart (V4SImode, op1)); |
1188cf5f | 6759 | rtx tmp2 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6760 | emit_insn (gen_sse2_pshufd (tmp2, tmp1, GEN_INT (0xff))); |
6761 | ||
6762 | rtx tmp3 = gen_reg_rtx (V4SImode); | |
6763 | emit_insn (gen_ashrv4si3 (tmp3, tmp2, GEN_INT (31))); | |
6764 | ||
51e9e8a2 RS |
6765 | rtx tmp4 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp1)); |
6766 | rtx tmp5 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp3)); | |
1188cf5f | 6767 | rtx tmp6 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
6768 | emit_insn (gen_vec_interleave_highv2di (tmp6, tmp4, tmp5)); |
6769 | ||
51e9e8a2 | 6770 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp6)); |
1188cf5f RS |
6771 | return; |
6772 | } | |
6773 | ||
6774 | if (bits == 96) | |
6775 | { | |
6776 | /* Three operations. */ | |
51e9e8a2 | 6777 | rtx tmp1 = force_reg(V4SImode, gen_lowpart (V4SImode, op1)); |
1188cf5f | 6778 | rtx tmp2 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6779 | emit_insn (gen_ashrv4si3 (tmp2, tmp1, GEN_INT (31))); |
6780 | ||
51e9e8a2 RS |
6781 | rtx tmp3 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp1)); |
6782 | rtx tmp4 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp2)); | |
1188cf5f | 6783 | rtx tmp5 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
6784 | emit_insn (gen_vec_interleave_highv2di (tmp5, tmp3, tmp4)); |
6785 | ||
51e9e8a2 | 6786 | rtx tmp6 = force_reg(V4SImode, gen_lowpart (V4SImode, tmp5)); |
1188cf5f | 6787 | rtx tmp7 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6788 | emit_insn (gen_sse2_pshufd (tmp7, tmp6, GEN_INT (0xfd))); |
6789 | ||
51e9e8a2 RS |
6790 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp7)); |
6791 | return; | |
6792 | } | |
6793 | ||
6794 | if (bits >= 111) | |
6795 | { | |
6796 | /* Three operations. */ | |
6797 | rtx tmp1 = force_reg (V4SImode, gen_lowpart (V4SImode, op1)); | |
6798 | rtx tmp2 = gen_reg_rtx (V4SImode); | |
6799 | emit_insn (gen_ashrv4si3 (tmp2, tmp1, GEN_INT (bits - 96))); | |
6800 | ||
6801 | rtx tmp3 = force_reg (V8HImode, gen_lowpart (V8HImode, tmp2)); | |
6802 | rtx tmp4 = gen_reg_rtx (V8HImode); | |
6803 | emit_insn (gen_sse2_pshufhw (tmp4, tmp3, GEN_INT (0xfe))); | |
6804 | ||
6805 | rtx tmp5 = force_reg (V4SImode, gen_lowpart (V4SImode, tmp4)); | |
6806 | rtx tmp6 = gen_reg_rtx (V4SImode); | |
6807 | emit_insn (gen_sse2_pshufd (tmp6, tmp5, GEN_INT (0xfe))); | |
6808 | ||
6809 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp6)); | |
1188cf5f RS |
6810 | return; |
6811 | } | |
6812 | ||
6813 | if (TARGET_AVX2 || TARGET_SSE4_1) | |
6814 | { | |
6815 | /* Three operations. */ | |
6816 | if (bits == 32) | |
6817 | { | |
51e9e8a2 | 6818 | rtx tmp1 = force_reg (V4SImode, gen_lowpart (V4SImode, op1)); |
1188cf5f | 6819 | rtx tmp2 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6820 | emit_insn (gen_ashrv4si3 (tmp2, tmp1, GEN_INT (31))); |
6821 | ||
6822 | rtx tmp3 = gen_reg_rtx (V1TImode); | |
6823 | emit_insn (gen_sse2_lshrv1ti3 (tmp3, op1, GEN_INT (32))); | |
6824 | ||
6825 | if (TARGET_AVX2) | |
6826 | { | |
51e9e8a2 | 6827 | rtx tmp4 = force_reg (V4SImode, gen_lowpart (V4SImode, tmp3)); |
1188cf5f | 6828 | rtx tmp5 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6829 | emit_insn (gen_avx2_pblenddv4si (tmp5, tmp2, tmp4, |
6830 | GEN_INT (7))); | |
6831 | ||
51e9e8a2 | 6832 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp5)); |
1188cf5f RS |
6833 | } |
6834 | else | |
6835 | { | |
51e9e8a2 RS |
6836 | rtx tmp4 = force_reg (V8HImode, gen_lowpart (V8HImode, tmp2)); |
6837 | rtx tmp5 = force_reg (V8HImode, gen_lowpart (V8HImode, tmp3)); | |
1188cf5f | 6838 | rtx tmp6 = gen_reg_rtx (V8HImode); |
1188cf5f RS |
6839 | emit_insn (gen_sse4_1_pblendw (tmp6, tmp4, tmp5, |
6840 | GEN_INT (0x3f))); | |
6841 | ||
51e9e8a2 | 6842 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp6)); |
1188cf5f RS |
6843 | } |
6844 | return; | |
6845 | } | |
6846 | ||
6847 | /* Three operations. */ | |
6848 | if (bits == 8 || bits == 16 || bits == 24) | |
6849 | { | |
51e9e8a2 | 6850 | rtx tmp1 = force_reg (V4SImode, gen_lowpart (V4SImode, op1)); |
1188cf5f | 6851 | rtx tmp2 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6852 | emit_insn (gen_ashrv4si3 (tmp2, tmp1, GEN_INT (bits))); |
6853 | ||
6854 | rtx tmp3 = gen_reg_rtx (V1TImode); | |
6855 | emit_insn (gen_sse2_lshrv1ti3 (tmp3, op1, GEN_INT (bits))); | |
6856 | ||
6857 | if (TARGET_AVX2) | |
6858 | { | |
51e9e8a2 | 6859 | rtx tmp4 = force_reg (V4SImode, gen_lowpart (V4SImode, tmp3)); |
1188cf5f | 6860 | rtx tmp5 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6861 | emit_insn (gen_avx2_pblenddv4si (tmp5, tmp2, tmp4, |
6862 | GEN_INT (7))); | |
6863 | ||
51e9e8a2 | 6864 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp5)); |
1188cf5f RS |
6865 | } |
6866 | else | |
6867 | { | |
51e9e8a2 RS |
6868 | rtx tmp4 = force_reg (V8HImode, gen_lowpart (V8HImode, tmp2)); |
6869 | rtx tmp5 = force_reg (V8HImode, gen_lowpart (V8HImode, tmp3)); | |
1188cf5f | 6870 | rtx tmp6 = gen_reg_rtx (V8HImode); |
1188cf5f RS |
6871 | emit_insn (gen_sse4_1_pblendw (tmp6, tmp4, tmp5, |
6872 | GEN_INT (0x3f))); | |
6873 | ||
51e9e8a2 | 6874 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp6)); |
1188cf5f RS |
6875 | } |
6876 | return; | |
6877 | } | |
6878 | } | |
6879 | ||
6880 | if (bits > 96) | |
6881 | { | |
6882 | /* Four operations. */ | |
51e9e8a2 | 6883 | rtx tmp1 = force_reg (V4SImode, gen_lowpart (V4SImode, op1)); |
1188cf5f | 6884 | rtx tmp2 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6885 | emit_insn (gen_ashrv4si3 (tmp2, tmp1, GEN_INT (bits - 96))); |
6886 | ||
6887 | rtx tmp3 = gen_reg_rtx (V4SImode); | |
6888 | emit_insn (gen_ashrv4si3 (tmp3, tmp1, GEN_INT (31))); | |
6889 | ||
51e9e8a2 RS |
6890 | rtx tmp4 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp2)); |
6891 | rtx tmp5 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp3)); | |
1188cf5f | 6892 | rtx tmp6 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
6893 | emit_insn (gen_vec_interleave_highv2di (tmp6, tmp4, tmp5)); |
6894 | ||
51e9e8a2 | 6895 | rtx tmp7 = force_reg (V4SImode, gen_lowpart (V4SImode, tmp6)); |
1188cf5f | 6896 | rtx tmp8 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6897 | emit_insn (gen_sse2_pshufd (tmp8, tmp7, GEN_INT (0xfd))); |
6898 | ||
51e9e8a2 | 6899 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp8)); |
1188cf5f RS |
6900 | return; |
6901 | } | |
6902 | ||
6903 | if (TARGET_SSE4_1 && (bits == 48 || bits == 80)) | |
6904 | { | |
6905 | /* Four operations. */ | |
51e9e8a2 | 6906 | rtx tmp1 = force_reg (V4SImode, gen_lowpart (V4SImode, op1)); |
1188cf5f | 6907 | rtx tmp2 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6908 | emit_insn (gen_sse2_pshufd (tmp2, tmp1, GEN_INT (0xff))); |
6909 | ||
6910 | rtx tmp3 = gen_reg_rtx (V4SImode); | |
6911 | emit_insn (gen_ashrv4si3 (tmp3, tmp2, GEN_INT (31))); | |
6912 | ||
6913 | rtx tmp4 = gen_reg_rtx (V1TImode); | |
6914 | emit_insn (gen_sse2_lshrv1ti3 (tmp4, op1, GEN_INT (bits))); | |
6915 | ||
51e9e8a2 RS |
6916 | rtx tmp5 = force_reg (V8HImode, gen_lowpart (V8HImode, tmp3)); |
6917 | rtx tmp6 = force_reg (V8HImode, gen_lowpart (V8HImode, tmp4)); | |
1188cf5f | 6918 | rtx tmp7 = gen_reg_rtx (V8HImode); |
1188cf5f RS |
6919 | emit_insn (gen_sse4_1_pblendw (tmp7, tmp5, tmp6, |
6920 | GEN_INT (bits == 48 ? 0x1f : 0x07))); | |
6921 | ||
51e9e8a2 | 6922 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp7)); |
1188cf5f RS |
6923 | return; |
6924 | } | |
6925 | ||
6926 | if ((bits & 7) == 0) | |
6927 | { | |
6928 | /* Five operations. */ | |
51e9e8a2 | 6929 | rtx tmp1 = force_reg (V4SImode, gen_lowpart (V4SImode, op1)); |
1188cf5f | 6930 | rtx tmp2 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6931 | emit_insn (gen_sse2_pshufd (tmp2, tmp1, GEN_INT (0xff))); |
6932 | ||
6933 | rtx tmp3 = gen_reg_rtx (V4SImode); | |
6934 | emit_insn (gen_ashrv4si3 (tmp3, tmp2, GEN_INT (31))); | |
6935 | ||
6936 | rtx tmp4 = gen_reg_rtx (V1TImode); | |
6937 | emit_insn (gen_sse2_lshrv1ti3 (tmp4, op1, GEN_INT (bits))); | |
6938 | ||
51e9e8a2 | 6939 | rtx tmp5 = force_reg (V1TImode, gen_lowpart (V1TImode, tmp3)); |
1188cf5f | 6940 | rtx tmp6 = gen_reg_rtx (V1TImode); |
1188cf5f RS |
6941 | emit_insn (gen_sse2_ashlv1ti3 (tmp6, tmp5, GEN_INT (128 - bits))); |
6942 | ||
51e9e8a2 RS |
6943 | rtx tmp7 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp4)); |
6944 | rtx tmp8 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp6)); | |
1188cf5f | 6945 | rtx tmp9 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
6946 | emit_insn (gen_iorv2di3 (tmp9, tmp7, tmp8)); |
6947 | ||
51e9e8a2 | 6948 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp9)); |
1188cf5f RS |
6949 | return; |
6950 | } | |
6951 | ||
6952 | if (TARGET_AVX2 && bits < 32) | |
6953 | { | |
6954 | /* Six operations. */ | |
51e9e8a2 | 6955 | rtx tmp1 = force_reg (V4SImode, gen_lowpart (V4SImode, op1)); |
1188cf5f | 6956 | rtx tmp2 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6957 | emit_insn (gen_ashrv4si3 (tmp2, tmp1, GEN_INT (bits))); |
6958 | ||
6959 | rtx tmp3 = gen_reg_rtx (V1TImode); | |
6960 | emit_insn (gen_sse2_lshrv1ti3 (tmp3, op1, GEN_INT (64))); | |
6961 | ||
51e9e8a2 | 6962 | rtx tmp4 = force_reg (V2DImode, gen_lowpart (V2DImode, op1)); |
1188cf5f | 6963 | rtx tmp5 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
6964 | emit_insn (gen_lshrv2di3 (tmp5, tmp4, GEN_INT (bits))); |
6965 | ||
51e9e8a2 | 6966 | rtx tmp6 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp3)); |
1188cf5f | 6967 | rtx tmp7 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
6968 | emit_insn (gen_ashlv2di3 (tmp7, tmp6, GEN_INT (64 - bits))); |
6969 | ||
6970 | rtx tmp8 = gen_reg_rtx (V2DImode); | |
6971 | emit_insn (gen_iorv2di3 (tmp8, tmp5, tmp7)); | |
6972 | ||
51e9e8a2 | 6973 | rtx tmp9 = force_reg (V4SImode, gen_lowpart (V4SImode, tmp8)); |
1188cf5f | 6974 | rtx tmp10 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6975 | emit_insn (gen_avx2_pblenddv4si (tmp10, tmp2, tmp9, GEN_INT (7))); |
6976 | ||
51e9e8a2 | 6977 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp10)); |
1188cf5f RS |
6978 | return; |
6979 | } | |
6980 | ||
6981 | if (TARGET_SSE4_1 && bits < 15) | |
6982 | { | |
6983 | /* Six operations. */ | |
51e9e8a2 | 6984 | rtx tmp1 = force_reg (V4SImode, gen_lowpart (V4SImode, op1)); |
1188cf5f | 6985 | rtx tmp2 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
6986 | emit_insn (gen_ashrv4si3 (tmp2, tmp1, GEN_INT (bits))); |
6987 | ||
6988 | rtx tmp3 = gen_reg_rtx (V1TImode); | |
6989 | emit_insn (gen_sse2_lshrv1ti3 (tmp3, op1, GEN_INT (64))); | |
6990 | ||
51e9e8a2 | 6991 | rtx tmp4 = force_reg (V2DImode, gen_lowpart (V2DImode, op1)); |
1188cf5f | 6992 | rtx tmp5 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
6993 | emit_insn (gen_lshrv2di3 (tmp5, tmp4, GEN_INT (bits))); |
6994 | ||
51e9e8a2 | 6995 | rtx tmp6 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp3)); |
1188cf5f | 6996 | rtx tmp7 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
6997 | emit_insn (gen_ashlv2di3 (tmp7, tmp6, GEN_INT (64 - bits))); |
6998 | ||
6999 | rtx tmp8 = gen_reg_rtx (V2DImode); | |
7000 | emit_insn (gen_iorv2di3 (tmp8, tmp5, tmp7)); | |
7001 | ||
51e9e8a2 RS |
7002 | rtx tmp9 = force_reg (V8HImode, gen_lowpart (V8HImode, tmp2)); |
7003 | rtx tmp10 = force_reg (V8HImode, gen_lowpart (V8HImode, tmp8)); | |
1188cf5f | 7004 | rtx tmp11 = gen_reg_rtx (V8HImode); |
1188cf5f RS |
7005 | emit_insn (gen_sse4_1_pblendw (tmp11, tmp9, tmp10, GEN_INT (0x3f))); |
7006 | ||
51e9e8a2 | 7007 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp11)); |
1188cf5f RS |
7008 | return; |
7009 | } | |
7010 | ||
7011 | if (bits == 1) | |
7012 | { | |
7013 | /* Eight operations. */ | |
7014 | rtx tmp1 = gen_reg_rtx (V1TImode); | |
7015 | emit_insn (gen_sse2_lshrv1ti3 (tmp1, op1, GEN_INT (64))); | |
7016 | ||
51e9e8a2 | 7017 | rtx tmp2 = force_reg (V2DImode, gen_lowpart (V2DImode, op1)); |
1188cf5f | 7018 | rtx tmp3 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
7019 | emit_insn (gen_lshrv2di3 (tmp3, tmp2, GEN_INT (1))); |
7020 | ||
51e9e8a2 | 7021 | rtx tmp4 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp1)); |
1188cf5f | 7022 | rtx tmp5 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
7023 | emit_insn (gen_ashlv2di3 (tmp5, tmp4, GEN_INT (63))); |
7024 | ||
7025 | rtx tmp6 = gen_reg_rtx (V2DImode); | |
7026 | emit_insn (gen_iorv2di3 (tmp6, tmp3, tmp5)); | |
7027 | ||
7028 | rtx tmp7 = gen_reg_rtx (V2DImode); | |
7029 | emit_insn (gen_lshrv2di3 (tmp7, tmp2, GEN_INT (63))); | |
7030 | ||
51e9e8a2 | 7031 | rtx tmp8 = force_reg (V4SImode, gen_lowpart (V4SImode, tmp7)); |
1188cf5f | 7032 | rtx tmp9 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
7033 | emit_insn (gen_sse2_pshufd (tmp9, tmp8, GEN_INT (0xbf))); |
7034 | ||
51e9e8a2 | 7035 | rtx tmp10 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp9)); |
1188cf5f | 7036 | rtx tmp11 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
7037 | emit_insn (gen_ashlv2di3 (tmp11, tmp10, GEN_INT (31))); |
7038 | ||
7039 | rtx tmp12 = gen_reg_rtx (V2DImode); | |
7040 | emit_insn (gen_iorv2di3 (tmp12, tmp6, tmp11)); | |
7041 | ||
51e9e8a2 | 7042 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp12)); |
1188cf5f RS |
7043 | return; |
7044 | } | |
7045 | ||
7046 | if (bits > 64) | |
7047 | { | |
7048 | /* Eight operations. */ | |
51e9e8a2 | 7049 | rtx tmp1 = force_reg (V4SImode, gen_lowpart (V4SImode, op1)); |
1188cf5f | 7050 | rtx tmp2 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
7051 | emit_insn (gen_sse2_pshufd (tmp2, tmp1, GEN_INT (0xff))); |
7052 | ||
7053 | rtx tmp3 = gen_reg_rtx (V4SImode); | |
7054 | emit_insn (gen_ashrv4si3 (tmp3, tmp2, GEN_INT (31))); | |
7055 | ||
7056 | rtx tmp4 = gen_reg_rtx (V1TImode); | |
7057 | emit_insn (gen_sse2_lshrv1ti3 (tmp4, op1, GEN_INT (64))); | |
7058 | ||
51e9e8a2 | 7059 | rtx tmp5 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp4)); |
1188cf5f | 7060 | rtx tmp6 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
7061 | emit_insn (gen_lshrv2di3 (tmp6, tmp5, GEN_INT (bits - 64))); |
7062 | ||
51e9e8a2 | 7063 | rtx tmp7 = force_reg (V1TImode, gen_lowpart (V1TImode, tmp3)); |
1188cf5f | 7064 | rtx tmp8 = gen_reg_rtx (V1TImode); |
1188cf5f RS |
7065 | emit_insn (gen_sse2_ashlv1ti3 (tmp8, tmp7, GEN_INT (64))); |
7066 | ||
51e9e8a2 | 7067 | rtx tmp9 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp3)); |
1188cf5f | 7068 | rtx tmp10 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
7069 | emit_insn (gen_ashlv2di3 (tmp10, tmp9, GEN_INT (128 - bits))); |
7070 | ||
51e9e8a2 | 7071 | rtx tmp11 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp8)); |
1188cf5f | 7072 | rtx tmp12 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
7073 | emit_insn (gen_iorv2di3 (tmp12, tmp10, tmp11)); |
7074 | ||
7075 | rtx tmp13 = gen_reg_rtx (V2DImode); | |
7076 | emit_insn (gen_iorv2di3 (tmp13, tmp6, tmp12)); | |
7077 | ||
51e9e8a2 | 7078 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp13)); |
1188cf5f RS |
7079 | } |
7080 | else | |
7081 | { | |
7082 | /* Nine operations. */ | |
51e9e8a2 | 7083 | rtx tmp1 = force_reg (V4SImode, gen_lowpart (V4SImode, op1)); |
1188cf5f | 7084 | rtx tmp2 = gen_reg_rtx (V4SImode); |
1188cf5f RS |
7085 | emit_insn (gen_sse2_pshufd (tmp2, tmp1, GEN_INT (0xff))); |
7086 | ||
7087 | rtx tmp3 = gen_reg_rtx (V4SImode); | |
7088 | emit_insn (gen_ashrv4si3 (tmp3, tmp2, GEN_INT (31))); | |
7089 | ||
7090 | rtx tmp4 = gen_reg_rtx (V1TImode); | |
7091 | emit_insn (gen_sse2_lshrv1ti3 (tmp4, op1, GEN_INT (64))); | |
7092 | ||
51e9e8a2 | 7093 | rtx tmp5 = force_reg (V2DImode, gen_lowpart (V2DImode, op1)); |
1188cf5f | 7094 | rtx tmp6 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
7095 | emit_insn (gen_lshrv2di3 (tmp6, tmp5, GEN_INT (bits))); |
7096 | ||
51e9e8a2 | 7097 | rtx tmp7 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp4)); |
1188cf5f | 7098 | rtx tmp8 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
7099 | emit_insn (gen_ashlv2di3 (tmp8, tmp7, GEN_INT (64 - bits))); |
7100 | ||
7101 | rtx tmp9 = gen_reg_rtx (V2DImode); | |
7102 | emit_insn (gen_iorv2di3 (tmp9, tmp6, tmp8)); | |
7103 | ||
51e9e8a2 | 7104 | rtx tmp10 = force_reg (V1TImode, gen_lowpart (V1TImode, tmp3)); |
1188cf5f | 7105 | rtx tmp11 = gen_reg_rtx (V1TImode); |
1188cf5f RS |
7106 | emit_insn (gen_sse2_ashlv1ti3 (tmp11, tmp10, GEN_INT (64))); |
7107 | ||
51e9e8a2 | 7108 | rtx tmp12 = force_reg (V2DImode, gen_lowpart (V2DImode, tmp11)); |
1188cf5f | 7109 | rtx tmp13 = gen_reg_rtx (V2DImode); |
1188cf5f RS |
7110 | emit_insn (gen_ashlv2di3 (tmp13, tmp12, GEN_INT (64 - bits))); |
7111 | ||
7112 | rtx tmp14 = gen_reg_rtx (V2DImode); | |
7113 | emit_insn (gen_iorv2di3 (tmp14, tmp9, tmp13)); | |
7114 | ||
51e9e8a2 | 7115 | emit_move_insn (operands[0], gen_lowpart (V1TImode, tmp14)); |
1188cf5f RS |
7116 | } |
7117 | } | |
7118 | ||
3c135697 JJ |
7119 | /* Replace all occurrences of REG FROM with REG TO in X, including |
7120 | occurrences with different modes. */ | |
7121 | ||
7122 | rtx | |
7123 | ix86_replace_reg_with_reg (rtx x, rtx from, rtx to) | |
7124 | { | |
7125 | gcc_checking_assert (REG_P (from) | |
7126 | && REG_P (to) | |
7127 | && GET_MODE (from) == GET_MODE (to)); | |
7128 | if (!reg_overlap_mentioned_p (from, x)) | |
7129 | return x; | |
7130 | rtx ret = copy_rtx (x); | |
7131 | subrtx_ptr_iterator::array_type array; | |
7132 | FOR_EACH_SUBRTX_PTR (iter, array, &ret, NONCONST) | |
7133 | { | |
7134 | rtx *loc = *iter; | |
7135 | x = *loc; | |
7136 | if (REG_P (x) && REGNO (x) == REGNO (from)) | |
7137 | { | |
7138 | if (x == from) | |
7139 | *loc = to; | |
7140 | else | |
7141 | { | |
7142 | gcc_checking_assert (REG_NREGS (x) == 1); | |
7143 | *loc = gen_rtx_REG (GET_MODE (x), REGNO (to)); | |
7144 | } | |
7145 | } | |
7146 | } | |
7147 | return ret; | |
7148 | } | |
7149 | ||
2bf6d935 ML |
7150 | /* Return mode for the memcpy/memset loop counter. Prefer SImode over |
7151 | DImode for constant loop counts. */ | |
7152 | ||
7153 | static machine_mode | |
7154 | counter_mode (rtx count_exp) | |
7155 | { | |
7156 | if (GET_MODE (count_exp) != VOIDmode) | |
7157 | return GET_MODE (count_exp); | |
7158 | if (!CONST_INT_P (count_exp)) | |
7159 | return Pmode; | |
7160 | if (TARGET_64BIT && (INTVAL (count_exp) & ~0xffffffff)) | |
7161 | return DImode; | |
7162 | return SImode; | |
7163 | } | |
7164 | ||
7165 | /* When ISSETMEM is FALSE, output simple loop to move memory pointer to SRCPTR | |
7166 | to DESTPTR via chunks of MODE unrolled UNROLL times, overall size is COUNT | |
7167 | specified in bytes. When ISSETMEM is TRUE, output the equivalent loop to set | |
7168 | memory by VALUE (supposed to be in MODE). | |
7169 | ||
7170 | The size is rounded down to whole number of chunk size moved at once. | |
7171 | SRCMEM and DESTMEM provide MEMrtx to feed proper aliasing info. */ | |
7172 | ||
7173 | ||
7174 | static void | |
76715c32 | 7175 | expand_set_or_cpymem_via_loop (rtx destmem, rtx srcmem, |
2bf6d935 ML |
7176 | rtx destptr, rtx srcptr, rtx value, |
7177 | rtx count, machine_mode mode, int unroll, | |
7178 | int expected_size, bool issetmem) | |
7179 | { | |
7180 | rtx_code_label *out_label, *top_label; | |
7181 | rtx iter, tmp; | |
7182 | machine_mode iter_mode = counter_mode (count); | |
7183 | int piece_size_n = GET_MODE_SIZE (mode) * unroll; | |
7184 | rtx piece_size = GEN_INT (piece_size_n); | |
7185 | rtx piece_size_mask = GEN_INT (~((GET_MODE_SIZE (mode) * unroll) - 1)); | |
7186 | rtx size; | |
7187 | int i; | |
7188 | ||
7189 | top_label = gen_label_rtx (); | |
7190 | out_label = gen_label_rtx (); | |
7191 | iter = gen_reg_rtx (iter_mode); | |
7192 | ||
7193 | size = expand_simple_binop (iter_mode, AND, count, piece_size_mask, | |
7194 | NULL, 1, OPTAB_DIRECT); | |
7195 | /* Those two should combine. */ | |
7196 | if (piece_size == const1_rtx) | |
7197 | { | |
7198 | emit_cmp_and_jump_insns (size, const0_rtx, EQ, NULL_RTX, iter_mode, | |
7199 | true, out_label); | |
7200 | predict_jump (REG_BR_PROB_BASE * 10 / 100); | |
7201 | } | |
7202 | emit_move_insn (iter, const0_rtx); | |
7203 | ||
7204 | emit_label (top_label); | |
7205 | ||
7206 | tmp = convert_modes (Pmode, iter_mode, iter, true); | |
7207 | ||
7208 | /* This assert could be relaxed - in this case we'll need to compute | |
7209 | smallest power of two, containing in PIECE_SIZE_N and pass it to | |
7210 | offset_address. */ | |
7211 | gcc_assert ((piece_size_n & (piece_size_n - 1)) == 0); | |
7212 | destmem = offset_address (destmem, tmp, piece_size_n); | |
7213 | destmem = adjust_address (destmem, mode, 0); | |
7214 | ||
7215 | if (!issetmem) | |
7216 | { | |
7217 | srcmem = offset_address (srcmem, copy_rtx (tmp), piece_size_n); | |
7218 | srcmem = adjust_address (srcmem, mode, 0); | |
7219 | ||
7220 | /* When unrolling for chips that reorder memory reads and writes, | |
7221 | we can save registers by using single temporary. | |
7222 | Also using 4 temporaries is overkill in 32bit mode. */ | |
7223 | if (!TARGET_64BIT && 0) | |
7224 | { | |
7225 | for (i = 0; i < unroll; i++) | |
7226 | { | |
7227 | if (i) | |
7228 | { | |
7229 | destmem = adjust_address (copy_rtx (destmem), mode, | |
7230 | GET_MODE_SIZE (mode)); | |
7231 | srcmem = adjust_address (copy_rtx (srcmem), mode, | |
7232 | GET_MODE_SIZE (mode)); | |
7233 | } | |
7234 | emit_move_insn (destmem, srcmem); | |
7235 | } | |
7236 | } | |
7237 | else | |
7238 | { | |
7239 | rtx tmpreg[4]; | |
7240 | gcc_assert (unroll <= 4); | |
7241 | for (i = 0; i < unroll; i++) | |
7242 | { | |
7243 | tmpreg[i] = gen_reg_rtx (mode); | |
7244 | if (i) | |
7245 | srcmem = adjust_address (copy_rtx (srcmem), mode, | |
7246 | GET_MODE_SIZE (mode)); | |
7247 | emit_move_insn (tmpreg[i], srcmem); | |
7248 | } | |
7249 | for (i = 0; i < unroll; i++) | |
7250 | { | |
7251 | if (i) | |
7252 | destmem = adjust_address (copy_rtx (destmem), mode, | |
7253 | GET_MODE_SIZE (mode)); | |
7254 | emit_move_insn (destmem, tmpreg[i]); | |
7255 | } | |
7256 | } | |
7257 | } | |
7258 | else | |
7259 | for (i = 0; i < unroll; i++) | |
7260 | { | |
7261 | if (i) | |
7262 | destmem = adjust_address (copy_rtx (destmem), mode, | |
7263 | GET_MODE_SIZE (mode)); | |
7264 | emit_move_insn (destmem, value); | |
7265 | } | |
7266 | ||
7267 | tmp = expand_simple_binop (iter_mode, PLUS, iter, piece_size, iter, | |
7268 | true, OPTAB_LIB_WIDEN); | |
7269 | if (tmp != iter) | |
7270 | emit_move_insn (iter, tmp); | |
7271 | ||
7272 | emit_cmp_and_jump_insns (iter, size, LT, NULL_RTX, iter_mode, | |
7273 | true, top_label); | |
7274 | if (expected_size != -1) | |
7275 | { | |
7276 | expected_size /= GET_MODE_SIZE (mode) * unroll; | |
7277 | if (expected_size == 0) | |
7278 | predict_jump (0); | |
7279 | else if (expected_size > REG_BR_PROB_BASE) | |
7280 | predict_jump (REG_BR_PROB_BASE - 1); | |
7281 | else | |
7282 | predict_jump (REG_BR_PROB_BASE - (REG_BR_PROB_BASE + expected_size / 2) | |
7283 | / expected_size); | |
7284 | } | |
7285 | else | |
7286 | predict_jump (REG_BR_PROB_BASE * 80 / 100); | |
7287 | iter = ix86_zero_extend_to_Pmode (iter); | |
7288 | tmp = expand_simple_binop (Pmode, PLUS, destptr, iter, destptr, | |
7289 | true, OPTAB_LIB_WIDEN); | |
7290 | if (tmp != destptr) | |
7291 | emit_move_insn (destptr, tmp); | |
7292 | if (!issetmem) | |
7293 | { | |
7294 | tmp = expand_simple_binop (Pmode, PLUS, srcptr, iter, srcptr, | |
7295 | true, OPTAB_LIB_WIDEN); | |
7296 | if (tmp != srcptr) | |
7297 | emit_move_insn (srcptr, tmp); | |
7298 | } | |
7299 | emit_label (out_label); | |
7300 | } | |
7301 | ||
7302 | /* Divide COUNTREG by SCALE. */ | |
7303 | static rtx | |
7304 | scale_counter (rtx countreg, int scale) | |
7305 | { | |
7306 | rtx sc; | |
7307 | ||
7308 | if (scale == 1) | |
7309 | return countreg; | |
7310 | if (CONST_INT_P (countreg)) | |
7311 | return GEN_INT (INTVAL (countreg) / scale); | |
7312 | gcc_assert (REG_P (countreg)); | |
7313 | ||
7314 | sc = expand_simple_binop (GET_MODE (countreg), LSHIFTRT, countreg, | |
7315 | GEN_INT (exact_log2 (scale)), | |
7316 | NULL, 1, OPTAB_DIRECT); | |
7317 | return sc; | |
7318 | } | |
7319 | ||
7320 | /* Output "rep; mov" or "rep; stos" instruction depending on ISSETMEM argument. | |
7321 | When ISSETMEM is true, arguments SRCMEM and SRCPTR are ignored. | |
7322 | When ISSETMEM is false, arguments VALUE and ORIG_VALUE are ignored. | |
7323 | For setmem case, VALUE is a promoted to a wider size ORIG_VALUE. | |
7324 | ORIG_VALUE is the original value passed to memset to fill the memory with. | |
7325 | Other arguments have same meaning as for previous function. */ | |
7326 | ||
7327 | static void | |
76715c32 | 7328 | expand_set_or_cpymem_via_rep (rtx destmem, rtx srcmem, |
2bf6d935 ML |
7329 | rtx destptr, rtx srcptr, rtx value, rtx orig_value, |
7330 | rtx count, | |
7331 | machine_mode mode, bool issetmem) | |
7332 | { | |
7333 | rtx destexp; | |
7334 | rtx srcexp; | |
7335 | rtx countreg; | |
7336 | HOST_WIDE_INT rounded_count; | |
7337 | ||
7338 | /* If possible, it is shorter to use rep movs. | |
7339 | TODO: Maybe it is better to move this logic to decide_alg. */ | |
7340 | if (mode == QImode && CONST_INT_P (count) && !(INTVAL (count) & 3) | |
bf24f4ec | 7341 | && !TARGET_PREFER_KNOWN_REP_MOVSB_STOSB |
2bf6d935 ML |
7342 | && (!issetmem || orig_value == const0_rtx)) |
7343 | mode = SImode; | |
7344 | ||
7345 | if (destptr != XEXP (destmem, 0) || GET_MODE (destmem) != BLKmode) | |
7346 | destmem = adjust_automodify_address_nv (destmem, BLKmode, destptr, 0); | |
7347 | ||
7348 | countreg = ix86_zero_extend_to_Pmode (scale_counter (count, | |
7349 | GET_MODE_SIZE (mode))); | |
7350 | if (mode != QImode) | |
7351 | { | |
7352 | destexp = gen_rtx_ASHIFT (Pmode, countreg, | |
7353 | GEN_INT (exact_log2 (GET_MODE_SIZE (mode)))); | |
7354 | destexp = gen_rtx_PLUS (Pmode, destexp, destptr); | |
7355 | } | |
7356 | else | |
7357 | destexp = gen_rtx_PLUS (Pmode, destptr, countreg); | |
7358 | if ((!issetmem || orig_value == const0_rtx) && CONST_INT_P (count)) | |
7359 | { | |
7360 | rounded_count | |
7361 | = ROUND_DOWN (INTVAL (count), (HOST_WIDE_INT) GET_MODE_SIZE (mode)); | |
7362 | destmem = shallow_copy_rtx (destmem); | |
7363 | set_mem_size (destmem, rounded_count); | |
7364 | } | |
7365 | else if (MEM_SIZE_KNOWN_P (destmem)) | |
7366 | clear_mem_size (destmem); | |
7367 | ||
7368 | if (issetmem) | |
7369 | { | |
7370 | value = force_reg (mode, gen_lowpart (mode, value)); | |
7371 | emit_insn (gen_rep_stos (destptr, countreg, destmem, value, destexp)); | |
7372 | } | |
7373 | else | |
7374 | { | |
7375 | if (srcptr != XEXP (srcmem, 0) || GET_MODE (srcmem) != BLKmode) | |
7376 | srcmem = adjust_automodify_address_nv (srcmem, BLKmode, srcptr, 0); | |
7377 | if (mode != QImode) | |
7378 | { | |
7379 | srcexp = gen_rtx_ASHIFT (Pmode, countreg, | |
7380 | GEN_INT (exact_log2 (GET_MODE_SIZE (mode)))); | |
7381 | srcexp = gen_rtx_PLUS (Pmode, srcexp, srcptr); | |
7382 | } | |
7383 | else | |
7384 | srcexp = gen_rtx_PLUS (Pmode, srcptr, countreg); | |
7385 | if (CONST_INT_P (count)) | |
7386 | { | |
7387 | rounded_count | |
7388 | = ROUND_DOWN (INTVAL (count), (HOST_WIDE_INT) GET_MODE_SIZE (mode)); | |
7389 | srcmem = shallow_copy_rtx (srcmem); | |
7390 | set_mem_size (srcmem, rounded_count); | |
7391 | } | |
7392 | else | |
7393 | { | |
7394 | if (MEM_SIZE_KNOWN_P (srcmem)) | |
7395 | clear_mem_size (srcmem); | |
7396 | } | |
7397 | emit_insn (gen_rep_mov (destptr, destmem, srcptr, srcmem, countreg, | |
7398 | destexp, srcexp)); | |
7399 | } | |
7400 | } | |
7401 | ||
7402 | /* This function emits moves to copy SIZE_TO_MOVE bytes from SRCMEM to | |
7403 | DESTMEM. | |
7404 | SRC is passed by pointer to be updated on return. | |
7405 | Return value is updated DST. */ | |
7406 | static rtx | |
7407 | emit_memmov (rtx destmem, rtx *srcmem, rtx destptr, rtx srcptr, | |
7408 | HOST_WIDE_INT size_to_move) | |
7409 | { | |
c3185b64 | 7410 | rtx dst = destmem, src = *srcmem, tempreg; |
2bf6d935 ML |
7411 | enum insn_code code; |
7412 | machine_mode move_mode; | |
7413 | int piece_size, i; | |
7414 | ||
7415 | /* Find the widest mode in which we could perform moves. | |
7416 | Start with the biggest power of 2 less than SIZE_TO_MOVE and half | |
7417 | it until move of such size is supported. */ | |
7418 | piece_size = 1 << floor_log2 (size_to_move); | |
7419 | while (!int_mode_for_size (piece_size * BITS_PER_UNIT, 0).exists (&move_mode) | |
7420 | || (code = optab_handler (mov_optab, move_mode)) == CODE_FOR_nothing) | |
7421 | { | |
7422 | gcc_assert (piece_size > 1); | |
7423 | piece_size >>= 1; | |
7424 | } | |
7425 | ||
7426 | /* Find the corresponding vector mode with the same size as MOVE_MODE. | |
7427 | MOVE_MODE is an integer mode at the moment (SI, DI, TI, etc.). */ | |
7428 | if (GET_MODE_SIZE (move_mode) > GET_MODE_SIZE (word_mode)) | |
7429 | { | |
7430 | int nunits = GET_MODE_SIZE (move_mode) / GET_MODE_SIZE (word_mode); | |
7431 | if (!mode_for_vector (word_mode, nunits).exists (&move_mode) | |
7432 | || (code = optab_handler (mov_optab, move_mode)) == CODE_FOR_nothing) | |
7433 | { | |
7434 | move_mode = word_mode; | |
7435 | piece_size = GET_MODE_SIZE (move_mode); | |
7436 | code = optab_handler (mov_optab, move_mode); | |
7437 | } | |
7438 | } | |
7439 | gcc_assert (code != CODE_FOR_nothing); | |
7440 | ||
7441 | dst = adjust_automodify_address_nv (dst, move_mode, destptr, 0); | |
7442 | src = adjust_automodify_address_nv (src, move_mode, srcptr, 0); | |
7443 | ||
7444 | /* Emit moves. We'll need SIZE_TO_MOVE/PIECE_SIZES moves. */ | |
7445 | gcc_assert (size_to_move % piece_size == 0); | |
c3185b64 | 7446 | |
2bf6d935 ML |
7447 | for (i = 0; i < size_to_move; i += piece_size) |
7448 | { | |
7449 | /* We move from memory to memory, so we'll need to do it via | |
7450 | a temporary register. */ | |
7451 | tempreg = gen_reg_rtx (move_mode); | |
7452 | emit_insn (GEN_FCN (code) (tempreg, src)); | |
7453 | emit_insn (GEN_FCN (code) (dst, tempreg)); | |
7454 | ||
7455 | emit_move_insn (destptr, | |
c3185b64 | 7456 | plus_constant (Pmode, copy_rtx (destptr), piece_size)); |
2bf6d935 | 7457 | emit_move_insn (srcptr, |
c3185b64 | 7458 | plus_constant (Pmode, copy_rtx (srcptr), piece_size)); |
2bf6d935 ML |
7459 | |
7460 | dst = adjust_automodify_address_nv (dst, move_mode, destptr, | |
7461 | piece_size); | |
7462 | src = adjust_automodify_address_nv (src, move_mode, srcptr, | |
7463 | piece_size); | |
7464 | } | |
7465 | ||
7466 | /* Update DST and SRC rtx. */ | |
7467 | *srcmem = src; | |
7468 | return dst; | |
7469 | } | |
7470 | ||
7471 | /* Helper function for the string operations below. Dest VARIABLE whether | |
7472 | it is aligned to VALUE bytes. If true, jump to the label. */ | |
7473 | ||
7474 | static rtx_code_label * | |
7475 | ix86_expand_aligntest (rtx variable, int value, bool epilogue) | |
7476 | { | |
7477 | rtx_code_label *label = gen_label_rtx (); | |
7478 | rtx tmpcount = gen_reg_rtx (GET_MODE (variable)); | |
7479 | if (GET_MODE (variable) == DImode) | |
7480 | emit_insn (gen_anddi3 (tmpcount, variable, GEN_INT (value))); | |
7481 | else | |
7482 | emit_insn (gen_andsi3 (tmpcount, variable, GEN_INT (value))); | |
7483 | emit_cmp_and_jump_insns (tmpcount, const0_rtx, EQ, 0, GET_MODE (variable), | |
7484 | 1, label); | |
7485 | if (epilogue) | |
7486 | predict_jump (REG_BR_PROB_BASE * 50 / 100); | |
7487 | else | |
7488 | predict_jump (REG_BR_PROB_BASE * 90 / 100); | |
7489 | return label; | |
7490 | } | |
7491 | ||
7492 | ||
7493 | /* Output code to copy at most count & (max_size - 1) bytes from SRC to DEST. */ | |
7494 | ||
7495 | static void | |
76715c32 | 7496 | expand_cpymem_epilogue (rtx destmem, rtx srcmem, |
2bf6d935 ML |
7497 | rtx destptr, rtx srcptr, rtx count, int max_size) |
7498 | { | |
7499 | rtx src, dest; | |
7500 | if (CONST_INT_P (count)) | |
7501 | { | |
7502 | HOST_WIDE_INT countval = INTVAL (count); | |
7503 | HOST_WIDE_INT epilogue_size = countval % max_size; | |
7504 | int i; | |
7505 | ||
7506 | /* For now MAX_SIZE should be a power of 2. This assert could be | |
7507 | relaxed, but it'll require a bit more complicated epilogue | |
7508 | expanding. */ | |
7509 | gcc_assert ((max_size & (max_size - 1)) == 0); | |
7510 | for (i = max_size; i >= 1; i >>= 1) | |
7511 | { | |
7512 | if (epilogue_size & i) | |
7513 | destmem = emit_memmov (destmem, &srcmem, destptr, srcptr, i); | |
7514 | } | |
7515 | return; | |
7516 | } | |
7517 | if (max_size > 8) | |
7518 | { | |
7519 | count = expand_simple_binop (GET_MODE (count), AND, count, GEN_INT (max_size - 1), | |
7520 | count, 1, OPTAB_DIRECT); | |
76715c32 | 7521 | expand_set_or_cpymem_via_loop (destmem, srcmem, destptr, srcptr, NULL, |
2bf6d935 ML |
7522 | count, QImode, 1, 4, false); |
7523 | return; | |
7524 | } | |
7525 | ||
7526 | /* When there are stringops, we can cheaply increase dest and src pointers. | |
7527 | Otherwise we save code size by maintaining offset (zero is readily | |
7528 | available from preceding rep operation) and using x86 addressing modes. | |
7529 | */ | |
7530 | if (TARGET_SINGLE_STRINGOP) | |
7531 | { | |
7532 | if (max_size > 4) | |
7533 | { | |
7534 | rtx_code_label *label = ix86_expand_aligntest (count, 4, true); | |
7535 | src = change_address (srcmem, SImode, srcptr); | |
7536 | dest = change_address (destmem, SImode, destptr); | |
7537 | emit_insn (gen_strmov (destptr, dest, srcptr, src)); | |
7538 | emit_label (label); | |
7539 | LABEL_NUSES (label) = 1; | |
7540 | } | |
7541 | if (max_size > 2) | |
7542 | { | |
7543 | rtx_code_label *label = ix86_expand_aligntest (count, 2, true); | |
7544 | src = change_address (srcmem, HImode, srcptr); | |
7545 | dest = change_address (destmem, HImode, destptr); | |
7546 | emit_insn (gen_strmov (destptr, dest, srcptr, src)); | |
7547 | emit_label (label); | |
7548 | LABEL_NUSES (label) = 1; | |
7549 | } | |
7550 | if (max_size > 1) | |
7551 | { | |
7552 | rtx_code_label *label = ix86_expand_aligntest (count, 1, true); | |
7553 | src = change_address (srcmem, QImode, srcptr); | |
7554 | dest = change_address (destmem, QImode, destptr); | |
7555 | emit_insn (gen_strmov (destptr, dest, srcptr, src)); | |
7556 | emit_label (label); | |
7557 | LABEL_NUSES (label) = 1; | |
7558 | } | |
7559 | } | |
7560 | else | |
7561 | { | |
7562 | rtx offset = force_reg (Pmode, const0_rtx); | |
7563 | rtx tmp; | |
7564 | ||
7565 | if (max_size > 4) | |
7566 | { | |
7567 | rtx_code_label *label = ix86_expand_aligntest (count, 4, true); | |
7568 | src = change_address (srcmem, SImode, srcptr); | |
7569 | dest = change_address (destmem, SImode, destptr); | |
7570 | emit_move_insn (dest, src); | |
7571 | tmp = expand_simple_binop (Pmode, PLUS, offset, GEN_INT (4), NULL, | |
7572 | true, OPTAB_LIB_WIDEN); | |
7573 | if (tmp != offset) | |
7574 | emit_move_insn (offset, tmp); | |
7575 | emit_label (label); | |
7576 | LABEL_NUSES (label) = 1; | |
7577 | } | |
7578 | if (max_size > 2) | |
7579 | { | |
7580 | rtx_code_label *label = ix86_expand_aligntest (count, 2, true); | |
7581 | tmp = gen_rtx_PLUS (Pmode, srcptr, offset); | |
7582 | src = change_address (srcmem, HImode, tmp); | |
7583 | tmp = gen_rtx_PLUS (Pmode, destptr, offset); | |
7584 | dest = change_address (destmem, HImode, tmp); | |
7585 | emit_move_insn (dest, src); | |
7586 | tmp = expand_simple_binop (Pmode, PLUS, offset, GEN_INT (2), tmp, | |
7587 | true, OPTAB_LIB_WIDEN); | |
7588 | if (tmp != offset) | |
7589 | emit_move_insn (offset, tmp); | |
7590 | emit_label (label); | |
7591 | LABEL_NUSES (label) = 1; | |
7592 | } | |
7593 | if (max_size > 1) | |
7594 | { | |
7595 | rtx_code_label *label = ix86_expand_aligntest (count, 1, true); | |
7596 | tmp = gen_rtx_PLUS (Pmode, srcptr, offset); | |
7597 | src = change_address (srcmem, QImode, tmp); | |
7598 | tmp = gen_rtx_PLUS (Pmode, destptr, offset); | |
7599 | dest = change_address (destmem, QImode, tmp); | |
7600 | emit_move_insn (dest, src); | |
7601 | emit_label (label); | |
7602 | LABEL_NUSES (label) = 1; | |
7603 | } | |
7604 | } | |
7605 | } | |
7606 | ||
7607 | /* This function emits moves to fill SIZE_TO_MOVE bytes starting from DESTMEM | |
7608 | with value PROMOTED_VAL. | |
7609 | SRC is passed by pointer to be updated on return. | |
7610 | Return value is updated DST. */ | |
7611 | static rtx | |
7612 | emit_memset (rtx destmem, rtx destptr, rtx promoted_val, | |
7613 | HOST_WIDE_INT size_to_move) | |
7614 | { | |
c3185b64 | 7615 | rtx dst = destmem; |
2bf6d935 ML |
7616 | enum insn_code code; |
7617 | machine_mode move_mode; | |
7618 | int piece_size, i; | |
7619 | ||
7620 | /* Find the widest mode in which we could perform moves. | |
7621 | Start with the biggest power of 2 less than SIZE_TO_MOVE and half | |
7622 | it until move of such size is supported. */ | |
7623 | move_mode = GET_MODE (promoted_val); | |
7624 | if (move_mode == VOIDmode) | |
7625 | move_mode = QImode; | |
7626 | if (size_to_move < GET_MODE_SIZE (move_mode)) | |
7627 | { | |
7628 | unsigned int move_bits = size_to_move * BITS_PER_UNIT; | |
7629 | move_mode = int_mode_for_size (move_bits, 0).require (); | |
7630 | promoted_val = gen_lowpart (move_mode, promoted_val); | |
7631 | } | |
7632 | piece_size = GET_MODE_SIZE (move_mode); | |
7633 | code = optab_handler (mov_optab, move_mode); | |
7634 | gcc_assert (code != CODE_FOR_nothing && promoted_val != NULL_RTX); | |
7635 | ||
7636 | dst = adjust_automodify_address_nv (dst, move_mode, destptr, 0); | |
7637 | ||
7638 | /* Emit moves. We'll need SIZE_TO_MOVE/PIECE_SIZES moves. */ | |
7639 | gcc_assert (size_to_move % piece_size == 0); | |
c3185b64 | 7640 | |
2bf6d935 ML |
7641 | for (i = 0; i < size_to_move; i += piece_size) |
7642 | { | |
7643 | if (piece_size <= GET_MODE_SIZE (word_mode)) | |
7644 | { | |
7645 | emit_insn (gen_strset (destptr, dst, promoted_val)); | |
7646 | dst = adjust_automodify_address_nv (dst, move_mode, destptr, | |
7647 | piece_size); | |
7648 | continue; | |
7649 | } | |
7650 | ||
7651 | emit_insn (GEN_FCN (code) (dst, promoted_val)); | |
7652 | ||
7653 | emit_move_insn (destptr, | |
c3185b64 | 7654 | plus_constant (Pmode, copy_rtx (destptr), piece_size)); |
2bf6d935 ML |
7655 | |
7656 | dst = adjust_automodify_address_nv (dst, move_mode, destptr, | |
7657 | piece_size); | |
7658 | } | |
7659 | ||
7660 | /* Update DST rtx. */ | |
7661 | return dst; | |
7662 | } | |
7663 | /* Output code to set at most count & (max_size - 1) bytes starting by DEST. */ | |
7664 | static void | |
7665 | expand_setmem_epilogue_via_loop (rtx destmem, rtx destptr, rtx value, | |
7666 | rtx count, int max_size) | |
7667 | { | |
7668 | count = expand_simple_binop (counter_mode (count), AND, count, | |
7669 | GEN_INT (max_size - 1), count, 1, OPTAB_DIRECT); | |
76715c32 | 7670 | expand_set_or_cpymem_via_loop (destmem, NULL, destptr, NULL, |
2bf6d935 ML |
7671 | gen_lowpart (QImode, value), count, QImode, |
7672 | 1, max_size / 2, true); | |
7673 | } | |
7674 | ||
7675 | /* Output code to set at most count & (max_size - 1) bytes starting by DEST. */ | |
7676 | static void | |
7677 | expand_setmem_epilogue (rtx destmem, rtx destptr, rtx value, rtx vec_value, | |
7678 | rtx count, int max_size) | |
7679 | { | |
7680 | rtx dest; | |
7681 | ||
7682 | if (CONST_INT_P (count)) | |
7683 | { | |
7684 | HOST_WIDE_INT countval = INTVAL (count); | |
7685 | HOST_WIDE_INT epilogue_size = countval % max_size; | |
7686 | int i; | |
7687 | ||
7688 | /* For now MAX_SIZE should be a power of 2. This assert could be | |
7689 | relaxed, but it'll require a bit more complicated epilogue | |
7690 | expanding. */ | |
7691 | gcc_assert ((max_size & (max_size - 1)) == 0); | |
7692 | for (i = max_size; i >= 1; i >>= 1) | |
7693 | { | |
7694 | if (epilogue_size & i) | |
7695 | { | |
7696 | if (vec_value && i > GET_MODE_SIZE (GET_MODE (value))) | |
7697 | destmem = emit_memset (destmem, destptr, vec_value, i); | |
7698 | else | |
7699 | destmem = emit_memset (destmem, destptr, value, i); | |
7700 | } | |
7701 | } | |
7702 | return; | |
7703 | } | |
7704 | if (max_size > 32) | |
7705 | { | |
7706 | expand_setmem_epilogue_via_loop (destmem, destptr, value, count, max_size); | |
7707 | return; | |
7708 | } | |
7709 | if (max_size > 16) | |
7710 | { | |
7711 | rtx_code_label *label = ix86_expand_aligntest (count, 16, true); | |
7712 | if (TARGET_64BIT) | |
7713 | { | |
7714 | dest = change_address (destmem, DImode, destptr); | |
7715 | emit_insn (gen_strset (destptr, dest, value)); | |
7716 | dest = adjust_automodify_address_nv (dest, DImode, destptr, 8); | |
7717 | emit_insn (gen_strset (destptr, dest, value)); | |
7718 | } | |
7719 | else | |
7720 | { | |
7721 | dest = change_address (destmem, SImode, destptr); | |
7722 | emit_insn (gen_strset (destptr, dest, value)); | |
7723 | dest = adjust_automodify_address_nv (dest, SImode, destptr, 4); | |
7724 | emit_insn (gen_strset (destptr, dest, value)); | |
7725 | dest = adjust_automodify_address_nv (dest, SImode, destptr, 8); | |
7726 | emit_insn (gen_strset (destptr, dest, value)); | |
7727 | dest = adjust_automodify_address_nv (dest, SImode, destptr, 12); | |
7728 | emit_insn (gen_strset (destptr, dest, value)); | |
7729 | } | |
7730 | emit_label (label); | |
7731 | LABEL_NUSES (label) = 1; | |
7732 | } | |
7733 | if (max_size > 8) | |
7734 | { | |
7735 | rtx_code_label *label = ix86_expand_aligntest (count, 8, true); | |
7736 | if (TARGET_64BIT) | |
7737 | { | |
7738 | dest = change_address (destmem, DImode, destptr); | |
7739 | emit_insn (gen_strset (destptr, dest, value)); | |
7740 | } | |
7741 | else | |
7742 | { | |
7743 | dest = change_address (destmem, SImode, destptr); | |
7744 | emit_insn (gen_strset (destptr, dest, value)); | |
7745 | dest = adjust_automodify_address_nv (dest, SImode, destptr, 4); | |
7746 | emit_insn (gen_strset (destptr, dest, value)); | |
7747 | } | |
7748 | emit_label (label); | |
7749 | LABEL_NUSES (label) = 1; | |
7750 | } | |
7751 | if (max_size > 4) | |
7752 | { | |
7753 | rtx_code_label *label = ix86_expand_aligntest (count, 4, true); | |
7754 | dest = change_address (destmem, SImode, destptr); | |
7755 | emit_insn (gen_strset (destptr, dest, gen_lowpart (SImode, value))); | |
7756 | emit_label (label); | |
7757 | LABEL_NUSES (label) = 1; | |
7758 | } | |
7759 | if (max_size > 2) | |
7760 | { | |
7761 | rtx_code_label *label = ix86_expand_aligntest (count, 2, true); | |
7762 | dest = change_address (destmem, HImode, destptr); | |
7763 | emit_insn (gen_strset (destptr, dest, gen_lowpart (HImode, value))); | |
7764 | emit_label (label); | |
7765 | LABEL_NUSES (label) = 1; | |
7766 | } | |
7767 | if (max_size > 1) | |
7768 | { | |
7769 | rtx_code_label *label = ix86_expand_aligntest (count, 1, true); | |
7770 | dest = change_address (destmem, QImode, destptr); | |
7771 | emit_insn (gen_strset (destptr, dest, gen_lowpart (QImode, value))); | |
7772 | emit_label (label); | |
7773 | LABEL_NUSES (label) = 1; | |
7774 | } | |
7775 | } | |
7776 | ||
7777 | /* Adjust COUNTER by the VALUE. */ | |
7778 | static void | |
7779 | ix86_adjust_counter (rtx countreg, HOST_WIDE_INT value) | |
7780 | { | |
83bc5e44 | 7781 | emit_insn (gen_add2_insn (countreg, GEN_INT (-value))); |
2bf6d935 ML |
7782 | } |
7783 | ||
7784 | /* Depending on ISSETMEM, copy enough from SRCMEM to DESTMEM or set enough to | |
7785 | DESTMEM to align it to DESIRED_ALIGNMENT. Original alignment is ALIGN. | |
7786 | Depending on ISSETMEM, either arguments SRCMEM/SRCPTR or VALUE/VEC_VALUE are | |
7787 | ignored. | |
7788 | Return value is updated DESTMEM. */ | |
7789 | ||
7790 | static rtx | |
76715c32 | 7791 | expand_set_or_cpymem_prologue (rtx destmem, rtx srcmem, |
2bf6d935 ML |
7792 | rtx destptr, rtx srcptr, rtx value, |
7793 | rtx vec_value, rtx count, int align, | |
7794 | int desired_alignment, bool issetmem) | |
7795 | { | |
7796 | int i; | |
7797 | for (i = 1; i < desired_alignment; i <<= 1) | |
7798 | { | |
7799 | if (align <= i) | |
7800 | { | |
7801 | rtx_code_label *label = ix86_expand_aligntest (destptr, i, false); | |
7802 | if (issetmem) | |
7803 | { | |
7804 | if (vec_value && i > GET_MODE_SIZE (GET_MODE (value))) | |
7805 | destmem = emit_memset (destmem, destptr, vec_value, i); | |
7806 | else | |
7807 | destmem = emit_memset (destmem, destptr, value, i); | |
7808 | } | |
7809 | else | |
7810 | destmem = emit_memmov (destmem, &srcmem, destptr, srcptr, i); | |
7811 | ix86_adjust_counter (count, i); | |
7812 | emit_label (label); | |
7813 | LABEL_NUSES (label) = 1; | |
7814 | set_mem_align (destmem, i * 2 * BITS_PER_UNIT); | |
7815 | } | |
7816 | } | |
7817 | return destmem; | |
7818 | } | |
7819 | ||
7820 | /* Test if COUNT&SIZE is nonzero and if so, expand movme | |
7821 | or setmem sequence that is valid for SIZE..2*SIZE-1 bytes | |
7822 | and jump to DONE_LABEL. */ | |
7823 | static void | |
76715c32 | 7824 | expand_small_cpymem_or_setmem (rtx destmem, rtx srcmem, |
2bf6d935 ML |
7825 | rtx destptr, rtx srcptr, |
7826 | rtx value, rtx vec_value, | |
7827 | rtx count, int size, | |
7828 | rtx done_label, bool issetmem) | |
7829 | { | |
7830 | rtx_code_label *label = ix86_expand_aligntest (count, size, false); | |
7831 | machine_mode mode = int_mode_for_size (size * BITS_PER_UNIT, 1).else_blk (); | |
7832 | rtx modesize; | |
7833 | int n; | |
7834 | ||
7835 | /* If we do not have vector value to copy, we must reduce size. */ | |
7836 | if (issetmem) | |
7837 | { | |
7838 | if (!vec_value) | |
7839 | { | |
7840 | if (GET_MODE (value) == VOIDmode && size > 8) | |
7841 | mode = Pmode; | |
7842 | else if (GET_MODE_SIZE (mode) > GET_MODE_SIZE (GET_MODE (value))) | |
7843 | mode = GET_MODE (value); | |
7844 | } | |
7845 | else | |
7846 | mode = GET_MODE (vec_value), value = vec_value; | |
7847 | } | |
7848 | else | |
7849 | { | |
7850 | /* Choose appropriate vector mode. */ | |
7851 | if (size >= 32) | |
7852 | mode = TARGET_AVX ? V32QImode : TARGET_SSE ? V16QImode : DImode; | |
7853 | else if (size >= 16) | |
7854 | mode = TARGET_SSE ? V16QImode : DImode; | |
7855 | srcmem = change_address (srcmem, mode, srcptr); | |
7856 | } | |
7857 | destmem = change_address (destmem, mode, destptr); | |
7858 | modesize = GEN_INT (GET_MODE_SIZE (mode)); | |
7859 | gcc_assert (GET_MODE_SIZE (mode) <= size); | |
7860 | for (n = 0; n * GET_MODE_SIZE (mode) < size; n++) | |
7861 | { | |
7862 | if (issetmem) | |
7863 | emit_move_insn (destmem, gen_lowpart (mode, value)); | |
7864 | else | |
7865 | { | |
7866 | emit_move_insn (destmem, srcmem); | |
7867 | srcmem = offset_address (srcmem, modesize, GET_MODE_SIZE (mode)); | |
7868 | } | |
7869 | destmem = offset_address (destmem, modesize, GET_MODE_SIZE (mode)); | |
7870 | } | |
7871 | ||
7872 | destmem = offset_address (destmem, count, 1); | |
7873 | destmem = offset_address (destmem, GEN_INT (-2 * size), | |
7874 | GET_MODE_SIZE (mode)); | |
7875 | if (!issetmem) | |
7876 | { | |
7877 | srcmem = offset_address (srcmem, count, 1); | |
7878 | srcmem = offset_address (srcmem, GEN_INT (-2 * size), | |
7879 | GET_MODE_SIZE (mode)); | |
7880 | } | |
7881 | for (n = 0; n * GET_MODE_SIZE (mode) < size; n++) | |
7882 | { | |
7883 | if (issetmem) | |
7884 | emit_move_insn (destmem, gen_lowpart (mode, value)); | |
7885 | else | |
7886 | { | |
7887 | emit_move_insn (destmem, srcmem); | |
7888 | srcmem = offset_address (srcmem, modesize, GET_MODE_SIZE (mode)); | |
7889 | } | |
7890 | destmem = offset_address (destmem, modesize, GET_MODE_SIZE (mode)); | |
7891 | } | |
7892 | emit_jump_insn (gen_jump (done_label)); | |
7893 | emit_barrier (); | |
7894 | ||
7895 | emit_label (label); | |
7896 | LABEL_NUSES (label) = 1; | |
7897 | } | |
7898 | ||
7899 | /* Handle small memcpy (up to SIZE that is supposed to be small power of 2. | |
7900 | and get ready for the main memcpy loop by copying iniital DESIRED_ALIGN-ALIGN | |
7901 | bytes and last SIZE bytes adjusitng DESTPTR/SRCPTR/COUNT in a way we can | |
7902 | proceed with an loop copying SIZE bytes at once. Do moves in MODE. | |
7903 | DONE_LABEL is a label after the whole copying sequence. The label is created | |
7904 | on demand if *DONE_LABEL is NULL. | |
7905 | MIN_SIZE is minimal size of block copied. This value gets adjusted for new | |
7906 | bounds after the initial copies. | |
7907 | ||
7908 | DESTMEM/SRCMEM are memory expressions pointing to the copies block, | |
7909 | DESTPTR/SRCPTR are pointers to the block. DYNAMIC_CHECK indicate whether | |
7910 | we will dispatch to a library call for large blocks. | |
7911 | ||
7912 | In pseudocode we do: | |
7913 | ||
7914 | if (COUNT < SIZE) | |
7915 | { | |
7916 | Assume that SIZE is 4. Bigger sizes are handled analogously | |
7917 | if (COUNT & 4) | |
7918 | { | |
7919 | copy 4 bytes from SRCPTR to DESTPTR | |
7920 | copy 4 bytes from SRCPTR + COUNT - 4 to DESTPTR + COUNT - 4 | |
7921 | goto done_label | |
7922 | } | |
7923 | if (!COUNT) | |
7924 | goto done_label; | |
7925 | copy 1 byte from SRCPTR to DESTPTR | |
7926 | if (COUNT & 2) | |
7927 | { | |
7928 | copy 2 bytes from SRCPTR to DESTPTR | |
7929 | copy 2 bytes from SRCPTR + COUNT - 2 to DESTPTR + COUNT - 2 | |
7930 | } | |
7931 | } | |
7932 | else | |
7933 | { | |
7934 | copy at least DESIRED_ALIGN-ALIGN bytes from SRCPTR to DESTPTR | |
7935 | copy SIZE bytes from SRCPTR + COUNT - SIZE to DESTPTR + COUNT -SIZE | |
7936 | ||
7937 | OLD_DESPTR = DESTPTR; | |
7938 | Align DESTPTR up to DESIRED_ALIGN | |
7939 | SRCPTR += DESTPTR - OLD_DESTPTR | |
7940 | COUNT -= DEST_PTR - OLD_DESTPTR | |
7941 | if (DYNAMIC_CHECK) | |
7942 | Round COUNT down to multiple of SIZE | |
7943 | << optional caller supplied zero size guard is here >> | |
7944 | << optional caller supplied dynamic check is here >> | |
7945 | << caller supplied main copy loop is here >> | |
7946 | } | |
7947 | done_label: | |
7948 | */ | |
7949 | static void | |
76715c32 | 7950 | expand_set_or_cpymem_prologue_epilogue_by_misaligned_moves (rtx destmem, rtx srcmem, |
2bf6d935 ML |
7951 | rtx *destptr, rtx *srcptr, |
7952 | machine_mode mode, | |
7953 | rtx value, rtx vec_value, | |
7954 | rtx *count, | |
7955 | rtx_code_label **done_label, | |
7956 | int size, | |
7957 | int desired_align, | |
7958 | int align, | |
7959 | unsigned HOST_WIDE_INT *min_size, | |
7960 | bool dynamic_check, | |
7961 | bool issetmem) | |
7962 | { | |
7963 | rtx_code_label *loop_label = NULL, *label; | |
7964 | int n; | |
7965 | rtx modesize; | |
7966 | int prolog_size = 0; | |
7967 | rtx mode_value; | |
7968 | ||
7969 | /* Chose proper value to copy. */ | |
7970 | if (issetmem && VECTOR_MODE_P (mode)) | |
7971 | mode_value = vec_value; | |
7972 | else | |
7973 | mode_value = value; | |
7974 | gcc_assert (GET_MODE_SIZE (mode) <= size); | |
7975 | ||
7976 | /* See if block is big or small, handle small blocks. */ | |
7977 | if (!CONST_INT_P (*count) && *min_size < (unsigned HOST_WIDE_INT)size) | |
7978 | { | |
7979 | int size2 = size; | |
7980 | loop_label = gen_label_rtx (); | |
7981 | ||
7982 | if (!*done_label) | |
7983 | *done_label = gen_label_rtx (); | |
7984 | ||
7985 | emit_cmp_and_jump_insns (*count, GEN_INT (size2), GE, 0, GET_MODE (*count), | |
7986 | 1, loop_label); | |
7987 | size2 >>= 1; | |
7988 | ||
7989 | /* Handle sizes > 3. */ | |
7990 | for (;size2 > 2; size2 >>= 1) | |
76715c32 | 7991 | expand_small_cpymem_or_setmem (destmem, srcmem, |
2bf6d935 ML |
7992 | *destptr, *srcptr, |
7993 | value, vec_value, | |
7994 | *count, | |
7995 | size2, *done_label, issetmem); | |
7996 | /* Nothing to copy? Jump to DONE_LABEL if so */ | |
7997 | emit_cmp_and_jump_insns (*count, const0_rtx, EQ, 0, GET_MODE (*count), | |
7998 | 1, *done_label); | |
7999 | ||
8000 | /* Do a byte copy. */ | |
8001 | destmem = change_address (destmem, QImode, *destptr); | |
8002 | if (issetmem) | |
8003 | emit_move_insn (destmem, gen_lowpart (QImode, value)); | |
8004 | else | |
8005 | { | |
8006 | srcmem = change_address (srcmem, QImode, *srcptr); | |
8007 | emit_move_insn (destmem, srcmem); | |
8008 | } | |
8009 | ||
8010 | /* Handle sizes 2 and 3. */ | |
8011 | label = ix86_expand_aligntest (*count, 2, false); | |
8012 | destmem = change_address (destmem, HImode, *destptr); | |
8013 | destmem = offset_address (destmem, *count, 1); | |
8014 | destmem = offset_address (destmem, GEN_INT (-2), 2); | |
8015 | if (issetmem) | |
8016 | emit_move_insn (destmem, gen_lowpart (HImode, value)); | |
8017 | else | |
8018 | { | |
8019 | srcmem = change_address (srcmem, HImode, *srcptr); | |
8020 | srcmem = offset_address (srcmem, *count, 1); | |
8021 | srcmem = offset_address (srcmem, GEN_INT (-2), 2); | |
8022 | emit_move_insn (destmem, srcmem); | |
8023 | } | |
8024 | ||
8025 | emit_label (label); | |
8026 | LABEL_NUSES (label) = 1; | |
8027 | emit_jump_insn (gen_jump (*done_label)); | |
8028 | emit_barrier (); | |
8029 | } | |
8030 | else | |
8031 | gcc_assert (*min_size >= (unsigned HOST_WIDE_INT)size | |
8032 | || UINTVAL (*count) >= (unsigned HOST_WIDE_INT)size); | |
8033 | ||
8034 | /* Start memcpy for COUNT >= SIZE. */ | |
8035 | if (loop_label) | |
8036 | { | |
8037 | emit_label (loop_label); | |
8038 | LABEL_NUSES (loop_label) = 1; | |
8039 | } | |
8040 | ||
8041 | /* Copy first desired_align bytes. */ | |
8042 | if (!issetmem) | |
8043 | srcmem = change_address (srcmem, mode, *srcptr); | |
8044 | destmem = change_address (destmem, mode, *destptr); | |
8045 | modesize = GEN_INT (GET_MODE_SIZE (mode)); | |
8046 | for (n = 0; prolog_size < desired_align - align; n++) | |
8047 | { | |
8048 | if (issetmem) | |
8049 | emit_move_insn (destmem, mode_value); | |
8050 | else | |
8051 | { | |
8052 | emit_move_insn (destmem, srcmem); | |
8053 | srcmem = offset_address (srcmem, modesize, GET_MODE_SIZE (mode)); | |
8054 | } | |
8055 | destmem = offset_address (destmem, modesize, GET_MODE_SIZE (mode)); | |
8056 | prolog_size += GET_MODE_SIZE (mode); | |
8057 | } | |
8058 | ||
8059 | ||
8060 | /* Copy last SIZE bytes. */ | |
8061 | destmem = offset_address (destmem, *count, 1); | |
8062 | destmem = offset_address (destmem, | |
8063 | GEN_INT (-size - prolog_size), | |
8064 | 1); | |
8065 | if (issetmem) | |
8066 | emit_move_insn (destmem, mode_value); | |
8067 | else | |
8068 | { | |
8069 | srcmem = offset_address (srcmem, *count, 1); | |
8070 | srcmem = offset_address (srcmem, | |
8071 | GEN_INT (-size - prolog_size), | |
8072 | 1); | |
8073 | emit_move_insn (destmem, srcmem); | |
8074 | } | |
8075 | for (n = 1; n * GET_MODE_SIZE (mode) < size; n++) | |
8076 | { | |
8077 | destmem = offset_address (destmem, modesize, 1); | |
8078 | if (issetmem) | |
8079 | emit_move_insn (destmem, mode_value); | |
8080 | else | |
8081 | { | |
8082 | srcmem = offset_address (srcmem, modesize, 1); | |
8083 | emit_move_insn (destmem, srcmem); | |
8084 | } | |
8085 | } | |
8086 | ||
8087 | /* Align destination. */ | |
8088 | if (desired_align > 1 && desired_align > align) | |
8089 | { | |
8090 | rtx saveddest = *destptr; | |
8091 | ||
8092 | gcc_assert (desired_align <= size); | |
8093 | /* Align destptr up, place it to new register. */ | |
8094 | *destptr = expand_simple_binop (GET_MODE (*destptr), PLUS, *destptr, | |
8095 | GEN_INT (prolog_size), | |
8096 | NULL_RTX, 1, OPTAB_DIRECT); | |
8097 | if (REG_P (*destptr) && REG_P (saveddest) && REG_POINTER (saveddest)) | |
8098 | REG_POINTER (*destptr) = 1; | |
8099 | *destptr = expand_simple_binop (GET_MODE (*destptr), AND, *destptr, | |
8100 | GEN_INT (-desired_align), | |
8101 | *destptr, 1, OPTAB_DIRECT); | |
8102 | /* See how many bytes we skipped. */ | |
8103 | saveddest = expand_simple_binop (GET_MODE (*destptr), MINUS, saveddest, | |
8104 | *destptr, | |
8105 | saveddest, 1, OPTAB_DIRECT); | |
8106 | /* Adjust srcptr and count. */ | |
8107 | if (!issetmem) | |
8108 | *srcptr = expand_simple_binop (GET_MODE (*srcptr), MINUS, *srcptr, | |
8109 | saveddest, *srcptr, 1, OPTAB_DIRECT); | |
8110 | *count = expand_simple_binop (GET_MODE (*count), PLUS, *count, | |
8111 | saveddest, *count, 1, OPTAB_DIRECT); | |
8112 | /* We copied at most size + prolog_size. */ | |
8113 | if (*min_size > (unsigned HOST_WIDE_INT)(size + prolog_size)) | |
8114 | *min_size | |
8115 | = ROUND_DOWN (*min_size - size, (unsigned HOST_WIDE_INT)size); | |
8116 | else | |
8117 | *min_size = 0; | |
8118 | ||
8119 | /* Our loops always round down the block size, but for dispatch to | |
8120 | library we need precise value. */ | |
8121 | if (dynamic_check) | |
8122 | *count = expand_simple_binop (GET_MODE (*count), AND, *count, | |
8123 | GEN_INT (-size), *count, 1, OPTAB_DIRECT); | |
8124 | } | |
8125 | else | |
8126 | { | |
8127 | gcc_assert (prolog_size == 0); | |
8128 | /* Decrease count, so we won't end up copying last word twice. */ | |
8129 | if (!CONST_INT_P (*count)) | |
8130 | *count = expand_simple_binop (GET_MODE (*count), PLUS, *count, | |
8131 | constm1_rtx, *count, 1, OPTAB_DIRECT); | |
8132 | else | |
8133 | *count = GEN_INT (ROUND_DOWN (UINTVAL (*count) - 1, | |
8134 | (unsigned HOST_WIDE_INT)size)); | |
8135 | if (*min_size) | |
8136 | *min_size = ROUND_DOWN (*min_size - 1, (unsigned HOST_WIDE_INT)size); | |
8137 | } | |
8138 | } | |
8139 | ||
8140 | ||
8141 | /* This function is like the previous one, except here we know how many bytes | |
8142 | need to be copied. That allows us to update alignment not only of DST, which | |
8143 | is returned, but also of SRC, which is passed as a pointer for that | |
8144 | reason. */ | |
8145 | static rtx | |
76715c32 | 8146 | expand_set_or_cpymem_constant_prologue (rtx dst, rtx *srcp, rtx destreg, |
2bf6d935 ML |
8147 | rtx srcreg, rtx value, rtx vec_value, |
8148 | int desired_align, int align_bytes, | |
8149 | bool issetmem) | |
8150 | { | |
8151 | rtx src = NULL; | |
8152 | rtx orig_dst = dst; | |
8153 | rtx orig_src = NULL; | |
8154 | int piece_size = 1; | |
8155 | int copied_bytes = 0; | |
8156 | ||
8157 | if (!issetmem) | |
8158 | { | |
8159 | gcc_assert (srcp != NULL); | |
8160 | src = *srcp; | |
8161 | orig_src = src; | |
8162 | } | |
8163 | ||
8164 | for (piece_size = 1; | |
8165 | piece_size <= desired_align && copied_bytes < align_bytes; | |
8166 | piece_size <<= 1) | |
8167 | { | |
8168 | if (align_bytes & piece_size) | |
8169 | { | |
8170 | if (issetmem) | |
8171 | { | |
8172 | if (vec_value && piece_size > GET_MODE_SIZE (GET_MODE (value))) | |
8173 | dst = emit_memset (dst, destreg, vec_value, piece_size); | |
8174 | else | |
8175 | dst = emit_memset (dst, destreg, value, piece_size); | |
8176 | } | |
8177 | else | |
8178 | dst = emit_memmov (dst, &src, destreg, srcreg, piece_size); | |
8179 | copied_bytes += piece_size; | |
8180 | } | |
8181 | } | |
8182 | if (MEM_ALIGN (dst) < (unsigned int) desired_align * BITS_PER_UNIT) | |
8183 | set_mem_align (dst, desired_align * BITS_PER_UNIT); | |
8184 | if (MEM_SIZE_KNOWN_P (orig_dst)) | |
8185 | set_mem_size (dst, MEM_SIZE (orig_dst) - align_bytes); | |
8186 | ||
8187 | if (!issetmem) | |
8188 | { | |
8189 | int src_align_bytes = get_mem_align_offset (src, desired_align | |
8190 | * BITS_PER_UNIT); | |
8191 | if (src_align_bytes >= 0) | |
8192 | src_align_bytes = desired_align - src_align_bytes; | |
8193 | if (src_align_bytes >= 0) | |
8194 | { | |
8195 | unsigned int src_align; | |
8196 | for (src_align = desired_align; src_align >= 2; src_align >>= 1) | |
8197 | { | |
8198 | if ((src_align_bytes & (src_align - 1)) | |
8199 | == (align_bytes & (src_align - 1))) | |
8200 | break; | |
8201 | } | |
8202 | if (src_align > (unsigned int) desired_align) | |
8203 | src_align = desired_align; | |
8204 | if (MEM_ALIGN (src) < src_align * BITS_PER_UNIT) | |
8205 | set_mem_align (src, src_align * BITS_PER_UNIT); | |
8206 | } | |
8207 | if (MEM_SIZE_KNOWN_P (orig_src)) | |
8208 | set_mem_size (src, MEM_SIZE (orig_src) - align_bytes); | |
8209 | *srcp = src; | |
8210 | } | |
8211 | ||
8212 | return dst; | |
8213 | } | |
8214 | ||
8215 | /* Return true if ALG can be used in current context. | |
8216 | Assume we expand memset if MEMSET is true. */ | |
8217 | static bool | |
8218 | alg_usable_p (enum stringop_alg alg, bool memset, bool have_as) | |
8219 | { | |
8220 | if (alg == no_stringop) | |
8221 | return false; | |
8222 | if (alg == vector_loop) | |
8223 | return TARGET_SSE || TARGET_AVX; | |
8224 | /* Algorithms using the rep prefix want at least edi and ecx; | |
8225 | additionally, memset wants eax and memcpy wants esi. Don't | |
8226 | consider such algorithms if the user has appropriated those | |
8227 | registers for their own purposes, or if we have a non-default | |
8228 | address space, since some string insns cannot override the segment. */ | |
8229 | if (alg == rep_prefix_1_byte | |
8230 | || alg == rep_prefix_4_byte | |
8231 | || alg == rep_prefix_8_byte) | |
8232 | { | |
8233 | if (have_as) | |
8234 | return false; | |
8235 | if (fixed_regs[CX_REG] | |
8236 | || fixed_regs[DI_REG] | |
8237 | || (memset ? fixed_regs[AX_REG] : fixed_regs[SI_REG])) | |
8238 | return false; | |
8239 | } | |
8240 | return true; | |
8241 | } | |
8242 | ||
8243 | /* Given COUNT and EXPECTED_SIZE, decide on codegen of string operation. */ | |
8244 | static enum stringop_alg | |
8245 | decide_alg (HOST_WIDE_INT count, HOST_WIDE_INT expected_size, | |
8246 | unsigned HOST_WIDE_INT min_size, unsigned HOST_WIDE_INT max_size, | |
8247 | bool memset, bool zero_memset, bool have_as, | |
8248 | int *dynamic_check, bool *noalign, bool recur) | |
8249 | { | |
8250 | const struct stringop_algs *algs; | |
8251 | bool optimize_for_speed; | |
8252 | int max = 0; | |
8253 | const struct processor_costs *cost; | |
8254 | int i; | |
8255 | bool any_alg_usable_p = false; | |
8256 | ||
8257 | *noalign = false; | |
8258 | *dynamic_check = -1; | |
8259 | ||
8260 | /* Even if the string operation call is cold, we still might spend a lot | |
8261 | of time processing large blocks. */ | |
8262 | if (optimize_function_for_size_p (cfun) | |
8263 | || (optimize_insn_for_size_p () | |
8264 | && (max_size < 256 | |
8265 | || (expected_size != -1 && expected_size < 256)))) | |
8266 | optimize_for_speed = false; | |
8267 | else | |
8268 | optimize_for_speed = true; | |
8269 | ||
8270 | cost = optimize_for_speed ? ix86_cost : &ix86_size_cost; | |
8271 | if (memset) | |
8272 | algs = &cost->memset[TARGET_64BIT != 0]; | |
8273 | else | |
8274 | algs = &cost->memcpy[TARGET_64BIT != 0]; | |
8275 | ||
8276 | /* See maximal size for user defined algorithm. */ | |
8277 | for (i = 0; i < MAX_STRINGOP_ALGS; i++) | |
8278 | { | |
8279 | enum stringop_alg candidate = algs->size[i].alg; | |
8280 | bool usable = alg_usable_p (candidate, memset, have_as); | |
8281 | any_alg_usable_p |= usable; | |
8282 | ||
8283 | if (candidate != libcall && candidate && usable) | |
8284 | max = algs->size[i].max; | |
8285 | } | |
8286 | ||
8287 | /* If expected size is not known but max size is small enough | |
8288 | so inline version is a win, set expected size into | |
8289 | the range. */ | |
8290 | if (((max > 1 && (unsigned HOST_WIDE_INT) max >= max_size) || max == -1) | |
8291 | && expected_size == -1) | |
8292 | expected_size = min_size / 2 + max_size / 2; | |
8293 | ||
8294 | /* If user specified the algorithm, honor it if possible. */ | |
8295 | if (ix86_stringop_alg != no_stringop | |
8296 | && alg_usable_p (ix86_stringop_alg, memset, have_as)) | |
8297 | return ix86_stringop_alg; | |
8298 | /* rep; movq or rep; movl is the smallest variant. */ | |
8299 | else if (!optimize_for_speed) | |
8300 | { | |
8301 | *noalign = true; | |
8302 | if (!count || (count & 3) || (memset && !zero_memset)) | |
8303 | return alg_usable_p (rep_prefix_1_byte, memset, have_as) | |
8304 | ? rep_prefix_1_byte : loop_1_byte; | |
8305 | else | |
8306 | return alg_usable_p (rep_prefix_4_byte, memset, have_as) | |
8307 | ? rep_prefix_4_byte : loop; | |
8308 | } | |
8309 | /* Very tiny blocks are best handled via the loop, REP is expensive to | |
8310 | setup. */ | |
8311 | else if (expected_size != -1 && expected_size < 4) | |
8312 | return loop_1_byte; | |
8313 | else if (expected_size != -1) | |
8314 | { | |
8315 | enum stringop_alg alg = libcall; | |
8316 | bool alg_noalign = false; | |
8317 | for (i = 0; i < MAX_STRINGOP_ALGS; i++) | |
8318 | { | |
8319 | /* We get here if the algorithms that were not libcall-based | |
8320 | were rep-prefix based and we are unable to use rep prefixes | |
8321 | based on global register usage. Break out of the loop and | |
8322 | use the heuristic below. */ | |
8323 | if (algs->size[i].max == 0) | |
8324 | break; | |
8325 | if (algs->size[i].max >= expected_size || algs->size[i].max == -1) | |
8326 | { | |
8327 | enum stringop_alg candidate = algs->size[i].alg; | |
8328 | ||
8329 | if (candidate != libcall | |
8330 | && alg_usable_p (candidate, memset, have_as)) | |
8331 | { | |
8332 | alg = candidate; | |
8333 | alg_noalign = algs->size[i].noalign; | |
8334 | } | |
8335 | /* Honor TARGET_INLINE_ALL_STRINGOPS by picking | |
8336 | last non-libcall inline algorithm. */ | |
8337 | if (TARGET_INLINE_ALL_STRINGOPS) | |
8338 | { | |
8339 | /* When the current size is best to be copied by a libcall, | |
8340 | but we are still forced to inline, run the heuristic below | |
8341 | that will pick code for medium sized blocks. */ | |
8342 | if (alg != libcall) | |
8343 | { | |
8344 | *noalign = alg_noalign; | |
8345 | return alg; | |
8346 | } | |
8347 | else if (!any_alg_usable_p) | |
8348 | break; | |
8349 | } | |
bf24f4ec L |
8350 | else if (alg_usable_p (candidate, memset, have_as) |
8351 | && !(TARGET_PREFER_KNOWN_REP_MOVSB_STOSB | |
8352 | && candidate == rep_prefix_1_byte | |
8353 | /* NB: If min_size != max_size, size is | |
8354 | unknown. */ | |
8355 | && min_size != max_size)) | |
2bf6d935 ML |
8356 | { |
8357 | *noalign = algs->size[i].noalign; | |
8358 | return candidate; | |
8359 | } | |
8360 | } | |
8361 | } | |
8362 | } | |
8363 | /* When asked to inline the call anyway, try to pick meaningful choice. | |
8364 | We look for maximal size of block that is faster to copy by hand and | |
8365 | take blocks of at most of that size guessing that average size will | |
8366 | be roughly half of the block. | |
8367 | ||
8368 | If this turns out to be bad, we might simply specify the preferred | |
8369 | choice in ix86_costs. */ | |
8370 | if ((TARGET_INLINE_ALL_STRINGOPS || TARGET_INLINE_STRINGOPS_DYNAMICALLY) | |
8371 | && (algs->unknown_size == libcall | |
8372 | || !alg_usable_p (algs->unknown_size, memset, have_as))) | |
8373 | { | |
8374 | enum stringop_alg alg; | |
8375 | HOST_WIDE_INT new_expected_size = (max > 0 ? max : 4096) / 2; | |
8376 | ||
8377 | /* If there aren't any usable algorithms or if recursing already, | |
8378 | then recursing on smaller sizes or same size isn't going to | |
8379 | find anything. Just return the simple byte-at-a-time copy loop. */ | |
8380 | if (!any_alg_usable_p || recur) | |
8381 | { | |
8382 | /* Pick something reasonable. */ | |
8383 | if (TARGET_INLINE_STRINGOPS_DYNAMICALLY && !recur) | |
8384 | *dynamic_check = 128; | |
8385 | return loop_1_byte; | |
8386 | } | |
8387 | alg = decide_alg (count, new_expected_size, min_size, max_size, memset, | |
8388 | zero_memset, have_as, dynamic_check, noalign, true); | |
8389 | gcc_assert (*dynamic_check == -1); | |
8390 | if (TARGET_INLINE_STRINGOPS_DYNAMICALLY) | |
8391 | *dynamic_check = max; | |
8392 | else | |
8393 | gcc_assert (alg != libcall); | |
8394 | return alg; | |
8395 | } | |
8396 | return (alg_usable_p (algs->unknown_size, memset, have_as) | |
8397 | ? algs->unknown_size : libcall); | |
8398 | } | |
8399 | ||
8400 | /* Decide on alignment. We know that the operand is already aligned to ALIGN | |
8401 | (ALIGN can be based on profile feedback and thus it is not 100% guaranteed). */ | |
8402 | static int | |
8403 | decide_alignment (int align, | |
8404 | enum stringop_alg alg, | |
8405 | int expected_size, | |
8406 | machine_mode move_mode) | |
8407 | { | |
8408 | int desired_align = 0; | |
8409 | ||
8410 | gcc_assert (alg != no_stringop); | |
8411 | ||
8412 | if (alg == libcall) | |
8413 | return 0; | |
8414 | if (move_mode == VOIDmode) | |
8415 | return 0; | |
8416 | ||
8417 | desired_align = GET_MODE_SIZE (move_mode); | |
8418 | /* PentiumPro has special logic triggering for 8 byte aligned blocks. | |
8419 | copying whole cacheline at once. */ | |
f23881fc | 8420 | if (TARGET_CPU_P (PENTIUMPRO) |
2bf6d935 ML |
8421 | && (alg == rep_prefix_4_byte || alg == rep_prefix_1_byte)) |
8422 | desired_align = 8; | |
8423 | ||
8424 | if (optimize_size) | |
8425 | desired_align = 1; | |
8426 | if (desired_align < align) | |
8427 | desired_align = align; | |
8428 | if (expected_size != -1 && expected_size < 4) | |
8429 | desired_align = align; | |
8430 | ||
8431 | return desired_align; | |
8432 | } | |
8433 | ||
8434 | ||
8435 | /* Helper function for memcpy. For QImode value 0xXY produce | |
8436 | 0xXYXYXYXY of wide specified by MODE. This is essentially | |
8437 | a * 0x10101010, but we can do slightly better than | |
8438 | synth_mult by unwinding the sequence by hand on CPUs with | |
8439 | slow multiply. */ | |
8440 | static rtx | |
8441 | promote_duplicated_reg (machine_mode mode, rtx val) | |
8442 | { | |
8443 | machine_mode valmode = GET_MODE (val); | |
8444 | rtx tmp; | |
8445 | int nops = mode == DImode ? 3 : 2; | |
8446 | ||
8447 | gcc_assert (mode == SImode || mode == DImode || val == const0_rtx); | |
8448 | if (val == const0_rtx) | |
8449 | return copy_to_mode_reg (mode, CONST0_RTX (mode)); | |
8450 | if (CONST_INT_P (val)) | |
8451 | { | |
8452 | HOST_WIDE_INT v = INTVAL (val) & 255; | |
8453 | ||
8454 | v |= v << 8; | |
8455 | v |= v << 16; | |
8456 | if (mode == DImode) | |
8457 | v |= (v << 16) << 16; | |
8458 | return copy_to_mode_reg (mode, gen_int_mode (v, mode)); | |
8459 | } | |
8460 | ||
8461 | if (valmode == VOIDmode) | |
8462 | valmode = QImode; | |
8463 | if (valmode != QImode) | |
8464 | val = gen_lowpart (QImode, val); | |
8465 | if (mode == QImode) | |
8466 | return val; | |
8467 | if (!TARGET_PARTIAL_REG_STALL) | |
8468 | nops--; | |
8469 | if (ix86_cost->mult_init[mode == DImode ? 3 : 2] | |
8470 | + ix86_cost->mult_bit * (mode == DImode ? 8 : 4) | |
8471 | <= (ix86_cost->shift_const + ix86_cost->add) * nops | |
8472 | + (COSTS_N_INSNS (TARGET_PARTIAL_REG_STALL == 0))) | |
8473 | { | |
8474 | rtx reg = convert_modes (mode, QImode, val, true); | |
8475 | tmp = promote_duplicated_reg (mode, const1_rtx); | |
8476 | return expand_simple_binop (mode, MULT, reg, tmp, NULL, 1, | |
8477 | OPTAB_DIRECT); | |
8478 | } | |
8479 | else | |
8480 | { | |
8481 | rtx reg = convert_modes (mode, QImode, val, true); | |
8482 | ||
8483 | if (!TARGET_PARTIAL_REG_STALL) | |
e9539592 | 8484 | emit_insn (gen_insv_1 (mode, reg, reg)); |
2bf6d935 ML |
8485 | else |
8486 | { | |
8487 | tmp = expand_simple_binop (mode, ASHIFT, reg, GEN_INT (8), | |
8488 | NULL, 1, OPTAB_DIRECT); | |
8489 | reg = expand_simple_binop (mode, IOR, reg, tmp, reg, 1, | |
8490 | OPTAB_DIRECT); | |
8491 | } | |
8492 | tmp = expand_simple_binop (mode, ASHIFT, reg, GEN_INT (16), | |
8493 | NULL, 1, OPTAB_DIRECT); | |
8494 | reg = expand_simple_binop (mode, IOR, reg, tmp, reg, 1, OPTAB_DIRECT); | |
8495 | if (mode == SImode) | |
8496 | return reg; | |
8497 | tmp = expand_simple_binop (mode, ASHIFT, reg, GEN_INT (32), | |
8498 | NULL, 1, OPTAB_DIRECT); | |
8499 | reg = expand_simple_binop (mode, IOR, reg, tmp, reg, 1, OPTAB_DIRECT); | |
8500 | return reg; | |
8501 | } | |
8502 | } | |
8503 | ||
8504 | /* Duplicate value VAL using promote_duplicated_reg into maximal size that will | |
8505 | be needed by main loop copying SIZE_NEEDED chunks and prologue getting | |
8506 | alignment from ALIGN to DESIRED_ALIGN. */ | |
8507 | static rtx | |
8508 | promote_duplicated_reg_to_size (rtx val, int size_needed, int desired_align, | |
8509 | int align) | |
8510 | { | |
8511 | rtx promoted_val; | |
8512 | ||
8513 | if (TARGET_64BIT | |
8514 | && (size_needed > 4 || (desired_align > align && desired_align > 4))) | |
8515 | promoted_val = promote_duplicated_reg (DImode, val); | |
8516 | else if (size_needed > 2 || (desired_align > align && desired_align > 2)) | |
8517 | promoted_val = promote_duplicated_reg (SImode, val); | |
8518 | else if (size_needed > 1 || (desired_align > align && desired_align > 1)) | |
8519 | promoted_val = promote_duplicated_reg (HImode, val); | |
8520 | else | |
8521 | promoted_val = val; | |
8522 | ||
8523 | return promoted_val; | |
8524 | } | |
8525 | ||
8526 | /* Copy the address to a Pmode register. This is used for x32 to | |
8527 | truncate DImode TLS address to a SImode register. */ | |
8528 | ||
8529 | static rtx | |
8530 | ix86_copy_addr_to_reg (rtx addr) | |
8531 | { | |
8532 | rtx reg; | |
8533 | if (GET_MODE (addr) == Pmode || GET_MODE (addr) == VOIDmode) | |
8534 | { | |
8535 | reg = copy_addr_to_reg (addr); | |
8536 | REG_POINTER (reg) = 1; | |
8537 | return reg; | |
8538 | } | |
8539 | else | |
8540 | { | |
8541 | gcc_assert (GET_MODE (addr) == DImode && Pmode == SImode); | |
8542 | reg = copy_to_mode_reg (DImode, addr); | |
8543 | REG_POINTER (reg) = 1; | |
8544 | return gen_rtx_SUBREG (SImode, reg, 0); | |
8545 | } | |
8546 | } | |
8547 | ||
8548 | /* Expand string move (memcpy) ot store (memset) operation. Use i386 string | |
8549 | operations when profitable. The code depends upon architecture, block size | |
8550 | and alignment, but always has one of the following overall structures: | |
8551 | ||
8552 | Aligned move sequence: | |
8553 | ||
8554 | 1) Prologue guard: Conditional that jumps up to epilogues for small | |
8555 | blocks that can be handled by epilogue alone. This is faster | |
8556 | but also needed for correctness, since prologue assume the block | |
8557 | is larger than the desired alignment. | |
8558 | ||
8559 | Optional dynamic check for size and libcall for large | |
8560 | blocks is emitted here too, with -minline-stringops-dynamically. | |
8561 | ||
8562 | 2) Prologue: copy first few bytes in order to get destination | |
8563 | aligned to DESIRED_ALIGN. It is emitted only when ALIGN is less | |
8564 | than DESIRED_ALIGN and up to DESIRED_ALIGN - ALIGN bytes can be | |
8565 | copied. We emit either a jump tree on power of two sized | |
8566 | blocks, or a byte loop. | |
8567 | ||
8568 | 3) Main body: the copying loop itself, copying in SIZE_NEEDED chunks | |
8569 | with specified algorithm. | |
8570 | ||
8571 | 4) Epilogue: code copying tail of the block that is too small to be | |
8572 | handled by main body (or up to size guarded by prologue guard). | |
8573 | ||
8574 | Misaligned move sequence | |
8575 | ||
8576 | 1) missaligned move prologue/epilogue containing: | |
8577 | a) Prologue handling small memory blocks and jumping to done_label | |
8578 | (skipped if blocks are known to be large enough) | |
8579 | b) Signle move copying first DESIRED_ALIGN-ALIGN bytes if alignment is | |
8580 | needed by single possibly misaligned move | |
8581 | (skipped if alignment is not needed) | |
8582 | c) Copy of last SIZE_NEEDED bytes by possibly misaligned moves | |
8583 | ||
8584 | 2) Zero size guard dispatching to done_label, if needed | |
8585 | ||
8586 | 3) dispatch to library call, if needed, | |
8587 | ||
8588 | 3) Main body: the copying loop itself, copying in SIZE_NEEDED chunks | |
8589 | with specified algorithm. */ | |
8590 | bool | |
76715c32 | 8591 | ix86_expand_set_or_cpymem (rtx dst, rtx src, rtx count_exp, rtx val_exp, |
2bf6d935 ML |
8592 | rtx align_exp, rtx expected_align_exp, |
8593 | rtx expected_size_exp, rtx min_size_exp, | |
8594 | rtx max_size_exp, rtx probable_max_size_exp, | |
8595 | bool issetmem) | |
8596 | { | |
8597 | rtx destreg; | |
8598 | rtx srcreg = NULL; | |
8599 | rtx_code_label *label = NULL; | |
8600 | rtx tmp; | |
8601 | rtx_code_label *jump_around_label = NULL; | |
8602 | HOST_WIDE_INT align = 1; | |
8603 | unsigned HOST_WIDE_INT count = 0; | |
8604 | HOST_WIDE_INT expected_size = -1; | |
8605 | int size_needed = 0, epilogue_size_needed; | |
8606 | int desired_align = 0, align_bytes = 0; | |
8607 | enum stringop_alg alg; | |
8608 | rtx promoted_val = NULL; | |
8609 | rtx vec_promoted_val = NULL; | |
8610 | bool force_loopy_epilogue = false; | |
8611 | int dynamic_check; | |
8612 | bool need_zero_guard = false; | |
8613 | bool noalign; | |
8614 | machine_mode move_mode = VOIDmode; | |
8615 | machine_mode wider_mode; | |
8616 | int unroll_factor = 1; | |
8617 | /* TODO: Once value ranges are available, fill in proper data. */ | |
8618 | unsigned HOST_WIDE_INT min_size = 0; | |
8619 | unsigned HOST_WIDE_INT max_size = -1; | |
8620 | unsigned HOST_WIDE_INT probable_max_size = -1; | |
8621 | bool misaligned_prologue_used = false; | |
8622 | bool have_as; | |
8623 | ||
8624 | if (CONST_INT_P (align_exp)) | |
8625 | align = INTVAL (align_exp); | |
8626 | /* i386 can do misaligned access on reasonably increased cost. */ | |
8627 | if (CONST_INT_P (expected_align_exp) | |
8628 | && INTVAL (expected_align_exp) > align) | |
8629 | align = INTVAL (expected_align_exp); | |
8630 | /* ALIGN is the minimum of destination and source alignment, but we care here | |
8631 | just about destination alignment. */ | |
8632 | else if (!issetmem | |
8633 | && MEM_ALIGN (dst) > (unsigned HOST_WIDE_INT) align * BITS_PER_UNIT) | |
8634 | align = MEM_ALIGN (dst) / BITS_PER_UNIT; | |
8635 | ||
8636 | if (CONST_INT_P (count_exp)) | |
8637 | { | |
8638 | min_size = max_size = probable_max_size = count = expected_size | |
8639 | = INTVAL (count_exp); | |
8640 | /* When COUNT is 0, there is nothing to do. */ | |
8641 | if (!count) | |
8642 | return true; | |
8643 | } | |
8644 | else | |
8645 | { | |
8646 | if (min_size_exp) | |
8647 | min_size = INTVAL (min_size_exp); | |
8648 | if (max_size_exp) | |
8649 | max_size = INTVAL (max_size_exp); | |
8650 | if (probable_max_size_exp) | |
8651 | probable_max_size = INTVAL (probable_max_size_exp); | |
8652 | if (CONST_INT_P (expected_size_exp)) | |
8653 | expected_size = INTVAL (expected_size_exp); | |
8654 | } | |
8655 | ||
8656 | /* Make sure we don't need to care about overflow later on. */ | |
8657 | if (count > (HOST_WIDE_INT_1U << 30)) | |
8658 | return false; | |
8659 | ||
8660 | have_as = !ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (dst)); | |
8661 | if (!issetmem) | |
8662 | have_as |= !ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (src)); | |
8663 | ||
8664 | /* Step 0: Decide on preferred algorithm, desired alignment and | |
8665 | size of chunks to be copied by main loop. */ | |
8666 | alg = decide_alg (count, expected_size, min_size, probable_max_size, | |
8667 | issetmem, | |
8668 | issetmem && val_exp == const0_rtx, have_as, | |
8669 | &dynamic_check, &noalign, false); | |
8670 | ||
8671 | if (dump_file) | |
8672 | fprintf (dump_file, "Selected stringop expansion strategy: %s\n", | |
8673 | stringop_alg_names[alg]); | |
8674 | ||
8675 | if (alg == libcall) | |
8676 | return false; | |
8677 | gcc_assert (alg != no_stringop); | |
8678 | ||
8679 | /* For now vector-version of memset is generated only for memory zeroing, as | |
8680 | creating of promoted vector value is very cheap in this case. */ | |
8681 | if (issetmem && alg == vector_loop && val_exp != const0_rtx) | |
8682 | alg = unrolled_loop; | |
8683 | ||
8684 | if (!count) | |
8685 | count_exp = copy_to_mode_reg (GET_MODE (count_exp), count_exp); | |
8686 | destreg = ix86_copy_addr_to_reg (XEXP (dst, 0)); | |
8687 | if (!issetmem) | |
8688 | srcreg = ix86_copy_addr_to_reg (XEXP (src, 0)); | |
8689 | ||
8690 | unroll_factor = 1; | |
8691 | move_mode = word_mode; | |
8692 | switch (alg) | |
8693 | { | |
8694 | case libcall: | |
8695 | case no_stringop: | |
8696 | case last_alg: | |
8697 | gcc_unreachable (); | |
8698 | case loop_1_byte: | |
8699 | need_zero_guard = true; | |
8700 | move_mode = QImode; | |
8701 | break; | |
8702 | case loop: | |
8703 | need_zero_guard = true; | |
8704 | break; | |
8705 | case unrolled_loop: | |
8706 | need_zero_guard = true; | |
8707 | unroll_factor = (TARGET_64BIT ? 4 : 2); | |
8708 | break; | |
8709 | case vector_loop: | |
8710 | need_zero_guard = true; | |
8711 | unroll_factor = 4; | |
8712 | /* Find the widest supported mode. */ | |
8713 | move_mode = word_mode; | |
8714 | while (GET_MODE_WIDER_MODE (move_mode).exists (&wider_mode) | |
8715 | && optab_handler (mov_optab, wider_mode) != CODE_FOR_nothing) | |
8716 | move_mode = wider_mode; | |
8717 | ||
586bbef1 | 8718 | if (TARGET_AVX256_SPLIT_REGS && GET_MODE_BITSIZE (move_mode) > 128) |
2bf6d935 | 8719 | move_mode = TImode; |
eef81eef JH |
8720 | if (TARGET_AVX512_SPLIT_REGS && GET_MODE_BITSIZE (move_mode) > 256) |
8721 | move_mode = OImode; | |
2bf6d935 ML |
8722 | |
8723 | /* Find the corresponding vector mode with the same size as MOVE_MODE. | |
8724 | MOVE_MODE is an integer mode at the moment (SI, DI, TI, etc.). */ | |
8725 | if (GET_MODE_SIZE (move_mode) > GET_MODE_SIZE (word_mode)) | |
8726 | { | |
8727 | int nunits = GET_MODE_SIZE (move_mode) / GET_MODE_SIZE (word_mode); | |
8728 | if (!mode_for_vector (word_mode, nunits).exists (&move_mode) | |
8729 | || optab_handler (mov_optab, move_mode) == CODE_FOR_nothing) | |
8730 | move_mode = word_mode; | |
8731 | } | |
8732 | gcc_assert (optab_handler (mov_optab, move_mode) != CODE_FOR_nothing); | |
8733 | break; | |
8734 | case rep_prefix_8_byte: | |
8735 | move_mode = DImode; | |
8736 | break; | |
8737 | case rep_prefix_4_byte: | |
8738 | move_mode = SImode; | |
8739 | break; | |
8740 | case rep_prefix_1_byte: | |
8741 | move_mode = QImode; | |
8742 | break; | |
8743 | } | |
8744 | size_needed = GET_MODE_SIZE (move_mode) * unroll_factor; | |
8745 | epilogue_size_needed = size_needed; | |
8746 | ||
8747 | /* If we are going to call any library calls conditionally, make sure any | |
8748 | pending stack adjustment happen before the first conditional branch, | |
8749 | otherwise they will be emitted before the library call only and won't | |
8750 | happen from the other branches. */ | |
8751 | if (dynamic_check != -1) | |
8752 | do_pending_stack_adjust (); | |
8753 | ||
8754 | desired_align = decide_alignment (align, alg, expected_size, move_mode); | |
8755 | if (!TARGET_ALIGN_STRINGOPS || noalign) | |
8756 | align = desired_align; | |
8757 | ||
8758 | /* Step 1: Prologue guard. */ | |
8759 | ||
8760 | /* Alignment code needs count to be in register. */ | |
8761 | if (CONST_INT_P (count_exp) && desired_align > align) | |
8762 | { | |
8763 | if (INTVAL (count_exp) > desired_align | |
8764 | && INTVAL (count_exp) > size_needed) | |
8765 | { | |
8766 | align_bytes | |
8767 | = get_mem_align_offset (dst, desired_align * BITS_PER_UNIT); | |
8768 | if (align_bytes <= 0) | |
8769 | align_bytes = 0; | |
8770 | else | |
8771 | align_bytes = desired_align - align_bytes; | |
8772 | } | |
8773 | if (align_bytes == 0) | |
8774 | count_exp = force_reg (counter_mode (count_exp), count_exp); | |
8775 | } | |
8776 | gcc_assert (desired_align >= 1 && align >= 1); | |
8777 | ||
8778 | /* Misaligned move sequences handle both prologue and epilogue at once. | |
8779 | Default code generation results in a smaller code for large alignments | |
8780 | and also avoids redundant job when sizes are known precisely. */ | |
8781 | misaligned_prologue_used | |
8782 | = (TARGET_MISALIGNED_MOVE_STRING_PRO_EPILOGUES | |
8783 | && MAX (desired_align, epilogue_size_needed) <= 32 | |
8784 | && desired_align <= epilogue_size_needed | |
8785 | && ((desired_align > align && !align_bytes) | |
8786 | || (!count && epilogue_size_needed > 1))); | |
8787 | ||
8788 | /* Do the cheap promotion to allow better CSE across the | |
8789 | main loop and epilogue (ie one load of the big constant in the | |
8790 | front of all code. | |
8791 | For now the misaligned move sequences do not have fast path | |
8792 | without broadcasting. */ | |
8793 | if (issetmem && ((CONST_INT_P (val_exp) || misaligned_prologue_used))) | |
8794 | { | |
8795 | if (alg == vector_loop) | |
8796 | { | |
8797 | gcc_assert (val_exp == const0_rtx); | |
8798 | vec_promoted_val = promote_duplicated_reg (move_mode, val_exp); | |
8799 | promoted_val = promote_duplicated_reg_to_size (val_exp, | |
8800 | GET_MODE_SIZE (word_mode), | |
8801 | desired_align, align); | |
8802 | } | |
8803 | else | |
8804 | { | |
8805 | promoted_val = promote_duplicated_reg_to_size (val_exp, size_needed, | |
8806 | desired_align, align); | |
8807 | } | |
8808 | } | |
8809 | /* Misaligned move sequences handles both prologues and epilogues at once. | |
8810 | Default code generation results in smaller code for large alignments and | |
8811 | also avoids redundant job when sizes are known precisely. */ | |
8812 | if (misaligned_prologue_used) | |
8813 | { | |
8814 | /* Misaligned move prologue handled small blocks by itself. */ | |
76715c32 | 8815 | expand_set_or_cpymem_prologue_epilogue_by_misaligned_moves |
2bf6d935 ML |
8816 | (dst, src, &destreg, &srcreg, |
8817 | move_mode, promoted_val, vec_promoted_val, | |
8818 | &count_exp, | |
8819 | &jump_around_label, | |
8820 | desired_align < align | |
8821 | ? MAX (desired_align, epilogue_size_needed) : epilogue_size_needed, | |
8822 | desired_align, align, &min_size, dynamic_check, issetmem); | |
8823 | if (!issetmem) | |
8824 | src = change_address (src, BLKmode, srcreg); | |
8825 | dst = change_address (dst, BLKmode, destreg); | |
8826 | set_mem_align (dst, desired_align * BITS_PER_UNIT); | |
8827 | epilogue_size_needed = 0; | |
8828 | if (need_zero_guard | |
8829 | && min_size < (unsigned HOST_WIDE_INT) size_needed) | |
8830 | { | |
8831 | /* It is possible that we copied enough so the main loop will not | |
8832 | execute. */ | |
8833 | gcc_assert (size_needed > 1); | |
8834 | if (jump_around_label == NULL_RTX) | |
8835 | jump_around_label = gen_label_rtx (); | |
8836 | emit_cmp_and_jump_insns (count_exp, | |
8837 | GEN_INT (size_needed), | |
8838 | LTU, 0, counter_mode (count_exp), 1, jump_around_label); | |
8839 | if (expected_size == -1 | |
8840 | || expected_size < (desired_align - align) / 2 + size_needed) | |
8841 | predict_jump (REG_BR_PROB_BASE * 20 / 100); | |
8842 | else | |
8843 | predict_jump (REG_BR_PROB_BASE * 60 / 100); | |
8844 | } | |
8845 | } | |
8846 | /* Ensure that alignment prologue won't copy past end of block. */ | |
8847 | else if (size_needed > 1 || (desired_align > 1 && desired_align > align)) | |
8848 | { | |
8849 | epilogue_size_needed = MAX (size_needed - 1, desired_align - align); | |
8850 | /* Epilogue always copies COUNT_EXP & EPILOGUE_SIZE_NEEDED bytes. | |
8851 | Make sure it is power of 2. */ | |
8852 | epilogue_size_needed = 1 << (floor_log2 (epilogue_size_needed) + 1); | |
8853 | ||
8854 | /* To improve performance of small blocks, we jump around the VAL | |
8855 | promoting mode. This mean that if the promoted VAL is not constant, | |
8856 | we might not use it in the epilogue and have to use byte | |
8857 | loop variant. */ | |
8858 | if (issetmem && epilogue_size_needed > 2 && !promoted_val) | |
8859 | force_loopy_epilogue = true; | |
8860 | if ((count && count < (unsigned HOST_WIDE_INT) epilogue_size_needed) | |
8861 | || max_size < (unsigned HOST_WIDE_INT) epilogue_size_needed) | |
8862 | { | |
8863 | /* If main algorithm works on QImode, no epilogue is needed. | |
8864 | For small sizes just don't align anything. */ | |
8865 | if (size_needed == 1) | |
8866 | desired_align = align; | |
8867 | else | |
8868 | goto epilogue; | |
8869 | } | |
8870 | else if (!count | |
8871 | && min_size < (unsigned HOST_WIDE_INT) epilogue_size_needed) | |
8872 | { | |
8873 | label = gen_label_rtx (); | |
8874 | emit_cmp_and_jump_insns (count_exp, | |
8875 | GEN_INT (epilogue_size_needed), | |
8876 | LTU, 0, counter_mode (count_exp), 1, label); | |
8877 | if (expected_size == -1 || expected_size < epilogue_size_needed) | |
8878 | predict_jump (REG_BR_PROB_BASE * 60 / 100); | |
8879 | else | |
8880 | predict_jump (REG_BR_PROB_BASE * 20 / 100); | |
8881 | } | |
8882 | } | |
8883 | ||
8884 | /* Emit code to decide on runtime whether library call or inline should be | |
8885 | used. */ | |
8886 | if (dynamic_check != -1) | |
8887 | { | |
8888 | if (!issetmem && CONST_INT_P (count_exp)) | |
8889 | { | |
8890 | if (UINTVAL (count_exp) >= (unsigned HOST_WIDE_INT)dynamic_check) | |
8891 | { | |
8892 | emit_block_copy_via_libcall (dst, src, count_exp); | |
8893 | count_exp = const0_rtx; | |
8894 | goto epilogue; | |
8895 | } | |
8896 | } | |
8897 | else | |
8898 | { | |
8899 | rtx_code_label *hot_label = gen_label_rtx (); | |
8900 | if (jump_around_label == NULL_RTX) | |
8901 | jump_around_label = gen_label_rtx (); | |
8902 | emit_cmp_and_jump_insns (count_exp, GEN_INT (dynamic_check - 1), | |
8903 | LEU, 0, counter_mode (count_exp), | |
8904 | 1, hot_label); | |
8905 | predict_jump (REG_BR_PROB_BASE * 90 / 100); | |
8906 | if (issetmem) | |
8907 | set_storage_via_libcall (dst, count_exp, val_exp); | |
8908 | else | |
8909 | emit_block_copy_via_libcall (dst, src, count_exp); | |
8910 | emit_jump (jump_around_label); | |
8911 | emit_label (hot_label); | |
8912 | } | |
8913 | } | |
8914 | ||
8915 | /* Step 2: Alignment prologue. */ | |
8916 | /* Do the expensive promotion once we branched off the small blocks. */ | |
8917 | if (issetmem && !promoted_val) | |
8918 | promoted_val = promote_duplicated_reg_to_size (val_exp, size_needed, | |
8919 | desired_align, align); | |
8920 | ||
8921 | if (desired_align > align && !misaligned_prologue_used) | |
8922 | { | |
8923 | if (align_bytes == 0) | |
8924 | { | |
8925 | /* Except for the first move in prologue, we no longer know | |
8926 | constant offset in aliasing info. It don't seems to worth | |
8927 | the pain to maintain it for the first move, so throw away | |
8928 | the info early. */ | |
8929 | dst = change_address (dst, BLKmode, destreg); | |
8930 | if (!issetmem) | |
8931 | src = change_address (src, BLKmode, srcreg); | |
76715c32 | 8932 | dst = expand_set_or_cpymem_prologue (dst, src, destreg, srcreg, |
2bf6d935 ML |
8933 | promoted_val, vec_promoted_val, |
8934 | count_exp, align, desired_align, | |
8935 | issetmem); | |
8936 | /* At most desired_align - align bytes are copied. */ | |
8937 | if (min_size < (unsigned)(desired_align - align)) | |
8938 | min_size = 0; | |
8939 | else | |
8940 | min_size -= desired_align - align; | |
8941 | } | |
8942 | else | |
8943 | { | |
8944 | /* If we know how many bytes need to be stored before dst is | |
8945 | sufficiently aligned, maintain aliasing info accurately. */ | |
76715c32 | 8946 | dst = expand_set_or_cpymem_constant_prologue (dst, &src, destreg, |
2bf6d935 ML |
8947 | srcreg, |
8948 | promoted_val, | |
8949 | vec_promoted_val, | |
8950 | desired_align, | |
8951 | align_bytes, | |
8952 | issetmem); | |
8953 | ||
8954 | count_exp = plus_constant (counter_mode (count_exp), | |
8955 | count_exp, -align_bytes); | |
8956 | count -= align_bytes; | |
8957 | min_size -= align_bytes; | |
8958 | max_size -= align_bytes; | |
8959 | } | |
8960 | if (need_zero_guard | |
8961 | && min_size < (unsigned HOST_WIDE_INT) size_needed | |
8962 | && (count < (unsigned HOST_WIDE_INT) size_needed | |
8963 | || (align_bytes == 0 | |
8964 | && count < ((unsigned HOST_WIDE_INT) size_needed | |
8965 | + desired_align - align)))) | |
8966 | { | |
8967 | /* It is possible that we copied enough so the main loop will not | |
8968 | execute. */ | |
8969 | gcc_assert (size_needed > 1); | |
8970 | if (label == NULL_RTX) | |
8971 | label = gen_label_rtx (); | |
8972 | emit_cmp_and_jump_insns (count_exp, | |
8973 | GEN_INT (size_needed), | |
8974 | LTU, 0, counter_mode (count_exp), 1, label); | |
8975 | if (expected_size == -1 | |
8976 | || expected_size < (desired_align - align) / 2 + size_needed) | |
8977 | predict_jump (REG_BR_PROB_BASE * 20 / 100); | |
8978 | else | |
8979 | predict_jump (REG_BR_PROB_BASE * 60 / 100); | |
8980 | } | |
8981 | } | |
8982 | if (label && size_needed == 1) | |
8983 | { | |
8984 | emit_label (label); | |
8985 | LABEL_NUSES (label) = 1; | |
8986 | label = NULL; | |
8987 | epilogue_size_needed = 1; | |
8988 | if (issetmem) | |
8989 | promoted_val = val_exp; | |
8990 | } | |
8991 | else if (label == NULL_RTX && !misaligned_prologue_used) | |
8992 | epilogue_size_needed = size_needed; | |
8993 | ||
8994 | /* Step 3: Main loop. */ | |
8995 | ||
8996 | switch (alg) | |
8997 | { | |
8998 | case libcall: | |
8999 | case no_stringop: | |
9000 | case last_alg: | |
9001 | gcc_unreachable (); | |
9002 | case loop_1_byte: | |
9003 | case loop: | |
9004 | case unrolled_loop: | |
76715c32 | 9005 | expand_set_or_cpymem_via_loop (dst, src, destreg, srcreg, promoted_val, |
2bf6d935 ML |
9006 | count_exp, move_mode, unroll_factor, |
9007 | expected_size, issetmem); | |
9008 | break; | |
9009 | case vector_loop: | |
76715c32 | 9010 | expand_set_or_cpymem_via_loop (dst, src, destreg, srcreg, |
2bf6d935 ML |
9011 | vec_promoted_val, count_exp, move_mode, |
9012 | unroll_factor, expected_size, issetmem); | |
9013 | break; | |
9014 | case rep_prefix_8_byte: | |
9015 | case rep_prefix_4_byte: | |
9016 | case rep_prefix_1_byte: | |
76715c32 | 9017 | expand_set_or_cpymem_via_rep (dst, src, destreg, srcreg, promoted_val, |
2bf6d935 ML |
9018 | val_exp, count_exp, move_mode, issetmem); |
9019 | break; | |
9020 | } | |
9021 | /* Adjust properly the offset of src and dest memory for aliasing. */ | |
9022 | if (CONST_INT_P (count_exp)) | |
9023 | { | |
9024 | if (!issetmem) | |
9025 | src = adjust_automodify_address_nv (src, BLKmode, srcreg, | |
9026 | (count / size_needed) * size_needed); | |
9027 | dst = adjust_automodify_address_nv (dst, BLKmode, destreg, | |
9028 | (count / size_needed) * size_needed); | |
9029 | } | |
9030 | else | |
9031 | { | |
9032 | if (!issetmem) | |
9033 | src = change_address (src, BLKmode, srcreg); | |
9034 | dst = change_address (dst, BLKmode, destreg); | |
9035 | } | |
9036 | ||
9037 | /* Step 4: Epilogue to copy the remaining bytes. */ | |
9038 | epilogue: | |
9039 | if (label) | |
9040 | { | |
9041 | /* When the main loop is done, COUNT_EXP might hold original count, | |
9042 | while we want to copy only COUNT_EXP & SIZE_NEEDED bytes. | |
9043 | Epilogue code will actually copy COUNT_EXP & EPILOGUE_SIZE_NEEDED | |
9044 | bytes. Compensate if needed. */ | |
9045 | ||
9046 | if (size_needed < epilogue_size_needed) | |
9047 | { | |
9048 | tmp = expand_simple_binop (counter_mode (count_exp), AND, count_exp, | |
9049 | GEN_INT (size_needed - 1), count_exp, 1, | |
9050 | OPTAB_DIRECT); | |
9051 | if (tmp != count_exp) | |
9052 | emit_move_insn (count_exp, tmp); | |
9053 | } | |
9054 | emit_label (label); | |
9055 | LABEL_NUSES (label) = 1; | |
9056 | } | |
9057 | ||
9058 | if (count_exp != const0_rtx && epilogue_size_needed > 1) | |
9059 | { | |
9060 | if (force_loopy_epilogue) | |
9061 | expand_setmem_epilogue_via_loop (dst, destreg, val_exp, count_exp, | |
9062 | epilogue_size_needed); | |
9063 | else | |
9064 | { | |
9065 | if (issetmem) | |
9066 | expand_setmem_epilogue (dst, destreg, promoted_val, | |
9067 | vec_promoted_val, count_exp, | |
9068 | epilogue_size_needed); | |
9069 | else | |
76715c32 | 9070 | expand_cpymem_epilogue (dst, src, destreg, srcreg, count_exp, |
2bf6d935 ML |
9071 | epilogue_size_needed); |
9072 | } | |
9073 | } | |
9074 | if (jump_around_label) | |
9075 | emit_label (jump_around_label); | |
9076 | return true; | |
9077 | } | |
9078 | ||
3edc21af L |
9079 | /* Expand cmpstrn or memcmp. */ |
9080 | ||
9081 | bool | |
9082 | ix86_expand_cmpstrn_or_cmpmem (rtx result, rtx src1, rtx src2, | |
9083 | rtx length, rtx align, bool is_cmpstrn) | |
9084 | { | |
4052c05e L |
9085 | /* Expand strncmp and memcmp only with -minline-all-stringops since |
9086 | "repz cmpsb" can be much slower than strncmp and memcmp functions | |
9087 | implemented with vector instructions, see | |
9088 | ||
9089 | https://gcc.gnu.org/bugzilla/show_bug.cgi?id=43052 | |
9090 | */ | |
9091 | if (!TARGET_INLINE_ALL_STRINGOPS) | |
3edc21af L |
9092 | return false; |
9093 | ||
9094 | /* Can't use this if the user has appropriated ecx, esi or edi. */ | |
9095 | if (fixed_regs[CX_REG] || fixed_regs[SI_REG] || fixed_regs[DI_REG]) | |
9096 | return false; | |
9097 | ||
9098 | if (is_cmpstrn) | |
9099 | { | |
9100 | /* For strncmp, length is the maximum length, which can be larger | |
9101 | than actual string lengths. We can expand the cmpstrn pattern | |
9102 | to "repz cmpsb" only if one of the strings is a constant so | |
9103 | that expand_builtin_strncmp() can write the length argument to | |
9104 | be the minimum of the const string length and the actual length | |
9105 | argument. Otherwise, "repz cmpsb" may pass the 0 byte. */ | |
9106 | tree t1 = MEM_EXPR (src1); | |
9107 | tree t2 = MEM_EXPR (src2); | |
9108 | if (!((t1 && TREE_CODE (t1) == MEM_REF | |
9109 | && TREE_CODE (TREE_OPERAND (t1, 0)) == ADDR_EXPR | |
9110 | && (TREE_CODE (TREE_OPERAND (TREE_OPERAND (t1, 0), 0)) | |
9111 | == STRING_CST)) | |
9112 | || (t2 && TREE_CODE (t2) == MEM_REF | |
9113 | && TREE_CODE (TREE_OPERAND (t2, 0)) == ADDR_EXPR | |
9114 | && (TREE_CODE (TREE_OPERAND (TREE_OPERAND (t2, 0), 0)) | |
9115 | == STRING_CST)))) | |
9116 | return false; | |
9117 | } | |
3edc21af L |
9118 | |
9119 | rtx addr1 = copy_addr_to_reg (XEXP (src1, 0)); | |
9120 | rtx addr2 = copy_addr_to_reg (XEXP (src2, 0)); | |
9121 | if (addr1 != XEXP (src1, 0)) | |
9122 | src1 = replace_equiv_address_nv (src1, addr1); | |
9123 | if (addr2 != XEXP (src2, 0)) | |
9124 | src2 = replace_equiv_address_nv (src2, addr2); | |
9125 | ||
9126 | /* NB: Make a copy of the data length to avoid changing the original | |
9127 | data length by cmpstrnqi patterns. */ | |
9128 | length = ix86_zero_extend_to_Pmode (length); | |
9129 | rtx lengthreg = gen_reg_rtx (Pmode); | |
9130 | emit_move_insn (lengthreg, length); | |
9131 | ||
9132 | /* If we are testing strict equality, we can use known alignment to | |
9133 | good advantage. This may be possible with combine, particularly | |
9134 | once cc0 is dead. */ | |
9135 | if (CONST_INT_P (length)) | |
9136 | { | |
9137 | if (length == const0_rtx) | |
9138 | { | |
9139 | emit_move_insn (result, const0_rtx); | |
9140 | return true; | |
9141 | } | |
9142 | emit_insn (gen_cmpstrnqi_nz_1 (addr1, addr2, lengthreg, align, | |
9143 | src1, src2)); | |
9144 | } | |
9145 | else | |
9146 | { | |
9147 | emit_insn (gen_cmp_1 (Pmode, lengthreg, lengthreg)); | |
9148 | emit_insn (gen_cmpstrnqi_1 (addr1, addr2, lengthreg, align, | |
9149 | src1, src2)); | |
9150 | } | |
9151 | ||
9152 | rtx out = gen_lowpart (QImode, result); | |
9153 | emit_insn (gen_cmpintqi (out)); | |
9154 | emit_move_insn (result, gen_rtx_SIGN_EXTEND (SImode, out)); | |
9155 | ||
9156 | return true; | |
9157 | } | |
2bf6d935 ML |
9158 | |
9159 | /* Expand the appropriate insns for doing strlen if not just doing | |
9160 | repnz; scasb | |
9161 | ||
9162 | out = result, initialized with the start address | |
9163 | align_rtx = alignment of the address. | |
9164 | scratch = scratch register, initialized with the startaddress when | |
9165 | not aligned, otherwise undefined | |
9166 | ||
9167 | This is just the body. It needs the initializations mentioned above and | |
9168 | some address computing at the end. These things are done in i386.md. */ | |
9169 | ||
9170 | static void | |
9171 | ix86_expand_strlensi_unroll_1 (rtx out, rtx src, rtx align_rtx) | |
9172 | { | |
9173 | int align; | |
9174 | rtx tmp; | |
9175 | rtx_code_label *align_2_label = NULL; | |
9176 | rtx_code_label *align_3_label = NULL; | |
9177 | rtx_code_label *align_4_label = gen_label_rtx (); | |
9178 | rtx_code_label *end_0_label = gen_label_rtx (); | |
9179 | rtx mem; | |
9180 | rtx tmpreg = gen_reg_rtx (SImode); | |
9181 | rtx scratch = gen_reg_rtx (SImode); | |
9182 | rtx cmp; | |
9183 | ||
9184 | align = 0; | |
9185 | if (CONST_INT_P (align_rtx)) | |
9186 | align = INTVAL (align_rtx); | |
9187 | ||
9188 | /* Loop to check 1..3 bytes for null to get an aligned pointer. */ | |
9189 | ||
9190 | /* Is there a known alignment and is it less than 4? */ | |
9191 | if (align < 4) | |
9192 | { | |
9193 | rtx scratch1 = gen_reg_rtx (Pmode); | |
9194 | emit_move_insn (scratch1, out); | |
9195 | /* Is there a known alignment and is it not 2? */ | |
9196 | if (align != 2) | |
9197 | { | |
9198 | align_3_label = gen_label_rtx (); /* Label when aligned to 3-byte */ | |
9199 | align_2_label = gen_label_rtx (); /* Label when aligned to 2-byte */ | |
9200 | ||
9201 | /* Leave just the 3 lower bits. */ | |
9202 | align_rtx = expand_binop (Pmode, and_optab, scratch1, GEN_INT (3), | |
9203 | NULL_RTX, 0, OPTAB_WIDEN); | |
9204 | ||
9205 | emit_cmp_and_jump_insns (align_rtx, const0_rtx, EQ, NULL, | |
9206 | Pmode, 1, align_4_label); | |
9207 | emit_cmp_and_jump_insns (align_rtx, const2_rtx, EQ, NULL, | |
9208 | Pmode, 1, align_2_label); | |
9209 | emit_cmp_and_jump_insns (align_rtx, const2_rtx, GTU, NULL, | |
9210 | Pmode, 1, align_3_label); | |
9211 | } | |
9212 | else | |
9213 | { | |
9214 | /* Since the alignment is 2, we have to check 2 or 0 bytes; | |
9215 | check if is aligned to 4 - byte. */ | |
9216 | ||
9217 | align_rtx = expand_binop (Pmode, and_optab, scratch1, const2_rtx, | |
9218 | NULL_RTX, 0, OPTAB_WIDEN); | |
9219 | ||
9220 | emit_cmp_and_jump_insns (align_rtx, const0_rtx, EQ, NULL, | |
9221 | Pmode, 1, align_4_label); | |
9222 | } | |
9223 | ||
9224 | mem = change_address (src, QImode, out); | |
9225 | ||
9226 | /* Now compare the bytes. */ | |
9227 | ||
9228 | /* Compare the first n unaligned byte on a byte per byte basis. */ | |
9229 | emit_cmp_and_jump_insns (mem, const0_rtx, EQ, NULL, | |
9230 | QImode, 1, end_0_label); | |
9231 | ||
9232 | /* Increment the address. */ | |
d9330fb5 | 9233 | emit_insn (gen_add2_insn (out, const1_rtx)); |
2bf6d935 ML |
9234 | |
9235 | /* Not needed with an alignment of 2 */ | |
9236 | if (align != 2) | |
9237 | { | |
9238 | emit_label (align_2_label); | |
9239 | ||
9240 | emit_cmp_and_jump_insns (mem, const0_rtx, EQ, NULL, QImode, 1, | |
9241 | end_0_label); | |
9242 | ||
d9330fb5 | 9243 | emit_insn (gen_add2_insn (out, const1_rtx)); |
2bf6d935 ML |
9244 | |
9245 | emit_label (align_3_label); | |
9246 | } | |
9247 | ||
9248 | emit_cmp_and_jump_insns (mem, const0_rtx, EQ, NULL, QImode, 1, | |
9249 | end_0_label); | |
9250 | ||
d9330fb5 | 9251 | emit_insn (gen_add2_insn (out, const1_rtx)); |
2bf6d935 ML |
9252 | } |
9253 | ||
9254 | /* Generate loop to check 4 bytes at a time. It is not a good idea to | |
9255 | align this loop. It gives only huge programs, but does not help to | |
9256 | speed up. */ | |
9257 | emit_label (align_4_label); | |
9258 | ||
9259 | mem = change_address (src, SImode, out); | |
9260 | emit_move_insn (scratch, mem); | |
d9330fb5 | 9261 | emit_insn (gen_add2_insn (out, GEN_INT (4))); |
2bf6d935 ML |
9262 | |
9263 | /* This formula yields a nonzero result iff one of the bytes is zero. | |
9264 | This saves three branches inside loop and many cycles. */ | |
9265 | ||
9266 | emit_insn (gen_addsi3 (tmpreg, scratch, GEN_INT (-0x01010101))); | |
9267 | emit_insn (gen_one_cmplsi2 (scratch, scratch)); | |
9268 | emit_insn (gen_andsi3 (tmpreg, tmpreg, scratch)); | |
9269 | emit_insn (gen_andsi3 (tmpreg, tmpreg, | |
9270 | gen_int_mode (0x80808080, SImode))); | |
9271 | emit_cmp_and_jump_insns (tmpreg, const0_rtx, EQ, 0, SImode, 1, | |
9272 | align_4_label); | |
9273 | ||
9274 | if (TARGET_CMOVE) | |
9275 | { | |
9276 | rtx reg = gen_reg_rtx (SImode); | |
9277 | rtx reg2 = gen_reg_rtx (Pmode); | |
9278 | emit_move_insn (reg, tmpreg); | |
9279 | emit_insn (gen_lshrsi3 (reg, reg, GEN_INT (16))); | |
9280 | ||
9281 | /* If zero is not in the first two bytes, move two bytes forward. */ | |
9282 | emit_insn (gen_testsi_ccno_1 (tmpreg, GEN_INT (0x8080))); | |
9283 | tmp = gen_rtx_REG (CCNOmode, FLAGS_REG); | |
9284 | tmp = gen_rtx_EQ (VOIDmode, tmp, const0_rtx); | |
9285 | emit_insn (gen_rtx_SET (tmpreg, | |
9286 | gen_rtx_IF_THEN_ELSE (SImode, tmp, | |
9287 | reg, | |
9288 | tmpreg))); | |
9289 | /* Emit lea manually to avoid clobbering of flags. */ | |
c3185b64 | 9290 | emit_insn (gen_rtx_SET (reg2, plus_constant (Pmode, out, 2))); |
2bf6d935 ML |
9291 | |
9292 | tmp = gen_rtx_REG (CCNOmode, FLAGS_REG); | |
9293 | tmp = gen_rtx_EQ (VOIDmode, tmp, const0_rtx); | |
9294 | emit_insn (gen_rtx_SET (out, | |
9295 | gen_rtx_IF_THEN_ELSE (Pmode, tmp, | |
9296 | reg2, | |
9297 | out))); | |
9298 | } | |
9299 | else | |
9300 | { | |
9301 | rtx_code_label *end_2_label = gen_label_rtx (); | |
9302 | /* Is zero in the first two bytes? */ | |
9303 | ||
9304 | emit_insn (gen_testsi_ccno_1 (tmpreg, GEN_INT (0x8080))); | |
9305 | tmp = gen_rtx_REG (CCNOmode, FLAGS_REG); | |
9306 | tmp = gen_rtx_NE (VOIDmode, tmp, const0_rtx); | |
9307 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp, | |
9308 | gen_rtx_LABEL_REF (VOIDmode, end_2_label), | |
9309 | pc_rtx); | |
9310 | tmp = emit_jump_insn (gen_rtx_SET (pc_rtx, tmp)); | |
9311 | JUMP_LABEL (tmp) = end_2_label; | |
9312 | ||
9313 | /* Not in the first two. Move two bytes forward. */ | |
9314 | emit_insn (gen_lshrsi3 (tmpreg, tmpreg, GEN_INT (16))); | |
d9330fb5 | 9315 | emit_insn (gen_add2_insn (out, const2_rtx)); |
2bf6d935 ML |
9316 | |
9317 | emit_label (end_2_label); | |
9318 | ||
9319 | } | |
9320 | ||
9321 | /* Avoid branch in fixing the byte. */ | |
9322 | tmpreg = gen_lowpart (QImode, tmpreg); | |
9323 | emit_insn (gen_addqi3_cconly_overflow (tmpreg, tmpreg)); | |
9324 | tmp = gen_rtx_REG (CCmode, FLAGS_REG); | |
9325 | cmp = gen_rtx_LTU (VOIDmode, tmp, const0_rtx); | |
d9330fb5 | 9326 | emit_insn (gen_sub3_carry (Pmode, out, out, GEN_INT (3), tmp, cmp)); |
2bf6d935 ML |
9327 | |
9328 | emit_label (end_0_label); | |
9329 | } | |
9330 | ||
9331 | /* Expand strlen. */ | |
9332 | ||
9333 | bool | |
9334 | ix86_expand_strlen (rtx out, rtx src, rtx eoschar, rtx align) | |
9335 | { | |
9336 | if (TARGET_UNROLL_STRLEN | |
9337 | && TARGET_INLINE_ALL_STRINGOPS | |
9338 | && eoschar == const0_rtx | |
9339 | && optimize > 1) | |
9340 | { | |
9341 | /* The generic case of strlen expander is long. Avoid it's | |
9342 | expanding unless TARGET_INLINE_ALL_STRINGOPS. */ | |
9343 | rtx addr = force_reg (Pmode, XEXP (src, 0)); | |
9344 | /* Well it seems that some optimizer does not combine a call like | |
9345 | foo(strlen(bar), strlen(bar)); | |
9346 | when the move and the subtraction is done here. It does calculate | |
9347 | the length just once when these instructions are done inside of | |
9348 | output_strlen_unroll(). But I think since &bar[strlen(bar)] is | |
9349 | often used and I use one fewer register for the lifetime of | |
9350 | output_strlen_unroll() this is better. */ | |
9351 | ||
9352 | emit_move_insn (out, addr); | |
9353 | ||
9354 | ix86_expand_strlensi_unroll_1 (out, src, align); | |
9355 | ||
9356 | /* strlensi_unroll_1 returns the address of the zero at the end of | |
9357 | the string, like memchr(), so compute the length by subtracting | |
9358 | the start address. */ | |
d9330fb5 | 9359 | emit_insn (gen_sub2_insn (out, addr)); |
2bf6d935 ML |
9360 | return true; |
9361 | } | |
9362 | else | |
9363 | return false; | |
9364 | } | |
9365 | ||
9366 | /* For given symbol (function) construct code to compute address of it's PLT | |
9367 | entry in large x86-64 PIC model. */ | |
9368 | ||
9369 | static rtx | |
9370 | construct_plt_address (rtx symbol) | |
9371 | { | |
9372 | rtx tmp, unspec; | |
9373 | ||
9374 | gcc_assert (GET_CODE (symbol) == SYMBOL_REF); | |
9375 | gcc_assert (ix86_cmodel == CM_LARGE_PIC && !TARGET_PECOFF); | |
9376 | gcc_assert (Pmode == DImode); | |
9377 | ||
9378 | tmp = gen_reg_rtx (Pmode); | |
9379 | unspec = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, symbol), UNSPEC_PLTOFF); | |
9380 | ||
9381 | emit_move_insn (tmp, gen_rtx_CONST (Pmode, unspec)); | |
d9330fb5 | 9382 | emit_insn (gen_add2_insn (tmp, pic_offset_table_rtx)); |
2bf6d935 ML |
9383 | return tmp; |
9384 | } | |
9385 | ||
9386 | /* Additional registers that are clobbered by SYSV calls. */ | |
9387 | ||
9388 | static int const x86_64_ms_sysv_extra_clobbered_registers | |
9389 | [NUM_X86_64_MS_CLOBBERED_REGS] = | |
9390 | { | |
9391 | SI_REG, DI_REG, | |
9392 | XMM6_REG, XMM7_REG, | |
9393 | XMM8_REG, XMM9_REG, XMM10_REG, XMM11_REG, | |
9394 | XMM12_REG, XMM13_REG, XMM14_REG, XMM15_REG | |
9395 | }; | |
9396 | ||
9397 | rtx_insn * | |
9398 | ix86_expand_call (rtx retval, rtx fnaddr, rtx callarg1, | |
9399 | rtx callarg2, | |
9400 | rtx pop, bool sibcall) | |
9401 | { | |
9402 | rtx vec[3]; | |
9403 | rtx use = NULL, call; | |
9404 | unsigned int vec_len = 0; | |
9405 | tree fndecl; | |
9406 | ||
9407 | if (GET_CODE (XEXP (fnaddr, 0)) == SYMBOL_REF) | |
9408 | { | |
9409 | fndecl = SYMBOL_REF_DECL (XEXP (fnaddr, 0)); | |
9410 | if (fndecl | |
9411 | && (lookup_attribute ("interrupt", | |
9412 | TYPE_ATTRIBUTES (TREE_TYPE (fndecl))))) | |
a9c697b8 | 9413 | error ("interrupt service routine cannot be called directly"); |
2bf6d935 ML |
9414 | } |
9415 | else | |
9416 | fndecl = NULL_TREE; | |
9417 | ||
9418 | if (pop == const0_rtx) | |
9419 | pop = NULL; | |
9420 | gcc_assert (!TARGET_64BIT || !pop); | |
9421 | ||
41bd1b19 | 9422 | rtx addr = XEXP (fnaddr, 0); |
2bf6d935 ML |
9423 | if (TARGET_MACHO && !TARGET_64BIT) |
9424 | { | |
9425 | #if TARGET_MACHO | |
9426 | if (flag_pic && GET_CODE (XEXP (fnaddr, 0)) == SYMBOL_REF) | |
9427 | fnaddr = machopic_indirect_call_target (fnaddr); | |
9428 | #endif | |
9429 | } | |
9430 | else | |
9431 | { | |
9432 | /* Static functions and indirect calls don't need the pic register. Also, | |
9433 | check if PLT was explicitly avoided via no-plt or "noplt" attribute, making | |
9434 | it an indirect call. */ | |
2bf6d935 ML |
9435 | if (flag_pic |
9436 | && GET_CODE (addr) == SYMBOL_REF | |
f7854b90 | 9437 | && ix86_call_use_plt_p (addr)) |
2bf6d935 ML |
9438 | { |
9439 | if (flag_plt | |
9440 | && (SYMBOL_REF_DECL (addr) == NULL_TREE | |
9441 | || !lookup_attribute ("noplt", | |
9442 | DECL_ATTRIBUTES (SYMBOL_REF_DECL (addr))))) | |
9443 | { | |
9444 | if (!TARGET_64BIT | |
9445 | || (ix86_cmodel == CM_LARGE_PIC | |
9446 | && DEFAULT_ABI != MS_ABI)) | |
9447 | { | |
9448 | use_reg (&use, gen_rtx_REG (Pmode, | |
9449 | REAL_PIC_OFFSET_TABLE_REGNUM)); | |
9450 | if (ix86_use_pseudo_pic_reg ()) | |
9451 | emit_move_insn (gen_rtx_REG (Pmode, | |
9452 | REAL_PIC_OFFSET_TABLE_REGNUM), | |
9453 | pic_offset_table_rtx); | |
9454 | } | |
9455 | } | |
9456 | else if (!TARGET_PECOFF && !TARGET_MACHO) | |
9457 | { | |
69157fe7 JJ |
9458 | if (TARGET_64BIT |
9459 | && ix86_cmodel == CM_LARGE_PIC | |
9460 | && DEFAULT_ABI != MS_ABI) | |
9461 | { | |
9462 | fnaddr = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, addr), | |
9463 | UNSPEC_GOT); | |
9464 | fnaddr = gen_rtx_CONST (Pmode, fnaddr); | |
9465 | fnaddr = force_reg (Pmode, fnaddr); | |
9466 | fnaddr = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, fnaddr); | |
9467 | } | |
9468 | else if (TARGET_64BIT) | |
2bf6d935 ML |
9469 | { |
9470 | fnaddr = gen_rtx_UNSPEC (Pmode, | |
9471 | gen_rtvec (1, addr), | |
9472 | UNSPEC_GOTPCREL); | |
9473 | fnaddr = gen_rtx_CONST (Pmode, fnaddr); | |
9474 | } | |
9475 | else | |
9476 | { | |
9477 | fnaddr = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, addr), | |
9478 | UNSPEC_GOT); | |
9479 | fnaddr = gen_rtx_CONST (Pmode, fnaddr); | |
9480 | fnaddr = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, | |
9481 | fnaddr); | |
9482 | } | |
9483 | fnaddr = gen_const_mem (Pmode, fnaddr); | |
9484 | /* Pmode may not be the same as word_mode for x32, which | |
9485 | doesn't support indirect branch via 32-bit memory slot. | |
9486 | Since x32 GOT slot is 64 bit with zero upper 32 bits, | |
9487 | indirect branch via x32 GOT slot is OK. */ | |
9488 | if (GET_MODE (fnaddr) != word_mode) | |
9489 | fnaddr = gen_rtx_ZERO_EXTEND (word_mode, fnaddr); | |
9490 | fnaddr = gen_rtx_MEM (QImode, fnaddr); | |
9491 | } | |
9492 | } | |
9493 | } | |
9494 | ||
9495 | /* Skip setting up RAX register for -mskip-rax-setup when there are no | |
9496 | parameters passed in vector registers. */ | |
9497 | if (TARGET_64BIT | |
9498 | && (INTVAL (callarg2) > 0 | |
9499 | || (INTVAL (callarg2) == 0 | |
9500 | && (TARGET_SSE || !flag_skip_rax_setup)))) | |
9501 | { | |
9502 | rtx al = gen_rtx_REG (QImode, AX_REG); | |
9503 | emit_move_insn (al, callarg2); | |
9504 | use_reg (&use, al); | |
9505 | } | |
9506 | ||
9507 | if (ix86_cmodel == CM_LARGE_PIC | |
9508 | && !TARGET_PECOFF | |
9509 | && MEM_P (fnaddr) | |
9510 | && GET_CODE (XEXP (fnaddr, 0)) == SYMBOL_REF | |
9511 | && !local_symbolic_operand (XEXP (fnaddr, 0), VOIDmode)) | |
9512 | fnaddr = gen_rtx_MEM (QImode, construct_plt_address (XEXP (fnaddr, 0))); | |
9513 | /* Since x32 GOT slot is 64 bit with zero upper 32 bits, indirect | |
9514 | branch via x32 GOT slot is OK. */ | |
9515 | else if (!(TARGET_X32 | |
9516 | && MEM_P (fnaddr) | |
9517 | && GET_CODE (XEXP (fnaddr, 0)) == ZERO_EXTEND | |
9518 | && GOT_memory_operand (XEXP (XEXP (fnaddr, 0), 0), Pmode)) | |
9519 | && (sibcall | |
9520 | ? !sibcall_insn_operand (XEXP (fnaddr, 0), word_mode) | |
9521 | : !call_insn_operand (XEXP (fnaddr, 0), word_mode))) | |
9522 | { | |
9523 | fnaddr = convert_to_mode (word_mode, XEXP (fnaddr, 0), 1); | |
9524 | fnaddr = gen_rtx_MEM (QImode, copy_to_mode_reg (word_mode, fnaddr)); | |
9525 | } | |
9526 | ||
bb576017 | 9527 | /* PR100665: Hwasan may tag code pointer which is not supported by LAM, |
9528 | mask off code pointers here. | |
9529 | TODO: also need to handle indirect jump. */ | |
9530 | if (ix86_memtag_can_tag_addresses () && !fndecl | |
9531 | && sanitize_flags_p (SANITIZE_HWADDRESS)) | |
9532 | { | |
9533 | rtx untagged_addr = ix86_memtag_untagged_pointer (XEXP (fnaddr, 0), | |
9534 | NULL_RTX); | |
9535 | fnaddr = gen_rtx_MEM (QImode, untagged_addr); | |
9536 | } | |
9537 | ||
2bf6d935 ML |
9538 | call = gen_rtx_CALL (VOIDmode, fnaddr, callarg1); |
9539 | ||
9540 | if (retval) | |
9541 | call = gen_rtx_SET (retval, call); | |
9542 | vec[vec_len++] = call; | |
9543 | ||
9544 | if (pop) | |
9545 | { | |
9546 | pop = gen_rtx_PLUS (Pmode, stack_pointer_rtx, pop); | |
9547 | pop = gen_rtx_SET (stack_pointer_rtx, pop); | |
9548 | vec[vec_len++] = pop; | |
9549 | } | |
9550 | ||
9551 | if (cfun->machine->no_caller_saved_registers | |
9552 | && (!fndecl | |
9553 | || (!TREE_THIS_VOLATILE (fndecl) | |
9554 | && !lookup_attribute ("no_caller_saved_registers", | |
9555 | TYPE_ATTRIBUTES (TREE_TYPE (fndecl)))))) | |
9556 | { | |
9557 | static const char ix86_call_used_regs[] = CALL_USED_REGISTERS; | |
9558 | bool is_64bit_ms_abi = (TARGET_64BIT | |
9559 | && ix86_function_abi (fndecl) == MS_ABI); | |
9560 | char c_mask = CALL_USED_REGISTERS_MASK (is_64bit_ms_abi); | |
9561 | ||
9562 | /* If there are no caller-saved registers, add all registers | |
9563 | that are clobbered by the call which returns. */ | |
9564 | for (int i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
9565 | if (!fixed_regs[i] | |
9566 | && (ix86_call_used_regs[i] == 1 | |
9567 | || (ix86_call_used_regs[i] & c_mask)) | |
9568 | && !STACK_REGNO_P (i) | |
9569 | && !MMX_REGNO_P (i)) | |
9570 | clobber_reg (&use, | |
9571 | gen_rtx_REG (GET_MODE (regno_reg_rtx[i]), i)); | |
9572 | } | |
9573 | else if (TARGET_64BIT_MS_ABI | |
9574 | && (!callarg2 || INTVAL (callarg2) != -2)) | |
9575 | { | |
9576 | unsigned i; | |
9577 | ||
9578 | for (i = 0; i < NUM_X86_64_MS_CLOBBERED_REGS; i++) | |
9579 | { | |
9580 | int regno = x86_64_ms_sysv_extra_clobbered_registers[i]; | |
9581 | machine_mode mode = SSE_REGNO_P (regno) ? TImode : DImode; | |
9582 | ||
9583 | clobber_reg (&use, gen_rtx_REG (mode, regno)); | |
9584 | } | |
9585 | ||
9586 | /* Set here, but it may get cleared later. */ | |
9587 | if (TARGET_CALL_MS2SYSV_XLOGUES) | |
9588 | { | |
9589 | if (!TARGET_SSE) | |
9590 | ; | |
9591 | ||
9592 | /* Don't break hot-patched functions. */ | |
9593 | else if (ix86_function_ms_hook_prologue (current_function_decl)) | |
9594 | ; | |
9595 | ||
9596 | /* TODO: Cases not yet examined. */ | |
9597 | else if (flag_split_stack) | |
9598 | warn_once_call_ms2sysv_xlogues ("-fsplit-stack"); | |
9599 | ||
9600 | else | |
9601 | { | |
9602 | gcc_assert (!reload_completed); | |
9603 | cfun->machine->call_ms2sysv = true; | |
9604 | } | |
9605 | } | |
9606 | } | |
9607 | ||
41bd1b19 IS |
9608 | if (TARGET_MACHO && TARGET_64BIT && !sibcall |
9609 | && ((GET_CODE (addr) == SYMBOL_REF && !SYMBOL_REF_LOCAL_P (addr)) | |
9610 | || !fndecl || TREE_PUBLIC (fndecl))) | |
9611 | { | |
9612 | /* We allow public functions defined in a TU to bind locally for PIC | |
9613 | code (the default) on 64bit Mach-O. | |
9614 | If such functions are not inlined, we cannot tell at compile-time if | |
9615 | they will be called via the lazy symbol resolver (this can depend on | |
9616 | options given at link-time). Therefore, we must assume that the lazy | |
9617 | resolver could be used which clobbers R11 and R10. */ | |
9618 | clobber_reg (&use, gen_rtx_REG (DImode, R11_REG)); | |
9619 | clobber_reg (&use, gen_rtx_REG (DImode, R10_REG)); | |
9620 | } | |
9621 | ||
2bf6d935 ML |
9622 | if (vec_len > 1) |
9623 | call = gen_rtx_PARALLEL (VOIDmode, gen_rtvec_v (vec_len, vec)); | |
9624 | rtx_insn *call_insn = emit_call_insn (call); | |
9625 | if (use) | |
9626 | CALL_INSN_FUNCTION_USAGE (call_insn) = use; | |
9627 | ||
9628 | return call_insn; | |
9629 | } | |
9630 | ||
9631 | /* Split simple return with popping POPC bytes from stack to indirect | |
9632 | branch with stack adjustment . */ | |
9633 | ||
9634 | void | |
9635 | ix86_split_simple_return_pop_internal (rtx popc) | |
9636 | { | |
9637 | struct machine_function *m = cfun->machine; | |
9638 | rtx ecx = gen_rtx_REG (SImode, CX_REG); | |
9639 | rtx_insn *insn; | |
9640 | ||
9641 | /* There is no "pascal" calling convention in any 64bit ABI. */ | |
9642 | gcc_assert (!TARGET_64BIT); | |
9643 | ||
9644 | insn = emit_insn (gen_pop (ecx)); | |
9645 | m->fs.cfa_offset -= UNITS_PER_WORD; | |
9646 | m->fs.sp_offset -= UNITS_PER_WORD; | |
9647 | ||
9648 | rtx x = plus_constant (Pmode, stack_pointer_rtx, UNITS_PER_WORD); | |
9649 | x = gen_rtx_SET (stack_pointer_rtx, x); | |
9650 | add_reg_note (insn, REG_CFA_ADJUST_CFA, x); | |
9651 | add_reg_note (insn, REG_CFA_REGISTER, gen_rtx_SET (ecx, pc_rtx)); | |
9652 | RTX_FRAME_RELATED_P (insn) = 1; | |
9653 | ||
9654 | x = gen_rtx_PLUS (Pmode, stack_pointer_rtx, popc); | |
9655 | x = gen_rtx_SET (stack_pointer_rtx, x); | |
9656 | insn = emit_insn (x); | |
9657 | add_reg_note (insn, REG_CFA_ADJUST_CFA, x); | |
9658 | RTX_FRAME_RELATED_P (insn) = 1; | |
9659 | ||
9660 | /* Now return address is in ECX. */ | |
9661 | emit_jump_insn (gen_simple_return_indirect_internal (ecx)); | |
9662 | } | |
9663 | ||
9664 | /* Errors in the source file can cause expand_expr to return const0_rtx | |
9665 | where we expect a vector. To avoid crashing, use one of the vector | |
9666 | clear instructions. */ | |
9667 | ||
9668 | static rtx | |
9669 | safe_vector_operand (rtx x, machine_mode mode) | |
9670 | { | |
9671 | if (x == const0_rtx) | |
9672 | x = CONST0_RTX (mode); | |
9673 | return x; | |
9674 | } | |
9675 | ||
9676 | /* Subroutine of ix86_expand_builtin to take care of binop insns. */ | |
9677 | ||
9678 | static rtx | |
9679 | ix86_expand_binop_builtin (enum insn_code icode, tree exp, rtx target) | |
9680 | { | |
9681 | rtx pat; | |
9682 | tree arg0 = CALL_EXPR_ARG (exp, 0); | |
9683 | tree arg1 = CALL_EXPR_ARG (exp, 1); | |
9684 | rtx op0 = expand_normal (arg0); | |
9685 | rtx op1 = expand_normal (arg1); | |
9686 | machine_mode tmode = insn_data[icode].operand[0].mode; | |
9687 | machine_mode mode0 = insn_data[icode].operand[1].mode; | |
9688 | machine_mode mode1 = insn_data[icode].operand[2].mode; | |
9689 | ||
9690 | if (VECTOR_MODE_P (mode0)) | |
9691 | op0 = safe_vector_operand (op0, mode0); | |
9692 | if (VECTOR_MODE_P (mode1)) | |
9693 | op1 = safe_vector_operand (op1, mode1); | |
9694 | ||
9695 | if (optimize || !target | |
9696 | || GET_MODE (target) != tmode | |
9697 | || !insn_data[icode].operand[0].predicate (target, tmode)) | |
9698 | target = gen_reg_rtx (tmode); | |
9699 | ||
9700 | if (GET_MODE (op1) == SImode && mode1 == TImode) | |
9701 | { | |
9702 | rtx x = gen_reg_rtx (V4SImode); | |
9703 | emit_insn (gen_sse2_loadd (x, op1)); | |
9704 | op1 = gen_lowpart (TImode, x); | |
9705 | } | |
9706 | ||
9707 | if (!insn_data[icode].operand[1].predicate (op0, mode0)) | |
9708 | op0 = copy_to_mode_reg (mode0, op0); | |
9709 | if (!insn_data[icode].operand[2].predicate (op1, mode1)) | |
9710 | op1 = copy_to_mode_reg (mode1, op1); | |
9711 | ||
9712 | pat = GEN_FCN (icode) (target, op0, op1); | |
9713 | if (! pat) | |
9714 | return 0; | |
9715 | ||
9716 | emit_insn (pat); | |
9717 | ||
9718 | return target; | |
9719 | } | |
9720 | ||
9721 | /* Subroutine of ix86_expand_builtin to take care of 2-4 argument insns. */ | |
9722 | ||
9723 | static rtx | |
9724 | ix86_expand_multi_arg_builtin (enum insn_code icode, tree exp, rtx target, | |
9725 | enum ix86_builtin_func_type m_type, | |
9726 | enum rtx_code sub_code) | |
9727 | { | |
9728 | rtx pat; | |
715a8bc8 | 9729 | unsigned int i, nargs; |
2bf6d935 ML |
9730 | bool comparison_p = false; |
9731 | bool tf_p = false; | |
9732 | bool last_arg_constant = false; | |
9733 | int num_memory = 0; | |
715a8bc8 | 9734 | rtx xops[4]; |
2bf6d935 ML |
9735 | |
9736 | machine_mode tmode = insn_data[icode].operand[0].mode; | |
9737 | ||
9738 | switch (m_type) | |
9739 | { | |
9740 | case MULTI_ARG_4_DF2_DI_I: | |
9741 | case MULTI_ARG_4_DF2_DI_I1: | |
9742 | case MULTI_ARG_4_SF2_SI_I: | |
9743 | case MULTI_ARG_4_SF2_SI_I1: | |
9744 | nargs = 4; | |
9745 | last_arg_constant = true; | |
9746 | break; | |
9747 | ||
9748 | case MULTI_ARG_3_SF: | |
9749 | case MULTI_ARG_3_DF: | |
9750 | case MULTI_ARG_3_SF2: | |
9751 | case MULTI_ARG_3_DF2: | |
9752 | case MULTI_ARG_3_DI: | |
9753 | case MULTI_ARG_3_SI: | |
9754 | case MULTI_ARG_3_SI_DI: | |
9755 | case MULTI_ARG_3_HI: | |
9756 | case MULTI_ARG_3_HI_SI: | |
9757 | case MULTI_ARG_3_QI: | |
9758 | case MULTI_ARG_3_DI2: | |
9759 | case MULTI_ARG_3_SI2: | |
9760 | case MULTI_ARG_3_HI2: | |
9761 | case MULTI_ARG_3_QI2: | |
9762 | nargs = 3; | |
9763 | break; | |
9764 | ||
9765 | case MULTI_ARG_2_SF: | |
9766 | case MULTI_ARG_2_DF: | |
9767 | case MULTI_ARG_2_DI: | |
9768 | case MULTI_ARG_2_SI: | |
9769 | case MULTI_ARG_2_HI: | |
9770 | case MULTI_ARG_2_QI: | |
9771 | nargs = 2; | |
9772 | break; | |
9773 | ||
9774 | case MULTI_ARG_2_DI_IMM: | |
9775 | case MULTI_ARG_2_SI_IMM: | |
9776 | case MULTI_ARG_2_HI_IMM: | |
9777 | case MULTI_ARG_2_QI_IMM: | |
9778 | nargs = 2; | |
9779 | last_arg_constant = true; | |
9780 | break; | |
9781 | ||
9782 | case MULTI_ARG_1_SF: | |
9783 | case MULTI_ARG_1_DF: | |
9784 | case MULTI_ARG_1_SF2: | |
9785 | case MULTI_ARG_1_DF2: | |
9786 | case MULTI_ARG_1_DI: | |
9787 | case MULTI_ARG_1_SI: | |
9788 | case MULTI_ARG_1_HI: | |
9789 | case MULTI_ARG_1_QI: | |
9790 | case MULTI_ARG_1_SI_DI: | |
9791 | case MULTI_ARG_1_HI_DI: | |
9792 | case MULTI_ARG_1_HI_SI: | |
9793 | case MULTI_ARG_1_QI_DI: | |
9794 | case MULTI_ARG_1_QI_SI: | |
9795 | case MULTI_ARG_1_QI_HI: | |
9796 | nargs = 1; | |
9797 | break; | |
9798 | ||
9799 | case MULTI_ARG_2_DI_CMP: | |
9800 | case MULTI_ARG_2_SI_CMP: | |
9801 | case MULTI_ARG_2_HI_CMP: | |
9802 | case MULTI_ARG_2_QI_CMP: | |
9803 | nargs = 2; | |
9804 | comparison_p = true; | |
9805 | break; | |
9806 | ||
9807 | case MULTI_ARG_2_SF_TF: | |
9808 | case MULTI_ARG_2_DF_TF: | |
9809 | case MULTI_ARG_2_DI_TF: | |
9810 | case MULTI_ARG_2_SI_TF: | |
9811 | case MULTI_ARG_2_HI_TF: | |
9812 | case MULTI_ARG_2_QI_TF: | |
9813 | nargs = 2; | |
9814 | tf_p = true; | |
9815 | break; | |
9816 | ||
9817 | default: | |
9818 | gcc_unreachable (); | |
9819 | } | |
9820 | ||
9821 | if (optimize || !target | |
9822 | || GET_MODE (target) != tmode | |
9823 | || !insn_data[icode].operand[0].predicate (target, tmode)) | |
9824 | target = gen_reg_rtx (tmode); | |
9825 | else if (memory_operand (target, tmode)) | |
9826 | num_memory++; | |
9827 | ||
715a8bc8 | 9828 | gcc_assert (nargs <= ARRAY_SIZE (xops)); |
2bf6d935 ML |
9829 | |
9830 | for (i = 0; i < nargs; i++) | |
9831 | { | |
9832 | tree arg = CALL_EXPR_ARG (exp, i); | |
9833 | rtx op = expand_normal (arg); | |
9834 | int adjust = (comparison_p) ? 1 : 0; | |
9835 | machine_mode mode = insn_data[icode].operand[i+adjust+1].mode; | |
9836 | ||
9837 | if (last_arg_constant && i == nargs - 1) | |
9838 | { | |
9839 | if (!insn_data[icode].operand[i + 1].predicate (op, mode)) | |
9840 | { | |
9841 | enum insn_code new_icode = icode; | |
9842 | switch (icode) | |
9843 | { | |
9844 | case CODE_FOR_xop_vpermil2v2df3: | |
9845 | case CODE_FOR_xop_vpermil2v4sf3: | |
9846 | case CODE_FOR_xop_vpermil2v4df3: | |
9847 | case CODE_FOR_xop_vpermil2v8sf3: | |
9848 | error ("the last argument must be a 2-bit immediate"); | |
9849 | return gen_reg_rtx (tmode); | |
9850 | case CODE_FOR_xop_rotlv2di3: | |
9851 | new_icode = CODE_FOR_rotlv2di3; | |
9852 | goto xop_rotl; | |
9853 | case CODE_FOR_xop_rotlv4si3: | |
9854 | new_icode = CODE_FOR_rotlv4si3; | |
9855 | goto xop_rotl; | |
9856 | case CODE_FOR_xop_rotlv8hi3: | |
9857 | new_icode = CODE_FOR_rotlv8hi3; | |
9858 | goto xop_rotl; | |
9859 | case CODE_FOR_xop_rotlv16qi3: | |
9860 | new_icode = CODE_FOR_rotlv16qi3; | |
9861 | xop_rotl: | |
9862 | if (CONST_INT_P (op)) | |
9863 | { | |
9864 | int mask = GET_MODE_UNIT_BITSIZE (tmode) - 1; | |
9865 | op = GEN_INT (INTVAL (op) & mask); | |
9866 | gcc_checking_assert | |
9867 | (insn_data[icode].operand[i + 1].predicate (op, mode)); | |
9868 | } | |
9869 | else | |
9870 | { | |
9871 | gcc_checking_assert | |
9872 | (nargs == 2 | |
9873 | && insn_data[new_icode].operand[0].mode == tmode | |
9874 | && insn_data[new_icode].operand[1].mode == tmode | |
9875 | && insn_data[new_icode].operand[2].mode == mode | |
9876 | && insn_data[new_icode].operand[0].predicate | |
9877 | == insn_data[icode].operand[0].predicate | |
9878 | && insn_data[new_icode].operand[1].predicate | |
9879 | == insn_data[icode].operand[1].predicate); | |
9880 | icode = new_icode; | |
9881 | goto non_constant; | |
9882 | } | |
9883 | break; | |
9884 | default: | |
9885 | gcc_unreachable (); | |
9886 | } | |
9887 | } | |
9888 | } | |
9889 | else | |
9890 | { | |
9891 | non_constant: | |
9892 | if (VECTOR_MODE_P (mode)) | |
9893 | op = safe_vector_operand (op, mode); | |
9894 | ||
9895 | /* If we aren't optimizing, only allow one memory operand to be | |
9896 | generated. */ | |
9897 | if (memory_operand (op, mode)) | |
9898 | num_memory++; | |
9899 | ||
9900 | gcc_assert (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode); | |
9901 | ||
9902 | if (optimize | |
9903 | || !insn_data[icode].operand[i+adjust+1].predicate (op, mode) | |
9904 | || num_memory > 1) | |
9905 | op = force_reg (mode, op); | |
9906 | } | |
9907 | ||
715a8bc8 | 9908 | xops[i] = op; |
2bf6d935 ML |
9909 | } |
9910 | ||
9911 | switch (nargs) | |
9912 | { | |
9913 | case 1: | |
715a8bc8 | 9914 | pat = GEN_FCN (icode) (target, xops[0]); |
2bf6d935 ML |
9915 | break; |
9916 | ||
9917 | case 2: | |
9918 | if (tf_p) | |
715a8bc8 | 9919 | pat = GEN_FCN (icode) (target, xops[0], xops[1], |
2bf6d935 ML |
9920 | GEN_INT ((int)sub_code)); |
9921 | else if (! comparison_p) | |
715a8bc8 | 9922 | pat = GEN_FCN (icode) (target, xops[0], xops[1]); |
2bf6d935 ML |
9923 | else |
9924 | { | |
9925 | rtx cmp_op = gen_rtx_fmt_ee (sub_code, GET_MODE (target), | |
715a8bc8 | 9926 | xops[0], xops[1]); |
2bf6d935 | 9927 | |
715a8bc8 | 9928 | pat = GEN_FCN (icode) (target, cmp_op, xops[0], xops[1]); |
2bf6d935 ML |
9929 | } |
9930 | break; | |
9931 | ||
9932 | case 3: | |
715a8bc8 | 9933 | pat = GEN_FCN (icode) (target, xops[0], xops[1], xops[2]); |
2bf6d935 ML |
9934 | break; |
9935 | ||
9936 | case 4: | |
715a8bc8 | 9937 | pat = GEN_FCN (icode) (target, xops[0], xops[1], xops[2], xops[3]); |
2bf6d935 ML |
9938 | break; |
9939 | ||
9940 | default: | |
9941 | gcc_unreachable (); | |
9942 | } | |
9943 | ||
9944 | if (! pat) | |
9945 | return 0; | |
9946 | ||
9947 | emit_insn (pat); | |
9948 | return target; | |
9949 | } | |
9950 | ||
9951 | /* Subroutine of ix86_expand_args_builtin to take care of scalar unop | |
9952 | insns with vec_merge. */ | |
9953 | ||
9954 | static rtx | |
9955 | ix86_expand_unop_vec_merge_builtin (enum insn_code icode, tree exp, | |
9956 | rtx target) | |
9957 | { | |
9958 | rtx pat; | |
9959 | tree arg0 = CALL_EXPR_ARG (exp, 0); | |
9960 | rtx op1, op0 = expand_normal (arg0); | |
9961 | machine_mode tmode = insn_data[icode].operand[0].mode; | |
9962 | machine_mode mode0 = insn_data[icode].operand[1].mode; | |
9963 | ||
9964 | if (optimize || !target | |
9965 | || GET_MODE (target) != tmode | |
9966 | || !insn_data[icode].operand[0].predicate (target, tmode)) | |
9967 | target = gen_reg_rtx (tmode); | |
9968 | ||
9969 | if (VECTOR_MODE_P (mode0)) | |
9970 | op0 = safe_vector_operand (op0, mode0); | |
9971 | ||
9972 | if ((optimize && !register_operand (op0, mode0)) | |
9973 | || !insn_data[icode].operand[1].predicate (op0, mode0)) | |
9974 | op0 = copy_to_mode_reg (mode0, op0); | |
9975 | ||
9976 | op1 = op0; | |
9977 | if (!insn_data[icode].operand[2].predicate (op1, mode0)) | |
9978 | op1 = copy_to_mode_reg (mode0, op1); | |
9979 | ||
9980 | pat = GEN_FCN (icode) (target, op0, op1); | |
9981 | if (! pat) | |
9982 | return 0; | |
9983 | emit_insn (pat); | |
9984 | return target; | |
9985 | } | |
9986 | ||
9987 | /* Subroutine of ix86_expand_builtin to take care of comparison insns. */ | |
9988 | ||
9989 | static rtx | |
9990 | ix86_expand_sse_compare (const struct builtin_description *d, | |
9991 | tree exp, rtx target, bool swap) | |
9992 | { | |
9993 | rtx pat; | |
9994 | tree arg0 = CALL_EXPR_ARG (exp, 0); | |
9995 | tree arg1 = CALL_EXPR_ARG (exp, 1); | |
9996 | rtx op0 = expand_normal (arg0); | |
9997 | rtx op1 = expand_normal (arg1); | |
9998 | rtx op2; | |
9999 | machine_mode tmode = insn_data[d->icode].operand[0].mode; | |
10000 | machine_mode mode0 = insn_data[d->icode].operand[1].mode; | |
10001 | machine_mode mode1 = insn_data[d->icode].operand[2].mode; | |
10002 | enum rtx_code comparison = d->comparison; | |
10003 | ||
10004 | if (VECTOR_MODE_P (mode0)) | |
10005 | op0 = safe_vector_operand (op0, mode0); | |
10006 | if (VECTOR_MODE_P (mode1)) | |
10007 | op1 = safe_vector_operand (op1, mode1); | |
10008 | ||
10009 | /* Swap operands if we have a comparison that isn't available in | |
10010 | hardware. */ | |
10011 | if (swap) | |
10012 | std::swap (op0, op1); | |
10013 | ||
10014 | if (optimize || !target | |
10015 | || GET_MODE (target) != tmode | |
10016 | || !insn_data[d->icode].operand[0].predicate (target, tmode)) | |
10017 | target = gen_reg_rtx (tmode); | |
10018 | ||
10019 | if ((optimize && !register_operand (op0, mode0)) | |
10020 | || !insn_data[d->icode].operand[1].predicate (op0, mode0)) | |
10021 | op0 = copy_to_mode_reg (mode0, op0); | |
10022 | if ((optimize && !register_operand (op1, mode1)) | |
10023 | || !insn_data[d->icode].operand[2].predicate (op1, mode1)) | |
10024 | op1 = copy_to_mode_reg (mode1, op1); | |
10025 | ||
10026 | op2 = gen_rtx_fmt_ee (comparison, mode0, op0, op1); | |
10027 | pat = GEN_FCN (d->icode) (target, op0, op1, op2); | |
10028 | if (! pat) | |
10029 | return 0; | |
10030 | emit_insn (pat); | |
10031 | return target; | |
10032 | } | |
10033 | ||
ae69e6f6 | 10034 | /* Subroutine of ix86_sse_comi and ix86_sse_comi_round to take care of |
10035 | * ordered EQ or unordered NE, generate PF jump. */ | |
10036 | ||
10037 | static rtx | |
10038 | ix86_ssecom_setcc (const enum rtx_code comparison, | |
10039 | bool check_unordered, machine_mode mode, | |
10040 | rtx set_dst, rtx target) | |
10041 | { | |
10042 | ||
10043 | rtx_code_label *label = NULL; | |
10044 | ||
10045 | /* NB: For ordered EQ or unordered NE, check ZF alone isn't sufficient | |
10046 | with NAN operands. */ | |
10047 | if (check_unordered) | |
10048 | { | |
10049 | gcc_assert (comparison == EQ || comparison == NE); | |
10050 | ||
10051 | rtx flag = gen_rtx_REG (CCFPmode, FLAGS_REG); | |
10052 | label = gen_label_rtx (); | |
10053 | rtx tmp = gen_rtx_fmt_ee (UNORDERED, VOIDmode, flag, const0_rtx); | |
10054 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp, | |
10055 | gen_rtx_LABEL_REF (VOIDmode, label), | |
10056 | pc_rtx); | |
10057 | emit_jump_insn (gen_rtx_SET (pc_rtx, tmp)); | |
10058 | } | |
10059 | ||
10060 | /* NB: Set CCFPmode and check a different CCmode which is in subset | |
10061 | of CCFPmode. */ | |
10062 | if (GET_MODE (set_dst) != mode) | |
10063 | { | |
10064 | gcc_assert (mode == CCAmode || mode == CCCmode | |
10065 | || mode == CCOmode || mode == CCPmode | |
10066 | || mode == CCSmode || mode == CCZmode); | |
10067 | set_dst = gen_rtx_REG (mode, FLAGS_REG); | |
10068 | } | |
10069 | ||
10070 | emit_insn (gen_rtx_SET (gen_rtx_STRICT_LOW_PART (VOIDmode, target), | |
10071 | gen_rtx_fmt_ee (comparison, QImode, | |
10072 | set_dst, | |
10073 | const0_rtx))); | |
10074 | ||
10075 | if (label) | |
10076 | emit_label (label); | |
10077 | ||
10078 | return SUBREG_REG (target); | |
10079 | } | |
10080 | ||
2bf6d935 ML |
10081 | /* Subroutine of ix86_expand_builtin to take care of comi insns. */ |
10082 | ||
10083 | static rtx | |
10084 | ix86_expand_sse_comi (const struct builtin_description *d, tree exp, | |
10085 | rtx target) | |
10086 | { | |
ae69e6f6 | 10087 | rtx pat, set_dst; |
2bf6d935 ML |
10088 | tree arg0 = CALL_EXPR_ARG (exp, 0); |
10089 | tree arg1 = CALL_EXPR_ARG (exp, 1); | |
10090 | rtx op0 = expand_normal (arg0); | |
10091 | rtx op1 = expand_normal (arg1); | |
ae69e6f6 | 10092 | enum insn_code icode = d->icode; |
10093 | const struct insn_data_d *insn_p = &insn_data[icode]; | |
10094 | machine_mode mode0 = insn_p->operand[0].mode; | |
10095 | machine_mode mode1 = insn_p->operand[1].mode; | |
2bf6d935 ML |
10096 | |
10097 | if (VECTOR_MODE_P (mode0)) | |
10098 | op0 = safe_vector_operand (op0, mode0); | |
10099 | if (VECTOR_MODE_P (mode1)) | |
10100 | op1 = safe_vector_operand (op1, mode1); | |
10101 | ||
ae69e6f6 | 10102 | enum rtx_code comparison = d->comparison; |
10103 | rtx const_val = const0_rtx; | |
10104 | ||
10105 | bool check_unordered = false; | |
10106 | machine_mode mode = CCFPmode; | |
10107 | switch (comparison) | |
10108 | { | |
10109 | case LE: /* -> GE */ | |
10110 | case LT: /* -> GT */ | |
10111 | std::swap (op0, op1); | |
10112 | comparison = swap_condition (comparison); | |
10113 | /* FALLTHRU */ | |
10114 | case GT: | |
10115 | case GE: | |
10116 | break; | |
10117 | case EQ: | |
10118 | check_unordered = true; | |
10119 | mode = CCZmode; | |
10120 | break; | |
10121 | case NE: | |
10122 | check_unordered = true; | |
10123 | mode = CCZmode; | |
10124 | const_val = const1_rtx; | |
10125 | break; | |
10126 | default: | |
10127 | gcc_unreachable (); | |
10128 | } | |
10129 | ||
2bf6d935 | 10130 | target = gen_reg_rtx (SImode); |
ae69e6f6 | 10131 | emit_move_insn (target, const_val); |
2bf6d935 ML |
10132 | target = gen_rtx_SUBREG (QImode, target, 0); |
10133 | ||
10134 | if ((optimize && !register_operand (op0, mode0)) | |
ae69e6f6 | 10135 | || !insn_p->operand[0].predicate (op0, mode0)) |
2bf6d935 ML |
10136 | op0 = copy_to_mode_reg (mode0, op0); |
10137 | if ((optimize && !register_operand (op1, mode1)) | |
ae69e6f6 | 10138 | || !insn_p->operand[1].predicate (op1, mode1)) |
2bf6d935 ML |
10139 | op1 = copy_to_mode_reg (mode1, op1); |
10140 | ||
ae69e6f6 | 10141 | pat = GEN_FCN (icode) (op0, op1); |
2bf6d935 ML |
10142 | if (! pat) |
10143 | return 0; | |
2bf6d935 | 10144 | |
ae69e6f6 | 10145 | set_dst = SET_DEST (pat); |
10146 | emit_insn (pat); | |
10147 | return ix86_ssecom_setcc (comparison, check_unordered, mode, | |
10148 | set_dst, target); | |
2bf6d935 ML |
10149 | } |
10150 | ||
10151 | /* Subroutines of ix86_expand_args_builtin to take care of round insns. */ | |
10152 | ||
10153 | static rtx | |
10154 | ix86_expand_sse_round (const struct builtin_description *d, tree exp, | |
10155 | rtx target) | |
10156 | { | |
10157 | rtx pat; | |
10158 | tree arg0 = CALL_EXPR_ARG (exp, 0); | |
10159 | rtx op1, op0 = expand_normal (arg0); | |
10160 | machine_mode tmode = insn_data[d->icode].operand[0].mode; | |
10161 | machine_mode mode0 = insn_data[d->icode].operand[1].mode; | |
10162 | ||
10163 | if (optimize || target == 0 | |
10164 | || GET_MODE (target) != tmode | |
10165 | || !insn_data[d->icode].operand[0].predicate (target, tmode)) | |
10166 | target = gen_reg_rtx (tmode); | |
10167 | ||
10168 | if (VECTOR_MODE_P (mode0)) | |
10169 | op0 = safe_vector_operand (op0, mode0); | |
10170 | ||
10171 | if ((optimize && !register_operand (op0, mode0)) | |
10172 | || !insn_data[d->icode].operand[0].predicate (op0, mode0)) | |
10173 | op0 = copy_to_mode_reg (mode0, op0); | |
10174 | ||
10175 | op1 = GEN_INT (d->comparison); | |
10176 | ||
10177 | pat = GEN_FCN (d->icode) (target, op0, op1); | |
10178 | if (! pat) | |
10179 | return 0; | |
10180 | emit_insn (pat); | |
10181 | return target; | |
10182 | } | |
10183 | ||
10184 | static rtx | |
10185 | ix86_expand_sse_round_vec_pack_sfix (const struct builtin_description *d, | |
10186 | tree exp, rtx target) | |
10187 | { | |
10188 | rtx pat; | |
10189 | tree arg0 = CALL_EXPR_ARG (exp, 0); | |
10190 | tree arg1 = CALL_EXPR_ARG (exp, 1); | |
10191 | rtx op0 = expand_normal (arg0); | |
10192 | rtx op1 = expand_normal (arg1); | |
10193 | rtx op2; | |
10194 | machine_mode tmode = insn_data[d->icode].operand[0].mode; | |
10195 | machine_mode mode0 = insn_data[d->icode].operand[1].mode; | |
10196 | machine_mode mode1 = insn_data[d->icode].operand[2].mode; | |
10197 | ||
10198 | if (optimize || target == 0 | |
10199 | || GET_MODE (target) != tmode | |
10200 | || !insn_data[d->icode].operand[0].predicate (target, tmode)) | |
10201 | target = gen_reg_rtx (tmode); | |
10202 | ||
10203 | op0 = safe_vector_operand (op0, mode0); | |
10204 | op1 = safe_vector_operand (op1, mode1); | |
10205 | ||
10206 | if ((optimize && !register_operand (op0, mode0)) | |
10207 | || !insn_data[d->icode].operand[0].predicate (op0, mode0)) | |
10208 | op0 = copy_to_mode_reg (mode0, op0); | |
10209 | if ((optimize && !register_operand (op1, mode1)) | |
10210 | || !insn_data[d->icode].operand[1].predicate (op1, mode1)) | |
10211 | op1 = copy_to_mode_reg (mode1, op1); | |
10212 | ||
10213 | op2 = GEN_INT (d->comparison); | |
10214 | ||
10215 | pat = GEN_FCN (d->icode) (target, op0, op1, op2); | |
10216 | if (! pat) | |
10217 | return 0; | |
10218 | emit_insn (pat); | |
10219 | return target; | |
10220 | } | |
10221 | ||
10222 | /* Subroutine of ix86_expand_builtin to take care of ptest insns. */ | |
10223 | ||
10224 | static rtx | |
10225 | ix86_expand_sse_ptest (const struct builtin_description *d, tree exp, | |
10226 | rtx target) | |
10227 | { | |
10228 | rtx pat; | |
10229 | tree arg0 = CALL_EXPR_ARG (exp, 0); | |
10230 | tree arg1 = CALL_EXPR_ARG (exp, 1); | |
10231 | rtx op0 = expand_normal (arg0); | |
10232 | rtx op1 = expand_normal (arg1); | |
10233 | machine_mode mode0 = insn_data[d->icode].operand[0].mode; | |
10234 | machine_mode mode1 = insn_data[d->icode].operand[1].mode; | |
10235 | enum rtx_code comparison = d->comparison; | |
10236 | ||
5322f009 RS |
10237 | /* ptest reg, reg sets the carry flag. */ |
10238 | if (comparison == LTU | |
10239 | && (d->code == IX86_BUILTIN_PTESTC | |
10240 | || d->code == IX86_BUILTIN_PTESTC256) | |
10241 | && rtx_equal_p (op0, op1)) | |
10242 | { | |
10243 | if (!target) | |
10244 | target = gen_reg_rtx (SImode); | |
10245 | emit_move_insn (target, const1_rtx); | |
10246 | return target; | |
10247 | } | |
10248 | ||
2bf6d935 ML |
10249 | if (VECTOR_MODE_P (mode0)) |
10250 | op0 = safe_vector_operand (op0, mode0); | |
10251 | if (VECTOR_MODE_P (mode1)) | |
10252 | op1 = safe_vector_operand (op1, mode1); | |
10253 | ||
10254 | target = gen_reg_rtx (SImode); | |
10255 | emit_move_insn (target, const0_rtx); | |
10256 | target = gen_rtx_SUBREG (QImode, target, 0); | |
10257 | ||
10258 | if ((optimize && !register_operand (op0, mode0)) | |
10259 | || !insn_data[d->icode].operand[0].predicate (op0, mode0)) | |
10260 | op0 = copy_to_mode_reg (mode0, op0); | |
10261 | if ((optimize && !register_operand (op1, mode1)) | |
10262 | || !insn_data[d->icode].operand[1].predicate (op1, mode1)) | |
10263 | op1 = copy_to_mode_reg (mode1, op1); | |
10264 | ||
10265 | pat = GEN_FCN (d->icode) (op0, op1); | |
10266 | if (! pat) | |
10267 | return 0; | |
10268 | emit_insn (pat); | |
10269 | emit_insn (gen_rtx_SET (gen_rtx_STRICT_LOW_PART (VOIDmode, target), | |
10270 | gen_rtx_fmt_ee (comparison, QImode, | |
10271 | SET_DEST (pat), | |
10272 | const0_rtx))); | |
10273 | ||
10274 | return SUBREG_REG (target); | |
10275 | } | |
10276 | ||
10277 | /* Subroutine of ix86_expand_builtin to take care of pcmpestr[im] insns. */ | |
10278 | ||
10279 | static rtx | |
10280 | ix86_expand_sse_pcmpestr (const struct builtin_description *d, | |
10281 | tree exp, rtx target) | |
10282 | { | |
10283 | rtx pat; | |
10284 | tree arg0 = CALL_EXPR_ARG (exp, 0); | |
10285 | tree arg1 = CALL_EXPR_ARG (exp, 1); | |
10286 | tree arg2 = CALL_EXPR_ARG (exp, 2); | |
10287 | tree arg3 = CALL_EXPR_ARG (exp, 3); | |
10288 | tree arg4 = CALL_EXPR_ARG (exp, 4); | |
10289 | rtx scratch0, scratch1; | |
10290 | rtx op0 = expand_normal (arg0); | |
10291 | rtx op1 = expand_normal (arg1); | |
10292 | rtx op2 = expand_normal (arg2); | |
10293 | rtx op3 = expand_normal (arg3); | |
10294 | rtx op4 = expand_normal (arg4); | |
10295 | machine_mode tmode0, tmode1, modev2, modei3, modev4, modei5, modeimm; | |
10296 | ||
10297 | tmode0 = insn_data[d->icode].operand[0].mode; | |
10298 | tmode1 = insn_data[d->icode].operand[1].mode; | |
10299 | modev2 = insn_data[d->icode].operand[2].mode; | |
10300 | modei3 = insn_data[d->icode].operand[3].mode; | |
10301 | modev4 = insn_data[d->icode].operand[4].mode; | |
10302 | modei5 = insn_data[d->icode].operand[5].mode; | |
10303 | modeimm = insn_data[d->icode].operand[6].mode; | |
10304 | ||
10305 | if (VECTOR_MODE_P (modev2)) | |
10306 | op0 = safe_vector_operand (op0, modev2); | |
10307 | if (VECTOR_MODE_P (modev4)) | |
10308 | op2 = safe_vector_operand (op2, modev4); | |
10309 | ||
10310 | if (!insn_data[d->icode].operand[2].predicate (op0, modev2)) | |
10311 | op0 = copy_to_mode_reg (modev2, op0); | |
10312 | if (!insn_data[d->icode].operand[3].predicate (op1, modei3)) | |
10313 | op1 = copy_to_mode_reg (modei3, op1); | |
10314 | if ((optimize && !register_operand (op2, modev4)) | |
10315 | || !insn_data[d->icode].operand[4].predicate (op2, modev4)) | |
10316 | op2 = copy_to_mode_reg (modev4, op2); | |
10317 | if (!insn_data[d->icode].operand[5].predicate (op3, modei5)) | |
10318 | op3 = copy_to_mode_reg (modei5, op3); | |
10319 | ||
10320 | if (!insn_data[d->icode].operand[6].predicate (op4, modeimm)) | |
10321 | { | |
10322 | error ("the fifth argument must be an 8-bit immediate"); | |
10323 | return const0_rtx; | |
10324 | } | |
10325 | ||
10326 | if (d->code == IX86_BUILTIN_PCMPESTRI128) | |
10327 | { | |
10328 | if (optimize || !target | |
10329 | || GET_MODE (target) != tmode0 | |
10330 | || !insn_data[d->icode].operand[0].predicate (target, tmode0)) | |
10331 | target = gen_reg_rtx (tmode0); | |
10332 | ||
10333 | scratch1 = gen_reg_rtx (tmode1); | |
10334 | ||
10335 | pat = GEN_FCN (d->icode) (target, scratch1, op0, op1, op2, op3, op4); | |
10336 | } | |
10337 | else if (d->code == IX86_BUILTIN_PCMPESTRM128) | |
10338 | { | |
10339 | if (optimize || !target | |
10340 | || GET_MODE (target) != tmode1 | |
10341 | || !insn_data[d->icode].operand[1].predicate (target, tmode1)) | |
10342 | target = gen_reg_rtx (tmode1); | |
10343 | ||
10344 | scratch0 = gen_reg_rtx (tmode0); | |
10345 | ||
10346 | pat = GEN_FCN (d->icode) (scratch0, target, op0, op1, op2, op3, op4); | |
10347 | } | |
10348 | else | |
10349 | { | |
10350 | gcc_assert (d->flag); | |
10351 | ||
10352 | scratch0 = gen_reg_rtx (tmode0); | |
10353 | scratch1 = gen_reg_rtx (tmode1); | |
10354 | ||
10355 | pat = GEN_FCN (d->icode) (scratch0, scratch1, op0, op1, op2, op3, op4); | |
10356 | } | |
10357 | ||
10358 | if (! pat) | |
10359 | return 0; | |
10360 | ||
10361 | emit_insn (pat); | |
10362 | ||
10363 | if (d->flag) | |
10364 | { | |
10365 | target = gen_reg_rtx (SImode); | |
10366 | emit_move_insn (target, const0_rtx); | |
10367 | target = gen_rtx_SUBREG (QImode, target, 0); | |
10368 | ||
10369 | emit_insn | |
10370 | (gen_rtx_SET (gen_rtx_STRICT_LOW_PART (VOIDmode, target), | |
10371 | gen_rtx_fmt_ee (EQ, QImode, | |
10372 | gen_rtx_REG ((machine_mode) d->flag, | |
10373 | FLAGS_REG), | |
10374 | const0_rtx))); | |
10375 | return SUBREG_REG (target); | |
10376 | } | |
10377 | else | |
10378 | return target; | |
10379 | } | |
10380 | ||
10381 | ||
10382 | /* Subroutine of ix86_expand_builtin to take care of pcmpistr[im] insns. */ | |
10383 | ||
10384 | static rtx | |
10385 | ix86_expand_sse_pcmpistr (const struct builtin_description *d, | |
10386 | tree exp, rtx target) | |
10387 | { | |
10388 | rtx pat; | |
10389 | tree arg0 = CALL_EXPR_ARG (exp, 0); | |
10390 | tree arg1 = CALL_EXPR_ARG (exp, 1); | |
10391 | tree arg2 = CALL_EXPR_ARG (exp, 2); | |
10392 | rtx scratch0, scratch1; | |
10393 | rtx op0 = expand_normal (arg0); | |
10394 | rtx op1 = expand_normal (arg1); | |
10395 | rtx op2 = expand_normal (arg2); | |
10396 | machine_mode tmode0, tmode1, modev2, modev3, modeimm; | |
10397 | ||
10398 | tmode0 = insn_data[d->icode].operand[0].mode; | |
10399 | tmode1 = insn_data[d->icode].operand[1].mode; | |
10400 | modev2 = insn_data[d->icode].operand[2].mode; | |
10401 | modev3 = insn_data[d->icode].operand[3].mode; | |
10402 | modeimm = insn_data[d->icode].operand[4].mode; | |
10403 | ||
10404 | if (VECTOR_MODE_P (modev2)) | |
10405 | op0 = safe_vector_operand (op0, modev2); | |
10406 | if (VECTOR_MODE_P (modev3)) | |
10407 | op1 = safe_vector_operand (op1, modev3); | |
10408 | ||
10409 | if (!insn_data[d->icode].operand[2].predicate (op0, modev2)) | |
10410 | op0 = copy_to_mode_reg (modev2, op0); | |
10411 | if ((optimize && !register_operand (op1, modev3)) | |
10412 | || !insn_data[d->icode].operand[3].predicate (op1, modev3)) | |
10413 | op1 = copy_to_mode_reg (modev3, op1); | |
10414 | ||
10415 | if (!insn_data[d->icode].operand[4].predicate (op2, modeimm)) | |
10416 | { | |
10417 | error ("the third argument must be an 8-bit immediate"); | |
10418 | return const0_rtx; | |
10419 | } | |
10420 | ||
10421 | if (d->code == IX86_BUILTIN_PCMPISTRI128) | |
10422 | { | |
10423 | if (optimize || !target | |
10424 | || GET_MODE (target) != tmode0 | |
10425 | || !insn_data[d->icode].operand[0].predicate (target, tmode0)) | |
10426 | target = gen_reg_rtx (tmode0); | |
10427 | ||
10428 | scratch1 = gen_reg_rtx (tmode1); | |
10429 | ||
10430 | pat = GEN_FCN (d->icode) (target, scratch1, op0, op1, op2); | |
10431 | } | |
10432 | else if (d->code == IX86_BUILTIN_PCMPISTRM128) | |
10433 | { | |
10434 | if (optimize || !target | |
10435 | || GET_MODE (target) != tmode1 | |
10436 | || !insn_data[d->icode].operand[1].predicate (target, tmode1)) | |
10437 | target = gen_reg_rtx (tmode1); | |
10438 | ||
10439 | scratch0 = gen_reg_rtx (tmode0); | |
10440 | ||
10441 | pat = GEN_FCN (d->icode) (scratch0, target, op0, op1, op2); | |
10442 | } | |
10443 | else | |
10444 | { | |
10445 | gcc_assert (d->flag); | |
10446 | ||
10447 | scratch0 = gen_reg_rtx (tmode0); | |
10448 | scratch1 = gen_reg_rtx (tmode1); | |
10449 | ||
10450 | pat = GEN_FCN (d->icode) (scratch0, scratch1, op0, op1, op2); | |
10451 | } | |
10452 | ||
10453 | if (! pat) | |
10454 | return 0; | |
10455 | ||
10456 | emit_insn (pat); | |
10457 | ||
10458 | if (d->flag) | |
10459 | { | |
10460 | target = gen_reg_rtx (SImode); | |
10461 | emit_move_insn (target, const0_rtx); | |
10462 | target = gen_rtx_SUBREG (QImode, target, 0); | |
10463 | ||
10464 | emit_insn | |
10465 | (gen_rtx_SET (gen_rtx_STRICT_LOW_PART (VOIDmode, target), | |
10466 | gen_rtx_fmt_ee (EQ, QImode, | |
10467 | gen_rtx_REG ((machine_mode) d->flag, | |
10468 | FLAGS_REG), | |
10469 | const0_rtx))); | |
10470 | return SUBREG_REG (target); | |
10471 | } | |
10472 | else | |
10473 | return target; | |
10474 | } | |
10475 | ||
10476 | /* Fixup modeless constants to fit required mode. */ | |
10477 | ||
10478 | static rtx | |
10479 | fixup_modeless_constant (rtx x, machine_mode mode) | |
10480 | { | |
10481 | if (GET_MODE (x) == VOIDmode) | |
10482 | x = convert_to_mode (mode, x, 1); | |
10483 | return x; | |
10484 | } | |
10485 | ||
10486 | /* Subroutine of ix86_expand_builtin to take care of insns with | |
10487 | variable number of operands. */ | |
10488 | ||
10489 | static rtx | |
10490 | ix86_expand_args_builtin (const struct builtin_description *d, | |
10491 | tree exp, rtx target) | |
10492 | { | |
10493 | rtx pat, real_target; | |
10494 | unsigned int i, nargs; | |
10495 | unsigned int nargs_constant = 0; | |
10496 | unsigned int mask_pos = 0; | |
10497 | int num_memory = 0; | |
715a8bc8 | 10498 | rtx xops[6]; |
2bf6d935 ML |
10499 | bool second_arg_count = false; |
10500 | enum insn_code icode = d->icode; | |
10501 | const struct insn_data_d *insn_p = &insn_data[icode]; | |
10502 | machine_mode tmode = insn_p->operand[0].mode; | |
10503 | machine_mode rmode = VOIDmode; | |
10504 | bool swap = false; | |
10505 | enum rtx_code comparison = d->comparison; | |
10506 | ||
10507 | switch ((enum ix86_builtin_func_type) d->flag) | |
10508 | { | |
10509 | case V2DF_FTYPE_V2DF_ROUND: | |
10510 | case V4DF_FTYPE_V4DF_ROUND: | |
10511 | case V8DF_FTYPE_V8DF_ROUND: | |
10512 | case V4SF_FTYPE_V4SF_ROUND: | |
10513 | case V8SF_FTYPE_V8SF_ROUND: | |
10514 | case V16SF_FTYPE_V16SF_ROUND: | |
84bcefd5 | 10515 | case V8HF_FTYPE_V8HF_ROUND: |
10516 | case V16HF_FTYPE_V16HF_ROUND: | |
10517 | case V32HF_FTYPE_V32HF_ROUND: | |
2bf6d935 ML |
10518 | case V4SI_FTYPE_V4SF_ROUND: |
10519 | case V8SI_FTYPE_V8SF_ROUND: | |
10520 | case V16SI_FTYPE_V16SF_ROUND: | |
10521 | return ix86_expand_sse_round (d, exp, target); | |
10522 | case V4SI_FTYPE_V2DF_V2DF_ROUND: | |
10523 | case V8SI_FTYPE_V4DF_V4DF_ROUND: | |
10524 | case V16SI_FTYPE_V8DF_V8DF_ROUND: | |
10525 | return ix86_expand_sse_round_vec_pack_sfix (d, exp, target); | |
10526 | case INT_FTYPE_V8SF_V8SF_PTEST: | |
10527 | case INT_FTYPE_V4DI_V4DI_PTEST: | |
10528 | case INT_FTYPE_V4DF_V4DF_PTEST: | |
10529 | case INT_FTYPE_V4SF_V4SF_PTEST: | |
10530 | case INT_FTYPE_V2DI_V2DI_PTEST: | |
10531 | case INT_FTYPE_V2DF_V2DF_PTEST: | |
10532 | return ix86_expand_sse_ptest (d, exp, target); | |
10533 | case FLOAT128_FTYPE_FLOAT128: | |
10534 | case FLOAT_FTYPE_FLOAT: | |
a1ecc560 | 10535 | case FLOAT_FTYPE_BFLOAT16: |
2bf6d935 ML |
10536 | case INT_FTYPE_INT: |
10537 | case UINT_FTYPE_UINT: | |
10538 | case UINT16_FTYPE_UINT16: | |
10539 | case UINT64_FTYPE_INT: | |
10540 | case UINT64_FTYPE_UINT64: | |
10541 | case INT64_FTYPE_INT64: | |
10542 | case INT64_FTYPE_V4SF: | |
10543 | case INT64_FTYPE_V2DF: | |
10544 | case INT_FTYPE_V16QI: | |
10545 | case INT_FTYPE_V8QI: | |
10546 | case INT_FTYPE_V8SF: | |
10547 | case INT_FTYPE_V4DF: | |
10548 | case INT_FTYPE_V4SF: | |
10549 | case INT_FTYPE_V2DF: | |
10550 | case INT_FTYPE_V32QI: | |
10551 | case V16QI_FTYPE_V16QI: | |
10552 | case V8SI_FTYPE_V8SF: | |
10553 | case V8SI_FTYPE_V4SI: | |
10554 | case V8HI_FTYPE_V8HI: | |
10555 | case V8HI_FTYPE_V16QI: | |
10556 | case V8QI_FTYPE_V8QI: | |
10557 | case V8SF_FTYPE_V8SF: | |
10558 | case V8SF_FTYPE_V8SI: | |
10559 | case V8SF_FTYPE_V4SF: | |
10560 | case V8SF_FTYPE_V8HI: | |
10561 | case V4SI_FTYPE_V4SI: | |
10562 | case V4SI_FTYPE_V16QI: | |
10563 | case V4SI_FTYPE_V4SF: | |
10564 | case V4SI_FTYPE_V8SI: | |
10565 | case V4SI_FTYPE_V8HI: | |
10566 | case V4SI_FTYPE_V4DF: | |
10567 | case V4SI_FTYPE_V2DF: | |
10568 | case V4HI_FTYPE_V4HI: | |
10569 | case V4DF_FTYPE_V4DF: | |
10570 | case V4DF_FTYPE_V4SI: | |
10571 | case V4DF_FTYPE_V4SF: | |
10572 | case V4DF_FTYPE_V2DF: | |
10573 | case V4SF_FTYPE_V4SF: | |
10574 | case V4SF_FTYPE_V4SI: | |
10575 | case V4SF_FTYPE_V8SF: | |
10576 | case V4SF_FTYPE_V4DF: | |
10577 | case V4SF_FTYPE_V8HI: | |
10578 | case V4SF_FTYPE_V2DF: | |
10579 | case V2DI_FTYPE_V2DI: | |
10580 | case V2DI_FTYPE_V16QI: | |
10581 | case V2DI_FTYPE_V8HI: | |
10582 | case V2DI_FTYPE_V4SI: | |
10583 | case V2DF_FTYPE_V2DF: | |
10584 | case V2DF_FTYPE_V4SI: | |
10585 | case V2DF_FTYPE_V4DF: | |
10586 | case V2DF_FTYPE_V4SF: | |
10587 | case V2DF_FTYPE_V2SI: | |
10588 | case V2SI_FTYPE_V2SI: | |
10589 | case V2SI_FTYPE_V4SF: | |
10590 | case V2SI_FTYPE_V2SF: | |
10591 | case V2SI_FTYPE_V2DF: | |
10592 | case V2SF_FTYPE_V2SF: | |
10593 | case V2SF_FTYPE_V2SI: | |
10594 | case V32QI_FTYPE_V32QI: | |
10595 | case V32QI_FTYPE_V16QI: | |
10596 | case V16HI_FTYPE_V16HI: | |
10597 | case V16HI_FTYPE_V8HI: | |
10598 | case V8SI_FTYPE_V8SI: | |
10599 | case V16HI_FTYPE_V16QI: | |
10600 | case V8SI_FTYPE_V16QI: | |
10601 | case V4DI_FTYPE_V16QI: | |
10602 | case V8SI_FTYPE_V8HI: | |
10603 | case V4DI_FTYPE_V8HI: | |
10604 | case V4DI_FTYPE_V4SI: | |
10605 | case V4DI_FTYPE_V2DI: | |
10606 | case UQI_FTYPE_UQI: | |
10607 | case UHI_FTYPE_UHI: | |
10608 | case USI_FTYPE_USI: | |
10609 | case USI_FTYPE_UQI: | |
10610 | case USI_FTYPE_UHI: | |
10611 | case UDI_FTYPE_UDI: | |
10612 | case UHI_FTYPE_V16QI: | |
10613 | case USI_FTYPE_V32QI: | |
10614 | case UDI_FTYPE_V64QI: | |
10615 | case V16QI_FTYPE_UHI: | |
10616 | case V32QI_FTYPE_USI: | |
10617 | case V64QI_FTYPE_UDI: | |
10618 | case V8HI_FTYPE_UQI: | |
10619 | case V16HI_FTYPE_UHI: | |
10620 | case V32HI_FTYPE_USI: | |
10621 | case V4SI_FTYPE_UQI: | |
10622 | case V8SI_FTYPE_UQI: | |
10623 | case V4SI_FTYPE_UHI: | |
10624 | case V8SI_FTYPE_UHI: | |
10625 | case UQI_FTYPE_V8HI: | |
10626 | case UHI_FTYPE_V16HI: | |
10627 | case USI_FTYPE_V32HI: | |
10628 | case UQI_FTYPE_V4SI: | |
10629 | case UQI_FTYPE_V8SI: | |
10630 | case UHI_FTYPE_V16SI: | |
10631 | case UQI_FTYPE_V2DI: | |
10632 | case UQI_FTYPE_V4DI: | |
10633 | case UQI_FTYPE_V8DI: | |
10634 | case V16SI_FTYPE_UHI: | |
10635 | case V2DI_FTYPE_UQI: | |
10636 | case V4DI_FTYPE_UQI: | |
10637 | case V16SI_FTYPE_INT: | |
10638 | case V16SF_FTYPE_V8SF: | |
10639 | case V16SI_FTYPE_V8SI: | |
10640 | case V16SF_FTYPE_V4SF: | |
10641 | case V16SI_FTYPE_V4SI: | |
10642 | case V16SI_FTYPE_V16SF: | |
10643 | case V16SI_FTYPE_V16SI: | |
10644 | case V64QI_FTYPE_V64QI: | |
10645 | case V32HI_FTYPE_V32HI: | |
10646 | case V16SF_FTYPE_V16SF: | |
10647 | case V8DI_FTYPE_UQI: | |
10648 | case V8DI_FTYPE_V8DI: | |
10649 | case V8DF_FTYPE_V4DF: | |
10650 | case V8DF_FTYPE_V2DF: | |
10651 | case V8DF_FTYPE_V8DF: | |
10652 | case V4DI_FTYPE_V4DI: | |
87235f1e | 10653 | case V16BF_FTYPE_V16SF: |
10654 | case V8BF_FTYPE_V8SF: | |
10655 | case V8BF_FTYPE_V4SF: | |
2bf6d935 ML |
10656 | nargs = 1; |
10657 | break; | |
10658 | case V4SF_FTYPE_V4SF_VEC_MERGE: | |
10659 | case V2DF_FTYPE_V2DF_VEC_MERGE: | |
10660 | return ix86_expand_unop_vec_merge_builtin (icode, exp, target); | |
10661 | case FLOAT128_FTYPE_FLOAT128_FLOAT128: | |
10662 | case V16QI_FTYPE_V16QI_V16QI: | |
10663 | case V16QI_FTYPE_V8HI_V8HI: | |
b96cb2ca | 10664 | case V16HF_FTYPE_V16HF_V16HF: |
2bf6d935 ML |
10665 | case V16SF_FTYPE_V16SF_V16SF: |
10666 | case V8QI_FTYPE_V8QI_V8QI: | |
10667 | case V8QI_FTYPE_V4HI_V4HI: | |
10668 | case V8HI_FTYPE_V8HI_V8HI: | |
10669 | case V8HI_FTYPE_V16QI_V16QI: | |
10670 | case V8HI_FTYPE_V4SI_V4SI: | |
b96cb2ca | 10671 | case V8HF_FTYPE_V8HF_V8HF: |
2bf6d935 ML |
10672 | case V8SF_FTYPE_V8SF_V8SF: |
10673 | case V8SF_FTYPE_V8SF_V8SI: | |
10674 | case V8DF_FTYPE_V8DF_V8DF: | |
10675 | case V4SI_FTYPE_V4SI_V4SI: | |
10676 | case V4SI_FTYPE_V8HI_V8HI: | |
10677 | case V4SI_FTYPE_V2DF_V2DF: | |
10678 | case V4HI_FTYPE_V4HI_V4HI: | |
10679 | case V4HI_FTYPE_V8QI_V8QI: | |
10680 | case V4HI_FTYPE_V2SI_V2SI: | |
10681 | case V4DF_FTYPE_V4DF_V4DF: | |
10682 | case V4DF_FTYPE_V4DF_V4DI: | |
10683 | case V4SF_FTYPE_V4SF_V4SF: | |
10684 | case V4SF_FTYPE_V4SF_V4SI: | |
10685 | case V4SF_FTYPE_V4SF_V2SI: | |
10686 | case V4SF_FTYPE_V4SF_V2DF: | |
10687 | case V4SF_FTYPE_V4SF_UINT: | |
10688 | case V4SF_FTYPE_V4SF_DI: | |
10689 | case V4SF_FTYPE_V4SF_SI: | |
10690 | case V2DI_FTYPE_V2DI_V2DI: | |
10691 | case V2DI_FTYPE_V16QI_V16QI: | |
10692 | case V2DI_FTYPE_V4SI_V4SI: | |
10693 | case V2DI_FTYPE_V2DI_V16QI: | |
10694 | case V2SI_FTYPE_V2SI_V2SI: | |
10695 | case V2SI_FTYPE_V4HI_V4HI: | |
10696 | case V2SI_FTYPE_V2SF_V2SF: | |
10697 | case V2DF_FTYPE_V2DF_V2DF: | |
10698 | case V2DF_FTYPE_V2DF_V4SF: | |
10699 | case V2DF_FTYPE_V2DF_V2DI: | |
10700 | case V2DF_FTYPE_V2DF_DI: | |
10701 | case V2DF_FTYPE_V2DF_SI: | |
10702 | case V2DF_FTYPE_V2DF_UINT: | |
10703 | case V2SF_FTYPE_V2SF_V2SF: | |
10704 | case V1DI_FTYPE_V1DI_V1DI: | |
10705 | case V1DI_FTYPE_V8QI_V8QI: | |
10706 | case V1DI_FTYPE_V2SI_V2SI: | |
10707 | case V32QI_FTYPE_V16HI_V16HI: | |
10708 | case V16HI_FTYPE_V8SI_V8SI: | |
10709 | case V64QI_FTYPE_V64QI_V64QI: | |
10710 | case V32QI_FTYPE_V32QI_V32QI: | |
10711 | case V16HI_FTYPE_V32QI_V32QI: | |
10712 | case V16HI_FTYPE_V16HI_V16HI: | |
10713 | case V8SI_FTYPE_V4DF_V4DF: | |
10714 | case V8SI_FTYPE_V8SI_V8SI: | |
10715 | case V8SI_FTYPE_V16HI_V16HI: | |
10716 | case V4DI_FTYPE_V4DI_V4DI: | |
10717 | case V4DI_FTYPE_V8SI_V8SI: | |
6bb0776e | 10718 | case V4DI_FTYPE_V32QI_V32QI: |
2bf6d935 ML |
10719 | case V8DI_FTYPE_V64QI_V64QI: |
10720 | if (comparison == UNKNOWN) | |
10721 | return ix86_expand_binop_builtin (icode, exp, target); | |
10722 | nargs = 2; | |
10723 | break; | |
10724 | case V4SF_FTYPE_V4SF_V4SF_SWAP: | |
10725 | case V2DF_FTYPE_V2DF_V2DF_SWAP: | |
10726 | gcc_assert (comparison != UNKNOWN); | |
10727 | nargs = 2; | |
10728 | swap = true; | |
10729 | break; | |
10730 | case V16HI_FTYPE_V16HI_V8HI_COUNT: | |
10731 | case V16HI_FTYPE_V16HI_SI_COUNT: | |
10732 | case V8SI_FTYPE_V8SI_V4SI_COUNT: | |
10733 | case V8SI_FTYPE_V8SI_SI_COUNT: | |
10734 | case V4DI_FTYPE_V4DI_V2DI_COUNT: | |
10735 | case V4DI_FTYPE_V4DI_INT_COUNT: | |
10736 | case V8HI_FTYPE_V8HI_V8HI_COUNT: | |
10737 | case V8HI_FTYPE_V8HI_SI_COUNT: | |
10738 | case V4SI_FTYPE_V4SI_V4SI_COUNT: | |
10739 | case V4SI_FTYPE_V4SI_SI_COUNT: | |
10740 | case V4HI_FTYPE_V4HI_V4HI_COUNT: | |
10741 | case V4HI_FTYPE_V4HI_SI_COUNT: | |
10742 | case V2DI_FTYPE_V2DI_V2DI_COUNT: | |
10743 | case V2DI_FTYPE_V2DI_SI_COUNT: | |
10744 | case V2SI_FTYPE_V2SI_V2SI_COUNT: | |
10745 | case V2SI_FTYPE_V2SI_SI_COUNT: | |
10746 | case V1DI_FTYPE_V1DI_V1DI_COUNT: | |
10747 | case V1DI_FTYPE_V1DI_SI_COUNT: | |
10748 | nargs = 2; | |
10749 | second_arg_count = true; | |
10750 | break; | |
10751 | case V16HI_FTYPE_V16HI_INT_V16HI_UHI_COUNT: | |
10752 | case V16HI_FTYPE_V16HI_V8HI_V16HI_UHI_COUNT: | |
10753 | case V16SI_FTYPE_V16SI_INT_V16SI_UHI_COUNT: | |
10754 | case V16SI_FTYPE_V16SI_V4SI_V16SI_UHI_COUNT: | |
10755 | case V2DI_FTYPE_V2DI_INT_V2DI_UQI_COUNT: | |
10756 | case V2DI_FTYPE_V2DI_V2DI_V2DI_UQI_COUNT: | |
10757 | case V32HI_FTYPE_V32HI_INT_V32HI_USI_COUNT: | |
10758 | case V32HI_FTYPE_V32HI_V8HI_V32HI_USI_COUNT: | |
10759 | case V4DI_FTYPE_V4DI_INT_V4DI_UQI_COUNT: | |
10760 | case V4DI_FTYPE_V4DI_V2DI_V4DI_UQI_COUNT: | |
10761 | case V4SI_FTYPE_V4SI_INT_V4SI_UQI_COUNT: | |
10762 | case V4SI_FTYPE_V4SI_V4SI_V4SI_UQI_COUNT: | |
10763 | case V8DI_FTYPE_V8DI_INT_V8DI_UQI_COUNT: | |
10764 | case V8DI_FTYPE_V8DI_V2DI_V8DI_UQI_COUNT: | |
10765 | case V8HI_FTYPE_V8HI_INT_V8HI_UQI_COUNT: | |
10766 | case V8HI_FTYPE_V8HI_V8HI_V8HI_UQI_COUNT: | |
10767 | case V8SI_FTYPE_V8SI_INT_V8SI_UQI_COUNT: | |
10768 | case V8SI_FTYPE_V8SI_V4SI_V8SI_UQI_COUNT: | |
10769 | nargs = 4; | |
10770 | second_arg_count = true; | |
10771 | break; | |
10772 | case UINT64_FTYPE_UINT64_UINT64: | |
10773 | case UINT_FTYPE_UINT_UINT: | |
10774 | case UINT_FTYPE_UINT_USHORT: | |
10775 | case UINT_FTYPE_UINT_UCHAR: | |
10776 | case UINT16_FTYPE_UINT16_INT: | |
10777 | case UINT8_FTYPE_UINT8_INT: | |
10778 | case UQI_FTYPE_UQI_UQI: | |
10779 | case UHI_FTYPE_UHI_UHI: | |
10780 | case USI_FTYPE_USI_USI: | |
10781 | case UDI_FTYPE_UDI_UDI: | |
10782 | case V16SI_FTYPE_V8DF_V8DF: | |
87235f1e | 10783 | case V32BF_FTYPE_V16SF_V16SF: |
10784 | case V16BF_FTYPE_V8SF_V8SF: | |
10785 | case V8BF_FTYPE_V4SF_V4SF: | |
10786 | case V16BF_FTYPE_V16SF_UHI: | |
10787 | case V8BF_FTYPE_V8SF_UQI: | |
10788 | case V8BF_FTYPE_V4SF_UQI: | |
2bf6d935 ML |
10789 | nargs = 2; |
10790 | break; | |
10791 | case V2DI_FTYPE_V2DI_INT_CONVERT: | |
10792 | nargs = 2; | |
10793 | rmode = V1TImode; | |
10794 | nargs_constant = 1; | |
10795 | break; | |
10796 | case V4DI_FTYPE_V4DI_INT_CONVERT: | |
10797 | nargs = 2; | |
10798 | rmode = V2TImode; | |
10799 | nargs_constant = 1; | |
10800 | break; | |
10801 | case V8DI_FTYPE_V8DI_INT_CONVERT: | |
10802 | nargs = 2; | |
10803 | rmode = V4TImode; | |
10804 | nargs_constant = 1; | |
10805 | break; | |
10806 | case V8HI_FTYPE_V8HI_INT: | |
10807 | case V8HI_FTYPE_V8SF_INT: | |
10808 | case V16HI_FTYPE_V16SF_INT: | |
10809 | case V8HI_FTYPE_V4SF_INT: | |
10810 | case V8SF_FTYPE_V8SF_INT: | |
10811 | case V4SF_FTYPE_V16SF_INT: | |
10812 | case V16SF_FTYPE_V16SF_INT: | |
10813 | case V4SI_FTYPE_V4SI_INT: | |
10814 | case V4SI_FTYPE_V8SI_INT: | |
10815 | case V4HI_FTYPE_V4HI_INT: | |
10816 | case V4DF_FTYPE_V4DF_INT: | |
10817 | case V4DF_FTYPE_V8DF_INT: | |
10818 | case V4SF_FTYPE_V4SF_INT: | |
10819 | case V4SF_FTYPE_V8SF_INT: | |
10820 | case V2DI_FTYPE_V2DI_INT: | |
10821 | case V2DF_FTYPE_V2DF_INT: | |
10822 | case V2DF_FTYPE_V4DF_INT: | |
10823 | case V16HI_FTYPE_V16HI_INT: | |
10824 | case V8SI_FTYPE_V8SI_INT: | |
10825 | case V16SI_FTYPE_V16SI_INT: | |
10826 | case V4SI_FTYPE_V16SI_INT: | |
10827 | case V4DI_FTYPE_V4DI_INT: | |
10828 | case V2DI_FTYPE_V4DI_INT: | |
10829 | case V4DI_FTYPE_V8DI_INT: | |
2bf6d935 ML |
10830 | case UQI_FTYPE_UQI_UQI_CONST: |
10831 | case UHI_FTYPE_UHI_UQI: | |
10832 | case USI_FTYPE_USI_UQI: | |
10833 | case UDI_FTYPE_UDI_UQI: | |
10834 | nargs = 2; | |
10835 | nargs_constant = 1; | |
10836 | break; | |
10837 | case V16QI_FTYPE_V16QI_V16QI_V16QI: | |
10838 | case V8SF_FTYPE_V8SF_V8SF_V8SF: | |
10839 | case V4DF_FTYPE_V4DF_V4DF_V4DF: | |
10840 | case V4SF_FTYPE_V4SF_V4SF_V4SF: | |
10841 | case V2DF_FTYPE_V2DF_V2DF_V2DF: | |
10842 | case V32QI_FTYPE_V32QI_V32QI_V32QI: | |
10843 | case UHI_FTYPE_V16SI_V16SI_UHI: | |
10844 | case UQI_FTYPE_V8DI_V8DI_UQI: | |
10845 | case V16HI_FTYPE_V16SI_V16HI_UHI: | |
10846 | case V16QI_FTYPE_V16SI_V16QI_UHI: | |
10847 | case V16QI_FTYPE_V8DI_V16QI_UQI: | |
4204740f | 10848 | case V32HF_FTYPE_V32HF_V32HF_USI: |
2bf6d935 ML |
10849 | case V16SF_FTYPE_V16SF_V16SF_UHI: |
10850 | case V16SF_FTYPE_V4SF_V16SF_UHI: | |
10851 | case V16SI_FTYPE_SI_V16SI_UHI: | |
10852 | case V16SI_FTYPE_V16HI_V16SI_UHI: | |
10853 | case V16SI_FTYPE_V16QI_V16SI_UHI: | |
10854 | case V8SF_FTYPE_V4SF_V8SF_UQI: | |
10855 | case V4DF_FTYPE_V2DF_V4DF_UQI: | |
10856 | case V8SI_FTYPE_V4SI_V8SI_UQI: | |
10857 | case V8SI_FTYPE_SI_V8SI_UQI: | |
10858 | case V4SI_FTYPE_V4SI_V4SI_UQI: | |
10859 | case V4SI_FTYPE_SI_V4SI_UQI: | |
10860 | case V4DI_FTYPE_V2DI_V4DI_UQI: | |
10861 | case V4DI_FTYPE_DI_V4DI_UQI: | |
10862 | case V2DI_FTYPE_V2DI_V2DI_UQI: | |
10863 | case V2DI_FTYPE_DI_V2DI_UQI: | |
10864 | case V64QI_FTYPE_V64QI_V64QI_UDI: | |
10865 | case V64QI_FTYPE_V16QI_V64QI_UDI: | |
10866 | case V64QI_FTYPE_QI_V64QI_UDI: | |
10867 | case V32QI_FTYPE_V32QI_V32QI_USI: | |
10868 | case V32QI_FTYPE_V16QI_V32QI_USI: | |
10869 | case V32QI_FTYPE_QI_V32QI_USI: | |
10870 | case V16QI_FTYPE_V16QI_V16QI_UHI: | |
10871 | case V16QI_FTYPE_QI_V16QI_UHI: | |
10872 | case V32HI_FTYPE_V8HI_V32HI_USI: | |
10873 | case V32HI_FTYPE_HI_V32HI_USI: | |
10874 | case V16HI_FTYPE_V8HI_V16HI_UHI: | |
10875 | case V16HI_FTYPE_HI_V16HI_UHI: | |
10876 | case V8HI_FTYPE_V8HI_V8HI_UQI: | |
10877 | case V8HI_FTYPE_HI_V8HI_UQI: | |
4204740f | 10878 | case V16HF_FTYPE_V16HF_V16HF_UHI: |
2bf6d935 ML |
10879 | case V8SF_FTYPE_V8HI_V8SF_UQI: |
10880 | case V4SF_FTYPE_V8HI_V4SF_UQI: | |
bd610db0 | 10881 | case V8SI_FTYPE_V8HF_V8SI_UQI: |
5a744e50 | 10882 | case V8SF_FTYPE_V8HF_V8SF_UQI: |
2bf6d935 ML |
10883 | case V8SI_FTYPE_V8SF_V8SI_UQI: |
10884 | case V4SI_FTYPE_V4SF_V4SI_UQI: | |
bd610db0 | 10885 | case V4SI_FTYPE_V8HF_V4SI_UQI: |
5a744e50 | 10886 | case V4SF_FTYPE_V8HF_V4SF_UQI: |
bd610db0 | 10887 | case V4DI_FTYPE_V8HF_V4DI_UQI: |
2bf6d935 | 10888 | case V4DI_FTYPE_V4SF_V4DI_UQI: |
bd610db0 | 10889 | case V2DI_FTYPE_V8HF_V2DI_UQI: |
2bf6d935 | 10890 | case V2DI_FTYPE_V4SF_V2DI_UQI: |
4204740f | 10891 | case V8HF_FTYPE_V8HF_V8HF_UQI: |
081070bc | 10892 | case V8HF_FTYPE_V8HF_V8HF_V8HF: |
be0e4c32 | 10893 | case V8HF_FTYPE_V8HI_V8HF_UQI: |
10894 | case V8HF_FTYPE_V8SI_V8HF_UQI: | |
5a744e50 | 10895 | case V8HF_FTYPE_V8SF_V8HF_UQI: |
be0e4c32 | 10896 | case V8HF_FTYPE_V4SI_V8HF_UQI: |
5a744e50 | 10897 | case V8HF_FTYPE_V4SF_V8HF_UQI: |
be0e4c32 | 10898 | case V8HF_FTYPE_V4DI_V8HF_UQI: |
5a744e50 | 10899 | case V8HF_FTYPE_V4DF_V8HF_UQI: |
be0e4c32 | 10900 | case V8HF_FTYPE_V2DI_V8HF_UQI: |
5a744e50 | 10901 | case V8HF_FTYPE_V2DF_V8HF_UQI: |
2bf6d935 ML |
10902 | case V4SF_FTYPE_V4DI_V4SF_UQI: |
10903 | case V4SF_FTYPE_V2DI_V4SF_UQI: | |
10904 | case V4DF_FTYPE_V4DI_V4DF_UQI: | |
5a744e50 | 10905 | case V4DF_FTYPE_V8HF_V4DF_UQI: |
10906 | case V2DF_FTYPE_V8HF_V2DF_UQI: | |
2bf6d935 ML |
10907 | case V2DF_FTYPE_V2DI_V2DF_UQI: |
10908 | case V16QI_FTYPE_V8HI_V16QI_UQI: | |
10909 | case V16QI_FTYPE_V16HI_V16QI_UHI: | |
10910 | case V16QI_FTYPE_V4SI_V16QI_UQI: | |
10911 | case V16QI_FTYPE_V8SI_V16QI_UQI: | |
bd610db0 | 10912 | case V8HI_FTYPE_V8HF_V8HI_UQI: |
2bf6d935 ML |
10913 | case V8HI_FTYPE_V4SI_V8HI_UQI: |
10914 | case V8HI_FTYPE_V8SI_V8HI_UQI: | |
10915 | case V16QI_FTYPE_V2DI_V16QI_UQI: | |
10916 | case V16QI_FTYPE_V4DI_V16QI_UQI: | |
10917 | case V8HI_FTYPE_V2DI_V8HI_UQI: | |
10918 | case V8HI_FTYPE_V4DI_V8HI_UQI: | |
10919 | case V4SI_FTYPE_V2DI_V4SI_UQI: | |
10920 | case V4SI_FTYPE_V4DI_V4SI_UQI: | |
10921 | case V32QI_FTYPE_V32HI_V32QI_USI: | |
10922 | case UHI_FTYPE_V16QI_V16QI_UHI: | |
10923 | case USI_FTYPE_V32QI_V32QI_USI: | |
10924 | case UDI_FTYPE_V64QI_V64QI_UDI: | |
10925 | case UQI_FTYPE_V8HI_V8HI_UQI: | |
10926 | case UHI_FTYPE_V16HI_V16HI_UHI: | |
10927 | case USI_FTYPE_V32HI_V32HI_USI: | |
10928 | case UQI_FTYPE_V4SI_V4SI_UQI: | |
10929 | case UQI_FTYPE_V8SI_V8SI_UQI: | |
10930 | case UQI_FTYPE_V2DI_V2DI_UQI: | |
10931 | case UQI_FTYPE_V4DI_V4DI_UQI: | |
10932 | case V4SF_FTYPE_V2DF_V4SF_UQI: | |
10933 | case V4SF_FTYPE_V4DF_V4SF_UQI: | |
10934 | case V16SI_FTYPE_V16SI_V16SI_UHI: | |
10935 | case V16SI_FTYPE_V4SI_V16SI_UHI: | |
10936 | case V2DI_FTYPE_V4SI_V2DI_UQI: | |
10937 | case V2DI_FTYPE_V8HI_V2DI_UQI: | |
10938 | case V2DI_FTYPE_V16QI_V2DI_UQI: | |
10939 | case V4DI_FTYPE_V4DI_V4DI_UQI: | |
10940 | case V4DI_FTYPE_V4SI_V4DI_UQI: | |
10941 | case V4DI_FTYPE_V8HI_V4DI_UQI: | |
10942 | case V4DI_FTYPE_V16QI_V4DI_UQI: | |
10943 | case V4DI_FTYPE_V4DF_V4DI_UQI: | |
10944 | case V2DI_FTYPE_V2DF_V2DI_UQI: | |
10945 | case V4SI_FTYPE_V4DF_V4SI_UQI: | |
10946 | case V4SI_FTYPE_V2DF_V4SI_UQI: | |
10947 | case V4SI_FTYPE_V8HI_V4SI_UQI: | |
10948 | case V4SI_FTYPE_V16QI_V4SI_UQI: | |
10949 | case V4DI_FTYPE_V4DI_V4DI_V4DI: | |
10950 | case V8DF_FTYPE_V2DF_V8DF_UQI: | |
10951 | case V8DF_FTYPE_V4DF_V8DF_UQI: | |
10952 | case V8DF_FTYPE_V8DF_V8DF_UQI: | |
10953 | case V8SF_FTYPE_V8SF_V8SF_UQI: | |
10954 | case V8SF_FTYPE_V8SI_V8SF_UQI: | |
10955 | case V4DF_FTYPE_V4DF_V4DF_UQI: | |
10956 | case V4SF_FTYPE_V4SF_V4SF_UQI: | |
10957 | case V2DF_FTYPE_V2DF_V2DF_UQI: | |
10958 | case V2DF_FTYPE_V4SF_V2DF_UQI: | |
10959 | case V2DF_FTYPE_V4SI_V2DF_UQI: | |
10960 | case V4SF_FTYPE_V4SI_V4SF_UQI: | |
10961 | case V4DF_FTYPE_V4SF_V4DF_UQI: | |
10962 | case V4DF_FTYPE_V4SI_V4DF_UQI: | |
10963 | case V8SI_FTYPE_V8SI_V8SI_UQI: | |
10964 | case V8SI_FTYPE_V8HI_V8SI_UQI: | |
10965 | case V8SI_FTYPE_V16QI_V8SI_UQI: | |
10966 | case V8DF_FTYPE_V8SI_V8DF_UQI: | |
10967 | case V8DI_FTYPE_DI_V8DI_UQI: | |
10968 | case V16SF_FTYPE_V8SF_V16SF_UHI: | |
10969 | case V16SI_FTYPE_V8SI_V16SI_UHI: | |
be0e4c32 | 10970 | case V16HF_FTYPE_V16HI_V16HF_UHI: |
081070bc | 10971 | case V16HF_FTYPE_V16HF_V16HF_V16HF: |
bd610db0 | 10972 | case V16HI_FTYPE_V16HF_V16HI_UHI: |
2bf6d935 ML |
10973 | case V16HI_FTYPE_V16HI_V16HI_UHI: |
10974 | case V8HI_FTYPE_V16QI_V8HI_UQI: | |
10975 | case V16HI_FTYPE_V16QI_V16HI_UHI: | |
10976 | case V32HI_FTYPE_V32HI_V32HI_USI: | |
10977 | case V32HI_FTYPE_V32QI_V32HI_USI: | |
10978 | case V8DI_FTYPE_V16QI_V8DI_UQI: | |
10979 | case V8DI_FTYPE_V2DI_V8DI_UQI: | |
10980 | case V8DI_FTYPE_V4DI_V8DI_UQI: | |
10981 | case V8DI_FTYPE_V8DI_V8DI_UQI: | |
10982 | case V8DI_FTYPE_V8HI_V8DI_UQI: | |
10983 | case V8DI_FTYPE_V8SI_V8DI_UQI: | |
10984 | case V8HI_FTYPE_V8DI_V8HI_UQI: | |
10985 | case V8SI_FTYPE_V8DI_V8SI_UQI: | |
10986 | case V4SI_FTYPE_V4SI_V4SI_V4SI: | |
10987 | case V16SI_FTYPE_V16SI_V16SI_V16SI: | |
10988 | case V8DI_FTYPE_V8DI_V8DI_V8DI: | |
10989 | case V32HI_FTYPE_V32HI_V32HI_V32HI: | |
10990 | case V2DI_FTYPE_V2DI_V2DI_V2DI: | |
10991 | case V16HI_FTYPE_V16HI_V16HI_V16HI: | |
10992 | case V8SI_FTYPE_V8SI_V8SI_V8SI: | |
10993 | case V8HI_FTYPE_V8HI_V8HI_V8HI: | |
87235f1e | 10994 | case V32BF_FTYPE_V16SF_V16SF_USI: |
10995 | case V16BF_FTYPE_V8SF_V8SF_UHI: | |
10996 | case V8BF_FTYPE_V4SF_V4SF_UQI: | |
10997 | case V16BF_FTYPE_V16SF_V16BF_UHI: | |
10998 | case V8BF_FTYPE_V8SF_V8BF_UQI: | |
10999 | case V8BF_FTYPE_V4SF_V8BF_UQI: | |
11000 | case V16SF_FTYPE_V16SF_V32BF_V32BF: | |
11001 | case V8SF_FTYPE_V8SF_V16BF_V16BF: | |
11002 | case V4SF_FTYPE_V4SF_V8BF_V8BF: | |
2bf6d935 ML |
11003 | nargs = 3; |
11004 | break; | |
11005 | case V32QI_FTYPE_V32QI_V32QI_INT: | |
11006 | case V16HI_FTYPE_V16HI_V16HI_INT: | |
11007 | case V16QI_FTYPE_V16QI_V16QI_INT: | |
11008 | case V4DI_FTYPE_V4DI_V4DI_INT: | |
11009 | case V8HI_FTYPE_V8HI_V8HI_INT: | |
11010 | case V8SI_FTYPE_V8SI_V8SI_INT: | |
11011 | case V8SI_FTYPE_V8SI_V4SI_INT: | |
11012 | case V8SF_FTYPE_V8SF_V8SF_INT: | |
11013 | case V8SF_FTYPE_V8SF_V4SF_INT: | |
11014 | case V4SI_FTYPE_V4SI_V4SI_INT: | |
11015 | case V4DF_FTYPE_V4DF_V4DF_INT: | |
11016 | case V16SF_FTYPE_V16SF_V16SF_INT: | |
11017 | case V16SF_FTYPE_V16SF_V4SF_INT: | |
11018 | case V16SI_FTYPE_V16SI_V4SI_INT: | |
11019 | case V4DF_FTYPE_V4DF_V2DF_INT: | |
11020 | case V4SF_FTYPE_V4SF_V4SF_INT: | |
11021 | case V2DI_FTYPE_V2DI_V2DI_INT: | |
11022 | case V4DI_FTYPE_V4DI_V2DI_INT: | |
11023 | case V2DF_FTYPE_V2DF_V2DF_INT: | |
11024 | case UQI_FTYPE_V8DI_V8UDI_INT: | |
11025 | case UQI_FTYPE_V8DF_V8DF_INT: | |
11026 | case UQI_FTYPE_V2DF_V2DF_INT: | |
11027 | case UQI_FTYPE_V4SF_V4SF_INT: | |
11028 | case UHI_FTYPE_V16SI_V16SI_INT: | |
11029 | case UHI_FTYPE_V16SF_V16SF_INT: | |
11030 | case V64QI_FTYPE_V64QI_V64QI_INT: | |
11031 | case V32HI_FTYPE_V32HI_V32HI_INT: | |
11032 | case V16SI_FTYPE_V16SI_V16SI_INT: | |
11033 | case V8DI_FTYPE_V8DI_V8DI_INT: | |
11034 | nargs = 3; | |
11035 | nargs_constant = 1; | |
11036 | break; | |
11037 | case V4DI_FTYPE_V4DI_V4DI_INT_CONVERT: | |
11038 | nargs = 3; | |
11039 | rmode = V4DImode; | |
11040 | nargs_constant = 1; | |
11041 | break; | |
11042 | case V2DI_FTYPE_V2DI_V2DI_INT_CONVERT: | |
11043 | nargs = 3; | |
11044 | rmode = V2DImode; | |
11045 | nargs_constant = 1; | |
11046 | break; | |
11047 | case V1DI_FTYPE_V1DI_V1DI_INT_CONVERT: | |
11048 | nargs = 3; | |
11049 | rmode = DImode; | |
11050 | nargs_constant = 1; | |
11051 | break; | |
11052 | case V2DI_FTYPE_V2DI_UINT_UINT: | |
11053 | nargs = 3; | |
11054 | nargs_constant = 2; | |
11055 | break; | |
11056 | case V8DI_FTYPE_V8DI_V8DI_INT_CONVERT: | |
11057 | nargs = 3; | |
11058 | rmode = V8DImode; | |
11059 | nargs_constant = 1; | |
11060 | break; | |
11061 | case V8DI_FTYPE_V8DI_V8DI_INT_V8DI_UDI_CONVERT: | |
11062 | nargs = 5; | |
11063 | rmode = V8DImode; | |
11064 | mask_pos = 2; | |
11065 | nargs_constant = 1; | |
11066 | break; | |
11067 | case QI_FTYPE_V8DF_INT_UQI: | |
11068 | case QI_FTYPE_V4DF_INT_UQI: | |
11069 | case QI_FTYPE_V2DF_INT_UQI: | |
11070 | case HI_FTYPE_V16SF_INT_UHI: | |
11071 | case QI_FTYPE_V8SF_INT_UQI: | |
11072 | case QI_FTYPE_V4SF_INT_UQI: | |
8486e9f2 | 11073 | case QI_FTYPE_V8HF_INT_UQI: |
11074 | case HI_FTYPE_V16HF_INT_UHI: | |
11075 | case SI_FTYPE_V32HF_INT_USI: | |
2bf6d935 ML |
11076 | case V4SI_FTYPE_V4SI_V4SI_UHI: |
11077 | case V8SI_FTYPE_V8SI_V8SI_UHI: | |
11078 | nargs = 3; | |
11079 | mask_pos = 1; | |
11080 | nargs_constant = 1; | |
11081 | break; | |
11082 | case V4DI_FTYPE_V4DI_V4DI_INT_V4DI_USI_CONVERT: | |
11083 | nargs = 5; | |
11084 | rmode = V4DImode; | |
11085 | mask_pos = 2; | |
11086 | nargs_constant = 1; | |
11087 | break; | |
11088 | case V2DI_FTYPE_V2DI_V2DI_INT_V2DI_UHI_CONVERT: | |
11089 | nargs = 5; | |
11090 | rmode = V2DImode; | |
11091 | mask_pos = 2; | |
11092 | nargs_constant = 1; | |
11093 | break; | |
11094 | case V32QI_FTYPE_V32QI_V32QI_V32QI_USI: | |
11095 | case V32HI_FTYPE_V32HI_V32HI_V32HI_USI: | |
11096 | case V32HI_FTYPE_V64QI_V64QI_V32HI_USI: | |
11097 | case V16SI_FTYPE_V32HI_V32HI_V16SI_UHI: | |
11098 | case V64QI_FTYPE_V64QI_V64QI_V64QI_UDI: | |
11099 | case V32HI_FTYPE_V32HI_V8HI_V32HI_USI: | |
11100 | case V16HI_FTYPE_V16HI_V8HI_V16HI_UHI: | |
11101 | case V8SI_FTYPE_V8SI_V4SI_V8SI_UQI: | |
11102 | case V4DI_FTYPE_V4DI_V2DI_V4DI_UQI: | |
11103 | case V64QI_FTYPE_V32HI_V32HI_V64QI_UDI: | |
11104 | case V32QI_FTYPE_V16HI_V16HI_V32QI_USI: | |
11105 | case V16QI_FTYPE_V8HI_V8HI_V16QI_UHI: | |
11106 | case V32HI_FTYPE_V16SI_V16SI_V32HI_USI: | |
11107 | case V16HI_FTYPE_V8SI_V8SI_V16HI_UHI: | |
11108 | case V8HI_FTYPE_V4SI_V4SI_V8HI_UQI: | |
11109 | case V4DF_FTYPE_V4DF_V4DI_V4DF_UQI: | |
bd7a34ef | 11110 | case V32HF_FTYPE_V32HF_V32HF_V32HF_USI: |
2bf6d935 ML |
11111 | case V8SF_FTYPE_V8SF_V8SI_V8SF_UQI: |
11112 | case V4SF_FTYPE_V4SF_V4SI_V4SF_UQI: | |
11113 | case V2DF_FTYPE_V2DF_V2DI_V2DF_UQI: | |
11114 | case V2DI_FTYPE_V4SI_V4SI_V2DI_UQI: | |
11115 | case V4DI_FTYPE_V8SI_V8SI_V4DI_UQI: | |
11116 | case V4DF_FTYPE_V4DI_V4DF_V4DF_UQI: | |
11117 | case V8SF_FTYPE_V8SI_V8SF_V8SF_UQI: | |
11118 | case V2DF_FTYPE_V2DI_V2DF_V2DF_UQI: | |
11119 | case V4SF_FTYPE_V4SI_V4SF_V4SF_UQI: | |
11120 | case V16SF_FTYPE_V16SF_V16SF_V16SF_UHI: | |
11121 | case V16SF_FTYPE_V16SF_V16SI_V16SF_UHI: | |
11122 | case V16SF_FTYPE_V16SI_V16SF_V16SF_UHI: | |
11123 | case V16SI_FTYPE_V16SI_V16SI_V16SI_UHI: | |
11124 | case V16SI_FTYPE_V16SI_V4SI_V16SI_UHI: | |
11125 | case V8HI_FTYPE_V8HI_V8HI_V8HI_UQI: | |
11126 | case V8SI_FTYPE_V8SI_V8SI_V8SI_UQI: | |
11127 | case V4SI_FTYPE_V4SI_V4SI_V4SI_UQI: | |
081070bc | 11128 | case V16HF_FTYPE_V16HF_V16HF_V16HF_UQI: |
bd7a34ef | 11129 | case V16HF_FTYPE_V16HF_V16HF_V16HF_UHI: |
2bf6d935 ML |
11130 | case V8SF_FTYPE_V8SF_V8SF_V8SF_UQI: |
11131 | case V16QI_FTYPE_V16QI_V16QI_V16QI_UHI: | |
11132 | case V16HI_FTYPE_V16HI_V16HI_V16HI_UHI: | |
11133 | case V2DI_FTYPE_V2DI_V2DI_V2DI_UQI: | |
11134 | case V2DF_FTYPE_V2DF_V2DF_V2DF_UQI: | |
11135 | case V4DI_FTYPE_V4DI_V4DI_V4DI_UQI: | |
11136 | case V4DF_FTYPE_V4DF_V4DF_V4DF_UQI: | |
bd7a34ef | 11137 | case V8HF_FTYPE_V8HF_V8HF_V8HF_UQI: |
2bf6d935 ML |
11138 | case V4SF_FTYPE_V4SF_V4SF_V4SF_UQI: |
11139 | case V8DF_FTYPE_V8DF_V8DF_V8DF_UQI: | |
11140 | case V8DF_FTYPE_V8DF_V8DI_V8DF_UQI: | |
11141 | case V8DF_FTYPE_V8DI_V8DF_V8DF_UQI: | |
11142 | case V8DI_FTYPE_V16SI_V16SI_V8DI_UQI: | |
11143 | case V8DI_FTYPE_V8DI_V2DI_V8DI_UQI: | |
11144 | case V8DI_FTYPE_V8DI_V8DI_V8DI_UQI: | |
11145 | case V8HI_FTYPE_V16QI_V16QI_V8HI_UQI: | |
11146 | case V16HI_FTYPE_V32QI_V32QI_V16HI_UHI: | |
11147 | case V8SI_FTYPE_V16HI_V16HI_V8SI_UQI: | |
11148 | case V4SI_FTYPE_V8HI_V8HI_V4SI_UQI: | |
87235f1e | 11149 | case V32BF_FTYPE_V16SF_V16SF_V32BF_USI: |
11150 | case V16BF_FTYPE_V8SF_V8SF_V16BF_UHI: | |
11151 | case V8BF_FTYPE_V4SF_V4SF_V8BF_UQI: | |
2bf6d935 ML |
11152 | nargs = 4; |
11153 | break; | |
11154 | case V2DF_FTYPE_V2DF_V2DF_V2DI_INT: | |
11155 | case V4DF_FTYPE_V4DF_V4DF_V4DI_INT: | |
11156 | case V4SF_FTYPE_V4SF_V4SF_V4SI_INT: | |
11157 | case V8SF_FTYPE_V8SF_V8SF_V8SI_INT: | |
11158 | case V16SF_FTYPE_V16SF_V16SF_V16SI_INT: | |
11159 | nargs = 4; | |
11160 | nargs_constant = 1; | |
11161 | break; | |
11162 | case UQI_FTYPE_V4DI_V4DI_INT_UQI: | |
11163 | case UQI_FTYPE_V8SI_V8SI_INT_UQI: | |
11164 | case QI_FTYPE_V4DF_V4DF_INT_UQI: | |
11165 | case QI_FTYPE_V8SF_V8SF_INT_UQI: | |
0f200733 | 11166 | case UHI_FTYPE_V16HF_V16HF_INT_UHI: |
2bf6d935 ML |
11167 | case UQI_FTYPE_V2DI_V2DI_INT_UQI: |
11168 | case UQI_FTYPE_V4SI_V4SI_INT_UQI: | |
11169 | case UQI_FTYPE_V2DF_V2DF_INT_UQI: | |
11170 | case UQI_FTYPE_V4SF_V4SF_INT_UQI: | |
0f200733 | 11171 | case UQI_FTYPE_V8HF_V8HF_INT_UQI: |
2bf6d935 ML |
11172 | case UDI_FTYPE_V64QI_V64QI_INT_UDI: |
11173 | case USI_FTYPE_V32QI_V32QI_INT_USI: | |
11174 | case UHI_FTYPE_V16QI_V16QI_INT_UHI: | |
11175 | case USI_FTYPE_V32HI_V32HI_INT_USI: | |
0f200733 | 11176 | case USI_FTYPE_V32HF_V32HF_INT_USI: |
2bf6d935 ML |
11177 | case UHI_FTYPE_V16HI_V16HI_INT_UHI: |
11178 | case UQI_FTYPE_V8HI_V8HI_INT_UQI: | |
2bf6d935 ML |
11179 | nargs = 4; |
11180 | mask_pos = 1; | |
11181 | nargs_constant = 1; | |
11182 | break; | |
11183 | case V2DI_FTYPE_V2DI_V2DI_UINT_UINT: | |
11184 | nargs = 4; | |
11185 | nargs_constant = 2; | |
11186 | break; | |
11187 | case UCHAR_FTYPE_UCHAR_UINT_UINT_PUNSIGNED: | |
11188 | case UCHAR_FTYPE_UCHAR_ULONGLONG_ULONGLONG_PULONGLONG: | |
87235f1e | 11189 | case V16SF_FTYPE_V16SF_V32BF_V32BF_UHI: |
11190 | case V8SF_FTYPE_V8SF_V16BF_V16BF_UQI: | |
11191 | case V4SF_FTYPE_V4SF_V8BF_V8BF_UQI: | |
2bf6d935 ML |
11192 | nargs = 4; |
11193 | break; | |
11194 | case UQI_FTYPE_V8DI_V8DI_INT_UQI: | |
11195 | case UHI_FTYPE_V16SI_V16SI_INT_UHI: | |
11196 | mask_pos = 1; | |
11197 | nargs = 4; | |
11198 | nargs_constant = 1; | |
11199 | break; | |
11200 | case V8SF_FTYPE_V8SF_INT_V8SF_UQI: | |
11201 | case V4SF_FTYPE_V4SF_INT_V4SF_UQI: | |
11202 | case V2DF_FTYPE_V4DF_INT_V2DF_UQI: | |
11203 | case V2DI_FTYPE_V4DI_INT_V2DI_UQI: | |
11204 | case V8SF_FTYPE_V16SF_INT_V8SF_UQI: | |
11205 | case V8SI_FTYPE_V16SI_INT_V8SI_UQI: | |
11206 | case V2DF_FTYPE_V8DF_INT_V2DF_UQI: | |
11207 | case V2DI_FTYPE_V8DI_INT_V2DI_UQI: | |
11208 | case V4SF_FTYPE_V8SF_INT_V4SF_UQI: | |
11209 | case V4SI_FTYPE_V8SI_INT_V4SI_UQI: | |
11210 | case V8HI_FTYPE_V8SF_INT_V8HI_UQI: | |
11211 | case V8HI_FTYPE_V4SF_INT_V8HI_UQI: | |
11212 | case V32HI_FTYPE_V32HI_INT_V32HI_USI: | |
11213 | case V16HI_FTYPE_V16HI_INT_V16HI_UHI: | |
11214 | case V8HI_FTYPE_V8HI_INT_V8HI_UQI: | |
11215 | case V4DI_FTYPE_V4DI_INT_V4DI_UQI: | |
11216 | case V2DI_FTYPE_V2DI_INT_V2DI_UQI: | |
11217 | case V8SI_FTYPE_V8SI_INT_V8SI_UQI: | |
11218 | case V4SI_FTYPE_V4SI_INT_V4SI_UQI: | |
11219 | case V4DF_FTYPE_V4DF_INT_V4DF_UQI: | |
11220 | case V2DF_FTYPE_V2DF_INT_V2DF_UQI: | |
11221 | case V8DF_FTYPE_V8DF_INT_V8DF_UQI: | |
11222 | case V16SF_FTYPE_V16SF_INT_V16SF_UHI: | |
11223 | case V16HI_FTYPE_V16SF_INT_V16HI_UHI: | |
11224 | case V16SI_FTYPE_V16SI_INT_V16SI_UHI: | |
8bed7617 | 11225 | case V16HF_FTYPE_V16HF_INT_V16HF_UHI: |
11226 | case V8HF_FTYPE_V8HF_INT_V8HF_UQI: | |
2bf6d935 ML |
11227 | case V4SI_FTYPE_V16SI_INT_V4SI_UQI: |
11228 | case V4DI_FTYPE_V8DI_INT_V4DI_UQI: | |
11229 | case V4DF_FTYPE_V8DF_INT_V4DF_UQI: | |
11230 | case V4SF_FTYPE_V16SF_INT_V4SF_UQI: | |
11231 | case V8DI_FTYPE_V8DI_INT_V8DI_UQI: | |
11232 | nargs = 4; | |
11233 | mask_pos = 2; | |
11234 | nargs_constant = 1; | |
11235 | break; | |
11236 | case V16SF_FTYPE_V16SF_V4SF_INT_V16SF_UHI: | |
11237 | case V16SI_FTYPE_V16SI_V4SI_INT_V16SI_UHI: | |
11238 | case V8DF_FTYPE_V8DF_V8DF_INT_V8DF_UQI: | |
11239 | case V8DI_FTYPE_V8DI_V8DI_INT_V8DI_UQI: | |
11240 | case V16SF_FTYPE_V16SF_V16SF_INT_V16SF_UHI: | |
11241 | case V16SI_FTYPE_V16SI_V16SI_INT_V16SI_UHI: | |
11242 | case V4SF_FTYPE_V4SF_V4SF_INT_V4SF_UQI: | |
11243 | case V2DF_FTYPE_V2DF_V2DF_INT_V2DF_UQI: | |
11244 | case V8DF_FTYPE_V8DF_V4DF_INT_V8DF_UQI: | |
11245 | case V8DI_FTYPE_V8DI_V4DI_INT_V8DI_UQI: | |
11246 | case V4DF_FTYPE_V4DF_V4DF_INT_V4DF_UQI: | |
11247 | case V8SF_FTYPE_V8SF_V8SF_INT_V8SF_UQI: | |
11248 | case V8DF_FTYPE_V8DF_V2DF_INT_V8DF_UQI: | |
11249 | case V8DI_FTYPE_V8DI_V2DI_INT_V8DI_UQI: | |
11250 | case V8SI_FTYPE_V8SI_V8SI_INT_V8SI_UQI: | |
11251 | case V4DI_FTYPE_V4DI_V4DI_INT_V4DI_UQI: | |
11252 | case V4SI_FTYPE_V4SI_V4SI_INT_V4SI_UQI: | |
11253 | case V2DI_FTYPE_V2DI_V2DI_INT_V2DI_UQI: | |
11254 | case V32HI_FTYPE_V64QI_V64QI_INT_V32HI_USI: | |
11255 | case V16HI_FTYPE_V32QI_V32QI_INT_V16HI_UHI: | |
11256 | case V8HI_FTYPE_V16QI_V16QI_INT_V8HI_UQI: | |
11257 | case V16SF_FTYPE_V16SF_V8SF_INT_V16SF_UHI: | |
11258 | case V16SI_FTYPE_V16SI_V8SI_INT_V16SI_UHI: | |
11259 | case V8SF_FTYPE_V8SF_V4SF_INT_V8SF_UQI: | |
11260 | case V8SI_FTYPE_V8SI_V4SI_INT_V8SI_UQI: | |
11261 | case V4DI_FTYPE_V4DI_V2DI_INT_V4DI_UQI: | |
11262 | case V4DF_FTYPE_V4DF_V2DF_INT_V4DF_UQI: | |
11263 | nargs = 5; | |
11264 | mask_pos = 2; | |
11265 | nargs_constant = 1; | |
11266 | break; | |
11267 | case V8DI_FTYPE_V8DI_V8DI_V8DI_INT_UQI: | |
11268 | case V16SI_FTYPE_V16SI_V16SI_V16SI_INT_UHI: | |
11269 | case V2DF_FTYPE_V2DF_V2DF_V2DI_INT_UQI: | |
11270 | case V4SF_FTYPE_V4SF_V4SF_V4SI_INT_UQI: | |
11271 | case V8SF_FTYPE_V8SF_V8SF_V8SI_INT_UQI: | |
11272 | case V8SI_FTYPE_V8SI_V8SI_V8SI_INT_UQI: | |
11273 | case V4DF_FTYPE_V4DF_V4DF_V4DI_INT_UQI: | |
11274 | case V4DI_FTYPE_V4DI_V4DI_V4DI_INT_UQI: | |
11275 | case V4SI_FTYPE_V4SI_V4SI_V4SI_INT_UQI: | |
11276 | case V2DI_FTYPE_V2DI_V2DI_V2DI_INT_UQI: | |
11277 | nargs = 5; | |
11278 | mask_pos = 1; | |
11279 | nargs_constant = 1; | |
11280 | break; | |
11281 | case V64QI_FTYPE_V64QI_V64QI_INT_V64QI_UDI: | |
11282 | case V32QI_FTYPE_V32QI_V32QI_INT_V32QI_USI: | |
11283 | case V16QI_FTYPE_V16QI_V16QI_INT_V16QI_UHI: | |
11284 | case V32HI_FTYPE_V32HI_V32HI_INT_V32HI_INT: | |
11285 | case V16SI_FTYPE_V16SI_V16SI_INT_V16SI_INT: | |
11286 | case V8DI_FTYPE_V8DI_V8DI_INT_V8DI_INT: | |
11287 | case V16HI_FTYPE_V16HI_V16HI_INT_V16HI_INT: | |
11288 | case V8SI_FTYPE_V8SI_V8SI_INT_V8SI_INT: | |
11289 | case V4DI_FTYPE_V4DI_V4DI_INT_V4DI_INT: | |
11290 | case V8HI_FTYPE_V8HI_V8HI_INT_V8HI_INT: | |
11291 | case V4SI_FTYPE_V4SI_V4SI_INT_V4SI_INT: | |
11292 | case V2DI_FTYPE_V2DI_V2DI_INT_V2DI_INT: | |
11293 | nargs = 5; | |
11294 | mask_pos = 1; | |
11295 | nargs_constant = 2; | |
11296 | break; | |
11297 | ||
11298 | default: | |
11299 | gcc_unreachable (); | |
11300 | } | |
11301 | ||
715a8bc8 | 11302 | gcc_assert (nargs <= ARRAY_SIZE (xops)); |
2bf6d935 ML |
11303 | |
11304 | if (comparison != UNKNOWN) | |
11305 | { | |
11306 | gcc_assert (nargs == 2); | |
11307 | return ix86_expand_sse_compare (d, exp, target, swap); | |
11308 | } | |
11309 | ||
11310 | if (rmode == VOIDmode || rmode == tmode) | |
11311 | { | |
11312 | if (optimize | |
11313 | || target == 0 | |
11314 | || GET_MODE (target) != tmode | |
11315 | || !insn_p->operand[0].predicate (target, tmode)) | |
11316 | target = gen_reg_rtx (tmode); | |
11317 | else if (memory_operand (target, tmode)) | |
11318 | num_memory++; | |
11319 | real_target = target; | |
11320 | } | |
11321 | else | |
11322 | { | |
11323 | real_target = gen_reg_rtx (tmode); | |
11324 | target = lowpart_subreg (rmode, real_target, tmode); | |
11325 | } | |
11326 | ||
11327 | for (i = 0; i < nargs; i++) | |
11328 | { | |
11329 | tree arg = CALL_EXPR_ARG (exp, i); | |
11330 | rtx op = expand_normal (arg); | |
11331 | machine_mode mode = insn_p->operand[i + 1].mode; | |
11332 | bool match = insn_p->operand[i + 1].predicate (op, mode); | |
11333 | ||
11334 | if (second_arg_count && i == 1) | |
11335 | { | |
11336 | /* SIMD shift insns take either an 8-bit immediate or | |
11337 | register as count. But builtin functions take int as | |
11338 | count. If count doesn't match, we put it in register. | |
11339 | The instructions are using 64-bit count, if op is just | |
11340 | 32-bit, zero-extend it, as negative shift counts | |
11341 | are undefined behavior and zero-extension is more | |
11342 | efficient. */ | |
11343 | if (!match) | |
11344 | { | |
11345 | if (SCALAR_INT_MODE_P (GET_MODE (op))) | |
11346 | op = convert_modes (mode, GET_MODE (op), op, 1); | |
11347 | else | |
11348 | op = lowpart_subreg (mode, op, GET_MODE (op)); | |
11349 | if (!insn_p->operand[i + 1].predicate (op, mode)) | |
11350 | op = copy_to_reg (op); | |
11351 | } | |
11352 | } | |
11353 | else if ((mask_pos && (nargs - i - mask_pos) == nargs_constant) || | |
11354 | (!mask_pos && (nargs - i) <= nargs_constant)) | |
11355 | { | |
11356 | if (!match) | |
11357 | switch (icode) | |
11358 | { | |
11359 | case CODE_FOR_avx_vinsertf128v4di: | |
11360 | case CODE_FOR_avx_vextractf128v4di: | |
11361 | error ("the last argument must be an 1-bit immediate"); | |
11362 | return const0_rtx; | |
11363 | ||
11364 | case CODE_FOR_avx512f_cmpv8di3_mask: | |
11365 | case CODE_FOR_avx512f_cmpv16si3_mask: | |
11366 | case CODE_FOR_avx512f_ucmpv8di3_mask: | |
11367 | case CODE_FOR_avx512f_ucmpv16si3_mask: | |
11368 | case CODE_FOR_avx512vl_cmpv4di3_mask: | |
11369 | case CODE_FOR_avx512vl_cmpv8si3_mask: | |
11370 | case CODE_FOR_avx512vl_ucmpv4di3_mask: | |
11371 | case CODE_FOR_avx512vl_ucmpv8si3_mask: | |
11372 | case CODE_FOR_avx512vl_cmpv2di3_mask: | |
11373 | case CODE_FOR_avx512vl_cmpv4si3_mask: | |
11374 | case CODE_FOR_avx512vl_ucmpv2di3_mask: | |
11375 | case CODE_FOR_avx512vl_ucmpv4si3_mask: | |
11376 | error ("the last argument must be a 3-bit immediate"); | |
11377 | return const0_rtx; | |
11378 | ||
11379 | case CODE_FOR_sse4_1_roundsd: | |
11380 | case CODE_FOR_sse4_1_roundss: | |
11381 | ||
11382 | case CODE_FOR_sse4_1_roundpd: | |
11383 | case CODE_FOR_sse4_1_roundps: | |
11384 | case CODE_FOR_avx_roundpd256: | |
11385 | case CODE_FOR_avx_roundps256: | |
11386 | ||
11387 | case CODE_FOR_sse4_1_roundpd_vec_pack_sfix: | |
11388 | case CODE_FOR_sse4_1_roundps_sfix: | |
11389 | case CODE_FOR_avx_roundpd_vec_pack_sfix256: | |
11390 | case CODE_FOR_avx_roundps_sfix256: | |
11391 | ||
11392 | case CODE_FOR_sse4_1_blendps: | |
11393 | case CODE_FOR_avx_blendpd256: | |
11394 | case CODE_FOR_avx_vpermilv4df: | |
11395 | case CODE_FOR_avx_vpermilv4df_mask: | |
11396 | case CODE_FOR_avx512f_getmantv8df_mask: | |
11397 | case CODE_FOR_avx512f_getmantv16sf_mask: | |
8486e9f2 | 11398 | case CODE_FOR_avx512vl_getmantv16hf_mask: |
2bf6d935 ML |
11399 | case CODE_FOR_avx512vl_getmantv8sf_mask: |
11400 | case CODE_FOR_avx512vl_getmantv4df_mask: | |
8486e9f2 | 11401 | case CODE_FOR_avx512fp16_getmantv8hf_mask: |
2bf6d935 ML |
11402 | case CODE_FOR_avx512vl_getmantv4sf_mask: |
11403 | case CODE_FOR_avx512vl_getmantv2df_mask: | |
11404 | case CODE_FOR_avx512dq_rangepv8df_mask_round: | |
11405 | case CODE_FOR_avx512dq_rangepv16sf_mask_round: | |
11406 | case CODE_FOR_avx512dq_rangepv4df_mask: | |
11407 | case CODE_FOR_avx512dq_rangepv8sf_mask: | |
11408 | case CODE_FOR_avx512dq_rangepv2df_mask: | |
11409 | case CODE_FOR_avx512dq_rangepv4sf_mask: | |
11410 | case CODE_FOR_avx_shufpd256_mask: | |
11411 | error ("the last argument must be a 4-bit immediate"); | |
11412 | return const0_rtx; | |
11413 | ||
11414 | case CODE_FOR_sha1rnds4: | |
11415 | case CODE_FOR_sse4_1_blendpd: | |
11416 | case CODE_FOR_avx_vpermilv2df: | |
11417 | case CODE_FOR_avx_vpermilv2df_mask: | |
11418 | case CODE_FOR_xop_vpermil2v2df3: | |
11419 | case CODE_FOR_xop_vpermil2v4sf3: | |
11420 | case CODE_FOR_xop_vpermil2v4df3: | |
11421 | case CODE_FOR_xop_vpermil2v8sf3: | |
11422 | case CODE_FOR_avx512f_vinsertf32x4_mask: | |
11423 | case CODE_FOR_avx512f_vinserti32x4_mask: | |
11424 | case CODE_FOR_avx512f_vextractf32x4_mask: | |
11425 | case CODE_FOR_avx512f_vextracti32x4_mask: | |
11426 | case CODE_FOR_sse2_shufpd: | |
11427 | case CODE_FOR_sse2_shufpd_mask: | |
11428 | case CODE_FOR_avx512dq_shuf_f64x2_mask: | |
11429 | case CODE_FOR_avx512dq_shuf_i64x2_mask: | |
11430 | case CODE_FOR_avx512vl_shuf_i32x4_mask: | |
11431 | case CODE_FOR_avx512vl_shuf_f32x4_mask: | |
11432 | error ("the last argument must be a 2-bit immediate"); | |
11433 | return const0_rtx; | |
11434 | ||
11435 | case CODE_FOR_avx_vextractf128v4df: | |
11436 | case CODE_FOR_avx_vextractf128v8sf: | |
11437 | case CODE_FOR_avx_vextractf128v8si: | |
11438 | case CODE_FOR_avx_vinsertf128v4df: | |
11439 | case CODE_FOR_avx_vinsertf128v8sf: | |
11440 | case CODE_FOR_avx_vinsertf128v8si: | |
11441 | case CODE_FOR_avx512f_vinsertf64x4_mask: | |
11442 | case CODE_FOR_avx512f_vinserti64x4_mask: | |
11443 | case CODE_FOR_avx512f_vextractf64x4_mask: | |
11444 | case CODE_FOR_avx512f_vextracti64x4_mask: | |
11445 | case CODE_FOR_avx512dq_vinsertf32x8_mask: | |
11446 | case CODE_FOR_avx512dq_vinserti32x8_mask: | |
11447 | case CODE_FOR_avx512vl_vinsertv4df: | |
11448 | case CODE_FOR_avx512vl_vinsertv4di: | |
11449 | case CODE_FOR_avx512vl_vinsertv8sf: | |
11450 | case CODE_FOR_avx512vl_vinsertv8si: | |
11451 | error ("the last argument must be a 1-bit immediate"); | |
11452 | return const0_rtx; | |
11453 | ||
11454 | case CODE_FOR_avx_vmcmpv2df3: | |
11455 | case CODE_FOR_avx_vmcmpv4sf3: | |
11456 | case CODE_FOR_avx_cmpv2df3: | |
11457 | case CODE_FOR_avx_cmpv4sf3: | |
11458 | case CODE_FOR_avx_cmpv4df3: | |
11459 | case CODE_FOR_avx_cmpv8sf3: | |
11460 | case CODE_FOR_avx512f_cmpv8df3_mask: | |
11461 | case CODE_FOR_avx512f_cmpv16sf3_mask: | |
11462 | case CODE_FOR_avx512f_vmcmpv2df3_mask: | |
11463 | case CODE_FOR_avx512f_vmcmpv4sf3_mask: | |
0f200733 | 11464 | case CODE_FOR_avx512bw_cmpv32hf3_mask: |
11465 | case CODE_FOR_avx512vl_cmpv16hf3_mask: | |
11466 | case CODE_FOR_avx512fp16_cmpv8hf3_mask: | |
2bf6d935 ML |
11467 | error ("the last argument must be a 5-bit immediate"); |
11468 | return const0_rtx; | |
11469 | ||
11470 | default: | |
11471 | switch (nargs_constant) | |
11472 | { | |
11473 | case 2: | |
11474 | if ((mask_pos && (nargs - i - mask_pos) == nargs_constant) || | |
11475 | (!mask_pos && (nargs - i) == nargs_constant)) | |
11476 | { | |
11477 | error ("the next to last argument must be an 8-bit immediate"); | |
11478 | break; | |
11479 | } | |
11480 | /* FALLTHRU */ | |
11481 | case 1: | |
11482 | error ("the last argument must be an 8-bit immediate"); | |
11483 | break; | |
11484 | default: | |
11485 | gcc_unreachable (); | |
11486 | } | |
11487 | return const0_rtx; | |
11488 | } | |
11489 | } | |
11490 | else | |
11491 | { | |
11492 | if (VECTOR_MODE_P (mode)) | |
11493 | op = safe_vector_operand (op, mode); | |
11494 | ||
11495 | /* If we aren't optimizing, only allow one memory operand to | |
11496 | be generated. */ | |
11497 | if (memory_operand (op, mode)) | |
11498 | num_memory++; | |
11499 | ||
11500 | op = fixup_modeless_constant (op, mode); | |
11501 | ||
11502 | if (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode) | |
11503 | { | |
11504 | if (optimize || !match || num_memory > 1) | |
11505 | op = copy_to_mode_reg (mode, op); | |
11506 | } | |
11507 | else | |
11508 | { | |
11509 | op = copy_to_reg (op); | |
11510 | op = lowpart_subreg (mode, op, GET_MODE (op)); | |
11511 | } | |
11512 | } | |
11513 | ||
715a8bc8 | 11514 | xops[i] = op; |
2bf6d935 ML |
11515 | } |
11516 | ||
11517 | switch (nargs) | |
11518 | { | |
11519 | case 1: | |
715a8bc8 | 11520 | pat = GEN_FCN (icode) (real_target, xops[0]); |
2bf6d935 ML |
11521 | break; |
11522 | case 2: | |
715a8bc8 | 11523 | pat = GEN_FCN (icode) (real_target, xops[0], xops[1]); |
2bf6d935 ML |
11524 | break; |
11525 | case 3: | |
715a8bc8 | 11526 | pat = GEN_FCN (icode) (real_target, xops[0], xops[1], xops[2]); |
2bf6d935 ML |
11527 | break; |
11528 | case 4: | |
715a8bc8 UB |
11529 | pat = GEN_FCN (icode) (real_target, xops[0], xops[1], |
11530 | xops[2], xops[3]); | |
2bf6d935 ML |
11531 | break; |
11532 | case 5: | |
715a8bc8 UB |
11533 | pat = GEN_FCN (icode) (real_target, xops[0], xops[1], |
11534 | xops[2], xops[3], xops[4]); | |
2bf6d935 ML |
11535 | break; |
11536 | case 6: | |
715a8bc8 UB |
11537 | pat = GEN_FCN (icode) (real_target, xops[0], xops[1], |
11538 | xops[2], xops[3], xops[4], xops[5]); | |
2bf6d935 ML |
11539 | break; |
11540 | default: | |
11541 | gcc_unreachable (); | |
11542 | } | |
11543 | ||
11544 | if (! pat) | |
11545 | return 0; | |
11546 | ||
11547 | emit_insn (pat); | |
11548 | return target; | |
11549 | } | |
11550 | ||
11551 | /* Transform pattern of following layout: | |
11552 | (set A | |
11553 | (unspec [B C] UNSPEC_EMBEDDED_ROUNDING)) | |
11554 | ) | |
11555 | into: | |
11556 | (set (A B)) */ | |
11557 | ||
11558 | static rtx | |
11559 | ix86_erase_embedded_rounding (rtx pat) | |
11560 | { | |
11561 | if (GET_CODE (pat) == INSN) | |
11562 | pat = PATTERN (pat); | |
11563 | ||
11564 | gcc_assert (GET_CODE (pat) == SET); | |
11565 | rtx src = SET_SRC (pat); | |
11566 | gcc_assert (XVECLEN (src, 0) == 2); | |
11567 | rtx p0 = XVECEXP (src, 0, 0); | |
11568 | gcc_assert (GET_CODE (src) == UNSPEC | |
11569 | && XINT (src, 1) == UNSPEC_EMBEDDED_ROUNDING); | |
11570 | rtx res = gen_rtx_SET (SET_DEST (pat), p0); | |
11571 | return res; | |
11572 | } | |
11573 | ||
11574 | /* Subroutine of ix86_expand_round_builtin to take care of comi insns | |
11575 | with rounding. */ | |
11576 | static rtx | |
11577 | ix86_expand_sse_comi_round (const struct builtin_description *d, | |
11578 | tree exp, rtx target) | |
11579 | { | |
11580 | rtx pat, set_dst; | |
11581 | tree arg0 = CALL_EXPR_ARG (exp, 0); | |
11582 | tree arg1 = CALL_EXPR_ARG (exp, 1); | |
11583 | tree arg2 = CALL_EXPR_ARG (exp, 2); | |
11584 | tree arg3 = CALL_EXPR_ARG (exp, 3); | |
11585 | rtx op0 = expand_normal (arg0); | |
11586 | rtx op1 = expand_normal (arg1); | |
11587 | rtx op2 = expand_normal (arg2); | |
11588 | rtx op3 = expand_normal (arg3); | |
11589 | enum insn_code icode = d->icode; | |
11590 | const struct insn_data_d *insn_p = &insn_data[icode]; | |
11591 | machine_mode mode0 = insn_p->operand[0].mode; | |
11592 | machine_mode mode1 = insn_p->operand[1].mode; | |
2bf6d935 ML |
11593 | |
11594 | /* See avxintrin.h for values. */ | |
467e9f38 | 11595 | static const enum rtx_code comparisons[32] = |
2bf6d935 | 11596 | { |
467e9f38 L |
11597 | EQ, LT, LE, UNORDERED, NE, UNGE, UNGT, ORDERED, |
11598 | UNEQ, UNLT, UNLE, UNORDERED, LTGT, GE, GT, ORDERED, | |
11599 | EQ, LT, LE, UNORDERED, NE, UNGE, UNGT, ORDERED, | |
11600 | UNEQ, UNLT, UNLE, UNORDERED, LTGT, GE, GT, ORDERED | |
2bf6d935 | 11601 | }; |
467e9f38 L |
11602 | static const bool ordereds[32] = |
11603 | { | |
11604 | true, true, true, false, false, false, false, true, | |
11605 | false, false, false, true, true, true, true, false, | |
11606 | true, true, true, false, false, false, false, true, | |
11607 | false, false, false, true, true, true, true, false | |
11608 | }; | |
11609 | static const bool non_signalings[32] = | |
2bf6d935 ML |
11610 | { |
11611 | true, false, false, true, true, false, false, true, | |
11612 | true, false, false, true, true, false, false, true, | |
11613 | false, true, true, false, false, true, true, false, | |
11614 | false, true, true, false, false, true, true, false | |
11615 | }; | |
11616 | ||
11617 | if (!CONST_INT_P (op2)) | |
11618 | { | |
11619 | error ("the third argument must be comparison constant"); | |
11620 | return const0_rtx; | |
11621 | } | |
11622 | if (INTVAL (op2) < 0 || INTVAL (op2) >= 32) | |
11623 | { | |
11624 | error ("incorrect comparison mode"); | |
11625 | return const0_rtx; | |
11626 | } | |
11627 | ||
11628 | if (!insn_p->operand[2].predicate (op3, SImode)) | |
11629 | { | |
11630 | error ("incorrect rounding operand"); | |
11631 | return const0_rtx; | |
11632 | } | |
11633 | ||
2bf6d935 ML |
11634 | if (VECTOR_MODE_P (mode0)) |
11635 | op0 = safe_vector_operand (op0, mode0); | |
11636 | if (VECTOR_MODE_P (mode1)) | |
11637 | op1 = safe_vector_operand (op1, mode1); | |
11638 | ||
467e9f38 L |
11639 | enum rtx_code comparison = comparisons[INTVAL (op2)]; |
11640 | bool ordered = ordereds[INTVAL (op2)]; | |
11641 | bool non_signaling = non_signalings[INTVAL (op2)]; | |
11642 | rtx const_val = const0_rtx; | |
11643 | ||
11644 | bool check_unordered = false; | |
11645 | machine_mode mode = CCFPmode; | |
11646 | switch (comparison) | |
11647 | { | |
11648 | case ORDERED: | |
11649 | if (!ordered) | |
11650 | { | |
11651 | /* NB: Use CCSmode/NE for _CMP_TRUE_UQ/_CMP_TRUE_US. */ | |
11652 | if (!non_signaling) | |
11653 | ordered = true; | |
11654 | mode = CCSmode; | |
11655 | } | |
11656 | else | |
11657 | { | |
11658 | /* NB: Use CCPmode/NE for _CMP_ORD_Q/_CMP_ORD_S. */ | |
11659 | if (non_signaling) | |
11660 | ordered = false; | |
11661 | mode = CCPmode; | |
11662 | } | |
11663 | comparison = NE; | |
11664 | break; | |
11665 | case UNORDERED: | |
11666 | if (ordered) | |
11667 | { | |
11668 | /* NB: Use CCSmode/EQ for _CMP_FALSE_OQ/_CMP_FALSE_OS. */ | |
11669 | if (non_signaling) | |
11670 | ordered = false; | |
11671 | mode = CCSmode; | |
11672 | } | |
11673 | else | |
11674 | { | |
11675 | /* NB: Use CCPmode/NE for _CMP_UNORD_Q/_CMP_UNORD_S. */ | |
11676 | if (!non_signaling) | |
11677 | ordered = true; | |
11678 | mode = CCPmode; | |
11679 | } | |
11680 | comparison = EQ; | |
11681 | break; | |
11682 | ||
11683 | case LE: /* -> GE */ | |
11684 | case LT: /* -> GT */ | |
11685 | case UNGE: /* -> UNLE */ | |
11686 | case UNGT: /* -> UNLT */ | |
11687 | std::swap (op0, op1); | |
11688 | comparison = swap_condition (comparison); | |
11689 | /* FALLTHRU */ | |
11690 | case GT: | |
11691 | case GE: | |
11692 | case UNEQ: | |
11693 | case UNLT: | |
11694 | case UNLE: | |
11695 | case LTGT: | |
11696 | /* These are supported by CCFPmode. NB: Use ordered/signaling | |
11697 | COMI or unordered/non-signaling UCOMI. Both set ZF, PF, CF | |
11698 | with NAN operands. */ | |
11699 | if (ordered == non_signaling) | |
11700 | ordered = !ordered; | |
11701 | break; | |
11702 | case EQ: | |
11703 | /* NB: COMI/UCOMI will set ZF with NAN operands. Use CCZmode for | |
11704 | _CMP_EQ_OQ/_CMP_EQ_OS. */ | |
11705 | check_unordered = true; | |
11706 | mode = CCZmode; | |
11707 | break; | |
11708 | case NE: | |
11709 | /* NB: COMI/UCOMI will set ZF with NAN operands. Use CCZmode for | |
11710 | _CMP_NEQ_UQ/_CMP_NEQ_US. */ | |
11711 | gcc_assert (!ordered); | |
11712 | check_unordered = true; | |
11713 | mode = CCZmode; | |
11714 | const_val = const1_rtx; | |
11715 | break; | |
11716 | default: | |
11717 | gcc_unreachable (); | |
11718 | } | |
11719 | ||
2bf6d935 | 11720 | target = gen_reg_rtx (SImode); |
467e9f38 | 11721 | emit_move_insn (target, const_val); |
2bf6d935 ML |
11722 | target = gen_rtx_SUBREG (QImode, target, 0); |
11723 | ||
11724 | if ((optimize && !register_operand (op0, mode0)) | |
11725 | || !insn_p->operand[0].predicate (op0, mode0)) | |
11726 | op0 = copy_to_mode_reg (mode0, op0); | |
11727 | if ((optimize && !register_operand (op1, mode1)) | |
11728 | || !insn_p->operand[1].predicate (op1, mode1)) | |
11729 | op1 = copy_to_mode_reg (mode1, op1); | |
11730 | ||
467e9f38 L |
11731 | /* |
11732 | 1. COMI: ordered and signaling. | |
11733 | 2. UCOMI: unordered and non-signaling. | |
11734 | */ | |
11735 | if (non_signaling) | |
11736 | icode = (icode == CODE_FOR_sse_comi_round | |
11737 | ? CODE_FOR_sse_ucomi_round | |
11738 | : CODE_FOR_sse2_ucomi_round); | |
2bf6d935 ML |
11739 | |
11740 | pat = GEN_FCN (icode) (op0, op1, op3); | |
11741 | if (! pat) | |
11742 | return 0; | |
11743 | ||
11744 | /* Rounding operand can be either NO_ROUND or ROUND_SAE at this point. */ | |
11745 | if (INTVAL (op3) == NO_ROUND) | |
11746 | { | |
11747 | pat = ix86_erase_embedded_rounding (pat); | |
11748 | if (! pat) | |
11749 | return 0; | |
11750 | ||
11751 | set_dst = SET_DEST (pat); | |
11752 | } | |
11753 | else | |
11754 | { | |
11755 | gcc_assert (GET_CODE (pat) == SET); | |
11756 | set_dst = SET_DEST (pat); | |
11757 | } | |
11758 | ||
11759 | emit_insn (pat); | |
467e9f38 | 11760 | |
ae69e6f6 | 11761 | return ix86_ssecom_setcc (comparison, check_unordered, mode, |
11762 | set_dst, target); | |
2bf6d935 ML |
11763 | } |
11764 | ||
11765 | static rtx | |
11766 | ix86_expand_round_builtin (const struct builtin_description *d, | |
11767 | tree exp, rtx target) | |
11768 | { | |
11769 | rtx pat; | |
11770 | unsigned int i, nargs; | |
715a8bc8 | 11771 | rtx xops[6]; |
2bf6d935 ML |
11772 | enum insn_code icode = d->icode; |
11773 | const struct insn_data_d *insn_p = &insn_data[icode]; | |
11774 | machine_mode tmode = insn_p->operand[0].mode; | |
11775 | unsigned int nargs_constant = 0; | |
11776 | unsigned int redundant_embed_rnd = 0; | |
11777 | ||
11778 | switch ((enum ix86_builtin_func_type) d->flag) | |
11779 | { | |
11780 | case UINT64_FTYPE_V2DF_INT: | |
11781 | case UINT64_FTYPE_V4SF_INT: | |
3069a2e5 | 11782 | case UINT64_FTYPE_V8HF_INT: |
2bf6d935 ML |
11783 | case UINT_FTYPE_V2DF_INT: |
11784 | case UINT_FTYPE_V4SF_INT: | |
3069a2e5 | 11785 | case UINT_FTYPE_V8HF_INT: |
2bf6d935 ML |
11786 | case INT64_FTYPE_V2DF_INT: |
11787 | case INT64_FTYPE_V4SF_INT: | |
3069a2e5 | 11788 | case INT64_FTYPE_V8HF_INT: |
2bf6d935 ML |
11789 | case INT_FTYPE_V2DF_INT: |
11790 | case INT_FTYPE_V4SF_INT: | |
3069a2e5 | 11791 | case INT_FTYPE_V8HF_INT: |
2bf6d935 ML |
11792 | nargs = 2; |
11793 | break; | |
bd7a34ef | 11794 | case V32HF_FTYPE_V32HF_V32HF_INT: |
71838266 | 11795 | case V8HF_FTYPE_V8HF_V8HF_INT: |
3069a2e5 | 11796 | case V8HF_FTYPE_V8HF_INT_INT: |
11797 | case V8HF_FTYPE_V8HF_UINT_INT: | |
11798 | case V8HF_FTYPE_V8HF_INT64_INT: | |
11799 | case V8HF_FTYPE_V8HF_UINT64_INT: | |
2bf6d935 ML |
11800 | case V4SF_FTYPE_V4SF_UINT_INT: |
11801 | case V4SF_FTYPE_V4SF_UINT64_INT: | |
11802 | case V2DF_FTYPE_V2DF_UINT64_INT: | |
11803 | case V4SF_FTYPE_V4SF_INT_INT: | |
11804 | case V4SF_FTYPE_V4SF_INT64_INT: | |
11805 | case V2DF_FTYPE_V2DF_INT64_INT: | |
11806 | case V4SF_FTYPE_V4SF_V4SF_INT: | |
11807 | case V2DF_FTYPE_V2DF_V2DF_INT: | |
11808 | case V4SF_FTYPE_V4SF_V2DF_INT: | |
11809 | case V2DF_FTYPE_V2DF_V4SF_INT: | |
11810 | nargs = 3; | |
11811 | break; | |
11812 | case V8SF_FTYPE_V8DF_V8SF_QI_INT: | |
11813 | case V8DF_FTYPE_V8DF_V8DF_QI_INT: | |
bd610db0 | 11814 | case V32HI_FTYPE_V32HF_V32HI_USI_INT: |
2bf6d935 | 11815 | case V8SI_FTYPE_V8DF_V8SI_QI_INT: |
bd610db0 | 11816 | case V8DI_FTYPE_V8HF_V8DI_UQI_INT: |
2bf6d935 ML |
11817 | case V8DI_FTYPE_V8DF_V8DI_QI_INT: |
11818 | case V8SF_FTYPE_V8DI_V8SF_QI_INT: | |
11819 | case V8DF_FTYPE_V8DI_V8DF_QI_INT: | |
5a744e50 | 11820 | case V8DF_FTYPE_V8HF_V8DF_UQI_INT: |
11821 | case V16SF_FTYPE_V16HF_V16SF_UHI_INT: | |
be0e4c32 | 11822 | case V32HF_FTYPE_V32HI_V32HF_USI_INT: |
4204740f | 11823 | case V32HF_FTYPE_V32HF_V32HF_USI_INT: |
081070bc | 11824 | case V32HF_FTYPE_V32HF_V32HF_V32HF_INT: |
2bf6d935 ML |
11825 | case V16SF_FTYPE_V16SF_V16SF_HI_INT: |
11826 | case V8DI_FTYPE_V8SF_V8DI_QI_INT: | |
11827 | case V16SF_FTYPE_V16SI_V16SF_HI_INT: | |
11828 | case V16SI_FTYPE_V16SF_V16SI_HI_INT: | |
bd610db0 | 11829 | case V16SI_FTYPE_V16HF_V16SI_UHI_INT: |
be0e4c32 | 11830 | case V16HF_FTYPE_V16SI_V16HF_UHI_INT: |
2bf6d935 ML |
11831 | case V8DF_FTYPE_V8SF_V8DF_QI_INT: |
11832 | case V16SF_FTYPE_V16HI_V16SF_HI_INT: | |
11833 | case V2DF_FTYPE_V2DF_V2DF_V2DF_INT: | |
11834 | case V4SF_FTYPE_V4SF_V4SF_V4SF_INT: | |
be0e4c32 | 11835 | case V8HF_FTYPE_V8DI_V8HF_UQI_INT: |
5a744e50 | 11836 | case V8HF_FTYPE_V8DF_V8HF_UQI_INT: |
11837 | case V16HF_FTYPE_V16SF_V16HF_UHI_INT: | |
3c9de0a9 | 11838 | case V8HF_FTYPE_V8HF_V8HF_V8HF_INT: |
2bf6d935 ML |
11839 | nargs = 4; |
11840 | break; | |
11841 | case V4SF_FTYPE_V4SF_V4SF_INT_INT: | |
11842 | case V2DF_FTYPE_V2DF_V2DF_INT_INT: | |
11843 | nargs_constant = 2; | |
11844 | nargs = 4; | |
11845 | break; | |
11846 | case INT_FTYPE_V4SF_V4SF_INT_INT: | |
11847 | case INT_FTYPE_V2DF_V2DF_INT_INT: | |
11848 | return ix86_expand_sse_comi_round (d, exp, target); | |
11849 | case V8DF_FTYPE_V8DF_V8DF_V8DF_UQI_INT: | |
11850 | case V2DF_FTYPE_V2DF_V2DF_V2DF_UQI_INT: | |
11851 | case V4SF_FTYPE_V4SF_V4SF_V4SF_UQI_INT: | |
90429b96 | 11852 | case V4SF_FTYPE_V8HF_V4SF_V4SF_UQI_INT: |
2bf6d935 | 11853 | case V16SF_FTYPE_V16SF_V16SF_V16SF_HI_INT: |
081070bc | 11854 | case V32HF_FTYPE_V32HF_V32HF_V32HF_UHI_INT: |
bd7a34ef | 11855 | case V32HF_FTYPE_V32HF_V32HF_V32HF_USI_INT: |
90429b96 | 11856 | case V2DF_FTYPE_V8HF_V2DF_V2DF_UQI_INT: |
2bf6d935 ML |
11857 | case V2DF_FTYPE_V2DF_V2DF_V2DF_QI_INT: |
11858 | case V2DF_FTYPE_V2DF_V4SF_V2DF_QI_INT: | |
93103603 | 11859 | case V2DF_FTYPE_V2DF_V4SF_V2DF_UQI_INT: |
2bf6d935 ML |
11860 | case V4SF_FTYPE_V4SF_V4SF_V4SF_QI_INT: |
11861 | case V4SF_FTYPE_V4SF_V2DF_V4SF_QI_INT: | |
93103603 | 11862 | case V4SF_FTYPE_V4SF_V2DF_V4SF_UQI_INT: |
71838266 | 11863 | case V8HF_FTYPE_V8HF_V8HF_V8HF_UQI_INT: |
90429b96 | 11864 | case V8HF_FTYPE_V2DF_V8HF_V8HF_UQI_INT: |
11865 | case V8HF_FTYPE_V4SF_V8HF_V8HF_UQI_INT: | |
2bf6d935 ML |
11866 | nargs = 5; |
11867 | break; | |
8bed7617 | 11868 | case V32HF_FTYPE_V32HF_INT_V32HF_USI_INT: |
2bf6d935 ML |
11869 | case V16SF_FTYPE_V16SF_INT_V16SF_HI_INT: |
11870 | case V8DF_FTYPE_V8DF_INT_V8DF_QI_INT: | |
93103603 SP |
11871 | case V8DF_FTYPE_V8DF_INT_V8DF_UQI_INT: |
11872 | case V16SF_FTYPE_V16SF_INT_V16SF_UHI_INT: | |
2bf6d935 ML |
11873 | nargs_constant = 4; |
11874 | nargs = 5; | |
11875 | break; | |
11876 | case UQI_FTYPE_V8DF_V8DF_INT_UQI_INT: | |
11877 | case UQI_FTYPE_V2DF_V2DF_INT_UQI_INT: | |
11878 | case UHI_FTYPE_V16SF_V16SF_INT_UHI_INT: | |
11879 | case UQI_FTYPE_V4SF_V4SF_INT_UQI_INT: | |
0f200733 | 11880 | case USI_FTYPE_V32HF_V32HF_INT_USI_INT: |
11881 | case UQI_FTYPE_V8HF_V8HF_INT_UQI_INT: | |
2bf6d935 ML |
11882 | nargs_constant = 3; |
11883 | nargs = 5; | |
11884 | break; | |
11885 | case V16SF_FTYPE_V16SF_V16SF_INT_V16SF_HI_INT: | |
11886 | case V8DF_FTYPE_V8DF_V8DF_INT_V8DF_QI_INT: | |
11887 | case V4SF_FTYPE_V4SF_V4SF_INT_V4SF_QI_INT: | |
11888 | case V2DF_FTYPE_V2DF_V2DF_INT_V2DF_QI_INT: | |
11889 | case V2DF_FTYPE_V2DF_V2DF_INT_V2DF_UQI_INT: | |
11890 | case V4SF_FTYPE_V4SF_V4SF_INT_V4SF_UQI_INT: | |
8bed7617 | 11891 | case V8HF_FTYPE_V8HF_V8HF_INT_V8HF_UQI_INT: |
2bf6d935 ML |
11892 | nargs = 6; |
11893 | nargs_constant = 4; | |
11894 | break; | |
11895 | case V8DF_FTYPE_V8DF_V8DF_V8DI_INT_QI_INT: | |
11896 | case V16SF_FTYPE_V16SF_V16SF_V16SI_INT_HI_INT: | |
11897 | case V2DF_FTYPE_V2DF_V2DF_V2DI_INT_QI_INT: | |
11898 | case V4SF_FTYPE_V4SF_V4SF_V4SI_INT_QI_INT: | |
11899 | nargs = 6; | |
11900 | nargs_constant = 3; | |
11901 | break; | |
11902 | default: | |
11903 | gcc_unreachable (); | |
11904 | } | |
715a8bc8 | 11905 | gcc_assert (nargs <= ARRAY_SIZE (xops)); |
2bf6d935 ML |
11906 | |
11907 | if (optimize | |
11908 | || target == 0 | |
11909 | || GET_MODE (target) != tmode | |
11910 | || !insn_p->operand[0].predicate (target, tmode)) | |
11911 | target = gen_reg_rtx (tmode); | |
11912 | ||
11913 | for (i = 0; i < nargs; i++) | |
11914 | { | |
11915 | tree arg = CALL_EXPR_ARG (exp, i); | |
11916 | rtx op = expand_normal (arg); | |
11917 | machine_mode mode = insn_p->operand[i + 1].mode; | |
11918 | bool match = insn_p->operand[i + 1].predicate (op, mode); | |
11919 | ||
11920 | if (i == nargs - nargs_constant) | |
11921 | { | |
11922 | if (!match) | |
11923 | { | |
11924 | switch (icode) | |
11925 | { | |
11926 | case CODE_FOR_avx512f_getmantv8df_mask_round: | |
11927 | case CODE_FOR_avx512f_getmantv16sf_mask_round: | |
8486e9f2 | 11928 | case CODE_FOR_avx512bw_getmantv32hf_mask_round: |
2bf6d935 ML |
11929 | case CODE_FOR_avx512f_vgetmantv2df_round: |
11930 | case CODE_FOR_avx512f_vgetmantv2df_mask_round: | |
11931 | case CODE_FOR_avx512f_vgetmantv4sf_round: | |
11932 | case CODE_FOR_avx512f_vgetmantv4sf_mask_round: | |
8486e9f2 | 11933 | case CODE_FOR_avx512f_vgetmantv8hf_mask_round: |
2bf6d935 ML |
11934 | error ("the immediate argument must be a 4-bit immediate"); |
11935 | return const0_rtx; | |
11936 | case CODE_FOR_avx512f_cmpv8df3_mask_round: | |
11937 | case CODE_FOR_avx512f_cmpv16sf3_mask_round: | |
11938 | case CODE_FOR_avx512f_vmcmpv2df3_mask_round: | |
11939 | case CODE_FOR_avx512f_vmcmpv4sf3_mask_round: | |
0f200733 | 11940 | case CODE_FOR_avx512f_vmcmpv8hf3_mask_round: |
11941 | case CODE_FOR_avx512bw_cmpv32hf3_mask_round: | |
2bf6d935 ML |
11942 | error ("the immediate argument must be a 5-bit immediate"); |
11943 | return const0_rtx; | |
11944 | default: | |
11945 | error ("the immediate argument must be an 8-bit immediate"); | |
11946 | return const0_rtx; | |
11947 | } | |
11948 | } | |
11949 | } | |
11950 | else if (i == nargs-1) | |
11951 | { | |
11952 | if (!insn_p->operand[nargs].predicate (op, SImode)) | |
11953 | { | |
11954 | error ("incorrect rounding operand"); | |
11955 | return const0_rtx; | |
11956 | } | |
11957 | ||
11958 | /* If there is no rounding use normal version of the pattern. */ | |
11959 | if (INTVAL (op) == NO_ROUND) | |
2f9529fc HW |
11960 | { |
11961 | /* Skip erasing embedded rounding for below expanders who | |
11962 | generates multiple insns. In ix86_erase_embedded_rounding | |
11963 | the pattern will be transformed to a single set, and emit_insn | |
11964 | appends the set insead of insert it to chain. So the insns | |
11965 | emitted inside define_expander would be ignored. */ | |
11966 | switch (icode) | |
11967 | { | |
11968 | case CODE_FOR_avx512bw_fmaddc_v32hf_mask1_round: | |
11969 | case CODE_FOR_avx512bw_fcmaddc_v32hf_mask1_round: | |
11970 | case CODE_FOR_avx512fp16_fmaddcsh_v8hf_mask1_round: | |
11971 | case CODE_FOR_avx512fp16_fcmaddcsh_v8hf_mask1_round: | |
11972 | case CODE_FOR_avx512fp16_fmaddcsh_v8hf_mask3_round: | |
11973 | case CODE_FOR_avx512fp16_fcmaddcsh_v8hf_mask3_round: | |
11974 | redundant_embed_rnd = 0; | |
11975 | break; | |
11976 | default: | |
11977 | redundant_embed_rnd = 1; | |
11978 | break; | |
11979 | } | |
11980 | } | |
2bf6d935 ML |
11981 | } |
11982 | else | |
11983 | { | |
11984 | if (VECTOR_MODE_P (mode)) | |
11985 | op = safe_vector_operand (op, mode); | |
11986 | ||
11987 | op = fixup_modeless_constant (op, mode); | |
11988 | ||
11989 | if (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode) | |
11990 | { | |
11991 | if (optimize || !match) | |
11992 | op = copy_to_mode_reg (mode, op); | |
11993 | } | |
11994 | else | |
11995 | { | |
11996 | op = copy_to_reg (op); | |
11997 | op = lowpart_subreg (mode, op, GET_MODE (op)); | |
11998 | } | |
11999 | } | |
12000 | ||
715a8bc8 | 12001 | xops[i] = op; |
2bf6d935 ML |
12002 | } |
12003 | ||
12004 | switch (nargs) | |
12005 | { | |
12006 | case 1: | |
715a8bc8 | 12007 | pat = GEN_FCN (icode) (target, xops[0]); |
2bf6d935 ML |
12008 | break; |
12009 | case 2: | |
715a8bc8 | 12010 | pat = GEN_FCN (icode) (target, xops[0], xops[1]); |
2bf6d935 ML |
12011 | break; |
12012 | case 3: | |
715a8bc8 | 12013 | pat = GEN_FCN (icode) (target, xops[0], xops[1], xops[2]); |
2bf6d935 ML |
12014 | break; |
12015 | case 4: | |
715a8bc8 UB |
12016 | pat = GEN_FCN (icode) (target, xops[0], xops[1], |
12017 | xops[2], xops[3]); | |
2bf6d935 ML |
12018 | break; |
12019 | case 5: | |
715a8bc8 UB |
12020 | pat = GEN_FCN (icode) (target, xops[0], xops[1], |
12021 | xops[2], xops[3], xops[4]); | |
2bf6d935 ML |
12022 | break; |
12023 | case 6: | |
715a8bc8 UB |
12024 | pat = GEN_FCN (icode) (target, xops[0], xops[1], |
12025 | xops[2], xops[3], xops[4], xops[5]); | |
2bf6d935 ML |
12026 | break; |
12027 | default: | |
12028 | gcc_unreachable (); | |
12029 | } | |
12030 | ||
12031 | if (!pat) | |
12032 | return 0; | |
12033 | ||
12034 | if (redundant_embed_rnd) | |
12035 | pat = ix86_erase_embedded_rounding (pat); | |
12036 | ||
12037 | emit_insn (pat); | |
12038 | return target; | |
12039 | } | |
12040 | ||
12041 | /* Subroutine of ix86_expand_builtin to take care of special insns | |
12042 | with variable number of operands. */ | |
12043 | ||
12044 | static rtx | |
12045 | ix86_expand_special_args_builtin (const struct builtin_description *d, | |
12046 | tree exp, rtx target) | |
12047 | { | |
12048 | tree arg; | |
12049 | rtx pat, op; | |
12050 | unsigned int i, nargs, arg_adjust, memory; | |
152834fe | 12051 | unsigned int constant = 100; |
2bf6d935 | 12052 | bool aligned_mem = false; |
152834fe | 12053 | rtx xops[4]; |
2bf6d935 | 12054 | enum insn_code icode = d->icode; |
2bf6d935 ML |
12055 | const struct insn_data_d *insn_p = &insn_data[icode]; |
12056 | machine_mode tmode = insn_p->operand[0].mode; | |
12057 | enum { load, store } klass; | |
12058 | ||
12059 | switch ((enum ix86_builtin_func_type) d->flag) | |
12060 | { | |
12061 | case VOID_FTYPE_VOID: | |
12062 | emit_insn (GEN_FCN (icode) (target)); | |
12063 | return 0; | |
12064 | case VOID_FTYPE_UINT64: | |
12065 | case VOID_FTYPE_UNSIGNED: | |
12066 | nargs = 0; | |
12067 | klass = store; | |
12068 | memory = 0; | |
12069 | break; | |
12070 | ||
12071 | case INT_FTYPE_VOID: | |
12072 | case USHORT_FTYPE_VOID: | |
12073 | case UINT64_FTYPE_VOID: | |
12074 | case UINT_FTYPE_VOID: | |
299a53d7 | 12075 | case UINT8_FTYPE_VOID: |
2bf6d935 ML |
12076 | case UNSIGNED_FTYPE_VOID: |
12077 | nargs = 0; | |
12078 | klass = load; | |
12079 | memory = 0; | |
12080 | break; | |
12081 | case UINT64_FTYPE_PUNSIGNED: | |
12082 | case V2DI_FTYPE_PV2DI: | |
12083 | case V4DI_FTYPE_PV4DI: | |
12084 | case V32QI_FTYPE_PCCHAR: | |
12085 | case V16QI_FTYPE_PCCHAR: | |
12086 | case V8SF_FTYPE_PCV4SF: | |
12087 | case V8SF_FTYPE_PCFLOAT: | |
12088 | case V4SF_FTYPE_PCFLOAT: | |
58685b93 | 12089 | case V4SF_FTYPE_PCFLOAT16: |
12090 | case V4SF_FTYPE_PCBFLOAT16: | |
12091 | case V4SF_FTYPE_PCV8BF: | |
12092 | case V4SF_FTYPE_PCV8HF: | |
12093 | case V8SF_FTYPE_PCFLOAT16: | |
12094 | case V8SF_FTYPE_PCBFLOAT16: | |
12095 | case V8SF_FTYPE_PCV16HF: | |
12096 | case V8SF_FTYPE_PCV16BF: | |
2bf6d935 ML |
12097 | case V4DF_FTYPE_PCV2DF: |
12098 | case V4DF_FTYPE_PCDOUBLE: | |
12099 | case V2DF_FTYPE_PCDOUBLE: | |
12100 | case VOID_FTYPE_PVOID: | |
12101 | case V8DI_FTYPE_PV8DI: | |
12102 | nargs = 1; | |
12103 | klass = load; | |
12104 | memory = 0; | |
12105 | switch (icode) | |
12106 | { | |
12107 | case CODE_FOR_sse4_1_movntdqa: | |
12108 | case CODE_FOR_avx2_movntdqa: | |
12109 | case CODE_FOR_avx512f_movntdqa: | |
12110 | aligned_mem = true; | |
12111 | break; | |
12112 | default: | |
12113 | break; | |
12114 | } | |
12115 | break; | |
12116 | case VOID_FTYPE_PV2SF_V4SF: | |
12117 | case VOID_FTYPE_PV8DI_V8DI: | |
12118 | case VOID_FTYPE_PV4DI_V4DI: | |
12119 | case VOID_FTYPE_PV2DI_V2DI: | |
12120 | case VOID_FTYPE_PCHAR_V32QI: | |
12121 | case VOID_FTYPE_PCHAR_V16QI: | |
12122 | case VOID_FTYPE_PFLOAT_V16SF: | |
12123 | case VOID_FTYPE_PFLOAT_V8SF: | |
12124 | case VOID_FTYPE_PFLOAT_V4SF: | |
12125 | case VOID_FTYPE_PDOUBLE_V8DF: | |
12126 | case VOID_FTYPE_PDOUBLE_V4DF: | |
12127 | case VOID_FTYPE_PDOUBLE_V2DF: | |
12128 | case VOID_FTYPE_PLONGLONG_LONGLONG: | |
12129 | case VOID_FTYPE_PULONGLONG_ULONGLONG: | |
12130 | case VOID_FTYPE_PUNSIGNED_UNSIGNED: | |
12131 | case VOID_FTYPE_PINT_INT: | |
12132 | nargs = 1; | |
12133 | klass = store; | |
12134 | /* Reserve memory operand for target. */ | |
715a8bc8 | 12135 | memory = ARRAY_SIZE (xops); |
2bf6d935 ML |
12136 | switch (icode) |
12137 | { | |
12138 | /* These builtins and instructions require the memory | |
12139 | to be properly aligned. */ | |
12140 | case CODE_FOR_avx_movntv4di: | |
12141 | case CODE_FOR_sse2_movntv2di: | |
12142 | case CODE_FOR_avx_movntv8sf: | |
12143 | case CODE_FOR_sse_movntv4sf: | |
12144 | case CODE_FOR_sse4a_vmmovntv4sf: | |
12145 | case CODE_FOR_avx_movntv4df: | |
12146 | case CODE_FOR_sse2_movntv2df: | |
12147 | case CODE_FOR_sse4a_vmmovntv2df: | |
12148 | case CODE_FOR_sse2_movntidi: | |
12149 | case CODE_FOR_sse_movntq: | |
12150 | case CODE_FOR_sse2_movntisi: | |
12151 | case CODE_FOR_avx512f_movntv16sf: | |
12152 | case CODE_FOR_avx512f_movntv8df: | |
12153 | case CODE_FOR_avx512f_movntv8di: | |
12154 | aligned_mem = true; | |
12155 | break; | |
12156 | default: | |
12157 | break; | |
12158 | } | |
12159 | break; | |
12160 | case VOID_FTYPE_PVOID_PCVOID: | |
12161 | nargs = 1; | |
12162 | klass = store; | |
12163 | memory = 0; | |
12164 | ||
12165 | break; | |
12166 | case V4SF_FTYPE_V4SF_PCV2SF: | |
12167 | case V2DF_FTYPE_V2DF_PCDOUBLE: | |
12168 | nargs = 2; | |
12169 | klass = load; | |
12170 | memory = 1; | |
12171 | break; | |
12172 | case V8SF_FTYPE_PCV8SF_V8SI: | |
12173 | case V4DF_FTYPE_PCV4DF_V4DI: | |
12174 | case V4SF_FTYPE_PCV4SF_V4SI: | |
12175 | case V2DF_FTYPE_PCV2DF_V2DI: | |
12176 | case V8SI_FTYPE_PCV8SI_V8SI: | |
12177 | case V4DI_FTYPE_PCV4DI_V4DI: | |
12178 | case V4SI_FTYPE_PCV4SI_V4SI: | |
12179 | case V2DI_FTYPE_PCV2DI_V2DI: | |
12180 | case VOID_FTYPE_INT_INT64: | |
12181 | nargs = 2; | |
12182 | klass = load; | |
12183 | memory = 0; | |
12184 | break; | |
12185 | case VOID_FTYPE_PV8DF_V8DF_UQI: | |
12186 | case VOID_FTYPE_PV4DF_V4DF_UQI: | |
12187 | case VOID_FTYPE_PV2DF_V2DF_UQI: | |
12188 | case VOID_FTYPE_PV16SF_V16SF_UHI: | |
12189 | case VOID_FTYPE_PV8SF_V8SF_UQI: | |
12190 | case VOID_FTYPE_PV4SF_V4SF_UQI: | |
12191 | case VOID_FTYPE_PV8DI_V8DI_UQI: | |
12192 | case VOID_FTYPE_PV4DI_V4DI_UQI: | |
12193 | case VOID_FTYPE_PV2DI_V2DI_UQI: | |
12194 | case VOID_FTYPE_PV16SI_V16SI_UHI: | |
12195 | case VOID_FTYPE_PV8SI_V8SI_UQI: | |
12196 | case VOID_FTYPE_PV4SI_V4SI_UQI: | |
12197 | case VOID_FTYPE_PV64QI_V64QI_UDI: | |
12198 | case VOID_FTYPE_PV32HI_V32HI_USI: | |
12199 | case VOID_FTYPE_PV32QI_V32QI_USI: | |
12200 | case VOID_FTYPE_PV16QI_V16QI_UHI: | |
12201 | case VOID_FTYPE_PV16HI_V16HI_UHI: | |
12202 | case VOID_FTYPE_PV8HI_V8HI_UQI: | |
12203 | switch (icode) | |
12204 | { | |
12205 | /* These builtins and instructions require the memory | |
12206 | to be properly aligned. */ | |
12207 | case CODE_FOR_avx512f_storev16sf_mask: | |
12208 | case CODE_FOR_avx512f_storev16si_mask: | |
12209 | case CODE_FOR_avx512f_storev8df_mask: | |
12210 | case CODE_FOR_avx512f_storev8di_mask: | |
12211 | case CODE_FOR_avx512vl_storev8sf_mask: | |
12212 | case CODE_FOR_avx512vl_storev8si_mask: | |
12213 | case CODE_FOR_avx512vl_storev4df_mask: | |
12214 | case CODE_FOR_avx512vl_storev4di_mask: | |
12215 | case CODE_FOR_avx512vl_storev4sf_mask: | |
12216 | case CODE_FOR_avx512vl_storev4si_mask: | |
12217 | case CODE_FOR_avx512vl_storev2df_mask: | |
12218 | case CODE_FOR_avx512vl_storev2di_mask: | |
12219 | aligned_mem = true; | |
12220 | break; | |
12221 | default: | |
12222 | break; | |
12223 | } | |
12224 | /* FALLTHRU */ | |
12225 | case VOID_FTYPE_PV8SF_V8SI_V8SF: | |
12226 | case VOID_FTYPE_PV4DF_V4DI_V4DF: | |
12227 | case VOID_FTYPE_PV4SF_V4SI_V4SF: | |
12228 | case VOID_FTYPE_PV2DF_V2DI_V2DF: | |
12229 | case VOID_FTYPE_PV8SI_V8SI_V8SI: | |
12230 | case VOID_FTYPE_PV4DI_V4DI_V4DI: | |
12231 | case VOID_FTYPE_PV4SI_V4SI_V4SI: | |
12232 | case VOID_FTYPE_PV2DI_V2DI_V2DI: | |
12233 | case VOID_FTYPE_PV8SI_V8DI_UQI: | |
12234 | case VOID_FTYPE_PV8HI_V8DI_UQI: | |
12235 | case VOID_FTYPE_PV16HI_V16SI_UHI: | |
4a948703 | 12236 | case VOID_FTYPE_PUDI_V8DI_UQI: |
2bf6d935 ML |
12237 | case VOID_FTYPE_PV16QI_V16SI_UHI: |
12238 | case VOID_FTYPE_PV4SI_V4DI_UQI: | |
4a948703 | 12239 | case VOID_FTYPE_PUDI_V2DI_UQI: |
12240 | case VOID_FTYPE_PUDI_V4DI_UQI: | |
12241 | case VOID_FTYPE_PUSI_V2DI_UQI: | |
2bf6d935 | 12242 | case VOID_FTYPE_PV8HI_V8SI_UQI: |
4a948703 | 12243 | case VOID_FTYPE_PUDI_V4SI_UQI: |
12244 | case VOID_FTYPE_PUSI_V4DI_UQI: | |
12245 | case VOID_FTYPE_PUHI_V2DI_UQI: | |
12246 | case VOID_FTYPE_PUDI_V8SI_UQI: | |
12247 | case VOID_FTYPE_PUSI_V4SI_UQI: | |
2bf6d935 ML |
12248 | case VOID_FTYPE_PCHAR_V64QI_UDI: |
12249 | case VOID_FTYPE_PCHAR_V32QI_USI: | |
12250 | case VOID_FTYPE_PCHAR_V16QI_UHI: | |
12251 | case VOID_FTYPE_PSHORT_V32HI_USI: | |
12252 | case VOID_FTYPE_PSHORT_V16HI_UHI: | |
12253 | case VOID_FTYPE_PSHORT_V8HI_UQI: | |
12254 | case VOID_FTYPE_PINT_V16SI_UHI: | |
12255 | case VOID_FTYPE_PINT_V8SI_UQI: | |
12256 | case VOID_FTYPE_PINT_V4SI_UQI: | |
12257 | case VOID_FTYPE_PINT64_V8DI_UQI: | |
12258 | case VOID_FTYPE_PINT64_V4DI_UQI: | |
12259 | case VOID_FTYPE_PINT64_V2DI_UQI: | |
12260 | case VOID_FTYPE_PDOUBLE_V8DF_UQI: | |
12261 | case VOID_FTYPE_PDOUBLE_V4DF_UQI: | |
12262 | case VOID_FTYPE_PDOUBLE_V2DF_UQI: | |
12263 | case VOID_FTYPE_PFLOAT_V16SF_UHI: | |
12264 | case VOID_FTYPE_PFLOAT_V8SF_UQI: | |
12265 | case VOID_FTYPE_PFLOAT_V4SF_UQI: | |
c4d423c7 | 12266 | case VOID_FTYPE_PCFLOAT16_V8HF_UQI: |
2bf6d935 ML |
12267 | case VOID_FTYPE_PV32QI_V32HI_USI: |
12268 | case VOID_FTYPE_PV16QI_V16HI_UHI: | |
4a948703 | 12269 | case VOID_FTYPE_PUDI_V8HI_UQI: |
2bf6d935 ML |
12270 | nargs = 2; |
12271 | klass = store; | |
12272 | /* Reserve memory operand for target. */ | |
715a8bc8 | 12273 | memory = ARRAY_SIZE (xops); |
2bf6d935 ML |
12274 | break; |
12275 | case V4SF_FTYPE_PCV4SF_V4SF_UQI: | |
12276 | case V8SF_FTYPE_PCV8SF_V8SF_UQI: | |
12277 | case V16SF_FTYPE_PCV16SF_V16SF_UHI: | |
12278 | case V4SI_FTYPE_PCV4SI_V4SI_UQI: | |
12279 | case V8SI_FTYPE_PCV8SI_V8SI_UQI: | |
12280 | case V16SI_FTYPE_PCV16SI_V16SI_UHI: | |
12281 | case V2DF_FTYPE_PCV2DF_V2DF_UQI: | |
12282 | case V4DF_FTYPE_PCV4DF_V4DF_UQI: | |
12283 | case V8DF_FTYPE_PCV8DF_V8DF_UQI: | |
12284 | case V2DI_FTYPE_PCV2DI_V2DI_UQI: | |
12285 | case V4DI_FTYPE_PCV4DI_V4DI_UQI: | |
12286 | case V8DI_FTYPE_PCV8DI_V8DI_UQI: | |
12287 | case V64QI_FTYPE_PCV64QI_V64QI_UDI: | |
12288 | case V32HI_FTYPE_PCV32HI_V32HI_USI: | |
12289 | case V32QI_FTYPE_PCV32QI_V32QI_USI: | |
12290 | case V16QI_FTYPE_PCV16QI_V16QI_UHI: | |
12291 | case V16HI_FTYPE_PCV16HI_V16HI_UHI: | |
12292 | case V8HI_FTYPE_PCV8HI_V8HI_UQI: | |
12293 | switch (icode) | |
12294 | { | |
12295 | /* These builtins and instructions require the memory | |
12296 | to be properly aligned. */ | |
12297 | case CODE_FOR_avx512f_loadv16sf_mask: | |
12298 | case CODE_FOR_avx512f_loadv16si_mask: | |
12299 | case CODE_FOR_avx512f_loadv8df_mask: | |
12300 | case CODE_FOR_avx512f_loadv8di_mask: | |
12301 | case CODE_FOR_avx512vl_loadv8sf_mask: | |
12302 | case CODE_FOR_avx512vl_loadv8si_mask: | |
12303 | case CODE_FOR_avx512vl_loadv4df_mask: | |
12304 | case CODE_FOR_avx512vl_loadv4di_mask: | |
12305 | case CODE_FOR_avx512vl_loadv4sf_mask: | |
12306 | case CODE_FOR_avx512vl_loadv4si_mask: | |
12307 | case CODE_FOR_avx512vl_loadv2df_mask: | |
12308 | case CODE_FOR_avx512vl_loadv2di_mask: | |
12309 | case CODE_FOR_avx512bw_loadv64qi_mask: | |
12310 | case CODE_FOR_avx512vl_loadv32qi_mask: | |
12311 | case CODE_FOR_avx512vl_loadv16qi_mask: | |
12312 | case CODE_FOR_avx512bw_loadv32hi_mask: | |
12313 | case CODE_FOR_avx512vl_loadv16hi_mask: | |
12314 | case CODE_FOR_avx512vl_loadv8hi_mask: | |
12315 | aligned_mem = true; | |
12316 | break; | |
12317 | default: | |
12318 | break; | |
12319 | } | |
12320 | /* FALLTHRU */ | |
12321 | case V64QI_FTYPE_PCCHAR_V64QI_UDI: | |
12322 | case V32QI_FTYPE_PCCHAR_V32QI_USI: | |
12323 | case V16QI_FTYPE_PCCHAR_V16QI_UHI: | |
12324 | case V32HI_FTYPE_PCSHORT_V32HI_USI: | |
12325 | case V16HI_FTYPE_PCSHORT_V16HI_UHI: | |
12326 | case V8HI_FTYPE_PCSHORT_V8HI_UQI: | |
12327 | case V16SI_FTYPE_PCINT_V16SI_UHI: | |
12328 | case V8SI_FTYPE_PCINT_V8SI_UQI: | |
12329 | case V4SI_FTYPE_PCINT_V4SI_UQI: | |
12330 | case V8DI_FTYPE_PCINT64_V8DI_UQI: | |
12331 | case V4DI_FTYPE_PCINT64_V4DI_UQI: | |
12332 | case V2DI_FTYPE_PCINT64_V2DI_UQI: | |
12333 | case V8DF_FTYPE_PCDOUBLE_V8DF_UQI: | |
12334 | case V4DF_FTYPE_PCDOUBLE_V4DF_UQI: | |
12335 | case V2DF_FTYPE_PCDOUBLE_V2DF_UQI: | |
12336 | case V16SF_FTYPE_PCFLOAT_V16SF_UHI: | |
12337 | case V8SF_FTYPE_PCFLOAT_V8SF_UQI: | |
12338 | case V4SF_FTYPE_PCFLOAT_V4SF_UQI: | |
c4d423c7 | 12339 | case V8HF_FTYPE_PCFLOAT16_V8HF_UQI: |
2bf6d935 ML |
12340 | nargs = 3; |
12341 | klass = load; | |
12342 | memory = 0; | |
12343 | break; | |
152834fe HJ |
12344 | case INT_FTYPE_PINT_INT_INT_INT: |
12345 | case LONGLONG_FTYPE_PLONGLONG_LONGLONG_LONGLONG_INT: | |
12346 | nargs = 4; | |
12347 | klass = load; | |
12348 | memory = 0; | |
12349 | constant = 3; | |
12350 | break; | |
2bf6d935 ML |
12351 | default: |
12352 | gcc_unreachable (); | |
12353 | } | |
12354 | ||
715a8bc8 | 12355 | gcc_assert (nargs <= ARRAY_SIZE (xops)); |
2bf6d935 ML |
12356 | |
12357 | if (klass == store) | |
12358 | { | |
12359 | arg = CALL_EXPR_ARG (exp, 0); | |
12360 | op = expand_normal (arg); | |
12361 | gcc_assert (target == 0); | |
12362 | if (memory) | |
12363 | { | |
12364 | op = ix86_zero_extend_to_Pmode (op); | |
12365 | target = gen_rtx_MEM (tmode, op); | |
12366 | /* target at this point has just BITS_PER_UNIT MEM_ALIGN | |
12367 | on it. Try to improve it using get_pointer_alignment, | |
12368 | and if the special builtin is one that requires strict | |
12369 | mode alignment, also from it's GET_MODE_ALIGNMENT. | |
12370 | Failure to do so could lead to ix86_legitimate_combined_insn | |
12371 | rejecting all changes to such insns. */ | |
12372 | unsigned int align = get_pointer_alignment (arg); | |
12373 | if (aligned_mem && align < GET_MODE_ALIGNMENT (tmode)) | |
12374 | align = GET_MODE_ALIGNMENT (tmode); | |
12375 | if (MEM_ALIGN (target) < align) | |
12376 | set_mem_align (target, align); | |
12377 | } | |
12378 | else | |
12379 | target = force_reg (tmode, op); | |
12380 | arg_adjust = 1; | |
12381 | } | |
12382 | else | |
12383 | { | |
12384 | arg_adjust = 0; | |
12385 | if (optimize | |
12386 | || target == 0 | |
12387 | || !register_operand (target, tmode) | |
12388 | || GET_MODE (target) != tmode) | |
12389 | target = gen_reg_rtx (tmode); | |
12390 | } | |
12391 | ||
12392 | for (i = 0; i < nargs; i++) | |
12393 | { | |
12394 | machine_mode mode = insn_p->operand[i + 1].mode; | |
2bf6d935 ML |
12395 | |
12396 | arg = CALL_EXPR_ARG (exp, i + arg_adjust); | |
12397 | op = expand_normal (arg); | |
2bf6d935 | 12398 | |
776a37f6 | 12399 | if (i == memory) |
2bf6d935 | 12400 | { |
776a37f6 | 12401 | /* This must be the memory operand. */ |
12402 | op = ix86_zero_extend_to_Pmode (op); | |
12403 | op = gen_rtx_MEM (mode, op); | |
12404 | /* op at this point has just BITS_PER_UNIT MEM_ALIGN | |
12405 | on it. Try to improve it using get_pointer_alignment, | |
12406 | and if the special builtin is one that requires strict | |
12407 | mode alignment, also from it's GET_MODE_ALIGNMENT. | |
12408 | Failure to do so could lead to ix86_legitimate_combined_insn | |
12409 | rejecting all changes to such insns. */ | |
12410 | unsigned int align = get_pointer_alignment (arg); | |
12411 | if (aligned_mem && align < GET_MODE_ALIGNMENT (mode)) | |
12412 | align = GET_MODE_ALIGNMENT (mode); | |
12413 | if (MEM_ALIGN (op) < align) | |
12414 | set_mem_align (op, align); | |
2bf6d935 | 12415 | } |
152834fe HJ |
12416 | else if (i == constant) |
12417 | { | |
12418 | /* This must be the constant. */ | |
12419 | if (!insn_p->operand[nargs].predicate(op, SImode)) | |
12420 | { | |
12421 | error ("the fourth argument must be one of enum %qs", "_CMPCCX_ENUM"); | |
12422 | return const0_rtx; | |
12423 | } | |
12424 | } | |
2bf6d935 ML |
12425 | else |
12426 | { | |
776a37f6 | 12427 | /* This must be register. */ |
12428 | if (VECTOR_MODE_P (mode)) | |
12429 | op = safe_vector_operand (op, mode); | |
2bf6d935 | 12430 | |
776a37f6 | 12431 | op = fixup_modeless_constant (op, mode); |
2bf6d935 | 12432 | |
b6efffa5 | 12433 | /* NB: 3-operands load implied it's a mask load or v{p}expand*, |
35c4c67e | 12434 | and that mask operand shoud be at the end. |
12435 | Keep all-ones mask which would be simplified by the expander. */ | |
12436 | if (nargs == 3 && i == 2 && klass == load | |
b6efffa5 | 12437 | && constm1_operand (op, mode) |
12438 | && insn_p->operand[i].predicate (op, mode)) | |
35c4c67e | 12439 | ; |
12440 | else if (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode) | |
776a37f6 | 12441 | op = copy_to_mode_reg (mode, op); |
12442 | else | |
12443 | { | |
12444 | op = copy_to_reg (op); | |
12445 | op = lowpart_subreg (mode, op, GET_MODE (op)); | |
2bf6d935 ML |
12446 | } |
12447 | } | |
12448 | ||
715a8bc8 | 12449 | xops[i]= op; |
2bf6d935 ML |
12450 | } |
12451 | ||
12452 | switch (nargs) | |
12453 | { | |
12454 | case 0: | |
12455 | pat = GEN_FCN (icode) (target); | |
12456 | break; | |
12457 | case 1: | |
715a8bc8 | 12458 | pat = GEN_FCN (icode) (target, xops[0]); |
2bf6d935 ML |
12459 | break; |
12460 | case 2: | |
715a8bc8 | 12461 | pat = GEN_FCN (icode) (target, xops[0], xops[1]); |
2bf6d935 ML |
12462 | break; |
12463 | case 3: | |
715a8bc8 | 12464 | pat = GEN_FCN (icode) (target, xops[0], xops[1], xops[2]); |
2bf6d935 | 12465 | break; |
152834fe HJ |
12466 | case 4: |
12467 | pat = GEN_FCN (icode) (target, xops[0], xops[1], xops[2], xops[3]); | |
12468 | break; | |
2bf6d935 ML |
12469 | default: |
12470 | gcc_unreachable (); | |
12471 | } | |
12472 | ||
12473 | if (! pat) | |
12474 | return 0; | |
715a8bc8 | 12475 | |
2bf6d935 ML |
12476 | emit_insn (pat); |
12477 | return klass == store ? 0 : target; | |
12478 | } | |
12479 | ||
12480 | /* Return the integer constant in ARG. Constrain it to be in the range | |
12481 | of the subparts of VEC_TYPE; issue an error if not. */ | |
12482 | ||
12483 | static int | |
12484 | get_element_number (tree vec_type, tree arg) | |
12485 | { | |
12486 | unsigned HOST_WIDE_INT elt, max = TYPE_VECTOR_SUBPARTS (vec_type) - 1; | |
12487 | ||
12488 | if (!tree_fits_uhwi_p (arg) | |
12489 | || (elt = tree_to_uhwi (arg), elt > max)) | |
12490 | { | |
a9c697b8 MS |
12491 | error ("selector must be an integer constant in the range " |
12492 | "[0, %wi]", max); | |
2bf6d935 ML |
12493 | return 0; |
12494 | } | |
12495 | ||
12496 | return elt; | |
12497 | } | |
12498 | ||
12499 | /* A subroutine of ix86_expand_builtin. These builtins are a wrapper around | |
12500 | ix86_expand_vector_init. We DO have language-level syntax for this, in | |
12501 | the form of (type){ init-list }. Except that since we can't place emms | |
12502 | instructions from inside the compiler, we can't allow the use of MMX | |
12503 | registers unless the user explicitly asks for it. So we do *not* define | |
12504 | vec_set/vec_extract/vec_init patterns for MMX modes in mmx.md. Instead | |
12505 | we have builtins invoked by mmintrin.h that gives us license to emit | |
12506 | these sorts of instructions. */ | |
12507 | ||
12508 | static rtx | |
12509 | ix86_expand_vec_init_builtin (tree type, tree exp, rtx target) | |
12510 | { | |
12511 | machine_mode tmode = TYPE_MODE (type); | |
12512 | machine_mode inner_mode = GET_MODE_INNER (tmode); | |
12513 | int i, n_elt = GET_MODE_NUNITS (tmode); | |
12514 | rtvec v = rtvec_alloc (n_elt); | |
12515 | ||
12516 | gcc_assert (VECTOR_MODE_P (tmode)); | |
12517 | gcc_assert (call_expr_nargs (exp) == n_elt); | |
12518 | ||
12519 | for (i = 0; i < n_elt; ++i) | |
12520 | { | |
12521 | rtx x = expand_normal (CALL_EXPR_ARG (exp, i)); | |
12522 | RTVEC_ELT (v, i) = gen_lowpart (inner_mode, x); | |
12523 | } | |
12524 | ||
12525 | if (!target || !register_operand (target, tmode)) | |
12526 | target = gen_reg_rtx (tmode); | |
12527 | ||
12528 | ix86_expand_vector_init (true, target, gen_rtx_PARALLEL (tmode, v)); | |
12529 | return target; | |
12530 | } | |
12531 | ||
12532 | /* A subroutine of ix86_expand_builtin. These builtins are a wrapper around | |
12533 | ix86_expand_vector_extract. They would be redundant (for non-MMX) if we | |
12534 | had a language-level syntax for referencing vector elements. */ | |
12535 | ||
12536 | static rtx | |
12537 | ix86_expand_vec_ext_builtin (tree exp, rtx target) | |
12538 | { | |
12539 | machine_mode tmode, mode0; | |
12540 | tree arg0, arg1; | |
12541 | int elt; | |
12542 | rtx op0; | |
12543 | ||
12544 | arg0 = CALL_EXPR_ARG (exp, 0); | |
12545 | arg1 = CALL_EXPR_ARG (exp, 1); | |
12546 | ||
12547 | op0 = expand_normal (arg0); | |
12548 | elt = get_element_number (TREE_TYPE (arg0), arg1); | |
12549 | ||
12550 | tmode = TYPE_MODE (TREE_TYPE (TREE_TYPE (arg0))); | |
12551 | mode0 = TYPE_MODE (TREE_TYPE (arg0)); | |
12552 | gcc_assert (VECTOR_MODE_P (mode0)); | |
12553 | ||
12554 | op0 = force_reg (mode0, op0); | |
12555 | ||
12556 | if (optimize || !target || !register_operand (target, tmode)) | |
12557 | target = gen_reg_rtx (tmode); | |
12558 | ||
12559 | ix86_expand_vector_extract (true, target, op0, elt); | |
12560 | ||
12561 | return target; | |
12562 | } | |
12563 | ||
12564 | /* A subroutine of ix86_expand_builtin. These builtins are a wrapper around | |
12565 | ix86_expand_vector_set. They would be redundant (for non-MMX) if we had | |
12566 | a language-level syntax for referencing vector elements. */ | |
12567 | ||
12568 | static rtx | |
12569 | ix86_expand_vec_set_builtin (tree exp) | |
12570 | { | |
12571 | machine_mode tmode, mode1; | |
12572 | tree arg0, arg1, arg2; | |
12573 | int elt; | |
12574 | rtx op0, op1, target; | |
12575 | ||
12576 | arg0 = CALL_EXPR_ARG (exp, 0); | |
12577 | arg1 = CALL_EXPR_ARG (exp, 1); | |
12578 | arg2 = CALL_EXPR_ARG (exp, 2); | |
12579 | ||
12580 | tmode = TYPE_MODE (TREE_TYPE (arg0)); | |
12581 | mode1 = TYPE_MODE (TREE_TYPE (TREE_TYPE (arg0))); | |
12582 | gcc_assert (VECTOR_MODE_P (tmode)); | |
12583 | ||
12584 | op0 = expand_expr (arg0, NULL_RTX, tmode, EXPAND_NORMAL); | |
12585 | op1 = expand_expr (arg1, NULL_RTX, mode1, EXPAND_NORMAL); | |
12586 | elt = get_element_number (TREE_TYPE (arg0), arg2); | |
12587 | ||
cda29c54 | 12588 | if (GET_MODE (op1) != mode1) |
2bf6d935 ML |
12589 | op1 = convert_modes (mode1, GET_MODE (op1), op1, true); |
12590 | ||
12591 | op0 = force_reg (tmode, op0); | |
12592 | op1 = force_reg (mode1, op1); | |
12593 | ||
12594 | /* OP0 is the source of these builtin functions and shouldn't be | |
12595 | modified. Create a copy, use it and return it as target. */ | |
12596 | target = gen_reg_rtx (tmode); | |
12597 | emit_move_insn (target, op0); | |
12598 | ix86_expand_vector_set (true, target, op1, elt); | |
12599 | ||
12600 | return target; | |
12601 | } | |
12602 | ||
823b3b79 | 12603 | /* Return true if the necessary isa options for this builtin exist, |
12604 | else false. | |
12605 | fcode = DECL_MD_FUNCTION_CODE (fndecl); */ | |
12606 | bool | |
12607 | ix86_check_builtin_isa_match (unsigned int fcode, | |
12608 | HOST_WIDE_INT* pbisa, | |
12609 | HOST_WIDE_INT* pbisa2) | |
2bf6d935 | 12610 | { |
2bf6d935 ML |
12611 | HOST_WIDE_INT isa = ix86_isa_flags; |
12612 | HOST_WIDE_INT isa2 = ix86_isa_flags2; | |
12613 | HOST_WIDE_INT bisa = ix86_builtins_isa[fcode].isa; | |
12614 | HOST_WIDE_INT bisa2 = ix86_builtins_isa[fcode].isa2; | |
5ebdbdb9 | 12615 | HOST_WIDE_INT tmp_isa = isa, tmp_isa2 = isa2; |
2bf6d935 ML |
12616 | /* The general case is we require all the ISAs specified in bisa{,2} |
12617 | to be enabled. | |
12618 | The exceptions are: | |
12619 | OPTION_MASK_ISA_SSE | OPTION_MASK_ISA_3DNOW_A | |
12620 | OPTION_MASK_ISA_SSE4_2 | OPTION_MASK_ISA_CRC32 | |
12621 | OPTION_MASK_ISA_FMA | OPTION_MASK_ISA_FMA4 | |
ca813880 | 12622 | (OPTION_MASK_ISA_AVX512VNNI | OPTION_MASK_ISA_AVX512VL) or |
12623 | OPTION_MASK_ISA2_AVXVNNI | |
5ebdbdb9 | 12624 | (OPTION_MASK_ISA_AVX512IFMA | OPTION_MASK_ISA_AVX512VL) or |
825d0041 | 12625 | OPTION_MASK_ISA2_AVXIFMA |
5ebdbdb9 | 12626 | (OPTION_MASK_ISA_AVX512VL | OPTION_MASK_ISA2_AVX512BF16) or |
58685b93 | 12627 | OPTION_MASK_ISA2_AVXNECONVERT |
a13d6ec8 JJ |
12628 | where for each such pair it is sufficient if either of the ISAs is |
12629 | enabled, plus if it is ored with other options also those others. | |
12630 | OPTION_MASK_ISA_MMX in bisa is satisfied also if TARGET_MMX_WITH_SSE. */ | |
db3f0d21 | 12631 | |
5ebdbdb9 HJ |
12632 | #define SHARE_BUILTIN(A1, A2, B1, B2) \ |
12633 | if ((((bisa & (A1)) == (A1) && (bisa2 & (A2)) == (A2)) \ | |
12634 | && ((bisa & (B1)) == (B1) && (bisa2 & (B2)) == (B2))) \ | |
12635 | && (((isa & (A1)) == (A1) && (isa2 & (A2)) == (A2)) \ | |
12636 | || ((isa & (B1)) == (B1) && (isa2 & (B2)) == (B2)))) \ | |
12637 | { \ | |
12638 | tmp_isa |= (A1) | (B1); \ | |
12639 | tmp_isa2 |= (A2) | (B2); \ | |
12640 | } | |
12641 | ||
12642 | SHARE_BUILTIN (OPTION_MASK_ISA_SSE, 0, OPTION_MASK_ISA_3DNOW_A, 0); | |
12643 | SHARE_BUILTIN (OPTION_MASK_ISA_SSE4_2, 0, OPTION_MASK_ISA_CRC32, 0); | |
12644 | SHARE_BUILTIN (OPTION_MASK_ISA_FMA, 0, OPTION_MASK_ISA_FMA4, 0); | |
12645 | SHARE_BUILTIN (OPTION_MASK_ISA_AVX512VNNI | OPTION_MASK_ISA_AVX512VL, 0, 0, | |
12646 | OPTION_MASK_ISA2_AVXVNNI); | |
12647 | SHARE_BUILTIN (OPTION_MASK_ISA_AVX512IFMA | OPTION_MASK_ISA_AVX512VL, 0, 0, | |
12648 | OPTION_MASK_ISA2_AVXIFMA); | |
12649 | SHARE_BUILTIN (OPTION_MASK_ISA_AVX512VL, OPTION_MASK_ISA2_AVX512BF16, 0, | |
12650 | OPTION_MASK_ISA2_AVXNECONVERT); | |
24a8acc1 | 12651 | SHARE_BUILTIN (OPTION_MASK_ISA_AES, 0, 0, OPTION_MASK_ISA2_VAES); |
5ebdbdb9 HJ |
12652 | isa = tmp_isa; |
12653 | isa2 = tmp_isa2; | |
58685b93 | 12654 | |
db3f0d21 UB |
12655 | if ((bisa & OPTION_MASK_ISA_MMX) && !TARGET_MMX && TARGET_MMX_WITH_SSE |
12656 | /* __builtin_ia32_maskmovq requires MMX registers. */ | |
6058b874 | 12657 | && fcode != IX86_BUILTIN_MASKMOVQ) |
a13d6ec8 JJ |
12658 | { |
12659 | bisa &= ~OPTION_MASK_ISA_MMX; | |
12660 | bisa |= OPTION_MASK_ISA_SSE2; | |
ecfdb16c | 12661 | } |
6058b874 | 12662 | |
823b3b79 | 12663 | if (pbisa) |
12664 | *pbisa = bisa; | |
12665 | if (pbisa2) | |
12666 | *pbisa2 = bisa2; | |
12667 | ||
12668 | return (bisa & isa) == bisa && (bisa2 & isa2) == bisa2; | |
12669 | } | |
12670 | ||
af29d0d6 RS |
12671 | /* Emit instructions to set the carry flag from ARG. */ |
12672 | ||
12673 | void | |
12674 | ix86_expand_carry (rtx arg) | |
12675 | { | |
12676 | if (!CONST_INT_P (arg) || arg == const0_rtx) | |
12677 | { | |
12678 | arg = convert_to_mode (QImode, arg, 1); | |
12679 | arg = copy_to_mode_reg (QImode, arg); | |
12680 | emit_insn (gen_addqi3_cconly_overflow (arg, constm1_rtx)); | |
12681 | } | |
12682 | else | |
12683 | emit_insn (gen_x86_stc ()); | |
12684 | } | |
12685 | ||
823b3b79 | 12686 | /* Expand an expression EXP that calls a built-in function, |
12687 | with result going to TARGET if that's convenient | |
12688 | (and in mode MODE if that's convenient). | |
12689 | SUBTARGET may be used as the target for computing one of EXP's operands. | |
12690 | IGNORE is nonzero if the value is to be ignored. */ | |
12691 | ||
12692 | rtx | |
12693 | ix86_expand_builtin (tree exp, rtx target, rtx subtarget, | |
12694 | machine_mode mode, int ignore) | |
12695 | { | |
12696 | size_t i; | |
12697 | enum insn_code icode, icode2; | |
12698 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); | |
12699 | tree arg0, arg1, arg2, arg3, arg4; | |
12700 | rtx op0, op1, op2, op3, op4, pat, pat2, insn; | |
12701 | machine_mode mode0, mode1, mode2, mode3, mode4; | |
12702 | unsigned int fcode = DECL_MD_FUNCTION_CODE (fndecl); | |
12703 | HOST_WIDE_INT bisa, bisa2; | |
12704 | ||
12705 | /* For CPU builtins that can be folded, fold first and expand the fold. */ | |
12706 | switch (fcode) | |
12707 | { | |
12708 | case IX86_BUILTIN_CPU_INIT: | |
12709 | { | |
12710 | /* Make it call __cpu_indicator_init in libgcc. */ | |
12711 | tree call_expr, fndecl, type; | |
12712 | type = build_function_type_list (integer_type_node, NULL_TREE); | |
12713 | fndecl = build_fn_decl ("__cpu_indicator_init", type); | |
12714 | call_expr = build_call_expr (fndecl, 0); | |
12715 | return expand_expr (call_expr, target, mode, EXPAND_NORMAL); | |
12716 | } | |
12717 | case IX86_BUILTIN_CPU_IS: | |
12718 | case IX86_BUILTIN_CPU_SUPPORTS: | |
12719 | { | |
12720 | tree arg0 = CALL_EXPR_ARG (exp, 0); | |
12721 | tree fold_expr = fold_builtin_cpu (fndecl, &arg0); | |
12722 | gcc_assert (fold_expr != NULL_TREE); | |
12723 | return expand_expr (fold_expr, target, mode, EXPAND_NORMAL); | |
12724 | } | |
12725 | } | |
12726 | ||
12727 | if (!ix86_check_builtin_isa_match (fcode, &bisa, &bisa2)) | |
2bf6d935 ML |
12728 | { |
12729 | bool add_abi_p = bisa & OPTION_MASK_ISA_64BIT; | |
12730 | if (TARGET_ABI_X32) | |
12731 | bisa |= OPTION_MASK_ABI_X32; | |
12732 | else | |
12733 | bisa |= OPTION_MASK_ABI_64; | |
12734 | char *opts = ix86_target_string (bisa, bisa2, 0, 0, NULL, NULL, | |
46e6341f JJ |
12735 | (enum fpmath_unit) 0, |
12736 | (enum prefer_vector_width) 0, | |
654cd743 | 12737 | PVW_NONE, PVW_NONE, |
46e6341f | 12738 | false, add_abi_p); |
2bf6d935 ML |
12739 | if (!opts) |
12740 | error ("%qE needs unknown isa option", fndecl); | |
12741 | else | |
12742 | { | |
12743 | gcc_assert (opts != NULL); | |
12744 | error ("%qE needs isa option %s", fndecl, opts); | |
12745 | free (opts); | |
12746 | } | |
12747 | return expand_call (exp, target, ignore); | |
12748 | } | |
12749 | ||
12750 | switch (fcode) | |
12751 | { | |
12752 | case IX86_BUILTIN_MASKMOVQ: | |
12753 | case IX86_BUILTIN_MASKMOVDQU: | |
12754 | icode = (fcode == IX86_BUILTIN_MASKMOVQ | |
12755 | ? CODE_FOR_mmx_maskmovq | |
12756 | : CODE_FOR_sse2_maskmovdqu); | |
12757 | /* Note the arg order is different from the operand order. */ | |
12758 | arg1 = CALL_EXPR_ARG (exp, 0); | |
12759 | arg2 = CALL_EXPR_ARG (exp, 1); | |
12760 | arg0 = CALL_EXPR_ARG (exp, 2); | |
12761 | op0 = expand_normal (arg0); | |
12762 | op1 = expand_normal (arg1); | |
12763 | op2 = expand_normal (arg2); | |
12764 | mode0 = insn_data[icode].operand[0].mode; | |
12765 | mode1 = insn_data[icode].operand[1].mode; | |
12766 | mode2 = insn_data[icode].operand[2].mode; | |
12767 | ||
12768 | op0 = ix86_zero_extend_to_Pmode (op0); | |
12769 | op0 = gen_rtx_MEM (mode1, op0); | |
12770 | ||
12771 | if (!insn_data[icode].operand[0].predicate (op0, mode0)) | |
12772 | op0 = copy_to_mode_reg (mode0, op0); | |
12773 | if (!insn_data[icode].operand[1].predicate (op1, mode1)) | |
12774 | op1 = copy_to_mode_reg (mode1, op1); | |
12775 | if (!insn_data[icode].operand[2].predicate (op2, mode2)) | |
12776 | op2 = copy_to_mode_reg (mode2, op2); | |
12777 | pat = GEN_FCN (icode) (op0, op1, op2); | |
12778 | if (! pat) | |
12779 | return 0; | |
12780 | emit_insn (pat); | |
12781 | return 0; | |
12782 | ||
12783 | case IX86_BUILTIN_LDMXCSR: | |
12784 | op0 = expand_normal (CALL_EXPR_ARG (exp, 0)); | |
12785 | target = assign_386_stack_local (SImode, SLOT_TEMP); | |
12786 | emit_move_insn (target, op0); | |
12787 | emit_insn (gen_sse_ldmxcsr (target)); | |
12788 | return 0; | |
12789 | ||
12790 | case IX86_BUILTIN_STMXCSR: | |
12791 | target = assign_386_stack_local (SImode, SLOT_TEMP); | |
12792 | emit_insn (gen_sse_stmxcsr (target)); | |
12793 | return copy_to_mode_reg (SImode, target); | |
12794 | ||
12795 | case IX86_BUILTIN_CLFLUSH: | |
12796 | arg0 = CALL_EXPR_ARG (exp, 0); | |
12797 | op0 = expand_normal (arg0); | |
12798 | icode = CODE_FOR_sse2_clflush; | |
12799 | if (!insn_data[icode].operand[0].predicate (op0, Pmode)) | |
12800 | op0 = ix86_zero_extend_to_Pmode (op0); | |
12801 | ||
12802 | emit_insn (gen_sse2_clflush (op0)); | |
12803 | return 0; | |
12804 | ||
12805 | case IX86_BUILTIN_CLWB: | |
12806 | arg0 = CALL_EXPR_ARG (exp, 0); | |
12807 | op0 = expand_normal (arg0); | |
12808 | icode = CODE_FOR_clwb; | |
12809 | if (!insn_data[icode].operand[0].predicate (op0, Pmode)) | |
12810 | op0 = ix86_zero_extend_to_Pmode (op0); | |
12811 | ||
12812 | emit_insn (gen_clwb (op0)); | |
12813 | return 0; | |
12814 | ||
12815 | case IX86_BUILTIN_CLFLUSHOPT: | |
12816 | arg0 = CALL_EXPR_ARG (exp, 0); | |
12817 | op0 = expand_normal (arg0); | |
12818 | icode = CODE_FOR_clflushopt; | |
12819 | if (!insn_data[icode].operand[0].predicate (op0, Pmode)) | |
12820 | op0 = ix86_zero_extend_to_Pmode (op0); | |
12821 | ||
12822 | emit_insn (gen_clflushopt (op0)); | |
12823 | return 0; | |
12824 | ||
12825 | case IX86_BUILTIN_MONITOR: | |
12826 | case IX86_BUILTIN_MONITORX: | |
12827 | arg0 = CALL_EXPR_ARG (exp, 0); | |
12828 | arg1 = CALL_EXPR_ARG (exp, 1); | |
12829 | arg2 = CALL_EXPR_ARG (exp, 2); | |
12830 | op0 = expand_normal (arg0); | |
12831 | op1 = expand_normal (arg1); | |
12832 | op2 = expand_normal (arg2); | |
12833 | if (!REG_P (op0)) | |
12834 | op0 = ix86_zero_extend_to_Pmode (op0); | |
12835 | if (!REG_P (op1)) | |
12836 | op1 = copy_to_mode_reg (SImode, op1); | |
12837 | if (!REG_P (op2)) | |
12838 | op2 = copy_to_mode_reg (SImode, op2); | |
12839 | ||
12840 | emit_insn (fcode == IX86_BUILTIN_MONITOR | |
a963ca40 UB |
12841 | ? gen_sse3_monitor (Pmode, op0, op1, op2) |
12842 | : gen_monitorx (Pmode, op0, op1, op2)); | |
2bf6d935 ML |
12843 | return 0; |
12844 | ||
12845 | case IX86_BUILTIN_MWAIT: | |
12846 | arg0 = CALL_EXPR_ARG (exp, 0); | |
12847 | arg1 = CALL_EXPR_ARG (exp, 1); | |
12848 | op0 = expand_normal (arg0); | |
12849 | op1 = expand_normal (arg1); | |
12850 | if (!REG_P (op0)) | |
12851 | op0 = copy_to_mode_reg (SImode, op0); | |
12852 | if (!REG_P (op1)) | |
12853 | op1 = copy_to_mode_reg (SImode, op1); | |
12854 | emit_insn (gen_sse3_mwait (op0, op1)); | |
12855 | return 0; | |
12856 | ||
12857 | case IX86_BUILTIN_MWAITX: | |
12858 | arg0 = CALL_EXPR_ARG (exp, 0); | |
12859 | arg1 = CALL_EXPR_ARG (exp, 1); | |
12860 | arg2 = CALL_EXPR_ARG (exp, 2); | |
12861 | op0 = expand_normal (arg0); | |
12862 | op1 = expand_normal (arg1); | |
12863 | op2 = expand_normal (arg2); | |
12864 | if (!REG_P (op0)) | |
12865 | op0 = copy_to_mode_reg (SImode, op0); | |
12866 | if (!REG_P (op1)) | |
12867 | op1 = copy_to_mode_reg (SImode, op1); | |
12868 | if (!REG_P (op2)) | |
12869 | op2 = copy_to_mode_reg (SImode, op2); | |
12870 | emit_insn (gen_mwaitx (op0, op1, op2)); | |
12871 | return 0; | |
12872 | ||
12873 | case IX86_BUILTIN_UMONITOR: | |
12874 | arg0 = CALL_EXPR_ARG (exp, 0); | |
12875 | op0 = expand_normal (arg0); | |
12876 | ||
12877 | op0 = ix86_zero_extend_to_Pmode (op0); | |
987a3082 | 12878 | emit_insn (gen_umonitor (Pmode, op0)); |
2bf6d935 ML |
12879 | return 0; |
12880 | ||
12881 | case IX86_BUILTIN_UMWAIT: | |
12882 | case IX86_BUILTIN_TPAUSE: | |
12883 | arg0 = CALL_EXPR_ARG (exp, 0); | |
12884 | arg1 = CALL_EXPR_ARG (exp, 1); | |
12885 | op0 = expand_normal (arg0); | |
12886 | op1 = expand_normal (arg1); | |
12887 | ||
12888 | if (!REG_P (op0)) | |
12889 | op0 = copy_to_mode_reg (SImode, op0); | |
12890 | ||
12891 | op1 = force_reg (DImode, op1); | |
12892 | ||
12893 | if (TARGET_64BIT) | |
12894 | { | |
12895 | op2 = expand_simple_binop (DImode, LSHIFTRT, op1, GEN_INT (32), | |
12896 | NULL, 1, OPTAB_DIRECT); | |
12897 | switch (fcode) | |
12898 | { | |
12899 | case IX86_BUILTIN_UMWAIT: | |
12900 | icode = CODE_FOR_umwait_rex64; | |
12901 | break; | |
12902 | case IX86_BUILTIN_TPAUSE: | |
12903 | icode = CODE_FOR_tpause_rex64; | |
12904 | break; | |
12905 | default: | |
12906 | gcc_unreachable (); | |
12907 | } | |
12908 | ||
12909 | op2 = gen_lowpart (SImode, op2); | |
12910 | op1 = gen_lowpart (SImode, op1); | |
12911 | pat = GEN_FCN (icode) (op0, op1, op2); | |
12912 | } | |
12913 | else | |
12914 | { | |
12915 | switch (fcode) | |
12916 | { | |
12917 | case IX86_BUILTIN_UMWAIT: | |
12918 | icode = CODE_FOR_umwait; | |
12919 | break; | |
12920 | case IX86_BUILTIN_TPAUSE: | |
12921 | icode = CODE_FOR_tpause; | |
12922 | break; | |
12923 | default: | |
12924 | gcc_unreachable (); | |
12925 | } | |
12926 | pat = GEN_FCN (icode) (op0, op1); | |
12927 | } | |
12928 | ||
12929 | if (!pat) | |
12930 | return 0; | |
12931 | ||
12932 | emit_insn (pat); | |
12933 | ||
12934 | if (target == 0 | |
12935 | || !register_operand (target, QImode)) | |
12936 | target = gen_reg_rtx (QImode); | |
12937 | ||
12938 | pat = gen_rtx_EQ (QImode, gen_rtx_REG (CCCmode, FLAGS_REG), | |
12939 | const0_rtx); | |
12940 | emit_insn (gen_rtx_SET (target, pat)); | |
12941 | ||
12942 | return target; | |
12943 | ||
299a53d7 | 12944 | case IX86_BUILTIN_TESTUI: |
12945 | emit_insn (gen_testui ()); | |
12946 | ||
12947 | if (target == 0 | |
12948 | || !register_operand (target, QImode)) | |
12949 | target = gen_reg_rtx (QImode); | |
12950 | ||
12951 | pat = gen_rtx_LTU (QImode, gen_rtx_REG (CCCmode, FLAGS_REG), | |
12952 | const0_rtx); | |
12953 | emit_insn (gen_rtx_SET (target, pat)); | |
12954 | ||
12955 | return target; | |
12956 | ||
2bf6d935 ML |
12957 | case IX86_BUILTIN_CLZERO: |
12958 | arg0 = CALL_EXPR_ARG (exp, 0); | |
12959 | op0 = expand_normal (arg0); | |
12960 | if (!REG_P (op0)) | |
12961 | op0 = ix86_zero_extend_to_Pmode (op0); | |
a963ca40 | 12962 | emit_insn (gen_clzero (Pmode, op0)); |
2bf6d935 ML |
12963 | return 0; |
12964 | ||
12965 | case IX86_BUILTIN_CLDEMOTE: | |
12966 | arg0 = CALL_EXPR_ARG (exp, 0); | |
12967 | op0 = expand_normal (arg0); | |
12968 | icode = CODE_FOR_cldemote; | |
12969 | if (!insn_data[icode].operand[0].predicate (op0, Pmode)) | |
12970 | op0 = ix86_zero_extend_to_Pmode (op0); | |
12971 | ||
12972 | emit_insn (gen_cldemote (op0)); | |
12973 | return 0; | |
12974 | ||
632a2f50 | 12975 | case IX86_BUILTIN_LOADIWKEY: |
12976 | { | |
12977 | arg0 = CALL_EXPR_ARG (exp, 0); | |
12978 | arg1 = CALL_EXPR_ARG (exp, 1); | |
12979 | arg2 = CALL_EXPR_ARG (exp, 2); | |
12980 | arg3 = CALL_EXPR_ARG (exp, 3); | |
12981 | ||
12982 | op0 = expand_normal (arg0); | |
12983 | op1 = expand_normal (arg1); | |
12984 | op2 = expand_normal (arg2); | |
12985 | op3 = expand_normal (arg3); | |
12986 | ||
12987 | if (!REG_P (op0)) | |
12988 | op0 = copy_to_mode_reg (V2DImode, op0); | |
12989 | if (!REG_P (op1)) | |
12990 | op1 = copy_to_mode_reg (V2DImode, op1); | |
12991 | if (!REG_P (op2)) | |
12992 | op2 = copy_to_mode_reg (V2DImode, op2); | |
12993 | if (!REG_P (op3)) | |
12994 | op3 = copy_to_mode_reg (SImode, op3); | |
12995 | ||
12996 | emit_insn (gen_loadiwkey (op0, op1, op2, op3)); | |
12997 | ||
12998 | return 0; | |
12999 | } | |
13000 | ||
13001 | case IX86_BUILTIN_AESDEC128KLU8: | |
13002 | icode = CODE_FOR_aesdec128klu8; | |
13003 | goto aesdecenc_expand; | |
13004 | ||
13005 | case IX86_BUILTIN_AESDEC256KLU8: | |
13006 | icode = CODE_FOR_aesdec256klu8; | |
13007 | goto aesdecenc_expand; | |
13008 | ||
13009 | case IX86_BUILTIN_AESENC128KLU8: | |
13010 | icode = CODE_FOR_aesenc128klu8; | |
13011 | goto aesdecenc_expand; | |
13012 | ||
13013 | case IX86_BUILTIN_AESENC256KLU8: | |
13014 | icode = CODE_FOR_aesenc256klu8; | |
13015 | ||
13016 | aesdecenc_expand: | |
13017 | ||
13018 | arg0 = CALL_EXPR_ARG (exp, 0); // __m128i *odata | |
13019 | arg1 = CALL_EXPR_ARG (exp, 1); // __m128i idata | |
13020 | arg2 = CALL_EXPR_ARG (exp, 2); // const void *p | |
13021 | ||
13022 | op0 = expand_normal (arg0); | |
13023 | op1 = expand_normal (arg1); | |
13024 | op2 = expand_normal (arg2); | |
13025 | ||
13026 | if (!address_operand (op0, V2DImode)) | |
13027 | { | |
13028 | op0 = convert_memory_address (Pmode, op0); | |
13029 | op0 = copy_addr_to_reg (op0); | |
13030 | } | |
13031 | op0 = gen_rtx_MEM (V2DImode, op0); | |
13032 | ||
13033 | if (!REG_P (op1)) | |
13034 | op1 = copy_to_mode_reg (V2DImode, op1); | |
13035 | ||
13036 | if (!address_operand (op2, VOIDmode)) | |
13037 | { | |
13038 | op2 = convert_memory_address (Pmode, op2); | |
13039 | op2 = copy_addr_to_reg (op2); | |
13040 | } | |
13041 | op2 = gen_rtx_MEM (BLKmode, op2); | |
13042 | ||
13043 | emit_insn (GEN_FCN (icode) (op1, op1, op2)); | |
13044 | ||
13045 | if (target == 0) | |
13046 | target = gen_reg_rtx (QImode); | |
13047 | ||
1aeefa57 HW |
13048 | /* NB: For aesenc/aesdec keylocker insn, ZF will be set when runtime |
13049 | error occurs. Then the output should be cleared for safety. */ | |
13050 | rtx_code_label *ok_label; | |
13051 | rtx tmp; | |
13052 | ||
13053 | tmp = gen_rtx_REG (CCZmode, FLAGS_REG); | |
13054 | pat = gen_rtx_EQ (QImode, tmp, const0_rtx); | |
13055 | ok_label = gen_label_rtx (); | |
13056 | emit_cmp_and_jump_insns (tmp, const0_rtx, NE, 0, GET_MODE (tmp), | |
13057 | true, ok_label); | |
13058 | /* Usually the runtime error seldom occur, so predict OK path as | |
13059 | hotspot to optimize it as fallthrough block. */ | |
13060 | predict_jump (REG_BR_PROB_BASE * 90 / 100); | |
13061 | ||
13062 | emit_insn (gen_rtx_SET (op1, const0_rtx)); | |
632a2f50 | 13063 | |
1aeefa57 HW |
13064 | emit_label (ok_label); |
13065 | emit_insn (gen_rtx_SET (target, pat)); | |
632a2f50 | 13066 | emit_insn (gen_rtx_SET (op0, op1)); |
13067 | ||
13068 | return target; | |
13069 | ||
13070 | case IX86_BUILTIN_AESDECWIDE128KLU8: | |
13071 | icode = CODE_FOR_aesdecwide128klu8; | |
13072 | goto wideaesdecenc_expand; | |
13073 | ||
13074 | case IX86_BUILTIN_AESDECWIDE256KLU8: | |
13075 | icode = CODE_FOR_aesdecwide256klu8; | |
13076 | goto wideaesdecenc_expand; | |
13077 | ||
13078 | case IX86_BUILTIN_AESENCWIDE128KLU8: | |
13079 | icode = CODE_FOR_aesencwide128klu8; | |
13080 | goto wideaesdecenc_expand; | |
13081 | ||
13082 | case IX86_BUILTIN_AESENCWIDE256KLU8: | |
13083 | icode = CODE_FOR_aesencwide256klu8; | |
13084 | ||
13085 | wideaesdecenc_expand: | |
13086 | ||
13087 | rtx xmm_regs[8]; | |
13088 | rtx op; | |
13089 | ||
13090 | arg0 = CALL_EXPR_ARG (exp, 0); // __m128i * odata | |
13091 | arg1 = CALL_EXPR_ARG (exp, 1); // const __m128i * idata | |
13092 | arg2 = CALL_EXPR_ARG (exp, 2); // const void *p | |
13093 | ||
13094 | op0 = expand_normal (arg0); | |
13095 | op1 = expand_normal (arg1); | |
13096 | op2 = expand_normal (arg2); | |
13097 | ||
13098 | if (!address_operand (op2, VOIDmode)) | |
13099 | { | |
13100 | op2 = convert_memory_address (Pmode, op2); | |
13101 | op2 = copy_addr_to_reg (op2); | |
13102 | } | |
13103 | op2 = gen_rtx_MEM (BLKmode, op2); | |
13104 | ||
13105 | for (i = 0; i < 8; i++) | |
13106 | { | |
13107 | xmm_regs[i] = gen_rtx_REG (V2DImode, GET_SSE_REGNO (i)); | |
13108 | ||
13109 | op = gen_rtx_MEM (V2DImode, | |
13110 | plus_constant (Pmode, op1, (i * 16))); | |
13111 | ||
13112 | emit_move_insn (xmm_regs[i], op); | |
13113 | } | |
13114 | ||
13115 | emit_insn (GEN_FCN (icode) (op2)); | |
13116 | ||
13117 | if (target == 0) | |
13118 | target = gen_reg_rtx (QImode); | |
13119 | ||
1aeefa57 HW |
13120 | tmp = gen_rtx_REG (CCZmode, FLAGS_REG); |
13121 | pat = gen_rtx_EQ (QImode, tmp, const0_rtx); | |
13122 | ok_label = gen_label_rtx (); | |
13123 | emit_cmp_and_jump_insns (tmp, const0_rtx, NE, 0, GET_MODE (tmp), | |
13124 | true, ok_label); | |
13125 | predict_jump (REG_BR_PROB_BASE * 90 / 100); | |
13126 | ||
13127 | for (i = 0; i < 8; i++) | |
13128 | emit_insn (gen_rtx_SET (xmm_regs[i], const0_rtx)); | |
13129 | ||
13130 | emit_label (ok_label); | |
632a2f50 | 13131 | emit_insn (gen_rtx_SET (target, pat)); |
13132 | ||
13133 | for (i = 0; i < 8; i++) | |
13134 | { | |
13135 | op = gen_rtx_MEM (V2DImode, | |
13136 | plus_constant (Pmode, op0, (i * 16))); | |
13137 | emit_move_insn (op, xmm_regs[i]); | |
13138 | } | |
13139 | ||
13140 | return target; | |
13141 | ||
13142 | case IX86_BUILTIN_ENCODEKEY128U32: | |
13143 | { | |
13144 | rtx op, xmm_regs[7]; | |
13145 | ||
13146 | arg0 = CALL_EXPR_ARG (exp, 0); // unsigned int htype | |
13147 | arg1 = CALL_EXPR_ARG (exp, 1); // __m128i key | |
13148 | arg2 = CALL_EXPR_ARG (exp, 2); // void *h | |
13149 | ||
13150 | op0 = expand_normal (arg0); | |
13151 | op1 = expand_normal (arg1); | |
13152 | op2 = expand_normal (arg2); | |
13153 | ||
13154 | if (!REG_P (op0)) | |
13155 | op0 = copy_to_mode_reg (SImode, op0); | |
13156 | ||
13157 | op = gen_rtx_REG (V2DImode, GET_SSE_REGNO (0)); | |
13158 | emit_move_insn (op, op1); | |
13159 | ||
13160 | for (i = 0; i < 3; i++) | |
13161 | xmm_regs[i] = gen_rtx_REG (V2DImode, GET_SSE_REGNO (i)); | |
13162 | ||
13163 | if (target == 0) | |
13164 | target = gen_reg_rtx (SImode); | |
13165 | ||
13166 | emit_insn (gen_encodekey128u32 (target, op0)); | |
13167 | ||
13168 | for (i = 0; i < 3; i++) | |
13169 | { | |
13170 | op = gen_rtx_MEM (V2DImode, | |
13171 | plus_constant (Pmode, op2, (i * 16))); | |
13172 | emit_move_insn (op, xmm_regs[i]); | |
13173 | } | |
13174 | ||
13175 | return target; | |
13176 | } | |
13177 | case IX86_BUILTIN_ENCODEKEY256U32: | |
13178 | { | |
13179 | rtx op, xmm_regs[7]; | |
13180 | ||
13181 | arg0 = CALL_EXPR_ARG (exp, 0); // unsigned int htype | |
13182 | arg1 = CALL_EXPR_ARG (exp, 1); // __m128i keylow | |
13183 | arg2 = CALL_EXPR_ARG (exp, 2); // __m128i keyhi | |
13184 | arg3 = CALL_EXPR_ARG (exp, 3); // void *h | |
13185 | ||
13186 | op0 = expand_normal (arg0); | |
13187 | op1 = expand_normal (arg1); | |
13188 | op2 = expand_normal (arg2); | |
13189 | op3 = expand_normal (arg3); | |
13190 | ||
13191 | if (!REG_P (op0)) | |
13192 | op0 = copy_to_mode_reg (SImode, op0); | |
13193 | ||
13194 | /* Force to use xmm0, xmm1 for keylow, keyhi*/ | |
13195 | op = gen_rtx_REG (V2DImode, GET_SSE_REGNO (0)); | |
13196 | emit_move_insn (op, op1); | |
13197 | op = gen_rtx_REG (V2DImode, GET_SSE_REGNO (1)); | |
13198 | emit_move_insn (op, op2); | |
13199 | ||
13200 | for (i = 0; i < 4; i++) | |
13201 | xmm_regs[i] = gen_rtx_REG (V2DImode, GET_SSE_REGNO (i)); | |
13202 | ||
13203 | if (target == 0) | |
13204 | target = gen_reg_rtx (SImode); | |
13205 | ||
13206 | emit_insn (gen_encodekey256u32 (target, op0)); | |
13207 | ||
13208 | for (i = 0; i < 4; i++) | |
13209 | { | |
13210 | op = gen_rtx_MEM (V2DImode, | |
13211 | plus_constant (Pmode, op3, (i * 16))); | |
13212 | emit_move_insn (op, xmm_regs[i]); | |
13213 | } | |
13214 | ||
13215 | return target; | |
13216 | } | |
13217 | ||
b384d9a0 HJ |
13218 | case IX86_BUILTIN_PREFETCH: |
13219 | { | |
13220 | arg0 = CALL_EXPR_ARG (exp, 0); // const void * | |
13221 | arg1 = CALL_EXPR_ARG (exp, 1); // const int | |
13222 | arg2 = CALL_EXPR_ARG (exp, 2); // const int | |
13223 | arg3 = CALL_EXPR_ARG (exp, 3); // const int | |
13224 | ||
13225 | op0 = expand_normal (arg0); | |
13226 | op1 = expand_normal (arg1); | |
13227 | op2 = expand_normal (arg2); | |
13228 | op3 = expand_normal (arg3); | |
13229 | ||
13230 | if (!CONST_INT_P (op1) || !CONST_INT_P (op2) || !CONST_INT_P (op3)) | |
13231 | { | |
13232 | error ("second, third and fourth argument must be a const"); | |
13233 | return const0_rtx; | |
13234 | } | |
13235 | ||
13236 | if (INTVAL (op3) == 1) | |
13237 | { | |
77a67e3a L |
13238 | if (INTVAL (op2) < 2 || INTVAL (op2) > 3) |
13239 | { | |
13240 | error ("invalid third argument"); | |
13241 | return const0_rtx; | |
13242 | } | |
13243 | ||
21de01f5 | 13244 | if (TARGET_64BIT && TARGET_PREFETCHI |
b384d9a0 HJ |
13245 | && local_func_symbolic_operand (op0, GET_MODE (op0))) |
13246 | emit_insn (gen_prefetchi (op0, op2)); | |
13247 | else | |
13248 | { | |
13249 | warning (0, "instruction prefetch applies when in 64-bit mode" | |
13250 | " with RIP-relative addressing and" | |
13251 | " option %<-mprefetchi%>;" | |
13252 | " they stay NOPs otherwise"); | |
13253 | emit_insn (gen_nop ()); | |
13254 | } | |
13255 | } | |
13256 | else | |
13257 | { | |
13258 | if (!address_operand (op0, VOIDmode)) | |
13259 | { | |
13260 | op0 = convert_memory_address (Pmode, op0); | |
13261 | op0 = copy_addr_to_reg (op0); | |
13262 | } | |
21de01f5 | 13263 | |
77a67e3a L |
13264 | if (INTVAL (op2) < 0 || INTVAL (op2) > 3) |
13265 | { | |
13266 | warning (0, "invalid third argument to %<__builtin_ia32_prefetch%>; using zero"); | |
13267 | op2 = const0_rtx; | |
13268 | } | |
13269 | ||
21de01f5 HJ |
13270 | if (TARGET_3DNOW || TARGET_PREFETCH_SSE |
13271 | || TARGET_PRFCHW || TARGET_PREFETCHWT1) | |
13272 | emit_insn (gen_prefetch (op0, op1, op2)); | |
13273 | else if (!MEM_P (op0) && side_effects_p (op0)) | |
13274 | /* Don't do anything with direct references to volatile memory, | |
13275 | but generate code to handle other side effects. */ | |
13276 | emit_insn (op0); | |
b384d9a0 HJ |
13277 | } |
13278 | ||
13279 | return 0; | |
13280 | } | |
13281 | ||
13282 | case IX86_BUILTIN_PREFETCHI: | |
13283 | { | |
13284 | arg0 = CALL_EXPR_ARG (exp, 0); // const void * | |
13285 | arg1 = CALL_EXPR_ARG (exp, 1); // const int | |
13286 | ||
13287 | op0 = expand_normal (arg0); | |
13288 | op1 = expand_normal (arg1); | |
13289 | ||
13290 | if (!CONST_INT_P (op1)) | |
13291 | { | |
13292 | error ("second argument must be a const"); | |
13293 | return const0_rtx; | |
13294 | } | |
13295 | ||
13296 | /* GOT/PLT_PIC should not be available for instruction prefetch. | |
13297 | It must be real instruction address. */ | |
13298 | if (TARGET_64BIT | |
13299 | && local_func_symbolic_operand (op0, GET_MODE (op0))) | |
13300 | emit_insn (gen_prefetchi (op0, op1)); | |
13301 | else | |
13302 | { | |
13303 | /* Ignore the hint. */ | |
13304 | warning (0, "instruction prefetch applies when in 64-bit mode" | |
13305 | " with RIP-relative addressing and" | |
13306 | " option %<-mprefetchi%>;" | |
13307 | " they stay NOPs otherwise"); | |
13308 | emit_insn (gen_nop ()); | |
13309 | } | |
13310 | ||
13311 | return 0; | |
13312 | } | |
13313 | ||
2bf6d935 ML |
13314 | case IX86_BUILTIN_VEC_INIT_V2SI: |
13315 | case IX86_BUILTIN_VEC_INIT_V4HI: | |
13316 | case IX86_BUILTIN_VEC_INIT_V8QI: | |
13317 | return ix86_expand_vec_init_builtin (TREE_TYPE (exp), exp, target); | |
13318 | ||
13319 | case IX86_BUILTIN_VEC_EXT_V2DF: | |
13320 | case IX86_BUILTIN_VEC_EXT_V2DI: | |
13321 | case IX86_BUILTIN_VEC_EXT_V4SF: | |
13322 | case IX86_BUILTIN_VEC_EXT_V4SI: | |
13323 | case IX86_BUILTIN_VEC_EXT_V8HI: | |
13324 | case IX86_BUILTIN_VEC_EXT_V2SI: | |
13325 | case IX86_BUILTIN_VEC_EXT_V4HI: | |
13326 | case IX86_BUILTIN_VEC_EXT_V16QI: | |
13327 | return ix86_expand_vec_ext_builtin (exp, target); | |
13328 | ||
13329 | case IX86_BUILTIN_VEC_SET_V2DI: | |
13330 | case IX86_BUILTIN_VEC_SET_V4SF: | |
13331 | case IX86_BUILTIN_VEC_SET_V4SI: | |
13332 | case IX86_BUILTIN_VEC_SET_V8HI: | |
13333 | case IX86_BUILTIN_VEC_SET_V4HI: | |
13334 | case IX86_BUILTIN_VEC_SET_V16QI: | |
13335 | return ix86_expand_vec_set_builtin (exp); | |
13336 | ||
13337 | case IX86_BUILTIN_NANQ: | |
13338 | case IX86_BUILTIN_NANSQ: | |
13339 | return expand_call (exp, target, ignore); | |
13340 | ||
13341 | case IX86_BUILTIN_RDPID: | |
13342 | ||
13343 | op0 = gen_reg_rtx (word_mode); | |
13344 | ||
13345 | if (TARGET_64BIT) | |
13346 | { | |
13347 | insn = gen_rdpid_rex64 (op0); | |
13348 | op0 = convert_to_mode (SImode, op0, 1); | |
13349 | } | |
13350 | else | |
13351 | insn = gen_rdpid (op0); | |
13352 | ||
13353 | emit_insn (insn); | |
13354 | ||
13355 | if (target == 0 | |
13356 | || !register_operand (target, SImode)) | |
13357 | target = gen_reg_rtx (SImode); | |
13358 | ||
13359 | emit_move_insn (target, op0); | |
13360 | return target; | |
13361 | ||
e21b52af HL |
13362 | case IX86_BUILTIN_2INTERSECTD512: |
13363 | case IX86_BUILTIN_2INTERSECTQ512: | |
13364 | case IX86_BUILTIN_2INTERSECTD256: | |
13365 | case IX86_BUILTIN_2INTERSECTQ256: | |
13366 | case IX86_BUILTIN_2INTERSECTD128: | |
13367 | case IX86_BUILTIN_2INTERSECTQ128: | |
13368 | arg0 = CALL_EXPR_ARG (exp, 0); | |
13369 | arg1 = CALL_EXPR_ARG (exp, 1); | |
13370 | arg2 = CALL_EXPR_ARG (exp, 2); | |
13371 | arg3 = CALL_EXPR_ARG (exp, 3); | |
13372 | op0 = expand_normal (arg0); | |
13373 | op1 = expand_normal (arg1); | |
13374 | op2 = expand_normal (arg2); | |
13375 | op3 = expand_normal (arg3); | |
13376 | ||
13377 | if (!address_operand (op0, VOIDmode)) | |
13378 | { | |
13379 | op0 = convert_memory_address (Pmode, op0); | |
13380 | op0 = copy_addr_to_reg (op0); | |
13381 | } | |
13382 | if (!address_operand (op1, VOIDmode)) | |
13383 | { | |
13384 | op1 = convert_memory_address (Pmode, op1); | |
13385 | op1 = copy_addr_to_reg (op1); | |
13386 | } | |
13387 | ||
13388 | switch (fcode) | |
13389 | { | |
13390 | case IX86_BUILTIN_2INTERSECTD512: | |
13391 | mode4 = P2HImode; | |
13392 | icode = CODE_FOR_avx512vp2intersect_2intersectv16si; | |
13393 | break; | |
13394 | case IX86_BUILTIN_2INTERSECTQ512: | |
13395 | mode4 = P2QImode; | |
13396 | icode = CODE_FOR_avx512vp2intersect_2intersectv8di; | |
13397 | break; | |
13398 | case IX86_BUILTIN_2INTERSECTD256: | |
13399 | mode4 = P2QImode; | |
13400 | icode = CODE_FOR_avx512vp2intersect_2intersectv8si; | |
13401 | break; | |
13402 | case IX86_BUILTIN_2INTERSECTQ256: | |
13403 | mode4 = P2QImode; | |
13404 | icode = CODE_FOR_avx512vp2intersect_2intersectv4di; | |
13405 | break; | |
13406 | case IX86_BUILTIN_2INTERSECTD128: | |
13407 | mode4 = P2QImode; | |
13408 | icode = CODE_FOR_avx512vp2intersect_2intersectv4si; | |
13409 | break; | |
13410 | case IX86_BUILTIN_2INTERSECTQ128: | |
13411 | mode4 = P2QImode; | |
13412 | icode = CODE_FOR_avx512vp2intersect_2intersectv2di; | |
13413 | break; | |
13414 | default: | |
13415 | gcc_unreachable (); | |
13416 | } | |
13417 | ||
13418 | mode2 = insn_data[icode].operand[1].mode; | |
13419 | mode3 = insn_data[icode].operand[2].mode; | |
13420 | if (!insn_data[icode].operand[1].predicate (op2, mode2)) | |
13421 | op2 = copy_to_mode_reg (mode2, op2); | |
13422 | if (!insn_data[icode].operand[2].predicate (op3, mode3)) | |
13423 | op3 = copy_to_mode_reg (mode3, op3); | |
13424 | ||
13425 | op4 = gen_reg_rtx (mode4); | |
13426 | emit_insn (GEN_FCN (icode) (op4, op2, op3)); | |
13427 | mode0 = mode4 == P2HImode ? HImode : QImode; | |
13428 | emit_move_insn (gen_rtx_MEM (mode0, op0), | |
13429 | gen_lowpart (mode0, op4)); | |
13430 | emit_move_insn (gen_rtx_MEM (mode0, op1), | |
13431 | gen_highpart (mode0, op4)); | |
13432 | ||
13433 | return 0; | |
13434 | ||
2bf6d935 ML |
13435 | case IX86_BUILTIN_RDPMC: |
13436 | case IX86_BUILTIN_RDTSC: | |
13437 | case IX86_BUILTIN_RDTSCP: | |
13438 | case IX86_BUILTIN_XGETBV: | |
13439 | ||
13440 | op0 = gen_reg_rtx (DImode); | |
13441 | op1 = gen_reg_rtx (DImode); | |
13442 | ||
13443 | if (fcode == IX86_BUILTIN_RDPMC) | |
13444 | { | |
13445 | arg0 = CALL_EXPR_ARG (exp, 0); | |
13446 | op2 = expand_normal (arg0); | |
13447 | if (!register_operand (op2, SImode)) | |
13448 | op2 = copy_to_mode_reg (SImode, op2); | |
13449 | ||
13450 | insn = (TARGET_64BIT | |
13451 | ? gen_rdpmc_rex64 (op0, op1, op2) | |
13452 | : gen_rdpmc (op0, op2)); | |
13453 | emit_insn (insn); | |
13454 | } | |
13455 | else if (fcode == IX86_BUILTIN_XGETBV) | |
13456 | { | |
13457 | arg0 = CALL_EXPR_ARG (exp, 0); | |
13458 | op2 = expand_normal (arg0); | |
13459 | if (!register_operand (op2, SImode)) | |
13460 | op2 = copy_to_mode_reg (SImode, op2); | |
13461 | ||
13462 | insn = (TARGET_64BIT | |
13463 | ? gen_xgetbv_rex64 (op0, op1, op2) | |
13464 | : gen_xgetbv (op0, op2)); | |
13465 | emit_insn (insn); | |
13466 | } | |
13467 | else if (fcode == IX86_BUILTIN_RDTSC) | |
13468 | { | |
13469 | insn = (TARGET_64BIT | |
13470 | ? gen_rdtsc_rex64 (op0, op1) | |
13471 | : gen_rdtsc (op0)); | |
13472 | emit_insn (insn); | |
13473 | } | |
13474 | else | |
13475 | { | |
13476 | op2 = gen_reg_rtx (SImode); | |
13477 | ||
13478 | insn = (TARGET_64BIT | |
13479 | ? gen_rdtscp_rex64 (op0, op1, op2) | |
13480 | : gen_rdtscp (op0, op2)); | |
13481 | emit_insn (insn); | |
13482 | ||
13483 | arg0 = CALL_EXPR_ARG (exp, 0); | |
13484 | op4 = expand_normal (arg0); | |
13485 | if (!address_operand (op4, VOIDmode)) | |
13486 | { | |
13487 | op4 = convert_memory_address (Pmode, op4); | |
13488 | op4 = copy_addr_to_reg (op4); | |
13489 | } | |
13490 | emit_move_insn (gen_rtx_MEM (SImode, op4), op2); | |
13491 | } | |
13492 | ||
13493 | if (target == 0 | |
13494 | || !register_operand (target, DImode)) | |
13495 | target = gen_reg_rtx (DImode); | |
13496 | ||
13497 | if (TARGET_64BIT) | |
13498 | { | |
13499 | op1 = expand_simple_binop (DImode, ASHIFT, op1, GEN_INT (32), | |
13500 | op1, 1, OPTAB_DIRECT); | |
13501 | op0 = expand_simple_binop (DImode, IOR, op0, op1, | |
13502 | op0, 1, OPTAB_DIRECT); | |
13503 | } | |
13504 | ||
13505 | emit_move_insn (target, op0); | |
13506 | return target; | |
13507 | ||
6a10feda XG |
13508 | case IX86_BUILTIN_ENQCMD: |
13509 | case IX86_BUILTIN_ENQCMDS: | |
2bf6d935 ML |
13510 | case IX86_BUILTIN_MOVDIR64B: |
13511 | ||
13512 | arg0 = CALL_EXPR_ARG (exp, 0); | |
13513 | arg1 = CALL_EXPR_ARG (exp, 1); | |
13514 | op0 = expand_normal (arg0); | |
13515 | op1 = expand_normal (arg1); | |
13516 | ||
13517 | op0 = ix86_zero_extend_to_Pmode (op0); | |
13518 | if (!address_operand (op1, VOIDmode)) | |
13519 | { | |
13520 | op1 = convert_memory_address (Pmode, op1); | |
13521 | op1 = copy_addr_to_reg (op1); | |
13522 | } | |
13523 | op1 = gen_rtx_MEM (XImode, op1); | |
13524 | ||
6a10feda XG |
13525 | if (fcode == IX86_BUILTIN_MOVDIR64B) |
13526 | { | |
13527 | emit_insn (gen_movdir64b (Pmode, op0, op1)); | |
13528 | return 0; | |
13529 | } | |
13530 | else | |
13531 | { | |
44320665 UB |
13532 | if (target == 0 |
13533 | || !register_operand (target, SImode)) | |
13534 | target = gen_reg_rtx (SImode); | |
6a10feda | 13535 | |
6a10feda XG |
13536 | emit_move_insn (target, const0_rtx); |
13537 | target = gen_rtx_SUBREG (QImode, target, 0); | |
13538 | ||
44320665 UB |
13539 | int unspecv = (fcode == IX86_BUILTIN_ENQCMD |
13540 | ? UNSPECV_ENQCMD | |
13541 | : UNSPECV_ENQCMDS); | |
13542 | icode = code_for_enqcmd (unspecv, Pmode); | |
13543 | emit_insn (GEN_FCN (icode) (op0, op1)); | |
6a10feda | 13544 | |
44320665 UB |
13545 | emit_insn |
13546 | (gen_rtx_SET (gen_rtx_STRICT_LOW_PART (VOIDmode, target), | |
13547 | gen_rtx_fmt_ee (EQ, QImode, | |
13548 | gen_rtx_REG (CCZmode, FLAGS_REG), | |
13549 | const0_rtx))); | |
6a10feda XG |
13550 | return SUBREG_REG (target); |
13551 | } | |
2bf6d935 ML |
13552 | |
13553 | case IX86_BUILTIN_FXSAVE: | |
13554 | case IX86_BUILTIN_FXRSTOR: | |
13555 | case IX86_BUILTIN_FXSAVE64: | |
13556 | case IX86_BUILTIN_FXRSTOR64: | |
13557 | case IX86_BUILTIN_FNSTENV: | |
13558 | case IX86_BUILTIN_FLDENV: | |
13559 | mode0 = BLKmode; | |
13560 | switch (fcode) | |
13561 | { | |
13562 | case IX86_BUILTIN_FXSAVE: | |
13563 | icode = CODE_FOR_fxsave; | |
13564 | break; | |
13565 | case IX86_BUILTIN_FXRSTOR: | |
13566 | icode = CODE_FOR_fxrstor; | |
13567 | break; | |
13568 | case IX86_BUILTIN_FXSAVE64: | |
13569 | icode = CODE_FOR_fxsave64; | |
13570 | break; | |
13571 | case IX86_BUILTIN_FXRSTOR64: | |
13572 | icode = CODE_FOR_fxrstor64; | |
13573 | break; | |
13574 | case IX86_BUILTIN_FNSTENV: | |
13575 | icode = CODE_FOR_fnstenv; | |
13576 | break; | |
13577 | case IX86_BUILTIN_FLDENV: | |
13578 | icode = CODE_FOR_fldenv; | |
13579 | break; | |
13580 | default: | |
13581 | gcc_unreachable (); | |
13582 | } | |
13583 | ||
13584 | arg0 = CALL_EXPR_ARG (exp, 0); | |
13585 | op0 = expand_normal (arg0); | |
13586 | ||
13587 | if (!address_operand (op0, VOIDmode)) | |
13588 | { | |
13589 | op0 = convert_memory_address (Pmode, op0); | |
13590 | op0 = copy_addr_to_reg (op0); | |
13591 | } | |
13592 | op0 = gen_rtx_MEM (mode0, op0); | |
13593 | ||
13594 | pat = GEN_FCN (icode) (op0); | |
13595 | if (pat) | |
13596 | emit_insn (pat); | |
13597 | return 0; | |
13598 | ||
13599 | case IX86_BUILTIN_XSETBV: | |
13600 | arg0 = CALL_EXPR_ARG (exp, 0); | |
13601 | arg1 = CALL_EXPR_ARG (exp, 1); | |
13602 | op0 = expand_normal (arg0); | |
13603 | op1 = expand_normal (arg1); | |
13604 | ||
13605 | if (!REG_P (op0)) | |
13606 | op0 = copy_to_mode_reg (SImode, op0); | |
13607 | ||
13608 | op1 = force_reg (DImode, op1); | |
13609 | ||
13610 | if (TARGET_64BIT) | |
13611 | { | |
13612 | op2 = expand_simple_binop (DImode, LSHIFTRT, op1, GEN_INT (32), | |
13613 | NULL, 1, OPTAB_DIRECT); | |
13614 | ||
13615 | icode = CODE_FOR_xsetbv_rex64; | |
13616 | ||
13617 | op2 = gen_lowpart (SImode, op2); | |
13618 | op1 = gen_lowpart (SImode, op1); | |
13619 | pat = GEN_FCN (icode) (op0, op1, op2); | |
13620 | } | |
13621 | else | |
13622 | { | |
13623 | icode = CODE_FOR_xsetbv; | |
13624 | ||
13625 | pat = GEN_FCN (icode) (op0, op1); | |
13626 | } | |
13627 | if (pat) | |
13628 | emit_insn (pat); | |
13629 | return 0; | |
13630 | ||
13631 | case IX86_BUILTIN_XSAVE: | |
13632 | case IX86_BUILTIN_XRSTOR: | |
13633 | case IX86_BUILTIN_XSAVE64: | |
13634 | case IX86_BUILTIN_XRSTOR64: | |
13635 | case IX86_BUILTIN_XSAVEOPT: | |
13636 | case IX86_BUILTIN_XSAVEOPT64: | |
13637 | case IX86_BUILTIN_XSAVES: | |
13638 | case IX86_BUILTIN_XRSTORS: | |
13639 | case IX86_BUILTIN_XSAVES64: | |
13640 | case IX86_BUILTIN_XRSTORS64: | |
13641 | case IX86_BUILTIN_XSAVEC: | |
13642 | case IX86_BUILTIN_XSAVEC64: | |
13643 | arg0 = CALL_EXPR_ARG (exp, 0); | |
13644 | arg1 = CALL_EXPR_ARG (exp, 1); | |
13645 | op0 = expand_normal (arg0); | |
13646 | op1 = expand_normal (arg1); | |
13647 | ||
13648 | if (!address_operand (op0, VOIDmode)) | |
13649 | { | |
13650 | op0 = convert_memory_address (Pmode, op0); | |
13651 | op0 = copy_addr_to_reg (op0); | |
13652 | } | |
13653 | op0 = gen_rtx_MEM (BLKmode, op0); | |
13654 | ||
13655 | op1 = force_reg (DImode, op1); | |
13656 | ||
13657 | if (TARGET_64BIT) | |
13658 | { | |
13659 | op2 = expand_simple_binop (DImode, LSHIFTRT, op1, GEN_INT (32), | |
13660 | NULL, 1, OPTAB_DIRECT); | |
13661 | switch (fcode) | |
13662 | { | |
13663 | case IX86_BUILTIN_XSAVE: | |
13664 | icode = CODE_FOR_xsave_rex64; | |
13665 | break; | |
13666 | case IX86_BUILTIN_XRSTOR: | |
13667 | icode = CODE_FOR_xrstor_rex64; | |
13668 | break; | |
13669 | case IX86_BUILTIN_XSAVE64: | |
13670 | icode = CODE_FOR_xsave64; | |
13671 | break; | |
13672 | case IX86_BUILTIN_XRSTOR64: | |
13673 | icode = CODE_FOR_xrstor64; | |
13674 | break; | |
13675 | case IX86_BUILTIN_XSAVEOPT: | |
13676 | icode = CODE_FOR_xsaveopt_rex64; | |
13677 | break; | |
13678 | case IX86_BUILTIN_XSAVEOPT64: | |
13679 | icode = CODE_FOR_xsaveopt64; | |
13680 | break; | |
13681 | case IX86_BUILTIN_XSAVES: | |
13682 | icode = CODE_FOR_xsaves_rex64; | |
13683 | break; | |
13684 | case IX86_BUILTIN_XRSTORS: | |
13685 | icode = CODE_FOR_xrstors_rex64; | |
13686 | break; | |
13687 | case IX86_BUILTIN_XSAVES64: | |
13688 | icode = CODE_FOR_xsaves64; | |
13689 | break; | |
13690 | case IX86_BUILTIN_XRSTORS64: | |
13691 | icode = CODE_FOR_xrstors64; | |
13692 | break; | |
13693 | case IX86_BUILTIN_XSAVEC: | |
13694 | icode = CODE_FOR_xsavec_rex64; | |
13695 | break; | |
13696 | case IX86_BUILTIN_XSAVEC64: | |
13697 | icode = CODE_FOR_xsavec64; | |
13698 | break; | |
13699 | default: | |
13700 | gcc_unreachable (); | |
13701 | } | |
13702 | ||
13703 | op2 = gen_lowpart (SImode, op2); | |
13704 | op1 = gen_lowpart (SImode, op1); | |
13705 | pat = GEN_FCN (icode) (op0, op1, op2); | |
13706 | } | |
13707 | else | |
13708 | { | |
13709 | switch (fcode) | |
13710 | { | |
13711 | case IX86_BUILTIN_XSAVE: | |
13712 | icode = CODE_FOR_xsave; | |
13713 | break; | |
13714 | case IX86_BUILTIN_XRSTOR: | |
13715 | icode = CODE_FOR_xrstor; | |
13716 | break; | |
13717 | case IX86_BUILTIN_XSAVEOPT: | |
13718 | icode = CODE_FOR_xsaveopt; | |
13719 | break; | |
13720 | case IX86_BUILTIN_XSAVES: | |
13721 | icode = CODE_FOR_xsaves; | |
13722 | break; | |
13723 | case IX86_BUILTIN_XRSTORS: | |
13724 | icode = CODE_FOR_xrstors; | |
13725 | break; | |
13726 | case IX86_BUILTIN_XSAVEC: | |
13727 | icode = CODE_FOR_xsavec; | |
13728 | break; | |
13729 | default: | |
13730 | gcc_unreachable (); | |
13731 | } | |
13732 | pat = GEN_FCN (icode) (op0, op1); | |
13733 | } | |
13734 | ||
13735 | if (pat) | |
13736 | emit_insn (pat); | |
13737 | return 0; | |
13738 | ||
13739 | case IX86_BUILTIN_LLWPCB: | |
13740 | arg0 = CALL_EXPR_ARG (exp, 0); | |
13741 | op0 = expand_normal (arg0); | |
2398c206 UB |
13742 | |
13743 | if (!register_operand (op0, Pmode)) | |
2bf6d935 | 13744 | op0 = ix86_zero_extend_to_Pmode (op0); |
2398c206 | 13745 | emit_insn (gen_lwp_llwpcb (Pmode, op0)); |
2bf6d935 ML |
13746 | return 0; |
13747 | ||
13748 | case IX86_BUILTIN_SLWPCB: | |
2bf6d935 | 13749 | if (!target |
2398c206 | 13750 | || !register_operand (target, Pmode)) |
2bf6d935 | 13751 | target = gen_reg_rtx (Pmode); |
2398c206 | 13752 | emit_insn (gen_lwp_slwpcb (Pmode, target)); |
2bf6d935 ML |
13753 | return target; |
13754 | ||
2398c206 UB |
13755 | case IX86_BUILTIN_LWPVAL32: |
13756 | case IX86_BUILTIN_LWPVAL64: | |
13757 | case IX86_BUILTIN_LWPINS32: | |
13758 | case IX86_BUILTIN_LWPINS64: | |
13759 | mode = ((fcode == IX86_BUILTIN_LWPVAL32 | |
13760 | || fcode == IX86_BUILTIN_LWPINS32) | |
13761 | ? SImode : DImode); | |
13762 | ||
13763 | if (fcode == IX86_BUILTIN_LWPVAL32 | |
13764 | || fcode == IX86_BUILTIN_LWPVAL64) | |
13765 | icode = code_for_lwp_lwpval (mode); | |
13766 | else | |
13767 | icode = code_for_lwp_lwpins (mode); | |
13768 | ||
13769 | arg0 = CALL_EXPR_ARG (exp, 0); | |
13770 | arg1 = CALL_EXPR_ARG (exp, 1); | |
13771 | arg2 = CALL_EXPR_ARG (exp, 2); | |
13772 | op0 = expand_normal (arg0); | |
13773 | op1 = expand_normal (arg1); | |
13774 | op2 = expand_normal (arg2); | |
13775 | mode0 = insn_data[icode].operand[0].mode; | |
13776 | ||
13777 | if (!insn_data[icode].operand[0].predicate (op0, mode0)) | |
13778 | op0 = copy_to_mode_reg (mode0, op0); | |
13779 | if (!insn_data[icode].operand[1].predicate (op1, SImode)) | |
13780 | op1 = copy_to_mode_reg (SImode, op1); | |
13781 | ||
13782 | if (!CONST_INT_P (op2)) | |
13783 | { | |
13784 | error ("the last argument must be a 32-bit immediate"); | |
13785 | return const0_rtx; | |
13786 | } | |
13787 | ||
13788 | emit_insn (GEN_FCN (icode) (op0, op1, op2)); | |
13789 | ||
13790 | if (fcode == IX86_BUILTIN_LWPINS32 | |
13791 | || fcode == IX86_BUILTIN_LWPINS64) | |
13792 | { | |
13793 | if (target == 0 | |
13794 | || !nonimmediate_operand (target, QImode)) | |
13795 | target = gen_reg_rtx (QImode); | |
13796 | ||
13797 | pat = gen_rtx_EQ (QImode, gen_rtx_REG (CCCmode, FLAGS_REG), | |
13798 | const0_rtx); | |
13799 | emit_insn (gen_rtx_SET (target, pat)); | |
13800 | ||
13801 | return target; | |
13802 | } | |
13803 | else | |
13804 | return 0; | |
13805 | ||
2bf6d935 ML |
13806 | case IX86_BUILTIN_BEXTRI32: |
13807 | case IX86_BUILTIN_BEXTRI64: | |
9e026191 UB |
13808 | mode = (fcode == IX86_BUILTIN_BEXTRI32 ? SImode : DImode); |
13809 | ||
2bf6d935 ML |
13810 | arg0 = CALL_EXPR_ARG (exp, 0); |
13811 | arg1 = CALL_EXPR_ARG (exp, 1); | |
13812 | op0 = expand_normal (arg0); | |
13813 | op1 = expand_normal (arg1); | |
9e026191 | 13814 | |
2bf6d935 | 13815 | if (!CONST_INT_P (op1)) |
9e026191 UB |
13816 | { |
13817 | error ("last argument must be an immediate"); | |
13818 | return const0_rtx; | |
13819 | } | |
2bf6d935 | 13820 | else |
9e026191 UB |
13821 | { |
13822 | unsigned char lsb_index = UINTVAL (op1); | |
13823 | unsigned char length = UINTVAL (op1) >> 8; | |
13824 | ||
13825 | unsigned char bitsize = GET_MODE_BITSIZE (mode); | |
13826 | ||
13827 | icode = code_for_tbm_bextri (mode); | |
2bf6d935 ML |
13828 | |
13829 | mode1 = insn_data[icode].operand[1].mode; | |
13830 | if (!insn_data[icode].operand[1].predicate (op0, mode1)) | |
13831 | op0 = copy_to_mode_reg (mode1, op0); | |
13832 | ||
13833 | mode0 = insn_data[icode].operand[0].mode; | |
13834 | if (target == 0 | |
13835 | || !register_operand (target, mode0)) | |
13836 | target = gen_reg_rtx (mode0); | |
13837 | ||
9e026191 UB |
13838 | if (length == 0 || lsb_index >= bitsize) |
13839 | { | |
13840 | emit_move_insn (target, const0_rtx); | |
13841 | return target; | |
13842 | } | |
13843 | ||
13844 | if (length + lsb_index > bitsize) | |
13845 | length = bitsize - lsb_index; | |
13846 | ||
13847 | op1 = GEN_INT (length); | |
13848 | op2 = GEN_INT (lsb_index); | |
13849 | ||
13850 | emit_insn (GEN_FCN (icode) (target, op0, op1, op2)); | |
13851 | return target; | |
13852 | } | |
2bf6d935 ML |
13853 | |
13854 | case IX86_BUILTIN_RDRAND16_STEP: | |
9e026191 | 13855 | mode = HImode; |
2bf6d935 ML |
13856 | goto rdrand_step; |
13857 | ||
13858 | case IX86_BUILTIN_RDRAND32_STEP: | |
9e026191 | 13859 | mode = SImode; |
2bf6d935 ML |
13860 | goto rdrand_step; |
13861 | ||
13862 | case IX86_BUILTIN_RDRAND64_STEP: | |
9e026191 | 13863 | mode = DImode; |
2bf6d935 ML |
13864 | |
13865 | rdrand_step: | |
13866 | arg0 = CALL_EXPR_ARG (exp, 0); | |
13867 | op1 = expand_normal (arg0); | |
13868 | if (!address_operand (op1, VOIDmode)) | |
13869 | { | |
13870 | op1 = convert_memory_address (Pmode, op1); | |
13871 | op1 = copy_addr_to_reg (op1); | |
13872 | } | |
13873 | ||
9e026191 UB |
13874 | op0 = gen_reg_rtx (mode); |
13875 | emit_insn (gen_rdrand (mode, op0)); | |
2bf6d935 | 13876 | |
9e026191 | 13877 | emit_move_insn (gen_rtx_MEM (mode, op1), op0); |
2bf6d935 | 13878 | |
9e026191 | 13879 | op1 = force_reg (SImode, const1_rtx); |
2bf6d935 ML |
13880 | |
13881 | /* Emit SImode conditional move. */ | |
9e026191 | 13882 | if (mode == HImode) |
2bf6d935 ML |
13883 | { |
13884 | if (TARGET_ZERO_EXTEND_WITH_AND | |
13885 | && optimize_function_for_speed_p (cfun)) | |
13886 | { | |
13887 | op2 = force_reg (SImode, const0_rtx); | |
13888 | ||
13889 | emit_insn (gen_movstricthi | |
13890 | (gen_lowpart (HImode, op2), op0)); | |
13891 | } | |
13892 | else | |
13893 | { | |
13894 | op2 = gen_reg_rtx (SImode); | |
13895 | ||
13896 | emit_insn (gen_zero_extendhisi2 (op2, op0)); | |
13897 | } | |
13898 | } | |
9e026191 | 13899 | else if (mode == SImode) |
2bf6d935 ML |
13900 | op2 = op0; |
13901 | else | |
13902 | op2 = gen_rtx_SUBREG (SImode, op0, 0); | |
13903 | ||
13904 | if (target == 0 | |
13905 | || !register_operand (target, SImode)) | |
13906 | target = gen_reg_rtx (SImode); | |
13907 | ||
13908 | pat = gen_rtx_GEU (VOIDmode, gen_rtx_REG (CCCmode, FLAGS_REG), | |
13909 | const0_rtx); | |
13910 | emit_insn (gen_rtx_SET (target, | |
13911 | gen_rtx_IF_THEN_ELSE (SImode, pat, op2, op1))); | |
13912 | return target; | |
13913 | ||
13914 | case IX86_BUILTIN_RDSEED16_STEP: | |
9e026191 | 13915 | mode = HImode; |
2bf6d935 ML |
13916 | goto rdseed_step; |
13917 | ||
13918 | case IX86_BUILTIN_RDSEED32_STEP: | |
9e026191 | 13919 | mode = SImode; |
2bf6d935 ML |
13920 | goto rdseed_step; |
13921 | ||
13922 | case IX86_BUILTIN_RDSEED64_STEP: | |
9e026191 | 13923 | mode = DImode; |
2bf6d935 ML |
13924 | |
13925 | rdseed_step: | |
13926 | arg0 = CALL_EXPR_ARG (exp, 0); | |
13927 | op1 = expand_normal (arg0); | |
13928 | if (!address_operand (op1, VOIDmode)) | |
13929 | { | |
13930 | op1 = convert_memory_address (Pmode, op1); | |
13931 | op1 = copy_addr_to_reg (op1); | |
13932 | } | |
13933 | ||
9e026191 UB |
13934 | op0 = gen_reg_rtx (mode); |
13935 | emit_insn (gen_rdseed (mode, op0)); | |
2bf6d935 | 13936 | |
9e026191 | 13937 | emit_move_insn (gen_rtx_MEM (mode, op1), op0); |
2bf6d935 ML |
13938 | |
13939 | op2 = gen_reg_rtx (QImode); | |
13940 | ||
13941 | pat = gen_rtx_LTU (QImode, gen_rtx_REG (CCCmode, FLAGS_REG), | |
13942 | const0_rtx); | |
13943 | emit_insn (gen_rtx_SET (op2, pat)); | |
13944 | ||
13945 | if (target == 0 | |
13946 | || !register_operand (target, SImode)) | |
13947 | target = gen_reg_rtx (SImode); | |
13948 | ||
13949 | emit_insn (gen_zero_extendqisi2 (target, op2)); | |
13950 | return target; | |
13951 | ||
13952 | case IX86_BUILTIN_SBB32: | |
13953 | icode = CODE_FOR_subborrowsi; | |
13954 | icode2 = CODE_FOR_subborrowsi_0; | |
13955 | mode0 = SImode; | |
13956 | mode1 = DImode; | |
13957 | mode2 = CCmode; | |
13958 | goto handlecarry; | |
13959 | ||
13960 | case IX86_BUILTIN_SBB64: | |
13961 | icode = CODE_FOR_subborrowdi; | |
13962 | icode2 = CODE_FOR_subborrowdi_0; | |
13963 | mode0 = DImode; | |
13964 | mode1 = TImode; | |
13965 | mode2 = CCmode; | |
13966 | goto handlecarry; | |
13967 | ||
13968 | case IX86_BUILTIN_ADDCARRYX32: | |
13969 | icode = CODE_FOR_addcarrysi; | |
13970 | icode2 = CODE_FOR_addcarrysi_0; | |
13971 | mode0 = SImode; | |
13972 | mode1 = DImode; | |
13973 | mode2 = CCCmode; | |
13974 | goto handlecarry; | |
13975 | ||
13976 | case IX86_BUILTIN_ADDCARRYX64: | |
13977 | icode = CODE_FOR_addcarrydi; | |
13978 | icode2 = CODE_FOR_addcarrydi_0; | |
13979 | mode0 = DImode; | |
13980 | mode1 = TImode; | |
13981 | mode2 = CCCmode; | |
13982 | ||
13983 | handlecarry: | |
13984 | arg0 = CALL_EXPR_ARG (exp, 0); /* unsigned char c_in. */ | |
13985 | arg1 = CALL_EXPR_ARG (exp, 1); /* unsigned int src1. */ | |
13986 | arg2 = CALL_EXPR_ARG (exp, 2); /* unsigned int src2. */ | |
13987 | arg3 = CALL_EXPR_ARG (exp, 3); /* unsigned int *sum_out. */ | |
13988 | ||
13989 | op1 = expand_normal (arg0); | |
2bf6d935 ML |
13990 | |
13991 | op2 = expand_normal (arg1); | |
13992 | if (!register_operand (op2, mode0)) | |
13993 | op2 = copy_to_mode_reg (mode0, op2); | |
13994 | ||
13995 | op3 = expand_normal (arg2); | |
13996 | if (!register_operand (op3, mode0)) | |
13997 | op3 = copy_to_mode_reg (mode0, op3); | |
13998 | ||
13999 | op4 = expand_normal (arg3); | |
14000 | if (!address_operand (op4, VOIDmode)) | |
14001 | { | |
14002 | op4 = convert_memory_address (Pmode, op4); | |
14003 | op4 = copy_addr_to_reg (op4); | |
14004 | } | |
14005 | ||
14006 | op0 = gen_reg_rtx (mode0); | |
eba3565c | 14007 | if (op1 == const0_rtx) |
2bf6d935 ML |
14008 | { |
14009 | /* If arg0 is 0, optimize right away into add or sub | |
14010 | instruction that sets CCCmode flags. */ | |
14011 | op1 = gen_rtx_REG (mode2, FLAGS_REG); | |
14012 | emit_insn (GEN_FCN (icode2) (op0, op2, op3)); | |
14013 | } | |
14014 | else | |
14015 | { | |
14016 | /* Generate CF from input operand. */ | |
af29d0d6 | 14017 | ix86_expand_carry (op1); |
2bf6d935 ML |
14018 | |
14019 | /* Generate instruction that consumes CF. */ | |
14020 | op1 = gen_rtx_REG (CCCmode, FLAGS_REG); | |
14021 | pat = gen_rtx_LTU (mode1, op1, const0_rtx); | |
14022 | pat2 = gen_rtx_LTU (mode0, op1, const0_rtx); | |
14023 | emit_insn (GEN_FCN (icode) (op0, op2, op3, op1, pat, pat2)); | |
14024 | } | |
14025 | ||
14026 | /* Return current CF value. */ | |
14027 | if (target == 0) | |
14028 | target = gen_reg_rtx (QImode); | |
14029 | ||
14030 | pat = gen_rtx_LTU (QImode, op1, const0_rtx); | |
14031 | emit_insn (gen_rtx_SET (target, pat)); | |
14032 | ||
14033 | /* Store the result. */ | |
14034 | emit_move_insn (gen_rtx_MEM (mode0, op4), op0); | |
14035 | ||
14036 | return target; | |
14037 | ||
14038 | case IX86_BUILTIN_READ_FLAGS: | |
b60bc913 JJ |
14039 | if (ignore) |
14040 | return const0_rtx; | |
14041 | ||
2bf6d935 ML |
14042 | emit_insn (gen_push (gen_rtx_REG (word_mode, FLAGS_REG))); |
14043 | ||
14044 | if (optimize | |
14045 | || target == NULL_RTX | |
14046 | || !nonimmediate_operand (target, word_mode) | |
14047 | || GET_MODE (target) != word_mode) | |
14048 | target = gen_reg_rtx (word_mode); | |
14049 | ||
14050 | emit_insn (gen_pop (target)); | |
14051 | return target; | |
14052 | ||
14053 | case IX86_BUILTIN_WRITE_FLAGS: | |
14054 | ||
14055 | arg0 = CALL_EXPR_ARG (exp, 0); | |
14056 | op0 = expand_normal (arg0); | |
14057 | if (!general_no_elim_operand (op0, word_mode)) | |
14058 | op0 = copy_to_mode_reg (word_mode, op0); | |
14059 | ||
14060 | emit_insn (gen_push (op0)); | |
14061 | emit_insn (gen_pop (gen_rtx_REG (word_mode, FLAGS_REG))); | |
14062 | return 0; | |
14063 | ||
14064 | case IX86_BUILTIN_KTESTC8: | |
14065 | icode = CODE_FOR_ktestqi; | |
14066 | mode3 = CCCmode; | |
14067 | goto kortest; | |
14068 | ||
14069 | case IX86_BUILTIN_KTESTZ8: | |
14070 | icode = CODE_FOR_ktestqi; | |
14071 | mode3 = CCZmode; | |
14072 | goto kortest; | |
14073 | ||
14074 | case IX86_BUILTIN_KTESTC16: | |
14075 | icode = CODE_FOR_ktesthi; | |
14076 | mode3 = CCCmode; | |
14077 | goto kortest; | |
14078 | ||
14079 | case IX86_BUILTIN_KTESTZ16: | |
14080 | icode = CODE_FOR_ktesthi; | |
14081 | mode3 = CCZmode; | |
14082 | goto kortest; | |
14083 | ||
14084 | case IX86_BUILTIN_KTESTC32: | |
14085 | icode = CODE_FOR_ktestsi; | |
14086 | mode3 = CCCmode; | |
14087 | goto kortest; | |
14088 | ||
14089 | case IX86_BUILTIN_KTESTZ32: | |
14090 | icode = CODE_FOR_ktestsi; | |
14091 | mode3 = CCZmode; | |
14092 | goto kortest; | |
14093 | ||
14094 | case IX86_BUILTIN_KTESTC64: | |
14095 | icode = CODE_FOR_ktestdi; | |
14096 | mode3 = CCCmode; | |
14097 | goto kortest; | |
14098 | ||
14099 | case IX86_BUILTIN_KTESTZ64: | |
14100 | icode = CODE_FOR_ktestdi; | |
14101 | mode3 = CCZmode; | |
14102 | goto kortest; | |
14103 | ||
14104 | case IX86_BUILTIN_KORTESTC8: | |
14105 | icode = CODE_FOR_kortestqi; | |
14106 | mode3 = CCCmode; | |
14107 | goto kortest; | |
14108 | ||
14109 | case IX86_BUILTIN_KORTESTZ8: | |
14110 | icode = CODE_FOR_kortestqi; | |
14111 | mode3 = CCZmode; | |
14112 | goto kortest; | |
14113 | ||
14114 | case IX86_BUILTIN_KORTESTC16: | |
14115 | icode = CODE_FOR_kortesthi; | |
14116 | mode3 = CCCmode; | |
14117 | goto kortest; | |
14118 | ||
14119 | case IX86_BUILTIN_KORTESTZ16: | |
14120 | icode = CODE_FOR_kortesthi; | |
14121 | mode3 = CCZmode; | |
14122 | goto kortest; | |
14123 | ||
14124 | case IX86_BUILTIN_KORTESTC32: | |
14125 | icode = CODE_FOR_kortestsi; | |
14126 | mode3 = CCCmode; | |
14127 | goto kortest; | |
14128 | ||
14129 | case IX86_BUILTIN_KORTESTZ32: | |
14130 | icode = CODE_FOR_kortestsi; | |
14131 | mode3 = CCZmode; | |
14132 | goto kortest; | |
14133 | ||
14134 | case IX86_BUILTIN_KORTESTC64: | |
14135 | icode = CODE_FOR_kortestdi; | |
14136 | mode3 = CCCmode; | |
14137 | goto kortest; | |
14138 | ||
14139 | case IX86_BUILTIN_KORTESTZ64: | |
14140 | icode = CODE_FOR_kortestdi; | |
14141 | mode3 = CCZmode; | |
14142 | ||
14143 | kortest: | |
14144 | arg0 = CALL_EXPR_ARG (exp, 0); /* Mask reg src1. */ | |
14145 | arg1 = CALL_EXPR_ARG (exp, 1); /* Mask reg src2. */ | |
14146 | op0 = expand_normal (arg0); | |
14147 | op1 = expand_normal (arg1); | |
14148 | ||
14149 | mode0 = insn_data[icode].operand[0].mode; | |
14150 | mode1 = insn_data[icode].operand[1].mode; | |
14151 | ||
14152 | if (GET_MODE (op0) != VOIDmode) | |
14153 | op0 = force_reg (GET_MODE (op0), op0); | |
14154 | ||
14155 | op0 = gen_lowpart (mode0, op0); | |
14156 | ||
14157 | if (!insn_data[icode].operand[0].predicate (op0, mode0)) | |
14158 | op0 = copy_to_mode_reg (mode0, op0); | |
14159 | ||
14160 | if (GET_MODE (op1) != VOIDmode) | |
14161 | op1 = force_reg (GET_MODE (op1), op1); | |
14162 | ||
14163 | op1 = gen_lowpart (mode1, op1); | |
14164 | ||
14165 | if (!insn_data[icode].operand[1].predicate (op1, mode1)) | |
14166 | op1 = copy_to_mode_reg (mode1, op1); | |
14167 | ||
14168 | target = gen_reg_rtx (QImode); | |
14169 | ||
14170 | /* Emit kortest. */ | |
14171 | emit_insn (GEN_FCN (icode) (op0, op1)); | |
14172 | /* And use setcc to return result from flags. */ | |
14173 | ix86_expand_setcc (target, EQ, | |
14174 | gen_rtx_REG (mode3, FLAGS_REG), const0_rtx); | |
14175 | return target; | |
14176 | ||
14177 | case IX86_BUILTIN_GATHERSIV2DF: | |
14178 | icode = CODE_FOR_avx2_gathersiv2df; | |
14179 | goto gather_gen; | |
14180 | case IX86_BUILTIN_GATHERSIV4DF: | |
14181 | icode = CODE_FOR_avx2_gathersiv4df; | |
14182 | goto gather_gen; | |
14183 | case IX86_BUILTIN_GATHERDIV2DF: | |
14184 | icode = CODE_FOR_avx2_gatherdiv2df; | |
14185 | goto gather_gen; | |
14186 | case IX86_BUILTIN_GATHERDIV4DF: | |
14187 | icode = CODE_FOR_avx2_gatherdiv4df; | |
14188 | goto gather_gen; | |
14189 | case IX86_BUILTIN_GATHERSIV4SF: | |
14190 | icode = CODE_FOR_avx2_gathersiv4sf; | |
14191 | goto gather_gen; | |
14192 | case IX86_BUILTIN_GATHERSIV8SF: | |
14193 | icode = CODE_FOR_avx2_gathersiv8sf; | |
14194 | goto gather_gen; | |
14195 | case IX86_BUILTIN_GATHERDIV4SF: | |
14196 | icode = CODE_FOR_avx2_gatherdiv4sf; | |
14197 | goto gather_gen; | |
14198 | case IX86_BUILTIN_GATHERDIV8SF: | |
14199 | icode = CODE_FOR_avx2_gatherdiv8sf; | |
14200 | goto gather_gen; | |
14201 | case IX86_BUILTIN_GATHERSIV2DI: | |
14202 | icode = CODE_FOR_avx2_gathersiv2di; | |
14203 | goto gather_gen; | |
14204 | case IX86_BUILTIN_GATHERSIV4DI: | |
14205 | icode = CODE_FOR_avx2_gathersiv4di; | |
14206 | goto gather_gen; | |
14207 | case IX86_BUILTIN_GATHERDIV2DI: | |
14208 | icode = CODE_FOR_avx2_gatherdiv2di; | |
14209 | goto gather_gen; | |
14210 | case IX86_BUILTIN_GATHERDIV4DI: | |
14211 | icode = CODE_FOR_avx2_gatherdiv4di; | |
14212 | goto gather_gen; | |
14213 | case IX86_BUILTIN_GATHERSIV4SI: | |
14214 | icode = CODE_FOR_avx2_gathersiv4si; | |
14215 | goto gather_gen; | |
14216 | case IX86_BUILTIN_GATHERSIV8SI: | |
14217 | icode = CODE_FOR_avx2_gathersiv8si; | |
14218 | goto gather_gen; | |
14219 | case IX86_BUILTIN_GATHERDIV4SI: | |
14220 | icode = CODE_FOR_avx2_gatherdiv4si; | |
14221 | goto gather_gen; | |
14222 | case IX86_BUILTIN_GATHERDIV8SI: | |
14223 | icode = CODE_FOR_avx2_gatherdiv8si; | |
14224 | goto gather_gen; | |
14225 | case IX86_BUILTIN_GATHERALTSIV4DF: | |
14226 | icode = CODE_FOR_avx2_gathersiv4df; | |
14227 | goto gather_gen; | |
14228 | case IX86_BUILTIN_GATHERALTDIV8SF: | |
14229 | icode = CODE_FOR_avx2_gatherdiv8sf; | |
14230 | goto gather_gen; | |
14231 | case IX86_BUILTIN_GATHERALTSIV4DI: | |
14232 | icode = CODE_FOR_avx2_gathersiv4di; | |
14233 | goto gather_gen; | |
14234 | case IX86_BUILTIN_GATHERALTDIV8SI: | |
14235 | icode = CODE_FOR_avx2_gatherdiv8si; | |
14236 | goto gather_gen; | |
14237 | case IX86_BUILTIN_GATHER3SIV16SF: | |
14238 | icode = CODE_FOR_avx512f_gathersiv16sf; | |
14239 | goto gather_gen; | |
14240 | case IX86_BUILTIN_GATHER3SIV8DF: | |
14241 | icode = CODE_FOR_avx512f_gathersiv8df; | |
14242 | goto gather_gen; | |
14243 | case IX86_BUILTIN_GATHER3DIV16SF: | |
14244 | icode = CODE_FOR_avx512f_gatherdiv16sf; | |
14245 | goto gather_gen; | |
14246 | case IX86_BUILTIN_GATHER3DIV8DF: | |
14247 | icode = CODE_FOR_avx512f_gatherdiv8df; | |
14248 | goto gather_gen; | |
14249 | case IX86_BUILTIN_GATHER3SIV16SI: | |
14250 | icode = CODE_FOR_avx512f_gathersiv16si; | |
14251 | goto gather_gen; | |
14252 | case IX86_BUILTIN_GATHER3SIV8DI: | |
14253 | icode = CODE_FOR_avx512f_gathersiv8di; | |
14254 | goto gather_gen; | |
14255 | case IX86_BUILTIN_GATHER3DIV16SI: | |
14256 | icode = CODE_FOR_avx512f_gatherdiv16si; | |
14257 | goto gather_gen; | |
14258 | case IX86_BUILTIN_GATHER3DIV8DI: | |
14259 | icode = CODE_FOR_avx512f_gatherdiv8di; | |
14260 | goto gather_gen; | |
14261 | case IX86_BUILTIN_GATHER3ALTSIV8DF: | |
14262 | icode = CODE_FOR_avx512f_gathersiv8df; | |
14263 | goto gather_gen; | |
14264 | case IX86_BUILTIN_GATHER3ALTDIV16SF: | |
14265 | icode = CODE_FOR_avx512f_gatherdiv16sf; | |
14266 | goto gather_gen; | |
14267 | case IX86_BUILTIN_GATHER3ALTSIV8DI: | |
14268 | icode = CODE_FOR_avx512f_gathersiv8di; | |
14269 | goto gather_gen; | |
14270 | case IX86_BUILTIN_GATHER3ALTDIV16SI: | |
14271 | icode = CODE_FOR_avx512f_gatherdiv16si; | |
14272 | goto gather_gen; | |
14273 | case IX86_BUILTIN_GATHER3SIV2DF: | |
14274 | icode = CODE_FOR_avx512vl_gathersiv2df; | |
14275 | goto gather_gen; | |
14276 | case IX86_BUILTIN_GATHER3SIV4DF: | |
14277 | icode = CODE_FOR_avx512vl_gathersiv4df; | |
14278 | goto gather_gen; | |
14279 | case IX86_BUILTIN_GATHER3DIV2DF: | |
14280 | icode = CODE_FOR_avx512vl_gatherdiv2df; | |
14281 | goto gather_gen; | |
14282 | case IX86_BUILTIN_GATHER3DIV4DF: | |
14283 | icode = CODE_FOR_avx512vl_gatherdiv4df; | |
14284 | goto gather_gen; | |
14285 | case IX86_BUILTIN_GATHER3SIV4SF: | |
14286 | icode = CODE_FOR_avx512vl_gathersiv4sf; | |
14287 | goto gather_gen; | |
14288 | case IX86_BUILTIN_GATHER3SIV8SF: | |
14289 | icode = CODE_FOR_avx512vl_gathersiv8sf; | |
14290 | goto gather_gen; | |
14291 | case IX86_BUILTIN_GATHER3DIV4SF: | |
14292 | icode = CODE_FOR_avx512vl_gatherdiv4sf; | |
14293 | goto gather_gen; | |
14294 | case IX86_BUILTIN_GATHER3DIV8SF: | |
14295 | icode = CODE_FOR_avx512vl_gatherdiv8sf; | |
14296 | goto gather_gen; | |
14297 | case IX86_BUILTIN_GATHER3SIV2DI: | |
14298 | icode = CODE_FOR_avx512vl_gathersiv2di; | |
14299 | goto gather_gen; | |
14300 | case IX86_BUILTIN_GATHER3SIV4DI: | |
14301 | icode = CODE_FOR_avx512vl_gathersiv4di; | |
14302 | goto gather_gen; | |
14303 | case IX86_BUILTIN_GATHER3DIV2DI: | |
14304 | icode = CODE_FOR_avx512vl_gatherdiv2di; | |
14305 | goto gather_gen; | |
14306 | case IX86_BUILTIN_GATHER3DIV4DI: | |
14307 | icode = CODE_FOR_avx512vl_gatherdiv4di; | |
14308 | goto gather_gen; | |
14309 | case IX86_BUILTIN_GATHER3SIV4SI: | |
14310 | icode = CODE_FOR_avx512vl_gathersiv4si; | |
14311 | goto gather_gen; | |
14312 | case IX86_BUILTIN_GATHER3SIV8SI: | |
14313 | icode = CODE_FOR_avx512vl_gathersiv8si; | |
14314 | goto gather_gen; | |
14315 | case IX86_BUILTIN_GATHER3DIV4SI: | |
14316 | icode = CODE_FOR_avx512vl_gatherdiv4si; | |
14317 | goto gather_gen; | |
14318 | case IX86_BUILTIN_GATHER3DIV8SI: | |
14319 | icode = CODE_FOR_avx512vl_gatherdiv8si; | |
14320 | goto gather_gen; | |
14321 | case IX86_BUILTIN_GATHER3ALTSIV4DF: | |
14322 | icode = CODE_FOR_avx512vl_gathersiv4df; | |
14323 | goto gather_gen; | |
14324 | case IX86_BUILTIN_GATHER3ALTDIV8SF: | |
14325 | icode = CODE_FOR_avx512vl_gatherdiv8sf; | |
14326 | goto gather_gen; | |
14327 | case IX86_BUILTIN_GATHER3ALTSIV4DI: | |
14328 | icode = CODE_FOR_avx512vl_gathersiv4di; | |
14329 | goto gather_gen; | |
14330 | case IX86_BUILTIN_GATHER3ALTDIV8SI: | |
14331 | icode = CODE_FOR_avx512vl_gatherdiv8si; | |
14332 | goto gather_gen; | |
14333 | case IX86_BUILTIN_SCATTERSIV16SF: | |
14334 | icode = CODE_FOR_avx512f_scattersiv16sf; | |
14335 | goto scatter_gen; | |
14336 | case IX86_BUILTIN_SCATTERSIV8DF: | |
14337 | icode = CODE_FOR_avx512f_scattersiv8df; | |
14338 | goto scatter_gen; | |
14339 | case IX86_BUILTIN_SCATTERDIV16SF: | |
14340 | icode = CODE_FOR_avx512f_scatterdiv16sf; | |
14341 | goto scatter_gen; | |
14342 | case IX86_BUILTIN_SCATTERDIV8DF: | |
14343 | icode = CODE_FOR_avx512f_scatterdiv8df; | |
14344 | goto scatter_gen; | |
14345 | case IX86_BUILTIN_SCATTERSIV16SI: | |
14346 | icode = CODE_FOR_avx512f_scattersiv16si; | |
14347 | goto scatter_gen; | |
14348 | case IX86_BUILTIN_SCATTERSIV8DI: | |
14349 | icode = CODE_FOR_avx512f_scattersiv8di; | |
14350 | goto scatter_gen; | |
14351 | case IX86_BUILTIN_SCATTERDIV16SI: | |
14352 | icode = CODE_FOR_avx512f_scatterdiv16si; | |
14353 | goto scatter_gen; | |
14354 | case IX86_BUILTIN_SCATTERDIV8DI: | |
14355 | icode = CODE_FOR_avx512f_scatterdiv8di; | |
14356 | goto scatter_gen; | |
14357 | case IX86_BUILTIN_SCATTERSIV8SF: | |
14358 | icode = CODE_FOR_avx512vl_scattersiv8sf; | |
14359 | goto scatter_gen; | |
14360 | case IX86_BUILTIN_SCATTERSIV4SF: | |
14361 | icode = CODE_FOR_avx512vl_scattersiv4sf; | |
14362 | goto scatter_gen; | |
14363 | case IX86_BUILTIN_SCATTERSIV4DF: | |
14364 | icode = CODE_FOR_avx512vl_scattersiv4df; | |
14365 | goto scatter_gen; | |
14366 | case IX86_BUILTIN_SCATTERSIV2DF: | |
14367 | icode = CODE_FOR_avx512vl_scattersiv2df; | |
14368 | goto scatter_gen; | |
14369 | case IX86_BUILTIN_SCATTERDIV8SF: | |
14370 | icode = CODE_FOR_avx512vl_scatterdiv8sf; | |
14371 | goto scatter_gen; | |
14372 | case IX86_BUILTIN_SCATTERDIV4SF: | |
14373 | icode = CODE_FOR_avx512vl_scatterdiv4sf; | |
14374 | goto scatter_gen; | |
14375 | case IX86_BUILTIN_SCATTERDIV4DF: | |
14376 | icode = CODE_FOR_avx512vl_scatterdiv4df; | |
14377 | goto scatter_gen; | |
14378 | case IX86_BUILTIN_SCATTERDIV2DF: | |
14379 | icode = CODE_FOR_avx512vl_scatterdiv2df; | |
14380 | goto scatter_gen; | |
14381 | case IX86_BUILTIN_SCATTERSIV8SI: | |
14382 | icode = CODE_FOR_avx512vl_scattersiv8si; | |
14383 | goto scatter_gen; | |
14384 | case IX86_BUILTIN_SCATTERSIV4SI: | |
14385 | icode = CODE_FOR_avx512vl_scattersiv4si; | |
14386 | goto scatter_gen; | |
14387 | case IX86_BUILTIN_SCATTERSIV4DI: | |
14388 | icode = CODE_FOR_avx512vl_scattersiv4di; | |
14389 | goto scatter_gen; | |
14390 | case IX86_BUILTIN_SCATTERSIV2DI: | |
14391 | icode = CODE_FOR_avx512vl_scattersiv2di; | |
14392 | goto scatter_gen; | |
14393 | case IX86_BUILTIN_SCATTERDIV8SI: | |
14394 | icode = CODE_FOR_avx512vl_scatterdiv8si; | |
14395 | goto scatter_gen; | |
14396 | case IX86_BUILTIN_SCATTERDIV4SI: | |
14397 | icode = CODE_FOR_avx512vl_scatterdiv4si; | |
14398 | goto scatter_gen; | |
14399 | case IX86_BUILTIN_SCATTERDIV4DI: | |
14400 | icode = CODE_FOR_avx512vl_scatterdiv4di; | |
14401 | goto scatter_gen; | |
14402 | case IX86_BUILTIN_SCATTERDIV2DI: | |
14403 | icode = CODE_FOR_avx512vl_scatterdiv2di; | |
14404 | goto scatter_gen; | |
14405 | case IX86_BUILTIN_GATHERPFDPD: | |
14406 | icode = CODE_FOR_avx512pf_gatherpfv8sidf; | |
14407 | goto vec_prefetch_gen; | |
14408 | case IX86_BUILTIN_SCATTERALTSIV8DF: | |
14409 | icode = CODE_FOR_avx512f_scattersiv8df; | |
14410 | goto scatter_gen; | |
14411 | case IX86_BUILTIN_SCATTERALTDIV16SF: | |
14412 | icode = CODE_FOR_avx512f_scatterdiv16sf; | |
14413 | goto scatter_gen; | |
14414 | case IX86_BUILTIN_SCATTERALTSIV8DI: | |
14415 | icode = CODE_FOR_avx512f_scattersiv8di; | |
14416 | goto scatter_gen; | |
14417 | case IX86_BUILTIN_SCATTERALTDIV16SI: | |
14418 | icode = CODE_FOR_avx512f_scatterdiv16si; | |
14419 | goto scatter_gen; | |
14420 | case IX86_BUILTIN_SCATTERALTSIV4DF: | |
14421 | icode = CODE_FOR_avx512vl_scattersiv4df; | |
14422 | goto scatter_gen; | |
14423 | case IX86_BUILTIN_SCATTERALTDIV8SF: | |
14424 | icode = CODE_FOR_avx512vl_scatterdiv8sf; | |
14425 | goto scatter_gen; | |
14426 | case IX86_BUILTIN_SCATTERALTSIV4DI: | |
14427 | icode = CODE_FOR_avx512vl_scattersiv4di; | |
14428 | goto scatter_gen; | |
14429 | case IX86_BUILTIN_SCATTERALTDIV8SI: | |
14430 | icode = CODE_FOR_avx512vl_scatterdiv8si; | |
14431 | goto scatter_gen; | |
14432 | case IX86_BUILTIN_SCATTERALTSIV2DF: | |
14433 | icode = CODE_FOR_avx512vl_scattersiv2df; | |
14434 | goto scatter_gen; | |
14435 | case IX86_BUILTIN_SCATTERALTDIV4SF: | |
14436 | icode = CODE_FOR_avx512vl_scatterdiv4sf; | |
14437 | goto scatter_gen; | |
14438 | case IX86_BUILTIN_SCATTERALTSIV2DI: | |
14439 | icode = CODE_FOR_avx512vl_scattersiv2di; | |
14440 | goto scatter_gen; | |
14441 | case IX86_BUILTIN_SCATTERALTDIV4SI: | |
14442 | icode = CODE_FOR_avx512vl_scatterdiv4si; | |
14443 | goto scatter_gen; | |
14444 | case IX86_BUILTIN_GATHERPFDPS: | |
14445 | icode = CODE_FOR_avx512pf_gatherpfv16sisf; | |
14446 | goto vec_prefetch_gen; | |
14447 | case IX86_BUILTIN_GATHERPFQPD: | |
14448 | icode = CODE_FOR_avx512pf_gatherpfv8didf; | |
14449 | goto vec_prefetch_gen; | |
14450 | case IX86_BUILTIN_GATHERPFQPS: | |
14451 | icode = CODE_FOR_avx512pf_gatherpfv8disf; | |
14452 | goto vec_prefetch_gen; | |
14453 | case IX86_BUILTIN_SCATTERPFDPD: | |
14454 | icode = CODE_FOR_avx512pf_scatterpfv8sidf; | |
14455 | goto vec_prefetch_gen; | |
14456 | case IX86_BUILTIN_SCATTERPFDPS: | |
14457 | icode = CODE_FOR_avx512pf_scatterpfv16sisf; | |
14458 | goto vec_prefetch_gen; | |
14459 | case IX86_BUILTIN_SCATTERPFQPD: | |
14460 | icode = CODE_FOR_avx512pf_scatterpfv8didf; | |
14461 | goto vec_prefetch_gen; | |
14462 | case IX86_BUILTIN_SCATTERPFQPS: | |
14463 | icode = CODE_FOR_avx512pf_scatterpfv8disf; | |
14464 | goto vec_prefetch_gen; | |
14465 | ||
14466 | gather_gen: | |
14467 | rtx half; | |
14468 | rtx (*gen) (rtx, rtx); | |
14469 | ||
14470 | arg0 = CALL_EXPR_ARG (exp, 0); | |
14471 | arg1 = CALL_EXPR_ARG (exp, 1); | |
14472 | arg2 = CALL_EXPR_ARG (exp, 2); | |
14473 | arg3 = CALL_EXPR_ARG (exp, 3); | |
14474 | arg4 = CALL_EXPR_ARG (exp, 4); | |
14475 | op0 = expand_normal (arg0); | |
14476 | op1 = expand_normal (arg1); | |
14477 | op2 = expand_normal (arg2); | |
14478 | op3 = expand_normal (arg3); | |
14479 | op4 = expand_normal (arg4); | |
14480 | /* Note the arg order is different from the operand order. */ | |
14481 | mode0 = insn_data[icode].operand[1].mode; | |
14482 | mode2 = insn_data[icode].operand[3].mode; | |
14483 | mode3 = insn_data[icode].operand[4].mode; | |
14484 | mode4 = insn_data[icode].operand[5].mode; | |
14485 | ||
14486 | if (target == NULL_RTX | |
14487 | || GET_MODE (target) != insn_data[icode].operand[0].mode | |
14488 | || !insn_data[icode].operand[0].predicate (target, | |
14489 | GET_MODE (target))) | |
14490 | subtarget = gen_reg_rtx (insn_data[icode].operand[0].mode); | |
14491 | else | |
14492 | subtarget = target; | |
14493 | ||
14494 | switch (fcode) | |
14495 | { | |
14496 | case IX86_BUILTIN_GATHER3ALTSIV8DF: | |
14497 | case IX86_BUILTIN_GATHER3ALTSIV8DI: | |
14498 | half = gen_reg_rtx (V8SImode); | |
14499 | if (!nonimmediate_operand (op2, V16SImode)) | |
14500 | op2 = copy_to_mode_reg (V16SImode, op2); | |
14501 | emit_insn (gen_vec_extract_lo_v16si (half, op2)); | |
14502 | op2 = half; | |
14503 | break; | |
14504 | case IX86_BUILTIN_GATHER3ALTSIV4DF: | |
14505 | case IX86_BUILTIN_GATHER3ALTSIV4DI: | |
14506 | case IX86_BUILTIN_GATHERALTSIV4DF: | |
14507 | case IX86_BUILTIN_GATHERALTSIV4DI: | |
14508 | half = gen_reg_rtx (V4SImode); | |
14509 | if (!nonimmediate_operand (op2, V8SImode)) | |
14510 | op2 = copy_to_mode_reg (V8SImode, op2); | |
14511 | emit_insn (gen_vec_extract_lo_v8si (half, op2)); | |
14512 | op2 = half; | |
14513 | break; | |
14514 | case IX86_BUILTIN_GATHER3ALTDIV16SF: | |
14515 | case IX86_BUILTIN_GATHER3ALTDIV16SI: | |
14516 | half = gen_reg_rtx (mode0); | |
14517 | if (mode0 == V8SFmode) | |
14518 | gen = gen_vec_extract_lo_v16sf; | |
14519 | else | |
14520 | gen = gen_vec_extract_lo_v16si; | |
14521 | if (!nonimmediate_operand (op0, GET_MODE (op0))) | |
14522 | op0 = copy_to_mode_reg (GET_MODE (op0), op0); | |
14523 | emit_insn (gen (half, op0)); | |
14524 | op0 = half; | |
14525 | op3 = lowpart_subreg (QImode, op3, HImode); | |
14526 | break; | |
14527 | case IX86_BUILTIN_GATHER3ALTDIV8SF: | |
14528 | case IX86_BUILTIN_GATHER3ALTDIV8SI: | |
14529 | case IX86_BUILTIN_GATHERALTDIV8SF: | |
14530 | case IX86_BUILTIN_GATHERALTDIV8SI: | |
14531 | half = gen_reg_rtx (mode0); | |
14532 | if (mode0 == V4SFmode) | |
14533 | gen = gen_vec_extract_lo_v8sf; | |
14534 | else | |
14535 | gen = gen_vec_extract_lo_v8si; | |
14536 | if (!nonimmediate_operand (op0, GET_MODE (op0))) | |
14537 | op0 = copy_to_mode_reg (GET_MODE (op0), op0); | |
14538 | emit_insn (gen (half, op0)); | |
14539 | op0 = half; | |
14540 | if (VECTOR_MODE_P (GET_MODE (op3))) | |
14541 | { | |
14542 | half = gen_reg_rtx (mode0); | |
14543 | if (!nonimmediate_operand (op3, GET_MODE (op3))) | |
14544 | op3 = copy_to_mode_reg (GET_MODE (op3), op3); | |
14545 | emit_insn (gen (half, op3)); | |
14546 | op3 = half; | |
14547 | } | |
14548 | break; | |
14549 | default: | |
14550 | break; | |
14551 | } | |
14552 | ||
14553 | /* Force memory operand only with base register here. But we | |
14554 | don't want to do it on memory operand for other builtin | |
14555 | functions. */ | |
14556 | op1 = ix86_zero_extend_to_Pmode (op1); | |
14557 | ||
14558 | if (!insn_data[icode].operand[1].predicate (op0, mode0)) | |
14559 | op0 = copy_to_mode_reg (mode0, op0); | |
14560 | if (!insn_data[icode].operand[2].predicate (op1, Pmode)) | |
14561 | op1 = copy_to_mode_reg (Pmode, op1); | |
14562 | if (!insn_data[icode].operand[3].predicate (op2, mode2)) | |
14563 | op2 = copy_to_mode_reg (mode2, op2); | |
14564 | ||
14565 | op3 = fixup_modeless_constant (op3, mode3); | |
14566 | ||
14567 | if (GET_MODE (op3) == mode3 || GET_MODE (op3) == VOIDmode) | |
14568 | { | |
14569 | if (!insn_data[icode].operand[4].predicate (op3, mode3)) | |
14570 | op3 = copy_to_mode_reg (mode3, op3); | |
14571 | } | |
14572 | else | |
14573 | { | |
14574 | op3 = copy_to_reg (op3); | |
14575 | op3 = lowpart_subreg (mode3, op3, GET_MODE (op3)); | |
14576 | } | |
14577 | if (!insn_data[icode].operand[5].predicate (op4, mode4)) | |
14578 | { | |
14579 | error ("the last argument must be scale 1, 2, 4, 8"); | |
14580 | return const0_rtx; | |
14581 | } | |
14582 | ||
14583 | /* Optimize. If mask is known to have all high bits set, | |
14584 | replace op0 with pc_rtx to signal that the instruction | |
14585 | overwrites the whole destination and doesn't use its | |
14586 | previous contents. */ | |
14587 | if (optimize) | |
14588 | { | |
14589 | if (TREE_CODE (arg3) == INTEGER_CST) | |
14590 | { | |
14591 | if (integer_all_onesp (arg3)) | |
14592 | op0 = pc_rtx; | |
14593 | } | |
14594 | else if (TREE_CODE (arg3) == VECTOR_CST) | |
14595 | { | |
14596 | unsigned int negative = 0; | |
14597 | for (i = 0; i < VECTOR_CST_NELTS (arg3); ++i) | |
14598 | { | |
14599 | tree cst = VECTOR_CST_ELT (arg3, i); | |
14600 | if (TREE_CODE (cst) == INTEGER_CST | |
14601 | && tree_int_cst_sign_bit (cst)) | |
14602 | negative++; | |
14603 | else if (TREE_CODE (cst) == REAL_CST | |
14604 | && REAL_VALUE_NEGATIVE (TREE_REAL_CST (cst))) | |
14605 | negative++; | |
14606 | } | |
14607 | if (negative == TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg3))) | |
14608 | op0 = pc_rtx; | |
14609 | } | |
14610 | else if (TREE_CODE (arg3) == SSA_NAME | |
9907413a | 14611 | && VECTOR_TYPE_P (TREE_TYPE (arg3))) |
2bf6d935 ML |
14612 | { |
14613 | /* Recognize also when mask is like: | |
14614 | __v2df src = _mm_setzero_pd (); | |
14615 | __v2df mask = _mm_cmpeq_pd (src, src); | |
14616 | or | |
14617 | __v8sf src = _mm256_setzero_ps (); | |
14618 | __v8sf mask = _mm256_cmp_ps (src, src, _CMP_EQ_OQ); | |
14619 | as that is a cheaper way to load all ones into | |
14620 | a register than having to load a constant from | |
14621 | memory. */ | |
14622 | gimple *def_stmt = SSA_NAME_DEF_STMT (arg3); | |
14623 | if (is_gimple_call (def_stmt)) | |
14624 | { | |
14625 | tree fndecl = gimple_call_fndecl (def_stmt); | |
14626 | if (fndecl | |
14627 | && fndecl_built_in_p (fndecl, BUILT_IN_MD)) | |
4d732405 | 14628 | switch (DECL_MD_FUNCTION_CODE (fndecl)) |
2bf6d935 ML |
14629 | { |
14630 | case IX86_BUILTIN_CMPPD: | |
14631 | case IX86_BUILTIN_CMPPS: | |
14632 | case IX86_BUILTIN_CMPPD256: | |
14633 | case IX86_BUILTIN_CMPPS256: | |
14634 | if (!integer_zerop (gimple_call_arg (def_stmt, 2))) | |
14635 | break; | |
14636 | /* FALLTHRU */ | |
14637 | case IX86_BUILTIN_CMPEQPD: | |
14638 | case IX86_BUILTIN_CMPEQPS: | |
14639 | if (initializer_zerop (gimple_call_arg (def_stmt, 0)) | |
14640 | && initializer_zerop (gimple_call_arg (def_stmt, | |
14641 | 1))) | |
14642 | op0 = pc_rtx; | |
14643 | break; | |
14644 | default: | |
14645 | break; | |
14646 | } | |
14647 | } | |
14648 | } | |
14649 | } | |
14650 | ||
14651 | pat = GEN_FCN (icode) (subtarget, op0, op1, op2, op3, op4); | |
14652 | if (! pat) | |
14653 | return const0_rtx; | |
14654 | emit_insn (pat); | |
14655 | ||
14656 | switch (fcode) | |
14657 | { | |
14658 | case IX86_BUILTIN_GATHER3DIV16SF: | |
14659 | if (target == NULL_RTX) | |
14660 | target = gen_reg_rtx (V8SFmode); | |
14661 | emit_insn (gen_vec_extract_lo_v16sf (target, subtarget)); | |
14662 | break; | |
14663 | case IX86_BUILTIN_GATHER3DIV16SI: | |
14664 | if (target == NULL_RTX) | |
14665 | target = gen_reg_rtx (V8SImode); | |
14666 | emit_insn (gen_vec_extract_lo_v16si (target, subtarget)); | |
14667 | break; | |
14668 | case IX86_BUILTIN_GATHER3DIV8SF: | |
14669 | case IX86_BUILTIN_GATHERDIV8SF: | |
14670 | if (target == NULL_RTX) | |
14671 | target = gen_reg_rtx (V4SFmode); | |
14672 | emit_insn (gen_vec_extract_lo_v8sf (target, subtarget)); | |
14673 | break; | |
14674 | case IX86_BUILTIN_GATHER3DIV8SI: | |
14675 | case IX86_BUILTIN_GATHERDIV8SI: | |
14676 | if (target == NULL_RTX) | |
14677 | target = gen_reg_rtx (V4SImode); | |
14678 | emit_insn (gen_vec_extract_lo_v8si (target, subtarget)); | |
14679 | break; | |
14680 | default: | |
14681 | target = subtarget; | |
14682 | break; | |
14683 | } | |
14684 | return target; | |
14685 | ||
14686 | scatter_gen: | |
14687 | arg0 = CALL_EXPR_ARG (exp, 0); | |
14688 | arg1 = CALL_EXPR_ARG (exp, 1); | |
14689 | arg2 = CALL_EXPR_ARG (exp, 2); | |
14690 | arg3 = CALL_EXPR_ARG (exp, 3); | |
14691 | arg4 = CALL_EXPR_ARG (exp, 4); | |
14692 | op0 = expand_normal (arg0); | |
14693 | op1 = expand_normal (arg1); | |
14694 | op2 = expand_normal (arg2); | |
14695 | op3 = expand_normal (arg3); | |
14696 | op4 = expand_normal (arg4); | |
14697 | mode1 = insn_data[icode].operand[1].mode; | |
14698 | mode2 = insn_data[icode].operand[2].mode; | |
14699 | mode3 = insn_data[icode].operand[3].mode; | |
14700 | mode4 = insn_data[icode].operand[4].mode; | |
14701 | ||
14702 | /* Scatter instruction stores operand op3 to memory with | |
14703 | indices from op2 and scale from op4 under writemask op1. | |
14704 | If index operand op2 has more elements then source operand | |
14705 | op3 one need to use only its low half. And vice versa. */ | |
14706 | switch (fcode) | |
14707 | { | |
14708 | case IX86_BUILTIN_SCATTERALTSIV8DF: | |
14709 | case IX86_BUILTIN_SCATTERALTSIV8DI: | |
14710 | half = gen_reg_rtx (V8SImode); | |
14711 | if (!nonimmediate_operand (op2, V16SImode)) | |
14712 | op2 = copy_to_mode_reg (V16SImode, op2); | |
14713 | emit_insn (gen_vec_extract_lo_v16si (half, op2)); | |
14714 | op2 = half; | |
14715 | break; | |
14716 | case IX86_BUILTIN_SCATTERALTDIV16SF: | |
14717 | case IX86_BUILTIN_SCATTERALTDIV16SI: | |
14718 | half = gen_reg_rtx (mode3); | |
14719 | if (mode3 == V8SFmode) | |
14720 | gen = gen_vec_extract_lo_v16sf; | |
14721 | else | |
14722 | gen = gen_vec_extract_lo_v16si; | |
14723 | if (!nonimmediate_operand (op3, GET_MODE (op3))) | |
14724 | op3 = copy_to_mode_reg (GET_MODE (op3), op3); | |
14725 | emit_insn (gen (half, op3)); | |
14726 | op3 = half; | |
14727 | break; | |
14728 | case IX86_BUILTIN_SCATTERALTSIV4DF: | |
14729 | case IX86_BUILTIN_SCATTERALTSIV4DI: | |
14730 | half = gen_reg_rtx (V4SImode); | |
14731 | if (!nonimmediate_operand (op2, V8SImode)) | |
14732 | op2 = copy_to_mode_reg (V8SImode, op2); | |
14733 | emit_insn (gen_vec_extract_lo_v8si (half, op2)); | |
14734 | op2 = half; | |
14735 | break; | |
14736 | case IX86_BUILTIN_SCATTERALTDIV8SF: | |
14737 | case IX86_BUILTIN_SCATTERALTDIV8SI: | |
14738 | half = gen_reg_rtx (mode3); | |
14739 | if (mode3 == V4SFmode) | |
14740 | gen = gen_vec_extract_lo_v8sf; | |
14741 | else | |
14742 | gen = gen_vec_extract_lo_v8si; | |
14743 | if (!nonimmediate_operand (op3, GET_MODE (op3))) | |
14744 | op3 = copy_to_mode_reg (GET_MODE (op3), op3); | |
14745 | emit_insn (gen (half, op3)); | |
14746 | op3 = half; | |
14747 | break; | |
14748 | case IX86_BUILTIN_SCATTERALTSIV2DF: | |
14749 | case IX86_BUILTIN_SCATTERALTSIV2DI: | |
14750 | if (!nonimmediate_operand (op2, V4SImode)) | |
14751 | op2 = copy_to_mode_reg (V4SImode, op2); | |
14752 | break; | |
14753 | case IX86_BUILTIN_SCATTERALTDIV4SF: | |
14754 | case IX86_BUILTIN_SCATTERALTDIV4SI: | |
14755 | if (!nonimmediate_operand (op3, GET_MODE (op3))) | |
14756 | op3 = copy_to_mode_reg (GET_MODE (op3), op3); | |
14757 | break; | |
14758 | default: | |
14759 | break; | |
14760 | } | |
14761 | ||
14762 | /* Force memory operand only with base register here. But we | |
14763 | don't want to do it on memory operand for other builtin | |
14764 | functions. */ | |
14765 | op0 = force_reg (Pmode, convert_to_mode (Pmode, op0, 1)); | |
14766 | ||
14767 | if (!insn_data[icode].operand[0].predicate (op0, Pmode)) | |
14768 | op0 = copy_to_mode_reg (Pmode, op0); | |
14769 | ||
14770 | op1 = fixup_modeless_constant (op1, mode1); | |
14771 | ||
14772 | if (GET_MODE (op1) == mode1 || GET_MODE (op1) == VOIDmode) | |
14773 | { | |
14774 | if (!insn_data[icode].operand[1].predicate (op1, mode1)) | |
14775 | op1 = copy_to_mode_reg (mode1, op1); | |
14776 | } | |
14777 | else | |
14778 | { | |
14779 | op1 = copy_to_reg (op1); | |
14780 | op1 = lowpart_subreg (mode1, op1, GET_MODE (op1)); | |
14781 | } | |
14782 | ||
14783 | if (!insn_data[icode].operand[2].predicate (op2, mode2)) | |
14784 | op2 = copy_to_mode_reg (mode2, op2); | |
14785 | ||
14786 | if (!insn_data[icode].operand[3].predicate (op3, mode3)) | |
14787 | op3 = copy_to_mode_reg (mode3, op3); | |
14788 | ||
14789 | if (!insn_data[icode].operand[4].predicate (op4, mode4)) | |
14790 | { | |
14791 | error ("the last argument must be scale 1, 2, 4, 8"); | |
14792 | return const0_rtx; | |
14793 | } | |
14794 | ||
14795 | pat = GEN_FCN (icode) (op0, op1, op2, op3, op4); | |
14796 | if (! pat) | |
14797 | return const0_rtx; | |
14798 | ||
14799 | emit_insn (pat); | |
14800 | return 0; | |
14801 | ||
14802 | vec_prefetch_gen: | |
14803 | arg0 = CALL_EXPR_ARG (exp, 0); | |
14804 | arg1 = CALL_EXPR_ARG (exp, 1); | |
14805 | arg2 = CALL_EXPR_ARG (exp, 2); | |
14806 | arg3 = CALL_EXPR_ARG (exp, 3); | |
14807 | arg4 = CALL_EXPR_ARG (exp, 4); | |
14808 | op0 = expand_normal (arg0); | |
14809 | op1 = expand_normal (arg1); | |
14810 | op2 = expand_normal (arg2); | |
14811 | op3 = expand_normal (arg3); | |
14812 | op4 = expand_normal (arg4); | |
14813 | mode0 = insn_data[icode].operand[0].mode; | |
14814 | mode1 = insn_data[icode].operand[1].mode; | |
14815 | mode3 = insn_data[icode].operand[3].mode; | |
14816 | mode4 = insn_data[icode].operand[4].mode; | |
14817 | ||
14818 | op0 = fixup_modeless_constant (op0, mode0); | |
14819 | ||
14820 | if (GET_MODE (op0) == mode0 || GET_MODE (op0) == VOIDmode) | |
14821 | { | |
14822 | if (!insn_data[icode].operand[0].predicate (op0, mode0)) | |
14823 | op0 = copy_to_mode_reg (mode0, op0); | |
14824 | } | |
14825 | else | |
14826 | { | |
14827 | op0 = copy_to_reg (op0); | |
14828 | op0 = lowpart_subreg (mode0, op0, GET_MODE (op0)); | |
14829 | } | |
14830 | ||
14831 | if (!insn_data[icode].operand[1].predicate (op1, mode1)) | |
14832 | op1 = copy_to_mode_reg (mode1, op1); | |
14833 | ||
14834 | /* Force memory operand only with base register here. But we | |
14835 | don't want to do it on memory operand for other builtin | |
14836 | functions. */ | |
14837 | op2 = force_reg (Pmode, convert_to_mode (Pmode, op2, 1)); | |
14838 | ||
14839 | if (!insn_data[icode].operand[2].predicate (op2, Pmode)) | |
14840 | op2 = copy_to_mode_reg (Pmode, op2); | |
14841 | ||
14842 | if (!insn_data[icode].operand[3].predicate (op3, mode3)) | |
14843 | { | |
14844 | error ("the forth argument must be scale 1, 2, 4, 8"); | |
14845 | return const0_rtx; | |
14846 | } | |
14847 | ||
14848 | if (!insn_data[icode].operand[4].predicate (op4, mode4)) | |
14849 | { | |
14850 | error ("incorrect hint operand"); | |
14851 | return const0_rtx; | |
14852 | } | |
14853 | ||
14854 | pat = GEN_FCN (icode) (op0, op1, op2, op3, op4); | |
14855 | if (! pat) | |
14856 | return const0_rtx; | |
14857 | ||
14858 | emit_insn (pat); | |
14859 | ||
14860 | return 0; | |
14861 | ||
14862 | case IX86_BUILTIN_XABORT: | |
14863 | icode = CODE_FOR_xabort; | |
14864 | arg0 = CALL_EXPR_ARG (exp, 0); | |
14865 | op0 = expand_normal (arg0); | |
14866 | mode0 = insn_data[icode].operand[0].mode; | |
14867 | if (!insn_data[icode].operand[0].predicate (op0, mode0)) | |
14868 | { | |
14869 | error ("the argument to %<xabort%> intrinsic must " | |
14870 | "be an 8-bit immediate"); | |
14871 | return const0_rtx; | |
14872 | } | |
14873 | emit_insn (gen_xabort (op0)); | |
14874 | return 0; | |
14875 | ||
b5034abb UB |
14876 | case IX86_BUILTIN_RDSSPD: |
14877 | case IX86_BUILTIN_RDSSPQ: | |
14878 | mode = (fcode == IX86_BUILTIN_RDSSPD ? SImode : DImode); | |
14879 | ||
14880 | if (target == 0 | |
14881 | || !register_operand (target, mode)) | |
14882 | target = gen_reg_rtx (mode); | |
14883 | ||
14884 | op0 = force_reg (mode, const0_rtx); | |
14885 | ||
14886 | emit_insn (gen_rdssp (mode, target, op0)); | |
14887 | return target; | |
14888 | ||
14889 | case IX86_BUILTIN_INCSSPD: | |
14890 | case IX86_BUILTIN_INCSSPQ: | |
14891 | mode = (fcode == IX86_BUILTIN_INCSSPD ? SImode : DImode); | |
14892 | ||
14893 | arg0 = CALL_EXPR_ARG (exp, 0); | |
14894 | op0 = expand_normal (arg0); | |
14895 | ||
14896 | op0 = force_reg (mode, op0); | |
14897 | ||
14898 | emit_insn (gen_incssp (mode, op0)); | |
14899 | return 0; | |
14900 | ||
83927c63 HW |
14901 | case IX86_BUILTIN_HRESET: |
14902 | icode = CODE_FOR_hreset; | |
14903 | arg0 = CALL_EXPR_ARG (exp, 0); | |
14904 | op0 = expand_normal (arg0); | |
14905 | op0 = force_reg (SImode, op0); | |
14906 | emit_insn (gen_hreset (op0)); | |
14907 | return 0; | |
14908 | ||
2bf6d935 ML |
14909 | case IX86_BUILTIN_RSTORSSP: |
14910 | case IX86_BUILTIN_CLRSSBSY: | |
14911 | arg0 = CALL_EXPR_ARG (exp, 0); | |
14912 | op0 = expand_normal (arg0); | |
14913 | icode = (fcode == IX86_BUILTIN_RSTORSSP | |
b5034abb UB |
14914 | ? CODE_FOR_rstorssp |
14915 | : CODE_FOR_clrssbsy); | |
14916 | ||
2bf6d935 ML |
14917 | if (!address_operand (op0, VOIDmode)) |
14918 | { | |
b5034abb UB |
14919 | op0 = convert_memory_address (Pmode, op0); |
14920 | op0 = copy_addr_to_reg (op0); | |
2bf6d935 | 14921 | } |
b5034abb | 14922 | emit_insn (GEN_FCN (icode) (gen_rtx_MEM (DImode, op0))); |
2bf6d935 ML |
14923 | return 0; |
14924 | ||
14925 | case IX86_BUILTIN_WRSSD: | |
14926 | case IX86_BUILTIN_WRSSQ: | |
14927 | case IX86_BUILTIN_WRUSSD: | |
14928 | case IX86_BUILTIN_WRUSSQ: | |
b5034abb UB |
14929 | mode = ((fcode == IX86_BUILTIN_WRSSD |
14930 | || fcode == IX86_BUILTIN_WRUSSD) | |
14931 | ? SImode : DImode); | |
14932 | ||
2bf6d935 ML |
14933 | arg0 = CALL_EXPR_ARG (exp, 0); |
14934 | op0 = expand_normal (arg0); | |
14935 | arg1 = CALL_EXPR_ARG (exp, 1); | |
14936 | op1 = expand_normal (arg1); | |
b5034abb | 14937 | |
2bf6d935 | 14938 | op0 = force_reg (mode, op0); |
b5034abb | 14939 | |
2bf6d935 ML |
14940 | if (!address_operand (op1, VOIDmode)) |
14941 | { | |
b5034abb UB |
14942 | op1 = convert_memory_address (Pmode, op1); |
14943 | op1 = copy_addr_to_reg (op1); | |
2bf6d935 | 14944 | } |
b5034abb UB |
14945 | op1 = gen_rtx_MEM (mode, op1); |
14946 | ||
44320665 UB |
14947 | icode = ((fcode == IX86_BUILTIN_WRSSD |
14948 | || fcode == IX86_BUILTIN_WRSSQ) | |
14949 | ? code_for_wrss (mode) | |
14950 | : code_for_wruss (mode)); | |
14951 | emit_insn (GEN_FCN (icode) (op0, op1)); | |
14952 | ||
2bf6d935 ML |
14953 | return 0; |
14954 | ||
14955 | default: | |
14956 | break; | |
14957 | } | |
14958 | ||
14959 | if (fcode >= IX86_BUILTIN__BDESC_SPECIAL_ARGS_FIRST | |
14960 | && fcode <= IX86_BUILTIN__BDESC_SPECIAL_ARGS_LAST) | |
14961 | { | |
14962 | i = fcode - IX86_BUILTIN__BDESC_SPECIAL_ARGS_FIRST; | |
14963 | return ix86_expand_special_args_builtin (bdesc_special_args + i, exp, | |
14964 | target); | |
14965 | } | |
14966 | ||
fd5d5794 UB |
14967 | if (fcode >= IX86_BUILTIN__BDESC_PURE_ARGS_FIRST |
14968 | && fcode <= IX86_BUILTIN__BDESC_PURE_ARGS_LAST) | |
14969 | { | |
14970 | i = fcode - IX86_BUILTIN__BDESC_PURE_ARGS_FIRST; | |
14971 | return ix86_expand_special_args_builtin (bdesc_pure_args + i, exp, | |
14972 | target); | |
14973 | } | |
14974 | ||
2bf6d935 ML |
14975 | if (fcode >= IX86_BUILTIN__BDESC_ARGS_FIRST |
14976 | && fcode <= IX86_BUILTIN__BDESC_ARGS_LAST) | |
14977 | { | |
14978 | i = fcode - IX86_BUILTIN__BDESC_ARGS_FIRST; | |
14979 | rtx (*fcn) (rtx, rtx, rtx, rtx) = NULL; | |
14980 | rtx (*fcn_mask) (rtx, rtx, rtx, rtx, rtx); | |
14981 | rtx (*fcn_maskz) (rtx, rtx, rtx, rtx, rtx, rtx); | |
14982 | int masked = 1; | |
14983 | machine_mode mode, wide_mode, nar_mode; | |
14984 | ||
14985 | nar_mode = V4SFmode; | |
14986 | mode = V16SFmode; | |
14987 | wide_mode = V64SFmode; | |
14988 | fcn_mask = gen_avx5124fmaddps_4fmaddps_mask; | |
14989 | fcn_maskz = gen_avx5124fmaddps_4fmaddps_maskz; | |
14990 | ||
14991 | switch (fcode) | |
14992 | { | |
14993 | case IX86_BUILTIN_4FMAPS: | |
14994 | fcn = gen_avx5124fmaddps_4fmaddps; | |
14995 | masked = 0; | |
14996 | goto v4fma_expand; | |
14997 | ||
14998 | case IX86_BUILTIN_4DPWSSD: | |
14999 | nar_mode = V4SImode; | |
15000 | mode = V16SImode; | |
15001 | wide_mode = V64SImode; | |
15002 | fcn = gen_avx5124vnniw_vp4dpwssd; | |
15003 | masked = 0; | |
15004 | goto v4fma_expand; | |
15005 | ||
15006 | case IX86_BUILTIN_4DPWSSDS: | |
15007 | nar_mode = V4SImode; | |
15008 | mode = V16SImode; | |
15009 | wide_mode = V64SImode; | |
15010 | fcn = gen_avx5124vnniw_vp4dpwssds; | |
15011 | masked = 0; | |
15012 | goto v4fma_expand; | |
15013 | ||
15014 | case IX86_BUILTIN_4FNMAPS: | |
15015 | fcn = gen_avx5124fmaddps_4fnmaddps; | |
15016 | masked = 0; | |
15017 | goto v4fma_expand; | |
15018 | ||
15019 | case IX86_BUILTIN_4FNMAPS_MASK: | |
15020 | fcn_mask = gen_avx5124fmaddps_4fnmaddps_mask; | |
15021 | fcn_maskz = gen_avx5124fmaddps_4fnmaddps_maskz; | |
15022 | goto v4fma_expand; | |
15023 | ||
15024 | case IX86_BUILTIN_4DPWSSD_MASK: | |
15025 | nar_mode = V4SImode; | |
15026 | mode = V16SImode; | |
15027 | wide_mode = V64SImode; | |
15028 | fcn_mask = gen_avx5124vnniw_vp4dpwssd_mask; | |
15029 | fcn_maskz = gen_avx5124vnniw_vp4dpwssd_maskz; | |
15030 | goto v4fma_expand; | |
15031 | ||
15032 | case IX86_BUILTIN_4DPWSSDS_MASK: | |
15033 | nar_mode = V4SImode; | |
15034 | mode = V16SImode; | |
15035 | wide_mode = V64SImode; | |
15036 | fcn_mask = gen_avx5124vnniw_vp4dpwssds_mask; | |
15037 | fcn_maskz = gen_avx5124vnniw_vp4dpwssds_maskz; | |
15038 | goto v4fma_expand; | |
15039 | ||
15040 | case IX86_BUILTIN_4FMAPS_MASK: | |
15041 | { | |
15042 | tree args[4]; | |
15043 | rtx ops[4]; | |
15044 | rtx wide_reg; | |
15045 | rtx accum; | |
15046 | rtx addr; | |
15047 | rtx mem; | |
15048 | ||
15049 | v4fma_expand: | |
15050 | wide_reg = gen_reg_rtx (wide_mode); | |
15051 | for (i = 0; i < 4; i++) | |
15052 | { | |
15053 | args[i] = CALL_EXPR_ARG (exp, i); | |
15054 | ops[i] = expand_normal (args[i]); | |
15055 | ||
15056 | emit_move_insn (gen_rtx_SUBREG (mode, wide_reg, i * 64), | |
15057 | ops[i]); | |
15058 | } | |
15059 | ||
15060 | accum = expand_normal (CALL_EXPR_ARG (exp, 4)); | |
15061 | accum = force_reg (mode, accum); | |
15062 | ||
15063 | addr = expand_normal (CALL_EXPR_ARG (exp, 5)); | |
15064 | addr = force_reg (Pmode, addr); | |
15065 | ||
15066 | mem = gen_rtx_MEM (nar_mode, addr); | |
15067 | ||
15068 | target = gen_reg_rtx (mode); | |
15069 | ||
15070 | emit_move_insn (target, accum); | |
15071 | ||
15072 | if (! masked) | |
15073 | emit_insn (fcn (target, accum, wide_reg, mem)); | |
15074 | else | |
15075 | { | |
15076 | rtx merge, mask; | |
15077 | merge = expand_normal (CALL_EXPR_ARG (exp, 6)); | |
15078 | ||
15079 | mask = expand_normal (CALL_EXPR_ARG (exp, 7)); | |
15080 | ||
15081 | if (CONST_INT_P (mask)) | |
15082 | mask = fixup_modeless_constant (mask, HImode); | |
15083 | ||
15084 | mask = force_reg (HImode, mask); | |
15085 | ||
15086 | if (GET_MODE (mask) != HImode) | |
15087 | mask = gen_rtx_SUBREG (HImode, mask, 0); | |
15088 | ||
15089 | /* If merge is 0 then we're about to emit z-masked variant. */ | |
15090 | if (const0_operand (merge, mode)) | |
15091 | emit_insn (fcn_maskz (target, accum, wide_reg, mem, merge, mask)); | |
15092 | /* If merge is the same as accum then emit merge-masked variant. */ | |
15093 | else if (CALL_EXPR_ARG (exp, 6) == CALL_EXPR_ARG (exp, 4)) | |
15094 | { | |
15095 | merge = force_reg (mode, merge); | |
15096 | emit_insn (fcn_mask (target, wide_reg, mem, merge, mask)); | |
15097 | } | |
15098 | /* Merge with something unknown might happen if we z-mask w/ -O0. */ | |
15099 | else | |
15100 | { | |
15101 | target = gen_reg_rtx (mode); | |
15102 | emit_move_insn (target, merge); | |
15103 | emit_insn (fcn_mask (target, wide_reg, mem, target, mask)); | |
15104 | } | |
15105 | } | |
15106 | return target; | |
15107 | } | |
15108 | ||
15109 | case IX86_BUILTIN_4FNMASS: | |
15110 | fcn = gen_avx5124fmaddps_4fnmaddss; | |
15111 | masked = 0; | |
15112 | goto s4fma_expand; | |
15113 | ||
15114 | case IX86_BUILTIN_4FMASS: | |
15115 | fcn = gen_avx5124fmaddps_4fmaddss; | |
15116 | masked = 0; | |
15117 | goto s4fma_expand; | |
15118 | ||
15119 | case IX86_BUILTIN_4FNMASS_MASK: | |
15120 | fcn_mask = gen_avx5124fmaddps_4fnmaddss_mask; | |
15121 | fcn_maskz = gen_avx5124fmaddps_4fnmaddss_maskz; | |
15122 | goto s4fma_expand; | |
15123 | ||
15124 | case IX86_BUILTIN_4FMASS_MASK: | |
15125 | { | |
15126 | tree args[4]; | |
15127 | rtx ops[4]; | |
15128 | rtx wide_reg; | |
15129 | rtx accum; | |
15130 | rtx addr; | |
15131 | rtx mem; | |
15132 | ||
15133 | fcn_mask = gen_avx5124fmaddps_4fmaddss_mask; | |
15134 | fcn_maskz = gen_avx5124fmaddps_4fmaddss_maskz; | |
15135 | ||
15136 | s4fma_expand: | |
15137 | mode = V4SFmode; | |
15138 | wide_reg = gen_reg_rtx (V64SFmode); | |
15139 | for (i = 0; i < 4; i++) | |
15140 | { | |
15141 | rtx tmp; | |
15142 | args[i] = CALL_EXPR_ARG (exp, i); | |
15143 | ops[i] = expand_normal (args[i]); | |
15144 | ||
15145 | tmp = gen_reg_rtx (SFmode); | |
15146 | emit_move_insn (tmp, gen_rtx_SUBREG (SFmode, ops[i], 0)); | |
15147 | ||
15148 | emit_move_insn (gen_rtx_SUBREG (V16SFmode, wide_reg, i * 64), | |
15149 | gen_rtx_SUBREG (V16SFmode, tmp, 0)); | |
15150 | } | |
15151 | ||
15152 | accum = expand_normal (CALL_EXPR_ARG (exp, 4)); | |
15153 | accum = force_reg (V4SFmode, accum); | |
15154 | ||
15155 | addr = expand_normal (CALL_EXPR_ARG (exp, 5)); | |
15156 | addr = force_reg (Pmode, addr); | |
15157 | ||
15158 | mem = gen_rtx_MEM (V4SFmode, addr); | |
15159 | ||
15160 | target = gen_reg_rtx (V4SFmode); | |
15161 | ||
15162 | emit_move_insn (target, accum); | |
15163 | ||
15164 | if (! masked) | |
15165 | emit_insn (fcn (target, accum, wide_reg, mem)); | |
15166 | else | |
15167 | { | |
15168 | rtx merge, mask; | |
15169 | merge = expand_normal (CALL_EXPR_ARG (exp, 6)); | |
15170 | ||
15171 | mask = expand_normal (CALL_EXPR_ARG (exp, 7)); | |
15172 | ||
15173 | if (CONST_INT_P (mask)) | |
15174 | mask = fixup_modeless_constant (mask, QImode); | |
15175 | ||
15176 | mask = force_reg (QImode, mask); | |
15177 | ||
15178 | if (GET_MODE (mask) != QImode) | |
15179 | mask = gen_rtx_SUBREG (QImode, mask, 0); | |
15180 | ||
15181 | /* If merge is 0 then we're about to emit z-masked variant. */ | |
15182 | if (const0_operand (merge, mode)) | |
15183 | emit_insn (fcn_maskz (target, accum, wide_reg, mem, merge, mask)); | |
15184 | /* If merge is the same as accum then emit merge-masked | |
15185 | variant. */ | |
15186 | else if (CALL_EXPR_ARG (exp, 6) == CALL_EXPR_ARG (exp, 4)) | |
15187 | { | |
15188 | merge = force_reg (mode, merge); | |
15189 | emit_insn (fcn_mask (target, wide_reg, mem, merge, mask)); | |
15190 | } | |
15191 | /* Merge with something unknown might happen if we z-mask | |
15192 | w/ -O0. */ | |
15193 | else | |
15194 | { | |
15195 | target = gen_reg_rtx (mode); | |
15196 | emit_move_insn (target, merge); | |
15197 | emit_insn (fcn_mask (target, wide_reg, mem, target, mask)); | |
15198 | } | |
15199 | } | |
15200 | return target; | |
15201 | } | |
15202 | case IX86_BUILTIN_RDPID: | |
15203 | return ix86_expand_special_args_builtin (bdesc_args + i, exp, | |
15204 | target); | |
15205 | case IX86_BUILTIN_FABSQ: | |
15206 | case IX86_BUILTIN_COPYSIGNQ: | |
15207 | if (!TARGET_SSE) | |
15208 | /* Emit a normal call if SSE isn't available. */ | |
15209 | return expand_call (exp, target, ignore); | |
15210 | /* FALLTHRU */ | |
15211 | default: | |
15212 | return ix86_expand_args_builtin (bdesc_args + i, exp, target); | |
15213 | } | |
15214 | } | |
15215 | ||
15216 | if (fcode >= IX86_BUILTIN__BDESC_COMI_FIRST | |
15217 | && fcode <= IX86_BUILTIN__BDESC_COMI_LAST) | |
15218 | { | |
15219 | i = fcode - IX86_BUILTIN__BDESC_COMI_FIRST; | |
15220 | return ix86_expand_sse_comi (bdesc_comi + i, exp, target); | |
15221 | } | |
15222 | ||
15223 | if (fcode >= IX86_BUILTIN__BDESC_ROUND_ARGS_FIRST | |
15224 | && fcode <= IX86_BUILTIN__BDESC_ROUND_ARGS_LAST) | |
15225 | { | |
15226 | i = fcode - IX86_BUILTIN__BDESC_ROUND_ARGS_FIRST; | |
15227 | return ix86_expand_round_builtin (bdesc_round_args + i, exp, target); | |
15228 | } | |
15229 | ||
15230 | if (fcode >= IX86_BUILTIN__BDESC_PCMPESTR_FIRST | |
15231 | && fcode <= IX86_BUILTIN__BDESC_PCMPESTR_LAST) | |
15232 | { | |
15233 | i = fcode - IX86_BUILTIN__BDESC_PCMPESTR_FIRST; | |
15234 | return ix86_expand_sse_pcmpestr (bdesc_pcmpestr + i, exp, target); | |
15235 | } | |
15236 | ||
15237 | if (fcode >= IX86_BUILTIN__BDESC_PCMPISTR_FIRST | |
15238 | && fcode <= IX86_BUILTIN__BDESC_PCMPISTR_LAST) | |
15239 | { | |
15240 | i = fcode - IX86_BUILTIN__BDESC_PCMPISTR_FIRST; | |
15241 | return ix86_expand_sse_pcmpistr (bdesc_pcmpistr + i, exp, target); | |
15242 | } | |
15243 | ||
15244 | if (fcode >= IX86_BUILTIN__BDESC_MULTI_ARG_FIRST | |
15245 | && fcode <= IX86_BUILTIN__BDESC_MULTI_ARG_LAST) | |
15246 | { | |
15247 | i = fcode - IX86_BUILTIN__BDESC_MULTI_ARG_FIRST; | |
15248 | const struct builtin_description *d = bdesc_multi_arg + i; | |
15249 | return ix86_expand_multi_arg_builtin (d->icode, exp, target, | |
15250 | (enum ix86_builtin_func_type) | |
15251 | d->flag, d->comparison); | |
15252 | } | |
15253 | ||
15254 | if (fcode >= IX86_BUILTIN__BDESC_CET_FIRST | |
15255 | && fcode <= IX86_BUILTIN__BDESC_CET_LAST) | |
15256 | { | |
15257 | i = fcode - IX86_BUILTIN__BDESC_CET_FIRST; | |
15258 | return ix86_expand_special_args_builtin (bdesc_cet + i, exp, | |
15259 | target); | |
15260 | } | |
15261 | ||
2bf6d935 ML |
15262 | gcc_unreachable (); |
15263 | } | |
15264 | ||
15265 | /* A subroutine of ix86_expand_vector_init_duplicate. Tries to | |
15266 | fill target with val via vec_duplicate. */ | |
15267 | ||
15268 | static bool | |
15269 | ix86_vector_duplicate_value (machine_mode mode, rtx target, rtx val) | |
15270 | { | |
15271 | bool ok; | |
15272 | rtx_insn *insn; | |
15273 | rtx dup; | |
b3237a2c JJ |
15274 | /* Save/restore recog_data in case this is called from splitters |
15275 | or other routines where recog_data needs to stay valid across | |
15276 | force_reg. See PR106577. */ | |
15277 | recog_data_d recog_data_save = recog_data; | |
2bf6d935 ML |
15278 | |
15279 | /* First attempt to recognize VAL as-is. */ | |
15280 | dup = gen_vec_duplicate (mode, val); | |
15281 | insn = emit_insn (gen_rtx_SET (target, dup)); | |
15282 | if (recog_memoized (insn) < 0) | |
15283 | { | |
15284 | rtx_insn *seq; | |
15285 | machine_mode innermode = GET_MODE_INNER (mode); | |
15286 | rtx reg; | |
15287 | ||
15288 | /* If that fails, force VAL into a register. */ | |
15289 | ||
15290 | start_sequence (); | |
15291 | reg = force_reg (innermode, val); | |
15292 | if (GET_MODE (reg) != innermode) | |
15293 | reg = gen_lowpart (innermode, reg); | |
15294 | SET_SRC (PATTERN (insn)) = gen_vec_duplicate (mode, reg); | |
15295 | seq = get_insns (); | |
15296 | end_sequence (); | |
15297 | if (seq) | |
15298 | emit_insn_before (seq, insn); | |
15299 | ||
15300 | ok = recog_memoized (insn) >= 0; | |
15301 | gcc_assert (ok); | |
15302 | } | |
b3237a2c | 15303 | recog_data = recog_data_save; |
2bf6d935 ML |
15304 | return true; |
15305 | } | |
15306 | ||
15307 | /* Get a vector mode of the same size as the original but with elements | |
15308 | twice as wide. This is only guaranteed to apply to integral vectors. */ | |
15309 | ||
15310 | static machine_mode | |
15311 | get_mode_wider_vector (machine_mode o) | |
15312 | { | |
e53b6e56 | 15313 | /* ??? Rely on the ordering that genmodes.cc gives to vectors. */ |
4b796619 | 15314 | machine_mode n = GET_MODE_NEXT_MODE (o).require (); |
2bf6d935 ML |
15315 | gcc_assert (GET_MODE_NUNITS (o) == GET_MODE_NUNITS (n) * 2); |
15316 | gcc_assert (GET_MODE_SIZE (o) == GET_MODE_SIZE (n)); | |
15317 | return n; | |
15318 | } | |
15319 | ||
15320 | static bool expand_vec_perm_broadcast_1 (struct expand_vec_perm_d *d); | |
15321 | static bool expand_vec_perm_1 (struct expand_vec_perm_d *d); | |
15322 | ||
15323 | /* A subroutine of ix86_expand_vector_init. Store into TARGET a vector | |
15324 | with all elements equal to VAR. Return true if successful. */ | |
15325 | ||
51c30227 | 15326 | bool |
2bf6d935 ML |
15327 | ix86_expand_vector_init_duplicate (bool mmx_ok, machine_mode mode, |
15328 | rtx target, rtx val) | |
15329 | { | |
15330 | bool ok; | |
15331 | ||
15332 | switch (mode) | |
15333 | { | |
15334 | case E_V2SImode: | |
15335 | case E_V2SFmode: | |
15336 | if (!mmx_ok) | |
15337 | return false; | |
15338 | /* FALLTHRU */ | |
15339 | ||
15340 | case E_V4DFmode: | |
15341 | case E_V4DImode: | |
15342 | case E_V8SFmode: | |
15343 | case E_V8SImode: | |
15344 | case E_V2DFmode: | |
15345 | case E_V2DImode: | |
15346 | case E_V4SFmode: | |
15347 | case E_V4SImode: | |
15348 | case E_V16SImode: | |
15349 | case E_V8DImode: | |
15350 | case E_V16SFmode: | |
15351 | case E_V8DFmode: | |
15352 | return ix86_vector_duplicate_value (mode, target, val); | |
15353 | ||
15354 | case E_V4HImode: | |
15355 | if (!mmx_ok) | |
15356 | return false; | |
15357 | if (TARGET_SSE || TARGET_3DNOW_A) | |
15358 | { | |
15359 | rtx x; | |
15360 | ||
15361 | val = gen_lowpart (SImode, val); | |
15362 | x = gen_rtx_TRUNCATE (HImode, val); | |
15363 | x = gen_rtx_VEC_DUPLICATE (mode, x); | |
15364 | emit_insn (gen_rtx_SET (target, x)); | |
15365 | return true; | |
15366 | } | |
15367 | goto widen; | |
15368 | ||
8d7dae0e UB |
15369 | case E_V2HImode: |
15370 | if (TARGET_SSE2) | |
15371 | { | |
15372 | rtx x; | |
15373 | ||
15374 | val = gen_lowpart (SImode, val); | |
15375 | x = gen_rtx_TRUNCATE (HImode, val); | |
15376 | x = gen_rtx_VEC_DUPLICATE (mode, x); | |
15377 | emit_insn (gen_rtx_SET (target, x)); | |
15378 | return true; | |
15379 | } | |
15380 | return false; | |
15381 | ||
2bf6d935 | 15382 | case E_V8QImode: |
64735dc9 | 15383 | case E_V4QImode: |
2bf6d935 ML |
15384 | if (!mmx_ok) |
15385 | return false; | |
15386 | goto widen; | |
15387 | ||
15388 | case E_V8HImode: | |
7a54d3de | 15389 | case E_V8HFmode: |
6910cad5 | 15390 | case E_V8BFmode: |
2bf6d935 ML |
15391 | if (TARGET_AVX2) |
15392 | return ix86_vector_duplicate_value (mode, target, val); | |
15393 | ||
15394 | if (TARGET_SSE2) | |
15395 | { | |
15396 | struct expand_vec_perm_d dperm; | |
15397 | rtx tmp1, tmp2; | |
15398 | ||
15399 | permute: | |
15400 | memset (&dperm, 0, sizeof (dperm)); | |
15401 | dperm.target = target; | |
15402 | dperm.vmode = mode; | |
15403 | dperm.nelt = GET_MODE_NUNITS (mode); | |
15404 | dperm.op0 = dperm.op1 = gen_reg_rtx (mode); | |
15405 | dperm.one_operand_p = true; | |
15406 | ||
092763fd | 15407 | if (mode == V8HFmode || mode == V8BFmode) |
e2385690 | 15408 | { |
092763fd | 15409 | tmp1 = force_reg (GET_MODE_INNER (mode), val); |
e2385690 | 15410 | tmp2 = gen_reg_rtx (mode); |
092763fd | 15411 | emit_insn (maybe_gen_vec_set_0 (mode, tmp2, |
15412 | CONST0_RTX (mode), tmp1)); | |
e2385690 HW |
15413 | tmp1 = gen_lowpart (mode, tmp2); |
15414 | } | |
7a54d3de UB |
15415 | else |
15416 | { | |
15417 | /* Extend to SImode using a paradoxical SUBREG. */ | |
15418 | tmp1 = gen_reg_rtx (SImode); | |
15419 | emit_move_insn (tmp1, gen_lowpart (SImode, val)); | |
15420 | ||
15421 | /* Insert the SImode value as | |
15422 | low element of a V4SImode vector. */ | |
15423 | tmp2 = gen_reg_rtx (V4SImode); | |
15424 | emit_insn (gen_vec_setv4si_0 (tmp2, CONST0_RTX (V4SImode), tmp1)); | |
15425 | tmp1 = gen_lowpart (mode, tmp2); | |
15426 | } | |
2bf6d935 | 15427 | |
7a54d3de | 15428 | emit_move_insn (dperm.op0, tmp1); |
2bf6d935 ML |
15429 | ok = (expand_vec_perm_1 (&dperm) |
15430 | || expand_vec_perm_broadcast_1 (&dperm)); | |
15431 | gcc_assert (ok); | |
15432 | return ok; | |
15433 | } | |
15434 | goto widen; | |
15435 | ||
15436 | case E_V16QImode: | |
15437 | if (TARGET_AVX2) | |
15438 | return ix86_vector_duplicate_value (mode, target, val); | |
15439 | ||
15440 | if (TARGET_SSE2) | |
15441 | goto permute; | |
15442 | goto widen; | |
15443 | ||
15444 | widen: | |
15445 | /* Replicate the value once into the next wider mode and recurse. */ | |
15446 | { | |
15447 | machine_mode smode, wsmode, wvmode; | |
15448 | rtx x; | |
15449 | ||
15450 | smode = GET_MODE_INNER (mode); | |
15451 | wvmode = get_mode_wider_vector (mode); | |
15452 | wsmode = GET_MODE_INNER (wvmode); | |
15453 | ||
15454 | val = convert_modes (wsmode, smode, val, true); | |
20a2c8ac UB |
15455 | |
15456 | if (smode == QImode && !TARGET_PARTIAL_REG_STALL) | |
15457 | emit_insn (gen_insv_1 (wsmode, val, val)); | |
15458 | else | |
15459 | { | |
15460 | x = expand_simple_binop (wsmode, ASHIFT, val, | |
15461 | GEN_INT (GET_MODE_BITSIZE (smode)), | |
15462 | NULL_RTX, 1, OPTAB_LIB_WIDEN); | |
15463 | val = expand_simple_binop (wsmode, IOR, val, x, x, 1, | |
15464 | OPTAB_LIB_WIDEN); | |
15465 | } | |
2bf6d935 ML |
15466 | |
15467 | x = gen_reg_rtx (wvmode); | |
15468 | ok = ix86_expand_vector_init_duplicate (mmx_ok, wvmode, x, val); | |
15469 | gcc_assert (ok); | |
15470 | emit_move_insn (target, gen_lowpart (GET_MODE (target), x)); | |
15471 | return ok; | |
15472 | } | |
15473 | ||
15474 | case E_V16HImode: | |
7a54d3de | 15475 | case E_V16HFmode: |
6910cad5 | 15476 | case E_V16BFmode: |
2bf6d935 ML |
15477 | case E_V32QImode: |
15478 | if (TARGET_AVX2) | |
15479 | return ix86_vector_duplicate_value (mode, target, val); | |
15480 | else | |
15481 | { | |
78260b9a | 15482 | machine_mode hvmode; |
15483 | switch (mode) | |
15484 | { | |
15485 | case V16HImode: | |
15486 | hvmode = V8HImode; | |
15487 | break; | |
15488 | case V16HFmode: | |
15489 | hvmode = V8HFmode; | |
15490 | break; | |
15491 | case V16BFmode: | |
15492 | hvmode = V8BFmode; | |
15493 | break; | |
15494 | case V32QImode: | |
15495 | hvmode = V16QImode; | |
15496 | break; | |
15497 | default: | |
15498 | gcc_unreachable (); | |
15499 | } | |
2bf6d935 ML |
15500 | rtx x = gen_reg_rtx (hvmode); |
15501 | ||
15502 | ok = ix86_expand_vector_init_duplicate (false, hvmode, x, val); | |
15503 | gcc_assert (ok); | |
15504 | ||
15505 | x = gen_rtx_VEC_CONCAT (mode, x, x); | |
15506 | emit_insn (gen_rtx_SET (target, x)); | |
15507 | } | |
15508 | return true; | |
15509 | ||
2bf6d935 | 15510 | case E_V32HImode: |
7a54d3de | 15511 | case E_V32HFmode: |
6910cad5 | 15512 | case E_V32BFmode: |
7a54d3de | 15513 | case E_V64QImode: |
2bf6d935 ML |
15514 | if (TARGET_AVX512BW) |
15515 | return ix86_vector_duplicate_value (mode, target, val); | |
15516 | else | |
15517 | { | |
78260b9a | 15518 | machine_mode hvmode; |
15519 | switch (mode) | |
15520 | { | |
15521 | case V32HImode: | |
15522 | hvmode = V16HImode; | |
15523 | break; | |
15524 | case V32HFmode: | |
15525 | hvmode = V16HFmode; | |
15526 | break; | |
15527 | case V32BFmode: | |
15528 | hvmode = V16BFmode; | |
15529 | break; | |
15530 | case V64QImode: | |
15531 | hvmode = V32QImode; | |
15532 | break; | |
15533 | default: | |
15534 | gcc_unreachable (); | |
15535 | } | |
2bf6d935 ML |
15536 | rtx x = gen_reg_rtx (hvmode); |
15537 | ||
15538 | ok = ix86_expand_vector_init_duplicate (false, hvmode, x, val); | |
15539 | gcc_assert (ok); | |
15540 | ||
15541 | x = gen_rtx_VEC_CONCAT (mode, x, x); | |
15542 | emit_insn (gen_rtx_SET (target, x)); | |
15543 | } | |
15544 | return true; | |
15545 | ||
15546 | default: | |
15547 | return false; | |
15548 | } | |
15549 | } | |
15550 | ||
15551 | /* A subroutine of ix86_expand_vector_init. Store into TARGET a vector | |
15552 | whose ONE_VAR element is VAR, and other elements are zero. Return true | |
15553 | if successful. */ | |
15554 | ||
15555 | static bool | |
15556 | ix86_expand_vector_init_one_nonzero (bool mmx_ok, machine_mode mode, | |
15557 | rtx target, rtx var, int one_var) | |
15558 | { | |
15559 | machine_mode vsimode; | |
15560 | rtx new_target; | |
15561 | rtx x, tmp; | |
15562 | bool use_vector_set = false; | |
15563 | rtx (*gen_vec_set_0) (rtx, rtx, rtx) = NULL; | |
15564 | ||
15565 | switch (mode) | |
15566 | { | |
15567 | case E_V2DImode: | |
15568 | /* For SSE4.1, we normally use vector set. But if the second | |
15569 | element is zero and inter-unit moves are OK, we use movq | |
15570 | instead. */ | |
15571 | use_vector_set = (TARGET_64BIT && TARGET_SSE4_1 | |
15572 | && !(TARGET_INTER_UNIT_MOVES_TO_VEC | |
15573 | && one_var == 0)); | |
15574 | break; | |
15575 | case E_V16QImode: | |
15576 | case E_V4SImode: | |
15577 | case E_V4SFmode: | |
15578 | use_vector_set = TARGET_SSE4_1; | |
15579 | break; | |
15580 | case E_V8HImode: | |
15581 | use_vector_set = TARGET_SSE2; | |
c4d423c7 | 15582 | gen_vec_set_0 = TARGET_AVX512FP16 && one_var == 0 |
15583 | ? gen_vec_setv8hi_0 : NULL; | |
2bf6d935 | 15584 | break; |
8a0eb0cd UB |
15585 | case E_V8QImode: |
15586 | use_vector_set = TARGET_MMX_WITH_SSE && TARGET_SSE4_1; | |
15587 | break; | |
2bf6d935 ML |
15588 | case E_V4HImode: |
15589 | use_vector_set = TARGET_SSE || TARGET_3DNOW_A; | |
15590 | break; | |
64735dc9 UB |
15591 | case E_V4QImode: |
15592 | use_vector_set = TARGET_SSE4_1; | |
15593 | break; | |
2bf6d935 | 15594 | case E_V32QImode: |
c4d423c7 | 15595 | use_vector_set = TARGET_AVX; |
15596 | break; | |
2bf6d935 ML |
15597 | case E_V16HImode: |
15598 | use_vector_set = TARGET_AVX; | |
c4d423c7 | 15599 | gen_vec_set_0 = TARGET_AVX512FP16 && one_var == 0 |
15600 | ? gen_vec_setv16hi_0 : NULL; | |
2bf6d935 ML |
15601 | break; |
15602 | case E_V8SImode: | |
15603 | use_vector_set = TARGET_AVX; | |
15604 | gen_vec_set_0 = gen_vec_setv8si_0; | |
15605 | break; | |
15606 | case E_V8SFmode: | |
15607 | use_vector_set = TARGET_AVX; | |
15608 | gen_vec_set_0 = gen_vec_setv8sf_0; | |
15609 | break; | |
15610 | case E_V4DFmode: | |
15611 | use_vector_set = TARGET_AVX; | |
15612 | gen_vec_set_0 = gen_vec_setv4df_0; | |
15613 | break; | |
15614 | case E_V4DImode: | |
15615 | /* Use ix86_expand_vector_set in 64bit mode only. */ | |
15616 | use_vector_set = TARGET_AVX && TARGET_64BIT; | |
15617 | gen_vec_set_0 = gen_vec_setv4di_0; | |
15618 | break; | |
15619 | case E_V16SImode: | |
15620 | use_vector_set = TARGET_AVX512F && one_var == 0; | |
15621 | gen_vec_set_0 = gen_vec_setv16si_0; | |
15622 | break; | |
15623 | case E_V16SFmode: | |
15624 | use_vector_set = TARGET_AVX512F && one_var == 0; | |
15625 | gen_vec_set_0 = gen_vec_setv16sf_0; | |
15626 | break; | |
15627 | case E_V8DFmode: | |
15628 | use_vector_set = TARGET_AVX512F && one_var == 0; | |
15629 | gen_vec_set_0 = gen_vec_setv8df_0; | |
15630 | break; | |
15631 | case E_V8DImode: | |
15632 | /* Use ix86_expand_vector_set in 64bit mode only. */ | |
15633 | use_vector_set = TARGET_AVX512F && TARGET_64BIT && one_var == 0; | |
15634 | gen_vec_set_0 = gen_vec_setv8di_0; | |
15635 | break; | |
9e2a82e1 | 15636 | case E_V8HFmode: |
15637 | use_vector_set = TARGET_AVX512FP16 && one_var == 0; | |
15638 | gen_vec_set_0 = gen_vec_setv8hf_0; | |
15639 | break; | |
15640 | case E_V16HFmode: | |
15641 | use_vector_set = TARGET_AVX512FP16 && one_var == 0; | |
15642 | gen_vec_set_0 = gen_vec_setv16hf_0; | |
15643 | break; | |
15644 | case E_V32HFmode: | |
15645 | use_vector_set = TARGET_AVX512FP16 && one_var == 0; | |
15646 | gen_vec_set_0 = gen_vec_setv32hf_0; | |
15647 | break; | |
6910cad5 | 15648 | case E_V8BFmode: |
15649 | use_vector_set = TARGET_AVX512FP16 && one_var == 0; | |
15650 | gen_vec_set_0 = gen_vec_setv8bf_0; | |
15651 | break; | |
15652 | case E_V16BFmode: | |
15653 | use_vector_set = TARGET_AVX512FP16 && one_var == 0; | |
15654 | gen_vec_set_0 = gen_vec_setv16bf_0; | |
15655 | break; | |
15656 | case E_V32BFmode: | |
15657 | use_vector_set = TARGET_AVX512FP16 && one_var == 0; | |
15658 | gen_vec_set_0 = gen_vec_setv32bf_0; | |
15659 | break; | |
c4d423c7 | 15660 | case E_V32HImode: |
15661 | use_vector_set = TARGET_AVX512FP16 && one_var == 0; | |
15662 | gen_vec_set_0 = gen_vec_setv32hi_0; | |
2bf6d935 ML |
15663 | default: |
15664 | break; | |
15665 | } | |
15666 | ||
15667 | if (use_vector_set) | |
15668 | { | |
15669 | if (gen_vec_set_0 && one_var == 0) | |
15670 | { | |
15671 | var = force_reg (GET_MODE_INNER (mode), var); | |
15672 | emit_insn (gen_vec_set_0 (target, CONST0_RTX (mode), var)); | |
15673 | return true; | |
15674 | } | |
15675 | emit_insn (gen_rtx_SET (target, CONST0_RTX (mode))); | |
15676 | var = force_reg (GET_MODE_INNER (mode), var); | |
15677 | ix86_expand_vector_set (mmx_ok, target, var, one_var); | |
15678 | return true; | |
15679 | } | |
15680 | ||
15681 | switch (mode) | |
15682 | { | |
15683 | case E_V2SFmode: | |
15684 | case E_V2SImode: | |
15685 | if (!mmx_ok) | |
15686 | return false; | |
15687 | /* FALLTHRU */ | |
15688 | ||
15689 | case E_V2DFmode: | |
15690 | case E_V2DImode: | |
15691 | if (one_var != 0) | |
15692 | return false; | |
15693 | var = force_reg (GET_MODE_INNER (mode), var); | |
15694 | x = gen_rtx_VEC_CONCAT (mode, var, CONST0_RTX (GET_MODE_INNER (mode))); | |
15695 | emit_insn (gen_rtx_SET (target, x)); | |
15696 | return true; | |
15697 | ||
15698 | case E_V4SFmode: | |
15699 | case E_V4SImode: | |
15700 | if (!REG_P (target) || REGNO (target) < FIRST_PSEUDO_REGISTER) | |
15701 | new_target = gen_reg_rtx (mode); | |
15702 | else | |
15703 | new_target = target; | |
15704 | var = force_reg (GET_MODE_INNER (mode), var); | |
15705 | x = gen_rtx_VEC_DUPLICATE (mode, var); | |
15706 | x = gen_rtx_VEC_MERGE (mode, x, CONST0_RTX (mode), const1_rtx); | |
15707 | emit_insn (gen_rtx_SET (new_target, x)); | |
15708 | if (one_var != 0) | |
15709 | { | |
15710 | /* We need to shuffle the value to the correct position, so | |
15711 | create a new pseudo to store the intermediate result. */ | |
15712 | ||
15713 | /* With SSE2, we can use the integer shuffle insns. */ | |
15714 | if (mode != V4SFmode && TARGET_SSE2) | |
15715 | { | |
15716 | emit_insn (gen_sse2_pshufd_1 (new_target, new_target, | |
15717 | const1_rtx, | |
15718 | GEN_INT (one_var == 1 ? 0 : 1), | |
15719 | GEN_INT (one_var == 2 ? 0 : 1), | |
15720 | GEN_INT (one_var == 3 ? 0 : 1))); | |
15721 | if (target != new_target) | |
15722 | emit_move_insn (target, new_target); | |
15723 | return true; | |
15724 | } | |
15725 | ||
15726 | /* Otherwise convert the intermediate result to V4SFmode and | |
15727 | use the SSE1 shuffle instructions. */ | |
15728 | if (mode != V4SFmode) | |
15729 | { | |
15730 | tmp = gen_reg_rtx (V4SFmode); | |
15731 | emit_move_insn (tmp, gen_lowpart (V4SFmode, new_target)); | |
15732 | } | |
15733 | else | |
15734 | tmp = new_target; | |
15735 | ||
15736 | emit_insn (gen_sse_shufps_v4sf (tmp, tmp, tmp, | |
15737 | const1_rtx, | |
15738 | GEN_INT (one_var == 1 ? 0 : 1), | |
15739 | GEN_INT (one_var == 2 ? 0+4 : 1+4), | |
15740 | GEN_INT (one_var == 3 ? 0+4 : 1+4))); | |
15741 | ||
15742 | if (mode != V4SFmode) | |
15743 | emit_move_insn (target, gen_lowpart (V4SImode, tmp)); | |
15744 | else if (tmp != target) | |
15745 | emit_move_insn (target, tmp); | |
15746 | } | |
15747 | else if (target != new_target) | |
15748 | emit_move_insn (target, new_target); | |
15749 | return true; | |
15750 | ||
15751 | case E_V8HImode: | |
15752 | case E_V16QImode: | |
15753 | vsimode = V4SImode; | |
15754 | goto widen; | |
15755 | case E_V4HImode: | |
15756 | case E_V8QImode: | |
15757 | if (!mmx_ok) | |
15758 | return false; | |
15759 | vsimode = V2SImode; | |
15760 | goto widen; | |
15761 | widen: | |
15762 | if (one_var != 0) | |
15763 | return false; | |
15764 | ||
15765 | /* Zero extend the variable element to SImode and recurse. */ | |
15766 | var = convert_modes (SImode, GET_MODE_INNER (mode), var, true); | |
15767 | ||
15768 | x = gen_reg_rtx (vsimode); | |
15769 | if (!ix86_expand_vector_init_one_nonzero (mmx_ok, vsimode, x, | |
15770 | var, one_var)) | |
15771 | gcc_unreachable (); | |
15772 | ||
15773 | emit_move_insn (target, gen_lowpart (mode, x)); | |
15774 | return true; | |
15775 | ||
15776 | default: | |
15777 | return false; | |
15778 | } | |
15779 | } | |
15780 | ||
15781 | /* A subroutine of ix86_expand_vector_init. Store into TARGET a vector | |
15782 | consisting of the values in VALS. It is known that all elements | |
15783 | except ONE_VAR are constants. Return true if successful. */ | |
15784 | ||
15785 | static bool | |
15786 | ix86_expand_vector_init_one_var (bool mmx_ok, machine_mode mode, | |
15787 | rtx target, rtx vals, int one_var) | |
15788 | { | |
15789 | rtx var = XVECEXP (vals, 0, one_var); | |
15790 | machine_mode wmode; | |
15791 | rtx const_vec, x; | |
15792 | ||
15793 | const_vec = copy_rtx (vals); | |
15794 | XVECEXP (const_vec, 0, one_var) = CONST0_RTX (GET_MODE_INNER (mode)); | |
15795 | const_vec = gen_rtx_CONST_VECTOR (mode, XVEC (const_vec, 0)); | |
15796 | ||
15797 | switch (mode) | |
15798 | { | |
15799 | case E_V2DFmode: | |
15800 | case E_V2DImode: | |
15801 | case E_V2SFmode: | |
15802 | case E_V2SImode: | |
15803 | /* For the two element vectors, it's just as easy to use | |
15804 | the general case. */ | |
15805 | return false; | |
15806 | ||
15807 | case E_V4DImode: | |
15808 | /* Use ix86_expand_vector_set in 64bit mode only. */ | |
15809 | if (!TARGET_64BIT) | |
15810 | return false; | |
15811 | /* FALLTHRU */ | |
9e2a82e1 | 15812 | case E_V8HFmode: |
15813 | case E_V16HFmode: | |
6910cad5 | 15814 | case E_V8BFmode: |
15815 | case E_V16BFmode: | |
2bf6d935 ML |
15816 | case E_V4DFmode: |
15817 | case E_V8SFmode: | |
15818 | case E_V8SImode: | |
15819 | case E_V16HImode: | |
15820 | case E_V32QImode: | |
15821 | case E_V4SFmode: | |
15822 | case E_V4SImode: | |
15823 | case E_V8HImode: | |
15824 | case E_V4HImode: | |
15825 | break; | |
15826 | ||
15827 | case E_V16QImode: | |
15828 | if (TARGET_SSE4_1) | |
15829 | break; | |
15830 | wmode = V8HImode; | |
15831 | goto widen; | |
15832 | case E_V8QImode: | |
8a0eb0cd UB |
15833 | if (TARGET_MMX_WITH_SSE && TARGET_SSE4_1) |
15834 | break; | |
2bf6d935 ML |
15835 | wmode = V4HImode; |
15836 | goto widen; | |
64735dc9 UB |
15837 | case E_V4QImode: |
15838 | if (TARGET_SSE4_1) | |
15839 | break; | |
15840 | wmode = V2HImode; | |
2bf6d935 ML |
15841 | widen: |
15842 | /* There's no way to set one QImode entry easily. Combine | |
15843 | the variable value with its adjacent constant value, and | |
15844 | promote to an HImode set. */ | |
15845 | x = XVECEXP (vals, 0, one_var ^ 1); | |
15846 | if (one_var & 1) | |
15847 | { | |
15848 | var = convert_modes (HImode, QImode, var, true); | |
15849 | var = expand_simple_binop (HImode, ASHIFT, var, GEN_INT (8), | |
15850 | NULL_RTX, 1, OPTAB_LIB_WIDEN); | |
15851 | x = GEN_INT (INTVAL (x) & 0xff); | |
15852 | } | |
15853 | else | |
15854 | { | |
15855 | var = convert_modes (HImode, QImode, var, true); | |
15856 | x = gen_int_mode (UINTVAL (x) << 8, HImode); | |
15857 | } | |
15858 | if (x != const0_rtx) | |
15859 | var = expand_simple_binop (HImode, IOR, var, x, var, | |
15860 | 1, OPTAB_LIB_WIDEN); | |
15861 | ||
15862 | x = gen_reg_rtx (wmode); | |
15863 | emit_move_insn (x, gen_lowpart (wmode, const_vec)); | |
15864 | ix86_expand_vector_set (mmx_ok, x, var, one_var >> 1); | |
15865 | ||
15866 | emit_move_insn (target, gen_lowpart (mode, x)); | |
15867 | return true; | |
15868 | ||
15869 | default: | |
15870 | return false; | |
15871 | } | |
15872 | ||
15873 | emit_move_insn (target, const_vec); | |
15874 | ix86_expand_vector_set (mmx_ok, target, var, one_var); | |
15875 | return true; | |
15876 | } | |
15877 | ||
15878 | /* A subroutine of ix86_expand_vector_init_general. Use vector | |
15879 | concatenate to handle the most general case: all values variable, | |
15880 | and none identical. */ | |
15881 | ||
15882 | static void | |
15883 | ix86_expand_vector_init_concat (machine_mode mode, | |
15884 | rtx target, rtx *ops, int n) | |
15885 | { | |
1aeecaf5 HL |
15886 | machine_mode half_mode = VOIDmode; |
15887 | rtx half[2]; | |
2bf6d935 ML |
15888 | rtvec v; |
15889 | int i, j; | |
15890 | ||
15891 | switch (n) | |
15892 | { | |
15893 | case 2: | |
15894 | switch (mode) | |
15895 | { | |
9e2a82e1 | 15896 | case E_V32HFmode: |
15897 | half_mode = V16HFmode; | |
15898 | break; | |
6910cad5 | 15899 | case E_V32BFmode: |
15900 | half_mode = V16BFmode; | |
15901 | break; | |
2bf6d935 | 15902 | case E_V16SImode: |
1aeecaf5 | 15903 | half_mode = V8SImode; |
2bf6d935 ML |
15904 | break; |
15905 | case E_V16SFmode: | |
1aeecaf5 | 15906 | half_mode = V8SFmode; |
2bf6d935 ML |
15907 | break; |
15908 | case E_V8DImode: | |
1aeecaf5 | 15909 | half_mode = V4DImode; |
2bf6d935 ML |
15910 | break; |
15911 | case E_V8DFmode: | |
1aeecaf5 | 15912 | half_mode = V4DFmode; |
2bf6d935 | 15913 | break; |
9e2a82e1 | 15914 | case E_V16HFmode: |
15915 | half_mode = V8HFmode; | |
15916 | break; | |
6910cad5 | 15917 | case E_V16BFmode: |
15918 | half_mode = V8BFmode; | |
15919 | break; | |
2bf6d935 | 15920 | case E_V8SImode: |
1aeecaf5 | 15921 | half_mode = V4SImode; |
2bf6d935 ML |
15922 | break; |
15923 | case E_V8SFmode: | |
1aeecaf5 | 15924 | half_mode = V4SFmode; |
2bf6d935 ML |
15925 | break; |
15926 | case E_V4DImode: | |
1aeecaf5 | 15927 | half_mode = V2DImode; |
2bf6d935 ML |
15928 | break; |
15929 | case E_V4DFmode: | |
1aeecaf5 | 15930 | half_mode = V2DFmode; |
2bf6d935 ML |
15931 | break; |
15932 | case E_V4SImode: | |
1aeecaf5 | 15933 | half_mode = V2SImode; |
2bf6d935 ML |
15934 | break; |
15935 | case E_V4SFmode: | |
1aeecaf5 | 15936 | half_mode = V2SFmode; |
2bf6d935 ML |
15937 | break; |
15938 | case E_V2DImode: | |
1aeecaf5 | 15939 | half_mode = DImode; |
2bf6d935 ML |
15940 | break; |
15941 | case E_V2SImode: | |
1aeecaf5 | 15942 | half_mode = SImode; |
2bf6d935 ML |
15943 | break; |
15944 | case E_V2DFmode: | |
1aeecaf5 | 15945 | half_mode = DFmode; |
2bf6d935 ML |
15946 | break; |
15947 | case E_V2SFmode: | |
1aeecaf5 | 15948 | half_mode = SFmode; |
2bf6d935 ML |
15949 | break; |
15950 | default: | |
15951 | gcc_unreachable (); | |
15952 | } | |
15953 | ||
1aeecaf5 HL |
15954 | if (!register_operand (ops[1], half_mode)) |
15955 | ops[1] = force_reg (half_mode, ops[1]); | |
15956 | if (!register_operand (ops[0], half_mode)) | |
15957 | ops[0] = force_reg (half_mode, ops[0]); | |
2bf6d935 ML |
15958 | emit_insn (gen_rtx_SET (target, gen_rtx_VEC_CONCAT (mode, ops[0], |
15959 | ops[1]))); | |
15960 | break; | |
15961 | ||
15962 | case 4: | |
15963 | switch (mode) | |
15964 | { | |
15965 | case E_V4DImode: | |
1aeecaf5 | 15966 | half_mode = V2DImode; |
2bf6d935 ML |
15967 | break; |
15968 | case E_V4DFmode: | |
1aeecaf5 | 15969 | half_mode = V2DFmode; |
2bf6d935 ML |
15970 | break; |
15971 | case E_V4SImode: | |
1aeecaf5 | 15972 | half_mode = V2SImode; |
2bf6d935 ML |
15973 | break; |
15974 | case E_V4SFmode: | |
1aeecaf5 | 15975 | half_mode = V2SFmode; |
2bf6d935 ML |
15976 | break; |
15977 | default: | |
15978 | gcc_unreachable (); | |
15979 | } | |
15980 | goto half; | |
15981 | ||
15982 | case 8: | |
15983 | switch (mode) | |
15984 | { | |
15985 | case E_V8DImode: | |
1aeecaf5 | 15986 | half_mode = V4DImode; |
2bf6d935 ML |
15987 | break; |
15988 | case E_V8DFmode: | |
1aeecaf5 | 15989 | half_mode = V4DFmode; |
2bf6d935 ML |
15990 | break; |
15991 | case E_V8SImode: | |
1aeecaf5 | 15992 | half_mode = V4SImode; |
2bf6d935 ML |
15993 | break; |
15994 | case E_V8SFmode: | |
1aeecaf5 | 15995 | half_mode = V4SFmode; |
2bf6d935 ML |
15996 | break; |
15997 | default: | |
15998 | gcc_unreachable (); | |
15999 | } | |
16000 | goto half; | |
16001 | ||
16002 | case 16: | |
16003 | switch (mode) | |
16004 | { | |
16005 | case E_V16SImode: | |
1aeecaf5 | 16006 | half_mode = V8SImode; |
2bf6d935 ML |
16007 | break; |
16008 | case E_V16SFmode: | |
1aeecaf5 | 16009 | half_mode = V8SFmode; |
2bf6d935 ML |
16010 | break; |
16011 | default: | |
16012 | gcc_unreachable (); | |
16013 | } | |
16014 | goto half; | |
16015 | ||
16016 | half: | |
16017 | /* FIXME: We process inputs backward to help RA. PR 36222. */ | |
16018 | i = n - 1; | |
1aeecaf5 | 16019 | for (j = 1; j != -1; j--) |
2bf6d935 | 16020 | { |
1aeecaf5 HL |
16021 | half[j] = gen_reg_rtx (half_mode); |
16022 | switch (n >> 1) | |
2bf6d935 | 16023 | { |
1aeecaf5 HL |
16024 | case 2: |
16025 | v = gen_rtvec (2, ops[i-1], ops[i]); | |
16026 | i -= 2; | |
16027 | break; | |
16028 | case 4: | |
16029 | v = gen_rtvec (4, ops[i-3], ops[i-2], ops[i-1], ops[i]); | |
16030 | i -= 4; | |
16031 | break; | |
16032 | case 8: | |
16033 | v = gen_rtvec (8, ops[i-7], ops[i-6], ops[i-5], ops[i-4], | |
16034 | ops[i-3], ops[i-2], ops[i-1], ops[i]); | |
16035 | i -= 8; | |
16036 | break; | |
16037 | default: | |
16038 | gcc_unreachable (); | |
2bf6d935 | 16039 | } |
1aeecaf5 HL |
16040 | ix86_expand_vector_init (false, half[j], |
16041 | gen_rtx_PARALLEL (half_mode, v)); | |
2bf6d935 | 16042 | } |
1aeecaf5 HL |
16043 | |
16044 | ix86_expand_vector_init_concat (mode, target, half, 2); | |
2bf6d935 ML |
16045 | break; |
16046 | ||
16047 | default: | |
16048 | gcc_unreachable (); | |
16049 | } | |
16050 | } | |
16051 | ||
16052 | /* A subroutine of ix86_expand_vector_init_general. Use vector | |
16053 | interleave to handle the most general case: all values variable, | |
16054 | and none identical. */ | |
16055 | ||
16056 | static void | |
16057 | ix86_expand_vector_init_interleave (machine_mode mode, | |
16058 | rtx target, rtx *ops, int n) | |
16059 | { | |
16060 | machine_mode first_imode, second_imode, third_imode, inner_mode; | |
16061 | int i, j; | |
9e2a82e1 | 16062 | rtx op, op0, op1; |
2bf6d935 ML |
16063 | rtx (*gen_load_even) (rtx, rtx, rtx); |
16064 | rtx (*gen_interleave_first_low) (rtx, rtx, rtx); | |
16065 | rtx (*gen_interleave_second_low) (rtx, rtx, rtx); | |
16066 | ||
16067 | switch (mode) | |
16068 | { | |
9e2a82e1 | 16069 | case E_V8HFmode: |
7fc4d600 | 16070 | gen_load_even = gen_vec_interleave_lowv8hf; |
9e2a82e1 | 16071 | gen_interleave_first_low = gen_vec_interleave_lowv4si; |
16072 | gen_interleave_second_low = gen_vec_interleave_lowv2di; | |
16073 | inner_mode = HFmode; | |
16074 | first_imode = V4SImode; | |
16075 | second_imode = V2DImode; | |
16076 | third_imode = VOIDmode; | |
16077 | break; | |
6910cad5 | 16078 | case E_V8BFmode: |
16079 | gen_load_even = gen_vec_interleave_lowv8bf; | |
16080 | gen_interleave_first_low = gen_vec_interleave_lowv4si; | |
16081 | gen_interleave_second_low = gen_vec_interleave_lowv2di; | |
16082 | inner_mode = BFmode; | |
16083 | first_imode = V4SImode; | |
16084 | second_imode = V2DImode; | |
16085 | third_imode = VOIDmode; | |
16086 | break; | |
2bf6d935 ML |
16087 | case E_V8HImode: |
16088 | gen_load_even = gen_vec_setv8hi; | |
16089 | gen_interleave_first_low = gen_vec_interleave_lowv4si; | |
16090 | gen_interleave_second_low = gen_vec_interleave_lowv2di; | |
16091 | inner_mode = HImode; | |
16092 | first_imode = V4SImode; | |
16093 | second_imode = V2DImode; | |
16094 | third_imode = VOIDmode; | |
16095 | break; | |
16096 | case E_V16QImode: | |
16097 | gen_load_even = gen_vec_setv16qi; | |
16098 | gen_interleave_first_low = gen_vec_interleave_lowv8hi; | |
16099 | gen_interleave_second_low = gen_vec_interleave_lowv4si; | |
16100 | inner_mode = QImode; | |
16101 | first_imode = V8HImode; | |
16102 | second_imode = V4SImode; | |
16103 | third_imode = V2DImode; | |
16104 | break; | |
16105 | default: | |
16106 | gcc_unreachable (); | |
16107 | } | |
16108 | ||
16109 | for (i = 0; i < n; i++) | |
16110 | { | |
9e2a82e1 | 16111 | op = ops [i + i]; |
6910cad5 | 16112 | if (inner_mode == HFmode || inner_mode == BFmode) |
9e2a82e1 | 16113 | { |
7fc4d600 | 16114 | rtx even, odd; |
6910cad5 | 16115 | /* Use vpuncklwd to pack 2 HFmode or BFmode. */ |
16116 | machine_mode vec_mode = | |
16117 | (inner_mode == HFmode) ? V8HFmode : V8BFmode; | |
16118 | op0 = gen_reg_rtx (vec_mode); | |
16119 | even = lowpart_subreg (vec_mode, | |
16120 | force_reg (inner_mode, op), inner_mode); | |
16121 | odd = lowpart_subreg (vec_mode, | |
16122 | force_reg (inner_mode, ops[i + i + 1]), | |
16123 | inner_mode); | |
7fc4d600 | 16124 | emit_insn (gen_load_even (op0, even, odd)); |
9e2a82e1 | 16125 | } |
7fc4d600 | 16126 | else |
16127 | { | |
16128 | /* Extend the odd elment to SImode using a paradoxical SUBREG. */ | |
16129 | op0 = gen_reg_rtx (SImode); | |
16130 | emit_move_insn (op0, gen_lowpart (SImode, op)); | |
9e2a82e1 | 16131 | |
7fc4d600 | 16132 | /* Insert the SImode value as low element of V4SImode vector. */ |
16133 | op1 = gen_reg_rtx (V4SImode); | |
16134 | op0 = gen_rtx_VEC_MERGE (V4SImode, | |
16135 | gen_rtx_VEC_DUPLICATE (V4SImode, | |
16136 | op0), | |
16137 | CONST0_RTX (V4SImode), | |
16138 | const1_rtx); | |
16139 | emit_insn (gen_rtx_SET (op1, op0)); | |
2bf6d935 | 16140 | |
7fc4d600 | 16141 | /* Cast the V4SImode vector back to a vector in orignal mode. */ |
16142 | op0 = gen_reg_rtx (mode); | |
16143 | emit_move_insn (op0, gen_lowpart (mode, op1)); | |
2bf6d935 | 16144 | |
7fc4d600 | 16145 | /* Load even elements into the second position. */ |
16146 | emit_insn (gen_load_even (op0, | |
16147 | force_reg (inner_mode, | |
16148 | ops[i + i + 1]), | |
16149 | const1_rtx)); | |
16150 | } | |
2bf6d935 ML |
16151 | |
16152 | /* Cast vector to FIRST_IMODE vector. */ | |
16153 | ops[i] = gen_reg_rtx (first_imode); | |
16154 | emit_move_insn (ops[i], gen_lowpart (first_imode, op0)); | |
16155 | } | |
16156 | ||
16157 | /* Interleave low FIRST_IMODE vectors. */ | |
16158 | for (i = j = 0; i < n; i += 2, j++) | |
16159 | { | |
16160 | op0 = gen_reg_rtx (first_imode); | |
16161 | emit_insn (gen_interleave_first_low (op0, ops[i], ops[i + 1])); | |
16162 | ||
16163 | /* Cast FIRST_IMODE vector to SECOND_IMODE vector. */ | |
16164 | ops[j] = gen_reg_rtx (second_imode); | |
16165 | emit_move_insn (ops[j], gen_lowpart (second_imode, op0)); | |
16166 | } | |
16167 | ||
16168 | /* Interleave low SECOND_IMODE vectors. */ | |
16169 | switch (second_imode) | |
16170 | { | |
16171 | case E_V4SImode: | |
16172 | for (i = j = 0; i < n / 2; i += 2, j++) | |
16173 | { | |
16174 | op0 = gen_reg_rtx (second_imode); | |
16175 | emit_insn (gen_interleave_second_low (op0, ops[i], | |
16176 | ops[i + 1])); | |
16177 | ||
16178 | /* Cast the SECOND_IMODE vector to the THIRD_IMODE | |
16179 | vector. */ | |
16180 | ops[j] = gen_reg_rtx (third_imode); | |
16181 | emit_move_insn (ops[j], gen_lowpart (third_imode, op0)); | |
16182 | } | |
16183 | second_imode = V2DImode; | |
16184 | gen_interleave_second_low = gen_vec_interleave_lowv2di; | |
16185 | /* FALLTHRU */ | |
16186 | ||
16187 | case E_V2DImode: | |
16188 | op0 = gen_reg_rtx (second_imode); | |
16189 | emit_insn (gen_interleave_second_low (op0, ops[0], | |
16190 | ops[1])); | |
16191 | ||
16192 | /* Cast the SECOND_IMODE vector back to a vector on original | |
16193 | mode. */ | |
16194 | emit_insn (gen_rtx_SET (target, gen_lowpart (mode, op0))); | |
16195 | break; | |
16196 | ||
16197 | default: | |
16198 | gcc_unreachable (); | |
16199 | } | |
16200 | } | |
16201 | ||
16202 | /* A subroutine of ix86_expand_vector_init. Handle the most general case: | |
16203 | all values variable, and none identical. */ | |
16204 | ||
16205 | static void | |
16206 | ix86_expand_vector_init_general (bool mmx_ok, machine_mode mode, | |
16207 | rtx target, rtx vals) | |
16208 | { | |
16209 | rtx ops[64], op0, op1, op2, op3, op4, op5; | |
16210 | machine_mode half_mode = VOIDmode; | |
16211 | machine_mode quarter_mode = VOIDmode; | |
16212 | int n, i; | |
16213 | ||
16214 | switch (mode) | |
16215 | { | |
16216 | case E_V2SFmode: | |
16217 | case E_V2SImode: | |
16218 | if (!mmx_ok && !TARGET_SSE) | |
16219 | break; | |
16220 | /* FALLTHRU */ | |
16221 | ||
16222 | case E_V16SImode: | |
16223 | case E_V16SFmode: | |
16224 | case E_V8DFmode: | |
16225 | case E_V8DImode: | |
16226 | case E_V8SFmode: | |
16227 | case E_V8SImode: | |
16228 | case E_V4DFmode: | |
16229 | case E_V4DImode: | |
16230 | case E_V4SFmode: | |
16231 | case E_V4SImode: | |
16232 | case E_V2DFmode: | |
16233 | case E_V2DImode: | |
16234 | n = GET_MODE_NUNITS (mode); | |
16235 | for (i = 0; i < n; i++) | |
16236 | ops[i] = XVECEXP (vals, 0, i); | |
16237 | ix86_expand_vector_init_concat (mode, target, ops, n); | |
16238 | return; | |
16239 | ||
16240 | case E_V2TImode: | |
16241 | for (i = 0; i < 2; i++) | |
16242 | ops[i] = gen_lowpart (V2DImode, XVECEXP (vals, 0, i)); | |
16243 | op0 = gen_reg_rtx (V4DImode); | |
16244 | ix86_expand_vector_init_concat (V4DImode, op0, ops, 2); | |
16245 | emit_move_insn (target, gen_lowpart (GET_MODE (target), op0)); | |
16246 | return; | |
16247 | ||
16248 | case E_V4TImode: | |
16249 | for (i = 0; i < 4; i++) | |
16250 | ops[i] = gen_lowpart (V2DImode, XVECEXP (vals, 0, i)); | |
16251 | ops[4] = gen_reg_rtx (V4DImode); | |
16252 | ix86_expand_vector_init_concat (V4DImode, ops[4], ops, 2); | |
16253 | ops[5] = gen_reg_rtx (V4DImode); | |
16254 | ix86_expand_vector_init_concat (V4DImode, ops[5], ops + 2, 2); | |
16255 | op0 = gen_reg_rtx (V8DImode); | |
16256 | ix86_expand_vector_init_concat (V8DImode, op0, ops + 4, 2); | |
16257 | emit_move_insn (target, gen_lowpart (GET_MODE (target), op0)); | |
16258 | return; | |
16259 | ||
16260 | case E_V32QImode: | |
16261 | half_mode = V16QImode; | |
16262 | goto half; | |
16263 | ||
16264 | case E_V16HImode: | |
16265 | half_mode = V8HImode; | |
16266 | goto half; | |
16267 | ||
9e2a82e1 | 16268 | case E_V16HFmode: |
16269 | half_mode = V8HFmode; | |
16270 | goto half; | |
16271 | ||
6910cad5 | 16272 | case E_V16BFmode: |
16273 | half_mode = V8BFmode; | |
16274 | goto half; | |
16275 | ||
2bf6d935 ML |
16276 | half: |
16277 | n = GET_MODE_NUNITS (mode); | |
16278 | for (i = 0; i < n; i++) | |
16279 | ops[i] = XVECEXP (vals, 0, i); | |
16280 | op0 = gen_reg_rtx (half_mode); | |
16281 | op1 = gen_reg_rtx (half_mode); | |
16282 | ix86_expand_vector_init_interleave (half_mode, op0, ops, | |
16283 | n >> 2); | |
16284 | ix86_expand_vector_init_interleave (half_mode, op1, | |
16285 | &ops [n >> 1], n >> 2); | |
16286 | emit_insn (gen_rtx_SET (target, gen_rtx_VEC_CONCAT (mode, op0, op1))); | |
16287 | return; | |
16288 | ||
16289 | case E_V64QImode: | |
16290 | quarter_mode = V16QImode; | |
16291 | half_mode = V32QImode; | |
16292 | goto quarter; | |
16293 | ||
16294 | case E_V32HImode: | |
16295 | quarter_mode = V8HImode; | |
16296 | half_mode = V16HImode; | |
16297 | goto quarter; | |
16298 | ||
9e2a82e1 | 16299 | case E_V32HFmode: |
16300 | quarter_mode = V8HFmode; | |
16301 | half_mode = V16HFmode; | |
16302 | goto quarter; | |
16303 | ||
6910cad5 | 16304 | case E_V32BFmode: |
16305 | quarter_mode = V8BFmode; | |
16306 | half_mode = V16BFmode; | |
16307 | goto quarter; | |
16308 | ||
2bf6d935 ML |
16309 | quarter: |
16310 | n = GET_MODE_NUNITS (mode); | |
16311 | for (i = 0; i < n; i++) | |
16312 | ops[i] = XVECEXP (vals, 0, i); | |
16313 | op0 = gen_reg_rtx (quarter_mode); | |
16314 | op1 = gen_reg_rtx (quarter_mode); | |
16315 | op2 = gen_reg_rtx (quarter_mode); | |
16316 | op3 = gen_reg_rtx (quarter_mode); | |
16317 | op4 = gen_reg_rtx (half_mode); | |
16318 | op5 = gen_reg_rtx (half_mode); | |
16319 | ix86_expand_vector_init_interleave (quarter_mode, op0, ops, | |
16320 | n >> 3); | |
16321 | ix86_expand_vector_init_interleave (quarter_mode, op1, | |
16322 | &ops [n >> 2], n >> 3); | |
16323 | ix86_expand_vector_init_interleave (quarter_mode, op2, | |
16324 | &ops [n >> 1], n >> 3); | |
16325 | ix86_expand_vector_init_interleave (quarter_mode, op3, | |
16326 | &ops [(n >> 1) | (n >> 2)], n >> 3); | |
16327 | emit_insn (gen_rtx_SET (op4, gen_rtx_VEC_CONCAT (half_mode, op0, op1))); | |
16328 | emit_insn (gen_rtx_SET (op5, gen_rtx_VEC_CONCAT (half_mode, op2, op3))); | |
16329 | emit_insn (gen_rtx_SET (target, gen_rtx_VEC_CONCAT (mode, op4, op5))); | |
16330 | return; | |
16331 | ||
16332 | case E_V16QImode: | |
16333 | if (!TARGET_SSE4_1) | |
16334 | break; | |
16335 | /* FALLTHRU */ | |
16336 | ||
16337 | case E_V8HImode: | |
16338 | if (!TARGET_SSE2) | |
16339 | break; | |
16340 | ||
16341 | /* Don't use ix86_expand_vector_init_interleave if we can't | |
16342 | move from GPR to SSE register directly. */ | |
16343 | if (!TARGET_INTER_UNIT_MOVES_TO_VEC) | |
16344 | break; | |
9e2a82e1 | 16345 | /* FALLTHRU */ |
16346 | ||
16347 | case E_V8HFmode: | |
6910cad5 | 16348 | case E_V8BFmode: |
2bf6d935 ML |
16349 | |
16350 | n = GET_MODE_NUNITS (mode); | |
16351 | for (i = 0; i < n; i++) | |
16352 | ops[i] = XVECEXP (vals, 0, i); | |
16353 | ix86_expand_vector_init_interleave (mode, target, ops, n >> 1); | |
16354 | return; | |
16355 | ||
16356 | case E_V4HImode: | |
16357 | case E_V8QImode: | |
8d7dae0e UB |
16358 | |
16359 | case E_V2HImode: | |
64735dc9 | 16360 | case E_V4QImode: |
2bf6d935 ML |
16361 | break; |
16362 | ||
16363 | default: | |
16364 | gcc_unreachable (); | |
16365 | } | |
16366 | ||
16367 | { | |
16368 | int i, j, n_elts, n_words, n_elt_per_word; | |
8d7dae0e | 16369 | machine_mode tmp_mode, inner_mode; |
2bf6d935 ML |
16370 | rtx words[4], shift; |
16371 | ||
8d7dae0e UB |
16372 | tmp_mode = (GET_MODE_SIZE (mode) < UNITS_PER_WORD) ? SImode : word_mode; |
16373 | ||
2bf6d935 ML |
16374 | inner_mode = GET_MODE_INNER (mode); |
16375 | n_elts = GET_MODE_NUNITS (mode); | |
8d7dae0e | 16376 | n_words = GET_MODE_SIZE (mode) / GET_MODE_SIZE (tmp_mode); |
2bf6d935 ML |
16377 | n_elt_per_word = n_elts / n_words; |
16378 | shift = GEN_INT (GET_MODE_BITSIZE (inner_mode)); | |
16379 | ||
16380 | for (i = 0; i < n_words; ++i) | |
16381 | { | |
16382 | rtx word = NULL_RTX; | |
16383 | ||
16384 | for (j = 0; j < n_elt_per_word; ++j) | |
16385 | { | |
16386 | rtx elt = XVECEXP (vals, 0, (i+1)*n_elt_per_word - j - 1); | |
8d7dae0e | 16387 | elt = convert_modes (tmp_mode, inner_mode, elt, true); |
2bf6d935 ML |
16388 | |
16389 | if (j == 0) | |
16390 | word = elt; | |
16391 | else | |
16392 | { | |
8d7dae0e | 16393 | word = expand_simple_binop (tmp_mode, ASHIFT, word, shift, |
e1a74058 | 16394 | NULL_RTX, 1, OPTAB_LIB_WIDEN); |
8d7dae0e | 16395 | word = expand_simple_binop (tmp_mode, IOR, word, elt, |
e1a74058 | 16396 | NULL_RTX, 1, OPTAB_LIB_WIDEN); |
2bf6d935 ML |
16397 | } |
16398 | } | |
16399 | ||
16400 | words[i] = word; | |
16401 | } | |
16402 | ||
16403 | if (n_words == 1) | |
16404 | emit_move_insn (target, gen_lowpart (mode, words[0])); | |
16405 | else if (n_words == 2) | |
16406 | { | |
affee7dc RB |
16407 | gcc_assert (tmp_mode == DImode || tmp_mode == SImode); |
16408 | machine_mode concat_mode = tmp_mode == DImode ? V2DImode : V2SImode; | |
16409 | rtx tmp = gen_reg_rtx (concat_mode); | |
16410 | vals = gen_rtx_PARALLEL (concat_mode, gen_rtvec_v (2, words)); | |
2b2bf793 | 16411 | ix86_expand_vector_init_general (mmx_ok, concat_mode, tmp, vals); |
affee7dc | 16412 | emit_move_insn (target, gen_lowpart (mode, tmp)); |
2bf6d935 ML |
16413 | } |
16414 | else if (n_words == 4) | |
16415 | { | |
16416 | rtx tmp = gen_reg_rtx (V4SImode); | |
8d7dae0e | 16417 | gcc_assert (tmp_mode == SImode); |
2bf6d935 ML |
16418 | vals = gen_rtx_PARALLEL (V4SImode, gen_rtvec_v (4, words)); |
16419 | ix86_expand_vector_init_general (false, V4SImode, tmp, vals); | |
16420 | emit_move_insn (target, gen_lowpart (mode, tmp)); | |
16421 | } | |
16422 | else | |
16423 | gcc_unreachable (); | |
16424 | } | |
16425 | } | |
16426 | ||
16427 | /* Initialize vector TARGET via VALS. Suppress the use of MMX | |
16428 | instructions unless MMX_OK is true. */ | |
16429 | ||
16430 | void | |
16431 | ix86_expand_vector_init (bool mmx_ok, rtx target, rtx vals) | |
16432 | { | |
16433 | machine_mode mode = GET_MODE (target); | |
16434 | machine_mode inner_mode = GET_MODE_INNER (mode); | |
16435 | int n_elts = GET_MODE_NUNITS (mode); | |
16436 | int n_var = 0, one_var = -1; | |
16437 | bool all_same = true, all_const_zero = true; | |
16438 | int i; | |
16439 | rtx x; | |
16440 | ||
16441 | /* Handle first initialization from vector elts. */ | |
16442 | if (n_elts != XVECLEN (vals, 0)) | |
16443 | { | |
16444 | rtx subtarget = target; | |
16445 | x = XVECEXP (vals, 0, 0); | |
16446 | gcc_assert (GET_MODE_INNER (GET_MODE (x)) == inner_mode); | |
16447 | if (GET_MODE_NUNITS (GET_MODE (x)) * 2 == n_elts) | |
16448 | { | |
16449 | rtx ops[2] = { XVECEXP (vals, 0, 0), XVECEXP (vals, 0, 1) }; | |
b7dd2e4e JJ |
16450 | if (inner_mode == QImode |
16451 | || inner_mode == HImode | |
575191b9 | 16452 | || inner_mode == TImode |
6910cad5 | 16453 | || inner_mode == HFmode |
16454 | || inner_mode == BFmode) | |
2bf6d935 ML |
16455 | { |
16456 | unsigned int n_bits = n_elts * GET_MODE_SIZE (inner_mode); | |
b7dd2e4e JJ |
16457 | scalar_mode elt_mode = inner_mode == TImode ? DImode : SImode; |
16458 | n_bits /= GET_MODE_SIZE (elt_mode); | |
16459 | mode = mode_for_vector (elt_mode, n_bits).require (); | |
16460 | inner_mode = mode_for_vector (elt_mode, n_bits / 2).require (); | |
2bf6d935 ML |
16461 | ops[0] = gen_lowpart (inner_mode, ops[0]); |
16462 | ops[1] = gen_lowpart (inner_mode, ops[1]); | |
16463 | subtarget = gen_reg_rtx (mode); | |
16464 | } | |
16465 | ix86_expand_vector_init_concat (mode, subtarget, ops, 2); | |
16466 | if (subtarget != target) | |
16467 | emit_move_insn (target, gen_lowpart (GET_MODE (target), subtarget)); | |
16468 | return; | |
16469 | } | |
16470 | gcc_unreachable (); | |
16471 | } | |
16472 | ||
16473 | for (i = 0; i < n_elts; ++i) | |
16474 | { | |
16475 | x = XVECEXP (vals, 0, i); | |
16476 | if (!(CONST_SCALAR_INT_P (x) | |
16477 | || CONST_DOUBLE_P (x) | |
16478 | || CONST_FIXED_P (x))) | |
16479 | n_var++, one_var = i; | |
16480 | else if (x != CONST0_RTX (inner_mode)) | |
16481 | all_const_zero = false; | |
16482 | if (i > 0 && !rtx_equal_p (x, XVECEXP (vals, 0, 0))) | |
16483 | all_same = false; | |
16484 | } | |
16485 | ||
16486 | /* Constants are best loaded from the constant pool. */ | |
16487 | if (n_var == 0) | |
16488 | { | |
16489 | emit_move_insn (target, gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0))); | |
16490 | return; | |
16491 | } | |
16492 | ||
16493 | /* If all values are identical, broadcast the value. */ | |
16494 | if (all_same | |
16495 | && ix86_expand_vector_init_duplicate (mmx_ok, mode, target, | |
16496 | XVECEXP (vals, 0, 0))) | |
16497 | return; | |
16498 | ||
16499 | /* Values where only one field is non-constant are best loaded from | |
16500 | the pool and overwritten via move later. */ | |
16501 | if (n_var == 1) | |
16502 | { | |
16503 | if (all_const_zero | |
16504 | && ix86_expand_vector_init_one_nonzero (mmx_ok, mode, target, | |
16505 | XVECEXP (vals, 0, one_var), | |
16506 | one_var)) | |
16507 | return; | |
16508 | ||
16509 | if (ix86_expand_vector_init_one_var (mmx_ok, mode, target, vals, one_var)) | |
16510 | return; | |
16511 | } | |
16512 | ||
16513 | ix86_expand_vector_init_general (mmx_ok, mode, target, vals); | |
16514 | } | |
16515 | ||
287cc750 | 16516 | /* Implemented as |
16517 | V setg (V v, int idx, T val) | |
16518 | { | |
16519 | V idxv = (V){idx, idx, idx, idx, idx, idx, idx, idx}; | |
16520 | V valv = (V){val, val, val, val, val, val, val, val}; | |
16521 | V mask = ((V){0, 1, 2, 3, 4, 5, 6, 7} == idxv); | |
16522 | v = (v & ~mask) | (valv & mask); | |
16523 | return v; | |
16524 | }. */ | |
16525 | void | |
16526 | ix86_expand_vector_set_var (rtx target, rtx val, rtx idx) | |
16527 | { | |
16528 | rtx vec[64]; | |
16529 | machine_mode mode = GET_MODE (target); | |
16530 | machine_mode cmp_mode = mode; | |
16531 | int n_elts = GET_MODE_NUNITS (mode); | |
16532 | rtx valv,idxv,constv,idx_tmp; | |
16533 | bool ok = false; | |
16534 | ||
16535 | /* 512-bits vector byte/word broadcast and comparison only available | |
16536 | under TARGET_AVX512BW, break 512-bits vector into two 256-bits vector | |
16537 | when without TARGET_AVX512BW. */ | |
6910cad5 | 16538 | if ((mode == V32HImode || mode == V32HFmode || mode == V32BFmode |
16539 | || mode == V64QImode) | |
7a54d3de | 16540 | && !TARGET_AVX512BW) |
287cc750 | 16541 | { |
16542 | gcc_assert (TARGET_AVX512F); | |
16543 | rtx vhi, vlo, idx_hi; | |
16544 | machine_mode half_mode; | |
16545 | rtx (*extract_hi)(rtx, rtx); | |
16546 | rtx (*extract_lo)(rtx, rtx); | |
16547 | ||
16548 | if (mode == V32HImode) | |
16549 | { | |
16550 | half_mode = V16HImode; | |
16551 | extract_hi = gen_vec_extract_hi_v32hi; | |
16552 | extract_lo = gen_vec_extract_lo_v32hi; | |
16553 | } | |
7a54d3de UB |
16554 | else if (mode == V32HFmode) |
16555 | { | |
16556 | half_mode = V16HFmode; | |
16557 | extract_hi = gen_vec_extract_hi_v32hf; | |
16558 | extract_lo = gen_vec_extract_lo_v32hf; | |
16559 | } | |
6910cad5 | 16560 | else if (mode == V32BFmode) |
16561 | { | |
16562 | half_mode = V16BFmode; | |
16563 | extract_hi = gen_vec_extract_hi_v32bf; | |
16564 | extract_lo = gen_vec_extract_lo_v32bf; | |
16565 | } | |
287cc750 | 16566 | else |
16567 | { | |
16568 | half_mode = V32QImode; | |
16569 | extract_hi = gen_vec_extract_hi_v64qi; | |
16570 | extract_lo = gen_vec_extract_lo_v64qi; | |
16571 | } | |
16572 | ||
16573 | vhi = gen_reg_rtx (half_mode); | |
16574 | vlo = gen_reg_rtx (half_mode); | |
16575 | idx_hi = gen_reg_rtx (GET_MODE (idx)); | |
16576 | emit_insn (extract_hi (vhi, target)); | |
16577 | emit_insn (extract_lo (vlo, target)); | |
16578 | vec[0] = idx_hi; | |
16579 | vec[1] = idx; | |
16580 | vec[2] = GEN_INT (n_elts/2); | |
16581 | ix86_expand_binary_operator (MINUS, GET_MODE (idx), vec); | |
16582 | ix86_expand_vector_set_var (vhi, val, idx_hi); | |
16583 | ix86_expand_vector_set_var (vlo, val, idx); | |
16584 | emit_insn (gen_rtx_SET (target, gen_rtx_VEC_CONCAT (mode, vlo, vhi))); | |
16585 | return; | |
16586 | } | |
16587 | ||
16588 | if (FLOAT_MODE_P (GET_MODE_INNER (mode))) | |
16589 | { | |
16590 | switch (mode) | |
16591 | { | |
16592 | case E_V2DFmode: | |
16593 | cmp_mode = V2DImode; | |
16594 | break; | |
16595 | case E_V4DFmode: | |
16596 | cmp_mode = V4DImode; | |
16597 | break; | |
16598 | case E_V8DFmode: | |
16599 | cmp_mode = V8DImode; | |
16600 | break; | |
20a2c8ac UB |
16601 | case E_V2SFmode: |
16602 | cmp_mode = V2SImode; | |
16603 | break; | |
287cc750 | 16604 | case E_V4SFmode: |
16605 | cmp_mode = V4SImode; | |
16606 | break; | |
16607 | case E_V8SFmode: | |
16608 | cmp_mode = V8SImode; | |
16609 | break; | |
16610 | case E_V16SFmode: | |
16611 | cmp_mode = V16SImode; | |
16612 | break; | |
9e2a82e1 | 16613 | case E_V8HFmode: |
16614 | cmp_mode = V8HImode; | |
16615 | break; | |
16616 | case E_V16HFmode: | |
16617 | cmp_mode = V16HImode; | |
16618 | break; | |
16619 | case E_V32HFmode: | |
16620 | cmp_mode = V32HImode; | |
16621 | break; | |
6910cad5 | 16622 | case E_V8BFmode: |
16623 | cmp_mode = V8HImode; | |
16624 | break; | |
16625 | case E_V16BFmode: | |
16626 | cmp_mode = V16HImode; | |
16627 | break; | |
16628 | case E_V32BFmode: | |
16629 | cmp_mode = V32HImode; | |
16630 | break; | |
287cc750 | 16631 | default: |
16632 | gcc_unreachable (); | |
16633 | } | |
16634 | } | |
16635 | ||
16636 | for (int i = 0; i != n_elts; i++) | |
16637 | vec[i] = GEN_INT (i); | |
16638 | constv = gen_rtx_CONST_VECTOR (cmp_mode, gen_rtvec_v (n_elts, vec)); | |
16639 | valv = gen_reg_rtx (mode); | |
16640 | idxv = gen_reg_rtx (cmp_mode); | |
16641 | idx_tmp = convert_to_mode (GET_MODE_INNER (cmp_mode), idx, 1); | |
16642 | ||
20a2c8ac UB |
16643 | ok = ix86_expand_vector_init_duplicate (TARGET_MMX_WITH_SSE, |
16644 | mode, valv, val); | |
287cc750 | 16645 | gcc_assert (ok); |
20a2c8ac UB |
16646 | ok = ix86_expand_vector_init_duplicate (TARGET_MMX_WITH_SSE, |
16647 | cmp_mode, idxv, idx_tmp); | |
287cc750 | 16648 | gcc_assert (ok); |
16649 | vec[0] = target; | |
16650 | vec[1] = valv; | |
16651 | vec[2] = target; | |
16652 | vec[3] = gen_rtx_EQ (mode, idxv, constv); | |
16653 | vec[4] = idxv; | |
16654 | vec[5] = constv; | |
16655 | ok = ix86_expand_int_vcond (vec); | |
16656 | gcc_assert (ok); | |
16657 | } | |
16658 | ||
2bf6d935 ML |
16659 | void |
16660 | ix86_expand_vector_set (bool mmx_ok, rtx target, rtx val, int elt) | |
16661 | { | |
16662 | machine_mode mode = GET_MODE (target); | |
16663 | machine_mode inner_mode = GET_MODE_INNER (mode); | |
16664 | machine_mode half_mode; | |
16665 | bool use_vec_merge = false; | |
7fc4d600 | 16666 | bool blendm_const = false; |
2bf6d935 | 16667 | rtx tmp; |
6910cad5 | 16668 | static rtx (*gen_extract[8][2]) (rtx, rtx) |
2bf6d935 ML |
16669 | = { |
16670 | { gen_vec_extract_lo_v32qi, gen_vec_extract_hi_v32qi }, | |
16671 | { gen_vec_extract_lo_v16hi, gen_vec_extract_hi_v16hi }, | |
16672 | { gen_vec_extract_lo_v8si, gen_vec_extract_hi_v8si }, | |
16673 | { gen_vec_extract_lo_v4di, gen_vec_extract_hi_v4di }, | |
16674 | { gen_vec_extract_lo_v8sf, gen_vec_extract_hi_v8sf }, | |
9e2a82e1 | 16675 | { gen_vec_extract_lo_v4df, gen_vec_extract_hi_v4df }, |
6910cad5 | 16676 | { gen_vec_extract_lo_v16hf, gen_vec_extract_hi_v16hf }, |
16677 | { gen_vec_extract_lo_v16bf, gen_vec_extract_hi_v16bf } | |
2bf6d935 | 16678 | }; |
6910cad5 | 16679 | static rtx (*gen_insert[8][2]) (rtx, rtx, rtx) |
2bf6d935 ML |
16680 | = { |
16681 | { gen_vec_set_lo_v32qi, gen_vec_set_hi_v32qi }, | |
16682 | { gen_vec_set_lo_v16hi, gen_vec_set_hi_v16hi }, | |
16683 | { gen_vec_set_lo_v8si, gen_vec_set_hi_v8si }, | |
16684 | { gen_vec_set_lo_v4di, gen_vec_set_hi_v4di }, | |
16685 | { gen_vec_set_lo_v8sf, gen_vec_set_hi_v8sf }, | |
9e2a82e1 | 16686 | { gen_vec_set_lo_v4df, gen_vec_set_hi_v4df }, |
16687 | { gen_vec_set_lo_v16hf, gen_vec_set_hi_v16hf }, | |
6910cad5 | 16688 | { gen_vec_set_lo_v16bf, gen_vec_set_hi_v16bf }, |
2bf6d935 ML |
16689 | }; |
16690 | int i, j, n; | |
16691 | machine_mode mmode = VOIDmode; | |
16692 | rtx (*gen_blendm) (rtx, rtx, rtx, rtx); | |
16693 | ||
16694 | switch (mode) | |
16695 | { | |
2bf6d935 | 16696 | case E_V2SImode: |
f15c7bd1 UB |
16697 | use_vec_merge = TARGET_MMX_WITH_SSE && TARGET_SSE4_1; |
16698 | if (use_vec_merge) | |
16699 | break; | |
16700 | /* FALLTHRU */ | |
16701 | ||
16702 | case E_V2SFmode: | |
2bf6d935 ML |
16703 | if (mmx_ok) |
16704 | { | |
16705 | tmp = gen_reg_rtx (GET_MODE_INNER (mode)); | |
16706 | ix86_expand_vector_extract (true, tmp, target, 1 - elt); | |
16707 | if (elt == 0) | |
16708 | tmp = gen_rtx_VEC_CONCAT (mode, val, tmp); | |
16709 | else | |
16710 | tmp = gen_rtx_VEC_CONCAT (mode, tmp, val); | |
16711 | emit_insn (gen_rtx_SET (target, tmp)); | |
16712 | return; | |
16713 | } | |
16714 | break; | |
16715 | ||
16716 | case E_V2DImode: | |
16717 | use_vec_merge = TARGET_SSE4_1 && TARGET_64BIT; | |
16718 | if (use_vec_merge) | |
16719 | break; | |
16720 | ||
16721 | tmp = gen_reg_rtx (GET_MODE_INNER (mode)); | |
16722 | ix86_expand_vector_extract (false, tmp, target, 1 - elt); | |
16723 | if (elt == 0) | |
16724 | tmp = gen_rtx_VEC_CONCAT (mode, val, tmp); | |
16725 | else | |
16726 | tmp = gen_rtx_VEC_CONCAT (mode, tmp, val); | |
16727 | emit_insn (gen_rtx_SET (target, tmp)); | |
16728 | return; | |
16729 | ||
16730 | case E_V2DFmode: | |
ac173024 L |
16731 | /* NB: For ELT == 0, use standard scalar operation patterns which |
16732 | preserve the rest of the vector for combiner: | |
16733 | ||
16734 | (vec_merge:V2DF | |
16735 | (vec_duplicate:V2DF (reg:DF)) | |
16736 | (reg:V2DF) | |
16737 | (const_int 1)) | |
16738 | */ | |
16739 | if (elt == 0) | |
16740 | goto do_vec_merge; | |
16741 | ||
2bf6d935 ML |
16742 | { |
16743 | rtx op0, op1; | |
16744 | ||
16745 | /* For the two element vectors, we implement a VEC_CONCAT with | |
16746 | the extraction of the other element. */ | |
16747 | ||
16748 | tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (1, GEN_INT (1 - elt))); | |
16749 | tmp = gen_rtx_VEC_SELECT (inner_mode, target, tmp); | |
16750 | ||
16751 | if (elt == 0) | |
16752 | op0 = val, op1 = tmp; | |
16753 | else | |
16754 | op0 = tmp, op1 = val; | |
16755 | ||
16756 | tmp = gen_rtx_VEC_CONCAT (mode, op0, op1); | |
16757 | emit_insn (gen_rtx_SET (target, tmp)); | |
16758 | } | |
16759 | return; | |
16760 | ||
16761 | case E_V4SFmode: | |
16762 | use_vec_merge = TARGET_SSE4_1; | |
16763 | if (use_vec_merge) | |
16764 | break; | |
16765 | ||
16766 | switch (elt) | |
16767 | { | |
16768 | case 0: | |
16769 | use_vec_merge = true; | |
16770 | break; | |
16771 | ||
16772 | case 1: | |
16773 | /* tmp = target = A B C D */ | |
16774 | tmp = copy_to_reg (target); | |
16775 | /* target = A A B B */ | |
16776 | emit_insn (gen_vec_interleave_lowv4sf (target, target, target)); | |
16777 | /* target = X A B B */ | |
16778 | ix86_expand_vector_set (false, target, val, 0); | |
16779 | /* target = A X C D */ | |
16780 | emit_insn (gen_sse_shufps_v4sf (target, target, tmp, | |
16781 | const1_rtx, const0_rtx, | |
16782 | GEN_INT (2+4), GEN_INT (3+4))); | |
16783 | return; | |
16784 | ||
16785 | case 2: | |
16786 | /* tmp = target = A B C D */ | |
16787 | tmp = copy_to_reg (target); | |
16788 | /* tmp = X B C D */ | |
16789 | ix86_expand_vector_set (false, tmp, val, 0); | |
16790 | /* target = A B X D */ | |
16791 | emit_insn (gen_sse_shufps_v4sf (target, target, tmp, | |
16792 | const0_rtx, const1_rtx, | |
16793 | GEN_INT (0+4), GEN_INT (3+4))); | |
16794 | return; | |
16795 | ||
16796 | case 3: | |
16797 | /* tmp = target = A B C D */ | |
16798 | tmp = copy_to_reg (target); | |
16799 | /* tmp = X B C D */ | |
16800 | ix86_expand_vector_set (false, tmp, val, 0); | |
16801 | /* target = A B X D */ | |
16802 | emit_insn (gen_sse_shufps_v4sf (target, target, tmp, | |
16803 | const0_rtx, const1_rtx, | |
16804 | GEN_INT (2+4), GEN_INT (0+4))); | |
16805 | return; | |
16806 | ||
16807 | default: | |
16808 | gcc_unreachable (); | |
16809 | } | |
16810 | break; | |
16811 | ||
16812 | case E_V4SImode: | |
16813 | use_vec_merge = TARGET_SSE4_1; | |
16814 | if (use_vec_merge) | |
16815 | break; | |
16816 | ||
16817 | /* Element 0 handled by vec_merge below. */ | |
16818 | if (elt == 0) | |
16819 | { | |
16820 | use_vec_merge = true; | |
16821 | break; | |
16822 | } | |
16823 | ||
16824 | if (TARGET_SSE2) | |
16825 | { | |
16826 | /* With SSE2, use integer shuffles to swap element 0 and ELT, | |
16827 | store into element 0, then shuffle them back. */ | |
16828 | ||
16829 | rtx order[4]; | |
16830 | ||
16831 | order[0] = GEN_INT (elt); | |
16832 | order[1] = const1_rtx; | |
16833 | order[2] = const2_rtx; | |
16834 | order[3] = GEN_INT (3); | |
16835 | order[elt] = const0_rtx; | |
16836 | ||
16837 | emit_insn (gen_sse2_pshufd_1 (target, target, order[0], | |
16838 | order[1], order[2], order[3])); | |
16839 | ||
16840 | ix86_expand_vector_set (false, target, val, 0); | |
16841 | ||
16842 | emit_insn (gen_sse2_pshufd_1 (target, target, order[0], | |
16843 | order[1], order[2], order[3])); | |
16844 | } | |
16845 | else | |
16846 | { | |
16847 | /* For SSE1, we have to reuse the V4SF code. */ | |
16848 | rtx t = gen_reg_rtx (V4SFmode); | |
16849 | emit_move_insn (t, gen_lowpart (V4SFmode, target)); | |
16850 | ix86_expand_vector_set (false, t, gen_lowpart (SFmode, val), elt); | |
16851 | emit_move_insn (target, gen_lowpart (mode, t)); | |
16852 | } | |
16853 | return; | |
16854 | ||
16855 | case E_V8HImode: | |
7eb961d8 | 16856 | case E_V8HFmode: |
6910cad5 | 16857 | case E_V8BFmode: |
5883e567 | 16858 | case E_V2HImode: |
2bf6d935 ML |
16859 | use_vec_merge = TARGET_SSE2; |
16860 | break; | |
16861 | case E_V4HImode: | |
16862 | use_vec_merge = mmx_ok && (TARGET_SSE || TARGET_3DNOW_A); | |
16863 | break; | |
16864 | ||
16865 | case E_V16QImode: | |
5883e567 | 16866 | case E_V4QImode: |
2bf6d935 ML |
16867 | use_vec_merge = TARGET_SSE4_1; |
16868 | break; | |
16869 | ||
16870 | case E_V8QImode: | |
f15c7bd1 | 16871 | use_vec_merge = TARGET_MMX_WITH_SSE && TARGET_SSE4_1; |
2bf6d935 ML |
16872 | break; |
16873 | ||
16874 | case E_V32QImode: | |
16875 | half_mode = V16QImode; | |
16876 | j = 0; | |
16877 | n = 16; | |
16878 | goto half; | |
16879 | ||
9e2a82e1 | 16880 | case E_V16HFmode: |
6910cad5 | 16881 | case E_V16BFmode: |
1f759dbd | 16882 | /* For ELT == 0, vec_setv8hf_0 can save 1 vpbroadcastw. */ |
16883 | if (TARGET_AVX2 && elt != 0) | |
7fc4d600 | 16884 | { |
16885 | mmode = SImode; | |
6910cad5 | 16886 | gen_blendm = ((mode == E_V16HFmode) ? gen_avx2_pblendph_1 |
16887 | : gen_avx2_pblendbf_1); | |
7fc4d600 | 16888 | blendm_const = true; |
16889 | break; | |
16890 | } | |
16891 | else | |
16892 | { | |
6910cad5 | 16893 | half_mode = ((mode == E_V16HFmode) ? V8HFmode : V8BFmode); |
16894 | j = ((mode == E_V16HFmode) ? 6 : 7); | |
7fc4d600 | 16895 | n = 8; |
16896 | goto half; | |
16897 | } | |
9e2a82e1 | 16898 | |
2bf6d935 ML |
16899 | case E_V16HImode: |
16900 | half_mode = V8HImode; | |
16901 | j = 1; | |
16902 | n = 8; | |
16903 | goto half; | |
16904 | ||
16905 | case E_V8SImode: | |
16906 | half_mode = V4SImode; | |
16907 | j = 2; | |
16908 | n = 4; | |
16909 | goto half; | |
16910 | ||
16911 | case E_V4DImode: | |
16912 | half_mode = V2DImode; | |
16913 | j = 3; | |
16914 | n = 2; | |
16915 | goto half; | |
16916 | ||
16917 | case E_V8SFmode: | |
16918 | half_mode = V4SFmode; | |
16919 | j = 4; | |
16920 | n = 4; | |
16921 | goto half; | |
16922 | ||
16923 | case E_V4DFmode: | |
16924 | half_mode = V2DFmode; | |
16925 | j = 5; | |
16926 | n = 2; | |
16927 | goto half; | |
16928 | ||
16929 | half: | |
16930 | /* Compute offset. */ | |
16931 | i = elt / n; | |
16932 | elt %= n; | |
16933 | ||
16934 | gcc_assert (i <= 1); | |
16935 | ||
16936 | /* Extract the half. */ | |
16937 | tmp = gen_reg_rtx (half_mode); | |
16938 | emit_insn (gen_extract[j][i] (tmp, target)); | |
16939 | ||
16940 | /* Put val in tmp at elt. */ | |
16941 | ix86_expand_vector_set (false, tmp, val, elt); | |
16942 | ||
16943 | /* Put it back. */ | |
16944 | emit_insn (gen_insert[j][i] (target, target, tmp)); | |
16945 | return; | |
16946 | ||
16947 | case E_V8DFmode: | |
16948 | if (TARGET_AVX512F) | |
16949 | { | |
16950 | mmode = QImode; | |
16951 | gen_blendm = gen_avx512f_blendmv8df; | |
16952 | } | |
16953 | break; | |
16954 | ||
16955 | case E_V8DImode: | |
16956 | if (TARGET_AVX512F) | |
16957 | { | |
16958 | mmode = QImode; | |
16959 | gen_blendm = gen_avx512f_blendmv8di; | |
16960 | } | |
16961 | break; | |
16962 | ||
16963 | case E_V16SFmode: | |
16964 | if (TARGET_AVX512F) | |
16965 | { | |
16966 | mmode = HImode; | |
16967 | gen_blendm = gen_avx512f_blendmv16sf; | |
16968 | } | |
16969 | break; | |
16970 | ||
16971 | case E_V16SImode: | |
16972 | if (TARGET_AVX512F) | |
16973 | { | |
16974 | mmode = HImode; | |
16975 | gen_blendm = gen_avx512f_blendmv16si; | |
16976 | } | |
16977 | break; | |
16978 | ||
9e2a82e1 | 16979 | case E_V32HFmode: |
16980 | if (TARGET_AVX512BW) | |
16981 | { | |
16982 | mmode = SImode; | |
16983 | gen_blendm = gen_avx512bw_blendmv32hf; | |
16984 | } | |
16985 | break; | |
6910cad5 | 16986 | case E_V32BFmode: |
16987 | if (TARGET_AVX512BW) | |
16988 | { | |
16989 | mmode = SImode; | |
16990 | gen_blendm = gen_avx512bw_blendmv32bf; | |
16991 | } | |
16992 | break; | |
2bf6d935 ML |
16993 | case E_V32HImode: |
16994 | if (TARGET_AVX512BW) | |
16995 | { | |
16996 | mmode = SImode; | |
16997 | gen_blendm = gen_avx512bw_blendmv32hi; | |
16998 | } | |
16999 | else if (TARGET_AVX512F) | |
17000 | { | |
17001 | half_mode = E_V8HImode; | |
17002 | n = 8; | |
17003 | goto quarter; | |
17004 | } | |
17005 | break; | |
17006 | ||
17007 | case E_V64QImode: | |
17008 | if (TARGET_AVX512BW) | |
17009 | { | |
17010 | mmode = DImode; | |
17011 | gen_blendm = gen_avx512bw_blendmv64qi; | |
17012 | } | |
17013 | else if (TARGET_AVX512F) | |
17014 | { | |
17015 | half_mode = E_V16QImode; | |
17016 | n = 16; | |
17017 | goto quarter; | |
17018 | } | |
17019 | break; | |
17020 | ||
17021 | quarter: | |
17022 | /* Compute offset. */ | |
17023 | i = elt / n; | |
17024 | elt %= n; | |
17025 | ||
17026 | gcc_assert (i <= 3); | |
17027 | ||
17028 | { | |
17029 | /* Extract the quarter. */ | |
17030 | tmp = gen_reg_rtx (V4SImode); | |
17031 | rtx tmp2 = gen_lowpart (V16SImode, target); | |
17032 | rtx mask = gen_reg_rtx (QImode); | |
17033 | ||
17034 | emit_move_insn (mask, constm1_rtx); | |
17035 | emit_insn (gen_avx512f_vextracti32x4_mask (tmp, tmp2, GEN_INT (i), | |
17036 | tmp, mask)); | |
17037 | ||
17038 | tmp2 = gen_reg_rtx (half_mode); | |
17039 | emit_move_insn (tmp2, gen_lowpart (half_mode, tmp)); | |
17040 | tmp = tmp2; | |
17041 | ||
17042 | /* Put val in tmp at elt. */ | |
17043 | ix86_expand_vector_set (false, tmp, val, elt); | |
17044 | ||
17045 | /* Put it back. */ | |
17046 | tmp2 = gen_reg_rtx (V16SImode); | |
17047 | rtx tmp3 = gen_lowpart (V16SImode, target); | |
17048 | mask = gen_reg_rtx (HImode); | |
17049 | emit_move_insn (mask, constm1_rtx); | |
17050 | tmp = gen_lowpart (V4SImode, tmp); | |
17051 | emit_insn (gen_avx512f_vinserti32x4_mask (tmp2, tmp3, tmp, GEN_INT (i), | |
17052 | tmp3, mask)); | |
17053 | emit_move_insn (target, gen_lowpart (mode, tmp2)); | |
17054 | } | |
17055 | return; | |
17056 | ||
17057 | default: | |
17058 | break; | |
17059 | } | |
17060 | ||
17061 | if (mmode != VOIDmode) | |
17062 | { | |
17063 | tmp = gen_reg_rtx (mode); | |
17064 | emit_insn (gen_rtx_SET (tmp, gen_rtx_VEC_DUPLICATE (mode, val))); | |
7fc4d600 | 17065 | rtx merge_mask = gen_int_mode (HOST_WIDE_INT_1U << elt, mmode); |
2bf6d935 ML |
17066 | /* The avx512*_blendm<mode> expanders have different operand order |
17067 | from VEC_MERGE. In VEC_MERGE, the first input operand is used for | |
17068 | elements where the mask is set and second input operand otherwise, | |
17069 | in {sse,avx}*_*blend* the first input operand is used for elements | |
17070 | where the mask is clear and second input operand otherwise. */ | |
7fc4d600 | 17071 | if (!blendm_const) |
17072 | merge_mask = force_reg (mmode, merge_mask); | |
17073 | emit_insn (gen_blendm (target, target, tmp, merge_mask)); | |
2bf6d935 ML |
17074 | } |
17075 | else if (use_vec_merge) | |
17076 | { | |
ac173024 | 17077 | do_vec_merge: |
2bf6d935 ML |
17078 | tmp = gen_rtx_VEC_DUPLICATE (mode, val); |
17079 | tmp = gen_rtx_VEC_MERGE (mode, tmp, target, | |
17080 | GEN_INT (HOST_WIDE_INT_1U << elt)); | |
17081 | emit_insn (gen_rtx_SET (target, tmp)); | |
17082 | } | |
17083 | else | |
17084 | { | |
17085 | rtx mem = assign_stack_temp (mode, GET_MODE_SIZE (mode)); | |
17086 | ||
17087 | emit_move_insn (mem, target); | |
17088 | ||
17089 | tmp = adjust_address (mem, inner_mode, elt * GET_MODE_SIZE (inner_mode)); | |
17090 | emit_move_insn (tmp, val); | |
17091 | ||
17092 | emit_move_insn (target, mem); | |
17093 | } | |
17094 | } | |
17095 | ||
17096 | void | |
17097 | ix86_expand_vector_extract (bool mmx_ok, rtx target, rtx vec, int elt) | |
17098 | { | |
17099 | machine_mode mode = GET_MODE (vec); | |
17100 | machine_mode inner_mode = GET_MODE_INNER (mode); | |
17101 | bool use_vec_extr = false; | |
17102 | rtx tmp; | |
17103 | ||
17104 | switch (mode) | |
17105 | { | |
17106 | case E_V2SImode: | |
5fbc8ab4 UB |
17107 | use_vec_extr = TARGET_MMX_WITH_SSE && TARGET_SSE4_1; |
17108 | if (use_vec_extr) | |
17109 | break; | |
17110 | /* FALLTHRU */ | |
17111 | ||
2bf6d935 ML |
17112 | case E_V2SFmode: |
17113 | if (!mmx_ok) | |
17114 | break; | |
17115 | /* FALLTHRU */ | |
17116 | ||
17117 | case E_V2DFmode: | |
17118 | case E_V2DImode: | |
17119 | case E_V2TImode: | |
17120 | case E_V4TImode: | |
17121 | use_vec_extr = true; | |
17122 | break; | |
17123 | ||
17124 | case E_V4SFmode: | |
17125 | use_vec_extr = TARGET_SSE4_1; | |
17126 | if (use_vec_extr) | |
17127 | break; | |
17128 | ||
17129 | switch (elt) | |
17130 | { | |
17131 | case 0: | |
17132 | tmp = vec; | |
17133 | break; | |
17134 | ||
17135 | case 1: | |
17136 | case 3: | |
17137 | tmp = gen_reg_rtx (mode); | |
17138 | emit_insn (gen_sse_shufps_v4sf (tmp, vec, vec, | |
17139 | GEN_INT (elt), GEN_INT (elt), | |
17140 | GEN_INT (elt+4), GEN_INT (elt+4))); | |
17141 | break; | |
17142 | ||
17143 | case 2: | |
17144 | tmp = gen_reg_rtx (mode); | |
17145 | emit_insn (gen_vec_interleave_highv4sf (tmp, vec, vec)); | |
17146 | break; | |
17147 | ||
17148 | default: | |
17149 | gcc_unreachable (); | |
17150 | } | |
17151 | vec = tmp; | |
17152 | use_vec_extr = true; | |
17153 | elt = 0; | |
17154 | break; | |
17155 | ||
17156 | case E_V4SImode: | |
17157 | use_vec_extr = TARGET_SSE4_1; | |
17158 | if (use_vec_extr) | |
17159 | break; | |
17160 | ||
17161 | if (TARGET_SSE2) | |
17162 | { | |
17163 | switch (elt) | |
17164 | { | |
17165 | case 0: | |
17166 | tmp = vec; | |
17167 | break; | |
17168 | ||
17169 | case 1: | |
17170 | case 3: | |
17171 | tmp = gen_reg_rtx (mode); | |
17172 | emit_insn (gen_sse2_pshufd_1 (tmp, vec, | |
17173 | GEN_INT (elt), GEN_INT (elt), | |
17174 | GEN_INT (elt), GEN_INT (elt))); | |
17175 | break; | |
17176 | ||
17177 | case 2: | |
17178 | tmp = gen_reg_rtx (mode); | |
17179 | emit_insn (gen_vec_interleave_highv4si (tmp, vec, vec)); | |
17180 | break; | |
17181 | ||
17182 | default: | |
17183 | gcc_unreachable (); | |
17184 | } | |
17185 | vec = tmp; | |
17186 | use_vec_extr = true; | |
17187 | elt = 0; | |
17188 | } | |
17189 | else | |
17190 | { | |
17191 | /* For SSE1, we have to reuse the V4SF code. */ | |
17192 | ix86_expand_vector_extract (false, gen_lowpart (SFmode, target), | |
17193 | gen_lowpart (V4SFmode, vec), elt); | |
17194 | return; | |
17195 | } | |
17196 | break; | |
17197 | ||
17198 | case E_V8HImode: | |
7a54d3de | 17199 | case E_V8HFmode: |
6910cad5 | 17200 | case E_V8BFmode: |
5883e567 | 17201 | case E_V2HImode: |
2bf6d935 ML |
17202 | use_vec_extr = TARGET_SSE2; |
17203 | break; | |
17204 | case E_V4HImode: | |
17205 | use_vec_extr = mmx_ok && (TARGET_SSE || TARGET_3DNOW_A); | |
17206 | break; | |
17207 | ||
17208 | case E_V16QImode: | |
17209 | use_vec_extr = TARGET_SSE4_1; | |
f66e6e2b JJ |
17210 | if (!use_vec_extr |
17211 | && TARGET_SSE2 | |
17212 | && elt == 0 | |
17213 | && (optimize_insn_for_size_p () || TARGET_INTER_UNIT_MOVES_FROM_VEC)) | |
17214 | { | |
17215 | tmp = gen_reg_rtx (SImode); | |
17216 | ix86_expand_vector_extract (false, tmp, gen_lowpart (V4SImode, vec), | |
17217 | 0); | |
17218 | emit_insn (gen_rtx_SET (target, gen_lowpart (QImode, tmp))); | |
17219 | return; | |
17220 | } | |
2bf6d935 | 17221 | break; |
5883e567 UB |
17222 | case E_V4QImode: |
17223 | use_vec_extr = TARGET_SSE4_1; | |
17224 | break; | |
2bf6d935 ML |
17225 | |
17226 | case E_V8SFmode: | |
17227 | if (TARGET_AVX) | |
17228 | { | |
17229 | tmp = gen_reg_rtx (V4SFmode); | |
17230 | if (elt < 4) | |
17231 | emit_insn (gen_vec_extract_lo_v8sf (tmp, vec)); | |
17232 | else | |
17233 | emit_insn (gen_vec_extract_hi_v8sf (tmp, vec)); | |
17234 | ix86_expand_vector_extract (false, target, tmp, elt & 3); | |
17235 | return; | |
17236 | } | |
17237 | break; | |
17238 | ||
17239 | case E_V4DFmode: | |
17240 | if (TARGET_AVX) | |
17241 | { | |
17242 | tmp = gen_reg_rtx (V2DFmode); | |
17243 | if (elt < 2) | |
17244 | emit_insn (gen_vec_extract_lo_v4df (tmp, vec)); | |
17245 | else | |
17246 | emit_insn (gen_vec_extract_hi_v4df (tmp, vec)); | |
17247 | ix86_expand_vector_extract (false, target, tmp, elt & 1); | |
17248 | return; | |
17249 | } | |
17250 | break; | |
17251 | ||
17252 | case E_V32QImode: | |
17253 | if (TARGET_AVX) | |
17254 | { | |
17255 | tmp = gen_reg_rtx (V16QImode); | |
17256 | if (elt < 16) | |
17257 | emit_insn (gen_vec_extract_lo_v32qi (tmp, vec)); | |
17258 | else | |
17259 | emit_insn (gen_vec_extract_hi_v32qi (tmp, vec)); | |
17260 | ix86_expand_vector_extract (false, target, tmp, elt & 15); | |
17261 | return; | |
17262 | } | |
17263 | break; | |
17264 | ||
17265 | case E_V16HImode: | |
17266 | if (TARGET_AVX) | |
17267 | { | |
17268 | tmp = gen_reg_rtx (V8HImode); | |
17269 | if (elt < 8) | |
17270 | emit_insn (gen_vec_extract_lo_v16hi (tmp, vec)); | |
17271 | else | |
17272 | emit_insn (gen_vec_extract_hi_v16hi (tmp, vec)); | |
17273 | ix86_expand_vector_extract (false, target, tmp, elt & 7); | |
17274 | return; | |
17275 | } | |
17276 | break; | |
17277 | ||
17278 | case E_V8SImode: | |
17279 | if (TARGET_AVX) | |
17280 | { | |
17281 | tmp = gen_reg_rtx (V4SImode); | |
17282 | if (elt < 4) | |
17283 | emit_insn (gen_vec_extract_lo_v8si (tmp, vec)); | |
17284 | else | |
17285 | emit_insn (gen_vec_extract_hi_v8si (tmp, vec)); | |
17286 | ix86_expand_vector_extract (false, target, tmp, elt & 3); | |
17287 | return; | |
17288 | } | |
17289 | break; | |
17290 | ||
17291 | case E_V4DImode: | |
17292 | if (TARGET_AVX) | |
17293 | { | |
17294 | tmp = gen_reg_rtx (V2DImode); | |
17295 | if (elt < 2) | |
17296 | emit_insn (gen_vec_extract_lo_v4di (tmp, vec)); | |
17297 | else | |
17298 | emit_insn (gen_vec_extract_hi_v4di (tmp, vec)); | |
17299 | ix86_expand_vector_extract (false, target, tmp, elt & 1); | |
17300 | return; | |
17301 | } | |
17302 | break; | |
17303 | ||
17304 | case E_V32HImode: | |
17305 | if (TARGET_AVX512BW) | |
17306 | { | |
17307 | tmp = gen_reg_rtx (V16HImode); | |
17308 | if (elt < 16) | |
17309 | emit_insn (gen_vec_extract_lo_v32hi (tmp, vec)); | |
17310 | else | |
17311 | emit_insn (gen_vec_extract_hi_v32hi (tmp, vec)); | |
17312 | ix86_expand_vector_extract (false, target, tmp, elt & 15); | |
17313 | return; | |
17314 | } | |
17315 | break; | |
17316 | ||
17317 | case E_V64QImode: | |
17318 | if (TARGET_AVX512BW) | |
17319 | { | |
17320 | tmp = gen_reg_rtx (V32QImode); | |
17321 | if (elt < 32) | |
17322 | emit_insn (gen_vec_extract_lo_v64qi (tmp, vec)); | |
17323 | else | |
17324 | emit_insn (gen_vec_extract_hi_v64qi (tmp, vec)); | |
17325 | ix86_expand_vector_extract (false, target, tmp, elt & 31); | |
17326 | return; | |
17327 | } | |
17328 | break; | |
17329 | ||
17330 | case E_V16SFmode: | |
17331 | tmp = gen_reg_rtx (V8SFmode); | |
17332 | if (elt < 8) | |
17333 | emit_insn (gen_vec_extract_lo_v16sf (tmp, vec)); | |
17334 | else | |
17335 | emit_insn (gen_vec_extract_hi_v16sf (tmp, vec)); | |
17336 | ix86_expand_vector_extract (false, target, tmp, elt & 7); | |
17337 | return; | |
17338 | ||
17339 | case E_V8DFmode: | |
17340 | tmp = gen_reg_rtx (V4DFmode); | |
17341 | if (elt < 4) | |
17342 | emit_insn (gen_vec_extract_lo_v8df (tmp, vec)); | |
17343 | else | |
17344 | emit_insn (gen_vec_extract_hi_v8df (tmp, vec)); | |
17345 | ix86_expand_vector_extract (false, target, tmp, elt & 3); | |
17346 | return; | |
17347 | ||
17348 | case E_V16SImode: | |
17349 | tmp = gen_reg_rtx (V8SImode); | |
17350 | if (elt < 8) | |
17351 | emit_insn (gen_vec_extract_lo_v16si (tmp, vec)); | |
17352 | else | |
17353 | emit_insn (gen_vec_extract_hi_v16si (tmp, vec)); | |
17354 | ix86_expand_vector_extract (false, target, tmp, elt & 7); | |
17355 | return; | |
17356 | ||
17357 | case E_V8DImode: | |
17358 | tmp = gen_reg_rtx (V4DImode); | |
17359 | if (elt < 4) | |
17360 | emit_insn (gen_vec_extract_lo_v8di (tmp, vec)); | |
17361 | else | |
17362 | emit_insn (gen_vec_extract_hi_v8di (tmp, vec)); | |
17363 | ix86_expand_vector_extract (false, target, tmp, elt & 3); | |
17364 | return; | |
17365 | ||
9e2a82e1 | 17366 | case E_V32HFmode: |
6910cad5 | 17367 | case E_V32BFmode: |
7a54d3de UB |
17368 | if (TARGET_AVX512BW) |
17369 | { | |
6910cad5 | 17370 | tmp = (mode == E_V32HFmode |
17371 | ? gen_reg_rtx (V16HFmode) | |
17372 | : gen_reg_rtx (V16BFmode)); | |
7a54d3de | 17373 | if (elt < 16) |
6910cad5 | 17374 | emit_insn (maybe_gen_vec_extract_lo (mode, tmp, vec)); |
7a54d3de | 17375 | else |
6910cad5 | 17376 | emit_insn (maybe_gen_vec_extract_hi (mode, tmp, vec)); |
7a54d3de UB |
17377 | ix86_expand_vector_extract (false, target, tmp, elt & 15); |
17378 | return; | |
17379 | } | |
17380 | break; | |
9e2a82e1 | 17381 | |
17382 | case E_V16HFmode: | |
6910cad5 | 17383 | case E_V16BFmode: |
7a54d3de UB |
17384 | if (TARGET_AVX) |
17385 | { | |
6910cad5 | 17386 | tmp = (mode == E_V16HFmode |
17387 | ? gen_reg_rtx (V8HFmode) | |
17388 | : gen_reg_rtx (V8BFmode)); | |
7a54d3de | 17389 | if (elt < 8) |
6910cad5 | 17390 | emit_insn (maybe_gen_vec_extract_lo (mode, tmp, vec)); |
7a54d3de | 17391 | else |
6910cad5 | 17392 | emit_insn (maybe_gen_vec_extract_hi (mode, tmp, vec)); |
7a54d3de UB |
17393 | ix86_expand_vector_extract (false, target, tmp, elt & 7); |
17394 | return; | |
17395 | } | |
9e2a82e1 | 17396 | break; |
17397 | ||
2bf6d935 | 17398 | case E_V8QImode: |
5fbc8ab4 | 17399 | use_vec_extr = TARGET_MMX_WITH_SSE && TARGET_SSE4_1; |
2bf6d935 | 17400 | /* ??? Could extract the appropriate HImode element and shift. */ |
5fbc8ab4 UB |
17401 | break; |
17402 | ||
2bf6d935 ML |
17403 | default: |
17404 | break; | |
17405 | } | |
17406 | ||
17407 | if (use_vec_extr) | |
17408 | { | |
17409 | tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (1, GEN_INT (elt))); | |
17410 | tmp = gen_rtx_VEC_SELECT (inner_mode, vec, tmp); | |
17411 | ||
17412 | /* Let the rtl optimizers know about the zero extension performed. */ | |
17413 | if (inner_mode == QImode || inner_mode == HImode) | |
17414 | { | |
97c32001 | 17415 | rtx reg = gen_reg_rtx (SImode); |
2bf6d935 | 17416 | tmp = gen_rtx_ZERO_EXTEND (SImode, tmp); |
97c32001 RS |
17417 | emit_move_insn (reg, tmp); |
17418 | tmp = gen_lowpart (inner_mode, reg); | |
17419 | SUBREG_PROMOTED_VAR_P (tmp) = 1; | |
17420 | SUBREG_PROMOTED_SET (tmp, 1); | |
2bf6d935 ML |
17421 | } |
17422 | ||
97c32001 | 17423 | emit_move_insn (target, tmp); |
2bf6d935 ML |
17424 | } |
17425 | else | |
17426 | { | |
17427 | rtx mem = assign_stack_temp (mode, GET_MODE_SIZE (mode)); | |
17428 | ||
17429 | emit_move_insn (mem, vec); | |
17430 | ||
17431 | tmp = adjust_address (mem, inner_mode, elt*GET_MODE_SIZE (inner_mode)); | |
17432 | emit_move_insn (target, tmp); | |
17433 | } | |
17434 | } | |
17435 | ||
17436 | /* Generate code to copy vector bits i / 2 ... i - 1 from vector SRC | |
17437 | to bits 0 ... i / 2 - 1 of vector DEST, which has the same mode. | |
17438 | The upper bits of DEST are undefined, though they shouldn't cause | |
17439 | exceptions (some bits from src or all zeros are ok). */ | |
17440 | ||
17441 | static void | |
17442 | emit_reduc_half (rtx dest, rtx src, int i) | |
17443 | { | |
17444 | rtx tem, d = dest; | |
17445 | switch (GET_MODE (src)) | |
17446 | { | |
17447 | case E_V4SFmode: | |
17448 | if (i == 128) | |
17449 | tem = gen_sse_movhlps (dest, src, src); | |
17450 | else | |
17451 | tem = gen_sse_shufps_v4sf (dest, src, src, const1_rtx, const1_rtx, | |
17452 | GEN_INT (1 + 4), GEN_INT (1 + 4)); | |
17453 | break; | |
17454 | case E_V2DFmode: | |
17455 | tem = gen_vec_interleave_highv2df (dest, src, src); | |
17456 | break; | |
73c535a0 | 17457 | case E_V4QImode: |
17458 | d = gen_reg_rtx (V1SImode); | |
17459 | tem = gen_mmx_lshrv1si3 (d, gen_lowpart (V1SImode, src), | |
17460 | GEN_INT (i / 2)); | |
17461 | break; | |
77ca2cfc | 17462 | case E_V4HImode: |
17463 | d = gen_reg_rtx (V1DImode); | |
17464 | tem = gen_mmx_lshrv1di3 (d, gen_lowpart (V1DImode, src), | |
17465 | GEN_INT (i / 2)); | |
17466 | break; | |
2bf6d935 ML |
17467 | case E_V16QImode: |
17468 | case E_V8HImode: | |
3540429b | 17469 | case E_V8HFmode: |
2bf6d935 ML |
17470 | case E_V4SImode: |
17471 | case E_V2DImode: | |
17472 | d = gen_reg_rtx (V1TImode); | |
17473 | tem = gen_sse2_lshrv1ti3 (d, gen_lowpart (V1TImode, src), | |
17474 | GEN_INT (i / 2)); | |
17475 | break; | |
17476 | case E_V8SFmode: | |
17477 | if (i == 256) | |
17478 | tem = gen_avx_vperm2f128v8sf3 (dest, src, src, const1_rtx); | |
17479 | else | |
17480 | tem = gen_avx_shufps256 (dest, src, src, | |
17481 | GEN_INT (i == 128 ? 2 + (3 << 2) : 1)); | |
17482 | break; | |
17483 | case E_V4DFmode: | |
17484 | if (i == 256) | |
17485 | tem = gen_avx_vperm2f128v4df3 (dest, src, src, const1_rtx); | |
17486 | else | |
17487 | tem = gen_avx_shufpd256 (dest, src, src, const1_rtx); | |
17488 | break; | |
17489 | case E_V32QImode: | |
17490 | case E_V16HImode: | |
3540429b | 17491 | case E_V16HFmode: |
2bf6d935 ML |
17492 | case E_V8SImode: |
17493 | case E_V4DImode: | |
17494 | if (i == 256) | |
17495 | { | |
17496 | if (GET_MODE (dest) != V4DImode) | |
17497 | d = gen_reg_rtx (V4DImode); | |
17498 | tem = gen_avx2_permv2ti (d, gen_lowpart (V4DImode, src), | |
17499 | gen_lowpart (V4DImode, src), | |
17500 | const1_rtx); | |
17501 | } | |
17502 | else | |
17503 | { | |
17504 | d = gen_reg_rtx (V2TImode); | |
17505 | tem = gen_avx2_lshrv2ti3 (d, gen_lowpart (V2TImode, src), | |
17506 | GEN_INT (i / 2)); | |
17507 | } | |
17508 | break; | |
17509 | case E_V64QImode: | |
17510 | case E_V32HImode: | |
3540429b | 17511 | case E_V32HFmode: |
bee27152 JJ |
17512 | if (i < 64) |
17513 | { | |
17514 | d = gen_reg_rtx (V4TImode); | |
17515 | tem = gen_avx512bw_lshrv4ti3 (d, gen_lowpart (V4TImode, src), | |
17516 | GEN_INT (i / 2)); | |
17517 | break; | |
17518 | } | |
17519 | /* FALLTHRU */ | |
2bf6d935 ML |
17520 | case E_V16SImode: |
17521 | case E_V16SFmode: | |
17522 | case E_V8DImode: | |
17523 | case E_V8DFmode: | |
17524 | if (i > 128) | |
17525 | tem = gen_avx512f_shuf_i32x4_1 (gen_lowpart (V16SImode, dest), | |
bee27152 JJ |
17526 | gen_lowpart (V16SImode, src), |
17527 | gen_lowpart (V16SImode, src), | |
17528 | GEN_INT (0x4 + (i == 512 ? 4 : 0)), | |
17529 | GEN_INT (0x5 + (i == 512 ? 4 : 0)), | |
17530 | GEN_INT (0x6 + (i == 512 ? 4 : 0)), | |
17531 | GEN_INT (0x7 + (i == 512 ? 4 : 0)), | |
17532 | GEN_INT (0xC), GEN_INT (0xD), | |
17533 | GEN_INT (0xE), GEN_INT (0xF), | |
17534 | GEN_INT (0x10), GEN_INT (0x11), | |
17535 | GEN_INT (0x12), GEN_INT (0x13), | |
17536 | GEN_INT (0x14), GEN_INT (0x15), | |
17537 | GEN_INT (0x16), GEN_INT (0x17)); | |
2bf6d935 ML |
17538 | else |
17539 | tem = gen_avx512f_pshufd_1 (gen_lowpart (V16SImode, dest), | |
bee27152 JJ |
17540 | gen_lowpart (V16SImode, src), |
17541 | GEN_INT (i == 128 ? 0x2 : 0x1), | |
17542 | GEN_INT (0x3), | |
17543 | GEN_INT (0x3), | |
17544 | GEN_INT (0x3), | |
17545 | GEN_INT (i == 128 ? 0x6 : 0x5), | |
17546 | GEN_INT (0x7), | |
17547 | GEN_INT (0x7), | |
17548 | GEN_INT (0x7), | |
17549 | GEN_INT (i == 128 ? 0xA : 0x9), | |
17550 | GEN_INT (0xB), | |
17551 | GEN_INT (0xB), | |
17552 | GEN_INT (0xB), | |
17553 | GEN_INT (i == 128 ? 0xE : 0xD), | |
17554 | GEN_INT (0xF), | |
17555 | GEN_INT (0xF), | |
17556 | GEN_INT (0xF)); | |
2bf6d935 ML |
17557 | break; |
17558 | default: | |
17559 | gcc_unreachable (); | |
17560 | } | |
17561 | emit_insn (tem); | |
17562 | if (d != dest) | |
17563 | emit_move_insn (dest, gen_lowpart (GET_MODE (dest), d)); | |
17564 | } | |
17565 | ||
17566 | /* Expand a vector reduction. FN is the binary pattern to reduce; | |
17567 | DEST is the destination; IN is the input vector. */ | |
17568 | ||
17569 | void | |
17570 | ix86_expand_reduc (rtx (*fn) (rtx, rtx, rtx), rtx dest, rtx in) | |
17571 | { | |
17572 | rtx half, dst, vec = in; | |
17573 | machine_mode mode = GET_MODE (in); | |
17574 | int i; | |
17575 | ||
17576 | /* SSE4 has a special instruction for V8HImode UMIN reduction. */ | |
17577 | if (TARGET_SSE4_1 | |
17578 | && mode == V8HImode | |
17579 | && fn == gen_uminv8hi3) | |
17580 | { | |
17581 | emit_insn (gen_sse4_1_phminposuw (dest, in)); | |
17582 | return; | |
17583 | } | |
17584 | ||
17585 | for (i = GET_MODE_BITSIZE (mode); | |
17586 | i > GET_MODE_UNIT_BITSIZE (mode); | |
17587 | i >>= 1) | |
17588 | { | |
17589 | half = gen_reg_rtx (mode); | |
17590 | emit_reduc_half (half, vec, i); | |
17591 | if (i == GET_MODE_UNIT_BITSIZE (mode) * 2) | |
17592 | dst = dest; | |
17593 | else | |
17594 | dst = gen_reg_rtx (mode); | |
17595 | emit_insn (fn (dst, half, vec)); | |
17596 | vec = dst; | |
17597 | } | |
17598 | } | |
17599 | ||
17600 | /* Output code to perform a conditional jump to LABEL, if C2 flag in | |
17601 | FP status register is set. */ | |
17602 | ||
17603 | void | |
17604 | ix86_emit_fp_unordered_jump (rtx label) | |
17605 | { | |
17606 | rtx reg = gen_reg_rtx (HImode); | |
17607 | rtx_insn *insn; | |
17608 | rtx temp; | |
17609 | ||
17610 | emit_insn (gen_x86_fnstsw_1 (reg)); | |
17611 | ||
17612 | if (TARGET_SAHF && (TARGET_USE_SAHF || optimize_insn_for_size_p ())) | |
17613 | { | |
17614 | emit_insn (gen_x86_sahf_1 (reg)); | |
17615 | ||
17616 | temp = gen_rtx_REG (CCmode, FLAGS_REG); | |
17617 | temp = gen_rtx_UNORDERED (VOIDmode, temp, const0_rtx); | |
17618 | } | |
17619 | else | |
17620 | { | |
17621 | emit_insn (gen_testqi_ext_1_ccno (reg, GEN_INT (0x04))); | |
17622 | ||
17623 | temp = gen_rtx_REG (CCNOmode, FLAGS_REG); | |
17624 | temp = gen_rtx_NE (VOIDmode, temp, const0_rtx); | |
17625 | } | |
17626 | ||
17627 | temp = gen_rtx_IF_THEN_ELSE (VOIDmode, temp, | |
17628 | gen_rtx_LABEL_REF (VOIDmode, label), | |
17629 | pc_rtx); | |
17630 | insn = emit_jump_insn (gen_rtx_SET (pc_rtx, temp)); | |
17631 | predict_jump (REG_BR_PROB_BASE * 10 / 100); | |
17632 | JUMP_LABEL (insn) = label; | |
17633 | } | |
17634 | ||
17635 | /* Output code to perform an sinh XFmode calculation. */ | |
17636 | ||
152f243f JJ |
17637 | void |
17638 | ix86_emit_i387_sinh (rtx op0, rtx op1) | |
2bf6d935 ML |
17639 | { |
17640 | rtx e1 = gen_reg_rtx (XFmode); | |
17641 | rtx e2 = gen_reg_rtx (XFmode); | |
17642 | rtx scratch = gen_reg_rtx (HImode); | |
17643 | rtx flags = gen_rtx_REG (CCNOmode, FLAGS_REG); | |
17644 | rtx half = const_double_from_real_value (dconsthalf, XFmode); | |
17645 | rtx cst1, tmp; | |
17646 | rtx_code_label *jump_label = gen_label_rtx (); | |
17647 | rtx_insn *insn; | |
17648 | ||
17649 | /* scratch = fxam (op1) */ | |
17650 | emit_insn (gen_fxamxf2_i387 (scratch, op1)); | |
17651 | ||
17652 | /* e1 = expm1 (|op1|) */ | |
17653 | emit_insn (gen_absxf2 (e2, op1)); | |
17654 | emit_insn (gen_expm1xf2 (e1, e2)); | |
17655 | ||
17656 | /* e2 = e1 / (e1 + 1.0) + e1 */ | |
17657 | cst1 = force_reg (XFmode, CONST1_RTX (XFmode)); | |
17658 | emit_insn (gen_addxf3 (e2, e1, cst1)); | |
17659 | emit_insn (gen_divxf3 (e2, e1, e2)); | |
17660 | emit_insn (gen_addxf3 (e2, e2, e1)); | |
17661 | ||
17662 | /* flags = signbit (op1) */ | |
17663 | emit_insn (gen_testqi_ext_1_ccno (scratch, GEN_INT (0x02))); | |
17664 | ||
17665 | /* if (flags) then e2 = -e2 */ | |
17666 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, | |
17667 | gen_rtx_EQ (VOIDmode, flags, const0_rtx), | |
17668 | gen_rtx_LABEL_REF (VOIDmode, jump_label), | |
17669 | pc_rtx); | |
17670 | insn = emit_jump_insn (gen_rtx_SET (pc_rtx, tmp)); | |
17671 | predict_jump (REG_BR_PROB_BASE * 50 / 100); | |
17672 | JUMP_LABEL (insn) = jump_label; | |
17673 | ||
17674 | emit_insn (gen_negxf2 (e2, e2)); | |
17675 | ||
17676 | emit_label (jump_label); | |
17677 | LABEL_NUSES (jump_label) = 1; | |
17678 | ||
17679 | /* op0 = 0.5 * e2 */ | |
17680 | half = force_reg (XFmode, half); | |
17681 | emit_insn (gen_mulxf3 (op0, e2, half)); | |
17682 | } | |
17683 | ||
17684 | /* Output code to perform an cosh XFmode calculation. */ | |
17685 | ||
152f243f JJ |
17686 | void |
17687 | ix86_emit_i387_cosh (rtx op0, rtx op1) | |
2bf6d935 ML |
17688 | { |
17689 | rtx e1 = gen_reg_rtx (XFmode); | |
17690 | rtx e2 = gen_reg_rtx (XFmode); | |
17691 | rtx half = const_double_from_real_value (dconsthalf, XFmode); | |
17692 | rtx cst1; | |
17693 | ||
17694 | /* e1 = exp (op1) */ | |
17695 | emit_insn (gen_expxf2 (e1, op1)); | |
17696 | ||
17697 | /* e2 = e1 + 1.0 / e1 */ | |
17698 | cst1 = force_reg (XFmode, CONST1_RTX (XFmode)); | |
17699 | emit_insn (gen_divxf3 (e2, cst1, e1)); | |
17700 | emit_insn (gen_addxf3 (e2, e1, e2)); | |
17701 | ||
17702 | /* op0 = 0.5 * e2 */ | |
17703 | half = force_reg (XFmode, half); | |
17704 | emit_insn (gen_mulxf3 (op0, e2, half)); | |
17705 | } | |
17706 | ||
17707 | /* Output code to perform an tanh XFmode calculation. */ | |
17708 | ||
152f243f JJ |
17709 | void |
17710 | ix86_emit_i387_tanh (rtx op0, rtx op1) | |
2bf6d935 ML |
17711 | { |
17712 | rtx e1 = gen_reg_rtx (XFmode); | |
17713 | rtx e2 = gen_reg_rtx (XFmode); | |
17714 | rtx scratch = gen_reg_rtx (HImode); | |
17715 | rtx flags = gen_rtx_REG (CCNOmode, FLAGS_REG); | |
17716 | rtx cst2, tmp; | |
17717 | rtx_code_label *jump_label = gen_label_rtx (); | |
17718 | rtx_insn *insn; | |
17719 | ||
17720 | /* scratch = fxam (op1) */ | |
17721 | emit_insn (gen_fxamxf2_i387 (scratch, op1)); | |
17722 | ||
17723 | /* e1 = expm1 (-|2 * op1|) */ | |
17724 | emit_insn (gen_addxf3 (e2, op1, op1)); | |
17725 | emit_insn (gen_absxf2 (e2, e2)); | |
17726 | emit_insn (gen_negxf2 (e2, e2)); | |
17727 | emit_insn (gen_expm1xf2 (e1, e2)); | |
17728 | ||
17729 | /* e2 = e1 / (e1 + 2.0) */ | |
17730 | cst2 = force_reg (XFmode, CONST2_RTX (XFmode)); | |
17731 | emit_insn (gen_addxf3 (e2, e1, cst2)); | |
17732 | emit_insn (gen_divxf3 (e2, e1, e2)); | |
17733 | ||
17734 | /* flags = signbit (op1) */ | |
17735 | emit_insn (gen_testqi_ext_1_ccno (scratch, GEN_INT (0x02))); | |
17736 | ||
17737 | /* if (!flags) then e2 = -e2 */ | |
17738 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, | |
17739 | gen_rtx_NE (VOIDmode, flags, const0_rtx), | |
17740 | gen_rtx_LABEL_REF (VOIDmode, jump_label), | |
17741 | pc_rtx); | |
17742 | insn = emit_jump_insn (gen_rtx_SET (pc_rtx, tmp)); | |
17743 | predict_jump (REG_BR_PROB_BASE * 50 / 100); | |
17744 | JUMP_LABEL (insn) = jump_label; | |
17745 | ||
17746 | emit_insn (gen_negxf2 (e2, e2)); | |
17747 | ||
17748 | emit_label (jump_label); | |
17749 | LABEL_NUSES (jump_label) = 1; | |
17750 | ||
17751 | emit_move_insn (op0, e2); | |
17752 | } | |
17753 | ||
17754 | /* Output code to perform an asinh XFmode calculation. */ | |
17755 | ||
152f243f JJ |
17756 | void |
17757 | ix86_emit_i387_asinh (rtx op0, rtx op1) | |
2bf6d935 ML |
17758 | { |
17759 | rtx e1 = gen_reg_rtx (XFmode); | |
17760 | rtx e2 = gen_reg_rtx (XFmode); | |
17761 | rtx scratch = gen_reg_rtx (HImode); | |
17762 | rtx flags = gen_rtx_REG (CCNOmode, FLAGS_REG); | |
17763 | rtx cst1, tmp; | |
17764 | rtx_code_label *jump_label = gen_label_rtx (); | |
17765 | rtx_insn *insn; | |
17766 | ||
17767 | /* e2 = sqrt (op1^2 + 1.0) + 1.0 */ | |
17768 | emit_insn (gen_mulxf3 (e1, op1, op1)); | |
17769 | cst1 = force_reg (XFmode, CONST1_RTX (XFmode)); | |
17770 | emit_insn (gen_addxf3 (e2, e1, cst1)); | |
17771 | emit_insn (gen_sqrtxf2 (e2, e2)); | |
17772 | emit_insn (gen_addxf3 (e2, e2, cst1)); | |
17773 | ||
17774 | /* e1 = e1 / e2 */ | |
17775 | emit_insn (gen_divxf3 (e1, e1, e2)); | |
17776 | ||
17777 | /* scratch = fxam (op1) */ | |
17778 | emit_insn (gen_fxamxf2_i387 (scratch, op1)); | |
17779 | ||
17780 | /* e1 = e1 + |op1| */ | |
17781 | emit_insn (gen_absxf2 (e2, op1)); | |
17782 | emit_insn (gen_addxf3 (e1, e1, e2)); | |
17783 | ||
17784 | /* e2 = log1p (e1) */ | |
17785 | ix86_emit_i387_log1p (e2, e1); | |
17786 | ||
17787 | /* flags = signbit (op1) */ | |
17788 | emit_insn (gen_testqi_ext_1_ccno (scratch, GEN_INT (0x02))); | |
17789 | ||
17790 | /* if (flags) then e2 = -e2 */ | |
17791 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, | |
17792 | gen_rtx_EQ (VOIDmode, flags, const0_rtx), | |
17793 | gen_rtx_LABEL_REF (VOIDmode, jump_label), | |
17794 | pc_rtx); | |
17795 | insn = emit_jump_insn (gen_rtx_SET (pc_rtx, tmp)); | |
17796 | predict_jump (REG_BR_PROB_BASE * 50 / 100); | |
17797 | JUMP_LABEL (insn) = jump_label; | |
17798 | ||
17799 | emit_insn (gen_negxf2 (e2, e2)); | |
17800 | ||
17801 | emit_label (jump_label); | |
17802 | LABEL_NUSES (jump_label) = 1; | |
17803 | ||
17804 | emit_move_insn (op0, e2); | |
17805 | } | |
17806 | ||
17807 | /* Output code to perform an acosh XFmode calculation. */ | |
17808 | ||
152f243f JJ |
17809 | void |
17810 | ix86_emit_i387_acosh (rtx op0, rtx op1) | |
2bf6d935 ML |
17811 | { |
17812 | rtx e1 = gen_reg_rtx (XFmode); | |
17813 | rtx e2 = gen_reg_rtx (XFmode); | |
17814 | rtx cst1 = force_reg (XFmode, CONST1_RTX (XFmode)); | |
17815 | ||
17816 | /* e2 = sqrt (op1 + 1.0) */ | |
17817 | emit_insn (gen_addxf3 (e2, op1, cst1)); | |
17818 | emit_insn (gen_sqrtxf2 (e2, e2)); | |
17819 | ||
17820 | /* e1 = sqrt (op1 - 1.0) */ | |
17821 | emit_insn (gen_subxf3 (e1, op1, cst1)); | |
17822 | emit_insn (gen_sqrtxf2 (e1, e1)); | |
17823 | ||
17824 | /* e1 = e1 * e2 */ | |
17825 | emit_insn (gen_mulxf3 (e1, e1, e2)); | |
17826 | ||
17827 | /* e1 = e1 + op1 */ | |
17828 | emit_insn (gen_addxf3 (e1, e1, op1)); | |
17829 | ||
17830 | /* op0 = log (e1) */ | |
17831 | emit_insn (gen_logxf2 (op0, e1)); | |
17832 | } | |
17833 | ||
17834 | /* Output code to perform an atanh XFmode calculation. */ | |
17835 | ||
152f243f JJ |
17836 | void |
17837 | ix86_emit_i387_atanh (rtx op0, rtx op1) | |
2bf6d935 ML |
17838 | { |
17839 | rtx e1 = gen_reg_rtx (XFmode); | |
17840 | rtx e2 = gen_reg_rtx (XFmode); | |
17841 | rtx scratch = gen_reg_rtx (HImode); | |
17842 | rtx flags = gen_rtx_REG (CCNOmode, FLAGS_REG); | |
17843 | rtx half = const_double_from_real_value (dconsthalf, XFmode); | |
17844 | rtx cst1, tmp; | |
17845 | rtx_code_label *jump_label = gen_label_rtx (); | |
17846 | rtx_insn *insn; | |
17847 | ||
17848 | /* scratch = fxam (op1) */ | |
17849 | emit_insn (gen_fxamxf2_i387 (scratch, op1)); | |
17850 | ||
17851 | /* e2 = |op1| */ | |
17852 | emit_insn (gen_absxf2 (e2, op1)); | |
17853 | ||
17854 | /* e1 = -(e2 + e2) / (e2 + 1.0) */ | |
17855 | cst1 = force_reg (XFmode, CONST1_RTX (XFmode)); | |
17856 | emit_insn (gen_addxf3 (e1, e2, cst1)); | |
17857 | emit_insn (gen_addxf3 (e2, e2, e2)); | |
17858 | emit_insn (gen_negxf2 (e2, e2)); | |
17859 | emit_insn (gen_divxf3 (e1, e2, e1)); | |
17860 | ||
17861 | /* e2 = log1p (e1) */ | |
17862 | ix86_emit_i387_log1p (e2, e1); | |
17863 | ||
17864 | /* flags = signbit (op1) */ | |
17865 | emit_insn (gen_testqi_ext_1_ccno (scratch, GEN_INT (0x02))); | |
17866 | ||
17867 | /* if (!flags) then e2 = -e2 */ | |
17868 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, | |
17869 | gen_rtx_NE (VOIDmode, flags, const0_rtx), | |
17870 | gen_rtx_LABEL_REF (VOIDmode, jump_label), | |
17871 | pc_rtx); | |
17872 | insn = emit_jump_insn (gen_rtx_SET (pc_rtx, tmp)); | |
17873 | predict_jump (REG_BR_PROB_BASE * 50 / 100); | |
17874 | JUMP_LABEL (insn) = jump_label; | |
17875 | ||
17876 | emit_insn (gen_negxf2 (e2, e2)); | |
17877 | ||
17878 | emit_label (jump_label); | |
17879 | LABEL_NUSES (jump_label) = 1; | |
17880 | ||
17881 | /* op0 = 0.5 * e2 */ | |
17882 | half = force_reg (XFmode, half); | |
17883 | emit_insn (gen_mulxf3 (op0, e2, half)); | |
17884 | } | |
17885 | ||
17886 | /* Output code to perform a log1p XFmode calculation. */ | |
17887 | ||
152f243f JJ |
17888 | void |
17889 | ix86_emit_i387_log1p (rtx op0, rtx op1) | |
2bf6d935 ML |
17890 | { |
17891 | rtx_code_label *label1 = gen_label_rtx (); | |
17892 | rtx_code_label *label2 = gen_label_rtx (); | |
17893 | ||
17894 | rtx tmp = gen_reg_rtx (XFmode); | |
17895 | rtx res = gen_reg_rtx (XFmode); | |
17896 | rtx cst, cstln2, cst1; | |
17897 | rtx_insn *insn; | |
17898 | ||
d481d137 JJ |
17899 | /* The emit_jump call emits pending stack adjust, make sure it is emitted |
17900 | before the conditional jump, otherwise the stack adjustment will be | |
17901 | only conditional. */ | |
17902 | do_pending_stack_adjust (); | |
17903 | ||
2bf6d935 ML |
17904 | cst = const_double_from_real_value |
17905 | (REAL_VALUE_ATOF ("0.29289321881345247561810596348408353", XFmode), XFmode); | |
17906 | cstln2 = force_reg (XFmode, standard_80387_constant_rtx (4)); /* fldln2 */ | |
17907 | ||
17908 | emit_insn (gen_absxf2 (tmp, op1)); | |
17909 | ||
17910 | cst = force_reg (XFmode, cst); | |
17911 | ix86_expand_branch (GE, tmp, cst, label1); | |
17912 | predict_jump (REG_BR_PROB_BASE * 10 / 100); | |
17913 | insn = get_last_insn (); | |
17914 | JUMP_LABEL (insn) = label1; | |
17915 | ||
17916 | emit_insn (gen_fyl2xp1xf3_i387 (res, op1, cstln2)); | |
17917 | emit_jump (label2); | |
17918 | ||
17919 | emit_label (label1); | |
17920 | LABEL_NUSES (label1) = 1; | |
17921 | ||
17922 | cst1 = force_reg (XFmode, CONST1_RTX (XFmode)); | |
17923 | emit_insn (gen_rtx_SET (tmp, gen_rtx_PLUS (XFmode, op1, cst1))); | |
17924 | emit_insn (gen_fyl2xxf3_i387 (res, tmp, cstln2)); | |
17925 | ||
17926 | emit_label (label2); | |
17927 | LABEL_NUSES (label2) = 1; | |
17928 | ||
17929 | emit_move_insn (op0, res); | |
17930 | } | |
17931 | ||
17932 | /* Emit code for round calculation. */ | |
152f243f JJ |
17933 | void |
17934 | ix86_emit_i387_round (rtx op0, rtx op1) | |
2bf6d935 ML |
17935 | { |
17936 | machine_mode inmode = GET_MODE (op1); | |
17937 | machine_mode outmode = GET_MODE (op0); | |
17938 | rtx e1 = gen_reg_rtx (XFmode); | |
17939 | rtx e2 = gen_reg_rtx (XFmode); | |
17940 | rtx scratch = gen_reg_rtx (HImode); | |
17941 | rtx flags = gen_rtx_REG (CCNOmode, FLAGS_REG); | |
17942 | rtx half = const_double_from_real_value (dconsthalf, XFmode); | |
17943 | rtx res = gen_reg_rtx (outmode); | |
17944 | rtx_code_label *jump_label = gen_label_rtx (); | |
17945 | rtx (*floor_insn) (rtx, rtx); | |
17946 | rtx (*neg_insn) (rtx, rtx); | |
17947 | rtx_insn *insn; | |
17948 | rtx tmp; | |
17949 | ||
17950 | switch (inmode) | |
17951 | { | |
17952 | case E_SFmode: | |
17953 | case E_DFmode: | |
17954 | tmp = gen_reg_rtx (XFmode); | |
17955 | ||
17956 | emit_insn (gen_rtx_SET (tmp, gen_rtx_FLOAT_EXTEND (XFmode, op1))); | |
17957 | op1 = tmp; | |
17958 | break; | |
17959 | case E_XFmode: | |
17960 | break; | |
17961 | default: | |
17962 | gcc_unreachable (); | |
17963 | } | |
17964 | ||
17965 | switch (outmode) | |
17966 | { | |
17967 | case E_SFmode: | |
17968 | floor_insn = gen_frndintxf2_floor; | |
17969 | neg_insn = gen_negsf2; | |
17970 | break; | |
17971 | case E_DFmode: | |
17972 | floor_insn = gen_frndintxf2_floor; | |
17973 | neg_insn = gen_negdf2; | |
17974 | break; | |
17975 | case E_XFmode: | |
17976 | floor_insn = gen_frndintxf2_floor; | |
17977 | neg_insn = gen_negxf2; | |
17978 | break; | |
17979 | case E_HImode: | |
17980 | floor_insn = gen_lfloorxfhi2; | |
17981 | neg_insn = gen_neghi2; | |
17982 | break; | |
17983 | case E_SImode: | |
17984 | floor_insn = gen_lfloorxfsi2; | |
17985 | neg_insn = gen_negsi2; | |
17986 | break; | |
17987 | case E_DImode: | |
17988 | floor_insn = gen_lfloorxfdi2; | |
17989 | neg_insn = gen_negdi2; | |
17990 | break; | |
17991 | default: | |
17992 | gcc_unreachable (); | |
17993 | } | |
17994 | ||
17995 | /* round(a) = sgn(a) * floor(fabs(a) + 0.5) */ | |
17996 | ||
17997 | /* scratch = fxam(op1) */ | |
17998 | emit_insn (gen_fxamxf2_i387 (scratch, op1)); | |
17999 | ||
18000 | /* e1 = fabs(op1) */ | |
18001 | emit_insn (gen_absxf2 (e1, op1)); | |
18002 | ||
18003 | /* e2 = e1 + 0.5 */ | |
18004 | half = force_reg (XFmode, half); | |
18005 | emit_insn (gen_rtx_SET (e2, gen_rtx_PLUS (XFmode, e1, half))); | |
18006 | ||
18007 | /* res = floor(e2) */ | |
18008 | switch (outmode) | |
18009 | { | |
18010 | case E_SFmode: | |
18011 | case E_DFmode: | |
18012 | { | |
18013 | tmp = gen_reg_rtx (XFmode); | |
18014 | ||
18015 | emit_insn (floor_insn (tmp, e2)); | |
18016 | emit_insn (gen_rtx_SET (res, | |
18017 | gen_rtx_UNSPEC (outmode, gen_rtvec (1, tmp), | |
18018 | UNSPEC_TRUNC_NOOP))); | |
18019 | } | |
18020 | break; | |
18021 | default: | |
18022 | emit_insn (floor_insn (res, e2)); | |
18023 | } | |
18024 | ||
18025 | /* flags = signbit(a) */ | |
18026 | emit_insn (gen_testqi_ext_1_ccno (scratch, GEN_INT (0x02))); | |
18027 | ||
18028 | /* if (flags) then res = -res */ | |
18029 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, | |
18030 | gen_rtx_EQ (VOIDmode, flags, const0_rtx), | |
18031 | gen_rtx_LABEL_REF (VOIDmode, jump_label), | |
18032 | pc_rtx); | |
18033 | insn = emit_jump_insn (gen_rtx_SET (pc_rtx, tmp)); | |
18034 | predict_jump (REG_BR_PROB_BASE * 50 / 100); | |
18035 | JUMP_LABEL (insn) = jump_label; | |
18036 | ||
18037 | emit_insn (neg_insn (res, res)); | |
18038 | ||
18039 | emit_label (jump_label); | |
18040 | LABEL_NUSES (jump_label) = 1; | |
18041 | ||
18042 | emit_move_insn (op0, res); | |
18043 | } | |
18044 | ||
18045 | /* Output code to perform a Newton-Rhapson approximation of a single precision | |
18046 | floating point divide [http://en.wikipedia.org/wiki/N-th_root_algorithm]. */ | |
18047 | ||
152f243f JJ |
18048 | void |
18049 | ix86_emit_swdivsf (rtx res, rtx a, rtx b, machine_mode mode) | |
2bf6d935 ML |
18050 | { |
18051 | rtx x0, x1, e0, e1; | |
18052 | ||
18053 | x0 = gen_reg_rtx (mode); | |
18054 | e0 = gen_reg_rtx (mode); | |
18055 | e1 = gen_reg_rtx (mode); | |
18056 | x1 = gen_reg_rtx (mode); | |
18057 | ||
18058 | /* a / b = a * ((rcp(b) + rcp(b)) - (b * rcp(b) * rcp (b))) */ | |
18059 | ||
18060 | b = force_reg (mode, b); | |
18061 | ||
18062 | /* x0 = rcp(b) estimate */ | |
18063 | if (mode == V16SFmode || mode == V8DFmode) | |
18064 | { | |
18065 | if (TARGET_AVX512ER) | |
18066 | { | |
18067 | emit_insn (gen_rtx_SET (x0, gen_rtx_UNSPEC (mode, gen_rtvec (1, b), | |
18068 | UNSPEC_RCP28))); | |
18069 | /* res = a * x0 */ | |
18070 | emit_insn (gen_rtx_SET (res, gen_rtx_MULT (mode, a, x0))); | |
18071 | return; | |
18072 | } | |
18073 | else | |
18074 | emit_insn (gen_rtx_SET (x0, gen_rtx_UNSPEC (mode, gen_rtvec (1, b), | |
18075 | UNSPEC_RCP14))); | |
18076 | } | |
18077 | else | |
18078 | emit_insn (gen_rtx_SET (x0, gen_rtx_UNSPEC (mode, gen_rtvec (1, b), | |
18079 | UNSPEC_RCP))); | |
18080 | ||
18081 | /* e0 = x0 * b */ | |
18082 | emit_insn (gen_rtx_SET (e0, gen_rtx_MULT (mode, x0, b))); | |
18083 | ||
18084 | /* e0 = x0 * e0 */ | |
18085 | emit_insn (gen_rtx_SET (e0, gen_rtx_MULT (mode, x0, e0))); | |
18086 | ||
18087 | /* e1 = x0 + x0 */ | |
18088 | emit_insn (gen_rtx_SET (e1, gen_rtx_PLUS (mode, x0, x0))); | |
18089 | ||
18090 | /* x1 = e1 - e0 */ | |
18091 | emit_insn (gen_rtx_SET (x1, gen_rtx_MINUS (mode, e1, e0))); | |
18092 | ||
18093 | /* res = a * x1 */ | |
18094 | emit_insn (gen_rtx_SET (res, gen_rtx_MULT (mode, a, x1))); | |
18095 | } | |
18096 | ||
18097 | /* Output code to perform a Newton-Rhapson approximation of a | |
18098 | single precision floating point [reciprocal] square root. */ | |
18099 | ||
152f243f JJ |
18100 | void |
18101 | ix86_emit_swsqrtsf (rtx res, rtx a, machine_mode mode, bool recip) | |
2bf6d935 ML |
18102 | { |
18103 | rtx x0, e0, e1, e2, e3, mthree, mhalf; | |
18104 | REAL_VALUE_TYPE r; | |
18105 | int unspec; | |
18106 | ||
18107 | x0 = gen_reg_rtx (mode); | |
18108 | e0 = gen_reg_rtx (mode); | |
18109 | e1 = gen_reg_rtx (mode); | |
18110 | e2 = gen_reg_rtx (mode); | |
18111 | e3 = gen_reg_rtx (mode); | |
18112 | ||
18113 | if (TARGET_AVX512ER && mode == V16SFmode) | |
18114 | { | |
18115 | if (recip) | |
18116 | /* res = rsqrt28(a) estimate */ | |
18117 | emit_insn (gen_rtx_SET (res, gen_rtx_UNSPEC (mode, gen_rtvec (1, a), | |
18118 | UNSPEC_RSQRT28))); | |
18119 | else | |
18120 | { | |
18121 | /* x0 = rsqrt28(a) estimate */ | |
18122 | emit_insn (gen_rtx_SET (x0, gen_rtx_UNSPEC (mode, gen_rtvec (1, a), | |
18123 | UNSPEC_RSQRT28))); | |
18124 | /* res = rcp28(x0) estimate */ | |
18125 | emit_insn (gen_rtx_SET (res, gen_rtx_UNSPEC (mode, gen_rtvec (1, x0), | |
18126 | UNSPEC_RCP28))); | |
18127 | } | |
18128 | return; | |
18129 | } | |
18130 | ||
18131 | real_from_integer (&r, VOIDmode, -3, SIGNED); | |
18132 | mthree = const_double_from_real_value (r, SFmode); | |
18133 | ||
18134 | real_arithmetic (&r, NEGATE_EXPR, &dconsthalf, NULL); | |
18135 | mhalf = const_double_from_real_value (r, SFmode); | |
18136 | unspec = UNSPEC_RSQRT; | |
18137 | ||
18138 | if (VECTOR_MODE_P (mode)) | |
18139 | { | |
18140 | mthree = ix86_build_const_vector (mode, true, mthree); | |
18141 | mhalf = ix86_build_const_vector (mode, true, mhalf); | |
18142 | /* There is no 512-bit rsqrt. There is however rsqrt14. */ | |
18143 | if (GET_MODE_SIZE (mode) == 64) | |
18144 | unspec = UNSPEC_RSQRT14; | |
18145 | } | |
18146 | ||
18147 | /* sqrt(a) = -0.5 * a * rsqrtss(a) * (a * rsqrtss(a) * rsqrtss(a) - 3.0) | |
18148 | rsqrt(a) = -0.5 * rsqrtss(a) * (a * rsqrtss(a) * rsqrtss(a) - 3.0) */ | |
18149 | ||
18150 | a = force_reg (mode, a); | |
18151 | ||
18152 | /* x0 = rsqrt(a) estimate */ | |
18153 | emit_insn (gen_rtx_SET (x0, gen_rtx_UNSPEC (mode, gen_rtvec (1, a), | |
18154 | unspec))); | |
18155 | ||
18156 | /* If (a == 0.0) Filter out infinity to prevent NaN for sqrt(0.0). */ | |
18157 | if (!recip) | |
18158 | { | |
18159 | rtx zero = force_reg (mode, CONST0_RTX(mode)); | |
18160 | rtx mask; | |
18161 | ||
18162 | /* Handle masked compare. */ | |
18163 | if (VECTOR_MODE_P (mode) && GET_MODE_SIZE (mode) == 64) | |
18164 | { | |
18165 | mask = gen_reg_rtx (HImode); | |
18166 | /* Imm value 0x4 corresponds to not-equal comparison. */ | |
18167 | emit_insn (gen_avx512f_cmpv16sf3 (mask, zero, a, GEN_INT (0x4))); | |
18168 | emit_insn (gen_avx512f_blendmv16sf (x0, zero, x0, mask)); | |
18169 | } | |
18170 | else | |
18171 | { | |
18172 | mask = gen_reg_rtx (mode); | |
18173 | emit_insn (gen_rtx_SET (mask, gen_rtx_NE (mode, zero, a))); | |
18174 | emit_insn (gen_rtx_SET (x0, gen_rtx_AND (mode, x0, mask))); | |
18175 | } | |
18176 | } | |
18177 | ||
fab263ab L |
18178 | mthree = force_reg (mode, mthree); |
18179 | ||
2bf6d935 ML |
18180 | /* e0 = x0 * a */ |
18181 | emit_insn (gen_rtx_SET (e0, gen_rtx_MULT (mode, x0, a))); | |
2bf6d935 | 18182 | |
a6645a82 L |
18183 | unsigned vector_size = GET_MODE_SIZE (mode); |
18184 | if (TARGET_FMA | |
18185 | || (TARGET_AVX512F && vector_size == 64) | |
18186 | || (TARGET_AVX512VL && (vector_size == 32 || vector_size == 16))) | |
fab263ab L |
18187 | emit_insn (gen_rtx_SET (e2, |
18188 | gen_rtx_FMA (mode, e0, x0, mthree))); | |
18189 | else | |
18190 | { | |
18191 | /* e1 = e0 * x0 */ | |
18192 | emit_insn (gen_rtx_SET (e1, gen_rtx_MULT (mode, e0, x0))); | |
18193 | ||
18194 | /* e2 = e1 - 3. */ | |
18195 | emit_insn (gen_rtx_SET (e2, gen_rtx_PLUS (mode, e1, mthree))); | |
18196 | } | |
2bf6d935 ML |
18197 | |
18198 | mhalf = force_reg (mode, mhalf); | |
18199 | if (recip) | |
18200 | /* e3 = -.5 * x0 */ | |
18201 | emit_insn (gen_rtx_SET (e3, gen_rtx_MULT (mode, x0, mhalf))); | |
18202 | else | |
18203 | /* e3 = -.5 * e0 */ | |
18204 | emit_insn (gen_rtx_SET (e3, gen_rtx_MULT (mode, e0, mhalf))); | |
18205 | /* ret = e2 * e3 */ | |
18206 | emit_insn (gen_rtx_SET (res, gen_rtx_MULT (mode, e2, e3))); | |
18207 | } | |
18208 | ||
18209 | /* Expand fabs (OP0) and return a new rtx that holds the result. The | |
18210 | mask for masking out the sign-bit is stored in *SMASK, if that is | |
18211 | non-null. */ | |
18212 | ||
18213 | static rtx | |
18214 | ix86_expand_sse_fabs (rtx op0, rtx *smask) | |
18215 | { | |
18216 | machine_mode vmode, mode = GET_MODE (op0); | |
18217 | rtx xa, mask; | |
18218 | ||
18219 | xa = gen_reg_rtx (mode); | |
18220 | if (mode == SFmode) | |
18221 | vmode = V4SFmode; | |
18222 | else if (mode == DFmode) | |
18223 | vmode = V2DFmode; | |
18224 | else | |
18225 | vmode = mode; | |
18226 | mask = ix86_build_signbit_mask (vmode, VECTOR_MODE_P (mode), true); | |
18227 | if (!VECTOR_MODE_P (mode)) | |
18228 | { | |
18229 | /* We need to generate a scalar mode mask in this case. */ | |
18230 | rtx tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (1, const0_rtx)); | |
18231 | tmp = gen_rtx_VEC_SELECT (mode, mask, tmp); | |
18232 | mask = gen_reg_rtx (mode); | |
18233 | emit_insn (gen_rtx_SET (mask, tmp)); | |
18234 | } | |
18235 | emit_insn (gen_rtx_SET (xa, gen_rtx_AND (mode, op0, mask))); | |
18236 | ||
18237 | if (smask) | |
18238 | *smask = mask; | |
18239 | ||
18240 | return xa; | |
18241 | } | |
18242 | ||
18243 | /* Expands a comparison of OP0 with OP1 using comparison code CODE, | |
18244 | swapping the operands if SWAP_OPERANDS is true. The expanded | |
18245 | code is a forward jump to a newly created label in case the | |
18246 | comparison is true. The generated label rtx is returned. */ | |
18247 | static rtx_code_label * | |
18248 | ix86_expand_sse_compare_and_jump (enum rtx_code code, rtx op0, rtx op1, | |
18249 | bool swap_operands) | |
18250 | { | |
18251 | bool unordered_compare = ix86_unordered_fp_compare (code); | |
18252 | rtx_code_label *label; | |
18253 | rtx tmp, reg; | |
18254 | ||
18255 | if (swap_operands) | |
18256 | std::swap (op0, op1); | |
18257 | ||
18258 | label = gen_label_rtx (); | |
18259 | tmp = gen_rtx_COMPARE (CCFPmode, op0, op1); | |
18260 | if (unordered_compare) | |
18261 | tmp = gen_rtx_UNSPEC (CCFPmode, gen_rtvec (1, tmp), UNSPEC_NOTRAP); | |
18262 | reg = gen_rtx_REG (CCFPmode, FLAGS_REG); | |
18263 | emit_insn (gen_rtx_SET (reg, tmp)); | |
18264 | tmp = gen_rtx_fmt_ee (code, VOIDmode, reg, const0_rtx); | |
18265 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp, | |
18266 | gen_rtx_LABEL_REF (VOIDmode, label), pc_rtx); | |
18267 | tmp = emit_jump_insn (gen_rtx_SET (pc_rtx, tmp)); | |
18268 | JUMP_LABEL (tmp) = label; | |
18269 | ||
18270 | return label; | |
18271 | } | |
18272 | ||
18273 | /* Expand a mask generating SSE comparison instruction comparing OP0 with OP1 | |
18274 | using comparison code CODE. Operands are swapped for the comparison if | |
18275 | SWAP_OPERANDS is true. Returns a rtx for the generated mask. */ | |
18276 | static rtx | |
18277 | ix86_expand_sse_compare_mask (enum rtx_code code, rtx op0, rtx op1, | |
18278 | bool swap_operands) | |
18279 | { | |
18280 | rtx (*insn)(rtx, rtx, rtx, rtx); | |
18281 | machine_mode mode = GET_MODE (op0); | |
18282 | rtx mask = gen_reg_rtx (mode); | |
18283 | ||
18284 | if (swap_operands) | |
18285 | std::swap (op0, op1); | |
18286 | ||
18287 | insn = mode == DFmode ? gen_setcc_df_sse : gen_setcc_sf_sse; | |
18288 | ||
18289 | emit_insn (insn (mask, op0, op1, | |
18290 | gen_rtx_fmt_ee (code, mode, op0, op1))); | |
18291 | return mask; | |
18292 | } | |
18293 | ||
18294 | /* Expand copysign from SIGN to the positive value ABS_VALUE | |
18295 | storing in RESULT. If MASK is non-null, it shall be a mask to mask out | |
18296 | the sign-bit. */ | |
18297 | ||
18298 | static void | |
18299 | ix86_sse_copysign_to_positive (rtx result, rtx abs_value, rtx sign, rtx mask) | |
18300 | { | |
18301 | machine_mode mode = GET_MODE (sign); | |
18302 | rtx sgn = gen_reg_rtx (mode); | |
18303 | if (mask == NULL_RTX) | |
18304 | { | |
18305 | machine_mode vmode; | |
18306 | ||
18307 | if (mode == SFmode) | |
18308 | vmode = V4SFmode; | |
18309 | else if (mode == DFmode) | |
18310 | vmode = V2DFmode; | |
18311 | else | |
18312 | vmode = mode; | |
18313 | ||
18314 | mask = ix86_build_signbit_mask (vmode, VECTOR_MODE_P (mode), false); | |
18315 | if (!VECTOR_MODE_P (mode)) | |
18316 | { | |
18317 | /* We need to generate a scalar mode mask in this case. */ | |
18318 | rtx tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (1, const0_rtx)); | |
18319 | tmp = gen_rtx_VEC_SELECT (mode, mask, tmp); | |
18320 | mask = gen_reg_rtx (mode); | |
18321 | emit_insn (gen_rtx_SET (mask, tmp)); | |
18322 | } | |
18323 | } | |
18324 | else | |
18325 | mask = gen_rtx_NOT (mode, mask); | |
18326 | emit_insn (gen_rtx_SET (sgn, gen_rtx_AND (mode, mask, sign))); | |
18327 | emit_insn (gen_rtx_SET (result, gen_rtx_IOR (mode, abs_value, sgn))); | |
18328 | } | |
18329 | ||
18330 | /* Expand SSE sequence for computing lround from OP1 storing | |
18331 | into OP0. */ | |
18332 | ||
18333 | void | |
18334 | ix86_expand_lround (rtx op0, rtx op1) | |
18335 | { | |
18336 | /* C code for the stuff we're doing below: | |
d2754fbb UB |
18337 | tmp = op1 + copysign (nextafter (0.5, 0.0), op1) |
18338 | return (long)tmp; | |
2bf6d935 ML |
18339 | */ |
18340 | machine_mode mode = GET_MODE (op1); | |
18341 | const struct real_format *fmt; | |
18342 | REAL_VALUE_TYPE pred_half, half_minus_pred_half; | |
18343 | rtx adj; | |
18344 | ||
18345 | /* load nextafter (0.5, 0.0) */ | |
18346 | fmt = REAL_MODE_FORMAT (mode); | |
18347 | real_2expN (&half_minus_pred_half, -(fmt->p) - 1, mode); | |
18348 | real_arithmetic (&pred_half, MINUS_EXPR, &dconsthalf, &half_minus_pred_half); | |
18349 | ||
18350 | /* adj = copysign (0.5, op1) */ | |
18351 | adj = force_reg (mode, const_double_from_real_value (pred_half, mode)); | |
18352 | ix86_sse_copysign_to_positive (adj, adj, force_reg (mode, op1), NULL_RTX); | |
18353 | ||
18354 | /* adj = op1 + adj */ | |
18355 | adj = expand_simple_binop (mode, PLUS, adj, op1, NULL_RTX, 0, OPTAB_DIRECT); | |
18356 | ||
18357 | /* op0 = (imode)adj */ | |
18358 | expand_fix (op0, adj, 0); | |
18359 | } | |
18360 | ||
18361 | /* Expand SSE2 sequence for computing lround from OPERAND1 storing | |
18362 | into OPERAND0. */ | |
18363 | ||
18364 | void | |
18365 | ix86_expand_lfloorceil (rtx op0, rtx op1, bool do_floor) | |
18366 | { | |
18367 | /* C code for the stuff we're doing below (for do_floor): | |
18368 | xi = (long)op1; | |
d2754fbb UB |
18369 | xi -= (double)xi > op1 ? 1 : 0; |
18370 | return xi; | |
2bf6d935 ML |
18371 | */ |
18372 | machine_mode fmode = GET_MODE (op1); | |
18373 | machine_mode imode = GET_MODE (op0); | |
18374 | rtx ireg, freg, tmp; | |
18375 | rtx_code_label *label; | |
18376 | ||
18377 | /* reg = (long)op1 */ | |
18378 | ireg = gen_reg_rtx (imode); | |
18379 | expand_fix (ireg, op1, 0); | |
18380 | ||
18381 | /* freg = (double)reg */ | |
18382 | freg = gen_reg_rtx (fmode); | |
18383 | expand_float (freg, ireg, 0); | |
18384 | ||
18385 | /* ireg = (freg > op1) ? ireg - 1 : ireg */ | |
18386 | label = ix86_expand_sse_compare_and_jump (UNLE, | |
18387 | freg, op1, !do_floor); | |
18388 | tmp = expand_simple_binop (imode, do_floor ? MINUS : PLUS, | |
18389 | ireg, const1_rtx, NULL_RTX, 0, OPTAB_DIRECT); | |
18390 | emit_move_insn (ireg, tmp); | |
18391 | ||
18392 | emit_label (label); | |
18393 | LABEL_NUSES (label) = 1; | |
18394 | ||
18395 | emit_move_insn (op0, ireg); | |
18396 | } | |
18397 | ||
18398 | /* Generate and return a rtx of mode MODE for 2**n where n is the number | |
18399 | of bits of the mantissa of MODE, which must be one of DFmode or SFmode. */ | |
18400 | ||
18401 | static rtx | |
18402 | ix86_gen_TWO52 (machine_mode mode) | |
18403 | { | |
d2754fbb | 18404 | const struct real_format *fmt; |
2bf6d935 ML |
18405 | REAL_VALUE_TYPE TWO52r; |
18406 | rtx TWO52; | |
18407 | ||
d2754fbb UB |
18408 | fmt = REAL_MODE_FORMAT (mode); |
18409 | real_2expN (&TWO52r, fmt->p - 1, mode); | |
2bf6d935 ML |
18410 | TWO52 = const_double_from_real_value (TWO52r, mode); |
18411 | TWO52 = force_reg (mode, TWO52); | |
18412 | ||
18413 | return TWO52; | |
18414 | } | |
18415 | ||
18416 | /* Expand rint rounding OPERAND1 and storing the result in OPERAND0. */ | |
18417 | ||
18418 | void | |
18419 | ix86_expand_rint (rtx operand0, rtx operand1) | |
18420 | { | |
18421 | /* C code for the stuff we're doing below: | |
18422 | xa = fabs (operand1); | |
d2754fbb | 18423 | if (!isless (xa, 2**52)) |
2bf6d935 | 18424 | return operand1; |
d2754fbb UB |
18425 | two52 = 2**52; |
18426 | if (flag_rounding_math) | |
2bf6d935 ML |
18427 | { |
18428 | two52 = copysign (two52, operand1); | |
18429 | xa = operand1; | |
18430 | } | |
d2754fbb UB |
18431 | xa = xa + two52 - two52; |
18432 | return copysign (xa, operand1); | |
2bf6d935 ML |
18433 | */ |
18434 | machine_mode mode = GET_MODE (operand0); | |
81615bb0 | 18435 | rtx res, xa, TWO52, mask; |
2bf6d935 ML |
18436 | rtx_code_label *label; |
18437 | ||
d2754fbb UB |
18438 | TWO52 = ix86_gen_TWO52 (mode); |
18439 | ||
18440 | /* Temporary for holding the result, initialized to the input | |
18441 | operand to ease control flow. */ | |
18442 | res = copy_to_reg (operand1); | |
2bf6d935 ML |
18443 | |
18444 | /* xa = abs (operand1) */ | |
18445 | xa = ix86_expand_sse_fabs (res, &mask); | |
18446 | ||
18447 | /* if (!isless (xa, TWO52)) goto label; */ | |
2bf6d935 ML |
18448 | label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false); |
18449 | ||
2bf6d935 ML |
18450 | if (flag_rounding_math) |
18451 | { | |
81615bb0 | 18452 | ix86_sse_copysign_to_positive (TWO52, TWO52, res, mask); |
2bf6d935 ML |
18453 | xa = res; |
18454 | } | |
18455 | ||
81615bb0 UB |
18456 | xa = expand_simple_binop (mode, PLUS, xa, TWO52, NULL_RTX, 0, OPTAB_DIRECT); |
18457 | xa = expand_simple_binop (mode, MINUS, xa, TWO52, xa, 0, OPTAB_DIRECT); | |
18458 | ||
18459 | /* Remove the sign with FE_DOWNWARD, where x - x = -0.0. */ | |
18460 | if (HONOR_SIGNED_ZEROS (mode) && flag_rounding_math) | |
18461 | xa = ix86_expand_sse_fabs (xa, NULL); | |
2bf6d935 ML |
18462 | |
18463 | ix86_sse_copysign_to_positive (res, xa, res, mask); | |
18464 | ||
18465 | emit_label (label); | |
18466 | LABEL_NUSES (label) = 1; | |
18467 | ||
18468 | emit_move_insn (operand0, res); | |
18469 | } | |
18470 | ||
36d387f2 UB |
18471 | /* Expand SSE2 sequence for computing floor or ceil |
18472 | from OPERAND1 storing into OPERAND0. */ | |
2bf6d935 ML |
18473 | void |
18474 | ix86_expand_floorceil (rtx operand0, rtx operand1, bool do_floor) | |
18475 | { | |
18476 | /* C code for the stuff we expand below. | |
18477 | double xa = fabs (x), x2; | |
d2754fbb UB |
18478 | if (!isless (xa, TWO52)) |
18479 | return x; | |
2bf6d935 | 18480 | x2 = (double)(long)x; |
337ed0eb | 18481 | |
2bf6d935 ML |
18482 | Compensate. Floor: |
18483 | if (x2 > x) | |
18484 | x2 -= 1; | |
18485 | Compensate. Ceil: | |
18486 | if (x2 < x) | |
18487 | x2 += 1; | |
337ed0eb | 18488 | |
2bf6d935 ML |
18489 | if (HONOR_SIGNED_ZEROS (mode)) |
18490 | return copysign (x2, x); | |
18491 | return x2; | |
18492 | */ | |
18493 | machine_mode mode = GET_MODE (operand0); | |
18494 | rtx xa, xi, TWO52, tmp, one, res, mask; | |
18495 | rtx_code_label *label; | |
18496 | ||
18497 | TWO52 = ix86_gen_TWO52 (mode); | |
18498 | ||
18499 | /* Temporary for holding the result, initialized to the input | |
18500 | operand to ease control flow. */ | |
d2754fbb | 18501 | res = copy_to_reg (operand1); |
2bf6d935 ML |
18502 | |
18503 | /* xa = abs (operand1) */ | |
18504 | xa = ix86_expand_sse_fabs (res, &mask); | |
18505 | ||
18506 | /* if (!isless (xa, TWO52)) goto label; */ | |
18507 | label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false); | |
18508 | ||
18509 | /* xa = (double)(long)x */ | |
d2754fbb | 18510 | xi = gen_reg_rtx (int_mode_for_mode (mode).require ()); |
2bf6d935 ML |
18511 | expand_fix (xi, res, 0); |
18512 | expand_float (xa, xi, 0); | |
18513 | ||
18514 | /* generate 1.0 */ | |
18515 | one = force_reg (mode, const_double_from_real_value (dconst1, mode)); | |
18516 | ||
18517 | /* Compensate: xa = xa - (xa > operand1 ? 1 : 0) */ | |
18518 | tmp = ix86_expand_sse_compare_mask (UNGT, xa, res, !do_floor); | |
18519 | emit_insn (gen_rtx_SET (tmp, gen_rtx_AND (mode, one, tmp))); | |
18520 | tmp = expand_simple_binop (mode, do_floor ? MINUS : PLUS, | |
18521 | xa, tmp, NULL_RTX, 0, OPTAB_DIRECT); | |
2bf6d935 | 18522 | if (HONOR_SIGNED_ZEROS (mode)) |
337ed0eb UB |
18523 | { |
18524 | /* Remove the sign with FE_DOWNWARD, where x - x = -0.0. */ | |
18525 | if (do_floor && flag_rounding_math) | |
18526 | tmp = ix86_expand_sse_fabs (tmp, NULL); | |
18527 | ||
18528 | ix86_sse_copysign_to_positive (tmp, tmp, res, mask); | |
18529 | } | |
18530 | emit_move_insn (res, tmp); | |
2bf6d935 ML |
18531 | |
18532 | emit_label (label); | |
18533 | LABEL_NUSES (label) = 1; | |
18534 | ||
18535 | emit_move_insn (operand0, res); | |
18536 | } | |
18537 | ||
36d387f2 UB |
18538 | /* Expand SSE2 sequence for computing floor or ceil from OPERAND1 storing |
18539 | into OPERAND0 without relying on DImode truncation via cvttsd2siq | |
18540 | that is only available on 64bit targets. */ | |
2bf6d935 | 18541 | void |
36d387f2 | 18542 | ix86_expand_floorceildf_32 (rtx operand0, rtx operand1, bool do_floor) |
2bf6d935 ML |
18543 | { |
18544 | /* C code for the stuff we expand below. | |
d2754fbb UB |
18545 | double xa = fabs (x), x2; |
18546 | if (!isless (xa, TWO52)) | |
18547 | return x; | |
18548 | xa = xa + TWO52 - TWO52; | |
18549 | x2 = copysign (xa, x); | |
337ed0eb | 18550 | |
36d387f2 | 18551 | Compensate. Floor: |
d2754fbb UB |
18552 | if (x2 > x) |
18553 | x2 -= 1; | |
36d387f2 | 18554 | Compensate. Ceil: |
d2754fbb UB |
18555 | if (x2 < x) |
18556 | x2 += 1; | |
337ed0eb | 18557 | |
36d387f2 UB |
18558 | if (HONOR_SIGNED_ZEROS (mode)) |
18559 | x2 = copysign (x2, x); | |
18560 | return x2; | |
2bf6d935 ML |
18561 | */ |
18562 | machine_mode mode = GET_MODE (operand0); | |
36d387f2 | 18563 | rtx xa, TWO52, tmp, one, res, mask; |
2bf6d935 ML |
18564 | rtx_code_label *label; |
18565 | ||
18566 | TWO52 = ix86_gen_TWO52 (mode); | |
18567 | ||
18568 | /* Temporary for holding the result, initialized to the input | |
18569 | operand to ease control flow. */ | |
d2754fbb | 18570 | res = copy_to_reg (operand1); |
2bf6d935 ML |
18571 | |
18572 | /* xa = abs (operand1) */ | |
18573 | xa = ix86_expand_sse_fabs (res, &mask); | |
18574 | ||
18575 | /* if (!isless (xa, TWO52)) goto label; */ | |
18576 | label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false); | |
18577 | ||
36d387f2 UB |
18578 | /* xa = xa + TWO52 - TWO52; */ |
18579 | xa = expand_simple_binop (mode, PLUS, xa, TWO52, NULL_RTX, 0, OPTAB_DIRECT); | |
18580 | xa = expand_simple_binop (mode, MINUS, xa, TWO52, xa, 0, OPTAB_DIRECT); | |
2bf6d935 | 18581 | |
36d387f2 UB |
18582 | /* xa = copysign (xa, operand1) */ |
18583 | ix86_sse_copysign_to_positive (xa, xa, res, mask); | |
2bf6d935 | 18584 | |
36d387f2 UB |
18585 | /* generate 1.0 */ |
18586 | one = force_reg (mode, const_double_from_real_value (dconst1, mode)); | |
2bf6d935 | 18587 | |
36d387f2 UB |
18588 | /* Compensate: xa = xa - (xa > operand1 ? 1 : 0) */ |
18589 | tmp = ix86_expand_sse_compare_mask (UNGT, xa, res, !do_floor); | |
18590 | emit_insn (gen_rtx_SET (tmp, gen_rtx_AND (mode, one, tmp))); | |
18591 | tmp = expand_simple_binop (mode, do_floor ? MINUS : PLUS, | |
18592 | xa, tmp, NULL_RTX, 0, OPTAB_DIRECT); | |
337ed0eb UB |
18593 | if (HONOR_SIGNED_ZEROS (mode)) |
18594 | { | |
18595 | /* Remove the sign with FE_DOWNWARD, where x - x = -0.0. */ | |
18596 | if (do_floor && flag_rounding_math) | |
18597 | tmp = ix86_expand_sse_fabs (tmp, NULL); | |
18598 | ||
18599 | ix86_sse_copysign_to_positive (tmp, tmp, res, mask); | |
18600 | } | |
36d387f2 | 18601 | emit_move_insn (res, tmp); |
2bf6d935 ML |
18602 | |
18603 | emit_label (label); | |
18604 | LABEL_NUSES (label) = 1; | |
18605 | ||
18606 | emit_move_insn (operand0, res); | |
18607 | } | |
18608 | ||
36d387f2 UB |
18609 | /* Expand SSE sequence for computing trunc |
18610 | from OPERAND1 storing into OPERAND0. */ | |
2bf6d935 ML |
18611 | void |
18612 | ix86_expand_trunc (rtx operand0, rtx operand1) | |
18613 | { | |
18614 | /* C code for SSE variant we expand below. | |
d2754fbb UB |
18615 | double xa = fabs (x), x2; |
18616 | if (!isless (xa, TWO52)) | |
18617 | return x; | |
18618 | x2 = (double)(long)x; | |
2bf6d935 ML |
18619 | if (HONOR_SIGNED_ZEROS (mode)) |
18620 | return copysign (x2, x); | |
18621 | return x2; | |
18622 | */ | |
18623 | machine_mode mode = GET_MODE (operand0); | |
18624 | rtx xa, xi, TWO52, res, mask; | |
18625 | rtx_code_label *label; | |
18626 | ||
18627 | TWO52 = ix86_gen_TWO52 (mode); | |
18628 | ||
18629 | /* Temporary for holding the result, initialized to the input | |
18630 | operand to ease control flow. */ | |
d2754fbb | 18631 | res = copy_to_reg (operand1); |
2bf6d935 ML |
18632 | |
18633 | /* xa = abs (operand1) */ | |
18634 | xa = ix86_expand_sse_fabs (res, &mask); | |
18635 | ||
18636 | /* if (!isless (xa, TWO52)) goto label; */ | |
18637 | label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false); | |
18638 | ||
97d3ddcf | 18639 | /* xa = (double)(long)x */ |
d2754fbb | 18640 | xi = gen_reg_rtx (int_mode_for_mode (mode).require ()); |
2bf6d935 | 18641 | expand_fix (xi, res, 0); |
97d3ddcf | 18642 | expand_float (xa, xi, 0); |
2bf6d935 ML |
18643 | |
18644 | if (HONOR_SIGNED_ZEROS (mode)) | |
97d3ddcf UB |
18645 | ix86_sse_copysign_to_positive (xa, xa, res, mask); |
18646 | ||
18647 | emit_move_insn (res, xa); | |
2bf6d935 ML |
18648 | |
18649 | emit_label (label); | |
18650 | LABEL_NUSES (label) = 1; | |
18651 | ||
18652 | emit_move_insn (operand0, res); | |
18653 | } | |
18654 | ||
18655 | /* Expand SSE sequence for computing trunc from OPERAND1 storing | |
36d387f2 UB |
18656 | into OPERAND0 without relying on DImode truncation via cvttsd2siq |
18657 | that is only available on 64bit targets. */ | |
2bf6d935 ML |
18658 | void |
18659 | ix86_expand_truncdf_32 (rtx operand0, rtx operand1) | |
18660 | { | |
18661 | machine_mode mode = GET_MODE (operand0); | |
c142ae5e | 18662 | rtx xa, xa2, TWO52, tmp, one, res, mask; |
2bf6d935 ML |
18663 | rtx_code_label *label; |
18664 | ||
18665 | /* C code for SSE variant we expand below. | |
d2754fbb UB |
18666 | double xa = fabs (x), x2; |
18667 | if (!isless (xa, TWO52)) | |
18668 | return x; | |
18669 | xa2 = xa + TWO52 - TWO52; | |
2bf6d935 | 18670 | Compensate: |
d2754fbb UB |
18671 | if (xa2 > xa) |
18672 | xa2 -= 1.0; | |
18673 | x2 = copysign (xa2, x); | |
18674 | return x2; | |
2bf6d935 ML |
18675 | */ |
18676 | ||
18677 | TWO52 = ix86_gen_TWO52 (mode); | |
18678 | ||
18679 | /* Temporary for holding the result, initialized to the input | |
18680 | operand to ease control flow. */ | |
d2754fbb | 18681 | res =copy_to_reg (operand1); |
2bf6d935 ML |
18682 | |
18683 | /* xa = abs (operand1) */ | |
c142ae5e | 18684 | xa = ix86_expand_sse_fabs (res, &mask); |
2bf6d935 ML |
18685 | |
18686 | /* if (!isless (xa, TWO52)) goto label; */ | |
18687 | label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false); | |
18688 | ||
c142ae5e UB |
18689 | /* xa2 = xa + TWO52 - TWO52; */ |
18690 | xa2 = expand_simple_binop (mode, PLUS, xa, TWO52, NULL_RTX, 0, OPTAB_DIRECT); | |
18691 | xa2 = expand_simple_binop (mode, MINUS, xa2, TWO52, xa2, 0, OPTAB_DIRECT); | |
2bf6d935 ML |
18692 | |
18693 | /* generate 1.0 */ | |
18694 | one = force_reg (mode, const_double_from_real_value (dconst1, mode)); | |
18695 | ||
c142ae5e UB |
18696 | /* Compensate: xa2 = xa2 - (xa2 > xa ? 1 : 0) */ |
18697 | tmp = ix86_expand_sse_compare_mask (UNGT, xa2, xa, false); | |
18698 | emit_insn (gen_rtx_SET (tmp, gen_rtx_AND (mode, one, tmp))); | |
2bf6d935 | 18699 | tmp = expand_simple_binop (mode, MINUS, |
c142ae5e UB |
18700 | xa2, tmp, NULL_RTX, 0, OPTAB_DIRECT); |
18701 | /* Remove the sign with FE_DOWNWARD, where x - x = -0.0. */ | |
81615bb0 | 18702 | if (HONOR_SIGNED_ZEROS (mode) && flag_rounding_math) |
c142ae5e | 18703 | tmp = ix86_expand_sse_fabs (tmp, NULL); |
2bf6d935 | 18704 | |
c142ae5e UB |
18705 | /* res = copysign (xa2, operand1) */ |
18706 | ix86_sse_copysign_to_positive (res, tmp, res, mask); | |
2bf6d935 ML |
18707 | |
18708 | emit_label (label); | |
18709 | LABEL_NUSES (label) = 1; | |
18710 | ||
18711 | emit_move_insn (operand0, res); | |
18712 | } | |
18713 | ||
36d387f2 UB |
18714 | /* Expand SSE sequence for computing round |
18715 | from OPERAND1 storing into OPERAND0. */ | |
2bf6d935 ML |
18716 | void |
18717 | ix86_expand_round (rtx operand0, rtx operand1) | |
18718 | { | |
18719 | /* C code for the stuff we're doing below: | |
d2754fbb UB |
18720 | double xa = fabs (x); |
18721 | if (!isless (xa, TWO52)) | |
18722 | return x; | |
18723 | xa = (double)(long)(xa + nextafter (0.5, 0.0)); | |
18724 | return copysign (xa, x); | |
2bf6d935 ML |
18725 | */ |
18726 | machine_mode mode = GET_MODE (operand0); | |
18727 | rtx res, TWO52, xa, xi, half, mask; | |
18728 | rtx_code_label *label; | |
18729 | const struct real_format *fmt; | |
18730 | REAL_VALUE_TYPE pred_half, half_minus_pred_half; | |
18731 | ||
18732 | /* Temporary for holding the result, initialized to the input | |
18733 | operand to ease control flow. */ | |
d2754fbb | 18734 | res = copy_to_reg (operand1); |
2bf6d935 ML |
18735 | |
18736 | TWO52 = ix86_gen_TWO52 (mode); | |
18737 | xa = ix86_expand_sse_fabs (res, &mask); | |
18738 | label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false); | |
18739 | ||
18740 | /* load nextafter (0.5, 0.0) */ | |
18741 | fmt = REAL_MODE_FORMAT (mode); | |
18742 | real_2expN (&half_minus_pred_half, -(fmt->p) - 1, mode); | |
18743 | real_arithmetic (&pred_half, MINUS_EXPR, &dconsthalf, &half_minus_pred_half); | |
18744 | ||
18745 | /* xa = xa + 0.5 */ | |
18746 | half = force_reg (mode, const_double_from_real_value (pred_half, mode)); | |
18747 | xa = expand_simple_binop (mode, PLUS, xa, half, NULL_RTX, 0, OPTAB_DIRECT); | |
18748 | ||
18749 | /* xa = (double)(int64_t)xa */ | |
d2754fbb | 18750 | xi = gen_reg_rtx (int_mode_for_mode (mode).require ()); |
2bf6d935 ML |
18751 | expand_fix (xi, xa, 0); |
18752 | expand_float (xa, xi, 0); | |
18753 | ||
18754 | /* res = copysign (xa, operand1) */ | |
97d3ddcf | 18755 | ix86_sse_copysign_to_positive (res, xa, res, mask); |
2bf6d935 ML |
18756 | |
18757 | emit_label (label); | |
18758 | LABEL_NUSES (label) = 1; | |
18759 | ||
18760 | emit_move_insn (operand0, res); | |
18761 | } | |
18762 | ||
36d387f2 UB |
18763 | /* Expand SSE sequence for computing round from OPERAND1 storing |
18764 | into OPERAND0 without relying on DImode truncation via cvttsd2siq | |
18765 | that is only available on 64bit targets. */ | |
18766 | void | |
18767 | ix86_expand_rounddf_32 (rtx operand0, rtx operand1) | |
18768 | { | |
18769 | /* C code for the stuff we expand below. | |
d2754fbb UB |
18770 | double xa = fabs (x), xa2, x2; |
18771 | if (!isless (xa, TWO52)) | |
18772 | return x; | |
36d387f2 UB |
18773 | Using the absolute value and copying back sign makes |
18774 | -0.0 -> -0.0 correct. | |
d2754fbb | 18775 | xa2 = xa + TWO52 - TWO52; |
36d387f2 UB |
18776 | Compensate. |
18777 | dxa = xa2 - xa; | |
d2754fbb UB |
18778 | if (dxa <= -0.5) |
18779 | xa2 += 1; | |
18780 | else if (dxa > 0.5) | |
18781 | xa2 -= 1; | |
18782 | x2 = copysign (xa2, x); | |
18783 | return x2; | |
36d387f2 UB |
18784 | */ |
18785 | machine_mode mode = GET_MODE (operand0); | |
18786 | rtx xa, xa2, dxa, TWO52, tmp, half, mhalf, one, res, mask; | |
18787 | rtx_code_label *label; | |
18788 | ||
18789 | TWO52 = ix86_gen_TWO52 (mode); | |
18790 | ||
18791 | /* Temporary for holding the result, initialized to the input | |
18792 | operand to ease control flow. */ | |
d2754fbb | 18793 | res = copy_to_reg (operand1); |
36d387f2 UB |
18794 | |
18795 | /* xa = abs (operand1) */ | |
18796 | xa = ix86_expand_sse_fabs (res, &mask); | |
18797 | ||
18798 | /* if (!isless (xa, TWO52)) goto label; */ | |
18799 | label = ix86_expand_sse_compare_and_jump (UNLE, TWO52, xa, false); | |
18800 | ||
18801 | /* xa2 = xa + TWO52 - TWO52; */ | |
18802 | xa2 = expand_simple_binop (mode, PLUS, xa, TWO52, NULL_RTX, 0, OPTAB_DIRECT); | |
18803 | xa2 = expand_simple_binop (mode, MINUS, xa2, TWO52, xa2, 0, OPTAB_DIRECT); | |
18804 | ||
18805 | /* dxa = xa2 - xa; */ | |
18806 | dxa = expand_simple_binop (mode, MINUS, xa2, xa, NULL_RTX, 0, OPTAB_DIRECT); | |
18807 | ||
18808 | /* generate 0.5, 1.0 and -0.5 */ | |
18809 | half = force_reg (mode, const_double_from_real_value (dconsthalf, mode)); | |
18810 | one = expand_simple_binop (mode, PLUS, half, half, NULL_RTX, 0, OPTAB_DIRECT); | |
18811 | mhalf = expand_simple_binop (mode, MINUS, half, one, NULL_RTX, | |
18812 | 0, OPTAB_DIRECT); | |
18813 | ||
18814 | /* Compensate. */ | |
18815 | /* xa2 = xa2 - (dxa > 0.5 ? 1 : 0) */ | |
18816 | tmp = ix86_expand_sse_compare_mask (UNGT, dxa, half, false); | |
18817 | emit_insn (gen_rtx_SET (tmp, gen_rtx_AND (mode, tmp, one))); | |
18818 | xa2 = expand_simple_binop (mode, MINUS, xa2, tmp, NULL_RTX, 0, OPTAB_DIRECT); | |
18819 | /* xa2 = xa2 + (dxa <= -0.5 ? 1 : 0) */ | |
18820 | tmp = ix86_expand_sse_compare_mask (UNGE, mhalf, dxa, false); | |
18821 | emit_insn (gen_rtx_SET (tmp, gen_rtx_AND (mode, tmp, one))); | |
18822 | xa2 = expand_simple_binop (mode, PLUS, xa2, tmp, NULL_RTX, 0, OPTAB_DIRECT); | |
18823 | ||
18824 | /* res = copysign (xa2, operand1) */ | |
97d3ddcf | 18825 | ix86_sse_copysign_to_positive (res, xa2, res, mask); |
36d387f2 UB |
18826 | |
18827 | emit_label (label); | |
18828 | LABEL_NUSES (label) = 1; | |
18829 | ||
18830 | emit_move_insn (operand0, res); | |
18831 | } | |
18832 | ||
2bf6d935 ML |
18833 | /* Expand SSE sequence for computing round |
18834 | from OP1 storing into OP0 using sse4 round insn. */ | |
18835 | void | |
18836 | ix86_expand_round_sse4 (rtx op0, rtx op1) | |
18837 | { | |
18838 | machine_mode mode = GET_MODE (op0); | |
18839 | rtx e1, e2, res, half; | |
18840 | const struct real_format *fmt; | |
18841 | REAL_VALUE_TYPE pred_half, half_minus_pred_half; | |
18842 | rtx (*gen_copysign) (rtx, rtx, rtx); | |
18843 | rtx (*gen_round) (rtx, rtx, rtx); | |
18844 | ||
18845 | switch (mode) | |
18846 | { | |
18847 | case E_SFmode: | |
18848 | gen_copysign = gen_copysignsf3; | |
18849 | gen_round = gen_sse4_1_roundsf2; | |
18850 | break; | |
18851 | case E_DFmode: | |
18852 | gen_copysign = gen_copysigndf3; | |
18853 | gen_round = gen_sse4_1_rounddf2; | |
18854 | break; | |
18855 | default: | |
18856 | gcc_unreachable (); | |
18857 | } | |
18858 | ||
18859 | /* round (a) = trunc (a + copysign (0.5, a)) */ | |
18860 | ||
18861 | /* load nextafter (0.5, 0.0) */ | |
18862 | fmt = REAL_MODE_FORMAT (mode); | |
18863 | real_2expN (&half_minus_pred_half, -(fmt->p) - 1, mode); | |
18864 | real_arithmetic (&pred_half, MINUS_EXPR, &dconsthalf, &half_minus_pred_half); | |
18865 | half = const_double_from_real_value (pred_half, mode); | |
18866 | ||
18867 | /* e1 = copysign (0.5, op1) */ | |
18868 | e1 = gen_reg_rtx (mode); | |
18869 | emit_insn (gen_copysign (e1, half, op1)); | |
18870 | ||
18871 | /* e2 = op1 + e1 */ | |
18872 | e2 = expand_simple_binop (mode, PLUS, op1, e1, NULL_RTX, 0, OPTAB_DIRECT); | |
18873 | ||
18874 | /* res = trunc (e2) */ | |
18875 | res = gen_reg_rtx (mode); | |
18876 | emit_insn (gen_round (res, e2, GEN_INT (ROUND_TRUNC))); | |
18877 | ||
18878 | emit_move_insn (op0, res); | |
18879 | } | |
18880 | ||
18881 | /* A cached (set (nil) (vselect (vconcat (nil) (nil)) (parallel []))) | |
18882 | insn, so that expand_vselect{,_vconcat} doesn't have to create a fresh | |
18883 | insn every time. */ | |
18884 | ||
18885 | static GTY(()) rtx_insn *vselect_insn; | |
18886 | ||
18887 | /* Initialize vselect_insn. */ | |
18888 | ||
18889 | static void | |
18890 | init_vselect_insn (void) | |
18891 | { | |
18892 | unsigned i; | |
18893 | rtx x; | |
18894 | ||
18895 | x = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (MAX_VECT_LEN)); | |
18896 | for (i = 0; i < MAX_VECT_LEN; ++i) | |
18897 | XVECEXP (x, 0, i) = const0_rtx; | |
18898 | x = gen_rtx_VEC_SELECT (V2DFmode, gen_rtx_VEC_CONCAT (V4DFmode, const0_rtx, | |
18899 | const0_rtx), x); | |
18900 | x = gen_rtx_SET (const0_rtx, x); | |
18901 | start_sequence (); | |
18902 | vselect_insn = emit_insn (x); | |
18903 | end_sequence (); | |
18904 | } | |
18905 | ||
18906 | /* Construct (set target (vec_select op0 (parallel perm))) and | |
18907 | return true if that's a valid instruction in the active ISA. */ | |
18908 | ||
18909 | static bool | |
18910 | expand_vselect (rtx target, rtx op0, const unsigned char *perm, | |
18911 | unsigned nelt, bool testing_p) | |
18912 | { | |
18913 | unsigned int i; | |
18914 | rtx x, save_vconcat; | |
18915 | int icode; | |
18916 | ||
18917 | if (vselect_insn == NULL_RTX) | |
18918 | init_vselect_insn (); | |
18919 | ||
18920 | x = XEXP (SET_SRC (PATTERN (vselect_insn)), 1); | |
18921 | PUT_NUM_ELEM (XVEC (x, 0), nelt); | |
18922 | for (i = 0; i < nelt; ++i) | |
18923 | XVECEXP (x, 0, i) = GEN_INT (perm[i]); | |
18924 | save_vconcat = XEXP (SET_SRC (PATTERN (vselect_insn)), 0); | |
18925 | XEXP (SET_SRC (PATTERN (vselect_insn)), 0) = op0; | |
18926 | PUT_MODE (SET_SRC (PATTERN (vselect_insn)), GET_MODE (target)); | |
18927 | SET_DEST (PATTERN (vselect_insn)) = target; | |
18928 | icode = recog_memoized (vselect_insn); | |
18929 | ||
18930 | if (icode >= 0 && !testing_p) | |
18931 | emit_insn (copy_rtx (PATTERN (vselect_insn))); | |
18932 | ||
18933 | SET_DEST (PATTERN (vselect_insn)) = const0_rtx; | |
18934 | XEXP (SET_SRC (PATTERN (vselect_insn)), 0) = save_vconcat; | |
18935 | INSN_CODE (vselect_insn) = -1; | |
18936 | ||
18937 | return icode >= 0; | |
18938 | } | |
18939 | ||
18940 | /* Similar, but generate a vec_concat from op0 and op1 as well. */ | |
18941 | ||
18942 | static bool | |
18943 | expand_vselect_vconcat (rtx target, rtx op0, rtx op1, | |
18944 | const unsigned char *perm, unsigned nelt, | |
18945 | bool testing_p) | |
18946 | { | |
18947 | machine_mode v2mode; | |
18948 | rtx x; | |
18949 | bool ok; | |
18950 | ||
18951 | if (vselect_insn == NULL_RTX) | |
18952 | init_vselect_insn (); | |
18953 | ||
18954 | if (!GET_MODE_2XWIDER_MODE (GET_MODE (op0)).exists (&v2mode)) | |
18955 | return false; | |
18956 | x = XEXP (SET_SRC (PATTERN (vselect_insn)), 0); | |
18957 | PUT_MODE (x, v2mode); | |
18958 | XEXP (x, 0) = op0; | |
18959 | XEXP (x, 1) = op1; | |
18960 | ok = expand_vselect (target, x, perm, nelt, testing_p); | |
18961 | XEXP (x, 0) = const0_rtx; | |
18962 | XEXP (x, 1) = const0_rtx; | |
18963 | return ok; | |
18964 | } | |
18965 | ||
4bf4c103 | 18966 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to implement D |
2bf6d935 ML |
18967 | using movss or movsd. */ |
18968 | static bool | |
18969 | expand_vec_perm_movs (struct expand_vec_perm_d *d) | |
18970 | { | |
18971 | machine_mode vmode = d->vmode; | |
18972 | unsigned i, nelt = d->nelt; | |
18973 | rtx x; | |
18974 | ||
18975 | if (d->one_operand_p) | |
18976 | return false; | |
18977 | ||
4bcc3b5c UB |
18978 | if (!(TARGET_SSE && (vmode == V4SFmode || vmode == V4SImode)) |
18979 | && !(TARGET_MMX_WITH_SSE && (vmode == V2SFmode || vmode == V2SImode)) | |
18980 | && !(TARGET_SSE2 && (vmode == V2DFmode || vmode == V2DImode))) | |
2bf6d935 ML |
18981 | return false; |
18982 | ||
18983 | /* Only the first element is changed. */ | |
18984 | if (d->perm[0] != nelt && d->perm[0] != 0) | |
18985 | return false; | |
18986 | for (i = 1; i < nelt; ++i) | |
18987 | if (d->perm[i] != i + nelt - d->perm[0]) | |
18988 | return false; | |
18989 | ||
18990 | if (d->testing_p) | |
18991 | return true; | |
18992 | ||
18993 | if (d->perm[0] == nelt) | |
18994 | x = gen_rtx_VEC_MERGE (vmode, d->op1, d->op0, GEN_INT (1)); | |
18995 | else | |
18996 | x = gen_rtx_VEC_MERGE (vmode, d->op0, d->op1, GEN_INT (1)); | |
18997 | ||
18998 | emit_insn (gen_rtx_SET (d->target, x)); | |
18999 | ||
19000 | return true; | |
19001 | } | |
19002 | ||
95b99e47 UB |
19003 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to implement D |
19004 | using insertps. */ | |
19005 | static bool | |
19006 | expand_vec_perm_insertps (struct expand_vec_perm_d *d) | |
19007 | { | |
19008 | machine_mode vmode = d->vmode; | |
19009 | unsigned i, cnt_s, nelt = d->nelt; | |
19010 | int cnt_d = -1; | |
19011 | rtx src, dst; | |
19012 | ||
19013 | if (d->one_operand_p) | |
19014 | return false; | |
19015 | ||
19016 | if (!(TARGET_SSE4_1 | |
19017 | && (vmode == V4SFmode || vmode == V4SImode | |
19018 | || (TARGET_MMX_WITH_SSE | |
19019 | && (vmode == V2SFmode || vmode == V2SImode))))) | |
19020 | return false; | |
19021 | ||
19022 | for (i = 0; i < nelt; ++i) | |
19023 | { | |
19024 | if (d->perm[i] == i) | |
19025 | continue; | |
19026 | if (cnt_d != -1) | |
19027 | { | |
19028 | cnt_d = -1; | |
19029 | break; | |
19030 | } | |
19031 | cnt_d = i; | |
19032 | } | |
19033 | ||
19034 | if (cnt_d == -1) | |
19035 | { | |
19036 | for (i = 0; i < nelt; ++i) | |
19037 | { | |
19038 | if (d->perm[i] == i + nelt) | |
19039 | continue; | |
19040 | if (cnt_d != -1) | |
19041 | return false; | |
19042 | cnt_d = i; | |
19043 | } | |
19044 | ||
19045 | if (cnt_d == -1) | |
19046 | return false; | |
19047 | } | |
19048 | ||
19049 | if (d->testing_p) | |
19050 | return true; | |
19051 | ||
19052 | gcc_assert (cnt_d != -1); | |
19053 | ||
19054 | cnt_s = d->perm[cnt_d]; | |
19055 | if (cnt_s < nelt) | |
19056 | { | |
19057 | src = d->op0; | |
19058 | dst = d->op1; | |
19059 | } | |
19060 | else | |
19061 | { | |
19062 | cnt_s -= nelt; | |
19063 | src = d->op1; | |
19064 | dst = d->op0; | |
19065 | } | |
19066 | gcc_assert (cnt_s < nelt); | |
19067 | ||
19068 | rtx x = gen_sse4_1_insertps (vmode, d->target, dst, src, | |
19069 | GEN_INT (cnt_s << 6 | cnt_d << 4)); | |
19070 | emit_insn (x); | |
19071 | ||
19072 | return true; | |
19073 | } | |
19074 | ||
4bf4c103 | 19075 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to implement D |
2bf6d935 ML |
19076 | in terms of blendp[sd] / pblendw / pblendvb / vpblendd. */ |
19077 | ||
19078 | static bool | |
19079 | expand_vec_perm_blend (struct expand_vec_perm_d *d) | |
19080 | { | |
19081 | machine_mode mmode, vmode = d->vmode; | |
fa2987ed JJ |
19082 | unsigned i, nelt = d->nelt; |
19083 | unsigned HOST_WIDE_INT mask; | |
2bf6d935 ML |
19084 | rtx target, op0, op1, maskop, x; |
19085 | rtx rperm[32], vperm; | |
19086 | ||
19087 | if (d->one_operand_p) | |
19088 | return false; | |
19089 | if (TARGET_AVX512F && GET_MODE_SIZE (vmode) == 64 | |
19090 | && (TARGET_AVX512BW | |
19091 | || GET_MODE_UNIT_SIZE (vmode) >= 4)) | |
19092 | ; | |
19093 | else if (TARGET_AVX2 && GET_MODE_SIZE (vmode) == 32) | |
19094 | ; | |
19095 | else if (TARGET_AVX && (vmode == V4DFmode || vmode == V8SFmode)) | |
19096 | ; | |
57052c6e UB |
19097 | else if (TARGET_SSE4_1 |
19098 | && (GET_MODE_SIZE (vmode) == 16 | |
19099 | || (TARGET_MMX_WITH_SSE && GET_MODE_SIZE (vmode) == 8) | |
19100 | || GET_MODE_SIZE (vmode) == 4)) | |
2bf6d935 ML |
19101 | ; |
19102 | else | |
19103 | return false; | |
19104 | ||
19105 | /* This is a blend, not a permute. Elements must stay in their | |
19106 | respective lanes. */ | |
19107 | for (i = 0; i < nelt; ++i) | |
19108 | { | |
19109 | unsigned e = d->perm[i]; | |
19110 | if (!(e == i || e == i + nelt)) | |
19111 | return false; | |
19112 | } | |
19113 | ||
19114 | if (d->testing_p) | |
19115 | return true; | |
19116 | ||
19117 | /* ??? Without SSE4.1, we could implement this with and/andn/or. This | |
19118 | decision should be extracted elsewhere, so that we only try that | |
19119 | sequence once all budget==3 options have been tried. */ | |
19120 | target = d->target; | |
19121 | op0 = d->op0; | |
19122 | op1 = d->op1; | |
19123 | mask = 0; | |
19124 | ||
19125 | switch (vmode) | |
19126 | { | |
19127 | case E_V8DFmode: | |
19128 | case E_V16SFmode: | |
19129 | case E_V4DFmode: | |
19130 | case E_V8SFmode: | |
19131 | case E_V2DFmode: | |
19132 | case E_V4SFmode: | |
57052c6e UB |
19133 | case E_V2SFmode: |
19134 | case E_V2HImode: | |
a325bdd1 | 19135 | case E_V4HImode: |
2bf6d935 ML |
19136 | case E_V8HImode: |
19137 | case E_V8SImode: | |
19138 | case E_V32HImode: | |
19139 | case E_V64QImode: | |
19140 | case E_V16SImode: | |
19141 | case E_V8DImode: | |
19142 | for (i = 0; i < nelt; ++i) | |
fa2987ed | 19143 | mask |= ((unsigned HOST_WIDE_INT) (d->perm[i] >= nelt)) << i; |
2bf6d935 ML |
19144 | break; |
19145 | ||
19146 | case E_V2DImode: | |
19147 | for (i = 0; i < 2; ++i) | |
19148 | mask |= (d->perm[i] >= 2 ? 15 : 0) << (i * 4); | |
19149 | vmode = V8HImode; | |
19150 | goto do_subreg; | |
19151 | ||
a325bdd1 PB |
19152 | case E_V2SImode: |
19153 | for (i = 0; i < 2; ++i) | |
19154 | mask |= (d->perm[i] >= 2 ? 3 : 0) << (i * 2); | |
19155 | vmode = V4HImode; | |
19156 | goto do_subreg; | |
19157 | ||
2bf6d935 | 19158 | case E_V4SImode: |
3588c8cb | 19159 | if (TARGET_AVX2) |
19160 | { | |
19161 | /* Use vpblendd instead of vpblendw. */ | |
19162 | for (i = 0; i < nelt; ++i) | |
19163 | mask |= ((unsigned HOST_WIDE_INT) (d->perm[i] >= nelt)) << i; | |
19164 | break; | |
19165 | } | |
19166 | else | |
19167 | { | |
19168 | for (i = 0; i < 4; ++i) | |
19169 | mask |= (d->perm[i] >= 4 ? 3 : 0) << (i * 2); | |
19170 | vmode = V8HImode; | |
19171 | goto do_subreg; | |
19172 | } | |
2bf6d935 ML |
19173 | |
19174 | case E_V16QImode: | |
19175 | /* See if bytes move in pairs so we can use pblendw with | |
19176 | an immediate argument, rather than pblendvb with a vector | |
19177 | argument. */ | |
19178 | for (i = 0; i < 16; i += 2) | |
19179 | if (d->perm[i] + 1 != d->perm[i + 1]) | |
19180 | { | |
19181 | use_pblendvb: | |
19182 | for (i = 0; i < nelt; ++i) | |
19183 | rperm[i] = (d->perm[i] < nelt ? const0_rtx : constm1_rtx); | |
19184 | ||
19185 | finish_pblendvb: | |
19186 | vperm = gen_rtx_CONST_VECTOR (vmode, gen_rtvec_v (nelt, rperm)); | |
19187 | vperm = force_reg (vmode, vperm); | |
19188 | ||
be8749f9 | 19189 | if (GET_MODE_SIZE (vmode) == 4) |
820ac79e | 19190 | emit_insn (gen_mmx_pblendvb_v4qi (target, op0, op1, vperm)); |
be8749f9 | 19191 | else if (GET_MODE_SIZE (vmode) == 8) |
820ac79e | 19192 | emit_insn (gen_mmx_pblendvb_v8qi (target, op0, op1, vperm)); |
a325bdd1 | 19193 | else if (GET_MODE_SIZE (vmode) == 16) |
2bf6d935 ML |
19194 | emit_insn (gen_sse4_1_pblendvb (target, op0, op1, vperm)); |
19195 | else | |
19196 | emit_insn (gen_avx2_pblendvb (target, op0, op1, vperm)); | |
19197 | if (target != d->target) | |
19198 | emit_move_insn (d->target, gen_lowpart (d->vmode, target)); | |
19199 | return true; | |
19200 | } | |
19201 | ||
19202 | for (i = 0; i < 8; ++i) | |
19203 | mask |= (d->perm[i * 2] >= 16) << i; | |
19204 | vmode = V8HImode; | |
19205 | /* FALLTHRU */ | |
19206 | ||
19207 | do_subreg: | |
19208 | target = gen_reg_rtx (vmode); | |
19209 | op0 = gen_lowpart (vmode, op0); | |
19210 | op1 = gen_lowpart (vmode, op1); | |
19211 | break; | |
19212 | ||
a325bdd1 PB |
19213 | case E_V8QImode: |
19214 | for (i = 0; i < 8; i += 2) | |
19215 | if (d->perm[i] + 1 != d->perm[i + 1]) | |
19216 | goto use_pblendvb; | |
19217 | ||
19218 | for (i = 0; i < 4; ++i) | |
19219 | mask |= (d->perm[i * 2] >= 8) << i; | |
19220 | vmode = V4HImode; | |
19221 | goto do_subreg; | |
19222 | ||
be8749f9 UB |
19223 | case E_V4QImode: |
19224 | for (i = 0; i < 4; i += 2) | |
19225 | if (d->perm[i] + 1 != d->perm[i + 1]) | |
19226 | goto use_pblendvb; | |
19227 | ||
19228 | for (i = 0; i < 2; ++i) | |
19229 | mask |= (d->perm[i * 2] >= 4) << i; | |
19230 | vmode = V2HImode; | |
19231 | goto do_subreg; | |
19232 | ||
2bf6d935 ML |
19233 | case E_V32QImode: |
19234 | /* See if bytes move in pairs. If not, vpblendvb must be used. */ | |
19235 | for (i = 0; i < 32; i += 2) | |
19236 | if (d->perm[i] + 1 != d->perm[i + 1]) | |
19237 | goto use_pblendvb; | |
19238 | /* See if bytes move in quadruplets. If yes, vpblendd | |
19239 | with immediate can be used. */ | |
19240 | for (i = 0; i < 32; i += 4) | |
19241 | if (d->perm[i] + 2 != d->perm[i + 2]) | |
19242 | break; | |
19243 | if (i < 32) | |
19244 | { | |
19245 | /* See if bytes move the same in both lanes. If yes, | |
19246 | vpblendw with immediate can be used. */ | |
19247 | for (i = 0; i < 16; i += 2) | |
19248 | if (d->perm[i] + 16 != d->perm[i + 16]) | |
19249 | goto use_pblendvb; | |
19250 | ||
19251 | /* Use vpblendw. */ | |
19252 | for (i = 0; i < 16; ++i) | |
19253 | mask |= (d->perm[i * 2] >= 32) << i; | |
19254 | vmode = V16HImode; | |
19255 | goto do_subreg; | |
19256 | } | |
19257 | ||
19258 | /* Use vpblendd. */ | |
19259 | for (i = 0; i < 8; ++i) | |
19260 | mask |= (d->perm[i * 4] >= 32) << i; | |
19261 | vmode = V8SImode; | |
19262 | goto do_subreg; | |
19263 | ||
19264 | case E_V16HImode: | |
19265 | /* See if words move in pairs. If yes, vpblendd can be used. */ | |
19266 | for (i = 0; i < 16; i += 2) | |
19267 | if (d->perm[i] + 1 != d->perm[i + 1]) | |
19268 | break; | |
19269 | if (i < 16) | |
19270 | { | |
19271 | /* See if words move the same in both lanes. If not, | |
19272 | vpblendvb must be used. */ | |
19273 | for (i = 0; i < 8; i++) | |
19274 | if (d->perm[i] + 8 != d->perm[i + 8]) | |
19275 | { | |
19276 | /* Use vpblendvb. */ | |
19277 | for (i = 0; i < 32; ++i) | |
19278 | rperm[i] = (d->perm[i / 2] < 16 ? const0_rtx : constm1_rtx); | |
19279 | ||
19280 | vmode = V32QImode; | |
19281 | nelt = 32; | |
19282 | target = gen_reg_rtx (vmode); | |
19283 | op0 = gen_lowpart (vmode, op0); | |
19284 | op1 = gen_lowpart (vmode, op1); | |
19285 | goto finish_pblendvb; | |
19286 | } | |
19287 | ||
19288 | /* Use vpblendw. */ | |
19289 | for (i = 0; i < 16; ++i) | |
19290 | mask |= (d->perm[i] >= 16) << i; | |
19291 | break; | |
19292 | } | |
19293 | ||
19294 | /* Use vpblendd. */ | |
19295 | for (i = 0; i < 8; ++i) | |
19296 | mask |= (d->perm[i * 2] >= 16) << i; | |
19297 | vmode = V8SImode; | |
19298 | goto do_subreg; | |
19299 | ||
19300 | case E_V4DImode: | |
19301 | /* Use vpblendd. */ | |
19302 | for (i = 0; i < 4; ++i) | |
19303 | mask |= (d->perm[i] >= 4 ? 3 : 0) << (i * 2); | |
19304 | vmode = V8SImode; | |
19305 | goto do_subreg; | |
19306 | ||
19307 | default: | |
19308 | gcc_unreachable (); | |
19309 | } | |
19310 | ||
19311 | switch (vmode) | |
19312 | { | |
19313 | case E_V8DFmode: | |
19314 | case E_V8DImode: | |
19315 | mmode = QImode; | |
19316 | break; | |
19317 | case E_V16SFmode: | |
19318 | case E_V16SImode: | |
19319 | mmode = HImode; | |
19320 | break; | |
19321 | case E_V32HImode: | |
19322 | mmode = SImode; | |
19323 | break; | |
19324 | case E_V64QImode: | |
19325 | mmode = DImode; | |
19326 | break; | |
19327 | default: | |
19328 | mmode = VOIDmode; | |
19329 | } | |
19330 | ||
19331 | if (mmode != VOIDmode) | |
19332 | maskop = force_reg (mmode, gen_int_mode (mask, mmode)); | |
19333 | else | |
19334 | maskop = GEN_INT (mask); | |
19335 | ||
19336 | /* This matches five different patterns with the different modes. */ | |
19337 | x = gen_rtx_VEC_MERGE (vmode, op1, op0, maskop); | |
19338 | x = gen_rtx_SET (target, x); | |
19339 | emit_insn (x); | |
19340 | if (target != d->target) | |
19341 | emit_move_insn (d->target, gen_lowpart (d->vmode, target)); | |
19342 | ||
19343 | return true; | |
19344 | } | |
19345 | ||
4bf4c103 | 19346 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to implement D |
2bf6d935 ML |
19347 | in terms of the variable form of vpermilps. |
19348 | ||
19349 | Note that we will have already failed the immediate input vpermilps, | |
19350 | which requires that the high and low part shuffle be identical; the | |
19351 | variable form doesn't require that. */ | |
19352 | ||
19353 | static bool | |
19354 | expand_vec_perm_vpermil (struct expand_vec_perm_d *d) | |
19355 | { | |
19356 | rtx rperm[8], vperm; | |
19357 | unsigned i; | |
19358 | ||
19359 | if (!TARGET_AVX || d->vmode != V8SFmode || !d->one_operand_p) | |
19360 | return false; | |
19361 | ||
19362 | /* We can only permute within the 128-bit lane. */ | |
19363 | for (i = 0; i < 8; ++i) | |
19364 | { | |
19365 | unsigned e = d->perm[i]; | |
19366 | if (i < 4 ? e >= 4 : e < 4) | |
19367 | return false; | |
19368 | } | |
19369 | ||
19370 | if (d->testing_p) | |
19371 | return true; | |
19372 | ||
19373 | for (i = 0; i < 8; ++i) | |
19374 | { | |
19375 | unsigned e = d->perm[i]; | |
19376 | ||
19377 | /* Within each 128-bit lane, the elements of op0 are numbered | |
19378 | from 0 and the elements of op1 are numbered from 4. */ | |
19379 | if (e >= 8 + 4) | |
19380 | e -= 8; | |
19381 | else if (e >= 4) | |
19382 | e -= 4; | |
19383 | ||
19384 | rperm[i] = GEN_INT (e); | |
19385 | } | |
19386 | ||
19387 | vperm = gen_rtx_CONST_VECTOR (V8SImode, gen_rtvec_v (8, rperm)); | |
19388 | vperm = force_reg (V8SImode, vperm); | |
19389 | emit_insn (gen_avx_vpermilvarv8sf3 (d->target, d->op0, vperm)); | |
19390 | ||
19391 | return true; | |
19392 | } | |
19393 | ||
1fa991d1 UB |
19394 | /* For V*[QHS]Imode permutations, check if the same permutation |
19395 | can't be performed in a 2x, 4x or 8x wider inner mode. */ | |
19396 | ||
19397 | static bool | |
19398 | canonicalize_vector_int_perm (const struct expand_vec_perm_d *d, | |
19399 | struct expand_vec_perm_d *nd) | |
19400 | { | |
19401 | int i; | |
19402 | machine_mode mode = VOIDmode; | |
19403 | ||
19404 | switch (d->vmode) | |
19405 | { | |
19406 | case E_V8QImode: mode = V4HImode; break; | |
19407 | case E_V16QImode: mode = V8HImode; break; | |
19408 | case E_V32QImode: mode = V16HImode; break; | |
19409 | case E_V64QImode: mode = V32HImode; break; | |
19410 | case E_V4HImode: mode = V2SImode; break; | |
19411 | case E_V8HImode: mode = V4SImode; break; | |
19412 | case E_V16HImode: mode = V8SImode; break; | |
19413 | case E_V32HImode: mode = V16SImode; break; | |
19414 | case E_V4SImode: mode = V2DImode; break; | |
19415 | case E_V8SImode: mode = V4DImode; break; | |
19416 | case E_V16SImode: mode = V8DImode; break; | |
19417 | default: return false; | |
19418 | } | |
19419 | for (i = 0; i < d->nelt; i += 2) | |
19420 | if ((d->perm[i] & 1) || d->perm[i + 1] != d->perm[i] + 1) | |
19421 | return false; | |
19422 | nd->vmode = mode; | |
19423 | nd->nelt = d->nelt / 2; | |
19424 | for (i = 0; i < nd->nelt; i++) | |
19425 | nd->perm[i] = d->perm[2 * i] / 2; | |
19426 | if (GET_MODE_INNER (mode) != DImode) | |
19427 | canonicalize_vector_int_perm (nd, nd); | |
19428 | if (nd != d) | |
19429 | { | |
19430 | nd->one_operand_p = d->one_operand_p; | |
19431 | nd->testing_p = d->testing_p; | |
19432 | if (d->op0 == d->op1) | |
19433 | nd->op0 = nd->op1 = gen_lowpart (nd->vmode, d->op0); | |
19434 | else | |
19435 | { | |
19436 | nd->op0 = gen_lowpart (nd->vmode, d->op0); | |
19437 | nd->op1 = gen_lowpart (nd->vmode, d->op1); | |
19438 | } | |
19439 | if (d->testing_p) | |
19440 | nd->target = gen_raw_REG (nd->vmode, LAST_VIRTUAL_REGISTER + 1); | |
19441 | else | |
19442 | nd->target = gen_reg_rtx (nd->vmode); | |
19443 | } | |
19444 | return true; | |
19445 | } | |
19446 | ||
2bf6d935 ML |
19447 | /* Return true if permutation D can be performed as VMODE permutation |
19448 | instead. */ | |
19449 | ||
19450 | static bool | |
19451 | valid_perm_using_mode_p (machine_mode vmode, struct expand_vec_perm_d *d) | |
19452 | { | |
19453 | unsigned int i, j, chunk; | |
19454 | ||
19455 | if (GET_MODE_CLASS (vmode) != MODE_VECTOR_INT | |
19456 | || GET_MODE_CLASS (d->vmode) != MODE_VECTOR_INT | |
19457 | || GET_MODE_SIZE (vmode) != GET_MODE_SIZE (d->vmode)) | |
19458 | return false; | |
19459 | ||
19460 | if (GET_MODE_NUNITS (vmode) >= d->nelt) | |
19461 | return true; | |
19462 | ||
19463 | chunk = d->nelt / GET_MODE_NUNITS (vmode); | |
19464 | for (i = 0; i < d->nelt; i += chunk) | |
19465 | if (d->perm[i] & (chunk - 1)) | |
19466 | return false; | |
19467 | else | |
19468 | for (j = 1; j < chunk; ++j) | |
19469 | if (d->perm[i] + j != d->perm[i + j]) | |
19470 | return false; | |
19471 | ||
19472 | return true; | |
19473 | } | |
19474 | ||
4bf4c103 | 19475 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to implement D |
2bf6d935 ML |
19476 | in terms of pshufb, vpperm, vpermq, vpermd, vpermps or vperm2i128. */ |
19477 | ||
19478 | static bool | |
19479 | expand_vec_perm_pshufb (struct expand_vec_perm_d *d) | |
19480 | { | |
19481 | unsigned i, nelt, eltsz, mask; | |
19482 | unsigned char perm[64]; | |
877c9e33 | 19483 | machine_mode vmode; |
1fa991d1 | 19484 | struct expand_vec_perm_d nd; |
2bf6d935 ML |
19485 | rtx rperm[64], vperm, target, op0, op1; |
19486 | ||
19487 | nelt = d->nelt; | |
19488 | ||
19489 | if (!d->one_operand_p) | |
be8749f9 UB |
19490 | switch (GET_MODE_SIZE (d->vmode)) |
19491 | { | |
19492 | case 4: | |
19493 | if (!TARGET_XOP) | |
19494 | return false; | |
19495 | vmode = V4QImode; | |
19496 | break; | |
37e93925 | 19497 | |
be8749f9 UB |
19498 | case 8: |
19499 | if (!TARGET_XOP) | |
19500 | return false; | |
19501 | vmode = V8QImode; | |
19502 | break; | |
2bf6d935 | 19503 | |
be8749f9 UB |
19504 | case 16: |
19505 | if (!TARGET_XOP) | |
2bf6d935 | 19506 | return false; |
877c9e33 | 19507 | vmode = V16QImode; |
be8749f9 UB |
19508 | break; |
19509 | ||
19510 | case 32: | |
19511 | if (!TARGET_AVX2) | |
19512 | return false; | |
19513 | ||
19514 | if (valid_perm_using_mode_p (V2TImode, d)) | |
19515 | { | |
19516 | if (d->testing_p) | |
19517 | return true; | |
19518 | ||
19519 | /* Use vperm2i128 insn. The pattern uses | |
19520 | V4DImode instead of V2TImode. */ | |
19521 | target = d->target; | |
19522 | if (d->vmode != V4DImode) | |
19523 | target = gen_reg_rtx (V4DImode); | |
19524 | op0 = gen_lowpart (V4DImode, d->op0); | |
19525 | op1 = gen_lowpart (V4DImode, d->op1); | |
19526 | rperm[0] | |
19527 | = GEN_INT ((d->perm[0] / (nelt / 2)) | |
19528 | | ((d->perm[nelt / 2] / (nelt / 2)) * 16)); | |
19529 | emit_insn (gen_avx2_permv2ti (target, op0, op1, rperm[0])); | |
19530 | if (target != d->target) | |
19531 | emit_move_insn (d->target, gen_lowpart (d->vmode, target)); | |
19532 | return true; | |
19533 | } | |
19534 | /* FALLTHRU */ | |
19535 | ||
19536 | default: | |
37e93925 | 19537 | return false; |
be8749f9 | 19538 | } |
2bf6d935 | 19539 | else |
be8749f9 UB |
19540 | switch (GET_MODE_SIZE (d->vmode)) |
19541 | { | |
19542 | case 4: | |
19543 | if (!TARGET_SSSE3) | |
19544 | return false; | |
19545 | vmode = V4QImode; | |
19546 | break; | |
2bf6d935 | 19547 | |
be8749f9 UB |
19548 | case 8: |
19549 | if (!TARGET_SSSE3) | |
19550 | return false; | |
19551 | vmode = V8QImode; | |
19552 | break; | |
2bf6d935 | 19553 | |
be8749f9 UB |
19554 | case 16: |
19555 | if (!TARGET_SSSE3) | |
19556 | return false; | |
877c9e33 | 19557 | vmode = V16QImode; |
be8749f9 UB |
19558 | break; |
19559 | ||
19560 | case 32: | |
19561 | if (!TARGET_AVX2) | |
19562 | return false; | |
19563 | ||
19564 | /* V4DImode should be already handled through | |
19565 | expand_vselect by vpermq instruction. */ | |
19566 | gcc_assert (d->vmode != V4DImode); | |
19567 | ||
19568 | vmode = V32QImode; | |
19569 | if (d->vmode == V8SImode | |
19570 | || d->vmode == V16HImode | |
19571 | || d->vmode == V32QImode) | |
19572 | { | |
19573 | /* First see if vpermq can be used for | |
19574 | V8SImode/V16HImode/V32QImode. */ | |
19575 | if (valid_perm_using_mode_p (V4DImode, d)) | |
19576 | { | |
19577 | for (i = 0; i < 4; i++) | |
19578 | perm[i] = (d->perm[i * nelt / 4] * 4 / nelt) & 3; | |
19579 | if (d->testing_p) | |
19580 | return true; | |
19581 | target = gen_reg_rtx (V4DImode); | |
19582 | if (expand_vselect (target, gen_lowpart (V4DImode, d->op0), | |
19583 | perm, 4, false)) | |
19584 | { | |
19585 | emit_move_insn (d->target, | |
19586 | gen_lowpart (d->vmode, target)); | |
2bf6d935 | 19587 | return true; |
be8749f9 UB |
19588 | } |
19589 | return false; | |
19590 | } | |
2bf6d935 | 19591 | |
be8749f9 UB |
19592 | /* Next see if vpermd can be used. */ |
19593 | if (valid_perm_using_mode_p (V8SImode, d)) | |
19594 | vmode = V8SImode; | |
19595 | } | |
19596 | /* Or if vpermps can be used. */ | |
19597 | else if (d->vmode == V8SFmode) | |
19598 | vmode = V8SImode; | |
2bf6d935 | 19599 | |
be8749f9 UB |
19600 | if (vmode == V32QImode) |
19601 | { | |
19602 | /* vpshufb only works intra lanes, it is not | |
19603 | possible to shuffle bytes in between the lanes. */ | |
19604 | for (i = 0; i < nelt; ++i) | |
19605 | if ((d->perm[i] ^ i) & (nelt / 2)) | |
19606 | return false; | |
19607 | } | |
19608 | break; | |
2bf6d935 | 19609 | |
be8749f9 UB |
19610 | case 64: |
19611 | if (!TARGET_AVX512BW) | |
19612 | return false; | |
2bf6d935 | 19613 | |
be8749f9 UB |
19614 | /* If vpermq didn't work, vpshufb won't work either. */ |
19615 | if (d->vmode == V8DFmode || d->vmode == V8DImode) | |
19616 | return false; | |
19617 | ||
19618 | vmode = V64QImode; | |
19619 | if (d->vmode == V16SImode | |
19620 | || d->vmode == V32HImode | |
19621 | || d->vmode == V64QImode) | |
19622 | { | |
19623 | /* First see if vpermq can be used for | |
19624 | V16SImode/V32HImode/V64QImode. */ | |
19625 | if (valid_perm_using_mode_p (V8DImode, d)) | |
19626 | { | |
19627 | for (i = 0; i < 8; i++) | |
19628 | perm[i] = (d->perm[i * nelt / 8] * 8 / nelt) & 7; | |
19629 | if (d->testing_p) | |
19630 | return true; | |
19631 | target = gen_reg_rtx (V8DImode); | |
19632 | if (expand_vselect (target, gen_lowpart (V8DImode, d->op0), | |
19633 | perm, 8, false)) | |
19634 | { | |
19635 | emit_move_insn (d->target, | |
19636 | gen_lowpart (d->vmode, target)); | |
2bf6d935 | 19637 | return true; |
be8749f9 UB |
19638 | } |
19639 | return false; | |
19640 | } | |
2bf6d935 | 19641 | |
be8749f9 UB |
19642 | /* Next see if vpermd can be used. */ |
19643 | if (valid_perm_using_mode_p (V16SImode, d)) | |
19644 | vmode = V16SImode; | |
19645 | } | |
19646 | /* Or if vpermps can be used. */ | |
19647 | else if (d->vmode == V16SFmode) | |
19648 | vmode = V16SImode; | |
877c9e33 | 19649 | |
be8749f9 UB |
19650 | if (vmode == V64QImode) |
19651 | { | |
19652 | /* vpshufb only works intra lanes, it is not | |
19653 | possible to shuffle bytes in between the lanes. */ | |
19654 | for (i = 0; i < nelt; ++i) | |
19655 | if ((d->perm[i] ^ i) & (3 * nelt / 4)) | |
19656 | return false; | |
19657 | } | |
19658 | break; | |
19659 | ||
19660 | default: | |
2bf6d935 | 19661 | return false; |
be8749f9 | 19662 | } |
2bf6d935 ML |
19663 | |
19664 | if (d->testing_p) | |
19665 | return true; | |
19666 | ||
681143b9 UB |
19667 | /* Try to avoid variable permutation instruction. */ |
19668 | if (canonicalize_vector_int_perm (d, &nd) && expand_vec_perm_1 (&nd)) | |
19669 | { | |
19670 | emit_move_insn (d->target, gen_lowpart (d->vmode, nd.target)); | |
19671 | return true; | |
19672 | } | |
19673 | ||
2bf6d935 ML |
19674 | if (vmode == V8SImode) |
19675 | for (i = 0; i < 8; ++i) | |
19676 | rperm[i] = GEN_INT ((d->perm[i * nelt / 8] * 8 / nelt) & 7); | |
19677 | else if (vmode == V16SImode) | |
19678 | for (i = 0; i < 16; ++i) | |
19679 | rperm[i] = GEN_INT ((d->perm[i * nelt / 16] * 16 / nelt) & 15); | |
19680 | else | |
19681 | { | |
19682 | eltsz = GET_MODE_UNIT_SIZE (d->vmode); | |
19683 | if (!d->one_operand_p) | |
19684 | mask = 2 * nelt - 1; | |
2bf6d935 ML |
19685 | else if (vmode == V64QImode) |
19686 | mask = nelt / 4 - 1; | |
a325bdd1 | 19687 | else if (vmode == V32QImode) |
2bf6d935 | 19688 | mask = nelt / 2 - 1; |
a325bdd1 PB |
19689 | else |
19690 | mask = nelt - 1; | |
2bf6d935 ML |
19691 | |
19692 | for (i = 0; i < nelt; ++i) | |
19693 | { | |
19694 | unsigned j, e = d->perm[i] & mask; | |
19695 | for (j = 0; j < eltsz; ++j) | |
19696 | rperm[i * eltsz + j] = GEN_INT (e * eltsz + j); | |
19697 | } | |
19698 | } | |
19699 | ||
a325bdd1 PB |
19700 | machine_mode vpmode = vmode; |
19701 | ||
877c9e33 UB |
19702 | nelt = GET_MODE_SIZE (vmode); |
19703 | ||
19704 | /* Emulate narrow modes with V16QI instructions. */ | |
19705 | if (nelt < 16) | |
a325bdd1 | 19706 | { |
dd835ec2 UB |
19707 | rtx m128 = GEN_INT (-128); |
19708 | ||
37e93925 | 19709 | /* Remap elements from the second operand, as we have to |
be8749f9 | 19710 | account for inactive top elements from the first operand. */ |
37e93925 | 19711 | if (!d->one_operand_p) |
be8749f9 | 19712 | { |
be8749f9 UB |
19713 | for (i = 0; i < nelt; ++i) |
19714 | { | |
877c9e33 UB |
19715 | unsigned ival = UINTVAL (rperm[i]); |
19716 | if (ival >= nelt) | |
19717 | rperm[i] = GEN_INT (ival + 16 - nelt); | |
be8749f9 UB |
19718 | } |
19719 | } | |
37e93925 | 19720 | |
877c9e33 | 19721 | /* Fill inactive elements in the top positions with zeros. */ |
a325bdd1 | 19722 | for (i = nelt; i < 16; ++i) |
dd835ec2 | 19723 | rperm[i] = m128; |
37e93925 | 19724 | |
a325bdd1 PB |
19725 | vpmode = V16QImode; |
19726 | } | |
19727 | ||
19728 | vperm = gen_rtx_CONST_VECTOR (vpmode, | |
19729 | gen_rtvec_v (GET_MODE_NUNITS (vpmode), rperm)); | |
19730 | vperm = force_reg (vpmode, vperm); | |
2bf6d935 | 19731 | |
37e93925 UB |
19732 | if (vmode == d->vmode) |
19733 | target = d->target; | |
19734 | else | |
2bf6d935 | 19735 | target = gen_reg_rtx (vmode); |
37e93925 | 19736 | |
2bf6d935 | 19737 | op0 = gen_lowpart (vmode, d->op0); |
37e93925 | 19738 | |
2bf6d935 ML |
19739 | if (d->one_operand_p) |
19740 | { | |
37e93925 UB |
19741 | rtx (*gen) (rtx, rtx, rtx); |
19742 | ||
be8749f9 UB |
19743 | if (vmode == V4QImode) |
19744 | gen = gen_mmx_pshufbv4qi3; | |
19745 | else if (vmode == V8QImode) | |
37e93925 | 19746 | gen = gen_mmx_pshufbv8qi3; |
a325bdd1 | 19747 | else if (vmode == V16QImode) |
37e93925 | 19748 | gen = gen_ssse3_pshufbv16qi3; |
2bf6d935 | 19749 | else if (vmode == V32QImode) |
37e93925 | 19750 | gen = gen_avx2_pshufbv32qi3; |
2bf6d935 | 19751 | else if (vmode == V64QImode) |
37e93925 | 19752 | gen = gen_avx512bw_pshufbv64qi3; |
2bf6d935 | 19753 | else if (vmode == V8SFmode) |
37e93925 | 19754 | gen = gen_avx2_permvarv8sf; |
2bf6d935 | 19755 | else if (vmode == V8SImode) |
37e93925 | 19756 | gen = gen_avx2_permvarv8si; |
2bf6d935 | 19757 | else if (vmode == V16SFmode) |
37e93925 | 19758 | gen = gen_avx512f_permvarv16sf; |
2bf6d935 | 19759 | else if (vmode == V16SImode) |
37e93925 | 19760 | gen = gen_avx512f_permvarv16si; |
2bf6d935 ML |
19761 | else |
19762 | gcc_unreachable (); | |
37e93925 UB |
19763 | |
19764 | emit_insn (gen (target, op0, vperm)); | |
2bf6d935 ML |
19765 | } |
19766 | else | |
19767 | { | |
37e93925 UB |
19768 | rtx (*gen) (rtx, rtx, rtx, rtx); |
19769 | ||
2bf6d935 | 19770 | op1 = gen_lowpart (vmode, d->op1); |
37e93925 | 19771 | |
be8749f9 UB |
19772 | if (vmode == V4QImode) |
19773 | gen = gen_mmx_ppermv32; | |
19774 | else if (vmode == V8QImode) | |
37e93925 UB |
19775 | gen = gen_mmx_ppermv64; |
19776 | else if (vmode == V16QImode) | |
19777 | gen = gen_xop_pperm; | |
19778 | else | |
19779 | gcc_unreachable (); | |
19780 | ||
19781 | emit_insn (gen (target, op0, op1, vperm)); | |
2bf6d935 | 19782 | } |
37e93925 | 19783 | |
2bf6d935 ML |
19784 | if (target != d->target) |
19785 | emit_move_insn (d->target, gen_lowpart (d->vmode, target)); | |
19786 | ||
19787 | return true; | |
19788 | } | |
19789 | ||
2bf6d935 ML |
19790 | /* Try to expand one-operand permutation with constant mask. */ |
19791 | ||
19792 | static bool | |
19793 | ix86_expand_vec_one_operand_perm_avx512 (struct expand_vec_perm_d *d) | |
19794 | { | |
19795 | machine_mode mode = GET_MODE (d->op0); | |
19796 | machine_mode maskmode = mode; | |
faf2b6bc | 19797 | unsigned inner_size = GET_MODE_SIZE (GET_MODE_INNER (mode)); |
2bf6d935 ML |
19798 | rtx (*gen) (rtx, rtx, rtx) = NULL; |
19799 | rtx target, op0, mask; | |
19800 | rtx vec[64]; | |
19801 | ||
19802 | if (!rtx_equal_p (d->op0, d->op1)) | |
19803 | return false; | |
19804 | ||
19805 | if (!TARGET_AVX512F) | |
19806 | return false; | |
19807 | ||
faf2b6bc | 19808 | /* Accept VNxHImode and VNxQImode now. */ |
19809 | if (!TARGET_AVX512VL && GET_MODE_SIZE (mode) < 64) | |
19810 | return false; | |
19811 | ||
19812 | /* vpermw. */ | |
19813 | if (!TARGET_AVX512BW && inner_size == 2) | |
19814 | return false; | |
19815 | ||
19816 | /* vpermb. */ | |
19817 | if (!TARGET_AVX512VBMI && inner_size == 1) | |
19818 | return false; | |
19819 | ||
2bf6d935 ML |
19820 | switch (mode) |
19821 | { | |
19822 | case E_V16SImode: | |
19823 | gen = gen_avx512f_permvarv16si; | |
19824 | break; | |
19825 | case E_V16SFmode: | |
19826 | gen = gen_avx512f_permvarv16sf; | |
19827 | maskmode = V16SImode; | |
19828 | break; | |
19829 | case E_V8DImode: | |
19830 | gen = gen_avx512f_permvarv8di; | |
19831 | break; | |
19832 | case E_V8DFmode: | |
19833 | gen = gen_avx512f_permvarv8df; | |
19834 | maskmode = V8DImode; | |
19835 | break; | |
faf2b6bc | 19836 | case E_V32HImode: |
19837 | gen = gen_avx512bw_permvarv32hi; | |
19838 | break; | |
19839 | case E_V16HImode: | |
19840 | gen = gen_avx512vl_permvarv16hi; | |
19841 | break; | |
19842 | case E_V8HImode: | |
19843 | gen = gen_avx512vl_permvarv8hi; | |
19844 | break; | |
19845 | case E_V64QImode: | |
19846 | gen = gen_avx512bw_permvarv64qi; | |
19847 | break; | |
19848 | case E_V32QImode: | |
19849 | gen = gen_avx512vl_permvarv32qi; | |
19850 | break; | |
19851 | case E_V16QImode: | |
19852 | gen = gen_avx512vl_permvarv16qi; | |
19853 | break; | |
19854 | ||
2bf6d935 ML |
19855 | default: |
19856 | return false; | |
19857 | } | |
19858 | ||
04b4f315 JJ |
19859 | if (d->testing_p) |
19860 | return true; | |
19861 | ||
2bf6d935 ML |
19862 | target = d->target; |
19863 | op0 = d->op0; | |
19864 | for (int i = 0; i < d->nelt; ++i) | |
19865 | vec[i] = GEN_INT (d->perm[i]); | |
19866 | mask = gen_rtx_CONST_VECTOR (maskmode, gen_rtvec_v (d->nelt, vec)); | |
19867 | emit_insn (gen (target, op0, force_reg (maskmode, mask))); | |
19868 | return true; | |
19869 | } | |
19870 | ||
19871 | static bool expand_vec_perm_palignr (struct expand_vec_perm_d *d, bool); | |
19872 | ||
4bf4c103 | 19873 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to instantiate D |
2bf6d935 ML |
19874 | in a single instruction. */ |
19875 | ||
19876 | static bool | |
19877 | expand_vec_perm_1 (struct expand_vec_perm_d *d) | |
19878 | { | |
19879 | unsigned i, nelt = d->nelt; | |
19880 | struct expand_vec_perm_d nd; | |
19881 | ||
19882 | /* Check plain VEC_SELECT first, because AVX has instructions that could | |
19883 | match both SEL and SEL+CONCAT, but the plain SEL will allow a memory | |
19884 | input where SEL+CONCAT may not. */ | |
19885 | if (d->one_operand_p) | |
19886 | { | |
19887 | int mask = nelt - 1; | |
19888 | bool identity_perm = true; | |
19889 | bool broadcast_perm = true; | |
19890 | ||
19891 | for (i = 0; i < nelt; i++) | |
19892 | { | |
19893 | nd.perm[i] = d->perm[i] & mask; | |
19894 | if (nd.perm[i] != i) | |
19895 | identity_perm = false; | |
19896 | if (nd.perm[i]) | |
19897 | broadcast_perm = false; | |
19898 | } | |
19899 | ||
19900 | if (identity_perm) | |
19901 | { | |
19902 | if (!d->testing_p) | |
19903 | emit_move_insn (d->target, d->op0); | |
19904 | return true; | |
19905 | } | |
19906 | else if (broadcast_perm && TARGET_AVX2) | |
19907 | { | |
19908 | /* Use vpbroadcast{b,w,d}. */ | |
19909 | rtx (*gen) (rtx, rtx) = NULL; | |
19910 | switch (d->vmode) | |
19911 | { | |
19912 | case E_V64QImode: | |
19913 | if (TARGET_AVX512BW) | |
19914 | gen = gen_avx512bw_vec_dupv64qi_1; | |
19915 | break; | |
19916 | case E_V32QImode: | |
19917 | gen = gen_avx2_pbroadcastv32qi_1; | |
19918 | break; | |
19919 | case E_V32HImode: | |
19920 | if (TARGET_AVX512BW) | |
19921 | gen = gen_avx512bw_vec_dupv32hi_1; | |
19922 | break; | |
19923 | case E_V16HImode: | |
19924 | gen = gen_avx2_pbroadcastv16hi_1; | |
19925 | break; | |
19926 | case E_V16SImode: | |
19927 | if (TARGET_AVX512F) | |
19928 | gen = gen_avx512f_vec_dupv16si_1; | |
19929 | break; | |
19930 | case E_V8SImode: | |
19931 | gen = gen_avx2_pbroadcastv8si_1; | |
19932 | break; | |
19933 | case E_V16QImode: | |
19934 | gen = gen_avx2_pbroadcastv16qi; | |
19935 | break; | |
19936 | case E_V8HImode: | |
19937 | gen = gen_avx2_pbroadcastv8hi; | |
19938 | break; | |
19939 | case E_V16SFmode: | |
19940 | if (TARGET_AVX512F) | |
19941 | gen = gen_avx512f_vec_dupv16sf_1; | |
19942 | break; | |
19943 | case E_V8SFmode: | |
19944 | gen = gen_avx2_vec_dupv8sf_1; | |
19945 | break; | |
19946 | case E_V8DFmode: | |
19947 | if (TARGET_AVX512F) | |
19948 | gen = gen_avx512f_vec_dupv8df_1; | |
19949 | break; | |
19950 | case E_V8DImode: | |
19951 | if (TARGET_AVX512F) | |
19952 | gen = gen_avx512f_vec_dupv8di_1; | |
19953 | break; | |
19954 | /* For other modes prefer other shuffles this function creates. */ | |
19955 | default: break; | |
19956 | } | |
19957 | if (gen != NULL) | |
19958 | { | |
19959 | if (!d->testing_p) | |
19960 | emit_insn (gen (d->target, d->op0)); | |
19961 | return true; | |
19962 | } | |
19963 | } | |
19964 | ||
19965 | if (expand_vselect (d->target, d->op0, nd.perm, nelt, d->testing_p)) | |
19966 | return true; | |
19967 | ||
19968 | /* There are plenty of patterns in sse.md that are written for | |
19969 | SEL+CONCAT and are not replicated for a single op. Perhaps | |
19970 | that should be changed, to avoid the nastiness here. */ | |
19971 | ||
19972 | /* Recognize interleave style patterns, which means incrementing | |
19973 | every other permutation operand. */ | |
19974 | for (i = 0; i < nelt; i += 2) | |
19975 | { | |
19976 | nd.perm[i] = d->perm[i] & mask; | |
19977 | nd.perm[i + 1] = (d->perm[i + 1] & mask) + nelt; | |
19978 | } | |
19979 | if (expand_vselect_vconcat (d->target, d->op0, d->op0, nd.perm, nelt, | |
19980 | d->testing_p)) | |
19981 | return true; | |
19982 | ||
19983 | /* Recognize shufps, which means adding {0, 0, nelt, nelt}. */ | |
19984 | if (nelt >= 4) | |
19985 | { | |
19986 | for (i = 0; i < nelt; i += 4) | |
19987 | { | |
19988 | nd.perm[i + 0] = d->perm[i + 0] & mask; | |
19989 | nd.perm[i + 1] = d->perm[i + 1] & mask; | |
19990 | nd.perm[i + 2] = (d->perm[i + 2] & mask) + nelt; | |
19991 | nd.perm[i + 3] = (d->perm[i + 3] & mask) + nelt; | |
19992 | } | |
19993 | ||
19994 | if (expand_vselect_vconcat (d->target, d->op0, d->op0, nd.perm, nelt, | |
19995 | d->testing_p)) | |
19996 | return true; | |
19997 | } | |
19998 | } | |
19999 | ||
57052c6e UB |
20000 | /* Try the SSE4.1 blend variable merge instructions. */ |
20001 | if (expand_vec_perm_blend (d)) | |
20002 | return true; | |
20003 | ||
2bf6d935 ML |
20004 | /* Try movss/movsd instructions. */ |
20005 | if (expand_vec_perm_movs (d)) | |
20006 | return true; | |
20007 | ||
95b99e47 UB |
20008 | /* Try the SSE4.1 insertps instruction. */ |
20009 | if (expand_vec_perm_insertps (d)) | |
20010 | return true; | |
20011 | ||
57052c6e | 20012 | /* Try the fully general two operand permute. */ |
2bf6d935 ML |
20013 | if (expand_vselect_vconcat (d->target, d->op0, d->op1, d->perm, nelt, |
20014 | d->testing_p)) | |
20015 | return true; | |
20016 | ||
20017 | /* Recognize interleave style patterns with reversed operands. */ | |
20018 | if (!d->one_operand_p) | |
20019 | { | |
20020 | for (i = 0; i < nelt; ++i) | |
20021 | { | |
20022 | unsigned e = d->perm[i]; | |
20023 | if (e >= nelt) | |
20024 | e -= nelt; | |
20025 | else | |
20026 | e += nelt; | |
20027 | nd.perm[i] = e; | |
20028 | } | |
20029 | ||
20030 | if (expand_vselect_vconcat (d->target, d->op1, d->op0, nd.perm, nelt, | |
20031 | d->testing_p)) | |
20032 | return true; | |
20033 | } | |
20034 | ||
2bf6d935 ML |
20035 | /* Try one of the AVX vpermil variable permutations. */ |
20036 | if (expand_vec_perm_vpermil (d)) | |
20037 | return true; | |
20038 | ||
20039 | /* Try the SSSE3 pshufb or XOP vpperm or AVX2 vperm2i128, | |
20040 | vpshufb, vpermd, vpermps or vpermq variable permutation. */ | |
20041 | if (expand_vec_perm_pshufb (d)) | |
20042 | return true; | |
20043 | ||
20044 | /* Try the AVX2 vpalignr instruction. */ | |
20045 | if (expand_vec_perm_palignr (d, true)) | |
20046 | return true; | |
20047 | ||
faf2b6bc | 20048 | /* Try the AVX512F vperm{w,b,s,d} instructions */ |
2bf6d935 ML |
20049 | if (ix86_expand_vec_one_operand_perm_avx512 (d)) |
20050 | return true; | |
20051 | ||
20052 | /* Try the AVX512F vpermt2/vpermi2 instructions. */ | |
20053 | if (ix86_expand_vec_perm_vpermt2 (NULL_RTX, NULL_RTX, NULL_RTX, NULL_RTX, d)) | |
20054 | return true; | |
20055 | ||
20056 | /* See if we can get the same permutation in different vector integer | |
20057 | mode. */ | |
20058 | if (canonicalize_vector_int_perm (d, &nd) && expand_vec_perm_1 (&nd)) | |
20059 | { | |
20060 | if (!d->testing_p) | |
20061 | emit_move_insn (d->target, gen_lowpart (d->vmode, nd.target)); | |
20062 | return true; | |
20063 | } | |
20064 | return false; | |
20065 | } | |
20066 | ||
1442e203 | 20067 | /* Canonicalize vec_perm index to make the first index |
20068 | always comes from the first vector. */ | |
20069 | static void | |
20070 | ix86_vec_perm_index_canon (struct expand_vec_perm_d *d) | |
20071 | { | |
20072 | unsigned nelt = d->nelt; | |
20073 | if (d->perm[0] < nelt) | |
20074 | return; | |
20075 | ||
20076 | for (unsigned i = 0; i != nelt; i++) | |
20077 | d->perm[i] = (d->perm[i] + nelt) % (2 * nelt); | |
20078 | ||
20079 | std::swap (d->op0, d->op1); | |
20080 | return; | |
20081 | } | |
20082 | ||
3db8e9c2 | 20083 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to implement D |
20084 | in terms of a pair of shufps+ shufps/pshufd instructions. */ | |
20085 | static bool | |
20086 | expand_vec_perm_shufps_shufps (struct expand_vec_perm_d *d) | |
20087 | { | |
20088 | unsigned char perm1[4]; | |
20089 | machine_mode vmode = d->vmode; | |
20090 | bool ok; | |
20091 | unsigned i, j, k, count = 0; | |
20092 | ||
20093 | if (d->one_operand_p | |
20094 | || (vmode != V4SImode && vmode != V4SFmode)) | |
20095 | return false; | |
20096 | ||
20097 | if (d->testing_p) | |
20098 | return true; | |
20099 | ||
1442e203 | 20100 | ix86_vec_perm_index_canon (d); |
3db8e9c2 | 20101 | for (i = 0; i < 4; ++i) |
20102 | count += d->perm[i] > 3 ? 1 : 0; | |
20103 | ||
20104 | gcc_assert (count & 3); | |
20105 | ||
20106 | rtx tmp = gen_reg_rtx (vmode); | |
20107 | /* 2 from op0 and 2 from op1. */ | |
20108 | if (count == 2) | |
20109 | { | |
20110 | unsigned char perm2[4]; | |
20111 | for (i = 0, j = 0, k = 2; i < 4; ++i) | |
20112 | if (d->perm[i] & 4) | |
20113 | { | |
20114 | perm1[k++] = d->perm[i]; | |
20115 | perm2[i] = k - 1; | |
20116 | } | |
20117 | else | |
20118 | { | |
20119 | perm1[j++] = d->perm[i]; | |
20120 | perm2[i] = j - 1; | |
20121 | } | |
20122 | ||
20123 | /* shufps. */ | |
20124 | ok = expand_vselect_vconcat (tmp, d->op0, d->op1, | |
20125 | perm1, d->nelt, false); | |
20126 | gcc_assert (ok); | |
20127 | if (vmode == V4SImode && TARGET_SSE2) | |
20128 | /* pshufd. */ | |
20129 | ok = expand_vselect (d->target, tmp, | |
20130 | perm2, d->nelt, false); | |
20131 | else | |
20132 | { | |
20133 | /* shufps. */ | |
20134 | perm2[2] += 4; | |
20135 | perm2[3] += 4; | |
20136 | ok = expand_vselect_vconcat (d->target, tmp, tmp, | |
20137 | perm2, d->nelt, false); | |
20138 | } | |
20139 | gcc_assert (ok); | |
20140 | } | |
20141 | /* 3 from one op and 1 from another. */ | |
20142 | else | |
20143 | { | |
20144 | unsigned pair_idx = 8, lone_idx = 8, shift; | |
20145 | ||
20146 | /* Find the lone index. */ | |
20147 | for (i = 0; i < 4; ++i) | |
20148 | if ((d->perm[i] > 3 && count == 1) | |
20149 | || (d->perm[i] < 4 && count == 3)) | |
20150 | lone_idx = i; | |
20151 | ||
20152 | /* When lone_idx is not 0, it must from second op(count == 1). */ | |
20153 | gcc_assert (count == (lone_idx ? 1 : 3)); | |
20154 | ||
20155 | /* Find the pair index that sits in the same half as the lone index. */ | |
20156 | shift = lone_idx & 2; | |
20157 | pair_idx = 1 - lone_idx + 2 * shift; | |
20158 | ||
20159 | /* First permutate lone index and pair index into the same vector as | |
20160 | [ lone, lone, pair, pair ]. */ | |
20161 | perm1[1] = perm1[0] | |
20162 | = (count == 3) ? d->perm[lone_idx] : d->perm[lone_idx] - 4; | |
20163 | perm1[3] = perm1[2] | |
20164 | = (count == 3) ? d->perm[pair_idx] : d->perm[pair_idx] + 4; | |
20165 | ||
20166 | /* Alway put the vector contains lone indx at the first. */ | |
20167 | if (count == 1) | |
20168 | std::swap (d->op0, d->op1); | |
20169 | ||
20170 | /* shufps. */ | |
20171 | ok = expand_vselect_vconcat (tmp, d->op0, d->op1, | |
20172 | perm1, d->nelt, false); | |
20173 | gcc_assert (ok); | |
20174 | ||
20175 | /* Refine lone and pair index to original order. */ | |
20176 | perm1[shift] = lone_idx << 1; | |
20177 | perm1[shift + 1] = pair_idx << 1; | |
20178 | ||
20179 | /* Select the remaining 2 elements in another vector. */ | |
20180 | for (i = 2 - shift; i < 4 - shift; ++i) | |
20181 | perm1[i] = lone_idx == 1 ? d->perm[i] + 4 : d->perm[i]; | |
20182 | ||
20183 | /* Adjust to original selector. */ | |
20184 | if (lone_idx > 1) | |
20185 | std::swap (tmp, d->op1); | |
20186 | ||
20187 | /* shufps. */ | |
20188 | ok = expand_vselect_vconcat (d->target, tmp, d->op1, | |
20189 | perm1, d->nelt, false); | |
20190 | ||
20191 | gcc_assert (ok); | |
20192 | } | |
20193 | ||
20194 | return true; | |
20195 | } | |
20196 | ||
4bf4c103 | 20197 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to implement D |
2bf6d935 ML |
20198 | in terms of a pair of pshuflw + pshufhw instructions. */ |
20199 | ||
20200 | static bool | |
20201 | expand_vec_perm_pshuflw_pshufhw (struct expand_vec_perm_d *d) | |
20202 | { | |
20203 | unsigned char perm2[MAX_VECT_LEN]; | |
20204 | unsigned i; | |
20205 | bool ok; | |
20206 | ||
20207 | if (d->vmode != V8HImode || !d->one_operand_p) | |
20208 | return false; | |
20209 | ||
20210 | /* The two permutations only operate in 64-bit lanes. */ | |
20211 | for (i = 0; i < 4; ++i) | |
20212 | if (d->perm[i] >= 4) | |
20213 | return false; | |
20214 | for (i = 4; i < 8; ++i) | |
20215 | if (d->perm[i] < 4) | |
20216 | return false; | |
20217 | ||
20218 | if (d->testing_p) | |
20219 | return true; | |
20220 | ||
20221 | /* Emit the pshuflw. */ | |
20222 | memcpy (perm2, d->perm, 4); | |
20223 | for (i = 4; i < 8; ++i) | |
20224 | perm2[i] = i; | |
20225 | ok = expand_vselect (d->target, d->op0, perm2, 8, d->testing_p); | |
20226 | gcc_assert (ok); | |
20227 | ||
20228 | /* Emit the pshufhw. */ | |
20229 | memcpy (perm2 + 4, d->perm + 4, 4); | |
20230 | for (i = 0; i < 4; ++i) | |
20231 | perm2[i] = i; | |
20232 | ok = expand_vselect (d->target, d->target, perm2, 8, d->testing_p); | |
20233 | gcc_assert (ok); | |
20234 | ||
20235 | return true; | |
20236 | } | |
20237 | ||
4bf4c103 | 20238 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to simplify |
2bf6d935 ML |
20239 | the permutation using the SSSE3 palignr instruction. This succeeds |
20240 | when all of the elements in PERM fit within one vector and we merely | |
20241 | need to shift them down so that a single vector permutation has a | |
20242 | chance to succeed. If SINGLE_INSN_ONLY_P, succeed if only | |
20243 | the vpalignr instruction itself can perform the requested permutation. */ | |
20244 | ||
20245 | static bool | |
20246 | expand_vec_perm_palignr (struct expand_vec_perm_d *d, bool single_insn_only_p) | |
20247 | { | |
20248 | unsigned i, nelt = d->nelt; | |
20249 | unsigned min, max, minswap, maxswap; | |
20250 | bool in_order, ok, swap = false; | |
20251 | rtx shift, target; | |
20252 | struct expand_vec_perm_d dcopy; | |
20253 | ||
20254 | /* Even with AVX, palignr only operates on 128-bit vectors, | |
20255 | in AVX2 palignr operates on both 128-bit lanes. */ | |
20256 | if ((!TARGET_SSSE3 || GET_MODE_SIZE (d->vmode) != 16) | |
20257 | && (!TARGET_AVX2 || GET_MODE_SIZE (d->vmode) != 32)) | |
20258 | return false; | |
20259 | ||
20260 | min = 2 * nelt; | |
20261 | max = 0; | |
20262 | minswap = 2 * nelt; | |
20263 | maxswap = 0; | |
20264 | for (i = 0; i < nelt; ++i) | |
20265 | { | |
20266 | unsigned e = d->perm[i]; | |
20267 | unsigned eswap = d->perm[i] ^ nelt; | |
20268 | if (GET_MODE_SIZE (d->vmode) == 32) | |
20269 | { | |
20270 | e = (e & ((nelt / 2) - 1)) | ((e & nelt) >> 1); | |
20271 | eswap = e ^ (nelt / 2); | |
20272 | } | |
20273 | if (e < min) | |
20274 | min = e; | |
20275 | if (e > max) | |
20276 | max = e; | |
20277 | if (eswap < minswap) | |
20278 | minswap = eswap; | |
20279 | if (eswap > maxswap) | |
20280 | maxswap = eswap; | |
20281 | } | |
20282 | if (min == 0 | |
20283 | || max - min >= (GET_MODE_SIZE (d->vmode) == 32 ? nelt / 2 : nelt)) | |
20284 | { | |
20285 | if (d->one_operand_p | |
20286 | || minswap == 0 | |
20287 | || maxswap - minswap >= (GET_MODE_SIZE (d->vmode) == 32 | |
20288 | ? nelt / 2 : nelt)) | |
20289 | return false; | |
20290 | swap = true; | |
20291 | min = minswap; | |
20292 | max = maxswap; | |
20293 | } | |
20294 | ||
20295 | /* Given that we have SSSE3, we know we'll be able to implement the | |
20296 | single operand permutation after the palignr with pshufb for | |
20297 | 128-bit vectors. If SINGLE_INSN_ONLY_P, in_order has to be computed | |
20298 | first. */ | |
20299 | if (d->testing_p && GET_MODE_SIZE (d->vmode) == 16 && !single_insn_only_p) | |
20300 | return true; | |
20301 | ||
20302 | dcopy = *d; | |
20303 | if (swap) | |
20304 | { | |
20305 | dcopy.op0 = d->op1; | |
20306 | dcopy.op1 = d->op0; | |
20307 | for (i = 0; i < nelt; ++i) | |
20308 | dcopy.perm[i] ^= nelt; | |
20309 | } | |
20310 | ||
20311 | in_order = true; | |
20312 | for (i = 0; i < nelt; ++i) | |
20313 | { | |
20314 | unsigned e = dcopy.perm[i]; | |
20315 | if (GET_MODE_SIZE (d->vmode) == 32 | |
20316 | && e >= nelt | |
20317 | && (e & (nelt / 2 - 1)) < min) | |
20318 | e = e - min - (nelt / 2); | |
20319 | else | |
20320 | e = e - min; | |
20321 | if (e != i) | |
20322 | in_order = false; | |
20323 | dcopy.perm[i] = e; | |
20324 | } | |
20325 | dcopy.one_operand_p = true; | |
20326 | ||
20327 | if (single_insn_only_p && !in_order) | |
20328 | return false; | |
20329 | ||
20330 | /* For AVX2, test whether we can permute the result in one instruction. */ | |
20331 | if (d->testing_p) | |
20332 | { | |
20333 | if (in_order) | |
20334 | return true; | |
20335 | dcopy.op1 = dcopy.op0; | |
20336 | return expand_vec_perm_1 (&dcopy); | |
20337 | } | |
20338 | ||
20339 | shift = GEN_INT (min * GET_MODE_UNIT_BITSIZE (d->vmode)); | |
20340 | if (GET_MODE_SIZE (d->vmode) == 16) | |
20341 | { | |
02e2e15e RS |
20342 | target = gen_reg_rtx (V1TImode); |
20343 | emit_insn (gen_ssse3_palignrv1ti (target, | |
20344 | gen_lowpart (V1TImode, dcopy.op1), | |
20345 | gen_lowpart (V1TImode, dcopy.op0), | |
20346 | shift)); | |
2bf6d935 ML |
20347 | } |
20348 | else | |
20349 | { | |
20350 | target = gen_reg_rtx (V2TImode); | |
20351 | emit_insn (gen_avx2_palignrv2ti (target, | |
20352 | gen_lowpart (V2TImode, dcopy.op1), | |
20353 | gen_lowpart (V2TImode, dcopy.op0), | |
20354 | shift)); | |
20355 | } | |
20356 | ||
20357 | dcopy.op0 = dcopy.op1 = gen_lowpart (d->vmode, target); | |
20358 | ||
20359 | /* Test for the degenerate case where the alignment by itself | |
20360 | produces the desired permutation. */ | |
20361 | if (in_order) | |
20362 | { | |
20363 | emit_move_insn (d->target, dcopy.op0); | |
20364 | return true; | |
20365 | } | |
20366 | ||
20367 | ok = expand_vec_perm_1 (&dcopy); | |
20368 | gcc_assert (ok || GET_MODE_SIZE (d->vmode) == 32); | |
20369 | ||
20370 | return ok; | |
20371 | } | |
20372 | ||
20373 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to simplify | |
20374 | the permutation using the SSE4_1 pblendv instruction. Potentially | |
20375 | reduces permutation from 2 pshufb and or to 1 pshufb and pblendv. */ | |
20376 | ||
20377 | static bool | |
20378 | expand_vec_perm_pblendv (struct expand_vec_perm_d *d) | |
20379 | { | |
20380 | unsigned i, which, nelt = d->nelt; | |
20381 | struct expand_vec_perm_d dcopy, dcopy1; | |
20382 | machine_mode vmode = d->vmode; | |
20383 | bool ok; | |
20384 | ||
20385 | /* Use the same checks as in expand_vec_perm_blend. */ | |
20386 | if (d->one_operand_p) | |
20387 | return false; | |
20388 | if (TARGET_AVX2 && GET_MODE_SIZE (vmode) == 32) | |
20389 | ; | |
20390 | else if (TARGET_AVX && (vmode == V4DFmode || vmode == V8SFmode)) | |
20391 | ; | |
acff89c7 UB |
20392 | else if (TARGET_SSE4_1 |
20393 | && (GET_MODE_SIZE (vmode) == 16 | |
20394 | || (TARGET_MMX_WITH_SSE && GET_MODE_SIZE (vmode) == 8) | |
20395 | || GET_MODE_SIZE (vmode) == 4)) | |
2bf6d935 ML |
20396 | ; |
20397 | else | |
20398 | return false; | |
20399 | ||
20400 | /* Figure out where permutation elements stay not in their | |
20401 | respective lanes. */ | |
20402 | for (i = 0, which = 0; i < nelt; ++i) | |
20403 | { | |
20404 | unsigned e = d->perm[i]; | |
20405 | if (e != i) | |
20406 | which |= (e < nelt ? 1 : 2); | |
20407 | } | |
20408 | /* We can pblend the part where elements stay not in their | |
20409 | respective lanes only when these elements are all in one | |
20410 | half of a permutation. | |
20411 | {0 1 8 3 4 5 9 7} is ok as 8, 9 are at not at their respective | |
20412 | lanes, but both 8 and 9 >= 8 | |
20413 | {0 1 8 3 4 5 2 7} is not ok as 2 and 8 are not at their | |
20414 | respective lanes and 8 >= 8, but 2 not. */ | |
20415 | if (which != 1 && which != 2) | |
20416 | return false; | |
20417 | if (d->testing_p && GET_MODE_SIZE (vmode) == 16) | |
20418 | return true; | |
20419 | ||
20420 | /* First we apply one operand permutation to the part where | |
20421 | elements stay not in their respective lanes. */ | |
20422 | dcopy = *d; | |
20423 | if (which == 2) | |
20424 | dcopy.op0 = dcopy.op1 = d->op1; | |
20425 | else | |
20426 | dcopy.op0 = dcopy.op1 = d->op0; | |
20427 | if (!d->testing_p) | |
20428 | dcopy.target = gen_reg_rtx (vmode); | |
20429 | dcopy.one_operand_p = true; | |
20430 | ||
20431 | for (i = 0; i < nelt; ++i) | |
20432 | dcopy.perm[i] = d->perm[i] & (nelt - 1); | |
20433 | ||
20434 | ok = expand_vec_perm_1 (&dcopy); | |
20435 | if (GET_MODE_SIZE (vmode) != 16 && !ok) | |
20436 | return false; | |
20437 | else | |
20438 | gcc_assert (ok); | |
20439 | if (d->testing_p) | |
20440 | return true; | |
20441 | ||
20442 | /* Next we put permuted elements into their positions. */ | |
20443 | dcopy1 = *d; | |
20444 | if (which == 2) | |
20445 | dcopy1.op1 = dcopy.target; | |
20446 | else | |
20447 | dcopy1.op0 = dcopy.target; | |
20448 | ||
20449 | for (i = 0; i < nelt; ++i) | |
20450 | dcopy1.perm[i] = ((d->perm[i] >= nelt) ? (nelt + i) : i); | |
20451 | ||
20452 | ok = expand_vec_perm_blend (&dcopy1); | |
20453 | gcc_assert (ok); | |
20454 | ||
20455 | return true; | |
20456 | } | |
20457 | ||
20458 | static bool expand_vec_perm_interleave3 (struct expand_vec_perm_d *d); | |
20459 | ||
4bf4c103 | 20460 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to simplify |
2bf6d935 ML |
20461 | a two vector permutation into a single vector permutation by using |
20462 | an interleave operation to merge the vectors. */ | |
20463 | ||
20464 | static bool | |
20465 | expand_vec_perm_interleave2 (struct expand_vec_perm_d *d) | |
20466 | { | |
20467 | struct expand_vec_perm_d dremap, dfinal; | |
20468 | unsigned i, nelt = d->nelt, nelt2 = nelt / 2; | |
20469 | unsigned HOST_WIDE_INT contents; | |
20470 | unsigned char remap[2 * MAX_VECT_LEN]; | |
20471 | rtx_insn *seq; | |
20472 | bool ok, same_halves = false; | |
20473 | ||
be8749f9 UB |
20474 | if (GET_MODE_SIZE (d->vmode) == 4 |
20475 | || GET_MODE_SIZE (d->vmode) == 8 | |
a325bdd1 | 20476 | || GET_MODE_SIZE (d->vmode) == 16) |
2bf6d935 ML |
20477 | { |
20478 | if (d->one_operand_p) | |
20479 | return false; | |
20480 | } | |
20481 | else if (GET_MODE_SIZE (d->vmode) == 32) | |
20482 | { | |
20483 | if (!TARGET_AVX) | |
20484 | return false; | |
20485 | /* For 32-byte modes allow even d->one_operand_p. | |
20486 | The lack of cross-lane shuffling in some instructions | |
20487 | might prevent a single insn shuffle. */ | |
20488 | dfinal = *d; | |
20489 | dfinal.testing_p = true; | |
20490 | /* If expand_vec_perm_interleave3 can expand this into | |
20491 | a 3 insn sequence, give up and let it be expanded as | |
20492 | 3 insn sequence. While that is one insn longer, | |
20493 | it doesn't need a memory operand and in the common | |
20494 | case that both interleave low and high permutations | |
20495 | with the same operands are adjacent needs 4 insns | |
20496 | for both after CSE. */ | |
20497 | if (expand_vec_perm_interleave3 (&dfinal)) | |
20498 | return false; | |
20499 | } | |
20500 | else | |
20501 | return false; | |
20502 | ||
20503 | /* Examine from whence the elements come. */ | |
20504 | contents = 0; | |
20505 | for (i = 0; i < nelt; ++i) | |
20506 | contents |= HOST_WIDE_INT_1U << d->perm[i]; | |
20507 | ||
20508 | memset (remap, 0xff, sizeof (remap)); | |
20509 | dremap = *d; | |
20510 | ||
be8749f9 UB |
20511 | if (GET_MODE_SIZE (d->vmode) == 4 |
20512 | || GET_MODE_SIZE (d->vmode) == 8) | |
a325bdd1 PB |
20513 | { |
20514 | unsigned HOST_WIDE_INT h1, h2, h3, h4; | |
20515 | ||
20516 | /* Split the two input vectors into 4 halves. */ | |
20517 | h1 = (HOST_WIDE_INT_1U << nelt2) - 1; | |
20518 | h2 = h1 << nelt2; | |
20519 | h3 = h2 << nelt2; | |
20520 | h4 = h3 << nelt2; | |
20521 | ||
20522 | /* If the elements from the low halves use interleave low, | |
20523 | and similarly for interleave high. */ | |
20524 | if ((contents & (h1 | h3)) == contents) | |
20525 | { | |
20526 | /* punpckl* */ | |
20527 | for (i = 0; i < nelt2; ++i) | |
20528 | { | |
20529 | remap[i] = i * 2; | |
20530 | remap[i + nelt] = i * 2 + 1; | |
20531 | dremap.perm[i * 2] = i; | |
20532 | dremap.perm[i * 2 + 1] = i + nelt; | |
20533 | } | |
20534 | } | |
20535 | else if ((contents & (h2 | h4)) == contents) | |
20536 | { | |
20537 | /* punpckh* */ | |
20538 | for (i = 0; i < nelt2; ++i) | |
20539 | { | |
20540 | remap[i + nelt2] = i * 2; | |
20541 | remap[i + nelt + nelt2] = i * 2 + 1; | |
20542 | dremap.perm[i * 2] = i + nelt2; | |
20543 | dremap.perm[i * 2 + 1] = i + nelt + nelt2; | |
20544 | } | |
20545 | } | |
20546 | else | |
20547 | return false; | |
20548 | } | |
20549 | else if (GET_MODE_SIZE (d->vmode) == 16) | |
2bf6d935 ML |
20550 | { |
20551 | unsigned HOST_WIDE_INT h1, h2, h3, h4; | |
20552 | ||
20553 | /* Split the two input vectors into 4 halves. */ | |
20554 | h1 = (HOST_WIDE_INT_1U << nelt2) - 1; | |
20555 | h2 = h1 << nelt2; | |
20556 | h3 = h2 << nelt2; | |
20557 | h4 = h3 << nelt2; | |
20558 | ||
20559 | /* If the elements from the low halves use interleave low, and similarly | |
20560 | for interleave high. If the elements are from mis-matched halves, we | |
20561 | can use shufps for V4SF/V4SI or do a DImode shuffle. */ | |
20562 | if ((contents & (h1 | h3)) == contents) | |
20563 | { | |
20564 | /* punpckl* */ | |
20565 | for (i = 0; i < nelt2; ++i) | |
20566 | { | |
20567 | remap[i] = i * 2; | |
20568 | remap[i + nelt] = i * 2 + 1; | |
20569 | dremap.perm[i * 2] = i; | |
20570 | dremap.perm[i * 2 + 1] = i + nelt; | |
20571 | } | |
20572 | if (!TARGET_SSE2 && d->vmode == V4SImode) | |
20573 | dremap.vmode = V4SFmode; | |
20574 | } | |
20575 | else if ((contents & (h2 | h4)) == contents) | |
20576 | { | |
20577 | /* punpckh* */ | |
20578 | for (i = 0; i < nelt2; ++i) | |
20579 | { | |
20580 | remap[i + nelt2] = i * 2; | |
20581 | remap[i + nelt + nelt2] = i * 2 + 1; | |
20582 | dremap.perm[i * 2] = i + nelt2; | |
20583 | dremap.perm[i * 2 + 1] = i + nelt + nelt2; | |
20584 | } | |
20585 | if (!TARGET_SSE2 && d->vmode == V4SImode) | |
20586 | dremap.vmode = V4SFmode; | |
20587 | } | |
20588 | else if ((contents & (h1 | h4)) == contents) | |
20589 | { | |
20590 | /* shufps */ | |
20591 | for (i = 0; i < nelt2; ++i) | |
20592 | { | |
20593 | remap[i] = i; | |
20594 | remap[i + nelt + nelt2] = i + nelt2; | |
20595 | dremap.perm[i] = i; | |
20596 | dremap.perm[i + nelt2] = i + nelt + nelt2; | |
20597 | } | |
20598 | if (nelt != 4) | |
20599 | { | |
20600 | /* shufpd */ | |
20601 | dremap.vmode = V2DImode; | |
20602 | dremap.nelt = 2; | |
20603 | dremap.perm[0] = 0; | |
20604 | dremap.perm[1] = 3; | |
20605 | } | |
20606 | } | |
20607 | else if ((contents & (h2 | h3)) == contents) | |
20608 | { | |
20609 | /* shufps */ | |
20610 | for (i = 0; i < nelt2; ++i) | |
20611 | { | |
20612 | remap[i + nelt2] = i; | |
20613 | remap[i + nelt] = i + nelt2; | |
20614 | dremap.perm[i] = i + nelt2; | |
20615 | dremap.perm[i + nelt2] = i + nelt; | |
20616 | } | |
20617 | if (nelt != 4) | |
20618 | { | |
20619 | /* shufpd */ | |
20620 | dremap.vmode = V2DImode; | |
20621 | dremap.nelt = 2; | |
20622 | dremap.perm[0] = 1; | |
20623 | dremap.perm[1] = 2; | |
20624 | } | |
20625 | } | |
20626 | else | |
20627 | return false; | |
20628 | } | |
20629 | else | |
20630 | { | |
20631 | unsigned int nelt4 = nelt / 4, nzcnt = 0; | |
20632 | unsigned HOST_WIDE_INT q[8]; | |
20633 | unsigned int nonzero_halves[4]; | |
20634 | ||
20635 | /* Split the two input vectors into 8 quarters. */ | |
20636 | q[0] = (HOST_WIDE_INT_1U << nelt4) - 1; | |
20637 | for (i = 1; i < 8; ++i) | |
20638 | q[i] = q[0] << (nelt4 * i); | |
20639 | for (i = 0; i < 4; ++i) | |
20640 | if (((q[2 * i] | q[2 * i + 1]) & contents) != 0) | |
20641 | { | |
20642 | nonzero_halves[nzcnt] = i; | |
20643 | ++nzcnt; | |
20644 | } | |
20645 | ||
20646 | if (nzcnt == 1) | |
20647 | { | |
20648 | gcc_assert (d->one_operand_p); | |
20649 | nonzero_halves[1] = nonzero_halves[0]; | |
20650 | same_halves = true; | |
20651 | } | |
20652 | else if (d->one_operand_p) | |
20653 | { | |
20654 | gcc_assert (nonzero_halves[0] == 0); | |
20655 | gcc_assert (nonzero_halves[1] == 1); | |
20656 | } | |
20657 | ||
20658 | if (nzcnt <= 2) | |
20659 | { | |
20660 | if (d->perm[0] / nelt2 == nonzero_halves[1]) | |
20661 | { | |
20662 | /* Attempt to increase the likelihood that dfinal | |
20663 | shuffle will be intra-lane. */ | |
20664 | std::swap (nonzero_halves[0], nonzero_halves[1]); | |
20665 | } | |
20666 | ||
20667 | /* vperm2f128 or vperm2i128. */ | |
20668 | for (i = 0; i < nelt2; ++i) | |
20669 | { | |
20670 | remap[i + nonzero_halves[1] * nelt2] = i + nelt2; | |
20671 | remap[i + nonzero_halves[0] * nelt2] = i; | |
20672 | dremap.perm[i + nelt2] = i + nonzero_halves[1] * nelt2; | |
20673 | dremap.perm[i] = i + nonzero_halves[0] * nelt2; | |
20674 | } | |
20675 | ||
20676 | if (d->vmode != V8SFmode | |
20677 | && d->vmode != V4DFmode | |
20678 | && d->vmode != V8SImode) | |
20679 | { | |
20680 | dremap.vmode = V8SImode; | |
20681 | dremap.nelt = 8; | |
20682 | for (i = 0; i < 4; ++i) | |
20683 | { | |
20684 | dremap.perm[i] = i + nonzero_halves[0] * 4; | |
20685 | dremap.perm[i + 4] = i + nonzero_halves[1] * 4; | |
20686 | } | |
20687 | } | |
20688 | } | |
20689 | else if (d->one_operand_p) | |
20690 | return false; | |
20691 | else if (TARGET_AVX2 | |
20692 | && (contents & (q[0] | q[2] | q[4] | q[6])) == contents) | |
20693 | { | |
20694 | /* vpunpckl* */ | |
20695 | for (i = 0; i < nelt4; ++i) | |
20696 | { | |
20697 | remap[i] = i * 2; | |
20698 | remap[i + nelt] = i * 2 + 1; | |
20699 | remap[i + nelt2] = i * 2 + nelt2; | |
20700 | remap[i + nelt + nelt2] = i * 2 + nelt2 + 1; | |
20701 | dremap.perm[i * 2] = i; | |
20702 | dremap.perm[i * 2 + 1] = i + nelt; | |
20703 | dremap.perm[i * 2 + nelt2] = i + nelt2; | |
20704 | dremap.perm[i * 2 + nelt2 + 1] = i + nelt + nelt2; | |
20705 | } | |
20706 | } | |
20707 | else if (TARGET_AVX2 | |
20708 | && (contents & (q[1] | q[3] | q[5] | q[7])) == contents) | |
20709 | { | |
20710 | /* vpunpckh* */ | |
20711 | for (i = 0; i < nelt4; ++i) | |
20712 | { | |
20713 | remap[i + nelt4] = i * 2; | |
20714 | remap[i + nelt + nelt4] = i * 2 + 1; | |
20715 | remap[i + nelt2 + nelt4] = i * 2 + nelt2; | |
20716 | remap[i + nelt + nelt2 + nelt4] = i * 2 + nelt2 + 1; | |
20717 | dremap.perm[i * 2] = i + nelt4; | |
20718 | dremap.perm[i * 2 + 1] = i + nelt + nelt4; | |
20719 | dremap.perm[i * 2 + nelt2] = i + nelt2 + nelt4; | |
20720 | dremap.perm[i * 2 + nelt2 + 1] = i + nelt + nelt2 + nelt4; | |
20721 | } | |
20722 | } | |
20723 | else | |
20724 | return false; | |
20725 | } | |
20726 | ||
20727 | /* Use the remapping array set up above to move the elements from their | |
20728 | swizzled locations into their final destinations. */ | |
20729 | dfinal = *d; | |
20730 | for (i = 0; i < nelt; ++i) | |
20731 | { | |
20732 | unsigned e = remap[d->perm[i]]; | |
20733 | gcc_assert (e < nelt); | |
20734 | /* If same_halves is true, both halves of the remapped vector are the | |
20735 | same. Avoid cross-lane accesses if possible. */ | |
20736 | if (same_halves && i >= nelt2) | |
20737 | { | |
20738 | gcc_assert (e < nelt2); | |
20739 | dfinal.perm[i] = e + nelt2; | |
20740 | } | |
20741 | else | |
20742 | dfinal.perm[i] = e; | |
20743 | } | |
20744 | if (!d->testing_p) | |
20745 | { | |
20746 | dremap.target = gen_reg_rtx (dremap.vmode); | |
20747 | dfinal.op0 = gen_lowpart (dfinal.vmode, dremap.target); | |
20748 | } | |
20749 | dfinal.op1 = dfinal.op0; | |
20750 | dfinal.one_operand_p = true; | |
20751 | ||
20752 | /* Test if the final remap can be done with a single insn. For V4SFmode or | |
20753 | V4SImode this *will* succeed. For V8HImode or V16QImode it may not. */ | |
20754 | start_sequence (); | |
20755 | ok = expand_vec_perm_1 (&dfinal); | |
20756 | seq = get_insns (); | |
20757 | end_sequence (); | |
20758 | ||
20759 | if (!ok) | |
20760 | return false; | |
20761 | ||
20762 | if (d->testing_p) | |
20763 | return true; | |
20764 | ||
20765 | if (dremap.vmode != dfinal.vmode) | |
20766 | { | |
20767 | dremap.op0 = gen_lowpart (dremap.vmode, dremap.op0); | |
20768 | dremap.op1 = gen_lowpart (dremap.vmode, dremap.op1); | |
20769 | } | |
20770 | ||
20771 | ok = expand_vec_perm_1 (&dremap); | |
20772 | gcc_assert (ok); | |
20773 | ||
20774 | emit_insn (seq); | |
20775 | return true; | |
20776 | } | |
20777 | ||
4bf4c103 | 20778 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to simplify |
2bf6d935 ML |
20779 | a single vector cross-lane permutation into vpermq followed |
20780 | by any of the single insn permutations. */ | |
20781 | ||
20782 | static bool | |
20783 | expand_vec_perm_vpermq_perm_1 (struct expand_vec_perm_d *d) | |
20784 | { | |
20785 | struct expand_vec_perm_d dremap, dfinal; | |
20786 | unsigned i, j, nelt = d->nelt, nelt2 = nelt / 2, nelt4 = nelt / 4; | |
20787 | unsigned contents[2]; | |
20788 | bool ok; | |
20789 | ||
20790 | if (!(TARGET_AVX2 | |
20791 | && (d->vmode == V32QImode || d->vmode == V16HImode) | |
20792 | && d->one_operand_p)) | |
20793 | return false; | |
20794 | ||
20795 | contents[0] = 0; | |
20796 | contents[1] = 0; | |
20797 | for (i = 0; i < nelt2; ++i) | |
20798 | { | |
20799 | contents[0] |= 1u << (d->perm[i] / nelt4); | |
20800 | contents[1] |= 1u << (d->perm[i + nelt2] / nelt4); | |
20801 | } | |
20802 | ||
20803 | for (i = 0; i < 2; ++i) | |
20804 | { | |
20805 | unsigned int cnt = 0; | |
20806 | for (j = 0; j < 4; ++j) | |
20807 | if ((contents[i] & (1u << j)) != 0 && ++cnt > 2) | |
20808 | return false; | |
20809 | } | |
20810 | ||
20811 | if (d->testing_p) | |
20812 | return true; | |
20813 | ||
20814 | dremap = *d; | |
20815 | dremap.vmode = V4DImode; | |
20816 | dremap.nelt = 4; | |
20817 | dremap.target = gen_reg_rtx (V4DImode); | |
20818 | dremap.op0 = gen_lowpart (V4DImode, d->op0); | |
20819 | dremap.op1 = dremap.op0; | |
20820 | dremap.one_operand_p = true; | |
20821 | for (i = 0; i < 2; ++i) | |
20822 | { | |
20823 | unsigned int cnt = 0; | |
20824 | for (j = 0; j < 4; ++j) | |
20825 | if ((contents[i] & (1u << j)) != 0) | |
20826 | dremap.perm[2 * i + cnt++] = j; | |
20827 | for (; cnt < 2; ++cnt) | |
20828 | dremap.perm[2 * i + cnt] = 0; | |
20829 | } | |
20830 | ||
20831 | dfinal = *d; | |
20832 | dfinal.op0 = gen_lowpart (dfinal.vmode, dremap.target); | |
20833 | dfinal.op1 = dfinal.op0; | |
20834 | dfinal.one_operand_p = true; | |
20835 | for (i = 0, j = 0; i < nelt; ++i) | |
20836 | { | |
20837 | if (i == nelt2) | |
20838 | j = 2; | |
20839 | dfinal.perm[i] = (d->perm[i] & (nelt4 - 1)) | (j ? nelt2 : 0); | |
20840 | if ((d->perm[i] / nelt4) == dremap.perm[j]) | |
20841 | ; | |
20842 | else if ((d->perm[i] / nelt4) == dremap.perm[j + 1]) | |
20843 | dfinal.perm[i] |= nelt4; | |
20844 | else | |
20845 | gcc_unreachable (); | |
20846 | } | |
20847 | ||
20848 | ok = expand_vec_perm_1 (&dremap); | |
20849 | gcc_assert (ok); | |
20850 | ||
20851 | ok = expand_vec_perm_1 (&dfinal); | |
20852 | gcc_assert (ok); | |
20853 | ||
20854 | return true; | |
20855 | } | |
20856 | ||
20857 | static bool canonicalize_perm (struct expand_vec_perm_d *d); | |
20858 | ||
4bf4c103 | 20859 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to expand |
2bf6d935 ML |
20860 | a vector permutation using two instructions, vperm2f128 resp. |
20861 | vperm2i128 followed by any single in-lane permutation. */ | |
20862 | ||
20863 | static bool | |
20864 | expand_vec_perm_vperm2f128 (struct expand_vec_perm_d *d) | |
20865 | { | |
20866 | struct expand_vec_perm_d dfirst, dsecond; | |
20867 | unsigned i, j, nelt = d->nelt, nelt2 = nelt / 2, perm; | |
20868 | bool ok; | |
20869 | ||
20870 | if (!TARGET_AVX | |
20871 | || GET_MODE_SIZE (d->vmode) != 32 | |
20872 | || (d->vmode != V8SFmode && d->vmode != V4DFmode && !TARGET_AVX2)) | |
20873 | return false; | |
20874 | ||
20875 | dsecond = *d; | |
20876 | dsecond.one_operand_p = false; | |
20877 | dsecond.testing_p = true; | |
20878 | ||
20879 | /* ((perm << 2)|perm) & 0x33 is the vperm2[fi]128 | |
20880 | immediate. For perm < 16 the second permutation uses | |
20881 | d->op0 as first operand, for perm >= 16 it uses d->op1 | |
20882 | as first operand. The second operand is the result of | |
20883 | vperm2[fi]128. */ | |
20884 | for (perm = 0; perm < 32; perm++) | |
20885 | { | |
20886 | /* Ignore permutations which do not move anything cross-lane. */ | |
20887 | if (perm < 16) | |
20888 | { | |
20889 | /* The second shuffle for e.g. V4DFmode has | |
20890 | 0123 and ABCD operands. | |
20891 | Ignore AB23, as 23 is already in the second lane | |
20892 | of the first operand. */ | |
20893 | if ((perm & 0xc) == (1 << 2)) continue; | |
20894 | /* And 01CD, as 01 is in the first lane of the first | |
20895 | operand. */ | |
20896 | if ((perm & 3) == 0) continue; | |
20897 | /* And 4567, as then the vperm2[fi]128 doesn't change | |
20898 | anything on the original 4567 second operand. */ | |
20899 | if ((perm & 0xf) == ((3 << 2) | 2)) continue; | |
20900 | } | |
20901 | else | |
20902 | { | |
20903 | /* The second shuffle for e.g. V4DFmode has | |
20904 | 4567 and ABCD operands. | |
20905 | Ignore AB67, as 67 is already in the second lane | |
20906 | of the first operand. */ | |
20907 | if ((perm & 0xc) == (3 << 2)) continue; | |
20908 | /* And 45CD, as 45 is in the first lane of the first | |
20909 | operand. */ | |
20910 | if ((perm & 3) == 2) continue; | |
20911 | /* And 0123, as then the vperm2[fi]128 doesn't change | |
20912 | anything on the original 0123 first operand. */ | |
20913 | if ((perm & 0xf) == (1 << 2)) continue; | |
20914 | } | |
20915 | ||
20916 | for (i = 0; i < nelt; i++) | |
20917 | { | |
20918 | j = d->perm[i] / nelt2; | |
20919 | if (j == ((perm >> (2 * (i >= nelt2))) & 3)) | |
20920 | dsecond.perm[i] = nelt + (i & nelt2) + (d->perm[i] & (nelt2 - 1)); | |
20921 | else if (j == (unsigned) (i >= nelt2) + 2 * (perm >= 16)) | |
20922 | dsecond.perm[i] = d->perm[i] & (nelt - 1); | |
20923 | else | |
20924 | break; | |
20925 | } | |
20926 | ||
20927 | if (i == nelt) | |
20928 | { | |
20929 | start_sequence (); | |
20930 | ok = expand_vec_perm_1 (&dsecond); | |
20931 | end_sequence (); | |
20932 | } | |
20933 | else | |
20934 | ok = false; | |
20935 | ||
20936 | if (ok) | |
20937 | { | |
20938 | if (d->testing_p) | |
20939 | return true; | |
20940 | ||
20941 | /* Found a usable second shuffle. dfirst will be | |
20942 | vperm2f128 on d->op0 and d->op1. */ | |
20943 | dsecond.testing_p = false; | |
20944 | dfirst = *d; | |
20945 | dfirst.target = gen_reg_rtx (d->vmode); | |
20946 | for (i = 0; i < nelt; i++) | |
20947 | dfirst.perm[i] = (i & (nelt2 - 1)) | |
20948 | + ((perm >> (2 * (i >= nelt2))) & 3) * nelt2; | |
20949 | ||
20950 | canonicalize_perm (&dfirst); | |
20951 | ok = expand_vec_perm_1 (&dfirst); | |
20952 | gcc_assert (ok); | |
20953 | ||
20954 | /* And dsecond is some single insn shuffle, taking | |
20955 | d->op0 and result of vperm2f128 (if perm < 16) or | |
20956 | d->op1 and result of vperm2f128 (otherwise). */ | |
20957 | if (perm >= 16) | |
20958 | dsecond.op0 = dsecond.op1; | |
20959 | dsecond.op1 = dfirst.target; | |
20960 | ||
20961 | ok = expand_vec_perm_1 (&dsecond); | |
20962 | gcc_assert (ok); | |
20963 | ||
20964 | return true; | |
20965 | } | |
20966 | ||
20967 | /* For one operand, the only useful vperm2f128 permutation is 0x01 | |
20968 | aka lanes swap. */ | |
20969 | if (d->one_operand_p) | |
20970 | return false; | |
20971 | } | |
20972 | ||
20973 | return false; | |
20974 | } | |
20975 | ||
4bf4c103 | 20976 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to simplify |
2bf6d935 ML |
20977 | a two vector permutation using 2 intra-lane interleave insns |
20978 | and cross-lane shuffle for 32-byte vectors. */ | |
20979 | ||
20980 | static bool | |
20981 | expand_vec_perm_interleave3 (struct expand_vec_perm_d *d) | |
20982 | { | |
20983 | unsigned i, nelt; | |
20984 | rtx (*gen) (rtx, rtx, rtx); | |
20985 | ||
20986 | if (d->one_operand_p) | |
20987 | return false; | |
20988 | if (TARGET_AVX2 && GET_MODE_SIZE (d->vmode) == 32) | |
20989 | ; | |
20990 | else if (TARGET_AVX && (d->vmode == V8SFmode || d->vmode == V4DFmode)) | |
20991 | ; | |
20992 | else | |
20993 | return false; | |
20994 | ||
20995 | nelt = d->nelt; | |
20996 | if (d->perm[0] != 0 && d->perm[0] != nelt / 2) | |
20997 | return false; | |
20998 | for (i = 0; i < nelt; i += 2) | |
20999 | if (d->perm[i] != d->perm[0] + i / 2 | |
21000 | || d->perm[i + 1] != d->perm[0] + i / 2 + nelt) | |
21001 | return false; | |
21002 | ||
21003 | if (d->testing_p) | |
21004 | return true; | |
21005 | ||
21006 | switch (d->vmode) | |
21007 | { | |
21008 | case E_V32QImode: | |
21009 | if (d->perm[0]) | |
21010 | gen = gen_vec_interleave_highv32qi; | |
21011 | else | |
21012 | gen = gen_vec_interleave_lowv32qi; | |
21013 | break; | |
21014 | case E_V16HImode: | |
21015 | if (d->perm[0]) | |
21016 | gen = gen_vec_interleave_highv16hi; | |
21017 | else | |
21018 | gen = gen_vec_interleave_lowv16hi; | |
21019 | break; | |
21020 | case E_V8SImode: | |
21021 | if (d->perm[0]) | |
21022 | gen = gen_vec_interleave_highv8si; | |
21023 | else | |
21024 | gen = gen_vec_interleave_lowv8si; | |
21025 | break; | |
21026 | case E_V4DImode: | |
21027 | if (d->perm[0]) | |
21028 | gen = gen_vec_interleave_highv4di; | |
21029 | else | |
21030 | gen = gen_vec_interleave_lowv4di; | |
21031 | break; | |
21032 | case E_V8SFmode: | |
21033 | if (d->perm[0]) | |
21034 | gen = gen_vec_interleave_highv8sf; | |
21035 | else | |
21036 | gen = gen_vec_interleave_lowv8sf; | |
21037 | break; | |
21038 | case E_V4DFmode: | |
21039 | if (d->perm[0]) | |
21040 | gen = gen_vec_interleave_highv4df; | |
21041 | else | |
21042 | gen = gen_vec_interleave_lowv4df; | |
21043 | break; | |
21044 | default: | |
21045 | gcc_unreachable (); | |
21046 | } | |
21047 | ||
21048 | emit_insn (gen (d->target, d->op0, d->op1)); | |
21049 | return true; | |
21050 | } | |
21051 | ||
4bf4c103 | 21052 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to implement |
2bf6d935 ML |
21053 | a single vector permutation using a single intra-lane vector |
21054 | permutation, vperm2f128 swapping the lanes and vblend* insn blending | |
21055 | the non-swapped and swapped vectors together. */ | |
21056 | ||
21057 | static bool | |
21058 | expand_vec_perm_vperm2f128_vblend (struct expand_vec_perm_d *d) | |
21059 | { | |
21060 | struct expand_vec_perm_d dfirst, dsecond; | |
21061 | unsigned i, j, msk, nelt = d->nelt, nelt2 = nelt / 2; | |
21062 | rtx_insn *seq; | |
21063 | bool ok; | |
21064 | rtx (*blend) (rtx, rtx, rtx, rtx) = NULL; | |
21065 | ||
21066 | if (!TARGET_AVX | |
21067 | || TARGET_AVX2 | |
21068 | || (d->vmode != V8SFmode && d->vmode != V4DFmode) | |
21069 | || !d->one_operand_p) | |
21070 | return false; | |
21071 | ||
21072 | dfirst = *d; | |
21073 | for (i = 0; i < nelt; i++) | |
21074 | dfirst.perm[i] = 0xff; | |
21075 | for (i = 0, msk = 0; i < nelt; i++) | |
21076 | { | |
21077 | j = (d->perm[i] & nelt2) ? i | nelt2 : i & ~nelt2; | |
21078 | if (dfirst.perm[j] != 0xff && dfirst.perm[j] != d->perm[i]) | |
21079 | return false; | |
21080 | dfirst.perm[j] = d->perm[i]; | |
21081 | if (j != i) | |
21082 | msk |= (1 << i); | |
21083 | } | |
21084 | for (i = 0; i < nelt; i++) | |
21085 | if (dfirst.perm[i] == 0xff) | |
21086 | dfirst.perm[i] = i; | |
21087 | ||
21088 | if (!d->testing_p) | |
21089 | dfirst.target = gen_reg_rtx (dfirst.vmode); | |
21090 | ||
21091 | start_sequence (); | |
21092 | ok = expand_vec_perm_1 (&dfirst); | |
21093 | seq = get_insns (); | |
21094 | end_sequence (); | |
21095 | ||
21096 | if (!ok) | |
21097 | return false; | |
21098 | ||
21099 | if (d->testing_p) | |
21100 | return true; | |
21101 | ||
21102 | emit_insn (seq); | |
21103 | ||
21104 | dsecond = *d; | |
21105 | dsecond.op0 = dfirst.target; | |
21106 | dsecond.op1 = dfirst.target; | |
21107 | dsecond.one_operand_p = true; | |
21108 | dsecond.target = gen_reg_rtx (dsecond.vmode); | |
21109 | for (i = 0; i < nelt; i++) | |
21110 | dsecond.perm[i] = i ^ nelt2; | |
21111 | ||
21112 | ok = expand_vec_perm_1 (&dsecond); | |
21113 | gcc_assert (ok); | |
21114 | ||
21115 | blend = d->vmode == V8SFmode ? gen_avx_blendps256 : gen_avx_blendpd256; | |
21116 | emit_insn (blend (d->target, dfirst.target, dsecond.target, GEN_INT (msk))); | |
21117 | return true; | |
21118 | } | |
21119 | ||
829c4bea JJ |
21120 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to implement |
21121 | a two vector permutation using two single vector permutations and | |
21122 | {,v}{,p}unpckl{ps,pd,bw,wd,dq}. If two_insn, succeed only if one | |
21123 | of dfirst or dsecond is identity permutation. */ | |
21124 | ||
21125 | static bool | |
21126 | expand_vec_perm_2perm_interleave (struct expand_vec_perm_d *d, bool two_insn) | |
21127 | { | |
21128 | unsigned i, nelt = d->nelt, nelt2 = nelt / 2, lane = nelt; | |
21129 | struct expand_vec_perm_d dfirst, dsecond, dfinal; | |
21130 | bool ident1 = true, ident2 = true; | |
21131 | ||
21132 | if (d->one_operand_p) | |
21133 | return false; | |
21134 | ||
21135 | if (GET_MODE_SIZE (d->vmode) == 16) | |
21136 | { | |
21137 | if (!TARGET_SSE) | |
21138 | return false; | |
21139 | if (d->vmode != V4SFmode && d->vmode != V2DFmode && !TARGET_SSE2) | |
21140 | return false; | |
21141 | } | |
21142 | else if (GET_MODE_SIZE (d->vmode) == 32) | |
21143 | { | |
21144 | if (!TARGET_AVX) | |
21145 | return false; | |
21146 | if (d->vmode != V8SFmode && d->vmode != V4DFmode && !TARGET_AVX2) | |
21147 | return false; | |
21148 | lane = nelt2; | |
21149 | } | |
21150 | else | |
21151 | return false; | |
21152 | ||
21153 | for (i = 1; i < nelt; i++) | |
21154 | if ((d->perm[i] >= nelt) != ((d->perm[0] >= nelt) ^ (i & 1))) | |
21155 | return false; | |
21156 | ||
21157 | dfirst = *d; | |
21158 | dsecond = *d; | |
21159 | dfinal = *d; | |
21160 | dfirst.op1 = dfirst.op0; | |
21161 | dfirst.one_operand_p = true; | |
21162 | dsecond.op0 = dsecond.op1; | |
21163 | dsecond.one_operand_p = true; | |
21164 | ||
21165 | for (i = 0; i < nelt; i++) | |
21166 | if (d->perm[i] >= nelt) | |
21167 | { | |
21168 | dsecond.perm[i / 2 + (i >= lane ? lane / 2 : 0)] = d->perm[i] - nelt; | |
21169 | if (d->perm[i] - nelt != i / 2 + (i >= lane ? lane / 2 : 0)) | |
21170 | ident2 = false; | |
21171 | dsecond.perm[i / 2 + (i >= lane ? lane : lane / 2)] | |
21172 | = d->perm[i] - nelt; | |
21173 | } | |
21174 | else | |
21175 | { | |
21176 | dfirst.perm[i / 2 + (i >= lane ? lane / 2 : 0)] = d->perm[i]; | |
21177 | if (d->perm[i] != i / 2 + (i >= lane ? lane / 2 : 0)) | |
21178 | ident1 = false; | |
21179 | dfirst.perm[i / 2 + (i >= lane ? lane : lane / 2)] = d->perm[i]; | |
21180 | } | |
21181 | ||
21182 | if (two_insn && !ident1 && !ident2) | |
21183 | return false; | |
21184 | ||
21185 | if (!d->testing_p) | |
21186 | { | |
21187 | if (!ident1) | |
21188 | dfinal.op0 = dfirst.target = gen_reg_rtx (d->vmode); | |
21189 | if (!ident2) | |
21190 | dfinal.op1 = dsecond.target = gen_reg_rtx (d->vmode); | |
21191 | if (d->perm[0] >= nelt) | |
21192 | std::swap (dfinal.op0, dfinal.op1); | |
21193 | } | |
21194 | ||
21195 | bool ok; | |
21196 | rtx_insn *seq1 = NULL, *seq2 = NULL; | |
21197 | ||
21198 | if (!ident1) | |
21199 | { | |
21200 | start_sequence (); | |
21201 | ok = expand_vec_perm_1 (&dfirst); | |
21202 | seq1 = get_insns (); | |
21203 | end_sequence (); | |
21204 | ||
21205 | if (!ok) | |
21206 | return false; | |
21207 | } | |
21208 | ||
21209 | if (!ident2) | |
21210 | { | |
21211 | start_sequence (); | |
21212 | ok = expand_vec_perm_1 (&dsecond); | |
21213 | seq2 = get_insns (); | |
21214 | end_sequence (); | |
21215 | ||
21216 | if (!ok) | |
21217 | return false; | |
21218 | } | |
21219 | ||
21220 | if (d->testing_p) | |
21221 | return true; | |
21222 | ||
21223 | for (i = 0; i < nelt; i++) | |
21224 | { | |
21225 | dfinal.perm[i] = i / 2; | |
21226 | if (i >= lane) | |
21227 | dfinal.perm[i] += lane / 2; | |
21228 | if ((i & 1) != 0) | |
21229 | dfinal.perm[i] += nelt; | |
21230 | } | |
21231 | emit_insn (seq1); | |
21232 | emit_insn (seq2); | |
21233 | ok = expand_vselect_vconcat (dfinal.target, dfinal.op0, dfinal.op1, | |
21234 | dfinal.perm, dfinal.nelt, false); | |
21235 | gcc_assert (ok); | |
21236 | return true; | |
21237 | } | |
21238 | ||
21239 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to simplify | |
21240 | the permutation using two single vector permutations and the SSE4_1 pblendv | |
21241 | instruction. If two_insn, succeed only if one of dfirst or dsecond is | |
21242 | identity permutation. */ | |
21243 | ||
21244 | static bool | |
21245 | expand_vec_perm_2perm_pblendv (struct expand_vec_perm_d *d, bool two_insn) | |
21246 | { | |
21247 | unsigned i, nelt = d->nelt; | |
21248 | struct expand_vec_perm_d dfirst, dsecond, dfinal; | |
21249 | machine_mode vmode = d->vmode; | |
21250 | bool ident1 = true, ident2 = true; | |
21251 | ||
21252 | /* Use the same checks as in expand_vec_perm_blend. */ | |
21253 | if (d->one_operand_p) | |
21254 | return false; | |
21255 | if (TARGET_AVX2 && GET_MODE_SIZE (vmode) == 32) | |
21256 | ; | |
21257 | else if (TARGET_AVX && (vmode == V4DFmode || vmode == V8SFmode)) | |
21258 | ; | |
acff89c7 UB |
21259 | else if (TARGET_SSE4_1 |
21260 | && (GET_MODE_SIZE (vmode) == 16 | |
21261 | || (TARGET_MMX_WITH_SSE && GET_MODE_SIZE (vmode) == 8) | |
21262 | || GET_MODE_SIZE (vmode) == 4)) | |
829c4bea JJ |
21263 | ; |
21264 | else | |
21265 | return false; | |
21266 | ||
21267 | dfirst = *d; | |
21268 | dsecond = *d; | |
21269 | dfinal = *d; | |
21270 | dfirst.op1 = dfirst.op0; | |
21271 | dfirst.one_operand_p = true; | |
21272 | dsecond.op0 = dsecond.op1; | |
21273 | dsecond.one_operand_p = true; | |
21274 | ||
21275 | for (i = 0; i < nelt; ++i) | |
21276 | if (d->perm[i] >= nelt) | |
21277 | { | |
21278 | dfirst.perm[i] = 0xff; | |
21279 | dsecond.perm[i] = d->perm[i] - nelt; | |
21280 | if (d->perm[i] != i + nelt) | |
21281 | ident2 = false; | |
21282 | } | |
21283 | else | |
21284 | { | |
21285 | dsecond.perm[i] = 0xff; | |
21286 | dfirst.perm[i] = d->perm[i]; | |
21287 | if (d->perm[i] != i) | |
21288 | ident1 = false; | |
21289 | } | |
21290 | ||
21291 | if (two_insn && !ident1 && !ident2) | |
21292 | return false; | |
21293 | ||
21294 | /* For now. Ideally treat 0xff as a wildcard. */ | |
21295 | for (i = 0; i < nelt; ++i) | |
21296 | if (dfirst.perm[i] == 0xff) | |
21297 | { | |
21298 | if (GET_MODE_SIZE (vmode) == 32 | |
21299 | && dfirst.perm[i ^ (nelt / 2)] != 0xff) | |
21300 | dfirst.perm[i] = dfirst.perm[i ^ (nelt / 2)] ^ (nelt / 2); | |
21301 | else | |
21302 | dfirst.perm[i] = i; | |
21303 | } | |
21304 | else | |
21305 | { | |
21306 | if (GET_MODE_SIZE (vmode) == 32 | |
21307 | && dsecond.perm[i ^ (nelt / 2)] != 0xff) | |
21308 | dsecond.perm[i] = dsecond.perm[i ^ (nelt / 2)] ^ (nelt / 2); | |
21309 | else | |
21310 | dsecond.perm[i] = i; | |
21311 | } | |
21312 | ||
21313 | if (!d->testing_p) | |
21314 | { | |
21315 | if (!ident1) | |
21316 | dfinal.op0 = dfirst.target = gen_reg_rtx (d->vmode); | |
21317 | if (!ident2) | |
21318 | dfinal.op1 = dsecond.target = gen_reg_rtx (d->vmode); | |
21319 | } | |
21320 | ||
21321 | bool ok; | |
21322 | rtx_insn *seq1 = NULL, *seq2 = NULL; | |
21323 | ||
21324 | if (!ident1) | |
21325 | { | |
21326 | start_sequence (); | |
21327 | ok = expand_vec_perm_1 (&dfirst); | |
21328 | seq1 = get_insns (); | |
21329 | end_sequence (); | |
21330 | ||
21331 | if (!ok) | |
21332 | return false; | |
21333 | } | |
21334 | ||
21335 | if (!ident2) | |
21336 | { | |
21337 | start_sequence (); | |
21338 | ok = expand_vec_perm_1 (&dsecond); | |
21339 | seq2 = get_insns (); | |
21340 | end_sequence (); | |
21341 | ||
21342 | if (!ok) | |
21343 | return false; | |
21344 | } | |
21345 | ||
21346 | if (d->testing_p) | |
21347 | return true; | |
21348 | ||
21349 | for (i = 0; i < nelt; ++i) | |
21350 | dfinal.perm[i] = (d->perm[i] >= nelt ? i + nelt : i); | |
21351 | ||
21352 | emit_insn (seq1); | |
21353 | emit_insn (seq2); | |
21354 | ok = expand_vec_perm_blend (&dfinal); | |
21355 | gcc_assert (ok); | |
21356 | return true; | |
21357 | } | |
21358 | ||
4bf4c103 | 21359 | /* A subroutine of ix86_expand_vec_perm_const_1. Implement a V4DF |
2bf6d935 ML |
21360 | permutation using two vperm2f128, followed by a vshufpd insn blending |
21361 | the two vectors together. */ | |
21362 | ||
21363 | static bool | |
21364 | expand_vec_perm_2vperm2f128_vshuf (struct expand_vec_perm_d *d) | |
21365 | { | |
21366 | struct expand_vec_perm_d dfirst, dsecond, dthird; | |
21367 | bool ok; | |
21368 | ||
21369 | if (!TARGET_AVX || (d->vmode != V4DFmode)) | |
21370 | return false; | |
21371 | ||
21372 | if (d->testing_p) | |
21373 | return true; | |
21374 | ||
21375 | dfirst = *d; | |
21376 | dsecond = *d; | |
21377 | dthird = *d; | |
21378 | ||
21379 | dfirst.perm[0] = (d->perm[0] & ~1); | |
21380 | dfirst.perm[1] = (d->perm[0] & ~1) + 1; | |
21381 | dfirst.perm[2] = (d->perm[2] & ~1); | |
21382 | dfirst.perm[3] = (d->perm[2] & ~1) + 1; | |
21383 | dsecond.perm[0] = (d->perm[1] & ~1); | |
21384 | dsecond.perm[1] = (d->perm[1] & ~1) + 1; | |
21385 | dsecond.perm[2] = (d->perm[3] & ~1); | |
21386 | dsecond.perm[3] = (d->perm[3] & ~1) + 1; | |
21387 | dthird.perm[0] = (d->perm[0] % 2); | |
21388 | dthird.perm[1] = (d->perm[1] % 2) + 4; | |
21389 | dthird.perm[2] = (d->perm[2] % 2) + 2; | |
21390 | dthird.perm[3] = (d->perm[3] % 2) + 6; | |
21391 | ||
21392 | dfirst.target = gen_reg_rtx (dfirst.vmode); | |
21393 | dsecond.target = gen_reg_rtx (dsecond.vmode); | |
21394 | dthird.op0 = dfirst.target; | |
21395 | dthird.op1 = dsecond.target; | |
21396 | dthird.one_operand_p = false; | |
21397 | ||
21398 | canonicalize_perm (&dfirst); | |
21399 | canonicalize_perm (&dsecond); | |
21400 | ||
21401 | ok = expand_vec_perm_1 (&dfirst) | |
21402 | && expand_vec_perm_1 (&dsecond) | |
21403 | && expand_vec_perm_1 (&dthird); | |
21404 | ||
21405 | gcc_assert (ok); | |
21406 | ||
21407 | return true; | |
21408 | } | |
21409 | ||
4bf4c103 JJ |
21410 | static bool ix86_expand_vec_perm_const_1 (struct expand_vec_perm_d *); |
21411 | ||
21412 | /* A subroutine of ix86_expand_vec_perm_const_1. Try to implement | |
21413 | a two vector permutation using two intra-lane vector | |
21414 | permutations, vperm2f128 swapping the lanes and vblend* insn blending | |
21415 | the non-swapped and swapped vectors together. */ | |
21416 | ||
21417 | static bool | |
21418 | expand_vec_perm2_vperm2f128_vblend (struct expand_vec_perm_d *d) | |
21419 | { | |
21420 | struct expand_vec_perm_d dfirst, dsecond, dthird; | |
21421 | unsigned i, j, msk, nelt = d->nelt, nelt2 = nelt / 2, which1 = 0, which2 = 0; | |
21422 | rtx_insn *seq1, *seq2; | |
21423 | bool ok; | |
21424 | rtx (*blend) (rtx, rtx, rtx, rtx) = NULL; | |
21425 | ||
21426 | if (!TARGET_AVX | |
21427 | || TARGET_AVX2 | |
21428 | || (d->vmode != V8SFmode && d->vmode != V4DFmode) | |
21429 | || d->one_operand_p) | |
21430 | return false; | |
21431 | ||
21432 | dfirst = *d; | |
21433 | dsecond = *d; | |
21434 | for (i = 0; i < nelt; i++) | |
21435 | { | |
21436 | dfirst.perm[i] = 0xff; | |
21437 | dsecond.perm[i] = 0xff; | |
21438 | } | |
21439 | for (i = 0, msk = 0; i < nelt; i++) | |
21440 | { | |
21441 | j = (d->perm[i] & nelt2) ? i | nelt2 : i & ~nelt2; | |
21442 | if (j == i) | |
21443 | { | |
21444 | dfirst.perm[j] = d->perm[i]; | |
21445 | which1 |= (d->perm[i] < nelt ? 1 : 2); | |
21446 | } | |
21447 | else | |
21448 | { | |
21449 | dsecond.perm[j] = d->perm[i]; | |
21450 | which2 |= (d->perm[i] < nelt ? 1 : 2); | |
21451 | msk |= (1U << i); | |
21452 | } | |
21453 | } | |
21454 | if (msk == 0 || msk == (1U << nelt) - 1) | |
21455 | return false; | |
21456 | ||
21457 | if (!d->testing_p) | |
21458 | { | |
21459 | dfirst.target = gen_reg_rtx (dfirst.vmode); | |
21460 | dsecond.target = gen_reg_rtx (dsecond.vmode); | |
21461 | } | |
21462 | ||
21463 | for (i = 0; i < nelt; i++) | |
21464 | { | |
21465 | if (dfirst.perm[i] == 0xff) | |
21466 | dfirst.perm[i] = (which1 == 2 ? i + nelt : i); | |
21467 | if (dsecond.perm[i] == 0xff) | |
21468 | dsecond.perm[i] = (which2 == 2 ? i + nelt : i); | |
21469 | } | |
21470 | canonicalize_perm (&dfirst); | |
21471 | start_sequence (); | |
21472 | ok = ix86_expand_vec_perm_const_1 (&dfirst); | |
21473 | seq1 = get_insns (); | |
21474 | end_sequence (); | |
21475 | ||
21476 | if (!ok) | |
21477 | return false; | |
21478 | ||
21479 | canonicalize_perm (&dsecond); | |
21480 | start_sequence (); | |
21481 | ok = ix86_expand_vec_perm_const_1 (&dsecond); | |
21482 | seq2 = get_insns (); | |
21483 | end_sequence (); | |
21484 | ||
21485 | if (!ok) | |
21486 | return false; | |
21487 | ||
21488 | if (d->testing_p) | |
21489 | return true; | |
21490 | ||
21491 | emit_insn (seq1); | |
21492 | emit_insn (seq2); | |
21493 | ||
21494 | dthird = *d; | |
21495 | dthird.op0 = dsecond.target; | |
21496 | dthird.op1 = dsecond.target; | |
21497 | dthird.one_operand_p = true; | |
21498 | dthird.target = gen_reg_rtx (dthird.vmode); | |
21499 | for (i = 0; i < nelt; i++) | |
21500 | dthird.perm[i] = i ^ nelt2; | |
21501 | ||
21502 | ok = expand_vec_perm_1 (&dthird); | |
21503 | gcc_assert (ok); | |
21504 | ||
21505 | blend = d->vmode == V8SFmode ? gen_avx_blendps256 : gen_avx_blendpd256; | |
21506 | emit_insn (blend (d->target, dfirst.target, dthird.target, GEN_INT (msk))); | |
21507 | return true; | |
21508 | } | |
21509 | ||
2bf6d935 ML |
21510 | /* A subroutine of expand_vec_perm_even_odd_1. Implement the double-word |
21511 | permutation with two pshufb insns and an ior. We should have already | |
21512 | failed all two instruction sequences. */ | |
21513 | ||
21514 | static bool | |
21515 | expand_vec_perm_pshufb2 (struct expand_vec_perm_d *d) | |
21516 | { | |
21517 | rtx rperm[2][16], vperm, l, h, op, m128; | |
21518 | unsigned int i, nelt, eltsz; | |
dd835ec2 UB |
21519 | machine_mode mode; |
21520 | rtx (*gen) (rtx, rtx, rtx); | |
2bf6d935 | 21521 | |
dd835ec2 | 21522 | if (!TARGET_SSSE3 || (GET_MODE_SIZE (d->vmode) != 16 |
be8749f9 UB |
21523 | && GET_MODE_SIZE (d->vmode) != 8 |
21524 | && GET_MODE_SIZE (d->vmode) != 4)) | |
2bf6d935 ML |
21525 | return false; |
21526 | gcc_assert (!d->one_operand_p); | |
21527 | ||
21528 | if (d->testing_p) | |
21529 | return true; | |
21530 | ||
dd835ec2 UB |
21531 | switch (GET_MODE_SIZE (d->vmode)) |
21532 | { | |
be8749f9 UB |
21533 | case 4: |
21534 | mode = V4QImode; | |
21535 | gen = gen_mmx_pshufbv4qi3; | |
21536 | break; | |
dd835ec2 UB |
21537 | case 8: |
21538 | mode = V8QImode; | |
21539 | gen = gen_mmx_pshufbv8qi3; | |
21540 | break; | |
21541 | case 16: | |
21542 | mode = V16QImode; | |
21543 | gen = gen_ssse3_pshufbv16qi3; | |
21544 | break; | |
21545 | default: | |
21546 | gcc_unreachable (); | |
21547 | } | |
21548 | ||
2bf6d935 ML |
21549 | nelt = d->nelt; |
21550 | eltsz = GET_MODE_UNIT_SIZE (d->vmode); | |
21551 | ||
21552 | /* Generate two permutation masks. If the required element is within | |
21553 | the given vector it is shuffled into the proper lane. If the required | |
21554 | element is in the other vector, force a zero into the lane by setting | |
21555 | bit 7 in the permutation mask. */ | |
21556 | m128 = GEN_INT (-128); | |
21557 | for (i = 0; i < nelt; ++i) | |
21558 | { | |
dd835ec2 | 21559 | unsigned j, k, e = d->perm[i]; |
2bf6d935 ML |
21560 | unsigned which = (e >= nelt); |
21561 | if (e >= nelt) | |
21562 | e -= nelt; | |
21563 | ||
21564 | for (j = 0; j < eltsz; ++j) | |
21565 | { | |
21566 | rperm[which][i*eltsz + j] = GEN_INT (e*eltsz + j); | |
21567 | rperm[1-which][i*eltsz + j] = m128; | |
21568 | } | |
dd835ec2 UB |
21569 | |
21570 | for (k = i*eltsz + j; k < 16; ++k) | |
21571 | rperm[0][k] = rperm[1][k] = m128; | |
2bf6d935 ML |
21572 | } |
21573 | ||
21574 | vperm = gen_rtx_CONST_VECTOR (V16QImode, gen_rtvec_v (16, rperm[0])); | |
21575 | vperm = force_reg (V16QImode, vperm); | |
21576 | ||
dd835ec2 UB |
21577 | l = gen_reg_rtx (mode); |
21578 | op = gen_lowpart (mode, d->op0); | |
21579 | emit_insn (gen (l, op, vperm)); | |
2bf6d935 ML |
21580 | |
21581 | vperm = gen_rtx_CONST_VECTOR (V16QImode, gen_rtvec_v (16, rperm[1])); | |
21582 | vperm = force_reg (V16QImode, vperm); | |
21583 | ||
dd835ec2 UB |
21584 | h = gen_reg_rtx (mode); |
21585 | op = gen_lowpart (mode, d->op1); | |
21586 | emit_insn (gen (h, op, vperm)); | |
2bf6d935 ML |
21587 | |
21588 | op = d->target; | |
dd835ec2 UB |
21589 | if (d->vmode != mode) |
21590 | op = gen_reg_rtx (mode); | |
b5193e35 | 21591 | ix86_emit_vec_binop (IOR, mode, op, l, h); |
2bf6d935 ML |
21592 | if (op != d->target) |
21593 | emit_move_insn (d->target, gen_lowpart (d->vmode, op)); | |
21594 | ||
21595 | return true; | |
21596 | } | |
21597 | ||
21598 | /* Implement arbitrary permutation of one V32QImode and V16QImode operand | |
21599 | with two vpshufb insns, vpermq and vpor. We should have already failed | |
21600 | all two or three instruction sequences. */ | |
21601 | ||
21602 | static bool | |
21603 | expand_vec_perm_vpshufb2_vpermq (struct expand_vec_perm_d *d) | |
21604 | { | |
21605 | rtx rperm[2][32], vperm, l, h, hp, op, m128; | |
21606 | unsigned int i, nelt, eltsz; | |
21607 | ||
21608 | if (!TARGET_AVX2 | |
21609 | || !d->one_operand_p | |
21610 | || (d->vmode != V32QImode && d->vmode != V16HImode)) | |
21611 | return false; | |
21612 | ||
21613 | if (d->testing_p) | |
21614 | return true; | |
21615 | ||
21616 | nelt = d->nelt; | |
21617 | eltsz = GET_MODE_UNIT_SIZE (d->vmode); | |
21618 | ||
21619 | /* Generate two permutation masks. If the required element is within | |
21620 | the same lane, it is shuffled in. If the required element from the | |
21621 | other lane, force a zero by setting bit 7 in the permutation mask. | |
21622 | In the other mask the mask has non-negative elements if element | |
21623 | is requested from the other lane, but also moved to the other lane, | |
21624 | so that the result of vpshufb can have the two V2TImode halves | |
21625 | swapped. */ | |
21626 | m128 = GEN_INT (-128); | |
21627 | for (i = 0; i < nelt; ++i) | |
21628 | { | |
21629 | unsigned j, e = d->perm[i] & (nelt / 2 - 1); | |
21630 | unsigned which = ((d->perm[i] ^ i) & (nelt / 2)) * eltsz; | |
21631 | ||
21632 | for (j = 0; j < eltsz; ++j) | |
21633 | { | |
21634 | rperm[!!which][(i * eltsz + j) ^ which] = GEN_INT (e * eltsz + j); | |
21635 | rperm[!which][(i * eltsz + j) ^ (which ^ 16)] = m128; | |
21636 | } | |
21637 | } | |
21638 | ||
21639 | vperm = gen_rtx_CONST_VECTOR (V32QImode, gen_rtvec_v (32, rperm[1])); | |
21640 | vperm = force_reg (V32QImode, vperm); | |
21641 | ||
21642 | h = gen_reg_rtx (V32QImode); | |
21643 | op = gen_lowpart (V32QImode, d->op0); | |
21644 | emit_insn (gen_avx2_pshufbv32qi3 (h, op, vperm)); | |
21645 | ||
21646 | /* Swap the 128-byte lanes of h into hp. */ | |
21647 | hp = gen_reg_rtx (V4DImode); | |
21648 | op = gen_lowpart (V4DImode, h); | |
21649 | emit_insn (gen_avx2_permv4di_1 (hp, op, const2_rtx, GEN_INT (3), const0_rtx, | |
21650 | const1_rtx)); | |
21651 | ||
21652 | vperm = gen_rtx_CONST_VECTOR (V32QImode, gen_rtvec_v (32, rperm[0])); | |
21653 | vperm = force_reg (V32QImode, vperm); | |
21654 | ||
21655 | l = gen_reg_rtx (V32QImode); | |
21656 | op = gen_lowpart (V32QImode, d->op0); | |
21657 | emit_insn (gen_avx2_pshufbv32qi3 (l, op, vperm)); | |
21658 | ||
21659 | op = d->target; | |
21660 | if (d->vmode != V32QImode) | |
21661 | op = gen_reg_rtx (V32QImode); | |
21662 | emit_insn (gen_iorv32qi3 (op, l, gen_lowpart (V32QImode, hp))); | |
21663 | if (op != d->target) | |
21664 | emit_move_insn (d->target, gen_lowpart (d->vmode, op)); | |
21665 | ||
21666 | return true; | |
21667 | } | |
21668 | ||
21669 | /* A subroutine of expand_vec_perm_even_odd_1. Implement extract-even | |
21670 | and extract-odd permutations of two V32QImode and V16QImode operand | |
21671 | with two vpshufb insns, vpor and vpermq. We should have already | |
21672 | failed all two or three instruction sequences. */ | |
21673 | ||
21674 | static bool | |
21675 | expand_vec_perm_vpshufb2_vpermq_even_odd (struct expand_vec_perm_d *d) | |
21676 | { | |
21677 | rtx rperm[2][32], vperm, l, h, ior, op, m128; | |
21678 | unsigned int i, nelt, eltsz; | |
21679 | ||
21680 | if (!TARGET_AVX2 | |
21681 | || d->one_operand_p | |
21682 | || (d->vmode != V32QImode && d->vmode != V16HImode)) | |
21683 | return false; | |
21684 | ||
21685 | for (i = 0; i < d->nelt; ++i) | |
21686 | if ((d->perm[i] ^ (i * 2)) & (3 * d->nelt / 2)) | |
21687 | return false; | |
21688 | ||
21689 | if (d->testing_p) | |
21690 | return true; | |
21691 | ||
21692 | nelt = d->nelt; | |
21693 | eltsz = GET_MODE_UNIT_SIZE (d->vmode); | |
21694 | ||
21695 | /* Generate two permutation masks. In the first permutation mask | |
21696 | the first quarter will contain indexes for the first half | |
21697 | of the op0, the second quarter will contain bit 7 set, third quarter | |
21698 | will contain indexes for the second half of the op0 and the | |
21699 | last quarter bit 7 set. In the second permutation mask | |
21700 | the first quarter will contain bit 7 set, the second quarter | |
21701 | indexes for the first half of the op1, the third quarter bit 7 set | |
21702 | and last quarter indexes for the second half of the op1. | |
21703 | I.e. the first mask e.g. for V32QImode extract even will be: | |
21704 | 0, 2, ..., 0xe, -128, ..., -128, 0, 2, ..., 0xe, -128, ..., -128 | |
21705 | (all values masked with 0xf except for -128) and second mask | |
21706 | for extract even will be | |
21707 | -128, ..., -128, 0, 2, ..., 0xe, -128, ..., -128, 0, 2, ..., 0xe. */ | |
21708 | m128 = GEN_INT (-128); | |
21709 | for (i = 0; i < nelt; ++i) | |
21710 | { | |
21711 | unsigned j, e = d->perm[i] & (nelt / 2 - 1); | |
21712 | unsigned which = d->perm[i] >= nelt; | |
21713 | unsigned xorv = (i >= nelt / 4 && i < 3 * nelt / 4) ? 24 : 0; | |
21714 | ||
21715 | for (j = 0; j < eltsz; ++j) | |
21716 | { | |
21717 | rperm[which][(i * eltsz + j) ^ xorv] = GEN_INT (e * eltsz + j); | |
21718 | rperm[1 - which][(i * eltsz + j) ^ xorv] = m128; | |
21719 | } | |
21720 | } | |
21721 | ||
21722 | vperm = gen_rtx_CONST_VECTOR (V32QImode, gen_rtvec_v (32, rperm[0])); | |
21723 | vperm = force_reg (V32QImode, vperm); | |
21724 | ||
21725 | l = gen_reg_rtx (V32QImode); | |
21726 | op = gen_lowpart (V32QImode, d->op0); | |
21727 | emit_insn (gen_avx2_pshufbv32qi3 (l, op, vperm)); | |
21728 | ||
21729 | vperm = gen_rtx_CONST_VECTOR (V32QImode, gen_rtvec_v (32, rperm[1])); | |
21730 | vperm = force_reg (V32QImode, vperm); | |
21731 | ||
21732 | h = gen_reg_rtx (V32QImode); | |
21733 | op = gen_lowpart (V32QImode, d->op1); | |
21734 | emit_insn (gen_avx2_pshufbv32qi3 (h, op, vperm)); | |
21735 | ||
21736 | ior = gen_reg_rtx (V32QImode); | |
21737 | emit_insn (gen_iorv32qi3 (ior, l, h)); | |
21738 | ||
21739 | /* Permute the V4DImode quarters using { 0, 2, 1, 3 } permutation. */ | |
21740 | op = gen_reg_rtx (V4DImode); | |
21741 | ior = gen_lowpart (V4DImode, ior); | |
21742 | emit_insn (gen_avx2_permv4di_1 (op, ior, const0_rtx, const2_rtx, | |
21743 | const1_rtx, GEN_INT (3))); | |
21744 | emit_move_insn (d->target, gen_lowpart (d->vmode, op)); | |
21745 | ||
21746 | return true; | |
21747 | } | |
21748 | ||
fcda0efc | 21749 | /* Implement permutation with pslldq + psrldq + por when pshufb is not |
21750 | available. */ | |
21751 | static bool | |
21752 | expand_vec_perm_pslldq_psrldq_por (struct expand_vec_perm_d *d, bool pandn) | |
21753 | { | |
21754 | unsigned i, nelt = d->nelt; | |
21755 | unsigned start1, end1 = -1; | |
21756 | machine_mode vmode = d->vmode, imode; | |
21757 | int start2 = -1; | |
21758 | bool clear_op0, clear_op1; | |
21759 | unsigned inner_size; | |
21760 | rtx op0, op1, dop1; | |
21761 | rtx (*gen_vec_shr) (rtx, rtx, rtx); | |
21762 | rtx (*gen_vec_shl) (rtx, rtx, rtx); | |
21763 | ||
21764 | /* pshufd can be used for V4SI/V2DI under TARGET_SSE2. */ | |
21765 | if (!TARGET_SSE2 || (vmode != E_V16QImode && vmode != E_V8HImode)) | |
21766 | return false; | |
21767 | ||
21768 | start1 = d->perm[0]; | |
21769 | for (i = 1; i < nelt; i++) | |
21770 | { | |
69c4b5c5 | 21771 | if (d->perm[i] != d->perm[i-1] + 1 |
21772 | || d->perm[i] == nelt) | |
fcda0efc | 21773 | { |
21774 | if (start2 == -1) | |
21775 | { | |
21776 | start2 = d->perm[i]; | |
21777 | end1 = d->perm[i-1]; | |
21778 | } | |
21779 | else | |
21780 | return false; | |
21781 | } | |
fcda0efc | 21782 | } |
21783 | ||
21784 | clear_op0 = end1 != nelt - 1; | |
21785 | clear_op1 = start2 % nelt != 0; | |
21786 | /* pandn/pand is needed to clear upper/lower bits of op0/op1. */ | |
21787 | if (!pandn && (clear_op0 || clear_op1)) | |
21788 | return false; | |
21789 | ||
21790 | if (d->testing_p) | |
21791 | return true; | |
21792 | ||
21793 | gen_vec_shr = vmode == E_V16QImode ? gen_vec_shr_v16qi : gen_vec_shr_v8hi; | |
21794 | gen_vec_shl = vmode == E_V16QImode ? gen_vec_shl_v16qi : gen_vec_shl_v8hi; | |
21795 | imode = GET_MODE_INNER (vmode); | |
21796 | inner_size = GET_MODE_BITSIZE (imode); | |
21797 | op0 = gen_reg_rtx (vmode); | |
21798 | op1 = gen_reg_rtx (vmode); | |
21799 | ||
21800 | if (start1) | |
21801 | emit_insn (gen_vec_shr (op0, d->op0, GEN_INT (start1 * inner_size))); | |
21802 | else | |
21803 | emit_move_insn (op0, d->op0); | |
21804 | ||
21805 | dop1 = d->op1; | |
21806 | if (d->one_operand_p) | |
21807 | dop1 = d->op0; | |
21808 | ||
21809 | int shl_offset = end1 - start1 + 1 - start2 % nelt; | |
21810 | if (shl_offset) | |
21811 | emit_insn (gen_vec_shl (op1, dop1, GEN_INT (shl_offset * inner_size))); | |
21812 | else | |
21813 | emit_move_insn (op1, dop1); | |
21814 | ||
21815 | /* Clear lower/upper bits for op0/op1. */ | |
21816 | if (clear_op0 || clear_op1) | |
21817 | { | |
21818 | rtx vec[16]; | |
21819 | rtx const_vec; | |
21820 | rtx clear; | |
21821 | for (i = 0; i != nelt; i++) | |
21822 | { | |
21823 | if (i < (end1 - start1 + 1)) | |
21824 | vec[i] = gen_int_mode ((HOST_WIDE_INT_1U << inner_size) - 1, imode); | |
21825 | else | |
21826 | vec[i] = CONST0_RTX (imode); | |
21827 | } | |
21828 | const_vec = gen_rtx_CONST_VECTOR (vmode, gen_rtvec_v (nelt, vec)); | |
21829 | const_vec = validize_mem (force_const_mem (vmode, const_vec)); | |
21830 | clear = force_reg (vmode, const_vec); | |
21831 | ||
21832 | if (clear_op0) | |
21833 | emit_move_insn (op0, gen_rtx_AND (vmode, op0, clear)); | |
21834 | if (clear_op1) | |
21835 | emit_move_insn (op1, gen_rtx_AND (vmode, | |
21836 | gen_rtx_NOT (vmode, clear), | |
21837 | op1)); | |
21838 | } | |
21839 | ||
21840 | emit_move_insn (d->target, gen_rtx_IOR (vmode, op0, op1)); | |
21841 | return true; | |
21842 | } | |
21843 | ||
2bf6d935 | 21844 | /* A subroutine of expand_vec_perm_even_odd_1. Implement extract-even |
a325bdd1 PB |
21845 | and extract-odd permutations of two V8QI, V8HI, V16QI, V16HI or V32QI |
21846 | operands with two "and" and "pack" or two "shift" and "pack" insns. | |
21847 | We should have already failed all two instruction sequences. */ | |
2bf6d935 ML |
21848 | |
21849 | static bool | |
21850 | expand_vec_perm_even_odd_pack (struct expand_vec_perm_d *d) | |
21851 | { | |
21852 | rtx op, dop0, dop1, t; | |
21853 | unsigned i, odd, c, s, nelt = d->nelt; | |
21854 | bool end_perm = false; | |
21855 | machine_mode half_mode; | |
21856 | rtx (*gen_and) (rtx, rtx, rtx); | |
21857 | rtx (*gen_pack) (rtx, rtx, rtx); | |
21858 | rtx (*gen_shift) (rtx, rtx, rtx); | |
21859 | ||
21860 | if (d->one_operand_p) | |
21861 | return false; | |
21862 | ||
21863 | switch (d->vmode) | |
21864 | { | |
dd835ec2 UB |
21865 | case E_V4HImode: |
21866 | /* Required for "pack". */ | |
21867 | if (!TARGET_SSE4_1) | |
21868 | return false; | |
21869 | c = 0xffff; | |
21870 | s = 16; | |
21871 | half_mode = V2SImode; | |
21872 | gen_and = gen_andv2si3; | |
21873 | gen_pack = gen_mmx_packusdw; | |
21874 | gen_shift = gen_lshrv2si3; | |
21875 | break; | |
2bf6d935 ML |
21876 | case E_V8HImode: |
21877 | /* Required for "pack". */ | |
21878 | if (!TARGET_SSE4_1) | |
21879 | return false; | |
21880 | c = 0xffff; | |
21881 | s = 16; | |
21882 | half_mode = V4SImode; | |
21883 | gen_and = gen_andv4si3; | |
21884 | gen_pack = gen_sse4_1_packusdw; | |
21885 | gen_shift = gen_lshrv4si3; | |
21886 | break; | |
a325bdd1 PB |
21887 | case E_V8QImode: |
21888 | /* No check as all instructions are SSE2. */ | |
21889 | c = 0xff; | |
21890 | s = 8; | |
21891 | half_mode = V4HImode; | |
21892 | gen_and = gen_andv4hi3; | |
21893 | gen_pack = gen_mmx_packuswb; | |
21894 | gen_shift = gen_lshrv4hi3; | |
21895 | break; | |
2bf6d935 ML |
21896 | case E_V16QImode: |
21897 | /* No check as all instructions are SSE2. */ | |
21898 | c = 0xff; | |
21899 | s = 8; | |
21900 | half_mode = V8HImode; | |
21901 | gen_and = gen_andv8hi3; | |
21902 | gen_pack = gen_sse2_packuswb; | |
21903 | gen_shift = gen_lshrv8hi3; | |
21904 | break; | |
21905 | case E_V16HImode: | |
21906 | if (!TARGET_AVX2) | |
21907 | return false; | |
21908 | c = 0xffff; | |
21909 | s = 16; | |
21910 | half_mode = V8SImode; | |
21911 | gen_and = gen_andv8si3; | |
21912 | gen_pack = gen_avx2_packusdw; | |
21913 | gen_shift = gen_lshrv8si3; | |
21914 | end_perm = true; | |
21915 | break; | |
21916 | case E_V32QImode: | |
21917 | if (!TARGET_AVX2) | |
21918 | return false; | |
21919 | c = 0xff; | |
21920 | s = 8; | |
21921 | half_mode = V16HImode; | |
21922 | gen_and = gen_andv16hi3; | |
21923 | gen_pack = gen_avx2_packuswb; | |
21924 | gen_shift = gen_lshrv16hi3; | |
21925 | end_perm = true; | |
21926 | break; | |
21927 | default: | |
dd835ec2 | 21928 | /* Only V4HI, V8QI, V8HI, V16QI, V16HI and V32QI modes |
a325bdd1 | 21929 | are more profitable than general shuffles. */ |
2bf6d935 ML |
21930 | return false; |
21931 | } | |
21932 | ||
21933 | /* Check that permutation is even or odd. */ | |
21934 | odd = d->perm[0]; | |
21935 | if (odd > 1) | |
21936 | return false; | |
21937 | ||
21938 | for (i = 1; i < nelt; ++i) | |
21939 | if (d->perm[i] != 2 * i + odd) | |
21940 | return false; | |
21941 | ||
21942 | if (d->testing_p) | |
21943 | return true; | |
21944 | ||
21945 | dop0 = gen_reg_rtx (half_mode); | |
21946 | dop1 = gen_reg_rtx (half_mode); | |
21947 | if (odd == 0) | |
21948 | { | |
21949 | t = gen_const_vec_duplicate (half_mode, GEN_INT (c)); | |
21950 | t = force_reg (half_mode, t); | |
21951 | emit_insn (gen_and (dop0, t, gen_lowpart (half_mode, d->op0))); | |
21952 | emit_insn (gen_and (dop1, t, gen_lowpart (half_mode, d->op1))); | |
21953 | } | |
21954 | else | |
21955 | { | |
21956 | emit_insn (gen_shift (dop0, | |
21957 | gen_lowpart (half_mode, d->op0), | |
21958 | GEN_INT (s))); | |
21959 | emit_insn (gen_shift (dop1, | |
21960 | gen_lowpart (half_mode, d->op1), | |
21961 | GEN_INT (s))); | |
21962 | } | |
21963 | /* In AVX2 for 256 bit case we need to permute pack result. */ | |
21964 | if (TARGET_AVX2 && end_perm) | |
21965 | { | |
21966 | op = gen_reg_rtx (d->vmode); | |
21967 | t = gen_reg_rtx (V4DImode); | |
21968 | emit_insn (gen_pack (op, dop0, dop1)); | |
21969 | emit_insn (gen_avx2_permv4di_1 (t, | |
21970 | gen_lowpart (V4DImode, op), | |
21971 | const0_rtx, | |
21972 | const2_rtx, | |
21973 | const1_rtx, | |
21974 | GEN_INT (3))); | |
21975 | emit_move_insn (d->target, gen_lowpart (d->vmode, t)); | |
21976 | } | |
21977 | else | |
21978 | emit_insn (gen_pack (d->target, dop0, dop1)); | |
21979 | ||
21980 | return true; | |
21981 | } | |
21982 | ||
21983 | /* A subroutine of expand_vec_perm_even_odd_1. Implement extract-even | |
21984 | and extract-odd permutations of two V64QI operands | |
21985 | with two "shifts", two "truncs" and one "concat" insns for "odd" | |
21986 | and two "truncs" and one concat insn for "even." | |
21987 | Have already failed all two instruction sequences. */ | |
21988 | ||
21989 | static bool | |
21990 | expand_vec_perm_even_odd_trunc (struct expand_vec_perm_d *d) | |
21991 | { | |
21992 | rtx t1, t2, t3, t4; | |
21993 | unsigned i, odd, nelt = d->nelt; | |
21994 | ||
21995 | if (!TARGET_AVX512BW | |
21996 | || d->one_operand_p | |
21997 | || d->vmode != V64QImode) | |
21998 | return false; | |
21999 | ||
22000 | /* Check that permutation is even or odd. */ | |
22001 | odd = d->perm[0]; | |
22002 | if (odd > 1) | |
22003 | return false; | |
22004 | ||
22005 | for (i = 1; i < nelt; ++i) | |
22006 | if (d->perm[i] != 2 * i + odd) | |
22007 | return false; | |
22008 | ||
22009 | if (d->testing_p) | |
22010 | return true; | |
22011 | ||
22012 | ||
22013 | if (odd) | |
22014 | { | |
22015 | t1 = gen_reg_rtx (V32HImode); | |
22016 | t2 = gen_reg_rtx (V32HImode); | |
22017 | emit_insn (gen_lshrv32hi3 (t1, | |
22018 | gen_lowpart (V32HImode, d->op0), | |
22019 | GEN_INT (8))); | |
22020 | emit_insn (gen_lshrv32hi3 (t2, | |
22021 | gen_lowpart (V32HImode, d->op1), | |
22022 | GEN_INT (8))); | |
22023 | } | |
22024 | else | |
22025 | { | |
22026 | t1 = gen_lowpart (V32HImode, d->op0); | |
22027 | t2 = gen_lowpart (V32HImode, d->op1); | |
22028 | } | |
22029 | ||
22030 | t3 = gen_reg_rtx (V32QImode); | |
22031 | t4 = gen_reg_rtx (V32QImode); | |
22032 | emit_insn (gen_avx512bw_truncatev32hiv32qi2 (t3, t1)); | |
22033 | emit_insn (gen_avx512bw_truncatev32hiv32qi2 (t4, t2)); | |
22034 | emit_insn (gen_avx_vec_concatv64qi (d->target, t3, t4)); | |
22035 | ||
22036 | return true; | |
22037 | } | |
22038 | ||
4bf4c103 | 22039 | /* A subroutine of ix86_expand_vec_perm_const_1. Implement extract-even |
2bf6d935 ML |
22040 | and extract-odd permutations. */ |
22041 | ||
22042 | static bool | |
22043 | expand_vec_perm_even_odd_1 (struct expand_vec_perm_d *d, unsigned odd) | |
22044 | { | |
22045 | rtx t1, t2, t3, t4, t5; | |
22046 | ||
22047 | switch (d->vmode) | |
22048 | { | |
22049 | case E_V4DFmode: | |
22050 | if (d->testing_p) | |
22051 | break; | |
22052 | t1 = gen_reg_rtx (V4DFmode); | |
22053 | t2 = gen_reg_rtx (V4DFmode); | |
22054 | ||
22055 | /* Shuffle the lanes around into { 0 1 4 5 } and { 2 3 6 7 }. */ | |
22056 | emit_insn (gen_avx_vperm2f128v4df3 (t1, d->op0, d->op1, GEN_INT (0x20))); | |
22057 | emit_insn (gen_avx_vperm2f128v4df3 (t2, d->op0, d->op1, GEN_INT (0x31))); | |
22058 | ||
22059 | /* Now an unpck[lh]pd will produce the result required. */ | |
22060 | if (odd) | |
22061 | t3 = gen_avx_unpckhpd256 (d->target, t1, t2); | |
22062 | else | |
22063 | t3 = gen_avx_unpcklpd256 (d->target, t1, t2); | |
22064 | emit_insn (t3); | |
22065 | break; | |
22066 | ||
22067 | case E_V8SFmode: | |
22068 | { | |
22069 | int mask = odd ? 0xdd : 0x88; | |
22070 | ||
22071 | if (d->testing_p) | |
22072 | break; | |
22073 | t1 = gen_reg_rtx (V8SFmode); | |
22074 | t2 = gen_reg_rtx (V8SFmode); | |
22075 | t3 = gen_reg_rtx (V8SFmode); | |
22076 | ||
22077 | /* Shuffle within the 128-bit lanes to produce: | |
22078 | { 0 2 8 a 4 6 c e } | { 1 3 9 b 5 7 d f }. */ | |
22079 | emit_insn (gen_avx_shufps256 (t1, d->op0, d->op1, | |
22080 | GEN_INT (mask))); | |
22081 | ||
22082 | /* Shuffle the lanes around to produce: | |
22083 | { 4 6 c e 0 2 8 a } and { 5 7 d f 1 3 9 b }. */ | |
22084 | emit_insn (gen_avx_vperm2f128v8sf3 (t2, t1, t1, | |
22085 | GEN_INT (0x3))); | |
22086 | ||
22087 | /* Shuffle within the 128-bit lanes to produce: | |
22088 | { 0 2 4 6 4 6 0 2 } | { 1 3 5 7 5 7 1 3 }. */ | |
22089 | emit_insn (gen_avx_shufps256 (t3, t1, t2, GEN_INT (0x44))); | |
22090 | ||
22091 | /* Shuffle within the 128-bit lanes to produce: | |
22092 | { 8 a c e c e 8 a } | { 9 b d f d f 9 b }. */ | |
22093 | emit_insn (gen_avx_shufps256 (t2, t1, t2, GEN_INT (0xee))); | |
22094 | ||
22095 | /* Shuffle the lanes around to produce: | |
22096 | { 0 2 4 6 8 a c e } | { 1 3 5 7 9 b d f }. */ | |
22097 | emit_insn (gen_avx_vperm2f128v8sf3 (d->target, t3, t2, | |
22098 | GEN_INT (0x20))); | |
22099 | } | |
22100 | break; | |
22101 | ||
22102 | case E_V2DFmode: | |
22103 | case E_V4SFmode: | |
22104 | case E_V2DImode: | |
9b8579a6 | 22105 | case E_V2SImode: |
2bf6d935 | 22106 | case E_V4SImode: |
8d7dae0e | 22107 | case E_V2HImode: |
2bf6d935 ML |
22108 | /* These are always directly implementable by expand_vec_perm_1. */ |
22109 | gcc_unreachable (); | |
22110 | ||
240198fe UB |
22111 | case E_V2SFmode: |
22112 | gcc_assert (TARGET_MMX_WITH_SSE); | |
22113 | /* We have no suitable instructions. */ | |
22114 | if (d->testing_p) | |
22115 | return false; | |
22116 | break; | |
22117 | ||
be8749f9 UB |
22118 | case E_V4QImode: |
22119 | if (TARGET_SSSE3 && !TARGET_SLOW_PSHUFB) | |
22120 | return expand_vec_perm_pshufb2 (d); | |
22121 | else | |
22122 | { | |
22123 | if (d->testing_p) | |
22124 | break; | |
22125 | /* We need 2*log2(N)-1 operations to achieve odd/even | |
22126 | with interleave. */ | |
22127 | t1 = gen_reg_rtx (V4QImode); | |
22128 | emit_insn (gen_mmx_punpckhbw_low (t1, d->op0, d->op1)); | |
22129 | emit_insn (gen_mmx_punpcklbw_low (d->target, d->op0, d->op1)); | |
22130 | if (odd) | |
22131 | t2 = gen_mmx_punpckhbw_low (d->target, d->target, t1); | |
22132 | else | |
22133 | t2 = gen_mmx_punpcklbw_low (d->target, d->target, t1); | |
22134 | emit_insn (t2); | |
22135 | } | |
22136 | break; | |
22137 | ||
9b8579a6 | 22138 | case E_V4HImode: |
dd835ec2 UB |
22139 | if (TARGET_SSE4_1) |
22140 | return expand_vec_perm_even_odd_pack (d); | |
22141 | else if (TARGET_SSSE3 && !TARGET_SLOW_PSHUFB) | |
22142 | return expand_vec_perm_pshufb2 (d); | |
9b8579a6 | 22143 | else |
dd835ec2 UB |
22144 | { |
22145 | if (d->testing_p) | |
22146 | break; | |
22147 | /* We need 2*log2(N)-1 operations to achieve odd/even | |
22148 | with interleave. */ | |
22149 | t1 = gen_reg_rtx (V4HImode); | |
22150 | emit_insn (gen_mmx_punpckhwd (t1, d->op0, d->op1)); | |
22151 | emit_insn (gen_mmx_punpcklwd (d->target, d->op0, d->op1)); | |
22152 | if (odd) | |
22153 | t2 = gen_mmx_punpckhwd (d->target, d->target, t1); | |
22154 | else | |
22155 | t2 = gen_mmx_punpcklwd (d->target, d->target, t1); | |
22156 | emit_insn (t2); | |
22157 | } | |
9b8579a6 UB |
22158 | break; |
22159 | ||
2bf6d935 ML |
22160 | case E_V8HImode: |
22161 | if (TARGET_SSE4_1) | |
22162 | return expand_vec_perm_even_odd_pack (d); | |
22163 | else if (TARGET_SSSE3 && !TARGET_SLOW_PSHUFB) | |
22164 | return expand_vec_perm_pshufb2 (d); | |
22165 | else | |
22166 | { | |
22167 | if (d->testing_p) | |
22168 | break; | |
22169 | /* We need 2*log2(N)-1 operations to achieve odd/even | |
22170 | with interleave. */ | |
22171 | t1 = gen_reg_rtx (V8HImode); | |
22172 | t2 = gen_reg_rtx (V8HImode); | |
22173 | emit_insn (gen_vec_interleave_highv8hi (t1, d->op0, d->op1)); | |
22174 | emit_insn (gen_vec_interleave_lowv8hi (d->target, d->op0, d->op1)); | |
22175 | emit_insn (gen_vec_interleave_highv8hi (t2, d->target, t1)); | |
22176 | emit_insn (gen_vec_interleave_lowv8hi (d->target, d->target, t1)); | |
22177 | if (odd) | |
22178 | t3 = gen_vec_interleave_highv8hi (d->target, d->target, t2); | |
22179 | else | |
22180 | t3 = gen_vec_interleave_lowv8hi (d->target, d->target, t2); | |
22181 | emit_insn (t3); | |
22182 | } | |
22183 | break; | |
22184 | ||
a325bdd1 | 22185 | case E_V8QImode: |
2bf6d935 ML |
22186 | case E_V16QImode: |
22187 | return expand_vec_perm_even_odd_pack (d); | |
22188 | ||
22189 | case E_V16HImode: | |
22190 | case E_V32QImode: | |
22191 | return expand_vec_perm_even_odd_pack (d); | |
22192 | ||
22193 | case E_V64QImode: | |
22194 | return expand_vec_perm_even_odd_trunc (d); | |
22195 | ||
22196 | case E_V4DImode: | |
22197 | if (!TARGET_AVX2) | |
22198 | { | |
22199 | struct expand_vec_perm_d d_copy = *d; | |
22200 | d_copy.vmode = V4DFmode; | |
22201 | if (d->testing_p) | |
22202 | d_copy.target = gen_raw_REG (V4DFmode, LAST_VIRTUAL_REGISTER + 1); | |
22203 | else | |
22204 | d_copy.target = gen_reg_rtx (V4DFmode); | |
22205 | d_copy.op0 = gen_lowpart (V4DFmode, d->op0); | |
22206 | d_copy.op1 = gen_lowpart (V4DFmode, d->op1); | |
22207 | if (expand_vec_perm_even_odd_1 (&d_copy, odd)) | |
22208 | { | |
22209 | if (!d->testing_p) | |
22210 | emit_move_insn (d->target, | |
22211 | gen_lowpart (V4DImode, d_copy.target)); | |
22212 | return true; | |
22213 | } | |
22214 | return false; | |
22215 | } | |
22216 | ||
22217 | if (d->testing_p) | |
22218 | break; | |
22219 | ||
22220 | t1 = gen_reg_rtx (V4DImode); | |
22221 | t2 = gen_reg_rtx (V4DImode); | |
22222 | ||
22223 | /* Shuffle the lanes around into { 0 1 4 5 } and { 2 3 6 7 }. */ | |
22224 | emit_insn (gen_avx2_permv2ti (t1, d->op0, d->op1, GEN_INT (0x20))); | |
22225 | emit_insn (gen_avx2_permv2ti (t2, d->op0, d->op1, GEN_INT (0x31))); | |
22226 | ||
22227 | /* Now an vpunpck[lh]qdq will produce the result required. */ | |
22228 | if (odd) | |
22229 | t3 = gen_avx2_interleave_highv4di (d->target, t1, t2); | |
22230 | else | |
22231 | t3 = gen_avx2_interleave_lowv4di (d->target, t1, t2); | |
22232 | emit_insn (t3); | |
22233 | break; | |
22234 | ||
22235 | case E_V8SImode: | |
22236 | if (!TARGET_AVX2) | |
22237 | { | |
22238 | struct expand_vec_perm_d d_copy = *d; | |
22239 | d_copy.vmode = V8SFmode; | |
22240 | if (d->testing_p) | |
22241 | d_copy.target = gen_raw_REG (V8SFmode, LAST_VIRTUAL_REGISTER + 1); | |
22242 | else | |
22243 | d_copy.target = gen_reg_rtx (V8SFmode); | |
22244 | d_copy.op0 = gen_lowpart (V8SFmode, d->op0); | |
22245 | d_copy.op1 = gen_lowpart (V8SFmode, d->op1); | |
22246 | if (expand_vec_perm_even_odd_1 (&d_copy, odd)) | |
22247 | { | |
22248 | if (!d->testing_p) | |
22249 | emit_move_insn (d->target, | |
22250 | gen_lowpart (V8SImode, d_copy.target)); | |
22251 | return true; | |
22252 | } | |
22253 | return false; | |
22254 | } | |
22255 | ||
22256 | if (d->testing_p) | |
22257 | break; | |
22258 | ||
22259 | t1 = gen_reg_rtx (V8SImode); | |
22260 | t2 = gen_reg_rtx (V8SImode); | |
22261 | t3 = gen_reg_rtx (V4DImode); | |
22262 | t4 = gen_reg_rtx (V4DImode); | |
22263 | t5 = gen_reg_rtx (V4DImode); | |
22264 | ||
22265 | /* Shuffle the lanes around into | |
22266 | { 0 1 2 3 8 9 a b } and { 4 5 6 7 c d e f }. */ | |
22267 | emit_insn (gen_avx2_permv2ti (t3, gen_lowpart (V4DImode, d->op0), | |
22268 | gen_lowpart (V4DImode, d->op1), | |
22269 | GEN_INT (0x20))); | |
22270 | emit_insn (gen_avx2_permv2ti (t4, gen_lowpart (V4DImode, d->op0), | |
22271 | gen_lowpart (V4DImode, d->op1), | |
22272 | GEN_INT (0x31))); | |
22273 | ||
22274 | /* Swap the 2nd and 3rd position in each lane into | |
22275 | { 0 2 1 3 8 a 9 b } and { 4 6 5 7 c e d f }. */ | |
22276 | emit_insn (gen_avx2_pshufdv3 (t1, gen_lowpart (V8SImode, t3), | |
22277 | GEN_INT (2 * 4 + 1 * 16 + 3 * 64))); | |
22278 | emit_insn (gen_avx2_pshufdv3 (t2, gen_lowpart (V8SImode, t4), | |
22279 | GEN_INT (2 * 4 + 1 * 16 + 3 * 64))); | |
22280 | ||
22281 | /* Now an vpunpck[lh]qdq will produce | |
22282 | { 0 2 4 6 8 a c e } resp. { 1 3 5 7 9 b d f }. */ | |
22283 | if (odd) | |
22284 | t3 = gen_avx2_interleave_highv4di (t5, gen_lowpart (V4DImode, t1), | |
22285 | gen_lowpart (V4DImode, t2)); | |
22286 | else | |
22287 | t3 = gen_avx2_interleave_lowv4di (t5, gen_lowpart (V4DImode, t1), | |
22288 | gen_lowpart (V4DImode, t2)); | |
22289 | emit_insn (t3); | |
22290 | emit_move_insn (d->target, gen_lowpart (V8SImode, t5)); | |
22291 | break; | |
22292 | ||
22293 | default: | |
22294 | gcc_unreachable (); | |
22295 | } | |
22296 | ||
22297 | return true; | |
22298 | } | |
22299 | ||
4bf4c103 | 22300 | /* A subroutine of ix86_expand_vec_perm_const_1. Pattern match |
2bf6d935 ML |
22301 | extract-even and extract-odd permutations. */ |
22302 | ||
22303 | static bool | |
22304 | expand_vec_perm_even_odd (struct expand_vec_perm_d *d) | |
22305 | { | |
22306 | unsigned i, odd, nelt = d->nelt; | |
22307 | ||
22308 | odd = d->perm[0]; | |
22309 | if (odd != 0 && odd != 1) | |
22310 | return false; | |
22311 | ||
22312 | for (i = 1; i < nelt; ++i) | |
22313 | if (d->perm[i] != 2 * i + odd) | |
22314 | return false; | |
22315 | ||
50b58779 JJ |
22316 | if (d->vmode == E_V32HImode |
22317 | && d->testing_p | |
22318 | && !TARGET_AVX512BW) | |
22319 | return false; | |
22320 | ||
2bf6d935 ML |
22321 | return expand_vec_perm_even_odd_1 (d, odd); |
22322 | } | |
22323 | ||
4bf4c103 | 22324 | /* A subroutine of ix86_expand_vec_perm_const_1. Implement broadcast |
2bf6d935 ML |
22325 | permutations. We assume that expand_vec_perm_1 has already failed. */ |
22326 | ||
22327 | static bool | |
22328 | expand_vec_perm_broadcast_1 (struct expand_vec_perm_d *d) | |
22329 | { | |
22330 | unsigned elt = d->perm[0], nelt2 = d->nelt / 2; | |
22331 | machine_mode vmode = d->vmode; | |
be8749f9 | 22332 | rtx (*gen) (rtx, rtx, rtx); |
2bf6d935 ML |
22333 | unsigned char perm2[4]; |
22334 | rtx op0 = d->op0, dest; | |
22335 | bool ok; | |
22336 | ||
22337 | switch (vmode) | |
22338 | { | |
22339 | case E_V4DFmode: | |
22340 | case E_V8SFmode: | |
22341 | /* These are special-cased in sse.md so that we can optionally | |
22342 | use the vbroadcast instruction. They expand to two insns | |
22343 | if the input happens to be in a register. */ | |
22344 | gcc_unreachable (); | |
22345 | ||
22346 | case E_V2DFmode: | |
240198fe | 22347 | case E_V2SFmode: |
2bf6d935 | 22348 | case E_V4SFmode: |
240198fe | 22349 | case E_V2DImode: |
9b8579a6 | 22350 | case E_V2SImode: |
2bf6d935 | 22351 | case E_V4SImode: |
8d7dae0e UB |
22352 | case E_V2HImode: |
22353 | case E_V4HImode: | |
2bf6d935 ML |
22354 | /* These are always implementable using standard shuffle patterns. */ |
22355 | gcc_unreachable (); | |
22356 | ||
be8749f9 UB |
22357 | case E_V4QImode: |
22358 | /* This can be implemented via interleave and pshuflw. */ | |
22359 | if (d->testing_p) | |
22360 | return true; | |
22361 | ||
22362 | if (elt >= nelt2) | |
22363 | { | |
22364 | gen = gen_mmx_punpckhbw_low; | |
22365 | elt -= nelt2; | |
22366 | } | |
22367 | else | |
22368 | gen = gen_mmx_punpcklbw_low; | |
22369 | ||
22370 | dest = gen_reg_rtx (vmode); | |
22371 | emit_insn (gen (dest, op0, op0)); | |
22372 | vmode = get_mode_wider_vector (vmode); | |
22373 | op0 = gen_lowpart (vmode, dest); | |
22374 | ||
22375 | memset (perm2, elt, 2); | |
22376 | dest = gen_reg_rtx (vmode); | |
22377 | ok = expand_vselect (dest, op0, perm2, 2, d->testing_p); | |
22378 | gcc_assert (ok); | |
22379 | ||
22380 | emit_move_insn (d->target, gen_lowpart (d->vmode, dest)); | |
22381 | return true; | |
22382 | ||
a325bdd1 | 22383 | case E_V8QImode: |
be8749f9 | 22384 | /* This can be implemented via interleave. We save one insn by |
a325bdd1 PB |
22385 | stopping once we have promoted to V2SImode and then use pshufd. */ |
22386 | if (d->testing_p) | |
22387 | return true; | |
22388 | do | |
22389 | { | |
a325bdd1 PB |
22390 | if (elt >= nelt2) |
22391 | { | |
22392 | gen = vmode == V8QImode ? gen_mmx_punpckhbw | |
22393 | : gen_mmx_punpckhwd; | |
22394 | elt -= nelt2; | |
22395 | } | |
be8749f9 UB |
22396 | else |
22397 | gen = vmode == V8QImode ? gen_mmx_punpcklbw | |
22398 | : gen_mmx_punpcklwd; | |
a325bdd1 PB |
22399 | nelt2 /= 2; |
22400 | ||
22401 | dest = gen_reg_rtx (vmode); | |
22402 | emit_insn (gen (dest, op0, op0)); | |
22403 | vmode = get_mode_wider_vector (vmode); | |
22404 | op0 = gen_lowpart (vmode, dest); | |
22405 | } | |
22406 | while (vmode != V2SImode); | |
22407 | ||
22408 | memset (perm2, elt, 2); | |
be8749f9 | 22409 | dest = gen_reg_rtx (vmode); |
a325bdd1 PB |
22410 | ok = expand_vselect (dest, op0, perm2, 2, d->testing_p); |
22411 | gcc_assert (ok); | |
be8749f9 UB |
22412 | |
22413 | emit_move_insn (d->target, gen_lowpart (d->vmode, dest)); | |
a325bdd1 PB |
22414 | return true; |
22415 | ||
2bf6d935 ML |
22416 | case E_V8HImode: |
22417 | case E_V16QImode: | |
22418 | /* These can be implemented via interleave. We save one insn by | |
22419 | stopping once we have promoted to V4SImode and then use pshufd. */ | |
22420 | if (d->testing_p) | |
22421 | return true; | |
22422 | do | |
22423 | { | |
2bf6d935 ML |
22424 | if (elt >= nelt2) |
22425 | { | |
22426 | gen = vmode == V16QImode ? gen_vec_interleave_highv16qi | |
22427 | : gen_vec_interleave_highv8hi; | |
22428 | elt -= nelt2; | |
22429 | } | |
be8749f9 UB |
22430 | else |
22431 | gen = vmode == V16QImode ? gen_vec_interleave_lowv16qi | |
22432 | : gen_vec_interleave_lowv8hi; | |
2bf6d935 ML |
22433 | nelt2 /= 2; |
22434 | ||
22435 | dest = gen_reg_rtx (vmode); | |
22436 | emit_insn (gen (dest, op0, op0)); | |
22437 | vmode = get_mode_wider_vector (vmode); | |
22438 | op0 = gen_lowpart (vmode, dest); | |
22439 | } | |
22440 | while (vmode != V4SImode); | |
22441 | ||
22442 | memset (perm2, elt, 4); | |
be8749f9 | 22443 | dest = gen_reg_rtx (vmode); |
2bf6d935 ML |
22444 | ok = expand_vselect (dest, op0, perm2, 4, d->testing_p); |
22445 | gcc_assert (ok); | |
be8749f9 UB |
22446 | |
22447 | emit_move_insn (d->target, gen_lowpart (d->vmode, dest)); | |
2bf6d935 ML |
22448 | return true; |
22449 | ||
7a54d3de | 22450 | case E_V8HFmode: |
092763fd | 22451 | case E_V8BFmode: |
7a54d3de UB |
22452 | /* This can be implemented via interleave and pshufd. */ |
22453 | if (d->testing_p) | |
22454 | return true; | |
22455 | ||
092763fd | 22456 | rtx (*maybe_gen) (machine_mode, int, rtx, rtx, rtx); |
7a54d3de UB |
22457 | if (elt >= nelt2) |
22458 | { | |
092763fd | 22459 | maybe_gen = maybe_gen_vec_interleave_high; |
7a54d3de UB |
22460 | elt -= nelt2; |
22461 | } | |
22462 | else | |
092763fd | 22463 | maybe_gen = maybe_gen_vec_interleave_low; |
7a54d3de UB |
22464 | nelt2 /= 2; |
22465 | ||
22466 | dest = gen_reg_rtx (vmode); | |
092763fd | 22467 | emit_insn (maybe_gen (vmode, 1, dest, op0, op0)); |
7a54d3de UB |
22468 | |
22469 | vmode = V4SImode; | |
22470 | op0 = gen_lowpart (vmode, dest); | |
22471 | ||
22472 | memset (perm2, elt, 4); | |
22473 | dest = gen_reg_rtx (vmode); | |
22474 | ok = expand_vselect (dest, op0, perm2, 4, d->testing_p); | |
22475 | gcc_assert (ok); | |
22476 | ||
22477 | emit_move_insn (d->target, gen_lowpart (d->vmode, dest)); | |
22478 | return true; | |
22479 | ||
2bf6d935 ML |
22480 | case E_V32QImode: |
22481 | case E_V16HImode: | |
22482 | case E_V8SImode: | |
22483 | case E_V4DImode: | |
22484 | /* For AVX2 broadcasts of the first element vpbroadcast* or | |
22485 | vpermq should be used by expand_vec_perm_1. */ | |
22486 | gcc_assert (!TARGET_AVX2 || d->perm[0]); | |
22487 | return false; | |
22488 | ||
240f0780 JJ |
22489 | case E_V64QImode: |
22490 | gcc_assert (!TARGET_AVX512BW || d->perm[0]); | |
22491 | return false; | |
22492 | ||
04b4f315 JJ |
22493 | case E_V32HImode: |
22494 | gcc_assert (!TARGET_AVX512BW); | |
22495 | return false; | |
22496 | ||
2bf6d935 ML |
22497 | default: |
22498 | gcc_unreachable (); | |
22499 | } | |
22500 | } | |
22501 | ||
4bf4c103 | 22502 | /* A subroutine of ix86_expand_vec_perm_const_1. Pattern match |
2bf6d935 ML |
22503 | broadcast permutations. */ |
22504 | ||
22505 | static bool | |
22506 | expand_vec_perm_broadcast (struct expand_vec_perm_d *d) | |
22507 | { | |
22508 | unsigned i, elt, nelt = d->nelt; | |
22509 | ||
22510 | if (!d->one_operand_p) | |
22511 | return false; | |
22512 | ||
22513 | elt = d->perm[0]; | |
22514 | for (i = 1; i < nelt; ++i) | |
22515 | if (d->perm[i] != elt) | |
22516 | return false; | |
22517 | ||
22518 | return expand_vec_perm_broadcast_1 (d); | |
22519 | } | |
22520 | ||
22521 | /* Implement arbitrary permutations of two V64QImode operands | |
22522 | with 2 vperm[it]2w, 2 vpshufb and one vpor instruction. */ | |
22523 | static bool | |
22524 | expand_vec_perm_vpermt2_vpshub2 (struct expand_vec_perm_d *d) | |
22525 | { | |
22526 | if (!TARGET_AVX512BW || !(d->vmode == V64QImode)) | |
22527 | return false; | |
22528 | ||
22529 | if (d->testing_p) | |
22530 | return true; | |
22531 | ||
22532 | struct expand_vec_perm_d ds[2]; | |
22533 | rtx rperm[128], vperm, target0, target1; | |
22534 | unsigned int i, nelt; | |
22535 | machine_mode vmode; | |
22536 | ||
22537 | nelt = d->nelt; | |
22538 | vmode = V64QImode; | |
22539 | ||
22540 | for (i = 0; i < 2; i++) | |
22541 | { | |
22542 | ds[i] = *d; | |
22543 | ds[i].vmode = V32HImode; | |
22544 | ds[i].nelt = 32; | |
22545 | ds[i].target = gen_reg_rtx (V32HImode); | |
22546 | ds[i].op0 = gen_lowpart (V32HImode, d->op0); | |
22547 | ds[i].op1 = gen_lowpart (V32HImode, d->op1); | |
22548 | } | |
22549 | ||
22550 | /* Prepare permutations such that the first one takes care of | |
22551 | putting the even bytes into the right positions or one higher | |
22552 | positions (ds[0]) and the second one takes care of | |
22553 | putting the odd bytes into the right positions or one below | |
22554 | (ds[1]). */ | |
22555 | ||
22556 | for (i = 0; i < nelt; i++) | |
22557 | { | |
22558 | ds[i & 1].perm[i / 2] = d->perm[i] / 2; | |
22559 | if (i & 1) | |
22560 | { | |
22561 | rperm[i] = constm1_rtx; | |
22562 | rperm[i + 64] = GEN_INT ((i & 14) + (d->perm[i] & 1)); | |
22563 | } | |
22564 | else | |
22565 | { | |
22566 | rperm[i] = GEN_INT ((i & 14) + (d->perm[i] & 1)); | |
22567 | rperm[i + 64] = constm1_rtx; | |
22568 | } | |
22569 | } | |
22570 | ||
22571 | bool ok = expand_vec_perm_1 (&ds[0]); | |
22572 | gcc_assert (ok); | |
22573 | ds[0].target = gen_lowpart (V64QImode, ds[0].target); | |
22574 | ||
22575 | ok = expand_vec_perm_1 (&ds[1]); | |
22576 | gcc_assert (ok); | |
22577 | ds[1].target = gen_lowpart (V64QImode, ds[1].target); | |
22578 | ||
22579 | vperm = gen_rtx_CONST_VECTOR (V64QImode, gen_rtvec_v (64, rperm)); | |
22580 | vperm = force_reg (vmode, vperm); | |
22581 | target0 = gen_reg_rtx (V64QImode); | |
22582 | emit_insn (gen_avx512bw_pshufbv64qi3 (target0, ds[0].target, vperm)); | |
22583 | ||
22584 | vperm = gen_rtx_CONST_VECTOR (V64QImode, gen_rtvec_v (64, rperm + 64)); | |
22585 | vperm = force_reg (vmode, vperm); | |
22586 | target1 = gen_reg_rtx (V64QImode); | |
22587 | emit_insn (gen_avx512bw_pshufbv64qi3 (target1, ds[1].target, vperm)); | |
22588 | ||
22589 | emit_insn (gen_iorv64qi3 (d->target, target0, target1)); | |
22590 | return true; | |
22591 | } | |
22592 | ||
22593 | /* Implement arbitrary permutation of two V32QImode and V16QImode operands | |
22594 | with 4 vpshufb insns, 2 vpermq and 3 vpor. We should have already failed | |
22595 | all the shorter instruction sequences. */ | |
22596 | ||
22597 | static bool | |
22598 | expand_vec_perm_vpshufb4_vpermq2 (struct expand_vec_perm_d *d) | |
22599 | { | |
22600 | rtx rperm[4][32], vperm, l[2], h[2], op, m128; | |
22601 | unsigned int i, nelt, eltsz; | |
22602 | bool used[4]; | |
22603 | ||
22604 | if (!TARGET_AVX2 | |
22605 | || d->one_operand_p | |
22606 | || (d->vmode != V32QImode && d->vmode != V16HImode)) | |
22607 | return false; | |
22608 | ||
22609 | if (d->testing_p) | |
22610 | return true; | |
22611 | ||
22612 | nelt = d->nelt; | |
22613 | eltsz = GET_MODE_UNIT_SIZE (d->vmode); | |
22614 | ||
22615 | /* Generate 4 permutation masks. If the required element is within | |
22616 | the same lane, it is shuffled in. If the required element from the | |
22617 | other lane, force a zero by setting bit 7 in the permutation mask. | |
22618 | In the other mask the mask has non-negative elements if element | |
22619 | is requested from the other lane, but also moved to the other lane, | |
22620 | so that the result of vpshufb can have the two V2TImode halves | |
22621 | swapped. */ | |
22622 | m128 = GEN_INT (-128); | |
22623 | for (i = 0; i < 32; ++i) | |
22624 | { | |
22625 | rperm[0][i] = m128; | |
22626 | rperm[1][i] = m128; | |
22627 | rperm[2][i] = m128; | |
22628 | rperm[3][i] = m128; | |
22629 | } | |
22630 | used[0] = false; | |
22631 | used[1] = false; | |
22632 | used[2] = false; | |
22633 | used[3] = false; | |
22634 | for (i = 0; i < nelt; ++i) | |
22635 | { | |
22636 | unsigned j, e = d->perm[i] & (nelt / 2 - 1); | |
22637 | unsigned xlane = ((d->perm[i] ^ i) & (nelt / 2)) * eltsz; | |
22638 | unsigned int which = ((d->perm[i] & nelt) ? 2 : 0) + (xlane ? 1 : 0); | |
22639 | ||
22640 | for (j = 0; j < eltsz; ++j) | |
22641 | rperm[which][(i * eltsz + j) ^ xlane] = GEN_INT (e * eltsz + j); | |
22642 | used[which] = true; | |
22643 | } | |
22644 | ||
22645 | for (i = 0; i < 2; ++i) | |
22646 | { | |
22647 | if (!used[2 * i + 1]) | |
22648 | { | |
22649 | h[i] = NULL_RTX; | |
22650 | continue; | |
22651 | } | |
22652 | vperm = gen_rtx_CONST_VECTOR (V32QImode, | |
22653 | gen_rtvec_v (32, rperm[2 * i + 1])); | |
22654 | vperm = force_reg (V32QImode, vperm); | |
22655 | h[i] = gen_reg_rtx (V32QImode); | |
22656 | op = gen_lowpart (V32QImode, i ? d->op1 : d->op0); | |
22657 | emit_insn (gen_avx2_pshufbv32qi3 (h[i], op, vperm)); | |
22658 | } | |
22659 | ||
22660 | /* Swap the 128-byte lanes of h[X]. */ | |
22661 | for (i = 0; i < 2; ++i) | |
22662 | { | |
22663 | if (h[i] == NULL_RTX) | |
22664 | continue; | |
22665 | op = gen_reg_rtx (V4DImode); | |
22666 | emit_insn (gen_avx2_permv4di_1 (op, gen_lowpart (V4DImode, h[i]), | |
22667 | const2_rtx, GEN_INT (3), const0_rtx, | |
22668 | const1_rtx)); | |
22669 | h[i] = gen_lowpart (V32QImode, op); | |
22670 | } | |
22671 | ||
22672 | for (i = 0; i < 2; ++i) | |
22673 | { | |
22674 | if (!used[2 * i]) | |
22675 | { | |
22676 | l[i] = NULL_RTX; | |
22677 | continue; | |
22678 | } | |
22679 | vperm = gen_rtx_CONST_VECTOR (V32QImode, gen_rtvec_v (32, rperm[2 * i])); | |
22680 | vperm = force_reg (V32QImode, vperm); | |
22681 | l[i] = gen_reg_rtx (V32QImode); | |
22682 | op = gen_lowpart (V32QImode, i ? d->op1 : d->op0); | |
22683 | emit_insn (gen_avx2_pshufbv32qi3 (l[i], op, vperm)); | |
22684 | } | |
22685 | ||
22686 | for (i = 0; i < 2; ++i) | |
22687 | { | |
22688 | if (h[i] && l[i]) | |
22689 | { | |
22690 | op = gen_reg_rtx (V32QImode); | |
22691 | emit_insn (gen_iorv32qi3 (op, l[i], h[i])); | |
22692 | l[i] = op; | |
22693 | } | |
22694 | else if (h[i]) | |
22695 | l[i] = h[i]; | |
22696 | } | |
22697 | ||
22698 | gcc_assert (l[0] && l[1]); | |
22699 | op = d->target; | |
22700 | if (d->vmode != V32QImode) | |
22701 | op = gen_reg_rtx (V32QImode); | |
22702 | emit_insn (gen_iorv32qi3 (op, l[0], l[1])); | |
22703 | if (op != d->target) | |
22704 | emit_move_insn (d->target, gen_lowpart (d->vmode, op)); | |
22705 | return true; | |
22706 | } | |
22707 | ||
22708 | /* The guts of ix86_vectorize_vec_perm_const. With all of the interface bits | |
22709 | taken care of, perform the expansion in D and return true on success. */ | |
22710 | ||
22711 | static bool | |
22712 | ix86_expand_vec_perm_const_1 (struct expand_vec_perm_d *d) | |
22713 | { | |
22714 | /* Try a single instruction expansion. */ | |
22715 | if (expand_vec_perm_1 (d)) | |
22716 | return true; | |
22717 | ||
22718 | /* Try sequences of two instructions. */ | |
22719 | ||
22720 | if (expand_vec_perm_pshuflw_pshufhw (d)) | |
22721 | return true; | |
22722 | ||
22723 | if (expand_vec_perm_palignr (d, false)) | |
22724 | return true; | |
22725 | ||
22726 | if (expand_vec_perm_interleave2 (d)) | |
22727 | return true; | |
22728 | ||
22729 | if (expand_vec_perm_broadcast (d)) | |
22730 | return true; | |
22731 | ||
22732 | if (expand_vec_perm_vpermq_perm_1 (d)) | |
22733 | return true; | |
22734 | ||
22735 | if (expand_vec_perm_vperm2f128 (d)) | |
22736 | return true; | |
22737 | ||
22738 | if (expand_vec_perm_pblendv (d)) | |
22739 | return true; | |
22740 | ||
829c4bea JJ |
22741 | if (expand_vec_perm_2perm_interleave (d, true)) |
22742 | return true; | |
22743 | ||
22744 | if (expand_vec_perm_2perm_pblendv (d, true)) | |
22745 | return true; | |
22746 | ||
3db8e9c2 | 22747 | if (expand_vec_perm_shufps_shufps (d)) |
22748 | return true; | |
22749 | ||
2bf6d935 ML |
22750 | /* Try sequences of three instructions. */ |
22751 | ||
22752 | if (expand_vec_perm_even_odd_pack (d)) | |
22753 | return true; | |
22754 | ||
22755 | if (expand_vec_perm_2vperm2f128_vshuf (d)) | |
22756 | return true; | |
22757 | ||
22758 | if (expand_vec_perm_pshufb2 (d)) | |
22759 | return true; | |
22760 | ||
fcda0efc | 22761 | if (expand_vec_perm_pslldq_psrldq_por (d, false)) |
22762 | return true; | |
22763 | ||
2bf6d935 ML |
22764 | if (expand_vec_perm_interleave3 (d)) |
22765 | return true; | |
22766 | ||
22767 | if (expand_vec_perm_vperm2f128_vblend (d)) | |
22768 | return true; | |
22769 | ||
829c4bea JJ |
22770 | if (expand_vec_perm_2perm_interleave (d, false)) |
22771 | return true; | |
22772 | ||
22773 | if (expand_vec_perm_2perm_pblendv (d, false)) | |
22774 | return true; | |
22775 | ||
2bf6d935 ML |
22776 | /* Try sequences of four instructions. */ |
22777 | ||
22778 | if (expand_vec_perm_even_odd_trunc (d)) | |
22779 | return true; | |
22780 | if (expand_vec_perm_vpshufb2_vpermq (d)) | |
22781 | return true; | |
22782 | ||
22783 | if (expand_vec_perm_vpshufb2_vpermq_even_odd (d)) | |
22784 | return true; | |
22785 | ||
22786 | if (expand_vec_perm_vpermt2_vpshub2 (d)) | |
22787 | return true; | |
22788 | ||
22789 | /* ??? Look for narrow permutations whose element orderings would | |
22790 | allow the promotion to a wider mode. */ | |
22791 | ||
22792 | /* ??? Look for sequences of interleave or a wider permute that place | |
22793 | the data into the correct lanes for a half-vector shuffle like | |
22794 | pshuf[lh]w or vpermilps. */ | |
22795 | ||
22796 | /* ??? Look for sequences of interleave that produce the desired results. | |
22797 | The combinatorics of punpck[lh] get pretty ugly... */ | |
22798 | ||
22799 | if (expand_vec_perm_even_odd (d)) | |
22800 | return true; | |
22801 | ||
fcda0efc | 22802 | /* Generate four or five instructions. */ |
22803 | if (expand_vec_perm_pslldq_psrldq_por (d, true)) | |
22804 | return true; | |
22805 | ||
2bf6d935 ML |
22806 | /* Even longer sequences. */ |
22807 | if (expand_vec_perm_vpshufb4_vpermq2 (d)) | |
22808 | return true; | |
22809 | ||
22810 | /* See if we can get the same permutation in different vector integer | |
22811 | mode. */ | |
22812 | struct expand_vec_perm_d nd; | |
22813 | if (canonicalize_vector_int_perm (d, &nd) && expand_vec_perm_1 (&nd)) | |
22814 | { | |
22815 | if (!d->testing_p) | |
22816 | emit_move_insn (d->target, gen_lowpart (d->vmode, nd.target)); | |
22817 | return true; | |
22818 | } | |
22819 | ||
4bf4c103 JJ |
22820 | /* Even longer, including recursion to ix86_expand_vec_perm_const_1. */ |
22821 | if (expand_vec_perm2_vperm2f128_vblend (d)) | |
22822 | return true; | |
22823 | ||
2bf6d935 ML |
22824 | return false; |
22825 | } | |
22826 | ||
22827 | /* If a permutation only uses one operand, make it clear. Returns true | |
22828 | if the permutation references both operands. */ | |
22829 | ||
22830 | static bool | |
22831 | canonicalize_perm (struct expand_vec_perm_d *d) | |
22832 | { | |
22833 | int i, which, nelt = d->nelt; | |
22834 | ||
22835 | for (i = which = 0; i < nelt; ++i) | |
4bf4c103 | 22836 | which |= (d->perm[i] < nelt ? 1 : 2); |
2bf6d935 ML |
22837 | |
22838 | d->one_operand_p = true; | |
22839 | switch (which) | |
22840 | { | |
22841 | default: | |
22842 | gcc_unreachable(); | |
22843 | ||
22844 | case 3: | |
22845 | if (!rtx_equal_p (d->op0, d->op1)) | |
22846 | { | |
22847 | d->one_operand_p = false; | |
22848 | break; | |
22849 | } | |
22850 | /* The elements of PERM do not suggest that only the first operand | |
22851 | is used, but both operands are identical. Allow easier matching | |
22852 | of the permutation by folding the permutation into the single | |
22853 | input vector. */ | |
22854 | /* FALLTHRU */ | |
22855 | ||
22856 | case 2: | |
22857 | for (i = 0; i < nelt; ++i) | |
22858 | d->perm[i] &= nelt - 1; | |
22859 | d->op0 = d->op1; | |
22860 | break; | |
22861 | ||
22862 | case 1: | |
22863 | d->op1 = d->op0; | |
22864 | break; | |
22865 | } | |
22866 | ||
22867 | return (which == 3); | |
22868 | } | |
22869 | ||
22870 | /* Implement TARGET_VECTORIZE_VEC_PERM_CONST. */ | |
22871 | ||
22872 | bool | |
ae8decf1 PK |
22873 | ix86_vectorize_vec_perm_const (machine_mode vmode, machine_mode op_mode, |
22874 | rtx target, rtx op0, rtx op1, | |
22875 | const vec_perm_indices &sel) | |
2bf6d935 | 22876 | { |
ae8decf1 PK |
22877 | if (vmode != op_mode) |
22878 | return false; | |
22879 | ||
2bf6d935 ML |
22880 | struct expand_vec_perm_d d; |
22881 | unsigned char perm[MAX_VECT_LEN]; | |
22882 | unsigned int i, nelt, which; | |
22883 | bool two_args; | |
22884 | ||
be072bfa HW |
22885 | /* For HF mode vector, convert it to HI using subreg. */ |
22886 | if (GET_MODE_INNER (vmode) == HFmode) | |
22887 | { | |
22888 | machine_mode orig_mode = vmode; | |
22889 | vmode = mode_for_vector (HImode, | |
22890 | GET_MODE_NUNITS (vmode)).require (); | |
22891 | if (target) | |
22892 | target = lowpart_subreg (vmode, target, orig_mode); | |
22893 | if (op0) | |
22894 | op0 = lowpart_subreg (vmode, op0, orig_mode); | |
22895 | if (op1) | |
22896 | op1 = lowpart_subreg (vmode, op1, orig_mode); | |
22897 | } | |
22898 | ||
2bf6d935 ML |
22899 | d.target = target; |
22900 | d.op0 = op0; | |
22901 | d.op1 = op1; | |
22902 | ||
22903 | d.vmode = vmode; | |
22904 | gcc_assert (VECTOR_MODE_P (d.vmode)); | |
22905 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
22906 | d.testing_p = !target; | |
22907 | ||
22908 | gcc_assert (sel.length () == nelt); | |
22909 | gcc_checking_assert (sizeof (d.perm) == sizeof (perm)); | |
22910 | ||
22911 | /* Given sufficient ISA support we can just return true here | |
22912 | for selected vector modes. */ | |
22913 | switch (d.vmode) | |
22914 | { | |
22915 | case E_V16SFmode: | |
22916 | case E_V16SImode: | |
22917 | case E_V8DImode: | |
22918 | case E_V8DFmode: | |
22919 | if (!TARGET_AVX512F) | |
22920 | return false; | |
22921 | /* All implementable with a single vperm[it]2 insn. */ | |
22922 | if (d.testing_p) | |
22923 | return true; | |
22924 | break; | |
22925 | case E_V32HImode: | |
50b58779 | 22926 | if (!TARGET_AVX512F) |
2bf6d935 | 22927 | return false; |
50b58779 | 22928 | if (d.testing_p && TARGET_AVX512BW) |
2bf6d935 ML |
22929 | /* All implementable with a single vperm[it]2 insn. */ |
22930 | return true; | |
22931 | break; | |
22932 | case E_V64QImode: | |
50b58779 | 22933 | if (!TARGET_AVX512F) |
2bf6d935 | 22934 | return false; |
50b58779 | 22935 | if (d.testing_p && TARGET_AVX512BW) |
2bf6d935 ML |
22936 | /* Implementable with 2 vperm[it]2, 2 vpshufb and 1 or insn. */ |
22937 | return true; | |
22938 | break; | |
22939 | case E_V8SImode: | |
22940 | case E_V8SFmode: | |
22941 | case E_V4DFmode: | |
22942 | case E_V4DImode: | |
22943 | if (!TARGET_AVX) | |
22944 | return false; | |
22945 | if (d.testing_p && TARGET_AVX512VL) | |
22946 | /* All implementable with a single vperm[it]2 insn. */ | |
22947 | return true; | |
22948 | break; | |
22949 | case E_V16HImode: | |
22950 | if (!TARGET_SSE2) | |
22951 | return false; | |
22952 | if (d.testing_p && TARGET_AVX2) | |
22953 | /* Implementable with 4 vpshufb insns, 2 vpermq and 3 vpor insns. */ | |
22954 | return true; | |
22955 | break; | |
22956 | case E_V32QImode: | |
22957 | if (!TARGET_SSE2) | |
22958 | return false; | |
22959 | if (d.testing_p && TARGET_AVX2) | |
22960 | /* Implementable with 4 vpshufb insns, 2 vpermq and 3 vpor insns. */ | |
22961 | return true; | |
22962 | break; | |
22963 | case E_V8HImode: | |
22964 | case E_V16QImode: | |
22965 | if (!TARGET_SSE2) | |
22966 | return false; | |
22967 | /* Fall through. */ | |
22968 | case E_V4SImode: | |
22969 | case E_V4SFmode: | |
22970 | if (!TARGET_SSE) | |
22971 | return false; | |
22972 | /* All implementable with a single vpperm insn. */ | |
22973 | if (d.testing_p && TARGET_XOP) | |
22974 | return true; | |
22975 | /* All implementable with 2 pshufb + 1 ior. */ | |
22976 | if (d.testing_p && TARGET_SSSE3) | |
22977 | return true; | |
22978 | break; | |
240198fe | 22979 | case E_V2SFmode: |
9b8579a6 UB |
22980 | case E_V2SImode: |
22981 | case E_V4HImode: | |
a325bdd1 | 22982 | case E_V8QImode: |
9b8579a6 UB |
22983 | if (!TARGET_MMX_WITH_SSE) |
22984 | return false; | |
22985 | break; | |
8d7dae0e | 22986 | case E_V2HImode: |
4986946f UB |
22987 | if (!TARGET_SSE2) |
22988 | return false; | |
22989 | /* All implementable with *punpckwd. */ | |
22990 | if (d.testing_p) | |
22991 | return true; | |
22992 | break; | |
be8749f9 UB |
22993 | case E_V4QImode: |
22994 | if (!TARGET_SSE2) | |
22995 | return false; | |
22996 | break; | |
2bf6d935 ML |
22997 | case E_V2DImode: |
22998 | case E_V2DFmode: | |
22999 | if (!TARGET_SSE) | |
23000 | return false; | |
23001 | /* All implementable with shufpd or unpck[lh]pd. */ | |
23002 | if (d.testing_p) | |
23003 | return true; | |
23004 | break; | |
23005 | default: | |
23006 | return false; | |
23007 | } | |
23008 | ||
23009 | for (i = which = 0; i < nelt; ++i) | |
23010 | { | |
23011 | unsigned char e = sel[i]; | |
23012 | gcc_assert (e < 2 * nelt); | |
23013 | d.perm[i] = e; | |
23014 | perm[i] = e; | |
23015 | which |= (e < nelt ? 1 : 2); | |
23016 | } | |
23017 | ||
23018 | if (d.testing_p) | |
23019 | { | |
23020 | /* For all elements from second vector, fold the elements to first. */ | |
23021 | if (which == 2) | |
23022 | for (i = 0; i < nelt; ++i) | |
23023 | d.perm[i] -= nelt; | |
23024 | ||
23025 | /* Check whether the mask can be applied to the vector type. */ | |
23026 | d.one_operand_p = (which != 3); | |
23027 | ||
8d7dae0e | 23028 | /* Implementable with shufps, pshufd or pshuflw. */ |
9b8579a6 | 23029 | if (d.one_operand_p |
240198fe | 23030 | && (d.vmode == V4SFmode || d.vmode == V2SFmode |
8d7dae0e UB |
23031 | || d.vmode == V4SImode || d.vmode == V2SImode |
23032 | || d.vmode == V4HImode || d.vmode == V2HImode)) | |
2bf6d935 ML |
23033 | return true; |
23034 | ||
23035 | /* Otherwise we have to go through the motions and see if we can | |
23036 | figure out how to generate the requested permutation. */ | |
23037 | d.target = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 1); | |
23038 | d.op1 = d.op0 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 2); | |
23039 | if (!d.one_operand_p) | |
23040 | d.op1 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 3); | |
23041 | ||
23042 | start_sequence (); | |
23043 | bool ret = ix86_expand_vec_perm_const_1 (&d); | |
23044 | end_sequence (); | |
23045 | ||
23046 | return ret; | |
23047 | } | |
23048 | ||
23049 | two_args = canonicalize_perm (&d); | |
23050 | ||
b1d1e2b5 JJ |
23051 | /* If one of the operands is a zero vector, try to match pmovzx. */ |
23052 | if (two_args && (d.op0 == CONST0_RTX (vmode) || d.op1 == CONST0_RTX (vmode))) | |
23053 | { | |
23054 | struct expand_vec_perm_d dzero = d; | |
23055 | if (d.op0 == CONST0_RTX (vmode)) | |
23056 | { | |
23057 | d.op1 = dzero.op1 = force_reg (vmode, d.op1); | |
23058 | std::swap (dzero.op0, dzero.op1); | |
23059 | for (i = 0; i < nelt; ++i) | |
23060 | dzero.perm[i] ^= nelt; | |
23061 | } | |
23062 | else | |
23063 | d.op0 = dzero.op0 = force_reg (vmode, d.op0); | |
23064 | ||
23065 | if (expand_vselect_vconcat (dzero.target, dzero.op0, dzero.op1, | |
23066 | dzero.perm, nelt, dzero.testing_p)) | |
23067 | return true; | |
23068 | } | |
23069 | ||
23070 | /* Force operands into registers. */ | |
23071 | rtx nop0 = force_reg (vmode, d.op0); | |
23072 | if (d.op0 == d.op1) | |
23073 | d.op1 = nop0; | |
23074 | d.op0 = nop0; | |
23075 | d.op1 = force_reg (vmode, d.op1); | |
23076 | ||
2bf6d935 ML |
23077 | if (ix86_expand_vec_perm_const_1 (&d)) |
23078 | return true; | |
23079 | ||
23080 | /* If the selector says both arguments are needed, but the operands are the | |
23081 | same, the above tried to expand with one_operand_p and flattened selector. | |
23082 | If that didn't work, retry without one_operand_p; we succeeded with that | |
23083 | during testing. */ | |
23084 | if (two_args && d.one_operand_p) | |
23085 | { | |
23086 | d.one_operand_p = false; | |
23087 | memcpy (d.perm, perm, sizeof (perm)); | |
23088 | return ix86_expand_vec_perm_const_1 (&d); | |
23089 | } | |
23090 | ||
23091 | return false; | |
23092 | } | |
23093 | ||
23094 | void | |
23095 | ix86_expand_vec_extract_even_odd (rtx targ, rtx op0, rtx op1, unsigned odd) | |
23096 | { | |
23097 | struct expand_vec_perm_d d; | |
23098 | unsigned i, nelt; | |
23099 | ||
23100 | d.target = targ; | |
23101 | d.op0 = op0; | |
23102 | d.op1 = op1; | |
23103 | d.vmode = GET_MODE (targ); | |
23104 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
23105 | d.one_operand_p = false; | |
23106 | d.testing_p = false; | |
23107 | ||
23108 | for (i = 0; i < nelt; ++i) | |
23109 | d.perm[i] = i * 2 + odd; | |
23110 | ||
23111 | /* We'll either be able to implement the permutation directly... */ | |
23112 | if (expand_vec_perm_1 (&d)) | |
23113 | return; | |
23114 | ||
23115 | /* ... or we use the special-case patterns. */ | |
23116 | expand_vec_perm_even_odd_1 (&d, odd); | |
23117 | } | |
23118 | ||
23119 | static void | |
23120 | ix86_expand_vec_interleave (rtx targ, rtx op0, rtx op1, bool high_p) | |
23121 | { | |
23122 | struct expand_vec_perm_d d; | |
23123 | unsigned i, nelt, base; | |
23124 | bool ok; | |
23125 | ||
23126 | d.target = targ; | |
23127 | d.op0 = op0; | |
23128 | d.op1 = op1; | |
23129 | d.vmode = GET_MODE (targ); | |
23130 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
23131 | d.one_operand_p = false; | |
23132 | d.testing_p = false; | |
23133 | ||
23134 | base = high_p ? nelt / 2 : 0; | |
23135 | for (i = 0; i < nelt / 2; ++i) | |
23136 | { | |
23137 | d.perm[i * 2] = i + base; | |
23138 | d.perm[i * 2 + 1] = i + base + nelt; | |
23139 | } | |
23140 | ||
23141 | /* Note that for AVX this isn't one instruction. */ | |
23142 | ok = ix86_expand_vec_perm_const_1 (&d); | |
23143 | gcc_assert (ok); | |
23144 | } | |
23145 | ||
c7199fb6 | 23146 | /* Expand a vector operation shift by constant for a V*QImode in terms of the |
23147 | same operation on V*HImode. Return true if success. */ | |
3bd86940 | 23148 | static bool |
23149 | ix86_expand_vec_shift_qihi_constant (enum rtx_code code, | |
23150 | rtx dest, rtx op1, rtx op2) | |
c7199fb6 | 23151 | { |
23152 | machine_mode qimode, himode; | |
c44c2a3b | 23153 | HOST_WIDE_INT and_constant, xor_constant; |
c7199fb6 | 23154 | HOST_WIDE_INT shift_amount; |
23155 | rtx vec_const_and, vec_const_xor; | |
23156 | rtx tmp, op1_subreg; | |
23157 | rtx (*gen_shift) (rtx, rtx, rtx); | |
23158 | rtx (*gen_and) (rtx, rtx, rtx); | |
23159 | rtx (*gen_xor) (rtx, rtx, rtx); | |
23160 | rtx (*gen_sub) (rtx, rtx, rtx); | |
23161 | ||
23162 | /* Only optimize shift by constant. */ | |
23163 | if (!CONST_INT_P (op2)) | |
23164 | return false; | |
23165 | ||
23166 | qimode = GET_MODE (dest); | |
23167 | shift_amount = INTVAL (op2); | |
23168 | /* Do nothing when shift amount greater equal 8. */ | |
23169 | if (shift_amount > 7) | |
23170 | return false; | |
23171 | ||
23172 | gcc_assert (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT); | |
23173 | /* Record sign bit. */ | |
23174 | xor_constant = 1 << (8 - shift_amount - 1); | |
23175 | ||
23176 | /* Zero upper/lower bits shift from left/right element. */ | |
23177 | and_constant | |
23178 | = (code == ASHIFT ? 256 - (1 << shift_amount) | |
23179 | : (1 << (8 - shift_amount)) - 1); | |
23180 | ||
23181 | switch (qimode) | |
23182 | { | |
23183 | case V16QImode: | |
23184 | himode = V8HImode; | |
23185 | gen_shift = | |
23186 | ((code == ASHIFT) | |
23187 | ? gen_ashlv8hi3 | |
23188 | : (code == ASHIFTRT) ? gen_ashrv8hi3 : gen_lshrv8hi3); | |
23189 | gen_and = gen_andv16qi3; | |
23190 | gen_xor = gen_xorv16qi3; | |
23191 | gen_sub = gen_subv16qi3; | |
23192 | break; | |
23193 | case V32QImode: | |
23194 | himode = V16HImode; | |
23195 | gen_shift = | |
23196 | ((code == ASHIFT) | |
23197 | ? gen_ashlv16hi3 | |
23198 | : (code == ASHIFTRT) ? gen_ashrv16hi3 : gen_lshrv16hi3); | |
23199 | gen_and = gen_andv32qi3; | |
23200 | gen_xor = gen_xorv32qi3; | |
23201 | gen_sub = gen_subv32qi3; | |
23202 | break; | |
23203 | case V64QImode: | |
23204 | himode = V32HImode; | |
23205 | gen_shift = | |
23206 | ((code == ASHIFT) | |
23207 | ? gen_ashlv32hi3 | |
23208 | : (code == ASHIFTRT) ? gen_ashrv32hi3 : gen_lshrv32hi3); | |
23209 | gen_and = gen_andv64qi3; | |
23210 | gen_xor = gen_xorv64qi3; | |
23211 | gen_sub = gen_subv64qi3; | |
23212 | break; | |
23213 | default: | |
23214 | gcc_unreachable (); | |
23215 | } | |
23216 | ||
23217 | tmp = gen_reg_rtx (himode); | |
23218 | vec_const_and = gen_reg_rtx (qimode); | |
23219 | op1_subreg = lowpart_subreg (himode, op1, qimode); | |
23220 | ||
23221 | /* For ASHIFT and LSHIFTRT, perform operation like | |
23222 | vpsllw/vpsrlw $shift_amount, %op1, %dest. | |
23223 | vpand %vec_const_and, %dest. */ | |
23224 | emit_insn (gen_shift (tmp, op1_subreg, op2)); | |
23225 | emit_move_insn (dest, simplify_gen_subreg (qimode, tmp, himode, 0)); | |
23226 | emit_move_insn (vec_const_and, | |
23227 | ix86_build_const_vector (qimode, true, | |
c44c2a3b | 23228 | gen_int_mode (and_constant, QImode))); |
c7199fb6 | 23229 | emit_insn (gen_and (dest, dest, vec_const_and)); |
23230 | ||
23231 | /* For ASHIFTRT, perform extra operation like | |
23232 | vpxor %vec_const_xor, %dest, %dest | |
23233 | vpsubb %vec_const_xor, %dest, %dest */ | |
23234 | if (code == ASHIFTRT) | |
23235 | { | |
23236 | vec_const_xor = gen_reg_rtx (qimode); | |
23237 | emit_move_insn (vec_const_xor, | |
23238 | ix86_build_const_vector (qimode, true, | |
c44c2a3b | 23239 | gen_int_mode (xor_constant, QImode))); |
c7199fb6 | 23240 | emit_insn (gen_xor (dest, dest, vec_const_xor)); |
23241 | emit_insn (gen_sub (dest, dest, vec_const_xor)); | |
23242 | } | |
23243 | return true; | |
23244 | } | |
23245 | ||
fe7b9c2e UB |
23246 | void |
23247 | ix86_expand_vecop_qihi_partial (enum rtx_code code, rtx dest, rtx op1, rtx op2) | |
23248 | { | |
23249 | machine_mode qimode = GET_MODE (dest); | |
52ff3f7b | 23250 | rtx qop1, qop2, hop1, hop2, qdest, hdest; |
fe7b9c2e | 23251 | bool op2vec = GET_MODE_CLASS (GET_MODE (op2)) == MODE_VECTOR_INT; |
52ff3f7b | 23252 | bool uns_p = code != ASHIFTRT; |
fe7b9c2e UB |
23253 | |
23254 | switch (qimode) | |
23255 | { | |
23256 | case E_V4QImode: | |
23257 | case E_V8QImode: | |
23258 | break; | |
23259 | default: | |
23260 | gcc_unreachable (); | |
23261 | } | |
23262 | ||
23263 | qop1 = lowpart_subreg (V16QImode, force_reg (qimode, op1), qimode); | |
23264 | ||
23265 | if (op2vec) | |
23266 | qop2 = lowpart_subreg (V16QImode, force_reg (qimode, op2), qimode); | |
23267 | else | |
23268 | qop2 = op2; | |
23269 | ||
c53f5100 UB |
23270 | qdest = gen_reg_rtx (V16QImode); |
23271 | ||
23272 | if (CONST_INT_P (op2) | |
23273 | && (code == ASHIFT || code == LSHIFTRT || code == ASHIFTRT) | |
23274 | && ix86_expand_vec_shift_qihi_constant (code, qdest, qop1, qop2)) | |
23275 | { | |
23276 | emit_move_insn (dest, gen_lowpart (qimode, qdest)); | |
23277 | return; | |
23278 | } | |
23279 | ||
fe7b9c2e UB |
23280 | switch (code) |
23281 | { | |
23282 | case MULT: | |
23283 | gcc_assert (op2vec); | |
52ff3f7b UB |
23284 | if (!TARGET_SSE4_1) |
23285 | { | |
23286 | /* Unpack data such that we've got a source byte in each low byte | |
23287 | of each word. We don't care what goes into the high byte of | |
23288 | each word. Rather than trying to get zero in there, most | |
23289 | convenient is to let it be a copy of the low byte. */ | |
23290 | hop1 = copy_to_reg (qop1); | |
23291 | hop2 = copy_to_reg (qop2); | |
23292 | emit_insn (gen_vec_interleave_lowv16qi (hop1, hop1, hop1)); | |
23293 | emit_insn (gen_vec_interleave_lowv16qi (hop2, hop2, hop2)); | |
23294 | break; | |
23295 | } | |
fe7b9c2e UB |
23296 | /* FALLTHRU */ |
23297 | case ASHIFT: | |
52ff3f7b | 23298 | case ASHIFTRT: |
fe7b9c2e UB |
23299 | case LSHIFTRT: |
23300 | hop1 = gen_reg_rtx (V8HImode); | |
23301 | ix86_expand_sse_unpack (hop1, qop1, uns_p, false); | |
52ff3f7b | 23302 | /* mult/vashr/vlshr/vashl */ |
fe7b9c2e UB |
23303 | if (op2vec) |
23304 | { | |
23305 | hop2 = gen_reg_rtx (V8HImode); | |
23306 | ix86_expand_sse_unpack (hop2, qop2, uns_p, false); | |
23307 | } | |
23308 | else | |
23309 | hop2 = qop2; | |
23310 | ||
23311 | break; | |
23312 | default: | |
23313 | gcc_unreachable (); | |
23314 | } | |
23315 | ||
23316 | if (code != MULT && op2vec) | |
23317 | { | |
23318 | /* Expand vashr/vlshr/vashl. */ | |
52ff3f7b UB |
23319 | hdest = gen_reg_rtx (V8HImode); |
23320 | emit_insn (gen_rtx_SET (hdest, | |
fe7b9c2e UB |
23321 | simplify_gen_binary (code, V8HImode, |
23322 | hop1, hop2))); | |
23323 | } | |
23324 | else | |
23325 | /* Expand mult/ashr/lshr/ashl. */ | |
52ff3f7b | 23326 | hdest = expand_simple_binop (V8HImode, code, hop1, hop2, |
fe7b9c2e UB |
23327 | NULL_RTX, 1, OPTAB_DIRECT); |
23328 | ||
23329 | if (TARGET_AVX512BW && TARGET_AVX512VL) | |
23330 | { | |
23331 | if (qimode == V8QImode) | |
23332 | qdest = dest; | |
23333 | else | |
23334 | qdest = gen_reg_rtx (V8QImode); | |
23335 | ||
52ff3f7b | 23336 | emit_insn (gen_truncv8hiv8qi2 (qdest, hdest)); |
fe7b9c2e UB |
23337 | } |
23338 | else | |
23339 | { | |
23340 | struct expand_vec_perm_d d; | |
52ff3f7b | 23341 | rtx qres = gen_lowpart (V16QImode, hdest); |
fe7b9c2e UB |
23342 | bool ok; |
23343 | int i; | |
23344 | ||
fe7b9c2e UB |
23345 | /* Merge the data back into the right place. */ |
23346 | d.target = qdest; | |
52ff3f7b | 23347 | d.op0 = d.op1 = qres; |
fe7b9c2e UB |
23348 | d.vmode = V16QImode; |
23349 | d.nelt = 16; | |
23350 | d.one_operand_p = false; | |
23351 | d.testing_p = false; | |
23352 | ||
23353 | for (i = 0; i < d.nelt; ++i) | |
23354 | d.perm[i] = i * 2; | |
23355 | ||
23356 | ok = ix86_expand_vec_perm_const_1 (&d); | |
23357 | gcc_assert (ok); | |
23358 | } | |
23359 | ||
23360 | if (qdest != dest) | |
23361 | emit_move_insn (dest, gen_lowpart (qimode, qdest)); | |
23362 | } | |
23363 | ||
52ff3f7b UB |
23364 | /* Emit instruction in 2x wider mode. For example, optimize |
23365 | vector MUL generation like | |
23366 | ||
23367 | vpmovzxbw ymm2, xmm0 | |
23368 | vpmovzxbw ymm3, xmm1 | |
23369 | vpmullw ymm4, ymm2, ymm3 | |
23370 | vpmovwb xmm0, ymm4 | |
23371 | ||
23372 | it would take less instructions than ix86_expand_vecop_qihi. | |
23373 | Return true if success. */ | |
23374 | ||
23375 | static bool | |
23376 | ix86_expand_vecop_qihi2 (enum rtx_code code, rtx dest, rtx op1, rtx op2) | |
23377 | { | |
23378 | machine_mode himode, qimode = GET_MODE (dest); | |
23379 | machine_mode wqimode; | |
23380 | rtx qop1, qop2, hop1, hop2, hdest; | |
23381 | rtx (*gen_truncate)(rtx, rtx) = NULL; | |
23382 | bool op2vec = GET_MODE_CLASS (GET_MODE (op2)) == MODE_VECTOR_INT; | |
23383 | bool uns_p = code != ASHIFTRT; | |
23384 | ||
23385 | if ((qimode == V16QImode && !TARGET_AVX2) | |
23386 | || (qimode == V32QImode && !TARGET_AVX512BW) | |
23387 | /* There are no V64HImode instructions. */ | |
23388 | || qimode == V64QImode) | |
23389 | return false; | |
23390 | ||
23391 | /* Do not generate ymm/zmm instructions when | |
23392 | target prefers 128/256 bit vector width. */ | |
23393 | if ((qimode == V16QImode && TARGET_PREFER_AVX128) | |
23394 | || (qimode == V32QImode && TARGET_PREFER_AVX256)) | |
23395 | return false; | |
23396 | ||
23397 | switch (qimode) | |
23398 | { | |
23399 | case E_V16QImode: | |
23400 | himode = V16HImode; | |
3c1e2b76 | 23401 | if (TARGET_AVX512VL && TARGET_AVX512BW) |
52ff3f7b UB |
23402 | gen_truncate = gen_truncv16hiv16qi2; |
23403 | break; | |
23404 | case E_V32QImode: | |
23405 | himode = V32HImode; | |
23406 | gen_truncate = gen_truncv32hiv32qi2; | |
23407 | break; | |
23408 | default: | |
23409 | gcc_unreachable (); | |
23410 | } | |
23411 | ||
23412 | wqimode = GET_MODE_2XWIDER_MODE (qimode).require (); | |
23413 | qop1 = lowpart_subreg (wqimode, force_reg (qimode, op1), qimode); | |
23414 | ||
23415 | if (op2vec) | |
23416 | qop2 = lowpart_subreg (wqimode, force_reg (qimode, op2), qimode); | |
23417 | else | |
23418 | qop2 = op2; | |
23419 | ||
23420 | hop1 = gen_reg_rtx (himode); | |
23421 | ix86_expand_sse_unpack (hop1, qop1, uns_p, false); | |
23422 | ||
23423 | if (op2vec) | |
23424 | { | |
23425 | hop2 = gen_reg_rtx (himode); | |
23426 | ix86_expand_sse_unpack (hop2, qop2, uns_p, false); | |
23427 | } | |
23428 | else | |
23429 | hop2 = qop2; | |
23430 | ||
2720bbd5 UB |
23431 | if (code != MULT && op2vec) |
23432 | { | |
23433 | /* Expand vashr/vlshr/vashl. */ | |
23434 | hdest = gen_reg_rtx (himode); | |
23435 | emit_insn (gen_rtx_SET (hdest, | |
23436 | simplify_gen_binary (code, himode, | |
23437 | hop1, hop2))); | |
23438 | } | |
23439 | else | |
23440 | /* Expand mult/ashr/lshr/ashl. */ | |
23441 | hdest = expand_simple_binop (himode, code, hop1, hop2, | |
23442 | NULL_RTX, 1, OPTAB_DIRECT); | |
52ff3f7b UB |
23443 | |
23444 | if (gen_truncate) | |
23445 | emit_insn (gen_truncate (dest, hdest)); | |
23446 | else | |
23447 | { | |
23448 | struct expand_vec_perm_d d; | |
23449 | rtx wqdest = gen_reg_rtx (wqimode); | |
23450 | rtx wqres = gen_lowpart (wqimode, hdest); | |
23451 | bool ok; | |
23452 | int i; | |
23453 | ||
23454 | /* Merge the data back into the right place. */ | |
23455 | d.target = wqdest; | |
23456 | d.op0 = d.op1 = wqres; | |
23457 | d.vmode = wqimode; | |
23458 | d.nelt = GET_MODE_NUNITS (wqimode); | |
23459 | d.one_operand_p = false; | |
23460 | d.testing_p = false; | |
23461 | ||
23462 | for (i = 0; i < d.nelt; ++i) | |
23463 | d.perm[i] = i * 2; | |
23464 | ||
23465 | ok = ix86_expand_vec_perm_const_1 (&d); | |
23466 | gcc_assert (ok); | |
23467 | ||
23468 | emit_move_insn (dest, gen_lowpart (qimode, wqdest)); | |
23469 | } | |
23470 | ||
23471 | return true; | |
23472 | } | |
23473 | ||
2bf6d935 ML |
23474 | /* Expand a vector operation CODE for a V*QImode in terms of the |
23475 | same operation on V*HImode. */ | |
23476 | ||
23477 | void | |
23478 | ix86_expand_vecop_qihi (enum rtx_code code, rtx dest, rtx op1, rtx op2) | |
23479 | { | |
23480 | machine_mode qimode = GET_MODE (dest); | |
23481 | machine_mode himode; | |
23482 | rtx (*gen_il) (rtx, rtx, rtx); | |
23483 | rtx (*gen_ih) (rtx, rtx, rtx); | |
23484 | rtx op1_l, op1_h, op2_l, op2_h, res_l, res_h; | |
fe7b9c2e | 23485 | bool op2vec = GET_MODE_CLASS (GET_MODE (op2)) == MODE_VECTOR_INT; |
2bf6d935 | 23486 | struct expand_vec_perm_d d; |
00fffa91 | 23487 | bool full_interleave = true; |
52ff3f7b | 23488 | bool uns_p = code != ASHIFTRT; |
00fffa91 | 23489 | bool ok; |
2bf6d935 ML |
23490 | int i; |
23491 | ||
3bd86940 | 23492 | if (CONST_INT_P (op2) |
23493 | && (code == ASHIFT || code == LSHIFTRT || code == ASHIFTRT) | |
23494 | && ix86_expand_vec_shift_qihi_constant (code, dest, op1, op2)) | |
23495 | return; | |
23496 | ||
0368fc54 | 23497 | if (ix86_expand_vecop_qihi2 (code, dest, op1, op2)) |
3bd86940 | 23498 | return; |
23499 | ||
2bf6d935 ML |
23500 | switch (qimode) |
23501 | { | |
23502 | case E_V16QImode: | |
23503 | himode = V8HImode; | |
2bf6d935 ML |
23504 | break; |
23505 | case E_V32QImode: | |
23506 | himode = V16HImode; | |
2bf6d935 ML |
23507 | break; |
23508 | case E_V64QImode: | |
23509 | himode = V32HImode; | |
2bf6d935 ML |
23510 | break; |
23511 | default: | |
23512 | gcc_unreachable (); | |
23513 | } | |
23514 | ||
2bf6d935 ML |
23515 | switch (code) |
23516 | { | |
23517 | case MULT: | |
fe7b9c2e | 23518 | gcc_assert (op2vec); |
2bf6d935 ML |
23519 | /* Unpack data such that we've got a source byte in each low byte of |
23520 | each word. We don't care what goes into the high byte of each word. | |
23521 | Rather than trying to get zero in there, most convenient is to let | |
23522 | it be a copy of the low byte. */ | |
00fffa91 UB |
23523 | switch (qimode) |
23524 | { | |
23525 | case E_V16QImode: | |
23526 | gen_il = gen_vec_interleave_lowv16qi; | |
23527 | gen_ih = gen_vec_interleave_highv16qi; | |
23528 | break; | |
23529 | case E_V32QImode: | |
23530 | gen_il = gen_avx2_interleave_lowv32qi; | |
23531 | gen_ih = gen_avx2_interleave_highv32qi; | |
23532 | full_interleave = false; | |
23533 | break; | |
23534 | case E_V64QImode: | |
23535 | gen_il = gen_avx512bw_interleave_lowv64qi; | |
23536 | gen_ih = gen_avx512bw_interleave_highv64qi; | |
23537 | full_interleave = false; | |
23538 | break; | |
23539 | default: | |
23540 | gcc_unreachable (); | |
23541 | } | |
23542 | ||
2bf6d935 ML |
23543 | op2_l = gen_reg_rtx (qimode); |
23544 | op2_h = gen_reg_rtx (qimode); | |
23545 | emit_insn (gen_il (op2_l, op2, op2)); | |
23546 | emit_insn (gen_ih (op2_h, op2, op2)); | |
23547 | ||
23548 | op1_l = gen_reg_rtx (qimode); | |
23549 | op1_h = gen_reg_rtx (qimode); | |
23550 | emit_insn (gen_il (op1_l, op1, op1)); | |
23551 | emit_insn (gen_ih (op1_h, op1, op1)); | |
2bf6d935 ML |
23552 | break; |
23553 | ||
23554 | case ASHIFT: | |
52ff3f7b | 23555 | case ASHIFTRT: |
2bf6d935 | 23556 | case LSHIFTRT: |
2bf6d935 ML |
23557 | op1_l = gen_reg_rtx (himode); |
23558 | op1_h = gen_reg_rtx (himode); | |
23559 | ix86_expand_sse_unpack (op1_l, op1, uns_p, false); | |
23560 | ix86_expand_sse_unpack (op1_h, op1, uns_p, true); | |
3bd86940 | 23561 | /* vashr/vlshr/vashl */ |
fe7b9c2e | 23562 | if (op2vec) |
3bd86940 | 23563 | { |
23564 | rtx tmp = force_reg (qimode, op2); | |
23565 | op2_l = gen_reg_rtx (himode); | |
23566 | op2_h = gen_reg_rtx (himode); | |
23567 | ix86_expand_sse_unpack (op2_l, tmp, uns_p, false); | |
23568 | ix86_expand_sse_unpack (op2_h, tmp, uns_p, true); | |
23569 | } | |
23570 | else | |
23571 | op2_l = op2_h = op2; | |
23572 | ||
2bf6d935 ML |
23573 | break; |
23574 | default: | |
23575 | gcc_unreachable (); | |
23576 | } | |
23577 | ||
fe7b9c2e | 23578 | if (code != MULT && op2vec) |
3bd86940 | 23579 | { |
00fffa91 | 23580 | /* Expand vashr/vlshr/vashl. */ |
3bd86940 | 23581 | res_l = gen_reg_rtx (himode); |
23582 | res_h = gen_reg_rtx (himode); | |
23583 | emit_insn (gen_rtx_SET (res_l, | |
23584 | simplify_gen_binary (code, himode, | |
23585 | op1_l, op2_l))); | |
23586 | emit_insn (gen_rtx_SET (res_h, | |
23587 | simplify_gen_binary (code, himode, | |
23588 | op1_h, op2_h))); | |
23589 | } | |
3bd86940 | 23590 | else |
23591 | { | |
00fffa91 | 23592 | /* Expand mult/ashr/lshr/ashl. */ |
3bd86940 | 23593 | res_l = expand_simple_binop (himode, code, op1_l, op2_l, NULL_RTX, |
23594 | 1, OPTAB_DIRECT); | |
23595 | res_h = expand_simple_binop (himode, code, op1_h, op2_h, NULL_RTX, | |
23596 | 1, OPTAB_DIRECT); | |
23597 | } | |
23598 | ||
2bf6d935 ML |
23599 | gcc_assert (res_l && res_h); |
23600 | ||
23601 | /* Merge the data back into the right place. */ | |
23602 | d.target = dest; | |
23603 | d.op0 = gen_lowpart (qimode, res_l); | |
23604 | d.op1 = gen_lowpart (qimode, res_h); | |
23605 | d.vmode = qimode; | |
23606 | d.nelt = GET_MODE_NUNITS (qimode); | |
23607 | d.one_operand_p = false; | |
23608 | d.testing_p = false; | |
23609 | ||
23610 | if (full_interleave) | |
23611 | { | |
00fffa91 | 23612 | /* We used the full interleave, the desired |
2bf6d935 ML |
23613 | results are in the even elements. */ |
23614 | for (i = 0; i < d.nelt; ++i) | |
23615 | d.perm[i] = i * 2; | |
23616 | } | |
23617 | else | |
23618 | { | |
23619 | /* For AVX, the interleave used above was not cross-lane. So the | |
23620 | extraction is evens but with the second and third quarter swapped. | |
23621 | Happily, that is even one insn shorter than even extraction. | |
23622 | For AVX512BW we have 4 lanes. We extract evens from within a lane, | |
23623 | always first from the first and then from the second source operand, | |
23624 | the index bits above the low 4 bits remains the same. | |
23625 | Thus, for d.nelt == 32 we want permutation | |
23626 | 0,2,4,..14, 32,34,36,..46, 16,18,20,..30, 48,50,52,..62 | |
23627 | and for d.nelt == 64 we want permutation | |
23628 | 0,2,4,..14, 64,66,68,..78, 16,18,20,..30, 80,82,84,..94, | |
23629 | 32,34,36,..46, 96,98,100,..110, 48,50,52,..62, 112,114,116,..126. */ | |
23630 | for (i = 0; i < d.nelt; ++i) | |
23631 | d.perm[i] = ((i * 2) & 14) + ((i & 8) ? d.nelt : 0) + (i & ~15); | |
23632 | } | |
23633 | ||
23634 | ok = ix86_expand_vec_perm_const_1 (&d); | |
23635 | gcc_assert (ok); | |
2bf6d935 ML |
23636 | } |
23637 | ||
23638 | /* Helper function of ix86_expand_mul_widen_evenodd. Return true | |
23639 | if op is CONST_VECTOR with all odd elements equal to their | |
23640 | preceding element. */ | |
23641 | ||
23642 | static bool | |
23643 | const_vector_equal_evenodd_p (rtx op) | |
23644 | { | |
23645 | machine_mode mode = GET_MODE (op); | |
23646 | int i, nunits = GET_MODE_NUNITS (mode); | |
23647 | if (GET_CODE (op) != CONST_VECTOR | |
23648 | || nunits != CONST_VECTOR_NUNITS (op)) | |
23649 | return false; | |
23650 | for (i = 0; i < nunits; i += 2) | |
23651 | if (CONST_VECTOR_ELT (op, i) != CONST_VECTOR_ELT (op, i + 1)) | |
23652 | return false; | |
23653 | return true; | |
23654 | } | |
23655 | ||
23656 | void | |
23657 | ix86_expand_mul_widen_evenodd (rtx dest, rtx op1, rtx op2, | |
23658 | bool uns_p, bool odd_p) | |
23659 | { | |
23660 | machine_mode mode = GET_MODE (op1); | |
23661 | machine_mode wmode = GET_MODE (dest); | |
23662 | rtx x; | |
23663 | rtx orig_op1 = op1, orig_op2 = op2; | |
23664 | ||
23665 | if (!nonimmediate_operand (op1, mode)) | |
23666 | op1 = force_reg (mode, op1); | |
23667 | if (!nonimmediate_operand (op2, mode)) | |
23668 | op2 = force_reg (mode, op2); | |
23669 | ||
23670 | /* We only play even/odd games with vectors of SImode. */ | |
23671 | gcc_assert (mode == V4SImode || mode == V8SImode || mode == V16SImode); | |
23672 | ||
23673 | /* If we're looking for the odd results, shift those members down to | |
23674 | the even slots. For some cpus this is faster than a PSHUFD. */ | |
23675 | if (odd_p) | |
23676 | { | |
23677 | /* For XOP use vpmacsdqh, but only for smult, as it is only | |
23678 | signed. */ | |
23679 | if (TARGET_XOP && mode == V4SImode && !uns_p) | |
23680 | { | |
23681 | x = force_reg (wmode, CONST0_RTX (wmode)); | |
23682 | emit_insn (gen_xop_pmacsdqh (dest, op1, op2, x)); | |
23683 | return; | |
23684 | } | |
23685 | ||
23686 | x = GEN_INT (GET_MODE_UNIT_BITSIZE (mode)); | |
23687 | if (!const_vector_equal_evenodd_p (orig_op1)) | |
23688 | op1 = expand_binop (wmode, lshr_optab, gen_lowpart (wmode, op1), | |
23689 | x, NULL, 1, OPTAB_DIRECT); | |
23690 | if (!const_vector_equal_evenodd_p (orig_op2)) | |
23691 | op2 = expand_binop (wmode, lshr_optab, gen_lowpart (wmode, op2), | |
23692 | x, NULL, 1, OPTAB_DIRECT); | |
23693 | op1 = gen_lowpart (mode, op1); | |
23694 | op2 = gen_lowpart (mode, op2); | |
23695 | } | |
23696 | ||
23697 | if (mode == V16SImode) | |
23698 | { | |
23699 | if (uns_p) | |
23700 | x = gen_vec_widen_umult_even_v16si (dest, op1, op2); | |
23701 | else | |
23702 | x = gen_vec_widen_smult_even_v16si (dest, op1, op2); | |
23703 | } | |
23704 | else if (mode == V8SImode) | |
23705 | { | |
23706 | if (uns_p) | |
23707 | x = gen_vec_widen_umult_even_v8si (dest, op1, op2); | |
23708 | else | |
23709 | x = gen_vec_widen_smult_even_v8si (dest, op1, op2); | |
23710 | } | |
23711 | else if (uns_p) | |
23712 | x = gen_vec_widen_umult_even_v4si (dest, op1, op2); | |
23713 | else if (TARGET_SSE4_1) | |
23714 | x = gen_sse4_1_mulv2siv2di3 (dest, op1, op2); | |
23715 | else | |
23716 | { | |
23717 | rtx s1, s2, t0, t1, t2; | |
23718 | ||
23719 | /* The easiest way to implement this without PMULDQ is to go through | |
23720 | the motions as if we are performing a full 64-bit multiply. With | |
23721 | the exception that we need to do less shuffling of the elements. */ | |
23722 | ||
23723 | /* Compute the sign-extension, aka highparts, of the two operands. */ | |
23724 | s1 = ix86_expand_sse_cmp (gen_reg_rtx (mode), GT, CONST0_RTX (mode), | |
23725 | op1, pc_rtx, pc_rtx); | |
23726 | s2 = ix86_expand_sse_cmp (gen_reg_rtx (mode), GT, CONST0_RTX (mode), | |
23727 | op2, pc_rtx, pc_rtx); | |
23728 | ||
23729 | /* Multiply LO(A) * HI(B), and vice-versa. */ | |
23730 | t1 = gen_reg_rtx (wmode); | |
23731 | t2 = gen_reg_rtx (wmode); | |
23732 | emit_insn (gen_vec_widen_umult_even_v4si (t1, s1, op2)); | |
23733 | emit_insn (gen_vec_widen_umult_even_v4si (t2, s2, op1)); | |
23734 | ||
23735 | /* Multiply LO(A) * LO(B). */ | |
23736 | t0 = gen_reg_rtx (wmode); | |
23737 | emit_insn (gen_vec_widen_umult_even_v4si (t0, op1, op2)); | |
23738 | ||
23739 | /* Combine and shift the highparts into place. */ | |
23740 | t1 = expand_binop (wmode, add_optab, t1, t2, t1, 1, OPTAB_DIRECT); | |
23741 | t1 = expand_binop (wmode, ashl_optab, t1, GEN_INT (32), t1, | |
23742 | 1, OPTAB_DIRECT); | |
23743 | ||
23744 | /* Combine high and low parts. */ | |
23745 | force_expand_binop (wmode, add_optab, t0, t1, dest, 1, OPTAB_DIRECT); | |
23746 | return; | |
23747 | } | |
23748 | emit_insn (x); | |
23749 | } | |
23750 | ||
23751 | void | |
23752 | ix86_expand_mul_widen_hilo (rtx dest, rtx op1, rtx op2, | |
23753 | bool uns_p, bool high_p) | |
23754 | { | |
23755 | machine_mode wmode = GET_MODE (dest); | |
23756 | machine_mode mode = GET_MODE (op1); | |
23757 | rtx t1, t2, t3, t4, mask; | |
23758 | ||
23759 | switch (mode) | |
23760 | { | |
23761 | case E_V4SImode: | |
23762 | t1 = gen_reg_rtx (mode); | |
23763 | t2 = gen_reg_rtx (mode); | |
23764 | if (TARGET_XOP && !uns_p) | |
23765 | { | |
23766 | /* With XOP, we have pmacsdqh, aka mul_widen_odd. In this case, | |
23767 | shuffle the elements once so that all elements are in the right | |
23768 | place for immediate use: { A C B D }. */ | |
23769 | emit_insn (gen_sse2_pshufd_1 (t1, op1, const0_rtx, const2_rtx, | |
23770 | const1_rtx, GEN_INT (3))); | |
23771 | emit_insn (gen_sse2_pshufd_1 (t2, op2, const0_rtx, const2_rtx, | |
23772 | const1_rtx, GEN_INT (3))); | |
23773 | } | |
23774 | else | |
23775 | { | |
23776 | /* Put the elements into place for the multiply. */ | |
23777 | ix86_expand_vec_interleave (t1, op1, op1, high_p); | |
23778 | ix86_expand_vec_interleave (t2, op2, op2, high_p); | |
23779 | high_p = false; | |
23780 | } | |
23781 | ix86_expand_mul_widen_evenodd (dest, t1, t2, uns_p, high_p); | |
23782 | break; | |
23783 | ||
23784 | case E_V8SImode: | |
23785 | /* Shuffle the elements between the lanes. After this we | |
23786 | have { A B E F | C D G H } for each operand. */ | |
23787 | t1 = gen_reg_rtx (V4DImode); | |
23788 | t2 = gen_reg_rtx (V4DImode); | |
23789 | emit_insn (gen_avx2_permv4di_1 (t1, gen_lowpart (V4DImode, op1), | |
23790 | const0_rtx, const2_rtx, | |
23791 | const1_rtx, GEN_INT (3))); | |
23792 | emit_insn (gen_avx2_permv4di_1 (t2, gen_lowpart (V4DImode, op2), | |
23793 | const0_rtx, const2_rtx, | |
23794 | const1_rtx, GEN_INT (3))); | |
23795 | ||
23796 | /* Shuffle the elements within the lanes. After this we | |
23797 | have { A A B B | C C D D } or { E E F F | G G H H }. */ | |
23798 | t3 = gen_reg_rtx (V8SImode); | |
23799 | t4 = gen_reg_rtx (V8SImode); | |
23800 | mask = GEN_INT (high_p | |
23801 | ? 2 + (2 << 2) + (3 << 4) + (3 << 6) | |
23802 | : 0 + (0 << 2) + (1 << 4) + (1 << 6)); | |
23803 | emit_insn (gen_avx2_pshufdv3 (t3, gen_lowpart (V8SImode, t1), mask)); | |
23804 | emit_insn (gen_avx2_pshufdv3 (t4, gen_lowpart (V8SImode, t2), mask)); | |
23805 | ||
23806 | ix86_expand_mul_widen_evenodd (dest, t3, t4, uns_p, false); | |
23807 | break; | |
23808 | ||
23809 | case E_V8HImode: | |
23810 | case E_V16HImode: | |
23811 | t1 = expand_binop (mode, smul_optab, op1, op2, NULL_RTX, | |
23812 | uns_p, OPTAB_DIRECT); | |
23813 | t2 = expand_binop (mode, | |
23814 | uns_p ? umul_highpart_optab : smul_highpart_optab, | |
23815 | op1, op2, NULL_RTX, uns_p, OPTAB_DIRECT); | |
23816 | gcc_assert (t1 && t2); | |
23817 | ||
23818 | t3 = gen_reg_rtx (mode); | |
23819 | ix86_expand_vec_interleave (t3, t1, t2, high_p); | |
23820 | emit_move_insn (dest, gen_lowpart (wmode, t3)); | |
23821 | break; | |
23822 | ||
23823 | case E_V16QImode: | |
23824 | case E_V32QImode: | |
23825 | case E_V32HImode: | |
23826 | case E_V16SImode: | |
23827 | case E_V64QImode: | |
23828 | t1 = gen_reg_rtx (wmode); | |
23829 | t2 = gen_reg_rtx (wmode); | |
23830 | ix86_expand_sse_unpack (t1, op1, uns_p, high_p); | |
23831 | ix86_expand_sse_unpack (t2, op2, uns_p, high_p); | |
23832 | ||
23833 | emit_insn (gen_rtx_SET (dest, gen_rtx_MULT (wmode, t1, t2))); | |
23834 | break; | |
23835 | ||
23836 | default: | |
23837 | gcc_unreachable (); | |
23838 | } | |
23839 | } | |
23840 | ||
23841 | void | |
23842 | ix86_expand_sse2_mulv4si3 (rtx op0, rtx op1, rtx op2) | |
23843 | { | |
23844 | rtx res_1, res_2, res_3, res_4; | |
23845 | ||
23846 | res_1 = gen_reg_rtx (V4SImode); | |
23847 | res_2 = gen_reg_rtx (V4SImode); | |
23848 | res_3 = gen_reg_rtx (V2DImode); | |
23849 | res_4 = gen_reg_rtx (V2DImode); | |
23850 | ix86_expand_mul_widen_evenodd (res_3, op1, op2, true, false); | |
23851 | ix86_expand_mul_widen_evenodd (res_4, op1, op2, true, true); | |
23852 | ||
23853 | /* Move the results in element 2 down to element 1; we don't care | |
23854 | what goes in elements 2 and 3. Then we can merge the parts | |
23855 | back together with an interleave. | |
23856 | ||
23857 | Note that two other sequences were tried: | |
23858 | (1) Use interleaves at the start instead of psrldq, which allows | |
23859 | us to use a single shufps to merge things back at the end. | |
23860 | (2) Use shufps here to combine the two vectors, then pshufd to | |
23861 | put the elements in the correct order. | |
23862 | In both cases the cost of the reformatting stall was too high | |
23863 | and the overall sequence slower. */ | |
23864 | ||
23865 | emit_insn (gen_sse2_pshufd_1 (res_1, gen_lowpart (V4SImode, res_3), | |
23866 | const0_rtx, const2_rtx, | |
23867 | const0_rtx, const0_rtx)); | |
23868 | emit_insn (gen_sse2_pshufd_1 (res_2, gen_lowpart (V4SImode, res_4), | |
23869 | const0_rtx, const2_rtx, | |
23870 | const0_rtx, const0_rtx)); | |
23871 | res_1 = emit_insn (gen_vec_interleave_lowv4si (op0, res_1, res_2)); | |
23872 | ||
23873 | set_unique_reg_note (res_1, REG_EQUAL, gen_rtx_MULT (V4SImode, op1, op2)); | |
23874 | } | |
23875 | ||
23876 | void | |
23877 | ix86_expand_sse2_mulvxdi3 (rtx op0, rtx op1, rtx op2) | |
23878 | { | |
23879 | machine_mode mode = GET_MODE (op0); | |
23880 | rtx t1, t2, t3, t4, t5, t6; | |
23881 | ||
23882 | if (TARGET_AVX512DQ && mode == V8DImode) | |
23883 | emit_insn (gen_avx512dq_mulv8di3 (op0, op1, op2)); | |
23884 | else if (TARGET_AVX512DQ && TARGET_AVX512VL && mode == V4DImode) | |
23885 | emit_insn (gen_avx512dq_mulv4di3 (op0, op1, op2)); | |
23886 | else if (TARGET_AVX512DQ && TARGET_AVX512VL && mode == V2DImode) | |
23887 | emit_insn (gen_avx512dq_mulv2di3 (op0, op1, op2)); | |
23888 | else if (TARGET_XOP && mode == V2DImode) | |
23889 | { | |
23890 | /* op1: A,B,C,D, op2: E,F,G,H */ | |
23891 | op1 = gen_lowpart (V4SImode, op1); | |
23892 | op2 = gen_lowpart (V4SImode, op2); | |
23893 | ||
23894 | t1 = gen_reg_rtx (V4SImode); | |
23895 | t2 = gen_reg_rtx (V4SImode); | |
23896 | t3 = gen_reg_rtx (V2DImode); | |
23897 | t4 = gen_reg_rtx (V2DImode); | |
23898 | ||
23899 | /* t1: B,A,D,C */ | |
23900 | emit_insn (gen_sse2_pshufd_1 (t1, op1, | |
23901 | GEN_INT (1), | |
23902 | GEN_INT (0), | |
23903 | GEN_INT (3), | |
23904 | GEN_INT (2))); | |
23905 | ||
23906 | /* t2: (B*E),(A*F),(D*G),(C*H) */ | |
23907 | emit_insn (gen_mulv4si3 (t2, t1, op2)); | |
23908 | ||
23909 | /* t3: (B*E)+(A*F), (D*G)+(C*H) */ | |
23910 | emit_insn (gen_xop_phadddq (t3, t2)); | |
23911 | ||
23912 | /* t4: ((B*E)+(A*F))<<32, ((D*G)+(C*H))<<32 */ | |
23913 | emit_insn (gen_ashlv2di3 (t4, t3, GEN_INT (32))); | |
23914 | ||
23915 | /* Multiply lower parts and add all */ | |
23916 | t5 = gen_reg_rtx (V2DImode); | |
23917 | emit_insn (gen_vec_widen_umult_even_v4si (t5, | |
23918 | gen_lowpart (V4SImode, op1), | |
23919 | gen_lowpart (V4SImode, op2))); | |
8ba6ea87 | 23920 | force_expand_binop (mode, add_optab, t5, t4, op0, 1, OPTAB_DIRECT); |
2bf6d935 ML |
23921 | } |
23922 | else | |
23923 | { | |
23924 | machine_mode nmode; | |
23925 | rtx (*umul) (rtx, rtx, rtx); | |
23926 | ||
23927 | if (mode == V2DImode) | |
23928 | { | |
23929 | umul = gen_vec_widen_umult_even_v4si; | |
23930 | nmode = V4SImode; | |
23931 | } | |
23932 | else if (mode == V4DImode) | |
23933 | { | |
23934 | umul = gen_vec_widen_umult_even_v8si; | |
23935 | nmode = V8SImode; | |
23936 | } | |
23937 | else if (mode == V8DImode) | |
23938 | { | |
23939 | umul = gen_vec_widen_umult_even_v16si; | |
23940 | nmode = V16SImode; | |
23941 | } | |
23942 | else | |
23943 | gcc_unreachable (); | |
23944 | ||
23945 | ||
23946 | /* Multiply low parts. */ | |
23947 | t1 = gen_reg_rtx (mode); | |
23948 | emit_insn (umul (t1, gen_lowpart (nmode, op1), gen_lowpart (nmode, op2))); | |
23949 | ||
23950 | /* Shift input vectors right 32 bits so we can multiply high parts. */ | |
23951 | t6 = GEN_INT (32); | |
23952 | t2 = expand_binop (mode, lshr_optab, op1, t6, NULL, 1, OPTAB_DIRECT); | |
23953 | t3 = expand_binop (mode, lshr_optab, op2, t6, NULL, 1, OPTAB_DIRECT); | |
23954 | ||
23955 | /* Multiply high parts by low parts. */ | |
23956 | t4 = gen_reg_rtx (mode); | |
23957 | t5 = gen_reg_rtx (mode); | |
23958 | emit_insn (umul (t4, gen_lowpart (nmode, t2), gen_lowpart (nmode, op2))); | |
23959 | emit_insn (umul (t5, gen_lowpart (nmode, t3), gen_lowpart (nmode, op1))); | |
23960 | ||
23961 | /* Combine and shift the highparts back. */ | |
23962 | t4 = expand_binop (mode, add_optab, t4, t5, t4, 1, OPTAB_DIRECT); | |
23963 | t4 = expand_binop (mode, ashl_optab, t4, t6, t4, 1, OPTAB_DIRECT); | |
23964 | ||
23965 | /* Combine high and low parts. */ | |
23966 | force_expand_binop (mode, add_optab, t1, t4, op0, 1, OPTAB_DIRECT); | |
23967 | } | |
23968 | ||
23969 | set_unique_reg_note (get_last_insn (), REG_EQUAL, | |
23970 | gen_rtx_MULT (mode, op1, op2)); | |
23971 | } | |
23972 | ||
23973 | /* Return 1 if control tansfer instruction INSN | |
23974 | should be encoded with notrack prefix. */ | |
23975 | ||
23976 | bool | |
e8b0314a | 23977 | ix86_notrack_prefixed_insn_p (rtx_insn *insn) |
2bf6d935 ML |
23978 | { |
23979 | if (!insn || !((flag_cf_protection & CF_BRANCH))) | |
23980 | return false; | |
23981 | ||
23982 | if (CALL_P (insn)) | |
23983 | { | |
23984 | rtx call = get_call_rtx_from (insn); | |
23985 | gcc_assert (call != NULL_RTX); | |
23986 | rtx addr = XEXP (call, 0); | |
23987 | ||
23988 | /* Do not emit 'notrack' if it's not an indirect call. */ | |
23989 | if (MEM_P (addr) | |
23990 | && GET_CODE (XEXP (addr, 0)) == SYMBOL_REF) | |
23991 | return false; | |
23992 | else | |
23993 | return find_reg_note (insn, REG_CALL_NOCF_CHECK, 0); | |
23994 | } | |
23995 | ||
23996 | if (JUMP_P (insn) && !flag_cet_switch) | |
23997 | { | |
23998 | rtx target = JUMP_LABEL (insn); | |
23999 | if (target == NULL_RTX || ANY_RETURN_P (target)) | |
24000 | return false; | |
24001 | ||
24002 | /* Check the jump is a switch table. */ | |
24003 | rtx_insn *label = as_a<rtx_insn *> (target); | |
24004 | rtx_insn *table = next_insn (label); | |
24005 | if (table == NULL_RTX || !JUMP_TABLE_DATA_P (table)) | |
24006 | return false; | |
24007 | else | |
24008 | return true; | |
24009 | } | |
24010 | return false; | |
24011 | } | |
24012 | ||
24013 | /* Calculate integer abs() using only SSE2 instructions. */ | |
24014 | ||
24015 | void | |
24016 | ix86_expand_sse2_abs (rtx target, rtx input) | |
24017 | { | |
24018 | machine_mode mode = GET_MODE (target); | |
24019 | rtx tmp0, tmp1, x; | |
24020 | ||
24021 | switch (mode) | |
24022 | { | |
24023 | case E_V2DImode: | |
24024 | case E_V4DImode: | |
24025 | /* For 64-bit signed integer X, with SSE4.2 use | |
24026 | pxor t0, t0; pcmpgtq X, t0; pxor t0, X; psubq t0, X. | |
24027 | Otherwise handle it similarly to V4SImode, except use 64 as W instead of | |
24028 | 32 and use logical instead of arithmetic right shift (which is | |
24029 | unimplemented) and subtract. */ | |
24030 | if (TARGET_SSE4_2) | |
24031 | { | |
24032 | tmp0 = gen_reg_rtx (mode); | |
24033 | tmp1 = gen_reg_rtx (mode); | |
24034 | emit_move_insn (tmp1, CONST0_RTX (mode)); | |
24035 | if (mode == E_V2DImode) | |
24036 | emit_insn (gen_sse4_2_gtv2di3 (tmp0, tmp1, input)); | |
24037 | else | |
24038 | emit_insn (gen_avx2_gtv4di3 (tmp0, tmp1, input)); | |
24039 | } | |
24040 | else | |
24041 | { | |
24042 | tmp0 = expand_simple_binop (mode, LSHIFTRT, input, | |
24043 | GEN_INT (GET_MODE_UNIT_BITSIZE (mode) | |
24044 | - 1), NULL, 0, OPTAB_DIRECT); | |
24045 | tmp0 = expand_simple_unop (mode, NEG, tmp0, NULL, false); | |
24046 | } | |
24047 | ||
24048 | tmp1 = expand_simple_binop (mode, XOR, tmp0, input, | |
24049 | NULL, 0, OPTAB_DIRECT); | |
24050 | x = expand_simple_binop (mode, MINUS, tmp1, tmp0, | |
24051 | target, 0, OPTAB_DIRECT); | |
24052 | break; | |
24053 | ||
24054 | case E_V4SImode: | |
24055 | /* For 32-bit signed integer X, the best way to calculate the absolute | |
24056 | value of X is (((signed) X >> (W-1)) ^ X) - ((signed) X >> (W-1)). */ | |
24057 | tmp0 = expand_simple_binop (mode, ASHIFTRT, input, | |
24058 | GEN_INT (GET_MODE_UNIT_BITSIZE (mode) - 1), | |
24059 | NULL, 0, OPTAB_DIRECT); | |
24060 | tmp1 = expand_simple_binop (mode, XOR, tmp0, input, | |
24061 | NULL, 0, OPTAB_DIRECT); | |
24062 | x = expand_simple_binop (mode, MINUS, tmp1, tmp0, | |
24063 | target, 0, OPTAB_DIRECT); | |
24064 | break; | |
24065 | ||
24066 | case E_V8HImode: | |
24067 | /* For 16-bit signed integer X, the best way to calculate the absolute | |
24068 | value of X is max (X, -X), as SSE2 provides the PMAXSW insn. */ | |
24069 | tmp0 = expand_unop (mode, neg_optab, input, NULL_RTX, 0); | |
24070 | ||
24071 | x = expand_simple_binop (mode, SMAX, tmp0, input, | |
24072 | target, 0, OPTAB_DIRECT); | |
24073 | break; | |
24074 | ||
24075 | case E_V16QImode: | |
24076 | /* For 8-bit signed integer X, the best way to calculate the absolute | |
24077 | value of X is min ((unsigned char) X, (unsigned char) (-X)), | |
24078 | as SSE2 provides the PMINUB insn. */ | |
24079 | tmp0 = expand_unop (mode, neg_optab, input, NULL_RTX, 0); | |
24080 | ||
24081 | x = expand_simple_binop (V16QImode, UMIN, tmp0, input, | |
24082 | target, 0, OPTAB_DIRECT); | |
24083 | break; | |
24084 | ||
24085 | default: | |
24086 | gcc_unreachable (); | |
24087 | } | |
24088 | ||
24089 | if (x != target) | |
24090 | emit_move_insn (target, x); | |
24091 | } | |
24092 | ||
24093 | /* Expand an extract from a vector register through pextr insn. | |
24094 | Return true if successful. */ | |
24095 | ||
24096 | bool | |
24097 | ix86_expand_pextr (rtx *operands) | |
24098 | { | |
24099 | rtx dst = operands[0]; | |
24100 | rtx src = operands[1]; | |
24101 | ||
24102 | unsigned int size = INTVAL (operands[2]); | |
24103 | unsigned int pos = INTVAL (operands[3]); | |
24104 | ||
24105 | if (SUBREG_P (dst)) | |
24106 | { | |
24107 | /* Reject non-lowpart subregs. */ | |
24108 | if (SUBREG_BYTE (dst) > 0) | |
24109 | return false; | |
24110 | dst = SUBREG_REG (dst); | |
24111 | } | |
24112 | ||
24113 | if (SUBREG_P (src)) | |
24114 | { | |
24115 | pos += SUBREG_BYTE (src) * BITS_PER_UNIT; | |
24116 | src = SUBREG_REG (src); | |
24117 | } | |
24118 | ||
24119 | switch (GET_MODE (src)) | |
24120 | { | |
24121 | case E_V16QImode: | |
24122 | case E_V8HImode: | |
24123 | case E_V4SImode: | |
24124 | case E_V2DImode: | |
24125 | case E_V1TImode: | |
2bf6d935 ML |
24126 | { |
24127 | machine_mode srcmode, dstmode; | |
24128 | rtx d, pat; | |
24129 | ||
24130 | if (!int_mode_for_size (size, 0).exists (&dstmode)) | |
24131 | return false; | |
24132 | ||
24133 | switch (dstmode) | |
24134 | { | |
24135 | case E_QImode: | |
24136 | if (!TARGET_SSE4_1) | |
24137 | return false; | |
24138 | srcmode = V16QImode; | |
24139 | break; | |
24140 | ||
24141 | case E_HImode: | |
24142 | if (!TARGET_SSE2) | |
24143 | return false; | |
24144 | srcmode = V8HImode; | |
24145 | break; | |
24146 | ||
24147 | case E_SImode: | |
24148 | if (!TARGET_SSE4_1) | |
24149 | return false; | |
24150 | srcmode = V4SImode; | |
24151 | break; | |
24152 | ||
24153 | case E_DImode: | |
24154 | gcc_assert (TARGET_64BIT); | |
24155 | if (!TARGET_SSE4_1) | |
24156 | return false; | |
24157 | srcmode = V2DImode; | |
24158 | break; | |
24159 | ||
24160 | default: | |
24161 | return false; | |
24162 | } | |
24163 | ||
24164 | /* Reject extractions from misaligned positions. */ | |
24165 | if (pos & (size-1)) | |
24166 | return false; | |
24167 | ||
24168 | if (GET_MODE (dst) == dstmode) | |
24169 | d = dst; | |
24170 | else | |
24171 | d = gen_reg_rtx (dstmode); | |
24172 | ||
24173 | /* Construct insn pattern. */ | |
24174 | pat = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (1, GEN_INT (pos / size))); | |
24175 | pat = gen_rtx_VEC_SELECT (dstmode, gen_lowpart (srcmode, src), pat); | |
24176 | ||
24177 | /* Let the rtl optimizers know about the zero extension performed. */ | |
24178 | if (dstmode == QImode || dstmode == HImode) | |
24179 | { | |
24180 | pat = gen_rtx_ZERO_EXTEND (SImode, pat); | |
24181 | d = gen_lowpart (SImode, d); | |
24182 | } | |
24183 | ||
24184 | emit_insn (gen_rtx_SET (d, pat)); | |
24185 | ||
24186 | if (d != dst) | |
24187 | emit_move_insn (dst, gen_lowpart (GET_MODE (dst), d)); | |
24188 | return true; | |
24189 | } | |
24190 | ||
24191 | default: | |
24192 | return false; | |
24193 | } | |
24194 | } | |
24195 | ||
24196 | /* Expand an insert into a vector register through pinsr insn. | |
24197 | Return true if successful. */ | |
24198 | ||
24199 | bool | |
24200 | ix86_expand_pinsr (rtx *operands) | |
24201 | { | |
24202 | rtx dst = operands[0]; | |
24203 | rtx src = operands[3]; | |
24204 | ||
24205 | unsigned int size = INTVAL (operands[1]); | |
24206 | unsigned int pos = INTVAL (operands[2]); | |
24207 | ||
24208 | if (SUBREG_P (dst)) | |
24209 | { | |
24210 | pos += SUBREG_BYTE (dst) * BITS_PER_UNIT; | |
24211 | dst = SUBREG_REG (dst); | |
24212 | } | |
24213 | ||
24214 | switch (GET_MODE (dst)) | |
24215 | { | |
24216 | case E_V16QImode: | |
24217 | case E_V8HImode: | |
24218 | case E_V4SImode: | |
24219 | case E_V2DImode: | |
24220 | case E_V1TImode: | |
2bf6d935 ML |
24221 | { |
24222 | machine_mode srcmode, dstmode; | |
24223 | rtx (*pinsr)(rtx, rtx, rtx, rtx); | |
24224 | rtx d; | |
24225 | ||
24226 | if (!int_mode_for_size (size, 0).exists (&srcmode)) | |
24227 | return false; | |
24228 | ||
24229 | switch (srcmode) | |
24230 | { | |
24231 | case E_QImode: | |
24232 | if (!TARGET_SSE4_1) | |
24233 | return false; | |
24234 | dstmode = V16QImode; | |
24235 | pinsr = gen_sse4_1_pinsrb; | |
24236 | break; | |
24237 | ||
24238 | case E_HImode: | |
24239 | if (!TARGET_SSE2) | |
24240 | return false; | |
24241 | dstmode = V8HImode; | |
24242 | pinsr = gen_sse2_pinsrw; | |
24243 | break; | |
24244 | ||
24245 | case E_SImode: | |
24246 | if (!TARGET_SSE4_1) | |
24247 | return false; | |
24248 | dstmode = V4SImode; | |
24249 | pinsr = gen_sse4_1_pinsrd; | |
24250 | break; | |
24251 | ||
24252 | case E_DImode: | |
24253 | gcc_assert (TARGET_64BIT); | |
24254 | if (!TARGET_SSE4_1) | |
24255 | return false; | |
24256 | dstmode = V2DImode; | |
24257 | pinsr = gen_sse4_1_pinsrq; | |
24258 | break; | |
24259 | ||
24260 | default: | |
24261 | return false; | |
24262 | } | |
24263 | ||
24264 | /* Reject insertions to misaligned positions. */ | |
24265 | if (pos & (size-1)) | |
24266 | return false; | |
24267 | ||
24268 | if (SUBREG_P (src)) | |
24269 | { | |
24270 | unsigned int srcpos = SUBREG_BYTE (src); | |
24271 | ||
24272 | if (srcpos > 0) | |
24273 | { | |
24274 | rtx extr_ops[4]; | |
24275 | ||
24276 | extr_ops[0] = gen_reg_rtx (srcmode); | |
24277 | extr_ops[1] = gen_lowpart (srcmode, SUBREG_REG (src)); | |
24278 | extr_ops[2] = GEN_INT (size); | |
24279 | extr_ops[3] = GEN_INT (srcpos * BITS_PER_UNIT); | |
24280 | ||
24281 | if (!ix86_expand_pextr (extr_ops)) | |
24282 | return false; | |
24283 | ||
24284 | src = extr_ops[0]; | |
24285 | } | |
24286 | else | |
24287 | src = gen_lowpart (srcmode, SUBREG_REG (src)); | |
24288 | } | |
24289 | ||
24290 | if (GET_MODE (dst) == dstmode) | |
24291 | d = dst; | |
24292 | else | |
24293 | d = gen_reg_rtx (dstmode); | |
24294 | ||
24295 | emit_insn (pinsr (d, gen_lowpart (dstmode, dst), | |
24296 | gen_lowpart (srcmode, src), | |
24297 | GEN_INT (1 << (pos / size)))); | |
24298 | if (d != dst) | |
24299 | emit_move_insn (dst, gen_lowpart (GET_MODE (dst), d)); | |
24300 | return true; | |
24301 | } | |
24302 | ||
24303 | default: | |
24304 | return false; | |
24305 | } | |
24306 | } | |
24307 | ||
24308 | /* All CPUs prefer to avoid cross-lane operations so perform reductions | |
24309 | upper against lower halves up to SSE reg size. */ | |
24310 | ||
24311 | machine_mode | |
24312 | ix86_split_reduction (machine_mode mode) | |
24313 | { | |
24314 | /* Reduce lowpart against highpart until we reach SSE reg width to | |
24315 | avoid cross-lane operations. */ | |
24316 | switch (mode) | |
24317 | { | |
24318 | case E_V8DImode: | |
24319 | case E_V4DImode: | |
24320 | return V2DImode; | |
24321 | case E_V16SImode: | |
24322 | case E_V8SImode: | |
24323 | return V4SImode; | |
24324 | case E_V32HImode: | |
24325 | case E_V16HImode: | |
24326 | return V8HImode; | |
24327 | case E_V64QImode: | |
24328 | case E_V32QImode: | |
24329 | return V16QImode; | |
24330 | case E_V16SFmode: | |
24331 | case E_V8SFmode: | |
24332 | return V4SFmode; | |
24333 | case E_V8DFmode: | |
24334 | case E_V4DFmode: | |
24335 | return V2DFmode; | |
24336 | default: | |
24337 | return mode; | |
24338 | } | |
24339 | } | |
24340 | ||
24341 | /* Generate call to __divmoddi4. */ | |
24342 | ||
24343 | void | |
24344 | ix86_expand_divmod_libfunc (rtx libfunc, machine_mode mode, | |
24345 | rtx op0, rtx op1, | |
24346 | rtx *quot_p, rtx *rem_p) | |
24347 | { | |
24348 | rtx rem = assign_386_stack_local (mode, SLOT_TEMP); | |
24349 | ||
24350 | rtx quot = emit_library_call_value (libfunc, NULL_RTX, LCT_NORMAL, | |
24351 | mode, op0, mode, op1, mode, | |
24352 | XEXP (rem, 0), Pmode); | |
24353 | *quot_p = quot; | |
24354 | *rem_p = rem; | |
24355 | } | |
24356 | ||
152f243f JJ |
24357 | void |
24358 | ix86_expand_atomic_fetch_op_loop (rtx target, rtx mem, rtx val, | |
24359 | enum rtx_code code, bool after, | |
24360 | bool doubleword) | |
4d281ff7 | 24361 | { |
0435b978 | 24362 | rtx old_reg, new_reg, old_mem, success; |
4d281ff7 | 24363 | machine_mode mode = GET_MODE (target); |
0435b978 | 24364 | rtx_code_label *loop_label = NULL; |
4d281ff7 HW |
24365 | |
24366 | old_reg = gen_reg_rtx (mode); | |
24367 | new_reg = old_reg; | |
4d281ff7 | 24368 | old_mem = copy_to_reg (mem); |
0435b978 | 24369 | loop_label = gen_label_rtx (); |
4d281ff7 HW |
24370 | emit_label (loop_label); |
24371 | emit_move_insn (old_reg, old_mem); | |
24372 | ||
24373 | /* return value for atomic_fetch_op. */ | |
24374 | if (!after) | |
24375 | emit_move_insn (target, old_reg); | |
24376 | ||
24377 | if (code == NOT) | |
24378 | { | |
24379 | new_reg = expand_simple_binop (mode, AND, new_reg, val, NULL_RTX, | |
24380 | true, OPTAB_LIB_WIDEN); | |
24381 | new_reg = expand_simple_unop (mode, code, new_reg, NULL_RTX, true); | |
24382 | } | |
24383 | else | |
24384 | new_reg = expand_simple_binop (mode, code, new_reg, val, NULL_RTX, | |
24385 | true, OPTAB_LIB_WIDEN); | |
24386 | ||
24387 | /* return value for atomic_op_fetch. */ | |
24388 | if (after) | |
24389 | emit_move_insn (target, new_reg); | |
24390 | ||
0435b978 HW |
24391 | success = NULL_RTX; |
24392 | ||
24393 | ix86_expand_cmpxchg_loop (&success, old_mem, mem, old_reg, new_reg, | |
24394 | gen_int_mode (MEMMODEL_SYNC_SEQ_CST, | |
24395 | SImode), | |
24396 | doubleword, loop_label); | |
24397 | } | |
24398 | ||
24399 | /* Relax cmpxchg instruction, param loop_label indicates whether | |
24400 | the instruction should be relaxed with a pause loop. If not, | |
24401 | it will be relaxed to an atomic load + compare, and skip | |
24402 | cmpxchg instruction if mem != exp_input. */ | |
24403 | ||
152f243f JJ |
24404 | void |
24405 | ix86_expand_cmpxchg_loop (rtx *ptarget_bool, rtx target_val, | |
24406 | rtx mem, rtx exp_input, rtx new_input, | |
24407 | rtx mem_model, bool doubleword, | |
24408 | rtx_code_label *loop_label) | |
0435b978 HW |
24409 | { |
24410 | rtx_code_label *cmp_label = NULL; | |
24411 | rtx_code_label *done_label = NULL; | |
24412 | rtx target_bool = NULL_RTX, new_mem = NULL_RTX; | |
24413 | rtx (*gen) (rtx, rtx, rtx, rtx, rtx) = NULL; | |
24414 | rtx (*gendw) (rtx, rtx, rtx, rtx, rtx, rtx) = NULL; | |
24415 | machine_mode mode = GET_MODE (target_val), hmode = mode; | |
24416 | ||
24417 | if (*ptarget_bool == NULL) | |
24418 | target_bool = gen_reg_rtx (QImode); | |
24419 | else | |
24420 | target_bool = *ptarget_bool; | |
24421 | ||
24422 | cmp_label = gen_label_rtx (); | |
24423 | done_label = gen_label_rtx (); | |
24424 | ||
24425 | new_mem = gen_reg_rtx (mode); | |
24426 | /* Load memory first. */ | |
24427 | expand_atomic_load (new_mem, mem, MEMMODEL_SEQ_CST); | |
24428 | ||
24429 | switch (mode) | |
24430 | { | |
9d1796d8 | 24431 | case E_TImode: |
0435b978 HW |
24432 | gendw = gen_atomic_compare_and_swapti_doubleword; |
24433 | hmode = DImode; | |
24434 | break; | |
9d1796d8 | 24435 | case E_DImode: |
0435b978 HW |
24436 | if (doubleword) |
24437 | { | |
24438 | gendw = gen_atomic_compare_and_swapdi_doubleword; | |
24439 | hmode = SImode; | |
24440 | } | |
24441 | else | |
24442 | gen = gen_atomic_compare_and_swapdi_1; | |
24443 | break; | |
9d1796d8 RS |
24444 | case E_SImode: |
24445 | gen = gen_atomic_compare_and_swapsi_1; | |
24446 | break; | |
24447 | case E_HImode: | |
24448 | gen = gen_atomic_compare_and_swaphi_1; | |
24449 | break; | |
24450 | case E_QImode: | |
24451 | gen = gen_atomic_compare_and_swapqi_1; | |
24452 | break; | |
0435b978 HW |
24453 | default: |
24454 | gcc_unreachable (); | |
24455 | } | |
4d281ff7 | 24456 | |
0435b978 | 24457 | /* Compare mem value with expected value. */ |
4d281ff7 HW |
24458 | if (doubleword) |
24459 | { | |
0435b978 HW |
24460 | rtx low_new_mem = gen_lowpart (hmode, new_mem); |
24461 | rtx low_exp_input = gen_lowpart (hmode, exp_input); | |
24462 | rtx high_new_mem = gen_highpart (hmode, new_mem); | |
24463 | rtx high_exp_input = gen_highpart (hmode, exp_input); | |
24464 | emit_cmp_and_jump_insns (low_new_mem, low_exp_input, NE, NULL_RTX, | |
24465 | hmode, 1, cmp_label, | |
4d281ff7 | 24466 | profile_probability::guessed_never ()); |
0435b978 HW |
24467 | emit_cmp_and_jump_insns (high_new_mem, high_exp_input, NE, NULL_RTX, |
24468 | hmode, 1, cmp_label, | |
4d281ff7 HW |
24469 | profile_probability::guessed_never ()); |
24470 | } | |
24471 | else | |
0435b978 HW |
24472 | emit_cmp_and_jump_insns (new_mem, exp_input, NE, NULL_RTX, |
24473 | GET_MODE (exp_input), 1, cmp_label, | |
4d281ff7 HW |
24474 | profile_probability::guessed_never ()); |
24475 | ||
0435b978 HW |
24476 | /* Directly emits cmpxchg here. */ |
24477 | if (doubleword) | |
24478 | emit_insn (gendw (target_val, mem, exp_input, | |
24479 | gen_lowpart (hmode, new_input), | |
24480 | gen_highpart (hmode, new_input), | |
24481 | mem_model)); | |
24482 | else | |
24483 | emit_insn (gen (target_val, mem, exp_input, new_input, mem_model)); | |
24484 | ||
24485 | if (!loop_label) | |
24486 | { | |
24487 | emit_jump_insn (gen_jump (done_label)); | |
24488 | emit_barrier (); | |
24489 | emit_label (cmp_label); | |
24490 | emit_move_insn (target_val, new_mem); | |
24491 | emit_label (done_label); | |
24492 | ix86_expand_setcc (target_bool, EQ, gen_rtx_REG (CCZmode, FLAGS_REG), | |
24493 | const0_rtx); | |
24494 | } | |
24495 | else | |
24496 | { | |
24497 | ix86_expand_setcc (target_bool, EQ, gen_rtx_REG (CCZmode, FLAGS_REG), | |
24498 | const0_rtx); | |
24499 | emit_cmp_and_jump_insns (target_bool, const0_rtx, EQ, const0_rtx, | |
24500 | GET_MODE (target_bool), 1, loop_label, | |
24501 | profile_probability::guessed_never ()); | |
24502 | emit_jump_insn (gen_jump (done_label)); | |
24503 | emit_barrier (); | |
24504 | ||
24505 | /* If mem is not expected, pause and loop back. */ | |
24506 | emit_label (cmp_label); | |
522f25e9 | 24507 | emit_move_insn (target_val, new_mem); |
0435b978 HW |
24508 | emit_insn (gen_pause ()); |
24509 | emit_jump_insn (gen_jump (loop_label)); | |
24510 | emit_barrier (); | |
24511 | emit_label (done_label); | |
24512 | } | |
24513 | ||
24514 | *ptarget_bool = target_bool; | |
4d281ff7 HW |
24515 | } |
24516 | ||
b1115dbf JJ |
24517 | /* Convert a BFmode VAL to SFmode without signaling sNaNs. |
24518 | This is done by returning SF SUBREG of ((HI SUBREG) (VAL)) << 16. */ | |
24519 | ||
24520 | rtx | |
24521 | ix86_expand_fast_convert_bf_to_sf (rtx val) | |
24522 | { | |
24523 | rtx op = gen_lowpart (HImode, val), ret; | |
24524 | if (CONST_INT_P (op)) | |
24525 | { | |
24526 | ret = simplify_const_unary_operation (FLOAT_EXTEND, SFmode, | |
24527 | val, BFmode); | |
24528 | if (ret) | |
24529 | return ret; | |
24530 | /* FLOAT_EXTEND simplification will fail if VAL is a sNaN. */ | |
24531 | ret = gen_reg_rtx (SImode); | |
24532 | emit_move_insn (ret, GEN_INT (INTVAL (op) & 0xffff)); | |
e55251f3 | 24533 | emit_insn (gen_ashlsi3 (ret, ret, GEN_INT (16))); |
24534 | return gen_lowpart (SFmode, ret); | |
b1115dbf | 24535 | } |
e55251f3 | 24536 | |
24537 | ret = gen_reg_rtx (SFmode); | |
24538 | emit_insn (gen_extendbfsf2_1 (ret, force_reg (BFmode, val))); | |
24539 | return ret; | |
b1115dbf JJ |
24540 | } |
24541 | ||
2bf6d935 | 24542 | #include "gt-i386-expand.h" |