]>
Commit | Line | Data |
---|---|---|
947ed59a | 1 | /* IR-agnostic target query functions relating to optabs |
fbd26352 | 2 | Copyright (C) 1987-2019 Free Software Foundation, Inc. |
947ed59a | 3 | |
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify it under | |
7 | the terms of the GNU General Public License as published by the Free | |
8 | Software Foundation; either version 3, or (at your option) any later | |
9 | version. | |
10 | ||
11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
12 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
17 | along with GCC; see the file COPYING3. If not see | |
18 | <http://www.gnu.org/licenses/>. */ | |
19 | ||
20 | ||
21 | #include "config.h" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
24 | #include "target.h" | |
25 | #include "insn-codes.h" | |
26 | #include "optabs-query.h" | |
27 | #include "optabs-libfuncs.h" | |
28 | #include "insn-config.h" | |
29 | #include "rtl.h" | |
30 | #include "recog.h" | |
d37760c5 | 31 | #include "vec-perm-indices.h" |
947ed59a | 32 | |
33 | struct target_optabs default_target_optabs; | |
34 | struct target_optabs *this_fn_optabs = &default_target_optabs; | |
35 | #if SWITCHABLE_TARGET | |
36 | struct target_optabs *this_target_optabs = &default_target_optabs; | |
37 | #endif | |
38 | ||
acdfe9e0 | 39 | /* Return the insn used to perform conversion OP from mode FROM_MODE |
40 | to mode TO_MODE; return CODE_FOR_nothing if the target does not have | |
41 | such an insn, or if it is unsuitable for optimization type OPT_TYPE. */ | |
42 | ||
43 | insn_code | |
44 | convert_optab_handler (convert_optab optab, machine_mode to_mode, | |
45 | machine_mode from_mode, optimization_type opt_type) | |
46 | { | |
47 | insn_code icode = convert_optab_handler (optab, to_mode, from_mode); | |
48 | if (icode == CODE_FOR_nothing | |
49 | || !targetm.optab_supported_p (optab, to_mode, from_mode, opt_type)) | |
50 | return CODE_FOR_nothing; | |
51 | return icode; | |
52 | } | |
53 | ||
54 | /* Return the insn used to implement mode MODE of OP; return | |
55 | CODE_FOR_nothing if the target does not have such an insn, | |
56 | or if it is unsuitable for optimization type OPT_TYPE. */ | |
57 | ||
58 | insn_code | |
59 | direct_optab_handler (convert_optab optab, machine_mode mode, | |
60 | optimization_type opt_type) | |
61 | { | |
62 | insn_code icode = direct_optab_handler (optab, mode); | |
63 | if (icode == CODE_FOR_nothing | |
64 | || !targetm.optab_supported_p (optab, mode, mode, opt_type)) | |
65 | return CODE_FOR_nothing; | |
66 | return icode; | |
67 | } | |
68 | ||
947ed59a | 69 | /* Enumerates the possible types of structure operand to an |
70 | extraction_insn. */ | |
71 | enum extraction_type { ET_unaligned_mem, ET_reg }; | |
72 | ||
73 | /* Check whether insv, extv or extzv pattern ICODE can be used for an | |
74 | insertion or extraction of type TYPE on a structure of mode MODE. | |
75 | Return true if so and fill in *INSN accordingly. STRUCT_OP is the | |
76 | operand number of the structure (the first sign_extract or zero_extract | |
77 | operand) and FIELD_OP is the operand number of the field (the other | |
78 | side of the set from the sign_extract or zero_extract). */ | |
79 | ||
80 | static bool | |
81 | get_traditional_extraction_insn (extraction_insn *insn, | |
82 | enum extraction_type type, | |
83 | machine_mode mode, | |
84 | enum insn_code icode, | |
85 | int struct_op, int field_op) | |
86 | { | |
87 | const struct insn_data_d *data = &insn_data[icode]; | |
88 | ||
89 | machine_mode struct_mode = data->operand[struct_op].mode; | |
90 | if (struct_mode == VOIDmode) | |
91 | struct_mode = word_mode; | |
92 | if (mode != struct_mode) | |
93 | return false; | |
94 | ||
95 | machine_mode field_mode = data->operand[field_op].mode; | |
96 | if (field_mode == VOIDmode) | |
97 | field_mode = word_mode; | |
98 | ||
99 | machine_mode pos_mode = data->operand[struct_op + 2].mode; | |
100 | if (pos_mode == VOIDmode) | |
101 | pos_mode = word_mode; | |
102 | ||
103 | insn->icode = icode; | |
54fea56d | 104 | insn->field_mode = as_a <scalar_int_mode> (field_mode); |
105 | if (type == ET_unaligned_mem) | |
106 | insn->struct_mode = byte_mode; | |
107 | else if (struct_mode == BLKmode) | |
108 | insn->struct_mode = opt_scalar_int_mode (); | |
109 | else | |
110 | insn->struct_mode = as_a <scalar_int_mode> (struct_mode); | |
111 | insn->pos_mode = as_a <scalar_int_mode> (pos_mode); | |
947ed59a | 112 | return true; |
113 | } | |
114 | ||
115 | /* Return true if an optab exists to perform an insertion or extraction | |
116 | of type TYPE in mode MODE. Describe the instruction in *INSN if so. | |
117 | ||
118 | REG_OPTAB is the optab to use for register structures and | |
119 | MISALIGN_OPTAB is the optab to use for misaligned memory structures. | |
120 | POS_OP is the operand number of the bit position. */ | |
121 | ||
122 | static bool | |
123 | get_optab_extraction_insn (struct extraction_insn *insn, | |
124 | enum extraction_type type, | |
125 | machine_mode mode, direct_optab reg_optab, | |
126 | direct_optab misalign_optab, int pos_op) | |
127 | { | |
128 | direct_optab optab = (type == ET_unaligned_mem ? misalign_optab : reg_optab); | |
129 | enum insn_code icode = direct_optab_handler (optab, mode); | |
130 | if (icode == CODE_FOR_nothing) | |
131 | return false; | |
132 | ||
133 | const struct insn_data_d *data = &insn_data[icode]; | |
134 | ||
54fea56d | 135 | machine_mode pos_mode = data->operand[pos_op].mode; |
136 | if (pos_mode == VOIDmode) | |
137 | pos_mode = word_mode; | |
138 | ||
947ed59a | 139 | insn->icode = icode; |
54fea56d | 140 | insn->field_mode = as_a <scalar_int_mode> (mode); |
141 | if (type == ET_unaligned_mem) | |
142 | insn->struct_mode = opt_scalar_int_mode (); | |
143 | else | |
144 | insn->struct_mode = insn->field_mode; | |
145 | insn->pos_mode = as_a <scalar_int_mode> (pos_mode); | |
947ed59a | 146 | return true; |
147 | } | |
148 | ||
149 | /* Return true if an instruction exists to perform an insertion or | |
150 | extraction (PATTERN says which) of type TYPE in mode MODE. | |
151 | Describe the instruction in *INSN if so. */ | |
152 | ||
153 | static bool | |
154 | get_extraction_insn (extraction_insn *insn, | |
155 | enum extraction_pattern pattern, | |
156 | enum extraction_type type, | |
157 | machine_mode mode) | |
158 | { | |
159 | switch (pattern) | |
160 | { | |
161 | case EP_insv: | |
162 | if (targetm.have_insv () | |
163 | && get_traditional_extraction_insn (insn, type, mode, | |
164 | targetm.code_for_insv, 0, 3)) | |
165 | return true; | |
166 | return get_optab_extraction_insn (insn, type, mode, insv_optab, | |
167 | insvmisalign_optab, 2); | |
168 | ||
169 | case EP_extv: | |
170 | if (targetm.have_extv () | |
171 | && get_traditional_extraction_insn (insn, type, mode, | |
172 | targetm.code_for_extv, 1, 0)) | |
173 | return true; | |
174 | return get_optab_extraction_insn (insn, type, mode, extv_optab, | |
175 | extvmisalign_optab, 3); | |
176 | ||
177 | case EP_extzv: | |
178 | if (targetm.have_extzv () | |
179 | && get_traditional_extraction_insn (insn, type, mode, | |
180 | targetm.code_for_extzv, 1, 0)) | |
181 | return true; | |
182 | return get_optab_extraction_insn (insn, type, mode, extzv_optab, | |
183 | extzvmisalign_optab, 3); | |
184 | ||
185 | default: | |
186 | gcc_unreachable (); | |
187 | } | |
188 | } | |
189 | ||
190 | /* Return true if an instruction exists to access a field of mode | |
191 | FIELDMODE in a structure that has STRUCT_BITS significant bits. | |
192 | Describe the "best" such instruction in *INSN if so. PATTERN and | |
193 | TYPE describe the type of insertion or extraction we want to perform. | |
194 | ||
195 | For an insertion, the number of significant structure bits includes | |
196 | all bits of the target. For an extraction, it need only include the | |
197 | most significant bit of the field. Larger widths are acceptable | |
198 | in both cases. */ | |
199 | ||
200 | static bool | |
201 | get_best_extraction_insn (extraction_insn *insn, | |
202 | enum extraction_pattern pattern, | |
203 | enum extraction_type type, | |
204 | unsigned HOST_WIDE_INT struct_bits, | |
205 | machine_mode field_mode) | |
206 | { | |
1a5d4b27 | 207 | opt_scalar_int_mode mode_iter; |
208 | FOR_EACH_MODE_FROM (mode_iter, smallest_int_mode_for_size (struct_bits)) | |
947ed59a | 209 | { |
1a5d4b27 | 210 | scalar_int_mode mode = mode_iter.require (); |
947ed59a | 211 | if (get_extraction_insn (insn, pattern, type, mode)) |
212 | { | |
1a5d4b27 | 213 | FOR_EACH_MODE_FROM (mode_iter, mode) |
947ed59a | 214 | { |
1a5d4b27 | 215 | mode = mode_iter.require (); |
52acb7ae | 216 | if (maybe_gt (GET_MODE_SIZE (mode), GET_MODE_SIZE (field_mode)) |
19a4dce4 | 217 | || TRULY_NOOP_TRUNCATION_MODES_P (insn->field_mode, |
218 | field_mode)) | |
219 | break; | |
947ed59a | 220 | get_extraction_insn (insn, pattern, type, mode); |
947ed59a | 221 | } |
222 | return true; | |
223 | } | |
947ed59a | 224 | } |
225 | return false; | |
226 | } | |
227 | ||
228 | /* Return true if an instruction exists to access a field of mode | |
229 | FIELDMODE in a register structure that has STRUCT_BITS significant bits. | |
230 | Describe the "best" such instruction in *INSN if so. PATTERN describes | |
231 | the type of insertion or extraction we want to perform. | |
232 | ||
233 | For an insertion, the number of significant structure bits includes | |
234 | all bits of the target. For an extraction, it need only include the | |
235 | most significant bit of the field. Larger widths are acceptable | |
236 | in both cases. */ | |
237 | ||
238 | bool | |
239 | get_best_reg_extraction_insn (extraction_insn *insn, | |
240 | enum extraction_pattern pattern, | |
241 | unsigned HOST_WIDE_INT struct_bits, | |
242 | machine_mode field_mode) | |
243 | { | |
244 | return get_best_extraction_insn (insn, pattern, ET_reg, struct_bits, | |
245 | field_mode); | |
246 | } | |
247 | ||
248 | /* Return true if an instruction exists to access a field of BITSIZE | |
249 | bits starting BITNUM bits into a memory structure. Describe the | |
250 | "best" such instruction in *INSN if so. PATTERN describes the type | |
251 | of insertion or extraction we want to perform and FIELDMODE is the | |
252 | natural mode of the extracted field. | |
253 | ||
254 | The instructions considered here only access bytes that overlap | |
255 | the bitfield; they do not touch any surrounding bytes. */ | |
256 | ||
257 | bool | |
258 | get_best_mem_extraction_insn (extraction_insn *insn, | |
259 | enum extraction_pattern pattern, | |
260 | HOST_WIDE_INT bitsize, HOST_WIDE_INT bitnum, | |
261 | machine_mode field_mode) | |
262 | { | |
263 | unsigned HOST_WIDE_INT struct_bits = (bitnum % BITS_PER_UNIT | |
264 | + bitsize | |
265 | + BITS_PER_UNIT - 1); | |
266 | struct_bits -= struct_bits % BITS_PER_UNIT; | |
267 | return get_best_extraction_insn (insn, pattern, ET_unaligned_mem, | |
268 | struct_bits, field_mode); | |
269 | } | |
270 | ||
271 | /* Return the insn code used to extend FROM_MODE to TO_MODE. | |
272 | UNSIGNEDP specifies zero-extension instead of sign-extension. If | |
273 | no such operation exists, CODE_FOR_nothing will be returned. */ | |
274 | ||
275 | enum insn_code | |
276 | can_extend_p (machine_mode to_mode, machine_mode from_mode, | |
277 | int unsignedp) | |
278 | { | |
279 | if (unsignedp < 0 && targetm.have_ptr_extend ()) | |
280 | return targetm.code_for_ptr_extend; | |
281 | ||
282 | convert_optab tab = unsignedp ? zext_optab : sext_optab; | |
283 | return convert_optab_handler (tab, to_mode, from_mode); | |
284 | } | |
285 | ||
286 | /* Return the insn code to convert fixed-point mode FIXMODE to floating-point | |
287 | mode FLTMODE, or CODE_FOR_nothing if no such instruction exists. | |
288 | UNSIGNEDP specifies whether FIXMODE is unsigned. */ | |
289 | ||
290 | enum insn_code | |
291 | can_float_p (machine_mode fltmode, machine_mode fixmode, | |
292 | int unsignedp) | |
293 | { | |
294 | convert_optab tab = unsignedp ? ufloat_optab : sfloat_optab; | |
295 | return convert_optab_handler (tab, fltmode, fixmode); | |
296 | } | |
297 | ||
298 | /* Return the insn code to convert floating-point mode FLTMODE to fixed-point | |
299 | mode FIXMODE, or CODE_FOR_nothing if no such instruction exists. | |
300 | UNSIGNEDP specifies whether FIXMODE is unsigned. | |
301 | ||
302 | On a successful return, set *TRUNCP_PTR to true if it is necessary to | |
303 | output an explicit FTRUNC before the instruction. */ | |
304 | ||
305 | enum insn_code | |
306 | can_fix_p (machine_mode fixmode, machine_mode fltmode, | |
307 | int unsignedp, bool *truncp_ptr) | |
308 | { | |
309 | convert_optab tab; | |
310 | enum insn_code icode; | |
311 | ||
312 | tab = unsignedp ? ufixtrunc_optab : sfixtrunc_optab; | |
313 | icode = convert_optab_handler (tab, fixmode, fltmode); | |
314 | if (icode != CODE_FOR_nothing) | |
315 | { | |
316 | *truncp_ptr = false; | |
317 | return icode; | |
318 | } | |
319 | ||
320 | /* FIXME: This requires a port to define both FIX and FTRUNC pattern | |
321 | for this to work. We need to rework the fix* and ftrunc* patterns | |
322 | and documentation. */ | |
323 | tab = unsignedp ? ufix_optab : sfix_optab; | |
324 | icode = convert_optab_handler (tab, fixmode, fltmode); | |
325 | if (icode != CODE_FOR_nothing | |
326 | && optab_handler (ftrunc_optab, fltmode) != CODE_FOR_nothing) | |
327 | { | |
328 | *truncp_ptr = true; | |
329 | return icode; | |
330 | } | |
331 | ||
332 | return CODE_FOR_nothing; | |
333 | } | |
334 | ||
335 | /* Return nonzero if a conditional move of mode MODE is supported. | |
336 | ||
337 | This function is for combine so it can tell whether an insn that looks | |
338 | like a conditional move is actually supported by the hardware. If we | |
339 | guess wrong we lose a bit on optimization, but that's it. */ | |
340 | /* ??? sparc64 supports conditionally moving integers values based on fp | |
341 | comparisons, and vice versa. How do we handle them? */ | |
342 | ||
343 | bool | |
344 | can_conditionally_move_p (machine_mode mode) | |
345 | { | |
346 | return direct_optab_handler (movcc_optab, mode) != CODE_FOR_nothing; | |
347 | } | |
348 | ||
2fda2c81 | 349 | /* If a target doesn't implement a permute on a vector with multibyte |
350 | elements, we can try to do the same permute on byte elements. | |
351 | If this makes sense for vector mode MODE then return the appropriate | |
352 | byte vector mode. */ | |
353 | ||
354 | opt_machine_mode | |
355 | qimode_for_vec_perm (machine_mode mode) | |
356 | { | |
357 | machine_mode qimode; | |
358 | if (GET_MODE_INNER (mode) != QImode | |
359 | && mode_for_vector (QImode, GET_MODE_SIZE (mode)).exists (&qimode) | |
360 | && VECTOR_MODE_P (qimode)) | |
361 | return qimode; | |
362 | return opt_machine_mode (); | |
363 | } | |
364 | ||
d37760c5 | 365 | /* Return true if selector SEL can be represented in the integer |
366 | equivalent of vector mode MODE. */ | |
367 | ||
368 | bool | |
369 | selector_fits_mode_p (machine_mode mode, const vec_perm_indices &sel) | |
370 | { | |
371 | unsigned HOST_WIDE_INT mask = GET_MODE_MASK (GET_MODE_INNER (mode)); | |
372 | return (mask == HOST_WIDE_INT_M1U | |
373 | || sel.all_in_range_p (0, mask + 1)); | |
374 | } | |
375 | ||
97f7d65e | 376 | /* Return true if VEC_PERM_EXPRs with variable selector operands can be |
377 | expanded using SIMD extensions of the CPU. MODE is the mode of the | |
378 | vectors being permuted. */ | |
947ed59a | 379 | |
380 | bool | |
97f7d65e | 381 | can_vec_perm_var_p (machine_mode mode) |
947ed59a | 382 | { |
947ed59a | 383 | /* If the target doesn't implement a vector mode for the vector type, |
384 | then no operations are supported. */ | |
385 | if (!VECTOR_MODE_P (mode)) | |
386 | return false; | |
387 | ||
947ed59a | 388 | if (direct_optab_handler (vec_perm_optab, mode) != CODE_FOR_nothing) |
389 | return true; | |
390 | ||
391 | /* We allow fallback to a QI vector mode, and adjust the mask. */ | |
97f7d65e | 392 | machine_mode qimode; |
fcd49bc5 | 393 | if (!qimode_for_vec_perm (mode).exists (&qimode) |
ba7efd65 | 394 | || maybe_gt (GET_MODE_NUNITS (qimode), GET_MODE_MASK (QImode) + 1)) |
947ed59a | 395 | return false; |
396 | ||
947ed59a | 397 | if (direct_optab_handler (vec_perm_optab, qimode) == CODE_FOR_nothing) |
398 | return false; | |
399 | ||
400 | /* In order to support the lowering of variable permutations, | |
401 | we need to support shifts and adds. */ | |
97f7d65e | 402 | if (GET_MODE_UNIT_SIZE (mode) > 2 |
403 | && optab_handler (ashl_optab, mode) == CODE_FOR_nothing | |
404 | && optab_handler (vashl_optab, mode) == CODE_FOR_nothing) | |
405 | return false; | |
406 | if (optab_handler (add_optab, qimode) == CODE_FOR_nothing) | |
407 | return false; | |
408 | ||
409 | return true; | |
410 | } | |
411 | ||
412 | /* Return true if the target directly supports VEC_PERM_EXPRs on vectors | |
413 | of mode MODE using the selector SEL. ALLOW_VARIABLE_P is true if it | |
414 | is acceptable to force the selector into a register and use a variable | |
415 | permute (if the target supports that). | |
416 | ||
417 | Note that additional permutations representing whole-vector shifts may | |
418 | also be handled via the vec_shr optab, but only where the second input | |
419 | vector is entirely constant zeroes; this case is not dealt with here. */ | |
420 | ||
421 | bool | |
422 | can_vec_perm_const_p (machine_mode mode, const vec_perm_indices &sel, | |
423 | bool allow_variable_p) | |
424 | { | |
425 | /* If the target doesn't implement a vector mode for the vector type, | |
426 | then no operations are supported. */ | |
427 | if (!VECTOR_MODE_P (mode)) | |
428 | return false; | |
429 | ||
430 | /* It's probably cheaper to test for the variable case first. */ | |
d37760c5 | 431 | if (allow_variable_p && selector_fits_mode_p (mode, sel)) |
97f7d65e | 432 | { |
433 | if (direct_optab_handler (vec_perm_optab, mode) != CODE_FOR_nothing) | |
434 | return true; | |
435 | ||
436 | /* Unlike can_vec_perm_var_p, we don't need to test for optabs | |
437 | related computing the QImode selector, since that happens at | |
438 | compile time. */ | |
439 | machine_mode qimode; | |
d37760c5 | 440 | if (qimode_for_vec_perm (mode).exists (&qimode)) |
441 | { | |
442 | vec_perm_indices qimode_indices; | |
443 | qimode_indices.new_expanded_vector (sel, GET_MODE_UNIT_SIZE (mode)); | |
444 | if (selector_fits_mode_p (qimode, qimode_indices) | |
445 | && (direct_optab_handler (vec_perm_optab, qimode) | |
446 | != CODE_FOR_nothing)) | |
447 | return true; | |
448 | } | |
97f7d65e | 449 | } |
450 | ||
d37760c5 | 451 | if (targetm.vectorize.vec_perm_const != NULL) |
947ed59a | 452 | { |
d37760c5 | 453 | if (targetm.vectorize.vec_perm_const (mode, NULL_RTX, NULL_RTX, |
454 | NULL_RTX, sel)) | |
97f7d65e | 455 | return true; |
456 | ||
457 | /* ??? For completeness, we ought to check the QImode version of | |
458 | vec_perm_const_optab. But all users of this implicit lowering | |
d37760c5 | 459 | feature implement the variable vec_perm_optab, and the ia64 |
460 | port specifically doesn't want us to lower V2SF operations | |
461 | into integer operations. */ | |
947ed59a | 462 | } |
463 | ||
97f7d65e | 464 | return false; |
947ed59a | 465 | } |
466 | ||
947ed59a | 467 | /* Find a widening optab even if it doesn't widen as much as we want. |
468 | E.g. if from_mode is HImode, and to_mode is DImode, and there is no | |
d2a1b453 | 469 | direct HI->SI insn, then return SI->DI, if that exists. */ |
947ed59a | 470 | |
471 | enum insn_code | |
472 | find_widening_optab_handler_and_mode (optab op, machine_mode to_mode, | |
473 | machine_mode from_mode, | |
947ed59a | 474 | machine_mode *found_mode) |
475 | { | |
13e61f3e | 476 | machine_mode limit_mode = to_mode; |
477 | if (is_a <scalar_int_mode> (from_mode)) | |
478 | { | |
479 | gcc_checking_assert (is_a <scalar_int_mode> (to_mode) | |
480 | && known_lt (GET_MODE_PRECISION (from_mode), | |
481 | GET_MODE_PRECISION (to_mode))); | |
482 | /* The modes after FROM_MODE are all MODE_INT, so the only | |
483 | MODE_PARTIAL_INT mode we consider is FROM_MODE itself. | |
484 | If LIMIT_MODE is MODE_PARTIAL_INT, stop at the containing | |
485 | MODE_INT. */ | |
486 | if (GET_MODE_CLASS (limit_mode) == MODE_PARTIAL_INT) | |
487 | limit_mode = GET_MODE_WIDER_MODE (limit_mode).require (); | |
488 | } | |
489 | else | |
490 | gcc_checking_assert (GET_MODE_CLASS (from_mode) == GET_MODE_CLASS (to_mode) | |
491 | && from_mode < to_mode); | |
492 | FOR_EACH_MODE (from_mode, from_mode, limit_mode) | |
947ed59a | 493 | { |
d2a1b453 | 494 | enum insn_code handler = convert_optab_handler (op, to_mode, from_mode); |
947ed59a | 495 | |
496 | if (handler != CODE_FOR_nothing) | |
497 | { | |
498 | if (found_mode) | |
499 | *found_mode = from_mode; | |
500 | return handler; | |
501 | } | |
502 | } | |
503 | ||
504 | return CODE_FOR_nothing; | |
505 | } | |
506 | ||
507 | /* Return non-zero if a highpart multiply is supported of can be synthisized. | |
508 | For the benefit of expand_mult_highpart, the return value is 1 for direct, | |
509 | 2 for even/odd widening, and 3 for hi/lo widening. */ | |
510 | ||
511 | int | |
512 | can_mult_highpart_p (machine_mode mode, bool uns_p) | |
513 | { | |
514 | optab op; | |
947ed59a | 515 | |
516 | op = uns_p ? umul_highpart_optab : smul_highpart_optab; | |
517 | if (optab_handler (op, mode) != CODE_FOR_nothing) | |
518 | return 1; | |
519 | ||
520 | /* If the mode is an integral vector, synth from widening operations. */ | |
521 | if (GET_MODE_CLASS (mode) != MODE_VECTOR_INT) | |
522 | return 0; | |
523 | ||
ba7efd65 | 524 | poly_int64 nunits = GET_MODE_NUNITS (mode); |
947ed59a | 525 | |
526 | op = uns_p ? vec_widen_umult_even_optab : vec_widen_smult_even_optab; | |
527 | if (optab_handler (op, mode) != CODE_FOR_nothing) | |
528 | { | |
529 | op = uns_p ? vec_widen_umult_odd_optab : vec_widen_smult_odd_optab; | |
530 | if (optab_handler (op, mode) != CODE_FOR_nothing) | |
531 | { | |
c3fa7fe9 | 532 | /* The encoding has 2 interleaved stepped patterns. */ |
533 | vec_perm_builder sel (nunits, 2, 3); | |
ba7efd65 | 534 | for (unsigned int i = 0; i < 6; ++i) |
282dc861 | 535 | sel.quick_push (!BYTES_BIG_ENDIAN |
536 | + (i & ~1) | |
537 | + ((i & 1) ? nunits : 0)); | |
1957c019 | 538 | vec_perm_indices indices (sel, 2, nunits); |
539 | if (can_vec_perm_const_p (mode, indices)) | |
947ed59a | 540 | return 2; |
541 | } | |
542 | } | |
543 | ||
544 | op = uns_p ? vec_widen_umult_hi_optab : vec_widen_smult_hi_optab; | |
545 | if (optab_handler (op, mode) != CODE_FOR_nothing) | |
546 | { | |
547 | op = uns_p ? vec_widen_umult_lo_optab : vec_widen_smult_lo_optab; | |
548 | if (optab_handler (op, mode) != CODE_FOR_nothing) | |
549 | { | |
c3fa7fe9 | 550 | /* The encoding has a single stepped pattern. */ |
551 | vec_perm_builder sel (nunits, 1, 3); | |
ba7efd65 | 552 | for (unsigned int i = 0; i < 3; ++i) |
282dc861 | 553 | sel.quick_push (2 * i + (BYTES_BIG_ENDIAN ? 0 : 1)); |
1957c019 | 554 | vec_perm_indices indices (sel, 2, nunits); |
555 | if (can_vec_perm_const_p (mode, indices)) | |
947ed59a | 556 | return 3; |
557 | } | |
558 | } | |
559 | ||
560 | return 0; | |
561 | } | |
562 | ||
563 | /* Return true if target supports vector masked load/store for mode. */ | |
564 | ||
565 | bool | |
f636f094 | 566 | can_vec_mask_load_store_p (machine_mode mode, |
567 | machine_mode mask_mode, | |
568 | bool is_load) | |
947ed59a | 569 | { |
570 | optab op = is_load ? maskload_optab : maskstore_optab; | |
571 | machine_mode vmode; | |
947ed59a | 572 | |
573 | /* If mode is vector mode, check it directly. */ | |
574 | if (VECTOR_MODE_P (mode)) | |
f636f094 | 575 | return convert_optab_handler (op, mode, mask_mode) != CODE_FOR_nothing; |
947ed59a | 576 | |
577 | /* Otherwise, return true if there is some vector mode with | |
578 | the mask load/store supported. */ | |
579 | ||
580 | /* See if there is any chance the mask load or store might be | |
581 | vectorized. If not, punt. */ | |
4c1a1be2 | 582 | scalar_mode smode; |
583 | if (!is_a <scalar_mode> (mode, &smode)) | |
584 | return false; | |
585 | ||
586 | vmode = targetm.vectorize.preferred_simd_mode (smode); | |
947ed59a | 587 | if (!VECTOR_MODE_P (vmode)) |
588 | return false; | |
589 | ||
38d5f204 | 590 | if ((targetm.vectorize.get_mask_mode |
591 | (GET_MODE_NUNITS (vmode), GET_MODE_SIZE (vmode)).exists (&mask_mode)) | |
592 | && convert_optab_handler (op, vmode, mask_mode) != CODE_FOR_nothing) | |
947ed59a | 593 | return true; |
594 | ||
3106770a | 595 | auto_vector_sizes vector_sizes; |
596 | targetm.vectorize.autovectorize_vector_sizes (&vector_sizes); | |
597 | for (unsigned int i = 0; i < vector_sizes.length (); ++i) | |
947ed59a | 598 | { |
3106770a | 599 | poly_uint64 cur = vector_sizes[i]; |
600 | poly_uint64 nunits; | |
601 | if (!multiple_p (cur, GET_MODE_SIZE (smode), &nunits)) | |
947ed59a | 602 | continue; |
ab53cba7 | 603 | if (mode_for_vector (smode, nunits).exists (&vmode) |
38d5f204 | 604 | && VECTOR_MODE_P (vmode) |
605 | && targetm.vectorize.get_mask_mode (nunits, cur).exists (&mask_mode) | |
606 | && convert_optab_handler (op, vmode, mask_mode) != CODE_FOR_nothing) | |
607 | return true; | |
947ed59a | 608 | } |
609 | return false; | |
610 | } | |
611 | ||
612 | /* Return true if there is a compare_and_swap pattern. */ | |
613 | ||
614 | bool | |
615 | can_compare_and_swap_p (machine_mode mode, bool allow_libcall) | |
616 | { | |
617 | enum insn_code icode; | |
618 | ||
619 | /* Check for __atomic_compare_and_swap. */ | |
620 | icode = direct_optab_handler (atomic_compare_and_swap_optab, mode); | |
621 | if (icode != CODE_FOR_nothing) | |
622 | return true; | |
623 | ||
624 | /* Check for __sync_compare_and_swap. */ | |
625 | icode = optab_handler (sync_compare_and_swap_optab, mode); | |
626 | if (icode != CODE_FOR_nothing) | |
627 | return true; | |
628 | if (allow_libcall && optab_libfunc (sync_compare_and_swap_optab, mode)) | |
629 | return true; | |
630 | ||
631 | /* No inline compare and swap. */ | |
632 | return false; | |
633 | } | |
634 | ||
635 | /* Return true if an atomic exchange can be performed. */ | |
636 | ||
637 | bool | |
638 | can_atomic_exchange_p (machine_mode mode, bool allow_libcall) | |
639 | { | |
640 | enum insn_code icode; | |
641 | ||
642 | /* Check for __atomic_exchange. */ | |
643 | icode = direct_optab_handler (atomic_exchange_optab, mode); | |
644 | if (icode != CODE_FOR_nothing) | |
645 | return true; | |
646 | ||
647 | /* Don't check __sync_test_and_set, as on some platforms that | |
648 | has reduced functionality. Targets that really do support | |
649 | a proper exchange should simply be updated to the __atomics. */ | |
650 | ||
651 | return can_compare_and_swap_p (mode, allow_libcall); | |
652 | } | |
653 | ||
d5f5fa27 | 654 | /* Return true if an atomic load can be performed without falling back to |
655 | a compare-and-swap. */ | |
656 | ||
657 | bool | |
658 | can_atomic_load_p (machine_mode mode) | |
659 | { | |
660 | enum insn_code icode; | |
661 | ||
662 | /* Does the target supports the load directly? */ | |
663 | icode = direct_optab_handler (atomic_load_optab, mode); | |
664 | if (icode != CODE_FOR_nothing) | |
665 | return true; | |
666 | ||
667 | /* If the size of the object is greater than word size on this target, | |
668 | then we assume that a load will not be atomic. Also see | |
669 | expand_atomic_load. */ | |
3ce67adc | 670 | return known_le (GET_MODE_PRECISION (mode), BITS_PER_WORD); |
d5f5fa27 | 671 | } |
672 | ||
947ed59a | 673 | /* Determine whether "1 << x" is relatively cheap in word_mode. */ |
674 | ||
675 | bool | |
676 | lshift_cheap_p (bool speed_p) | |
677 | { | |
678 | /* FIXME: This should be made target dependent via this "this_target" | |
679 | mechanism, similar to e.g. can_copy_init_p in gcse.c. */ | |
680 | static bool init[2] = { false, false }; | |
681 | static bool cheap[2] = { true, true }; | |
682 | ||
683 | /* If the targer has no lshift in word_mode, the operation will most | |
684 | probably not be cheap. ??? Does GCC even work for such targets? */ | |
685 | if (optab_handler (ashl_optab, word_mode) == CODE_FOR_nothing) | |
686 | return false; | |
687 | ||
688 | if (!init[speed_p]) | |
689 | { | |
690 | rtx reg = gen_raw_REG (word_mode, 10000); | |
691 | int cost = set_src_cost (gen_rtx_ASHIFT (word_mode, const1_rtx, reg), | |
692 | word_mode, speed_p); | |
693 | cheap[speed_p] = cost < COSTS_N_INSNS (3); | |
694 | init[speed_p] = true; | |
695 | } | |
696 | ||
697 | return cheap[speed_p]; | |
698 | } | |
1619606c | 699 | |
700 | /* Return true if optab OP supports at least one mode. */ | |
701 | ||
702 | static bool | |
703 | supports_at_least_one_mode_p (optab op) | |
704 | { | |
705 | for (int i = 0; i < NUM_MACHINE_MODES; ++i) | |
706 | if (direct_optab_handler (op, (machine_mode) i) != CODE_FOR_nothing) | |
707 | return true; | |
708 | ||
709 | return false; | |
710 | } | |
711 | ||
712 | /* Return true if vec_gather_load is available for at least one vector | |
713 | mode. */ | |
714 | ||
715 | bool | |
716 | supports_vec_gather_load_p () | |
717 | { | |
718 | if (this_fn_optabs->supports_vec_gather_load_cached) | |
719 | return this_fn_optabs->supports_vec_gather_load; | |
720 | ||
721 | this_fn_optabs->supports_vec_gather_load_cached = true; | |
722 | ||
723 | this_fn_optabs->supports_vec_gather_load | |
724 | = supports_at_least_one_mode_p (gather_load_optab); | |
725 | ||
726 | return this_fn_optabs->supports_vec_gather_load; | |
727 | } | |
0bf8b382 | 728 | |
729 | /* Return true if vec_scatter_store is available for at least one vector | |
730 | mode. */ | |
731 | ||
732 | bool | |
733 | supports_vec_scatter_store_p () | |
734 | { | |
735 | if (this_fn_optabs->supports_vec_scatter_store_cached) | |
736 | return this_fn_optabs->supports_vec_scatter_store; | |
737 | ||
738 | this_fn_optabs->supports_vec_scatter_store_cached = true; | |
739 | ||
740 | this_fn_optabs->supports_vec_scatter_store | |
741 | = supports_at_least_one_mode_p (scatter_store_optab); | |
742 | ||
743 | return this_fn_optabs->supports_vec_scatter_store; | |
744 | } | |
745 |