]>
Commit | Line | Data |
---|---|---|
1 | /* Machine description for AArch64 architecture. | |
2 | Copyright (C) 2009-2014 Free Software Foundation, Inc. | |
3 | Contributed by ARM Ltd. | |
4 | ||
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it | |
8 | under the terms of the GNU General Public License as published by | |
9 | the Free Software Foundation; either version 3, or (at your option) | |
10 | any later version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but | |
13 | WITHOUT ANY WARRANTY; without even the implied warranty of | |
14 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
15 | General Public License for more details. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
18 | along with GCC; see the file COPYING3. If not see | |
19 | <http://www.gnu.org/licenses/>. */ | |
20 | ||
21 | #include "config.h" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
24 | #include "tm.h" | |
25 | #include "insn-codes.h" | |
26 | #include "rtl.h" | |
27 | #include "insn-attr.h" | |
28 | #include "tree.h" | |
29 | #include "stringpool.h" | |
30 | #include "stor-layout.h" | |
31 | #include "calls.h" | |
32 | #include "varasm.h" | |
33 | #include "regs.h" | |
34 | #include "df.h" | |
35 | #include "hard-reg-set.h" | |
36 | #include "output.h" | |
37 | #include "expr.h" | |
38 | #include "reload.h" | |
39 | #include "toplev.h" | |
40 | #include "target.h" | |
41 | #include "target-def.h" | |
42 | #include "targhooks.h" | |
43 | #include "ggc.h" | |
44 | #include "function.h" | |
45 | #include "tm_p.h" | |
46 | #include "recog.h" | |
47 | #include "langhooks.h" | |
48 | #include "diagnostic-core.h" | |
49 | #include "pointer-set.h" | |
50 | #include "hash-table.h" | |
51 | #include "vec.h" | |
52 | #include "basic-block.h" | |
53 | #include "tree-ssa-alias.h" | |
54 | #include "internal-fn.h" | |
55 | #include "gimple-fold.h" | |
56 | #include "tree-eh.h" | |
57 | #include "gimple-expr.h" | |
58 | #include "is-a.h" | |
59 | #include "gimple.h" | |
60 | #include "gimplify.h" | |
61 | #include "optabs.h" | |
62 | #include "dwarf2.h" | |
63 | #include "cfgloop.h" | |
64 | #include "tree-vectorizer.h" | |
65 | #include "config/arm/aarch-cost-tables.h" | |
66 | #include "dumpfile.h" | |
67 | ||
68 | /* Defined for convenience. */ | |
69 | #define POINTER_BYTES (POINTER_SIZE / BITS_PER_UNIT) | |
70 | ||
71 | /* Classifies an address. | |
72 | ||
73 | ADDRESS_REG_IMM | |
74 | A simple base register plus immediate offset. | |
75 | ||
76 | ADDRESS_REG_WB | |
77 | A base register indexed by immediate offset with writeback. | |
78 | ||
79 | ADDRESS_REG_REG | |
80 | A base register indexed by (optionally scaled) register. | |
81 | ||
82 | ADDRESS_REG_UXTW | |
83 | A base register indexed by (optionally scaled) zero-extended register. | |
84 | ||
85 | ADDRESS_REG_SXTW | |
86 | A base register indexed by (optionally scaled) sign-extended register. | |
87 | ||
88 | ADDRESS_LO_SUM | |
89 | A LO_SUM rtx with a base register and "LO12" symbol relocation. | |
90 | ||
91 | ADDRESS_SYMBOLIC: | |
92 | A constant symbolic address, in pc-relative literal pool. */ | |
93 | ||
94 | enum aarch64_address_type { | |
95 | ADDRESS_REG_IMM, | |
96 | ADDRESS_REG_WB, | |
97 | ADDRESS_REG_REG, | |
98 | ADDRESS_REG_UXTW, | |
99 | ADDRESS_REG_SXTW, | |
100 | ADDRESS_LO_SUM, | |
101 | ADDRESS_SYMBOLIC | |
102 | }; | |
103 | ||
104 | struct aarch64_address_info { | |
105 | enum aarch64_address_type type; | |
106 | rtx base; | |
107 | rtx offset; | |
108 | int shift; | |
109 | enum aarch64_symbol_type symbol_type; | |
110 | }; | |
111 | ||
112 | struct simd_immediate_info | |
113 | { | |
114 | rtx value; | |
115 | int shift; | |
116 | int element_width; | |
117 | bool mvn; | |
118 | bool msl; | |
119 | }; | |
120 | ||
121 | /* The current code model. */ | |
122 | enum aarch64_code_model aarch64_cmodel; | |
123 | ||
124 | #ifdef HAVE_AS_TLS | |
125 | #undef TARGET_HAVE_TLS | |
126 | #define TARGET_HAVE_TLS 1 | |
127 | #endif | |
128 | ||
129 | static bool aarch64_lra_p (void); | |
130 | static bool aarch64_composite_type_p (const_tree, enum machine_mode); | |
131 | static bool aarch64_vfp_is_call_or_return_candidate (enum machine_mode, | |
132 | const_tree, | |
133 | enum machine_mode *, int *, | |
134 | bool *); | |
135 | static void aarch64_elf_asm_constructor (rtx, int) ATTRIBUTE_UNUSED; | |
136 | static void aarch64_elf_asm_destructor (rtx, int) ATTRIBUTE_UNUSED; | |
137 | static void aarch64_override_options_after_change (void); | |
138 | static bool aarch64_vector_mode_supported_p (enum machine_mode); | |
139 | static unsigned bit_count (unsigned HOST_WIDE_INT); | |
140 | static bool aarch64_const_vec_all_same_int_p (rtx, | |
141 | HOST_WIDE_INT, HOST_WIDE_INT); | |
142 | ||
143 | static bool aarch64_vectorize_vec_perm_const_ok (enum machine_mode vmode, | |
144 | const unsigned char *sel); | |
145 | static int aarch64_address_cost (rtx, enum machine_mode, addr_space_t, bool); | |
146 | ||
147 | /* The processor for which instructions should be scheduled. */ | |
148 | enum aarch64_processor aarch64_tune = cortexa53; | |
149 | ||
150 | /* The current tuning set. */ | |
151 | const struct tune_params *aarch64_tune_params; | |
152 | ||
153 | /* Mask to specify which instructions we are allowed to generate. */ | |
154 | unsigned long aarch64_isa_flags = 0; | |
155 | ||
156 | /* Mask to specify which instruction scheduling options should be used. */ | |
157 | unsigned long aarch64_tune_flags = 0; | |
158 | ||
159 | /* Tuning parameters. */ | |
160 | ||
161 | #if HAVE_DESIGNATED_INITIALIZERS | |
162 | #define NAMED_PARAM(NAME, VAL) .NAME = (VAL) | |
163 | #else | |
164 | #define NAMED_PARAM(NAME, VAL) (VAL) | |
165 | #endif | |
166 | ||
167 | #if HAVE_DESIGNATED_INITIALIZERS && GCC_VERSION >= 2007 | |
168 | __extension__ | |
169 | #endif | |
170 | ||
171 | #if HAVE_DESIGNATED_INITIALIZERS && GCC_VERSION >= 2007 | |
172 | __extension__ | |
173 | #endif | |
174 | static const struct cpu_addrcost_table generic_addrcost_table = | |
175 | { | |
176 | #if HAVE_DESIGNATED_INITIALIZERS | |
177 | .addr_scale_costs = | |
178 | #endif | |
179 | { | |
180 | NAMED_PARAM (qi, 0), | |
181 | NAMED_PARAM (hi, 0), | |
182 | NAMED_PARAM (si, 0), | |
183 | NAMED_PARAM (ti, 0), | |
184 | }, | |
185 | NAMED_PARAM (pre_modify, 0), | |
186 | NAMED_PARAM (post_modify, 0), | |
187 | NAMED_PARAM (register_offset, 0), | |
188 | NAMED_PARAM (register_extend, 0), | |
189 | NAMED_PARAM (imm_offset, 0) | |
190 | }; | |
191 | ||
192 | #if HAVE_DESIGNATED_INITIALIZERS && GCC_VERSION >= 2007 | |
193 | __extension__ | |
194 | #endif | |
195 | static const struct cpu_addrcost_table cortexa57_addrcost_table = | |
196 | { | |
197 | #if HAVE_DESIGNATED_INITIALIZERS | |
198 | .addr_scale_costs = | |
199 | #endif | |
200 | { | |
201 | NAMED_PARAM (qi, 0), | |
202 | NAMED_PARAM (hi, 1), | |
203 | NAMED_PARAM (si, 0), | |
204 | NAMED_PARAM (ti, 1), | |
205 | }, | |
206 | NAMED_PARAM (pre_modify, 0), | |
207 | NAMED_PARAM (post_modify, 0), | |
208 | NAMED_PARAM (register_offset, 0), | |
209 | NAMED_PARAM (register_extend, 0), | |
210 | NAMED_PARAM (imm_offset, 0), | |
211 | }; | |
212 | ||
213 | #if HAVE_DESIGNATED_INITIALIZERS && GCC_VERSION >= 2007 | |
214 | __extension__ | |
215 | #endif | |
216 | static const struct cpu_regmove_cost generic_regmove_cost = | |
217 | { | |
218 | NAMED_PARAM (GP2GP, 1), | |
219 | NAMED_PARAM (GP2FP, 2), | |
220 | NAMED_PARAM (FP2GP, 2), | |
221 | /* We currently do not provide direct support for TFmode Q->Q move. | |
222 | Therefore we need to raise the cost above 2 in order to have | |
223 | reload handle the situation. */ | |
224 | NAMED_PARAM (FP2FP, 4) | |
225 | }; | |
226 | ||
227 | /* Generic costs for vector insn classes. */ | |
228 | #if HAVE_DESIGNATED_INITIALIZERS && GCC_VERSION >= 2007 | |
229 | __extension__ | |
230 | #endif | |
231 | static const struct cpu_vector_cost generic_vector_cost = | |
232 | { | |
233 | NAMED_PARAM (scalar_stmt_cost, 1), | |
234 | NAMED_PARAM (scalar_load_cost, 1), | |
235 | NAMED_PARAM (scalar_store_cost, 1), | |
236 | NAMED_PARAM (vec_stmt_cost, 1), | |
237 | NAMED_PARAM (vec_to_scalar_cost, 1), | |
238 | NAMED_PARAM (scalar_to_vec_cost, 1), | |
239 | NAMED_PARAM (vec_align_load_cost, 1), | |
240 | NAMED_PARAM (vec_unalign_load_cost, 1), | |
241 | NAMED_PARAM (vec_unalign_store_cost, 1), | |
242 | NAMED_PARAM (vec_store_cost, 1), | |
243 | NAMED_PARAM (cond_taken_branch_cost, 3), | |
244 | NAMED_PARAM (cond_not_taken_branch_cost, 1) | |
245 | }; | |
246 | ||
247 | /* Generic costs for vector insn classes. */ | |
248 | #if HAVE_DESIGNATED_INITIALIZERS && GCC_VERSION >= 2007 | |
249 | __extension__ | |
250 | #endif | |
251 | static const struct cpu_vector_cost cortexa57_vector_cost = | |
252 | { | |
253 | NAMED_PARAM (scalar_stmt_cost, 1), | |
254 | NAMED_PARAM (scalar_load_cost, 4), | |
255 | NAMED_PARAM (scalar_store_cost, 1), | |
256 | NAMED_PARAM (vec_stmt_cost, 3), | |
257 | NAMED_PARAM (vec_to_scalar_cost, 8), | |
258 | NAMED_PARAM (scalar_to_vec_cost, 8), | |
259 | NAMED_PARAM (vec_align_load_cost, 5), | |
260 | NAMED_PARAM (vec_unalign_load_cost, 5), | |
261 | NAMED_PARAM (vec_unalign_store_cost, 1), | |
262 | NAMED_PARAM (vec_store_cost, 1), | |
263 | NAMED_PARAM (cond_taken_branch_cost, 1), | |
264 | NAMED_PARAM (cond_not_taken_branch_cost, 1) | |
265 | }; | |
266 | ||
267 | #if HAVE_DESIGNATED_INITIALIZERS && GCC_VERSION >= 2007 | |
268 | __extension__ | |
269 | #endif | |
270 | static const struct tune_params generic_tunings = | |
271 | { | |
272 | &cortexa57_extra_costs, | |
273 | &generic_addrcost_table, | |
274 | &generic_regmove_cost, | |
275 | &generic_vector_cost, | |
276 | NAMED_PARAM (memmov_cost, 4), | |
277 | NAMED_PARAM (issue_rate, 2) | |
278 | }; | |
279 | ||
280 | static const struct tune_params cortexa53_tunings = | |
281 | { | |
282 | &cortexa53_extra_costs, | |
283 | &generic_addrcost_table, | |
284 | &generic_regmove_cost, | |
285 | &generic_vector_cost, | |
286 | NAMED_PARAM (memmov_cost, 4), | |
287 | NAMED_PARAM (issue_rate, 2) | |
288 | }; | |
289 | ||
290 | static const struct tune_params cortexa57_tunings = | |
291 | { | |
292 | &cortexa57_extra_costs, | |
293 | &cortexa57_addrcost_table, | |
294 | &generic_regmove_cost, | |
295 | &cortexa57_vector_cost, | |
296 | NAMED_PARAM (memmov_cost, 4), | |
297 | NAMED_PARAM (issue_rate, 3) | |
298 | }; | |
299 | ||
300 | /* A processor implementing AArch64. */ | |
301 | struct processor | |
302 | { | |
303 | const char *const name; | |
304 | enum aarch64_processor core; | |
305 | const char *arch; | |
306 | const unsigned long flags; | |
307 | const struct tune_params *const tune; | |
308 | }; | |
309 | ||
310 | /* Processor cores implementing AArch64. */ | |
311 | static const struct processor all_cores[] = | |
312 | { | |
313 | #define AARCH64_CORE(NAME, X, IDENT, ARCH, FLAGS, COSTS) \ | |
314 | {NAME, IDENT, #ARCH, FLAGS | AARCH64_FL_FOR_ARCH##ARCH, &COSTS##_tunings}, | |
315 | #include "aarch64-cores.def" | |
316 | #undef AARCH64_CORE | |
317 | {"generic", cortexa53, "8", AARCH64_FL_FPSIMD | AARCH64_FL_FOR_ARCH8, &generic_tunings}, | |
318 | {NULL, aarch64_none, NULL, 0, NULL} | |
319 | }; | |
320 | ||
321 | /* Architectures implementing AArch64. */ | |
322 | static const struct processor all_architectures[] = | |
323 | { | |
324 | #define AARCH64_ARCH(NAME, CORE, ARCH, FLAGS) \ | |
325 | {NAME, CORE, #ARCH, FLAGS, NULL}, | |
326 | #include "aarch64-arches.def" | |
327 | #undef AARCH64_ARCH | |
328 | {NULL, aarch64_none, NULL, 0, NULL} | |
329 | }; | |
330 | ||
331 | /* Target specification. These are populated as commandline arguments | |
332 | are processed, or NULL if not specified. */ | |
333 | static const struct processor *selected_arch; | |
334 | static const struct processor *selected_cpu; | |
335 | static const struct processor *selected_tune; | |
336 | ||
337 | #define AARCH64_CPU_DEFAULT_FLAGS ((selected_cpu) ? selected_cpu->flags : 0) | |
338 | ||
339 | /* An ISA extension in the co-processor and main instruction set space. */ | |
340 | struct aarch64_option_extension | |
341 | { | |
342 | const char *const name; | |
343 | const unsigned long flags_on; | |
344 | const unsigned long flags_off; | |
345 | }; | |
346 | ||
347 | /* ISA extensions in AArch64. */ | |
348 | static const struct aarch64_option_extension all_extensions[] = | |
349 | { | |
350 | #define AARCH64_OPT_EXTENSION(NAME, FLAGS_ON, FLAGS_OFF) \ | |
351 | {NAME, FLAGS_ON, FLAGS_OFF}, | |
352 | #include "aarch64-option-extensions.def" | |
353 | #undef AARCH64_OPT_EXTENSION | |
354 | {NULL, 0, 0} | |
355 | }; | |
356 | ||
357 | /* Used to track the size of an address when generating a pre/post | |
358 | increment address. */ | |
359 | static enum machine_mode aarch64_memory_reference_mode; | |
360 | ||
361 | /* Used to force GTY into this file. */ | |
362 | static GTY(()) int gty_dummy; | |
363 | ||
364 | /* A table of valid AArch64 "bitmask immediate" values for | |
365 | logical instructions. */ | |
366 | ||
367 | #define AARCH64_NUM_BITMASKS 5334 | |
368 | static unsigned HOST_WIDE_INT aarch64_bitmasks[AARCH64_NUM_BITMASKS]; | |
369 | ||
370 | typedef enum aarch64_cond_code | |
371 | { | |
372 | AARCH64_EQ = 0, AARCH64_NE, AARCH64_CS, AARCH64_CC, AARCH64_MI, AARCH64_PL, | |
373 | AARCH64_VS, AARCH64_VC, AARCH64_HI, AARCH64_LS, AARCH64_GE, AARCH64_LT, | |
374 | AARCH64_GT, AARCH64_LE, AARCH64_AL, AARCH64_NV | |
375 | } | |
376 | aarch64_cc; | |
377 | ||
378 | #define AARCH64_INVERSE_CONDITION_CODE(X) ((aarch64_cc) (((int) X) ^ 1)) | |
379 | ||
380 | /* The condition codes of the processor, and the inverse function. */ | |
381 | static const char * const aarch64_condition_codes[] = | |
382 | { | |
383 | "eq", "ne", "cs", "cc", "mi", "pl", "vs", "vc", | |
384 | "hi", "ls", "ge", "lt", "gt", "le", "al", "nv" | |
385 | }; | |
386 | ||
387 | /* Provide a mapping from gcc register numbers to dwarf register numbers. */ | |
388 | unsigned | |
389 | aarch64_dbx_register_number (unsigned regno) | |
390 | { | |
391 | if (GP_REGNUM_P (regno)) | |
392 | return AARCH64_DWARF_R0 + regno - R0_REGNUM; | |
393 | else if (regno == SP_REGNUM) | |
394 | return AARCH64_DWARF_SP; | |
395 | else if (FP_REGNUM_P (regno)) | |
396 | return AARCH64_DWARF_V0 + regno - V0_REGNUM; | |
397 | ||
398 | /* Return values >= DWARF_FRAME_REGISTERS indicate that there is no | |
399 | equivalent DWARF register. */ | |
400 | return DWARF_FRAME_REGISTERS; | |
401 | } | |
402 | ||
403 | /* Return TRUE if MODE is any of the large INT modes. */ | |
404 | static bool | |
405 | aarch64_vect_struct_mode_p (enum machine_mode mode) | |
406 | { | |
407 | return mode == OImode || mode == CImode || mode == XImode; | |
408 | } | |
409 | ||
410 | /* Return TRUE if MODE is any of the vector modes. */ | |
411 | static bool | |
412 | aarch64_vector_mode_p (enum machine_mode mode) | |
413 | { | |
414 | return aarch64_vector_mode_supported_p (mode) | |
415 | || aarch64_vect_struct_mode_p (mode); | |
416 | } | |
417 | ||
418 | /* Implement target hook TARGET_ARRAY_MODE_SUPPORTED_P. */ | |
419 | static bool | |
420 | aarch64_array_mode_supported_p (enum machine_mode mode, | |
421 | unsigned HOST_WIDE_INT nelems) | |
422 | { | |
423 | if (TARGET_SIMD | |
424 | && AARCH64_VALID_SIMD_QREG_MODE (mode) | |
425 | && (nelems >= 2 && nelems <= 4)) | |
426 | return true; | |
427 | ||
428 | return false; | |
429 | } | |
430 | ||
431 | /* Implement HARD_REGNO_NREGS. */ | |
432 | ||
433 | int | |
434 | aarch64_hard_regno_nregs (unsigned regno, enum machine_mode mode) | |
435 | { | |
436 | switch (aarch64_regno_regclass (regno)) | |
437 | { | |
438 | case FP_REGS: | |
439 | case FP_LO_REGS: | |
440 | return (GET_MODE_SIZE (mode) + UNITS_PER_VREG - 1) / UNITS_PER_VREG; | |
441 | default: | |
442 | return (GET_MODE_SIZE (mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
443 | } | |
444 | gcc_unreachable (); | |
445 | } | |
446 | ||
447 | /* Implement HARD_REGNO_MODE_OK. */ | |
448 | ||
449 | int | |
450 | aarch64_hard_regno_mode_ok (unsigned regno, enum machine_mode mode) | |
451 | { | |
452 | if (GET_MODE_CLASS (mode) == MODE_CC) | |
453 | return regno == CC_REGNUM; | |
454 | ||
455 | if (regno == SP_REGNUM) | |
456 | /* The purpose of comparing with ptr_mode is to support the | |
457 | global register variable associated with the stack pointer | |
458 | register via the syntax of asm ("wsp") in ILP32. */ | |
459 | return mode == Pmode || mode == ptr_mode; | |
460 | ||
461 | if (regno == FRAME_POINTER_REGNUM || regno == ARG_POINTER_REGNUM) | |
462 | return mode == Pmode; | |
463 | ||
464 | if (GP_REGNUM_P (regno) && ! aarch64_vect_struct_mode_p (mode)) | |
465 | return 1; | |
466 | ||
467 | if (FP_REGNUM_P (regno)) | |
468 | { | |
469 | if (aarch64_vect_struct_mode_p (mode)) | |
470 | return | |
471 | (regno + aarch64_hard_regno_nregs (regno, mode) - 1) <= V31_REGNUM; | |
472 | else | |
473 | return 1; | |
474 | } | |
475 | ||
476 | return 0; | |
477 | } | |
478 | ||
479 | /* Implement HARD_REGNO_CALLER_SAVE_MODE. */ | |
480 | enum machine_mode | |
481 | aarch64_hard_regno_caller_save_mode (unsigned regno, unsigned nregs, | |
482 | enum machine_mode mode) | |
483 | { | |
484 | /* Handle modes that fit within single registers. */ | |
485 | if (nregs == 1 && GET_MODE_SIZE (mode) <= 16) | |
486 | { | |
487 | if (GET_MODE_SIZE (mode) >= 4) | |
488 | return mode; | |
489 | else | |
490 | return SImode; | |
491 | } | |
492 | /* Fall back to generic for multi-reg and very large modes. */ | |
493 | else | |
494 | return choose_hard_reg_mode (regno, nregs, false); | |
495 | } | |
496 | ||
497 | /* Return true if calls to DECL should be treated as | |
498 | long-calls (ie called via a register). */ | |
499 | static bool | |
500 | aarch64_decl_is_long_call_p (const_tree decl ATTRIBUTE_UNUSED) | |
501 | { | |
502 | return false; | |
503 | } | |
504 | ||
505 | /* Return true if calls to symbol-ref SYM should be treated as | |
506 | long-calls (ie called via a register). */ | |
507 | bool | |
508 | aarch64_is_long_call_p (rtx sym) | |
509 | { | |
510 | return aarch64_decl_is_long_call_p (SYMBOL_REF_DECL (sym)); | |
511 | } | |
512 | ||
513 | /* Return true if the offsets to a zero/sign-extract operation | |
514 | represent an expression that matches an extend operation. The | |
515 | operands represent the paramters from | |
516 | ||
517 | (extract:MODE (mult (reg) (MULT_IMM)) (EXTRACT_IMM) (const_int 0)). */ | |
518 | bool | |
519 | aarch64_is_extend_from_extract (enum machine_mode mode, rtx mult_imm, | |
520 | rtx extract_imm) | |
521 | { | |
522 | HOST_WIDE_INT mult_val, extract_val; | |
523 | ||
524 | if (! CONST_INT_P (mult_imm) || ! CONST_INT_P (extract_imm)) | |
525 | return false; | |
526 | ||
527 | mult_val = INTVAL (mult_imm); | |
528 | extract_val = INTVAL (extract_imm); | |
529 | ||
530 | if (extract_val > 8 | |
531 | && extract_val < GET_MODE_BITSIZE (mode) | |
532 | && exact_log2 (extract_val & ~7) > 0 | |
533 | && (extract_val & 7) <= 4 | |
534 | && mult_val == (1 << (extract_val & 7))) | |
535 | return true; | |
536 | ||
537 | return false; | |
538 | } | |
539 | ||
540 | /* Emit an insn that's a simple single-set. Both the operands must be | |
541 | known to be valid. */ | |
542 | inline static rtx | |
543 | emit_set_insn (rtx x, rtx y) | |
544 | { | |
545 | return emit_insn (gen_rtx_SET (VOIDmode, x, y)); | |
546 | } | |
547 | ||
548 | /* X and Y are two things to compare using CODE. Emit the compare insn and | |
549 | return the rtx for register 0 in the proper mode. */ | |
550 | rtx | |
551 | aarch64_gen_compare_reg (RTX_CODE code, rtx x, rtx y) | |
552 | { | |
553 | enum machine_mode mode = SELECT_CC_MODE (code, x, y); | |
554 | rtx cc_reg = gen_rtx_REG (mode, CC_REGNUM); | |
555 | ||
556 | emit_set_insn (cc_reg, gen_rtx_COMPARE (mode, x, y)); | |
557 | return cc_reg; | |
558 | } | |
559 | ||
560 | /* Build the SYMBOL_REF for __tls_get_addr. */ | |
561 | ||
562 | static GTY(()) rtx tls_get_addr_libfunc; | |
563 | ||
564 | rtx | |
565 | aarch64_tls_get_addr (void) | |
566 | { | |
567 | if (!tls_get_addr_libfunc) | |
568 | tls_get_addr_libfunc = init_one_libfunc ("__tls_get_addr"); | |
569 | return tls_get_addr_libfunc; | |
570 | } | |
571 | ||
572 | /* Return the TLS model to use for ADDR. */ | |
573 | ||
574 | static enum tls_model | |
575 | tls_symbolic_operand_type (rtx addr) | |
576 | { | |
577 | enum tls_model tls_kind = TLS_MODEL_NONE; | |
578 | rtx sym, addend; | |
579 | ||
580 | if (GET_CODE (addr) == CONST) | |
581 | { | |
582 | split_const (addr, &sym, &addend); | |
583 | if (GET_CODE (sym) == SYMBOL_REF) | |
584 | tls_kind = SYMBOL_REF_TLS_MODEL (sym); | |
585 | } | |
586 | else if (GET_CODE (addr) == SYMBOL_REF) | |
587 | tls_kind = SYMBOL_REF_TLS_MODEL (addr); | |
588 | ||
589 | return tls_kind; | |
590 | } | |
591 | ||
592 | /* We'll allow lo_sum's in addresses in our legitimate addresses | |
593 | so that combine would take care of combining addresses where | |
594 | necessary, but for generation purposes, we'll generate the address | |
595 | as : | |
596 | RTL Absolute | |
597 | tmp = hi (symbol_ref); adrp x1, foo | |
598 | dest = lo_sum (tmp, symbol_ref); add dest, x1, :lo_12:foo | |
599 | nop | |
600 | ||
601 | PIC TLS | |
602 | adrp x1, :got:foo adrp tmp, :tlsgd:foo | |
603 | ldr x1, [:got_lo12:foo] add dest, tmp, :tlsgd_lo12:foo | |
604 | bl __tls_get_addr | |
605 | nop | |
606 | ||
607 | Load TLS symbol, depending on TLS mechanism and TLS access model. | |
608 | ||
609 | Global Dynamic - Traditional TLS: | |
610 | adrp tmp, :tlsgd:imm | |
611 | add dest, tmp, #:tlsgd_lo12:imm | |
612 | bl __tls_get_addr | |
613 | ||
614 | Global Dynamic - TLS Descriptors: | |
615 | adrp dest, :tlsdesc:imm | |
616 | ldr tmp, [dest, #:tlsdesc_lo12:imm] | |
617 | add dest, dest, #:tlsdesc_lo12:imm | |
618 | blr tmp | |
619 | mrs tp, tpidr_el0 | |
620 | add dest, dest, tp | |
621 | ||
622 | Initial Exec: | |
623 | mrs tp, tpidr_el0 | |
624 | adrp tmp, :gottprel:imm | |
625 | ldr dest, [tmp, #:gottprel_lo12:imm] | |
626 | add dest, dest, tp | |
627 | ||
628 | Local Exec: | |
629 | mrs tp, tpidr_el0 | |
630 | add t0, tp, #:tprel_hi12:imm | |
631 | add t0, #:tprel_lo12_nc:imm | |
632 | */ | |
633 | ||
634 | static void | |
635 | aarch64_load_symref_appropriately (rtx dest, rtx imm, | |
636 | enum aarch64_symbol_type type) | |
637 | { | |
638 | switch (type) | |
639 | { | |
640 | case SYMBOL_SMALL_ABSOLUTE: | |
641 | { | |
642 | /* In ILP32, the mode of dest can be either SImode or DImode. */ | |
643 | rtx tmp_reg = dest; | |
644 | enum machine_mode mode = GET_MODE (dest); | |
645 | ||
646 | gcc_assert (mode == Pmode || mode == ptr_mode); | |
647 | ||
648 | if (can_create_pseudo_p ()) | |
649 | tmp_reg = gen_reg_rtx (mode); | |
650 | ||
651 | emit_move_insn (tmp_reg, gen_rtx_HIGH (mode, imm)); | |
652 | emit_insn (gen_add_losym (dest, tmp_reg, imm)); | |
653 | return; | |
654 | } | |
655 | ||
656 | case SYMBOL_TINY_ABSOLUTE: | |
657 | emit_insn (gen_rtx_SET (Pmode, dest, imm)); | |
658 | return; | |
659 | ||
660 | case SYMBOL_SMALL_GOT: | |
661 | { | |
662 | /* In ILP32, the mode of dest can be either SImode or DImode, | |
663 | while the got entry is always of SImode size. The mode of | |
664 | dest depends on how dest is used: if dest is assigned to a | |
665 | pointer (e.g. in the memory), it has SImode; it may have | |
666 | DImode if dest is dereferenced to access the memeory. | |
667 | This is why we have to handle three different ldr_got_small | |
668 | patterns here (two patterns for ILP32). */ | |
669 | rtx tmp_reg = dest; | |
670 | enum machine_mode mode = GET_MODE (dest); | |
671 | ||
672 | if (can_create_pseudo_p ()) | |
673 | tmp_reg = gen_reg_rtx (mode); | |
674 | ||
675 | emit_move_insn (tmp_reg, gen_rtx_HIGH (mode, imm)); | |
676 | if (mode == ptr_mode) | |
677 | { | |
678 | if (mode == DImode) | |
679 | emit_insn (gen_ldr_got_small_di (dest, tmp_reg, imm)); | |
680 | else | |
681 | emit_insn (gen_ldr_got_small_si (dest, tmp_reg, imm)); | |
682 | } | |
683 | else | |
684 | { | |
685 | gcc_assert (mode == Pmode); | |
686 | emit_insn (gen_ldr_got_small_sidi (dest, tmp_reg, imm)); | |
687 | } | |
688 | ||
689 | return; | |
690 | } | |
691 | ||
692 | case SYMBOL_SMALL_TLSGD: | |
693 | { | |
694 | rtx insns; | |
695 | rtx result = gen_rtx_REG (Pmode, R0_REGNUM); | |
696 | ||
697 | start_sequence (); | |
698 | emit_call_insn (gen_tlsgd_small (result, imm)); | |
699 | insns = get_insns (); | |
700 | end_sequence (); | |
701 | ||
702 | RTL_CONST_CALL_P (insns) = 1; | |
703 | emit_libcall_block (insns, dest, result, imm); | |
704 | return; | |
705 | } | |
706 | ||
707 | case SYMBOL_SMALL_TLSDESC: | |
708 | { | |
709 | enum machine_mode mode = GET_MODE (dest); | |
710 | rtx x0 = gen_rtx_REG (mode, R0_REGNUM); | |
711 | rtx tp; | |
712 | ||
713 | gcc_assert (mode == Pmode || mode == ptr_mode); | |
714 | ||
715 | /* In ILP32, the got entry is always of SImode size. Unlike | |
716 | small GOT, the dest is fixed at reg 0. */ | |
717 | if (TARGET_ILP32) | |
718 | emit_insn (gen_tlsdesc_small_si (imm)); | |
719 | else | |
720 | emit_insn (gen_tlsdesc_small_di (imm)); | |
721 | tp = aarch64_load_tp (NULL); | |
722 | ||
723 | if (mode != Pmode) | |
724 | tp = gen_lowpart (mode, tp); | |
725 | ||
726 | emit_insn (gen_rtx_SET (mode, dest, gen_rtx_PLUS (mode, tp, x0))); | |
727 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); | |
728 | return; | |
729 | } | |
730 | ||
731 | case SYMBOL_SMALL_GOTTPREL: | |
732 | { | |
733 | /* In ILP32, the mode of dest can be either SImode or DImode, | |
734 | while the got entry is always of SImode size. The mode of | |
735 | dest depends on how dest is used: if dest is assigned to a | |
736 | pointer (e.g. in the memory), it has SImode; it may have | |
737 | DImode if dest is dereferenced to access the memeory. | |
738 | This is why we have to handle three different tlsie_small | |
739 | patterns here (two patterns for ILP32). */ | |
740 | enum machine_mode mode = GET_MODE (dest); | |
741 | rtx tmp_reg = gen_reg_rtx (mode); | |
742 | rtx tp = aarch64_load_tp (NULL); | |
743 | ||
744 | if (mode == ptr_mode) | |
745 | { | |
746 | if (mode == DImode) | |
747 | emit_insn (gen_tlsie_small_di (tmp_reg, imm)); | |
748 | else | |
749 | { | |
750 | emit_insn (gen_tlsie_small_si (tmp_reg, imm)); | |
751 | tp = gen_lowpart (mode, tp); | |
752 | } | |
753 | } | |
754 | else | |
755 | { | |
756 | gcc_assert (mode == Pmode); | |
757 | emit_insn (gen_tlsie_small_sidi (tmp_reg, imm)); | |
758 | } | |
759 | ||
760 | emit_insn (gen_rtx_SET (mode, dest, gen_rtx_PLUS (mode, tp, tmp_reg))); | |
761 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); | |
762 | return; | |
763 | } | |
764 | ||
765 | case SYMBOL_SMALL_TPREL: | |
766 | { | |
767 | rtx tp = aarch64_load_tp (NULL); | |
768 | emit_insn (gen_tlsle_small (dest, tp, imm)); | |
769 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); | |
770 | return; | |
771 | } | |
772 | ||
773 | case SYMBOL_TINY_GOT: | |
774 | emit_insn (gen_ldr_got_tiny (dest, imm)); | |
775 | return; | |
776 | ||
777 | default: | |
778 | gcc_unreachable (); | |
779 | } | |
780 | } | |
781 | ||
782 | /* Emit a move from SRC to DEST. Assume that the move expanders can | |
783 | handle all moves if !can_create_pseudo_p (). The distinction is | |
784 | important because, unlike emit_move_insn, the move expanders know | |
785 | how to force Pmode objects into the constant pool even when the | |
786 | constant pool address is not itself legitimate. */ | |
787 | static rtx | |
788 | aarch64_emit_move (rtx dest, rtx src) | |
789 | { | |
790 | return (can_create_pseudo_p () | |
791 | ? emit_move_insn (dest, src) | |
792 | : emit_move_insn_1 (dest, src)); | |
793 | } | |
794 | ||
795 | /* Split a 128-bit move operation into two 64-bit move operations, | |
796 | taking care to handle partial overlap of register to register | |
797 | copies. Special cases are needed when moving between GP regs and | |
798 | FP regs. SRC can be a register, constant or memory; DST a register | |
799 | or memory. If either operand is memory it must not have any side | |
800 | effects. */ | |
801 | void | |
802 | aarch64_split_128bit_move (rtx dst, rtx src) | |
803 | { | |
804 | rtx dst_lo, dst_hi; | |
805 | rtx src_lo, src_hi; | |
806 | ||
807 | enum machine_mode mode = GET_MODE (dst); | |
808 | ||
809 | gcc_assert (mode == TImode || mode == TFmode); | |
810 | gcc_assert (!(side_effects_p (src) || side_effects_p (dst))); | |
811 | gcc_assert (mode == GET_MODE (src) || GET_MODE (src) == VOIDmode); | |
812 | ||
813 | if (REG_P (dst) && REG_P (src)) | |
814 | { | |
815 | int src_regno = REGNO (src); | |
816 | int dst_regno = REGNO (dst); | |
817 | ||
818 | /* Handle FP <-> GP regs. */ | |
819 | if (FP_REGNUM_P (dst_regno) && GP_REGNUM_P (src_regno)) | |
820 | { | |
821 | src_lo = gen_lowpart (word_mode, src); | |
822 | src_hi = gen_highpart (word_mode, src); | |
823 | ||
824 | if (mode == TImode) | |
825 | { | |
826 | emit_insn (gen_aarch64_movtilow_di (dst, src_lo)); | |
827 | emit_insn (gen_aarch64_movtihigh_di (dst, src_hi)); | |
828 | } | |
829 | else | |
830 | { | |
831 | emit_insn (gen_aarch64_movtflow_di (dst, src_lo)); | |
832 | emit_insn (gen_aarch64_movtfhigh_di (dst, src_hi)); | |
833 | } | |
834 | return; | |
835 | } | |
836 | else if (GP_REGNUM_P (dst_regno) && FP_REGNUM_P (src_regno)) | |
837 | { | |
838 | dst_lo = gen_lowpart (word_mode, dst); | |
839 | dst_hi = gen_highpart (word_mode, dst); | |
840 | ||
841 | if (mode == TImode) | |
842 | { | |
843 | emit_insn (gen_aarch64_movdi_tilow (dst_lo, src)); | |
844 | emit_insn (gen_aarch64_movdi_tihigh (dst_hi, src)); | |
845 | } | |
846 | else | |
847 | { | |
848 | emit_insn (gen_aarch64_movdi_tflow (dst_lo, src)); | |
849 | emit_insn (gen_aarch64_movdi_tfhigh (dst_hi, src)); | |
850 | } | |
851 | return; | |
852 | } | |
853 | } | |
854 | ||
855 | dst_lo = gen_lowpart (word_mode, dst); | |
856 | dst_hi = gen_highpart (word_mode, dst); | |
857 | src_lo = gen_lowpart (word_mode, src); | |
858 | src_hi = gen_highpart_mode (word_mode, mode, src); | |
859 | ||
860 | /* At most one pairing may overlap. */ | |
861 | if (reg_overlap_mentioned_p (dst_lo, src_hi)) | |
862 | { | |
863 | aarch64_emit_move (dst_hi, src_hi); | |
864 | aarch64_emit_move (dst_lo, src_lo); | |
865 | } | |
866 | else | |
867 | { | |
868 | aarch64_emit_move (dst_lo, src_lo); | |
869 | aarch64_emit_move (dst_hi, src_hi); | |
870 | } | |
871 | } | |
872 | ||
873 | bool | |
874 | aarch64_split_128bit_move_p (rtx dst, rtx src) | |
875 | { | |
876 | return (! REG_P (src) | |
877 | || ! (FP_REGNUM_P (REGNO (dst)) && FP_REGNUM_P (REGNO (src)))); | |
878 | } | |
879 | ||
880 | /* Split a complex SIMD combine. */ | |
881 | ||
882 | void | |
883 | aarch64_split_simd_combine (rtx dst, rtx src1, rtx src2) | |
884 | { | |
885 | enum machine_mode src_mode = GET_MODE (src1); | |
886 | enum machine_mode dst_mode = GET_MODE (dst); | |
887 | ||
888 | gcc_assert (VECTOR_MODE_P (dst_mode)); | |
889 | ||
890 | if (REG_P (dst) && REG_P (src1) && REG_P (src2)) | |
891 | { | |
892 | rtx (*gen) (rtx, rtx, rtx); | |
893 | ||
894 | switch (src_mode) | |
895 | { | |
896 | case V8QImode: | |
897 | gen = gen_aarch64_simd_combinev8qi; | |
898 | break; | |
899 | case V4HImode: | |
900 | gen = gen_aarch64_simd_combinev4hi; | |
901 | break; | |
902 | case V2SImode: | |
903 | gen = gen_aarch64_simd_combinev2si; | |
904 | break; | |
905 | case V2SFmode: | |
906 | gen = gen_aarch64_simd_combinev2sf; | |
907 | break; | |
908 | case DImode: | |
909 | gen = gen_aarch64_simd_combinedi; | |
910 | break; | |
911 | case DFmode: | |
912 | gen = gen_aarch64_simd_combinedf; | |
913 | break; | |
914 | default: | |
915 | gcc_unreachable (); | |
916 | } | |
917 | ||
918 | emit_insn (gen (dst, src1, src2)); | |
919 | return; | |
920 | } | |
921 | } | |
922 | ||
923 | /* Split a complex SIMD move. */ | |
924 | ||
925 | void | |
926 | aarch64_split_simd_move (rtx dst, rtx src) | |
927 | { | |
928 | enum machine_mode src_mode = GET_MODE (src); | |
929 | enum machine_mode dst_mode = GET_MODE (dst); | |
930 | ||
931 | gcc_assert (VECTOR_MODE_P (dst_mode)); | |
932 | ||
933 | if (REG_P (dst) && REG_P (src)) | |
934 | { | |
935 | rtx (*gen) (rtx, rtx); | |
936 | ||
937 | gcc_assert (VECTOR_MODE_P (src_mode)); | |
938 | ||
939 | switch (src_mode) | |
940 | { | |
941 | case V16QImode: | |
942 | gen = gen_aarch64_split_simd_movv16qi; | |
943 | break; | |
944 | case V8HImode: | |
945 | gen = gen_aarch64_split_simd_movv8hi; | |
946 | break; | |
947 | case V4SImode: | |
948 | gen = gen_aarch64_split_simd_movv4si; | |
949 | break; | |
950 | case V2DImode: | |
951 | gen = gen_aarch64_split_simd_movv2di; | |
952 | break; | |
953 | case V4SFmode: | |
954 | gen = gen_aarch64_split_simd_movv4sf; | |
955 | break; | |
956 | case V2DFmode: | |
957 | gen = gen_aarch64_split_simd_movv2df; | |
958 | break; | |
959 | default: | |
960 | gcc_unreachable (); | |
961 | } | |
962 | ||
963 | emit_insn (gen (dst, src)); | |
964 | return; | |
965 | } | |
966 | } | |
967 | ||
968 | static rtx | |
969 | aarch64_force_temporary (enum machine_mode mode, rtx x, rtx value) | |
970 | { | |
971 | if (can_create_pseudo_p ()) | |
972 | return force_reg (mode, value); | |
973 | else | |
974 | { | |
975 | x = aarch64_emit_move (x, value); | |
976 | return x; | |
977 | } | |
978 | } | |
979 | ||
980 | ||
981 | static rtx | |
982 | aarch64_add_offset (enum machine_mode mode, rtx temp, rtx reg, HOST_WIDE_INT offset) | |
983 | { | |
984 | if (!aarch64_plus_immediate (GEN_INT (offset), mode)) | |
985 | { | |
986 | rtx high; | |
987 | /* Load the full offset into a register. This | |
988 | might be improvable in the future. */ | |
989 | high = GEN_INT (offset); | |
990 | offset = 0; | |
991 | high = aarch64_force_temporary (mode, temp, high); | |
992 | reg = aarch64_force_temporary (mode, temp, | |
993 | gen_rtx_PLUS (mode, high, reg)); | |
994 | } | |
995 | return plus_constant (mode, reg, offset); | |
996 | } | |
997 | ||
998 | void | |
999 | aarch64_expand_mov_immediate (rtx dest, rtx imm) | |
1000 | { | |
1001 | enum machine_mode mode = GET_MODE (dest); | |
1002 | unsigned HOST_WIDE_INT mask; | |
1003 | int i; | |
1004 | bool first; | |
1005 | unsigned HOST_WIDE_INT val; | |
1006 | bool subtargets; | |
1007 | rtx subtarget; | |
1008 | int one_match, zero_match; | |
1009 | ||
1010 | gcc_assert (mode == SImode || mode == DImode); | |
1011 | ||
1012 | /* Check on what type of symbol it is. */ | |
1013 | if (GET_CODE (imm) == SYMBOL_REF | |
1014 | || GET_CODE (imm) == LABEL_REF | |
1015 | || GET_CODE (imm) == CONST) | |
1016 | { | |
1017 | rtx mem, base, offset; | |
1018 | enum aarch64_symbol_type sty; | |
1019 | ||
1020 | /* If we have (const (plus symbol offset)), separate out the offset | |
1021 | before we start classifying the symbol. */ | |
1022 | split_const (imm, &base, &offset); | |
1023 | ||
1024 | sty = aarch64_classify_symbol (base, SYMBOL_CONTEXT_ADR); | |
1025 | switch (sty) | |
1026 | { | |
1027 | case SYMBOL_FORCE_TO_MEM: | |
1028 | if (offset != const0_rtx | |
1029 | && targetm.cannot_force_const_mem (mode, imm)) | |
1030 | { | |
1031 | gcc_assert (can_create_pseudo_p ()); | |
1032 | base = aarch64_force_temporary (mode, dest, base); | |
1033 | base = aarch64_add_offset (mode, NULL, base, INTVAL (offset)); | |
1034 | aarch64_emit_move (dest, base); | |
1035 | return; | |
1036 | } | |
1037 | mem = force_const_mem (ptr_mode, imm); | |
1038 | gcc_assert (mem); | |
1039 | if (mode != ptr_mode) | |
1040 | mem = gen_rtx_ZERO_EXTEND (mode, mem); | |
1041 | emit_insn (gen_rtx_SET (VOIDmode, dest, mem)); | |
1042 | return; | |
1043 | ||
1044 | case SYMBOL_SMALL_TLSGD: | |
1045 | case SYMBOL_SMALL_TLSDESC: | |
1046 | case SYMBOL_SMALL_GOTTPREL: | |
1047 | case SYMBOL_SMALL_GOT: | |
1048 | case SYMBOL_TINY_GOT: | |
1049 | if (offset != const0_rtx) | |
1050 | { | |
1051 | gcc_assert(can_create_pseudo_p ()); | |
1052 | base = aarch64_force_temporary (mode, dest, base); | |
1053 | base = aarch64_add_offset (mode, NULL, base, INTVAL (offset)); | |
1054 | aarch64_emit_move (dest, base); | |
1055 | return; | |
1056 | } | |
1057 | /* FALLTHRU */ | |
1058 | ||
1059 | case SYMBOL_SMALL_TPREL: | |
1060 | case SYMBOL_SMALL_ABSOLUTE: | |
1061 | case SYMBOL_TINY_ABSOLUTE: | |
1062 | aarch64_load_symref_appropriately (dest, imm, sty); | |
1063 | return; | |
1064 | ||
1065 | default: | |
1066 | gcc_unreachable (); | |
1067 | } | |
1068 | } | |
1069 | ||
1070 | if (CONST_INT_P (imm) && aarch64_move_imm (INTVAL (imm), mode)) | |
1071 | { | |
1072 | emit_insn (gen_rtx_SET (VOIDmode, dest, imm)); | |
1073 | return; | |
1074 | } | |
1075 | ||
1076 | if (!CONST_INT_P (imm)) | |
1077 | { | |
1078 | if (GET_CODE (imm) == HIGH) | |
1079 | emit_insn (gen_rtx_SET (VOIDmode, dest, imm)); | |
1080 | else | |
1081 | { | |
1082 | rtx mem = force_const_mem (mode, imm); | |
1083 | gcc_assert (mem); | |
1084 | emit_insn (gen_rtx_SET (VOIDmode, dest, mem)); | |
1085 | } | |
1086 | ||
1087 | return; | |
1088 | } | |
1089 | ||
1090 | if (mode == SImode) | |
1091 | { | |
1092 | /* We know we can't do this in 1 insn, and we must be able to do it | |
1093 | in two; so don't mess around looking for sequences that don't buy | |
1094 | us anything. */ | |
1095 | emit_insn (gen_rtx_SET (VOIDmode, dest, GEN_INT (INTVAL (imm) & 0xffff))); | |
1096 | emit_insn (gen_insv_immsi (dest, GEN_INT (16), | |
1097 | GEN_INT ((INTVAL (imm) >> 16) & 0xffff))); | |
1098 | return; | |
1099 | } | |
1100 | ||
1101 | /* Remaining cases are all for DImode. */ | |
1102 | ||
1103 | val = INTVAL (imm); | |
1104 | subtargets = optimize && can_create_pseudo_p (); | |
1105 | ||
1106 | one_match = 0; | |
1107 | zero_match = 0; | |
1108 | mask = 0xffff; | |
1109 | ||
1110 | for (i = 0; i < 64; i += 16, mask <<= 16) | |
1111 | { | |
1112 | if ((val & mask) == 0) | |
1113 | zero_match++; | |
1114 | else if ((val & mask) == mask) | |
1115 | one_match++; | |
1116 | } | |
1117 | ||
1118 | if (one_match == 2) | |
1119 | { | |
1120 | mask = 0xffff; | |
1121 | for (i = 0; i < 64; i += 16, mask <<= 16) | |
1122 | { | |
1123 | if ((val & mask) != mask) | |
1124 | { | |
1125 | emit_insn (gen_rtx_SET (VOIDmode, dest, GEN_INT (val | mask))); | |
1126 | emit_insn (gen_insv_immdi (dest, GEN_INT (i), | |
1127 | GEN_INT ((val >> i) & 0xffff))); | |
1128 | return; | |
1129 | } | |
1130 | } | |
1131 | gcc_unreachable (); | |
1132 | } | |
1133 | ||
1134 | if (zero_match == 2) | |
1135 | goto simple_sequence; | |
1136 | ||
1137 | mask = 0x0ffff0000UL; | |
1138 | for (i = 16; i < 64; i += 16, mask <<= 16) | |
1139 | { | |
1140 | HOST_WIDE_INT comp = mask & ~(mask - 1); | |
1141 | ||
1142 | if (aarch64_uimm12_shift (val - (val & mask))) | |
1143 | { | |
1144 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1145 | ||
1146 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, GEN_INT (val & mask))); | |
1147 | emit_insn (gen_adddi3 (dest, subtarget, | |
1148 | GEN_INT (val - (val & mask)))); | |
1149 | return; | |
1150 | } | |
1151 | else if (aarch64_uimm12_shift (-(val - ((val + comp) & mask)))) | |
1152 | { | |
1153 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1154 | ||
1155 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1156 | GEN_INT ((val + comp) & mask))); | |
1157 | emit_insn (gen_adddi3 (dest, subtarget, | |
1158 | GEN_INT (val - ((val + comp) & mask)))); | |
1159 | return; | |
1160 | } | |
1161 | else if (aarch64_uimm12_shift (val - ((val - comp) | ~mask))) | |
1162 | { | |
1163 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1164 | ||
1165 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1166 | GEN_INT ((val - comp) | ~mask))); | |
1167 | emit_insn (gen_adddi3 (dest, subtarget, | |
1168 | GEN_INT (val - ((val - comp) | ~mask)))); | |
1169 | return; | |
1170 | } | |
1171 | else if (aarch64_uimm12_shift (-(val - (val | ~mask)))) | |
1172 | { | |
1173 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1174 | ||
1175 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1176 | GEN_INT (val | ~mask))); | |
1177 | emit_insn (gen_adddi3 (dest, subtarget, | |
1178 | GEN_INT (val - (val | ~mask)))); | |
1179 | return; | |
1180 | } | |
1181 | } | |
1182 | ||
1183 | /* See if we can do it by arithmetically combining two | |
1184 | immediates. */ | |
1185 | for (i = 0; i < AARCH64_NUM_BITMASKS; i++) | |
1186 | { | |
1187 | int j; | |
1188 | mask = 0xffff; | |
1189 | ||
1190 | if (aarch64_uimm12_shift (val - aarch64_bitmasks[i]) | |
1191 | || aarch64_uimm12_shift (-val + aarch64_bitmasks[i])) | |
1192 | { | |
1193 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1194 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1195 | GEN_INT (aarch64_bitmasks[i]))); | |
1196 | emit_insn (gen_adddi3 (dest, subtarget, | |
1197 | GEN_INT (val - aarch64_bitmasks[i]))); | |
1198 | return; | |
1199 | } | |
1200 | ||
1201 | for (j = 0; j < 64; j += 16, mask <<= 16) | |
1202 | { | |
1203 | if ((aarch64_bitmasks[i] & ~mask) == (val & ~mask)) | |
1204 | { | |
1205 | emit_insn (gen_rtx_SET (VOIDmode, dest, | |
1206 | GEN_INT (aarch64_bitmasks[i]))); | |
1207 | emit_insn (gen_insv_immdi (dest, GEN_INT (j), | |
1208 | GEN_INT ((val >> j) & 0xffff))); | |
1209 | return; | |
1210 | } | |
1211 | } | |
1212 | } | |
1213 | ||
1214 | /* See if we can do it by logically combining two immediates. */ | |
1215 | for (i = 0; i < AARCH64_NUM_BITMASKS; i++) | |
1216 | { | |
1217 | if ((aarch64_bitmasks[i] & val) == aarch64_bitmasks[i]) | |
1218 | { | |
1219 | int j; | |
1220 | ||
1221 | for (j = i + 1; j < AARCH64_NUM_BITMASKS; j++) | |
1222 | if (val == (aarch64_bitmasks[i] | aarch64_bitmasks[j])) | |
1223 | { | |
1224 | subtarget = subtargets ? gen_reg_rtx (mode) : dest; | |
1225 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1226 | GEN_INT (aarch64_bitmasks[i]))); | |
1227 | emit_insn (gen_iordi3 (dest, subtarget, | |
1228 | GEN_INT (aarch64_bitmasks[j]))); | |
1229 | return; | |
1230 | } | |
1231 | } | |
1232 | else if ((val & aarch64_bitmasks[i]) == val) | |
1233 | { | |
1234 | int j; | |
1235 | ||
1236 | for (j = i + 1; j < AARCH64_NUM_BITMASKS; j++) | |
1237 | if (val == (aarch64_bitmasks[j] & aarch64_bitmasks[i])) | |
1238 | { | |
1239 | ||
1240 | subtarget = subtargets ? gen_reg_rtx (mode) : dest; | |
1241 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1242 | GEN_INT (aarch64_bitmasks[j]))); | |
1243 | emit_insn (gen_anddi3 (dest, subtarget, | |
1244 | GEN_INT (aarch64_bitmasks[i]))); | |
1245 | return; | |
1246 | } | |
1247 | } | |
1248 | } | |
1249 | ||
1250 | simple_sequence: | |
1251 | first = true; | |
1252 | mask = 0xffff; | |
1253 | for (i = 0; i < 64; i += 16, mask <<= 16) | |
1254 | { | |
1255 | if ((val & mask) != 0) | |
1256 | { | |
1257 | if (first) | |
1258 | { | |
1259 | emit_insn (gen_rtx_SET (VOIDmode, dest, | |
1260 | GEN_INT (val & mask))); | |
1261 | first = false; | |
1262 | } | |
1263 | else | |
1264 | emit_insn (gen_insv_immdi (dest, GEN_INT (i), | |
1265 | GEN_INT ((val >> i) & 0xffff))); | |
1266 | } | |
1267 | } | |
1268 | } | |
1269 | ||
1270 | static bool | |
1271 | aarch64_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED, | |
1272 | tree exp ATTRIBUTE_UNUSED) | |
1273 | { | |
1274 | /* Currently, always true. */ | |
1275 | return true; | |
1276 | } | |
1277 | ||
1278 | /* Implement TARGET_PASS_BY_REFERENCE. */ | |
1279 | ||
1280 | static bool | |
1281 | aarch64_pass_by_reference (cumulative_args_t pcum ATTRIBUTE_UNUSED, | |
1282 | enum machine_mode mode, | |
1283 | const_tree type, | |
1284 | bool named ATTRIBUTE_UNUSED) | |
1285 | { | |
1286 | HOST_WIDE_INT size; | |
1287 | enum machine_mode dummymode; | |
1288 | int nregs; | |
1289 | ||
1290 | /* GET_MODE_SIZE (BLKmode) is useless since it is 0. */ | |
1291 | size = (mode == BLKmode && type) | |
1292 | ? int_size_in_bytes (type) : (int) GET_MODE_SIZE (mode); | |
1293 | ||
1294 | /* Aggregates are passed by reference based on their size. */ | |
1295 | if (type && AGGREGATE_TYPE_P (type)) | |
1296 | { | |
1297 | size = int_size_in_bytes (type); | |
1298 | } | |
1299 | ||
1300 | /* Variable sized arguments are always returned by reference. */ | |
1301 | if (size < 0) | |
1302 | return true; | |
1303 | ||
1304 | /* Can this be a candidate to be passed in fp/simd register(s)? */ | |
1305 | if (aarch64_vfp_is_call_or_return_candidate (mode, type, | |
1306 | &dummymode, &nregs, | |
1307 | NULL)) | |
1308 | return false; | |
1309 | ||
1310 | /* Arguments which are variable sized or larger than 2 registers are | |
1311 | passed by reference unless they are a homogenous floating point | |
1312 | aggregate. */ | |
1313 | return size > 2 * UNITS_PER_WORD; | |
1314 | } | |
1315 | ||
1316 | /* Return TRUE if VALTYPE is padded to its least significant bits. */ | |
1317 | static bool | |
1318 | aarch64_return_in_msb (const_tree valtype) | |
1319 | { | |
1320 | enum machine_mode dummy_mode; | |
1321 | int dummy_int; | |
1322 | ||
1323 | /* Never happens in little-endian mode. */ | |
1324 | if (!BYTES_BIG_ENDIAN) | |
1325 | return false; | |
1326 | ||
1327 | /* Only composite types smaller than or equal to 16 bytes can | |
1328 | be potentially returned in registers. */ | |
1329 | if (!aarch64_composite_type_p (valtype, TYPE_MODE (valtype)) | |
1330 | || int_size_in_bytes (valtype) <= 0 | |
1331 | || int_size_in_bytes (valtype) > 16) | |
1332 | return false; | |
1333 | ||
1334 | /* But not a composite that is an HFA (Homogeneous Floating-point Aggregate) | |
1335 | or an HVA (Homogeneous Short-Vector Aggregate); such a special composite | |
1336 | is always passed/returned in the least significant bits of fp/simd | |
1337 | register(s). */ | |
1338 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (valtype), valtype, | |
1339 | &dummy_mode, &dummy_int, NULL)) | |
1340 | return false; | |
1341 | ||
1342 | return true; | |
1343 | } | |
1344 | ||
1345 | /* Implement TARGET_FUNCTION_VALUE. | |
1346 | Define how to find the value returned by a function. */ | |
1347 | ||
1348 | static rtx | |
1349 | aarch64_function_value (const_tree type, const_tree func, | |
1350 | bool outgoing ATTRIBUTE_UNUSED) | |
1351 | { | |
1352 | enum machine_mode mode; | |
1353 | int unsignedp; | |
1354 | int count; | |
1355 | enum machine_mode ag_mode; | |
1356 | ||
1357 | mode = TYPE_MODE (type); | |
1358 | if (INTEGRAL_TYPE_P (type)) | |
1359 | mode = promote_function_mode (type, mode, &unsignedp, func, 1); | |
1360 | ||
1361 | if (aarch64_return_in_msb (type)) | |
1362 | { | |
1363 | HOST_WIDE_INT size = int_size_in_bytes (type); | |
1364 | ||
1365 | if (size % UNITS_PER_WORD != 0) | |
1366 | { | |
1367 | size += UNITS_PER_WORD - size % UNITS_PER_WORD; | |
1368 | mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0); | |
1369 | } | |
1370 | } | |
1371 | ||
1372 | if (aarch64_vfp_is_call_or_return_candidate (mode, type, | |
1373 | &ag_mode, &count, NULL)) | |
1374 | { | |
1375 | if (!aarch64_composite_type_p (type, mode)) | |
1376 | { | |
1377 | gcc_assert (count == 1 && mode == ag_mode); | |
1378 | return gen_rtx_REG (mode, V0_REGNUM); | |
1379 | } | |
1380 | else | |
1381 | { | |
1382 | int i; | |
1383 | rtx par; | |
1384 | ||
1385 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (count)); | |
1386 | for (i = 0; i < count; i++) | |
1387 | { | |
1388 | rtx tmp = gen_rtx_REG (ag_mode, V0_REGNUM + i); | |
1389 | tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp, | |
1390 | GEN_INT (i * GET_MODE_SIZE (ag_mode))); | |
1391 | XVECEXP (par, 0, i) = tmp; | |
1392 | } | |
1393 | return par; | |
1394 | } | |
1395 | } | |
1396 | else | |
1397 | return gen_rtx_REG (mode, R0_REGNUM); | |
1398 | } | |
1399 | ||
1400 | /* Implements TARGET_FUNCTION_VALUE_REGNO_P. | |
1401 | Return true if REGNO is the number of a hard register in which the values | |
1402 | of called function may come back. */ | |
1403 | ||
1404 | static bool | |
1405 | aarch64_function_value_regno_p (const unsigned int regno) | |
1406 | { | |
1407 | /* Maximum of 16 bytes can be returned in the general registers. Examples | |
1408 | of 16-byte return values are: 128-bit integers and 16-byte small | |
1409 | structures (excluding homogeneous floating-point aggregates). */ | |
1410 | if (regno == R0_REGNUM || regno == R1_REGNUM) | |
1411 | return true; | |
1412 | ||
1413 | /* Up to four fp/simd registers can return a function value, e.g. a | |
1414 | homogeneous floating-point aggregate having four members. */ | |
1415 | if (regno >= V0_REGNUM && regno < V0_REGNUM + HA_MAX_NUM_FLDS) | |
1416 | return !TARGET_GENERAL_REGS_ONLY; | |
1417 | ||
1418 | return false; | |
1419 | } | |
1420 | ||
1421 | /* Implement TARGET_RETURN_IN_MEMORY. | |
1422 | ||
1423 | If the type T of the result of a function is such that | |
1424 | void func (T arg) | |
1425 | would require that arg be passed as a value in a register (or set of | |
1426 | registers) according to the parameter passing rules, then the result | |
1427 | is returned in the same registers as would be used for such an | |
1428 | argument. */ | |
1429 | ||
1430 | static bool | |
1431 | aarch64_return_in_memory (const_tree type, const_tree fndecl ATTRIBUTE_UNUSED) | |
1432 | { | |
1433 | HOST_WIDE_INT size; | |
1434 | enum machine_mode ag_mode; | |
1435 | int count; | |
1436 | ||
1437 | if (!AGGREGATE_TYPE_P (type) | |
1438 | && TREE_CODE (type) != COMPLEX_TYPE | |
1439 | && TREE_CODE (type) != VECTOR_TYPE) | |
1440 | /* Simple scalar types always returned in registers. */ | |
1441 | return false; | |
1442 | ||
1443 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (type), | |
1444 | type, | |
1445 | &ag_mode, | |
1446 | &count, | |
1447 | NULL)) | |
1448 | return false; | |
1449 | ||
1450 | /* Types larger than 2 registers returned in memory. */ | |
1451 | size = int_size_in_bytes (type); | |
1452 | return (size < 0 || size > 2 * UNITS_PER_WORD); | |
1453 | } | |
1454 | ||
1455 | static bool | |
1456 | aarch64_vfp_is_call_candidate (cumulative_args_t pcum_v, enum machine_mode mode, | |
1457 | const_tree type, int *nregs) | |
1458 | { | |
1459 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
1460 | return aarch64_vfp_is_call_or_return_candidate (mode, | |
1461 | type, | |
1462 | &pcum->aapcs_vfp_rmode, | |
1463 | nregs, | |
1464 | NULL); | |
1465 | } | |
1466 | ||
1467 | /* Given MODE and TYPE of a function argument, return the alignment in | |
1468 | bits. The idea is to suppress any stronger alignment requested by | |
1469 | the user and opt for the natural alignment (specified in AAPCS64 \S 4.1). | |
1470 | This is a helper function for local use only. */ | |
1471 | ||
1472 | static unsigned int | |
1473 | aarch64_function_arg_alignment (enum machine_mode mode, const_tree type) | |
1474 | { | |
1475 | unsigned int alignment; | |
1476 | ||
1477 | if (type) | |
1478 | { | |
1479 | if (!integer_zerop (TYPE_SIZE (type))) | |
1480 | { | |
1481 | if (TYPE_MODE (type) == mode) | |
1482 | alignment = TYPE_ALIGN (type); | |
1483 | else | |
1484 | alignment = GET_MODE_ALIGNMENT (mode); | |
1485 | } | |
1486 | else | |
1487 | alignment = 0; | |
1488 | } | |
1489 | else | |
1490 | alignment = GET_MODE_ALIGNMENT (mode); | |
1491 | ||
1492 | return alignment; | |
1493 | } | |
1494 | ||
1495 | /* Layout a function argument according to the AAPCS64 rules. The rule | |
1496 | numbers refer to the rule numbers in the AAPCS64. */ | |
1497 | ||
1498 | static void | |
1499 | aarch64_layout_arg (cumulative_args_t pcum_v, enum machine_mode mode, | |
1500 | const_tree type, | |
1501 | bool named ATTRIBUTE_UNUSED) | |
1502 | { | |
1503 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
1504 | int ncrn, nvrn, nregs; | |
1505 | bool allocate_ncrn, allocate_nvrn; | |
1506 | ||
1507 | /* We need to do this once per argument. */ | |
1508 | if (pcum->aapcs_arg_processed) | |
1509 | return; | |
1510 | ||
1511 | pcum->aapcs_arg_processed = true; | |
1512 | ||
1513 | allocate_ncrn = (type) ? !(FLOAT_TYPE_P (type)) : !FLOAT_MODE_P (mode); | |
1514 | allocate_nvrn = aarch64_vfp_is_call_candidate (pcum_v, | |
1515 | mode, | |
1516 | type, | |
1517 | &nregs); | |
1518 | ||
1519 | /* allocate_ncrn may be false-positive, but allocate_nvrn is quite reliable. | |
1520 | The following code thus handles passing by SIMD/FP registers first. */ | |
1521 | ||
1522 | nvrn = pcum->aapcs_nvrn; | |
1523 | ||
1524 | /* C1 - C5 for floating point, homogenous floating point aggregates (HFA) | |
1525 | and homogenous short-vector aggregates (HVA). */ | |
1526 | if (allocate_nvrn) | |
1527 | { | |
1528 | if (nvrn + nregs <= NUM_FP_ARG_REGS) | |
1529 | { | |
1530 | pcum->aapcs_nextnvrn = nvrn + nregs; | |
1531 | if (!aarch64_composite_type_p (type, mode)) | |
1532 | { | |
1533 | gcc_assert (nregs == 1); | |
1534 | pcum->aapcs_reg = gen_rtx_REG (mode, V0_REGNUM + nvrn); | |
1535 | } | |
1536 | else | |
1537 | { | |
1538 | rtx par; | |
1539 | int i; | |
1540 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (nregs)); | |
1541 | for (i = 0; i < nregs; i++) | |
1542 | { | |
1543 | rtx tmp = gen_rtx_REG (pcum->aapcs_vfp_rmode, | |
1544 | V0_REGNUM + nvrn + i); | |
1545 | tmp = gen_rtx_EXPR_LIST | |
1546 | (VOIDmode, tmp, | |
1547 | GEN_INT (i * GET_MODE_SIZE (pcum->aapcs_vfp_rmode))); | |
1548 | XVECEXP (par, 0, i) = tmp; | |
1549 | } | |
1550 | pcum->aapcs_reg = par; | |
1551 | } | |
1552 | return; | |
1553 | } | |
1554 | else | |
1555 | { | |
1556 | /* C.3 NSRN is set to 8. */ | |
1557 | pcum->aapcs_nextnvrn = NUM_FP_ARG_REGS; | |
1558 | goto on_stack; | |
1559 | } | |
1560 | } | |
1561 | ||
1562 | ncrn = pcum->aapcs_ncrn; | |
1563 | nregs = ((type ? int_size_in_bytes (type) : GET_MODE_SIZE (mode)) | |
1564 | + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
1565 | ||
1566 | ||
1567 | /* C6 - C9. though the sign and zero extension semantics are | |
1568 | handled elsewhere. This is the case where the argument fits | |
1569 | entirely general registers. */ | |
1570 | if (allocate_ncrn && (ncrn + nregs <= NUM_ARG_REGS)) | |
1571 | { | |
1572 | unsigned int alignment = aarch64_function_arg_alignment (mode, type); | |
1573 | ||
1574 | gcc_assert (nregs == 0 || nregs == 1 || nregs == 2); | |
1575 | ||
1576 | /* C.8 if the argument has an alignment of 16 then the NGRN is | |
1577 | rounded up to the next even number. */ | |
1578 | if (nregs == 2 && alignment == 16 * BITS_PER_UNIT && ncrn % 2) | |
1579 | { | |
1580 | ++ncrn; | |
1581 | gcc_assert (ncrn + nregs <= NUM_ARG_REGS); | |
1582 | } | |
1583 | /* NREGS can be 0 when e.g. an empty structure is to be passed. | |
1584 | A reg is still generated for it, but the caller should be smart | |
1585 | enough not to use it. */ | |
1586 | if (nregs == 0 || nregs == 1 || GET_MODE_CLASS (mode) == MODE_INT) | |
1587 | { | |
1588 | pcum->aapcs_reg = gen_rtx_REG (mode, R0_REGNUM + ncrn); | |
1589 | } | |
1590 | else | |
1591 | { | |
1592 | rtx par; | |
1593 | int i; | |
1594 | ||
1595 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (nregs)); | |
1596 | for (i = 0; i < nregs; i++) | |
1597 | { | |
1598 | rtx tmp = gen_rtx_REG (word_mode, R0_REGNUM + ncrn + i); | |
1599 | tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp, | |
1600 | GEN_INT (i * UNITS_PER_WORD)); | |
1601 | XVECEXP (par, 0, i) = tmp; | |
1602 | } | |
1603 | pcum->aapcs_reg = par; | |
1604 | } | |
1605 | ||
1606 | pcum->aapcs_nextncrn = ncrn + nregs; | |
1607 | return; | |
1608 | } | |
1609 | ||
1610 | /* C.11 */ | |
1611 | pcum->aapcs_nextncrn = NUM_ARG_REGS; | |
1612 | ||
1613 | /* The argument is passed on stack; record the needed number of words for | |
1614 | this argument (we can re-use NREGS) and align the total size if | |
1615 | necessary. */ | |
1616 | on_stack: | |
1617 | pcum->aapcs_stack_words = nregs; | |
1618 | if (aarch64_function_arg_alignment (mode, type) == 16 * BITS_PER_UNIT) | |
1619 | pcum->aapcs_stack_size = AARCH64_ROUND_UP (pcum->aapcs_stack_size, | |
1620 | 16 / UNITS_PER_WORD) + 1; | |
1621 | return; | |
1622 | } | |
1623 | ||
1624 | /* Implement TARGET_FUNCTION_ARG. */ | |
1625 | ||
1626 | static rtx | |
1627 | aarch64_function_arg (cumulative_args_t pcum_v, enum machine_mode mode, | |
1628 | const_tree type, bool named) | |
1629 | { | |
1630 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
1631 | gcc_assert (pcum->pcs_variant == ARM_PCS_AAPCS64); | |
1632 | ||
1633 | if (mode == VOIDmode) | |
1634 | return NULL_RTX; | |
1635 | ||
1636 | aarch64_layout_arg (pcum_v, mode, type, named); | |
1637 | return pcum->aapcs_reg; | |
1638 | } | |
1639 | ||
1640 | void | |
1641 | aarch64_init_cumulative_args (CUMULATIVE_ARGS *pcum, | |
1642 | const_tree fntype ATTRIBUTE_UNUSED, | |
1643 | rtx libname ATTRIBUTE_UNUSED, | |
1644 | const_tree fndecl ATTRIBUTE_UNUSED, | |
1645 | unsigned n_named ATTRIBUTE_UNUSED) | |
1646 | { | |
1647 | pcum->aapcs_ncrn = 0; | |
1648 | pcum->aapcs_nvrn = 0; | |
1649 | pcum->aapcs_nextncrn = 0; | |
1650 | pcum->aapcs_nextnvrn = 0; | |
1651 | pcum->pcs_variant = ARM_PCS_AAPCS64; | |
1652 | pcum->aapcs_reg = NULL_RTX; | |
1653 | pcum->aapcs_arg_processed = false; | |
1654 | pcum->aapcs_stack_words = 0; | |
1655 | pcum->aapcs_stack_size = 0; | |
1656 | ||
1657 | return; | |
1658 | } | |
1659 | ||
1660 | static void | |
1661 | aarch64_function_arg_advance (cumulative_args_t pcum_v, | |
1662 | enum machine_mode mode, | |
1663 | const_tree type, | |
1664 | bool named) | |
1665 | { | |
1666 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
1667 | if (pcum->pcs_variant == ARM_PCS_AAPCS64) | |
1668 | { | |
1669 | aarch64_layout_arg (pcum_v, mode, type, named); | |
1670 | gcc_assert ((pcum->aapcs_reg != NULL_RTX) | |
1671 | != (pcum->aapcs_stack_words != 0)); | |
1672 | pcum->aapcs_arg_processed = false; | |
1673 | pcum->aapcs_ncrn = pcum->aapcs_nextncrn; | |
1674 | pcum->aapcs_nvrn = pcum->aapcs_nextnvrn; | |
1675 | pcum->aapcs_stack_size += pcum->aapcs_stack_words; | |
1676 | pcum->aapcs_stack_words = 0; | |
1677 | pcum->aapcs_reg = NULL_RTX; | |
1678 | } | |
1679 | } | |
1680 | ||
1681 | bool | |
1682 | aarch64_function_arg_regno_p (unsigned regno) | |
1683 | { | |
1684 | return ((GP_REGNUM_P (regno) && regno < R0_REGNUM + NUM_ARG_REGS) | |
1685 | || (FP_REGNUM_P (regno) && regno < V0_REGNUM + NUM_FP_ARG_REGS)); | |
1686 | } | |
1687 | ||
1688 | /* Implement FUNCTION_ARG_BOUNDARY. Every parameter gets at least | |
1689 | PARM_BOUNDARY bits of alignment, but will be given anything up | |
1690 | to STACK_BOUNDARY bits if the type requires it. This makes sure | |
1691 | that both before and after the layout of each argument, the Next | |
1692 | Stacked Argument Address (NSAA) will have a minimum alignment of | |
1693 | 8 bytes. */ | |
1694 | ||
1695 | static unsigned int | |
1696 | aarch64_function_arg_boundary (enum machine_mode mode, const_tree type) | |
1697 | { | |
1698 | unsigned int alignment = aarch64_function_arg_alignment (mode, type); | |
1699 | ||
1700 | if (alignment < PARM_BOUNDARY) | |
1701 | alignment = PARM_BOUNDARY; | |
1702 | if (alignment > STACK_BOUNDARY) | |
1703 | alignment = STACK_BOUNDARY; | |
1704 | return alignment; | |
1705 | } | |
1706 | ||
1707 | /* For use by FUNCTION_ARG_PADDING (MODE, TYPE). | |
1708 | ||
1709 | Return true if an argument passed on the stack should be padded upwards, | |
1710 | i.e. if the least-significant byte of the stack slot has useful data. | |
1711 | ||
1712 | Small aggregate types are placed in the lowest memory address. | |
1713 | ||
1714 | The related parameter passing rules are B.4, C.3, C.5 and C.14. */ | |
1715 | ||
1716 | bool | |
1717 | aarch64_pad_arg_upward (enum machine_mode mode, const_tree type) | |
1718 | { | |
1719 | /* On little-endian targets, the least significant byte of every stack | |
1720 | argument is passed at the lowest byte address of the stack slot. */ | |
1721 | if (!BYTES_BIG_ENDIAN) | |
1722 | return true; | |
1723 | ||
1724 | /* Otherwise, integral, floating-point and pointer types are padded downward: | |
1725 | the least significant byte of a stack argument is passed at the highest | |
1726 | byte address of the stack slot. */ | |
1727 | if (type | |
1728 | ? (INTEGRAL_TYPE_P (type) || SCALAR_FLOAT_TYPE_P (type) | |
1729 | || POINTER_TYPE_P (type)) | |
1730 | : (SCALAR_INT_MODE_P (mode) || SCALAR_FLOAT_MODE_P (mode))) | |
1731 | return false; | |
1732 | ||
1733 | /* Everything else padded upward, i.e. data in first byte of stack slot. */ | |
1734 | return true; | |
1735 | } | |
1736 | ||
1737 | /* Similarly, for use by BLOCK_REG_PADDING (MODE, TYPE, FIRST). | |
1738 | ||
1739 | It specifies padding for the last (may also be the only) | |
1740 | element of a block move between registers and memory. If | |
1741 | assuming the block is in the memory, padding upward means that | |
1742 | the last element is padded after its highest significant byte, | |
1743 | while in downward padding, the last element is padded at the | |
1744 | its least significant byte side. | |
1745 | ||
1746 | Small aggregates and small complex types are always padded | |
1747 | upwards. | |
1748 | ||
1749 | We don't need to worry about homogeneous floating-point or | |
1750 | short-vector aggregates; their move is not affected by the | |
1751 | padding direction determined here. Regardless of endianness, | |
1752 | each element of such an aggregate is put in the least | |
1753 | significant bits of a fp/simd register. | |
1754 | ||
1755 | Return !BYTES_BIG_ENDIAN if the least significant byte of the | |
1756 | register has useful data, and return the opposite if the most | |
1757 | significant byte does. */ | |
1758 | ||
1759 | bool | |
1760 | aarch64_pad_reg_upward (enum machine_mode mode, const_tree type, | |
1761 | bool first ATTRIBUTE_UNUSED) | |
1762 | { | |
1763 | ||
1764 | /* Small composite types are always padded upward. */ | |
1765 | if (BYTES_BIG_ENDIAN && aarch64_composite_type_p (type, mode)) | |
1766 | { | |
1767 | HOST_WIDE_INT size = (type ? int_size_in_bytes (type) | |
1768 | : GET_MODE_SIZE (mode)); | |
1769 | if (size < 2 * UNITS_PER_WORD) | |
1770 | return true; | |
1771 | } | |
1772 | ||
1773 | /* Otherwise, use the default padding. */ | |
1774 | return !BYTES_BIG_ENDIAN; | |
1775 | } | |
1776 | ||
1777 | static enum machine_mode | |
1778 | aarch64_libgcc_cmp_return_mode (void) | |
1779 | { | |
1780 | return SImode; | |
1781 | } | |
1782 | ||
1783 | static bool | |
1784 | aarch64_frame_pointer_required (void) | |
1785 | { | |
1786 | /* If the function contains dynamic stack allocations, we need to | |
1787 | use the frame pointer to access the static parts of the frame. */ | |
1788 | if (cfun->calls_alloca) | |
1789 | return true; | |
1790 | ||
1791 | /* In aarch64_override_options_after_change | |
1792 | flag_omit_leaf_frame_pointer turns off the frame pointer by | |
1793 | default. Turn it back on now if we've not got a leaf | |
1794 | function. */ | |
1795 | if (flag_omit_leaf_frame_pointer | |
1796 | && (!crtl->is_leaf || df_regs_ever_live_p (LR_REGNUM))) | |
1797 | return true; | |
1798 | ||
1799 | return false; | |
1800 | } | |
1801 | ||
1802 | /* Mark the registers that need to be saved by the callee and calculate | |
1803 | the size of the callee-saved registers area and frame record (both FP | |
1804 | and LR may be omitted). */ | |
1805 | static void | |
1806 | aarch64_layout_frame (void) | |
1807 | { | |
1808 | HOST_WIDE_INT offset = 0; | |
1809 | int regno; | |
1810 | ||
1811 | if (reload_completed && cfun->machine->frame.laid_out) | |
1812 | return; | |
1813 | ||
1814 | /* First mark all the registers that really need to be saved... */ | |
1815 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) | |
1816 | cfun->machine->frame.reg_offset[regno] = -1; | |
1817 | ||
1818 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
1819 | cfun->machine->frame.reg_offset[regno] = -1; | |
1820 | ||
1821 | /* ... that includes the eh data registers (if needed)... */ | |
1822 | if (crtl->calls_eh_return) | |
1823 | for (regno = 0; EH_RETURN_DATA_REGNO (regno) != INVALID_REGNUM; regno++) | |
1824 | cfun->machine->frame.reg_offset[EH_RETURN_DATA_REGNO (regno)] = 0; | |
1825 | ||
1826 | /* ... and any callee saved register that dataflow says is live. */ | |
1827 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) | |
1828 | if (df_regs_ever_live_p (regno) | |
1829 | && !call_used_regs[regno]) | |
1830 | cfun->machine->frame.reg_offset[regno] = 0; | |
1831 | ||
1832 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
1833 | if (df_regs_ever_live_p (regno) | |
1834 | && !call_used_regs[regno]) | |
1835 | cfun->machine->frame.reg_offset[regno] = 0; | |
1836 | ||
1837 | if (frame_pointer_needed) | |
1838 | { | |
1839 | cfun->machine->frame.reg_offset[R30_REGNUM] = 0; | |
1840 | cfun->machine->frame.reg_offset[R29_REGNUM] = 0; | |
1841 | cfun->machine->frame.hardfp_offset = 2 * UNITS_PER_WORD; | |
1842 | } | |
1843 | ||
1844 | /* Now assign stack slots for them. */ | |
1845 | for (regno = R0_REGNUM; regno <= R28_REGNUM; regno++) | |
1846 | if (cfun->machine->frame.reg_offset[regno] != -1) | |
1847 | { | |
1848 | cfun->machine->frame.reg_offset[regno] = offset; | |
1849 | offset += UNITS_PER_WORD; | |
1850 | } | |
1851 | ||
1852 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
1853 | if (cfun->machine->frame.reg_offset[regno] != -1) | |
1854 | { | |
1855 | cfun->machine->frame.reg_offset[regno] = offset; | |
1856 | offset += UNITS_PER_WORD; | |
1857 | } | |
1858 | ||
1859 | if (frame_pointer_needed) | |
1860 | { | |
1861 | cfun->machine->frame.reg_offset[R29_REGNUM] = offset; | |
1862 | offset += UNITS_PER_WORD; | |
1863 | } | |
1864 | ||
1865 | if (cfun->machine->frame.reg_offset[R30_REGNUM] != -1) | |
1866 | { | |
1867 | cfun->machine->frame.reg_offset[R30_REGNUM] = offset; | |
1868 | offset += UNITS_PER_WORD; | |
1869 | } | |
1870 | ||
1871 | cfun->machine->frame.padding0 = | |
1872 | (AARCH64_ROUND_UP (offset, STACK_BOUNDARY / BITS_PER_UNIT) - offset); | |
1873 | offset = AARCH64_ROUND_UP (offset, STACK_BOUNDARY / BITS_PER_UNIT); | |
1874 | ||
1875 | cfun->machine->frame.saved_regs_size = offset; | |
1876 | cfun->machine->frame.laid_out = true; | |
1877 | } | |
1878 | ||
1879 | /* Make the last instruction frame-related and note that it performs | |
1880 | the operation described by FRAME_PATTERN. */ | |
1881 | ||
1882 | static void | |
1883 | aarch64_set_frame_expr (rtx frame_pattern) | |
1884 | { | |
1885 | rtx insn; | |
1886 | ||
1887 | insn = get_last_insn (); | |
1888 | RTX_FRAME_RELATED_P (insn) = 1; | |
1889 | RTX_FRAME_RELATED_P (frame_pattern) = 1; | |
1890 | REG_NOTES (insn) = alloc_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
1891 | frame_pattern, | |
1892 | REG_NOTES (insn)); | |
1893 | } | |
1894 | ||
1895 | static bool | |
1896 | aarch64_register_saved_on_entry (int regno) | |
1897 | { | |
1898 | return cfun->machine->frame.reg_offset[regno] != -1; | |
1899 | } | |
1900 | ||
1901 | ||
1902 | static void | |
1903 | aarch64_save_or_restore_fprs (int start_offset, int increment, | |
1904 | bool restore, rtx base_rtx) | |
1905 | ||
1906 | { | |
1907 | unsigned regno; | |
1908 | unsigned regno2; | |
1909 | rtx insn; | |
1910 | rtx (*gen_mem_ref)(enum machine_mode, rtx) | |
1911 | = (frame_pointer_needed)? gen_frame_mem : gen_rtx_MEM; | |
1912 | ||
1913 | ||
1914 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
1915 | { | |
1916 | if (aarch64_register_saved_on_entry (regno)) | |
1917 | { | |
1918 | rtx mem; | |
1919 | mem = gen_mem_ref (DFmode, | |
1920 | plus_constant (Pmode, | |
1921 | base_rtx, | |
1922 | start_offset)); | |
1923 | ||
1924 | for (regno2 = regno + 1; | |
1925 | regno2 <= V31_REGNUM | |
1926 | && !aarch64_register_saved_on_entry (regno2); | |
1927 | regno2++) | |
1928 | { | |
1929 | /* Empty loop. */ | |
1930 | } | |
1931 | if (regno2 <= V31_REGNUM && | |
1932 | aarch64_register_saved_on_entry (regno2)) | |
1933 | { | |
1934 | rtx mem2; | |
1935 | /* Next highest register to be saved. */ | |
1936 | mem2 = gen_mem_ref (DFmode, | |
1937 | plus_constant | |
1938 | (Pmode, | |
1939 | base_rtx, | |
1940 | start_offset + increment)); | |
1941 | if (restore == false) | |
1942 | { | |
1943 | insn = emit_insn | |
1944 | ( gen_store_pairdf (mem, gen_rtx_REG (DFmode, regno), | |
1945 | mem2, gen_rtx_REG (DFmode, regno2))); | |
1946 | ||
1947 | } | |
1948 | else | |
1949 | { | |
1950 | insn = emit_insn | |
1951 | ( gen_load_pairdf (gen_rtx_REG (DFmode, regno), mem, | |
1952 | gen_rtx_REG (DFmode, regno2), mem2)); | |
1953 | ||
1954 | add_reg_note (insn, REG_CFA_RESTORE, | |
1955 | gen_rtx_REG (DFmode, regno)); | |
1956 | add_reg_note (insn, REG_CFA_RESTORE, | |
1957 | gen_rtx_REG (DFmode, regno2)); | |
1958 | } | |
1959 | ||
1960 | /* The first part of a frame-related parallel insn | |
1961 | is always assumed to be relevant to the frame | |
1962 | calculations; subsequent parts, are only | |
1963 | frame-related if explicitly marked. */ | |
1964 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; | |
1965 | regno = regno2; | |
1966 | start_offset += increment * 2; | |
1967 | } | |
1968 | else | |
1969 | { | |
1970 | if (restore == false) | |
1971 | insn = emit_move_insn (mem, gen_rtx_REG (DFmode, regno)); | |
1972 | else | |
1973 | { | |
1974 | insn = emit_move_insn (gen_rtx_REG (DFmode, regno), mem); | |
1975 | add_reg_note (insn, REG_CFA_RESTORE, | |
1976 | gen_rtx_REG (DImode, regno)); | |
1977 | } | |
1978 | start_offset += increment; | |
1979 | } | |
1980 | RTX_FRAME_RELATED_P (insn) = 1; | |
1981 | } | |
1982 | } | |
1983 | ||
1984 | } | |
1985 | ||
1986 | ||
1987 | /* offset from the stack pointer of where the saves and | |
1988 | restore's have to happen. */ | |
1989 | static void | |
1990 | aarch64_save_or_restore_callee_save_registers (HOST_WIDE_INT offset, | |
1991 | bool restore) | |
1992 | { | |
1993 | rtx insn; | |
1994 | rtx base_rtx = stack_pointer_rtx; | |
1995 | HOST_WIDE_INT start_offset = offset; | |
1996 | HOST_WIDE_INT increment = UNITS_PER_WORD; | |
1997 | rtx (*gen_mem_ref)(enum machine_mode, rtx) = (frame_pointer_needed)? gen_frame_mem : gen_rtx_MEM; | |
1998 | unsigned limit = (frame_pointer_needed)? R28_REGNUM: R30_REGNUM; | |
1999 | unsigned regno; | |
2000 | unsigned regno2; | |
2001 | ||
2002 | for (regno = R0_REGNUM; regno <= limit; regno++) | |
2003 | { | |
2004 | if (aarch64_register_saved_on_entry (regno)) | |
2005 | { | |
2006 | rtx mem; | |
2007 | mem = gen_mem_ref (Pmode, | |
2008 | plus_constant (Pmode, | |
2009 | base_rtx, | |
2010 | start_offset)); | |
2011 | ||
2012 | for (regno2 = regno + 1; | |
2013 | regno2 <= limit | |
2014 | && !aarch64_register_saved_on_entry (regno2); | |
2015 | regno2++) | |
2016 | { | |
2017 | /* Empty loop. */ | |
2018 | } | |
2019 | if (regno2 <= limit && | |
2020 | aarch64_register_saved_on_entry (regno2)) | |
2021 | { | |
2022 | rtx mem2; | |
2023 | /* Next highest register to be saved. */ | |
2024 | mem2 = gen_mem_ref (Pmode, | |
2025 | plus_constant | |
2026 | (Pmode, | |
2027 | base_rtx, | |
2028 | start_offset + increment)); | |
2029 | if (restore == false) | |
2030 | { | |
2031 | insn = emit_insn | |
2032 | ( gen_store_pairdi (mem, gen_rtx_REG (DImode, regno), | |
2033 | mem2, gen_rtx_REG (DImode, regno2))); | |
2034 | ||
2035 | } | |
2036 | else | |
2037 | { | |
2038 | insn = emit_insn | |
2039 | ( gen_load_pairdi (gen_rtx_REG (DImode, regno), mem, | |
2040 | gen_rtx_REG (DImode, regno2), mem2)); | |
2041 | ||
2042 | add_reg_note (insn, REG_CFA_RESTORE, gen_rtx_REG (DImode, regno)); | |
2043 | add_reg_note (insn, REG_CFA_RESTORE, gen_rtx_REG (DImode, regno2)); | |
2044 | } | |
2045 | ||
2046 | /* The first part of a frame-related parallel insn | |
2047 | is always assumed to be relevant to the frame | |
2048 | calculations; subsequent parts, are only | |
2049 | frame-related if explicitly marked. */ | |
2050 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, | |
2051 | 1)) = 1; | |
2052 | regno = regno2; | |
2053 | start_offset += increment * 2; | |
2054 | } | |
2055 | else | |
2056 | { | |
2057 | if (restore == false) | |
2058 | insn = emit_move_insn (mem, gen_rtx_REG (DImode, regno)); | |
2059 | else | |
2060 | { | |
2061 | insn = emit_move_insn (gen_rtx_REG (DImode, regno), mem); | |
2062 | add_reg_note (insn, REG_CFA_RESTORE, gen_rtx_REG (DImode, regno)); | |
2063 | } | |
2064 | start_offset += increment; | |
2065 | } | |
2066 | RTX_FRAME_RELATED_P (insn) = 1; | |
2067 | } | |
2068 | } | |
2069 | ||
2070 | aarch64_save_or_restore_fprs (start_offset, increment, restore, base_rtx); | |
2071 | ||
2072 | } | |
2073 | ||
2074 | /* AArch64 stack frames generated by this compiler look like: | |
2075 | ||
2076 | +-------------------------------+ | |
2077 | | | | |
2078 | | incoming stack arguments | | |
2079 | | | | |
2080 | +-------------------------------+ <-- arg_pointer_rtx | |
2081 | | | | |
2082 | | callee-allocated save area | | |
2083 | | for register varargs | | |
2084 | | | | |
2085 | +-------------------------------+ <-- frame_pointer_rtx | |
2086 | | | | |
2087 | | local variables | | |
2088 | | | | |
2089 | +-------------------------------+ | |
2090 | | padding0 | \ | |
2091 | +-------------------------------+ | | |
2092 | | | | | |
2093 | | | | | |
2094 | | callee-saved registers | | frame.saved_regs_size | |
2095 | | | | | |
2096 | +-------------------------------+ | | |
2097 | | LR' | | | |
2098 | +-------------------------------+ | | |
2099 | | FP' | / | |
2100 | P +-------------------------------+ <-- hard_frame_pointer_rtx | |
2101 | | dynamic allocation | | |
2102 | +-------------------------------+ | |
2103 | | | | |
2104 | | outgoing stack arguments | | |
2105 | | | | |
2106 | +-------------------------------+ <-- stack_pointer_rtx | |
2107 | ||
2108 | Dynamic stack allocations such as alloca insert data at point P. | |
2109 | They decrease stack_pointer_rtx but leave frame_pointer_rtx and | |
2110 | hard_frame_pointer_rtx unchanged. */ | |
2111 | ||
2112 | /* Generate the prologue instructions for entry into a function. | |
2113 | Establish the stack frame by decreasing the stack pointer with a | |
2114 | properly calculated size and, if necessary, create a frame record | |
2115 | filled with the values of LR and previous frame pointer. The | |
2116 | current FP is also set up if it is in use. */ | |
2117 | ||
2118 | void | |
2119 | aarch64_expand_prologue (void) | |
2120 | { | |
2121 | /* sub sp, sp, #<frame_size> | |
2122 | stp {fp, lr}, [sp, #<frame_size> - 16] | |
2123 | add fp, sp, #<frame_size> - hardfp_offset | |
2124 | stp {cs_reg}, [fp, #-16] etc. | |
2125 | ||
2126 | sub sp, sp, <final_adjustment_if_any> | |
2127 | */ | |
2128 | HOST_WIDE_INT original_frame_size; /* local variables + vararg save */ | |
2129 | HOST_WIDE_INT frame_size, offset; | |
2130 | HOST_WIDE_INT fp_offset; /* FP offset from SP */ | |
2131 | rtx insn; | |
2132 | ||
2133 | aarch64_layout_frame (); | |
2134 | original_frame_size = get_frame_size () + cfun->machine->saved_varargs_size; | |
2135 | gcc_assert ((!cfun->machine->saved_varargs_size || cfun->stdarg) | |
2136 | && (cfun->stdarg || !cfun->machine->saved_varargs_size)); | |
2137 | frame_size = (original_frame_size + cfun->machine->frame.saved_regs_size | |
2138 | + crtl->outgoing_args_size); | |
2139 | offset = frame_size = AARCH64_ROUND_UP (frame_size, | |
2140 | STACK_BOUNDARY / BITS_PER_UNIT); | |
2141 | ||
2142 | if (flag_stack_usage_info) | |
2143 | current_function_static_stack_size = frame_size; | |
2144 | ||
2145 | fp_offset = (offset | |
2146 | - original_frame_size | |
2147 | - cfun->machine->frame.saved_regs_size); | |
2148 | ||
2149 | /* Store pairs and load pairs have a range only -512 to 504. */ | |
2150 | if (offset >= 512) | |
2151 | { | |
2152 | /* When the frame has a large size, an initial decrease is done on | |
2153 | the stack pointer to jump over the callee-allocated save area for | |
2154 | register varargs, the local variable area and/or the callee-saved | |
2155 | register area. This will allow the pre-index write-back | |
2156 | store pair instructions to be used for setting up the stack frame | |
2157 | efficiently. */ | |
2158 | offset = original_frame_size + cfun->machine->frame.saved_regs_size; | |
2159 | if (offset >= 512) | |
2160 | offset = cfun->machine->frame.saved_regs_size; | |
2161 | ||
2162 | frame_size -= (offset + crtl->outgoing_args_size); | |
2163 | fp_offset = 0; | |
2164 | ||
2165 | if (frame_size >= 0x1000000) | |
2166 | { | |
2167 | rtx op0 = gen_rtx_REG (Pmode, IP0_REGNUM); | |
2168 | emit_move_insn (op0, GEN_INT (-frame_size)); | |
2169 | emit_insn (gen_add2_insn (stack_pointer_rtx, op0)); | |
2170 | aarch64_set_frame_expr (gen_rtx_SET | |
2171 | (Pmode, stack_pointer_rtx, | |
2172 | plus_constant (Pmode, | |
2173 | stack_pointer_rtx, | |
2174 | -frame_size))); | |
2175 | } | |
2176 | else if (frame_size > 0) | |
2177 | { | |
2178 | if ((frame_size & 0xfff) != frame_size) | |
2179 | { | |
2180 | insn = emit_insn (gen_add2_insn | |
2181 | (stack_pointer_rtx, | |
2182 | GEN_INT (-(frame_size | |
2183 | & ~(HOST_WIDE_INT)0xfff)))); | |
2184 | RTX_FRAME_RELATED_P (insn) = 1; | |
2185 | } | |
2186 | if ((frame_size & 0xfff) != 0) | |
2187 | { | |
2188 | insn = emit_insn (gen_add2_insn | |
2189 | (stack_pointer_rtx, | |
2190 | GEN_INT (-(frame_size | |
2191 | & (HOST_WIDE_INT)0xfff)))); | |
2192 | RTX_FRAME_RELATED_P (insn) = 1; | |
2193 | } | |
2194 | } | |
2195 | } | |
2196 | else | |
2197 | frame_size = -1; | |
2198 | ||
2199 | if (offset > 0) | |
2200 | { | |
2201 | /* Save the frame pointer and lr if the frame pointer is needed | |
2202 | first. Make the frame pointer point to the location of the | |
2203 | old frame pointer on the stack. */ | |
2204 | if (frame_pointer_needed) | |
2205 | { | |
2206 | rtx mem_fp, mem_lr; | |
2207 | ||
2208 | if (fp_offset) | |
2209 | { | |
2210 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, | |
2211 | GEN_INT (-offset))); | |
2212 | RTX_FRAME_RELATED_P (insn) = 1; | |
2213 | aarch64_set_frame_expr (gen_rtx_SET | |
2214 | (Pmode, stack_pointer_rtx, | |
2215 | gen_rtx_MINUS (Pmode, | |
2216 | stack_pointer_rtx, | |
2217 | GEN_INT (offset)))); | |
2218 | mem_fp = gen_frame_mem (DImode, | |
2219 | plus_constant (Pmode, | |
2220 | stack_pointer_rtx, | |
2221 | fp_offset)); | |
2222 | mem_lr = gen_frame_mem (DImode, | |
2223 | plus_constant (Pmode, | |
2224 | stack_pointer_rtx, | |
2225 | fp_offset | |
2226 | + UNITS_PER_WORD)); | |
2227 | insn = emit_insn (gen_store_pairdi (mem_fp, | |
2228 | hard_frame_pointer_rtx, | |
2229 | mem_lr, | |
2230 | gen_rtx_REG (DImode, | |
2231 | LR_REGNUM))); | |
2232 | } | |
2233 | else | |
2234 | { | |
2235 | insn = emit_insn (gen_storewb_pairdi_di | |
2236 | (stack_pointer_rtx, stack_pointer_rtx, | |
2237 | hard_frame_pointer_rtx, | |
2238 | gen_rtx_REG (DImode, LR_REGNUM), | |
2239 | GEN_INT (-offset), | |
2240 | GEN_INT (GET_MODE_SIZE (DImode) - offset))); | |
2241 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 2)) = 1; | |
2242 | } | |
2243 | ||
2244 | /* The first part of a frame-related parallel insn is always | |
2245 | assumed to be relevant to the frame calculations; | |
2246 | subsequent parts, are only frame-related if explicitly | |
2247 | marked. */ | |
2248 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; | |
2249 | RTX_FRAME_RELATED_P (insn) = 1; | |
2250 | ||
2251 | /* Set up frame pointer to point to the location of the | |
2252 | previous frame pointer on the stack. */ | |
2253 | insn = emit_insn (gen_add3_insn (hard_frame_pointer_rtx, | |
2254 | stack_pointer_rtx, | |
2255 | GEN_INT (fp_offset))); | |
2256 | aarch64_set_frame_expr (gen_rtx_SET | |
2257 | (Pmode, hard_frame_pointer_rtx, | |
2258 | plus_constant (Pmode, | |
2259 | stack_pointer_rtx, | |
2260 | fp_offset))); | |
2261 | RTX_FRAME_RELATED_P (insn) = 1; | |
2262 | insn = emit_insn (gen_stack_tie (stack_pointer_rtx, | |
2263 | hard_frame_pointer_rtx)); | |
2264 | } | |
2265 | else | |
2266 | { | |
2267 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, | |
2268 | GEN_INT (-offset))); | |
2269 | RTX_FRAME_RELATED_P (insn) = 1; | |
2270 | } | |
2271 | ||
2272 | aarch64_save_or_restore_callee_save_registers | |
2273 | (fp_offset + cfun->machine->frame.hardfp_offset, 0); | |
2274 | } | |
2275 | ||
2276 | /* when offset >= 512, | |
2277 | sub sp, sp, #<outgoing_args_size> */ | |
2278 | if (frame_size > -1) | |
2279 | { | |
2280 | if (crtl->outgoing_args_size > 0) | |
2281 | { | |
2282 | insn = emit_insn (gen_add2_insn | |
2283 | (stack_pointer_rtx, | |
2284 | GEN_INT (- crtl->outgoing_args_size))); | |
2285 | RTX_FRAME_RELATED_P (insn) = 1; | |
2286 | } | |
2287 | } | |
2288 | } | |
2289 | ||
2290 | /* Generate the epilogue instructions for returning from a function. */ | |
2291 | void | |
2292 | aarch64_expand_epilogue (bool for_sibcall) | |
2293 | { | |
2294 | HOST_WIDE_INT original_frame_size, frame_size, offset; | |
2295 | HOST_WIDE_INT fp_offset; | |
2296 | rtx insn; | |
2297 | rtx cfa_reg; | |
2298 | ||
2299 | aarch64_layout_frame (); | |
2300 | original_frame_size = get_frame_size () + cfun->machine->saved_varargs_size; | |
2301 | frame_size = (original_frame_size + cfun->machine->frame.saved_regs_size | |
2302 | + crtl->outgoing_args_size); | |
2303 | offset = frame_size = AARCH64_ROUND_UP (frame_size, | |
2304 | STACK_BOUNDARY / BITS_PER_UNIT); | |
2305 | ||
2306 | fp_offset = (offset | |
2307 | - original_frame_size | |
2308 | - cfun->machine->frame.saved_regs_size); | |
2309 | ||
2310 | cfa_reg = frame_pointer_needed ? hard_frame_pointer_rtx : stack_pointer_rtx; | |
2311 | ||
2312 | /* Store pairs and load pairs have a range only -512 to 504. */ | |
2313 | if (offset >= 512) | |
2314 | { | |
2315 | offset = original_frame_size + cfun->machine->frame.saved_regs_size; | |
2316 | if (offset >= 512) | |
2317 | offset = cfun->machine->frame.saved_regs_size; | |
2318 | ||
2319 | frame_size -= (offset + crtl->outgoing_args_size); | |
2320 | fp_offset = 0; | |
2321 | if (!frame_pointer_needed && crtl->outgoing_args_size > 0) | |
2322 | { | |
2323 | insn = emit_insn (gen_add2_insn | |
2324 | (stack_pointer_rtx, | |
2325 | GEN_INT (crtl->outgoing_args_size))); | |
2326 | RTX_FRAME_RELATED_P (insn) = 1; | |
2327 | } | |
2328 | } | |
2329 | else | |
2330 | frame_size = -1; | |
2331 | ||
2332 | /* If there were outgoing arguments or we've done dynamic stack | |
2333 | allocation, then restore the stack pointer from the frame | |
2334 | pointer. This is at most one insn and more efficient than using | |
2335 | GCC's internal mechanism. */ | |
2336 | if (frame_pointer_needed | |
2337 | && (crtl->outgoing_args_size || cfun->calls_alloca)) | |
2338 | { | |
2339 | insn = emit_insn (gen_add3_insn (stack_pointer_rtx, | |
2340 | hard_frame_pointer_rtx, | |
2341 | GEN_INT (- fp_offset))); | |
2342 | RTX_FRAME_RELATED_P (insn) = 1; | |
2343 | /* As SP is set to (FP - fp_offset), according to the rules in | |
2344 | dwarf2cfi.c:dwarf2out_frame_debug_expr, CFA should be calculated | |
2345 | from the value of SP from now on. */ | |
2346 | cfa_reg = stack_pointer_rtx; | |
2347 | } | |
2348 | ||
2349 | aarch64_save_or_restore_callee_save_registers | |
2350 | (fp_offset + cfun->machine->frame.hardfp_offset, 1); | |
2351 | ||
2352 | /* Restore the frame pointer and lr if the frame pointer is needed. */ | |
2353 | if (offset > 0) | |
2354 | { | |
2355 | if (frame_pointer_needed) | |
2356 | { | |
2357 | rtx mem_fp, mem_lr; | |
2358 | ||
2359 | if (fp_offset) | |
2360 | { | |
2361 | mem_fp = gen_frame_mem (DImode, | |
2362 | plus_constant (Pmode, | |
2363 | stack_pointer_rtx, | |
2364 | fp_offset)); | |
2365 | mem_lr = gen_frame_mem (DImode, | |
2366 | plus_constant (Pmode, | |
2367 | stack_pointer_rtx, | |
2368 | fp_offset | |
2369 | + UNITS_PER_WORD)); | |
2370 | insn = emit_insn (gen_load_pairdi (hard_frame_pointer_rtx, | |
2371 | mem_fp, | |
2372 | gen_rtx_REG (DImode, | |
2373 | LR_REGNUM), | |
2374 | mem_lr)); | |
2375 | } | |
2376 | else | |
2377 | { | |
2378 | insn = emit_insn (gen_loadwb_pairdi_di | |
2379 | (stack_pointer_rtx, | |
2380 | stack_pointer_rtx, | |
2381 | hard_frame_pointer_rtx, | |
2382 | gen_rtx_REG (DImode, LR_REGNUM), | |
2383 | GEN_INT (offset), | |
2384 | GEN_INT (GET_MODE_SIZE (DImode) + offset))); | |
2385 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 2)) = 1; | |
2386 | add_reg_note (insn, REG_CFA_ADJUST_CFA, | |
2387 | (gen_rtx_SET (Pmode, stack_pointer_rtx, | |
2388 | plus_constant (Pmode, cfa_reg, | |
2389 | offset)))); | |
2390 | } | |
2391 | ||
2392 | /* The first part of a frame-related parallel insn | |
2393 | is always assumed to be relevant to the frame | |
2394 | calculations; subsequent parts, are only | |
2395 | frame-related if explicitly marked. */ | |
2396 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; | |
2397 | RTX_FRAME_RELATED_P (insn) = 1; | |
2398 | add_reg_note (insn, REG_CFA_RESTORE, hard_frame_pointer_rtx); | |
2399 | add_reg_note (insn, REG_CFA_RESTORE, | |
2400 | gen_rtx_REG (DImode, LR_REGNUM)); | |
2401 | ||
2402 | if (fp_offset) | |
2403 | { | |
2404 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, | |
2405 | GEN_INT (offset))); | |
2406 | RTX_FRAME_RELATED_P (insn) = 1; | |
2407 | } | |
2408 | } | |
2409 | else | |
2410 | { | |
2411 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, | |
2412 | GEN_INT (offset))); | |
2413 | RTX_FRAME_RELATED_P (insn) = 1; | |
2414 | } | |
2415 | } | |
2416 | ||
2417 | /* Stack adjustment for exception handler. */ | |
2418 | if (crtl->calls_eh_return) | |
2419 | { | |
2420 | /* We need to unwind the stack by the offset computed by | |
2421 | EH_RETURN_STACKADJ_RTX. However, at this point the CFA is | |
2422 | based on SP. Ideally we would update the SP and define the | |
2423 | CFA along the lines of: | |
2424 | ||
2425 | SP = SP + EH_RETURN_STACKADJ_RTX | |
2426 | (regnote CFA = SP - EH_RETURN_STACKADJ_RTX) | |
2427 | ||
2428 | However the dwarf emitter only understands a constant | |
2429 | register offset. | |
2430 | ||
2431 | The solution chosen here is to use the otherwise unused IP0 | |
2432 | as a temporary register to hold the current SP value. The | |
2433 | CFA is described using IP0 then SP is modified. */ | |
2434 | ||
2435 | rtx ip0 = gen_rtx_REG (DImode, IP0_REGNUM); | |
2436 | ||
2437 | insn = emit_move_insn (ip0, stack_pointer_rtx); | |
2438 | add_reg_note (insn, REG_CFA_DEF_CFA, ip0); | |
2439 | RTX_FRAME_RELATED_P (insn) = 1; | |
2440 | ||
2441 | emit_insn (gen_add2_insn (stack_pointer_rtx, EH_RETURN_STACKADJ_RTX)); | |
2442 | ||
2443 | /* Ensure the assignment to IP0 does not get optimized away. */ | |
2444 | emit_use (ip0); | |
2445 | } | |
2446 | ||
2447 | if (frame_size > -1) | |
2448 | { | |
2449 | if (frame_size >= 0x1000000) | |
2450 | { | |
2451 | rtx op0 = gen_rtx_REG (Pmode, IP0_REGNUM); | |
2452 | emit_move_insn (op0, GEN_INT (frame_size)); | |
2453 | emit_insn (gen_add2_insn (stack_pointer_rtx, op0)); | |
2454 | aarch64_set_frame_expr (gen_rtx_SET | |
2455 | (Pmode, stack_pointer_rtx, | |
2456 | plus_constant (Pmode, | |
2457 | stack_pointer_rtx, | |
2458 | frame_size))); | |
2459 | } | |
2460 | else if (frame_size > 0) | |
2461 | { | |
2462 | if ((frame_size & 0xfff) != 0) | |
2463 | { | |
2464 | insn = emit_insn (gen_add2_insn | |
2465 | (stack_pointer_rtx, | |
2466 | GEN_INT ((frame_size | |
2467 | & (HOST_WIDE_INT) 0xfff)))); | |
2468 | RTX_FRAME_RELATED_P (insn) = 1; | |
2469 | } | |
2470 | if ((frame_size & 0xfff) != frame_size) | |
2471 | { | |
2472 | insn = emit_insn (gen_add2_insn | |
2473 | (stack_pointer_rtx, | |
2474 | GEN_INT ((frame_size | |
2475 | & ~ (HOST_WIDE_INT) 0xfff)))); | |
2476 | RTX_FRAME_RELATED_P (insn) = 1; | |
2477 | } | |
2478 | } | |
2479 | ||
2480 | aarch64_set_frame_expr (gen_rtx_SET (Pmode, stack_pointer_rtx, | |
2481 | plus_constant (Pmode, | |
2482 | stack_pointer_rtx, | |
2483 | offset))); | |
2484 | } | |
2485 | ||
2486 | emit_use (gen_rtx_REG (DImode, LR_REGNUM)); | |
2487 | if (!for_sibcall) | |
2488 | emit_jump_insn (ret_rtx); | |
2489 | } | |
2490 | ||
2491 | /* Return the place to copy the exception unwinding return address to. | |
2492 | This will probably be a stack slot, but could (in theory be the | |
2493 | return register). */ | |
2494 | rtx | |
2495 | aarch64_final_eh_return_addr (void) | |
2496 | { | |
2497 | HOST_WIDE_INT original_frame_size, frame_size, offset, fp_offset; | |
2498 | aarch64_layout_frame (); | |
2499 | original_frame_size = get_frame_size () + cfun->machine->saved_varargs_size; | |
2500 | frame_size = (original_frame_size + cfun->machine->frame.saved_regs_size | |
2501 | + crtl->outgoing_args_size); | |
2502 | offset = frame_size = AARCH64_ROUND_UP (frame_size, | |
2503 | STACK_BOUNDARY / BITS_PER_UNIT); | |
2504 | fp_offset = offset | |
2505 | - original_frame_size | |
2506 | - cfun->machine->frame.saved_regs_size; | |
2507 | ||
2508 | if (cfun->machine->frame.reg_offset[LR_REGNUM] < 0) | |
2509 | return gen_rtx_REG (DImode, LR_REGNUM); | |
2510 | ||
2511 | /* DSE and CSELIB do not detect an alias between sp+k1 and fp+k2. This can | |
2512 | result in a store to save LR introduced by builtin_eh_return () being | |
2513 | incorrectly deleted because the alias is not detected. | |
2514 | So in the calculation of the address to copy the exception unwinding | |
2515 | return address to, we note 2 cases. | |
2516 | If FP is needed and the fp_offset is 0, it means that SP = FP and hence | |
2517 | we return a SP-relative location since all the addresses are SP-relative | |
2518 | in this case. This prevents the store from being optimized away. | |
2519 | If the fp_offset is not 0, then the addresses will be FP-relative and | |
2520 | therefore we return a FP-relative location. */ | |
2521 | ||
2522 | if (frame_pointer_needed) | |
2523 | { | |
2524 | if (fp_offset) | |
2525 | return gen_frame_mem (DImode, | |
2526 | plus_constant (Pmode, hard_frame_pointer_rtx, UNITS_PER_WORD)); | |
2527 | else | |
2528 | return gen_frame_mem (DImode, | |
2529 | plus_constant (Pmode, stack_pointer_rtx, UNITS_PER_WORD)); | |
2530 | } | |
2531 | ||
2532 | /* If FP is not needed, we calculate the location of LR, which would be | |
2533 | at the top of the saved registers block. */ | |
2534 | ||
2535 | return gen_frame_mem (DImode, | |
2536 | plus_constant (Pmode, | |
2537 | stack_pointer_rtx, | |
2538 | fp_offset | |
2539 | + cfun->machine->frame.saved_regs_size | |
2540 | - 2 * UNITS_PER_WORD)); | |
2541 | } | |
2542 | ||
2543 | /* Possibly output code to build up a constant in a register. For | |
2544 | the benefit of the costs infrastructure, returns the number of | |
2545 | instructions which would be emitted. GENERATE inhibits or | |
2546 | enables code generation. */ | |
2547 | ||
2548 | static int | |
2549 | aarch64_build_constant (int regnum, HOST_WIDE_INT val, bool generate) | |
2550 | { | |
2551 | int insns = 0; | |
2552 | ||
2553 | if (aarch64_bitmask_imm (val, DImode)) | |
2554 | { | |
2555 | if (generate) | |
2556 | emit_move_insn (gen_rtx_REG (Pmode, regnum), GEN_INT (val)); | |
2557 | insns = 1; | |
2558 | } | |
2559 | else | |
2560 | { | |
2561 | int i; | |
2562 | int ncount = 0; | |
2563 | int zcount = 0; | |
2564 | HOST_WIDE_INT valp = val >> 16; | |
2565 | HOST_WIDE_INT valm; | |
2566 | HOST_WIDE_INT tval; | |
2567 | ||
2568 | for (i = 16; i < 64; i += 16) | |
2569 | { | |
2570 | valm = (valp & 0xffff); | |
2571 | ||
2572 | if (valm != 0) | |
2573 | ++ zcount; | |
2574 | ||
2575 | if (valm != 0xffff) | |
2576 | ++ ncount; | |
2577 | ||
2578 | valp >>= 16; | |
2579 | } | |
2580 | ||
2581 | /* zcount contains the number of additional MOVK instructions | |
2582 | required if the constant is built up with an initial MOVZ instruction, | |
2583 | while ncount is the number of MOVK instructions required if starting | |
2584 | with a MOVN instruction. Choose the sequence that yields the fewest | |
2585 | number of instructions, preferring MOVZ instructions when they are both | |
2586 | the same. */ | |
2587 | if (ncount < zcount) | |
2588 | { | |
2589 | if (generate) | |
2590 | emit_move_insn (gen_rtx_REG (Pmode, regnum), | |
2591 | GEN_INT (val | ~(HOST_WIDE_INT) 0xffff)); | |
2592 | tval = 0xffff; | |
2593 | insns++; | |
2594 | } | |
2595 | else | |
2596 | { | |
2597 | if (generate) | |
2598 | emit_move_insn (gen_rtx_REG (Pmode, regnum), | |
2599 | GEN_INT (val & 0xffff)); | |
2600 | tval = 0; | |
2601 | insns++; | |
2602 | } | |
2603 | ||
2604 | val >>= 16; | |
2605 | ||
2606 | for (i = 16; i < 64; i += 16) | |
2607 | { | |
2608 | if ((val & 0xffff) != tval) | |
2609 | { | |
2610 | if (generate) | |
2611 | emit_insn (gen_insv_immdi (gen_rtx_REG (Pmode, regnum), | |
2612 | GEN_INT (i), | |
2613 | GEN_INT (val & 0xffff))); | |
2614 | insns++; | |
2615 | } | |
2616 | val >>= 16; | |
2617 | } | |
2618 | } | |
2619 | return insns; | |
2620 | } | |
2621 | ||
2622 | static void | |
2623 | aarch64_add_constant (int regnum, int scratchreg, HOST_WIDE_INT delta) | |
2624 | { | |
2625 | HOST_WIDE_INT mdelta = delta; | |
2626 | rtx this_rtx = gen_rtx_REG (Pmode, regnum); | |
2627 | rtx scratch_rtx = gen_rtx_REG (Pmode, scratchreg); | |
2628 | ||
2629 | if (mdelta < 0) | |
2630 | mdelta = -mdelta; | |
2631 | ||
2632 | if (mdelta >= 4096 * 4096) | |
2633 | { | |
2634 | (void) aarch64_build_constant (scratchreg, delta, true); | |
2635 | emit_insn (gen_add3_insn (this_rtx, this_rtx, scratch_rtx)); | |
2636 | } | |
2637 | else if (mdelta > 0) | |
2638 | { | |
2639 | if (mdelta >= 4096) | |
2640 | { | |
2641 | emit_insn (gen_rtx_SET (Pmode, scratch_rtx, GEN_INT (mdelta / 4096))); | |
2642 | rtx shift = gen_rtx_ASHIFT (Pmode, scratch_rtx, GEN_INT (12)); | |
2643 | if (delta < 0) | |
2644 | emit_insn (gen_rtx_SET (Pmode, this_rtx, | |
2645 | gen_rtx_MINUS (Pmode, this_rtx, shift))); | |
2646 | else | |
2647 | emit_insn (gen_rtx_SET (Pmode, this_rtx, | |
2648 | gen_rtx_PLUS (Pmode, this_rtx, shift))); | |
2649 | } | |
2650 | if (mdelta % 4096 != 0) | |
2651 | { | |
2652 | scratch_rtx = GEN_INT ((delta < 0 ? -1 : 1) * (mdelta % 4096)); | |
2653 | emit_insn (gen_rtx_SET (Pmode, this_rtx, | |
2654 | gen_rtx_PLUS (Pmode, this_rtx, scratch_rtx))); | |
2655 | } | |
2656 | } | |
2657 | } | |
2658 | ||
2659 | /* Output code to add DELTA to the first argument, and then jump | |
2660 | to FUNCTION. Used for C++ multiple inheritance. */ | |
2661 | static void | |
2662 | aarch64_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED, | |
2663 | HOST_WIDE_INT delta, | |
2664 | HOST_WIDE_INT vcall_offset, | |
2665 | tree function) | |
2666 | { | |
2667 | /* The this pointer is always in x0. Note that this differs from | |
2668 | Arm where the this pointer maybe bumped to r1 if r0 is required | |
2669 | to return a pointer to an aggregate. On AArch64 a result value | |
2670 | pointer will be in x8. */ | |
2671 | int this_regno = R0_REGNUM; | |
2672 | rtx this_rtx, temp0, temp1, addr, insn, funexp; | |
2673 | ||
2674 | reload_completed = 1; | |
2675 | emit_note (NOTE_INSN_PROLOGUE_END); | |
2676 | ||
2677 | if (vcall_offset == 0) | |
2678 | aarch64_add_constant (this_regno, IP1_REGNUM, delta); | |
2679 | else | |
2680 | { | |
2681 | gcc_assert ((vcall_offset & (POINTER_BYTES - 1)) == 0); | |
2682 | ||
2683 | this_rtx = gen_rtx_REG (Pmode, this_regno); | |
2684 | temp0 = gen_rtx_REG (Pmode, IP0_REGNUM); | |
2685 | temp1 = gen_rtx_REG (Pmode, IP1_REGNUM); | |
2686 | ||
2687 | addr = this_rtx; | |
2688 | if (delta != 0) | |
2689 | { | |
2690 | if (delta >= -256 && delta < 256) | |
2691 | addr = gen_rtx_PRE_MODIFY (Pmode, this_rtx, | |
2692 | plus_constant (Pmode, this_rtx, delta)); | |
2693 | else | |
2694 | aarch64_add_constant (this_regno, IP1_REGNUM, delta); | |
2695 | } | |
2696 | ||
2697 | if (Pmode == ptr_mode) | |
2698 | aarch64_emit_move (temp0, gen_rtx_MEM (ptr_mode, addr)); | |
2699 | else | |
2700 | aarch64_emit_move (temp0, | |
2701 | gen_rtx_ZERO_EXTEND (Pmode, | |
2702 | gen_rtx_MEM (ptr_mode, addr))); | |
2703 | ||
2704 | if (vcall_offset >= -256 && vcall_offset < 4096 * POINTER_BYTES) | |
2705 | addr = plus_constant (Pmode, temp0, vcall_offset); | |
2706 | else | |
2707 | { | |
2708 | (void) aarch64_build_constant (IP1_REGNUM, vcall_offset, true); | |
2709 | addr = gen_rtx_PLUS (Pmode, temp0, temp1); | |
2710 | } | |
2711 | ||
2712 | if (Pmode == ptr_mode) | |
2713 | aarch64_emit_move (temp1, gen_rtx_MEM (ptr_mode,addr)); | |
2714 | else | |
2715 | aarch64_emit_move (temp1, | |
2716 | gen_rtx_SIGN_EXTEND (Pmode, | |
2717 | gen_rtx_MEM (ptr_mode, addr))); | |
2718 | ||
2719 | emit_insn (gen_add2_insn (this_rtx, temp1)); | |
2720 | } | |
2721 | ||
2722 | /* Generate a tail call to the target function. */ | |
2723 | if (!TREE_USED (function)) | |
2724 | { | |
2725 | assemble_external (function); | |
2726 | TREE_USED (function) = 1; | |
2727 | } | |
2728 | funexp = XEXP (DECL_RTL (function), 0); | |
2729 | funexp = gen_rtx_MEM (FUNCTION_MODE, funexp); | |
2730 | insn = emit_call_insn (gen_sibcall (funexp, const0_rtx, NULL_RTX)); | |
2731 | SIBLING_CALL_P (insn) = 1; | |
2732 | ||
2733 | insn = get_insns (); | |
2734 | shorten_branches (insn); | |
2735 | final_start_function (insn, file, 1); | |
2736 | final (insn, file, 1); | |
2737 | final_end_function (); | |
2738 | ||
2739 | /* Stop pretending to be a post-reload pass. */ | |
2740 | reload_completed = 0; | |
2741 | } | |
2742 | ||
2743 | static int | |
2744 | aarch64_tls_operand_p_1 (rtx *x, void *data ATTRIBUTE_UNUSED) | |
2745 | { | |
2746 | if (GET_CODE (*x) == SYMBOL_REF) | |
2747 | return SYMBOL_REF_TLS_MODEL (*x) != 0; | |
2748 | ||
2749 | /* Don't recurse into UNSPEC_TLS looking for TLS symbols; these are | |
2750 | TLS offsets, not real symbol references. */ | |
2751 | if (GET_CODE (*x) == UNSPEC | |
2752 | && XINT (*x, 1) == UNSPEC_TLS) | |
2753 | return -1; | |
2754 | ||
2755 | return 0; | |
2756 | } | |
2757 | ||
2758 | static bool | |
2759 | aarch64_tls_referenced_p (rtx x) | |
2760 | { | |
2761 | if (!TARGET_HAVE_TLS) | |
2762 | return false; | |
2763 | ||
2764 | return for_each_rtx (&x, aarch64_tls_operand_p_1, NULL); | |
2765 | } | |
2766 | ||
2767 | ||
2768 | static int | |
2769 | aarch64_bitmasks_cmp (const void *i1, const void *i2) | |
2770 | { | |
2771 | const unsigned HOST_WIDE_INT *imm1 = (const unsigned HOST_WIDE_INT *) i1; | |
2772 | const unsigned HOST_WIDE_INT *imm2 = (const unsigned HOST_WIDE_INT *) i2; | |
2773 | ||
2774 | if (*imm1 < *imm2) | |
2775 | return -1; | |
2776 | if (*imm1 > *imm2) | |
2777 | return +1; | |
2778 | return 0; | |
2779 | } | |
2780 | ||
2781 | ||
2782 | static void | |
2783 | aarch64_build_bitmask_table (void) | |
2784 | { | |
2785 | unsigned HOST_WIDE_INT mask, imm; | |
2786 | unsigned int log_e, e, s, r; | |
2787 | unsigned int nimms = 0; | |
2788 | ||
2789 | for (log_e = 1; log_e <= 6; log_e++) | |
2790 | { | |
2791 | e = 1 << log_e; | |
2792 | if (e == 64) | |
2793 | mask = ~(HOST_WIDE_INT) 0; | |
2794 | else | |
2795 | mask = ((HOST_WIDE_INT) 1 << e) - 1; | |
2796 | for (s = 1; s < e; s++) | |
2797 | { | |
2798 | for (r = 0; r < e; r++) | |
2799 | { | |
2800 | /* set s consecutive bits to 1 (s < 64) */ | |
2801 | imm = ((unsigned HOST_WIDE_INT)1 << s) - 1; | |
2802 | /* rotate right by r */ | |
2803 | if (r != 0) | |
2804 | imm = ((imm >> r) | (imm << (e - r))) & mask; | |
2805 | /* replicate the constant depending on SIMD size */ | |
2806 | switch (log_e) { | |
2807 | case 1: imm |= (imm << 2); | |
2808 | case 2: imm |= (imm << 4); | |
2809 | case 3: imm |= (imm << 8); | |
2810 | case 4: imm |= (imm << 16); | |
2811 | case 5: imm |= (imm << 32); | |
2812 | case 6: | |
2813 | break; | |
2814 | default: | |
2815 | gcc_unreachable (); | |
2816 | } | |
2817 | gcc_assert (nimms < AARCH64_NUM_BITMASKS); | |
2818 | aarch64_bitmasks[nimms++] = imm; | |
2819 | } | |
2820 | } | |
2821 | } | |
2822 | ||
2823 | gcc_assert (nimms == AARCH64_NUM_BITMASKS); | |
2824 | qsort (aarch64_bitmasks, nimms, sizeof (aarch64_bitmasks[0]), | |
2825 | aarch64_bitmasks_cmp); | |
2826 | } | |
2827 | ||
2828 | ||
2829 | /* Return true if val can be encoded as a 12-bit unsigned immediate with | |
2830 | a left shift of 0 or 12 bits. */ | |
2831 | bool | |
2832 | aarch64_uimm12_shift (HOST_WIDE_INT val) | |
2833 | { | |
2834 | return ((val & (((HOST_WIDE_INT) 0xfff) << 0)) == val | |
2835 | || (val & (((HOST_WIDE_INT) 0xfff) << 12)) == val | |
2836 | ); | |
2837 | } | |
2838 | ||
2839 | ||
2840 | /* Return true if val is an immediate that can be loaded into a | |
2841 | register by a MOVZ instruction. */ | |
2842 | static bool | |
2843 | aarch64_movw_imm (HOST_WIDE_INT val, enum machine_mode mode) | |
2844 | { | |
2845 | if (GET_MODE_SIZE (mode) > 4) | |
2846 | { | |
2847 | if ((val & (((HOST_WIDE_INT) 0xffff) << 32)) == val | |
2848 | || (val & (((HOST_WIDE_INT) 0xffff) << 48)) == val) | |
2849 | return 1; | |
2850 | } | |
2851 | else | |
2852 | { | |
2853 | /* Ignore sign extension. */ | |
2854 | val &= (HOST_WIDE_INT) 0xffffffff; | |
2855 | } | |
2856 | return ((val & (((HOST_WIDE_INT) 0xffff) << 0)) == val | |
2857 | || (val & (((HOST_WIDE_INT) 0xffff) << 16)) == val); | |
2858 | } | |
2859 | ||
2860 | ||
2861 | /* Return true if val is a valid bitmask immediate. */ | |
2862 | bool | |
2863 | aarch64_bitmask_imm (HOST_WIDE_INT val, enum machine_mode mode) | |
2864 | { | |
2865 | if (GET_MODE_SIZE (mode) < 8) | |
2866 | { | |
2867 | /* Replicate bit pattern. */ | |
2868 | val &= (HOST_WIDE_INT) 0xffffffff; | |
2869 | val |= val << 32; | |
2870 | } | |
2871 | return bsearch (&val, aarch64_bitmasks, AARCH64_NUM_BITMASKS, | |
2872 | sizeof (aarch64_bitmasks[0]), aarch64_bitmasks_cmp) != NULL; | |
2873 | } | |
2874 | ||
2875 | ||
2876 | /* Return true if val is an immediate that can be loaded into a | |
2877 | register in a single instruction. */ | |
2878 | bool | |
2879 | aarch64_move_imm (HOST_WIDE_INT val, enum machine_mode mode) | |
2880 | { | |
2881 | if (aarch64_movw_imm (val, mode) || aarch64_movw_imm (~val, mode)) | |
2882 | return 1; | |
2883 | return aarch64_bitmask_imm (val, mode); | |
2884 | } | |
2885 | ||
2886 | static bool | |
2887 | aarch64_cannot_force_const_mem (enum machine_mode mode ATTRIBUTE_UNUSED, rtx x) | |
2888 | { | |
2889 | rtx base, offset; | |
2890 | ||
2891 | if (GET_CODE (x) == HIGH) | |
2892 | return true; | |
2893 | ||
2894 | split_const (x, &base, &offset); | |
2895 | if (GET_CODE (base) == SYMBOL_REF || GET_CODE (base) == LABEL_REF) | |
2896 | { | |
2897 | if (aarch64_classify_symbol (base, SYMBOL_CONTEXT_ADR) | |
2898 | != SYMBOL_FORCE_TO_MEM) | |
2899 | return true; | |
2900 | else | |
2901 | /* Avoid generating a 64-bit relocation in ILP32; leave | |
2902 | to aarch64_expand_mov_immediate to handle it properly. */ | |
2903 | return mode != ptr_mode; | |
2904 | } | |
2905 | ||
2906 | return aarch64_tls_referenced_p (x); | |
2907 | } | |
2908 | ||
2909 | /* Return true if register REGNO is a valid index register. | |
2910 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
2911 | ||
2912 | bool | |
2913 | aarch64_regno_ok_for_index_p (int regno, bool strict_p) | |
2914 | { | |
2915 | if (!HARD_REGISTER_NUM_P (regno)) | |
2916 | { | |
2917 | if (!strict_p) | |
2918 | return true; | |
2919 | ||
2920 | if (!reg_renumber) | |
2921 | return false; | |
2922 | ||
2923 | regno = reg_renumber[regno]; | |
2924 | } | |
2925 | return GP_REGNUM_P (regno); | |
2926 | } | |
2927 | ||
2928 | /* Return true if register REGNO is a valid base register for mode MODE. | |
2929 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
2930 | ||
2931 | bool | |
2932 | aarch64_regno_ok_for_base_p (int regno, bool strict_p) | |
2933 | { | |
2934 | if (!HARD_REGISTER_NUM_P (regno)) | |
2935 | { | |
2936 | if (!strict_p) | |
2937 | return true; | |
2938 | ||
2939 | if (!reg_renumber) | |
2940 | return false; | |
2941 | ||
2942 | regno = reg_renumber[regno]; | |
2943 | } | |
2944 | ||
2945 | /* The fake registers will be eliminated to either the stack or | |
2946 | hard frame pointer, both of which are usually valid base registers. | |
2947 | Reload deals with the cases where the eliminated form isn't valid. */ | |
2948 | return (GP_REGNUM_P (regno) | |
2949 | || regno == SP_REGNUM | |
2950 | || regno == FRAME_POINTER_REGNUM | |
2951 | || regno == ARG_POINTER_REGNUM); | |
2952 | } | |
2953 | ||
2954 | /* Return true if X is a valid base register for mode MODE. | |
2955 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
2956 | ||
2957 | static bool | |
2958 | aarch64_base_register_rtx_p (rtx x, bool strict_p) | |
2959 | { | |
2960 | if (!strict_p && GET_CODE (x) == SUBREG) | |
2961 | x = SUBREG_REG (x); | |
2962 | ||
2963 | return (REG_P (x) && aarch64_regno_ok_for_base_p (REGNO (x), strict_p)); | |
2964 | } | |
2965 | ||
2966 | /* Return true if address offset is a valid index. If it is, fill in INFO | |
2967 | appropriately. STRICT_P is true if REG_OK_STRICT is in effect. */ | |
2968 | ||
2969 | static bool | |
2970 | aarch64_classify_index (struct aarch64_address_info *info, rtx x, | |
2971 | enum machine_mode mode, bool strict_p) | |
2972 | { | |
2973 | enum aarch64_address_type type; | |
2974 | rtx index; | |
2975 | int shift; | |
2976 | ||
2977 | /* (reg:P) */ | |
2978 | if ((REG_P (x) || GET_CODE (x) == SUBREG) | |
2979 | && GET_MODE (x) == Pmode) | |
2980 | { | |
2981 | type = ADDRESS_REG_REG; | |
2982 | index = x; | |
2983 | shift = 0; | |
2984 | } | |
2985 | /* (sign_extend:DI (reg:SI)) */ | |
2986 | else if ((GET_CODE (x) == SIGN_EXTEND | |
2987 | || GET_CODE (x) == ZERO_EXTEND) | |
2988 | && GET_MODE (x) == DImode | |
2989 | && GET_MODE (XEXP (x, 0)) == SImode) | |
2990 | { | |
2991 | type = (GET_CODE (x) == SIGN_EXTEND) | |
2992 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
2993 | index = XEXP (x, 0); | |
2994 | shift = 0; | |
2995 | } | |
2996 | /* (mult:DI (sign_extend:DI (reg:SI)) (const_int scale)) */ | |
2997 | else if (GET_CODE (x) == MULT | |
2998 | && (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
2999 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
3000 | && GET_MODE (XEXP (x, 0)) == DImode | |
3001 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == SImode | |
3002 | && CONST_INT_P (XEXP (x, 1))) | |
3003 | { | |
3004 | type = (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
3005 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3006 | index = XEXP (XEXP (x, 0), 0); | |
3007 | shift = exact_log2 (INTVAL (XEXP (x, 1))); | |
3008 | } | |
3009 | /* (ashift:DI (sign_extend:DI (reg:SI)) (const_int shift)) */ | |
3010 | else if (GET_CODE (x) == ASHIFT | |
3011 | && (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
3012 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
3013 | && GET_MODE (XEXP (x, 0)) == DImode | |
3014 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == SImode | |
3015 | && CONST_INT_P (XEXP (x, 1))) | |
3016 | { | |
3017 | type = (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
3018 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3019 | index = XEXP (XEXP (x, 0), 0); | |
3020 | shift = INTVAL (XEXP (x, 1)); | |
3021 | } | |
3022 | /* (sign_extract:DI (mult:DI (reg:DI) (const_int scale)) 32+shift 0) */ | |
3023 | else if ((GET_CODE (x) == SIGN_EXTRACT | |
3024 | || GET_CODE (x) == ZERO_EXTRACT) | |
3025 | && GET_MODE (x) == DImode | |
3026 | && GET_CODE (XEXP (x, 0)) == MULT | |
3027 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3028 | && CONST_INT_P (XEXP (XEXP (x, 0), 1))) | |
3029 | { | |
3030 | type = (GET_CODE (x) == SIGN_EXTRACT) | |
3031 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3032 | index = XEXP (XEXP (x, 0), 0); | |
3033 | shift = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1))); | |
3034 | if (INTVAL (XEXP (x, 1)) != 32 + shift | |
3035 | || INTVAL (XEXP (x, 2)) != 0) | |
3036 | shift = -1; | |
3037 | } | |
3038 | /* (and:DI (mult:DI (reg:DI) (const_int scale)) | |
3039 | (const_int 0xffffffff<<shift)) */ | |
3040 | else if (GET_CODE (x) == AND | |
3041 | && GET_MODE (x) == DImode | |
3042 | && GET_CODE (XEXP (x, 0)) == MULT | |
3043 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3044 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
3045 | && CONST_INT_P (XEXP (x, 1))) | |
3046 | { | |
3047 | type = ADDRESS_REG_UXTW; | |
3048 | index = XEXP (XEXP (x, 0), 0); | |
3049 | shift = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1))); | |
3050 | if (INTVAL (XEXP (x, 1)) != (HOST_WIDE_INT)0xffffffff << shift) | |
3051 | shift = -1; | |
3052 | } | |
3053 | /* (sign_extract:DI (ashift:DI (reg:DI) (const_int shift)) 32+shift 0) */ | |
3054 | else if ((GET_CODE (x) == SIGN_EXTRACT | |
3055 | || GET_CODE (x) == ZERO_EXTRACT) | |
3056 | && GET_MODE (x) == DImode | |
3057 | && GET_CODE (XEXP (x, 0)) == ASHIFT | |
3058 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3059 | && CONST_INT_P (XEXP (XEXP (x, 0), 1))) | |
3060 | { | |
3061 | type = (GET_CODE (x) == SIGN_EXTRACT) | |
3062 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3063 | index = XEXP (XEXP (x, 0), 0); | |
3064 | shift = INTVAL (XEXP (XEXP (x, 0), 1)); | |
3065 | if (INTVAL (XEXP (x, 1)) != 32 + shift | |
3066 | || INTVAL (XEXP (x, 2)) != 0) | |
3067 | shift = -1; | |
3068 | } | |
3069 | /* (and:DI (ashift:DI (reg:DI) (const_int shift)) | |
3070 | (const_int 0xffffffff<<shift)) */ | |
3071 | else if (GET_CODE (x) == AND | |
3072 | && GET_MODE (x) == DImode | |
3073 | && GET_CODE (XEXP (x, 0)) == ASHIFT | |
3074 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3075 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
3076 | && CONST_INT_P (XEXP (x, 1))) | |
3077 | { | |
3078 | type = ADDRESS_REG_UXTW; | |
3079 | index = XEXP (XEXP (x, 0), 0); | |
3080 | shift = INTVAL (XEXP (XEXP (x, 0), 1)); | |
3081 | if (INTVAL (XEXP (x, 1)) != (HOST_WIDE_INT)0xffffffff << shift) | |
3082 | shift = -1; | |
3083 | } | |
3084 | /* (mult:P (reg:P) (const_int scale)) */ | |
3085 | else if (GET_CODE (x) == MULT | |
3086 | && GET_MODE (x) == Pmode | |
3087 | && GET_MODE (XEXP (x, 0)) == Pmode | |
3088 | && CONST_INT_P (XEXP (x, 1))) | |
3089 | { | |
3090 | type = ADDRESS_REG_REG; | |
3091 | index = XEXP (x, 0); | |
3092 | shift = exact_log2 (INTVAL (XEXP (x, 1))); | |
3093 | } | |
3094 | /* (ashift:P (reg:P) (const_int shift)) */ | |
3095 | else if (GET_CODE (x) == ASHIFT | |
3096 | && GET_MODE (x) == Pmode | |
3097 | && GET_MODE (XEXP (x, 0)) == Pmode | |
3098 | && CONST_INT_P (XEXP (x, 1))) | |
3099 | { | |
3100 | type = ADDRESS_REG_REG; | |
3101 | index = XEXP (x, 0); | |
3102 | shift = INTVAL (XEXP (x, 1)); | |
3103 | } | |
3104 | else | |
3105 | return false; | |
3106 | ||
3107 | if (GET_CODE (index) == SUBREG) | |
3108 | index = SUBREG_REG (index); | |
3109 | ||
3110 | if ((shift == 0 || | |
3111 | (shift > 0 && shift <= 3 | |
3112 | && (1 << shift) == GET_MODE_SIZE (mode))) | |
3113 | && REG_P (index) | |
3114 | && aarch64_regno_ok_for_index_p (REGNO (index), strict_p)) | |
3115 | { | |
3116 | info->type = type; | |
3117 | info->offset = index; | |
3118 | info->shift = shift; | |
3119 | return true; | |
3120 | } | |
3121 | ||
3122 | return false; | |
3123 | } | |
3124 | ||
3125 | static inline bool | |
3126 | offset_7bit_signed_scaled_p (enum machine_mode mode, HOST_WIDE_INT offset) | |
3127 | { | |
3128 | return (offset >= -64 * GET_MODE_SIZE (mode) | |
3129 | && offset < 64 * GET_MODE_SIZE (mode) | |
3130 | && offset % GET_MODE_SIZE (mode) == 0); | |
3131 | } | |
3132 | ||
3133 | static inline bool | |
3134 | offset_9bit_signed_unscaled_p (enum machine_mode mode ATTRIBUTE_UNUSED, | |
3135 | HOST_WIDE_INT offset) | |
3136 | { | |
3137 | return offset >= -256 && offset < 256; | |
3138 | } | |
3139 | ||
3140 | static inline bool | |
3141 | offset_12bit_unsigned_scaled_p (enum machine_mode mode, HOST_WIDE_INT offset) | |
3142 | { | |
3143 | return (offset >= 0 | |
3144 | && offset < 4096 * GET_MODE_SIZE (mode) | |
3145 | && offset % GET_MODE_SIZE (mode) == 0); | |
3146 | } | |
3147 | ||
3148 | /* Return true if X is a valid address for machine mode MODE. If it is, | |
3149 | fill in INFO appropriately. STRICT_P is true if REG_OK_STRICT is in | |
3150 | effect. OUTER_CODE is PARALLEL for a load/store pair. */ | |
3151 | ||
3152 | static bool | |
3153 | aarch64_classify_address (struct aarch64_address_info *info, | |
3154 | rtx x, enum machine_mode mode, | |
3155 | RTX_CODE outer_code, bool strict_p) | |
3156 | { | |
3157 | enum rtx_code code = GET_CODE (x); | |
3158 | rtx op0, op1; | |
3159 | bool allow_reg_index_p = | |
3160 | outer_code != PARALLEL && GET_MODE_SIZE(mode) != 16; | |
3161 | ||
3162 | /* Don't support anything other than POST_INC or REG addressing for | |
3163 | AdvSIMD. */ | |
3164 | if (aarch64_vector_mode_p (mode) | |
3165 | && (code != POST_INC && code != REG)) | |
3166 | return false; | |
3167 | ||
3168 | switch (code) | |
3169 | { | |
3170 | case REG: | |
3171 | case SUBREG: | |
3172 | info->type = ADDRESS_REG_IMM; | |
3173 | info->base = x; | |
3174 | info->offset = const0_rtx; | |
3175 | return aarch64_base_register_rtx_p (x, strict_p); | |
3176 | ||
3177 | case PLUS: | |
3178 | op0 = XEXP (x, 0); | |
3179 | op1 = XEXP (x, 1); | |
3180 | if (GET_MODE_SIZE (mode) != 0 | |
3181 | && CONST_INT_P (op1) | |
3182 | && aarch64_base_register_rtx_p (op0, strict_p)) | |
3183 | { | |
3184 | HOST_WIDE_INT offset = INTVAL (op1); | |
3185 | ||
3186 | info->type = ADDRESS_REG_IMM; | |
3187 | info->base = op0; | |
3188 | info->offset = op1; | |
3189 | ||
3190 | /* TImode and TFmode values are allowed in both pairs of X | |
3191 | registers and individual Q registers. The available | |
3192 | address modes are: | |
3193 | X,X: 7-bit signed scaled offset | |
3194 | Q: 9-bit signed offset | |
3195 | We conservatively require an offset representable in either mode. | |
3196 | */ | |
3197 | if (mode == TImode || mode == TFmode) | |
3198 | return (offset_7bit_signed_scaled_p (mode, offset) | |
3199 | && offset_9bit_signed_unscaled_p (mode, offset)); | |
3200 | ||
3201 | if (outer_code == PARALLEL) | |
3202 | return ((GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) | |
3203 | && offset_7bit_signed_scaled_p (mode, offset)); | |
3204 | else | |
3205 | return (offset_9bit_signed_unscaled_p (mode, offset) | |
3206 | || offset_12bit_unsigned_scaled_p (mode, offset)); | |
3207 | } | |
3208 | ||
3209 | if (allow_reg_index_p) | |
3210 | { | |
3211 | /* Look for base + (scaled/extended) index register. */ | |
3212 | if (aarch64_base_register_rtx_p (op0, strict_p) | |
3213 | && aarch64_classify_index (info, op1, mode, strict_p)) | |
3214 | { | |
3215 | info->base = op0; | |
3216 | return true; | |
3217 | } | |
3218 | if (aarch64_base_register_rtx_p (op1, strict_p) | |
3219 | && aarch64_classify_index (info, op0, mode, strict_p)) | |
3220 | { | |
3221 | info->base = op1; | |
3222 | return true; | |
3223 | } | |
3224 | } | |
3225 | ||
3226 | return false; | |
3227 | ||
3228 | case POST_INC: | |
3229 | case POST_DEC: | |
3230 | case PRE_INC: | |
3231 | case PRE_DEC: | |
3232 | info->type = ADDRESS_REG_WB; | |
3233 | info->base = XEXP (x, 0); | |
3234 | info->offset = NULL_RTX; | |
3235 | return aarch64_base_register_rtx_p (info->base, strict_p); | |
3236 | ||
3237 | case POST_MODIFY: | |
3238 | case PRE_MODIFY: | |
3239 | info->type = ADDRESS_REG_WB; | |
3240 | info->base = XEXP (x, 0); | |
3241 | if (GET_CODE (XEXP (x, 1)) == PLUS | |
3242 | && CONST_INT_P (XEXP (XEXP (x, 1), 1)) | |
3243 | && rtx_equal_p (XEXP (XEXP (x, 1), 0), info->base) | |
3244 | && aarch64_base_register_rtx_p (info->base, strict_p)) | |
3245 | { | |
3246 | HOST_WIDE_INT offset; | |
3247 | info->offset = XEXP (XEXP (x, 1), 1); | |
3248 | offset = INTVAL (info->offset); | |
3249 | ||
3250 | /* TImode and TFmode values are allowed in both pairs of X | |
3251 | registers and individual Q registers. The available | |
3252 | address modes are: | |
3253 | X,X: 7-bit signed scaled offset | |
3254 | Q: 9-bit signed offset | |
3255 | We conservatively require an offset representable in either mode. | |
3256 | */ | |
3257 | if (mode == TImode || mode == TFmode) | |
3258 | return (offset_7bit_signed_scaled_p (mode, offset) | |
3259 | && offset_9bit_signed_unscaled_p (mode, offset)); | |
3260 | ||
3261 | if (outer_code == PARALLEL) | |
3262 | return ((GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) | |
3263 | && offset_7bit_signed_scaled_p (mode, offset)); | |
3264 | else | |
3265 | return offset_9bit_signed_unscaled_p (mode, offset); | |
3266 | } | |
3267 | return false; | |
3268 | ||
3269 | case CONST: | |
3270 | case SYMBOL_REF: | |
3271 | case LABEL_REF: | |
3272 | /* load literal: pc-relative constant pool entry. Only supported | |
3273 | for SI mode or larger. */ | |
3274 | info->type = ADDRESS_SYMBOLIC; | |
3275 | if (outer_code != PARALLEL && GET_MODE_SIZE (mode) >= 4) | |
3276 | { | |
3277 | rtx sym, addend; | |
3278 | ||
3279 | split_const (x, &sym, &addend); | |
3280 | return (GET_CODE (sym) == LABEL_REF | |
3281 | || (GET_CODE (sym) == SYMBOL_REF | |
3282 | && CONSTANT_POOL_ADDRESS_P (sym))); | |
3283 | } | |
3284 | return false; | |
3285 | ||
3286 | case LO_SUM: | |
3287 | info->type = ADDRESS_LO_SUM; | |
3288 | info->base = XEXP (x, 0); | |
3289 | info->offset = XEXP (x, 1); | |
3290 | if (allow_reg_index_p | |
3291 | && aarch64_base_register_rtx_p (info->base, strict_p)) | |
3292 | { | |
3293 | rtx sym, offs; | |
3294 | split_const (info->offset, &sym, &offs); | |
3295 | if (GET_CODE (sym) == SYMBOL_REF | |
3296 | && (aarch64_classify_symbol (sym, SYMBOL_CONTEXT_MEM) | |
3297 | == SYMBOL_SMALL_ABSOLUTE)) | |
3298 | { | |
3299 | /* The symbol and offset must be aligned to the access size. */ | |
3300 | unsigned int align; | |
3301 | unsigned int ref_size; | |
3302 | ||
3303 | if (CONSTANT_POOL_ADDRESS_P (sym)) | |
3304 | align = GET_MODE_ALIGNMENT (get_pool_mode (sym)); | |
3305 | else if (TREE_CONSTANT_POOL_ADDRESS_P (sym)) | |
3306 | { | |
3307 | tree exp = SYMBOL_REF_DECL (sym); | |
3308 | align = TYPE_ALIGN (TREE_TYPE (exp)); | |
3309 | align = CONSTANT_ALIGNMENT (exp, align); | |
3310 | } | |
3311 | else if (SYMBOL_REF_DECL (sym)) | |
3312 | align = DECL_ALIGN (SYMBOL_REF_DECL (sym)); | |
3313 | else if (SYMBOL_REF_HAS_BLOCK_INFO_P (sym) | |
3314 | && SYMBOL_REF_BLOCK (sym) != NULL) | |
3315 | align = SYMBOL_REF_BLOCK (sym)->alignment; | |
3316 | else | |
3317 | align = BITS_PER_UNIT; | |
3318 | ||
3319 | ref_size = GET_MODE_SIZE (mode); | |
3320 | if (ref_size == 0) | |
3321 | ref_size = GET_MODE_SIZE (DImode); | |
3322 | ||
3323 | return ((INTVAL (offs) & (ref_size - 1)) == 0 | |
3324 | && ((align / BITS_PER_UNIT) & (ref_size - 1)) == 0); | |
3325 | } | |
3326 | } | |
3327 | return false; | |
3328 | ||
3329 | default: | |
3330 | return false; | |
3331 | } | |
3332 | } | |
3333 | ||
3334 | bool | |
3335 | aarch64_symbolic_address_p (rtx x) | |
3336 | { | |
3337 | rtx offset; | |
3338 | ||
3339 | split_const (x, &x, &offset); | |
3340 | return GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == LABEL_REF; | |
3341 | } | |
3342 | ||
3343 | /* Classify the base of symbolic expression X, given that X appears in | |
3344 | context CONTEXT. */ | |
3345 | ||
3346 | enum aarch64_symbol_type | |
3347 | aarch64_classify_symbolic_expression (rtx x, | |
3348 | enum aarch64_symbol_context context) | |
3349 | { | |
3350 | rtx offset; | |
3351 | ||
3352 | split_const (x, &x, &offset); | |
3353 | return aarch64_classify_symbol (x, context); | |
3354 | } | |
3355 | ||
3356 | ||
3357 | /* Return TRUE if X is a legitimate address for accessing memory in | |
3358 | mode MODE. */ | |
3359 | static bool | |
3360 | aarch64_legitimate_address_hook_p (enum machine_mode mode, rtx x, bool strict_p) | |
3361 | { | |
3362 | struct aarch64_address_info addr; | |
3363 | ||
3364 | return aarch64_classify_address (&addr, x, mode, MEM, strict_p); | |
3365 | } | |
3366 | ||
3367 | /* Return TRUE if X is a legitimate address for accessing memory in | |
3368 | mode MODE. OUTER_CODE will be PARALLEL if this is a load/store | |
3369 | pair operation. */ | |
3370 | bool | |
3371 | aarch64_legitimate_address_p (enum machine_mode mode, rtx x, | |
3372 | RTX_CODE outer_code, bool strict_p) | |
3373 | { | |
3374 | struct aarch64_address_info addr; | |
3375 | ||
3376 | return aarch64_classify_address (&addr, x, mode, outer_code, strict_p); | |
3377 | } | |
3378 | ||
3379 | /* Return TRUE if rtx X is immediate constant 0.0 */ | |
3380 | bool | |
3381 | aarch64_float_const_zero_rtx_p (rtx x) | |
3382 | { | |
3383 | REAL_VALUE_TYPE r; | |
3384 | ||
3385 | if (GET_MODE (x) == VOIDmode) | |
3386 | return false; | |
3387 | ||
3388 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
3389 | if (REAL_VALUE_MINUS_ZERO (r)) | |
3390 | return !HONOR_SIGNED_ZEROS (GET_MODE (x)); | |
3391 | return REAL_VALUES_EQUAL (r, dconst0); | |
3392 | } | |
3393 | ||
3394 | /* Return the fixed registers used for condition codes. */ | |
3395 | ||
3396 | static bool | |
3397 | aarch64_fixed_condition_code_regs (unsigned int *p1, unsigned int *p2) | |
3398 | { | |
3399 | *p1 = CC_REGNUM; | |
3400 | *p2 = INVALID_REGNUM; | |
3401 | return true; | |
3402 | } | |
3403 | ||
3404 | enum machine_mode | |
3405 | aarch64_select_cc_mode (RTX_CODE code, rtx x, rtx y) | |
3406 | { | |
3407 | /* All floating point compares return CCFP if it is an equality | |
3408 | comparison, and CCFPE otherwise. */ | |
3409 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
3410 | { | |
3411 | switch (code) | |
3412 | { | |
3413 | case EQ: | |
3414 | case NE: | |
3415 | case UNORDERED: | |
3416 | case ORDERED: | |
3417 | case UNLT: | |
3418 | case UNLE: | |
3419 | case UNGT: | |
3420 | case UNGE: | |
3421 | case UNEQ: | |
3422 | case LTGT: | |
3423 | return CCFPmode; | |
3424 | ||
3425 | case LT: | |
3426 | case LE: | |
3427 | case GT: | |
3428 | case GE: | |
3429 | return CCFPEmode; | |
3430 | ||
3431 | default: | |
3432 | gcc_unreachable (); | |
3433 | } | |
3434 | } | |
3435 | ||
3436 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
3437 | && y == const0_rtx | |
3438 | && (code == EQ || code == NE || code == LT || code == GE) | |
3439 | && (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS || GET_CODE (x) == AND | |
3440 | || GET_CODE (x) == NEG)) | |
3441 | return CC_NZmode; | |
3442 | ||
3443 | /* A compare with a shifted operand. Because of canonicalization, | |
3444 | the comparison will have to be swapped when we emit the assembly | |
3445 | code. */ | |
3446 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
3447 | && (GET_CODE (y) == REG || GET_CODE (y) == SUBREG) | |
3448 | && (GET_CODE (x) == ASHIFT || GET_CODE (x) == ASHIFTRT | |
3449 | || GET_CODE (x) == LSHIFTRT | |
3450 | || GET_CODE (x) == ZERO_EXTEND || GET_CODE (x) == SIGN_EXTEND)) | |
3451 | return CC_SWPmode; | |
3452 | ||
3453 | /* Similarly for a negated operand, but we can only do this for | |
3454 | equalities. */ | |
3455 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
3456 | && (GET_CODE (y) == REG || GET_CODE (y) == SUBREG) | |
3457 | && (code == EQ || code == NE) | |
3458 | && GET_CODE (x) == NEG) | |
3459 | return CC_Zmode; | |
3460 | ||
3461 | /* A compare of a mode narrower than SI mode against zero can be done | |
3462 | by extending the value in the comparison. */ | |
3463 | if ((GET_MODE (x) == QImode || GET_MODE (x) == HImode) | |
3464 | && y == const0_rtx) | |
3465 | /* Only use sign-extension if we really need it. */ | |
3466 | return ((code == GT || code == GE || code == LE || code == LT) | |
3467 | ? CC_SESWPmode : CC_ZESWPmode); | |
3468 | ||
3469 | /* For everything else, return CCmode. */ | |
3470 | return CCmode; | |
3471 | } | |
3472 | ||
3473 | static unsigned | |
3474 | aarch64_get_condition_code (rtx x) | |
3475 | { | |
3476 | enum machine_mode mode = GET_MODE (XEXP (x, 0)); | |
3477 | enum rtx_code comp_code = GET_CODE (x); | |
3478 | ||
3479 | if (GET_MODE_CLASS (mode) != MODE_CC) | |
3480 | mode = SELECT_CC_MODE (comp_code, XEXP (x, 0), XEXP (x, 1)); | |
3481 | ||
3482 | switch (mode) | |
3483 | { | |
3484 | case CCFPmode: | |
3485 | case CCFPEmode: | |
3486 | switch (comp_code) | |
3487 | { | |
3488 | case GE: return AARCH64_GE; | |
3489 | case GT: return AARCH64_GT; | |
3490 | case LE: return AARCH64_LS; | |
3491 | case LT: return AARCH64_MI; | |
3492 | case NE: return AARCH64_NE; | |
3493 | case EQ: return AARCH64_EQ; | |
3494 | case ORDERED: return AARCH64_VC; | |
3495 | case UNORDERED: return AARCH64_VS; | |
3496 | case UNLT: return AARCH64_LT; | |
3497 | case UNLE: return AARCH64_LE; | |
3498 | case UNGT: return AARCH64_HI; | |
3499 | case UNGE: return AARCH64_PL; | |
3500 | default: gcc_unreachable (); | |
3501 | } | |
3502 | break; | |
3503 | ||
3504 | case CCmode: | |
3505 | switch (comp_code) | |
3506 | { | |
3507 | case NE: return AARCH64_NE; | |
3508 | case EQ: return AARCH64_EQ; | |
3509 | case GE: return AARCH64_GE; | |
3510 | case GT: return AARCH64_GT; | |
3511 | case LE: return AARCH64_LE; | |
3512 | case LT: return AARCH64_LT; | |
3513 | case GEU: return AARCH64_CS; | |
3514 | case GTU: return AARCH64_HI; | |
3515 | case LEU: return AARCH64_LS; | |
3516 | case LTU: return AARCH64_CC; | |
3517 | default: gcc_unreachable (); | |
3518 | } | |
3519 | break; | |
3520 | ||
3521 | case CC_SWPmode: | |
3522 | case CC_ZESWPmode: | |
3523 | case CC_SESWPmode: | |
3524 | switch (comp_code) | |
3525 | { | |
3526 | case NE: return AARCH64_NE; | |
3527 | case EQ: return AARCH64_EQ; | |
3528 | case GE: return AARCH64_LE; | |
3529 | case GT: return AARCH64_LT; | |
3530 | case LE: return AARCH64_GE; | |
3531 | case LT: return AARCH64_GT; | |
3532 | case GEU: return AARCH64_LS; | |
3533 | case GTU: return AARCH64_CC; | |
3534 | case LEU: return AARCH64_CS; | |
3535 | case LTU: return AARCH64_HI; | |
3536 | default: gcc_unreachable (); | |
3537 | } | |
3538 | break; | |
3539 | ||
3540 | case CC_NZmode: | |
3541 | switch (comp_code) | |
3542 | { | |
3543 | case NE: return AARCH64_NE; | |
3544 | case EQ: return AARCH64_EQ; | |
3545 | case GE: return AARCH64_PL; | |
3546 | case LT: return AARCH64_MI; | |
3547 | default: gcc_unreachable (); | |
3548 | } | |
3549 | break; | |
3550 | ||
3551 | case CC_Zmode: | |
3552 | switch (comp_code) | |
3553 | { | |
3554 | case NE: return AARCH64_NE; | |
3555 | case EQ: return AARCH64_EQ; | |
3556 | default: gcc_unreachable (); | |
3557 | } | |
3558 | break; | |
3559 | ||
3560 | default: | |
3561 | gcc_unreachable (); | |
3562 | break; | |
3563 | } | |
3564 | } | |
3565 | ||
3566 | static unsigned | |
3567 | bit_count (unsigned HOST_WIDE_INT value) | |
3568 | { | |
3569 | unsigned count = 0; | |
3570 | ||
3571 | while (value) | |
3572 | { | |
3573 | count++; | |
3574 | value &= value - 1; | |
3575 | } | |
3576 | ||
3577 | return count; | |
3578 | } | |
3579 | ||
3580 | void | |
3581 | aarch64_print_operand (FILE *f, rtx x, char code) | |
3582 | { | |
3583 | switch (code) | |
3584 | { | |
3585 | /* An integer or symbol address without a preceding # sign. */ | |
3586 | case 'c': | |
3587 | switch (GET_CODE (x)) | |
3588 | { | |
3589 | case CONST_INT: | |
3590 | fprintf (f, HOST_WIDE_INT_PRINT_DEC, INTVAL (x)); | |
3591 | break; | |
3592 | ||
3593 | case SYMBOL_REF: | |
3594 | output_addr_const (f, x); | |
3595 | break; | |
3596 | ||
3597 | case CONST: | |
3598 | if (GET_CODE (XEXP (x, 0)) == PLUS | |
3599 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF) | |
3600 | { | |
3601 | output_addr_const (f, x); | |
3602 | break; | |
3603 | } | |
3604 | /* Fall through. */ | |
3605 | ||
3606 | default: | |
3607 | output_operand_lossage ("Unsupported operand for code '%c'", code); | |
3608 | } | |
3609 | break; | |
3610 | ||
3611 | case 'e': | |
3612 | /* Print the sign/zero-extend size as a character 8->b, 16->h, 32->w. */ | |
3613 | { | |
3614 | int n; | |
3615 | ||
3616 | if (GET_CODE (x) != CONST_INT | |
3617 | || (n = exact_log2 (INTVAL (x) & ~7)) <= 0) | |
3618 | { | |
3619 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3620 | return; | |
3621 | } | |
3622 | ||
3623 | switch (n) | |
3624 | { | |
3625 | case 3: | |
3626 | fputc ('b', f); | |
3627 | break; | |
3628 | case 4: | |
3629 | fputc ('h', f); | |
3630 | break; | |
3631 | case 5: | |
3632 | fputc ('w', f); | |
3633 | break; | |
3634 | default: | |
3635 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3636 | return; | |
3637 | } | |
3638 | } | |
3639 | break; | |
3640 | ||
3641 | case 'p': | |
3642 | { | |
3643 | int n; | |
3644 | ||
3645 | /* Print N such that 2^N == X. */ | |
3646 | if (GET_CODE (x) != CONST_INT || (n = exact_log2 (INTVAL (x))) < 0) | |
3647 | { | |
3648 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3649 | return; | |
3650 | } | |
3651 | ||
3652 | asm_fprintf (f, "%d", n); | |
3653 | } | |
3654 | break; | |
3655 | ||
3656 | case 'P': | |
3657 | /* Print the number of non-zero bits in X (a const_int). */ | |
3658 | if (GET_CODE (x) != CONST_INT) | |
3659 | { | |
3660 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3661 | return; | |
3662 | } | |
3663 | ||
3664 | asm_fprintf (f, "%u", bit_count (INTVAL (x))); | |
3665 | break; | |
3666 | ||
3667 | case 'H': | |
3668 | /* Print the higher numbered register of a pair (TImode) of regs. */ | |
3669 | if (GET_CODE (x) != REG || !GP_REGNUM_P (REGNO (x) + 1)) | |
3670 | { | |
3671 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3672 | return; | |
3673 | } | |
3674 | ||
3675 | asm_fprintf (f, "%s", reg_names [REGNO (x) + 1]); | |
3676 | break; | |
3677 | ||
3678 | case 'm': | |
3679 | /* Print a condition (eq, ne, etc). */ | |
3680 | ||
3681 | /* CONST_TRUE_RTX means always -- that's the default. */ | |
3682 | if (x == const_true_rtx) | |
3683 | return; | |
3684 | ||
3685 | if (!COMPARISON_P (x)) | |
3686 | { | |
3687 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3688 | return; | |
3689 | } | |
3690 | ||
3691 | fputs (aarch64_condition_codes[aarch64_get_condition_code (x)], f); | |
3692 | break; | |
3693 | ||
3694 | case 'M': | |
3695 | /* Print the inverse of a condition (eq <-> ne, etc). */ | |
3696 | ||
3697 | /* CONST_TRUE_RTX means never -- that's the default. */ | |
3698 | if (x == const_true_rtx) | |
3699 | { | |
3700 | fputs ("nv", f); | |
3701 | return; | |
3702 | } | |
3703 | ||
3704 | if (!COMPARISON_P (x)) | |
3705 | { | |
3706 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3707 | return; | |
3708 | } | |
3709 | ||
3710 | fputs (aarch64_condition_codes[AARCH64_INVERSE_CONDITION_CODE | |
3711 | (aarch64_get_condition_code (x))], f); | |
3712 | break; | |
3713 | ||
3714 | case 'b': | |
3715 | case 'h': | |
3716 | case 's': | |
3717 | case 'd': | |
3718 | case 'q': | |
3719 | /* Print a scalar FP/SIMD register name. */ | |
3720 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) | |
3721 | { | |
3722 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
3723 | return; | |
3724 | } | |
3725 | asm_fprintf (f, "%c%d", code, REGNO (x) - V0_REGNUM); | |
3726 | break; | |
3727 | ||
3728 | case 'S': | |
3729 | case 'T': | |
3730 | case 'U': | |
3731 | case 'V': | |
3732 | /* Print the first FP/SIMD register name in a list. */ | |
3733 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) | |
3734 | { | |
3735 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
3736 | return; | |
3737 | } | |
3738 | asm_fprintf (f, "v%d", REGNO (x) - V0_REGNUM + (code - 'S')); | |
3739 | break; | |
3740 | ||
3741 | case 'X': | |
3742 | /* Print bottom 16 bits of integer constant in hex. */ | |
3743 | if (GET_CODE (x) != CONST_INT) | |
3744 | { | |
3745 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3746 | return; | |
3747 | } | |
3748 | asm_fprintf (f, "0x%wx", UINTVAL (x) & 0xffff); | |
3749 | break; | |
3750 | ||
3751 | case 'w': | |
3752 | case 'x': | |
3753 | /* Print a general register name or the zero register (32-bit or | |
3754 | 64-bit). */ | |
3755 | if (x == const0_rtx | |
3756 | || (CONST_DOUBLE_P (x) && aarch64_float_const_zero_rtx_p (x))) | |
3757 | { | |
3758 | asm_fprintf (f, "%czr", code); | |
3759 | break; | |
3760 | } | |
3761 | ||
3762 | if (REG_P (x) && GP_REGNUM_P (REGNO (x))) | |
3763 | { | |
3764 | asm_fprintf (f, "%c%d", code, REGNO (x) - R0_REGNUM); | |
3765 | break; | |
3766 | } | |
3767 | ||
3768 | if (REG_P (x) && REGNO (x) == SP_REGNUM) | |
3769 | { | |
3770 | asm_fprintf (f, "%ssp", code == 'w' ? "w" : ""); | |
3771 | break; | |
3772 | } | |
3773 | ||
3774 | /* Fall through */ | |
3775 | ||
3776 | case 0: | |
3777 | /* Print a normal operand, if it's a general register, then we | |
3778 | assume DImode. */ | |
3779 | if (x == NULL) | |
3780 | { | |
3781 | output_operand_lossage ("missing operand"); | |
3782 | return; | |
3783 | } | |
3784 | ||
3785 | switch (GET_CODE (x)) | |
3786 | { | |
3787 | case REG: | |
3788 | asm_fprintf (f, "%s", reg_names [REGNO (x)]); | |
3789 | break; | |
3790 | ||
3791 | case MEM: | |
3792 | aarch64_memory_reference_mode = GET_MODE (x); | |
3793 | output_address (XEXP (x, 0)); | |
3794 | break; | |
3795 | ||
3796 | case LABEL_REF: | |
3797 | case SYMBOL_REF: | |
3798 | output_addr_const (asm_out_file, x); | |
3799 | break; | |
3800 | ||
3801 | case CONST_INT: | |
3802 | asm_fprintf (f, "%wd", INTVAL (x)); | |
3803 | break; | |
3804 | ||
3805 | case CONST_VECTOR: | |
3806 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_VECTOR_INT) | |
3807 | { | |
3808 | gcc_assert (aarch64_const_vec_all_same_int_p (x, | |
3809 | HOST_WIDE_INT_MIN, | |
3810 | HOST_WIDE_INT_MAX)); | |
3811 | asm_fprintf (f, "%wd", INTVAL (CONST_VECTOR_ELT (x, 0))); | |
3812 | } | |
3813 | else if (aarch64_simd_imm_zero_p (x, GET_MODE (x))) | |
3814 | { | |
3815 | fputc ('0', f); | |
3816 | } | |
3817 | else | |
3818 | gcc_unreachable (); | |
3819 | break; | |
3820 | ||
3821 | case CONST_DOUBLE: | |
3822 | /* CONST_DOUBLE can represent a double-width integer. | |
3823 | In this case, the mode of x is VOIDmode. */ | |
3824 | if (GET_MODE (x) == VOIDmode) | |
3825 | ; /* Do Nothing. */ | |
3826 | else if (aarch64_float_const_zero_rtx_p (x)) | |
3827 | { | |
3828 | fputc ('0', f); | |
3829 | break; | |
3830 | } | |
3831 | else if (aarch64_float_const_representable_p (x)) | |
3832 | { | |
3833 | #define buf_size 20 | |
3834 | char float_buf[buf_size] = {'\0'}; | |
3835 | REAL_VALUE_TYPE r; | |
3836 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
3837 | real_to_decimal_for_mode (float_buf, &r, | |
3838 | buf_size, buf_size, | |
3839 | 1, GET_MODE (x)); | |
3840 | asm_fprintf (asm_out_file, "%s", float_buf); | |
3841 | break; | |
3842 | #undef buf_size | |
3843 | } | |
3844 | output_operand_lossage ("invalid constant"); | |
3845 | return; | |
3846 | default: | |
3847 | output_operand_lossage ("invalid operand"); | |
3848 | return; | |
3849 | } | |
3850 | break; | |
3851 | ||
3852 | case 'A': | |
3853 | if (GET_CODE (x) == HIGH) | |
3854 | x = XEXP (x, 0); | |
3855 | ||
3856 | switch (aarch64_classify_symbolic_expression (x, SYMBOL_CONTEXT_ADR)) | |
3857 | { | |
3858 | case SYMBOL_SMALL_GOT: | |
3859 | asm_fprintf (asm_out_file, ":got:"); | |
3860 | break; | |
3861 | ||
3862 | case SYMBOL_SMALL_TLSGD: | |
3863 | asm_fprintf (asm_out_file, ":tlsgd:"); | |
3864 | break; | |
3865 | ||
3866 | case SYMBOL_SMALL_TLSDESC: | |
3867 | asm_fprintf (asm_out_file, ":tlsdesc:"); | |
3868 | break; | |
3869 | ||
3870 | case SYMBOL_SMALL_GOTTPREL: | |
3871 | asm_fprintf (asm_out_file, ":gottprel:"); | |
3872 | break; | |
3873 | ||
3874 | case SYMBOL_SMALL_TPREL: | |
3875 | asm_fprintf (asm_out_file, ":tprel:"); | |
3876 | break; | |
3877 | ||
3878 | case SYMBOL_TINY_GOT: | |
3879 | gcc_unreachable (); | |
3880 | break; | |
3881 | ||
3882 | default: | |
3883 | break; | |
3884 | } | |
3885 | output_addr_const (asm_out_file, x); | |
3886 | break; | |
3887 | ||
3888 | case 'L': | |
3889 | switch (aarch64_classify_symbolic_expression (x, SYMBOL_CONTEXT_ADR)) | |
3890 | { | |
3891 | case SYMBOL_SMALL_GOT: | |
3892 | asm_fprintf (asm_out_file, ":lo12:"); | |
3893 | break; | |
3894 | ||
3895 | case SYMBOL_SMALL_TLSGD: | |
3896 | asm_fprintf (asm_out_file, ":tlsgd_lo12:"); | |
3897 | break; | |
3898 | ||
3899 | case SYMBOL_SMALL_TLSDESC: | |
3900 | asm_fprintf (asm_out_file, ":tlsdesc_lo12:"); | |
3901 | break; | |
3902 | ||
3903 | case SYMBOL_SMALL_GOTTPREL: | |
3904 | asm_fprintf (asm_out_file, ":gottprel_lo12:"); | |
3905 | break; | |
3906 | ||
3907 | case SYMBOL_SMALL_TPREL: | |
3908 | asm_fprintf (asm_out_file, ":tprel_lo12_nc:"); | |
3909 | break; | |
3910 | ||
3911 | case SYMBOL_TINY_GOT: | |
3912 | asm_fprintf (asm_out_file, ":got:"); | |
3913 | break; | |
3914 | ||
3915 | default: | |
3916 | break; | |
3917 | } | |
3918 | output_addr_const (asm_out_file, x); | |
3919 | break; | |
3920 | ||
3921 | case 'G': | |
3922 | ||
3923 | switch (aarch64_classify_symbolic_expression (x, SYMBOL_CONTEXT_ADR)) | |
3924 | { | |
3925 | case SYMBOL_SMALL_TPREL: | |
3926 | asm_fprintf (asm_out_file, ":tprel_hi12:"); | |
3927 | break; | |
3928 | default: | |
3929 | break; | |
3930 | } | |
3931 | output_addr_const (asm_out_file, x); | |
3932 | break; | |
3933 | ||
3934 | default: | |
3935 | output_operand_lossage ("invalid operand prefix '%%%c'", code); | |
3936 | return; | |
3937 | } | |
3938 | } | |
3939 | ||
3940 | void | |
3941 | aarch64_print_operand_address (FILE *f, rtx x) | |
3942 | { | |
3943 | struct aarch64_address_info addr; | |
3944 | ||
3945 | if (aarch64_classify_address (&addr, x, aarch64_memory_reference_mode, | |
3946 | MEM, true)) | |
3947 | switch (addr.type) | |
3948 | { | |
3949 | case ADDRESS_REG_IMM: | |
3950 | if (addr.offset == const0_rtx) | |
3951 | asm_fprintf (f, "[%s]", reg_names [REGNO (addr.base)]); | |
3952 | else | |
3953 | asm_fprintf (f, "[%s, %wd]", reg_names [REGNO (addr.base)], | |
3954 | INTVAL (addr.offset)); | |
3955 | return; | |
3956 | ||
3957 | case ADDRESS_REG_REG: | |
3958 | if (addr.shift == 0) | |
3959 | asm_fprintf (f, "[%s, %s]", reg_names [REGNO (addr.base)], | |
3960 | reg_names [REGNO (addr.offset)]); | |
3961 | else | |
3962 | asm_fprintf (f, "[%s, %s, lsl %u]", reg_names [REGNO (addr.base)], | |
3963 | reg_names [REGNO (addr.offset)], addr.shift); | |
3964 | return; | |
3965 | ||
3966 | case ADDRESS_REG_UXTW: | |
3967 | if (addr.shift == 0) | |
3968 | asm_fprintf (f, "[%s, w%d, uxtw]", reg_names [REGNO (addr.base)], | |
3969 | REGNO (addr.offset) - R0_REGNUM); | |
3970 | else | |
3971 | asm_fprintf (f, "[%s, w%d, uxtw %u]", reg_names [REGNO (addr.base)], | |
3972 | REGNO (addr.offset) - R0_REGNUM, addr.shift); | |
3973 | return; | |
3974 | ||
3975 | case ADDRESS_REG_SXTW: | |
3976 | if (addr.shift == 0) | |
3977 | asm_fprintf (f, "[%s, w%d, sxtw]", reg_names [REGNO (addr.base)], | |
3978 | REGNO (addr.offset) - R0_REGNUM); | |
3979 | else | |
3980 | asm_fprintf (f, "[%s, w%d, sxtw %u]", reg_names [REGNO (addr.base)], | |
3981 | REGNO (addr.offset) - R0_REGNUM, addr.shift); | |
3982 | return; | |
3983 | ||
3984 | case ADDRESS_REG_WB: | |
3985 | switch (GET_CODE (x)) | |
3986 | { | |
3987 | case PRE_INC: | |
3988 | asm_fprintf (f, "[%s, %d]!", reg_names [REGNO (addr.base)], | |
3989 | GET_MODE_SIZE (aarch64_memory_reference_mode)); | |
3990 | return; | |
3991 | case POST_INC: | |
3992 | asm_fprintf (f, "[%s], %d", reg_names [REGNO (addr.base)], | |
3993 | GET_MODE_SIZE (aarch64_memory_reference_mode)); | |
3994 | return; | |
3995 | case PRE_DEC: | |
3996 | asm_fprintf (f, "[%s, -%d]!", reg_names [REGNO (addr.base)], | |
3997 | GET_MODE_SIZE (aarch64_memory_reference_mode)); | |
3998 | return; | |
3999 | case POST_DEC: | |
4000 | asm_fprintf (f, "[%s], -%d", reg_names [REGNO (addr.base)], | |
4001 | GET_MODE_SIZE (aarch64_memory_reference_mode)); | |
4002 | return; | |
4003 | case PRE_MODIFY: | |
4004 | asm_fprintf (f, "[%s, %wd]!", reg_names [REGNO (addr.base)], | |
4005 | INTVAL (addr.offset)); | |
4006 | return; | |
4007 | case POST_MODIFY: | |
4008 | asm_fprintf (f, "[%s], %wd", reg_names [REGNO (addr.base)], | |
4009 | INTVAL (addr.offset)); | |
4010 | return; | |
4011 | default: | |
4012 | break; | |
4013 | } | |
4014 | break; | |
4015 | ||
4016 | case ADDRESS_LO_SUM: | |
4017 | asm_fprintf (f, "[%s, #:lo12:", reg_names [REGNO (addr.base)]); | |
4018 | output_addr_const (f, addr.offset); | |
4019 | asm_fprintf (f, "]"); | |
4020 | return; | |
4021 | ||
4022 | case ADDRESS_SYMBOLIC: | |
4023 | break; | |
4024 | } | |
4025 | ||
4026 | output_addr_const (f, x); | |
4027 | } | |
4028 | ||
4029 | bool | |
4030 | aarch64_label_mentioned_p (rtx x) | |
4031 | { | |
4032 | const char *fmt; | |
4033 | int i; | |
4034 | ||
4035 | if (GET_CODE (x) == LABEL_REF) | |
4036 | return true; | |
4037 | ||
4038 | /* UNSPEC_TLS entries for a symbol include a LABEL_REF for the | |
4039 | referencing instruction, but they are constant offsets, not | |
4040 | symbols. */ | |
4041 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
4042 | return false; | |
4043 | ||
4044 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
4045 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
4046 | { | |
4047 | if (fmt[i] == 'E') | |
4048 | { | |
4049 | int j; | |
4050 | ||
4051 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
4052 | if (aarch64_label_mentioned_p (XVECEXP (x, i, j))) | |
4053 | return 1; | |
4054 | } | |
4055 | else if (fmt[i] == 'e' && aarch64_label_mentioned_p (XEXP (x, i))) | |
4056 | return 1; | |
4057 | } | |
4058 | ||
4059 | return 0; | |
4060 | } | |
4061 | ||
4062 | /* Implement REGNO_REG_CLASS. */ | |
4063 | ||
4064 | enum reg_class | |
4065 | aarch64_regno_regclass (unsigned regno) | |
4066 | { | |
4067 | if (GP_REGNUM_P (regno)) | |
4068 | return GENERAL_REGS; | |
4069 | ||
4070 | if (regno == SP_REGNUM) | |
4071 | return STACK_REG; | |
4072 | ||
4073 | if (regno == FRAME_POINTER_REGNUM | |
4074 | || regno == ARG_POINTER_REGNUM) | |
4075 | return POINTER_REGS; | |
4076 | ||
4077 | if (FP_REGNUM_P (regno)) | |
4078 | return FP_LO_REGNUM_P (regno) ? FP_LO_REGS : FP_REGS; | |
4079 | ||
4080 | return NO_REGS; | |
4081 | } | |
4082 | ||
4083 | /* Try a machine-dependent way of reloading an illegitimate address | |
4084 | operand. If we find one, push the reload and return the new rtx. */ | |
4085 | ||
4086 | rtx | |
4087 | aarch64_legitimize_reload_address (rtx *x_p, | |
4088 | enum machine_mode mode, | |
4089 | int opnum, int type, | |
4090 | int ind_levels ATTRIBUTE_UNUSED) | |
4091 | { | |
4092 | rtx x = *x_p; | |
4093 | ||
4094 | /* Do not allow mem (plus (reg, const)) if vector mode. */ | |
4095 | if (aarch64_vector_mode_p (mode) | |
4096 | && GET_CODE (x) == PLUS | |
4097 | && REG_P (XEXP (x, 0)) | |
4098 | && CONST_INT_P (XEXP (x, 1))) | |
4099 | { | |
4100 | rtx orig_rtx = x; | |
4101 | x = copy_rtx (x); | |
4102 | push_reload (orig_rtx, NULL_RTX, x_p, NULL, | |
4103 | BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0, | |
4104 | opnum, (enum reload_type) type); | |
4105 | return x; | |
4106 | } | |
4107 | ||
4108 | /* We must recognize output that we have already generated ourselves. */ | |
4109 | if (GET_CODE (x) == PLUS | |
4110 | && GET_CODE (XEXP (x, 0)) == PLUS | |
4111 | && REG_P (XEXP (XEXP (x, 0), 0)) | |
4112 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
4113 | && CONST_INT_P (XEXP (x, 1))) | |
4114 | { | |
4115 | push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL, | |
4116 | BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0, | |
4117 | opnum, (enum reload_type) type); | |
4118 | return x; | |
4119 | } | |
4120 | ||
4121 | /* We wish to handle large displacements off a base register by splitting | |
4122 | the addend across an add and the mem insn. This can cut the number of | |
4123 | extra insns needed from 3 to 1. It is only useful for load/store of a | |
4124 | single register with 12 bit offset field. */ | |
4125 | if (GET_CODE (x) == PLUS | |
4126 | && REG_P (XEXP (x, 0)) | |
4127 | && CONST_INT_P (XEXP (x, 1)) | |
4128 | && HARD_REGISTER_P (XEXP (x, 0)) | |
4129 | && mode != TImode | |
4130 | && mode != TFmode | |
4131 | && aarch64_regno_ok_for_base_p (REGNO (XEXP (x, 0)), true)) | |
4132 | { | |
4133 | HOST_WIDE_INT val = INTVAL (XEXP (x, 1)); | |
4134 | HOST_WIDE_INT low = val & 0xfff; | |
4135 | HOST_WIDE_INT high = val - low; | |
4136 | HOST_WIDE_INT offs; | |
4137 | rtx cst; | |
4138 | enum machine_mode xmode = GET_MODE (x); | |
4139 | ||
4140 | /* In ILP32, xmode can be either DImode or SImode. */ | |
4141 | gcc_assert (xmode == DImode || xmode == SImode); | |
4142 | ||
4143 | /* Reload non-zero BLKmode offsets. This is because we cannot ascertain | |
4144 | BLKmode alignment. */ | |
4145 | if (GET_MODE_SIZE (mode) == 0) | |
4146 | return NULL_RTX; | |
4147 | ||
4148 | offs = low % GET_MODE_SIZE (mode); | |
4149 | ||
4150 | /* Align misaligned offset by adjusting high part to compensate. */ | |
4151 | if (offs != 0) | |
4152 | { | |
4153 | if (aarch64_uimm12_shift (high + offs)) | |
4154 | { | |
4155 | /* Align down. */ | |
4156 | low = low - offs; | |
4157 | high = high + offs; | |
4158 | } | |
4159 | else | |
4160 | { | |
4161 | /* Align up. */ | |
4162 | offs = GET_MODE_SIZE (mode) - offs; | |
4163 | low = low + offs; | |
4164 | high = high + (low & 0x1000) - offs; | |
4165 | low &= 0xfff; | |
4166 | } | |
4167 | } | |
4168 | ||
4169 | /* Check for overflow. */ | |
4170 | if (high + low != val) | |
4171 | return NULL_RTX; | |
4172 | ||
4173 | cst = GEN_INT (high); | |
4174 | if (!aarch64_uimm12_shift (high)) | |
4175 | cst = force_const_mem (xmode, cst); | |
4176 | ||
4177 | /* Reload high part into base reg, leaving the low part | |
4178 | in the mem instruction. | |
4179 | Note that replacing this gen_rtx_PLUS with plus_constant is | |
4180 | wrong in this case because we rely on the | |
4181 | (plus (plus reg c1) c2) structure being preserved so that | |
4182 | XEXP (*p, 0) in push_reload below uses the correct term. */ | |
4183 | x = gen_rtx_PLUS (xmode, | |
4184 | gen_rtx_PLUS (xmode, XEXP (x, 0), cst), | |
4185 | GEN_INT (low)); | |
4186 | ||
4187 | push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL, | |
4188 | BASE_REG_CLASS, xmode, VOIDmode, 0, 0, | |
4189 | opnum, (enum reload_type) type); | |
4190 | return x; | |
4191 | } | |
4192 | ||
4193 | return NULL_RTX; | |
4194 | } | |
4195 | ||
4196 | ||
4197 | static reg_class_t | |
4198 | aarch64_secondary_reload (bool in_p ATTRIBUTE_UNUSED, rtx x, | |
4199 | reg_class_t rclass, | |
4200 | enum machine_mode mode, | |
4201 | secondary_reload_info *sri) | |
4202 | { | |
4203 | /* Without the TARGET_SIMD instructions we cannot move a Q register | |
4204 | to a Q register directly. We need a scratch. */ | |
4205 | if (REG_P (x) && (mode == TFmode || mode == TImode) && mode == GET_MODE (x) | |
4206 | && FP_REGNUM_P (REGNO (x)) && !TARGET_SIMD | |
4207 | && reg_class_subset_p (rclass, FP_REGS)) | |
4208 | { | |
4209 | if (mode == TFmode) | |
4210 | sri->icode = CODE_FOR_aarch64_reload_movtf; | |
4211 | else if (mode == TImode) | |
4212 | sri->icode = CODE_FOR_aarch64_reload_movti; | |
4213 | return NO_REGS; | |
4214 | } | |
4215 | ||
4216 | /* A TFmode or TImode memory access should be handled via an FP_REGS | |
4217 | because AArch64 has richer addressing modes for LDR/STR instructions | |
4218 | than LDP/STP instructions. */ | |
4219 | if (!TARGET_GENERAL_REGS_ONLY && rclass == GENERAL_REGS | |
4220 | && GET_MODE_SIZE (mode) == 16 && MEM_P (x)) | |
4221 | return FP_REGS; | |
4222 | ||
4223 | if (rclass == FP_REGS && (mode == TImode || mode == TFmode) && CONSTANT_P(x)) | |
4224 | return GENERAL_REGS; | |
4225 | ||
4226 | return NO_REGS; | |
4227 | } | |
4228 | ||
4229 | static bool | |
4230 | aarch64_can_eliminate (const int from, const int to) | |
4231 | { | |
4232 | /* If we need a frame pointer, we must eliminate FRAME_POINTER_REGNUM into | |
4233 | HARD_FRAME_POINTER_REGNUM and not into STACK_POINTER_REGNUM. */ | |
4234 | ||
4235 | if (frame_pointer_needed) | |
4236 | { | |
4237 | if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
4238 | return true; | |
4239 | if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) | |
4240 | return false; | |
4241 | if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM | |
4242 | && !cfun->calls_alloca) | |
4243 | return true; | |
4244 | if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
4245 | return true; | |
4246 | ||
4247 | return false; | |
4248 | } | |
4249 | ||
4250 | return true; | |
4251 | } | |
4252 | ||
4253 | HOST_WIDE_INT | |
4254 | aarch64_initial_elimination_offset (unsigned from, unsigned to) | |
4255 | { | |
4256 | HOST_WIDE_INT frame_size; | |
4257 | HOST_WIDE_INT offset; | |
4258 | ||
4259 | aarch64_layout_frame (); | |
4260 | frame_size = (get_frame_size () + cfun->machine->frame.saved_regs_size | |
4261 | + crtl->outgoing_args_size | |
4262 | + cfun->machine->saved_varargs_size); | |
4263 | ||
4264 | frame_size = AARCH64_ROUND_UP (frame_size, STACK_BOUNDARY / BITS_PER_UNIT); | |
4265 | offset = frame_size; | |
4266 | ||
4267 | if (to == HARD_FRAME_POINTER_REGNUM) | |
4268 | { | |
4269 | if (from == ARG_POINTER_REGNUM) | |
4270 | return offset - crtl->outgoing_args_size; | |
4271 | ||
4272 | if (from == FRAME_POINTER_REGNUM) | |
4273 | return cfun->machine->frame.saved_regs_size + get_frame_size (); | |
4274 | } | |
4275 | ||
4276 | if (to == STACK_POINTER_REGNUM) | |
4277 | { | |
4278 | if (from == FRAME_POINTER_REGNUM) | |
4279 | { | |
4280 | HOST_WIDE_INT elim = crtl->outgoing_args_size | |
4281 | + cfun->machine->frame.saved_regs_size | |
4282 | + get_frame_size (); | |
4283 | elim = AARCH64_ROUND_UP (elim, STACK_BOUNDARY / BITS_PER_UNIT); | |
4284 | return elim; | |
4285 | } | |
4286 | } | |
4287 | ||
4288 | return offset; | |
4289 | } | |
4290 | ||
4291 | ||
4292 | /* Implement RETURN_ADDR_RTX. We do not support moving back to a | |
4293 | previous frame. */ | |
4294 | ||
4295 | rtx | |
4296 | aarch64_return_addr (int count, rtx frame ATTRIBUTE_UNUSED) | |
4297 | { | |
4298 | if (count != 0) | |
4299 | return const0_rtx; | |
4300 | return get_hard_reg_initial_val (Pmode, LR_REGNUM); | |
4301 | } | |
4302 | ||
4303 | ||
4304 | static void | |
4305 | aarch64_asm_trampoline_template (FILE *f) | |
4306 | { | |
4307 | if (TARGET_ILP32) | |
4308 | { | |
4309 | asm_fprintf (f, "\tldr\tw%d, .+16\n", IP1_REGNUM - R0_REGNUM); | |
4310 | asm_fprintf (f, "\tldr\tw%d, .+16\n", STATIC_CHAIN_REGNUM - R0_REGNUM); | |
4311 | } | |
4312 | else | |
4313 | { | |
4314 | asm_fprintf (f, "\tldr\t%s, .+16\n", reg_names [IP1_REGNUM]); | |
4315 | asm_fprintf (f, "\tldr\t%s, .+20\n", reg_names [STATIC_CHAIN_REGNUM]); | |
4316 | } | |
4317 | asm_fprintf (f, "\tbr\t%s\n", reg_names [IP1_REGNUM]); | |
4318 | assemble_aligned_integer (4, const0_rtx); | |
4319 | assemble_aligned_integer (POINTER_BYTES, const0_rtx); | |
4320 | assemble_aligned_integer (POINTER_BYTES, const0_rtx); | |
4321 | } | |
4322 | ||
4323 | static void | |
4324 | aarch64_trampoline_init (rtx m_tramp, tree fndecl, rtx chain_value) | |
4325 | { | |
4326 | rtx fnaddr, mem, a_tramp; | |
4327 | const int tramp_code_sz = 16; | |
4328 | ||
4329 | /* Don't need to copy the trailing D-words, we fill those in below. */ | |
4330 | emit_block_move (m_tramp, assemble_trampoline_template (), | |
4331 | GEN_INT (tramp_code_sz), BLOCK_OP_NORMAL); | |
4332 | mem = adjust_address (m_tramp, ptr_mode, tramp_code_sz); | |
4333 | fnaddr = XEXP (DECL_RTL (fndecl), 0); | |
4334 | if (GET_MODE (fnaddr) != ptr_mode) | |
4335 | fnaddr = convert_memory_address (ptr_mode, fnaddr); | |
4336 | emit_move_insn (mem, fnaddr); | |
4337 | ||
4338 | mem = adjust_address (m_tramp, ptr_mode, tramp_code_sz + POINTER_BYTES); | |
4339 | emit_move_insn (mem, chain_value); | |
4340 | ||
4341 | /* XXX We should really define a "clear_cache" pattern and use | |
4342 | gen_clear_cache(). */ | |
4343 | a_tramp = XEXP (m_tramp, 0); | |
4344 | emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__clear_cache"), | |
4345 | LCT_NORMAL, VOIDmode, 2, a_tramp, ptr_mode, | |
4346 | plus_constant (ptr_mode, a_tramp, TRAMPOLINE_SIZE), | |
4347 | ptr_mode); | |
4348 | } | |
4349 | ||
4350 | static unsigned char | |
4351 | aarch64_class_max_nregs (reg_class_t regclass, enum machine_mode mode) | |
4352 | { | |
4353 | switch (regclass) | |
4354 | { | |
4355 | case CALLER_SAVE_REGS: | |
4356 | case POINTER_REGS: | |
4357 | case GENERAL_REGS: | |
4358 | case ALL_REGS: | |
4359 | case FP_REGS: | |
4360 | case FP_LO_REGS: | |
4361 | return | |
4362 | aarch64_vector_mode_p (mode) ? (GET_MODE_SIZE (mode) + 15) / 16 : | |
4363 | (GET_MODE_SIZE (mode) + 7) / 8; | |
4364 | case STACK_REG: | |
4365 | return 1; | |
4366 | ||
4367 | case NO_REGS: | |
4368 | return 0; | |
4369 | ||
4370 | default: | |
4371 | break; | |
4372 | } | |
4373 | gcc_unreachable (); | |
4374 | } | |
4375 | ||
4376 | static reg_class_t | |
4377 | aarch64_preferred_reload_class (rtx x, reg_class_t regclass) | |
4378 | { | |
4379 | if (regclass == POINTER_REGS) | |
4380 | return GENERAL_REGS; | |
4381 | ||
4382 | if (regclass == STACK_REG) | |
4383 | { | |
4384 | if (REG_P(x) | |
4385 | && reg_class_subset_p (REGNO_REG_CLASS (REGNO (x)), POINTER_REGS)) | |
4386 | return regclass; | |
4387 | ||
4388 | return NO_REGS; | |
4389 | } | |
4390 | ||
4391 | /* If it's an integer immediate that MOVI can't handle, then | |
4392 | FP_REGS is not an option, so we return NO_REGS instead. */ | |
4393 | if (CONST_INT_P (x) && reg_class_subset_p (regclass, FP_REGS) | |
4394 | && !aarch64_simd_imm_scalar_p (x, GET_MODE (x))) | |
4395 | return NO_REGS; | |
4396 | ||
4397 | /* Register eliminiation can result in a request for | |
4398 | SP+constant->FP_REGS. We cannot support such operations which | |
4399 | use SP as source and an FP_REG as destination, so reject out | |
4400 | right now. */ | |
4401 | if (! reg_class_subset_p (regclass, GENERAL_REGS) && GET_CODE (x) == PLUS) | |
4402 | { | |
4403 | rtx lhs = XEXP (x, 0); | |
4404 | ||
4405 | /* Look through a possible SUBREG introduced by ILP32. */ | |
4406 | if (GET_CODE (lhs) == SUBREG) | |
4407 | lhs = SUBREG_REG (lhs); | |
4408 | ||
4409 | gcc_assert (REG_P (lhs)); | |
4410 | gcc_assert (reg_class_subset_p (REGNO_REG_CLASS (REGNO (lhs)), | |
4411 | POINTER_REGS)); | |
4412 | return NO_REGS; | |
4413 | } | |
4414 | ||
4415 | return regclass; | |
4416 | } | |
4417 | ||
4418 | void | |
4419 | aarch64_asm_output_labelref (FILE* f, const char *name) | |
4420 | { | |
4421 | asm_fprintf (f, "%U%s", name); | |
4422 | } | |
4423 | ||
4424 | static void | |
4425 | aarch64_elf_asm_constructor (rtx symbol, int priority) | |
4426 | { | |
4427 | if (priority == DEFAULT_INIT_PRIORITY) | |
4428 | default_ctor_section_asm_out_constructor (symbol, priority); | |
4429 | else | |
4430 | { | |
4431 | section *s; | |
4432 | char buf[18]; | |
4433 | snprintf (buf, sizeof (buf), ".init_array.%.5u", priority); | |
4434 | s = get_section (buf, SECTION_WRITE, NULL); | |
4435 | switch_to_section (s); | |
4436 | assemble_align (POINTER_SIZE); | |
4437 | assemble_aligned_integer (POINTER_BYTES, symbol); | |
4438 | } | |
4439 | } | |
4440 | ||
4441 | static void | |
4442 | aarch64_elf_asm_destructor (rtx symbol, int priority) | |
4443 | { | |
4444 | if (priority == DEFAULT_INIT_PRIORITY) | |
4445 | default_dtor_section_asm_out_destructor (symbol, priority); | |
4446 | else | |
4447 | { | |
4448 | section *s; | |
4449 | char buf[18]; | |
4450 | snprintf (buf, sizeof (buf), ".fini_array.%.5u", priority); | |
4451 | s = get_section (buf, SECTION_WRITE, NULL); | |
4452 | switch_to_section (s); | |
4453 | assemble_align (POINTER_SIZE); | |
4454 | assemble_aligned_integer (POINTER_BYTES, symbol); | |
4455 | } | |
4456 | } | |
4457 | ||
4458 | const char* | |
4459 | aarch64_output_casesi (rtx *operands) | |
4460 | { | |
4461 | char buf[100]; | |
4462 | char label[100]; | |
4463 | rtx diff_vec = PATTERN (NEXT_INSN (operands[2])); | |
4464 | int index; | |
4465 | static const char *const patterns[4][2] = | |
4466 | { | |
4467 | { | |
4468 | "ldrb\t%w3, [%0,%w1,uxtw]", | |
4469 | "add\t%3, %4, %w3, sxtb #2" | |
4470 | }, | |
4471 | { | |
4472 | "ldrh\t%w3, [%0,%w1,uxtw #1]", | |
4473 | "add\t%3, %4, %w3, sxth #2" | |
4474 | }, | |
4475 | { | |
4476 | "ldr\t%w3, [%0,%w1,uxtw #2]", | |
4477 | "add\t%3, %4, %w3, sxtw #2" | |
4478 | }, | |
4479 | /* We assume that DImode is only generated when not optimizing and | |
4480 | that we don't really need 64-bit address offsets. That would | |
4481 | imply an object file with 8GB of code in a single function! */ | |
4482 | { | |
4483 | "ldr\t%w3, [%0,%w1,uxtw #2]", | |
4484 | "add\t%3, %4, %w3, sxtw #2" | |
4485 | } | |
4486 | }; | |
4487 | ||
4488 | gcc_assert (GET_CODE (diff_vec) == ADDR_DIFF_VEC); | |
4489 | ||
4490 | index = exact_log2 (GET_MODE_SIZE (GET_MODE (diff_vec))); | |
4491 | ||
4492 | gcc_assert (index >= 0 && index <= 3); | |
4493 | ||
4494 | /* Need to implement table size reduction, by chaning the code below. */ | |
4495 | output_asm_insn (patterns[index][0], operands); | |
4496 | ASM_GENERATE_INTERNAL_LABEL (label, "Lrtx", CODE_LABEL_NUMBER (operands[2])); | |
4497 | snprintf (buf, sizeof (buf), | |
4498 | "adr\t%%4, %s", targetm.strip_name_encoding (label)); | |
4499 | output_asm_insn (buf, operands); | |
4500 | output_asm_insn (patterns[index][1], operands); | |
4501 | output_asm_insn ("br\t%3", operands); | |
4502 | assemble_label (asm_out_file, label); | |
4503 | return ""; | |
4504 | } | |
4505 | ||
4506 | ||
4507 | /* Return size in bits of an arithmetic operand which is shifted/scaled and | |
4508 | masked such that it is suitable for a UXTB, UXTH, or UXTW extend | |
4509 | operator. */ | |
4510 | ||
4511 | int | |
4512 | aarch64_uxt_size (int shift, HOST_WIDE_INT mask) | |
4513 | { | |
4514 | if (shift >= 0 && shift <= 3) | |
4515 | { | |
4516 | int size; | |
4517 | for (size = 8; size <= 32; size *= 2) | |
4518 | { | |
4519 | HOST_WIDE_INT bits = ((HOST_WIDE_INT)1U << size) - 1; | |
4520 | if (mask == bits << shift) | |
4521 | return size; | |
4522 | } | |
4523 | } | |
4524 | return 0; | |
4525 | } | |
4526 | ||
4527 | static bool | |
4528 | aarch64_use_blocks_for_constant_p (enum machine_mode mode ATTRIBUTE_UNUSED, | |
4529 | const_rtx x ATTRIBUTE_UNUSED) | |
4530 | { | |
4531 | /* We can't use blocks for constants when we're using a per-function | |
4532 | constant pool. */ | |
4533 | return false; | |
4534 | } | |
4535 | ||
4536 | static section * | |
4537 | aarch64_select_rtx_section (enum machine_mode mode ATTRIBUTE_UNUSED, | |
4538 | rtx x ATTRIBUTE_UNUSED, | |
4539 | unsigned HOST_WIDE_INT align ATTRIBUTE_UNUSED) | |
4540 | { | |
4541 | /* Force all constant pool entries into the current function section. */ | |
4542 | return function_section (current_function_decl); | |
4543 | } | |
4544 | ||
4545 | ||
4546 | /* Costs. */ | |
4547 | ||
4548 | /* Helper function for rtx cost calculation. Strip a shift expression | |
4549 | from X. Returns the inner operand if successful, or the original | |
4550 | expression on failure. */ | |
4551 | static rtx | |
4552 | aarch64_strip_shift (rtx x) | |
4553 | { | |
4554 | rtx op = x; | |
4555 | ||
4556 | /* We accept both ROTATERT and ROTATE: since the RHS must be a constant | |
4557 | we can convert both to ROR during final output. */ | |
4558 | if ((GET_CODE (op) == ASHIFT | |
4559 | || GET_CODE (op) == ASHIFTRT | |
4560 | || GET_CODE (op) == LSHIFTRT | |
4561 | || GET_CODE (op) == ROTATERT | |
4562 | || GET_CODE (op) == ROTATE) | |
4563 | && CONST_INT_P (XEXP (op, 1))) | |
4564 | return XEXP (op, 0); | |
4565 | ||
4566 | if (GET_CODE (op) == MULT | |
4567 | && CONST_INT_P (XEXP (op, 1)) | |
4568 | && ((unsigned) exact_log2 (INTVAL (XEXP (op, 1)))) < 64) | |
4569 | return XEXP (op, 0); | |
4570 | ||
4571 | return x; | |
4572 | } | |
4573 | ||
4574 | /* Helper function for rtx cost calculation. Strip an extend | |
4575 | expression from X. Returns the inner operand if successful, or the | |
4576 | original expression on failure. We deal with a number of possible | |
4577 | canonicalization variations here. */ | |
4578 | static rtx | |
4579 | aarch64_strip_extend (rtx x) | |
4580 | { | |
4581 | rtx op = x; | |
4582 | ||
4583 | /* Zero and sign extraction of a widened value. */ | |
4584 | if ((GET_CODE (op) == ZERO_EXTRACT || GET_CODE (op) == SIGN_EXTRACT) | |
4585 | && XEXP (op, 2) == const0_rtx | |
4586 | && GET_CODE (XEXP (op, 0)) == MULT | |
4587 | && aarch64_is_extend_from_extract (GET_MODE (op), XEXP (XEXP (op, 0), 1), | |
4588 | XEXP (op, 1))) | |
4589 | return XEXP (XEXP (op, 0), 0); | |
4590 | ||
4591 | /* It can also be represented (for zero-extend) as an AND with an | |
4592 | immediate. */ | |
4593 | if (GET_CODE (op) == AND | |
4594 | && GET_CODE (XEXP (op, 0)) == MULT | |
4595 | && CONST_INT_P (XEXP (XEXP (op, 0), 1)) | |
4596 | && CONST_INT_P (XEXP (op, 1)) | |
4597 | && aarch64_uxt_size (exact_log2 (INTVAL (XEXP (XEXP (op, 0), 1))), | |
4598 | INTVAL (XEXP (op, 1))) != 0) | |
4599 | return XEXP (XEXP (op, 0), 0); | |
4600 | ||
4601 | /* Now handle extended register, as this may also have an optional | |
4602 | left shift by 1..4. */ | |
4603 | if (GET_CODE (op) == ASHIFT | |
4604 | && CONST_INT_P (XEXP (op, 1)) | |
4605 | && ((unsigned HOST_WIDE_INT) INTVAL (XEXP (op, 1))) <= 4) | |
4606 | op = XEXP (op, 0); | |
4607 | ||
4608 | if (GET_CODE (op) == ZERO_EXTEND | |
4609 | || GET_CODE (op) == SIGN_EXTEND) | |
4610 | op = XEXP (op, 0); | |
4611 | ||
4612 | if (op != x) | |
4613 | return op; | |
4614 | ||
4615 | return x; | |
4616 | } | |
4617 | ||
4618 | /* Helper function for rtx cost calculation. Calculate the cost of | |
4619 | a MULT, which may be part of a multiply-accumulate rtx. Return | |
4620 | the calculated cost of the expression, recursing manually in to | |
4621 | operands where needed. */ | |
4622 | ||
4623 | static int | |
4624 | aarch64_rtx_mult_cost (rtx x, int code, int outer, bool speed) | |
4625 | { | |
4626 | rtx op0, op1; | |
4627 | const struct cpu_cost_table *extra_cost | |
4628 | = aarch64_tune_params->insn_extra_cost; | |
4629 | int cost = 0; | |
4630 | bool maybe_fma = (outer == PLUS || outer == MINUS); | |
4631 | enum machine_mode mode = GET_MODE (x); | |
4632 | ||
4633 | gcc_checking_assert (code == MULT); | |
4634 | ||
4635 | op0 = XEXP (x, 0); | |
4636 | op1 = XEXP (x, 1); | |
4637 | ||
4638 | if (VECTOR_MODE_P (mode)) | |
4639 | mode = GET_MODE_INNER (mode); | |
4640 | ||
4641 | /* Integer multiply/fma. */ | |
4642 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
4643 | { | |
4644 | /* The multiply will be canonicalized as a shift, cost it as such. */ | |
4645 | if (CONST_INT_P (op1) | |
4646 | && exact_log2 (INTVAL (op1)) > 0) | |
4647 | { | |
4648 | if (speed) | |
4649 | { | |
4650 | if (maybe_fma) | |
4651 | /* ADD (shifted register). */ | |
4652 | cost += extra_cost->alu.arith_shift; | |
4653 | else | |
4654 | /* LSL (immediate). */ | |
4655 | cost += extra_cost->alu.shift; | |
4656 | } | |
4657 | ||
4658 | cost += rtx_cost (op0, GET_CODE (op0), 0, speed); | |
4659 | ||
4660 | return cost; | |
4661 | } | |
4662 | ||
4663 | /* Integer multiplies or FMAs have zero/sign extending variants. */ | |
4664 | if ((GET_CODE (op0) == ZERO_EXTEND | |
4665 | && GET_CODE (op1) == ZERO_EXTEND) | |
4666 | || (GET_CODE (op0) == SIGN_EXTEND | |
4667 | && GET_CODE (op1) == SIGN_EXTEND)) | |
4668 | { | |
4669 | cost += rtx_cost (XEXP (op0, 0), MULT, 0, speed) | |
4670 | + rtx_cost (XEXP (op1, 0), MULT, 1, speed); | |
4671 | ||
4672 | if (speed) | |
4673 | { | |
4674 | if (maybe_fma) | |
4675 | /* MADD/SMADDL/UMADDL. */ | |
4676 | cost += extra_cost->mult[0].extend_add; | |
4677 | else | |
4678 | /* MUL/SMULL/UMULL. */ | |
4679 | cost += extra_cost->mult[0].extend; | |
4680 | } | |
4681 | ||
4682 | return cost; | |
4683 | } | |
4684 | ||
4685 | /* This is either an integer multiply or an FMA. In both cases | |
4686 | we want to recurse and cost the operands. */ | |
4687 | cost += rtx_cost (op0, MULT, 0, speed) | |
4688 | + rtx_cost (op1, MULT, 1, speed); | |
4689 | ||
4690 | if (speed) | |
4691 | { | |
4692 | if (maybe_fma) | |
4693 | /* MADD. */ | |
4694 | cost += extra_cost->mult[mode == DImode].add; | |
4695 | else | |
4696 | /* MUL. */ | |
4697 | cost += extra_cost->mult[mode == DImode].simple; | |
4698 | } | |
4699 | ||
4700 | return cost; | |
4701 | } | |
4702 | else | |
4703 | { | |
4704 | if (speed) | |
4705 | { | |
4706 | /* Floating-point FMA/FMUL can also support negations of the | |
4707 | operands. */ | |
4708 | if (GET_CODE (op0) == NEG) | |
4709 | op0 = XEXP (op0, 0); | |
4710 | if (GET_CODE (op1) == NEG) | |
4711 | op1 = XEXP (op1, 0); | |
4712 | ||
4713 | if (maybe_fma) | |
4714 | /* FMADD/FNMADD/FNMSUB/FMSUB. */ | |
4715 | cost += extra_cost->fp[mode == DFmode].fma; | |
4716 | else | |
4717 | /* FMUL/FNMUL. */ | |
4718 | cost += extra_cost->fp[mode == DFmode].mult; | |
4719 | } | |
4720 | ||
4721 | cost += rtx_cost (op0, MULT, 0, speed) | |
4722 | + rtx_cost (op1, MULT, 1, speed); | |
4723 | return cost; | |
4724 | } | |
4725 | } | |
4726 | ||
4727 | static int | |
4728 | aarch64_address_cost (rtx x, | |
4729 | enum machine_mode mode, | |
4730 | addr_space_t as ATTRIBUTE_UNUSED, | |
4731 | bool speed) | |
4732 | { | |
4733 | enum rtx_code c = GET_CODE (x); | |
4734 | const struct cpu_addrcost_table *addr_cost = aarch64_tune_params->addr_cost; | |
4735 | struct aarch64_address_info info; | |
4736 | int cost = 0; | |
4737 | info.shift = 0; | |
4738 | ||
4739 | if (!aarch64_classify_address (&info, x, mode, c, false)) | |
4740 | { | |
4741 | if (GET_CODE (x) == CONST || GET_CODE (x) == SYMBOL_REF) | |
4742 | { | |
4743 | /* This is a CONST or SYMBOL ref which will be split | |
4744 | in a different way depending on the code model in use. | |
4745 | Cost it through the generic infrastructure. */ | |
4746 | int cost_symbol_ref = rtx_cost (x, MEM, 1, speed); | |
4747 | /* Divide through by the cost of one instruction to | |
4748 | bring it to the same units as the address costs. */ | |
4749 | cost_symbol_ref /= COSTS_N_INSNS (1); | |
4750 | /* The cost is then the cost of preparing the address, | |
4751 | followed by an immediate (possibly 0) offset. */ | |
4752 | return cost_symbol_ref + addr_cost->imm_offset; | |
4753 | } | |
4754 | else | |
4755 | { | |
4756 | /* This is most likely a jump table from a case | |
4757 | statement. */ | |
4758 | return addr_cost->register_offset; | |
4759 | } | |
4760 | } | |
4761 | ||
4762 | switch (info.type) | |
4763 | { | |
4764 | case ADDRESS_LO_SUM: | |
4765 | case ADDRESS_SYMBOLIC: | |
4766 | case ADDRESS_REG_IMM: | |
4767 | cost += addr_cost->imm_offset; | |
4768 | break; | |
4769 | ||
4770 | case ADDRESS_REG_WB: | |
4771 | if (c == PRE_INC || c == PRE_DEC || c == PRE_MODIFY) | |
4772 | cost += addr_cost->pre_modify; | |
4773 | else if (c == POST_INC || c == POST_DEC || c == POST_MODIFY) | |
4774 | cost += addr_cost->post_modify; | |
4775 | else | |
4776 | gcc_unreachable (); | |
4777 | ||
4778 | break; | |
4779 | ||
4780 | case ADDRESS_REG_REG: | |
4781 | cost += addr_cost->register_offset; | |
4782 | break; | |
4783 | ||
4784 | case ADDRESS_REG_UXTW: | |
4785 | case ADDRESS_REG_SXTW: | |
4786 | cost += addr_cost->register_extend; | |
4787 | break; | |
4788 | ||
4789 | default: | |
4790 | gcc_unreachable (); | |
4791 | } | |
4792 | ||
4793 | ||
4794 | if (info.shift > 0) | |
4795 | { | |
4796 | /* For the sake of calculating the cost of the shifted register | |
4797 | component, we can treat same sized modes in the same way. */ | |
4798 | switch (GET_MODE_BITSIZE (mode)) | |
4799 | { | |
4800 | case 16: | |
4801 | cost += addr_cost->addr_scale_costs.hi; | |
4802 | break; | |
4803 | ||
4804 | case 32: | |
4805 | cost += addr_cost->addr_scale_costs.si; | |
4806 | break; | |
4807 | ||
4808 | case 64: | |
4809 | cost += addr_cost->addr_scale_costs.di; | |
4810 | break; | |
4811 | ||
4812 | /* We can't tell, or this is a 128-bit vector. */ | |
4813 | default: | |
4814 | cost += addr_cost->addr_scale_costs.ti; | |
4815 | break; | |
4816 | } | |
4817 | } | |
4818 | ||
4819 | return cost; | |
4820 | } | |
4821 | ||
4822 | /* Return true if the RTX X in mode MODE is a zero or sign extract | |
4823 | usable in an ADD or SUB (extended register) instruction. */ | |
4824 | static bool | |
4825 | aarch64_rtx_arith_op_extract_p (rtx x, enum machine_mode mode) | |
4826 | { | |
4827 | /* Catch add with a sign extract. | |
4828 | This is add_<optab><mode>_multp2. */ | |
4829 | if (GET_CODE (x) == SIGN_EXTRACT | |
4830 | || GET_CODE (x) == ZERO_EXTRACT) | |
4831 | { | |
4832 | rtx op0 = XEXP (x, 0); | |
4833 | rtx op1 = XEXP (x, 1); | |
4834 | rtx op2 = XEXP (x, 2); | |
4835 | ||
4836 | if (GET_CODE (op0) == MULT | |
4837 | && CONST_INT_P (op1) | |
4838 | && op2 == const0_rtx | |
4839 | && CONST_INT_P (XEXP (op0, 1)) | |
4840 | && aarch64_is_extend_from_extract (mode, | |
4841 | XEXP (op0, 1), | |
4842 | op1)) | |
4843 | { | |
4844 | return true; | |
4845 | } | |
4846 | } | |
4847 | ||
4848 | return false; | |
4849 | } | |
4850 | ||
4851 | /* Calculate the cost of calculating X, storing it in *COST. Result | |
4852 | is true if the total cost of the operation has now been calculated. */ | |
4853 | static bool | |
4854 | aarch64_rtx_costs (rtx x, int code, int outer ATTRIBUTE_UNUSED, | |
4855 | int param ATTRIBUTE_UNUSED, int *cost, bool speed) | |
4856 | { | |
4857 | rtx op0, op1, op2; | |
4858 | const struct cpu_cost_table *extra_cost | |
4859 | = aarch64_tune_params->insn_extra_cost; | |
4860 | enum machine_mode mode = GET_MODE (x); | |
4861 | ||
4862 | /* By default, assume that everything has equivalent cost to the | |
4863 | cheapest instruction. Any additional costs are applied as a delta | |
4864 | above this default. */ | |
4865 | *cost = COSTS_N_INSNS (1); | |
4866 | ||
4867 | /* TODO: The cost infrastructure currently does not handle | |
4868 | vector operations. Assume that all vector operations | |
4869 | are equally expensive. */ | |
4870 | if (VECTOR_MODE_P (mode)) | |
4871 | { | |
4872 | if (speed) | |
4873 | *cost += extra_cost->vect.alu; | |
4874 | return true; | |
4875 | } | |
4876 | ||
4877 | switch (code) | |
4878 | { | |
4879 | case SET: | |
4880 | /* The cost depends entirely on the operands to SET. */ | |
4881 | *cost = 0; | |
4882 | op0 = SET_DEST (x); | |
4883 | op1 = SET_SRC (x); | |
4884 | ||
4885 | switch (GET_CODE (op0)) | |
4886 | { | |
4887 | case MEM: | |
4888 | if (speed) | |
4889 | { | |
4890 | rtx address = XEXP (op0, 0); | |
4891 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
4892 | *cost += extra_cost->ldst.store; | |
4893 | else if (mode == SFmode) | |
4894 | *cost += extra_cost->ldst.storef; | |
4895 | else if (mode == DFmode) | |
4896 | *cost += extra_cost->ldst.stored; | |
4897 | ||
4898 | *cost += | |
4899 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
4900 | 0, speed)); | |
4901 | } | |
4902 | ||
4903 | *cost += rtx_cost (op1, SET, 1, speed); | |
4904 | return true; | |
4905 | ||
4906 | case SUBREG: | |
4907 | if (! REG_P (SUBREG_REG (op0))) | |
4908 | *cost += rtx_cost (SUBREG_REG (op0), SET, 0, speed); | |
4909 | ||
4910 | /* Fall through. */ | |
4911 | case REG: | |
4912 | /* const0_rtx is in general free, but we will use an | |
4913 | instruction to set a register to 0. */ | |
4914 | if (REG_P (op1) || op1 == const0_rtx) | |
4915 | { | |
4916 | /* The cost is 1 per register copied. */ | |
4917 | int n_minus_1 = (GET_MODE_SIZE (GET_MODE (op0)) - 1) | |
4918 | / UNITS_PER_WORD; | |
4919 | *cost = COSTS_N_INSNS (n_minus_1 + 1); | |
4920 | } | |
4921 | else | |
4922 | /* Cost is just the cost of the RHS of the set. */ | |
4923 | *cost += rtx_cost (op1, SET, 1, speed); | |
4924 | return true; | |
4925 | ||
4926 | case ZERO_EXTRACT: | |
4927 | case SIGN_EXTRACT: | |
4928 | /* Bit-field insertion. Strip any redundant widening of | |
4929 | the RHS to meet the width of the target. */ | |
4930 | if (GET_CODE (op1) == SUBREG) | |
4931 | op1 = SUBREG_REG (op1); | |
4932 | if ((GET_CODE (op1) == ZERO_EXTEND | |
4933 | || GET_CODE (op1) == SIGN_EXTEND) | |
4934 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
4935 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (op1, 0))) | |
4936 | >= INTVAL (XEXP (op0, 1)))) | |
4937 | op1 = XEXP (op1, 0); | |
4938 | ||
4939 | if (CONST_INT_P (op1)) | |
4940 | { | |
4941 | /* MOV immediate is assumed to always be cheap. */ | |
4942 | *cost = COSTS_N_INSNS (1); | |
4943 | } | |
4944 | else | |
4945 | { | |
4946 | /* BFM. */ | |
4947 | if (speed) | |
4948 | *cost += extra_cost->alu.bfi; | |
4949 | *cost += rtx_cost (op1, (enum rtx_code) code, 1, speed); | |
4950 | } | |
4951 | ||
4952 | return true; | |
4953 | ||
4954 | default: | |
4955 | /* We can't make sense of this, assume default cost. */ | |
4956 | *cost = COSTS_N_INSNS (1); | |
4957 | break; | |
4958 | } | |
4959 | return false; | |
4960 | ||
4961 | case CONST_INT: | |
4962 | /* If an instruction can incorporate a constant within the | |
4963 | instruction, the instruction's expression avoids calling | |
4964 | rtx_cost() on the constant. If rtx_cost() is called on a | |
4965 | constant, then it is usually because the constant must be | |
4966 | moved into a register by one or more instructions. | |
4967 | ||
4968 | The exception is constant 0, which can be expressed | |
4969 | as XZR/WZR and is therefore free. The exception to this is | |
4970 | if we have (set (reg) (const0_rtx)) in which case we must cost | |
4971 | the move. However, we can catch that when we cost the SET, so | |
4972 | we don't need to consider that here. */ | |
4973 | if (x == const0_rtx) | |
4974 | *cost = 0; | |
4975 | else | |
4976 | { | |
4977 | /* To an approximation, building any other constant is | |
4978 | proportionally expensive to the number of instructions | |
4979 | required to build that constant. This is true whether we | |
4980 | are compiling for SPEED or otherwise. */ | |
4981 | *cost = COSTS_N_INSNS (aarch64_build_constant (0, | |
4982 | INTVAL (x), | |
4983 | false)); | |
4984 | } | |
4985 | return true; | |
4986 | ||
4987 | case CONST_DOUBLE: | |
4988 | if (speed) | |
4989 | { | |
4990 | /* mov[df,sf]_aarch64. */ | |
4991 | if (aarch64_float_const_representable_p (x)) | |
4992 | /* FMOV (scalar immediate). */ | |
4993 | *cost += extra_cost->fp[mode == DFmode].fpconst; | |
4994 | else if (!aarch64_float_const_zero_rtx_p (x)) | |
4995 | { | |
4996 | /* This will be a load from memory. */ | |
4997 | if (mode == DFmode) | |
4998 | *cost += extra_cost->ldst.loadd; | |
4999 | else | |
5000 | *cost += extra_cost->ldst.loadf; | |
5001 | } | |
5002 | else | |
5003 | /* Otherwise this is +0.0. We get this using MOVI d0, #0 | |
5004 | or MOV v0.s[0], wzr - neither of which are modeled by the | |
5005 | cost tables. Just use the default cost. */ | |
5006 | { | |
5007 | } | |
5008 | } | |
5009 | ||
5010 | return true; | |
5011 | ||
5012 | case MEM: | |
5013 | if (speed) | |
5014 | { | |
5015 | /* For loads we want the base cost of a load, plus an | |
5016 | approximation for the additional cost of the addressing | |
5017 | mode. */ | |
5018 | rtx address = XEXP (x, 0); | |
5019 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
5020 | *cost += extra_cost->ldst.load; | |
5021 | else if (mode == SFmode) | |
5022 | *cost += extra_cost->ldst.loadf; | |
5023 | else if (mode == DFmode) | |
5024 | *cost += extra_cost->ldst.loadd; | |
5025 | ||
5026 | *cost += | |
5027 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
5028 | 0, speed)); | |
5029 | } | |
5030 | ||
5031 | return true; | |
5032 | ||
5033 | case NEG: | |
5034 | op0 = XEXP (x, 0); | |
5035 | ||
5036 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT) | |
5037 | { | |
5038 | if (GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMPARE | |
5039 | || GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMM_COMPARE) | |
5040 | { | |
5041 | /* CSETM. */ | |
5042 | *cost += rtx_cost (XEXP (op0, 0), NEG, 0, speed); | |
5043 | return true; | |
5044 | } | |
5045 | ||
5046 | /* Cost this as SUB wzr, X. */ | |
5047 | op0 = CONST0_RTX (GET_MODE (x)); | |
5048 | op1 = XEXP (x, 0); | |
5049 | goto cost_minus; | |
5050 | } | |
5051 | ||
5052 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
5053 | { | |
5054 | /* Support (neg(fma...)) as a single instruction only if | |
5055 | sign of zeros is unimportant. This matches the decision | |
5056 | making in aarch64.md. */ | |
5057 | if (GET_CODE (op0) == FMA && !HONOR_SIGNED_ZEROS (GET_MODE (op0))) | |
5058 | { | |
5059 | /* FNMADD. */ | |
5060 | *cost = rtx_cost (op0, NEG, 0, speed); | |
5061 | return true; | |
5062 | } | |
5063 | if (speed) | |
5064 | /* FNEG. */ | |
5065 | *cost += extra_cost->fp[mode == DFmode].neg; | |
5066 | return false; | |
5067 | } | |
5068 | ||
5069 | return false; | |
5070 | ||
5071 | case COMPARE: | |
5072 | op0 = XEXP (x, 0); | |
5073 | op1 = XEXP (x, 1); | |
5074 | ||
5075 | if (op1 == const0_rtx | |
5076 | && GET_CODE (op0) == AND) | |
5077 | { | |
5078 | x = op0; | |
5079 | goto cost_logic; | |
5080 | } | |
5081 | ||
5082 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT) | |
5083 | { | |
5084 | /* TODO: A write to the CC flags possibly costs extra, this | |
5085 | needs encoding in the cost tables. */ | |
5086 | ||
5087 | /* CC_ZESWPmode supports zero extend for free. */ | |
5088 | if (GET_MODE (x) == CC_ZESWPmode && GET_CODE (op0) == ZERO_EXTEND) | |
5089 | op0 = XEXP (op0, 0); | |
5090 | ||
5091 | /* ANDS. */ | |
5092 | if (GET_CODE (op0) == AND) | |
5093 | { | |
5094 | x = op0; | |
5095 | goto cost_logic; | |
5096 | } | |
5097 | ||
5098 | if (GET_CODE (op0) == PLUS) | |
5099 | { | |
5100 | /* ADDS (and CMN alias). */ | |
5101 | x = op0; | |
5102 | goto cost_plus; | |
5103 | } | |
5104 | ||
5105 | if (GET_CODE (op0) == MINUS) | |
5106 | { | |
5107 | /* SUBS. */ | |
5108 | x = op0; | |
5109 | goto cost_minus; | |
5110 | } | |
5111 | ||
5112 | if (GET_CODE (op1) == NEG) | |
5113 | { | |
5114 | /* CMN. */ | |
5115 | if (speed) | |
5116 | *cost += extra_cost->alu.arith; | |
5117 | ||
5118 | *cost += rtx_cost (op0, COMPARE, 0, speed); | |
5119 | *cost += rtx_cost (XEXP (op1, 0), NEG, 1, speed); | |
5120 | return true; | |
5121 | } | |
5122 | ||
5123 | /* CMP. | |
5124 | ||
5125 | Compare can freely swap the order of operands, and | |
5126 | canonicalization puts the more complex operation first. | |
5127 | But the integer MINUS logic expects the shift/extend | |
5128 | operation in op1. */ | |
5129 | if (! (REG_P (op0) | |
5130 | || (GET_CODE (op0) == SUBREG && REG_P (SUBREG_REG (op0))))) | |
5131 | { | |
5132 | op0 = XEXP (x, 1); | |
5133 | op1 = XEXP (x, 0); | |
5134 | } | |
5135 | goto cost_minus; | |
5136 | } | |
5137 | ||
5138 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT) | |
5139 | { | |
5140 | /* FCMP. */ | |
5141 | if (speed) | |
5142 | *cost += extra_cost->fp[mode == DFmode].compare; | |
5143 | ||
5144 | if (CONST_DOUBLE_P (op1) && aarch64_float_const_zero_rtx_p (op1)) | |
5145 | { | |
5146 | /* FCMP supports constant 0.0 for no extra cost. */ | |
5147 | return true; | |
5148 | } | |
5149 | return false; | |
5150 | } | |
5151 | ||
5152 | return false; | |
5153 | ||
5154 | case MINUS: | |
5155 | { | |
5156 | op0 = XEXP (x, 0); | |
5157 | op1 = XEXP (x, 1); | |
5158 | ||
5159 | cost_minus: | |
5160 | /* Detect valid immediates. */ | |
5161 | if ((GET_MODE_CLASS (mode) == MODE_INT | |
5162 | || (GET_MODE_CLASS (mode) == MODE_CC | |
5163 | && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT)) | |
5164 | && CONST_INT_P (op1) | |
5165 | && aarch64_uimm12_shift (INTVAL (op1))) | |
5166 | { | |
5167 | *cost += rtx_cost (op0, MINUS, 0, speed); | |
5168 | ||
5169 | if (speed) | |
5170 | /* SUB(S) (immediate). */ | |
5171 | *cost += extra_cost->alu.arith; | |
5172 | return true; | |
5173 | ||
5174 | } | |
5175 | ||
5176 | /* Look for SUB (extended register). */ | |
5177 | if (aarch64_rtx_arith_op_extract_p (op1, mode)) | |
5178 | { | |
5179 | if (speed) | |
5180 | *cost += extra_cost->alu.arith_shift; | |
5181 | ||
5182 | *cost += rtx_cost (XEXP (XEXP (op1, 0), 0), | |
5183 | (enum rtx_code) GET_CODE (op1), | |
5184 | 0, speed); | |
5185 | return true; | |
5186 | } | |
5187 | ||
5188 | rtx new_op1 = aarch64_strip_extend (op1); | |
5189 | ||
5190 | /* Cost this as an FMA-alike operation. */ | |
5191 | if ((GET_CODE (new_op1) == MULT | |
5192 | || GET_CODE (new_op1) == ASHIFT) | |
5193 | && code != COMPARE) | |
5194 | { | |
5195 | *cost += aarch64_rtx_mult_cost (new_op1, MULT, | |
5196 | (enum rtx_code) code, | |
5197 | speed); | |
5198 | *cost += rtx_cost (op0, MINUS, 0, speed); | |
5199 | return true; | |
5200 | } | |
5201 | ||
5202 | *cost += rtx_cost (new_op1, MINUS, 1, speed); | |
5203 | ||
5204 | if (speed) | |
5205 | { | |
5206 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
5207 | /* SUB(S). */ | |
5208 | *cost += extra_cost->alu.arith; | |
5209 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5210 | /* FSUB. */ | |
5211 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
5212 | } | |
5213 | return true; | |
5214 | } | |
5215 | ||
5216 | case PLUS: | |
5217 | { | |
5218 | rtx new_op0; | |
5219 | ||
5220 | op0 = XEXP (x, 0); | |
5221 | op1 = XEXP (x, 1); | |
5222 | ||
5223 | cost_plus: | |
5224 | if (GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMPARE | |
5225 | || GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMM_COMPARE) | |
5226 | { | |
5227 | /* CSINC. */ | |
5228 | *cost += rtx_cost (XEXP (op0, 0), PLUS, 0, speed); | |
5229 | *cost += rtx_cost (op1, PLUS, 1, speed); | |
5230 | return true; | |
5231 | } | |
5232 | ||
5233 | if (GET_MODE_CLASS (mode) == MODE_INT | |
5234 | && CONST_INT_P (op1) | |
5235 | && aarch64_uimm12_shift (INTVAL (op1))) | |
5236 | { | |
5237 | *cost += rtx_cost (op0, PLUS, 0, speed); | |
5238 | ||
5239 | if (speed) | |
5240 | /* ADD (immediate). */ | |
5241 | *cost += extra_cost->alu.arith; | |
5242 | return true; | |
5243 | } | |
5244 | ||
5245 | /* Look for ADD (extended register). */ | |
5246 | if (aarch64_rtx_arith_op_extract_p (op0, mode)) | |
5247 | { | |
5248 | if (speed) | |
5249 | *cost += extra_cost->alu.arith_shift; | |
5250 | ||
5251 | *cost += rtx_cost (XEXP (XEXP (op0, 0), 0), | |
5252 | (enum rtx_code) GET_CODE (op0), | |
5253 | 0, speed); | |
5254 | return true; | |
5255 | } | |
5256 | ||
5257 | /* Strip any extend, leave shifts behind as we will | |
5258 | cost them through mult_cost. */ | |
5259 | new_op0 = aarch64_strip_extend (op0); | |
5260 | ||
5261 | if (GET_CODE (new_op0) == MULT | |
5262 | || GET_CODE (new_op0) == ASHIFT) | |
5263 | { | |
5264 | *cost += aarch64_rtx_mult_cost (new_op0, MULT, PLUS, | |
5265 | speed); | |
5266 | *cost += rtx_cost (op1, PLUS, 1, speed); | |
5267 | return true; | |
5268 | } | |
5269 | ||
5270 | *cost += (rtx_cost (new_op0, PLUS, 0, speed) | |
5271 | + rtx_cost (op1, PLUS, 1, speed)); | |
5272 | ||
5273 | if (speed) | |
5274 | { | |
5275 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
5276 | /* ADD. */ | |
5277 | *cost += extra_cost->alu.arith; | |
5278 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5279 | /* FADD. */ | |
5280 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
5281 | } | |
5282 | return true; | |
5283 | } | |
5284 | ||
5285 | case BSWAP: | |
5286 | *cost = COSTS_N_INSNS (1); | |
5287 | ||
5288 | if (speed) | |
5289 | *cost += extra_cost->alu.rev; | |
5290 | ||
5291 | return false; | |
5292 | ||
5293 | case IOR: | |
5294 | if (aarch_rev16_p (x)) | |
5295 | { | |
5296 | *cost = COSTS_N_INSNS (1); | |
5297 | ||
5298 | if (speed) | |
5299 | *cost += extra_cost->alu.rev; | |
5300 | ||
5301 | return true; | |
5302 | } | |
5303 | /* Fall through. */ | |
5304 | case XOR: | |
5305 | case AND: | |
5306 | cost_logic: | |
5307 | op0 = XEXP (x, 0); | |
5308 | op1 = XEXP (x, 1); | |
5309 | ||
5310 | if (code == AND | |
5311 | && GET_CODE (op0) == MULT | |
5312 | && CONST_INT_P (XEXP (op0, 1)) | |
5313 | && CONST_INT_P (op1) | |
5314 | && aarch64_uxt_size (exact_log2 (INTVAL (XEXP (op0, 1))), | |
5315 | INTVAL (op1)) != 0) | |
5316 | { | |
5317 | /* This is a UBFM/SBFM. */ | |
5318 | *cost += rtx_cost (XEXP (op0, 0), ZERO_EXTRACT, 0, speed); | |
5319 | if (speed) | |
5320 | *cost += extra_cost->alu.bfx; | |
5321 | return true; | |
5322 | } | |
5323 | ||
5324 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT) | |
5325 | { | |
5326 | /* We possibly get the immediate for free, this is not | |
5327 | modelled. */ | |
5328 | if (CONST_INT_P (op1) | |
5329 | && aarch64_bitmask_imm (INTVAL (op1), GET_MODE (x))) | |
5330 | { | |
5331 | *cost += rtx_cost (op0, (enum rtx_code) code, 0, speed); | |
5332 | ||
5333 | if (speed) | |
5334 | *cost += extra_cost->alu.logical; | |
5335 | ||
5336 | return true; | |
5337 | } | |
5338 | else | |
5339 | { | |
5340 | rtx new_op0 = op0; | |
5341 | ||
5342 | /* Handle ORN, EON, or BIC. */ | |
5343 | if (GET_CODE (op0) == NOT) | |
5344 | op0 = XEXP (op0, 0); | |
5345 | ||
5346 | new_op0 = aarch64_strip_shift (op0); | |
5347 | ||
5348 | /* If we had a shift on op0 then this is a logical-shift- | |
5349 | by-register/immediate operation. Otherwise, this is just | |
5350 | a logical operation. */ | |
5351 | if (speed) | |
5352 | { | |
5353 | if (new_op0 != op0) | |
5354 | { | |
5355 | /* Shift by immediate. */ | |
5356 | if (CONST_INT_P (XEXP (op0, 1))) | |
5357 | *cost += extra_cost->alu.log_shift; | |
5358 | else | |
5359 | *cost += extra_cost->alu.log_shift_reg; | |
5360 | } | |
5361 | else | |
5362 | *cost += extra_cost->alu.logical; | |
5363 | } | |
5364 | ||
5365 | /* In both cases we want to cost both operands. */ | |
5366 | *cost += rtx_cost (new_op0, (enum rtx_code) code, 0, speed) | |
5367 | + rtx_cost (op1, (enum rtx_code) code, 1, speed); | |
5368 | ||
5369 | return true; | |
5370 | } | |
5371 | } | |
5372 | return false; | |
5373 | ||
5374 | case NOT: | |
5375 | /* MVN. */ | |
5376 | if (speed) | |
5377 | *cost += extra_cost->alu.logical; | |
5378 | ||
5379 | /* The logical instruction could have the shifted register form, | |
5380 | but the cost is the same if the shift is processed as a separate | |
5381 | instruction, so we don't bother with it here. */ | |
5382 | return false; | |
5383 | ||
5384 | case ZERO_EXTEND: | |
5385 | ||
5386 | op0 = XEXP (x, 0); | |
5387 | /* If a value is written in SI mode, then zero extended to DI | |
5388 | mode, the operation will in general be free as a write to | |
5389 | a 'w' register implicitly zeroes the upper bits of an 'x' | |
5390 | register. However, if this is | |
5391 | ||
5392 | (set (reg) (zero_extend (reg))) | |
5393 | ||
5394 | we must cost the explicit register move. */ | |
5395 | if (mode == DImode | |
5396 | && GET_MODE (op0) == SImode | |
5397 | && outer == SET) | |
5398 | { | |
5399 | int op_cost = rtx_cost (XEXP (x, 0), ZERO_EXTEND, 0, speed); | |
5400 | ||
5401 | if (!op_cost && speed) | |
5402 | /* MOV. */ | |
5403 | *cost += extra_cost->alu.extend; | |
5404 | else | |
5405 | /* Free, the cost is that of the SI mode operation. */ | |
5406 | *cost = op_cost; | |
5407 | ||
5408 | return true; | |
5409 | } | |
5410 | else if (MEM_P (XEXP (x, 0))) | |
5411 | { | |
5412 | /* All loads can zero extend to any size for free. */ | |
5413 | *cost = rtx_cost (XEXP (x, 0), ZERO_EXTEND, param, speed); | |
5414 | return true; | |
5415 | } | |
5416 | ||
5417 | /* UXTB/UXTH. */ | |
5418 | if (speed) | |
5419 | *cost += extra_cost->alu.extend; | |
5420 | ||
5421 | return false; | |
5422 | ||
5423 | case SIGN_EXTEND: | |
5424 | if (MEM_P (XEXP (x, 0))) | |
5425 | { | |
5426 | /* LDRSH. */ | |
5427 | if (speed) | |
5428 | { | |
5429 | rtx address = XEXP (XEXP (x, 0), 0); | |
5430 | *cost += extra_cost->ldst.load_sign_extend; | |
5431 | ||
5432 | *cost += | |
5433 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
5434 | 0, speed)); | |
5435 | } | |
5436 | return true; | |
5437 | } | |
5438 | ||
5439 | if (speed) | |
5440 | *cost += extra_cost->alu.extend; | |
5441 | return false; | |
5442 | ||
5443 | case ASHIFT: | |
5444 | op0 = XEXP (x, 0); | |
5445 | op1 = XEXP (x, 1); | |
5446 | ||
5447 | if (CONST_INT_P (op1)) | |
5448 | { | |
5449 | /* LSL (immediate), UBMF, UBFIZ and friends. These are all | |
5450 | aliases. */ | |
5451 | if (speed) | |
5452 | *cost += extra_cost->alu.shift; | |
5453 | ||
5454 | /* We can incorporate zero/sign extend for free. */ | |
5455 | if (GET_CODE (op0) == ZERO_EXTEND | |
5456 | || GET_CODE (op0) == SIGN_EXTEND) | |
5457 | op0 = XEXP (op0, 0); | |
5458 | ||
5459 | *cost += rtx_cost (op0, ASHIFT, 0, speed); | |
5460 | return true; | |
5461 | } | |
5462 | else | |
5463 | { | |
5464 | /* LSLV. */ | |
5465 | if (speed) | |
5466 | *cost += extra_cost->alu.shift_reg; | |
5467 | ||
5468 | return false; /* All arguments need to be in registers. */ | |
5469 | } | |
5470 | ||
5471 | case ROTATE: | |
5472 | case ROTATERT: | |
5473 | case LSHIFTRT: | |
5474 | case ASHIFTRT: | |
5475 | op0 = XEXP (x, 0); | |
5476 | op1 = XEXP (x, 1); | |
5477 | ||
5478 | if (CONST_INT_P (op1)) | |
5479 | { | |
5480 | /* ASR (immediate) and friends. */ | |
5481 | if (speed) | |
5482 | *cost += extra_cost->alu.shift; | |
5483 | ||
5484 | *cost += rtx_cost (op0, (enum rtx_code) code, 0, speed); | |
5485 | return true; | |
5486 | } | |
5487 | else | |
5488 | { | |
5489 | ||
5490 | /* ASR (register) and friends. */ | |
5491 | if (speed) | |
5492 | *cost += extra_cost->alu.shift_reg; | |
5493 | ||
5494 | return false; /* All arguments need to be in registers. */ | |
5495 | } | |
5496 | ||
5497 | case SYMBOL_REF: | |
5498 | ||
5499 | if (aarch64_cmodel == AARCH64_CMODEL_LARGE) | |
5500 | { | |
5501 | /* LDR. */ | |
5502 | if (speed) | |
5503 | *cost += extra_cost->ldst.load; | |
5504 | } | |
5505 | else if (aarch64_cmodel == AARCH64_CMODEL_SMALL | |
5506 | || aarch64_cmodel == AARCH64_CMODEL_SMALL_PIC) | |
5507 | { | |
5508 | /* ADRP, followed by ADD. */ | |
5509 | *cost += COSTS_N_INSNS (1); | |
5510 | if (speed) | |
5511 | *cost += 2 * extra_cost->alu.arith; | |
5512 | } | |
5513 | else if (aarch64_cmodel == AARCH64_CMODEL_TINY | |
5514 | || aarch64_cmodel == AARCH64_CMODEL_TINY_PIC) | |
5515 | { | |
5516 | /* ADR. */ | |
5517 | if (speed) | |
5518 | *cost += extra_cost->alu.arith; | |
5519 | } | |
5520 | ||
5521 | if (flag_pic) | |
5522 | { | |
5523 | /* One extra load instruction, after accessing the GOT. */ | |
5524 | *cost += COSTS_N_INSNS (1); | |
5525 | if (speed) | |
5526 | *cost += extra_cost->ldst.load; | |
5527 | } | |
5528 | return true; | |
5529 | ||
5530 | case HIGH: | |
5531 | case LO_SUM: | |
5532 | /* ADRP/ADD (immediate). */ | |
5533 | if (speed) | |
5534 | *cost += extra_cost->alu.arith; | |
5535 | return true; | |
5536 | ||
5537 | case ZERO_EXTRACT: | |
5538 | case SIGN_EXTRACT: | |
5539 | /* UBFX/SBFX. */ | |
5540 | if (speed) | |
5541 | *cost += extra_cost->alu.bfx; | |
5542 | ||
5543 | /* We can trust that the immediates used will be correct (there | |
5544 | are no by-register forms), so we need only cost op0. */ | |
5545 | *cost += rtx_cost (XEXP (x, 0), (enum rtx_code) code, 0, speed); | |
5546 | return true; | |
5547 | ||
5548 | case MULT: | |
5549 | *cost += aarch64_rtx_mult_cost (x, MULT, 0, speed); | |
5550 | /* aarch64_rtx_mult_cost always handles recursion to its | |
5551 | operands. */ | |
5552 | return true; | |
5553 | ||
5554 | case MOD: | |
5555 | case UMOD: | |
5556 | if (speed) | |
5557 | { | |
5558 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT) | |
5559 | *cost += (extra_cost->mult[GET_MODE (x) == DImode].add | |
5560 | + extra_cost->mult[GET_MODE (x) == DImode].idiv); | |
5561 | else if (GET_MODE (x) == DFmode) | |
5562 | *cost += (extra_cost->fp[1].mult | |
5563 | + extra_cost->fp[1].div); | |
5564 | else if (GET_MODE (x) == SFmode) | |
5565 | *cost += (extra_cost->fp[0].mult | |
5566 | + extra_cost->fp[0].div); | |
5567 | } | |
5568 | return false; /* All arguments need to be in registers. */ | |
5569 | ||
5570 | case DIV: | |
5571 | case UDIV: | |
5572 | case SQRT: | |
5573 | if (speed) | |
5574 | { | |
5575 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
5576 | /* There is no integer SQRT, so only DIV and UDIV can get | |
5577 | here. */ | |
5578 | *cost += extra_cost->mult[mode == DImode].idiv; | |
5579 | else | |
5580 | *cost += extra_cost->fp[mode == DFmode].div; | |
5581 | } | |
5582 | return false; /* All arguments need to be in registers. */ | |
5583 | ||
5584 | case IF_THEN_ELSE: | |
5585 | op2 = XEXP (x, 2); | |
5586 | op0 = XEXP (x, 0); | |
5587 | op1 = XEXP (x, 1); | |
5588 | ||
5589 | if (GET_CODE (op1) == PC || GET_CODE (op2) == PC) | |
5590 | { | |
5591 | /* Conditional branch. */ | |
5592 | if (GET_MODE_CLASS (GET_MODE (XEXP (op0, 0))) == MODE_CC) | |
5593 | return true; | |
5594 | else | |
5595 | { | |
5596 | if (GET_CODE (op0) == NE | |
5597 | || GET_CODE (op0) == EQ) | |
5598 | { | |
5599 | rtx inner = XEXP (op0, 0); | |
5600 | rtx comparator = XEXP (op0, 1); | |
5601 | ||
5602 | if (comparator == const0_rtx) | |
5603 | { | |
5604 | /* TBZ/TBNZ/CBZ/CBNZ. */ | |
5605 | if (GET_CODE (inner) == ZERO_EXTRACT) | |
5606 | /* TBZ/TBNZ. */ | |
5607 | *cost += rtx_cost (XEXP (inner, 0), ZERO_EXTRACT, | |
5608 | 0, speed); | |
5609 | else | |
5610 | /* CBZ/CBNZ. */ | |
5611 | *cost += rtx_cost (inner, GET_CODE (op0), 0, speed); | |
5612 | ||
5613 | return true; | |
5614 | } | |
5615 | } | |
5616 | else if (GET_CODE (op0) == LT | |
5617 | || GET_CODE (op0) == GE) | |
5618 | { | |
5619 | rtx comparator = XEXP (op0, 1); | |
5620 | ||
5621 | /* TBZ/TBNZ. */ | |
5622 | if (comparator == const0_rtx) | |
5623 | return true; | |
5624 | } | |
5625 | } | |
5626 | } | |
5627 | else if (GET_MODE_CLASS (GET_MODE (XEXP (op0, 0))) == MODE_CC) | |
5628 | { | |
5629 | /* It's a conditional operation based on the status flags, | |
5630 | so it must be some flavor of CSEL. */ | |
5631 | ||
5632 | /* CSNEG, CSINV, and CSINC are handled for free as part of CSEL. */ | |
5633 | if (GET_CODE (op1) == NEG | |
5634 | || GET_CODE (op1) == NOT | |
5635 | || (GET_CODE (op1) == PLUS && XEXP (op1, 1) == const1_rtx)) | |
5636 | op1 = XEXP (op1, 0); | |
5637 | ||
5638 | *cost += rtx_cost (op1, IF_THEN_ELSE, 1, speed); | |
5639 | *cost += rtx_cost (op2, IF_THEN_ELSE, 2, speed); | |
5640 | return true; | |
5641 | } | |
5642 | ||
5643 | /* We don't know what this is, cost all operands. */ | |
5644 | return false; | |
5645 | ||
5646 | case EQ: | |
5647 | case NE: | |
5648 | case GT: | |
5649 | case GTU: | |
5650 | case LT: | |
5651 | case LTU: | |
5652 | case GE: | |
5653 | case GEU: | |
5654 | case LE: | |
5655 | case LEU: | |
5656 | ||
5657 | return false; /* All arguments must be in registers. */ | |
5658 | ||
5659 | case FMA: | |
5660 | op0 = XEXP (x, 0); | |
5661 | op1 = XEXP (x, 1); | |
5662 | op2 = XEXP (x, 2); | |
5663 | ||
5664 | if (speed) | |
5665 | *cost += extra_cost->fp[mode == DFmode].fma; | |
5666 | ||
5667 | /* FMSUB, FNMADD, and FNMSUB are free. */ | |
5668 | if (GET_CODE (op0) == NEG) | |
5669 | op0 = XEXP (op0, 0); | |
5670 | ||
5671 | if (GET_CODE (op2) == NEG) | |
5672 | op2 = XEXP (op2, 0); | |
5673 | ||
5674 | /* aarch64_fnma4_elt_to_64v2df has the NEG as operand 1, | |
5675 | and the by-element operand as operand 0. */ | |
5676 | if (GET_CODE (op1) == NEG) | |
5677 | op1 = XEXP (op1, 0); | |
5678 | ||
5679 | /* Catch vector-by-element operations. The by-element operand can | |
5680 | either be (vec_duplicate (vec_select (x))) or just | |
5681 | (vec_select (x)), depending on whether we are multiplying by | |
5682 | a vector or a scalar. | |
5683 | ||
5684 | Canonicalization is not very good in these cases, FMA4 will put the | |
5685 | by-element operand as operand 0, FNMA4 will have it as operand 1. */ | |
5686 | if (GET_CODE (op0) == VEC_DUPLICATE) | |
5687 | op0 = XEXP (op0, 0); | |
5688 | else if (GET_CODE (op1) == VEC_DUPLICATE) | |
5689 | op1 = XEXP (op1, 0); | |
5690 | ||
5691 | if (GET_CODE (op0) == VEC_SELECT) | |
5692 | op0 = XEXP (op0, 0); | |
5693 | else if (GET_CODE (op1) == VEC_SELECT) | |
5694 | op1 = XEXP (op1, 0); | |
5695 | ||
5696 | /* If the remaining parameters are not registers, | |
5697 | get the cost to put them into registers. */ | |
5698 | *cost += rtx_cost (op0, FMA, 0, speed); | |
5699 | *cost += rtx_cost (op1, FMA, 1, speed); | |
5700 | *cost += rtx_cost (op2, FMA, 2, speed); | |
5701 | return true; | |
5702 | ||
5703 | case FLOAT_EXTEND: | |
5704 | if (speed) | |
5705 | *cost += extra_cost->fp[mode == DFmode].widen; | |
5706 | return false; | |
5707 | ||
5708 | case FLOAT_TRUNCATE: | |
5709 | if (speed) | |
5710 | *cost += extra_cost->fp[mode == DFmode].narrow; | |
5711 | return false; | |
5712 | ||
5713 | case ABS: | |
5714 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5715 | { | |
5716 | /* FABS and FNEG are analogous. */ | |
5717 | if (speed) | |
5718 | *cost += extra_cost->fp[mode == DFmode].neg; | |
5719 | } | |
5720 | else | |
5721 | { | |
5722 | /* Integer ABS will either be split to | |
5723 | two arithmetic instructions, or will be an ABS | |
5724 | (scalar), which we don't model. */ | |
5725 | *cost = COSTS_N_INSNS (2); | |
5726 | if (speed) | |
5727 | *cost += 2 * extra_cost->alu.arith; | |
5728 | } | |
5729 | return false; | |
5730 | ||
5731 | case SMAX: | |
5732 | case SMIN: | |
5733 | if (speed) | |
5734 | { | |
5735 | /* FMAXNM/FMINNM/FMAX/FMIN. | |
5736 | TODO: This may not be accurate for all implementations, but | |
5737 | we do not model this in the cost tables. */ | |
5738 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
5739 | } | |
5740 | return false; | |
5741 | ||
5742 | case TRUNCATE: | |
5743 | ||
5744 | /* Decompose <su>muldi3_highpart. */ | |
5745 | if (/* (truncate:DI */ | |
5746 | mode == DImode | |
5747 | /* (lshiftrt:TI */ | |
5748 | && GET_MODE (XEXP (x, 0)) == TImode | |
5749 | && GET_CODE (XEXP (x, 0)) == LSHIFTRT | |
5750 | /* (mult:TI */ | |
5751 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT | |
5752 | /* (ANY_EXTEND:TI (reg:DI)) | |
5753 | (ANY_EXTEND:TI (reg:DI))) */ | |
5754 | && ((GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == ZERO_EXTEND | |
5755 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == ZERO_EXTEND) | |
5756 | || (GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == SIGN_EXTEND | |
5757 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == SIGN_EXTEND)) | |
5758 | && GET_MODE (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 0)) == DImode | |
5759 | && GET_MODE (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 1), 0)) == DImode | |
5760 | /* (const_int 64) */ | |
5761 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
5762 | && UINTVAL (XEXP (XEXP (x, 0), 1)) == 64) | |
5763 | { | |
5764 | /* UMULH/SMULH. */ | |
5765 | if (speed) | |
5766 | *cost += extra_cost->mult[mode == DImode].extend; | |
5767 | *cost += rtx_cost (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 0), | |
5768 | MULT, 0, speed); | |
5769 | *cost += rtx_cost (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 1), 0), | |
5770 | MULT, 1, speed); | |
5771 | return true; | |
5772 | } | |
5773 | ||
5774 | /* Fall through. */ | |
5775 | default: | |
5776 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
5777 | fprintf (dump_file, | |
5778 | "\nFailed to cost RTX. Assuming default cost.\n"); | |
5779 | ||
5780 | return true; | |
5781 | } | |
5782 | return false; | |
5783 | } | |
5784 | ||
5785 | /* Wrapper around aarch64_rtx_costs, dumps the partial, or total cost | |
5786 | calculated for X. This cost is stored in *COST. Returns true | |
5787 | if the total cost of X was calculated. */ | |
5788 | static bool | |
5789 | aarch64_rtx_costs_wrapper (rtx x, int code, int outer, | |
5790 | int param, int *cost, bool speed) | |
5791 | { | |
5792 | bool result = aarch64_rtx_costs (x, code, outer, param, cost, speed); | |
5793 | ||
5794 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
5795 | { | |
5796 | print_rtl_single (dump_file, x); | |
5797 | fprintf (dump_file, "\n%s cost: %d (%s)\n", | |
5798 | speed ? "Hot" : "Cold", | |
5799 | *cost, result ? "final" : "partial"); | |
5800 | } | |
5801 | ||
5802 | return result; | |
5803 | } | |
5804 | ||
5805 | static int | |
5806 | aarch64_register_move_cost (enum machine_mode mode, | |
5807 | reg_class_t from_i, reg_class_t to_i) | |
5808 | { | |
5809 | enum reg_class from = (enum reg_class) from_i; | |
5810 | enum reg_class to = (enum reg_class) to_i; | |
5811 | const struct cpu_regmove_cost *regmove_cost | |
5812 | = aarch64_tune_params->regmove_cost; | |
5813 | ||
5814 | /* Moving between GPR and stack cost is the same as GP2GP. */ | |
5815 | if ((from == GENERAL_REGS && to == STACK_REG) | |
5816 | || (to == GENERAL_REGS && from == STACK_REG)) | |
5817 | return regmove_cost->GP2GP; | |
5818 | ||
5819 | /* To/From the stack register, we move via the gprs. */ | |
5820 | if (to == STACK_REG || from == STACK_REG) | |
5821 | return aarch64_register_move_cost (mode, from, GENERAL_REGS) | |
5822 | + aarch64_register_move_cost (mode, GENERAL_REGS, to); | |
5823 | ||
5824 | if (from == GENERAL_REGS && to == GENERAL_REGS) | |
5825 | return regmove_cost->GP2GP; | |
5826 | else if (from == GENERAL_REGS) | |
5827 | return regmove_cost->GP2FP; | |
5828 | else if (to == GENERAL_REGS) | |
5829 | return regmove_cost->FP2GP; | |
5830 | ||
5831 | /* When AdvSIMD instructions are disabled it is not possible to move | |
5832 | a 128-bit value directly between Q registers. This is handled in | |
5833 | secondary reload. A general register is used as a scratch to move | |
5834 | the upper DI value and the lower DI value is moved directly, | |
5835 | hence the cost is the sum of three moves. */ | |
5836 | if (! TARGET_SIMD && GET_MODE_SIZE (mode) == 128) | |
5837 | return regmove_cost->GP2FP + regmove_cost->FP2GP + regmove_cost->FP2FP; | |
5838 | ||
5839 | return regmove_cost->FP2FP; | |
5840 | } | |
5841 | ||
5842 | static int | |
5843 | aarch64_memory_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED, | |
5844 | reg_class_t rclass ATTRIBUTE_UNUSED, | |
5845 | bool in ATTRIBUTE_UNUSED) | |
5846 | { | |
5847 | return aarch64_tune_params->memmov_cost; | |
5848 | } | |
5849 | ||
5850 | /* Return the number of instructions that can be issued per cycle. */ | |
5851 | static int | |
5852 | aarch64_sched_issue_rate (void) | |
5853 | { | |
5854 | return aarch64_tune_params->issue_rate; | |
5855 | } | |
5856 | ||
5857 | /* Vectorizer cost model target hooks. */ | |
5858 | ||
5859 | /* Implement targetm.vectorize.builtin_vectorization_cost. */ | |
5860 | static int | |
5861 | aarch64_builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost, | |
5862 | tree vectype, | |
5863 | int misalign ATTRIBUTE_UNUSED) | |
5864 | { | |
5865 | unsigned elements; | |
5866 | ||
5867 | switch (type_of_cost) | |
5868 | { | |
5869 | case scalar_stmt: | |
5870 | return aarch64_tune_params->vec_costs->scalar_stmt_cost; | |
5871 | ||
5872 | case scalar_load: | |
5873 | return aarch64_tune_params->vec_costs->scalar_load_cost; | |
5874 | ||
5875 | case scalar_store: | |
5876 | return aarch64_tune_params->vec_costs->scalar_store_cost; | |
5877 | ||
5878 | case vector_stmt: | |
5879 | return aarch64_tune_params->vec_costs->vec_stmt_cost; | |
5880 | ||
5881 | case vector_load: | |
5882 | return aarch64_tune_params->vec_costs->vec_align_load_cost; | |
5883 | ||
5884 | case vector_store: | |
5885 | return aarch64_tune_params->vec_costs->vec_store_cost; | |
5886 | ||
5887 | case vec_to_scalar: | |
5888 | return aarch64_tune_params->vec_costs->vec_to_scalar_cost; | |
5889 | ||
5890 | case scalar_to_vec: | |
5891 | return aarch64_tune_params->vec_costs->scalar_to_vec_cost; | |
5892 | ||
5893 | case unaligned_load: | |
5894 | return aarch64_tune_params->vec_costs->vec_unalign_load_cost; | |
5895 | ||
5896 | case unaligned_store: | |
5897 | return aarch64_tune_params->vec_costs->vec_unalign_store_cost; | |
5898 | ||
5899 | case cond_branch_taken: | |
5900 | return aarch64_tune_params->vec_costs->cond_taken_branch_cost; | |
5901 | ||
5902 | case cond_branch_not_taken: | |
5903 | return aarch64_tune_params->vec_costs->cond_not_taken_branch_cost; | |
5904 | ||
5905 | case vec_perm: | |
5906 | case vec_promote_demote: | |
5907 | return aarch64_tune_params->vec_costs->vec_stmt_cost; | |
5908 | ||
5909 | case vec_construct: | |
5910 | elements = TYPE_VECTOR_SUBPARTS (vectype); | |
5911 | return elements / 2 + 1; | |
5912 | ||
5913 | default: | |
5914 | gcc_unreachable (); | |
5915 | } | |
5916 | } | |
5917 | ||
5918 | /* Implement targetm.vectorize.add_stmt_cost. */ | |
5919 | static unsigned | |
5920 | aarch64_add_stmt_cost (void *data, int count, enum vect_cost_for_stmt kind, | |
5921 | struct _stmt_vec_info *stmt_info, int misalign, | |
5922 | enum vect_cost_model_location where) | |
5923 | { | |
5924 | unsigned *cost = (unsigned *) data; | |
5925 | unsigned retval = 0; | |
5926 | ||
5927 | if (flag_vect_cost_model) | |
5928 | { | |
5929 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; | |
5930 | int stmt_cost = | |
5931 | aarch64_builtin_vectorization_cost (kind, vectype, misalign); | |
5932 | ||
5933 | /* Statements in an inner loop relative to the loop being | |
5934 | vectorized are weighted more heavily. The value here is | |
5935 | a function (linear for now) of the loop nest level. */ | |
5936 | if (where == vect_body && stmt_info && stmt_in_inner_loop_p (stmt_info)) | |
5937 | { | |
5938 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
5939 | struct loop *loop = LOOP_VINFO_LOOP (loop_info); | |
5940 | unsigned nest_level = loop_depth (loop); | |
5941 | ||
5942 | count *= nest_level; | |
5943 | } | |
5944 | ||
5945 | retval = (unsigned) (count * stmt_cost); | |
5946 | cost[where] += retval; | |
5947 | } | |
5948 | ||
5949 | return retval; | |
5950 | } | |
5951 | ||
5952 | static void initialize_aarch64_code_model (void); | |
5953 | ||
5954 | /* Parse the architecture extension string. */ | |
5955 | ||
5956 | static void | |
5957 | aarch64_parse_extension (char *str) | |
5958 | { | |
5959 | /* The extension string is parsed left to right. */ | |
5960 | const struct aarch64_option_extension *opt = NULL; | |
5961 | ||
5962 | /* Flag to say whether we are adding or removing an extension. */ | |
5963 | int adding_ext = -1; | |
5964 | ||
5965 | while (str != NULL && *str != 0) | |
5966 | { | |
5967 | char *ext; | |
5968 | size_t len; | |
5969 | ||
5970 | str++; | |
5971 | ext = strchr (str, '+'); | |
5972 | ||
5973 | if (ext != NULL) | |
5974 | len = ext - str; | |
5975 | else | |
5976 | len = strlen (str); | |
5977 | ||
5978 | if (len >= 2 && strncmp (str, "no", 2) == 0) | |
5979 | { | |
5980 | adding_ext = 0; | |
5981 | len -= 2; | |
5982 | str += 2; | |
5983 | } | |
5984 | else if (len > 0) | |
5985 | adding_ext = 1; | |
5986 | ||
5987 | if (len == 0) | |
5988 | { | |
5989 | error ("missing feature modifier after %qs", "+no"); | |
5990 | return; | |
5991 | } | |
5992 | ||
5993 | /* Scan over the extensions table trying to find an exact match. */ | |
5994 | for (opt = all_extensions; opt->name != NULL; opt++) | |
5995 | { | |
5996 | if (strlen (opt->name) == len && strncmp (opt->name, str, len) == 0) | |
5997 | { | |
5998 | /* Add or remove the extension. */ | |
5999 | if (adding_ext) | |
6000 | aarch64_isa_flags |= opt->flags_on; | |
6001 | else | |
6002 | aarch64_isa_flags &= ~(opt->flags_off); | |
6003 | break; | |
6004 | } | |
6005 | } | |
6006 | ||
6007 | if (opt->name == NULL) | |
6008 | { | |
6009 | /* Extension not found in list. */ | |
6010 | error ("unknown feature modifier %qs", str); | |
6011 | return; | |
6012 | } | |
6013 | ||
6014 | str = ext; | |
6015 | }; | |
6016 | ||
6017 | return; | |
6018 | } | |
6019 | ||
6020 | /* Parse the ARCH string. */ | |
6021 | ||
6022 | static void | |
6023 | aarch64_parse_arch (void) | |
6024 | { | |
6025 | char *ext; | |
6026 | const struct processor *arch; | |
6027 | char *str = (char *) alloca (strlen (aarch64_arch_string) + 1); | |
6028 | size_t len; | |
6029 | ||
6030 | strcpy (str, aarch64_arch_string); | |
6031 | ||
6032 | ext = strchr (str, '+'); | |
6033 | ||
6034 | if (ext != NULL) | |
6035 | len = ext - str; | |
6036 | else | |
6037 | len = strlen (str); | |
6038 | ||
6039 | if (len == 0) | |
6040 | { | |
6041 | error ("missing arch name in -march=%qs", str); | |
6042 | return; | |
6043 | } | |
6044 | ||
6045 | /* Loop through the list of supported ARCHs to find a match. */ | |
6046 | for (arch = all_architectures; arch->name != NULL; arch++) | |
6047 | { | |
6048 | if (strlen (arch->name) == len && strncmp (arch->name, str, len) == 0) | |
6049 | { | |
6050 | selected_arch = arch; | |
6051 | aarch64_isa_flags = selected_arch->flags; | |
6052 | ||
6053 | if (!selected_cpu) | |
6054 | selected_cpu = &all_cores[selected_arch->core]; | |
6055 | ||
6056 | if (ext != NULL) | |
6057 | { | |
6058 | /* ARCH string contains at least one extension. */ | |
6059 | aarch64_parse_extension (ext); | |
6060 | } | |
6061 | ||
6062 | if (strcmp (selected_arch->arch, selected_cpu->arch)) | |
6063 | { | |
6064 | warning (0, "switch -mcpu=%s conflicts with -march=%s switch", | |
6065 | selected_cpu->name, selected_arch->name); | |
6066 | } | |
6067 | ||
6068 | return; | |
6069 | } | |
6070 | } | |
6071 | ||
6072 | /* ARCH name not found in list. */ | |
6073 | error ("unknown value %qs for -march", str); | |
6074 | return; | |
6075 | } | |
6076 | ||
6077 | /* Parse the CPU string. */ | |
6078 | ||
6079 | static void | |
6080 | aarch64_parse_cpu (void) | |
6081 | { | |
6082 | char *ext; | |
6083 | const struct processor *cpu; | |
6084 | char *str = (char *) alloca (strlen (aarch64_cpu_string) + 1); | |
6085 | size_t len; | |
6086 | ||
6087 | strcpy (str, aarch64_cpu_string); | |
6088 | ||
6089 | ext = strchr (str, '+'); | |
6090 | ||
6091 | if (ext != NULL) | |
6092 | len = ext - str; | |
6093 | else | |
6094 | len = strlen (str); | |
6095 | ||
6096 | if (len == 0) | |
6097 | { | |
6098 | error ("missing cpu name in -mcpu=%qs", str); | |
6099 | return; | |
6100 | } | |
6101 | ||
6102 | /* Loop through the list of supported CPUs to find a match. */ | |
6103 | for (cpu = all_cores; cpu->name != NULL; cpu++) | |
6104 | { | |
6105 | if (strlen (cpu->name) == len && strncmp (cpu->name, str, len) == 0) | |
6106 | { | |
6107 | selected_cpu = cpu; | |
6108 | selected_tune = cpu; | |
6109 | aarch64_isa_flags = selected_cpu->flags; | |
6110 | ||
6111 | if (ext != NULL) | |
6112 | { | |
6113 | /* CPU string contains at least one extension. */ | |
6114 | aarch64_parse_extension (ext); | |
6115 | } | |
6116 | ||
6117 | return; | |
6118 | } | |
6119 | } | |
6120 | ||
6121 | /* CPU name not found in list. */ | |
6122 | error ("unknown value %qs for -mcpu", str); | |
6123 | return; | |
6124 | } | |
6125 | ||
6126 | /* Parse the TUNE string. */ | |
6127 | ||
6128 | static void | |
6129 | aarch64_parse_tune (void) | |
6130 | { | |
6131 | const struct processor *cpu; | |
6132 | char *str = (char *) alloca (strlen (aarch64_tune_string) + 1); | |
6133 | strcpy (str, aarch64_tune_string); | |
6134 | ||
6135 | /* Loop through the list of supported CPUs to find a match. */ | |
6136 | for (cpu = all_cores; cpu->name != NULL; cpu++) | |
6137 | { | |
6138 | if (strcmp (cpu->name, str) == 0) | |
6139 | { | |
6140 | selected_tune = cpu; | |
6141 | return; | |
6142 | } | |
6143 | } | |
6144 | ||
6145 | /* CPU name not found in list. */ | |
6146 | error ("unknown value %qs for -mtune", str); | |
6147 | return; | |
6148 | } | |
6149 | ||
6150 | ||
6151 | /* Implement TARGET_OPTION_OVERRIDE. */ | |
6152 | ||
6153 | static void | |
6154 | aarch64_override_options (void) | |
6155 | { | |
6156 | /* -mcpu=CPU is shorthand for -march=ARCH_FOR_CPU, -mtune=CPU. | |
6157 | If either of -march or -mtune is given, they override their | |
6158 | respective component of -mcpu. | |
6159 | ||
6160 | So, first parse AARCH64_CPU_STRING, then the others, be careful | |
6161 | with -march as, if -mcpu is not present on the command line, march | |
6162 | must set a sensible default CPU. */ | |
6163 | if (aarch64_cpu_string) | |
6164 | { | |
6165 | aarch64_parse_cpu (); | |
6166 | } | |
6167 | ||
6168 | if (aarch64_arch_string) | |
6169 | { | |
6170 | aarch64_parse_arch (); | |
6171 | } | |
6172 | ||
6173 | if (aarch64_tune_string) | |
6174 | { | |
6175 | aarch64_parse_tune (); | |
6176 | } | |
6177 | ||
6178 | #ifndef HAVE_AS_MABI_OPTION | |
6179 | /* The compiler may have been configured with 2.23.* binutils, which does | |
6180 | not have support for ILP32. */ | |
6181 | if (TARGET_ILP32) | |
6182 | error ("Assembler does not support -mabi=ilp32"); | |
6183 | #endif | |
6184 | ||
6185 | initialize_aarch64_code_model (); | |
6186 | ||
6187 | aarch64_build_bitmask_table (); | |
6188 | ||
6189 | /* This target defaults to strict volatile bitfields. */ | |
6190 | if (flag_strict_volatile_bitfields < 0 && abi_version_at_least (2)) | |
6191 | flag_strict_volatile_bitfields = 1; | |
6192 | ||
6193 | /* If the user did not specify a processor, choose the default | |
6194 | one for them. This will be the CPU set during configuration using | |
6195 | --with-cpu, otherwise it is "generic". */ | |
6196 | if (!selected_cpu) | |
6197 | { | |
6198 | selected_cpu = &all_cores[TARGET_CPU_DEFAULT & 0x3f]; | |
6199 | aarch64_isa_flags = TARGET_CPU_DEFAULT >> 6; | |
6200 | } | |
6201 | ||
6202 | gcc_assert (selected_cpu); | |
6203 | ||
6204 | /* The selected cpu may be an architecture, so lookup tuning by core ID. */ | |
6205 | if (!selected_tune) | |
6206 | selected_tune = &all_cores[selected_cpu->core]; | |
6207 | ||
6208 | aarch64_tune_flags = selected_tune->flags; | |
6209 | aarch64_tune = selected_tune->core; | |
6210 | aarch64_tune_params = selected_tune->tune; | |
6211 | ||
6212 | aarch64_override_options_after_change (); | |
6213 | } | |
6214 | ||
6215 | /* Implement targetm.override_options_after_change. */ | |
6216 | ||
6217 | static void | |
6218 | aarch64_override_options_after_change (void) | |
6219 | { | |
6220 | if (flag_omit_frame_pointer) | |
6221 | flag_omit_leaf_frame_pointer = false; | |
6222 | else if (flag_omit_leaf_frame_pointer) | |
6223 | flag_omit_frame_pointer = true; | |
6224 | } | |
6225 | ||
6226 | static struct machine_function * | |
6227 | aarch64_init_machine_status (void) | |
6228 | { | |
6229 | struct machine_function *machine; | |
6230 | machine = ggc_cleared_alloc<machine_function> (); | |
6231 | return machine; | |
6232 | } | |
6233 | ||
6234 | void | |
6235 | aarch64_init_expanders (void) | |
6236 | { | |
6237 | init_machine_status = aarch64_init_machine_status; | |
6238 | } | |
6239 | ||
6240 | /* A checking mechanism for the implementation of the various code models. */ | |
6241 | static void | |
6242 | initialize_aarch64_code_model (void) | |
6243 | { | |
6244 | if (flag_pic) | |
6245 | { | |
6246 | switch (aarch64_cmodel_var) | |
6247 | { | |
6248 | case AARCH64_CMODEL_TINY: | |
6249 | aarch64_cmodel = AARCH64_CMODEL_TINY_PIC; | |
6250 | break; | |
6251 | case AARCH64_CMODEL_SMALL: | |
6252 | aarch64_cmodel = AARCH64_CMODEL_SMALL_PIC; | |
6253 | break; | |
6254 | case AARCH64_CMODEL_LARGE: | |
6255 | sorry ("code model %qs with -f%s", "large", | |
6256 | flag_pic > 1 ? "PIC" : "pic"); | |
6257 | default: | |
6258 | gcc_unreachable (); | |
6259 | } | |
6260 | } | |
6261 | else | |
6262 | aarch64_cmodel = aarch64_cmodel_var; | |
6263 | } | |
6264 | ||
6265 | /* Return true if SYMBOL_REF X binds locally. */ | |
6266 | ||
6267 | static bool | |
6268 | aarch64_symbol_binds_local_p (const_rtx x) | |
6269 | { | |
6270 | return (SYMBOL_REF_DECL (x) | |
6271 | ? targetm.binds_local_p (SYMBOL_REF_DECL (x)) | |
6272 | : SYMBOL_REF_LOCAL_P (x)); | |
6273 | } | |
6274 | ||
6275 | /* Return true if SYMBOL_REF X is thread local */ | |
6276 | static bool | |
6277 | aarch64_tls_symbol_p (rtx x) | |
6278 | { | |
6279 | if (! TARGET_HAVE_TLS) | |
6280 | return false; | |
6281 | ||
6282 | if (GET_CODE (x) != SYMBOL_REF) | |
6283 | return false; | |
6284 | ||
6285 | return SYMBOL_REF_TLS_MODEL (x) != 0; | |
6286 | } | |
6287 | ||
6288 | /* Classify a TLS symbol into one of the TLS kinds. */ | |
6289 | enum aarch64_symbol_type | |
6290 | aarch64_classify_tls_symbol (rtx x) | |
6291 | { | |
6292 | enum tls_model tls_kind = tls_symbolic_operand_type (x); | |
6293 | ||
6294 | switch (tls_kind) | |
6295 | { | |
6296 | case TLS_MODEL_GLOBAL_DYNAMIC: | |
6297 | case TLS_MODEL_LOCAL_DYNAMIC: | |
6298 | return TARGET_TLS_DESC ? SYMBOL_SMALL_TLSDESC : SYMBOL_SMALL_TLSGD; | |
6299 | ||
6300 | case TLS_MODEL_INITIAL_EXEC: | |
6301 | return SYMBOL_SMALL_GOTTPREL; | |
6302 | ||
6303 | case TLS_MODEL_LOCAL_EXEC: | |
6304 | return SYMBOL_SMALL_TPREL; | |
6305 | ||
6306 | case TLS_MODEL_EMULATED: | |
6307 | case TLS_MODEL_NONE: | |
6308 | return SYMBOL_FORCE_TO_MEM; | |
6309 | ||
6310 | default: | |
6311 | gcc_unreachable (); | |
6312 | } | |
6313 | } | |
6314 | ||
6315 | /* Return the method that should be used to access SYMBOL_REF or | |
6316 | LABEL_REF X in context CONTEXT. */ | |
6317 | ||
6318 | enum aarch64_symbol_type | |
6319 | aarch64_classify_symbol (rtx x, | |
6320 | enum aarch64_symbol_context context ATTRIBUTE_UNUSED) | |
6321 | { | |
6322 | if (GET_CODE (x) == LABEL_REF) | |
6323 | { | |
6324 | switch (aarch64_cmodel) | |
6325 | { | |
6326 | case AARCH64_CMODEL_LARGE: | |
6327 | return SYMBOL_FORCE_TO_MEM; | |
6328 | ||
6329 | case AARCH64_CMODEL_TINY_PIC: | |
6330 | case AARCH64_CMODEL_TINY: | |
6331 | return SYMBOL_TINY_ABSOLUTE; | |
6332 | ||
6333 | case AARCH64_CMODEL_SMALL_PIC: | |
6334 | case AARCH64_CMODEL_SMALL: | |
6335 | return SYMBOL_SMALL_ABSOLUTE; | |
6336 | ||
6337 | default: | |
6338 | gcc_unreachable (); | |
6339 | } | |
6340 | } | |
6341 | ||
6342 | if (GET_CODE (x) == SYMBOL_REF) | |
6343 | { | |
6344 | if (aarch64_cmodel == AARCH64_CMODEL_LARGE) | |
6345 | return SYMBOL_FORCE_TO_MEM; | |
6346 | ||
6347 | if (aarch64_tls_symbol_p (x)) | |
6348 | return aarch64_classify_tls_symbol (x); | |
6349 | ||
6350 | switch (aarch64_cmodel) | |
6351 | { | |
6352 | case AARCH64_CMODEL_TINY: | |
6353 | if (SYMBOL_REF_WEAK (x)) | |
6354 | return SYMBOL_FORCE_TO_MEM; | |
6355 | return SYMBOL_TINY_ABSOLUTE; | |
6356 | ||
6357 | case AARCH64_CMODEL_SMALL: | |
6358 | if (SYMBOL_REF_WEAK (x)) | |
6359 | return SYMBOL_FORCE_TO_MEM; | |
6360 | return SYMBOL_SMALL_ABSOLUTE; | |
6361 | ||
6362 | case AARCH64_CMODEL_TINY_PIC: | |
6363 | if (!aarch64_symbol_binds_local_p (x)) | |
6364 | return SYMBOL_TINY_GOT; | |
6365 | return SYMBOL_TINY_ABSOLUTE; | |
6366 | ||
6367 | case AARCH64_CMODEL_SMALL_PIC: | |
6368 | if (!aarch64_symbol_binds_local_p (x)) | |
6369 | return SYMBOL_SMALL_GOT; | |
6370 | return SYMBOL_SMALL_ABSOLUTE; | |
6371 | ||
6372 | default: | |
6373 | gcc_unreachable (); | |
6374 | } | |
6375 | } | |
6376 | ||
6377 | /* By default push everything into the constant pool. */ | |
6378 | return SYMBOL_FORCE_TO_MEM; | |
6379 | } | |
6380 | ||
6381 | bool | |
6382 | aarch64_constant_address_p (rtx x) | |
6383 | { | |
6384 | return (CONSTANT_P (x) && memory_address_p (DImode, x)); | |
6385 | } | |
6386 | ||
6387 | bool | |
6388 | aarch64_legitimate_pic_operand_p (rtx x) | |
6389 | { | |
6390 | if (GET_CODE (x) == SYMBOL_REF | |
6391 | || (GET_CODE (x) == CONST | |
6392 | && GET_CODE (XEXP (x, 0)) == PLUS | |
6393 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF)) | |
6394 | return false; | |
6395 | ||
6396 | return true; | |
6397 | } | |
6398 | ||
6399 | /* Return true if X holds either a quarter-precision or | |
6400 | floating-point +0.0 constant. */ | |
6401 | static bool | |
6402 | aarch64_valid_floating_const (enum machine_mode mode, rtx x) | |
6403 | { | |
6404 | if (!CONST_DOUBLE_P (x)) | |
6405 | return false; | |
6406 | ||
6407 | /* TODO: We could handle moving 0.0 to a TFmode register, | |
6408 | but first we would like to refactor the movtf_aarch64 | |
6409 | to be more amicable to split moves properly and | |
6410 | correctly gate on TARGET_SIMD. For now - reject all | |
6411 | constants which are not to SFmode or DFmode registers. */ | |
6412 | if (!(mode == SFmode || mode == DFmode)) | |
6413 | return false; | |
6414 | ||
6415 | if (aarch64_float_const_zero_rtx_p (x)) | |
6416 | return true; | |
6417 | return aarch64_float_const_representable_p (x); | |
6418 | } | |
6419 | ||
6420 | static bool | |
6421 | aarch64_legitimate_constant_p (enum machine_mode mode, rtx x) | |
6422 | { | |
6423 | /* Do not allow vector struct mode constants. We could support | |
6424 | 0 and -1 easily, but they need support in aarch64-simd.md. */ | |
6425 | if (TARGET_SIMD && aarch64_vect_struct_mode_p (mode)) | |
6426 | return false; | |
6427 | ||
6428 | /* This could probably go away because | |
6429 | we now decompose CONST_INTs according to expand_mov_immediate. */ | |
6430 | if ((GET_CODE (x) == CONST_VECTOR | |
6431 | && aarch64_simd_valid_immediate (x, mode, false, NULL)) | |
6432 | || CONST_INT_P (x) || aarch64_valid_floating_const (mode, x)) | |
6433 | return !targetm.cannot_force_const_mem (mode, x); | |
6434 | ||
6435 | if (GET_CODE (x) == HIGH | |
6436 | && aarch64_valid_symref (XEXP (x, 0), GET_MODE (XEXP (x, 0)))) | |
6437 | return true; | |
6438 | ||
6439 | return aarch64_constant_address_p (x); | |
6440 | } | |
6441 | ||
6442 | rtx | |
6443 | aarch64_load_tp (rtx target) | |
6444 | { | |
6445 | if (!target | |
6446 | || GET_MODE (target) != Pmode | |
6447 | || !register_operand (target, Pmode)) | |
6448 | target = gen_reg_rtx (Pmode); | |
6449 | ||
6450 | /* Can return in any reg. */ | |
6451 | emit_insn (gen_aarch64_load_tp_hard (target)); | |
6452 | return target; | |
6453 | } | |
6454 | ||
6455 | /* On AAPCS systems, this is the "struct __va_list". */ | |
6456 | static GTY(()) tree va_list_type; | |
6457 | ||
6458 | /* Implement TARGET_BUILD_BUILTIN_VA_LIST. | |
6459 | Return the type to use as __builtin_va_list. | |
6460 | ||
6461 | AAPCS64 \S 7.1.4 requires that va_list be a typedef for a type defined as: | |
6462 | ||
6463 | struct __va_list | |
6464 | { | |
6465 | void *__stack; | |
6466 | void *__gr_top; | |
6467 | void *__vr_top; | |
6468 | int __gr_offs; | |
6469 | int __vr_offs; | |
6470 | }; */ | |
6471 | ||
6472 | static tree | |
6473 | aarch64_build_builtin_va_list (void) | |
6474 | { | |
6475 | tree va_list_name; | |
6476 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
6477 | ||
6478 | /* Create the type. */ | |
6479 | va_list_type = lang_hooks.types.make_type (RECORD_TYPE); | |
6480 | /* Give it the required name. */ | |
6481 | va_list_name = build_decl (BUILTINS_LOCATION, | |
6482 | TYPE_DECL, | |
6483 | get_identifier ("__va_list"), | |
6484 | va_list_type); | |
6485 | DECL_ARTIFICIAL (va_list_name) = 1; | |
6486 | TYPE_NAME (va_list_type) = va_list_name; | |
6487 | TYPE_STUB_DECL (va_list_type) = va_list_name; | |
6488 | ||
6489 | /* Create the fields. */ | |
6490 | f_stack = build_decl (BUILTINS_LOCATION, | |
6491 | FIELD_DECL, get_identifier ("__stack"), | |
6492 | ptr_type_node); | |
6493 | f_grtop = build_decl (BUILTINS_LOCATION, | |
6494 | FIELD_DECL, get_identifier ("__gr_top"), | |
6495 | ptr_type_node); | |
6496 | f_vrtop = build_decl (BUILTINS_LOCATION, | |
6497 | FIELD_DECL, get_identifier ("__vr_top"), | |
6498 | ptr_type_node); | |
6499 | f_groff = build_decl (BUILTINS_LOCATION, | |
6500 | FIELD_DECL, get_identifier ("__gr_offs"), | |
6501 | integer_type_node); | |
6502 | f_vroff = build_decl (BUILTINS_LOCATION, | |
6503 | FIELD_DECL, get_identifier ("__vr_offs"), | |
6504 | integer_type_node); | |
6505 | ||
6506 | DECL_ARTIFICIAL (f_stack) = 1; | |
6507 | DECL_ARTIFICIAL (f_grtop) = 1; | |
6508 | DECL_ARTIFICIAL (f_vrtop) = 1; | |
6509 | DECL_ARTIFICIAL (f_groff) = 1; | |
6510 | DECL_ARTIFICIAL (f_vroff) = 1; | |
6511 | ||
6512 | DECL_FIELD_CONTEXT (f_stack) = va_list_type; | |
6513 | DECL_FIELD_CONTEXT (f_grtop) = va_list_type; | |
6514 | DECL_FIELD_CONTEXT (f_vrtop) = va_list_type; | |
6515 | DECL_FIELD_CONTEXT (f_groff) = va_list_type; | |
6516 | DECL_FIELD_CONTEXT (f_vroff) = va_list_type; | |
6517 | ||
6518 | TYPE_FIELDS (va_list_type) = f_stack; | |
6519 | DECL_CHAIN (f_stack) = f_grtop; | |
6520 | DECL_CHAIN (f_grtop) = f_vrtop; | |
6521 | DECL_CHAIN (f_vrtop) = f_groff; | |
6522 | DECL_CHAIN (f_groff) = f_vroff; | |
6523 | ||
6524 | /* Compute its layout. */ | |
6525 | layout_type (va_list_type); | |
6526 | ||
6527 | return va_list_type; | |
6528 | } | |
6529 | ||
6530 | /* Implement TARGET_EXPAND_BUILTIN_VA_START. */ | |
6531 | static void | |
6532 | aarch64_expand_builtin_va_start (tree valist, rtx nextarg ATTRIBUTE_UNUSED) | |
6533 | { | |
6534 | const CUMULATIVE_ARGS *cum; | |
6535 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
6536 | tree stack, grtop, vrtop, groff, vroff; | |
6537 | tree t; | |
6538 | int gr_save_area_size; | |
6539 | int vr_save_area_size; | |
6540 | int vr_offset; | |
6541 | ||
6542 | cum = &crtl->args.info; | |
6543 | gr_save_area_size | |
6544 | = (NUM_ARG_REGS - cum->aapcs_ncrn) * UNITS_PER_WORD; | |
6545 | vr_save_area_size | |
6546 | = (NUM_FP_ARG_REGS - cum->aapcs_nvrn) * UNITS_PER_VREG; | |
6547 | ||
6548 | if (TARGET_GENERAL_REGS_ONLY) | |
6549 | { | |
6550 | if (cum->aapcs_nvrn > 0) | |
6551 | sorry ("%qs and floating point or vector arguments", | |
6552 | "-mgeneral-regs-only"); | |
6553 | vr_save_area_size = 0; | |
6554 | } | |
6555 | ||
6556 | f_stack = TYPE_FIELDS (va_list_type_node); | |
6557 | f_grtop = DECL_CHAIN (f_stack); | |
6558 | f_vrtop = DECL_CHAIN (f_grtop); | |
6559 | f_groff = DECL_CHAIN (f_vrtop); | |
6560 | f_vroff = DECL_CHAIN (f_groff); | |
6561 | ||
6562 | stack = build3 (COMPONENT_REF, TREE_TYPE (f_stack), valist, f_stack, | |
6563 | NULL_TREE); | |
6564 | grtop = build3 (COMPONENT_REF, TREE_TYPE (f_grtop), valist, f_grtop, | |
6565 | NULL_TREE); | |
6566 | vrtop = build3 (COMPONENT_REF, TREE_TYPE (f_vrtop), valist, f_vrtop, | |
6567 | NULL_TREE); | |
6568 | groff = build3 (COMPONENT_REF, TREE_TYPE (f_groff), valist, f_groff, | |
6569 | NULL_TREE); | |
6570 | vroff = build3 (COMPONENT_REF, TREE_TYPE (f_vroff), valist, f_vroff, | |
6571 | NULL_TREE); | |
6572 | ||
6573 | /* Emit code to initialize STACK, which points to the next varargs stack | |
6574 | argument. CUM->AAPCS_STACK_SIZE gives the number of stack words used | |
6575 | by named arguments. STACK is 8-byte aligned. */ | |
6576 | t = make_tree (TREE_TYPE (stack), virtual_incoming_args_rtx); | |
6577 | if (cum->aapcs_stack_size > 0) | |
6578 | t = fold_build_pointer_plus_hwi (t, cum->aapcs_stack_size * UNITS_PER_WORD); | |
6579 | t = build2 (MODIFY_EXPR, TREE_TYPE (stack), stack, t); | |
6580 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
6581 | ||
6582 | /* Emit code to initialize GRTOP, the top of the GR save area. | |
6583 | virtual_incoming_args_rtx should have been 16 byte aligned. */ | |
6584 | t = make_tree (TREE_TYPE (grtop), virtual_incoming_args_rtx); | |
6585 | t = build2 (MODIFY_EXPR, TREE_TYPE (grtop), grtop, t); | |
6586 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
6587 | ||
6588 | /* Emit code to initialize VRTOP, the top of the VR save area. | |
6589 | This address is gr_save_area_bytes below GRTOP, rounded | |
6590 | down to the next 16-byte boundary. */ | |
6591 | t = make_tree (TREE_TYPE (vrtop), virtual_incoming_args_rtx); | |
6592 | vr_offset = AARCH64_ROUND_UP (gr_save_area_size, | |
6593 | STACK_BOUNDARY / BITS_PER_UNIT); | |
6594 | ||
6595 | if (vr_offset) | |
6596 | t = fold_build_pointer_plus_hwi (t, -vr_offset); | |
6597 | t = build2 (MODIFY_EXPR, TREE_TYPE (vrtop), vrtop, t); | |
6598 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
6599 | ||
6600 | /* Emit code to initialize GROFF, the offset from GRTOP of the | |
6601 | next GPR argument. */ | |
6602 | t = build2 (MODIFY_EXPR, TREE_TYPE (groff), groff, | |
6603 | build_int_cst (TREE_TYPE (groff), -gr_save_area_size)); | |
6604 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
6605 | ||
6606 | /* Likewise emit code to initialize VROFF, the offset from FTOP | |
6607 | of the next VR argument. */ | |
6608 | t = build2 (MODIFY_EXPR, TREE_TYPE (vroff), vroff, | |
6609 | build_int_cst (TREE_TYPE (vroff), -vr_save_area_size)); | |
6610 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
6611 | } | |
6612 | ||
6613 | /* Implement TARGET_GIMPLIFY_VA_ARG_EXPR. */ | |
6614 | ||
6615 | static tree | |
6616 | aarch64_gimplify_va_arg_expr (tree valist, tree type, gimple_seq *pre_p, | |
6617 | gimple_seq *post_p ATTRIBUTE_UNUSED) | |
6618 | { | |
6619 | tree addr; | |
6620 | bool indirect_p; | |
6621 | bool is_ha; /* is HFA or HVA. */ | |
6622 | bool dw_align; /* double-word align. */ | |
6623 | enum machine_mode ag_mode = VOIDmode; | |
6624 | int nregs; | |
6625 | enum machine_mode mode; | |
6626 | ||
6627 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
6628 | tree stack, f_top, f_off, off, arg, roundup, on_stack; | |
6629 | HOST_WIDE_INT size, rsize, adjust, align; | |
6630 | tree t, u, cond1, cond2; | |
6631 | ||
6632 | indirect_p = pass_by_reference (NULL, TYPE_MODE (type), type, false); | |
6633 | if (indirect_p) | |
6634 | type = build_pointer_type (type); | |
6635 | ||
6636 | mode = TYPE_MODE (type); | |
6637 | ||
6638 | f_stack = TYPE_FIELDS (va_list_type_node); | |
6639 | f_grtop = DECL_CHAIN (f_stack); | |
6640 | f_vrtop = DECL_CHAIN (f_grtop); | |
6641 | f_groff = DECL_CHAIN (f_vrtop); | |
6642 | f_vroff = DECL_CHAIN (f_groff); | |
6643 | ||
6644 | stack = build3 (COMPONENT_REF, TREE_TYPE (f_stack), unshare_expr (valist), | |
6645 | f_stack, NULL_TREE); | |
6646 | size = int_size_in_bytes (type); | |
6647 | align = aarch64_function_arg_alignment (mode, type) / BITS_PER_UNIT; | |
6648 | ||
6649 | dw_align = false; | |
6650 | adjust = 0; | |
6651 | if (aarch64_vfp_is_call_or_return_candidate (mode, | |
6652 | type, | |
6653 | &ag_mode, | |
6654 | &nregs, | |
6655 | &is_ha)) | |
6656 | { | |
6657 | /* TYPE passed in fp/simd registers. */ | |
6658 | if (TARGET_GENERAL_REGS_ONLY) | |
6659 | sorry ("%qs and floating point or vector arguments", | |
6660 | "-mgeneral-regs-only"); | |
6661 | ||
6662 | f_top = build3 (COMPONENT_REF, TREE_TYPE (f_vrtop), | |
6663 | unshare_expr (valist), f_vrtop, NULL_TREE); | |
6664 | f_off = build3 (COMPONENT_REF, TREE_TYPE (f_vroff), | |
6665 | unshare_expr (valist), f_vroff, NULL_TREE); | |
6666 | ||
6667 | rsize = nregs * UNITS_PER_VREG; | |
6668 | ||
6669 | if (is_ha) | |
6670 | { | |
6671 | if (BYTES_BIG_ENDIAN && GET_MODE_SIZE (ag_mode) < UNITS_PER_VREG) | |
6672 | adjust = UNITS_PER_VREG - GET_MODE_SIZE (ag_mode); | |
6673 | } | |
6674 | else if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
6675 | && size < UNITS_PER_VREG) | |
6676 | { | |
6677 | adjust = UNITS_PER_VREG - size; | |
6678 | } | |
6679 | } | |
6680 | else | |
6681 | { | |
6682 | /* TYPE passed in general registers. */ | |
6683 | f_top = build3 (COMPONENT_REF, TREE_TYPE (f_grtop), | |
6684 | unshare_expr (valist), f_grtop, NULL_TREE); | |
6685 | f_off = build3 (COMPONENT_REF, TREE_TYPE (f_groff), | |
6686 | unshare_expr (valist), f_groff, NULL_TREE); | |
6687 | rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD; | |
6688 | nregs = rsize / UNITS_PER_WORD; | |
6689 | ||
6690 | if (align > 8) | |
6691 | dw_align = true; | |
6692 | ||
6693 | if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
6694 | && size < UNITS_PER_WORD) | |
6695 | { | |
6696 | adjust = UNITS_PER_WORD - size; | |
6697 | } | |
6698 | } | |
6699 | ||
6700 | /* Get a local temporary for the field value. */ | |
6701 | off = get_initialized_tmp_var (f_off, pre_p, NULL); | |
6702 | ||
6703 | /* Emit code to branch if off >= 0. */ | |
6704 | t = build2 (GE_EXPR, boolean_type_node, off, | |
6705 | build_int_cst (TREE_TYPE (off), 0)); | |
6706 | cond1 = build3 (COND_EXPR, ptr_type_node, t, NULL_TREE, NULL_TREE); | |
6707 | ||
6708 | if (dw_align) | |
6709 | { | |
6710 | /* Emit: offs = (offs + 15) & -16. */ | |
6711 | t = build2 (PLUS_EXPR, TREE_TYPE (off), off, | |
6712 | build_int_cst (TREE_TYPE (off), 15)); | |
6713 | t = build2 (BIT_AND_EXPR, TREE_TYPE (off), t, | |
6714 | build_int_cst (TREE_TYPE (off), -16)); | |
6715 | roundup = build2 (MODIFY_EXPR, TREE_TYPE (off), off, t); | |
6716 | } | |
6717 | else | |
6718 | roundup = NULL; | |
6719 | ||
6720 | /* Update ap.__[g|v]r_offs */ | |
6721 | t = build2 (PLUS_EXPR, TREE_TYPE (off), off, | |
6722 | build_int_cst (TREE_TYPE (off), rsize)); | |
6723 | t = build2 (MODIFY_EXPR, TREE_TYPE (f_off), unshare_expr (f_off), t); | |
6724 | ||
6725 | /* String up. */ | |
6726 | if (roundup) | |
6727 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), roundup, t); | |
6728 | ||
6729 | /* [cond2] if (ap.__[g|v]r_offs > 0) */ | |
6730 | u = build2 (GT_EXPR, boolean_type_node, unshare_expr (f_off), | |
6731 | build_int_cst (TREE_TYPE (f_off), 0)); | |
6732 | cond2 = build3 (COND_EXPR, ptr_type_node, u, NULL_TREE, NULL_TREE); | |
6733 | ||
6734 | /* String up: make sure the assignment happens before the use. */ | |
6735 | t = build2 (COMPOUND_EXPR, TREE_TYPE (cond2), t, cond2); | |
6736 | COND_EXPR_ELSE (cond1) = t; | |
6737 | ||
6738 | /* Prepare the trees handling the argument that is passed on the stack; | |
6739 | the top level node will store in ON_STACK. */ | |
6740 | arg = get_initialized_tmp_var (stack, pre_p, NULL); | |
6741 | if (align > 8) | |
6742 | { | |
6743 | /* if (alignof(type) > 8) (arg = arg + 15) & -16; */ | |
6744 | t = fold_convert (intDI_type_node, arg); | |
6745 | t = build2 (PLUS_EXPR, TREE_TYPE (t), t, | |
6746 | build_int_cst (TREE_TYPE (t), 15)); | |
6747 | t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t, | |
6748 | build_int_cst (TREE_TYPE (t), -16)); | |
6749 | t = fold_convert (TREE_TYPE (arg), t); | |
6750 | roundup = build2 (MODIFY_EXPR, TREE_TYPE (arg), arg, t); | |
6751 | } | |
6752 | else | |
6753 | roundup = NULL; | |
6754 | /* Advance ap.__stack */ | |
6755 | t = fold_convert (intDI_type_node, arg); | |
6756 | t = build2 (PLUS_EXPR, TREE_TYPE (t), t, | |
6757 | build_int_cst (TREE_TYPE (t), size + 7)); | |
6758 | t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t, | |
6759 | build_int_cst (TREE_TYPE (t), -8)); | |
6760 | t = fold_convert (TREE_TYPE (arg), t); | |
6761 | t = build2 (MODIFY_EXPR, TREE_TYPE (stack), unshare_expr (stack), t); | |
6762 | /* String up roundup and advance. */ | |
6763 | if (roundup) | |
6764 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), roundup, t); | |
6765 | /* String up with arg */ | |
6766 | on_stack = build2 (COMPOUND_EXPR, TREE_TYPE (arg), t, arg); | |
6767 | /* Big-endianness related address adjustment. */ | |
6768 | if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
6769 | && size < UNITS_PER_WORD) | |
6770 | { | |
6771 | t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (arg), arg, | |
6772 | size_int (UNITS_PER_WORD - size)); | |
6773 | on_stack = build2 (COMPOUND_EXPR, TREE_TYPE (arg), on_stack, t); | |
6774 | } | |
6775 | ||
6776 | COND_EXPR_THEN (cond1) = unshare_expr (on_stack); | |
6777 | COND_EXPR_THEN (cond2) = unshare_expr (on_stack); | |
6778 | ||
6779 | /* Adjustment to OFFSET in the case of BIG_ENDIAN. */ | |
6780 | t = off; | |
6781 | if (adjust) | |
6782 | t = build2 (PREINCREMENT_EXPR, TREE_TYPE (off), off, | |
6783 | build_int_cst (TREE_TYPE (off), adjust)); | |
6784 | ||
6785 | t = fold_convert (sizetype, t); | |
6786 | t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (f_top), f_top, t); | |
6787 | ||
6788 | if (is_ha) | |
6789 | { | |
6790 | /* type ha; // treat as "struct {ftype field[n];}" | |
6791 | ... [computing offs] | |
6792 | for (i = 0; i <nregs; ++i, offs += 16) | |
6793 | ha.field[i] = *((ftype *)(ap.__vr_top + offs)); | |
6794 | return ha; */ | |
6795 | int i; | |
6796 | tree tmp_ha, field_t, field_ptr_t; | |
6797 | ||
6798 | /* Declare a local variable. */ | |
6799 | tmp_ha = create_tmp_var_raw (type, "ha"); | |
6800 | gimple_add_tmp_var (tmp_ha); | |
6801 | ||
6802 | /* Establish the base type. */ | |
6803 | switch (ag_mode) | |
6804 | { | |
6805 | case SFmode: | |
6806 | field_t = float_type_node; | |
6807 | field_ptr_t = float_ptr_type_node; | |
6808 | break; | |
6809 | case DFmode: | |
6810 | field_t = double_type_node; | |
6811 | field_ptr_t = double_ptr_type_node; | |
6812 | break; | |
6813 | case TFmode: | |
6814 | field_t = long_double_type_node; | |
6815 | field_ptr_t = long_double_ptr_type_node; | |
6816 | break; | |
6817 | /* The half precision and quad precision are not fully supported yet. Enable | |
6818 | the following code after the support is complete. Need to find the correct | |
6819 | type node for __fp16 *. */ | |
6820 | #if 0 | |
6821 | case HFmode: | |
6822 | field_t = float_type_node; | |
6823 | field_ptr_t = float_ptr_type_node; | |
6824 | break; | |
6825 | #endif | |
6826 | case V2SImode: | |
6827 | case V4SImode: | |
6828 | { | |
6829 | tree innertype = make_signed_type (GET_MODE_PRECISION (SImode)); | |
6830 | field_t = build_vector_type_for_mode (innertype, ag_mode); | |
6831 | field_ptr_t = build_pointer_type (field_t); | |
6832 | } | |
6833 | break; | |
6834 | default: | |
6835 | gcc_assert (0); | |
6836 | } | |
6837 | ||
6838 | /* *(field_ptr_t)&ha = *((field_ptr_t)vr_saved_area */ | |
6839 | tmp_ha = build1 (ADDR_EXPR, field_ptr_t, tmp_ha); | |
6840 | addr = t; | |
6841 | t = fold_convert (field_ptr_t, addr); | |
6842 | t = build2 (MODIFY_EXPR, field_t, | |
6843 | build1 (INDIRECT_REF, field_t, tmp_ha), | |
6844 | build1 (INDIRECT_REF, field_t, t)); | |
6845 | ||
6846 | /* ha.field[i] = *((field_ptr_t)vr_saved_area + i) */ | |
6847 | for (i = 1; i < nregs; ++i) | |
6848 | { | |
6849 | addr = fold_build_pointer_plus_hwi (addr, UNITS_PER_VREG); | |
6850 | u = fold_convert (field_ptr_t, addr); | |
6851 | u = build2 (MODIFY_EXPR, field_t, | |
6852 | build2 (MEM_REF, field_t, tmp_ha, | |
6853 | build_int_cst (field_ptr_t, | |
6854 | (i * | |
6855 | int_size_in_bytes (field_t)))), | |
6856 | build1 (INDIRECT_REF, field_t, u)); | |
6857 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), t, u); | |
6858 | } | |
6859 | ||
6860 | u = fold_convert (TREE_TYPE (f_top), tmp_ha); | |
6861 | t = build2 (COMPOUND_EXPR, TREE_TYPE (f_top), t, u); | |
6862 | } | |
6863 | ||
6864 | COND_EXPR_ELSE (cond2) = t; | |
6865 | addr = fold_convert (build_pointer_type (type), cond1); | |
6866 | addr = build_va_arg_indirect_ref (addr); | |
6867 | ||
6868 | if (indirect_p) | |
6869 | addr = build_va_arg_indirect_ref (addr); | |
6870 | ||
6871 | return addr; | |
6872 | } | |
6873 | ||
6874 | /* Implement TARGET_SETUP_INCOMING_VARARGS. */ | |
6875 | ||
6876 | static void | |
6877 | aarch64_setup_incoming_varargs (cumulative_args_t cum_v, enum machine_mode mode, | |
6878 | tree type, int *pretend_size ATTRIBUTE_UNUSED, | |
6879 | int no_rtl) | |
6880 | { | |
6881 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
6882 | CUMULATIVE_ARGS local_cum; | |
6883 | int gr_saved, vr_saved; | |
6884 | ||
6885 | /* The caller has advanced CUM up to, but not beyond, the last named | |
6886 | argument. Advance a local copy of CUM past the last "real" named | |
6887 | argument, to find out how many registers are left over. */ | |
6888 | local_cum = *cum; | |
6889 | aarch64_function_arg_advance (pack_cumulative_args(&local_cum), mode, type, true); | |
6890 | ||
6891 | /* Found out how many registers we need to save. */ | |
6892 | gr_saved = NUM_ARG_REGS - local_cum.aapcs_ncrn; | |
6893 | vr_saved = NUM_FP_ARG_REGS - local_cum.aapcs_nvrn; | |
6894 | ||
6895 | if (TARGET_GENERAL_REGS_ONLY) | |
6896 | { | |
6897 | if (local_cum.aapcs_nvrn > 0) | |
6898 | sorry ("%qs and floating point or vector arguments", | |
6899 | "-mgeneral-regs-only"); | |
6900 | vr_saved = 0; | |
6901 | } | |
6902 | ||
6903 | if (!no_rtl) | |
6904 | { | |
6905 | if (gr_saved > 0) | |
6906 | { | |
6907 | rtx ptr, mem; | |
6908 | ||
6909 | /* virtual_incoming_args_rtx should have been 16-byte aligned. */ | |
6910 | ptr = plus_constant (Pmode, virtual_incoming_args_rtx, | |
6911 | - gr_saved * UNITS_PER_WORD); | |
6912 | mem = gen_frame_mem (BLKmode, ptr); | |
6913 | set_mem_alias_set (mem, get_varargs_alias_set ()); | |
6914 | ||
6915 | move_block_from_reg (local_cum.aapcs_ncrn + R0_REGNUM, | |
6916 | mem, gr_saved); | |
6917 | } | |
6918 | if (vr_saved > 0) | |
6919 | { | |
6920 | /* We can't use move_block_from_reg, because it will use | |
6921 | the wrong mode, storing D regs only. */ | |
6922 | enum machine_mode mode = TImode; | |
6923 | int off, i; | |
6924 | ||
6925 | /* Set OFF to the offset from virtual_incoming_args_rtx of | |
6926 | the first vector register. The VR save area lies below | |
6927 | the GR one, and is aligned to 16 bytes. */ | |
6928 | off = -AARCH64_ROUND_UP (gr_saved * UNITS_PER_WORD, | |
6929 | STACK_BOUNDARY / BITS_PER_UNIT); | |
6930 | off -= vr_saved * UNITS_PER_VREG; | |
6931 | ||
6932 | for (i = local_cum.aapcs_nvrn; i < NUM_FP_ARG_REGS; ++i) | |
6933 | { | |
6934 | rtx ptr, mem; | |
6935 | ||
6936 | ptr = plus_constant (Pmode, virtual_incoming_args_rtx, off); | |
6937 | mem = gen_frame_mem (mode, ptr); | |
6938 | set_mem_alias_set (mem, get_varargs_alias_set ()); | |
6939 | aarch64_emit_move (mem, gen_rtx_REG (mode, V0_REGNUM + i)); | |
6940 | off += UNITS_PER_VREG; | |
6941 | } | |
6942 | } | |
6943 | } | |
6944 | ||
6945 | /* We don't save the size into *PRETEND_SIZE because we want to avoid | |
6946 | any complication of having crtl->args.pretend_args_size changed. */ | |
6947 | cfun->machine->saved_varargs_size | |
6948 | = (AARCH64_ROUND_UP (gr_saved * UNITS_PER_WORD, | |
6949 | STACK_BOUNDARY / BITS_PER_UNIT) | |
6950 | + vr_saved * UNITS_PER_VREG); | |
6951 | } | |
6952 | ||
6953 | static void | |
6954 | aarch64_conditional_register_usage (void) | |
6955 | { | |
6956 | int i; | |
6957 | if (!TARGET_FLOAT) | |
6958 | { | |
6959 | for (i = V0_REGNUM; i <= V31_REGNUM; i++) | |
6960 | { | |
6961 | fixed_regs[i] = 1; | |
6962 | call_used_regs[i] = 1; | |
6963 | } | |
6964 | } | |
6965 | } | |
6966 | ||
6967 | /* Walk down the type tree of TYPE counting consecutive base elements. | |
6968 | If *MODEP is VOIDmode, then set it to the first valid floating point | |
6969 | type. If a non-floating point type is found, or if a floating point | |
6970 | type that doesn't match a non-VOIDmode *MODEP is found, then return -1, | |
6971 | otherwise return the count in the sub-tree. */ | |
6972 | static int | |
6973 | aapcs_vfp_sub_candidate (const_tree type, enum machine_mode *modep) | |
6974 | { | |
6975 | enum machine_mode mode; | |
6976 | HOST_WIDE_INT size; | |
6977 | ||
6978 | switch (TREE_CODE (type)) | |
6979 | { | |
6980 | case REAL_TYPE: | |
6981 | mode = TYPE_MODE (type); | |
6982 | if (mode != DFmode && mode != SFmode && mode != TFmode) | |
6983 | return -1; | |
6984 | ||
6985 | if (*modep == VOIDmode) | |
6986 | *modep = mode; | |
6987 | ||
6988 | if (*modep == mode) | |
6989 | return 1; | |
6990 | ||
6991 | break; | |
6992 | ||
6993 | case COMPLEX_TYPE: | |
6994 | mode = TYPE_MODE (TREE_TYPE (type)); | |
6995 | if (mode != DFmode && mode != SFmode && mode != TFmode) | |
6996 | return -1; | |
6997 | ||
6998 | if (*modep == VOIDmode) | |
6999 | *modep = mode; | |
7000 | ||
7001 | if (*modep == mode) | |
7002 | return 2; | |
7003 | ||
7004 | break; | |
7005 | ||
7006 | case VECTOR_TYPE: | |
7007 | /* Use V2SImode and V4SImode as representatives of all 64-bit | |
7008 | and 128-bit vector types. */ | |
7009 | size = int_size_in_bytes (type); | |
7010 | switch (size) | |
7011 | { | |
7012 | case 8: | |
7013 | mode = V2SImode; | |
7014 | break; | |
7015 | case 16: | |
7016 | mode = V4SImode; | |
7017 | break; | |
7018 | default: | |
7019 | return -1; | |
7020 | } | |
7021 | ||
7022 | if (*modep == VOIDmode) | |
7023 | *modep = mode; | |
7024 | ||
7025 | /* Vector modes are considered to be opaque: two vectors are | |
7026 | equivalent for the purposes of being homogeneous aggregates | |
7027 | if they are the same size. */ | |
7028 | if (*modep == mode) | |
7029 | return 1; | |
7030 | ||
7031 | break; | |
7032 | ||
7033 | case ARRAY_TYPE: | |
7034 | { | |
7035 | int count; | |
7036 | tree index = TYPE_DOMAIN (type); | |
7037 | ||
7038 | /* Can't handle incomplete types nor sizes that are not | |
7039 | fixed. */ | |
7040 | if (!COMPLETE_TYPE_P (type) | |
7041 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
7042 | return -1; | |
7043 | ||
7044 | count = aapcs_vfp_sub_candidate (TREE_TYPE (type), modep); | |
7045 | if (count == -1 | |
7046 | || !index | |
7047 | || !TYPE_MAX_VALUE (index) | |
7048 | || !tree_fits_uhwi_p (TYPE_MAX_VALUE (index)) | |
7049 | || !TYPE_MIN_VALUE (index) | |
7050 | || !tree_fits_uhwi_p (TYPE_MIN_VALUE (index)) | |
7051 | || count < 0) | |
7052 | return -1; | |
7053 | ||
7054 | count *= (1 + tree_to_uhwi (TYPE_MAX_VALUE (index)) | |
7055 | - tree_to_uhwi (TYPE_MIN_VALUE (index))); | |
7056 | ||
7057 | /* There must be no padding. */ | |
7058 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) | |
7059 | return -1; | |
7060 | ||
7061 | return count; | |
7062 | } | |
7063 | ||
7064 | case RECORD_TYPE: | |
7065 | { | |
7066 | int count = 0; | |
7067 | int sub_count; | |
7068 | tree field; | |
7069 | ||
7070 | /* Can't handle incomplete types nor sizes that are not | |
7071 | fixed. */ | |
7072 | if (!COMPLETE_TYPE_P (type) | |
7073 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
7074 | return -1; | |
7075 | ||
7076 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
7077 | { | |
7078 | if (TREE_CODE (field) != FIELD_DECL) | |
7079 | continue; | |
7080 | ||
7081 | sub_count = aapcs_vfp_sub_candidate (TREE_TYPE (field), modep); | |
7082 | if (sub_count < 0) | |
7083 | return -1; | |
7084 | count += sub_count; | |
7085 | } | |
7086 | ||
7087 | /* There must be no padding. */ | |
7088 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) | |
7089 | return -1; | |
7090 | ||
7091 | return count; | |
7092 | } | |
7093 | ||
7094 | case UNION_TYPE: | |
7095 | case QUAL_UNION_TYPE: | |
7096 | { | |
7097 | /* These aren't very interesting except in a degenerate case. */ | |
7098 | int count = 0; | |
7099 | int sub_count; | |
7100 | tree field; | |
7101 | ||
7102 | /* Can't handle incomplete types nor sizes that are not | |
7103 | fixed. */ | |
7104 | if (!COMPLETE_TYPE_P (type) | |
7105 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
7106 | return -1; | |
7107 | ||
7108 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
7109 | { | |
7110 | if (TREE_CODE (field) != FIELD_DECL) | |
7111 | continue; | |
7112 | ||
7113 | sub_count = aapcs_vfp_sub_candidate (TREE_TYPE (field), modep); | |
7114 | if (sub_count < 0) | |
7115 | return -1; | |
7116 | count = count > sub_count ? count : sub_count; | |
7117 | } | |
7118 | ||
7119 | /* There must be no padding. */ | |
7120 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) | |
7121 | return -1; | |
7122 | ||
7123 | return count; | |
7124 | } | |
7125 | ||
7126 | default: | |
7127 | break; | |
7128 | } | |
7129 | ||
7130 | return -1; | |
7131 | } | |
7132 | ||
7133 | /* Return true if we use LRA instead of reload pass. */ | |
7134 | static bool | |
7135 | aarch64_lra_p (void) | |
7136 | { | |
7137 | return aarch64_lra_flag; | |
7138 | } | |
7139 | ||
7140 | /* Return TRUE if the type, as described by TYPE and MODE, is a composite | |
7141 | type as described in AAPCS64 \S 4.3. This includes aggregate, union and | |
7142 | array types. The C99 floating-point complex types are also considered | |
7143 | as composite types, according to AAPCS64 \S 7.1.1. The complex integer | |
7144 | types, which are GCC extensions and out of the scope of AAPCS64, are | |
7145 | treated as composite types here as well. | |
7146 | ||
7147 | Note that MODE itself is not sufficient in determining whether a type | |
7148 | is such a composite type or not. This is because | |
7149 | stor-layout.c:compute_record_mode may have already changed the MODE | |
7150 | (BLKmode) of a RECORD_TYPE TYPE to some other mode. For example, a | |
7151 | structure with only one field may have its MODE set to the mode of the | |
7152 | field. Also an integer mode whose size matches the size of the | |
7153 | RECORD_TYPE type may be used to substitute the original mode | |
7154 | (i.e. BLKmode) in certain circumstances. In other words, MODE cannot be | |
7155 | solely relied on. */ | |
7156 | ||
7157 | static bool | |
7158 | aarch64_composite_type_p (const_tree type, | |
7159 | enum machine_mode mode) | |
7160 | { | |
7161 | if (type && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == COMPLEX_TYPE)) | |
7162 | return true; | |
7163 | ||
7164 | if (mode == BLKmode | |
7165 | || GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT | |
7166 | || GET_MODE_CLASS (mode) == MODE_COMPLEX_INT) | |
7167 | return true; | |
7168 | ||
7169 | return false; | |
7170 | } | |
7171 | ||
7172 | /* Return TRUE if the type, as described by TYPE and MODE, is a short vector | |
7173 | type as described in AAPCS64 \S 4.1.2. | |
7174 | ||
7175 | See the comment above aarch64_composite_type_p for the notes on MODE. */ | |
7176 | ||
7177 | static bool | |
7178 | aarch64_short_vector_p (const_tree type, | |
7179 | enum machine_mode mode) | |
7180 | { | |
7181 | HOST_WIDE_INT size = -1; | |
7182 | ||
7183 | if (type && TREE_CODE (type) == VECTOR_TYPE) | |
7184 | size = int_size_in_bytes (type); | |
7185 | else if (!aarch64_composite_type_p (type, mode) | |
7186 | && (GET_MODE_CLASS (mode) == MODE_VECTOR_INT | |
7187 | || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT)) | |
7188 | size = GET_MODE_SIZE (mode); | |
7189 | ||
7190 | return (size == 8 || size == 16) ? true : false; | |
7191 | } | |
7192 | ||
7193 | /* Return TRUE if an argument, whose type is described by TYPE and MODE, | |
7194 | shall be passed or returned in simd/fp register(s) (providing these | |
7195 | parameter passing registers are available). | |
7196 | ||
7197 | Upon successful return, *COUNT returns the number of needed registers, | |
7198 | *BASE_MODE returns the mode of the individual register and when IS_HAF | |
7199 | is not NULL, *IS_HA indicates whether or not the argument is a homogeneous | |
7200 | floating-point aggregate or a homogeneous short-vector aggregate. */ | |
7201 | ||
7202 | static bool | |
7203 | aarch64_vfp_is_call_or_return_candidate (enum machine_mode mode, | |
7204 | const_tree type, | |
7205 | enum machine_mode *base_mode, | |
7206 | int *count, | |
7207 | bool *is_ha) | |
7208 | { | |
7209 | enum machine_mode new_mode = VOIDmode; | |
7210 | bool composite_p = aarch64_composite_type_p (type, mode); | |
7211 | ||
7212 | if (is_ha != NULL) *is_ha = false; | |
7213 | ||
7214 | if ((!composite_p && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
7215 | || aarch64_short_vector_p (type, mode)) | |
7216 | { | |
7217 | *count = 1; | |
7218 | new_mode = mode; | |
7219 | } | |
7220 | else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT) | |
7221 | { | |
7222 | if (is_ha != NULL) *is_ha = true; | |
7223 | *count = 2; | |
7224 | new_mode = GET_MODE_INNER (mode); | |
7225 | } | |
7226 | else if (type && composite_p) | |
7227 | { | |
7228 | int ag_count = aapcs_vfp_sub_candidate (type, &new_mode); | |
7229 | ||
7230 | if (ag_count > 0 && ag_count <= HA_MAX_NUM_FLDS) | |
7231 | { | |
7232 | if (is_ha != NULL) *is_ha = true; | |
7233 | *count = ag_count; | |
7234 | } | |
7235 | else | |
7236 | return false; | |
7237 | } | |
7238 | else | |
7239 | return false; | |
7240 | ||
7241 | *base_mode = new_mode; | |
7242 | return true; | |
7243 | } | |
7244 | ||
7245 | /* Implement TARGET_STRUCT_VALUE_RTX. */ | |
7246 | ||
7247 | static rtx | |
7248 | aarch64_struct_value_rtx (tree fndecl ATTRIBUTE_UNUSED, | |
7249 | int incoming ATTRIBUTE_UNUSED) | |
7250 | { | |
7251 | return gen_rtx_REG (Pmode, AARCH64_STRUCT_VALUE_REGNUM); | |
7252 | } | |
7253 | ||
7254 | /* Implements target hook vector_mode_supported_p. */ | |
7255 | static bool | |
7256 | aarch64_vector_mode_supported_p (enum machine_mode mode) | |
7257 | { | |
7258 | if (TARGET_SIMD | |
7259 | && (mode == V4SImode || mode == V8HImode | |
7260 | || mode == V16QImode || mode == V2DImode | |
7261 | || mode == V2SImode || mode == V4HImode | |
7262 | || mode == V8QImode || mode == V2SFmode | |
7263 | || mode == V4SFmode || mode == V2DFmode)) | |
7264 | return true; | |
7265 | ||
7266 | return false; | |
7267 | } | |
7268 | ||
7269 | /* Return appropriate SIMD container | |
7270 | for MODE within a vector of WIDTH bits. */ | |
7271 | static enum machine_mode | |
7272 | aarch64_simd_container_mode (enum machine_mode mode, unsigned width) | |
7273 | { | |
7274 | gcc_assert (width == 64 || width == 128); | |
7275 | if (TARGET_SIMD) | |
7276 | { | |
7277 | if (width == 128) | |
7278 | switch (mode) | |
7279 | { | |
7280 | case DFmode: | |
7281 | return V2DFmode; | |
7282 | case SFmode: | |
7283 | return V4SFmode; | |
7284 | case SImode: | |
7285 | return V4SImode; | |
7286 | case HImode: | |
7287 | return V8HImode; | |
7288 | case QImode: | |
7289 | return V16QImode; | |
7290 | case DImode: | |
7291 | return V2DImode; | |
7292 | default: | |
7293 | break; | |
7294 | } | |
7295 | else | |
7296 | switch (mode) | |
7297 | { | |
7298 | case SFmode: | |
7299 | return V2SFmode; | |
7300 | case SImode: | |
7301 | return V2SImode; | |
7302 | case HImode: | |
7303 | return V4HImode; | |
7304 | case QImode: | |
7305 | return V8QImode; | |
7306 | default: | |
7307 | break; | |
7308 | } | |
7309 | } | |
7310 | return word_mode; | |
7311 | } | |
7312 | ||
7313 | /* Return 128-bit container as the preferred SIMD mode for MODE. */ | |
7314 | static enum machine_mode | |
7315 | aarch64_preferred_simd_mode (enum machine_mode mode) | |
7316 | { | |
7317 | return aarch64_simd_container_mode (mode, 128); | |
7318 | } | |
7319 | ||
7320 | /* Return the bitmask of possible vector sizes for the vectorizer | |
7321 | to iterate over. */ | |
7322 | static unsigned int | |
7323 | aarch64_autovectorize_vector_sizes (void) | |
7324 | { | |
7325 | return (16 | 8); | |
7326 | } | |
7327 | ||
7328 | /* A table to help perform AArch64-specific name mangling for AdvSIMD | |
7329 | vector types in order to conform to the AAPCS64 (see "Procedure | |
7330 | Call Standard for the ARM 64-bit Architecture", Appendix A). To | |
7331 | qualify for emission with the mangled names defined in that document, | |
7332 | a vector type must not only be of the correct mode but also be | |
7333 | composed of AdvSIMD vector element types (e.g. | |
7334 | _builtin_aarch64_simd_qi); these types are registered by | |
7335 | aarch64_init_simd_builtins (). In other words, vector types defined | |
7336 | in other ways e.g. via vector_size attribute will get default | |
7337 | mangled names. */ | |
7338 | typedef struct | |
7339 | { | |
7340 | enum machine_mode mode; | |
7341 | const char *element_type_name; | |
7342 | const char *mangled_name; | |
7343 | } aarch64_simd_mangle_map_entry; | |
7344 | ||
7345 | static aarch64_simd_mangle_map_entry aarch64_simd_mangle_map[] = { | |
7346 | /* 64-bit containerized types. */ | |
7347 | { V8QImode, "__builtin_aarch64_simd_qi", "10__Int8x8_t" }, | |
7348 | { V8QImode, "__builtin_aarch64_simd_uqi", "11__Uint8x8_t" }, | |
7349 | { V4HImode, "__builtin_aarch64_simd_hi", "11__Int16x4_t" }, | |
7350 | { V4HImode, "__builtin_aarch64_simd_uhi", "12__Uint16x4_t" }, | |
7351 | { V2SImode, "__builtin_aarch64_simd_si", "11__Int32x2_t" }, | |
7352 | { V2SImode, "__builtin_aarch64_simd_usi", "12__Uint32x2_t" }, | |
7353 | { V2SFmode, "__builtin_aarch64_simd_sf", "13__Float32x2_t" }, | |
7354 | { V8QImode, "__builtin_aarch64_simd_poly8", "11__Poly8x8_t" }, | |
7355 | { V4HImode, "__builtin_aarch64_simd_poly16", "12__Poly16x4_t" }, | |
7356 | /* 128-bit containerized types. */ | |
7357 | { V16QImode, "__builtin_aarch64_simd_qi", "11__Int8x16_t" }, | |
7358 | { V16QImode, "__builtin_aarch64_simd_uqi", "12__Uint8x16_t" }, | |
7359 | { V8HImode, "__builtin_aarch64_simd_hi", "11__Int16x8_t" }, | |
7360 | { V8HImode, "__builtin_aarch64_simd_uhi", "12__Uint16x8_t" }, | |
7361 | { V4SImode, "__builtin_aarch64_simd_si", "11__Int32x4_t" }, | |
7362 | { V4SImode, "__builtin_aarch64_simd_usi", "12__Uint32x4_t" }, | |
7363 | { V2DImode, "__builtin_aarch64_simd_di", "11__Int64x2_t" }, | |
7364 | { V2DImode, "__builtin_aarch64_simd_udi", "12__Uint64x2_t" }, | |
7365 | { V4SFmode, "__builtin_aarch64_simd_sf", "13__Float32x4_t" }, | |
7366 | { V2DFmode, "__builtin_aarch64_simd_df", "13__Float64x2_t" }, | |
7367 | { V16QImode, "__builtin_aarch64_simd_poly8", "12__Poly8x16_t" }, | |
7368 | { V8HImode, "__builtin_aarch64_simd_poly16", "12__Poly16x8_t" }, | |
7369 | { V2DImode, "__builtin_aarch64_simd_poly64", "12__Poly64x2_t" }, | |
7370 | { VOIDmode, NULL, NULL } | |
7371 | }; | |
7372 | ||
7373 | /* Implement TARGET_MANGLE_TYPE. */ | |
7374 | ||
7375 | static const char * | |
7376 | aarch64_mangle_type (const_tree type) | |
7377 | { | |
7378 | /* The AArch64 ABI documents say that "__va_list" has to be | |
7379 | managled as if it is in the "std" namespace. */ | |
7380 | if (lang_hooks.types_compatible_p (CONST_CAST_TREE (type), va_list_type)) | |
7381 | return "St9__va_list"; | |
7382 | ||
7383 | /* Check the mode of the vector type, and the name of the vector | |
7384 | element type, against the table. */ | |
7385 | if (TREE_CODE (type) == VECTOR_TYPE) | |
7386 | { | |
7387 | aarch64_simd_mangle_map_entry *pos = aarch64_simd_mangle_map; | |
7388 | ||
7389 | while (pos->mode != VOIDmode) | |
7390 | { | |
7391 | tree elt_type = TREE_TYPE (type); | |
7392 | ||
7393 | if (pos->mode == TYPE_MODE (type) | |
7394 | && TREE_CODE (TYPE_NAME (elt_type)) == TYPE_DECL | |
7395 | && !strcmp (IDENTIFIER_POINTER (DECL_NAME (TYPE_NAME (elt_type))), | |
7396 | pos->element_type_name)) | |
7397 | return pos->mangled_name; | |
7398 | ||
7399 | pos++; | |
7400 | } | |
7401 | } | |
7402 | ||
7403 | /* Use the default mangling. */ | |
7404 | return NULL; | |
7405 | } | |
7406 | ||
7407 | /* Return the equivalent letter for size. */ | |
7408 | static char | |
7409 | sizetochar (int size) | |
7410 | { | |
7411 | switch (size) | |
7412 | { | |
7413 | case 64: return 'd'; | |
7414 | case 32: return 's'; | |
7415 | case 16: return 'h'; | |
7416 | case 8 : return 'b'; | |
7417 | default: gcc_unreachable (); | |
7418 | } | |
7419 | } | |
7420 | ||
7421 | /* Return true iff x is a uniform vector of floating-point | |
7422 | constants, and the constant can be represented in | |
7423 | quarter-precision form. Note, as aarch64_float_const_representable | |
7424 | rejects both +0.0 and -0.0, we will also reject +0.0 and -0.0. */ | |
7425 | static bool | |
7426 | aarch64_vect_float_const_representable_p (rtx x) | |
7427 | { | |
7428 | int i = 0; | |
7429 | REAL_VALUE_TYPE r0, ri; | |
7430 | rtx x0, xi; | |
7431 | ||
7432 | if (GET_MODE_CLASS (GET_MODE (x)) != MODE_VECTOR_FLOAT) | |
7433 | return false; | |
7434 | ||
7435 | x0 = CONST_VECTOR_ELT (x, 0); | |
7436 | if (!CONST_DOUBLE_P (x0)) | |
7437 | return false; | |
7438 | ||
7439 | REAL_VALUE_FROM_CONST_DOUBLE (r0, x0); | |
7440 | ||
7441 | for (i = 1; i < CONST_VECTOR_NUNITS (x); i++) | |
7442 | { | |
7443 | xi = CONST_VECTOR_ELT (x, i); | |
7444 | if (!CONST_DOUBLE_P (xi)) | |
7445 | return false; | |
7446 | ||
7447 | REAL_VALUE_FROM_CONST_DOUBLE (ri, xi); | |
7448 | if (!REAL_VALUES_EQUAL (r0, ri)) | |
7449 | return false; | |
7450 | } | |
7451 | ||
7452 | return aarch64_float_const_representable_p (x0); | |
7453 | } | |
7454 | ||
7455 | /* Return true for valid and false for invalid. */ | |
7456 | bool | |
7457 | aarch64_simd_valid_immediate (rtx op, enum machine_mode mode, bool inverse, | |
7458 | struct simd_immediate_info *info) | |
7459 | { | |
7460 | #define CHECK(STRIDE, ELSIZE, CLASS, TEST, SHIFT, NEG) \ | |
7461 | matches = 1; \ | |
7462 | for (i = 0; i < idx; i += (STRIDE)) \ | |
7463 | if (!(TEST)) \ | |
7464 | matches = 0; \ | |
7465 | if (matches) \ | |
7466 | { \ | |
7467 | immtype = (CLASS); \ | |
7468 | elsize = (ELSIZE); \ | |
7469 | eshift = (SHIFT); \ | |
7470 | emvn = (NEG); \ | |
7471 | break; \ | |
7472 | } | |
7473 | ||
7474 | unsigned int i, elsize = 0, idx = 0, n_elts = CONST_VECTOR_NUNITS (op); | |
7475 | unsigned int innersize = GET_MODE_SIZE (GET_MODE_INNER (mode)); | |
7476 | unsigned char bytes[16]; | |
7477 | int immtype = -1, matches; | |
7478 | unsigned int invmask = inverse ? 0xff : 0; | |
7479 | int eshift, emvn; | |
7480 | ||
7481 | if (GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) | |
7482 | { | |
7483 | if (! (aarch64_simd_imm_zero_p (op, mode) | |
7484 | || aarch64_vect_float_const_representable_p (op))) | |
7485 | return false; | |
7486 | ||
7487 | if (info) | |
7488 | { | |
7489 | info->value = CONST_VECTOR_ELT (op, 0); | |
7490 | info->element_width = GET_MODE_BITSIZE (GET_MODE (info->value)); | |
7491 | info->mvn = false; | |
7492 | info->shift = 0; | |
7493 | } | |
7494 | ||
7495 | return true; | |
7496 | } | |
7497 | ||
7498 | /* Splat vector constant out into a byte vector. */ | |
7499 | for (i = 0; i < n_elts; i++) | |
7500 | { | |
7501 | /* The vector is provided in gcc endian-neutral fashion. For aarch64_be, | |
7502 | it must be laid out in the vector register in reverse order. */ | |
7503 | rtx el = CONST_VECTOR_ELT (op, BYTES_BIG_ENDIAN ? (n_elts - 1 - i) : i); | |
7504 | unsigned HOST_WIDE_INT elpart; | |
7505 | unsigned int part, parts; | |
7506 | ||
7507 | if (GET_CODE (el) == CONST_INT) | |
7508 | { | |
7509 | elpart = INTVAL (el); | |
7510 | parts = 1; | |
7511 | } | |
7512 | else if (GET_CODE (el) == CONST_DOUBLE) | |
7513 | { | |
7514 | elpart = CONST_DOUBLE_LOW (el); | |
7515 | parts = 2; | |
7516 | } | |
7517 | else | |
7518 | gcc_unreachable (); | |
7519 | ||
7520 | for (part = 0; part < parts; part++) | |
7521 | { | |
7522 | unsigned int byte; | |
7523 | for (byte = 0; byte < innersize; byte++) | |
7524 | { | |
7525 | bytes[idx++] = (elpart & 0xff) ^ invmask; | |
7526 | elpart >>= BITS_PER_UNIT; | |
7527 | } | |
7528 | if (GET_CODE (el) == CONST_DOUBLE) | |
7529 | elpart = CONST_DOUBLE_HIGH (el); | |
7530 | } | |
7531 | } | |
7532 | ||
7533 | /* Sanity check. */ | |
7534 | gcc_assert (idx == GET_MODE_SIZE (mode)); | |
7535 | ||
7536 | do | |
7537 | { | |
7538 | CHECK (4, 32, 0, bytes[i] == bytes[0] && bytes[i + 1] == 0 | |
7539 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 0, 0); | |
7540 | ||
7541 | CHECK (4, 32, 1, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
7542 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 8, 0); | |
7543 | ||
7544 | CHECK (4, 32, 2, bytes[i] == 0 && bytes[i + 1] == 0 | |
7545 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0, 16, 0); | |
7546 | ||
7547 | CHECK (4, 32, 3, bytes[i] == 0 && bytes[i + 1] == 0 | |
7548 | && bytes[i + 2] == 0 && bytes[i + 3] == bytes[3], 24, 0); | |
7549 | ||
7550 | CHECK (2, 16, 4, bytes[i] == bytes[0] && bytes[i + 1] == 0, 0, 0); | |
7551 | ||
7552 | CHECK (2, 16, 5, bytes[i] == 0 && bytes[i + 1] == bytes[1], 8, 0); | |
7553 | ||
7554 | CHECK (4, 32, 6, bytes[i] == bytes[0] && bytes[i + 1] == 0xff | |
7555 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 0, 1); | |
7556 | ||
7557 | CHECK (4, 32, 7, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
7558 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 8, 1); | |
7559 | ||
7560 | CHECK (4, 32, 8, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
7561 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff, 16, 1); | |
7562 | ||
7563 | CHECK (4, 32, 9, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
7564 | && bytes[i + 2] == 0xff && bytes[i + 3] == bytes[3], 24, 1); | |
7565 | ||
7566 | CHECK (2, 16, 10, bytes[i] == bytes[0] && bytes[i + 1] == 0xff, 0, 1); | |
7567 | ||
7568 | CHECK (2, 16, 11, bytes[i] == 0xff && bytes[i + 1] == bytes[1], 8, 1); | |
7569 | ||
7570 | CHECK (4, 32, 12, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
7571 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 8, 0); | |
7572 | ||
7573 | CHECK (4, 32, 13, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
7574 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 8, 1); | |
7575 | ||
7576 | CHECK (4, 32, 14, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
7577 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0, 16, 0); | |
7578 | ||
7579 | CHECK (4, 32, 15, bytes[i] == 0 && bytes[i + 1] == 0 | |
7580 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff, 16, 1); | |
7581 | ||
7582 | CHECK (1, 8, 16, bytes[i] == bytes[0], 0, 0); | |
7583 | ||
7584 | CHECK (1, 64, 17, (bytes[i] == 0 || bytes[i] == 0xff) | |
7585 | && bytes[i] == bytes[(i + 8) % idx], 0, 0); | |
7586 | } | |
7587 | while (0); | |
7588 | ||
7589 | if (immtype == -1) | |
7590 | return false; | |
7591 | ||
7592 | if (info) | |
7593 | { | |
7594 | info->element_width = elsize; | |
7595 | info->mvn = emvn != 0; | |
7596 | info->shift = eshift; | |
7597 | ||
7598 | unsigned HOST_WIDE_INT imm = 0; | |
7599 | ||
7600 | if (immtype >= 12 && immtype <= 15) | |
7601 | info->msl = true; | |
7602 | ||
7603 | /* Un-invert bytes of recognized vector, if necessary. */ | |
7604 | if (invmask != 0) | |
7605 | for (i = 0; i < idx; i++) | |
7606 | bytes[i] ^= invmask; | |
7607 | ||
7608 | if (immtype == 17) | |
7609 | { | |
7610 | /* FIXME: Broken on 32-bit H_W_I hosts. */ | |
7611 | gcc_assert (sizeof (HOST_WIDE_INT) == 8); | |
7612 | ||
7613 | for (i = 0; i < 8; i++) | |
7614 | imm |= (unsigned HOST_WIDE_INT) (bytes[i] ? 0xff : 0) | |
7615 | << (i * BITS_PER_UNIT); | |
7616 | ||
7617 | ||
7618 | info->value = GEN_INT (imm); | |
7619 | } | |
7620 | else | |
7621 | { | |
7622 | for (i = 0; i < elsize / BITS_PER_UNIT; i++) | |
7623 | imm |= (unsigned HOST_WIDE_INT) bytes[i] << (i * BITS_PER_UNIT); | |
7624 | ||
7625 | /* Construct 'abcdefgh' because the assembler cannot handle | |
7626 | generic constants. */ | |
7627 | if (info->mvn) | |
7628 | imm = ~imm; | |
7629 | imm = (imm >> info->shift) & 0xff; | |
7630 | info->value = GEN_INT (imm); | |
7631 | } | |
7632 | } | |
7633 | ||
7634 | return true; | |
7635 | #undef CHECK | |
7636 | } | |
7637 | ||
7638 | static bool | |
7639 | aarch64_const_vec_all_same_int_p (rtx x, | |
7640 | HOST_WIDE_INT minval, | |
7641 | HOST_WIDE_INT maxval) | |
7642 | { | |
7643 | HOST_WIDE_INT firstval; | |
7644 | int count, i; | |
7645 | ||
7646 | if (GET_CODE (x) != CONST_VECTOR | |
7647 | || GET_MODE_CLASS (GET_MODE (x)) != MODE_VECTOR_INT) | |
7648 | return false; | |
7649 | ||
7650 | firstval = INTVAL (CONST_VECTOR_ELT (x, 0)); | |
7651 | if (firstval < minval || firstval > maxval) | |
7652 | return false; | |
7653 | ||
7654 | count = CONST_VECTOR_NUNITS (x); | |
7655 | for (i = 1; i < count; i++) | |
7656 | if (INTVAL (CONST_VECTOR_ELT (x, i)) != firstval) | |
7657 | return false; | |
7658 | ||
7659 | return true; | |
7660 | } | |
7661 | ||
7662 | /* Check of immediate shift constants are within range. */ | |
7663 | bool | |
7664 | aarch64_simd_shift_imm_p (rtx x, enum machine_mode mode, bool left) | |
7665 | { | |
7666 | int bit_width = GET_MODE_UNIT_SIZE (mode) * BITS_PER_UNIT; | |
7667 | if (left) | |
7668 | return aarch64_const_vec_all_same_int_p (x, 0, bit_width - 1); | |
7669 | else | |
7670 | return aarch64_const_vec_all_same_int_p (x, 1, bit_width); | |
7671 | } | |
7672 | ||
7673 | /* Return true if X is a uniform vector where all elements | |
7674 | are either the floating-point constant 0.0 or the | |
7675 | integer constant 0. */ | |
7676 | bool | |
7677 | aarch64_simd_imm_zero_p (rtx x, enum machine_mode mode) | |
7678 | { | |
7679 | return x == CONST0_RTX (mode); | |
7680 | } | |
7681 | ||
7682 | bool | |
7683 | aarch64_simd_imm_scalar_p (rtx x, enum machine_mode mode ATTRIBUTE_UNUSED) | |
7684 | { | |
7685 | HOST_WIDE_INT imm = INTVAL (x); | |
7686 | int i; | |
7687 | ||
7688 | for (i = 0; i < 8; i++) | |
7689 | { | |
7690 | unsigned int byte = imm & 0xff; | |
7691 | if (byte != 0xff && byte != 0) | |
7692 | return false; | |
7693 | imm >>= 8; | |
7694 | } | |
7695 | ||
7696 | return true; | |
7697 | } | |
7698 | ||
7699 | bool | |
7700 | aarch64_mov_operand_p (rtx x, | |
7701 | enum aarch64_symbol_context context, | |
7702 | enum machine_mode mode) | |
7703 | { | |
7704 | if (GET_CODE (x) == HIGH | |
7705 | && aarch64_valid_symref (XEXP (x, 0), GET_MODE (XEXP (x, 0)))) | |
7706 | return true; | |
7707 | ||
7708 | if (CONST_INT_P (x) && aarch64_move_imm (INTVAL (x), mode)) | |
7709 | return true; | |
7710 | ||
7711 | if (GET_CODE (x) == SYMBOL_REF && mode == DImode && CONSTANT_ADDRESS_P (x)) | |
7712 | return true; | |
7713 | ||
7714 | return aarch64_classify_symbolic_expression (x, context) | |
7715 | == SYMBOL_TINY_ABSOLUTE; | |
7716 | } | |
7717 | ||
7718 | /* Return a const_int vector of VAL. */ | |
7719 | rtx | |
7720 | aarch64_simd_gen_const_vector_dup (enum machine_mode mode, int val) | |
7721 | { | |
7722 | int nunits = GET_MODE_NUNITS (mode); | |
7723 | rtvec v = rtvec_alloc (nunits); | |
7724 | int i; | |
7725 | ||
7726 | for (i=0; i < nunits; i++) | |
7727 | RTVEC_ELT (v, i) = GEN_INT (val); | |
7728 | ||
7729 | return gen_rtx_CONST_VECTOR (mode, v); | |
7730 | } | |
7731 | ||
7732 | /* Check OP is a legal scalar immediate for the MOVI instruction. */ | |
7733 | ||
7734 | bool | |
7735 | aarch64_simd_scalar_immediate_valid_for_move (rtx op, enum machine_mode mode) | |
7736 | { | |
7737 | enum machine_mode vmode; | |
7738 | ||
7739 | gcc_assert (!VECTOR_MODE_P (mode)); | |
7740 | vmode = aarch64_preferred_simd_mode (mode); | |
7741 | rtx op_v = aarch64_simd_gen_const_vector_dup (vmode, INTVAL (op)); | |
7742 | return aarch64_simd_valid_immediate (op_v, vmode, false, NULL); | |
7743 | } | |
7744 | ||
7745 | /* Construct and return a PARALLEL RTX vector. */ | |
7746 | rtx | |
7747 | aarch64_simd_vect_par_cnst_half (enum machine_mode mode, bool high) | |
7748 | { | |
7749 | int nunits = GET_MODE_NUNITS (mode); | |
7750 | rtvec v = rtvec_alloc (nunits / 2); | |
7751 | int base = high ? nunits / 2 : 0; | |
7752 | rtx t1; | |
7753 | int i; | |
7754 | ||
7755 | for (i=0; i < nunits / 2; i++) | |
7756 | RTVEC_ELT (v, i) = GEN_INT (base + i); | |
7757 | ||
7758 | t1 = gen_rtx_PARALLEL (mode, v); | |
7759 | return t1; | |
7760 | } | |
7761 | ||
7762 | /* Bounds-check lanes. Ensure OPERAND lies between LOW (inclusive) and | |
7763 | HIGH (exclusive). */ | |
7764 | void | |
7765 | aarch64_simd_lane_bounds (rtx operand, HOST_WIDE_INT low, HOST_WIDE_INT high) | |
7766 | { | |
7767 | HOST_WIDE_INT lane; | |
7768 | gcc_assert (GET_CODE (operand) == CONST_INT); | |
7769 | lane = INTVAL (operand); | |
7770 | ||
7771 | if (lane < low || lane >= high) | |
7772 | error ("lane out of range"); | |
7773 | } | |
7774 | ||
7775 | void | |
7776 | aarch64_simd_const_bounds (rtx operand, HOST_WIDE_INT low, HOST_WIDE_INT high) | |
7777 | { | |
7778 | gcc_assert (GET_CODE (operand) == CONST_INT); | |
7779 | HOST_WIDE_INT lane = INTVAL (operand); | |
7780 | ||
7781 | if (lane < low || lane >= high) | |
7782 | error ("constant out of range"); | |
7783 | } | |
7784 | ||
7785 | /* Emit code to reinterpret one AdvSIMD type as another, | |
7786 | without altering bits. */ | |
7787 | void | |
7788 | aarch64_simd_reinterpret (rtx dest, rtx src) | |
7789 | { | |
7790 | emit_move_insn (dest, gen_lowpart (GET_MODE (dest), src)); | |
7791 | } | |
7792 | ||
7793 | /* Emit code to place a AdvSIMD pair result in memory locations (with equal | |
7794 | registers). */ | |
7795 | void | |
7796 | aarch64_simd_emit_pair_result_insn (enum machine_mode mode, | |
7797 | rtx (*intfn) (rtx, rtx, rtx), rtx destaddr, | |
7798 | rtx op1) | |
7799 | { | |
7800 | rtx mem = gen_rtx_MEM (mode, destaddr); | |
7801 | rtx tmp1 = gen_reg_rtx (mode); | |
7802 | rtx tmp2 = gen_reg_rtx (mode); | |
7803 | ||
7804 | emit_insn (intfn (tmp1, op1, tmp2)); | |
7805 | ||
7806 | emit_move_insn (mem, tmp1); | |
7807 | mem = adjust_address (mem, mode, GET_MODE_SIZE (mode)); | |
7808 | emit_move_insn (mem, tmp2); | |
7809 | } | |
7810 | ||
7811 | /* Return TRUE if OP is a valid vector addressing mode. */ | |
7812 | bool | |
7813 | aarch64_simd_mem_operand_p (rtx op) | |
7814 | { | |
7815 | return MEM_P (op) && (GET_CODE (XEXP (op, 0)) == POST_INC | |
7816 | || GET_CODE (XEXP (op, 0)) == REG); | |
7817 | } | |
7818 | ||
7819 | /* Set up OPERANDS for a register copy from SRC to DEST, taking care | |
7820 | not to early-clobber SRC registers in the process. | |
7821 | ||
7822 | We assume that the operands described by SRC and DEST represent a | |
7823 | decomposed copy of OPERANDS[1] into OPERANDS[0]. COUNT is the | |
7824 | number of components into which the copy has been decomposed. */ | |
7825 | void | |
7826 | aarch64_simd_disambiguate_copy (rtx *operands, rtx *dest, | |
7827 | rtx *src, unsigned int count) | |
7828 | { | |
7829 | unsigned int i; | |
7830 | ||
7831 | if (!reg_overlap_mentioned_p (operands[0], operands[1]) | |
7832 | || REGNO (operands[0]) < REGNO (operands[1])) | |
7833 | { | |
7834 | for (i = 0; i < count; i++) | |
7835 | { | |
7836 | operands[2 * i] = dest[i]; | |
7837 | operands[2 * i + 1] = src[i]; | |
7838 | } | |
7839 | } | |
7840 | else | |
7841 | { | |
7842 | for (i = 0; i < count; i++) | |
7843 | { | |
7844 | operands[2 * i] = dest[count - i - 1]; | |
7845 | operands[2 * i + 1] = src[count - i - 1]; | |
7846 | } | |
7847 | } | |
7848 | } | |
7849 | ||
7850 | /* Compute and return the length of aarch64_simd_mov<mode>, where <mode> is | |
7851 | one of VSTRUCT modes: OI, CI or XI. */ | |
7852 | int | |
7853 | aarch64_simd_attr_length_move (rtx insn) | |
7854 | { | |
7855 | enum machine_mode mode; | |
7856 | ||
7857 | extract_insn_cached (insn); | |
7858 | ||
7859 | if (REG_P (recog_data.operand[0]) && REG_P (recog_data.operand[1])) | |
7860 | { | |
7861 | mode = GET_MODE (recog_data.operand[0]); | |
7862 | switch (mode) | |
7863 | { | |
7864 | case OImode: | |
7865 | return 8; | |
7866 | case CImode: | |
7867 | return 12; | |
7868 | case XImode: | |
7869 | return 16; | |
7870 | default: | |
7871 | gcc_unreachable (); | |
7872 | } | |
7873 | } | |
7874 | return 4; | |
7875 | } | |
7876 | ||
7877 | /* Implement target hook TARGET_VECTOR_ALIGNMENT. The AAPCS64 sets the maximum | |
7878 | alignment of a vector to 128 bits. */ | |
7879 | static HOST_WIDE_INT | |
7880 | aarch64_simd_vector_alignment (const_tree type) | |
7881 | { | |
7882 | HOST_WIDE_INT align = tree_to_shwi (TYPE_SIZE (type)); | |
7883 | return MIN (align, 128); | |
7884 | } | |
7885 | ||
7886 | /* Implement target hook TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE. */ | |
7887 | static bool | |
7888 | aarch64_simd_vector_alignment_reachable (const_tree type, bool is_packed) | |
7889 | { | |
7890 | if (is_packed) | |
7891 | return false; | |
7892 | ||
7893 | /* We guarantee alignment for vectors up to 128-bits. */ | |
7894 | if (tree_int_cst_compare (TYPE_SIZE (type), | |
7895 | bitsize_int (BIGGEST_ALIGNMENT)) > 0) | |
7896 | return false; | |
7897 | ||
7898 | /* Vectors whose size is <= BIGGEST_ALIGNMENT are naturally aligned. */ | |
7899 | return true; | |
7900 | } | |
7901 | ||
7902 | /* If VALS is a vector constant that can be loaded into a register | |
7903 | using DUP, generate instructions to do so and return an RTX to | |
7904 | assign to the register. Otherwise return NULL_RTX. */ | |
7905 | static rtx | |
7906 | aarch64_simd_dup_constant (rtx vals) | |
7907 | { | |
7908 | enum machine_mode mode = GET_MODE (vals); | |
7909 | enum machine_mode inner_mode = GET_MODE_INNER (mode); | |
7910 | int n_elts = GET_MODE_NUNITS (mode); | |
7911 | bool all_same = true; | |
7912 | rtx x; | |
7913 | int i; | |
7914 | ||
7915 | if (GET_CODE (vals) != CONST_VECTOR) | |
7916 | return NULL_RTX; | |
7917 | ||
7918 | for (i = 1; i < n_elts; ++i) | |
7919 | { | |
7920 | x = CONST_VECTOR_ELT (vals, i); | |
7921 | if (!rtx_equal_p (x, CONST_VECTOR_ELT (vals, 0))) | |
7922 | all_same = false; | |
7923 | } | |
7924 | ||
7925 | if (!all_same) | |
7926 | return NULL_RTX; | |
7927 | ||
7928 | /* We can load this constant by using DUP and a constant in a | |
7929 | single ARM register. This will be cheaper than a vector | |
7930 | load. */ | |
7931 | x = copy_to_mode_reg (inner_mode, CONST_VECTOR_ELT (vals, 0)); | |
7932 | return gen_rtx_VEC_DUPLICATE (mode, x); | |
7933 | } | |
7934 | ||
7935 | ||
7936 | /* Generate code to load VALS, which is a PARALLEL containing only | |
7937 | constants (for vec_init) or CONST_VECTOR, efficiently into a | |
7938 | register. Returns an RTX to copy into the register, or NULL_RTX | |
7939 | for a PARALLEL that can not be converted into a CONST_VECTOR. */ | |
7940 | static rtx | |
7941 | aarch64_simd_make_constant (rtx vals) | |
7942 | { | |
7943 | enum machine_mode mode = GET_MODE (vals); | |
7944 | rtx const_dup; | |
7945 | rtx const_vec = NULL_RTX; | |
7946 | int n_elts = GET_MODE_NUNITS (mode); | |
7947 | int n_const = 0; | |
7948 | int i; | |
7949 | ||
7950 | if (GET_CODE (vals) == CONST_VECTOR) | |
7951 | const_vec = vals; | |
7952 | else if (GET_CODE (vals) == PARALLEL) | |
7953 | { | |
7954 | /* A CONST_VECTOR must contain only CONST_INTs and | |
7955 | CONST_DOUBLEs, but CONSTANT_P allows more (e.g. SYMBOL_REF). | |
7956 | Only store valid constants in a CONST_VECTOR. */ | |
7957 | for (i = 0; i < n_elts; ++i) | |
7958 | { | |
7959 | rtx x = XVECEXP (vals, 0, i); | |
7960 | if (CONST_INT_P (x) || CONST_DOUBLE_P (x)) | |
7961 | n_const++; | |
7962 | } | |
7963 | if (n_const == n_elts) | |
7964 | const_vec = gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0)); | |
7965 | } | |
7966 | else | |
7967 | gcc_unreachable (); | |
7968 | ||
7969 | if (const_vec != NULL_RTX | |
7970 | && aarch64_simd_valid_immediate (const_vec, mode, false, NULL)) | |
7971 | /* Load using MOVI/MVNI. */ | |
7972 | return const_vec; | |
7973 | else if ((const_dup = aarch64_simd_dup_constant (vals)) != NULL_RTX) | |
7974 | /* Loaded using DUP. */ | |
7975 | return const_dup; | |
7976 | else if (const_vec != NULL_RTX) | |
7977 | /* Load from constant pool. We can not take advantage of single-cycle | |
7978 | LD1 because we need a PC-relative addressing mode. */ | |
7979 | return const_vec; | |
7980 | else | |
7981 | /* A PARALLEL containing something not valid inside CONST_VECTOR. | |
7982 | We can not construct an initializer. */ | |
7983 | return NULL_RTX; | |
7984 | } | |
7985 | ||
7986 | void | |
7987 | aarch64_expand_vector_init (rtx target, rtx vals) | |
7988 | { | |
7989 | enum machine_mode mode = GET_MODE (target); | |
7990 | enum machine_mode inner_mode = GET_MODE_INNER (mode); | |
7991 | int n_elts = GET_MODE_NUNITS (mode); | |
7992 | int n_var = 0, one_var = -1; | |
7993 | bool all_same = true; | |
7994 | rtx x, mem; | |
7995 | int i; | |
7996 | ||
7997 | x = XVECEXP (vals, 0, 0); | |
7998 | if (!CONST_INT_P (x) && !CONST_DOUBLE_P (x)) | |
7999 | n_var = 1, one_var = 0; | |
8000 | ||
8001 | for (i = 1; i < n_elts; ++i) | |
8002 | { | |
8003 | x = XVECEXP (vals, 0, i); | |
8004 | if (!CONST_INT_P (x) && !CONST_DOUBLE_P (x)) | |
8005 | ++n_var, one_var = i; | |
8006 | ||
8007 | if (!rtx_equal_p (x, XVECEXP (vals, 0, 0))) | |
8008 | all_same = false; | |
8009 | } | |
8010 | ||
8011 | if (n_var == 0) | |
8012 | { | |
8013 | rtx constant = aarch64_simd_make_constant (vals); | |
8014 | if (constant != NULL_RTX) | |
8015 | { | |
8016 | emit_move_insn (target, constant); | |
8017 | return; | |
8018 | } | |
8019 | } | |
8020 | ||
8021 | /* Splat a single non-constant element if we can. */ | |
8022 | if (all_same) | |
8023 | { | |
8024 | x = copy_to_mode_reg (inner_mode, XVECEXP (vals, 0, 0)); | |
8025 | aarch64_emit_move (target, gen_rtx_VEC_DUPLICATE (mode, x)); | |
8026 | return; | |
8027 | } | |
8028 | ||
8029 | /* One field is non-constant. Load constant then overwrite varying | |
8030 | field. This is more efficient than using the stack. */ | |
8031 | if (n_var == 1) | |
8032 | { | |
8033 | rtx copy = copy_rtx (vals); | |
8034 | rtx index = GEN_INT (one_var); | |
8035 | enum insn_code icode; | |
8036 | ||
8037 | /* Load constant part of vector, substitute neighboring value for | |
8038 | varying element. */ | |
8039 | XVECEXP (copy, 0, one_var) = XVECEXP (vals, 0, one_var ^ 1); | |
8040 | aarch64_expand_vector_init (target, copy); | |
8041 | ||
8042 | /* Insert variable. */ | |
8043 | x = copy_to_mode_reg (inner_mode, XVECEXP (vals, 0, one_var)); | |
8044 | icode = optab_handler (vec_set_optab, mode); | |
8045 | gcc_assert (icode != CODE_FOR_nothing); | |
8046 | emit_insn (GEN_FCN (icode) (target, x, index)); | |
8047 | return; | |
8048 | } | |
8049 | ||
8050 | /* Construct the vector in memory one field at a time | |
8051 | and load the whole vector. */ | |
8052 | mem = assign_stack_temp (mode, GET_MODE_SIZE (mode)); | |
8053 | for (i = 0; i < n_elts; i++) | |
8054 | emit_move_insn (adjust_address_nv (mem, inner_mode, | |
8055 | i * GET_MODE_SIZE (inner_mode)), | |
8056 | XVECEXP (vals, 0, i)); | |
8057 | emit_move_insn (target, mem); | |
8058 | ||
8059 | } | |
8060 | ||
8061 | static unsigned HOST_WIDE_INT | |
8062 | aarch64_shift_truncation_mask (enum machine_mode mode) | |
8063 | { | |
8064 | return | |
8065 | (aarch64_vector_mode_supported_p (mode) | |
8066 | || aarch64_vect_struct_mode_p (mode)) ? 0 : (GET_MODE_BITSIZE (mode) - 1); | |
8067 | } | |
8068 | ||
8069 | #ifndef TLS_SECTION_ASM_FLAG | |
8070 | #define TLS_SECTION_ASM_FLAG 'T' | |
8071 | #endif | |
8072 | ||
8073 | void | |
8074 | aarch64_elf_asm_named_section (const char *name, unsigned int flags, | |
8075 | tree decl ATTRIBUTE_UNUSED) | |
8076 | { | |
8077 | char flagchars[10], *f = flagchars; | |
8078 | ||
8079 | /* If we have already declared this section, we can use an | |
8080 | abbreviated form to switch back to it -- unless this section is | |
8081 | part of a COMDAT groups, in which case GAS requires the full | |
8082 | declaration every time. */ | |
8083 | if (!(HAVE_COMDAT_GROUP && (flags & SECTION_LINKONCE)) | |
8084 | && (flags & SECTION_DECLARED)) | |
8085 | { | |
8086 | fprintf (asm_out_file, "\t.section\t%s\n", name); | |
8087 | return; | |
8088 | } | |
8089 | ||
8090 | if (!(flags & SECTION_DEBUG)) | |
8091 | *f++ = 'a'; | |
8092 | if (flags & SECTION_WRITE) | |
8093 | *f++ = 'w'; | |
8094 | if (flags & SECTION_CODE) | |
8095 | *f++ = 'x'; | |
8096 | if (flags & SECTION_SMALL) | |
8097 | *f++ = 's'; | |
8098 | if (flags & SECTION_MERGE) | |
8099 | *f++ = 'M'; | |
8100 | if (flags & SECTION_STRINGS) | |
8101 | *f++ = 'S'; | |
8102 | if (flags & SECTION_TLS) | |
8103 | *f++ = TLS_SECTION_ASM_FLAG; | |
8104 | if (HAVE_COMDAT_GROUP && (flags & SECTION_LINKONCE)) | |
8105 | *f++ = 'G'; | |
8106 | *f = '\0'; | |
8107 | ||
8108 | fprintf (asm_out_file, "\t.section\t%s,\"%s\"", name, flagchars); | |
8109 | ||
8110 | if (!(flags & SECTION_NOTYPE)) | |
8111 | { | |
8112 | const char *type; | |
8113 | const char *format; | |
8114 | ||
8115 | if (flags & SECTION_BSS) | |
8116 | type = "nobits"; | |
8117 | else | |
8118 | type = "progbits"; | |
8119 | ||
8120 | #ifdef TYPE_OPERAND_FMT | |
8121 | format = "," TYPE_OPERAND_FMT; | |
8122 | #else | |
8123 | format = ",@%s"; | |
8124 | #endif | |
8125 | ||
8126 | fprintf (asm_out_file, format, type); | |
8127 | ||
8128 | if (flags & SECTION_ENTSIZE) | |
8129 | fprintf (asm_out_file, ",%d", flags & SECTION_ENTSIZE); | |
8130 | if (HAVE_COMDAT_GROUP && (flags & SECTION_LINKONCE)) | |
8131 | { | |
8132 | if (TREE_CODE (decl) == IDENTIFIER_NODE) | |
8133 | fprintf (asm_out_file, ",%s,comdat", IDENTIFIER_POINTER (decl)); | |
8134 | else | |
8135 | fprintf (asm_out_file, ",%s,comdat", | |
8136 | IDENTIFIER_POINTER (DECL_COMDAT_GROUP (decl))); | |
8137 | } | |
8138 | } | |
8139 | ||
8140 | putc ('\n', asm_out_file); | |
8141 | } | |
8142 | ||
8143 | /* Select a format to encode pointers in exception handling data. */ | |
8144 | int | |
8145 | aarch64_asm_preferred_eh_data_format (int code ATTRIBUTE_UNUSED, int global) | |
8146 | { | |
8147 | int type; | |
8148 | switch (aarch64_cmodel) | |
8149 | { | |
8150 | case AARCH64_CMODEL_TINY: | |
8151 | case AARCH64_CMODEL_TINY_PIC: | |
8152 | case AARCH64_CMODEL_SMALL: | |
8153 | case AARCH64_CMODEL_SMALL_PIC: | |
8154 | /* text+got+data < 4Gb. 4-byte signed relocs are sufficient | |
8155 | for everything. */ | |
8156 | type = DW_EH_PE_sdata4; | |
8157 | break; | |
8158 | default: | |
8159 | /* No assumptions here. 8-byte relocs required. */ | |
8160 | type = DW_EH_PE_sdata8; | |
8161 | break; | |
8162 | } | |
8163 | return (global ? DW_EH_PE_indirect : 0) | DW_EH_PE_pcrel | type; | |
8164 | } | |
8165 | ||
8166 | /* Emit load exclusive. */ | |
8167 | ||
8168 | static void | |
8169 | aarch64_emit_load_exclusive (enum machine_mode mode, rtx rval, | |
8170 | rtx mem, rtx model_rtx) | |
8171 | { | |
8172 | rtx (*gen) (rtx, rtx, rtx); | |
8173 | ||
8174 | switch (mode) | |
8175 | { | |
8176 | case QImode: gen = gen_aarch64_load_exclusiveqi; break; | |
8177 | case HImode: gen = gen_aarch64_load_exclusivehi; break; | |
8178 | case SImode: gen = gen_aarch64_load_exclusivesi; break; | |
8179 | case DImode: gen = gen_aarch64_load_exclusivedi; break; | |
8180 | default: | |
8181 | gcc_unreachable (); | |
8182 | } | |
8183 | ||
8184 | emit_insn (gen (rval, mem, model_rtx)); | |
8185 | } | |
8186 | ||
8187 | /* Emit store exclusive. */ | |
8188 | ||
8189 | static void | |
8190 | aarch64_emit_store_exclusive (enum machine_mode mode, rtx bval, | |
8191 | rtx rval, rtx mem, rtx model_rtx) | |
8192 | { | |
8193 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
8194 | ||
8195 | switch (mode) | |
8196 | { | |
8197 | case QImode: gen = gen_aarch64_store_exclusiveqi; break; | |
8198 | case HImode: gen = gen_aarch64_store_exclusivehi; break; | |
8199 | case SImode: gen = gen_aarch64_store_exclusivesi; break; | |
8200 | case DImode: gen = gen_aarch64_store_exclusivedi; break; | |
8201 | default: | |
8202 | gcc_unreachable (); | |
8203 | } | |
8204 | ||
8205 | emit_insn (gen (bval, rval, mem, model_rtx)); | |
8206 | } | |
8207 | ||
8208 | /* Mark the previous jump instruction as unlikely. */ | |
8209 | ||
8210 | static void | |
8211 | aarch64_emit_unlikely_jump (rtx insn) | |
8212 | { | |
8213 | int very_unlikely = REG_BR_PROB_BASE / 100 - 1; | |
8214 | ||
8215 | insn = emit_jump_insn (insn); | |
8216 | add_int_reg_note (insn, REG_BR_PROB, very_unlikely); | |
8217 | } | |
8218 | ||
8219 | /* Expand a compare and swap pattern. */ | |
8220 | ||
8221 | void | |
8222 | aarch64_expand_compare_and_swap (rtx operands[]) | |
8223 | { | |
8224 | rtx bval, rval, mem, oldval, newval, is_weak, mod_s, mod_f, x; | |
8225 | enum machine_mode mode, cmp_mode; | |
8226 | rtx (*gen) (rtx, rtx, rtx, rtx, rtx, rtx, rtx); | |
8227 | ||
8228 | bval = operands[0]; | |
8229 | rval = operands[1]; | |
8230 | mem = operands[2]; | |
8231 | oldval = operands[3]; | |
8232 | newval = operands[4]; | |
8233 | is_weak = operands[5]; | |
8234 | mod_s = operands[6]; | |
8235 | mod_f = operands[7]; | |
8236 | mode = GET_MODE (mem); | |
8237 | cmp_mode = mode; | |
8238 | ||
8239 | /* Normally the succ memory model must be stronger than fail, but in the | |
8240 | unlikely event of fail being ACQUIRE and succ being RELEASE we need to | |
8241 | promote succ to ACQ_REL so that we don't lose the acquire semantics. */ | |
8242 | ||
8243 | if (INTVAL (mod_f) == MEMMODEL_ACQUIRE | |
8244 | && INTVAL (mod_s) == MEMMODEL_RELEASE) | |
8245 | mod_s = GEN_INT (MEMMODEL_ACQ_REL); | |
8246 | ||
8247 | switch (mode) | |
8248 | { | |
8249 | case QImode: | |
8250 | case HImode: | |
8251 | /* For short modes, we're going to perform the comparison in SImode, | |
8252 | so do the zero-extension now. */ | |
8253 | cmp_mode = SImode; | |
8254 | rval = gen_reg_rtx (SImode); | |
8255 | oldval = convert_modes (SImode, mode, oldval, true); | |
8256 | /* Fall through. */ | |
8257 | ||
8258 | case SImode: | |
8259 | case DImode: | |
8260 | /* Force the value into a register if needed. */ | |
8261 | if (!aarch64_plus_operand (oldval, mode)) | |
8262 | oldval = force_reg (cmp_mode, oldval); | |
8263 | break; | |
8264 | ||
8265 | default: | |
8266 | gcc_unreachable (); | |
8267 | } | |
8268 | ||
8269 | switch (mode) | |
8270 | { | |
8271 | case QImode: gen = gen_atomic_compare_and_swapqi_1; break; | |
8272 | case HImode: gen = gen_atomic_compare_and_swaphi_1; break; | |
8273 | case SImode: gen = gen_atomic_compare_and_swapsi_1; break; | |
8274 | case DImode: gen = gen_atomic_compare_and_swapdi_1; break; | |
8275 | default: | |
8276 | gcc_unreachable (); | |
8277 | } | |
8278 | ||
8279 | emit_insn (gen (rval, mem, oldval, newval, is_weak, mod_s, mod_f)); | |
8280 | ||
8281 | if (mode == QImode || mode == HImode) | |
8282 | emit_move_insn (operands[1], gen_lowpart (mode, rval)); | |
8283 | ||
8284 | x = gen_rtx_REG (CCmode, CC_REGNUM); | |
8285 | x = gen_rtx_EQ (SImode, x, const0_rtx); | |
8286 | emit_insn (gen_rtx_SET (VOIDmode, bval, x)); | |
8287 | } | |
8288 | ||
8289 | /* Split a compare and swap pattern. */ | |
8290 | ||
8291 | void | |
8292 | aarch64_split_compare_and_swap (rtx operands[]) | |
8293 | { | |
8294 | rtx rval, mem, oldval, newval, scratch; | |
8295 | enum machine_mode mode; | |
8296 | bool is_weak; | |
8297 | rtx label1, label2, x, cond; | |
8298 | ||
8299 | rval = operands[0]; | |
8300 | mem = operands[1]; | |
8301 | oldval = operands[2]; | |
8302 | newval = operands[3]; | |
8303 | is_weak = (operands[4] != const0_rtx); | |
8304 | scratch = operands[7]; | |
8305 | mode = GET_MODE (mem); | |
8306 | ||
8307 | label1 = NULL_RTX; | |
8308 | if (!is_weak) | |
8309 | { | |
8310 | label1 = gen_label_rtx (); | |
8311 | emit_label (label1); | |
8312 | } | |
8313 | label2 = gen_label_rtx (); | |
8314 | ||
8315 | aarch64_emit_load_exclusive (mode, rval, mem, operands[5]); | |
8316 | ||
8317 | cond = aarch64_gen_compare_reg (NE, rval, oldval); | |
8318 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
8319 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
8320 | gen_rtx_LABEL_REF (Pmode, label2), pc_rtx); | |
8321 | aarch64_emit_unlikely_jump (gen_rtx_SET (VOIDmode, pc_rtx, x)); | |
8322 | ||
8323 | aarch64_emit_store_exclusive (mode, scratch, mem, newval, operands[5]); | |
8324 | ||
8325 | if (!is_weak) | |
8326 | { | |
8327 | x = gen_rtx_NE (VOIDmode, scratch, const0_rtx); | |
8328 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
8329 | gen_rtx_LABEL_REF (Pmode, label1), pc_rtx); | |
8330 | aarch64_emit_unlikely_jump (gen_rtx_SET (VOIDmode, pc_rtx, x)); | |
8331 | } | |
8332 | else | |
8333 | { | |
8334 | cond = gen_rtx_REG (CCmode, CC_REGNUM); | |
8335 | x = gen_rtx_COMPARE (CCmode, scratch, const0_rtx); | |
8336 | emit_insn (gen_rtx_SET (VOIDmode, cond, x)); | |
8337 | } | |
8338 | ||
8339 | emit_label (label2); | |
8340 | } | |
8341 | ||
8342 | /* Split an atomic operation. */ | |
8343 | ||
8344 | void | |
8345 | aarch64_split_atomic_op (enum rtx_code code, rtx old_out, rtx new_out, rtx mem, | |
8346 | rtx value, rtx model_rtx, rtx cond) | |
8347 | { | |
8348 | enum machine_mode mode = GET_MODE (mem); | |
8349 | enum machine_mode wmode = (mode == DImode ? DImode : SImode); | |
8350 | rtx label, x; | |
8351 | ||
8352 | label = gen_label_rtx (); | |
8353 | emit_label (label); | |
8354 | ||
8355 | if (new_out) | |
8356 | new_out = gen_lowpart (wmode, new_out); | |
8357 | if (old_out) | |
8358 | old_out = gen_lowpart (wmode, old_out); | |
8359 | else | |
8360 | old_out = new_out; | |
8361 | value = simplify_gen_subreg (wmode, value, mode, 0); | |
8362 | ||
8363 | aarch64_emit_load_exclusive (mode, old_out, mem, model_rtx); | |
8364 | ||
8365 | switch (code) | |
8366 | { | |
8367 | case SET: | |
8368 | new_out = value; | |
8369 | break; | |
8370 | ||
8371 | case NOT: | |
8372 | x = gen_rtx_AND (wmode, old_out, value); | |
8373 | emit_insn (gen_rtx_SET (VOIDmode, new_out, x)); | |
8374 | x = gen_rtx_NOT (wmode, new_out); | |
8375 | emit_insn (gen_rtx_SET (VOIDmode, new_out, x)); | |
8376 | break; | |
8377 | ||
8378 | case MINUS: | |
8379 | if (CONST_INT_P (value)) | |
8380 | { | |
8381 | value = GEN_INT (-INTVAL (value)); | |
8382 | code = PLUS; | |
8383 | } | |
8384 | /* Fall through. */ | |
8385 | ||
8386 | default: | |
8387 | x = gen_rtx_fmt_ee (code, wmode, old_out, value); | |
8388 | emit_insn (gen_rtx_SET (VOIDmode, new_out, x)); | |
8389 | break; | |
8390 | } | |
8391 | ||
8392 | aarch64_emit_store_exclusive (mode, cond, mem, | |
8393 | gen_lowpart (mode, new_out), model_rtx); | |
8394 | ||
8395 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
8396 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
8397 | gen_rtx_LABEL_REF (Pmode, label), pc_rtx); | |
8398 | aarch64_emit_unlikely_jump (gen_rtx_SET (VOIDmode, pc_rtx, x)); | |
8399 | } | |
8400 | ||
8401 | static void | |
8402 | aarch64_print_extension (void) | |
8403 | { | |
8404 | const struct aarch64_option_extension *opt = NULL; | |
8405 | ||
8406 | for (opt = all_extensions; opt->name != NULL; opt++) | |
8407 | if ((aarch64_isa_flags & opt->flags_on) == opt->flags_on) | |
8408 | asm_fprintf (asm_out_file, "+%s", opt->name); | |
8409 | ||
8410 | asm_fprintf (asm_out_file, "\n"); | |
8411 | } | |
8412 | ||
8413 | static void | |
8414 | aarch64_start_file (void) | |
8415 | { | |
8416 | if (selected_arch) | |
8417 | { | |
8418 | asm_fprintf (asm_out_file, "\t.arch %s", selected_arch->name); | |
8419 | aarch64_print_extension (); | |
8420 | } | |
8421 | else if (selected_cpu) | |
8422 | { | |
8423 | const char *truncated_name | |
8424 | = aarch64_rewrite_selected_cpu (selected_cpu->name); | |
8425 | asm_fprintf (asm_out_file, "\t.cpu %s", truncated_name); | |
8426 | aarch64_print_extension (); | |
8427 | } | |
8428 | default_file_start(); | |
8429 | } | |
8430 | ||
8431 | /* Target hook for c_mode_for_suffix. */ | |
8432 | static enum machine_mode | |
8433 | aarch64_c_mode_for_suffix (char suffix) | |
8434 | { | |
8435 | if (suffix == 'q') | |
8436 | return TFmode; | |
8437 | ||
8438 | return VOIDmode; | |
8439 | } | |
8440 | ||
8441 | /* We can only represent floating point constants which will fit in | |
8442 | "quarter-precision" values. These values are characterised by | |
8443 | a sign bit, a 4-bit mantissa and a 3-bit exponent. And are given | |
8444 | by: | |
8445 | ||
8446 | (-1)^s * (n/16) * 2^r | |
8447 | ||
8448 | Where: | |
8449 | 's' is the sign bit. | |
8450 | 'n' is an integer in the range 16 <= n <= 31. | |
8451 | 'r' is an integer in the range -3 <= r <= 4. */ | |
8452 | ||
8453 | /* Return true iff X can be represented by a quarter-precision | |
8454 | floating point immediate operand X. Note, we cannot represent 0.0. */ | |
8455 | bool | |
8456 | aarch64_float_const_representable_p (rtx x) | |
8457 | { | |
8458 | /* This represents our current view of how many bits | |
8459 | make up the mantissa. */ | |
8460 | int point_pos = 2 * HOST_BITS_PER_WIDE_INT - 1; | |
8461 | int exponent; | |
8462 | unsigned HOST_WIDE_INT mantissa, mask; | |
8463 | REAL_VALUE_TYPE r, m; | |
8464 | bool fail; | |
8465 | ||
8466 | if (!CONST_DOUBLE_P (x)) | |
8467 | return false; | |
8468 | ||
8469 | if (GET_MODE (x) == VOIDmode) | |
8470 | return false; | |
8471 | ||
8472 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
8473 | ||
8474 | /* We cannot represent infinities, NaNs or +/-zero. We won't | |
8475 | know if we have +zero until we analyse the mantissa, but we | |
8476 | can reject the other invalid values. */ | |
8477 | if (REAL_VALUE_ISINF (r) || REAL_VALUE_ISNAN (r) | |
8478 | || REAL_VALUE_MINUS_ZERO (r)) | |
8479 | return false; | |
8480 | ||
8481 | /* Extract exponent. */ | |
8482 | r = real_value_abs (&r); | |
8483 | exponent = REAL_EXP (&r); | |
8484 | ||
8485 | /* For the mantissa, we expand into two HOST_WIDE_INTS, apart from the | |
8486 | highest (sign) bit, with a fixed binary point at bit point_pos. | |
8487 | m1 holds the low part of the mantissa, m2 the high part. | |
8488 | WARNING: If we ever have a representation using more than 2 * H_W_I - 1 | |
8489 | bits for the mantissa, this can fail (low bits will be lost). */ | |
8490 | real_ldexp (&m, &r, point_pos - exponent); | |
8491 | wide_int w = real_to_integer (&m, &fail, HOST_BITS_PER_WIDE_INT * 2); | |
8492 | ||
8493 | /* If the low part of the mantissa has bits set we cannot represent | |
8494 | the value. */ | |
8495 | if (w.elt (0) != 0) | |
8496 | return false; | |
8497 | /* We have rejected the lower HOST_WIDE_INT, so update our | |
8498 | understanding of how many bits lie in the mantissa and | |
8499 | look only at the high HOST_WIDE_INT. */ | |
8500 | mantissa = w.elt (1); | |
8501 | point_pos -= HOST_BITS_PER_WIDE_INT; | |
8502 | ||
8503 | /* We can only represent values with a mantissa of the form 1.xxxx. */ | |
8504 | mask = ((unsigned HOST_WIDE_INT)1 << (point_pos - 5)) - 1; | |
8505 | if ((mantissa & mask) != 0) | |
8506 | return false; | |
8507 | ||
8508 | /* Having filtered unrepresentable values, we may now remove all | |
8509 | but the highest 5 bits. */ | |
8510 | mantissa >>= point_pos - 5; | |
8511 | ||
8512 | /* We cannot represent the value 0.0, so reject it. This is handled | |
8513 | elsewhere. */ | |
8514 | if (mantissa == 0) | |
8515 | return false; | |
8516 | ||
8517 | /* Then, as bit 4 is always set, we can mask it off, leaving | |
8518 | the mantissa in the range [0, 15]. */ | |
8519 | mantissa &= ~(1 << 4); | |
8520 | gcc_assert (mantissa <= 15); | |
8521 | ||
8522 | /* GCC internally does not use IEEE754-like encoding (where normalized | |
8523 | significands are in the range [1, 2). GCC uses [0.5, 1) (see real.c). | |
8524 | Our mantissa values are shifted 4 places to the left relative to | |
8525 | normalized IEEE754 so we must modify the exponent returned by REAL_EXP | |
8526 | by 5 places to correct for GCC's representation. */ | |
8527 | exponent = 5 - exponent; | |
8528 | ||
8529 | return (exponent >= 0 && exponent <= 7); | |
8530 | } | |
8531 | ||
8532 | char* | |
8533 | aarch64_output_simd_mov_immediate (rtx const_vector, | |
8534 | enum machine_mode mode, | |
8535 | unsigned width) | |
8536 | { | |
8537 | bool is_valid; | |
8538 | static char templ[40]; | |
8539 | const char *mnemonic; | |
8540 | const char *shift_op; | |
8541 | unsigned int lane_count = 0; | |
8542 | char element_char; | |
8543 | ||
8544 | struct simd_immediate_info info = { NULL_RTX, 0, 0, false, false }; | |
8545 | ||
8546 | /* This will return true to show const_vector is legal for use as either | |
8547 | a AdvSIMD MOVI instruction (or, implicitly, MVNI) immediate. It will | |
8548 | also update INFO to show how the immediate should be generated. */ | |
8549 | is_valid = aarch64_simd_valid_immediate (const_vector, mode, false, &info); | |
8550 | gcc_assert (is_valid); | |
8551 | ||
8552 | element_char = sizetochar (info.element_width); | |
8553 | lane_count = width / info.element_width; | |
8554 | ||
8555 | mode = GET_MODE_INNER (mode); | |
8556 | if (mode == SFmode || mode == DFmode) | |
8557 | { | |
8558 | gcc_assert (info.shift == 0 && ! info.mvn); | |
8559 | if (aarch64_float_const_zero_rtx_p (info.value)) | |
8560 | info.value = GEN_INT (0); | |
8561 | else | |
8562 | { | |
8563 | #define buf_size 20 | |
8564 | REAL_VALUE_TYPE r; | |
8565 | REAL_VALUE_FROM_CONST_DOUBLE (r, info.value); | |
8566 | char float_buf[buf_size] = {'\0'}; | |
8567 | real_to_decimal_for_mode (float_buf, &r, buf_size, buf_size, 1, mode); | |
8568 | #undef buf_size | |
8569 | ||
8570 | if (lane_count == 1) | |
8571 | snprintf (templ, sizeof (templ), "fmov\t%%d0, %s", float_buf); | |
8572 | else | |
8573 | snprintf (templ, sizeof (templ), "fmov\t%%0.%d%c, %s", | |
8574 | lane_count, element_char, float_buf); | |
8575 | return templ; | |
8576 | } | |
8577 | } | |
8578 | ||
8579 | mnemonic = info.mvn ? "mvni" : "movi"; | |
8580 | shift_op = info.msl ? "msl" : "lsl"; | |
8581 | ||
8582 | if (lane_count == 1) | |
8583 | snprintf (templ, sizeof (templ), "%s\t%%d0, " HOST_WIDE_INT_PRINT_HEX, | |
8584 | mnemonic, UINTVAL (info.value)); | |
8585 | else if (info.shift) | |
8586 | snprintf (templ, sizeof (templ), "%s\t%%0.%d%c, " HOST_WIDE_INT_PRINT_HEX | |
8587 | ", %s %d", mnemonic, lane_count, element_char, | |
8588 | UINTVAL (info.value), shift_op, info.shift); | |
8589 | else | |
8590 | snprintf (templ, sizeof (templ), "%s\t%%0.%d%c, " HOST_WIDE_INT_PRINT_HEX, | |
8591 | mnemonic, lane_count, element_char, UINTVAL (info.value)); | |
8592 | return templ; | |
8593 | } | |
8594 | ||
8595 | char* | |
8596 | aarch64_output_scalar_simd_mov_immediate (rtx immediate, | |
8597 | enum machine_mode mode) | |
8598 | { | |
8599 | enum machine_mode vmode; | |
8600 | ||
8601 | gcc_assert (!VECTOR_MODE_P (mode)); | |
8602 | vmode = aarch64_simd_container_mode (mode, 64); | |
8603 | rtx v_op = aarch64_simd_gen_const_vector_dup (vmode, INTVAL (immediate)); | |
8604 | return aarch64_output_simd_mov_immediate (v_op, vmode, 64); | |
8605 | } | |
8606 | ||
8607 | /* Split operands into moves from op[1] + op[2] into op[0]. */ | |
8608 | ||
8609 | void | |
8610 | aarch64_split_combinev16qi (rtx operands[3]) | |
8611 | { | |
8612 | unsigned int dest = REGNO (operands[0]); | |
8613 | unsigned int src1 = REGNO (operands[1]); | |
8614 | unsigned int src2 = REGNO (operands[2]); | |
8615 | enum machine_mode halfmode = GET_MODE (operands[1]); | |
8616 | unsigned int halfregs = HARD_REGNO_NREGS (src1, halfmode); | |
8617 | rtx destlo, desthi; | |
8618 | ||
8619 | gcc_assert (halfmode == V16QImode); | |
8620 | ||
8621 | if (src1 == dest && src2 == dest + halfregs) | |
8622 | { | |
8623 | /* No-op move. Can't split to nothing; emit something. */ | |
8624 | emit_note (NOTE_INSN_DELETED); | |
8625 | return; | |
8626 | } | |
8627 | ||
8628 | /* Preserve register attributes for variable tracking. */ | |
8629 | destlo = gen_rtx_REG_offset (operands[0], halfmode, dest, 0); | |
8630 | desthi = gen_rtx_REG_offset (operands[0], halfmode, dest + halfregs, | |
8631 | GET_MODE_SIZE (halfmode)); | |
8632 | ||
8633 | /* Special case of reversed high/low parts. */ | |
8634 | if (reg_overlap_mentioned_p (operands[2], destlo) | |
8635 | && reg_overlap_mentioned_p (operands[1], desthi)) | |
8636 | { | |
8637 | emit_insn (gen_xorv16qi3 (operands[1], operands[1], operands[2])); | |
8638 | emit_insn (gen_xorv16qi3 (operands[2], operands[1], operands[2])); | |
8639 | emit_insn (gen_xorv16qi3 (operands[1], operands[1], operands[2])); | |
8640 | } | |
8641 | else if (!reg_overlap_mentioned_p (operands[2], destlo)) | |
8642 | { | |
8643 | /* Try to avoid unnecessary moves if part of the result | |
8644 | is in the right place already. */ | |
8645 | if (src1 != dest) | |
8646 | emit_move_insn (destlo, operands[1]); | |
8647 | if (src2 != dest + halfregs) | |
8648 | emit_move_insn (desthi, operands[2]); | |
8649 | } | |
8650 | else | |
8651 | { | |
8652 | if (src2 != dest + halfregs) | |
8653 | emit_move_insn (desthi, operands[2]); | |
8654 | if (src1 != dest) | |
8655 | emit_move_insn (destlo, operands[1]); | |
8656 | } | |
8657 | } | |
8658 | ||
8659 | /* vec_perm support. */ | |
8660 | ||
8661 | #define MAX_VECT_LEN 16 | |
8662 | ||
8663 | struct expand_vec_perm_d | |
8664 | { | |
8665 | rtx target, op0, op1; | |
8666 | unsigned char perm[MAX_VECT_LEN]; | |
8667 | enum machine_mode vmode; | |
8668 | unsigned char nelt; | |
8669 | bool one_vector_p; | |
8670 | bool testing_p; | |
8671 | }; | |
8672 | ||
8673 | /* Generate a variable permutation. */ | |
8674 | ||
8675 | static void | |
8676 | aarch64_expand_vec_perm_1 (rtx target, rtx op0, rtx op1, rtx sel) | |
8677 | { | |
8678 | enum machine_mode vmode = GET_MODE (target); | |
8679 | bool one_vector_p = rtx_equal_p (op0, op1); | |
8680 | ||
8681 | gcc_checking_assert (vmode == V8QImode || vmode == V16QImode); | |
8682 | gcc_checking_assert (GET_MODE (op0) == vmode); | |
8683 | gcc_checking_assert (GET_MODE (op1) == vmode); | |
8684 | gcc_checking_assert (GET_MODE (sel) == vmode); | |
8685 | gcc_checking_assert (TARGET_SIMD); | |
8686 | ||
8687 | if (one_vector_p) | |
8688 | { | |
8689 | if (vmode == V8QImode) | |
8690 | { | |
8691 | /* Expand the argument to a V16QI mode by duplicating it. */ | |
8692 | rtx pair = gen_reg_rtx (V16QImode); | |
8693 | emit_insn (gen_aarch64_combinev8qi (pair, op0, op0)); | |
8694 | emit_insn (gen_aarch64_tbl1v8qi (target, pair, sel)); | |
8695 | } | |
8696 | else | |
8697 | { | |
8698 | emit_insn (gen_aarch64_tbl1v16qi (target, op0, sel)); | |
8699 | } | |
8700 | } | |
8701 | else | |
8702 | { | |
8703 | rtx pair; | |
8704 | ||
8705 | if (vmode == V8QImode) | |
8706 | { | |
8707 | pair = gen_reg_rtx (V16QImode); | |
8708 | emit_insn (gen_aarch64_combinev8qi (pair, op0, op1)); | |
8709 | emit_insn (gen_aarch64_tbl1v8qi (target, pair, sel)); | |
8710 | } | |
8711 | else | |
8712 | { | |
8713 | pair = gen_reg_rtx (OImode); | |
8714 | emit_insn (gen_aarch64_combinev16qi (pair, op0, op1)); | |
8715 | emit_insn (gen_aarch64_tbl2v16qi (target, pair, sel)); | |
8716 | } | |
8717 | } | |
8718 | } | |
8719 | ||
8720 | void | |
8721 | aarch64_expand_vec_perm (rtx target, rtx op0, rtx op1, rtx sel) | |
8722 | { | |
8723 | enum machine_mode vmode = GET_MODE (target); | |
8724 | unsigned int i, nelt = GET_MODE_NUNITS (vmode); | |
8725 | bool one_vector_p = rtx_equal_p (op0, op1); | |
8726 | rtx rmask[MAX_VECT_LEN], mask; | |
8727 | ||
8728 | gcc_checking_assert (!BYTES_BIG_ENDIAN); | |
8729 | ||
8730 | /* The TBL instruction does not use a modulo index, so we must take care | |
8731 | of that ourselves. */ | |
8732 | mask = GEN_INT (one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
8733 | for (i = 0; i < nelt; ++i) | |
8734 | rmask[i] = mask; | |
8735 | mask = gen_rtx_CONST_VECTOR (vmode, gen_rtvec_v (nelt, rmask)); | |
8736 | sel = expand_simple_binop (vmode, AND, sel, mask, NULL, 0, OPTAB_LIB_WIDEN); | |
8737 | ||
8738 | aarch64_expand_vec_perm_1 (target, op0, op1, sel); | |
8739 | } | |
8740 | ||
8741 | /* Recognize patterns suitable for the TRN instructions. */ | |
8742 | static bool | |
8743 | aarch64_evpc_trn (struct expand_vec_perm_d *d) | |
8744 | { | |
8745 | unsigned int i, odd, mask, nelt = d->nelt; | |
8746 | rtx out, in0, in1, x; | |
8747 | rtx (*gen) (rtx, rtx, rtx); | |
8748 | enum machine_mode vmode = d->vmode; | |
8749 | ||
8750 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
8751 | return false; | |
8752 | ||
8753 | /* Note that these are little-endian tests. | |
8754 | We correct for big-endian later. */ | |
8755 | if (d->perm[0] == 0) | |
8756 | odd = 0; | |
8757 | else if (d->perm[0] == 1) | |
8758 | odd = 1; | |
8759 | else | |
8760 | return false; | |
8761 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
8762 | ||
8763 | for (i = 0; i < nelt; i += 2) | |
8764 | { | |
8765 | if (d->perm[i] != i + odd) | |
8766 | return false; | |
8767 | if (d->perm[i + 1] != ((i + nelt + odd) & mask)) | |
8768 | return false; | |
8769 | } | |
8770 | ||
8771 | /* Success! */ | |
8772 | if (d->testing_p) | |
8773 | return true; | |
8774 | ||
8775 | in0 = d->op0; | |
8776 | in1 = d->op1; | |
8777 | if (BYTES_BIG_ENDIAN) | |
8778 | { | |
8779 | x = in0, in0 = in1, in1 = x; | |
8780 | odd = !odd; | |
8781 | } | |
8782 | out = d->target; | |
8783 | ||
8784 | if (odd) | |
8785 | { | |
8786 | switch (vmode) | |
8787 | { | |
8788 | case V16QImode: gen = gen_aarch64_trn2v16qi; break; | |
8789 | case V8QImode: gen = gen_aarch64_trn2v8qi; break; | |
8790 | case V8HImode: gen = gen_aarch64_trn2v8hi; break; | |
8791 | case V4HImode: gen = gen_aarch64_trn2v4hi; break; | |
8792 | case V4SImode: gen = gen_aarch64_trn2v4si; break; | |
8793 | case V2SImode: gen = gen_aarch64_trn2v2si; break; | |
8794 | case V2DImode: gen = gen_aarch64_trn2v2di; break; | |
8795 | case V4SFmode: gen = gen_aarch64_trn2v4sf; break; | |
8796 | case V2SFmode: gen = gen_aarch64_trn2v2sf; break; | |
8797 | case V2DFmode: gen = gen_aarch64_trn2v2df; break; | |
8798 | default: | |
8799 | return false; | |
8800 | } | |
8801 | } | |
8802 | else | |
8803 | { | |
8804 | switch (vmode) | |
8805 | { | |
8806 | case V16QImode: gen = gen_aarch64_trn1v16qi; break; | |
8807 | case V8QImode: gen = gen_aarch64_trn1v8qi; break; | |
8808 | case V8HImode: gen = gen_aarch64_trn1v8hi; break; | |
8809 | case V4HImode: gen = gen_aarch64_trn1v4hi; break; | |
8810 | case V4SImode: gen = gen_aarch64_trn1v4si; break; | |
8811 | case V2SImode: gen = gen_aarch64_trn1v2si; break; | |
8812 | case V2DImode: gen = gen_aarch64_trn1v2di; break; | |
8813 | case V4SFmode: gen = gen_aarch64_trn1v4sf; break; | |
8814 | case V2SFmode: gen = gen_aarch64_trn1v2sf; break; | |
8815 | case V2DFmode: gen = gen_aarch64_trn1v2df; break; | |
8816 | default: | |
8817 | return false; | |
8818 | } | |
8819 | } | |
8820 | ||
8821 | emit_insn (gen (out, in0, in1)); | |
8822 | return true; | |
8823 | } | |
8824 | ||
8825 | /* Recognize patterns suitable for the UZP instructions. */ | |
8826 | static bool | |
8827 | aarch64_evpc_uzp (struct expand_vec_perm_d *d) | |
8828 | { | |
8829 | unsigned int i, odd, mask, nelt = d->nelt; | |
8830 | rtx out, in0, in1, x; | |
8831 | rtx (*gen) (rtx, rtx, rtx); | |
8832 | enum machine_mode vmode = d->vmode; | |
8833 | ||
8834 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
8835 | return false; | |
8836 | ||
8837 | /* Note that these are little-endian tests. | |
8838 | We correct for big-endian later. */ | |
8839 | if (d->perm[0] == 0) | |
8840 | odd = 0; | |
8841 | else if (d->perm[0] == 1) | |
8842 | odd = 1; | |
8843 | else | |
8844 | return false; | |
8845 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
8846 | ||
8847 | for (i = 0; i < nelt; i++) | |
8848 | { | |
8849 | unsigned elt = (i * 2 + odd) & mask; | |
8850 | if (d->perm[i] != elt) | |
8851 | return false; | |
8852 | } | |
8853 | ||
8854 | /* Success! */ | |
8855 | if (d->testing_p) | |
8856 | return true; | |
8857 | ||
8858 | in0 = d->op0; | |
8859 | in1 = d->op1; | |
8860 | if (BYTES_BIG_ENDIAN) | |
8861 | { | |
8862 | x = in0, in0 = in1, in1 = x; | |
8863 | odd = !odd; | |
8864 | } | |
8865 | out = d->target; | |
8866 | ||
8867 | if (odd) | |
8868 | { | |
8869 | switch (vmode) | |
8870 | { | |
8871 | case V16QImode: gen = gen_aarch64_uzp2v16qi; break; | |
8872 | case V8QImode: gen = gen_aarch64_uzp2v8qi; break; | |
8873 | case V8HImode: gen = gen_aarch64_uzp2v8hi; break; | |
8874 | case V4HImode: gen = gen_aarch64_uzp2v4hi; break; | |
8875 | case V4SImode: gen = gen_aarch64_uzp2v4si; break; | |
8876 | case V2SImode: gen = gen_aarch64_uzp2v2si; break; | |
8877 | case V2DImode: gen = gen_aarch64_uzp2v2di; break; | |
8878 | case V4SFmode: gen = gen_aarch64_uzp2v4sf; break; | |
8879 | case V2SFmode: gen = gen_aarch64_uzp2v2sf; break; | |
8880 | case V2DFmode: gen = gen_aarch64_uzp2v2df; break; | |
8881 | default: | |
8882 | return false; | |
8883 | } | |
8884 | } | |
8885 | else | |
8886 | { | |
8887 | switch (vmode) | |
8888 | { | |
8889 | case V16QImode: gen = gen_aarch64_uzp1v16qi; break; | |
8890 | case V8QImode: gen = gen_aarch64_uzp1v8qi; break; | |
8891 | case V8HImode: gen = gen_aarch64_uzp1v8hi; break; | |
8892 | case V4HImode: gen = gen_aarch64_uzp1v4hi; break; | |
8893 | case V4SImode: gen = gen_aarch64_uzp1v4si; break; | |
8894 | case V2SImode: gen = gen_aarch64_uzp1v2si; break; | |
8895 | case V2DImode: gen = gen_aarch64_uzp1v2di; break; | |
8896 | case V4SFmode: gen = gen_aarch64_uzp1v4sf; break; | |
8897 | case V2SFmode: gen = gen_aarch64_uzp1v2sf; break; | |
8898 | case V2DFmode: gen = gen_aarch64_uzp1v2df; break; | |
8899 | default: | |
8900 | return false; | |
8901 | } | |
8902 | } | |
8903 | ||
8904 | emit_insn (gen (out, in0, in1)); | |
8905 | return true; | |
8906 | } | |
8907 | ||
8908 | /* Recognize patterns suitable for the ZIP instructions. */ | |
8909 | static bool | |
8910 | aarch64_evpc_zip (struct expand_vec_perm_d *d) | |
8911 | { | |
8912 | unsigned int i, high, mask, nelt = d->nelt; | |
8913 | rtx out, in0, in1, x; | |
8914 | rtx (*gen) (rtx, rtx, rtx); | |
8915 | enum machine_mode vmode = d->vmode; | |
8916 | ||
8917 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
8918 | return false; | |
8919 | ||
8920 | /* Note that these are little-endian tests. | |
8921 | We correct for big-endian later. */ | |
8922 | high = nelt / 2; | |
8923 | if (d->perm[0] == high) | |
8924 | /* Do Nothing. */ | |
8925 | ; | |
8926 | else if (d->perm[0] == 0) | |
8927 | high = 0; | |
8928 | else | |
8929 | return false; | |
8930 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
8931 | ||
8932 | for (i = 0; i < nelt / 2; i++) | |
8933 | { | |
8934 | unsigned elt = (i + high) & mask; | |
8935 | if (d->perm[i * 2] != elt) | |
8936 | return false; | |
8937 | elt = (elt + nelt) & mask; | |
8938 | if (d->perm[i * 2 + 1] != elt) | |
8939 | return false; | |
8940 | } | |
8941 | ||
8942 | /* Success! */ | |
8943 | if (d->testing_p) | |
8944 | return true; | |
8945 | ||
8946 | in0 = d->op0; | |
8947 | in1 = d->op1; | |
8948 | if (BYTES_BIG_ENDIAN) | |
8949 | { | |
8950 | x = in0, in0 = in1, in1 = x; | |
8951 | high = !high; | |
8952 | } | |
8953 | out = d->target; | |
8954 | ||
8955 | if (high) | |
8956 | { | |
8957 | switch (vmode) | |
8958 | { | |
8959 | case V16QImode: gen = gen_aarch64_zip2v16qi; break; | |
8960 | case V8QImode: gen = gen_aarch64_zip2v8qi; break; | |
8961 | case V8HImode: gen = gen_aarch64_zip2v8hi; break; | |
8962 | case V4HImode: gen = gen_aarch64_zip2v4hi; break; | |
8963 | case V4SImode: gen = gen_aarch64_zip2v4si; break; | |
8964 | case V2SImode: gen = gen_aarch64_zip2v2si; break; | |
8965 | case V2DImode: gen = gen_aarch64_zip2v2di; break; | |
8966 | case V4SFmode: gen = gen_aarch64_zip2v4sf; break; | |
8967 | case V2SFmode: gen = gen_aarch64_zip2v2sf; break; | |
8968 | case V2DFmode: gen = gen_aarch64_zip2v2df; break; | |
8969 | default: | |
8970 | return false; | |
8971 | } | |
8972 | } | |
8973 | else | |
8974 | { | |
8975 | switch (vmode) | |
8976 | { | |
8977 | case V16QImode: gen = gen_aarch64_zip1v16qi; break; | |
8978 | case V8QImode: gen = gen_aarch64_zip1v8qi; break; | |
8979 | case V8HImode: gen = gen_aarch64_zip1v8hi; break; | |
8980 | case V4HImode: gen = gen_aarch64_zip1v4hi; break; | |
8981 | case V4SImode: gen = gen_aarch64_zip1v4si; break; | |
8982 | case V2SImode: gen = gen_aarch64_zip1v2si; break; | |
8983 | case V2DImode: gen = gen_aarch64_zip1v2di; break; | |
8984 | case V4SFmode: gen = gen_aarch64_zip1v4sf; break; | |
8985 | case V2SFmode: gen = gen_aarch64_zip1v2sf; break; | |
8986 | case V2DFmode: gen = gen_aarch64_zip1v2df; break; | |
8987 | default: | |
8988 | return false; | |
8989 | } | |
8990 | } | |
8991 | ||
8992 | emit_insn (gen (out, in0, in1)); | |
8993 | return true; | |
8994 | } | |
8995 | ||
8996 | /* Recognize patterns for the EXT insn. */ | |
8997 | ||
8998 | static bool | |
8999 | aarch64_evpc_ext (struct expand_vec_perm_d *d) | |
9000 | { | |
9001 | unsigned int i, nelt = d->nelt; | |
9002 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
9003 | rtx offset; | |
9004 | ||
9005 | unsigned int location = d->perm[0]; /* Always < nelt. */ | |
9006 | ||
9007 | /* Check if the extracted indices are increasing by one. */ | |
9008 | for (i = 1; i < nelt; i++) | |
9009 | { | |
9010 | unsigned int required = location + i; | |
9011 | if (d->one_vector_p) | |
9012 | { | |
9013 | /* We'll pass the same vector in twice, so allow indices to wrap. */ | |
9014 | required &= (nelt - 1); | |
9015 | } | |
9016 | if (d->perm[i] != required) | |
9017 | return false; | |
9018 | } | |
9019 | ||
9020 | /* The mid-end handles masks that just return one of the input vectors. */ | |
9021 | gcc_assert (location != 0); | |
9022 | ||
9023 | switch (d->vmode) | |
9024 | { | |
9025 | case V16QImode: gen = gen_aarch64_extv16qi; break; | |
9026 | case V8QImode: gen = gen_aarch64_extv8qi; break; | |
9027 | case V4HImode: gen = gen_aarch64_extv4hi; break; | |
9028 | case V8HImode: gen = gen_aarch64_extv8hi; break; | |
9029 | case V2SImode: gen = gen_aarch64_extv2si; break; | |
9030 | case V4SImode: gen = gen_aarch64_extv4si; break; | |
9031 | case V2SFmode: gen = gen_aarch64_extv2sf; break; | |
9032 | case V4SFmode: gen = gen_aarch64_extv4sf; break; | |
9033 | case V2DImode: gen = gen_aarch64_extv2di; break; | |
9034 | case V2DFmode: gen = gen_aarch64_extv2df; break; | |
9035 | default: | |
9036 | return false; | |
9037 | } | |
9038 | ||
9039 | /* Success! */ | |
9040 | if (d->testing_p) | |
9041 | return true; | |
9042 | ||
9043 | if (BYTES_BIG_ENDIAN) | |
9044 | { | |
9045 | /* After setup, we want the high elements of the first vector (stored | |
9046 | at the LSB end of the register), and the low elements of the second | |
9047 | vector (stored at the MSB end of the register). So swap. */ | |
9048 | rtx temp = d->op0; | |
9049 | d->op0 = d->op1; | |
9050 | d->op1 = temp; | |
9051 | /* location != 0 (above), so safe to assume (nelt - location) < nelt. */ | |
9052 | location = nelt - location; | |
9053 | } | |
9054 | ||
9055 | offset = GEN_INT (location); | |
9056 | emit_insn (gen (d->target, d->op0, d->op1, offset)); | |
9057 | return true; | |
9058 | } | |
9059 | ||
9060 | static bool | |
9061 | aarch64_evpc_dup (struct expand_vec_perm_d *d) | |
9062 | { | |
9063 | rtx (*gen) (rtx, rtx, rtx); | |
9064 | rtx out = d->target; | |
9065 | rtx in0; | |
9066 | enum machine_mode vmode = d->vmode; | |
9067 | unsigned int i, elt, nelt = d->nelt; | |
9068 | rtx lane; | |
9069 | ||
9070 | /* TODO: This may not be big-endian safe. */ | |
9071 | if (BYTES_BIG_ENDIAN) | |
9072 | return false; | |
9073 | ||
9074 | elt = d->perm[0]; | |
9075 | for (i = 1; i < nelt; i++) | |
9076 | { | |
9077 | if (elt != d->perm[i]) | |
9078 | return false; | |
9079 | } | |
9080 | ||
9081 | /* The generic preparation in aarch64_expand_vec_perm_const_1 | |
9082 | swaps the operand order and the permute indices if it finds | |
9083 | d->perm[0] to be in the second operand. Thus, we can always | |
9084 | use d->op0 and need not do any extra arithmetic to get the | |
9085 | correct lane number. */ | |
9086 | in0 = d->op0; | |
9087 | lane = GEN_INT (elt); | |
9088 | ||
9089 | switch (vmode) | |
9090 | { | |
9091 | case V16QImode: gen = gen_aarch64_dup_lanev16qi; break; | |
9092 | case V8QImode: gen = gen_aarch64_dup_lanev8qi; break; | |
9093 | case V8HImode: gen = gen_aarch64_dup_lanev8hi; break; | |
9094 | case V4HImode: gen = gen_aarch64_dup_lanev4hi; break; | |
9095 | case V4SImode: gen = gen_aarch64_dup_lanev4si; break; | |
9096 | case V2SImode: gen = gen_aarch64_dup_lanev2si; break; | |
9097 | case V2DImode: gen = gen_aarch64_dup_lanev2di; break; | |
9098 | case V4SFmode: gen = gen_aarch64_dup_lanev4sf; break; | |
9099 | case V2SFmode: gen = gen_aarch64_dup_lanev2sf; break; | |
9100 | case V2DFmode: gen = gen_aarch64_dup_lanev2df; break; | |
9101 | default: | |
9102 | return false; | |
9103 | } | |
9104 | ||
9105 | emit_insn (gen (out, in0, lane)); | |
9106 | return true; | |
9107 | } | |
9108 | ||
9109 | static bool | |
9110 | aarch64_evpc_tbl (struct expand_vec_perm_d *d) | |
9111 | { | |
9112 | rtx rperm[MAX_VECT_LEN], sel; | |
9113 | enum machine_mode vmode = d->vmode; | |
9114 | unsigned int i, nelt = d->nelt; | |
9115 | ||
9116 | if (d->testing_p) | |
9117 | return true; | |
9118 | ||
9119 | /* Generic code will try constant permutation twice. Once with the | |
9120 | original mode and again with the elements lowered to QImode. | |
9121 | So wait and don't do the selector expansion ourselves. */ | |
9122 | if (vmode != V8QImode && vmode != V16QImode) | |
9123 | return false; | |
9124 | ||
9125 | for (i = 0; i < nelt; ++i) | |
9126 | { | |
9127 | int nunits = GET_MODE_NUNITS (vmode); | |
9128 | ||
9129 | /* If big-endian and two vectors we end up with a weird mixed-endian | |
9130 | mode on NEON. Reverse the index within each word but not the word | |
9131 | itself. */ | |
9132 | rperm[i] = GEN_INT (BYTES_BIG_ENDIAN ? d->perm[i] ^ (nunits - 1) | |
9133 | : d->perm[i]); | |
9134 | } | |
9135 | sel = gen_rtx_CONST_VECTOR (vmode, gen_rtvec_v (nelt, rperm)); | |
9136 | sel = force_reg (vmode, sel); | |
9137 | ||
9138 | aarch64_expand_vec_perm_1 (d->target, d->op0, d->op1, sel); | |
9139 | return true; | |
9140 | } | |
9141 | ||
9142 | static bool | |
9143 | aarch64_expand_vec_perm_const_1 (struct expand_vec_perm_d *d) | |
9144 | { | |
9145 | /* The pattern matching functions above are written to look for a small | |
9146 | number to begin the sequence (0, 1, N/2). If we begin with an index | |
9147 | from the second operand, we can swap the operands. */ | |
9148 | if (d->perm[0] >= d->nelt) | |
9149 | { | |
9150 | unsigned i, nelt = d->nelt; | |
9151 | rtx x; | |
9152 | ||
9153 | gcc_assert (nelt == (nelt & -nelt)); | |
9154 | for (i = 0; i < nelt; ++i) | |
9155 | d->perm[i] ^= nelt; /* Keep the same index, but in the other vector. */ | |
9156 | ||
9157 | x = d->op0; | |
9158 | d->op0 = d->op1; | |
9159 | d->op1 = x; | |
9160 | } | |
9161 | ||
9162 | if (TARGET_SIMD) | |
9163 | { | |
9164 | if (aarch64_evpc_ext (d)) | |
9165 | return true; | |
9166 | else if (aarch64_evpc_zip (d)) | |
9167 | return true; | |
9168 | else if (aarch64_evpc_uzp (d)) | |
9169 | return true; | |
9170 | else if (aarch64_evpc_trn (d)) | |
9171 | return true; | |
9172 | else if (aarch64_evpc_dup (d)) | |
9173 | return true; | |
9174 | return aarch64_evpc_tbl (d); | |
9175 | } | |
9176 | return false; | |
9177 | } | |
9178 | ||
9179 | /* Expand a vec_perm_const pattern. */ | |
9180 | ||
9181 | bool | |
9182 | aarch64_expand_vec_perm_const (rtx target, rtx op0, rtx op1, rtx sel) | |
9183 | { | |
9184 | struct expand_vec_perm_d d; | |
9185 | int i, nelt, which; | |
9186 | ||
9187 | d.target = target; | |
9188 | d.op0 = op0; | |
9189 | d.op1 = op1; | |
9190 | ||
9191 | d.vmode = GET_MODE (target); | |
9192 | gcc_assert (VECTOR_MODE_P (d.vmode)); | |
9193 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
9194 | d.testing_p = false; | |
9195 | ||
9196 | for (i = which = 0; i < nelt; ++i) | |
9197 | { | |
9198 | rtx e = XVECEXP (sel, 0, i); | |
9199 | int ei = INTVAL (e) & (2 * nelt - 1); | |
9200 | which |= (ei < nelt ? 1 : 2); | |
9201 | d.perm[i] = ei; | |
9202 | } | |
9203 | ||
9204 | switch (which) | |
9205 | { | |
9206 | default: | |
9207 | gcc_unreachable (); | |
9208 | ||
9209 | case 3: | |
9210 | d.one_vector_p = false; | |
9211 | if (!rtx_equal_p (op0, op1)) | |
9212 | break; | |
9213 | ||
9214 | /* The elements of PERM do not suggest that only the first operand | |
9215 | is used, but both operands are identical. Allow easier matching | |
9216 | of the permutation by folding the permutation into the single | |
9217 | input vector. */ | |
9218 | /* Fall Through. */ | |
9219 | case 2: | |
9220 | for (i = 0; i < nelt; ++i) | |
9221 | d.perm[i] &= nelt - 1; | |
9222 | d.op0 = op1; | |
9223 | d.one_vector_p = true; | |
9224 | break; | |
9225 | ||
9226 | case 1: | |
9227 | d.op1 = op0; | |
9228 | d.one_vector_p = true; | |
9229 | break; | |
9230 | } | |
9231 | ||
9232 | return aarch64_expand_vec_perm_const_1 (&d); | |
9233 | } | |
9234 | ||
9235 | static bool | |
9236 | aarch64_vectorize_vec_perm_const_ok (enum machine_mode vmode, | |
9237 | const unsigned char *sel) | |
9238 | { | |
9239 | struct expand_vec_perm_d d; | |
9240 | unsigned int i, nelt, which; | |
9241 | bool ret; | |
9242 | ||
9243 | d.vmode = vmode; | |
9244 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
9245 | d.testing_p = true; | |
9246 | memcpy (d.perm, sel, nelt); | |
9247 | ||
9248 | /* Calculate whether all elements are in one vector. */ | |
9249 | for (i = which = 0; i < nelt; ++i) | |
9250 | { | |
9251 | unsigned char e = d.perm[i]; | |
9252 | gcc_assert (e < 2 * nelt); | |
9253 | which |= (e < nelt ? 1 : 2); | |
9254 | } | |
9255 | ||
9256 | /* If all elements are from the second vector, reindex as if from the | |
9257 | first vector. */ | |
9258 | if (which == 2) | |
9259 | for (i = 0; i < nelt; ++i) | |
9260 | d.perm[i] -= nelt; | |
9261 | ||
9262 | /* Check whether the mask can be applied to a single vector. */ | |
9263 | d.one_vector_p = (which != 3); | |
9264 | ||
9265 | d.target = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 1); | |
9266 | d.op1 = d.op0 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 2); | |
9267 | if (!d.one_vector_p) | |
9268 | d.op1 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 3); | |
9269 | ||
9270 | start_sequence (); | |
9271 | ret = aarch64_expand_vec_perm_const_1 (&d); | |
9272 | end_sequence (); | |
9273 | ||
9274 | return ret; | |
9275 | } | |
9276 | ||
9277 | /* Implement target hook CANNOT_CHANGE_MODE_CLASS. */ | |
9278 | bool | |
9279 | aarch64_cannot_change_mode_class (enum machine_mode from, | |
9280 | enum machine_mode to, | |
9281 | enum reg_class rclass) | |
9282 | { | |
9283 | /* Full-reg subregs are allowed on general regs or any class if they are | |
9284 | the same size. */ | |
9285 | if (GET_MODE_SIZE (from) == GET_MODE_SIZE (to) | |
9286 | || !reg_classes_intersect_p (FP_REGS, rclass)) | |
9287 | return false; | |
9288 | ||
9289 | /* Limited combinations of subregs are safe on FPREGs. Particularly, | |
9290 | 1. Vector Mode to Scalar mode where 1 unit of the vector is accessed. | |
9291 | 2. Scalar to Scalar for integer modes or same size float modes. | |
9292 | 3. Vector to Vector modes. | |
9293 | 4. On little-endian only, Vector-Structure to Vector modes. */ | |
9294 | if (GET_MODE_SIZE (from) > GET_MODE_SIZE (to)) | |
9295 | { | |
9296 | if (aarch64_vector_mode_supported_p (from) | |
9297 | && GET_MODE_SIZE (GET_MODE_INNER (from)) == GET_MODE_SIZE (to)) | |
9298 | return false; | |
9299 | ||
9300 | if (GET_MODE_NUNITS (from) == 1 | |
9301 | && GET_MODE_NUNITS (to) == 1 | |
9302 | && (GET_MODE_CLASS (from) == MODE_INT | |
9303 | || from == to)) | |
9304 | return false; | |
9305 | ||
9306 | if (aarch64_vector_mode_supported_p (from) | |
9307 | && aarch64_vector_mode_supported_p (to)) | |
9308 | return false; | |
9309 | ||
9310 | /* Within an vector structure straddling multiple vector registers | |
9311 | we are in a mixed-endian representation. As such, we can't | |
9312 | easily change modes for BYTES_BIG_ENDIAN. Otherwise, we can | |
9313 | switch between vectors and vector structures cheaply. */ | |
9314 | if (!BYTES_BIG_ENDIAN) | |
9315 | if ((aarch64_vector_mode_supported_p (from) | |
9316 | && aarch64_vect_struct_mode_p (to)) | |
9317 | || (aarch64_vector_mode_supported_p (to) | |
9318 | && aarch64_vect_struct_mode_p (from))) | |
9319 | return false; | |
9320 | } | |
9321 | ||
9322 | return true; | |
9323 | } | |
9324 | ||
9325 | /* Implement MODES_TIEABLE_P. */ | |
9326 | ||
9327 | bool | |
9328 | aarch64_modes_tieable_p (enum machine_mode mode1, enum machine_mode mode2) | |
9329 | { | |
9330 | if (GET_MODE_CLASS (mode1) == GET_MODE_CLASS (mode2)) | |
9331 | return true; | |
9332 | ||
9333 | /* We specifically want to allow elements of "structure" modes to | |
9334 | be tieable to the structure. This more general condition allows | |
9335 | other rarer situations too. */ | |
9336 | if (TARGET_SIMD | |
9337 | && aarch64_vector_mode_p (mode1) | |
9338 | && aarch64_vector_mode_p (mode2)) | |
9339 | return true; | |
9340 | ||
9341 | return false; | |
9342 | } | |
9343 | ||
9344 | #undef TARGET_ADDRESS_COST | |
9345 | #define TARGET_ADDRESS_COST aarch64_address_cost | |
9346 | ||
9347 | /* This hook will determines whether unnamed bitfields affect the alignment | |
9348 | of the containing structure. The hook returns true if the structure | |
9349 | should inherit the alignment requirements of an unnamed bitfield's | |
9350 | type. */ | |
9351 | #undef TARGET_ALIGN_ANON_BITFIELD | |
9352 | #define TARGET_ALIGN_ANON_BITFIELD hook_bool_void_true | |
9353 | ||
9354 | #undef TARGET_ASM_ALIGNED_DI_OP | |
9355 | #define TARGET_ASM_ALIGNED_DI_OP "\t.xword\t" | |
9356 | ||
9357 | #undef TARGET_ASM_ALIGNED_HI_OP | |
9358 | #define TARGET_ASM_ALIGNED_HI_OP "\t.hword\t" | |
9359 | ||
9360 | #undef TARGET_ASM_ALIGNED_SI_OP | |
9361 | #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t" | |
9362 | ||
9363 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
9364 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK \ | |
9365 | hook_bool_const_tree_hwi_hwi_const_tree_true | |
9366 | ||
9367 | #undef TARGET_ASM_FILE_START | |
9368 | #define TARGET_ASM_FILE_START aarch64_start_file | |
9369 | ||
9370 | #undef TARGET_ASM_OUTPUT_MI_THUNK | |
9371 | #define TARGET_ASM_OUTPUT_MI_THUNK aarch64_output_mi_thunk | |
9372 | ||
9373 | #undef TARGET_ASM_SELECT_RTX_SECTION | |
9374 | #define TARGET_ASM_SELECT_RTX_SECTION aarch64_select_rtx_section | |
9375 | ||
9376 | #undef TARGET_ASM_TRAMPOLINE_TEMPLATE | |
9377 | #define TARGET_ASM_TRAMPOLINE_TEMPLATE aarch64_asm_trampoline_template | |
9378 | ||
9379 | #undef TARGET_BUILD_BUILTIN_VA_LIST | |
9380 | #define TARGET_BUILD_BUILTIN_VA_LIST aarch64_build_builtin_va_list | |
9381 | ||
9382 | #undef TARGET_CALLEE_COPIES | |
9383 | #define TARGET_CALLEE_COPIES hook_bool_CUMULATIVE_ARGS_mode_tree_bool_false | |
9384 | ||
9385 | #undef TARGET_CAN_ELIMINATE | |
9386 | #define TARGET_CAN_ELIMINATE aarch64_can_eliminate | |
9387 | ||
9388 | #undef TARGET_CANNOT_FORCE_CONST_MEM | |
9389 | #define TARGET_CANNOT_FORCE_CONST_MEM aarch64_cannot_force_const_mem | |
9390 | ||
9391 | #undef TARGET_CONDITIONAL_REGISTER_USAGE | |
9392 | #define TARGET_CONDITIONAL_REGISTER_USAGE aarch64_conditional_register_usage | |
9393 | ||
9394 | /* Only the least significant bit is used for initialization guard | |
9395 | variables. */ | |
9396 | #undef TARGET_CXX_GUARD_MASK_BIT | |
9397 | #define TARGET_CXX_GUARD_MASK_BIT hook_bool_void_true | |
9398 | ||
9399 | #undef TARGET_C_MODE_FOR_SUFFIX | |
9400 | #define TARGET_C_MODE_FOR_SUFFIX aarch64_c_mode_for_suffix | |
9401 | ||
9402 | #ifdef TARGET_BIG_ENDIAN_DEFAULT | |
9403 | #undef TARGET_DEFAULT_TARGET_FLAGS | |
9404 | #define TARGET_DEFAULT_TARGET_FLAGS (MASK_BIG_END) | |
9405 | #endif | |
9406 | ||
9407 | #undef TARGET_CLASS_MAX_NREGS | |
9408 | #define TARGET_CLASS_MAX_NREGS aarch64_class_max_nregs | |
9409 | ||
9410 | #undef TARGET_BUILTIN_DECL | |
9411 | #define TARGET_BUILTIN_DECL aarch64_builtin_decl | |
9412 | ||
9413 | #undef TARGET_EXPAND_BUILTIN | |
9414 | #define TARGET_EXPAND_BUILTIN aarch64_expand_builtin | |
9415 | ||
9416 | #undef TARGET_EXPAND_BUILTIN_VA_START | |
9417 | #define TARGET_EXPAND_BUILTIN_VA_START aarch64_expand_builtin_va_start | |
9418 | ||
9419 | #undef TARGET_FOLD_BUILTIN | |
9420 | #define TARGET_FOLD_BUILTIN aarch64_fold_builtin | |
9421 | ||
9422 | #undef TARGET_FUNCTION_ARG | |
9423 | #define TARGET_FUNCTION_ARG aarch64_function_arg | |
9424 | ||
9425 | #undef TARGET_FUNCTION_ARG_ADVANCE | |
9426 | #define TARGET_FUNCTION_ARG_ADVANCE aarch64_function_arg_advance | |
9427 | ||
9428 | #undef TARGET_FUNCTION_ARG_BOUNDARY | |
9429 | #define TARGET_FUNCTION_ARG_BOUNDARY aarch64_function_arg_boundary | |
9430 | ||
9431 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL | |
9432 | #define TARGET_FUNCTION_OK_FOR_SIBCALL aarch64_function_ok_for_sibcall | |
9433 | ||
9434 | #undef TARGET_FUNCTION_VALUE | |
9435 | #define TARGET_FUNCTION_VALUE aarch64_function_value | |
9436 | ||
9437 | #undef TARGET_FUNCTION_VALUE_REGNO_P | |
9438 | #define TARGET_FUNCTION_VALUE_REGNO_P aarch64_function_value_regno_p | |
9439 | ||
9440 | #undef TARGET_FRAME_POINTER_REQUIRED | |
9441 | #define TARGET_FRAME_POINTER_REQUIRED aarch64_frame_pointer_required | |
9442 | ||
9443 | #undef TARGET_GIMPLE_FOLD_BUILTIN | |
9444 | #define TARGET_GIMPLE_FOLD_BUILTIN aarch64_gimple_fold_builtin | |
9445 | ||
9446 | #undef TARGET_GIMPLIFY_VA_ARG_EXPR | |
9447 | #define TARGET_GIMPLIFY_VA_ARG_EXPR aarch64_gimplify_va_arg_expr | |
9448 | ||
9449 | #undef TARGET_INIT_BUILTINS | |
9450 | #define TARGET_INIT_BUILTINS aarch64_init_builtins | |
9451 | ||
9452 | #undef TARGET_LEGITIMATE_ADDRESS_P | |
9453 | #define TARGET_LEGITIMATE_ADDRESS_P aarch64_legitimate_address_hook_p | |
9454 | ||
9455 | #undef TARGET_LEGITIMATE_CONSTANT_P | |
9456 | #define TARGET_LEGITIMATE_CONSTANT_P aarch64_legitimate_constant_p | |
9457 | ||
9458 | #undef TARGET_LIBGCC_CMP_RETURN_MODE | |
9459 | #define TARGET_LIBGCC_CMP_RETURN_MODE aarch64_libgcc_cmp_return_mode | |
9460 | ||
9461 | #undef TARGET_LRA_P | |
9462 | #define TARGET_LRA_P aarch64_lra_p | |
9463 | ||
9464 | #undef TARGET_MANGLE_TYPE | |
9465 | #define TARGET_MANGLE_TYPE aarch64_mangle_type | |
9466 | ||
9467 | #undef TARGET_MEMORY_MOVE_COST | |
9468 | #define TARGET_MEMORY_MOVE_COST aarch64_memory_move_cost | |
9469 | ||
9470 | #undef TARGET_MUST_PASS_IN_STACK | |
9471 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
9472 | ||
9473 | /* This target hook should return true if accesses to volatile bitfields | |
9474 | should use the narrowest mode possible. It should return false if these | |
9475 | accesses should use the bitfield container type. */ | |
9476 | #undef TARGET_NARROW_VOLATILE_BITFIELD | |
9477 | #define TARGET_NARROW_VOLATILE_BITFIELD hook_bool_void_false | |
9478 | ||
9479 | #undef TARGET_OPTION_OVERRIDE | |
9480 | #define TARGET_OPTION_OVERRIDE aarch64_override_options | |
9481 | ||
9482 | #undef TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE | |
9483 | #define TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE \ | |
9484 | aarch64_override_options_after_change | |
9485 | ||
9486 | #undef TARGET_PASS_BY_REFERENCE | |
9487 | #define TARGET_PASS_BY_REFERENCE aarch64_pass_by_reference | |
9488 | ||
9489 | #undef TARGET_PREFERRED_RELOAD_CLASS | |
9490 | #define TARGET_PREFERRED_RELOAD_CLASS aarch64_preferred_reload_class | |
9491 | ||
9492 | #undef TARGET_SECONDARY_RELOAD | |
9493 | #define TARGET_SECONDARY_RELOAD aarch64_secondary_reload | |
9494 | ||
9495 | #undef TARGET_SHIFT_TRUNCATION_MASK | |
9496 | #define TARGET_SHIFT_TRUNCATION_MASK aarch64_shift_truncation_mask | |
9497 | ||
9498 | #undef TARGET_SETUP_INCOMING_VARARGS | |
9499 | #define TARGET_SETUP_INCOMING_VARARGS aarch64_setup_incoming_varargs | |
9500 | ||
9501 | #undef TARGET_STRUCT_VALUE_RTX | |
9502 | #define TARGET_STRUCT_VALUE_RTX aarch64_struct_value_rtx | |
9503 | ||
9504 | #undef TARGET_REGISTER_MOVE_COST | |
9505 | #define TARGET_REGISTER_MOVE_COST aarch64_register_move_cost | |
9506 | ||
9507 | #undef TARGET_RETURN_IN_MEMORY | |
9508 | #define TARGET_RETURN_IN_MEMORY aarch64_return_in_memory | |
9509 | ||
9510 | #undef TARGET_RETURN_IN_MSB | |
9511 | #define TARGET_RETURN_IN_MSB aarch64_return_in_msb | |
9512 | ||
9513 | #undef TARGET_RTX_COSTS | |
9514 | #define TARGET_RTX_COSTS aarch64_rtx_costs_wrapper | |
9515 | ||
9516 | #undef TARGET_SCHED_ISSUE_RATE | |
9517 | #define TARGET_SCHED_ISSUE_RATE aarch64_sched_issue_rate | |
9518 | ||
9519 | #undef TARGET_TRAMPOLINE_INIT | |
9520 | #define TARGET_TRAMPOLINE_INIT aarch64_trampoline_init | |
9521 | ||
9522 | #undef TARGET_USE_BLOCKS_FOR_CONSTANT_P | |
9523 | #define TARGET_USE_BLOCKS_FOR_CONSTANT_P aarch64_use_blocks_for_constant_p | |
9524 | ||
9525 | #undef TARGET_VECTOR_MODE_SUPPORTED_P | |
9526 | #define TARGET_VECTOR_MODE_SUPPORTED_P aarch64_vector_mode_supported_p | |
9527 | ||
9528 | #undef TARGET_ARRAY_MODE_SUPPORTED_P | |
9529 | #define TARGET_ARRAY_MODE_SUPPORTED_P aarch64_array_mode_supported_p | |
9530 | ||
9531 | #undef TARGET_VECTORIZE_ADD_STMT_COST | |
9532 | #define TARGET_VECTORIZE_ADD_STMT_COST aarch64_add_stmt_cost | |
9533 | ||
9534 | #undef TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST | |
9535 | #define TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST \ | |
9536 | aarch64_builtin_vectorization_cost | |
9537 | ||
9538 | #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE | |
9539 | #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE aarch64_preferred_simd_mode | |
9540 | ||
9541 | #undef TARGET_VECTORIZE_BUILTINS | |
9542 | #define TARGET_VECTORIZE_BUILTINS | |
9543 | ||
9544 | #undef TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION | |
9545 | #define TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION \ | |
9546 | aarch64_builtin_vectorized_function | |
9547 | ||
9548 | #undef TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES | |
9549 | #define TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES \ | |
9550 | aarch64_autovectorize_vector_sizes | |
9551 | ||
9552 | #undef TARGET_ATOMIC_ASSIGN_EXPAND_FENV | |
9553 | #define TARGET_ATOMIC_ASSIGN_EXPAND_FENV \ | |
9554 | aarch64_atomic_assign_expand_fenv | |
9555 | ||
9556 | /* Section anchor support. */ | |
9557 | ||
9558 | #undef TARGET_MIN_ANCHOR_OFFSET | |
9559 | #define TARGET_MIN_ANCHOR_OFFSET -256 | |
9560 | ||
9561 | /* Limit the maximum anchor offset to 4k-1, since that's the limit for a | |
9562 | byte offset; we can do much more for larger data types, but have no way | |
9563 | to determine the size of the access. We assume accesses are aligned. */ | |
9564 | #undef TARGET_MAX_ANCHOR_OFFSET | |
9565 | #define TARGET_MAX_ANCHOR_OFFSET 4095 | |
9566 | ||
9567 | #undef TARGET_VECTOR_ALIGNMENT | |
9568 | #define TARGET_VECTOR_ALIGNMENT aarch64_simd_vector_alignment | |
9569 | ||
9570 | #undef TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE | |
9571 | #define TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE \ | |
9572 | aarch64_simd_vector_alignment_reachable | |
9573 | ||
9574 | /* vec_perm support. */ | |
9575 | ||
9576 | #undef TARGET_VECTORIZE_VEC_PERM_CONST_OK | |
9577 | #define TARGET_VECTORIZE_VEC_PERM_CONST_OK \ | |
9578 | aarch64_vectorize_vec_perm_const_ok | |
9579 | ||
9580 | ||
9581 | #undef TARGET_FIXED_CONDITION_CODE_REGS | |
9582 | #define TARGET_FIXED_CONDITION_CODE_REGS aarch64_fixed_condition_code_regs | |
9583 | ||
9584 | #undef TARGET_FLAGS_REGNUM | |
9585 | #define TARGET_FLAGS_REGNUM CC_REGNUM | |
9586 | ||
9587 | struct gcc_target targetm = TARGET_INITIALIZER; | |
9588 | ||
9589 | #include "gt-aarch64.h" |