]>
Commit | Line | Data |
---|---|---|
bdb7bf8a | 1 | /* Machine description for AArch64 architecture. |
818ab71a | 2 | Copyright (C) 2009-2016 Free Software Foundation, Inc. |
43e9d192 IB |
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" | |
01736018 | 22 | #define INCLUDE_STRING |
43e9d192 IB |
23 | #include "system.h" |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
e11c4407 AM |
26 | #include "target.h" |
27 | #include "rtl.h" | |
c7131fb2 AM |
28 | #include "tree.h" |
29 | #include "gimple.h" | |
e11c4407 AM |
30 | #include "cfghooks.h" |
31 | #include "cfgloop.h" | |
c7131fb2 | 32 | #include "df.h" |
e11c4407 AM |
33 | #include "tm_p.h" |
34 | #include "stringpool.h" | |
35 | #include "optabs.h" | |
36 | #include "regs.h" | |
37 | #include "emit-rtl.h" | |
38 | #include "recog.h" | |
39 | #include "diagnostic.h" | |
43e9d192 | 40 | #include "insn-attr.h" |
40e23961 | 41 | #include "alias.h" |
40e23961 | 42 | #include "fold-const.h" |
d8a2d370 DN |
43 | #include "stor-layout.h" |
44 | #include "calls.h" | |
45 | #include "varasm.h" | |
43e9d192 | 46 | #include "output.h" |
36566b39 | 47 | #include "flags.h" |
36566b39 | 48 | #include "explow.h" |
43e9d192 IB |
49 | #include "expr.h" |
50 | #include "reload.h" | |
43e9d192 | 51 | #include "langhooks.h" |
5a2c8331 | 52 | #include "opts.h" |
2d6bc7fa | 53 | #include "params.h" |
45b0be94 | 54 | #include "gimplify.h" |
43e9d192 | 55 | #include "dwarf2.h" |
61d371eb | 56 | #include "gimple-iterator.h" |
8990e73a | 57 | #include "tree-vectorizer.h" |
d1bcc29f | 58 | #include "aarch64-cost-tables.h" |
0ee859b5 | 59 | #include "dumpfile.h" |
9b2b7279 | 60 | #include "builtins.h" |
8baff86e | 61 | #include "rtl-iter.h" |
9bbe08fe | 62 | #include "tm-constrs.h" |
d03f7e44 | 63 | #include "sched-int.h" |
fde9b31b | 64 | #include "cortex-a57-fma-steering.h" |
d78006d9 | 65 | #include "target-globals.h" |
a3eb8a52 | 66 | #include "common/common-target.h" |
43e9d192 | 67 | |
994c5d85 | 68 | /* This file should be included last. */ |
d58627a0 RS |
69 | #include "target-def.h" |
70 | ||
28514dda YZ |
71 | /* Defined for convenience. */ |
72 | #define POINTER_BYTES (POINTER_SIZE / BITS_PER_UNIT) | |
73 | ||
43e9d192 IB |
74 | /* Classifies an address. |
75 | ||
76 | ADDRESS_REG_IMM | |
77 | A simple base register plus immediate offset. | |
78 | ||
79 | ADDRESS_REG_WB | |
80 | A base register indexed by immediate offset with writeback. | |
81 | ||
82 | ADDRESS_REG_REG | |
83 | A base register indexed by (optionally scaled) register. | |
84 | ||
85 | ADDRESS_REG_UXTW | |
86 | A base register indexed by (optionally scaled) zero-extended register. | |
87 | ||
88 | ADDRESS_REG_SXTW | |
89 | A base register indexed by (optionally scaled) sign-extended register. | |
90 | ||
91 | ADDRESS_LO_SUM | |
92 | A LO_SUM rtx with a base register and "LO12" symbol relocation. | |
93 | ||
94 | ADDRESS_SYMBOLIC: | |
95 | A constant symbolic address, in pc-relative literal pool. */ | |
96 | ||
97 | enum aarch64_address_type { | |
98 | ADDRESS_REG_IMM, | |
99 | ADDRESS_REG_WB, | |
100 | ADDRESS_REG_REG, | |
101 | ADDRESS_REG_UXTW, | |
102 | ADDRESS_REG_SXTW, | |
103 | ADDRESS_LO_SUM, | |
104 | ADDRESS_SYMBOLIC | |
105 | }; | |
106 | ||
107 | struct aarch64_address_info { | |
108 | enum aarch64_address_type type; | |
109 | rtx base; | |
110 | rtx offset; | |
111 | int shift; | |
112 | enum aarch64_symbol_type symbol_type; | |
113 | }; | |
114 | ||
48063b9d IB |
115 | struct simd_immediate_info |
116 | { | |
117 | rtx value; | |
118 | int shift; | |
119 | int element_width; | |
48063b9d | 120 | bool mvn; |
e4f0f84d | 121 | bool msl; |
48063b9d IB |
122 | }; |
123 | ||
43e9d192 IB |
124 | /* The current code model. */ |
125 | enum aarch64_code_model aarch64_cmodel; | |
126 | ||
127 | #ifdef HAVE_AS_TLS | |
128 | #undef TARGET_HAVE_TLS | |
129 | #define TARGET_HAVE_TLS 1 | |
130 | #endif | |
131 | ||
ef4bddc2 RS |
132 | static bool aarch64_composite_type_p (const_tree, machine_mode); |
133 | static bool aarch64_vfp_is_call_or_return_candidate (machine_mode, | |
43e9d192 | 134 | const_tree, |
ef4bddc2 | 135 | machine_mode *, int *, |
43e9d192 IB |
136 | bool *); |
137 | static void aarch64_elf_asm_constructor (rtx, int) ATTRIBUTE_UNUSED; | |
138 | static void aarch64_elf_asm_destructor (rtx, int) ATTRIBUTE_UNUSED; | |
43e9d192 | 139 | static void aarch64_override_options_after_change (void); |
ef4bddc2 | 140 | static bool aarch64_vector_mode_supported_p (machine_mode); |
ef4bddc2 | 141 | static bool aarch64_vectorize_vec_perm_const_ok (machine_mode vmode, |
88b08073 | 142 | const unsigned char *sel); |
ef4bddc2 | 143 | static int aarch64_address_cost (rtx, machine_mode, addr_space_t, bool); |
88b08073 | 144 | |
0c6caaf8 RL |
145 | /* Major revision number of the ARM Architecture implemented by the target. */ |
146 | unsigned aarch64_architecture_version; | |
147 | ||
43e9d192 | 148 | /* The processor for which instructions should be scheduled. */ |
02fdbd5b | 149 | enum aarch64_processor aarch64_tune = cortexa53; |
43e9d192 | 150 | |
43e9d192 IB |
151 | /* Mask to specify which instruction scheduling options should be used. */ |
152 | unsigned long aarch64_tune_flags = 0; | |
153 | ||
1be34295 | 154 | /* Global flag for PC relative loads. */ |
9ee6540a | 155 | bool aarch64_pcrelative_literal_loads; |
1be34295 | 156 | |
8dec06f2 JG |
157 | /* Support for command line parsing of boolean flags in the tuning |
158 | structures. */ | |
159 | struct aarch64_flag_desc | |
160 | { | |
161 | const char* name; | |
162 | unsigned int flag; | |
163 | }; | |
164 | ||
ed9fa8d2 | 165 | #define AARCH64_FUSION_PAIR(name, internal_name) \ |
8dec06f2 JG |
166 | { name, AARCH64_FUSE_##internal_name }, |
167 | static const struct aarch64_flag_desc aarch64_fusible_pairs[] = | |
168 | { | |
169 | { "none", AARCH64_FUSE_NOTHING }, | |
170 | #include "aarch64-fusion-pairs.def" | |
171 | { "all", AARCH64_FUSE_ALL }, | |
172 | { NULL, AARCH64_FUSE_NOTHING } | |
173 | }; | |
174 | #undef AARCH64_FUION_PAIR | |
175 | ||
a339a01c | 176 | #define AARCH64_EXTRA_TUNING_OPTION(name, internal_name) \ |
8dec06f2 JG |
177 | { name, AARCH64_EXTRA_TUNE_##internal_name }, |
178 | static const struct aarch64_flag_desc aarch64_tuning_flags[] = | |
179 | { | |
180 | { "none", AARCH64_EXTRA_TUNE_NONE }, | |
181 | #include "aarch64-tuning-flags.def" | |
182 | { "all", AARCH64_EXTRA_TUNE_ALL }, | |
183 | { NULL, AARCH64_EXTRA_TUNE_NONE } | |
184 | }; | |
185 | #undef AARCH64_EXTRA_TUNING_OPTION | |
186 | ||
43e9d192 IB |
187 | /* Tuning parameters. */ |
188 | ||
43e9d192 IB |
189 | static const struct cpu_addrcost_table generic_addrcost_table = |
190 | { | |
67747367 | 191 | { |
bd95e655 JG |
192 | 0, /* hi */ |
193 | 0, /* si */ | |
194 | 0, /* di */ | |
195 | 0, /* ti */ | |
67747367 | 196 | }, |
bd95e655 JG |
197 | 0, /* pre_modify */ |
198 | 0, /* post_modify */ | |
199 | 0, /* register_offset */ | |
783879e6 EM |
200 | 0, /* register_sextend */ |
201 | 0, /* register_zextend */ | |
bd95e655 | 202 | 0 /* imm_offset */ |
43e9d192 IB |
203 | }; |
204 | ||
60bff090 JG |
205 | static const struct cpu_addrcost_table cortexa57_addrcost_table = |
206 | { | |
60bff090 | 207 | { |
bd95e655 JG |
208 | 1, /* hi */ |
209 | 0, /* si */ | |
210 | 0, /* di */ | |
211 | 1, /* ti */ | |
60bff090 | 212 | }, |
bd95e655 JG |
213 | 0, /* pre_modify */ |
214 | 0, /* post_modify */ | |
215 | 0, /* register_offset */ | |
783879e6 EM |
216 | 0, /* register_sextend */ |
217 | 0, /* register_zextend */ | |
bd95e655 | 218 | 0, /* imm_offset */ |
60bff090 JG |
219 | }; |
220 | ||
5ec1ae3b EM |
221 | static const struct cpu_addrcost_table exynosm1_addrcost_table = |
222 | { | |
223 | { | |
224 | 0, /* hi */ | |
225 | 0, /* si */ | |
226 | 0, /* di */ | |
227 | 2, /* ti */ | |
228 | }, | |
229 | 0, /* pre_modify */ | |
230 | 0, /* post_modify */ | |
231 | 1, /* register_offset */ | |
232 | 1, /* register_sextend */ | |
233 | 2, /* register_zextend */ | |
234 | 0, /* imm_offset */ | |
235 | }; | |
236 | ||
381e27aa PT |
237 | static const struct cpu_addrcost_table xgene1_addrcost_table = |
238 | { | |
381e27aa | 239 | { |
bd95e655 JG |
240 | 1, /* hi */ |
241 | 0, /* si */ | |
242 | 0, /* di */ | |
243 | 1, /* ti */ | |
381e27aa | 244 | }, |
bd95e655 JG |
245 | 1, /* pre_modify */ |
246 | 0, /* post_modify */ | |
247 | 0, /* register_offset */ | |
783879e6 EM |
248 | 1, /* register_sextend */ |
249 | 1, /* register_zextend */ | |
bd95e655 | 250 | 0, /* imm_offset */ |
381e27aa PT |
251 | }; |
252 | ||
ee446d9f JW |
253 | static const struct cpu_addrcost_table qdf24xx_addrcost_table = |
254 | { | |
255 | { | |
256 | 1, /* hi */ | |
257 | 0, /* si */ | |
258 | 0, /* di */ | |
259 | 1, /* ti */ | |
260 | }, | |
261 | 0, /* pre_modify */ | |
262 | 0, /* post_modify */ | |
263 | 0, /* register_offset */ | |
264 | 0, /* register_sextend */ | |
265 | 0, /* register_zextend */ | |
266 | 0 /* imm_offset */ | |
267 | }; | |
268 | ||
ad611a4c VP |
269 | static const struct cpu_addrcost_table vulcan_addrcost_table = |
270 | { | |
271 | { | |
272 | 0, /* hi */ | |
273 | 0, /* si */ | |
274 | 0, /* di */ | |
275 | 2, /* ti */ | |
276 | }, | |
277 | 0, /* pre_modify */ | |
278 | 0, /* post_modify */ | |
279 | 2, /* register_offset */ | |
280 | 3, /* register_sextend */ | |
281 | 3, /* register_zextend */ | |
282 | 0, /* imm_offset */ | |
283 | }; | |
284 | ||
43e9d192 IB |
285 | static const struct cpu_regmove_cost generic_regmove_cost = |
286 | { | |
bd95e655 | 287 | 1, /* GP2GP */ |
3969c510 WD |
288 | /* Avoid the use of slow int<->fp moves for spilling by setting |
289 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
290 | 5, /* GP2FP */ |
291 | 5, /* FP2GP */ | |
292 | 2 /* FP2FP */ | |
43e9d192 IB |
293 | }; |
294 | ||
e4a9c55a WD |
295 | static const struct cpu_regmove_cost cortexa57_regmove_cost = |
296 | { | |
bd95e655 | 297 | 1, /* GP2GP */ |
e4a9c55a WD |
298 | /* Avoid the use of slow int<->fp moves for spilling by setting |
299 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
300 | 5, /* GP2FP */ |
301 | 5, /* FP2GP */ | |
302 | 2 /* FP2FP */ | |
e4a9c55a WD |
303 | }; |
304 | ||
305 | static const struct cpu_regmove_cost cortexa53_regmove_cost = | |
306 | { | |
bd95e655 | 307 | 1, /* GP2GP */ |
e4a9c55a WD |
308 | /* Avoid the use of slow int<->fp moves for spilling by setting |
309 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
310 | 5, /* GP2FP */ |
311 | 5, /* FP2GP */ | |
312 | 2 /* FP2FP */ | |
e4a9c55a WD |
313 | }; |
314 | ||
5ec1ae3b EM |
315 | static const struct cpu_regmove_cost exynosm1_regmove_cost = |
316 | { | |
317 | 1, /* GP2GP */ | |
318 | /* Avoid the use of slow int<->fp moves for spilling by setting | |
319 | their cost higher than memmov_cost (actual, 4 and 9). */ | |
320 | 9, /* GP2FP */ | |
321 | 9, /* FP2GP */ | |
322 | 1 /* FP2FP */ | |
323 | }; | |
324 | ||
d1bcc29f AP |
325 | static const struct cpu_regmove_cost thunderx_regmove_cost = |
326 | { | |
bd95e655 JG |
327 | 2, /* GP2GP */ |
328 | 2, /* GP2FP */ | |
329 | 6, /* FP2GP */ | |
330 | 4 /* FP2FP */ | |
d1bcc29f AP |
331 | }; |
332 | ||
381e27aa PT |
333 | static const struct cpu_regmove_cost xgene1_regmove_cost = |
334 | { | |
bd95e655 | 335 | 1, /* GP2GP */ |
381e27aa PT |
336 | /* Avoid the use of slow int<->fp moves for spilling by setting |
337 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
338 | 8, /* GP2FP */ |
339 | 8, /* FP2GP */ | |
340 | 2 /* FP2FP */ | |
381e27aa PT |
341 | }; |
342 | ||
ee446d9f JW |
343 | static const struct cpu_regmove_cost qdf24xx_regmove_cost = |
344 | { | |
345 | 2, /* GP2GP */ | |
346 | /* Avoid the use of int<->fp moves for spilling. */ | |
347 | 6, /* GP2FP */ | |
348 | 6, /* FP2GP */ | |
349 | 4 /* FP2FP */ | |
350 | }; | |
351 | ||
ad611a4c VP |
352 | static const struct cpu_regmove_cost vulcan_regmove_cost = |
353 | { | |
354 | 1, /* GP2GP */ | |
355 | /* Avoid the use of int<->fp moves for spilling. */ | |
356 | 8, /* GP2FP */ | |
357 | 8, /* FP2GP */ | |
358 | 4 /* FP2FP */ | |
359 | }; | |
360 | ||
8990e73a | 361 | /* Generic costs for vector insn classes. */ |
8990e73a TB |
362 | static const struct cpu_vector_cost generic_vector_cost = |
363 | { | |
bd95e655 JG |
364 | 1, /* scalar_stmt_cost */ |
365 | 1, /* scalar_load_cost */ | |
366 | 1, /* scalar_store_cost */ | |
367 | 1, /* vec_stmt_cost */ | |
c428f91c | 368 | 2, /* vec_permute_cost */ |
bd95e655 JG |
369 | 1, /* vec_to_scalar_cost */ |
370 | 1, /* scalar_to_vec_cost */ | |
371 | 1, /* vec_align_load_cost */ | |
372 | 1, /* vec_unalign_load_cost */ | |
373 | 1, /* vec_unalign_store_cost */ | |
374 | 1, /* vec_store_cost */ | |
375 | 3, /* cond_taken_branch_cost */ | |
376 | 1 /* cond_not_taken_branch_cost */ | |
8990e73a TB |
377 | }; |
378 | ||
60bff090 | 379 | /* Generic costs for vector insn classes. */ |
60bff090 JG |
380 | static const struct cpu_vector_cost cortexa57_vector_cost = |
381 | { | |
bd95e655 JG |
382 | 1, /* scalar_stmt_cost */ |
383 | 4, /* scalar_load_cost */ | |
384 | 1, /* scalar_store_cost */ | |
385 | 3, /* vec_stmt_cost */ | |
c428f91c | 386 | 3, /* vec_permute_cost */ |
bd95e655 JG |
387 | 8, /* vec_to_scalar_cost */ |
388 | 8, /* scalar_to_vec_cost */ | |
389 | 5, /* vec_align_load_cost */ | |
390 | 5, /* vec_unalign_load_cost */ | |
391 | 1, /* vec_unalign_store_cost */ | |
392 | 1, /* vec_store_cost */ | |
393 | 1, /* cond_taken_branch_cost */ | |
394 | 1 /* cond_not_taken_branch_cost */ | |
60bff090 JG |
395 | }; |
396 | ||
5ec1ae3b EM |
397 | static const struct cpu_vector_cost exynosm1_vector_cost = |
398 | { | |
399 | 1, /* scalar_stmt_cost */ | |
400 | 5, /* scalar_load_cost */ | |
401 | 1, /* scalar_store_cost */ | |
402 | 3, /* vec_stmt_cost */ | |
c428f91c | 403 | 3, /* vec_permute_cost */ |
5ec1ae3b EM |
404 | 3, /* vec_to_scalar_cost */ |
405 | 3, /* scalar_to_vec_cost */ | |
406 | 5, /* vec_align_load_cost */ | |
407 | 5, /* vec_unalign_load_cost */ | |
408 | 1, /* vec_unalign_store_cost */ | |
409 | 1, /* vec_store_cost */ | |
410 | 1, /* cond_taken_branch_cost */ | |
411 | 1 /* cond_not_taken_branch_cost */ | |
412 | }; | |
413 | ||
381e27aa | 414 | /* Generic costs for vector insn classes. */ |
381e27aa PT |
415 | static const struct cpu_vector_cost xgene1_vector_cost = |
416 | { | |
bd95e655 JG |
417 | 1, /* scalar_stmt_cost */ |
418 | 5, /* scalar_load_cost */ | |
419 | 1, /* scalar_store_cost */ | |
420 | 2, /* vec_stmt_cost */ | |
c428f91c | 421 | 2, /* vec_permute_cost */ |
bd95e655 JG |
422 | 4, /* vec_to_scalar_cost */ |
423 | 4, /* scalar_to_vec_cost */ | |
424 | 10, /* vec_align_load_cost */ | |
425 | 10, /* vec_unalign_load_cost */ | |
426 | 2, /* vec_unalign_store_cost */ | |
427 | 2, /* vec_store_cost */ | |
428 | 2, /* cond_taken_branch_cost */ | |
429 | 1 /* cond_not_taken_branch_cost */ | |
381e27aa PT |
430 | }; |
431 | ||
ad611a4c VP |
432 | /* Costs for vector insn classes for Vulcan. */ |
433 | static const struct cpu_vector_cost vulcan_vector_cost = | |
434 | { | |
435 | 6, /* scalar_stmt_cost */ | |
436 | 4, /* scalar_load_cost */ | |
437 | 1, /* scalar_store_cost */ | |
438 | 6, /* vec_stmt_cost */ | |
439 | 3, /* vec_permute_cost */ | |
440 | 6, /* vec_to_scalar_cost */ | |
441 | 5, /* scalar_to_vec_cost */ | |
442 | 8, /* vec_align_load_cost */ | |
443 | 8, /* vec_unalign_load_cost */ | |
444 | 4, /* vec_unalign_store_cost */ | |
445 | 4, /* vec_store_cost */ | |
446 | 2, /* cond_taken_branch_cost */ | |
447 | 1 /* cond_not_taken_branch_cost */ | |
448 | }; | |
449 | ||
b9066f5a MW |
450 | /* Generic costs for branch instructions. */ |
451 | static const struct cpu_branch_cost generic_branch_cost = | |
452 | { | |
453 | 2, /* Predictable. */ | |
454 | 2 /* Unpredictable. */ | |
455 | }; | |
456 | ||
67707f65 JG |
457 | /* Branch costs for Cortex-A57. */ |
458 | static const struct cpu_branch_cost cortexa57_branch_cost = | |
459 | { | |
460 | 1, /* Predictable. */ | |
461 | 3 /* Unpredictable. */ | |
462 | }; | |
463 | ||
ad611a4c VP |
464 | /* Branch costs for Vulcan. */ |
465 | static const struct cpu_branch_cost vulcan_branch_cost = | |
466 | { | |
467 | 1, /* Predictable. */ | |
468 | 3 /* Unpredictable. */ | |
469 | }; | |
470 | ||
9acc9cbe EM |
471 | /* Generic approximation modes. */ |
472 | static const cpu_approx_modes generic_approx_modes = | |
473 | { | |
79a2bc2d | 474 | AARCH64_APPROX_NONE, /* division */ |
98daafa0 | 475 | AARCH64_APPROX_NONE, /* sqrt */ |
9acc9cbe EM |
476 | AARCH64_APPROX_NONE /* recip_sqrt */ |
477 | }; | |
478 | ||
479 | /* Approximation modes for Exynos M1. */ | |
480 | static const cpu_approx_modes exynosm1_approx_modes = | |
481 | { | |
79a2bc2d | 482 | AARCH64_APPROX_NONE, /* division */ |
98daafa0 | 483 | AARCH64_APPROX_ALL, /* sqrt */ |
9acc9cbe EM |
484 | AARCH64_APPROX_ALL /* recip_sqrt */ |
485 | }; | |
486 | ||
487 | /* Approximation modes for X-Gene 1. */ | |
488 | static const cpu_approx_modes xgene1_approx_modes = | |
489 | { | |
79a2bc2d | 490 | AARCH64_APPROX_NONE, /* division */ |
98daafa0 | 491 | AARCH64_APPROX_NONE, /* sqrt */ |
9acc9cbe EM |
492 | AARCH64_APPROX_ALL /* recip_sqrt */ |
493 | }; | |
494 | ||
43e9d192 IB |
495 | static const struct tune_params generic_tunings = |
496 | { | |
4e2cd668 | 497 | &cortexa57_extra_costs, |
43e9d192 IB |
498 | &generic_addrcost_table, |
499 | &generic_regmove_cost, | |
8990e73a | 500 | &generic_vector_cost, |
b9066f5a | 501 | &generic_branch_cost, |
9acc9cbe | 502 | &generic_approx_modes, |
bd95e655 JG |
503 | 4, /* memmov_cost */ |
504 | 2, /* issue_rate */ | |
e9a3a175 | 505 | AARCH64_FUSE_NOTHING, /* fusible_ops */ |
0b82a5a2 WD |
506 | 8, /* function_align. */ |
507 | 8, /* jump_align. */ | |
508 | 4, /* loop_align. */ | |
cee66c68 WD |
509 | 2, /* int_reassoc_width. */ |
510 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
511 | 1, /* vec_reassoc_width. */ |
512 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 513 | 2, /* min_div_recip_mul_df. */ |
50487d79 EM |
514 | 0, /* max_case_values. */ |
515 | 0, /* cache_line_size. */ | |
2d6bc7fa | 516 | tune_params::AUTOPREFETCHER_OFF, /* autoprefetcher_model. */ |
dfba575f | 517 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ |
43e9d192 IB |
518 | }; |
519 | ||
1c72a3ca JG |
520 | static const struct tune_params cortexa35_tunings = |
521 | { | |
522 | &cortexa53_extra_costs, | |
523 | &generic_addrcost_table, | |
524 | &cortexa53_regmove_cost, | |
525 | &generic_vector_cost, | |
0bc24338 | 526 | &cortexa57_branch_cost, |
9acc9cbe | 527 | &generic_approx_modes, |
1c72a3ca JG |
528 | 4, /* memmov_cost */ |
529 | 1, /* issue_rate */ | |
0bc24338 | 530 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD |
1c72a3ca | 531 | | AARCH64_FUSE_MOVK_MOVK | AARCH64_FUSE_ADRP_LDR), /* fusible_ops */ |
d4407370 | 532 | 16, /* function_align. */ |
1c72a3ca | 533 | 8, /* jump_align. */ |
d4407370 | 534 | 8, /* loop_align. */ |
1c72a3ca JG |
535 | 2, /* int_reassoc_width. */ |
536 | 4, /* fp_reassoc_width. */ | |
537 | 1, /* vec_reassoc_width. */ | |
538 | 2, /* min_div_recip_mul_sf. */ | |
539 | 2, /* min_div_recip_mul_df. */ | |
540 | 0, /* max_case_values. */ | |
541 | 0, /* cache_line_size. */ | |
542 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ | |
543 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ | |
544 | }; | |
545 | ||
984239ad KT |
546 | static const struct tune_params cortexa53_tunings = |
547 | { | |
548 | &cortexa53_extra_costs, | |
549 | &generic_addrcost_table, | |
e4a9c55a | 550 | &cortexa53_regmove_cost, |
984239ad | 551 | &generic_vector_cost, |
0bc24338 | 552 | &cortexa57_branch_cost, |
9acc9cbe | 553 | &generic_approx_modes, |
bd95e655 JG |
554 | 4, /* memmov_cost */ |
555 | 2, /* issue_rate */ | |
00a8574a | 556 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD |
e9a3a175 | 557 | | AARCH64_FUSE_MOVK_MOVK | AARCH64_FUSE_ADRP_LDR), /* fusible_ops */ |
d4407370 | 558 | 16, /* function_align. */ |
0b82a5a2 | 559 | 8, /* jump_align. */ |
d4407370 | 560 | 8, /* loop_align. */ |
cee66c68 WD |
561 | 2, /* int_reassoc_width. */ |
562 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
563 | 1, /* vec_reassoc_width. */ |
564 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 565 | 2, /* min_div_recip_mul_df. */ |
50487d79 EM |
566 | 0, /* max_case_values. */ |
567 | 0, /* cache_line_size. */ | |
2d6bc7fa | 568 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
dfba575f | 569 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ |
984239ad KT |
570 | }; |
571 | ||
4fd92af6 KT |
572 | static const struct tune_params cortexa57_tunings = |
573 | { | |
574 | &cortexa57_extra_costs, | |
60bff090 | 575 | &cortexa57_addrcost_table, |
e4a9c55a | 576 | &cortexa57_regmove_cost, |
60bff090 | 577 | &cortexa57_vector_cost, |
67707f65 | 578 | &cortexa57_branch_cost, |
9acc9cbe | 579 | &generic_approx_modes, |
bd95e655 JG |
580 | 4, /* memmov_cost */ |
581 | 3, /* issue_rate */ | |
00a8574a | 582 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD |
e9a3a175 | 583 | | AARCH64_FUSE_MOVK_MOVK), /* fusible_ops */ |
0b82a5a2 WD |
584 | 16, /* function_align. */ |
585 | 8, /* jump_align. */ | |
d4407370 | 586 | 8, /* loop_align. */ |
cee66c68 WD |
587 | 2, /* int_reassoc_width. */ |
588 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
589 | 1, /* vec_reassoc_width. */ |
590 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 591 | 2, /* min_div_recip_mul_df. */ |
50487d79 EM |
592 | 0, /* max_case_values. */ |
593 | 0, /* cache_line_size. */ | |
2d6bc7fa | 594 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
7c175186 | 595 | (AARCH64_EXTRA_TUNE_RENAME_FMA_REGS) /* tune_flags. */ |
dfba575f JG |
596 | }; |
597 | ||
598 | static const struct tune_params cortexa72_tunings = | |
599 | { | |
600 | &cortexa57_extra_costs, | |
601 | &cortexa57_addrcost_table, | |
602 | &cortexa57_regmove_cost, | |
603 | &cortexa57_vector_cost, | |
0bc24338 | 604 | &cortexa57_branch_cost, |
9acc9cbe | 605 | &generic_approx_modes, |
dfba575f JG |
606 | 4, /* memmov_cost */ |
607 | 3, /* issue_rate */ | |
00a8574a | 608 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD |
dfba575f JG |
609 | | AARCH64_FUSE_MOVK_MOVK), /* fusible_ops */ |
610 | 16, /* function_align. */ | |
611 | 8, /* jump_align. */ | |
d4407370 | 612 | 8, /* loop_align. */ |
dfba575f JG |
613 | 2, /* int_reassoc_width. */ |
614 | 4, /* fp_reassoc_width. */ | |
615 | 1, /* vec_reassoc_width. */ | |
616 | 2, /* min_div_recip_mul_sf. */ | |
617 | 2, /* min_div_recip_mul_df. */ | |
50487d79 EM |
618 | 0, /* max_case_values. */ |
619 | 0, /* cache_line_size. */ | |
0bc24338 | 620 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
dfba575f | 621 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ |
4fd92af6 KT |
622 | }; |
623 | ||
4fb570c4 KT |
624 | static const struct tune_params cortexa73_tunings = |
625 | { | |
626 | &cortexa57_extra_costs, | |
627 | &cortexa57_addrcost_table, | |
628 | &cortexa57_regmove_cost, | |
629 | &cortexa57_vector_cost, | |
0bc24338 | 630 | &cortexa57_branch_cost, |
4fb570c4 KT |
631 | &generic_approx_modes, |
632 | 4, /* memmov_cost. */ | |
633 | 2, /* issue_rate. */ | |
634 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD | |
635 | | AARCH64_FUSE_MOVK_MOVK | AARCH64_FUSE_ADRP_LDR), /* fusible_ops */ | |
636 | 16, /* function_align. */ | |
637 | 8, /* jump_align. */ | |
d4407370 | 638 | 8, /* loop_align. */ |
4fb570c4 KT |
639 | 2, /* int_reassoc_width. */ |
640 | 4, /* fp_reassoc_width. */ | |
641 | 1, /* vec_reassoc_width. */ | |
642 | 2, /* min_div_recip_mul_sf. */ | |
643 | 2, /* min_div_recip_mul_df. */ | |
644 | 0, /* max_case_values. */ | |
645 | 0, /* cache_line_size. */ | |
646 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ | |
647 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ | |
648 | }; | |
649 | ||
5ec1ae3b EM |
650 | static const struct tune_params exynosm1_tunings = |
651 | { | |
652 | &exynosm1_extra_costs, | |
653 | &exynosm1_addrcost_table, | |
654 | &exynosm1_regmove_cost, | |
655 | &exynosm1_vector_cost, | |
656 | &generic_branch_cost, | |
9acc9cbe | 657 | &exynosm1_approx_modes, |
5ec1ae3b EM |
658 | 4, /* memmov_cost */ |
659 | 3, /* issue_rate */ | |
25cc2199 | 660 | (AARCH64_FUSE_AES_AESMC), /* fusible_ops */ |
5ec1ae3b EM |
661 | 4, /* function_align. */ |
662 | 4, /* jump_align. */ | |
663 | 4, /* loop_align. */ | |
664 | 2, /* int_reassoc_width. */ | |
665 | 4, /* fp_reassoc_width. */ | |
666 | 1, /* vec_reassoc_width. */ | |
667 | 2, /* min_div_recip_mul_sf. */ | |
668 | 2, /* min_div_recip_mul_df. */ | |
669 | 48, /* max_case_values. */ | |
670 | 64, /* cache_line_size. */ | |
220379df | 671 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
9acc9cbe | 672 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ |
5ec1ae3b EM |
673 | }; |
674 | ||
d1bcc29f AP |
675 | static const struct tune_params thunderx_tunings = |
676 | { | |
677 | &thunderx_extra_costs, | |
678 | &generic_addrcost_table, | |
679 | &thunderx_regmove_cost, | |
680 | &generic_vector_cost, | |
b9066f5a | 681 | &generic_branch_cost, |
9acc9cbe | 682 | &generic_approx_modes, |
bd95e655 JG |
683 | 6, /* memmov_cost */ |
684 | 2, /* issue_rate */ | |
e9a3a175 | 685 | AARCH64_FUSE_CMP_BRANCH, /* fusible_ops */ |
0b82a5a2 WD |
686 | 8, /* function_align. */ |
687 | 8, /* jump_align. */ | |
688 | 8, /* loop_align. */ | |
cee66c68 WD |
689 | 2, /* int_reassoc_width. */ |
690 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
691 | 1, /* vec_reassoc_width. */ |
692 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 693 | 2, /* min_div_recip_mul_df. */ |
50487d79 EM |
694 | 0, /* max_case_values. */ |
695 | 0, /* cache_line_size. */ | |
2d6bc7fa | 696 | tune_params::AUTOPREFETCHER_OFF, /* autoprefetcher_model. */ |
dfba575f | 697 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ |
d1bcc29f AP |
698 | }; |
699 | ||
381e27aa PT |
700 | static const struct tune_params xgene1_tunings = |
701 | { | |
702 | &xgene1_extra_costs, | |
703 | &xgene1_addrcost_table, | |
704 | &xgene1_regmove_cost, | |
705 | &xgene1_vector_cost, | |
b9066f5a | 706 | &generic_branch_cost, |
9acc9cbe | 707 | &xgene1_approx_modes, |
bd95e655 JG |
708 | 6, /* memmov_cost */ |
709 | 4, /* issue_rate */ | |
e9a3a175 | 710 | AARCH64_FUSE_NOTHING, /* fusible_ops */ |
381e27aa PT |
711 | 16, /* function_align. */ |
712 | 8, /* jump_align. */ | |
713 | 16, /* loop_align. */ | |
714 | 2, /* int_reassoc_width. */ | |
715 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
716 | 1, /* vec_reassoc_width. */ |
717 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 718 | 2, /* min_div_recip_mul_df. */ |
50487d79 EM |
719 | 0, /* max_case_values. */ |
720 | 0, /* cache_line_size. */ | |
2d6bc7fa | 721 | tune_params::AUTOPREFETCHER_OFF, /* autoprefetcher_model. */ |
9acc9cbe | 722 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ |
381e27aa PT |
723 | }; |
724 | ||
ee446d9f JW |
725 | static const struct tune_params qdf24xx_tunings = |
726 | { | |
727 | &qdf24xx_extra_costs, | |
728 | &qdf24xx_addrcost_table, | |
729 | &qdf24xx_regmove_cost, | |
730 | &generic_vector_cost, | |
731 | &generic_branch_cost, | |
732 | &generic_approx_modes, | |
733 | 4, /* memmov_cost */ | |
734 | 4, /* issue_rate */ | |
735 | (AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD | |
736 | | AARCH64_FUSE_MOVK_MOVK), /* fuseable_ops */ | |
737 | 16, /* function_align. */ | |
738 | 8, /* jump_align. */ | |
739 | 16, /* loop_align. */ | |
740 | 2, /* int_reassoc_width. */ | |
741 | 4, /* fp_reassoc_width. */ | |
742 | 1, /* vec_reassoc_width. */ | |
743 | 2, /* min_div_recip_mul_sf. */ | |
744 | 2, /* min_div_recip_mul_df. */ | |
745 | 0, /* max_case_values. */ | |
746 | 64, /* cache_line_size. */ | |
747 | tune_params::AUTOPREFETCHER_STRONG, /* autoprefetcher_model. */ | |
748 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ | |
749 | }; | |
750 | ||
ad611a4c VP |
751 | static const struct tune_params vulcan_tunings = |
752 | { | |
753 | &vulcan_extra_costs, | |
754 | &vulcan_addrcost_table, | |
755 | &vulcan_regmove_cost, | |
756 | &vulcan_vector_cost, | |
757 | &vulcan_branch_cost, | |
758 | &generic_approx_modes, | |
759 | 4, /* memmov_cost. */ | |
760 | 4, /* issue_rate. */ | |
761 | AARCH64_FUSE_NOTHING, /* fuseable_ops. */ | |
762 | 16, /* function_align. */ | |
763 | 8, /* jump_align. */ | |
764 | 16, /* loop_align. */ | |
765 | 3, /* int_reassoc_width. */ | |
766 | 2, /* fp_reassoc_width. */ | |
767 | 2, /* vec_reassoc_width. */ | |
768 | 2, /* min_div_recip_mul_sf. */ | |
769 | 2, /* min_div_recip_mul_df. */ | |
770 | 0, /* max_case_values. */ | |
771 | 0, /* cache_line_size. */ | |
772 | tune_params::AUTOPREFETCHER_OFF, /* autoprefetcher_model. */ | |
773 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ | |
774 | }; | |
775 | ||
8dec06f2 JG |
776 | /* Support for fine-grained override of the tuning structures. */ |
777 | struct aarch64_tuning_override_function | |
778 | { | |
779 | const char* name; | |
780 | void (*parse_override)(const char*, struct tune_params*); | |
781 | }; | |
782 | ||
783 | static void aarch64_parse_fuse_string (const char*, struct tune_params*); | |
784 | static void aarch64_parse_tune_string (const char*, struct tune_params*); | |
785 | ||
786 | static const struct aarch64_tuning_override_function | |
787 | aarch64_tuning_override_functions[] = | |
788 | { | |
789 | { "fuse", aarch64_parse_fuse_string }, | |
790 | { "tune", aarch64_parse_tune_string }, | |
791 | { NULL, NULL } | |
792 | }; | |
793 | ||
43e9d192 IB |
794 | /* A processor implementing AArch64. */ |
795 | struct processor | |
796 | { | |
797 | const char *const name; | |
46806c44 KT |
798 | enum aarch64_processor ident; |
799 | enum aarch64_processor sched_core; | |
393ae126 | 800 | enum aarch64_arch arch; |
0c6caaf8 | 801 | unsigned architecture_version; |
43e9d192 IB |
802 | const unsigned long flags; |
803 | const struct tune_params *const tune; | |
804 | }; | |
805 | ||
393ae126 KT |
806 | /* Architectures implementing AArch64. */ |
807 | static const struct processor all_architectures[] = | |
808 | { | |
809 | #define AARCH64_ARCH(NAME, CORE, ARCH_IDENT, ARCH_REV, FLAGS) \ | |
810 | {NAME, CORE, CORE, AARCH64_ARCH_##ARCH_IDENT, ARCH_REV, FLAGS, NULL}, | |
811 | #include "aarch64-arches.def" | |
812 | #undef AARCH64_ARCH | |
813 | {NULL, aarch64_none, aarch64_none, aarch64_no_arch, 0, 0, NULL} | |
814 | }; | |
815 | ||
43e9d192 IB |
816 | /* Processor cores implementing AArch64. */ |
817 | static const struct processor all_cores[] = | |
818 | { | |
7e1bcce3 | 819 | #define AARCH64_CORE(NAME, IDENT, SCHED, ARCH, FLAGS, COSTS, IMP, PART) \ |
393ae126 KT |
820 | {NAME, IDENT, SCHED, AARCH64_ARCH_##ARCH, \ |
821 | all_architectures[AARCH64_ARCH_##ARCH].architecture_version, \ | |
822 | FLAGS, &COSTS##_tunings}, | |
43e9d192 IB |
823 | #include "aarch64-cores.def" |
824 | #undef AARCH64_CORE | |
393ae126 KT |
825 | {"generic", generic, cortexa53, AARCH64_ARCH_8A, 8, |
826 | AARCH64_FL_FOR_ARCH8, &generic_tunings}, | |
827 | {NULL, aarch64_none, aarch64_none, aarch64_no_arch, 0, 0, NULL} | |
43e9d192 IB |
828 | }; |
829 | ||
43e9d192 | 830 | |
361fb3ee KT |
831 | /* Target specification. These are populated by the -march, -mtune, -mcpu |
832 | handling code or by target attributes. */ | |
43e9d192 IB |
833 | static const struct processor *selected_arch; |
834 | static const struct processor *selected_cpu; | |
835 | static const struct processor *selected_tune; | |
836 | ||
b175b679 JG |
837 | /* The current tuning set. */ |
838 | struct tune_params aarch64_tune_params = generic_tunings; | |
839 | ||
43e9d192 IB |
840 | #define AARCH64_CPU_DEFAULT_FLAGS ((selected_cpu) ? selected_cpu->flags : 0) |
841 | ||
842 | /* An ISA extension in the co-processor and main instruction set space. */ | |
843 | struct aarch64_option_extension | |
844 | { | |
845 | const char *const name; | |
846 | const unsigned long flags_on; | |
847 | const unsigned long flags_off; | |
848 | }; | |
849 | ||
43e9d192 IB |
850 | typedef enum aarch64_cond_code |
851 | { | |
852 | AARCH64_EQ = 0, AARCH64_NE, AARCH64_CS, AARCH64_CC, AARCH64_MI, AARCH64_PL, | |
853 | AARCH64_VS, AARCH64_VC, AARCH64_HI, AARCH64_LS, AARCH64_GE, AARCH64_LT, | |
854 | AARCH64_GT, AARCH64_LE, AARCH64_AL, AARCH64_NV | |
855 | } | |
856 | aarch64_cc; | |
857 | ||
858 | #define AARCH64_INVERSE_CONDITION_CODE(X) ((aarch64_cc) (((int) X) ^ 1)) | |
859 | ||
860 | /* The condition codes of the processor, and the inverse function. */ | |
861 | static const char * const aarch64_condition_codes[] = | |
862 | { | |
863 | "eq", "ne", "cs", "cc", "mi", "pl", "vs", "vc", | |
864 | "hi", "ls", "ge", "lt", "gt", "le", "al", "nv" | |
865 | }; | |
866 | ||
973d2e01 TP |
867 | /* Generate code to enable conditional branches in functions over 1 MiB. */ |
868 | const char * | |
869 | aarch64_gen_far_branch (rtx * operands, int pos_label, const char * dest, | |
870 | const char * branch_format) | |
871 | { | |
872 | rtx_code_label * tmp_label = gen_label_rtx (); | |
873 | char label_buf[256]; | |
874 | char buffer[128]; | |
875 | ASM_GENERATE_INTERNAL_LABEL (label_buf, dest, | |
876 | CODE_LABEL_NUMBER (tmp_label)); | |
877 | const char *label_ptr = targetm.strip_name_encoding (label_buf); | |
878 | rtx dest_label = operands[pos_label]; | |
879 | operands[pos_label] = tmp_label; | |
880 | ||
881 | snprintf (buffer, sizeof (buffer), "%s%s", branch_format, label_ptr); | |
882 | output_asm_insn (buffer, operands); | |
883 | ||
884 | snprintf (buffer, sizeof (buffer), "b\t%%l%d\n%s:", pos_label, label_ptr); | |
885 | operands[pos_label] = dest_label; | |
886 | output_asm_insn (buffer, operands); | |
887 | return ""; | |
888 | } | |
889 | ||
261fb553 AL |
890 | void |
891 | aarch64_err_no_fpadvsimd (machine_mode mode, const char *msg) | |
892 | { | |
893 | const char *mc = FLOAT_MODE_P (mode) ? "floating-point" : "vector"; | |
894 | if (TARGET_GENERAL_REGS_ONLY) | |
895 | error ("%qs is incompatible with %s %s", "-mgeneral-regs-only", mc, msg); | |
896 | else | |
897 | error ("%qs feature modifier is incompatible with %s %s", "+nofp", mc, msg); | |
898 | } | |
899 | ||
c64f7d37 WD |
900 | /* Implement TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS. |
901 | The register allocator chooses ALL_REGS if FP_REGS and GENERAL_REGS have | |
31e2b5a3 WD |
902 | the same cost even if ALL_REGS has a much larger cost. ALL_REGS is also |
903 | used if the cost of both FP_REGS and GENERAL_REGS is lower than the memory | |
904 | cost (in this case the best class is the lowest cost one). Using ALL_REGS | |
905 | irrespectively of its cost results in bad allocations with many redundant | |
906 | int<->FP moves which are expensive on various cores. | |
907 | To avoid this we don't allow ALL_REGS as the allocno class, but force a | |
908 | decision between FP_REGS and GENERAL_REGS. We use the allocno class if it | |
909 | isn't ALL_REGS. Similarly, use the best class if it isn't ALL_REGS. | |
910 | Otherwise set the allocno class depending on the mode. | |
911 | The result of this is that it is no longer inefficient to have a higher | |
912 | memory move cost than the register move cost. | |
913 | */ | |
c64f7d37 WD |
914 | |
915 | static reg_class_t | |
31e2b5a3 WD |
916 | aarch64_ira_change_pseudo_allocno_class (int regno, reg_class_t allocno_class, |
917 | reg_class_t best_class) | |
c64f7d37 WD |
918 | { |
919 | enum machine_mode mode; | |
920 | ||
921 | if (allocno_class != ALL_REGS) | |
922 | return allocno_class; | |
923 | ||
31e2b5a3 WD |
924 | if (best_class != ALL_REGS) |
925 | return best_class; | |
926 | ||
c64f7d37 WD |
927 | mode = PSEUDO_REGNO_MODE (regno); |
928 | return FLOAT_MODE_P (mode) || VECTOR_MODE_P (mode) ? FP_REGS : GENERAL_REGS; | |
929 | } | |
930 | ||
26e0ff94 | 931 | static unsigned int |
50093a33 | 932 | aarch64_min_divisions_for_recip_mul (enum machine_mode mode) |
26e0ff94 | 933 | { |
50093a33 | 934 | if (GET_MODE_UNIT_SIZE (mode) == 4) |
b175b679 JG |
935 | return aarch64_tune_params.min_div_recip_mul_sf; |
936 | return aarch64_tune_params.min_div_recip_mul_df; | |
26e0ff94 WD |
937 | } |
938 | ||
cee66c68 WD |
939 | static int |
940 | aarch64_reassociation_width (unsigned opc ATTRIBUTE_UNUSED, | |
941 | enum machine_mode mode) | |
942 | { | |
943 | if (VECTOR_MODE_P (mode)) | |
b175b679 | 944 | return aarch64_tune_params.vec_reassoc_width; |
cee66c68 | 945 | if (INTEGRAL_MODE_P (mode)) |
b175b679 | 946 | return aarch64_tune_params.int_reassoc_width; |
cee66c68 | 947 | if (FLOAT_MODE_P (mode)) |
b175b679 | 948 | return aarch64_tune_params.fp_reassoc_width; |
cee66c68 WD |
949 | return 1; |
950 | } | |
951 | ||
43e9d192 IB |
952 | /* Provide a mapping from gcc register numbers to dwarf register numbers. */ |
953 | unsigned | |
954 | aarch64_dbx_register_number (unsigned regno) | |
955 | { | |
956 | if (GP_REGNUM_P (regno)) | |
957 | return AARCH64_DWARF_R0 + regno - R0_REGNUM; | |
958 | else if (regno == SP_REGNUM) | |
959 | return AARCH64_DWARF_SP; | |
960 | else if (FP_REGNUM_P (regno)) | |
961 | return AARCH64_DWARF_V0 + regno - V0_REGNUM; | |
962 | ||
963 | /* Return values >= DWARF_FRAME_REGISTERS indicate that there is no | |
964 | equivalent DWARF register. */ | |
965 | return DWARF_FRAME_REGISTERS; | |
966 | } | |
967 | ||
968 | /* Return TRUE if MODE is any of the large INT modes. */ | |
969 | static bool | |
ef4bddc2 | 970 | aarch64_vect_struct_mode_p (machine_mode mode) |
43e9d192 IB |
971 | { |
972 | return mode == OImode || mode == CImode || mode == XImode; | |
973 | } | |
974 | ||
975 | /* Return TRUE if MODE is any of the vector modes. */ | |
976 | static bool | |
ef4bddc2 | 977 | aarch64_vector_mode_p (machine_mode mode) |
43e9d192 IB |
978 | { |
979 | return aarch64_vector_mode_supported_p (mode) | |
980 | || aarch64_vect_struct_mode_p (mode); | |
981 | } | |
982 | ||
983 | /* Implement target hook TARGET_ARRAY_MODE_SUPPORTED_P. */ | |
984 | static bool | |
ef4bddc2 | 985 | aarch64_array_mode_supported_p (machine_mode mode, |
43e9d192 IB |
986 | unsigned HOST_WIDE_INT nelems) |
987 | { | |
988 | if (TARGET_SIMD | |
635e66fe AL |
989 | && (AARCH64_VALID_SIMD_QREG_MODE (mode) |
990 | || AARCH64_VALID_SIMD_DREG_MODE (mode)) | |
43e9d192 IB |
991 | && (nelems >= 2 && nelems <= 4)) |
992 | return true; | |
993 | ||
994 | return false; | |
995 | } | |
996 | ||
997 | /* Implement HARD_REGNO_NREGS. */ | |
998 | ||
999 | int | |
ef4bddc2 | 1000 | aarch64_hard_regno_nregs (unsigned regno, machine_mode mode) |
43e9d192 IB |
1001 | { |
1002 | switch (aarch64_regno_regclass (regno)) | |
1003 | { | |
1004 | case FP_REGS: | |
1005 | case FP_LO_REGS: | |
1006 | return (GET_MODE_SIZE (mode) + UNITS_PER_VREG - 1) / UNITS_PER_VREG; | |
1007 | default: | |
1008 | return (GET_MODE_SIZE (mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
1009 | } | |
1010 | gcc_unreachable (); | |
1011 | } | |
1012 | ||
1013 | /* Implement HARD_REGNO_MODE_OK. */ | |
1014 | ||
1015 | int | |
ef4bddc2 | 1016 | aarch64_hard_regno_mode_ok (unsigned regno, machine_mode mode) |
43e9d192 IB |
1017 | { |
1018 | if (GET_MODE_CLASS (mode) == MODE_CC) | |
1019 | return regno == CC_REGNUM; | |
1020 | ||
9259db42 YZ |
1021 | if (regno == SP_REGNUM) |
1022 | /* The purpose of comparing with ptr_mode is to support the | |
1023 | global register variable associated with the stack pointer | |
1024 | register via the syntax of asm ("wsp") in ILP32. */ | |
1025 | return mode == Pmode || mode == ptr_mode; | |
1026 | ||
1027 | if (regno == FRAME_POINTER_REGNUM || regno == ARG_POINTER_REGNUM) | |
43e9d192 IB |
1028 | return mode == Pmode; |
1029 | ||
1030 | if (GP_REGNUM_P (regno) && ! aarch64_vect_struct_mode_p (mode)) | |
1031 | return 1; | |
1032 | ||
1033 | if (FP_REGNUM_P (regno)) | |
1034 | { | |
1035 | if (aarch64_vect_struct_mode_p (mode)) | |
1036 | return | |
1037 | (regno + aarch64_hard_regno_nregs (regno, mode) - 1) <= V31_REGNUM; | |
1038 | else | |
1039 | return 1; | |
1040 | } | |
1041 | ||
1042 | return 0; | |
1043 | } | |
1044 | ||
73d9ac6a | 1045 | /* Implement HARD_REGNO_CALLER_SAVE_MODE. */ |
ef4bddc2 | 1046 | machine_mode |
73d9ac6a | 1047 | aarch64_hard_regno_caller_save_mode (unsigned regno, unsigned nregs, |
ef4bddc2 | 1048 | machine_mode mode) |
73d9ac6a IB |
1049 | { |
1050 | /* Handle modes that fit within single registers. */ | |
1051 | if (nregs == 1 && GET_MODE_SIZE (mode) <= 16) | |
1052 | { | |
1053 | if (GET_MODE_SIZE (mode) >= 4) | |
1054 | return mode; | |
1055 | else | |
1056 | return SImode; | |
1057 | } | |
1058 | /* Fall back to generic for multi-reg and very large modes. */ | |
1059 | else | |
1060 | return choose_hard_reg_mode (regno, nregs, false); | |
1061 | } | |
1062 | ||
43e9d192 IB |
1063 | /* Return true if calls to DECL should be treated as |
1064 | long-calls (ie called via a register). */ | |
1065 | static bool | |
1066 | aarch64_decl_is_long_call_p (const_tree decl ATTRIBUTE_UNUSED) | |
1067 | { | |
1068 | return false; | |
1069 | } | |
1070 | ||
1071 | /* Return true if calls to symbol-ref SYM should be treated as | |
1072 | long-calls (ie called via a register). */ | |
1073 | bool | |
1074 | aarch64_is_long_call_p (rtx sym) | |
1075 | { | |
1076 | return aarch64_decl_is_long_call_p (SYMBOL_REF_DECL (sym)); | |
1077 | } | |
1078 | ||
b60d63cb JW |
1079 | /* Return true if calls to symbol-ref SYM should not go through |
1080 | plt stubs. */ | |
1081 | ||
1082 | bool | |
1083 | aarch64_is_noplt_call_p (rtx sym) | |
1084 | { | |
1085 | const_tree decl = SYMBOL_REF_DECL (sym); | |
1086 | ||
1087 | if (flag_pic | |
1088 | && decl | |
1089 | && (!flag_plt | |
1090 | || lookup_attribute ("noplt", DECL_ATTRIBUTES (decl))) | |
1091 | && !targetm.binds_local_p (decl)) | |
1092 | return true; | |
1093 | ||
1094 | return false; | |
1095 | } | |
1096 | ||
43e9d192 IB |
1097 | /* Return true if the offsets to a zero/sign-extract operation |
1098 | represent an expression that matches an extend operation. The | |
1099 | operands represent the paramters from | |
1100 | ||
4745e701 | 1101 | (extract:MODE (mult (reg) (MULT_IMM)) (EXTRACT_IMM) (const_int 0)). */ |
43e9d192 | 1102 | bool |
ef4bddc2 | 1103 | aarch64_is_extend_from_extract (machine_mode mode, rtx mult_imm, |
43e9d192 IB |
1104 | rtx extract_imm) |
1105 | { | |
1106 | HOST_WIDE_INT mult_val, extract_val; | |
1107 | ||
1108 | if (! CONST_INT_P (mult_imm) || ! CONST_INT_P (extract_imm)) | |
1109 | return false; | |
1110 | ||
1111 | mult_val = INTVAL (mult_imm); | |
1112 | extract_val = INTVAL (extract_imm); | |
1113 | ||
1114 | if (extract_val > 8 | |
1115 | && extract_val < GET_MODE_BITSIZE (mode) | |
1116 | && exact_log2 (extract_val & ~7) > 0 | |
1117 | && (extract_val & 7) <= 4 | |
1118 | && mult_val == (1 << (extract_val & 7))) | |
1119 | return true; | |
1120 | ||
1121 | return false; | |
1122 | } | |
1123 | ||
1124 | /* Emit an insn that's a simple single-set. Both the operands must be | |
1125 | known to be valid. */ | |
1126 | inline static rtx | |
1127 | emit_set_insn (rtx x, rtx y) | |
1128 | { | |
f7df4a84 | 1129 | return emit_insn (gen_rtx_SET (x, y)); |
43e9d192 IB |
1130 | } |
1131 | ||
1132 | /* X and Y are two things to compare using CODE. Emit the compare insn and | |
1133 | return the rtx for register 0 in the proper mode. */ | |
1134 | rtx | |
1135 | aarch64_gen_compare_reg (RTX_CODE code, rtx x, rtx y) | |
1136 | { | |
ef4bddc2 | 1137 | machine_mode mode = SELECT_CC_MODE (code, x, y); |
43e9d192 IB |
1138 | rtx cc_reg = gen_rtx_REG (mode, CC_REGNUM); |
1139 | ||
1140 | emit_set_insn (cc_reg, gen_rtx_COMPARE (mode, x, y)); | |
1141 | return cc_reg; | |
1142 | } | |
1143 | ||
1144 | /* Build the SYMBOL_REF for __tls_get_addr. */ | |
1145 | ||
1146 | static GTY(()) rtx tls_get_addr_libfunc; | |
1147 | ||
1148 | rtx | |
1149 | aarch64_tls_get_addr (void) | |
1150 | { | |
1151 | if (!tls_get_addr_libfunc) | |
1152 | tls_get_addr_libfunc = init_one_libfunc ("__tls_get_addr"); | |
1153 | return tls_get_addr_libfunc; | |
1154 | } | |
1155 | ||
1156 | /* Return the TLS model to use for ADDR. */ | |
1157 | ||
1158 | static enum tls_model | |
1159 | tls_symbolic_operand_type (rtx addr) | |
1160 | { | |
1161 | enum tls_model tls_kind = TLS_MODEL_NONE; | |
1162 | rtx sym, addend; | |
1163 | ||
1164 | if (GET_CODE (addr) == CONST) | |
1165 | { | |
1166 | split_const (addr, &sym, &addend); | |
1167 | if (GET_CODE (sym) == SYMBOL_REF) | |
1168 | tls_kind = SYMBOL_REF_TLS_MODEL (sym); | |
1169 | } | |
1170 | else if (GET_CODE (addr) == SYMBOL_REF) | |
1171 | tls_kind = SYMBOL_REF_TLS_MODEL (addr); | |
1172 | ||
1173 | return tls_kind; | |
1174 | } | |
1175 | ||
1176 | /* We'll allow lo_sum's in addresses in our legitimate addresses | |
1177 | so that combine would take care of combining addresses where | |
1178 | necessary, but for generation purposes, we'll generate the address | |
1179 | as : | |
1180 | RTL Absolute | |
1181 | tmp = hi (symbol_ref); adrp x1, foo | |
1182 | dest = lo_sum (tmp, symbol_ref); add dest, x1, :lo_12:foo | |
1183 | nop | |
1184 | ||
1185 | PIC TLS | |
1186 | adrp x1, :got:foo adrp tmp, :tlsgd:foo | |
1187 | ldr x1, [:got_lo12:foo] add dest, tmp, :tlsgd_lo12:foo | |
1188 | bl __tls_get_addr | |
1189 | nop | |
1190 | ||
1191 | Load TLS symbol, depending on TLS mechanism and TLS access model. | |
1192 | ||
1193 | Global Dynamic - Traditional TLS: | |
1194 | adrp tmp, :tlsgd:imm | |
1195 | add dest, tmp, #:tlsgd_lo12:imm | |
1196 | bl __tls_get_addr | |
1197 | ||
1198 | Global Dynamic - TLS Descriptors: | |
1199 | adrp dest, :tlsdesc:imm | |
1200 | ldr tmp, [dest, #:tlsdesc_lo12:imm] | |
1201 | add dest, dest, #:tlsdesc_lo12:imm | |
1202 | blr tmp | |
1203 | mrs tp, tpidr_el0 | |
1204 | add dest, dest, tp | |
1205 | ||
1206 | Initial Exec: | |
1207 | mrs tp, tpidr_el0 | |
1208 | adrp tmp, :gottprel:imm | |
1209 | ldr dest, [tmp, #:gottprel_lo12:imm] | |
1210 | add dest, dest, tp | |
1211 | ||
1212 | Local Exec: | |
1213 | mrs tp, tpidr_el0 | |
0699caae RL |
1214 | add t0, tp, #:tprel_hi12:imm, lsl #12 |
1215 | add t0, t0, #:tprel_lo12_nc:imm | |
43e9d192 IB |
1216 | */ |
1217 | ||
1218 | static void | |
1219 | aarch64_load_symref_appropriately (rtx dest, rtx imm, | |
1220 | enum aarch64_symbol_type type) | |
1221 | { | |
1222 | switch (type) | |
1223 | { | |
1224 | case SYMBOL_SMALL_ABSOLUTE: | |
1225 | { | |
28514dda | 1226 | /* In ILP32, the mode of dest can be either SImode or DImode. */ |
43e9d192 | 1227 | rtx tmp_reg = dest; |
ef4bddc2 | 1228 | machine_mode mode = GET_MODE (dest); |
28514dda YZ |
1229 | |
1230 | gcc_assert (mode == Pmode || mode == ptr_mode); | |
1231 | ||
43e9d192 | 1232 | if (can_create_pseudo_p ()) |
28514dda | 1233 | tmp_reg = gen_reg_rtx (mode); |
43e9d192 | 1234 | |
28514dda | 1235 | emit_move_insn (tmp_reg, gen_rtx_HIGH (mode, imm)); |
43e9d192 IB |
1236 | emit_insn (gen_add_losym (dest, tmp_reg, imm)); |
1237 | return; | |
1238 | } | |
1239 | ||
a5350ddc | 1240 | case SYMBOL_TINY_ABSOLUTE: |
f7df4a84 | 1241 | emit_insn (gen_rtx_SET (dest, imm)); |
a5350ddc CSS |
1242 | return; |
1243 | ||
1b1e81f8 JW |
1244 | case SYMBOL_SMALL_GOT_28K: |
1245 | { | |
1246 | machine_mode mode = GET_MODE (dest); | |
1247 | rtx gp_rtx = pic_offset_table_rtx; | |
53021678 JW |
1248 | rtx insn; |
1249 | rtx mem; | |
1b1e81f8 JW |
1250 | |
1251 | /* NOTE: pic_offset_table_rtx can be NULL_RTX, because we can reach | |
1252 | here before rtl expand. Tree IVOPT will generate rtl pattern to | |
1253 | decide rtx costs, in which case pic_offset_table_rtx is not | |
1254 | initialized. For that case no need to generate the first adrp | |
026c3cfd | 1255 | instruction as the final cost for global variable access is |
1b1e81f8 JW |
1256 | one instruction. */ |
1257 | if (gp_rtx != NULL) | |
1258 | { | |
1259 | /* -fpic for -mcmodel=small allow 32K GOT table size (but we are | |
1260 | using the page base as GOT base, the first page may be wasted, | |
1261 | in the worst scenario, there is only 28K space for GOT). | |
1262 | ||
1263 | The generate instruction sequence for accessing global variable | |
1264 | is: | |
1265 | ||
a3957742 | 1266 | ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym] |
1b1e81f8 JW |
1267 | |
1268 | Only one instruction needed. But we must initialize | |
1269 | pic_offset_table_rtx properly. We generate initialize insn for | |
1270 | every global access, and allow CSE to remove all redundant. | |
1271 | ||
1272 | The final instruction sequences will look like the following | |
1273 | for multiply global variables access. | |
1274 | ||
a3957742 | 1275 | adrp pic_offset_table_rtx, _GLOBAL_OFFSET_TABLE_ |
1b1e81f8 | 1276 | |
a3957742 JW |
1277 | ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym1] |
1278 | ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym2] | |
1279 | ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym3] | |
1280 | ... */ | |
1b1e81f8 JW |
1281 | |
1282 | rtx s = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_"); | |
1283 | crtl->uses_pic_offset_table = 1; | |
1284 | emit_move_insn (gp_rtx, gen_rtx_HIGH (Pmode, s)); | |
1285 | ||
1286 | if (mode != GET_MODE (gp_rtx)) | |
1287 | gp_rtx = simplify_gen_subreg (mode, gp_rtx, GET_MODE (gp_rtx), 0); | |
1288 | } | |
1289 | ||
1290 | if (mode == ptr_mode) | |
1291 | { | |
1292 | if (mode == DImode) | |
53021678 | 1293 | insn = gen_ldr_got_small_28k_di (dest, gp_rtx, imm); |
1b1e81f8 | 1294 | else |
53021678 JW |
1295 | insn = gen_ldr_got_small_28k_si (dest, gp_rtx, imm); |
1296 | ||
1297 | mem = XVECEXP (SET_SRC (insn), 0, 0); | |
1b1e81f8 JW |
1298 | } |
1299 | else | |
1300 | { | |
1301 | gcc_assert (mode == Pmode); | |
53021678 JW |
1302 | |
1303 | insn = gen_ldr_got_small_28k_sidi (dest, gp_rtx, imm); | |
1304 | mem = XVECEXP (XEXP (SET_SRC (insn), 0), 0, 0); | |
1b1e81f8 JW |
1305 | } |
1306 | ||
53021678 JW |
1307 | /* The operand is expected to be MEM. Whenever the related insn |
1308 | pattern changed, above code which calculate mem should be | |
1309 | updated. */ | |
1310 | gcc_assert (GET_CODE (mem) == MEM); | |
1311 | MEM_READONLY_P (mem) = 1; | |
1312 | MEM_NOTRAP_P (mem) = 1; | |
1313 | emit_insn (insn); | |
1b1e81f8 JW |
1314 | return; |
1315 | } | |
1316 | ||
6642bdb4 | 1317 | case SYMBOL_SMALL_GOT_4G: |
43e9d192 | 1318 | { |
28514dda YZ |
1319 | /* In ILP32, the mode of dest can be either SImode or DImode, |
1320 | while the got entry is always of SImode size. The mode of | |
1321 | dest depends on how dest is used: if dest is assigned to a | |
1322 | pointer (e.g. in the memory), it has SImode; it may have | |
1323 | DImode if dest is dereferenced to access the memeory. | |
1324 | This is why we have to handle three different ldr_got_small | |
1325 | patterns here (two patterns for ILP32). */ | |
53021678 JW |
1326 | |
1327 | rtx insn; | |
1328 | rtx mem; | |
43e9d192 | 1329 | rtx tmp_reg = dest; |
ef4bddc2 | 1330 | machine_mode mode = GET_MODE (dest); |
28514dda | 1331 | |
43e9d192 | 1332 | if (can_create_pseudo_p ()) |
28514dda YZ |
1333 | tmp_reg = gen_reg_rtx (mode); |
1334 | ||
1335 | emit_move_insn (tmp_reg, gen_rtx_HIGH (mode, imm)); | |
1336 | if (mode == ptr_mode) | |
1337 | { | |
1338 | if (mode == DImode) | |
53021678 | 1339 | insn = gen_ldr_got_small_di (dest, tmp_reg, imm); |
28514dda | 1340 | else |
53021678 JW |
1341 | insn = gen_ldr_got_small_si (dest, tmp_reg, imm); |
1342 | ||
1343 | mem = XVECEXP (SET_SRC (insn), 0, 0); | |
28514dda YZ |
1344 | } |
1345 | else | |
1346 | { | |
1347 | gcc_assert (mode == Pmode); | |
53021678 JW |
1348 | |
1349 | insn = gen_ldr_got_small_sidi (dest, tmp_reg, imm); | |
1350 | mem = XVECEXP (XEXP (SET_SRC (insn), 0), 0, 0); | |
28514dda YZ |
1351 | } |
1352 | ||
53021678 JW |
1353 | gcc_assert (GET_CODE (mem) == MEM); |
1354 | MEM_READONLY_P (mem) = 1; | |
1355 | MEM_NOTRAP_P (mem) = 1; | |
1356 | emit_insn (insn); | |
43e9d192 IB |
1357 | return; |
1358 | } | |
1359 | ||
1360 | case SYMBOL_SMALL_TLSGD: | |
1361 | { | |
5d8a22a5 | 1362 | rtx_insn *insns; |
43e9d192 IB |
1363 | rtx result = gen_rtx_REG (Pmode, R0_REGNUM); |
1364 | ||
1365 | start_sequence (); | |
78607708 | 1366 | aarch64_emit_call_insn (gen_tlsgd_small (result, imm)); |
43e9d192 IB |
1367 | insns = get_insns (); |
1368 | end_sequence (); | |
1369 | ||
1370 | RTL_CONST_CALL_P (insns) = 1; | |
1371 | emit_libcall_block (insns, dest, result, imm); | |
1372 | return; | |
1373 | } | |
1374 | ||
1375 | case SYMBOL_SMALL_TLSDESC: | |
1376 | { | |
ef4bddc2 | 1377 | machine_mode mode = GET_MODE (dest); |
621ad2de | 1378 | rtx x0 = gen_rtx_REG (mode, R0_REGNUM); |
43e9d192 IB |
1379 | rtx tp; |
1380 | ||
621ad2de AP |
1381 | gcc_assert (mode == Pmode || mode == ptr_mode); |
1382 | ||
2876a13f JW |
1383 | /* In ILP32, the got entry is always of SImode size. Unlike |
1384 | small GOT, the dest is fixed at reg 0. */ | |
1385 | if (TARGET_ILP32) | |
1386 | emit_insn (gen_tlsdesc_small_si (imm)); | |
621ad2de | 1387 | else |
2876a13f | 1388 | emit_insn (gen_tlsdesc_small_di (imm)); |
43e9d192 | 1389 | tp = aarch64_load_tp (NULL); |
621ad2de AP |
1390 | |
1391 | if (mode != Pmode) | |
1392 | tp = gen_lowpart (mode, tp); | |
1393 | ||
2876a13f | 1394 | emit_insn (gen_rtx_SET (dest, gen_rtx_PLUS (mode, tp, x0))); |
43e9d192 IB |
1395 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); |
1396 | return; | |
1397 | } | |
1398 | ||
79496620 | 1399 | case SYMBOL_SMALL_TLSIE: |
43e9d192 | 1400 | { |
621ad2de AP |
1401 | /* In ILP32, the mode of dest can be either SImode or DImode, |
1402 | while the got entry is always of SImode size. The mode of | |
1403 | dest depends on how dest is used: if dest is assigned to a | |
1404 | pointer (e.g. in the memory), it has SImode; it may have | |
1405 | DImode if dest is dereferenced to access the memeory. | |
1406 | This is why we have to handle three different tlsie_small | |
1407 | patterns here (two patterns for ILP32). */ | |
ef4bddc2 | 1408 | machine_mode mode = GET_MODE (dest); |
621ad2de | 1409 | rtx tmp_reg = gen_reg_rtx (mode); |
43e9d192 | 1410 | rtx tp = aarch64_load_tp (NULL); |
621ad2de AP |
1411 | |
1412 | if (mode == ptr_mode) | |
1413 | { | |
1414 | if (mode == DImode) | |
1415 | emit_insn (gen_tlsie_small_di (tmp_reg, imm)); | |
1416 | else | |
1417 | { | |
1418 | emit_insn (gen_tlsie_small_si (tmp_reg, imm)); | |
1419 | tp = gen_lowpart (mode, tp); | |
1420 | } | |
1421 | } | |
1422 | else | |
1423 | { | |
1424 | gcc_assert (mode == Pmode); | |
1425 | emit_insn (gen_tlsie_small_sidi (tmp_reg, imm)); | |
1426 | } | |
1427 | ||
f7df4a84 | 1428 | emit_insn (gen_rtx_SET (dest, gen_rtx_PLUS (mode, tp, tmp_reg))); |
43e9d192 IB |
1429 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); |
1430 | return; | |
1431 | } | |
1432 | ||
cbf5629e | 1433 | case SYMBOL_TLSLE12: |
d18ba284 | 1434 | case SYMBOL_TLSLE24: |
cbf5629e JW |
1435 | case SYMBOL_TLSLE32: |
1436 | case SYMBOL_TLSLE48: | |
43e9d192 | 1437 | { |
cbf5629e | 1438 | machine_mode mode = GET_MODE (dest); |
43e9d192 | 1439 | rtx tp = aarch64_load_tp (NULL); |
e6f7f0e9 | 1440 | |
cbf5629e JW |
1441 | if (mode != Pmode) |
1442 | tp = gen_lowpart (mode, tp); | |
1443 | ||
1444 | switch (type) | |
1445 | { | |
1446 | case SYMBOL_TLSLE12: | |
1447 | emit_insn ((mode == DImode ? gen_tlsle12_di : gen_tlsle12_si) | |
1448 | (dest, tp, imm)); | |
1449 | break; | |
1450 | case SYMBOL_TLSLE24: | |
1451 | emit_insn ((mode == DImode ? gen_tlsle24_di : gen_tlsle24_si) | |
1452 | (dest, tp, imm)); | |
1453 | break; | |
1454 | case SYMBOL_TLSLE32: | |
1455 | emit_insn ((mode == DImode ? gen_tlsle32_di : gen_tlsle32_si) | |
1456 | (dest, imm)); | |
1457 | emit_insn ((mode == DImode ? gen_adddi3 : gen_addsi3) | |
1458 | (dest, dest, tp)); | |
1459 | break; | |
1460 | case SYMBOL_TLSLE48: | |
1461 | emit_insn ((mode == DImode ? gen_tlsle48_di : gen_tlsle48_si) | |
1462 | (dest, imm)); | |
1463 | emit_insn ((mode == DImode ? gen_adddi3 : gen_addsi3) | |
1464 | (dest, dest, tp)); | |
1465 | break; | |
1466 | default: | |
1467 | gcc_unreachable (); | |
1468 | } | |
e6f7f0e9 | 1469 | |
43e9d192 IB |
1470 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); |
1471 | return; | |
1472 | } | |
1473 | ||
87dd8ab0 MS |
1474 | case SYMBOL_TINY_GOT: |
1475 | emit_insn (gen_ldr_got_tiny (dest, imm)); | |
1476 | return; | |
1477 | ||
5ae7caad JW |
1478 | case SYMBOL_TINY_TLSIE: |
1479 | { | |
1480 | machine_mode mode = GET_MODE (dest); | |
1481 | rtx tp = aarch64_load_tp (NULL); | |
1482 | ||
1483 | if (mode == ptr_mode) | |
1484 | { | |
1485 | if (mode == DImode) | |
1486 | emit_insn (gen_tlsie_tiny_di (dest, imm, tp)); | |
1487 | else | |
1488 | { | |
1489 | tp = gen_lowpart (mode, tp); | |
1490 | emit_insn (gen_tlsie_tiny_si (dest, imm, tp)); | |
1491 | } | |
1492 | } | |
1493 | else | |
1494 | { | |
1495 | gcc_assert (mode == Pmode); | |
1496 | emit_insn (gen_tlsie_tiny_sidi (dest, imm, tp)); | |
1497 | } | |
1498 | ||
1499 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); | |
1500 | return; | |
1501 | } | |
1502 | ||
43e9d192 IB |
1503 | default: |
1504 | gcc_unreachable (); | |
1505 | } | |
1506 | } | |
1507 | ||
1508 | /* Emit a move from SRC to DEST. Assume that the move expanders can | |
1509 | handle all moves if !can_create_pseudo_p (). The distinction is | |
1510 | important because, unlike emit_move_insn, the move expanders know | |
1511 | how to force Pmode objects into the constant pool even when the | |
1512 | constant pool address is not itself legitimate. */ | |
1513 | static rtx | |
1514 | aarch64_emit_move (rtx dest, rtx src) | |
1515 | { | |
1516 | return (can_create_pseudo_p () | |
1517 | ? emit_move_insn (dest, src) | |
1518 | : emit_move_insn_1 (dest, src)); | |
1519 | } | |
1520 | ||
030d03b8 RE |
1521 | /* Split a 128-bit move operation into two 64-bit move operations, |
1522 | taking care to handle partial overlap of register to register | |
1523 | copies. Special cases are needed when moving between GP regs and | |
1524 | FP regs. SRC can be a register, constant or memory; DST a register | |
1525 | or memory. If either operand is memory it must not have any side | |
1526 | effects. */ | |
43e9d192 IB |
1527 | void |
1528 | aarch64_split_128bit_move (rtx dst, rtx src) | |
1529 | { | |
030d03b8 RE |
1530 | rtx dst_lo, dst_hi; |
1531 | rtx src_lo, src_hi; | |
43e9d192 | 1532 | |
ef4bddc2 | 1533 | machine_mode mode = GET_MODE (dst); |
12dc6974 | 1534 | |
030d03b8 RE |
1535 | gcc_assert (mode == TImode || mode == TFmode); |
1536 | gcc_assert (!(side_effects_p (src) || side_effects_p (dst))); | |
1537 | gcc_assert (mode == GET_MODE (src) || GET_MODE (src) == VOIDmode); | |
43e9d192 IB |
1538 | |
1539 | if (REG_P (dst) && REG_P (src)) | |
1540 | { | |
030d03b8 RE |
1541 | int src_regno = REGNO (src); |
1542 | int dst_regno = REGNO (dst); | |
43e9d192 | 1543 | |
030d03b8 | 1544 | /* Handle FP <-> GP regs. */ |
43e9d192 IB |
1545 | if (FP_REGNUM_P (dst_regno) && GP_REGNUM_P (src_regno)) |
1546 | { | |
030d03b8 RE |
1547 | src_lo = gen_lowpart (word_mode, src); |
1548 | src_hi = gen_highpart (word_mode, src); | |
1549 | ||
1550 | if (mode == TImode) | |
1551 | { | |
1552 | emit_insn (gen_aarch64_movtilow_di (dst, src_lo)); | |
1553 | emit_insn (gen_aarch64_movtihigh_di (dst, src_hi)); | |
1554 | } | |
1555 | else | |
1556 | { | |
1557 | emit_insn (gen_aarch64_movtflow_di (dst, src_lo)); | |
1558 | emit_insn (gen_aarch64_movtfhigh_di (dst, src_hi)); | |
1559 | } | |
1560 | return; | |
43e9d192 IB |
1561 | } |
1562 | else if (GP_REGNUM_P (dst_regno) && FP_REGNUM_P (src_regno)) | |
1563 | { | |
030d03b8 RE |
1564 | dst_lo = gen_lowpart (word_mode, dst); |
1565 | dst_hi = gen_highpart (word_mode, dst); | |
1566 | ||
1567 | if (mode == TImode) | |
1568 | { | |
1569 | emit_insn (gen_aarch64_movdi_tilow (dst_lo, src)); | |
1570 | emit_insn (gen_aarch64_movdi_tihigh (dst_hi, src)); | |
1571 | } | |
1572 | else | |
1573 | { | |
1574 | emit_insn (gen_aarch64_movdi_tflow (dst_lo, src)); | |
1575 | emit_insn (gen_aarch64_movdi_tfhigh (dst_hi, src)); | |
1576 | } | |
1577 | return; | |
43e9d192 | 1578 | } |
43e9d192 IB |
1579 | } |
1580 | ||
030d03b8 RE |
1581 | dst_lo = gen_lowpart (word_mode, dst); |
1582 | dst_hi = gen_highpart (word_mode, dst); | |
1583 | src_lo = gen_lowpart (word_mode, src); | |
1584 | src_hi = gen_highpart_mode (word_mode, mode, src); | |
1585 | ||
1586 | /* At most one pairing may overlap. */ | |
1587 | if (reg_overlap_mentioned_p (dst_lo, src_hi)) | |
1588 | { | |
1589 | aarch64_emit_move (dst_hi, src_hi); | |
1590 | aarch64_emit_move (dst_lo, src_lo); | |
1591 | } | |
1592 | else | |
1593 | { | |
1594 | aarch64_emit_move (dst_lo, src_lo); | |
1595 | aarch64_emit_move (dst_hi, src_hi); | |
1596 | } | |
43e9d192 IB |
1597 | } |
1598 | ||
1599 | bool | |
1600 | aarch64_split_128bit_move_p (rtx dst, rtx src) | |
1601 | { | |
1602 | return (! REG_P (src) | |
1603 | || ! (FP_REGNUM_P (REGNO (dst)) && FP_REGNUM_P (REGNO (src)))); | |
1604 | } | |
1605 | ||
8b033a8a SN |
1606 | /* Split a complex SIMD combine. */ |
1607 | ||
1608 | void | |
1609 | aarch64_split_simd_combine (rtx dst, rtx src1, rtx src2) | |
1610 | { | |
ef4bddc2 RS |
1611 | machine_mode src_mode = GET_MODE (src1); |
1612 | machine_mode dst_mode = GET_MODE (dst); | |
8b033a8a SN |
1613 | |
1614 | gcc_assert (VECTOR_MODE_P (dst_mode)); | |
1615 | ||
1616 | if (REG_P (dst) && REG_P (src1) && REG_P (src2)) | |
1617 | { | |
1618 | rtx (*gen) (rtx, rtx, rtx); | |
1619 | ||
1620 | switch (src_mode) | |
1621 | { | |
1622 | case V8QImode: | |
1623 | gen = gen_aarch64_simd_combinev8qi; | |
1624 | break; | |
1625 | case V4HImode: | |
1626 | gen = gen_aarch64_simd_combinev4hi; | |
1627 | break; | |
1628 | case V2SImode: | |
1629 | gen = gen_aarch64_simd_combinev2si; | |
1630 | break; | |
7c369485 AL |
1631 | case V4HFmode: |
1632 | gen = gen_aarch64_simd_combinev4hf; | |
1633 | break; | |
8b033a8a SN |
1634 | case V2SFmode: |
1635 | gen = gen_aarch64_simd_combinev2sf; | |
1636 | break; | |
1637 | case DImode: | |
1638 | gen = gen_aarch64_simd_combinedi; | |
1639 | break; | |
1640 | case DFmode: | |
1641 | gen = gen_aarch64_simd_combinedf; | |
1642 | break; | |
1643 | default: | |
1644 | gcc_unreachable (); | |
1645 | } | |
1646 | ||
1647 | emit_insn (gen (dst, src1, src2)); | |
1648 | return; | |
1649 | } | |
1650 | } | |
1651 | ||
fd4842cd SN |
1652 | /* Split a complex SIMD move. */ |
1653 | ||
1654 | void | |
1655 | aarch64_split_simd_move (rtx dst, rtx src) | |
1656 | { | |
ef4bddc2 RS |
1657 | machine_mode src_mode = GET_MODE (src); |
1658 | machine_mode dst_mode = GET_MODE (dst); | |
fd4842cd SN |
1659 | |
1660 | gcc_assert (VECTOR_MODE_P (dst_mode)); | |
1661 | ||
1662 | if (REG_P (dst) && REG_P (src)) | |
1663 | { | |
c59b7e28 SN |
1664 | rtx (*gen) (rtx, rtx); |
1665 | ||
fd4842cd SN |
1666 | gcc_assert (VECTOR_MODE_P (src_mode)); |
1667 | ||
1668 | switch (src_mode) | |
1669 | { | |
1670 | case V16QImode: | |
c59b7e28 | 1671 | gen = gen_aarch64_split_simd_movv16qi; |
fd4842cd SN |
1672 | break; |
1673 | case V8HImode: | |
c59b7e28 | 1674 | gen = gen_aarch64_split_simd_movv8hi; |
fd4842cd SN |
1675 | break; |
1676 | case V4SImode: | |
c59b7e28 | 1677 | gen = gen_aarch64_split_simd_movv4si; |
fd4842cd SN |
1678 | break; |
1679 | case V2DImode: | |
c59b7e28 | 1680 | gen = gen_aarch64_split_simd_movv2di; |
fd4842cd | 1681 | break; |
71a11456 AL |
1682 | case V8HFmode: |
1683 | gen = gen_aarch64_split_simd_movv8hf; | |
1684 | break; | |
fd4842cd | 1685 | case V4SFmode: |
c59b7e28 | 1686 | gen = gen_aarch64_split_simd_movv4sf; |
fd4842cd SN |
1687 | break; |
1688 | case V2DFmode: | |
c59b7e28 | 1689 | gen = gen_aarch64_split_simd_movv2df; |
fd4842cd SN |
1690 | break; |
1691 | default: | |
1692 | gcc_unreachable (); | |
1693 | } | |
c59b7e28 SN |
1694 | |
1695 | emit_insn (gen (dst, src)); | |
fd4842cd SN |
1696 | return; |
1697 | } | |
1698 | } | |
1699 | ||
ef22810a RH |
1700 | bool |
1701 | aarch64_zero_extend_const_eq (machine_mode xmode, rtx x, | |
1702 | machine_mode ymode, rtx y) | |
1703 | { | |
1704 | rtx r = simplify_const_unary_operation (ZERO_EXTEND, xmode, y, ymode); | |
1705 | gcc_assert (r != NULL); | |
1706 | return rtx_equal_p (x, r); | |
1707 | } | |
1708 | ||
1709 | ||
43e9d192 | 1710 | static rtx |
ef4bddc2 | 1711 | aarch64_force_temporary (machine_mode mode, rtx x, rtx value) |
43e9d192 IB |
1712 | { |
1713 | if (can_create_pseudo_p ()) | |
e18b4a81 | 1714 | return force_reg (mode, value); |
43e9d192 IB |
1715 | else |
1716 | { | |
1717 | x = aarch64_emit_move (x, value); | |
1718 | return x; | |
1719 | } | |
1720 | } | |
1721 | ||
1722 | ||
1723 | static rtx | |
ef4bddc2 | 1724 | aarch64_add_offset (machine_mode mode, rtx temp, rtx reg, HOST_WIDE_INT offset) |
43e9d192 | 1725 | { |
9c023bf0 | 1726 | if (!aarch64_plus_immediate (GEN_INT (offset), mode)) |
43e9d192 IB |
1727 | { |
1728 | rtx high; | |
1729 | /* Load the full offset into a register. This | |
1730 | might be improvable in the future. */ | |
1731 | high = GEN_INT (offset); | |
1732 | offset = 0; | |
e18b4a81 YZ |
1733 | high = aarch64_force_temporary (mode, temp, high); |
1734 | reg = aarch64_force_temporary (mode, temp, | |
1735 | gen_rtx_PLUS (mode, high, reg)); | |
43e9d192 IB |
1736 | } |
1737 | return plus_constant (mode, reg, offset); | |
1738 | } | |
1739 | ||
82614948 RR |
1740 | static int |
1741 | aarch64_internal_mov_immediate (rtx dest, rtx imm, bool generate, | |
1742 | machine_mode mode) | |
43e9d192 | 1743 | { |
43e9d192 | 1744 | int i; |
9a4865db WD |
1745 | unsigned HOST_WIDE_INT val, val2, mask; |
1746 | int one_match, zero_match; | |
1747 | int num_insns; | |
43e9d192 | 1748 | |
9a4865db WD |
1749 | val = INTVAL (imm); |
1750 | ||
1751 | if (aarch64_move_imm (val, mode)) | |
43e9d192 | 1752 | { |
82614948 | 1753 | if (generate) |
f7df4a84 | 1754 | emit_insn (gen_rtx_SET (dest, imm)); |
9a4865db | 1755 | return 1; |
43e9d192 IB |
1756 | } |
1757 | ||
9a4865db | 1758 | if ((val >> 32) == 0 || mode == SImode) |
43e9d192 | 1759 | { |
82614948 RR |
1760 | if (generate) |
1761 | { | |
9a4865db WD |
1762 | emit_insn (gen_rtx_SET (dest, GEN_INT (val & 0xffff))); |
1763 | if (mode == SImode) | |
1764 | emit_insn (gen_insv_immsi (dest, GEN_INT (16), | |
1765 | GEN_INT ((val >> 16) & 0xffff))); | |
1766 | else | |
1767 | emit_insn (gen_insv_immdi (dest, GEN_INT (16), | |
1768 | GEN_INT ((val >> 16) & 0xffff))); | |
82614948 | 1769 | } |
9a4865db | 1770 | return 2; |
43e9d192 IB |
1771 | } |
1772 | ||
1773 | /* Remaining cases are all for DImode. */ | |
1774 | ||
43e9d192 | 1775 | mask = 0xffff; |
9a4865db WD |
1776 | zero_match = ((val & mask) == 0) + ((val & (mask << 16)) == 0) + |
1777 | ((val & (mask << 32)) == 0) + ((val & (mask << 48)) == 0); | |
1778 | one_match = ((~val & mask) == 0) + ((~val & (mask << 16)) == 0) + | |
1779 | ((~val & (mask << 32)) == 0) + ((~val & (mask << 48)) == 0); | |
43e9d192 | 1780 | |
62c8d76c | 1781 | if (zero_match != 2 && one_match != 2) |
43e9d192 | 1782 | { |
62c8d76c WD |
1783 | /* Try emitting a bitmask immediate with a movk replacing 16 bits. |
1784 | For a 64-bit bitmask try whether changing 16 bits to all ones or | |
1785 | zeroes creates a valid bitmask. To check any repeated bitmask, | |
1786 | try using 16 bits from the other 32-bit half of val. */ | |
43e9d192 | 1787 | |
62c8d76c | 1788 | for (i = 0; i < 64; i += 16, mask <<= 16) |
43e9d192 | 1789 | { |
62c8d76c WD |
1790 | val2 = val & ~mask; |
1791 | if (val2 != val && aarch64_bitmask_imm (val2, mode)) | |
1792 | break; | |
1793 | val2 = val | mask; | |
1794 | if (val2 != val && aarch64_bitmask_imm (val2, mode)) | |
1795 | break; | |
1796 | val2 = val2 & ~mask; | |
1797 | val2 = val2 | (((val2 >> 32) | (val2 << 32)) & mask); | |
1798 | if (val2 != val && aarch64_bitmask_imm (val2, mode)) | |
1799 | break; | |
43e9d192 | 1800 | } |
62c8d76c | 1801 | if (i != 64) |
43e9d192 | 1802 | { |
62c8d76c | 1803 | if (generate) |
43e9d192 | 1804 | { |
62c8d76c WD |
1805 | emit_insn (gen_rtx_SET (dest, GEN_INT (val2))); |
1806 | emit_insn (gen_insv_immdi (dest, GEN_INT (i), | |
9a4865db | 1807 | GEN_INT ((val >> i) & 0xffff))); |
43e9d192 | 1808 | } |
1312b1ba | 1809 | return 2; |
43e9d192 IB |
1810 | } |
1811 | } | |
1812 | ||
9a4865db WD |
1813 | /* Generate 2-4 instructions, skipping 16 bits of all zeroes or ones which |
1814 | are emitted by the initial mov. If one_match > zero_match, skip set bits, | |
1815 | otherwise skip zero bits. */ | |
2c274197 | 1816 | |
9a4865db | 1817 | num_insns = 1; |
43e9d192 | 1818 | mask = 0xffff; |
9a4865db WD |
1819 | val2 = one_match > zero_match ? ~val : val; |
1820 | i = (val2 & mask) != 0 ? 0 : (val2 & (mask << 16)) != 0 ? 16 : 32; | |
1821 | ||
1822 | if (generate) | |
1823 | emit_insn (gen_rtx_SET (dest, GEN_INT (one_match > zero_match | |
1824 | ? (val | ~(mask << i)) | |
1825 | : (val & (mask << i))))); | |
1826 | for (i += 16; i < 64; i += 16) | |
43e9d192 | 1827 | { |
9a4865db WD |
1828 | if ((val2 & (mask << i)) == 0) |
1829 | continue; | |
1830 | if (generate) | |
1831 | emit_insn (gen_insv_immdi (dest, GEN_INT (i), | |
1832 | GEN_INT ((val >> i) & 0xffff))); | |
1833 | num_insns ++; | |
82614948 RR |
1834 | } |
1835 | ||
1836 | return num_insns; | |
1837 | } | |
1838 | ||
1839 | ||
1840 | void | |
1841 | aarch64_expand_mov_immediate (rtx dest, rtx imm) | |
1842 | { | |
1843 | machine_mode mode = GET_MODE (dest); | |
1844 | ||
1845 | gcc_assert (mode == SImode || mode == DImode); | |
1846 | ||
1847 | /* Check on what type of symbol it is. */ | |
1848 | if (GET_CODE (imm) == SYMBOL_REF | |
1849 | || GET_CODE (imm) == LABEL_REF | |
1850 | || GET_CODE (imm) == CONST) | |
1851 | { | |
1852 | rtx mem, base, offset; | |
1853 | enum aarch64_symbol_type sty; | |
1854 | ||
1855 | /* If we have (const (plus symbol offset)), separate out the offset | |
1856 | before we start classifying the symbol. */ | |
1857 | split_const (imm, &base, &offset); | |
1858 | ||
a6e0bfa7 | 1859 | sty = aarch64_classify_symbol (base, offset); |
82614948 RR |
1860 | switch (sty) |
1861 | { | |
1862 | case SYMBOL_FORCE_TO_MEM: | |
1863 | if (offset != const0_rtx | |
1864 | && targetm.cannot_force_const_mem (mode, imm)) | |
1865 | { | |
1866 | gcc_assert (can_create_pseudo_p ()); | |
1867 | base = aarch64_force_temporary (mode, dest, base); | |
1868 | base = aarch64_add_offset (mode, NULL, base, INTVAL (offset)); | |
1869 | aarch64_emit_move (dest, base); | |
1870 | return; | |
1871 | } | |
b4f50fd4 | 1872 | |
82614948 RR |
1873 | mem = force_const_mem (ptr_mode, imm); |
1874 | gcc_assert (mem); | |
b4f50fd4 RR |
1875 | |
1876 | /* If we aren't generating PC relative literals, then | |
1877 | we need to expand the literal pool access carefully. | |
1878 | This is something that needs to be done in a number | |
1879 | of places, so could well live as a separate function. */ | |
9ee6540a | 1880 | if (!aarch64_pcrelative_literal_loads) |
b4f50fd4 RR |
1881 | { |
1882 | gcc_assert (can_create_pseudo_p ()); | |
1883 | base = gen_reg_rtx (ptr_mode); | |
1884 | aarch64_expand_mov_immediate (base, XEXP (mem, 0)); | |
1885 | mem = gen_rtx_MEM (ptr_mode, base); | |
1886 | } | |
1887 | ||
82614948 RR |
1888 | if (mode != ptr_mode) |
1889 | mem = gen_rtx_ZERO_EXTEND (mode, mem); | |
b4f50fd4 | 1890 | |
f7df4a84 | 1891 | emit_insn (gen_rtx_SET (dest, mem)); |
b4f50fd4 | 1892 | |
82614948 RR |
1893 | return; |
1894 | ||
1895 | case SYMBOL_SMALL_TLSGD: | |
1896 | case SYMBOL_SMALL_TLSDESC: | |
79496620 | 1897 | case SYMBOL_SMALL_TLSIE: |
1b1e81f8 | 1898 | case SYMBOL_SMALL_GOT_28K: |
6642bdb4 | 1899 | case SYMBOL_SMALL_GOT_4G: |
82614948 | 1900 | case SYMBOL_TINY_GOT: |
5ae7caad | 1901 | case SYMBOL_TINY_TLSIE: |
82614948 RR |
1902 | if (offset != const0_rtx) |
1903 | { | |
1904 | gcc_assert(can_create_pseudo_p ()); | |
1905 | base = aarch64_force_temporary (mode, dest, base); | |
1906 | base = aarch64_add_offset (mode, NULL, base, INTVAL (offset)); | |
1907 | aarch64_emit_move (dest, base); | |
1908 | return; | |
1909 | } | |
1910 | /* FALLTHRU */ | |
1911 | ||
82614948 RR |
1912 | case SYMBOL_SMALL_ABSOLUTE: |
1913 | case SYMBOL_TINY_ABSOLUTE: | |
cbf5629e | 1914 | case SYMBOL_TLSLE12: |
d18ba284 | 1915 | case SYMBOL_TLSLE24: |
cbf5629e JW |
1916 | case SYMBOL_TLSLE32: |
1917 | case SYMBOL_TLSLE48: | |
82614948 RR |
1918 | aarch64_load_symref_appropriately (dest, imm, sty); |
1919 | return; | |
1920 | ||
1921 | default: | |
1922 | gcc_unreachable (); | |
1923 | } | |
1924 | } | |
1925 | ||
1926 | if (!CONST_INT_P (imm)) | |
1927 | { | |
1928 | if (GET_CODE (imm) == HIGH) | |
f7df4a84 | 1929 | emit_insn (gen_rtx_SET (dest, imm)); |
82614948 RR |
1930 | else |
1931 | { | |
1932 | rtx mem = force_const_mem (mode, imm); | |
1933 | gcc_assert (mem); | |
f7df4a84 | 1934 | emit_insn (gen_rtx_SET (dest, mem)); |
43e9d192 | 1935 | } |
82614948 RR |
1936 | |
1937 | return; | |
43e9d192 | 1938 | } |
82614948 RR |
1939 | |
1940 | aarch64_internal_mov_immediate (dest, imm, true, GET_MODE (dest)); | |
43e9d192 IB |
1941 | } |
1942 | ||
1943 | static bool | |
fee9ba42 JW |
1944 | aarch64_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED, |
1945 | tree exp ATTRIBUTE_UNUSED) | |
43e9d192 | 1946 | { |
fee9ba42 | 1947 | /* Currently, always true. */ |
43e9d192 IB |
1948 | return true; |
1949 | } | |
1950 | ||
1951 | /* Implement TARGET_PASS_BY_REFERENCE. */ | |
1952 | ||
1953 | static bool | |
1954 | aarch64_pass_by_reference (cumulative_args_t pcum ATTRIBUTE_UNUSED, | |
ef4bddc2 | 1955 | machine_mode mode, |
43e9d192 IB |
1956 | const_tree type, |
1957 | bool named ATTRIBUTE_UNUSED) | |
1958 | { | |
1959 | HOST_WIDE_INT size; | |
ef4bddc2 | 1960 | machine_mode dummymode; |
43e9d192 IB |
1961 | int nregs; |
1962 | ||
1963 | /* GET_MODE_SIZE (BLKmode) is useless since it is 0. */ | |
1964 | size = (mode == BLKmode && type) | |
1965 | ? int_size_in_bytes (type) : (int) GET_MODE_SIZE (mode); | |
1966 | ||
aadc1c43 MHD |
1967 | /* Aggregates are passed by reference based on their size. */ |
1968 | if (type && AGGREGATE_TYPE_P (type)) | |
43e9d192 | 1969 | { |
aadc1c43 | 1970 | size = int_size_in_bytes (type); |
43e9d192 IB |
1971 | } |
1972 | ||
1973 | /* Variable sized arguments are always returned by reference. */ | |
1974 | if (size < 0) | |
1975 | return true; | |
1976 | ||
1977 | /* Can this be a candidate to be passed in fp/simd register(s)? */ | |
1978 | if (aarch64_vfp_is_call_or_return_candidate (mode, type, | |
1979 | &dummymode, &nregs, | |
1980 | NULL)) | |
1981 | return false; | |
1982 | ||
1983 | /* Arguments which are variable sized or larger than 2 registers are | |
1984 | passed by reference unless they are a homogenous floating point | |
1985 | aggregate. */ | |
1986 | return size > 2 * UNITS_PER_WORD; | |
1987 | } | |
1988 | ||
1989 | /* Return TRUE if VALTYPE is padded to its least significant bits. */ | |
1990 | static bool | |
1991 | aarch64_return_in_msb (const_tree valtype) | |
1992 | { | |
ef4bddc2 | 1993 | machine_mode dummy_mode; |
43e9d192 IB |
1994 | int dummy_int; |
1995 | ||
1996 | /* Never happens in little-endian mode. */ | |
1997 | if (!BYTES_BIG_ENDIAN) | |
1998 | return false; | |
1999 | ||
2000 | /* Only composite types smaller than or equal to 16 bytes can | |
2001 | be potentially returned in registers. */ | |
2002 | if (!aarch64_composite_type_p (valtype, TYPE_MODE (valtype)) | |
2003 | || int_size_in_bytes (valtype) <= 0 | |
2004 | || int_size_in_bytes (valtype) > 16) | |
2005 | return false; | |
2006 | ||
2007 | /* But not a composite that is an HFA (Homogeneous Floating-point Aggregate) | |
2008 | or an HVA (Homogeneous Short-Vector Aggregate); such a special composite | |
2009 | is always passed/returned in the least significant bits of fp/simd | |
2010 | register(s). */ | |
2011 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (valtype), valtype, | |
2012 | &dummy_mode, &dummy_int, NULL)) | |
2013 | return false; | |
2014 | ||
2015 | return true; | |
2016 | } | |
2017 | ||
2018 | /* Implement TARGET_FUNCTION_VALUE. | |
2019 | Define how to find the value returned by a function. */ | |
2020 | ||
2021 | static rtx | |
2022 | aarch64_function_value (const_tree type, const_tree func, | |
2023 | bool outgoing ATTRIBUTE_UNUSED) | |
2024 | { | |
ef4bddc2 | 2025 | machine_mode mode; |
43e9d192 IB |
2026 | int unsignedp; |
2027 | int count; | |
ef4bddc2 | 2028 | machine_mode ag_mode; |
43e9d192 IB |
2029 | |
2030 | mode = TYPE_MODE (type); | |
2031 | if (INTEGRAL_TYPE_P (type)) | |
2032 | mode = promote_function_mode (type, mode, &unsignedp, func, 1); | |
2033 | ||
2034 | if (aarch64_return_in_msb (type)) | |
2035 | { | |
2036 | HOST_WIDE_INT size = int_size_in_bytes (type); | |
2037 | ||
2038 | if (size % UNITS_PER_WORD != 0) | |
2039 | { | |
2040 | size += UNITS_PER_WORD - size % UNITS_PER_WORD; | |
2041 | mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0); | |
2042 | } | |
2043 | } | |
2044 | ||
2045 | if (aarch64_vfp_is_call_or_return_candidate (mode, type, | |
2046 | &ag_mode, &count, NULL)) | |
2047 | { | |
2048 | if (!aarch64_composite_type_p (type, mode)) | |
2049 | { | |
2050 | gcc_assert (count == 1 && mode == ag_mode); | |
2051 | return gen_rtx_REG (mode, V0_REGNUM); | |
2052 | } | |
2053 | else | |
2054 | { | |
2055 | int i; | |
2056 | rtx par; | |
2057 | ||
2058 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (count)); | |
2059 | for (i = 0; i < count; i++) | |
2060 | { | |
2061 | rtx tmp = gen_rtx_REG (ag_mode, V0_REGNUM + i); | |
2062 | tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp, | |
2063 | GEN_INT (i * GET_MODE_SIZE (ag_mode))); | |
2064 | XVECEXP (par, 0, i) = tmp; | |
2065 | } | |
2066 | return par; | |
2067 | } | |
2068 | } | |
2069 | else | |
2070 | return gen_rtx_REG (mode, R0_REGNUM); | |
2071 | } | |
2072 | ||
2073 | /* Implements TARGET_FUNCTION_VALUE_REGNO_P. | |
2074 | Return true if REGNO is the number of a hard register in which the values | |
2075 | of called function may come back. */ | |
2076 | ||
2077 | static bool | |
2078 | aarch64_function_value_regno_p (const unsigned int regno) | |
2079 | { | |
2080 | /* Maximum of 16 bytes can be returned in the general registers. Examples | |
2081 | of 16-byte return values are: 128-bit integers and 16-byte small | |
2082 | structures (excluding homogeneous floating-point aggregates). */ | |
2083 | if (regno == R0_REGNUM || regno == R1_REGNUM) | |
2084 | return true; | |
2085 | ||
2086 | /* Up to four fp/simd registers can return a function value, e.g. a | |
2087 | homogeneous floating-point aggregate having four members. */ | |
2088 | if (regno >= V0_REGNUM && regno < V0_REGNUM + HA_MAX_NUM_FLDS) | |
d5726973 | 2089 | return TARGET_FLOAT; |
43e9d192 IB |
2090 | |
2091 | return false; | |
2092 | } | |
2093 | ||
2094 | /* Implement TARGET_RETURN_IN_MEMORY. | |
2095 | ||
2096 | If the type T of the result of a function is such that | |
2097 | void func (T arg) | |
2098 | would require that arg be passed as a value in a register (or set of | |
2099 | registers) according to the parameter passing rules, then the result | |
2100 | is returned in the same registers as would be used for such an | |
2101 | argument. */ | |
2102 | ||
2103 | static bool | |
2104 | aarch64_return_in_memory (const_tree type, const_tree fndecl ATTRIBUTE_UNUSED) | |
2105 | { | |
2106 | HOST_WIDE_INT size; | |
ef4bddc2 | 2107 | machine_mode ag_mode; |
43e9d192 IB |
2108 | int count; |
2109 | ||
2110 | if (!AGGREGATE_TYPE_P (type) | |
2111 | && TREE_CODE (type) != COMPLEX_TYPE | |
2112 | && TREE_CODE (type) != VECTOR_TYPE) | |
2113 | /* Simple scalar types always returned in registers. */ | |
2114 | return false; | |
2115 | ||
2116 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (type), | |
2117 | type, | |
2118 | &ag_mode, | |
2119 | &count, | |
2120 | NULL)) | |
2121 | return false; | |
2122 | ||
2123 | /* Types larger than 2 registers returned in memory. */ | |
2124 | size = int_size_in_bytes (type); | |
2125 | return (size < 0 || size > 2 * UNITS_PER_WORD); | |
2126 | } | |
2127 | ||
2128 | static bool | |
ef4bddc2 | 2129 | aarch64_vfp_is_call_candidate (cumulative_args_t pcum_v, machine_mode mode, |
43e9d192 IB |
2130 | const_tree type, int *nregs) |
2131 | { | |
2132 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
2133 | return aarch64_vfp_is_call_or_return_candidate (mode, | |
2134 | type, | |
2135 | &pcum->aapcs_vfp_rmode, | |
2136 | nregs, | |
2137 | NULL); | |
2138 | } | |
2139 | ||
2140 | /* Given MODE and TYPE of a function argument, return the alignment in | |
2141 | bits. The idea is to suppress any stronger alignment requested by | |
2142 | the user and opt for the natural alignment (specified in AAPCS64 \S 4.1). | |
2143 | This is a helper function for local use only. */ | |
2144 | ||
2145 | static unsigned int | |
ef4bddc2 | 2146 | aarch64_function_arg_alignment (machine_mode mode, const_tree type) |
43e9d192 | 2147 | { |
75d6cc81 AL |
2148 | if (!type) |
2149 | return GET_MODE_ALIGNMENT (mode); | |
2150 | if (integer_zerop (TYPE_SIZE (type))) | |
2151 | return 0; | |
43e9d192 | 2152 | |
75d6cc81 AL |
2153 | gcc_assert (TYPE_MODE (type) == mode); |
2154 | ||
2155 | if (!AGGREGATE_TYPE_P (type)) | |
2156 | return TYPE_ALIGN (TYPE_MAIN_VARIANT (type)); | |
2157 | ||
2158 | if (TREE_CODE (type) == ARRAY_TYPE) | |
2159 | return TYPE_ALIGN (TREE_TYPE (type)); | |
2160 | ||
2161 | unsigned int alignment = 0; | |
2162 | ||
2163 | for (tree field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field)) | |
2164 | alignment = std::max (alignment, DECL_ALIGN (field)); | |
43e9d192 IB |
2165 | |
2166 | return alignment; | |
2167 | } | |
2168 | ||
2169 | /* Layout a function argument according to the AAPCS64 rules. The rule | |
2170 | numbers refer to the rule numbers in the AAPCS64. */ | |
2171 | ||
2172 | static void | |
ef4bddc2 | 2173 | aarch64_layout_arg (cumulative_args_t pcum_v, machine_mode mode, |
43e9d192 IB |
2174 | const_tree type, |
2175 | bool named ATTRIBUTE_UNUSED) | |
2176 | { | |
2177 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
2178 | int ncrn, nvrn, nregs; | |
2179 | bool allocate_ncrn, allocate_nvrn; | |
3abf17cf | 2180 | HOST_WIDE_INT size; |
43e9d192 IB |
2181 | |
2182 | /* We need to do this once per argument. */ | |
2183 | if (pcum->aapcs_arg_processed) | |
2184 | return; | |
2185 | ||
2186 | pcum->aapcs_arg_processed = true; | |
2187 | ||
3abf17cf YZ |
2188 | /* Size in bytes, rounded to the nearest multiple of 8 bytes. */ |
2189 | size | |
4f59f9f2 UB |
2190 | = ROUND_UP (type ? int_size_in_bytes (type) : GET_MODE_SIZE (mode), |
2191 | UNITS_PER_WORD); | |
3abf17cf | 2192 | |
43e9d192 IB |
2193 | allocate_ncrn = (type) ? !(FLOAT_TYPE_P (type)) : !FLOAT_MODE_P (mode); |
2194 | allocate_nvrn = aarch64_vfp_is_call_candidate (pcum_v, | |
2195 | mode, | |
2196 | type, | |
2197 | &nregs); | |
2198 | ||
2199 | /* allocate_ncrn may be false-positive, but allocate_nvrn is quite reliable. | |
2200 | The following code thus handles passing by SIMD/FP registers first. */ | |
2201 | ||
2202 | nvrn = pcum->aapcs_nvrn; | |
2203 | ||
2204 | /* C1 - C5 for floating point, homogenous floating point aggregates (HFA) | |
2205 | and homogenous short-vector aggregates (HVA). */ | |
2206 | if (allocate_nvrn) | |
2207 | { | |
261fb553 AL |
2208 | if (!TARGET_FLOAT) |
2209 | aarch64_err_no_fpadvsimd (mode, "argument"); | |
2210 | ||
43e9d192 IB |
2211 | if (nvrn + nregs <= NUM_FP_ARG_REGS) |
2212 | { | |
2213 | pcum->aapcs_nextnvrn = nvrn + nregs; | |
2214 | if (!aarch64_composite_type_p (type, mode)) | |
2215 | { | |
2216 | gcc_assert (nregs == 1); | |
2217 | pcum->aapcs_reg = gen_rtx_REG (mode, V0_REGNUM + nvrn); | |
2218 | } | |
2219 | else | |
2220 | { | |
2221 | rtx par; | |
2222 | int i; | |
2223 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (nregs)); | |
2224 | for (i = 0; i < nregs; i++) | |
2225 | { | |
2226 | rtx tmp = gen_rtx_REG (pcum->aapcs_vfp_rmode, | |
2227 | V0_REGNUM + nvrn + i); | |
2228 | tmp = gen_rtx_EXPR_LIST | |
2229 | (VOIDmode, tmp, | |
2230 | GEN_INT (i * GET_MODE_SIZE (pcum->aapcs_vfp_rmode))); | |
2231 | XVECEXP (par, 0, i) = tmp; | |
2232 | } | |
2233 | pcum->aapcs_reg = par; | |
2234 | } | |
2235 | return; | |
2236 | } | |
2237 | else | |
2238 | { | |
2239 | /* C.3 NSRN is set to 8. */ | |
2240 | pcum->aapcs_nextnvrn = NUM_FP_ARG_REGS; | |
2241 | goto on_stack; | |
2242 | } | |
2243 | } | |
2244 | ||
2245 | ncrn = pcum->aapcs_ncrn; | |
3abf17cf | 2246 | nregs = size / UNITS_PER_WORD; |
43e9d192 IB |
2247 | |
2248 | /* C6 - C9. though the sign and zero extension semantics are | |
2249 | handled elsewhere. This is the case where the argument fits | |
2250 | entirely general registers. */ | |
2251 | if (allocate_ncrn && (ncrn + nregs <= NUM_ARG_REGS)) | |
2252 | { | |
2253 | unsigned int alignment = aarch64_function_arg_alignment (mode, type); | |
2254 | ||
2255 | gcc_assert (nregs == 0 || nregs == 1 || nregs == 2); | |
2256 | ||
2257 | /* C.8 if the argument has an alignment of 16 then the NGRN is | |
2258 | rounded up to the next even number. */ | |
2259 | if (nregs == 2 && alignment == 16 * BITS_PER_UNIT && ncrn % 2) | |
2260 | { | |
2261 | ++ncrn; | |
2262 | gcc_assert (ncrn + nregs <= NUM_ARG_REGS); | |
2263 | } | |
2264 | /* NREGS can be 0 when e.g. an empty structure is to be passed. | |
2265 | A reg is still generated for it, but the caller should be smart | |
2266 | enough not to use it. */ | |
2267 | if (nregs == 0 || nregs == 1 || GET_MODE_CLASS (mode) == MODE_INT) | |
2268 | { | |
2269 | pcum->aapcs_reg = gen_rtx_REG (mode, R0_REGNUM + ncrn); | |
2270 | } | |
2271 | else | |
2272 | { | |
2273 | rtx par; | |
2274 | int i; | |
2275 | ||
2276 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (nregs)); | |
2277 | for (i = 0; i < nregs; i++) | |
2278 | { | |
2279 | rtx tmp = gen_rtx_REG (word_mode, R0_REGNUM + ncrn + i); | |
2280 | tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp, | |
2281 | GEN_INT (i * UNITS_PER_WORD)); | |
2282 | XVECEXP (par, 0, i) = tmp; | |
2283 | } | |
2284 | pcum->aapcs_reg = par; | |
2285 | } | |
2286 | ||
2287 | pcum->aapcs_nextncrn = ncrn + nregs; | |
2288 | return; | |
2289 | } | |
2290 | ||
2291 | /* C.11 */ | |
2292 | pcum->aapcs_nextncrn = NUM_ARG_REGS; | |
2293 | ||
2294 | /* The argument is passed on stack; record the needed number of words for | |
3abf17cf | 2295 | this argument and align the total size if necessary. */ |
43e9d192 | 2296 | on_stack: |
3abf17cf | 2297 | pcum->aapcs_stack_words = size / UNITS_PER_WORD; |
43e9d192 | 2298 | if (aarch64_function_arg_alignment (mode, type) == 16 * BITS_PER_UNIT) |
4f59f9f2 UB |
2299 | pcum->aapcs_stack_size = ROUND_UP (pcum->aapcs_stack_size, |
2300 | 16 / UNITS_PER_WORD); | |
43e9d192 IB |
2301 | return; |
2302 | } | |
2303 | ||
2304 | /* Implement TARGET_FUNCTION_ARG. */ | |
2305 | ||
2306 | static rtx | |
ef4bddc2 | 2307 | aarch64_function_arg (cumulative_args_t pcum_v, machine_mode mode, |
43e9d192 IB |
2308 | const_tree type, bool named) |
2309 | { | |
2310 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
2311 | gcc_assert (pcum->pcs_variant == ARM_PCS_AAPCS64); | |
2312 | ||
2313 | if (mode == VOIDmode) | |
2314 | return NULL_RTX; | |
2315 | ||
2316 | aarch64_layout_arg (pcum_v, mode, type, named); | |
2317 | return pcum->aapcs_reg; | |
2318 | } | |
2319 | ||
2320 | void | |
2321 | aarch64_init_cumulative_args (CUMULATIVE_ARGS *pcum, | |
2322 | const_tree fntype ATTRIBUTE_UNUSED, | |
2323 | rtx libname ATTRIBUTE_UNUSED, | |
2324 | const_tree fndecl ATTRIBUTE_UNUSED, | |
2325 | unsigned n_named ATTRIBUTE_UNUSED) | |
2326 | { | |
2327 | pcum->aapcs_ncrn = 0; | |
2328 | pcum->aapcs_nvrn = 0; | |
2329 | pcum->aapcs_nextncrn = 0; | |
2330 | pcum->aapcs_nextnvrn = 0; | |
2331 | pcum->pcs_variant = ARM_PCS_AAPCS64; | |
2332 | pcum->aapcs_reg = NULL_RTX; | |
2333 | pcum->aapcs_arg_processed = false; | |
2334 | pcum->aapcs_stack_words = 0; | |
2335 | pcum->aapcs_stack_size = 0; | |
2336 | ||
261fb553 AL |
2337 | if (!TARGET_FLOAT |
2338 | && fndecl && TREE_PUBLIC (fndecl) | |
2339 | && fntype && fntype != error_mark_node) | |
2340 | { | |
2341 | const_tree type = TREE_TYPE (fntype); | |
2342 | machine_mode mode ATTRIBUTE_UNUSED; /* To pass pointer as argument. */ | |
2343 | int nregs ATTRIBUTE_UNUSED; /* Likewise. */ | |
2344 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (type), type, | |
2345 | &mode, &nregs, NULL)) | |
2346 | aarch64_err_no_fpadvsimd (TYPE_MODE (type), "return type"); | |
2347 | } | |
43e9d192 IB |
2348 | return; |
2349 | } | |
2350 | ||
2351 | static void | |
2352 | aarch64_function_arg_advance (cumulative_args_t pcum_v, | |
ef4bddc2 | 2353 | machine_mode mode, |
43e9d192 IB |
2354 | const_tree type, |
2355 | bool named) | |
2356 | { | |
2357 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
2358 | if (pcum->pcs_variant == ARM_PCS_AAPCS64) | |
2359 | { | |
2360 | aarch64_layout_arg (pcum_v, mode, type, named); | |
2361 | gcc_assert ((pcum->aapcs_reg != NULL_RTX) | |
2362 | != (pcum->aapcs_stack_words != 0)); | |
2363 | pcum->aapcs_arg_processed = false; | |
2364 | pcum->aapcs_ncrn = pcum->aapcs_nextncrn; | |
2365 | pcum->aapcs_nvrn = pcum->aapcs_nextnvrn; | |
2366 | pcum->aapcs_stack_size += pcum->aapcs_stack_words; | |
2367 | pcum->aapcs_stack_words = 0; | |
2368 | pcum->aapcs_reg = NULL_RTX; | |
2369 | } | |
2370 | } | |
2371 | ||
2372 | bool | |
2373 | aarch64_function_arg_regno_p (unsigned regno) | |
2374 | { | |
2375 | return ((GP_REGNUM_P (regno) && regno < R0_REGNUM + NUM_ARG_REGS) | |
2376 | || (FP_REGNUM_P (regno) && regno < V0_REGNUM + NUM_FP_ARG_REGS)); | |
2377 | } | |
2378 | ||
2379 | /* Implement FUNCTION_ARG_BOUNDARY. Every parameter gets at least | |
2380 | PARM_BOUNDARY bits of alignment, but will be given anything up | |
2381 | to STACK_BOUNDARY bits if the type requires it. This makes sure | |
2382 | that both before and after the layout of each argument, the Next | |
2383 | Stacked Argument Address (NSAA) will have a minimum alignment of | |
2384 | 8 bytes. */ | |
2385 | ||
2386 | static unsigned int | |
ef4bddc2 | 2387 | aarch64_function_arg_boundary (machine_mode mode, const_tree type) |
43e9d192 IB |
2388 | { |
2389 | unsigned int alignment = aarch64_function_arg_alignment (mode, type); | |
2390 | ||
2391 | if (alignment < PARM_BOUNDARY) | |
2392 | alignment = PARM_BOUNDARY; | |
2393 | if (alignment > STACK_BOUNDARY) | |
2394 | alignment = STACK_BOUNDARY; | |
2395 | return alignment; | |
2396 | } | |
2397 | ||
2398 | /* For use by FUNCTION_ARG_PADDING (MODE, TYPE). | |
2399 | ||
2400 | Return true if an argument passed on the stack should be padded upwards, | |
2401 | i.e. if the least-significant byte of the stack slot has useful data. | |
2402 | ||
2403 | Small aggregate types are placed in the lowest memory address. | |
2404 | ||
2405 | The related parameter passing rules are B.4, C.3, C.5 and C.14. */ | |
2406 | ||
2407 | bool | |
ef4bddc2 | 2408 | aarch64_pad_arg_upward (machine_mode mode, const_tree type) |
43e9d192 IB |
2409 | { |
2410 | /* On little-endian targets, the least significant byte of every stack | |
2411 | argument is passed at the lowest byte address of the stack slot. */ | |
2412 | if (!BYTES_BIG_ENDIAN) | |
2413 | return true; | |
2414 | ||
00edcfbe | 2415 | /* Otherwise, integral, floating-point and pointer types are padded downward: |
43e9d192 IB |
2416 | the least significant byte of a stack argument is passed at the highest |
2417 | byte address of the stack slot. */ | |
2418 | if (type | |
00edcfbe YZ |
2419 | ? (INTEGRAL_TYPE_P (type) || SCALAR_FLOAT_TYPE_P (type) |
2420 | || POINTER_TYPE_P (type)) | |
43e9d192 IB |
2421 | : (SCALAR_INT_MODE_P (mode) || SCALAR_FLOAT_MODE_P (mode))) |
2422 | return false; | |
2423 | ||
2424 | /* Everything else padded upward, i.e. data in first byte of stack slot. */ | |
2425 | return true; | |
2426 | } | |
2427 | ||
2428 | /* Similarly, for use by BLOCK_REG_PADDING (MODE, TYPE, FIRST). | |
2429 | ||
2430 | It specifies padding for the last (may also be the only) | |
2431 | element of a block move between registers and memory. If | |
2432 | assuming the block is in the memory, padding upward means that | |
2433 | the last element is padded after its highest significant byte, | |
2434 | while in downward padding, the last element is padded at the | |
2435 | its least significant byte side. | |
2436 | ||
2437 | Small aggregates and small complex types are always padded | |
2438 | upwards. | |
2439 | ||
2440 | We don't need to worry about homogeneous floating-point or | |
2441 | short-vector aggregates; their move is not affected by the | |
2442 | padding direction determined here. Regardless of endianness, | |
2443 | each element of such an aggregate is put in the least | |
2444 | significant bits of a fp/simd register. | |
2445 | ||
2446 | Return !BYTES_BIG_ENDIAN if the least significant byte of the | |
2447 | register has useful data, and return the opposite if the most | |
2448 | significant byte does. */ | |
2449 | ||
2450 | bool | |
ef4bddc2 | 2451 | aarch64_pad_reg_upward (machine_mode mode, const_tree type, |
43e9d192 IB |
2452 | bool first ATTRIBUTE_UNUSED) |
2453 | { | |
2454 | ||
2455 | /* Small composite types are always padded upward. */ | |
2456 | if (BYTES_BIG_ENDIAN && aarch64_composite_type_p (type, mode)) | |
2457 | { | |
2458 | HOST_WIDE_INT size = (type ? int_size_in_bytes (type) | |
2459 | : GET_MODE_SIZE (mode)); | |
2460 | if (size < 2 * UNITS_PER_WORD) | |
2461 | return true; | |
2462 | } | |
2463 | ||
2464 | /* Otherwise, use the default padding. */ | |
2465 | return !BYTES_BIG_ENDIAN; | |
2466 | } | |
2467 | ||
ef4bddc2 | 2468 | static machine_mode |
43e9d192 IB |
2469 | aarch64_libgcc_cmp_return_mode (void) |
2470 | { | |
2471 | return SImode; | |
2472 | } | |
2473 | ||
a3eb8a52 EB |
2474 | #define PROBE_INTERVAL (1 << STACK_CHECK_PROBE_INTERVAL_EXP) |
2475 | ||
2476 | /* We use the 12-bit shifted immediate arithmetic instructions so values | |
2477 | must be multiple of (1 << 12), i.e. 4096. */ | |
2478 | #define ARITH_FACTOR 4096 | |
2479 | ||
2480 | #if (PROBE_INTERVAL % ARITH_FACTOR) != 0 | |
2481 | #error Cannot use simple address calculation for stack probing | |
2482 | #endif | |
2483 | ||
2484 | /* The pair of scratch registers used for stack probing. */ | |
2485 | #define PROBE_STACK_FIRST_REG 9 | |
2486 | #define PROBE_STACK_SECOND_REG 10 | |
2487 | ||
2488 | /* Emit code to probe a range of stack addresses from FIRST to FIRST+SIZE, | |
2489 | inclusive. These are offsets from the current stack pointer. */ | |
2490 | ||
2491 | static void | |
2492 | aarch64_emit_probe_stack_range (HOST_WIDE_INT first, HOST_WIDE_INT size) | |
2493 | { | |
2494 | rtx reg1 = gen_rtx_REG (ptr_mode, PROBE_STACK_FIRST_REG); | |
2495 | ||
2496 | /* See the same assertion on PROBE_INTERVAL above. */ | |
2497 | gcc_assert ((first % ARITH_FACTOR) == 0); | |
2498 | ||
2499 | /* See if we have a constant small number of probes to generate. If so, | |
2500 | that's the easy case. */ | |
2501 | if (size <= PROBE_INTERVAL) | |
2502 | { | |
2503 | const HOST_WIDE_INT base = ROUND_UP (size, ARITH_FACTOR); | |
2504 | ||
2505 | emit_set_insn (reg1, | |
2506 | plus_constant (ptr_mode, | |
2507 | stack_pointer_rtx, -(first + base))); | |
2508 | emit_stack_probe (plus_constant (ptr_mode, reg1, base - size)); | |
2509 | } | |
2510 | ||
2511 | /* The run-time loop is made up of 8 insns in the generic case while the | |
2512 | compile-time loop is made up of 4+2*(n-2) insns for n # of intervals. */ | |
2513 | else if (size <= 4 * PROBE_INTERVAL) | |
2514 | { | |
2515 | HOST_WIDE_INT i, rem; | |
2516 | ||
2517 | emit_set_insn (reg1, | |
2518 | plus_constant (ptr_mode, | |
2519 | stack_pointer_rtx, | |
2520 | -(first + PROBE_INTERVAL))); | |
2521 | emit_stack_probe (reg1); | |
2522 | ||
2523 | /* Probe at FIRST + N * PROBE_INTERVAL for values of N from 2 until | |
2524 | it exceeds SIZE. If only two probes are needed, this will not | |
2525 | generate any code. Then probe at FIRST + SIZE. */ | |
2526 | for (i = 2 * PROBE_INTERVAL; i < size; i += PROBE_INTERVAL) | |
2527 | { | |
2528 | emit_set_insn (reg1, | |
2529 | plus_constant (ptr_mode, reg1, -PROBE_INTERVAL)); | |
2530 | emit_stack_probe (reg1); | |
2531 | } | |
2532 | ||
2533 | rem = size - (i - PROBE_INTERVAL); | |
2534 | if (rem > 256) | |
2535 | { | |
2536 | const HOST_WIDE_INT base = ROUND_UP (rem, ARITH_FACTOR); | |
2537 | ||
2538 | emit_set_insn (reg1, plus_constant (ptr_mode, reg1, -base)); | |
2539 | emit_stack_probe (plus_constant (ptr_mode, reg1, base - rem)); | |
2540 | } | |
2541 | else | |
2542 | emit_stack_probe (plus_constant (ptr_mode, reg1, -rem)); | |
2543 | } | |
2544 | ||
2545 | /* Otherwise, do the same as above, but in a loop. Note that we must be | |
2546 | extra careful with variables wrapping around because we might be at | |
2547 | the very top (or the very bottom) of the address space and we have | |
2548 | to be able to handle this case properly; in particular, we use an | |
2549 | equality test for the loop condition. */ | |
2550 | else | |
2551 | { | |
2552 | rtx reg2 = gen_rtx_REG (ptr_mode, PROBE_STACK_SECOND_REG); | |
2553 | ||
2554 | /* Step 1: round SIZE to the previous multiple of the interval. */ | |
2555 | ||
2556 | HOST_WIDE_INT rounded_size = size & -PROBE_INTERVAL; | |
2557 | ||
2558 | ||
2559 | /* Step 2: compute initial and final value of the loop counter. */ | |
2560 | ||
2561 | /* TEST_ADDR = SP + FIRST. */ | |
2562 | emit_set_insn (reg1, | |
2563 | plus_constant (ptr_mode, stack_pointer_rtx, -first)); | |
2564 | ||
2565 | /* LAST_ADDR = SP + FIRST + ROUNDED_SIZE. */ | |
2566 | emit_set_insn (reg2, | |
2567 | plus_constant (ptr_mode, stack_pointer_rtx, | |
2568 | -(first + rounded_size))); | |
2569 | ||
2570 | ||
2571 | /* Step 3: the loop | |
2572 | ||
2573 | do | |
2574 | { | |
2575 | TEST_ADDR = TEST_ADDR + PROBE_INTERVAL | |
2576 | probe at TEST_ADDR | |
2577 | } | |
2578 | while (TEST_ADDR != LAST_ADDR) | |
2579 | ||
2580 | probes at FIRST + N * PROBE_INTERVAL for values of N from 1 | |
2581 | until it is equal to ROUNDED_SIZE. */ | |
2582 | ||
2583 | if (ptr_mode == DImode) | |
2584 | emit_insn (gen_probe_stack_range_di (reg1, reg1, reg2)); | |
2585 | else | |
2586 | emit_insn (gen_probe_stack_range_si (reg1, reg1, reg2)); | |
2587 | ||
2588 | ||
2589 | /* Step 4: probe at FIRST + SIZE if we cannot assert at compile-time | |
2590 | that SIZE is equal to ROUNDED_SIZE. */ | |
2591 | ||
2592 | if (size != rounded_size) | |
2593 | { | |
2594 | HOST_WIDE_INT rem = size - rounded_size; | |
2595 | ||
2596 | if (rem > 256) | |
2597 | { | |
2598 | const HOST_WIDE_INT base = ROUND_UP (rem, ARITH_FACTOR); | |
2599 | ||
2600 | emit_set_insn (reg2, plus_constant (ptr_mode, reg2, -base)); | |
2601 | emit_stack_probe (plus_constant (ptr_mode, reg2, base - rem)); | |
2602 | } | |
2603 | else | |
2604 | emit_stack_probe (plus_constant (ptr_mode, reg2, -rem)); | |
2605 | } | |
2606 | } | |
2607 | ||
2608 | /* Make sure nothing is scheduled before we are done. */ | |
2609 | emit_insn (gen_blockage ()); | |
2610 | } | |
2611 | ||
2612 | /* Probe a range of stack addresses from REG1 to REG2 inclusive. These are | |
2613 | absolute addresses. */ | |
2614 | ||
2615 | const char * | |
2616 | aarch64_output_probe_stack_range (rtx reg1, rtx reg2) | |
2617 | { | |
2618 | static int labelno = 0; | |
2619 | char loop_lab[32]; | |
2620 | rtx xops[2]; | |
2621 | ||
2622 | ASM_GENERATE_INTERNAL_LABEL (loop_lab, "LPSRL", labelno++); | |
2623 | ||
2624 | /* Loop. */ | |
2625 | ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, loop_lab); | |
2626 | ||
2627 | /* TEST_ADDR = TEST_ADDR + PROBE_INTERVAL. */ | |
2628 | xops[0] = reg1; | |
2629 | xops[1] = GEN_INT (PROBE_INTERVAL); | |
2630 | output_asm_insn ("sub\t%0, %0, %1", xops); | |
2631 | ||
2632 | /* Probe at TEST_ADDR. */ | |
2633 | output_asm_insn ("str\txzr, [%0]", xops); | |
2634 | ||
2635 | /* Test if TEST_ADDR == LAST_ADDR. */ | |
2636 | xops[1] = reg2; | |
2637 | output_asm_insn ("cmp\t%0, %1", xops); | |
2638 | ||
2639 | /* Branch. */ | |
2640 | fputs ("\tb.ne\t", asm_out_file); | |
2641 | assemble_name_raw (asm_out_file, loop_lab); | |
2642 | fputc ('\n', asm_out_file); | |
2643 | ||
2644 | return ""; | |
2645 | } | |
2646 | ||
43e9d192 IB |
2647 | static bool |
2648 | aarch64_frame_pointer_required (void) | |
2649 | { | |
0b7f8166 MS |
2650 | /* In aarch64_override_options_after_change |
2651 | flag_omit_leaf_frame_pointer turns off the frame pointer by | |
2652 | default. Turn it back on now if we've not got a leaf | |
2653 | function. */ | |
2654 | if (flag_omit_leaf_frame_pointer | |
2655 | && (!crtl->is_leaf || df_regs_ever_live_p (LR_REGNUM))) | |
2656 | return true; | |
43e9d192 | 2657 | |
0b7f8166 | 2658 | return false; |
43e9d192 IB |
2659 | } |
2660 | ||
2661 | /* Mark the registers that need to be saved by the callee and calculate | |
2662 | the size of the callee-saved registers area and frame record (both FP | |
2663 | and LR may be omitted). */ | |
2664 | static void | |
2665 | aarch64_layout_frame (void) | |
2666 | { | |
2667 | HOST_WIDE_INT offset = 0; | |
2668 | int regno; | |
2669 | ||
2670 | if (reload_completed && cfun->machine->frame.laid_out) | |
2671 | return; | |
2672 | ||
97826595 MS |
2673 | #define SLOT_NOT_REQUIRED (-2) |
2674 | #define SLOT_REQUIRED (-1) | |
2675 | ||
363ffa50 JW |
2676 | cfun->machine->frame.wb_candidate1 = FIRST_PSEUDO_REGISTER; |
2677 | cfun->machine->frame.wb_candidate2 = FIRST_PSEUDO_REGISTER; | |
2678 | ||
43e9d192 IB |
2679 | /* First mark all the registers that really need to be saved... */ |
2680 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) | |
97826595 | 2681 | cfun->machine->frame.reg_offset[regno] = SLOT_NOT_REQUIRED; |
43e9d192 IB |
2682 | |
2683 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
97826595 | 2684 | cfun->machine->frame.reg_offset[regno] = SLOT_NOT_REQUIRED; |
43e9d192 IB |
2685 | |
2686 | /* ... that includes the eh data registers (if needed)... */ | |
2687 | if (crtl->calls_eh_return) | |
2688 | for (regno = 0; EH_RETURN_DATA_REGNO (regno) != INVALID_REGNUM; regno++) | |
97826595 MS |
2689 | cfun->machine->frame.reg_offset[EH_RETURN_DATA_REGNO (regno)] |
2690 | = SLOT_REQUIRED; | |
43e9d192 IB |
2691 | |
2692 | /* ... and any callee saved register that dataflow says is live. */ | |
2693 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) | |
2694 | if (df_regs_ever_live_p (regno) | |
1c923b60 JW |
2695 | && (regno == R30_REGNUM |
2696 | || !call_used_regs[regno])) | |
97826595 | 2697 | cfun->machine->frame.reg_offset[regno] = SLOT_REQUIRED; |
43e9d192 IB |
2698 | |
2699 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
2700 | if (df_regs_ever_live_p (regno) | |
2701 | && !call_used_regs[regno]) | |
97826595 | 2702 | cfun->machine->frame.reg_offset[regno] = SLOT_REQUIRED; |
43e9d192 IB |
2703 | |
2704 | if (frame_pointer_needed) | |
2705 | { | |
2e1cdae5 | 2706 | /* FP and LR are placed in the linkage record. */ |
43e9d192 | 2707 | cfun->machine->frame.reg_offset[R29_REGNUM] = 0; |
363ffa50 | 2708 | cfun->machine->frame.wb_candidate1 = R29_REGNUM; |
2e1cdae5 | 2709 | cfun->machine->frame.reg_offset[R30_REGNUM] = UNITS_PER_WORD; |
363ffa50 | 2710 | cfun->machine->frame.wb_candidate2 = R30_REGNUM; |
43e9d192 | 2711 | cfun->machine->frame.hardfp_offset = 2 * UNITS_PER_WORD; |
2e1cdae5 | 2712 | offset += 2 * UNITS_PER_WORD; |
43e9d192 IB |
2713 | } |
2714 | ||
2715 | /* Now assign stack slots for them. */ | |
2e1cdae5 | 2716 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) |
97826595 | 2717 | if (cfun->machine->frame.reg_offset[regno] == SLOT_REQUIRED) |
43e9d192 IB |
2718 | { |
2719 | cfun->machine->frame.reg_offset[regno] = offset; | |
363ffa50 JW |
2720 | if (cfun->machine->frame.wb_candidate1 == FIRST_PSEUDO_REGISTER) |
2721 | cfun->machine->frame.wb_candidate1 = regno; | |
2722 | else if (cfun->machine->frame.wb_candidate2 == FIRST_PSEUDO_REGISTER) | |
2723 | cfun->machine->frame.wb_candidate2 = regno; | |
43e9d192 IB |
2724 | offset += UNITS_PER_WORD; |
2725 | } | |
2726 | ||
2727 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
97826595 | 2728 | if (cfun->machine->frame.reg_offset[regno] == SLOT_REQUIRED) |
43e9d192 IB |
2729 | { |
2730 | cfun->machine->frame.reg_offset[regno] = offset; | |
363ffa50 JW |
2731 | if (cfun->machine->frame.wb_candidate1 == FIRST_PSEUDO_REGISTER) |
2732 | cfun->machine->frame.wb_candidate1 = regno; | |
2733 | else if (cfun->machine->frame.wb_candidate2 == FIRST_PSEUDO_REGISTER | |
2734 | && cfun->machine->frame.wb_candidate1 >= V0_REGNUM) | |
2735 | cfun->machine->frame.wb_candidate2 = regno; | |
43e9d192 IB |
2736 | offset += UNITS_PER_WORD; |
2737 | } | |
2738 | ||
43e9d192 | 2739 | cfun->machine->frame.padding0 = |
4f59f9f2 UB |
2740 | (ROUND_UP (offset, STACK_BOUNDARY / BITS_PER_UNIT) - offset); |
2741 | offset = ROUND_UP (offset, STACK_BOUNDARY / BITS_PER_UNIT); | |
43e9d192 IB |
2742 | |
2743 | cfun->machine->frame.saved_regs_size = offset; | |
1c960e02 MS |
2744 | |
2745 | cfun->machine->frame.hard_fp_offset | |
4f59f9f2 UB |
2746 | = ROUND_UP (cfun->machine->frame.saved_varargs_size |
2747 | + get_frame_size () | |
2748 | + cfun->machine->frame.saved_regs_size, | |
2749 | STACK_BOUNDARY / BITS_PER_UNIT); | |
1c960e02 MS |
2750 | |
2751 | cfun->machine->frame.frame_size | |
4f59f9f2 UB |
2752 | = ROUND_UP (cfun->machine->frame.hard_fp_offset |
2753 | + crtl->outgoing_args_size, | |
2754 | STACK_BOUNDARY / BITS_PER_UNIT); | |
1c960e02 | 2755 | |
43e9d192 IB |
2756 | cfun->machine->frame.laid_out = true; |
2757 | } | |
2758 | ||
43e9d192 IB |
2759 | static bool |
2760 | aarch64_register_saved_on_entry (int regno) | |
2761 | { | |
97826595 | 2762 | return cfun->machine->frame.reg_offset[regno] >= 0; |
43e9d192 IB |
2763 | } |
2764 | ||
64dedd72 JW |
2765 | static unsigned |
2766 | aarch64_next_callee_save (unsigned regno, unsigned limit) | |
2767 | { | |
2768 | while (regno <= limit && !aarch64_register_saved_on_entry (regno)) | |
2769 | regno ++; | |
2770 | return regno; | |
2771 | } | |
43e9d192 | 2772 | |
c5e1f66e | 2773 | static void |
ef4bddc2 | 2774 | aarch64_pushwb_single_reg (machine_mode mode, unsigned regno, |
c5e1f66e JW |
2775 | HOST_WIDE_INT adjustment) |
2776 | { | |
2777 | rtx base_rtx = stack_pointer_rtx; | |
2778 | rtx insn, reg, mem; | |
2779 | ||
2780 | reg = gen_rtx_REG (mode, regno); | |
2781 | mem = gen_rtx_PRE_MODIFY (Pmode, base_rtx, | |
2782 | plus_constant (Pmode, base_rtx, -adjustment)); | |
2783 | mem = gen_rtx_MEM (mode, mem); | |
2784 | ||
2785 | insn = emit_move_insn (mem, reg); | |
2786 | RTX_FRAME_RELATED_P (insn) = 1; | |
2787 | } | |
2788 | ||
80c11907 | 2789 | static rtx |
ef4bddc2 | 2790 | aarch64_gen_storewb_pair (machine_mode mode, rtx base, rtx reg, rtx reg2, |
80c11907 JW |
2791 | HOST_WIDE_INT adjustment) |
2792 | { | |
2793 | switch (mode) | |
2794 | { | |
2795 | case DImode: | |
2796 | return gen_storewb_pairdi_di (base, base, reg, reg2, | |
2797 | GEN_INT (-adjustment), | |
2798 | GEN_INT (UNITS_PER_WORD - adjustment)); | |
2799 | case DFmode: | |
2800 | return gen_storewb_pairdf_di (base, base, reg, reg2, | |
2801 | GEN_INT (-adjustment), | |
2802 | GEN_INT (UNITS_PER_WORD - adjustment)); | |
2803 | default: | |
2804 | gcc_unreachable (); | |
2805 | } | |
2806 | } | |
2807 | ||
2808 | static void | |
ef4bddc2 | 2809 | aarch64_pushwb_pair_reg (machine_mode mode, unsigned regno1, |
80c11907 JW |
2810 | unsigned regno2, HOST_WIDE_INT adjustment) |
2811 | { | |
5d8a22a5 | 2812 | rtx_insn *insn; |
80c11907 JW |
2813 | rtx reg1 = gen_rtx_REG (mode, regno1); |
2814 | rtx reg2 = gen_rtx_REG (mode, regno2); | |
2815 | ||
2816 | insn = emit_insn (aarch64_gen_storewb_pair (mode, stack_pointer_rtx, reg1, | |
2817 | reg2, adjustment)); | |
2818 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 2)) = 1; | |
80c11907 JW |
2819 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; |
2820 | RTX_FRAME_RELATED_P (insn) = 1; | |
2821 | } | |
2822 | ||
159313d9 | 2823 | static rtx |
ef4bddc2 | 2824 | aarch64_gen_loadwb_pair (machine_mode mode, rtx base, rtx reg, rtx reg2, |
159313d9 JW |
2825 | HOST_WIDE_INT adjustment) |
2826 | { | |
2827 | switch (mode) | |
2828 | { | |
2829 | case DImode: | |
2830 | return gen_loadwb_pairdi_di (base, base, reg, reg2, GEN_INT (adjustment), | |
3e322b3f | 2831 | GEN_INT (UNITS_PER_WORD)); |
159313d9 JW |
2832 | case DFmode: |
2833 | return gen_loadwb_pairdf_di (base, base, reg, reg2, GEN_INT (adjustment), | |
3e322b3f | 2834 | GEN_INT (UNITS_PER_WORD)); |
159313d9 JW |
2835 | default: |
2836 | gcc_unreachable (); | |
2837 | } | |
2838 | } | |
2839 | ||
72df5c1f | 2840 | static rtx |
ef4bddc2 | 2841 | aarch64_gen_store_pair (machine_mode mode, rtx mem1, rtx reg1, rtx mem2, |
72df5c1f JW |
2842 | rtx reg2) |
2843 | { | |
2844 | switch (mode) | |
2845 | { | |
2846 | case DImode: | |
2847 | return gen_store_pairdi (mem1, reg1, mem2, reg2); | |
2848 | ||
2849 | case DFmode: | |
2850 | return gen_store_pairdf (mem1, reg1, mem2, reg2); | |
2851 | ||
2852 | default: | |
2853 | gcc_unreachable (); | |
2854 | } | |
2855 | } | |
2856 | ||
2857 | static rtx | |
ef4bddc2 | 2858 | aarch64_gen_load_pair (machine_mode mode, rtx reg1, rtx mem1, rtx reg2, |
72df5c1f JW |
2859 | rtx mem2) |
2860 | { | |
2861 | switch (mode) | |
2862 | { | |
2863 | case DImode: | |
2864 | return gen_load_pairdi (reg1, mem1, reg2, mem2); | |
2865 | ||
2866 | case DFmode: | |
2867 | return gen_load_pairdf (reg1, mem1, reg2, mem2); | |
2868 | ||
2869 | default: | |
2870 | gcc_unreachable (); | |
2871 | } | |
2872 | } | |
2873 | ||
43e9d192 | 2874 | |
43e9d192 | 2875 | static void |
ef4bddc2 | 2876 | aarch64_save_callee_saves (machine_mode mode, HOST_WIDE_INT start_offset, |
ae13fce3 | 2877 | unsigned start, unsigned limit, bool skip_wb) |
43e9d192 | 2878 | { |
5d8a22a5 | 2879 | rtx_insn *insn; |
ef4bddc2 | 2880 | rtx (*gen_mem_ref) (machine_mode, rtx) = (frame_pointer_needed |
a007a21c | 2881 | ? gen_frame_mem : gen_rtx_MEM); |
43e9d192 IB |
2882 | unsigned regno; |
2883 | unsigned regno2; | |
2884 | ||
0ec74a1e | 2885 | for (regno = aarch64_next_callee_save (start, limit); |
64dedd72 JW |
2886 | regno <= limit; |
2887 | regno = aarch64_next_callee_save (regno + 1, limit)) | |
43e9d192 | 2888 | { |
ae13fce3 JW |
2889 | rtx reg, mem; |
2890 | HOST_WIDE_INT offset; | |
64dedd72 | 2891 | |
ae13fce3 JW |
2892 | if (skip_wb |
2893 | && (regno == cfun->machine->frame.wb_candidate1 | |
2894 | || regno == cfun->machine->frame.wb_candidate2)) | |
2895 | continue; | |
2896 | ||
2897 | reg = gen_rtx_REG (mode, regno); | |
2898 | offset = start_offset + cfun->machine->frame.reg_offset[regno]; | |
0ec74a1e JW |
2899 | mem = gen_mem_ref (mode, plus_constant (Pmode, stack_pointer_rtx, |
2900 | offset)); | |
64dedd72 JW |
2901 | |
2902 | regno2 = aarch64_next_callee_save (regno + 1, limit); | |
2903 | ||
2904 | if (regno2 <= limit | |
2905 | && ((cfun->machine->frame.reg_offset[regno] + UNITS_PER_WORD) | |
2906 | == cfun->machine->frame.reg_offset[regno2])) | |
2907 | ||
43e9d192 | 2908 | { |
0ec74a1e | 2909 | rtx reg2 = gen_rtx_REG (mode, regno2); |
64dedd72 JW |
2910 | rtx mem2; |
2911 | ||
2912 | offset = start_offset + cfun->machine->frame.reg_offset[regno2]; | |
8ed2fc62 JW |
2913 | mem2 = gen_mem_ref (mode, plus_constant (Pmode, stack_pointer_rtx, |
2914 | offset)); | |
2915 | insn = emit_insn (aarch64_gen_store_pair (mode, mem, reg, mem2, | |
2916 | reg2)); | |
0b4a9743 | 2917 | |
64dedd72 JW |
2918 | /* The first part of a frame-related parallel insn is |
2919 | always assumed to be relevant to the frame | |
2920 | calculations; subsequent parts, are only | |
2921 | frame-related if explicitly marked. */ | |
2922 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; | |
2923 | regno = regno2; | |
2924 | } | |
2925 | else | |
8ed2fc62 JW |
2926 | insn = emit_move_insn (mem, reg); |
2927 | ||
2928 | RTX_FRAME_RELATED_P (insn) = 1; | |
2929 | } | |
2930 | } | |
2931 | ||
2932 | static void | |
ef4bddc2 | 2933 | aarch64_restore_callee_saves (machine_mode mode, |
8ed2fc62 | 2934 | HOST_WIDE_INT start_offset, unsigned start, |
dd991abb | 2935 | unsigned limit, bool skip_wb, rtx *cfi_ops) |
8ed2fc62 | 2936 | { |
8ed2fc62 | 2937 | rtx base_rtx = stack_pointer_rtx; |
ef4bddc2 | 2938 | rtx (*gen_mem_ref) (machine_mode, rtx) = (frame_pointer_needed |
8ed2fc62 JW |
2939 | ? gen_frame_mem : gen_rtx_MEM); |
2940 | unsigned regno; | |
2941 | unsigned regno2; | |
2942 | HOST_WIDE_INT offset; | |
2943 | ||
2944 | for (regno = aarch64_next_callee_save (start, limit); | |
2945 | regno <= limit; | |
2946 | regno = aarch64_next_callee_save (regno + 1, limit)) | |
2947 | { | |
ae13fce3 | 2948 | rtx reg, mem; |
8ed2fc62 | 2949 | |
ae13fce3 JW |
2950 | if (skip_wb |
2951 | && (regno == cfun->machine->frame.wb_candidate1 | |
2952 | || regno == cfun->machine->frame.wb_candidate2)) | |
2953 | continue; | |
2954 | ||
2955 | reg = gen_rtx_REG (mode, regno); | |
8ed2fc62 JW |
2956 | offset = start_offset + cfun->machine->frame.reg_offset[regno]; |
2957 | mem = gen_mem_ref (mode, plus_constant (Pmode, base_rtx, offset)); | |
2958 | ||
2959 | regno2 = aarch64_next_callee_save (regno + 1, limit); | |
2960 | ||
2961 | if (regno2 <= limit | |
2962 | && ((cfun->machine->frame.reg_offset[regno] + UNITS_PER_WORD) | |
2963 | == cfun->machine->frame.reg_offset[regno2])) | |
64dedd72 | 2964 | { |
8ed2fc62 JW |
2965 | rtx reg2 = gen_rtx_REG (mode, regno2); |
2966 | rtx mem2; | |
2967 | ||
2968 | offset = start_offset + cfun->machine->frame.reg_offset[regno2]; | |
2969 | mem2 = gen_mem_ref (mode, plus_constant (Pmode, base_rtx, offset)); | |
dd991abb | 2970 | emit_insn (aarch64_gen_load_pair (mode, reg, mem, reg2, mem2)); |
8ed2fc62 | 2971 | |
dd991abb | 2972 | *cfi_ops = alloc_reg_note (REG_CFA_RESTORE, reg2, *cfi_ops); |
8ed2fc62 | 2973 | regno = regno2; |
43e9d192 | 2974 | } |
8ed2fc62 | 2975 | else |
dd991abb RH |
2976 | emit_move_insn (reg, mem); |
2977 | *cfi_ops = alloc_reg_note (REG_CFA_RESTORE, reg, *cfi_ops); | |
43e9d192 | 2978 | } |
43e9d192 IB |
2979 | } |
2980 | ||
2981 | /* AArch64 stack frames generated by this compiler look like: | |
2982 | ||
2983 | +-------------------------------+ | |
2984 | | | | |
2985 | | incoming stack arguments | | |
2986 | | | | |
34834420 MS |
2987 | +-------------------------------+ |
2988 | | | <-- incoming stack pointer (aligned) | |
43e9d192 IB |
2989 | | callee-allocated save area | |
2990 | | for register varargs | | |
2991 | | | | |
34834420 MS |
2992 | +-------------------------------+ |
2993 | | local variables | <-- frame_pointer_rtx | |
43e9d192 IB |
2994 | | | |
2995 | +-------------------------------+ | |
454fdba9 RL |
2996 | | padding0 | \ |
2997 | +-------------------------------+ | | |
454fdba9 | 2998 | | callee-saved registers | | frame.saved_regs_size |
454fdba9 RL |
2999 | +-------------------------------+ | |
3000 | | LR' | | | |
3001 | +-------------------------------+ | | |
34834420 MS |
3002 | | FP' | / <- hard_frame_pointer_rtx (aligned) |
3003 | +-------------------------------+ | |
43e9d192 IB |
3004 | | dynamic allocation | |
3005 | +-------------------------------+ | |
34834420 MS |
3006 | | padding | |
3007 | +-------------------------------+ | |
3008 | | outgoing stack arguments | <-- arg_pointer | |
3009 | | | | |
3010 | +-------------------------------+ | |
3011 | | | <-- stack_pointer_rtx (aligned) | |
43e9d192 | 3012 | |
34834420 MS |
3013 | Dynamic stack allocations via alloca() decrease stack_pointer_rtx |
3014 | but leave frame_pointer_rtx and hard_frame_pointer_rtx | |
3015 | unchanged. */ | |
43e9d192 IB |
3016 | |
3017 | /* Generate the prologue instructions for entry into a function. | |
3018 | Establish the stack frame by decreasing the stack pointer with a | |
3019 | properly calculated size and, if necessary, create a frame record | |
3020 | filled with the values of LR and previous frame pointer. The | |
6991c977 | 3021 | current FP is also set up if it is in use. */ |
43e9d192 IB |
3022 | |
3023 | void | |
3024 | aarch64_expand_prologue (void) | |
3025 | { | |
3026 | /* sub sp, sp, #<frame_size> | |
3027 | stp {fp, lr}, [sp, #<frame_size> - 16] | |
3028 | add fp, sp, #<frame_size> - hardfp_offset | |
3029 | stp {cs_reg}, [fp, #-16] etc. | |
3030 | ||
3031 | sub sp, sp, <final_adjustment_if_any> | |
3032 | */ | |
43e9d192 | 3033 | HOST_WIDE_INT frame_size, offset; |
1c960e02 | 3034 | HOST_WIDE_INT fp_offset; /* Offset from hard FP to SP. */ |
dd991abb | 3035 | HOST_WIDE_INT hard_fp_offset; |
5d8a22a5 | 3036 | rtx_insn *insn; |
43e9d192 IB |
3037 | |
3038 | aarch64_layout_frame (); | |
43e9d192 | 3039 | |
dd991abb RH |
3040 | offset = frame_size = cfun->machine->frame.frame_size; |
3041 | hard_fp_offset = cfun->machine->frame.hard_fp_offset; | |
3042 | fp_offset = frame_size - hard_fp_offset; | |
43e9d192 | 3043 | |
dd991abb RH |
3044 | if (flag_stack_usage_info) |
3045 | current_function_static_stack_size = frame_size; | |
43e9d192 | 3046 | |
a3eb8a52 EB |
3047 | if (flag_stack_check == STATIC_BUILTIN_STACK_CHECK) |
3048 | { | |
3049 | if (crtl->is_leaf && !cfun->calls_alloca) | |
3050 | { | |
3051 | if (frame_size > PROBE_INTERVAL && frame_size > STACK_CHECK_PROTECT) | |
3052 | aarch64_emit_probe_stack_range (STACK_CHECK_PROTECT, | |
3053 | frame_size - STACK_CHECK_PROTECT); | |
3054 | } | |
3055 | else if (frame_size > 0) | |
3056 | aarch64_emit_probe_stack_range (STACK_CHECK_PROTECT, frame_size); | |
3057 | } | |
3058 | ||
44c0e7b9 | 3059 | /* Store pairs and load pairs have a range only -512 to 504. */ |
43e9d192 IB |
3060 | if (offset >= 512) |
3061 | { | |
3062 | /* When the frame has a large size, an initial decrease is done on | |
3063 | the stack pointer to jump over the callee-allocated save area for | |
3064 | register varargs, the local variable area and/or the callee-saved | |
3065 | register area. This will allow the pre-index write-back | |
3066 | store pair instructions to be used for setting up the stack frame | |
3067 | efficiently. */ | |
dd991abb | 3068 | offset = hard_fp_offset; |
43e9d192 IB |
3069 | if (offset >= 512) |
3070 | offset = cfun->machine->frame.saved_regs_size; | |
3071 | ||
3072 | frame_size -= (offset + crtl->outgoing_args_size); | |
3073 | fp_offset = 0; | |
3074 | ||
3075 | if (frame_size >= 0x1000000) | |
3076 | { | |
3077 | rtx op0 = gen_rtx_REG (Pmode, IP0_REGNUM); | |
3078 | emit_move_insn (op0, GEN_INT (-frame_size)); | |
dd991abb RH |
3079 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, op0)); |
3080 | ||
3081 | add_reg_note (insn, REG_CFA_ADJUST_CFA, | |
f7df4a84 | 3082 | gen_rtx_SET (stack_pointer_rtx, |
dd991abb RH |
3083 | plus_constant (Pmode, stack_pointer_rtx, |
3084 | -frame_size))); | |
3085 | RTX_FRAME_RELATED_P (insn) = 1; | |
43e9d192 IB |
3086 | } |
3087 | else if (frame_size > 0) | |
3088 | { | |
dd991abb RH |
3089 | int hi_ofs = frame_size & 0xfff000; |
3090 | int lo_ofs = frame_size & 0x000fff; | |
3091 | ||
3092 | if (hi_ofs) | |
43e9d192 IB |
3093 | { |
3094 | insn = emit_insn (gen_add2_insn | |
dd991abb | 3095 | (stack_pointer_rtx, GEN_INT (-hi_ofs))); |
43e9d192 IB |
3096 | RTX_FRAME_RELATED_P (insn) = 1; |
3097 | } | |
dd991abb | 3098 | if (lo_ofs) |
43e9d192 IB |
3099 | { |
3100 | insn = emit_insn (gen_add2_insn | |
dd991abb | 3101 | (stack_pointer_rtx, GEN_INT (-lo_ofs))); |
43e9d192 IB |
3102 | RTX_FRAME_RELATED_P (insn) = 1; |
3103 | } | |
3104 | } | |
3105 | } | |
3106 | else | |
3107 | frame_size = -1; | |
3108 | ||
3109 | if (offset > 0) | |
3110 | { | |
ae13fce3 JW |
3111 | bool skip_wb = false; |
3112 | ||
43e9d192 IB |
3113 | if (frame_pointer_needed) |
3114 | { | |
c5e1f66e JW |
3115 | skip_wb = true; |
3116 | ||
43e9d192 IB |
3117 | if (fp_offset) |
3118 | { | |
3119 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, | |
3120 | GEN_INT (-offset))); | |
3121 | RTX_FRAME_RELATED_P (insn) = 1; | |
80c11907 JW |
3122 | |
3123 | aarch64_save_callee_saves (DImode, fp_offset, R29_REGNUM, | |
c5e1f66e | 3124 | R30_REGNUM, false); |
43e9d192 IB |
3125 | } |
3126 | else | |
80c11907 | 3127 | aarch64_pushwb_pair_reg (DImode, R29_REGNUM, R30_REGNUM, offset); |
43e9d192 IB |
3128 | |
3129 | /* Set up frame pointer to point to the location of the | |
3130 | previous frame pointer on the stack. */ | |
3131 | insn = emit_insn (gen_add3_insn (hard_frame_pointer_rtx, | |
3132 | stack_pointer_rtx, | |
3133 | GEN_INT (fp_offset))); | |
43e9d192 | 3134 | RTX_FRAME_RELATED_P (insn) = 1; |
dd991abb | 3135 | emit_insn (gen_stack_tie (stack_pointer_rtx, hard_frame_pointer_rtx)); |
43e9d192 IB |
3136 | } |
3137 | else | |
3138 | { | |
c5e1f66e JW |
3139 | unsigned reg1 = cfun->machine->frame.wb_candidate1; |
3140 | unsigned reg2 = cfun->machine->frame.wb_candidate2; | |
80c11907 | 3141 | |
c5e1f66e JW |
3142 | if (fp_offset |
3143 | || reg1 == FIRST_PSEUDO_REGISTER | |
3144 | || (reg2 == FIRST_PSEUDO_REGISTER | |
3145 | && offset >= 256)) | |
3146 | { | |
3147 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, | |
3148 | GEN_INT (-offset))); | |
3149 | RTX_FRAME_RELATED_P (insn) = 1; | |
3150 | } | |
3151 | else | |
3152 | { | |
ef4bddc2 | 3153 | machine_mode mode1 = (reg1 <= R30_REGNUM) ? DImode : DFmode; |
c5e1f66e JW |
3154 | |
3155 | skip_wb = true; | |
3156 | ||
3157 | if (reg2 == FIRST_PSEUDO_REGISTER) | |
3158 | aarch64_pushwb_single_reg (mode1, reg1, offset); | |
3159 | else | |
3160 | aarch64_pushwb_pair_reg (mode1, reg1, reg2, offset); | |
3161 | } | |
43e9d192 IB |
3162 | } |
3163 | ||
c5e1f66e JW |
3164 | aarch64_save_callee_saves (DImode, fp_offset, R0_REGNUM, R30_REGNUM, |
3165 | skip_wb); | |
ae13fce3 JW |
3166 | aarch64_save_callee_saves (DFmode, fp_offset, V0_REGNUM, V31_REGNUM, |
3167 | skip_wb); | |
43e9d192 IB |
3168 | } |
3169 | ||
3170 | /* when offset >= 512, | |
3171 | sub sp, sp, #<outgoing_args_size> */ | |
3172 | if (frame_size > -1) | |
3173 | { | |
3174 | if (crtl->outgoing_args_size > 0) | |
3175 | { | |
3176 | insn = emit_insn (gen_add2_insn | |
3177 | (stack_pointer_rtx, | |
3178 | GEN_INT (- crtl->outgoing_args_size))); | |
3179 | RTX_FRAME_RELATED_P (insn) = 1; | |
3180 | } | |
3181 | } | |
3182 | } | |
3183 | ||
4f942779 RL |
3184 | /* Return TRUE if we can use a simple_return insn. |
3185 | ||
3186 | This function checks whether the callee saved stack is empty, which | |
3187 | means no restore actions are need. The pro_and_epilogue will use | |
3188 | this to check whether shrink-wrapping opt is feasible. */ | |
3189 | ||
3190 | bool | |
3191 | aarch64_use_return_insn_p (void) | |
3192 | { | |
3193 | if (!reload_completed) | |
3194 | return false; | |
3195 | ||
3196 | if (crtl->profile) | |
3197 | return false; | |
3198 | ||
3199 | aarch64_layout_frame (); | |
3200 | ||
3201 | return cfun->machine->frame.frame_size == 0; | |
3202 | } | |
3203 | ||
43e9d192 IB |
3204 | /* Generate the epilogue instructions for returning from a function. */ |
3205 | void | |
3206 | aarch64_expand_epilogue (bool for_sibcall) | |
3207 | { | |
1c960e02 | 3208 | HOST_WIDE_INT frame_size, offset; |
43e9d192 | 3209 | HOST_WIDE_INT fp_offset; |
dd991abb | 3210 | HOST_WIDE_INT hard_fp_offset; |
5d8a22a5 | 3211 | rtx_insn *insn; |
7e8c2bd5 JW |
3212 | /* We need to add memory barrier to prevent read from deallocated stack. */ |
3213 | bool need_barrier_p = (get_frame_size () != 0 | |
3214 | || cfun->machine->frame.saved_varargs_size); | |
43e9d192 IB |
3215 | |
3216 | aarch64_layout_frame (); | |
43e9d192 | 3217 | |
1c960e02 | 3218 | offset = frame_size = cfun->machine->frame.frame_size; |
dd991abb RH |
3219 | hard_fp_offset = cfun->machine->frame.hard_fp_offset; |
3220 | fp_offset = frame_size - hard_fp_offset; | |
44c0e7b9 YZ |
3221 | |
3222 | /* Store pairs and load pairs have a range only -512 to 504. */ | |
43e9d192 IB |
3223 | if (offset >= 512) |
3224 | { | |
dd991abb | 3225 | offset = hard_fp_offset; |
43e9d192 IB |
3226 | if (offset >= 512) |
3227 | offset = cfun->machine->frame.saved_regs_size; | |
3228 | ||
3229 | frame_size -= (offset + crtl->outgoing_args_size); | |
3230 | fp_offset = 0; | |
3231 | if (!frame_pointer_needed && crtl->outgoing_args_size > 0) | |
3232 | { | |
3233 | insn = emit_insn (gen_add2_insn | |
3234 | (stack_pointer_rtx, | |
3235 | GEN_INT (crtl->outgoing_args_size))); | |
3236 | RTX_FRAME_RELATED_P (insn) = 1; | |
3237 | } | |
3238 | } | |
3239 | else | |
3240 | frame_size = -1; | |
3241 | ||
3242 | /* If there were outgoing arguments or we've done dynamic stack | |
3243 | allocation, then restore the stack pointer from the frame | |
3244 | pointer. This is at most one insn and more efficient than using | |
3245 | GCC's internal mechanism. */ | |
3246 | if (frame_pointer_needed | |
3247 | && (crtl->outgoing_args_size || cfun->calls_alloca)) | |
3248 | { | |
7e8c2bd5 JW |
3249 | if (cfun->calls_alloca) |
3250 | emit_insn (gen_stack_tie (stack_pointer_rtx, stack_pointer_rtx)); | |
3251 | ||
43e9d192 IB |
3252 | insn = emit_insn (gen_add3_insn (stack_pointer_rtx, |
3253 | hard_frame_pointer_rtx, | |
8f454e9f JW |
3254 | GEN_INT (0))); |
3255 | offset = offset - fp_offset; | |
43e9d192 IB |
3256 | } |
3257 | ||
43e9d192 IB |
3258 | if (offset > 0) |
3259 | { | |
4b92caa1 JW |
3260 | unsigned reg1 = cfun->machine->frame.wb_candidate1; |
3261 | unsigned reg2 = cfun->machine->frame.wb_candidate2; | |
3262 | bool skip_wb = true; | |
dd991abb | 3263 | rtx cfi_ops = NULL; |
4b92caa1 | 3264 | |
43e9d192 | 3265 | if (frame_pointer_needed) |
4b92caa1 JW |
3266 | fp_offset = 0; |
3267 | else if (fp_offset | |
3268 | || reg1 == FIRST_PSEUDO_REGISTER | |
3269 | || (reg2 == FIRST_PSEUDO_REGISTER | |
3270 | && offset >= 256)) | |
3271 | skip_wb = false; | |
3272 | ||
3273 | aarch64_restore_callee_saves (DImode, fp_offset, R0_REGNUM, R30_REGNUM, | |
dd991abb | 3274 | skip_wb, &cfi_ops); |
4b92caa1 | 3275 | aarch64_restore_callee_saves (DFmode, fp_offset, V0_REGNUM, V31_REGNUM, |
dd991abb | 3276 | skip_wb, &cfi_ops); |
4b92caa1 | 3277 | |
7e8c2bd5 JW |
3278 | if (need_barrier_p) |
3279 | emit_insn (gen_stack_tie (stack_pointer_rtx, stack_pointer_rtx)); | |
3280 | ||
4b92caa1 | 3281 | if (skip_wb) |
43e9d192 | 3282 | { |
ef4bddc2 | 3283 | machine_mode mode1 = (reg1 <= R30_REGNUM) ? DImode : DFmode; |
dd991abb | 3284 | rtx rreg1 = gen_rtx_REG (mode1, reg1); |
4b92caa1 | 3285 | |
dd991abb | 3286 | cfi_ops = alloc_reg_note (REG_CFA_RESTORE, rreg1, cfi_ops); |
4b92caa1 | 3287 | if (reg2 == FIRST_PSEUDO_REGISTER) |
dd991abb RH |
3288 | { |
3289 | rtx mem = plus_constant (Pmode, stack_pointer_rtx, offset); | |
3290 | mem = gen_rtx_POST_MODIFY (Pmode, stack_pointer_rtx, mem); | |
3291 | mem = gen_rtx_MEM (mode1, mem); | |
3292 | insn = emit_move_insn (rreg1, mem); | |
3293 | } | |
4b92caa1 JW |
3294 | else |
3295 | { | |
dd991abb | 3296 | rtx rreg2 = gen_rtx_REG (mode1, reg2); |
4b92caa1 | 3297 | |
dd991abb RH |
3298 | cfi_ops = alloc_reg_note (REG_CFA_RESTORE, rreg2, cfi_ops); |
3299 | insn = emit_insn (aarch64_gen_loadwb_pair | |
3300 | (mode1, stack_pointer_rtx, rreg1, | |
3301 | rreg2, offset)); | |
4b92caa1 | 3302 | } |
43e9d192 | 3303 | } |
43e9d192 IB |
3304 | else |
3305 | { | |
3306 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, | |
3307 | GEN_INT (offset))); | |
43e9d192 | 3308 | } |
43e9d192 | 3309 | |
dd991abb RH |
3310 | /* Reset the CFA to be SP + FRAME_SIZE. */ |
3311 | rtx new_cfa = stack_pointer_rtx; | |
3312 | if (frame_size > 0) | |
3313 | new_cfa = plus_constant (Pmode, new_cfa, frame_size); | |
3314 | cfi_ops = alloc_reg_note (REG_CFA_DEF_CFA, new_cfa, cfi_ops); | |
3315 | REG_NOTES (insn) = cfi_ops; | |
43e9d192 | 3316 | RTX_FRAME_RELATED_P (insn) = 1; |
43e9d192 IB |
3317 | } |
3318 | ||
dd991abb | 3319 | if (frame_size > 0) |
43e9d192 | 3320 | { |
7e8c2bd5 JW |
3321 | if (need_barrier_p) |
3322 | emit_insn (gen_stack_tie (stack_pointer_rtx, stack_pointer_rtx)); | |
3323 | ||
43e9d192 IB |
3324 | if (frame_size >= 0x1000000) |
3325 | { | |
3326 | rtx op0 = gen_rtx_REG (Pmode, IP0_REGNUM); | |
3327 | emit_move_insn (op0, GEN_INT (frame_size)); | |
dd991abb | 3328 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, op0)); |
43e9d192 | 3329 | } |
dd991abb | 3330 | else |
43e9d192 | 3331 | { |
dd991abb RH |
3332 | int hi_ofs = frame_size & 0xfff000; |
3333 | int lo_ofs = frame_size & 0x000fff; | |
3334 | ||
3335 | if (hi_ofs && lo_ofs) | |
43e9d192 IB |
3336 | { |
3337 | insn = emit_insn (gen_add2_insn | |
dd991abb | 3338 | (stack_pointer_rtx, GEN_INT (hi_ofs))); |
43e9d192 | 3339 | RTX_FRAME_RELATED_P (insn) = 1; |
dd991abb | 3340 | frame_size = lo_ofs; |
43e9d192 | 3341 | } |
dd991abb RH |
3342 | insn = emit_insn (gen_add2_insn |
3343 | (stack_pointer_rtx, GEN_INT (frame_size))); | |
43e9d192 IB |
3344 | } |
3345 | ||
dd991abb RH |
3346 | /* Reset the CFA to be SP + 0. */ |
3347 | add_reg_note (insn, REG_CFA_DEF_CFA, stack_pointer_rtx); | |
3348 | RTX_FRAME_RELATED_P (insn) = 1; | |
3349 | } | |
3350 | ||
3351 | /* Stack adjustment for exception handler. */ | |
3352 | if (crtl->calls_eh_return) | |
3353 | { | |
3354 | /* We need to unwind the stack by the offset computed by | |
3355 | EH_RETURN_STACKADJ_RTX. We have already reset the CFA | |
3356 | to be SP; letting the CFA move during this adjustment | |
3357 | is just as correct as retaining the CFA from the body | |
3358 | of the function. Therefore, do nothing special. */ | |
3359 | emit_insn (gen_add2_insn (stack_pointer_rtx, EH_RETURN_STACKADJ_RTX)); | |
43e9d192 IB |
3360 | } |
3361 | ||
3362 | emit_use (gen_rtx_REG (DImode, LR_REGNUM)); | |
3363 | if (!for_sibcall) | |
3364 | emit_jump_insn (ret_rtx); | |
3365 | } | |
3366 | ||
3367 | /* Return the place to copy the exception unwinding return address to. | |
3368 | This will probably be a stack slot, but could (in theory be the | |
3369 | return register). */ | |
3370 | rtx | |
3371 | aarch64_final_eh_return_addr (void) | |
3372 | { | |
1c960e02 MS |
3373 | HOST_WIDE_INT fp_offset; |
3374 | ||
43e9d192 | 3375 | aarch64_layout_frame (); |
1c960e02 MS |
3376 | |
3377 | fp_offset = cfun->machine->frame.frame_size | |
3378 | - cfun->machine->frame.hard_fp_offset; | |
43e9d192 IB |
3379 | |
3380 | if (cfun->machine->frame.reg_offset[LR_REGNUM] < 0) | |
3381 | return gen_rtx_REG (DImode, LR_REGNUM); | |
3382 | ||
3383 | /* DSE and CSELIB do not detect an alias between sp+k1 and fp+k2. This can | |
3384 | result in a store to save LR introduced by builtin_eh_return () being | |
3385 | incorrectly deleted because the alias is not detected. | |
3386 | So in the calculation of the address to copy the exception unwinding | |
3387 | return address to, we note 2 cases. | |
3388 | If FP is needed and the fp_offset is 0, it means that SP = FP and hence | |
3389 | we return a SP-relative location since all the addresses are SP-relative | |
3390 | in this case. This prevents the store from being optimized away. | |
3391 | If the fp_offset is not 0, then the addresses will be FP-relative and | |
3392 | therefore we return a FP-relative location. */ | |
3393 | ||
3394 | if (frame_pointer_needed) | |
3395 | { | |
3396 | if (fp_offset) | |
3397 | return gen_frame_mem (DImode, | |
3398 | plus_constant (Pmode, hard_frame_pointer_rtx, UNITS_PER_WORD)); | |
3399 | else | |
3400 | return gen_frame_mem (DImode, | |
3401 | plus_constant (Pmode, stack_pointer_rtx, UNITS_PER_WORD)); | |
3402 | } | |
3403 | ||
3404 | /* If FP is not needed, we calculate the location of LR, which would be | |
3405 | at the top of the saved registers block. */ | |
3406 | ||
3407 | return gen_frame_mem (DImode, | |
3408 | plus_constant (Pmode, | |
3409 | stack_pointer_rtx, | |
3410 | fp_offset | |
3411 | + cfun->machine->frame.saved_regs_size | |
3412 | - 2 * UNITS_PER_WORD)); | |
3413 | } | |
3414 | ||
43e9d192 | 3415 | static void |
f43657b4 JW |
3416 | aarch64_add_constant (machine_mode mode, int regnum, int scratchreg, |
3417 | HOST_WIDE_INT delta) | |
43e9d192 IB |
3418 | { |
3419 | HOST_WIDE_INT mdelta = delta; | |
f43657b4 JW |
3420 | rtx this_rtx = gen_rtx_REG (mode, regnum); |
3421 | rtx scratch_rtx = gen_rtx_REG (mode, scratchreg); | |
43e9d192 IB |
3422 | |
3423 | if (mdelta < 0) | |
3424 | mdelta = -mdelta; | |
3425 | ||
3426 | if (mdelta >= 4096 * 4096) | |
3427 | { | |
f43657b4 | 3428 | aarch64_internal_mov_immediate (scratch_rtx, GEN_INT (delta), true, mode); |
d9600ae5 | 3429 | emit_insn (gen_add3_insn (this_rtx, this_rtx, scratch_rtx)); |
43e9d192 IB |
3430 | } |
3431 | else if (mdelta > 0) | |
3432 | { | |
43e9d192 | 3433 | if (mdelta >= 4096) |
d9600ae5 | 3434 | { |
f7df4a84 | 3435 | emit_insn (gen_rtx_SET (scratch_rtx, GEN_INT (mdelta / 4096))); |
f43657b4 | 3436 | rtx shift = gen_rtx_ASHIFT (mode, scratch_rtx, GEN_INT (12)); |
d9600ae5 | 3437 | if (delta < 0) |
f7df4a84 | 3438 | emit_insn (gen_rtx_SET (this_rtx, |
f43657b4 | 3439 | gen_rtx_MINUS (mode, this_rtx, shift))); |
d9600ae5 | 3440 | else |
f7df4a84 | 3441 | emit_insn (gen_rtx_SET (this_rtx, |
f43657b4 | 3442 | gen_rtx_PLUS (mode, this_rtx, shift))); |
d9600ae5 | 3443 | } |
43e9d192 | 3444 | if (mdelta % 4096 != 0) |
d9600ae5 SN |
3445 | { |
3446 | scratch_rtx = GEN_INT ((delta < 0 ? -1 : 1) * (mdelta % 4096)); | |
f7df4a84 | 3447 | emit_insn (gen_rtx_SET (this_rtx, |
f43657b4 | 3448 | gen_rtx_PLUS (mode, this_rtx, scratch_rtx))); |
d9600ae5 | 3449 | } |
43e9d192 IB |
3450 | } |
3451 | } | |
3452 | ||
3453 | /* Output code to add DELTA to the first argument, and then jump | |
3454 | to FUNCTION. Used for C++ multiple inheritance. */ | |
3455 | static void | |
3456 | aarch64_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED, | |
3457 | HOST_WIDE_INT delta, | |
3458 | HOST_WIDE_INT vcall_offset, | |
3459 | tree function) | |
3460 | { | |
3461 | /* The this pointer is always in x0. Note that this differs from | |
3462 | Arm where the this pointer maybe bumped to r1 if r0 is required | |
3463 | to return a pointer to an aggregate. On AArch64 a result value | |
3464 | pointer will be in x8. */ | |
3465 | int this_regno = R0_REGNUM; | |
5d8a22a5 DM |
3466 | rtx this_rtx, temp0, temp1, addr, funexp; |
3467 | rtx_insn *insn; | |
43e9d192 | 3468 | |
75f1d6fc SN |
3469 | reload_completed = 1; |
3470 | emit_note (NOTE_INSN_PROLOGUE_END); | |
43e9d192 IB |
3471 | |
3472 | if (vcall_offset == 0) | |
f43657b4 | 3473 | aarch64_add_constant (Pmode, this_regno, IP1_REGNUM, delta); |
43e9d192 IB |
3474 | else |
3475 | { | |
28514dda | 3476 | gcc_assert ((vcall_offset & (POINTER_BYTES - 1)) == 0); |
43e9d192 | 3477 | |
75f1d6fc SN |
3478 | this_rtx = gen_rtx_REG (Pmode, this_regno); |
3479 | temp0 = gen_rtx_REG (Pmode, IP0_REGNUM); | |
3480 | temp1 = gen_rtx_REG (Pmode, IP1_REGNUM); | |
43e9d192 | 3481 | |
75f1d6fc SN |
3482 | addr = this_rtx; |
3483 | if (delta != 0) | |
3484 | { | |
3485 | if (delta >= -256 && delta < 256) | |
3486 | addr = gen_rtx_PRE_MODIFY (Pmode, this_rtx, | |
3487 | plus_constant (Pmode, this_rtx, delta)); | |
3488 | else | |
f43657b4 | 3489 | aarch64_add_constant (Pmode, this_regno, IP1_REGNUM, delta); |
43e9d192 IB |
3490 | } |
3491 | ||
28514dda YZ |
3492 | if (Pmode == ptr_mode) |
3493 | aarch64_emit_move (temp0, gen_rtx_MEM (ptr_mode, addr)); | |
3494 | else | |
3495 | aarch64_emit_move (temp0, | |
3496 | gen_rtx_ZERO_EXTEND (Pmode, | |
3497 | gen_rtx_MEM (ptr_mode, addr))); | |
75f1d6fc | 3498 | |
28514dda | 3499 | if (vcall_offset >= -256 && vcall_offset < 4096 * POINTER_BYTES) |
75f1d6fc | 3500 | addr = plus_constant (Pmode, temp0, vcall_offset); |
43e9d192 IB |
3501 | else |
3502 | { | |
f43657b4 JW |
3503 | aarch64_internal_mov_immediate (temp1, GEN_INT (vcall_offset), true, |
3504 | Pmode); | |
75f1d6fc | 3505 | addr = gen_rtx_PLUS (Pmode, temp0, temp1); |
43e9d192 IB |
3506 | } |
3507 | ||
28514dda YZ |
3508 | if (Pmode == ptr_mode) |
3509 | aarch64_emit_move (temp1, gen_rtx_MEM (ptr_mode,addr)); | |
3510 | else | |
3511 | aarch64_emit_move (temp1, | |
3512 | gen_rtx_SIGN_EXTEND (Pmode, | |
3513 | gen_rtx_MEM (ptr_mode, addr))); | |
3514 | ||
75f1d6fc | 3515 | emit_insn (gen_add2_insn (this_rtx, temp1)); |
43e9d192 IB |
3516 | } |
3517 | ||
75f1d6fc SN |
3518 | /* Generate a tail call to the target function. */ |
3519 | if (!TREE_USED (function)) | |
3520 | { | |
3521 | assemble_external (function); | |
3522 | TREE_USED (function) = 1; | |
3523 | } | |
3524 | funexp = XEXP (DECL_RTL (function), 0); | |
3525 | funexp = gen_rtx_MEM (FUNCTION_MODE, funexp); | |
3526 | insn = emit_call_insn (gen_sibcall (funexp, const0_rtx, NULL_RTX)); | |
3527 | SIBLING_CALL_P (insn) = 1; | |
3528 | ||
3529 | insn = get_insns (); | |
3530 | shorten_branches (insn); | |
3531 | final_start_function (insn, file, 1); | |
3532 | final (insn, file, 1); | |
43e9d192 | 3533 | final_end_function (); |
75f1d6fc SN |
3534 | |
3535 | /* Stop pretending to be a post-reload pass. */ | |
3536 | reload_completed = 0; | |
43e9d192 IB |
3537 | } |
3538 | ||
43e9d192 IB |
3539 | static bool |
3540 | aarch64_tls_referenced_p (rtx x) | |
3541 | { | |
3542 | if (!TARGET_HAVE_TLS) | |
3543 | return false; | |
e7de8563 RS |
3544 | subrtx_iterator::array_type array; |
3545 | FOR_EACH_SUBRTX (iter, array, x, ALL) | |
3546 | { | |
3547 | const_rtx x = *iter; | |
3548 | if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (x) != 0) | |
3549 | return true; | |
3550 | /* Don't recurse into UNSPEC_TLS looking for TLS symbols; these are | |
3551 | TLS offsets, not real symbol references. */ | |
3552 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
3553 | iter.skip_subrtxes (); | |
3554 | } | |
3555 | return false; | |
43e9d192 IB |
3556 | } |
3557 | ||
3558 | ||
43e9d192 IB |
3559 | /* Return true if val can be encoded as a 12-bit unsigned immediate with |
3560 | a left shift of 0 or 12 bits. */ | |
3561 | bool | |
3562 | aarch64_uimm12_shift (HOST_WIDE_INT val) | |
3563 | { | |
3564 | return ((val & (((HOST_WIDE_INT) 0xfff) << 0)) == val | |
3565 | || (val & (((HOST_WIDE_INT) 0xfff) << 12)) == val | |
3566 | ); | |
3567 | } | |
3568 | ||
3569 | ||
3570 | /* Return true if val is an immediate that can be loaded into a | |
3571 | register by a MOVZ instruction. */ | |
3572 | static bool | |
ef4bddc2 | 3573 | aarch64_movw_imm (HOST_WIDE_INT val, machine_mode mode) |
43e9d192 IB |
3574 | { |
3575 | if (GET_MODE_SIZE (mode) > 4) | |
3576 | { | |
3577 | if ((val & (((HOST_WIDE_INT) 0xffff) << 32)) == val | |
3578 | || (val & (((HOST_WIDE_INT) 0xffff) << 48)) == val) | |
3579 | return 1; | |
3580 | } | |
3581 | else | |
3582 | { | |
3583 | /* Ignore sign extension. */ | |
3584 | val &= (HOST_WIDE_INT) 0xffffffff; | |
3585 | } | |
3586 | return ((val & (((HOST_WIDE_INT) 0xffff) << 0)) == val | |
3587 | || (val & (((HOST_WIDE_INT) 0xffff) << 16)) == val); | |
3588 | } | |
3589 | ||
a64c73a2 WD |
3590 | /* Multipliers for repeating bitmasks of width 32, 16, 8, 4, and 2. */ |
3591 | ||
3592 | static const unsigned HOST_WIDE_INT bitmask_imm_mul[] = | |
3593 | { | |
3594 | 0x0000000100000001ull, | |
3595 | 0x0001000100010001ull, | |
3596 | 0x0101010101010101ull, | |
3597 | 0x1111111111111111ull, | |
3598 | 0x5555555555555555ull, | |
3599 | }; | |
3600 | ||
43e9d192 IB |
3601 | |
3602 | /* Return true if val is a valid bitmask immediate. */ | |
a64c73a2 | 3603 | |
43e9d192 | 3604 | bool |
a64c73a2 | 3605 | aarch64_bitmask_imm (HOST_WIDE_INT val_in, machine_mode mode) |
43e9d192 | 3606 | { |
a64c73a2 WD |
3607 | unsigned HOST_WIDE_INT val, tmp, mask, first_one, next_one; |
3608 | int bits; | |
3609 | ||
3610 | /* Check for a single sequence of one bits and return quickly if so. | |
3611 | The special cases of all ones and all zeroes returns false. */ | |
3612 | val = (unsigned HOST_WIDE_INT) val_in; | |
3613 | tmp = val + (val & -val); | |
3614 | ||
3615 | if (tmp == (tmp & -tmp)) | |
3616 | return (val + 1) > 1; | |
3617 | ||
3618 | /* Replicate 32-bit immediates so we can treat them as 64-bit. */ | |
3619 | if (mode == SImode) | |
3620 | val = (val << 32) | (val & 0xffffffff); | |
3621 | ||
3622 | /* Invert if the immediate doesn't start with a zero bit - this means we | |
3623 | only need to search for sequences of one bits. */ | |
3624 | if (val & 1) | |
3625 | val = ~val; | |
3626 | ||
3627 | /* Find the first set bit and set tmp to val with the first sequence of one | |
3628 | bits removed. Return success if there is a single sequence of ones. */ | |
3629 | first_one = val & -val; | |
3630 | tmp = val & (val + first_one); | |
3631 | ||
3632 | if (tmp == 0) | |
3633 | return true; | |
3634 | ||
3635 | /* Find the next set bit and compute the difference in bit position. */ | |
3636 | next_one = tmp & -tmp; | |
3637 | bits = clz_hwi (first_one) - clz_hwi (next_one); | |
3638 | mask = val ^ tmp; | |
3639 | ||
3640 | /* Check the bit position difference is a power of 2, and that the first | |
3641 | sequence of one bits fits within 'bits' bits. */ | |
3642 | if ((mask >> bits) != 0 || bits != (bits & -bits)) | |
3643 | return false; | |
3644 | ||
3645 | /* Check the sequence of one bits is repeated 64/bits times. */ | |
3646 | return val == mask * bitmask_imm_mul[__builtin_clz (bits) - 26]; | |
43e9d192 IB |
3647 | } |
3648 | ||
3649 | ||
3650 | /* Return true if val is an immediate that can be loaded into a | |
3651 | register in a single instruction. */ | |
3652 | bool | |
ef4bddc2 | 3653 | aarch64_move_imm (HOST_WIDE_INT val, machine_mode mode) |
43e9d192 IB |
3654 | { |
3655 | if (aarch64_movw_imm (val, mode) || aarch64_movw_imm (~val, mode)) | |
3656 | return 1; | |
3657 | return aarch64_bitmask_imm (val, mode); | |
3658 | } | |
3659 | ||
3660 | static bool | |
ef4bddc2 | 3661 | aarch64_cannot_force_const_mem (machine_mode mode ATTRIBUTE_UNUSED, rtx x) |
43e9d192 IB |
3662 | { |
3663 | rtx base, offset; | |
7eda14e1 | 3664 | |
43e9d192 IB |
3665 | if (GET_CODE (x) == HIGH) |
3666 | return true; | |
3667 | ||
3668 | split_const (x, &base, &offset); | |
3669 | if (GET_CODE (base) == SYMBOL_REF || GET_CODE (base) == LABEL_REF) | |
28514dda | 3670 | { |
a6e0bfa7 | 3671 | if (aarch64_classify_symbol (base, offset) |
28514dda YZ |
3672 | != SYMBOL_FORCE_TO_MEM) |
3673 | return true; | |
3674 | else | |
3675 | /* Avoid generating a 64-bit relocation in ILP32; leave | |
3676 | to aarch64_expand_mov_immediate to handle it properly. */ | |
3677 | return mode != ptr_mode; | |
3678 | } | |
43e9d192 IB |
3679 | |
3680 | return aarch64_tls_referenced_p (x); | |
3681 | } | |
3682 | ||
e79136e4 WD |
3683 | /* Implement TARGET_CASE_VALUES_THRESHOLD. |
3684 | The expansion for a table switch is quite expensive due to the number | |
3685 | of instructions, the table lookup and hard to predict indirect jump. | |
3686 | When optimizing for speed, and -O3 enabled, use the per-core tuning if | |
3687 | set, otherwise use tables for > 16 cases as a tradeoff between size and | |
3688 | performance. When optimizing for size, use the default setting. */ | |
50487d79 EM |
3689 | |
3690 | static unsigned int | |
3691 | aarch64_case_values_threshold (void) | |
3692 | { | |
3693 | /* Use the specified limit for the number of cases before using jump | |
3694 | tables at higher optimization levels. */ | |
3695 | if (optimize > 2 | |
3696 | && selected_cpu->tune->max_case_values != 0) | |
3697 | return selected_cpu->tune->max_case_values; | |
3698 | else | |
e79136e4 | 3699 | return optimize_size ? default_case_values_threshold () : 17; |
50487d79 EM |
3700 | } |
3701 | ||
43e9d192 IB |
3702 | /* Return true if register REGNO is a valid index register. |
3703 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
3704 | ||
3705 | bool | |
3706 | aarch64_regno_ok_for_index_p (int regno, bool strict_p) | |
3707 | { | |
3708 | if (!HARD_REGISTER_NUM_P (regno)) | |
3709 | { | |
3710 | if (!strict_p) | |
3711 | return true; | |
3712 | ||
3713 | if (!reg_renumber) | |
3714 | return false; | |
3715 | ||
3716 | regno = reg_renumber[regno]; | |
3717 | } | |
3718 | return GP_REGNUM_P (regno); | |
3719 | } | |
3720 | ||
3721 | /* Return true if register REGNO is a valid base register for mode MODE. | |
3722 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
3723 | ||
3724 | bool | |
3725 | aarch64_regno_ok_for_base_p (int regno, bool strict_p) | |
3726 | { | |
3727 | if (!HARD_REGISTER_NUM_P (regno)) | |
3728 | { | |
3729 | if (!strict_p) | |
3730 | return true; | |
3731 | ||
3732 | if (!reg_renumber) | |
3733 | return false; | |
3734 | ||
3735 | regno = reg_renumber[regno]; | |
3736 | } | |
3737 | ||
3738 | /* The fake registers will be eliminated to either the stack or | |
3739 | hard frame pointer, both of which are usually valid base registers. | |
3740 | Reload deals with the cases where the eliminated form isn't valid. */ | |
3741 | return (GP_REGNUM_P (regno) | |
3742 | || regno == SP_REGNUM | |
3743 | || regno == FRAME_POINTER_REGNUM | |
3744 | || regno == ARG_POINTER_REGNUM); | |
3745 | } | |
3746 | ||
3747 | /* Return true if X is a valid base register for mode MODE. | |
3748 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
3749 | ||
3750 | static bool | |
3751 | aarch64_base_register_rtx_p (rtx x, bool strict_p) | |
3752 | { | |
3753 | if (!strict_p && GET_CODE (x) == SUBREG) | |
3754 | x = SUBREG_REG (x); | |
3755 | ||
3756 | return (REG_P (x) && aarch64_regno_ok_for_base_p (REGNO (x), strict_p)); | |
3757 | } | |
3758 | ||
3759 | /* Return true if address offset is a valid index. If it is, fill in INFO | |
3760 | appropriately. STRICT_P is true if REG_OK_STRICT is in effect. */ | |
3761 | ||
3762 | static bool | |
3763 | aarch64_classify_index (struct aarch64_address_info *info, rtx x, | |
ef4bddc2 | 3764 | machine_mode mode, bool strict_p) |
43e9d192 IB |
3765 | { |
3766 | enum aarch64_address_type type; | |
3767 | rtx index; | |
3768 | int shift; | |
3769 | ||
3770 | /* (reg:P) */ | |
3771 | if ((REG_P (x) || GET_CODE (x) == SUBREG) | |
3772 | && GET_MODE (x) == Pmode) | |
3773 | { | |
3774 | type = ADDRESS_REG_REG; | |
3775 | index = x; | |
3776 | shift = 0; | |
3777 | } | |
3778 | /* (sign_extend:DI (reg:SI)) */ | |
3779 | else if ((GET_CODE (x) == SIGN_EXTEND | |
3780 | || GET_CODE (x) == ZERO_EXTEND) | |
3781 | && GET_MODE (x) == DImode | |
3782 | && GET_MODE (XEXP (x, 0)) == SImode) | |
3783 | { | |
3784 | type = (GET_CODE (x) == SIGN_EXTEND) | |
3785 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3786 | index = XEXP (x, 0); | |
3787 | shift = 0; | |
3788 | } | |
3789 | /* (mult:DI (sign_extend:DI (reg:SI)) (const_int scale)) */ | |
3790 | else if (GET_CODE (x) == MULT | |
3791 | && (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
3792 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
3793 | && GET_MODE (XEXP (x, 0)) == DImode | |
3794 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == SImode | |
3795 | && CONST_INT_P (XEXP (x, 1))) | |
3796 | { | |
3797 | type = (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
3798 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3799 | index = XEXP (XEXP (x, 0), 0); | |
3800 | shift = exact_log2 (INTVAL (XEXP (x, 1))); | |
3801 | } | |
3802 | /* (ashift:DI (sign_extend:DI (reg:SI)) (const_int shift)) */ | |
3803 | else if (GET_CODE (x) == ASHIFT | |
3804 | && (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
3805 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
3806 | && GET_MODE (XEXP (x, 0)) == DImode | |
3807 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == SImode | |
3808 | && CONST_INT_P (XEXP (x, 1))) | |
3809 | { | |
3810 | type = (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
3811 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3812 | index = XEXP (XEXP (x, 0), 0); | |
3813 | shift = INTVAL (XEXP (x, 1)); | |
3814 | } | |
3815 | /* (sign_extract:DI (mult:DI (reg:DI) (const_int scale)) 32+shift 0) */ | |
3816 | else if ((GET_CODE (x) == SIGN_EXTRACT | |
3817 | || GET_CODE (x) == ZERO_EXTRACT) | |
3818 | && GET_MODE (x) == DImode | |
3819 | && GET_CODE (XEXP (x, 0)) == MULT | |
3820 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3821 | && CONST_INT_P (XEXP (XEXP (x, 0), 1))) | |
3822 | { | |
3823 | type = (GET_CODE (x) == SIGN_EXTRACT) | |
3824 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3825 | index = XEXP (XEXP (x, 0), 0); | |
3826 | shift = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1))); | |
3827 | if (INTVAL (XEXP (x, 1)) != 32 + shift | |
3828 | || INTVAL (XEXP (x, 2)) != 0) | |
3829 | shift = -1; | |
3830 | } | |
3831 | /* (and:DI (mult:DI (reg:DI) (const_int scale)) | |
3832 | (const_int 0xffffffff<<shift)) */ | |
3833 | else if (GET_CODE (x) == AND | |
3834 | && GET_MODE (x) == DImode | |
3835 | && GET_CODE (XEXP (x, 0)) == MULT | |
3836 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3837 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
3838 | && CONST_INT_P (XEXP (x, 1))) | |
3839 | { | |
3840 | type = ADDRESS_REG_UXTW; | |
3841 | index = XEXP (XEXP (x, 0), 0); | |
3842 | shift = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1))); | |
3843 | if (INTVAL (XEXP (x, 1)) != (HOST_WIDE_INT)0xffffffff << shift) | |
3844 | shift = -1; | |
3845 | } | |
3846 | /* (sign_extract:DI (ashift:DI (reg:DI) (const_int shift)) 32+shift 0) */ | |
3847 | else if ((GET_CODE (x) == SIGN_EXTRACT | |
3848 | || GET_CODE (x) == ZERO_EXTRACT) | |
3849 | && GET_MODE (x) == DImode | |
3850 | && GET_CODE (XEXP (x, 0)) == ASHIFT | |
3851 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3852 | && CONST_INT_P (XEXP (XEXP (x, 0), 1))) | |
3853 | { | |
3854 | type = (GET_CODE (x) == SIGN_EXTRACT) | |
3855 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3856 | index = XEXP (XEXP (x, 0), 0); | |
3857 | shift = INTVAL (XEXP (XEXP (x, 0), 1)); | |
3858 | if (INTVAL (XEXP (x, 1)) != 32 + shift | |
3859 | || INTVAL (XEXP (x, 2)) != 0) | |
3860 | shift = -1; | |
3861 | } | |
3862 | /* (and:DI (ashift:DI (reg:DI) (const_int shift)) | |
3863 | (const_int 0xffffffff<<shift)) */ | |
3864 | else if (GET_CODE (x) == AND | |
3865 | && GET_MODE (x) == DImode | |
3866 | && GET_CODE (XEXP (x, 0)) == ASHIFT | |
3867 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3868 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
3869 | && CONST_INT_P (XEXP (x, 1))) | |
3870 | { | |
3871 | type = ADDRESS_REG_UXTW; | |
3872 | index = XEXP (XEXP (x, 0), 0); | |
3873 | shift = INTVAL (XEXP (XEXP (x, 0), 1)); | |
3874 | if (INTVAL (XEXP (x, 1)) != (HOST_WIDE_INT)0xffffffff << shift) | |
3875 | shift = -1; | |
3876 | } | |
3877 | /* (mult:P (reg:P) (const_int scale)) */ | |
3878 | else if (GET_CODE (x) == MULT | |
3879 | && GET_MODE (x) == Pmode | |
3880 | && GET_MODE (XEXP (x, 0)) == Pmode | |
3881 | && CONST_INT_P (XEXP (x, 1))) | |
3882 | { | |
3883 | type = ADDRESS_REG_REG; | |
3884 | index = XEXP (x, 0); | |
3885 | shift = exact_log2 (INTVAL (XEXP (x, 1))); | |
3886 | } | |
3887 | /* (ashift:P (reg:P) (const_int shift)) */ | |
3888 | else if (GET_CODE (x) == ASHIFT | |
3889 | && GET_MODE (x) == Pmode | |
3890 | && GET_MODE (XEXP (x, 0)) == Pmode | |
3891 | && CONST_INT_P (XEXP (x, 1))) | |
3892 | { | |
3893 | type = ADDRESS_REG_REG; | |
3894 | index = XEXP (x, 0); | |
3895 | shift = INTVAL (XEXP (x, 1)); | |
3896 | } | |
3897 | else | |
3898 | return false; | |
3899 | ||
3900 | if (GET_CODE (index) == SUBREG) | |
3901 | index = SUBREG_REG (index); | |
3902 | ||
3903 | if ((shift == 0 || | |
3904 | (shift > 0 && shift <= 3 | |
3905 | && (1 << shift) == GET_MODE_SIZE (mode))) | |
3906 | && REG_P (index) | |
3907 | && aarch64_regno_ok_for_index_p (REGNO (index), strict_p)) | |
3908 | { | |
3909 | info->type = type; | |
3910 | info->offset = index; | |
3911 | info->shift = shift; | |
3912 | return true; | |
3913 | } | |
3914 | ||
3915 | return false; | |
3916 | } | |
3917 | ||
44707478 | 3918 | bool |
ef4bddc2 | 3919 | aarch64_offset_7bit_signed_scaled_p (machine_mode mode, HOST_WIDE_INT offset) |
43e9d192 IB |
3920 | { |
3921 | return (offset >= -64 * GET_MODE_SIZE (mode) | |
3922 | && offset < 64 * GET_MODE_SIZE (mode) | |
3923 | && offset % GET_MODE_SIZE (mode) == 0); | |
3924 | } | |
3925 | ||
3926 | static inline bool | |
ef4bddc2 | 3927 | offset_9bit_signed_unscaled_p (machine_mode mode ATTRIBUTE_UNUSED, |
43e9d192 IB |
3928 | HOST_WIDE_INT offset) |
3929 | { | |
3930 | return offset >= -256 && offset < 256; | |
3931 | } | |
3932 | ||
3933 | static inline bool | |
ef4bddc2 | 3934 | offset_12bit_unsigned_scaled_p (machine_mode mode, HOST_WIDE_INT offset) |
43e9d192 IB |
3935 | { |
3936 | return (offset >= 0 | |
3937 | && offset < 4096 * GET_MODE_SIZE (mode) | |
3938 | && offset % GET_MODE_SIZE (mode) == 0); | |
3939 | } | |
3940 | ||
abc52318 KT |
3941 | /* Return true if MODE is one of the modes for which we |
3942 | support LDP/STP operations. */ | |
3943 | ||
3944 | static bool | |
3945 | aarch64_mode_valid_for_sched_fusion_p (machine_mode mode) | |
3946 | { | |
3947 | return mode == SImode || mode == DImode | |
3948 | || mode == SFmode || mode == DFmode | |
3949 | || (aarch64_vector_mode_supported_p (mode) | |
3950 | && GET_MODE_SIZE (mode) == 8); | |
3951 | } | |
3952 | ||
9e0218fc RH |
3953 | /* Return true if REGNO is a virtual pointer register, or an eliminable |
3954 | "soft" frame register. Like REGNO_PTR_FRAME_P except that we don't | |
3955 | include stack_pointer or hard_frame_pointer. */ | |
3956 | static bool | |
3957 | virt_or_elim_regno_p (unsigned regno) | |
3958 | { | |
3959 | return ((regno >= FIRST_VIRTUAL_REGISTER | |
3960 | && regno <= LAST_VIRTUAL_POINTER_REGISTER) | |
3961 | || regno == FRAME_POINTER_REGNUM | |
3962 | || regno == ARG_POINTER_REGNUM); | |
3963 | } | |
3964 | ||
43e9d192 IB |
3965 | /* Return true if X is a valid address for machine mode MODE. If it is, |
3966 | fill in INFO appropriately. STRICT_P is true if REG_OK_STRICT is in | |
3967 | effect. OUTER_CODE is PARALLEL for a load/store pair. */ | |
3968 | ||
3969 | static bool | |
3970 | aarch64_classify_address (struct aarch64_address_info *info, | |
ef4bddc2 | 3971 | rtx x, machine_mode mode, |
43e9d192 IB |
3972 | RTX_CODE outer_code, bool strict_p) |
3973 | { | |
3974 | enum rtx_code code = GET_CODE (x); | |
3975 | rtx op0, op1; | |
2d8c6dc1 AH |
3976 | |
3977 | /* On BE, we use load/store pair for all large int mode load/stores. */ | |
3978 | bool load_store_pair_p = (outer_code == PARALLEL | |
3979 | || (BYTES_BIG_ENDIAN | |
3980 | && aarch64_vect_struct_mode_p (mode))); | |
3981 | ||
43e9d192 | 3982 | bool allow_reg_index_p = |
2d8c6dc1 AH |
3983 | !load_store_pair_p |
3984 | && (GET_MODE_SIZE (mode) != 16 || aarch64_vector_mode_supported_p (mode)) | |
3985 | && !aarch64_vect_struct_mode_p (mode); | |
3986 | ||
3987 | /* On LE, for AdvSIMD, don't support anything other than POST_INC or | |
3988 | REG addressing. */ | |
3989 | if (aarch64_vect_struct_mode_p (mode) && !BYTES_BIG_ENDIAN | |
43e9d192 IB |
3990 | && (code != POST_INC && code != REG)) |
3991 | return false; | |
3992 | ||
3993 | switch (code) | |
3994 | { | |
3995 | case REG: | |
3996 | case SUBREG: | |
3997 | info->type = ADDRESS_REG_IMM; | |
3998 | info->base = x; | |
3999 | info->offset = const0_rtx; | |
4000 | return aarch64_base_register_rtx_p (x, strict_p); | |
4001 | ||
4002 | case PLUS: | |
4003 | op0 = XEXP (x, 0); | |
4004 | op1 = XEXP (x, 1); | |
15c0c5c9 JW |
4005 | |
4006 | if (! strict_p | |
4aa81c2e | 4007 | && REG_P (op0) |
9e0218fc | 4008 | && virt_or_elim_regno_p (REGNO (op0)) |
4aa81c2e | 4009 | && CONST_INT_P (op1)) |
15c0c5c9 JW |
4010 | { |
4011 | info->type = ADDRESS_REG_IMM; | |
4012 | info->base = op0; | |
4013 | info->offset = op1; | |
4014 | ||
4015 | return true; | |
4016 | } | |
4017 | ||
43e9d192 IB |
4018 | if (GET_MODE_SIZE (mode) != 0 |
4019 | && CONST_INT_P (op1) | |
4020 | && aarch64_base_register_rtx_p (op0, strict_p)) | |
4021 | { | |
4022 | HOST_WIDE_INT offset = INTVAL (op1); | |
4023 | ||
4024 | info->type = ADDRESS_REG_IMM; | |
4025 | info->base = op0; | |
4026 | info->offset = op1; | |
4027 | ||
4028 | /* TImode and TFmode values are allowed in both pairs of X | |
4029 | registers and individual Q registers. The available | |
4030 | address modes are: | |
4031 | X,X: 7-bit signed scaled offset | |
4032 | Q: 9-bit signed offset | |
4033 | We conservatively require an offset representable in either mode. | |
4034 | */ | |
4035 | if (mode == TImode || mode == TFmode) | |
44707478 | 4036 | return (aarch64_offset_7bit_signed_scaled_p (mode, offset) |
43e9d192 IB |
4037 | && offset_9bit_signed_unscaled_p (mode, offset)); |
4038 | ||
2d8c6dc1 AH |
4039 | /* A 7bit offset check because OImode will emit a ldp/stp |
4040 | instruction (only big endian will get here). | |
4041 | For ldp/stp instructions, the offset is scaled for the size of a | |
4042 | single element of the pair. */ | |
4043 | if (mode == OImode) | |
4044 | return aarch64_offset_7bit_signed_scaled_p (TImode, offset); | |
4045 | ||
4046 | /* Three 9/12 bit offsets checks because CImode will emit three | |
4047 | ldr/str instructions (only big endian will get here). */ | |
4048 | if (mode == CImode) | |
4049 | return (aarch64_offset_7bit_signed_scaled_p (TImode, offset) | |
4050 | && (offset_9bit_signed_unscaled_p (V16QImode, offset + 32) | |
4051 | || offset_12bit_unsigned_scaled_p (V16QImode, | |
4052 | offset + 32))); | |
4053 | ||
4054 | /* Two 7bit offsets checks because XImode will emit two ldp/stp | |
4055 | instructions (only big endian will get here). */ | |
4056 | if (mode == XImode) | |
4057 | return (aarch64_offset_7bit_signed_scaled_p (TImode, offset) | |
4058 | && aarch64_offset_7bit_signed_scaled_p (TImode, | |
4059 | offset + 32)); | |
4060 | ||
4061 | if (load_store_pair_p) | |
43e9d192 | 4062 | return ((GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) |
44707478 | 4063 | && aarch64_offset_7bit_signed_scaled_p (mode, offset)); |
43e9d192 IB |
4064 | else |
4065 | return (offset_9bit_signed_unscaled_p (mode, offset) | |
4066 | || offset_12bit_unsigned_scaled_p (mode, offset)); | |
4067 | } | |
4068 | ||
4069 | if (allow_reg_index_p) | |
4070 | { | |
4071 | /* Look for base + (scaled/extended) index register. */ | |
4072 | if (aarch64_base_register_rtx_p (op0, strict_p) | |
4073 | && aarch64_classify_index (info, op1, mode, strict_p)) | |
4074 | { | |
4075 | info->base = op0; | |
4076 | return true; | |
4077 | } | |
4078 | if (aarch64_base_register_rtx_p (op1, strict_p) | |
4079 | && aarch64_classify_index (info, op0, mode, strict_p)) | |
4080 | { | |
4081 | info->base = op1; | |
4082 | return true; | |
4083 | } | |
4084 | } | |
4085 | ||
4086 | return false; | |
4087 | ||
4088 | case POST_INC: | |
4089 | case POST_DEC: | |
4090 | case PRE_INC: | |
4091 | case PRE_DEC: | |
4092 | info->type = ADDRESS_REG_WB; | |
4093 | info->base = XEXP (x, 0); | |
4094 | info->offset = NULL_RTX; | |
4095 | return aarch64_base_register_rtx_p (info->base, strict_p); | |
4096 | ||
4097 | case POST_MODIFY: | |
4098 | case PRE_MODIFY: | |
4099 | info->type = ADDRESS_REG_WB; | |
4100 | info->base = XEXP (x, 0); | |
4101 | if (GET_CODE (XEXP (x, 1)) == PLUS | |
4102 | && CONST_INT_P (XEXP (XEXP (x, 1), 1)) | |
4103 | && rtx_equal_p (XEXP (XEXP (x, 1), 0), info->base) | |
4104 | && aarch64_base_register_rtx_p (info->base, strict_p)) | |
4105 | { | |
4106 | HOST_WIDE_INT offset; | |
4107 | info->offset = XEXP (XEXP (x, 1), 1); | |
4108 | offset = INTVAL (info->offset); | |
4109 | ||
4110 | /* TImode and TFmode values are allowed in both pairs of X | |
4111 | registers and individual Q registers. The available | |
4112 | address modes are: | |
4113 | X,X: 7-bit signed scaled offset | |
4114 | Q: 9-bit signed offset | |
4115 | We conservatively require an offset representable in either mode. | |
4116 | */ | |
4117 | if (mode == TImode || mode == TFmode) | |
44707478 | 4118 | return (aarch64_offset_7bit_signed_scaled_p (mode, offset) |
43e9d192 IB |
4119 | && offset_9bit_signed_unscaled_p (mode, offset)); |
4120 | ||
2d8c6dc1 | 4121 | if (load_store_pair_p) |
43e9d192 | 4122 | return ((GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) |
44707478 | 4123 | && aarch64_offset_7bit_signed_scaled_p (mode, offset)); |
43e9d192 IB |
4124 | else |
4125 | return offset_9bit_signed_unscaled_p (mode, offset); | |
4126 | } | |
4127 | return false; | |
4128 | ||
4129 | case CONST: | |
4130 | case SYMBOL_REF: | |
4131 | case LABEL_REF: | |
79517551 SN |
4132 | /* load literal: pc-relative constant pool entry. Only supported |
4133 | for SI mode or larger. */ | |
43e9d192 | 4134 | info->type = ADDRESS_SYMBOLIC; |
2d8c6dc1 AH |
4135 | |
4136 | if (!load_store_pair_p && GET_MODE_SIZE (mode) >= 4) | |
43e9d192 IB |
4137 | { |
4138 | rtx sym, addend; | |
4139 | ||
4140 | split_const (x, &sym, &addend); | |
b4f50fd4 RR |
4141 | return ((GET_CODE (sym) == LABEL_REF |
4142 | || (GET_CODE (sym) == SYMBOL_REF | |
4143 | && CONSTANT_POOL_ADDRESS_P (sym) | |
9ee6540a | 4144 | && aarch64_pcrelative_literal_loads))); |
43e9d192 IB |
4145 | } |
4146 | return false; | |
4147 | ||
4148 | case LO_SUM: | |
4149 | info->type = ADDRESS_LO_SUM; | |
4150 | info->base = XEXP (x, 0); | |
4151 | info->offset = XEXP (x, 1); | |
4152 | if (allow_reg_index_p | |
4153 | && aarch64_base_register_rtx_p (info->base, strict_p)) | |
4154 | { | |
4155 | rtx sym, offs; | |
4156 | split_const (info->offset, &sym, &offs); | |
4157 | if (GET_CODE (sym) == SYMBOL_REF | |
a6e0bfa7 | 4158 | && (aarch64_classify_symbol (sym, offs) == SYMBOL_SMALL_ABSOLUTE)) |
43e9d192 IB |
4159 | { |
4160 | /* The symbol and offset must be aligned to the access size. */ | |
4161 | unsigned int align; | |
4162 | unsigned int ref_size; | |
4163 | ||
4164 | if (CONSTANT_POOL_ADDRESS_P (sym)) | |
4165 | align = GET_MODE_ALIGNMENT (get_pool_mode (sym)); | |
4166 | else if (TREE_CONSTANT_POOL_ADDRESS_P (sym)) | |
4167 | { | |
4168 | tree exp = SYMBOL_REF_DECL (sym); | |
4169 | align = TYPE_ALIGN (TREE_TYPE (exp)); | |
4170 | align = CONSTANT_ALIGNMENT (exp, align); | |
4171 | } | |
4172 | else if (SYMBOL_REF_DECL (sym)) | |
4173 | align = DECL_ALIGN (SYMBOL_REF_DECL (sym)); | |
6c031d8d KV |
4174 | else if (SYMBOL_REF_HAS_BLOCK_INFO_P (sym) |
4175 | && SYMBOL_REF_BLOCK (sym) != NULL) | |
4176 | align = SYMBOL_REF_BLOCK (sym)->alignment; | |
43e9d192 IB |
4177 | else |
4178 | align = BITS_PER_UNIT; | |
4179 | ||
4180 | ref_size = GET_MODE_SIZE (mode); | |
4181 | if (ref_size == 0) | |
4182 | ref_size = GET_MODE_SIZE (DImode); | |
4183 | ||
4184 | return ((INTVAL (offs) & (ref_size - 1)) == 0 | |
4185 | && ((align / BITS_PER_UNIT) & (ref_size - 1)) == 0); | |
4186 | } | |
4187 | } | |
4188 | return false; | |
4189 | ||
4190 | default: | |
4191 | return false; | |
4192 | } | |
4193 | } | |
4194 | ||
4195 | bool | |
4196 | aarch64_symbolic_address_p (rtx x) | |
4197 | { | |
4198 | rtx offset; | |
4199 | ||
4200 | split_const (x, &x, &offset); | |
4201 | return GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == LABEL_REF; | |
4202 | } | |
4203 | ||
a6e0bfa7 | 4204 | /* Classify the base of symbolic expression X. */ |
da4f13a4 MS |
4205 | |
4206 | enum aarch64_symbol_type | |
a6e0bfa7 | 4207 | aarch64_classify_symbolic_expression (rtx x) |
43e9d192 IB |
4208 | { |
4209 | rtx offset; | |
da4f13a4 | 4210 | |
43e9d192 | 4211 | split_const (x, &x, &offset); |
a6e0bfa7 | 4212 | return aarch64_classify_symbol (x, offset); |
43e9d192 IB |
4213 | } |
4214 | ||
4215 | ||
4216 | /* Return TRUE if X is a legitimate address for accessing memory in | |
4217 | mode MODE. */ | |
4218 | static bool | |
ef4bddc2 | 4219 | aarch64_legitimate_address_hook_p (machine_mode mode, rtx x, bool strict_p) |
43e9d192 IB |
4220 | { |
4221 | struct aarch64_address_info addr; | |
4222 | ||
4223 | return aarch64_classify_address (&addr, x, mode, MEM, strict_p); | |
4224 | } | |
4225 | ||
4226 | /* Return TRUE if X is a legitimate address for accessing memory in | |
4227 | mode MODE. OUTER_CODE will be PARALLEL if this is a load/store | |
4228 | pair operation. */ | |
4229 | bool | |
ef4bddc2 | 4230 | aarch64_legitimate_address_p (machine_mode mode, rtx x, |
aef66c94 | 4231 | RTX_CODE outer_code, bool strict_p) |
43e9d192 IB |
4232 | { |
4233 | struct aarch64_address_info addr; | |
4234 | ||
4235 | return aarch64_classify_address (&addr, x, mode, outer_code, strict_p); | |
4236 | } | |
4237 | ||
4238 | /* Return TRUE if rtx X is immediate constant 0.0 */ | |
4239 | bool | |
3520f7cc | 4240 | aarch64_float_const_zero_rtx_p (rtx x) |
43e9d192 | 4241 | { |
43e9d192 IB |
4242 | if (GET_MODE (x) == VOIDmode) |
4243 | return false; | |
4244 | ||
34a72c33 | 4245 | if (REAL_VALUE_MINUS_ZERO (*CONST_DOUBLE_REAL_VALUE (x))) |
43e9d192 | 4246 | return !HONOR_SIGNED_ZEROS (GET_MODE (x)); |
34a72c33 | 4247 | return real_equal (CONST_DOUBLE_REAL_VALUE (x), &dconst0); |
43e9d192 IB |
4248 | } |
4249 | ||
70f09188 AP |
4250 | /* Return the fixed registers used for condition codes. */ |
4251 | ||
4252 | static bool | |
4253 | aarch64_fixed_condition_code_regs (unsigned int *p1, unsigned int *p2) | |
4254 | { | |
4255 | *p1 = CC_REGNUM; | |
4256 | *p2 = INVALID_REGNUM; | |
4257 | return true; | |
4258 | } | |
4259 | ||
78607708 TV |
4260 | /* Emit call insn with PAT and do aarch64-specific handling. */ |
4261 | ||
d07a3fed | 4262 | void |
78607708 TV |
4263 | aarch64_emit_call_insn (rtx pat) |
4264 | { | |
4265 | rtx insn = emit_call_insn (pat); | |
4266 | ||
4267 | rtx *fusage = &CALL_INSN_FUNCTION_USAGE (insn); | |
4268 | clobber_reg (fusage, gen_rtx_REG (word_mode, IP0_REGNUM)); | |
4269 | clobber_reg (fusage, gen_rtx_REG (word_mode, IP1_REGNUM)); | |
4270 | } | |
4271 | ||
ef4bddc2 | 4272 | machine_mode |
43e9d192 IB |
4273 | aarch64_select_cc_mode (RTX_CODE code, rtx x, rtx y) |
4274 | { | |
4275 | /* All floating point compares return CCFP if it is an equality | |
4276 | comparison, and CCFPE otherwise. */ | |
4277 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
4278 | { | |
4279 | switch (code) | |
4280 | { | |
4281 | case EQ: | |
4282 | case NE: | |
4283 | case UNORDERED: | |
4284 | case ORDERED: | |
4285 | case UNLT: | |
4286 | case UNLE: | |
4287 | case UNGT: | |
4288 | case UNGE: | |
4289 | case UNEQ: | |
4290 | case LTGT: | |
4291 | return CCFPmode; | |
4292 | ||
4293 | case LT: | |
4294 | case LE: | |
4295 | case GT: | |
4296 | case GE: | |
4297 | return CCFPEmode; | |
4298 | ||
4299 | default: | |
4300 | gcc_unreachable (); | |
4301 | } | |
4302 | } | |
4303 | ||
2b8568fe KT |
4304 | /* Equality comparisons of short modes against zero can be performed |
4305 | using the TST instruction with the appropriate bitmask. */ | |
4306 | if (y == const0_rtx && REG_P (x) | |
4307 | && (code == EQ || code == NE) | |
4308 | && (GET_MODE (x) == HImode || GET_MODE (x) == QImode)) | |
4309 | return CC_NZmode; | |
4310 | ||
43e9d192 IB |
4311 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) |
4312 | && y == const0_rtx | |
4313 | && (code == EQ || code == NE || code == LT || code == GE) | |
b056c910 | 4314 | && (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS || GET_CODE (x) == AND |
7325d85a KT |
4315 | || GET_CODE (x) == NEG |
4316 | || (GET_CODE (x) == ZERO_EXTRACT && CONST_INT_P (XEXP (x, 1)) | |
4317 | && CONST_INT_P (XEXP (x, 2))))) | |
43e9d192 IB |
4318 | return CC_NZmode; |
4319 | ||
1c992d1e | 4320 | /* A compare with a shifted operand. Because of canonicalization, |
43e9d192 IB |
4321 | the comparison will have to be swapped when we emit the assembly |
4322 | code. */ | |
4323 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
4aa81c2e | 4324 | && (REG_P (y) || GET_CODE (y) == SUBREG) |
43e9d192 IB |
4325 | && (GET_CODE (x) == ASHIFT || GET_CODE (x) == ASHIFTRT |
4326 | || GET_CODE (x) == LSHIFTRT | |
1c992d1e | 4327 | || GET_CODE (x) == ZERO_EXTEND || GET_CODE (x) == SIGN_EXTEND)) |
43e9d192 IB |
4328 | return CC_SWPmode; |
4329 | ||
1c992d1e RE |
4330 | /* Similarly for a negated operand, but we can only do this for |
4331 | equalities. */ | |
4332 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
4aa81c2e | 4333 | && (REG_P (y) || GET_CODE (y) == SUBREG) |
1c992d1e RE |
4334 | && (code == EQ || code == NE) |
4335 | && GET_CODE (x) == NEG) | |
4336 | return CC_Zmode; | |
4337 | ||
ef22810a RH |
4338 | /* A test for unsigned overflow. */ |
4339 | if ((GET_MODE (x) == DImode || GET_MODE (x) == TImode) | |
4340 | && code == NE | |
4341 | && GET_CODE (x) == PLUS | |
4342 | && GET_CODE (y) == ZERO_EXTEND) | |
4343 | return CC_Cmode; | |
4344 | ||
43e9d192 IB |
4345 | /* For everything else, return CCmode. */ |
4346 | return CCmode; | |
4347 | } | |
4348 | ||
3dfa7055 ZC |
4349 | static int |
4350 | aarch64_get_condition_code_1 (enum machine_mode, enum rtx_code); | |
4351 | ||
cd5660ab | 4352 | int |
43e9d192 IB |
4353 | aarch64_get_condition_code (rtx x) |
4354 | { | |
ef4bddc2 | 4355 | machine_mode mode = GET_MODE (XEXP (x, 0)); |
43e9d192 IB |
4356 | enum rtx_code comp_code = GET_CODE (x); |
4357 | ||
4358 | if (GET_MODE_CLASS (mode) != MODE_CC) | |
4359 | mode = SELECT_CC_MODE (comp_code, XEXP (x, 0), XEXP (x, 1)); | |
3dfa7055 ZC |
4360 | return aarch64_get_condition_code_1 (mode, comp_code); |
4361 | } | |
43e9d192 | 4362 | |
3dfa7055 ZC |
4363 | static int |
4364 | aarch64_get_condition_code_1 (enum machine_mode mode, enum rtx_code comp_code) | |
4365 | { | |
43e9d192 IB |
4366 | switch (mode) |
4367 | { | |
4368 | case CCFPmode: | |
4369 | case CCFPEmode: | |
4370 | switch (comp_code) | |
4371 | { | |
4372 | case GE: return AARCH64_GE; | |
4373 | case GT: return AARCH64_GT; | |
4374 | case LE: return AARCH64_LS; | |
4375 | case LT: return AARCH64_MI; | |
4376 | case NE: return AARCH64_NE; | |
4377 | case EQ: return AARCH64_EQ; | |
4378 | case ORDERED: return AARCH64_VC; | |
4379 | case UNORDERED: return AARCH64_VS; | |
4380 | case UNLT: return AARCH64_LT; | |
4381 | case UNLE: return AARCH64_LE; | |
4382 | case UNGT: return AARCH64_HI; | |
4383 | case UNGE: return AARCH64_PL; | |
cd5660ab | 4384 | default: return -1; |
43e9d192 IB |
4385 | } |
4386 | break; | |
4387 | ||
4388 | case CCmode: | |
4389 | switch (comp_code) | |
4390 | { | |
4391 | case NE: return AARCH64_NE; | |
4392 | case EQ: return AARCH64_EQ; | |
4393 | case GE: return AARCH64_GE; | |
4394 | case GT: return AARCH64_GT; | |
4395 | case LE: return AARCH64_LE; | |
4396 | case LT: return AARCH64_LT; | |
4397 | case GEU: return AARCH64_CS; | |
4398 | case GTU: return AARCH64_HI; | |
4399 | case LEU: return AARCH64_LS; | |
4400 | case LTU: return AARCH64_CC; | |
cd5660ab | 4401 | default: return -1; |
43e9d192 IB |
4402 | } |
4403 | break; | |
4404 | ||
4405 | case CC_SWPmode: | |
43e9d192 IB |
4406 | switch (comp_code) |
4407 | { | |
4408 | case NE: return AARCH64_NE; | |
4409 | case EQ: return AARCH64_EQ; | |
4410 | case GE: return AARCH64_LE; | |
4411 | case GT: return AARCH64_LT; | |
4412 | case LE: return AARCH64_GE; | |
4413 | case LT: return AARCH64_GT; | |
4414 | case GEU: return AARCH64_LS; | |
4415 | case GTU: return AARCH64_CC; | |
4416 | case LEU: return AARCH64_CS; | |
4417 | case LTU: return AARCH64_HI; | |
cd5660ab | 4418 | default: return -1; |
43e9d192 IB |
4419 | } |
4420 | break; | |
4421 | ||
4422 | case CC_NZmode: | |
4423 | switch (comp_code) | |
4424 | { | |
4425 | case NE: return AARCH64_NE; | |
4426 | case EQ: return AARCH64_EQ; | |
4427 | case GE: return AARCH64_PL; | |
4428 | case LT: return AARCH64_MI; | |
cd5660ab | 4429 | default: return -1; |
43e9d192 IB |
4430 | } |
4431 | break; | |
4432 | ||
1c992d1e RE |
4433 | case CC_Zmode: |
4434 | switch (comp_code) | |
4435 | { | |
4436 | case NE: return AARCH64_NE; | |
4437 | case EQ: return AARCH64_EQ; | |
cd5660ab | 4438 | default: return -1; |
1c992d1e RE |
4439 | } |
4440 | break; | |
4441 | ||
ef22810a RH |
4442 | case CC_Cmode: |
4443 | switch (comp_code) | |
4444 | { | |
4445 | case NE: return AARCH64_CS; | |
4446 | case EQ: return AARCH64_CC; | |
4447 | default: return -1; | |
4448 | } | |
4449 | break; | |
4450 | ||
43e9d192 | 4451 | default: |
cd5660ab | 4452 | return -1; |
43e9d192 IB |
4453 | break; |
4454 | } | |
3dfa7055 | 4455 | |
3dfa7055 | 4456 | return -1; |
43e9d192 IB |
4457 | } |
4458 | ||
ddeabd3e AL |
4459 | bool |
4460 | aarch64_const_vec_all_same_in_range_p (rtx x, | |
4461 | HOST_WIDE_INT minval, | |
4462 | HOST_WIDE_INT maxval) | |
4463 | { | |
4464 | HOST_WIDE_INT firstval; | |
4465 | int count, i; | |
4466 | ||
4467 | if (GET_CODE (x) != CONST_VECTOR | |
4468 | || GET_MODE_CLASS (GET_MODE (x)) != MODE_VECTOR_INT) | |
4469 | return false; | |
4470 | ||
4471 | firstval = INTVAL (CONST_VECTOR_ELT (x, 0)); | |
4472 | if (firstval < minval || firstval > maxval) | |
4473 | return false; | |
4474 | ||
4475 | count = CONST_VECTOR_NUNITS (x); | |
4476 | for (i = 1; i < count; i++) | |
4477 | if (INTVAL (CONST_VECTOR_ELT (x, i)) != firstval) | |
4478 | return false; | |
4479 | ||
4480 | return true; | |
4481 | } | |
4482 | ||
4483 | bool | |
4484 | aarch64_const_vec_all_same_int_p (rtx x, HOST_WIDE_INT val) | |
4485 | { | |
4486 | return aarch64_const_vec_all_same_in_range_p (x, val, val); | |
4487 | } | |
4488 | ||
43e9d192 | 4489 | |
cf670503 ZC |
4490 | /* N Z C V. */ |
4491 | #define AARCH64_CC_V 1 | |
4492 | #define AARCH64_CC_C (1 << 1) | |
4493 | #define AARCH64_CC_Z (1 << 2) | |
4494 | #define AARCH64_CC_N (1 << 3) | |
4495 | ||
c8012fbc WD |
4496 | /* N Z C V flags for ccmp. Indexed by AARCH64_COND_CODE. */ |
4497 | static const int aarch64_nzcv_codes[] = | |
4498 | { | |
4499 | 0, /* EQ, Z == 1. */ | |
4500 | AARCH64_CC_Z, /* NE, Z == 0. */ | |
4501 | 0, /* CS, C == 1. */ | |
4502 | AARCH64_CC_C, /* CC, C == 0. */ | |
4503 | 0, /* MI, N == 1. */ | |
4504 | AARCH64_CC_N, /* PL, N == 0. */ | |
4505 | 0, /* VS, V == 1. */ | |
4506 | AARCH64_CC_V, /* VC, V == 0. */ | |
4507 | 0, /* HI, C ==1 && Z == 0. */ | |
4508 | AARCH64_CC_C, /* LS, !(C == 1 && Z == 0). */ | |
4509 | AARCH64_CC_V, /* GE, N == V. */ | |
4510 | 0, /* LT, N != V. */ | |
4511 | AARCH64_CC_Z, /* GT, Z == 0 && N == V. */ | |
4512 | 0, /* LE, !(Z == 0 && N == V). */ | |
4513 | 0, /* AL, Any. */ | |
4514 | 0 /* NV, Any. */ | |
cf670503 ZC |
4515 | }; |
4516 | ||
cc8ca59e JB |
4517 | static void |
4518 | aarch64_print_operand (FILE *f, rtx x, int code) | |
43e9d192 IB |
4519 | { |
4520 | switch (code) | |
4521 | { | |
f541a481 KT |
4522 | /* An integer or symbol address without a preceding # sign. */ |
4523 | case 'c': | |
4524 | switch (GET_CODE (x)) | |
4525 | { | |
4526 | case CONST_INT: | |
4527 | fprintf (f, HOST_WIDE_INT_PRINT_DEC, INTVAL (x)); | |
4528 | break; | |
4529 | ||
4530 | case SYMBOL_REF: | |
4531 | output_addr_const (f, x); | |
4532 | break; | |
4533 | ||
4534 | case CONST: | |
4535 | if (GET_CODE (XEXP (x, 0)) == PLUS | |
4536 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF) | |
4537 | { | |
4538 | output_addr_const (f, x); | |
4539 | break; | |
4540 | } | |
4541 | /* Fall through. */ | |
4542 | ||
4543 | default: | |
4544 | output_operand_lossage ("Unsupported operand for code '%c'", code); | |
4545 | } | |
4546 | break; | |
4547 | ||
43e9d192 IB |
4548 | case 'e': |
4549 | /* Print the sign/zero-extend size as a character 8->b, 16->h, 32->w. */ | |
4550 | { | |
4551 | int n; | |
4552 | ||
4aa81c2e | 4553 | if (!CONST_INT_P (x) |
43e9d192 IB |
4554 | || (n = exact_log2 (INTVAL (x) & ~7)) <= 0) |
4555 | { | |
4556 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4557 | return; | |
4558 | } | |
4559 | ||
4560 | switch (n) | |
4561 | { | |
4562 | case 3: | |
4563 | fputc ('b', f); | |
4564 | break; | |
4565 | case 4: | |
4566 | fputc ('h', f); | |
4567 | break; | |
4568 | case 5: | |
4569 | fputc ('w', f); | |
4570 | break; | |
4571 | default: | |
4572 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4573 | return; | |
4574 | } | |
4575 | } | |
4576 | break; | |
4577 | ||
4578 | case 'p': | |
4579 | { | |
4580 | int n; | |
4581 | ||
4582 | /* Print N such that 2^N == X. */ | |
4aa81c2e | 4583 | if (!CONST_INT_P (x) || (n = exact_log2 (INTVAL (x))) < 0) |
43e9d192 IB |
4584 | { |
4585 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4586 | return; | |
4587 | } | |
4588 | ||
4589 | asm_fprintf (f, "%d", n); | |
4590 | } | |
4591 | break; | |
4592 | ||
4593 | case 'P': | |
4594 | /* Print the number of non-zero bits in X (a const_int). */ | |
4aa81c2e | 4595 | if (!CONST_INT_P (x)) |
43e9d192 IB |
4596 | { |
4597 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4598 | return; | |
4599 | } | |
4600 | ||
8d55c61b | 4601 | asm_fprintf (f, "%u", popcount_hwi (INTVAL (x))); |
43e9d192 IB |
4602 | break; |
4603 | ||
4604 | case 'H': | |
4605 | /* Print the higher numbered register of a pair (TImode) of regs. */ | |
4aa81c2e | 4606 | if (!REG_P (x) || !GP_REGNUM_P (REGNO (x) + 1)) |
43e9d192 IB |
4607 | { |
4608 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4609 | return; | |
4610 | } | |
4611 | ||
01a3a324 | 4612 | asm_fprintf (f, "%s", reg_names [REGNO (x) + 1]); |
43e9d192 IB |
4613 | break; |
4614 | ||
43e9d192 | 4615 | case 'M': |
c8012fbc | 4616 | case 'm': |
cd5660ab KT |
4617 | { |
4618 | int cond_code; | |
c8012fbc | 4619 | /* Print a condition (eq, ne, etc) or its inverse. */ |
43e9d192 | 4620 | |
c8012fbc WD |
4621 | /* CONST_TRUE_RTX means al/nv (al is the default, don't print it). */ |
4622 | if (x == const_true_rtx) | |
cd5660ab | 4623 | { |
c8012fbc WD |
4624 | if (code == 'M') |
4625 | fputs ("nv", f); | |
cd5660ab KT |
4626 | return; |
4627 | } | |
43e9d192 | 4628 | |
cd5660ab KT |
4629 | if (!COMPARISON_P (x)) |
4630 | { | |
4631 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4632 | return; | |
4633 | } | |
c8012fbc | 4634 | |
cd5660ab KT |
4635 | cond_code = aarch64_get_condition_code (x); |
4636 | gcc_assert (cond_code >= 0); | |
c8012fbc WD |
4637 | if (code == 'M') |
4638 | cond_code = AARCH64_INVERSE_CONDITION_CODE (cond_code); | |
4639 | fputs (aarch64_condition_codes[cond_code], f); | |
cd5660ab | 4640 | } |
43e9d192 IB |
4641 | break; |
4642 | ||
4643 | case 'b': | |
4644 | case 'h': | |
4645 | case 's': | |
4646 | case 'd': | |
4647 | case 'q': | |
4648 | /* Print a scalar FP/SIMD register name. */ | |
4649 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) | |
4650 | { | |
4651 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
4652 | return; | |
4653 | } | |
50ce6f88 | 4654 | asm_fprintf (f, "%c%d", code, REGNO (x) - V0_REGNUM); |
43e9d192 IB |
4655 | break; |
4656 | ||
4657 | case 'S': | |
4658 | case 'T': | |
4659 | case 'U': | |
4660 | case 'V': | |
4661 | /* Print the first FP/SIMD register name in a list. */ | |
4662 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) | |
4663 | { | |
4664 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
4665 | return; | |
4666 | } | |
50ce6f88 | 4667 | asm_fprintf (f, "v%d", REGNO (x) - V0_REGNUM + (code - 'S')); |
43e9d192 IB |
4668 | break; |
4669 | ||
2d8c6dc1 AH |
4670 | case 'R': |
4671 | /* Print a scalar FP/SIMD register name + 1. */ | |
4672 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) | |
4673 | { | |
4674 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
4675 | return; | |
4676 | } | |
4677 | asm_fprintf (f, "q%d", REGNO (x) - V0_REGNUM + 1); | |
4678 | break; | |
4679 | ||
a05c0ddf | 4680 | case 'X': |
50d38551 | 4681 | /* Print bottom 16 bits of integer constant in hex. */ |
4aa81c2e | 4682 | if (!CONST_INT_P (x)) |
a05c0ddf IB |
4683 | { |
4684 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4685 | return; | |
4686 | } | |
50d38551 | 4687 | asm_fprintf (f, "0x%wx", UINTVAL (x) & 0xffff); |
a05c0ddf IB |
4688 | break; |
4689 | ||
43e9d192 IB |
4690 | case 'w': |
4691 | case 'x': | |
4692 | /* Print a general register name or the zero register (32-bit or | |
4693 | 64-bit). */ | |
3520f7cc JG |
4694 | if (x == const0_rtx |
4695 | || (CONST_DOUBLE_P (x) && aarch64_float_const_zero_rtx_p (x))) | |
43e9d192 | 4696 | { |
50ce6f88 | 4697 | asm_fprintf (f, "%czr", code); |
43e9d192 IB |
4698 | break; |
4699 | } | |
4700 | ||
4701 | if (REG_P (x) && GP_REGNUM_P (REGNO (x))) | |
4702 | { | |
50ce6f88 | 4703 | asm_fprintf (f, "%c%d", code, REGNO (x) - R0_REGNUM); |
43e9d192 IB |
4704 | break; |
4705 | } | |
4706 | ||
4707 | if (REG_P (x) && REGNO (x) == SP_REGNUM) | |
4708 | { | |
50ce6f88 | 4709 | asm_fprintf (f, "%ssp", code == 'w' ? "w" : ""); |
43e9d192 IB |
4710 | break; |
4711 | } | |
4712 | ||
4713 | /* Fall through */ | |
4714 | ||
4715 | case 0: | |
4716 | /* Print a normal operand, if it's a general register, then we | |
4717 | assume DImode. */ | |
4718 | if (x == NULL) | |
4719 | { | |
4720 | output_operand_lossage ("missing operand"); | |
4721 | return; | |
4722 | } | |
4723 | ||
4724 | switch (GET_CODE (x)) | |
4725 | { | |
4726 | case REG: | |
01a3a324 | 4727 | asm_fprintf (f, "%s", reg_names [REGNO (x)]); |
43e9d192 IB |
4728 | break; |
4729 | ||
4730 | case MEM: | |
cc8ca59e | 4731 | output_address (GET_MODE (x), XEXP (x, 0)); |
43e9d192 IB |
4732 | break; |
4733 | ||
2af16a7c | 4734 | case CONST: |
43e9d192 IB |
4735 | case LABEL_REF: |
4736 | case SYMBOL_REF: | |
4737 | output_addr_const (asm_out_file, x); | |
4738 | break; | |
4739 | ||
4740 | case CONST_INT: | |
4741 | asm_fprintf (f, "%wd", INTVAL (x)); | |
4742 | break; | |
4743 | ||
4744 | case CONST_VECTOR: | |
3520f7cc JG |
4745 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_VECTOR_INT) |
4746 | { | |
ddeabd3e AL |
4747 | gcc_assert ( |
4748 | aarch64_const_vec_all_same_in_range_p (x, | |
4749 | HOST_WIDE_INT_MIN, | |
4750 | HOST_WIDE_INT_MAX)); | |
3520f7cc JG |
4751 | asm_fprintf (f, "%wd", INTVAL (CONST_VECTOR_ELT (x, 0))); |
4752 | } | |
4753 | else if (aarch64_simd_imm_zero_p (x, GET_MODE (x))) | |
4754 | { | |
4755 | fputc ('0', f); | |
4756 | } | |
4757 | else | |
4758 | gcc_unreachable (); | |
43e9d192 IB |
4759 | break; |
4760 | ||
3520f7cc | 4761 | case CONST_DOUBLE: |
2ca5b430 KT |
4762 | /* Since we define TARGET_SUPPORTS_WIDE_INT we shouldn't ever |
4763 | be getting CONST_DOUBLEs holding integers. */ | |
4764 | gcc_assert (GET_MODE (x) != VOIDmode); | |
4765 | if (aarch64_float_const_zero_rtx_p (x)) | |
3520f7cc JG |
4766 | { |
4767 | fputc ('0', f); | |
4768 | break; | |
4769 | } | |
4770 | else if (aarch64_float_const_representable_p (x)) | |
4771 | { | |
4772 | #define buf_size 20 | |
4773 | char float_buf[buf_size] = {'\0'}; | |
34a72c33 RS |
4774 | real_to_decimal_for_mode (float_buf, |
4775 | CONST_DOUBLE_REAL_VALUE (x), | |
3520f7cc JG |
4776 | buf_size, buf_size, |
4777 | 1, GET_MODE (x)); | |
4778 | asm_fprintf (asm_out_file, "%s", float_buf); | |
4779 | break; | |
4780 | #undef buf_size | |
4781 | } | |
4782 | output_operand_lossage ("invalid constant"); | |
4783 | return; | |
43e9d192 IB |
4784 | default: |
4785 | output_operand_lossage ("invalid operand"); | |
4786 | return; | |
4787 | } | |
4788 | break; | |
4789 | ||
4790 | case 'A': | |
4791 | if (GET_CODE (x) == HIGH) | |
4792 | x = XEXP (x, 0); | |
4793 | ||
a6e0bfa7 | 4794 | switch (aarch64_classify_symbolic_expression (x)) |
43e9d192 | 4795 | { |
6642bdb4 | 4796 | case SYMBOL_SMALL_GOT_4G: |
43e9d192 IB |
4797 | asm_fprintf (asm_out_file, ":got:"); |
4798 | break; | |
4799 | ||
4800 | case SYMBOL_SMALL_TLSGD: | |
4801 | asm_fprintf (asm_out_file, ":tlsgd:"); | |
4802 | break; | |
4803 | ||
4804 | case SYMBOL_SMALL_TLSDESC: | |
4805 | asm_fprintf (asm_out_file, ":tlsdesc:"); | |
4806 | break; | |
4807 | ||
79496620 | 4808 | case SYMBOL_SMALL_TLSIE: |
43e9d192 IB |
4809 | asm_fprintf (asm_out_file, ":gottprel:"); |
4810 | break; | |
4811 | ||
d18ba284 | 4812 | case SYMBOL_TLSLE24: |
43e9d192 IB |
4813 | asm_fprintf (asm_out_file, ":tprel:"); |
4814 | break; | |
4815 | ||
87dd8ab0 MS |
4816 | case SYMBOL_TINY_GOT: |
4817 | gcc_unreachable (); | |
4818 | break; | |
4819 | ||
43e9d192 IB |
4820 | default: |
4821 | break; | |
4822 | } | |
4823 | output_addr_const (asm_out_file, x); | |
4824 | break; | |
4825 | ||
4826 | case 'L': | |
a6e0bfa7 | 4827 | switch (aarch64_classify_symbolic_expression (x)) |
43e9d192 | 4828 | { |
6642bdb4 | 4829 | case SYMBOL_SMALL_GOT_4G: |
43e9d192 IB |
4830 | asm_fprintf (asm_out_file, ":lo12:"); |
4831 | break; | |
4832 | ||
4833 | case SYMBOL_SMALL_TLSGD: | |
4834 | asm_fprintf (asm_out_file, ":tlsgd_lo12:"); | |
4835 | break; | |
4836 | ||
4837 | case SYMBOL_SMALL_TLSDESC: | |
4838 | asm_fprintf (asm_out_file, ":tlsdesc_lo12:"); | |
4839 | break; | |
4840 | ||
79496620 | 4841 | case SYMBOL_SMALL_TLSIE: |
43e9d192 IB |
4842 | asm_fprintf (asm_out_file, ":gottprel_lo12:"); |
4843 | break; | |
4844 | ||
cbf5629e JW |
4845 | case SYMBOL_TLSLE12: |
4846 | asm_fprintf (asm_out_file, ":tprel_lo12:"); | |
4847 | break; | |
4848 | ||
d18ba284 | 4849 | case SYMBOL_TLSLE24: |
43e9d192 IB |
4850 | asm_fprintf (asm_out_file, ":tprel_lo12_nc:"); |
4851 | break; | |
4852 | ||
87dd8ab0 MS |
4853 | case SYMBOL_TINY_GOT: |
4854 | asm_fprintf (asm_out_file, ":got:"); | |
4855 | break; | |
4856 | ||
5ae7caad JW |
4857 | case SYMBOL_TINY_TLSIE: |
4858 | asm_fprintf (asm_out_file, ":gottprel:"); | |
4859 | break; | |
4860 | ||
43e9d192 IB |
4861 | default: |
4862 | break; | |
4863 | } | |
4864 | output_addr_const (asm_out_file, x); | |
4865 | break; | |
4866 | ||
4867 | case 'G': | |
4868 | ||
a6e0bfa7 | 4869 | switch (aarch64_classify_symbolic_expression (x)) |
43e9d192 | 4870 | { |
d18ba284 | 4871 | case SYMBOL_TLSLE24: |
43e9d192 IB |
4872 | asm_fprintf (asm_out_file, ":tprel_hi12:"); |
4873 | break; | |
4874 | default: | |
4875 | break; | |
4876 | } | |
4877 | output_addr_const (asm_out_file, x); | |
4878 | break; | |
4879 | ||
cf670503 ZC |
4880 | case 'k': |
4881 | { | |
c8012fbc | 4882 | HOST_WIDE_INT cond_code; |
cf670503 ZC |
4883 | /* Print nzcv. */ |
4884 | ||
c8012fbc | 4885 | if (!CONST_INT_P (x)) |
cf670503 ZC |
4886 | { |
4887 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4888 | return; | |
4889 | } | |
4890 | ||
c8012fbc WD |
4891 | cond_code = INTVAL (x); |
4892 | gcc_assert (cond_code >= 0 && cond_code <= AARCH64_NV); | |
4893 | asm_fprintf (f, "%d", aarch64_nzcv_codes[cond_code]); | |
cf670503 ZC |
4894 | } |
4895 | break; | |
4896 | ||
43e9d192 IB |
4897 | default: |
4898 | output_operand_lossage ("invalid operand prefix '%%%c'", code); | |
4899 | return; | |
4900 | } | |
4901 | } | |
4902 | ||
cc8ca59e JB |
4903 | static void |
4904 | aarch64_print_operand_address (FILE *f, machine_mode mode, rtx x) | |
43e9d192 IB |
4905 | { |
4906 | struct aarch64_address_info addr; | |
4907 | ||
cc8ca59e | 4908 | if (aarch64_classify_address (&addr, x, mode, MEM, true)) |
43e9d192 IB |
4909 | switch (addr.type) |
4910 | { | |
4911 | case ADDRESS_REG_IMM: | |
4912 | if (addr.offset == const0_rtx) | |
01a3a324 | 4913 | asm_fprintf (f, "[%s]", reg_names [REGNO (addr.base)]); |
43e9d192 | 4914 | else |
16a3246f | 4915 | asm_fprintf (f, "[%s, %wd]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4916 | INTVAL (addr.offset)); |
4917 | return; | |
4918 | ||
4919 | case ADDRESS_REG_REG: | |
4920 | if (addr.shift == 0) | |
16a3246f | 4921 | asm_fprintf (f, "[%s, %s]", reg_names [REGNO (addr.base)], |
01a3a324 | 4922 | reg_names [REGNO (addr.offset)]); |
43e9d192 | 4923 | else |
16a3246f | 4924 | asm_fprintf (f, "[%s, %s, lsl %u]", reg_names [REGNO (addr.base)], |
01a3a324 | 4925 | reg_names [REGNO (addr.offset)], addr.shift); |
43e9d192 IB |
4926 | return; |
4927 | ||
4928 | case ADDRESS_REG_UXTW: | |
4929 | if (addr.shift == 0) | |
16a3246f | 4930 | asm_fprintf (f, "[%s, w%d, uxtw]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4931 | REGNO (addr.offset) - R0_REGNUM); |
4932 | else | |
16a3246f | 4933 | asm_fprintf (f, "[%s, w%d, uxtw %u]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4934 | REGNO (addr.offset) - R0_REGNUM, addr.shift); |
4935 | return; | |
4936 | ||
4937 | case ADDRESS_REG_SXTW: | |
4938 | if (addr.shift == 0) | |
16a3246f | 4939 | asm_fprintf (f, "[%s, w%d, sxtw]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4940 | REGNO (addr.offset) - R0_REGNUM); |
4941 | else | |
16a3246f | 4942 | asm_fprintf (f, "[%s, w%d, sxtw %u]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4943 | REGNO (addr.offset) - R0_REGNUM, addr.shift); |
4944 | return; | |
4945 | ||
4946 | case ADDRESS_REG_WB: | |
4947 | switch (GET_CODE (x)) | |
4948 | { | |
4949 | case PRE_INC: | |
16a3246f | 4950 | asm_fprintf (f, "[%s, %d]!", reg_names [REGNO (addr.base)], |
cc8ca59e | 4951 | GET_MODE_SIZE (mode)); |
43e9d192 IB |
4952 | return; |
4953 | case POST_INC: | |
16a3246f | 4954 | asm_fprintf (f, "[%s], %d", reg_names [REGNO (addr.base)], |
cc8ca59e | 4955 | GET_MODE_SIZE (mode)); |
43e9d192 IB |
4956 | return; |
4957 | case PRE_DEC: | |
16a3246f | 4958 | asm_fprintf (f, "[%s, -%d]!", reg_names [REGNO (addr.base)], |
cc8ca59e | 4959 | GET_MODE_SIZE (mode)); |
43e9d192 IB |
4960 | return; |
4961 | case POST_DEC: | |
16a3246f | 4962 | asm_fprintf (f, "[%s], -%d", reg_names [REGNO (addr.base)], |
cc8ca59e | 4963 | GET_MODE_SIZE (mode)); |
43e9d192 IB |
4964 | return; |
4965 | case PRE_MODIFY: | |
16a3246f | 4966 | asm_fprintf (f, "[%s, %wd]!", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4967 | INTVAL (addr.offset)); |
4968 | return; | |
4969 | case POST_MODIFY: | |
16a3246f | 4970 | asm_fprintf (f, "[%s], %wd", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4971 | INTVAL (addr.offset)); |
4972 | return; | |
4973 | default: | |
4974 | break; | |
4975 | } | |
4976 | break; | |
4977 | ||
4978 | case ADDRESS_LO_SUM: | |
16a3246f | 4979 | asm_fprintf (f, "[%s, #:lo12:", reg_names [REGNO (addr.base)]); |
43e9d192 IB |
4980 | output_addr_const (f, addr.offset); |
4981 | asm_fprintf (f, "]"); | |
4982 | return; | |
4983 | ||
4984 | case ADDRESS_SYMBOLIC: | |
4985 | break; | |
4986 | } | |
4987 | ||
4988 | output_addr_const (f, x); | |
4989 | } | |
4990 | ||
43e9d192 IB |
4991 | bool |
4992 | aarch64_label_mentioned_p (rtx x) | |
4993 | { | |
4994 | const char *fmt; | |
4995 | int i; | |
4996 | ||
4997 | if (GET_CODE (x) == LABEL_REF) | |
4998 | return true; | |
4999 | ||
5000 | /* UNSPEC_TLS entries for a symbol include a LABEL_REF for the | |
5001 | referencing instruction, but they are constant offsets, not | |
5002 | symbols. */ | |
5003 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
5004 | return false; | |
5005 | ||
5006 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
5007 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
5008 | { | |
5009 | if (fmt[i] == 'E') | |
5010 | { | |
5011 | int j; | |
5012 | ||
5013 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
5014 | if (aarch64_label_mentioned_p (XVECEXP (x, i, j))) | |
5015 | return 1; | |
5016 | } | |
5017 | else if (fmt[i] == 'e' && aarch64_label_mentioned_p (XEXP (x, i))) | |
5018 | return 1; | |
5019 | } | |
5020 | ||
5021 | return 0; | |
5022 | } | |
5023 | ||
5024 | /* Implement REGNO_REG_CLASS. */ | |
5025 | ||
5026 | enum reg_class | |
5027 | aarch64_regno_regclass (unsigned regno) | |
5028 | { | |
5029 | if (GP_REGNUM_P (regno)) | |
a4a182c6 | 5030 | return GENERAL_REGS; |
43e9d192 IB |
5031 | |
5032 | if (regno == SP_REGNUM) | |
5033 | return STACK_REG; | |
5034 | ||
5035 | if (regno == FRAME_POINTER_REGNUM | |
5036 | || regno == ARG_POINTER_REGNUM) | |
f24bb080 | 5037 | return POINTER_REGS; |
43e9d192 IB |
5038 | |
5039 | if (FP_REGNUM_P (regno)) | |
5040 | return FP_LO_REGNUM_P (regno) ? FP_LO_REGS : FP_REGS; | |
5041 | ||
5042 | return NO_REGS; | |
5043 | } | |
5044 | ||
0c4ec427 | 5045 | static rtx |
ef4bddc2 | 5046 | aarch64_legitimize_address (rtx x, rtx /* orig_x */, machine_mode mode) |
0c4ec427 RE |
5047 | { |
5048 | /* Try to split X+CONST into Y=X+(CONST & ~mask), Y+(CONST&mask), | |
5049 | where mask is selected by alignment and size of the offset. | |
5050 | We try to pick as large a range for the offset as possible to | |
5051 | maximize the chance of a CSE. However, for aligned addresses | |
5052 | we limit the range to 4k so that structures with different sized | |
e8426e0a BC |
5053 | elements are likely to use the same base. We need to be careful |
5054 | not to split a CONST for some forms of address expression, otherwise | |
5055 | it will generate sub-optimal code. */ | |
0c4ec427 RE |
5056 | |
5057 | if (GET_CODE (x) == PLUS && CONST_INT_P (XEXP (x, 1))) | |
5058 | { | |
9e0218fc | 5059 | rtx base = XEXP (x, 0); |
17d7bdd8 | 5060 | rtx offset_rtx = XEXP (x, 1); |
9e0218fc | 5061 | HOST_WIDE_INT offset = INTVAL (offset_rtx); |
0c4ec427 | 5062 | |
9e0218fc | 5063 | if (GET_CODE (base) == PLUS) |
e8426e0a | 5064 | { |
9e0218fc RH |
5065 | rtx op0 = XEXP (base, 0); |
5066 | rtx op1 = XEXP (base, 1); | |
5067 | ||
5068 | /* Force any scaling into a temp for CSE. */ | |
5069 | op0 = force_reg (Pmode, op0); | |
5070 | op1 = force_reg (Pmode, op1); | |
5071 | ||
5072 | /* Let the pointer register be in op0. */ | |
5073 | if (REG_POINTER (op1)) | |
5074 | std::swap (op0, op1); | |
5075 | ||
5076 | /* If the pointer is virtual or frame related, then we know that | |
5077 | virtual register instantiation or register elimination is going | |
5078 | to apply a second constant. We want the two constants folded | |
5079 | together easily. Therefore, emit as (OP0 + CONST) + OP1. */ | |
5080 | if (virt_or_elim_regno_p (REGNO (op0))) | |
e8426e0a | 5081 | { |
9e0218fc RH |
5082 | base = expand_binop (Pmode, add_optab, op0, offset_rtx, |
5083 | NULL_RTX, true, OPTAB_DIRECT); | |
5084 | return gen_rtx_PLUS (Pmode, base, op1); | |
e8426e0a | 5085 | } |
e8426e0a | 5086 | |
9e0218fc RH |
5087 | /* Otherwise, in order to encourage CSE (and thence loop strength |
5088 | reduce) scaled addresses, emit as (OP0 + OP1) + CONST. */ | |
5089 | base = expand_binop (Pmode, add_optab, op0, op1, | |
5090 | NULL_RTX, true, OPTAB_DIRECT); | |
5091 | x = gen_rtx_PLUS (Pmode, base, offset_rtx); | |
e8426e0a BC |
5092 | } |
5093 | ||
0c4ec427 | 5094 | /* Does it look like we'll need a load/store-pair operation? */ |
9e0218fc | 5095 | HOST_WIDE_INT base_offset; |
0c4ec427 RE |
5096 | if (GET_MODE_SIZE (mode) > 16 |
5097 | || mode == TImode) | |
5098 | base_offset = ((offset + 64 * GET_MODE_SIZE (mode)) | |
5099 | & ~((128 * GET_MODE_SIZE (mode)) - 1)); | |
5100 | /* For offsets aren't a multiple of the access size, the limit is | |
5101 | -256...255. */ | |
5102 | else if (offset & (GET_MODE_SIZE (mode) - 1)) | |
5103 | base_offset = (offset + 0x100) & ~0x1ff; | |
5104 | else | |
5105 | base_offset = offset & ~0xfff; | |
5106 | ||
9e0218fc RH |
5107 | if (base_offset != 0) |
5108 | { | |
5109 | base = plus_constant (Pmode, base, base_offset); | |
5110 | base = force_operand (base, NULL_RTX); | |
5111 | return plus_constant (Pmode, base, offset - base_offset); | |
5112 | } | |
0c4ec427 RE |
5113 | } |
5114 | ||
5115 | return x; | |
5116 | } | |
5117 | ||
b4f50fd4 RR |
5118 | /* Return the reload icode required for a constant pool in mode. */ |
5119 | static enum insn_code | |
5120 | aarch64_constant_pool_reload_icode (machine_mode mode) | |
5121 | { | |
5122 | switch (mode) | |
5123 | { | |
5124 | case SFmode: | |
5125 | return CODE_FOR_aarch64_reload_movcpsfdi; | |
5126 | ||
5127 | case DFmode: | |
5128 | return CODE_FOR_aarch64_reload_movcpdfdi; | |
5129 | ||
5130 | case TFmode: | |
5131 | return CODE_FOR_aarch64_reload_movcptfdi; | |
5132 | ||
5133 | case V8QImode: | |
5134 | return CODE_FOR_aarch64_reload_movcpv8qidi; | |
5135 | ||
5136 | case V16QImode: | |
5137 | return CODE_FOR_aarch64_reload_movcpv16qidi; | |
5138 | ||
5139 | case V4HImode: | |
5140 | return CODE_FOR_aarch64_reload_movcpv4hidi; | |
5141 | ||
5142 | case V8HImode: | |
5143 | return CODE_FOR_aarch64_reload_movcpv8hidi; | |
5144 | ||
5145 | case V2SImode: | |
5146 | return CODE_FOR_aarch64_reload_movcpv2sidi; | |
5147 | ||
5148 | case V4SImode: | |
5149 | return CODE_FOR_aarch64_reload_movcpv4sidi; | |
5150 | ||
5151 | case V2DImode: | |
5152 | return CODE_FOR_aarch64_reload_movcpv2didi; | |
5153 | ||
5154 | case V2DFmode: | |
5155 | return CODE_FOR_aarch64_reload_movcpv2dfdi; | |
5156 | ||
5157 | default: | |
5158 | gcc_unreachable (); | |
5159 | } | |
5160 | ||
5161 | gcc_unreachable (); | |
5162 | } | |
43e9d192 IB |
5163 | static reg_class_t |
5164 | aarch64_secondary_reload (bool in_p ATTRIBUTE_UNUSED, rtx x, | |
5165 | reg_class_t rclass, | |
ef4bddc2 | 5166 | machine_mode mode, |
43e9d192 IB |
5167 | secondary_reload_info *sri) |
5168 | { | |
b4f50fd4 RR |
5169 | |
5170 | /* If we have to disable direct literal pool loads and stores because the | |
5171 | function is too big, then we need a scratch register. */ | |
5172 | if (MEM_P (x) && GET_CODE (x) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (x) | |
5173 | && (SCALAR_FLOAT_MODE_P (GET_MODE (x)) | |
5174 | || targetm.vector_mode_supported_p (GET_MODE (x))) | |
9ee6540a | 5175 | && !aarch64_pcrelative_literal_loads) |
b4f50fd4 RR |
5176 | { |
5177 | sri->icode = aarch64_constant_pool_reload_icode (mode); | |
5178 | return NO_REGS; | |
5179 | } | |
5180 | ||
43e9d192 IB |
5181 | /* Without the TARGET_SIMD instructions we cannot move a Q register |
5182 | to a Q register directly. We need a scratch. */ | |
5183 | if (REG_P (x) && (mode == TFmode || mode == TImode) && mode == GET_MODE (x) | |
5184 | && FP_REGNUM_P (REGNO (x)) && !TARGET_SIMD | |
5185 | && reg_class_subset_p (rclass, FP_REGS)) | |
5186 | { | |
5187 | if (mode == TFmode) | |
5188 | sri->icode = CODE_FOR_aarch64_reload_movtf; | |
5189 | else if (mode == TImode) | |
5190 | sri->icode = CODE_FOR_aarch64_reload_movti; | |
5191 | return NO_REGS; | |
5192 | } | |
5193 | ||
5194 | /* A TFmode or TImode memory access should be handled via an FP_REGS | |
5195 | because AArch64 has richer addressing modes for LDR/STR instructions | |
5196 | than LDP/STP instructions. */ | |
d5726973 | 5197 | if (TARGET_FLOAT && rclass == GENERAL_REGS |
43e9d192 IB |
5198 | && GET_MODE_SIZE (mode) == 16 && MEM_P (x)) |
5199 | return FP_REGS; | |
5200 | ||
5201 | if (rclass == FP_REGS && (mode == TImode || mode == TFmode) && CONSTANT_P(x)) | |
a4a182c6 | 5202 | return GENERAL_REGS; |
43e9d192 IB |
5203 | |
5204 | return NO_REGS; | |
5205 | } | |
5206 | ||
5207 | static bool | |
5208 | aarch64_can_eliminate (const int from, const int to) | |
5209 | { | |
5210 | /* If we need a frame pointer, we must eliminate FRAME_POINTER_REGNUM into | |
5211 | HARD_FRAME_POINTER_REGNUM and not into STACK_POINTER_REGNUM. */ | |
5212 | ||
5213 | if (frame_pointer_needed) | |
5214 | { | |
5215 | if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
5216 | return true; | |
5217 | if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) | |
5218 | return false; | |
5219 | if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM | |
5220 | && !cfun->calls_alloca) | |
5221 | return true; | |
5222 | if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
5223 | return true; | |
0b7f8166 MS |
5224 | |
5225 | return false; | |
43e9d192 | 5226 | } |
1c923b60 JW |
5227 | else |
5228 | { | |
5229 | /* If we decided that we didn't need a leaf frame pointer but then used | |
5230 | LR in the function, then we'll want a frame pointer after all, so | |
5231 | prevent this elimination to ensure a frame pointer is used. */ | |
5232 | if (to == STACK_POINTER_REGNUM | |
5233 | && flag_omit_leaf_frame_pointer | |
5234 | && df_regs_ever_live_p (LR_REGNUM)) | |
5235 | return false; | |
5236 | } | |
777e6976 | 5237 | |
43e9d192 IB |
5238 | return true; |
5239 | } | |
5240 | ||
5241 | HOST_WIDE_INT | |
5242 | aarch64_initial_elimination_offset (unsigned from, unsigned to) | |
5243 | { | |
43e9d192 | 5244 | aarch64_layout_frame (); |
78c29983 MS |
5245 | |
5246 | if (to == HARD_FRAME_POINTER_REGNUM) | |
5247 | { | |
5248 | if (from == ARG_POINTER_REGNUM) | |
1c960e02 | 5249 | return cfun->machine->frame.frame_size - crtl->outgoing_args_size; |
78c29983 MS |
5250 | |
5251 | if (from == FRAME_POINTER_REGNUM) | |
1c960e02 MS |
5252 | return (cfun->machine->frame.hard_fp_offset |
5253 | - cfun->machine->frame.saved_varargs_size); | |
78c29983 MS |
5254 | } |
5255 | ||
5256 | if (to == STACK_POINTER_REGNUM) | |
5257 | { | |
5258 | if (from == FRAME_POINTER_REGNUM) | |
1c960e02 MS |
5259 | return (cfun->machine->frame.frame_size |
5260 | - cfun->machine->frame.saved_varargs_size); | |
78c29983 MS |
5261 | } |
5262 | ||
1c960e02 | 5263 | return cfun->machine->frame.frame_size; |
43e9d192 IB |
5264 | } |
5265 | ||
43e9d192 IB |
5266 | /* Implement RETURN_ADDR_RTX. We do not support moving back to a |
5267 | previous frame. */ | |
5268 | ||
5269 | rtx | |
5270 | aarch64_return_addr (int count, rtx frame ATTRIBUTE_UNUSED) | |
5271 | { | |
5272 | if (count != 0) | |
5273 | return const0_rtx; | |
5274 | return get_hard_reg_initial_val (Pmode, LR_REGNUM); | |
5275 | } | |
5276 | ||
5277 | ||
5278 | static void | |
5279 | aarch64_asm_trampoline_template (FILE *f) | |
5280 | { | |
28514dda YZ |
5281 | if (TARGET_ILP32) |
5282 | { | |
5283 | asm_fprintf (f, "\tldr\tw%d, .+16\n", IP1_REGNUM - R0_REGNUM); | |
5284 | asm_fprintf (f, "\tldr\tw%d, .+16\n", STATIC_CHAIN_REGNUM - R0_REGNUM); | |
5285 | } | |
5286 | else | |
5287 | { | |
5288 | asm_fprintf (f, "\tldr\t%s, .+16\n", reg_names [IP1_REGNUM]); | |
5289 | asm_fprintf (f, "\tldr\t%s, .+20\n", reg_names [STATIC_CHAIN_REGNUM]); | |
5290 | } | |
01a3a324 | 5291 | asm_fprintf (f, "\tbr\t%s\n", reg_names [IP1_REGNUM]); |
43e9d192 | 5292 | assemble_aligned_integer (4, const0_rtx); |
28514dda YZ |
5293 | assemble_aligned_integer (POINTER_BYTES, const0_rtx); |
5294 | assemble_aligned_integer (POINTER_BYTES, const0_rtx); | |
43e9d192 IB |
5295 | } |
5296 | ||
5297 | static void | |
5298 | aarch64_trampoline_init (rtx m_tramp, tree fndecl, rtx chain_value) | |
5299 | { | |
5300 | rtx fnaddr, mem, a_tramp; | |
28514dda | 5301 | const int tramp_code_sz = 16; |
43e9d192 IB |
5302 | |
5303 | /* Don't need to copy the trailing D-words, we fill those in below. */ | |
5304 | emit_block_move (m_tramp, assemble_trampoline_template (), | |
28514dda YZ |
5305 | GEN_INT (tramp_code_sz), BLOCK_OP_NORMAL); |
5306 | mem = adjust_address (m_tramp, ptr_mode, tramp_code_sz); | |
43e9d192 | 5307 | fnaddr = XEXP (DECL_RTL (fndecl), 0); |
28514dda YZ |
5308 | if (GET_MODE (fnaddr) != ptr_mode) |
5309 | fnaddr = convert_memory_address (ptr_mode, fnaddr); | |
43e9d192 IB |
5310 | emit_move_insn (mem, fnaddr); |
5311 | ||
28514dda | 5312 | mem = adjust_address (m_tramp, ptr_mode, tramp_code_sz + POINTER_BYTES); |
43e9d192 IB |
5313 | emit_move_insn (mem, chain_value); |
5314 | ||
5315 | /* XXX We should really define a "clear_cache" pattern and use | |
5316 | gen_clear_cache(). */ | |
5317 | a_tramp = XEXP (m_tramp, 0); | |
5318 | emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__clear_cache"), | |
28514dda YZ |
5319 | LCT_NORMAL, VOIDmode, 2, a_tramp, ptr_mode, |
5320 | plus_constant (ptr_mode, a_tramp, TRAMPOLINE_SIZE), | |
5321 | ptr_mode); | |
43e9d192 IB |
5322 | } |
5323 | ||
5324 | static unsigned char | |
ef4bddc2 | 5325 | aarch64_class_max_nregs (reg_class_t regclass, machine_mode mode) |
43e9d192 IB |
5326 | { |
5327 | switch (regclass) | |
5328 | { | |
fee9ba42 | 5329 | case CALLER_SAVE_REGS: |
43e9d192 IB |
5330 | case POINTER_REGS: |
5331 | case GENERAL_REGS: | |
5332 | case ALL_REGS: | |
5333 | case FP_REGS: | |
5334 | case FP_LO_REGS: | |
5335 | return | |
7bd11911 KT |
5336 | aarch64_vector_mode_p (mode) |
5337 | ? (GET_MODE_SIZE (mode) + UNITS_PER_VREG - 1) / UNITS_PER_VREG | |
5338 | : (GET_MODE_SIZE (mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
43e9d192 IB |
5339 | case STACK_REG: |
5340 | return 1; | |
5341 | ||
5342 | case NO_REGS: | |
5343 | return 0; | |
5344 | ||
5345 | default: | |
5346 | break; | |
5347 | } | |
5348 | gcc_unreachable (); | |
5349 | } | |
5350 | ||
5351 | static reg_class_t | |
78d8b9f0 | 5352 | aarch64_preferred_reload_class (rtx x, reg_class_t regclass) |
43e9d192 | 5353 | { |
51bb310d | 5354 | if (regclass == POINTER_REGS) |
78d8b9f0 IB |
5355 | return GENERAL_REGS; |
5356 | ||
51bb310d MS |
5357 | if (regclass == STACK_REG) |
5358 | { | |
5359 | if (REG_P(x) | |
5360 | && reg_class_subset_p (REGNO_REG_CLASS (REGNO (x)), POINTER_REGS)) | |
5361 | return regclass; | |
5362 | ||
5363 | return NO_REGS; | |
5364 | } | |
5365 | ||
78d8b9f0 IB |
5366 | /* If it's an integer immediate that MOVI can't handle, then |
5367 | FP_REGS is not an option, so we return NO_REGS instead. */ | |
5368 | if (CONST_INT_P (x) && reg_class_subset_p (regclass, FP_REGS) | |
5369 | && !aarch64_simd_imm_scalar_p (x, GET_MODE (x))) | |
5370 | return NO_REGS; | |
5371 | ||
27bd251b IB |
5372 | /* Register eliminiation can result in a request for |
5373 | SP+constant->FP_REGS. We cannot support such operations which | |
5374 | use SP as source and an FP_REG as destination, so reject out | |
5375 | right now. */ | |
5376 | if (! reg_class_subset_p (regclass, GENERAL_REGS) && GET_CODE (x) == PLUS) | |
5377 | { | |
5378 | rtx lhs = XEXP (x, 0); | |
5379 | ||
5380 | /* Look through a possible SUBREG introduced by ILP32. */ | |
5381 | if (GET_CODE (lhs) == SUBREG) | |
5382 | lhs = SUBREG_REG (lhs); | |
5383 | ||
5384 | gcc_assert (REG_P (lhs)); | |
5385 | gcc_assert (reg_class_subset_p (REGNO_REG_CLASS (REGNO (lhs)), | |
5386 | POINTER_REGS)); | |
5387 | return NO_REGS; | |
5388 | } | |
5389 | ||
78d8b9f0 | 5390 | return regclass; |
43e9d192 IB |
5391 | } |
5392 | ||
5393 | void | |
5394 | aarch64_asm_output_labelref (FILE* f, const char *name) | |
5395 | { | |
5396 | asm_fprintf (f, "%U%s", name); | |
5397 | } | |
5398 | ||
5399 | static void | |
5400 | aarch64_elf_asm_constructor (rtx symbol, int priority) | |
5401 | { | |
5402 | if (priority == DEFAULT_INIT_PRIORITY) | |
5403 | default_ctor_section_asm_out_constructor (symbol, priority); | |
5404 | else | |
5405 | { | |
5406 | section *s; | |
5407 | char buf[18]; | |
5408 | snprintf (buf, sizeof (buf), ".init_array.%.5u", priority); | |
5409 | s = get_section (buf, SECTION_WRITE, NULL); | |
5410 | switch_to_section (s); | |
5411 | assemble_align (POINTER_SIZE); | |
28514dda | 5412 | assemble_aligned_integer (POINTER_BYTES, symbol); |
43e9d192 IB |
5413 | } |
5414 | } | |
5415 | ||
5416 | static void | |
5417 | aarch64_elf_asm_destructor (rtx symbol, int priority) | |
5418 | { | |
5419 | if (priority == DEFAULT_INIT_PRIORITY) | |
5420 | default_dtor_section_asm_out_destructor (symbol, priority); | |
5421 | else | |
5422 | { | |
5423 | section *s; | |
5424 | char buf[18]; | |
5425 | snprintf (buf, sizeof (buf), ".fini_array.%.5u", priority); | |
5426 | s = get_section (buf, SECTION_WRITE, NULL); | |
5427 | switch_to_section (s); | |
5428 | assemble_align (POINTER_SIZE); | |
28514dda | 5429 | assemble_aligned_integer (POINTER_BYTES, symbol); |
43e9d192 IB |
5430 | } |
5431 | } | |
5432 | ||
5433 | const char* | |
5434 | aarch64_output_casesi (rtx *operands) | |
5435 | { | |
5436 | char buf[100]; | |
5437 | char label[100]; | |
b32d5189 | 5438 | rtx diff_vec = PATTERN (NEXT_INSN (as_a <rtx_insn *> (operands[2]))); |
43e9d192 IB |
5439 | int index; |
5440 | static const char *const patterns[4][2] = | |
5441 | { | |
5442 | { | |
5443 | "ldrb\t%w3, [%0,%w1,uxtw]", | |
5444 | "add\t%3, %4, %w3, sxtb #2" | |
5445 | }, | |
5446 | { | |
5447 | "ldrh\t%w3, [%0,%w1,uxtw #1]", | |
5448 | "add\t%3, %4, %w3, sxth #2" | |
5449 | }, | |
5450 | { | |
5451 | "ldr\t%w3, [%0,%w1,uxtw #2]", | |
5452 | "add\t%3, %4, %w3, sxtw #2" | |
5453 | }, | |
5454 | /* We assume that DImode is only generated when not optimizing and | |
5455 | that we don't really need 64-bit address offsets. That would | |
5456 | imply an object file with 8GB of code in a single function! */ | |
5457 | { | |
5458 | "ldr\t%w3, [%0,%w1,uxtw #2]", | |
5459 | "add\t%3, %4, %w3, sxtw #2" | |
5460 | } | |
5461 | }; | |
5462 | ||
5463 | gcc_assert (GET_CODE (diff_vec) == ADDR_DIFF_VEC); | |
5464 | ||
5465 | index = exact_log2 (GET_MODE_SIZE (GET_MODE (diff_vec))); | |
5466 | ||
5467 | gcc_assert (index >= 0 && index <= 3); | |
5468 | ||
5469 | /* Need to implement table size reduction, by chaning the code below. */ | |
5470 | output_asm_insn (patterns[index][0], operands); | |
5471 | ASM_GENERATE_INTERNAL_LABEL (label, "Lrtx", CODE_LABEL_NUMBER (operands[2])); | |
5472 | snprintf (buf, sizeof (buf), | |
5473 | "adr\t%%4, %s", targetm.strip_name_encoding (label)); | |
5474 | output_asm_insn (buf, operands); | |
5475 | output_asm_insn (patterns[index][1], operands); | |
5476 | output_asm_insn ("br\t%3", operands); | |
5477 | assemble_label (asm_out_file, label); | |
5478 | return ""; | |
5479 | } | |
5480 | ||
5481 | ||
5482 | /* Return size in bits of an arithmetic operand which is shifted/scaled and | |
5483 | masked such that it is suitable for a UXTB, UXTH, or UXTW extend | |
5484 | operator. */ | |
5485 | ||
5486 | int | |
5487 | aarch64_uxt_size (int shift, HOST_WIDE_INT mask) | |
5488 | { | |
5489 | if (shift >= 0 && shift <= 3) | |
5490 | { | |
5491 | int size; | |
5492 | for (size = 8; size <= 32; size *= 2) | |
5493 | { | |
5494 | HOST_WIDE_INT bits = ((HOST_WIDE_INT)1U << size) - 1; | |
5495 | if (mask == bits << shift) | |
5496 | return size; | |
5497 | } | |
5498 | } | |
5499 | return 0; | |
5500 | } | |
5501 | ||
e78d485e RR |
5502 | /* Constant pools are per function only when PC relative |
5503 | literal loads are true or we are in the large memory | |
5504 | model. */ | |
5505 | ||
5506 | static inline bool | |
5507 | aarch64_can_use_per_function_literal_pools_p (void) | |
5508 | { | |
9ee6540a | 5509 | return (aarch64_pcrelative_literal_loads |
e78d485e RR |
5510 | || aarch64_cmodel == AARCH64_CMODEL_LARGE); |
5511 | } | |
5512 | ||
43e9d192 | 5513 | static bool |
e78d485e | 5514 | aarch64_use_blocks_for_constant_p (machine_mode, const_rtx) |
43e9d192 | 5515 | { |
3eece53d RR |
5516 | /* Fixme:: In an ideal world this would work similar |
5517 | to the logic in aarch64_select_rtx_section but this | |
5518 | breaks bootstrap in gcc go. For now we workaround | |
5519 | this by returning false here. */ | |
5520 | return false; | |
43e9d192 IB |
5521 | } |
5522 | ||
e78d485e RR |
5523 | /* Select appropriate section for constants depending |
5524 | on where we place literal pools. */ | |
5525 | ||
43e9d192 | 5526 | static section * |
e78d485e RR |
5527 | aarch64_select_rtx_section (machine_mode mode, |
5528 | rtx x, | |
5529 | unsigned HOST_WIDE_INT align) | |
43e9d192 | 5530 | { |
e78d485e RR |
5531 | if (aarch64_can_use_per_function_literal_pools_p ()) |
5532 | return function_section (current_function_decl); | |
43e9d192 | 5533 | |
e78d485e RR |
5534 | return default_elf_select_rtx_section (mode, x, align); |
5535 | } | |
43e9d192 | 5536 | |
5fca7b66 RH |
5537 | /* Implement ASM_OUTPUT_POOL_EPILOGUE. */ |
5538 | void | |
5539 | aarch64_asm_output_pool_epilogue (FILE *f, const char *, tree, | |
5540 | HOST_WIDE_INT offset) | |
5541 | { | |
5542 | /* When using per-function literal pools, we must ensure that any code | |
5543 | section is aligned to the minimal instruction length, lest we get | |
5544 | errors from the assembler re "unaligned instructions". */ | |
5545 | if ((offset & 3) && aarch64_can_use_per_function_literal_pools_p ()) | |
5546 | ASM_OUTPUT_ALIGN (f, 2); | |
5547 | } | |
5548 | ||
43e9d192 IB |
5549 | /* Costs. */ |
5550 | ||
5551 | /* Helper function for rtx cost calculation. Strip a shift expression | |
5552 | from X. Returns the inner operand if successful, or the original | |
5553 | expression on failure. */ | |
5554 | static rtx | |
5555 | aarch64_strip_shift (rtx x) | |
5556 | { | |
5557 | rtx op = x; | |
5558 | ||
57b77d46 RE |
5559 | /* We accept both ROTATERT and ROTATE: since the RHS must be a constant |
5560 | we can convert both to ROR during final output. */ | |
43e9d192 IB |
5561 | if ((GET_CODE (op) == ASHIFT |
5562 | || GET_CODE (op) == ASHIFTRT | |
57b77d46 RE |
5563 | || GET_CODE (op) == LSHIFTRT |
5564 | || GET_CODE (op) == ROTATERT | |
5565 | || GET_CODE (op) == ROTATE) | |
43e9d192 IB |
5566 | && CONST_INT_P (XEXP (op, 1))) |
5567 | return XEXP (op, 0); | |
5568 | ||
5569 | if (GET_CODE (op) == MULT | |
5570 | && CONST_INT_P (XEXP (op, 1)) | |
5571 | && ((unsigned) exact_log2 (INTVAL (XEXP (op, 1)))) < 64) | |
5572 | return XEXP (op, 0); | |
5573 | ||
5574 | return x; | |
5575 | } | |
5576 | ||
4745e701 | 5577 | /* Helper function for rtx cost calculation. Strip an extend |
43e9d192 IB |
5578 | expression from X. Returns the inner operand if successful, or the |
5579 | original expression on failure. We deal with a number of possible | |
5580 | canonicalization variations here. */ | |
5581 | static rtx | |
4745e701 | 5582 | aarch64_strip_extend (rtx x) |
43e9d192 IB |
5583 | { |
5584 | rtx op = x; | |
5585 | ||
5586 | /* Zero and sign extraction of a widened value. */ | |
5587 | if ((GET_CODE (op) == ZERO_EXTRACT || GET_CODE (op) == SIGN_EXTRACT) | |
5588 | && XEXP (op, 2) == const0_rtx | |
4745e701 | 5589 | && GET_CODE (XEXP (op, 0)) == MULT |
43e9d192 IB |
5590 | && aarch64_is_extend_from_extract (GET_MODE (op), XEXP (XEXP (op, 0), 1), |
5591 | XEXP (op, 1))) | |
5592 | return XEXP (XEXP (op, 0), 0); | |
5593 | ||
5594 | /* It can also be represented (for zero-extend) as an AND with an | |
5595 | immediate. */ | |
5596 | if (GET_CODE (op) == AND | |
5597 | && GET_CODE (XEXP (op, 0)) == MULT | |
5598 | && CONST_INT_P (XEXP (XEXP (op, 0), 1)) | |
5599 | && CONST_INT_P (XEXP (op, 1)) | |
5600 | && aarch64_uxt_size (exact_log2 (INTVAL (XEXP (XEXP (op, 0), 1))), | |
5601 | INTVAL (XEXP (op, 1))) != 0) | |
5602 | return XEXP (XEXP (op, 0), 0); | |
5603 | ||
5604 | /* Now handle extended register, as this may also have an optional | |
5605 | left shift by 1..4. */ | |
5606 | if (GET_CODE (op) == ASHIFT | |
5607 | && CONST_INT_P (XEXP (op, 1)) | |
5608 | && ((unsigned HOST_WIDE_INT) INTVAL (XEXP (op, 1))) <= 4) | |
5609 | op = XEXP (op, 0); | |
5610 | ||
5611 | if (GET_CODE (op) == ZERO_EXTEND | |
5612 | || GET_CODE (op) == SIGN_EXTEND) | |
5613 | op = XEXP (op, 0); | |
5614 | ||
5615 | if (op != x) | |
5616 | return op; | |
5617 | ||
4745e701 JG |
5618 | return x; |
5619 | } | |
5620 | ||
0a78ebe4 KT |
5621 | /* Return true iff CODE is a shift supported in combination |
5622 | with arithmetic instructions. */ | |
4d1919ed | 5623 | |
0a78ebe4 KT |
5624 | static bool |
5625 | aarch64_shift_p (enum rtx_code code) | |
5626 | { | |
5627 | return code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT; | |
5628 | } | |
5629 | ||
4745e701 | 5630 | /* Helper function for rtx cost calculation. Calculate the cost of |
0a78ebe4 KT |
5631 | a MULT or ASHIFT, which may be part of a compound PLUS/MINUS rtx. |
5632 | Return the calculated cost of the expression, recursing manually in to | |
4745e701 JG |
5633 | operands where needed. */ |
5634 | ||
5635 | static int | |
e548c9df | 5636 | aarch64_rtx_mult_cost (rtx x, enum rtx_code code, int outer, bool speed) |
4745e701 JG |
5637 | { |
5638 | rtx op0, op1; | |
5639 | const struct cpu_cost_table *extra_cost | |
b175b679 | 5640 | = aarch64_tune_params.insn_extra_cost; |
4745e701 | 5641 | int cost = 0; |
0a78ebe4 | 5642 | bool compound_p = (outer == PLUS || outer == MINUS); |
ef4bddc2 | 5643 | machine_mode mode = GET_MODE (x); |
4745e701 JG |
5644 | |
5645 | gcc_checking_assert (code == MULT); | |
5646 | ||
5647 | op0 = XEXP (x, 0); | |
5648 | op1 = XEXP (x, 1); | |
5649 | ||
5650 | if (VECTOR_MODE_P (mode)) | |
5651 | mode = GET_MODE_INNER (mode); | |
5652 | ||
5653 | /* Integer multiply/fma. */ | |
5654 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
5655 | { | |
5656 | /* The multiply will be canonicalized as a shift, cost it as such. */ | |
0a78ebe4 KT |
5657 | if (aarch64_shift_p (GET_CODE (x)) |
5658 | || (CONST_INT_P (op1) | |
5659 | && exact_log2 (INTVAL (op1)) > 0)) | |
4745e701 | 5660 | { |
0a78ebe4 KT |
5661 | bool is_extend = GET_CODE (op0) == ZERO_EXTEND |
5662 | || GET_CODE (op0) == SIGN_EXTEND; | |
4745e701 JG |
5663 | if (speed) |
5664 | { | |
0a78ebe4 KT |
5665 | if (compound_p) |
5666 | { | |
5667 | if (REG_P (op1)) | |
5668 | /* ARITH + shift-by-register. */ | |
5669 | cost += extra_cost->alu.arith_shift_reg; | |
5670 | else if (is_extend) | |
5671 | /* ARITH + extended register. We don't have a cost field | |
5672 | for ARITH+EXTEND+SHIFT, so use extend_arith here. */ | |
5673 | cost += extra_cost->alu.extend_arith; | |
5674 | else | |
5675 | /* ARITH + shift-by-immediate. */ | |
5676 | cost += extra_cost->alu.arith_shift; | |
5677 | } | |
4745e701 JG |
5678 | else |
5679 | /* LSL (immediate). */ | |
0a78ebe4 KT |
5680 | cost += extra_cost->alu.shift; |
5681 | ||
4745e701 | 5682 | } |
0a78ebe4 KT |
5683 | /* Strip extends as we will have costed them in the case above. */ |
5684 | if (is_extend) | |
5685 | op0 = aarch64_strip_extend (op0); | |
4745e701 | 5686 | |
e548c9df | 5687 | cost += rtx_cost (op0, VOIDmode, code, 0, speed); |
4745e701 JG |
5688 | |
5689 | return cost; | |
5690 | } | |
5691 | ||
d2ac256b KT |
5692 | /* MNEG or [US]MNEGL. Extract the NEG operand and indicate that it's a |
5693 | compound and let the below cases handle it. After all, MNEG is a | |
5694 | special-case alias of MSUB. */ | |
5695 | if (GET_CODE (op0) == NEG) | |
5696 | { | |
5697 | op0 = XEXP (op0, 0); | |
5698 | compound_p = true; | |
5699 | } | |
5700 | ||
4745e701 JG |
5701 | /* Integer multiplies or FMAs have zero/sign extending variants. */ |
5702 | if ((GET_CODE (op0) == ZERO_EXTEND | |
5703 | && GET_CODE (op1) == ZERO_EXTEND) | |
5704 | || (GET_CODE (op0) == SIGN_EXTEND | |
5705 | && GET_CODE (op1) == SIGN_EXTEND)) | |
5706 | { | |
e548c9df AM |
5707 | cost += rtx_cost (XEXP (op0, 0), VOIDmode, MULT, 0, speed); |
5708 | cost += rtx_cost (XEXP (op1, 0), VOIDmode, MULT, 1, speed); | |
4745e701 JG |
5709 | |
5710 | if (speed) | |
5711 | { | |
0a78ebe4 | 5712 | if (compound_p) |
d2ac256b | 5713 | /* SMADDL/UMADDL/UMSUBL/SMSUBL. */ |
4745e701 JG |
5714 | cost += extra_cost->mult[0].extend_add; |
5715 | else | |
5716 | /* MUL/SMULL/UMULL. */ | |
5717 | cost += extra_cost->mult[0].extend; | |
5718 | } | |
5719 | ||
5720 | return cost; | |
5721 | } | |
5722 | ||
d2ac256b | 5723 | /* This is either an integer multiply or a MADD. In both cases |
4745e701 | 5724 | we want to recurse and cost the operands. */ |
e548c9df AM |
5725 | cost += rtx_cost (op0, mode, MULT, 0, speed); |
5726 | cost += rtx_cost (op1, mode, MULT, 1, speed); | |
4745e701 JG |
5727 | |
5728 | if (speed) | |
5729 | { | |
0a78ebe4 | 5730 | if (compound_p) |
d2ac256b | 5731 | /* MADD/MSUB. */ |
4745e701 JG |
5732 | cost += extra_cost->mult[mode == DImode].add; |
5733 | else | |
5734 | /* MUL. */ | |
5735 | cost += extra_cost->mult[mode == DImode].simple; | |
5736 | } | |
5737 | ||
5738 | return cost; | |
5739 | } | |
5740 | else | |
5741 | { | |
5742 | if (speed) | |
5743 | { | |
3d840f7d | 5744 | /* Floating-point FMA/FMUL can also support negations of the |
d318517d SN |
5745 | operands, unless the rounding mode is upward or downward in |
5746 | which case FNMUL is different than FMUL with operand negation. */ | |
5747 | bool neg0 = GET_CODE (op0) == NEG; | |
5748 | bool neg1 = GET_CODE (op1) == NEG; | |
5749 | if (compound_p || !flag_rounding_math || (neg0 && neg1)) | |
5750 | { | |
5751 | if (neg0) | |
5752 | op0 = XEXP (op0, 0); | |
5753 | if (neg1) | |
5754 | op1 = XEXP (op1, 0); | |
5755 | } | |
4745e701 | 5756 | |
0a78ebe4 | 5757 | if (compound_p) |
4745e701 JG |
5758 | /* FMADD/FNMADD/FNMSUB/FMSUB. */ |
5759 | cost += extra_cost->fp[mode == DFmode].fma; | |
5760 | else | |
3d840f7d | 5761 | /* FMUL/FNMUL. */ |
4745e701 JG |
5762 | cost += extra_cost->fp[mode == DFmode].mult; |
5763 | } | |
5764 | ||
e548c9df AM |
5765 | cost += rtx_cost (op0, mode, MULT, 0, speed); |
5766 | cost += rtx_cost (op1, mode, MULT, 1, speed); | |
4745e701 JG |
5767 | return cost; |
5768 | } | |
43e9d192 IB |
5769 | } |
5770 | ||
67747367 JG |
5771 | static int |
5772 | aarch64_address_cost (rtx x, | |
ef4bddc2 | 5773 | machine_mode mode, |
67747367 JG |
5774 | addr_space_t as ATTRIBUTE_UNUSED, |
5775 | bool speed) | |
5776 | { | |
5777 | enum rtx_code c = GET_CODE (x); | |
b175b679 | 5778 | const struct cpu_addrcost_table *addr_cost = aarch64_tune_params.addr_cost; |
67747367 JG |
5779 | struct aarch64_address_info info; |
5780 | int cost = 0; | |
5781 | info.shift = 0; | |
5782 | ||
5783 | if (!aarch64_classify_address (&info, x, mode, c, false)) | |
5784 | { | |
5785 | if (GET_CODE (x) == CONST || GET_CODE (x) == SYMBOL_REF) | |
5786 | { | |
5787 | /* This is a CONST or SYMBOL ref which will be split | |
5788 | in a different way depending on the code model in use. | |
5789 | Cost it through the generic infrastructure. */ | |
e548c9df | 5790 | int cost_symbol_ref = rtx_cost (x, Pmode, MEM, 1, speed); |
67747367 JG |
5791 | /* Divide through by the cost of one instruction to |
5792 | bring it to the same units as the address costs. */ | |
5793 | cost_symbol_ref /= COSTS_N_INSNS (1); | |
5794 | /* The cost is then the cost of preparing the address, | |
5795 | followed by an immediate (possibly 0) offset. */ | |
5796 | return cost_symbol_ref + addr_cost->imm_offset; | |
5797 | } | |
5798 | else | |
5799 | { | |
5800 | /* This is most likely a jump table from a case | |
5801 | statement. */ | |
5802 | return addr_cost->register_offset; | |
5803 | } | |
5804 | } | |
5805 | ||
5806 | switch (info.type) | |
5807 | { | |
5808 | case ADDRESS_LO_SUM: | |
5809 | case ADDRESS_SYMBOLIC: | |
5810 | case ADDRESS_REG_IMM: | |
5811 | cost += addr_cost->imm_offset; | |
5812 | break; | |
5813 | ||
5814 | case ADDRESS_REG_WB: | |
5815 | if (c == PRE_INC || c == PRE_DEC || c == PRE_MODIFY) | |
5816 | cost += addr_cost->pre_modify; | |
5817 | else if (c == POST_INC || c == POST_DEC || c == POST_MODIFY) | |
5818 | cost += addr_cost->post_modify; | |
5819 | else | |
5820 | gcc_unreachable (); | |
5821 | ||
5822 | break; | |
5823 | ||
5824 | case ADDRESS_REG_REG: | |
5825 | cost += addr_cost->register_offset; | |
5826 | break; | |
5827 | ||
67747367 | 5828 | case ADDRESS_REG_SXTW: |
783879e6 EM |
5829 | cost += addr_cost->register_sextend; |
5830 | break; | |
5831 | ||
5832 | case ADDRESS_REG_UXTW: | |
5833 | cost += addr_cost->register_zextend; | |
67747367 JG |
5834 | break; |
5835 | ||
5836 | default: | |
5837 | gcc_unreachable (); | |
5838 | } | |
5839 | ||
5840 | ||
5841 | if (info.shift > 0) | |
5842 | { | |
5843 | /* For the sake of calculating the cost of the shifted register | |
5844 | component, we can treat same sized modes in the same way. */ | |
5845 | switch (GET_MODE_BITSIZE (mode)) | |
5846 | { | |
5847 | case 16: | |
5848 | cost += addr_cost->addr_scale_costs.hi; | |
5849 | break; | |
5850 | ||
5851 | case 32: | |
5852 | cost += addr_cost->addr_scale_costs.si; | |
5853 | break; | |
5854 | ||
5855 | case 64: | |
5856 | cost += addr_cost->addr_scale_costs.di; | |
5857 | break; | |
5858 | ||
5859 | /* We can't tell, or this is a 128-bit vector. */ | |
5860 | default: | |
5861 | cost += addr_cost->addr_scale_costs.ti; | |
5862 | break; | |
5863 | } | |
5864 | } | |
5865 | ||
5866 | return cost; | |
5867 | } | |
5868 | ||
b9066f5a MW |
5869 | /* Return the cost of a branch. If SPEED_P is true then the compiler is |
5870 | optimizing for speed. If PREDICTABLE_P is true then the branch is predicted | |
5871 | to be taken. */ | |
5872 | ||
5873 | int | |
5874 | aarch64_branch_cost (bool speed_p, bool predictable_p) | |
5875 | { | |
5876 | /* When optimizing for speed, use the cost of unpredictable branches. */ | |
5877 | const struct cpu_branch_cost *branch_costs = | |
b175b679 | 5878 | aarch64_tune_params.branch_costs; |
b9066f5a MW |
5879 | |
5880 | if (!speed_p || predictable_p) | |
5881 | return branch_costs->predictable; | |
5882 | else | |
5883 | return branch_costs->unpredictable; | |
5884 | } | |
5885 | ||
7cc2145f JG |
5886 | /* Return true if the RTX X in mode MODE is a zero or sign extract |
5887 | usable in an ADD or SUB (extended register) instruction. */ | |
5888 | static bool | |
ef4bddc2 | 5889 | aarch64_rtx_arith_op_extract_p (rtx x, machine_mode mode) |
7cc2145f JG |
5890 | { |
5891 | /* Catch add with a sign extract. | |
5892 | This is add_<optab><mode>_multp2. */ | |
5893 | if (GET_CODE (x) == SIGN_EXTRACT | |
5894 | || GET_CODE (x) == ZERO_EXTRACT) | |
5895 | { | |
5896 | rtx op0 = XEXP (x, 0); | |
5897 | rtx op1 = XEXP (x, 1); | |
5898 | rtx op2 = XEXP (x, 2); | |
5899 | ||
5900 | if (GET_CODE (op0) == MULT | |
5901 | && CONST_INT_P (op1) | |
5902 | && op2 == const0_rtx | |
5903 | && CONST_INT_P (XEXP (op0, 1)) | |
5904 | && aarch64_is_extend_from_extract (mode, | |
5905 | XEXP (op0, 1), | |
5906 | op1)) | |
5907 | { | |
5908 | return true; | |
5909 | } | |
5910 | } | |
e47c4031 KT |
5911 | /* The simple case <ARITH>, XD, XN, XM, [us]xt. |
5912 | No shift. */ | |
5913 | else if (GET_CODE (x) == SIGN_EXTEND | |
5914 | || GET_CODE (x) == ZERO_EXTEND) | |
5915 | return REG_P (XEXP (x, 0)); | |
7cc2145f JG |
5916 | |
5917 | return false; | |
5918 | } | |
5919 | ||
61263118 KT |
5920 | static bool |
5921 | aarch64_frint_unspec_p (unsigned int u) | |
5922 | { | |
5923 | switch (u) | |
5924 | { | |
5925 | case UNSPEC_FRINTZ: | |
5926 | case UNSPEC_FRINTP: | |
5927 | case UNSPEC_FRINTM: | |
5928 | case UNSPEC_FRINTA: | |
5929 | case UNSPEC_FRINTN: | |
5930 | case UNSPEC_FRINTX: | |
5931 | case UNSPEC_FRINTI: | |
5932 | return true; | |
5933 | ||
5934 | default: | |
5935 | return false; | |
5936 | } | |
5937 | } | |
5938 | ||
fb0cb7fa KT |
5939 | /* Return true iff X is an rtx that will match an extr instruction |
5940 | i.e. as described in the *extr<mode>5_insn family of patterns. | |
5941 | OP0 and OP1 will be set to the operands of the shifts involved | |
5942 | on success and will be NULL_RTX otherwise. */ | |
5943 | ||
5944 | static bool | |
5945 | aarch64_extr_rtx_p (rtx x, rtx *res_op0, rtx *res_op1) | |
5946 | { | |
5947 | rtx op0, op1; | |
5948 | machine_mode mode = GET_MODE (x); | |
5949 | ||
5950 | *res_op0 = NULL_RTX; | |
5951 | *res_op1 = NULL_RTX; | |
5952 | ||
5953 | if (GET_CODE (x) != IOR) | |
5954 | return false; | |
5955 | ||
5956 | op0 = XEXP (x, 0); | |
5957 | op1 = XEXP (x, 1); | |
5958 | ||
5959 | if ((GET_CODE (op0) == ASHIFT && GET_CODE (op1) == LSHIFTRT) | |
5960 | || (GET_CODE (op1) == ASHIFT && GET_CODE (op0) == LSHIFTRT)) | |
5961 | { | |
5962 | /* Canonicalise locally to ashift in op0, lshiftrt in op1. */ | |
5963 | if (GET_CODE (op1) == ASHIFT) | |
5964 | std::swap (op0, op1); | |
5965 | ||
5966 | if (!CONST_INT_P (XEXP (op0, 1)) || !CONST_INT_P (XEXP (op1, 1))) | |
5967 | return false; | |
5968 | ||
5969 | unsigned HOST_WIDE_INT shft_amnt_0 = UINTVAL (XEXP (op0, 1)); | |
5970 | unsigned HOST_WIDE_INT shft_amnt_1 = UINTVAL (XEXP (op1, 1)); | |
5971 | ||
5972 | if (shft_amnt_0 < GET_MODE_BITSIZE (mode) | |
5973 | && shft_amnt_0 + shft_amnt_1 == GET_MODE_BITSIZE (mode)) | |
5974 | { | |
5975 | *res_op0 = XEXP (op0, 0); | |
5976 | *res_op1 = XEXP (op1, 0); | |
5977 | return true; | |
5978 | } | |
5979 | } | |
5980 | ||
5981 | return false; | |
5982 | } | |
5983 | ||
2d5ffe46 AP |
5984 | /* Calculate the cost of calculating (if_then_else (OP0) (OP1) (OP2)), |
5985 | storing it in *COST. Result is true if the total cost of the operation | |
5986 | has now been calculated. */ | |
5987 | static bool | |
5988 | aarch64_if_then_else_costs (rtx op0, rtx op1, rtx op2, int *cost, bool speed) | |
5989 | { | |
b9e3afe9 AP |
5990 | rtx inner; |
5991 | rtx comparator; | |
5992 | enum rtx_code cmpcode; | |
5993 | ||
5994 | if (COMPARISON_P (op0)) | |
5995 | { | |
5996 | inner = XEXP (op0, 0); | |
5997 | comparator = XEXP (op0, 1); | |
5998 | cmpcode = GET_CODE (op0); | |
5999 | } | |
6000 | else | |
6001 | { | |
6002 | inner = op0; | |
6003 | comparator = const0_rtx; | |
6004 | cmpcode = NE; | |
6005 | } | |
6006 | ||
2d5ffe46 AP |
6007 | if (GET_CODE (op1) == PC || GET_CODE (op2) == PC) |
6008 | { | |
6009 | /* Conditional branch. */ | |
b9e3afe9 | 6010 | if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_CC) |
2d5ffe46 AP |
6011 | return true; |
6012 | else | |
6013 | { | |
b9e3afe9 | 6014 | if (cmpcode == NE || cmpcode == EQ) |
2d5ffe46 | 6015 | { |
2d5ffe46 AP |
6016 | if (comparator == const0_rtx) |
6017 | { | |
6018 | /* TBZ/TBNZ/CBZ/CBNZ. */ | |
6019 | if (GET_CODE (inner) == ZERO_EXTRACT) | |
6020 | /* TBZ/TBNZ. */ | |
e548c9df AM |
6021 | *cost += rtx_cost (XEXP (inner, 0), VOIDmode, |
6022 | ZERO_EXTRACT, 0, speed); | |
6023 | else | |
6024 | /* CBZ/CBNZ. */ | |
6025 | *cost += rtx_cost (inner, VOIDmode, cmpcode, 0, speed); | |
2d5ffe46 AP |
6026 | |
6027 | return true; | |
6028 | } | |
6029 | } | |
b9e3afe9 | 6030 | else if (cmpcode == LT || cmpcode == GE) |
2d5ffe46 | 6031 | { |
2d5ffe46 AP |
6032 | /* TBZ/TBNZ. */ |
6033 | if (comparator == const0_rtx) | |
6034 | return true; | |
6035 | } | |
6036 | } | |
6037 | } | |
b9e3afe9 | 6038 | else if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_CC) |
2d5ffe46 | 6039 | { |
786298dc | 6040 | /* CCMP. */ |
6dfeb7ce | 6041 | if (GET_CODE (op1) == COMPARE) |
786298dc WD |
6042 | { |
6043 | /* Increase cost of CCMP reg, 0, imm, CC to prefer CMP reg, 0. */ | |
6044 | if (XEXP (op1, 1) == const0_rtx) | |
6045 | *cost += 1; | |
6046 | if (speed) | |
6047 | { | |
6048 | machine_mode mode = GET_MODE (XEXP (op1, 0)); | |
6049 | const struct cpu_cost_table *extra_cost | |
6050 | = aarch64_tune_params.insn_extra_cost; | |
6051 | ||
6052 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
6053 | *cost += extra_cost->alu.arith; | |
6054 | else | |
6055 | *cost += extra_cost->fp[mode == DFmode].compare; | |
6056 | } | |
6057 | return true; | |
6058 | } | |
6059 | ||
2d5ffe46 AP |
6060 | /* It's a conditional operation based on the status flags, |
6061 | so it must be some flavor of CSEL. */ | |
6062 | ||
6063 | /* CSNEG, CSINV, and CSINC are handled for free as part of CSEL. */ | |
6064 | if (GET_CODE (op1) == NEG | |
6065 | || GET_CODE (op1) == NOT | |
6066 | || (GET_CODE (op1) == PLUS && XEXP (op1, 1) == const1_rtx)) | |
6067 | op1 = XEXP (op1, 0); | |
bad00732 KT |
6068 | else if (GET_CODE (op1) == ZERO_EXTEND && GET_CODE (op2) == ZERO_EXTEND) |
6069 | { | |
6070 | /* CSEL with zero-extension (*cmovdi_insn_uxtw). */ | |
6071 | op1 = XEXP (op1, 0); | |
6072 | op2 = XEXP (op2, 0); | |
6073 | } | |
2d5ffe46 | 6074 | |
e548c9df AM |
6075 | *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed); |
6076 | *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed); | |
2d5ffe46 AP |
6077 | return true; |
6078 | } | |
6079 | ||
6080 | /* We don't know what this is, cost all operands. */ | |
6081 | return false; | |
6082 | } | |
6083 | ||
283b6c85 KT |
6084 | /* Check whether X is a bitfield operation of the form shift + extend that |
6085 | maps down to a UBFIZ/SBFIZ/UBFX/SBFX instruction. If so, return the | |
6086 | operand to which the bitfield operation is applied. Otherwise return | |
6087 | NULL_RTX. */ | |
6088 | ||
6089 | static rtx | |
6090 | aarch64_extend_bitfield_pattern_p (rtx x) | |
6091 | { | |
6092 | rtx_code outer_code = GET_CODE (x); | |
6093 | machine_mode outer_mode = GET_MODE (x); | |
6094 | ||
6095 | if (outer_code != ZERO_EXTEND && outer_code != SIGN_EXTEND | |
6096 | && outer_mode != SImode && outer_mode != DImode) | |
6097 | return NULL_RTX; | |
6098 | ||
6099 | rtx inner = XEXP (x, 0); | |
6100 | rtx_code inner_code = GET_CODE (inner); | |
6101 | machine_mode inner_mode = GET_MODE (inner); | |
6102 | rtx op = NULL_RTX; | |
6103 | ||
6104 | switch (inner_code) | |
6105 | { | |
6106 | case ASHIFT: | |
6107 | if (CONST_INT_P (XEXP (inner, 1)) | |
6108 | && (inner_mode == QImode || inner_mode == HImode)) | |
6109 | op = XEXP (inner, 0); | |
6110 | break; | |
6111 | case LSHIFTRT: | |
6112 | if (outer_code == ZERO_EXTEND && CONST_INT_P (XEXP (inner, 1)) | |
6113 | && (inner_mode == QImode || inner_mode == HImode)) | |
6114 | op = XEXP (inner, 0); | |
6115 | break; | |
6116 | case ASHIFTRT: | |
6117 | if (outer_code == SIGN_EXTEND && CONST_INT_P (XEXP (inner, 1)) | |
6118 | && (inner_mode == QImode || inner_mode == HImode)) | |
6119 | op = XEXP (inner, 0); | |
6120 | break; | |
6121 | default: | |
6122 | break; | |
6123 | } | |
6124 | ||
6125 | return op; | |
6126 | } | |
6127 | ||
8c83f71d KT |
6128 | /* Return true if the mask and a shift amount from an RTX of the form |
6129 | (x << SHFT_AMNT) & MASK are valid to combine into a UBFIZ instruction of | |
6130 | mode MODE. See the *andim_ashift<mode>_bfiz pattern. */ | |
6131 | ||
6132 | bool | |
6133 | aarch64_mask_and_shift_for_ubfiz_p (machine_mode mode, rtx mask, rtx shft_amnt) | |
6134 | { | |
6135 | return CONST_INT_P (mask) && CONST_INT_P (shft_amnt) | |
6136 | && INTVAL (shft_amnt) < GET_MODE_BITSIZE (mode) | |
6137 | && exact_log2 ((INTVAL (mask) >> INTVAL (shft_amnt)) + 1) >= 0 | |
6138 | && (INTVAL (mask) & ((1 << INTVAL (shft_amnt)) - 1)) == 0; | |
6139 | } | |
6140 | ||
43e9d192 IB |
6141 | /* Calculate the cost of calculating X, storing it in *COST. Result |
6142 | is true if the total cost of the operation has now been calculated. */ | |
6143 | static bool | |
e548c9df | 6144 | aarch64_rtx_costs (rtx x, machine_mode mode, int outer ATTRIBUTE_UNUSED, |
43e9d192 IB |
6145 | int param ATTRIBUTE_UNUSED, int *cost, bool speed) |
6146 | { | |
a8eecd00 | 6147 | rtx op0, op1, op2; |
73250c4c | 6148 | const struct cpu_cost_table *extra_cost |
b175b679 | 6149 | = aarch64_tune_params.insn_extra_cost; |
e548c9df | 6150 | int code = GET_CODE (x); |
43e9d192 | 6151 | |
7fc5ef02 JG |
6152 | /* By default, assume that everything has equivalent cost to the |
6153 | cheapest instruction. Any additional costs are applied as a delta | |
6154 | above this default. */ | |
6155 | *cost = COSTS_N_INSNS (1); | |
6156 | ||
43e9d192 IB |
6157 | switch (code) |
6158 | { | |
6159 | case SET: | |
ba123b0d JG |
6160 | /* The cost depends entirely on the operands to SET. */ |
6161 | *cost = 0; | |
43e9d192 IB |
6162 | op0 = SET_DEST (x); |
6163 | op1 = SET_SRC (x); | |
6164 | ||
6165 | switch (GET_CODE (op0)) | |
6166 | { | |
6167 | case MEM: | |
6168 | if (speed) | |
2961177e JG |
6169 | { |
6170 | rtx address = XEXP (op0, 0); | |
b6875aac KV |
6171 | if (VECTOR_MODE_P (mode)) |
6172 | *cost += extra_cost->ldst.storev; | |
6173 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
2961177e JG |
6174 | *cost += extra_cost->ldst.store; |
6175 | else if (mode == SFmode) | |
6176 | *cost += extra_cost->ldst.storef; | |
6177 | else if (mode == DFmode) | |
6178 | *cost += extra_cost->ldst.stored; | |
6179 | ||
6180 | *cost += | |
6181 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
6182 | 0, speed)); | |
6183 | } | |
43e9d192 | 6184 | |
e548c9df | 6185 | *cost += rtx_cost (op1, mode, SET, 1, speed); |
43e9d192 IB |
6186 | return true; |
6187 | ||
6188 | case SUBREG: | |
6189 | if (! REG_P (SUBREG_REG (op0))) | |
e548c9df | 6190 | *cost += rtx_cost (SUBREG_REG (op0), VOIDmode, SET, 0, speed); |
ba123b0d | 6191 | |
43e9d192 IB |
6192 | /* Fall through. */ |
6193 | case REG: | |
b6875aac KV |
6194 | /* The cost is one per vector-register copied. */ |
6195 | if (VECTOR_MODE_P (GET_MODE (op0)) && REG_P (op1)) | |
6196 | { | |
6197 | int n_minus_1 = (GET_MODE_SIZE (GET_MODE (op0)) - 1) | |
6198 | / GET_MODE_SIZE (V4SImode); | |
6199 | *cost = COSTS_N_INSNS (n_minus_1 + 1); | |
6200 | } | |
ba123b0d JG |
6201 | /* const0_rtx is in general free, but we will use an |
6202 | instruction to set a register to 0. */ | |
b6875aac KV |
6203 | else if (REG_P (op1) || op1 == const0_rtx) |
6204 | { | |
6205 | /* The cost is 1 per register copied. */ | |
6206 | int n_minus_1 = (GET_MODE_SIZE (GET_MODE (op0)) - 1) | |
ba123b0d | 6207 | / UNITS_PER_WORD; |
b6875aac KV |
6208 | *cost = COSTS_N_INSNS (n_minus_1 + 1); |
6209 | } | |
ba123b0d JG |
6210 | else |
6211 | /* Cost is just the cost of the RHS of the set. */ | |
e548c9df | 6212 | *cost += rtx_cost (op1, mode, SET, 1, speed); |
43e9d192 IB |
6213 | return true; |
6214 | ||
ba123b0d | 6215 | case ZERO_EXTRACT: |
43e9d192 | 6216 | case SIGN_EXTRACT: |
ba123b0d JG |
6217 | /* Bit-field insertion. Strip any redundant widening of |
6218 | the RHS to meet the width of the target. */ | |
43e9d192 IB |
6219 | if (GET_CODE (op1) == SUBREG) |
6220 | op1 = SUBREG_REG (op1); | |
6221 | if ((GET_CODE (op1) == ZERO_EXTEND | |
6222 | || GET_CODE (op1) == SIGN_EXTEND) | |
4aa81c2e | 6223 | && CONST_INT_P (XEXP (op0, 1)) |
43e9d192 IB |
6224 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (op1, 0))) |
6225 | >= INTVAL (XEXP (op0, 1)))) | |
6226 | op1 = XEXP (op1, 0); | |
ba123b0d JG |
6227 | |
6228 | if (CONST_INT_P (op1)) | |
6229 | { | |
6230 | /* MOV immediate is assumed to always be cheap. */ | |
6231 | *cost = COSTS_N_INSNS (1); | |
6232 | } | |
6233 | else | |
6234 | { | |
6235 | /* BFM. */ | |
6236 | if (speed) | |
6237 | *cost += extra_cost->alu.bfi; | |
e548c9df | 6238 | *cost += rtx_cost (op1, VOIDmode, (enum rtx_code) code, 1, speed); |
ba123b0d JG |
6239 | } |
6240 | ||
43e9d192 IB |
6241 | return true; |
6242 | ||
6243 | default: | |
ba123b0d JG |
6244 | /* We can't make sense of this, assume default cost. */ |
6245 | *cost = COSTS_N_INSNS (1); | |
61263118 | 6246 | return false; |
43e9d192 IB |
6247 | } |
6248 | return false; | |
6249 | ||
9dfc162c JG |
6250 | case CONST_INT: |
6251 | /* If an instruction can incorporate a constant within the | |
6252 | instruction, the instruction's expression avoids calling | |
6253 | rtx_cost() on the constant. If rtx_cost() is called on a | |
6254 | constant, then it is usually because the constant must be | |
6255 | moved into a register by one or more instructions. | |
6256 | ||
6257 | The exception is constant 0, which can be expressed | |
6258 | as XZR/WZR and is therefore free. The exception to this is | |
6259 | if we have (set (reg) (const0_rtx)) in which case we must cost | |
6260 | the move. However, we can catch that when we cost the SET, so | |
6261 | we don't need to consider that here. */ | |
6262 | if (x == const0_rtx) | |
6263 | *cost = 0; | |
6264 | else | |
6265 | { | |
6266 | /* To an approximation, building any other constant is | |
6267 | proportionally expensive to the number of instructions | |
6268 | required to build that constant. This is true whether we | |
6269 | are compiling for SPEED or otherwise. */ | |
82614948 RR |
6270 | *cost = COSTS_N_INSNS (aarch64_internal_mov_immediate |
6271 | (NULL_RTX, x, false, mode)); | |
9dfc162c JG |
6272 | } |
6273 | return true; | |
6274 | ||
6275 | case CONST_DOUBLE: | |
6276 | if (speed) | |
6277 | { | |
6278 | /* mov[df,sf]_aarch64. */ | |
6279 | if (aarch64_float_const_representable_p (x)) | |
6280 | /* FMOV (scalar immediate). */ | |
6281 | *cost += extra_cost->fp[mode == DFmode].fpconst; | |
6282 | else if (!aarch64_float_const_zero_rtx_p (x)) | |
6283 | { | |
6284 | /* This will be a load from memory. */ | |
6285 | if (mode == DFmode) | |
6286 | *cost += extra_cost->ldst.loadd; | |
6287 | else | |
6288 | *cost += extra_cost->ldst.loadf; | |
6289 | } | |
6290 | else | |
6291 | /* Otherwise this is +0.0. We get this using MOVI d0, #0 | |
6292 | or MOV v0.s[0], wzr - neither of which are modeled by the | |
6293 | cost tables. Just use the default cost. */ | |
6294 | { | |
6295 | } | |
6296 | } | |
6297 | ||
6298 | return true; | |
6299 | ||
43e9d192 IB |
6300 | case MEM: |
6301 | if (speed) | |
2961177e JG |
6302 | { |
6303 | /* For loads we want the base cost of a load, plus an | |
6304 | approximation for the additional cost of the addressing | |
6305 | mode. */ | |
6306 | rtx address = XEXP (x, 0); | |
b6875aac KV |
6307 | if (VECTOR_MODE_P (mode)) |
6308 | *cost += extra_cost->ldst.loadv; | |
6309 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
2961177e JG |
6310 | *cost += extra_cost->ldst.load; |
6311 | else if (mode == SFmode) | |
6312 | *cost += extra_cost->ldst.loadf; | |
6313 | else if (mode == DFmode) | |
6314 | *cost += extra_cost->ldst.loadd; | |
6315 | ||
6316 | *cost += | |
6317 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
6318 | 0, speed)); | |
6319 | } | |
43e9d192 IB |
6320 | |
6321 | return true; | |
6322 | ||
6323 | case NEG: | |
4745e701 JG |
6324 | op0 = XEXP (x, 0); |
6325 | ||
b6875aac KV |
6326 | if (VECTOR_MODE_P (mode)) |
6327 | { | |
6328 | if (speed) | |
6329 | { | |
6330 | /* FNEG. */ | |
6331 | *cost += extra_cost->vect.alu; | |
6332 | } | |
6333 | return false; | |
6334 | } | |
6335 | ||
e548c9df AM |
6336 | if (GET_MODE_CLASS (mode) == MODE_INT) |
6337 | { | |
4745e701 JG |
6338 | if (GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMPARE |
6339 | || GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMM_COMPARE) | |
6340 | { | |
6341 | /* CSETM. */ | |
e548c9df | 6342 | *cost += rtx_cost (XEXP (op0, 0), VOIDmode, NEG, 0, speed); |
4745e701 JG |
6343 | return true; |
6344 | } | |
6345 | ||
6346 | /* Cost this as SUB wzr, X. */ | |
e548c9df | 6347 | op0 = CONST0_RTX (mode); |
4745e701 JG |
6348 | op1 = XEXP (x, 0); |
6349 | goto cost_minus; | |
6350 | } | |
6351 | ||
e548c9df | 6352 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
4745e701 JG |
6353 | { |
6354 | /* Support (neg(fma...)) as a single instruction only if | |
6355 | sign of zeros is unimportant. This matches the decision | |
6356 | making in aarch64.md. */ | |
6357 | if (GET_CODE (op0) == FMA && !HONOR_SIGNED_ZEROS (GET_MODE (op0))) | |
6358 | { | |
6359 | /* FNMADD. */ | |
e548c9df | 6360 | *cost = rtx_cost (op0, mode, NEG, 0, speed); |
4745e701 JG |
6361 | return true; |
6362 | } | |
d318517d SN |
6363 | if (GET_CODE (op0) == MULT) |
6364 | { | |
6365 | /* FNMUL. */ | |
6366 | *cost = rtx_cost (op0, mode, NEG, 0, speed); | |
6367 | return true; | |
6368 | } | |
4745e701 JG |
6369 | if (speed) |
6370 | /* FNEG. */ | |
6371 | *cost += extra_cost->fp[mode == DFmode].neg; | |
6372 | return false; | |
6373 | } | |
6374 | ||
6375 | return false; | |
43e9d192 | 6376 | |
781aeb73 KT |
6377 | case CLRSB: |
6378 | case CLZ: | |
6379 | if (speed) | |
b6875aac KV |
6380 | { |
6381 | if (VECTOR_MODE_P (mode)) | |
6382 | *cost += extra_cost->vect.alu; | |
6383 | else | |
6384 | *cost += extra_cost->alu.clz; | |
6385 | } | |
781aeb73 KT |
6386 | |
6387 | return false; | |
6388 | ||
43e9d192 IB |
6389 | case COMPARE: |
6390 | op0 = XEXP (x, 0); | |
6391 | op1 = XEXP (x, 1); | |
6392 | ||
6393 | if (op1 == const0_rtx | |
6394 | && GET_CODE (op0) == AND) | |
6395 | { | |
6396 | x = op0; | |
e548c9df | 6397 | mode = GET_MODE (op0); |
43e9d192 IB |
6398 | goto cost_logic; |
6399 | } | |
6400 | ||
a8eecd00 JG |
6401 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT) |
6402 | { | |
6403 | /* TODO: A write to the CC flags possibly costs extra, this | |
6404 | needs encoding in the cost tables. */ | |
6405 | ||
e548c9df | 6406 | mode = GET_MODE (op0); |
a8eecd00 JG |
6407 | /* ANDS. */ |
6408 | if (GET_CODE (op0) == AND) | |
6409 | { | |
6410 | x = op0; | |
6411 | goto cost_logic; | |
6412 | } | |
6413 | ||
6414 | if (GET_CODE (op0) == PLUS) | |
6415 | { | |
6416 | /* ADDS (and CMN alias). */ | |
6417 | x = op0; | |
6418 | goto cost_plus; | |
6419 | } | |
6420 | ||
6421 | if (GET_CODE (op0) == MINUS) | |
6422 | { | |
6423 | /* SUBS. */ | |
6424 | x = op0; | |
6425 | goto cost_minus; | |
6426 | } | |
6427 | ||
345854d8 KT |
6428 | if (GET_CODE (op0) == ZERO_EXTRACT && op1 == const0_rtx |
6429 | && GET_MODE (x) == CC_NZmode && CONST_INT_P (XEXP (op0, 1)) | |
6430 | && CONST_INT_P (XEXP (op0, 2))) | |
6431 | { | |
6432 | /* COMPARE of ZERO_EXTRACT form of TST-immediate. | |
6433 | Handle it here directly rather than going to cost_logic | |
6434 | since we know the immediate generated for the TST is valid | |
6435 | so we can avoid creating an intermediate rtx for it only | |
6436 | for costing purposes. */ | |
6437 | if (speed) | |
6438 | *cost += extra_cost->alu.logical; | |
6439 | ||
6440 | *cost += rtx_cost (XEXP (op0, 0), GET_MODE (op0), | |
6441 | ZERO_EXTRACT, 0, speed); | |
6442 | return true; | |
6443 | } | |
6444 | ||
a8eecd00 JG |
6445 | if (GET_CODE (op1) == NEG) |
6446 | { | |
6447 | /* CMN. */ | |
6448 | if (speed) | |
6449 | *cost += extra_cost->alu.arith; | |
6450 | ||
e548c9df AM |
6451 | *cost += rtx_cost (op0, mode, COMPARE, 0, speed); |
6452 | *cost += rtx_cost (XEXP (op1, 0), mode, NEG, 1, speed); | |
a8eecd00 JG |
6453 | return true; |
6454 | } | |
6455 | ||
6456 | /* CMP. | |
6457 | ||
6458 | Compare can freely swap the order of operands, and | |
6459 | canonicalization puts the more complex operation first. | |
6460 | But the integer MINUS logic expects the shift/extend | |
6461 | operation in op1. */ | |
6462 | if (! (REG_P (op0) | |
6463 | || (GET_CODE (op0) == SUBREG && REG_P (SUBREG_REG (op0))))) | |
6464 | { | |
6465 | op0 = XEXP (x, 1); | |
6466 | op1 = XEXP (x, 0); | |
6467 | } | |
6468 | goto cost_minus; | |
6469 | } | |
6470 | ||
6471 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT) | |
6472 | { | |
6473 | /* FCMP. */ | |
6474 | if (speed) | |
6475 | *cost += extra_cost->fp[mode == DFmode].compare; | |
6476 | ||
6477 | if (CONST_DOUBLE_P (op1) && aarch64_float_const_zero_rtx_p (op1)) | |
6478 | { | |
e548c9df | 6479 | *cost += rtx_cost (op0, VOIDmode, COMPARE, 0, speed); |
a8eecd00 JG |
6480 | /* FCMP supports constant 0.0 for no extra cost. */ |
6481 | return true; | |
6482 | } | |
6483 | return false; | |
6484 | } | |
6485 | ||
b6875aac KV |
6486 | if (VECTOR_MODE_P (mode)) |
6487 | { | |
6488 | /* Vector compare. */ | |
6489 | if (speed) | |
6490 | *cost += extra_cost->vect.alu; | |
6491 | ||
6492 | if (aarch64_float_const_zero_rtx_p (op1)) | |
6493 | { | |
6494 | /* Vector cm (eq|ge|gt|lt|le) supports constant 0.0 for no extra | |
6495 | cost. */ | |
6496 | return true; | |
6497 | } | |
6498 | return false; | |
6499 | } | |
a8eecd00 | 6500 | return false; |
43e9d192 IB |
6501 | |
6502 | case MINUS: | |
4745e701 JG |
6503 | { |
6504 | op0 = XEXP (x, 0); | |
6505 | op1 = XEXP (x, 1); | |
6506 | ||
6507 | cost_minus: | |
e548c9df | 6508 | *cost += rtx_cost (op0, mode, MINUS, 0, speed); |
23cb6618 | 6509 | |
4745e701 JG |
6510 | /* Detect valid immediates. */ |
6511 | if ((GET_MODE_CLASS (mode) == MODE_INT | |
6512 | || (GET_MODE_CLASS (mode) == MODE_CC | |
6513 | && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT)) | |
6514 | && CONST_INT_P (op1) | |
6515 | && aarch64_uimm12_shift (INTVAL (op1))) | |
6516 | { | |
4745e701 JG |
6517 | if (speed) |
6518 | /* SUB(S) (immediate). */ | |
6519 | *cost += extra_cost->alu.arith; | |
6520 | return true; | |
4745e701 JG |
6521 | } |
6522 | ||
7cc2145f JG |
6523 | /* Look for SUB (extended register). */ |
6524 | if (aarch64_rtx_arith_op_extract_p (op1, mode)) | |
6525 | { | |
6526 | if (speed) | |
2533c820 | 6527 | *cost += extra_cost->alu.extend_arith; |
7cc2145f | 6528 | |
e47c4031 KT |
6529 | op1 = aarch64_strip_extend (op1); |
6530 | *cost += rtx_cost (op1, VOIDmode, | |
e548c9df | 6531 | (enum rtx_code) GET_CODE (op1), 0, speed); |
7cc2145f JG |
6532 | return true; |
6533 | } | |
6534 | ||
4745e701 JG |
6535 | rtx new_op1 = aarch64_strip_extend (op1); |
6536 | ||
6537 | /* Cost this as an FMA-alike operation. */ | |
6538 | if ((GET_CODE (new_op1) == MULT | |
0a78ebe4 | 6539 | || aarch64_shift_p (GET_CODE (new_op1))) |
4745e701 JG |
6540 | && code != COMPARE) |
6541 | { | |
6542 | *cost += aarch64_rtx_mult_cost (new_op1, MULT, | |
6543 | (enum rtx_code) code, | |
6544 | speed); | |
4745e701 JG |
6545 | return true; |
6546 | } | |
43e9d192 | 6547 | |
e548c9df | 6548 | *cost += rtx_cost (new_op1, VOIDmode, MINUS, 1, speed); |
43e9d192 | 6549 | |
4745e701 JG |
6550 | if (speed) |
6551 | { | |
b6875aac KV |
6552 | if (VECTOR_MODE_P (mode)) |
6553 | { | |
6554 | /* Vector SUB. */ | |
6555 | *cost += extra_cost->vect.alu; | |
6556 | } | |
6557 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
6558 | { | |
6559 | /* SUB(S). */ | |
6560 | *cost += extra_cost->alu.arith; | |
6561 | } | |
4745e701 | 6562 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
b6875aac KV |
6563 | { |
6564 | /* FSUB. */ | |
6565 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
6566 | } | |
4745e701 JG |
6567 | } |
6568 | return true; | |
6569 | } | |
43e9d192 IB |
6570 | |
6571 | case PLUS: | |
4745e701 JG |
6572 | { |
6573 | rtx new_op0; | |
43e9d192 | 6574 | |
4745e701 JG |
6575 | op0 = XEXP (x, 0); |
6576 | op1 = XEXP (x, 1); | |
43e9d192 | 6577 | |
a8eecd00 | 6578 | cost_plus: |
4745e701 JG |
6579 | if (GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMPARE |
6580 | || GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMM_COMPARE) | |
6581 | { | |
6582 | /* CSINC. */ | |
e548c9df AM |
6583 | *cost += rtx_cost (XEXP (op0, 0), mode, PLUS, 0, speed); |
6584 | *cost += rtx_cost (op1, mode, PLUS, 1, speed); | |
4745e701 JG |
6585 | return true; |
6586 | } | |
43e9d192 | 6587 | |
4745e701 JG |
6588 | if (GET_MODE_CLASS (mode) == MODE_INT |
6589 | && CONST_INT_P (op1) | |
6590 | && aarch64_uimm12_shift (INTVAL (op1))) | |
6591 | { | |
e548c9df | 6592 | *cost += rtx_cost (op0, mode, PLUS, 0, speed); |
43e9d192 | 6593 | |
4745e701 JG |
6594 | if (speed) |
6595 | /* ADD (immediate). */ | |
6596 | *cost += extra_cost->alu.arith; | |
6597 | return true; | |
6598 | } | |
6599 | ||
e548c9df | 6600 | *cost += rtx_cost (op1, mode, PLUS, 1, speed); |
23cb6618 | 6601 | |
7cc2145f JG |
6602 | /* Look for ADD (extended register). */ |
6603 | if (aarch64_rtx_arith_op_extract_p (op0, mode)) | |
6604 | { | |
6605 | if (speed) | |
2533c820 | 6606 | *cost += extra_cost->alu.extend_arith; |
7cc2145f | 6607 | |
e47c4031 KT |
6608 | op0 = aarch64_strip_extend (op0); |
6609 | *cost += rtx_cost (op0, VOIDmode, | |
e548c9df | 6610 | (enum rtx_code) GET_CODE (op0), 0, speed); |
7cc2145f JG |
6611 | return true; |
6612 | } | |
6613 | ||
4745e701 JG |
6614 | /* Strip any extend, leave shifts behind as we will |
6615 | cost them through mult_cost. */ | |
6616 | new_op0 = aarch64_strip_extend (op0); | |
6617 | ||
6618 | if (GET_CODE (new_op0) == MULT | |
0a78ebe4 | 6619 | || aarch64_shift_p (GET_CODE (new_op0))) |
4745e701 JG |
6620 | { |
6621 | *cost += aarch64_rtx_mult_cost (new_op0, MULT, PLUS, | |
6622 | speed); | |
4745e701 JG |
6623 | return true; |
6624 | } | |
6625 | ||
e548c9df | 6626 | *cost += rtx_cost (new_op0, VOIDmode, PLUS, 0, speed); |
4745e701 JG |
6627 | |
6628 | if (speed) | |
6629 | { | |
b6875aac KV |
6630 | if (VECTOR_MODE_P (mode)) |
6631 | { | |
6632 | /* Vector ADD. */ | |
6633 | *cost += extra_cost->vect.alu; | |
6634 | } | |
6635 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
6636 | { | |
6637 | /* ADD. */ | |
6638 | *cost += extra_cost->alu.arith; | |
6639 | } | |
4745e701 | 6640 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
b6875aac KV |
6641 | { |
6642 | /* FADD. */ | |
6643 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
6644 | } | |
4745e701 JG |
6645 | } |
6646 | return true; | |
6647 | } | |
43e9d192 | 6648 | |
18b42b2a KT |
6649 | case BSWAP: |
6650 | *cost = COSTS_N_INSNS (1); | |
6651 | ||
6652 | if (speed) | |
b6875aac KV |
6653 | { |
6654 | if (VECTOR_MODE_P (mode)) | |
6655 | *cost += extra_cost->vect.alu; | |
6656 | else | |
6657 | *cost += extra_cost->alu.rev; | |
6658 | } | |
18b42b2a KT |
6659 | return false; |
6660 | ||
43e9d192 | 6661 | case IOR: |
f7d5cf8d KT |
6662 | if (aarch_rev16_p (x)) |
6663 | { | |
6664 | *cost = COSTS_N_INSNS (1); | |
6665 | ||
b6875aac KV |
6666 | if (speed) |
6667 | { | |
6668 | if (VECTOR_MODE_P (mode)) | |
6669 | *cost += extra_cost->vect.alu; | |
6670 | else | |
6671 | *cost += extra_cost->alu.rev; | |
6672 | } | |
6673 | return true; | |
f7d5cf8d | 6674 | } |
fb0cb7fa KT |
6675 | |
6676 | if (aarch64_extr_rtx_p (x, &op0, &op1)) | |
6677 | { | |
e548c9df AM |
6678 | *cost += rtx_cost (op0, mode, IOR, 0, speed); |
6679 | *cost += rtx_cost (op1, mode, IOR, 1, speed); | |
fb0cb7fa KT |
6680 | if (speed) |
6681 | *cost += extra_cost->alu.shift; | |
6682 | ||
6683 | return true; | |
6684 | } | |
f7d5cf8d | 6685 | /* Fall through. */ |
43e9d192 IB |
6686 | case XOR: |
6687 | case AND: | |
6688 | cost_logic: | |
6689 | op0 = XEXP (x, 0); | |
6690 | op1 = XEXP (x, 1); | |
6691 | ||
b6875aac KV |
6692 | if (VECTOR_MODE_P (mode)) |
6693 | { | |
6694 | if (speed) | |
6695 | *cost += extra_cost->vect.alu; | |
6696 | return true; | |
6697 | } | |
6698 | ||
268c3b47 JG |
6699 | if (code == AND |
6700 | && GET_CODE (op0) == MULT | |
6701 | && CONST_INT_P (XEXP (op0, 1)) | |
6702 | && CONST_INT_P (op1) | |
6703 | && aarch64_uxt_size (exact_log2 (INTVAL (XEXP (op0, 1))), | |
6704 | INTVAL (op1)) != 0) | |
6705 | { | |
6706 | /* This is a UBFM/SBFM. */ | |
e548c9df | 6707 | *cost += rtx_cost (XEXP (op0, 0), mode, ZERO_EXTRACT, 0, speed); |
268c3b47 JG |
6708 | if (speed) |
6709 | *cost += extra_cost->alu.bfx; | |
6710 | return true; | |
6711 | } | |
6712 | ||
e548c9df | 6713 | if (GET_MODE_CLASS (mode) == MODE_INT) |
43e9d192 | 6714 | { |
8c83f71d | 6715 | if (CONST_INT_P (op1)) |
43e9d192 | 6716 | { |
8c83f71d KT |
6717 | /* We have a mask + shift version of a UBFIZ |
6718 | i.e. the *andim_ashift<mode>_bfiz pattern. */ | |
6719 | if (GET_CODE (op0) == ASHIFT | |
6720 | && aarch64_mask_and_shift_for_ubfiz_p (mode, op1, | |
6721 | XEXP (op0, 1))) | |
6722 | { | |
6723 | *cost += rtx_cost (XEXP (op0, 0), mode, | |
6724 | (enum rtx_code) code, 0, speed); | |
6725 | if (speed) | |
6726 | *cost += extra_cost->alu.bfx; | |
268c3b47 | 6727 | |
8c83f71d KT |
6728 | return true; |
6729 | } | |
6730 | else if (aarch64_bitmask_imm (INTVAL (op1), mode)) | |
6731 | { | |
6732 | /* We possibly get the immediate for free, this is not | |
6733 | modelled. */ | |
6734 | *cost += rtx_cost (op0, mode, (enum rtx_code) code, 0, speed); | |
6735 | if (speed) | |
6736 | *cost += extra_cost->alu.logical; | |
268c3b47 | 6737 | |
8c83f71d KT |
6738 | return true; |
6739 | } | |
43e9d192 IB |
6740 | } |
6741 | else | |
6742 | { | |
268c3b47 JG |
6743 | rtx new_op0 = op0; |
6744 | ||
6745 | /* Handle ORN, EON, or BIC. */ | |
43e9d192 IB |
6746 | if (GET_CODE (op0) == NOT) |
6747 | op0 = XEXP (op0, 0); | |
268c3b47 JG |
6748 | |
6749 | new_op0 = aarch64_strip_shift (op0); | |
6750 | ||
6751 | /* If we had a shift on op0 then this is a logical-shift- | |
6752 | by-register/immediate operation. Otherwise, this is just | |
6753 | a logical operation. */ | |
6754 | if (speed) | |
6755 | { | |
6756 | if (new_op0 != op0) | |
6757 | { | |
6758 | /* Shift by immediate. */ | |
6759 | if (CONST_INT_P (XEXP (op0, 1))) | |
6760 | *cost += extra_cost->alu.log_shift; | |
6761 | else | |
6762 | *cost += extra_cost->alu.log_shift_reg; | |
6763 | } | |
6764 | else | |
6765 | *cost += extra_cost->alu.logical; | |
6766 | } | |
6767 | ||
6768 | /* In both cases we want to cost both operands. */ | |
e548c9df AM |
6769 | *cost += rtx_cost (new_op0, mode, (enum rtx_code) code, 0, speed); |
6770 | *cost += rtx_cost (op1, mode, (enum rtx_code) code, 1, speed); | |
268c3b47 JG |
6771 | |
6772 | return true; | |
43e9d192 | 6773 | } |
43e9d192 IB |
6774 | } |
6775 | return false; | |
6776 | ||
268c3b47 | 6777 | case NOT: |
6365da9e KT |
6778 | x = XEXP (x, 0); |
6779 | op0 = aarch64_strip_shift (x); | |
6780 | ||
b6875aac KV |
6781 | if (VECTOR_MODE_P (mode)) |
6782 | { | |
6783 | /* Vector NOT. */ | |
6784 | *cost += extra_cost->vect.alu; | |
6785 | return false; | |
6786 | } | |
6787 | ||
6365da9e KT |
6788 | /* MVN-shifted-reg. */ |
6789 | if (op0 != x) | |
6790 | { | |
e548c9df | 6791 | *cost += rtx_cost (op0, mode, (enum rtx_code) code, 0, speed); |
6365da9e KT |
6792 | |
6793 | if (speed) | |
6794 | *cost += extra_cost->alu.log_shift; | |
6795 | ||
6796 | return true; | |
6797 | } | |
6798 | /* EON can have two forms: (xor (not a) b) but also (not (xor a b)). | |
6799 | Handle the second form here taking care that 'a' in the above can | |
6800 | be a shift. */ | |
6801 | else if (GET_CODE (op0) == XOR) | |
6802 | { | |
6803 | rtx newop0 = XEXP (op0, 0); | |
6804 | rtx newop1 = XEXP (op0, 1); | |
6805 | rtx op0_stripped = aarch64_strip_shift (newop0); | |
6806 | ||
e548c9df AM |
6807 | *cost += rtx_cost (newop1, mode, (enum rtx_code) code, 1, speed); |
6808 | *cost += rtx_cost (op0_stripped, mode, XOR, 0, speed); | |
6365da9e KT |
6809 | |
6810 | if (speed) | |
6811 | { | |
6812 | if (op0_stripped != newop0) | |
6813 | *cost += extra_cost->alu.log_shift; | |
6814 | else | |
6815 | *cost += extra_cost->alu.logical; | |
6816 | } | |
6817 | ||
6818 | return true; | |
6819 | } | |
268c3b47 JG |
6820 | /* MVN. */ |
6821 | if (speed) | |
6822 | *cost += extra_cost->alu.logical; | |
6823 | ||
268c3b47 JG |
6824 | return false; |
6825 | ||
43e9d192 | 6826 | case ZERO_EXTEND: |
b1685e62 JG |
6827 | |
6828 | op0 = XEXP (x, 0); | |
6829 | /* If a value is written in SI mode, then zero extended to DI | |
6830 | mode, the operation will in general be free as a write to | |
6831 | a 'w' register implicitly zeroes the upper bits of an 'x' | |
6832 | register. However, if this is | |
6833 | ||
6834 | (set (reg) (zero_extend (reg))) | |
6835 | ||
6836 | we must cost the explicit register move. */ | |
6837 | if (mode == DImode | |
6838 | && GET_MODE (op0) == SImode | |
6839 | && outer == SET) | |
6840 | { | |
e548c9df | 6841 | int op_cost = rtx_cost (op0, VOIDmode, ZERO_EXTEND, 0, speed); |
b1685e62 JG |
6842 | |
6843 | if (!op_cost && speed) | |
6844 | /* MOV. */ | |
6845 | *cost += extra_cost->alu.extend; | |
6846 | else | |
6847 | /* Free, the cost is that of the SI mode operation. */ | |
6848 | *cost = op_cost; | |
6849 | ||
6850 | return true; | |
6851 | } | |
e548c9df | 6852 | else if (MEM_P (op0)) |
43e9d192 | 6853 | { |
b1685e62 | 6854 | /* All loads can zero extend to any size for free. */ |
e548c9df | 6855 | *cost = rtx_cost (op0, VOIDmode, ZERO_EXTEND, param, speed); |
43e9d192 IB |
6856 | return true; |
6857 | } | |
b1685e62 | 6858 | |
283b6c85 KT |
6859 | op0 = aarch64_extend_bitfield_pattern_p (x); |
6860 | if (op0) | |
6861 | { | |
6862 | *cost += rtx_cost (op0, mode, ZERO_EXTEND, 0, speed); | |
6863 | if (speed) | |
6864 | *cost += extra_cost->alu.bfx; | |
6865 | return true; | |
6866 | } | |
6867 | ||
b1685e62 | 6868 | if (speed) |
b6875aac KV |
6869 | { |
6870 | if (VECTOR_MODE_P (mode)) | |
6871 | { | |
6872 | /* UMOV. */ | |
6873 | *cost += extra_cost->vect.alu; | |
6874 | } | |
6875 | else | |
6876 | { | |
6877 | /* UXTB/UXTH. */ | |
6878 | *cost += extra_cost->alu.extend; | |
6879 | } | |
6880 | } | |
43e9d192 IB |
6881 | return false; |
6882 | ||
6883 | case SIGN_EXTEND: | |
b1685e62 | 6884 | if (MEM_P (XEXP (x, 0))) |
43e9d192 | 6885 | { |
b1685e62 JG |
6886 | /* LDRSH. */ |
6887 | if (speed) | |
6888 | { | |
6889 | rtx address = XEXP (XEXP (x, 0), 0); | |
6890 | *cost += extra_cost->ldst.load_sign_extend; | |
6891 | ||
6892 | *cost += | |
6893 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
6894 | 0, speed)); | |
6895 | } | |
43e9d192 IB |
6896 | return true; |
6897 | } | |
b1685e62 | 6898 | |
283b6c85 KT |
6899 | op0 = aarch64_extend_bitfield_pattern_p (x); |
6900 | if (op0) | |
6901 | { | |
6902 | *cost += rtx_cost (op0, mode, SIGN_EXTEND, 0, speed); | |
6903 | if (speed) | |
6904 | *cost += extra_cost->alu.bfx; | |
6905 | return true; | |
6906 | } | |
6907 | ||
b1685e62 | 6908 | if (speed) |
b6875aac KV |
6909 | { |
6910 | if (VECTOR_MODE_P (mode)) | |
6911 | *cost += extra_cost->vect.alu; | |
6912 | else | |
6913 | *cost += extra_cost->alu.extend; | |
6914 | } | |
43e9d192 IB |
6915 | return false; |
6916 | ||
ba0cfa17 JG |
6917 | case ASHIFT: |
6918 | op0 = XEXP (x, 0); | |
6919 | op1 = XEXP (x, 1); | |
6920 | ||
6921 | if (CONST_INT_P (op1)) | |
6922 | { | |
ba0cfa17 | 6923 | if (speed) |
b6875aac KV |
6924 | { |
6925 | if (VECTOR_MODE_P (mode)) | |
6926 | { | |
6927 | /* Vector shift (immediate). */ | |
6928 | *cost += extra_cost->vect.alu; | |
6929 | } | |
6930 | else | |
6931 | { | |
6932 | /* LSL (immediate), UBMF, UBFIZ and friends. These are all | |
6933 | aliases. */ | |
6934 | *cost += extra_cost->alu.shift; | |
6935 | } | |
6936 | } | |
ba0cfa17 JG |
6937 | |
6938 | /* We can incorporate zero/sign extend for free. */ | |
6939 | if (GET_CODE (op0) == ZERO_EXTEND | |
6940 | || GET_CODE (op0) == SIGN_EXTEND) | |
6941 | op0 = XEXP (op0, 0); | |
6942 | ||
e548c9df | 6943 | *cost += rtx_cost (op0, VOIDmode, ASHIFT, 0, speed); |
ba0cfa17 JG |
6944 | return true; |
6945 | } | |
6946 | else | |
6947 | { | |
ba0cfa17 | 6948 | if (speed) |
b6875aac KV |
6949 | { |
6950 | if (VECTOR_MODE_P (mode)) | |
6951 | { | |
6952 | /* Vector shift (register). */ | |
6953 | *cost += extra_cost->vect.alu; | |
6954 | } | |
6955 | else | |
6956 | { | |
6957 | /* LSLV. */ | |
6958 | *cost += extra_cost->alu.shift_reg; | |
6959 | } | |
6960 | } | |
ba0cfa17 JG |
6961 | return false; /* All arguments need to be in registers. */ |
6962 | } | |
6963 | ||
43e9d192 | 6964 | case ROTATE: |
43e9d192 IB |
6965 | case ROTATERT: |
6966 | case LSHIFTRT: | |
43e9d192 | 6967 | case ASHIFTRT: |
ba0cfa17 JG |
6968 | op0 = XEXP (x, 0); |
6969 | op1 = XEXP (x, 1); | |
43e9d192 | 6970 | |
ba0cfa17 JG |
6971 | if (CONST_INT_P (op1)) |
6972 | { | |
6973 | /* ASR (immediate) and friends. */ | |
6974 | if (speed) | |
b6875aac KV |
6975 | { |
6976 | if (VECTOR_MODE_P (mode)) | |
6977 | *cost += extra_cost->vect.alu; | |
6978 | else | |
6979 | *cost += extra_cost->alu.shift; | |
6980 | } | |
43e9d192 | 6981 | |
e548c9df | 6982 | *cost += rtx_cost (op0, mode, (enum rtx_code) code, 0, speed); |
ba0cfa17 JG |
6983 | return true; |
6984 | } | |
6985 | else | |
6986 | { | |
6987 | ||
6988 | /* ASR (register) and friends. */ | |
6989 | if (speed) | |
b6875aac KV |
6990 | { |
6991 | if (VECTOR_MODE_P (mode)) | |
6992 | *cost += extra_cost->vect.alu; | |
6993 | else | |
6994 | *cost += extra_cost->alu.shift_reg; | |
6995 | } | |
ba0cfa17 JG |
6996 | return false; /* All arguments need to be in registers. */ |
6997 | } | |
43e9d192 | 6998 | |
909734be JG |
6999 | case SYMBOL_REF: |
7000 | ||
1b1e81f8 JW |
7001 | if (aarch64_cmodel == AARCH64_CMODEL_LARGE |
7002 | || aarch64_cmodel == AARCH64_CMODEL_SMALL_SPIC) | |
909734be JG |
7003 | { |
7004 | /* LDR. */ | |
7005 | if (speed) | |
7006 | *cost += extra_cost->ldst.load; | |
7007 | } | |
7008 | else if (aarch64_cmodel == AARCH64_CMODEL_SMALL | |
7009 | || aarch64_cmodel == AARCH64_CMODEL_SMALL_PIC) | |
7010 | { | |
7011 | /* ADRP, followed by ADD. */ | |
7012 | *cost += COSTS_N_INSNS (1); | |
7013 | if (speed) | |
7014 | *cost += 2 * extra_cost->alu.arith; | |
7015 | } | |
7016 | else if (aarch64_cmodel == AARCH64_CMODEL_TINY | |
7017 | || aarch64_cmodel == AARCH64_CMODEL_TINY_PIC) | |
7018 | { | |
7019 | /* ADR. */ | |
7020 | if (speed) | |
7021 | *cost += extra_cost->alu.arith; | |
7022 | } | |
7023 | ||
7024 | if (flag_pic) | |
7025 | { | |
7026 | /* One extra load instruction, after accessing the GOT. */ | |
7027 | *cost += COSTS_N_INSNS (1); | |
7028 | if (speed) | |
7029 | *cost += extra_cost->ldst.load; | |
7030 | } | |
43e9d192 IB |
7031 | return true; |
7032 | ||
909734be | 7033 | case HIGH: |
43e9d192 | 7034 | case LO_SUM: |
909734be JG |
7035 | /* ADRP/ADD (immediate). */ |
7036 | if (speed) | |
7037 | *cost += extra_cost->alu.arith; | |
43e9d192 IB |
7038 | return true; |
7039 | ||
7040 | case ZERO_EXTRACT: | |
7041 | case SIGN_EXTRACT: | |
7cc2145f JG |
7042 | /* UBFX/SBFX. */ |
7043 | if (speed) | |
b6875aac KV |
7044 | { |
7045 | if (VECTOR_MODE_P (mode)) | |
7046 | *cost += extra_cost->vect.alu; | |
7047 | else | |
7048 | *cost += extra_cost->alu.bfx; | |
7049 | } | |
7cc2145f JG |
7050 | |
7051 | /* We can trust that the immediates used will be correct (there | |
7052 | are no by-register forms), so we need only cost op0. */ | |
e548c9df | 7053 | *cost += rtx_cost (XEXP (x, 0), VOIDmode, (enum rtx_code) code, 0, speed); |
43e9d192 IB |
7054 | return true; |
7055 | ||
7056 | case MULT: | |
4745e701 JG |
7057 | *cost += aarch64_rtx_mult_cost (x, MULT, 0, speed); |
7058 | /* aarch64_rtx_mult_cost always handles recursion to its | |
7059 | operands. */ | |
7060 | return true; | |
43e9d192 IB |
7061 | |
7062 | case MOD: | |
4f58fe36 KT |
7063 | /* We can expand signed mod by power of 2 using a NEGS, two parallel |
7064 | ANDs and a CSNEG. Assume here that CSNEG is the same as the cost of | |
7065 | an unconditional negate. This case should only ever be reached through | |
7066 | the set_smod_pow2_cheap check in expmed.c. */ | |
7067 | if (CONST_INT_P (XEXP (x, 1)) | |
7068 | && exact_log2 (INTVAL (XEXP (x, 1))) > 0 | |
7069 | && (mode == SImode || mode == DImode)) | |
7070 | { | |
7071 | /* We expand to 4 instructions. Reset the baseline. */ | |
7072 | *cost = COSTS_N_INSNS (4); | |
7073 | ||
7074 | if (speed) | |
7075 | *cost += 2 * extra_cost->alu.logical | |
7076 | + 2 * extra_cost->alu.arith; | |
7077 | ||
7078 | return true; | |
7079 | } | |
7080 | ||
7081 | /* Fall-through. */ | |
43e9d192 | 7082 | case UMOD: |
43e9d192 IB |
7083 | if (speed) |
7084 | { | |
b6875aac KV |
7085 | if (VECTOR_MODE_P (mode)) |
7086 | *cost += extra_cost->vect.alu; | |
e548c9df AM |
7087 | else if (GET_MODE_CLASS (mode) == MODE_INT) |
7088 | *cost += (extra_cost->mult[mode == DImode].add | |
7089 | + extra_cost->mult[mode == DImode].idiv); | |
7090 | else if (mode == DFmode) | |
73250c4c KT |
7091 | *cost += (extra_cost->fp[1].mult |
7092 | + extra_cost->fp[1].div); | |
e548c9df | 7093 | else if (mode == SFmode) |
73250c4c KT |
7094 | *cost += (extra_cost->fp[0].mult |
7095 | + extra_cost->fp[0].div); | |
43e9d192 IB |
7096 | } |
7097 | return false; /* All arguments need to be in registers. */ | |
7098 | ||
7099 | case DIV: | |
7100 | case UDIV: | |
4105fe38 | 7101 | case SQRT: |
43e9d192 IB |
7102 | if (speed) |
7103 | { | |
b6875aac KV |
7104 | if (VECTOR_MODE_P (mode)) |
7105 | *cost += extra_cost->vect.alu; | |
7106 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
4105fe38 JG |
7107 | /* There is no integer SQRT, so only DIV and UDIV can get |
7108 | here. */ | |
7109 | *cost += extra_cost->mult[mode == DImode].idiv; | |
7110 | else | |
7111 | *cost += extra_cost->fp[mode == DFmode].div; | |
43e9d192 IB |
7112 | } |
7113 | return false; /* All arguments need to be in registers. */ | |
7114 | ||
a8eecd00 | 7115 | case IF_THEN_ELSE: |
2d5ffe46 AP |
7116 | return aarch64_if_then_else_costs (XEXP (x, 0), XEXP (x, 1), |
7117 | XEXP (x, 2), cost, speed); | |
a8eecd00 JG |
7118 | |
7119 | case EQ: | |
7120 | case NE: | |
7121 | case GT: | |
7122 | case GTU: | |
7123 | case LT: | |
7124 | case LTU: | |
7125 | case GE: | |
7126 | case GEU: | |
7127 | case LE: | |
7128 | case LEU: | |
7129 | ||
7130 | return false; /* All arguments must be in registers. */ | |
7131 | ||
b292109f JG |
7132 | case FMA: |
7133 | op0 = XEXP (x, 0); | |
7134 | op1 = XEXP (x, 1); | |
7135 | op2 = XEXP (x, 2); | |
7136 | ||
7137 | if (speed) | |
b6875aac KV |
7138 | { |
7139 | if (VECTOR_MODE_P (mode)) | |
7140 | *cost += extra_cost->vect.alu; | |
7141 | else | |
7142 | *cost += extra_cost->fp[mode == DFmode].fma; | |
7143 | } | |
b292109f JG |
7144 | |
7145 | /* FMSUB, FNMADD, and FNMSUB are free. */ | |
7146 | if (GET_CODE (op0) == NEG) | |
7147 | op0 = XEXP (op0, 0); | |
7148 | ||
7149 | if (GET_CODE (op2) == NEG) | |
7150 | op2 = XEXP (op2, 0); | |
7151 | ||
7152 | /* aarch64_fnma4_elt_to_64v2df has the NEG as operand 1, | |
7153 | and the by-element operand as operand 0. */ | |
7154 | if (GET_CODE (op1) == NEG) | |
7155 | op1 = XEXP (op1, 0); | |
7156 | ||
7157 | /* Catch vector-by-element operations. The by-element operand can | |
7158 | either be (vec_duplicate (vec_select (x))) or just | |
7159 | (vec_select (x)), depending on whether we are multiplying by | |
7160 | a vector or a scalar. | |
7161 | ||
7162 | Canonicalization is not very good in these cases, FMA4 will put the | |
7163 | by-element operand as operand 0, FNMA4 will have it as operand 1. */ | |
7164 | if (GET_CODE (op0) == VEC_DUPLICATE) | |
7165 | op0 = XEXP (op0, 0); | |
7166 | else if (GET_CODE (op1) == VEC_DUPLICATE) | |
7167 | op1 = XEXP (op1, 0); | |
7168 | ||
7169 | if (GET_CODE (op0) == VEC_SELECT) | |
7170 | op0 = XEXP (op0, 0); | |
7171 | else if (GET_CODE (op1) == VEC_SELECT) | |
7172 | op1 = XEXP (op1, 0); | |
7173 | ||
7174 | /* If the remaining parameters are not registers, | |
7175 | get the cost to put them into registers. */ | |
e548c9df AM |
7176 | *cost += rtx_cost (op0, mode, FMA, 0, speed); |
7177 | *cost += rtx_cost (op1, mode, FMA, 1, speed); | |
7178 | *cost += rtx_cost (op2, mode, FMA, 2, speed); | |
b292109f JG |
7179 | return true; |
7180 | ||
5e2a765b KT |
7181 | case FLOAT: |
7182 | case UNSIGNED_FLOAT: | |
7183 | if (speed) | |
7184 | *cost += extra_cost->fp[mode == DFmode].fromint; | |
7185 | return false; | |
7186 | ||
b292109f JG |
7187 | case FLOAT_EXTEND: |
7188 | if (speed) | |
b6875aac KV |
7189 | { |
7190 | if (VECTOR_MODE_P (mode)) | |
7191 | { | |
7192 | /*Vector truncate. */ | |
7193 | *cost += extra_cost->vect.alu; | |
7194 | } | |
7195 | else | |
7196 | *cost += extra_cost->fp[mode == DFmode].widen; | |
7197 | } | |
b292109f JG |
7198 | return false; |
7199 | ||
7200 | case FLOAT_TRUNCATE: | |
7201 | if (speed) | |
b6875aac KV |
7202 | { |
7203 | if (VECTOR_MODE_P (mode)) | |
7204 | { | |
7205 | /*Vector conversion. */ | |
7206 | *cost += extra_cost->vect.alu; | |
7207 | } | |
7208 | else | |
7209 | *cost += extra_cost->fp[mode == DFmode].narrow; | |
7210 | } | |
b292109f JG |
7211 | return false; |
7212 | ||
61263118 KT |
7213 | case FIX: |
7214 | case UNSIGNED_FIX: | |
7215 | x = XEXP (x, 0); | |
7216 | /* Strip the rounding part. They will all be implemented | |
7217 | by the fcvt* family of instructions anyway. */ | |
7218 | if (GET_CODE (x) == UNSPEC) | |
7219 | { | |
7220 | unsigned int uns_code = XINT (x, 1); | |
7221 | ||
7222 | if (uns_code == UNSPEC_FRINTA | |
7223 | || uns_code == UNSPEC_FRINTM | |
7224 | || uns_code == UNSPEC_FRINTN | |
7225 | || uns_code == UNSPEC_FRINTP | |
7226 | || uns_code == UNSPEC_FRINTZ) | |
7227 | x = XVECEXP (x, 0, 0); | |
7228 | } | |
7229 | ||
7230 | if (speed) | |
b6875aac KV |
7231 | { |
7232 | if (VECTOR_MODE_P (mode)) | |
7233 | *cost += extra_cost->vect.alu; | |
7234 | else | |
7235 | *cost += extra_cost->fp[GET_MODE (x) == DFmode].toint; | |
7236 | } | |
39252973 KT |
7237 | |
7238 | /* We can combine fmul by a power of 2 followed by a fcvt into a single | |
7239 | fixed-point fcvt. */ | |
7240 | if (GET_CODE (x) == MULT | |
7241 | && ((VECTOR_MODE_P (mode) | |
7242 | && aarch64_vec_fpconst_pow_of_2 (XEXP (x, 1)) > 0) | |
7243 | || aarch64_fpconst_pow_of_2 (XEXP (x, 1)) > 0)) | |
7244 | { | |
7245 | *cost += rtx_cost (XEXP (x, 0), VOIDmode, (rtx_code) code, | |
7246 | 0, speed); | |
7247 | return true; | |
7248 | } | |
7249 | ||
e548c9df | 7250 | *cost += rtx_cost (x, VOIDmode, (enum rtx_code) code, 0, speed); |
61263118 KT |
7251 | return true; |
7252 | ||
b292109f | 7253 | case ABS: |
b6875aac KV |
7254 | if (VECTOR_MODE_P (mode)) |
7255 | { | |
7256 | /* ABS (vector). */ | |
7257 | if (speed) | |
7258 | *cost += extra_cost->vect.alu; | |
7259 | } | |
7260 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
b292109f | 7261 | { |
19261b99 KT |
7262 | op0 = XEXP (x, 0); |
7263 | ||
7264 | /* FABD, which is analogous to FADD. */ | |
7265 | if (GET_CODE (op0) == MINUS) | |
7266 | { | |
e548c9df AM |
7267 | *cost += rtx_cost (XEXP (op0, 0), mode, MINUS, 0, speed); |
7268 | *cost += rtx_cost (XEXP (op0, 1), mode, MINUS, 1, speed); | |
19261b99 KT |
7269 | if (speed) |
7270 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
7271 | ||
7272 | return true; | |
7273 | } | |
7274 | /* Simple FABS is analogous to FNEG. */ | |
b292109f JG |
7275 | if (speed) |
7276 | *cost += extra_cost->fp[mode == DFmode].neg; | |
7277 | } | |
7278 | else | |
7279 | { | |
7280 | /* Integer ABS will either be split to | |
7281 | two arithmetic instructions, or will be an ABS | |
7282 | (scalar), which we don't model. */ | |
7283 | *cost = COSTS_N_INSNS (2); | |
7284 | if (speed) | |
7285 | *cost += 2 * extra_cost->alu.arith; | |
7286 | } | |
7287 | return false; | |
7288 | ||
7289 | case SMAX: | |
7290 | case SMIN: | |
7291 | if (speed) | |
7292 | { | |
b6875aac KV |
7293 | if (VECTOR_MODE_P (mode)) |
7294 | *cost += extra_cost->vect.alu; | |
7295 | else | |
7296 | { | |
7297 | /* FMAXNM/FMINNM/FMAX/FMIN. | |
7298 | TODO: This may not be accurate for all implementations, but | |
7299 | we do not model this in the cost tables. */ | |
7300 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
7301 | } | |
b292109f JG |
7302 | } |
7303 | return false; | |
7304 | ||
61263118 KT |
7305 | case UNSPEC: |
7306 | /* The floating point round to integer frint* instructions. */ | |
7307 | if (aarch64_frint_unspec_p (XINT (x, 1))) | |
7308 | { | |
7309 | if (speed) | |
7310 | *cost += extra_cost->fp[mode == DFmode].roundint; | |
7311 | ||
7312 | return false; | |
7313 | } | |
781aeb73 KT |
7314 | |
7315 | if (XINT (x, 1) == UNSPEC_RBIT) | |
7316 | { | |
7317 | if (speed) | |
7318 | *cost += extra_cost->alu.rev; | |
7319 | ||
7320 | return false; | |
7321 | } | |
61263118 KT |
7322 | break; |
7323 | ||
fb620c4a JG |
7324 | case TRUNCATE: |
7325 | ||
7326 | /* Decompose <su>muldi3_highpart. */ | |
7327 | if (/* (truncate:DI */ | |
7328 | mode == DImode | |
7329 | /* (lshiftrt:TI */ | |
7330 | && GET_MODE (XEXP (x, 0)) == TImode | |
7331 | && GET_CODE (XEXP (x, 0)) == LSHIFTRT | |
7332 | /* (mult:TI */ | |
7333 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT | |
7334 | /* (ANY_EXTEND:TI (reg:DI)) | |
7335 | (ANY_EXTEND:TI (reg:DI))) */ | |
7336 | && ((GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == ZERO_EXTEND | |
7337 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == ZERO_EXTEND) | |
7338 | || (GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == SIGN_EXTEND | |
7339 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == SIGN_EXTEND)) | |
7340 | && GET_MODE (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 0)) == DImode | |
7341 | && GET_MODE (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 1), 0)) == DImode | |
7342 | /* (const_int 64) */ | |
7343 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
7344 | && UINTVAL (XEXP (XEXP (x, 0), 1)) == 64) | |
7345 | { | |
7346 | /* UMULH/SMULH. */ | |
7347 | if (speed) | |
7348 | *cost += extra_cost->mult[mode == DImode].extend; | |
e548c9df AM |
7349 | *cost += rtx_cost (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 0), |
7350 | mode, MULT, 0, speed); | |
7351 | *cost += rtx_cost (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 1), 0), | |
7352 | mode, MULT, 1, speed); | |
fb620c4a JG |
7353 | return true; |
7354 | } | |
7355 | ||
7356 | /* Fall through. */ | |
43e9d192 | 7357 | default: |
61263118 | 7358 | break; |
43e9d192 | 7359 | } |
61263118 KT |
7360 | |
7361 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
7362 | fprintf (dump_file, | |
7363 | "\nFailed to cost RTX. Assuming default cost.\n"); | |
7364 | ||
7365 | return true; | |
43e9d192 IB |
7366 | } |
7367 | ||
0ee859b5 JG |
7368 | /* Wrapper around aarch64_rtx_costs, dumps the partial, or total cost |
7369 | calculated for X. This cost is stored in *COST. Returns true | |
7370 | if the total cost of X was calculated. */ | |
7371 | static bool | |
e548c9df | 7372 | aarch64_rtx_costs_wrapper (rtx x, machine_mode mode, int outer, |
0ee859b5 JG |
7373 | int param, int *cost, bool speed) |
7374 | { | |
e548c9df | 7375 | bool result = aarch64_rtx_costs (x, mode, outer, param, cost, speed); |
0ee859b5 JG |
7376 | |
7377 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
7378 | { | |
7379 | print_rtl_single (dump_file, x); | |
7380 | fprintf (dump_file, "\n%s cost: %d (%s)\n", | |
7381 | speed ? "Hot" : "Cold", | |
7382 | *cost, result ? "final" : "partial"); | |
7383 | } | |
7384 | ||
7385 | return result; | |
7386 | } | |
7387 | ||
43e9d192 | 7388 | static int |
ef4bddc2 | 7389 | aarch64_register_move_cost (machine_mode mode, |
8a3a7e67 | 7390 | reg_class_t from_i, reg_class_t to_i) |
43e9d192 | 7391 | { |
8a3a7e67 RH |
7392 | enum reg_class from = (enum reg_class) from_i; |
7393 | enum reg_class to = (enum reg_class) to_i; | |
43e9d192 | 7394 | const struct cpu_regmove_cost *regmove_cost |
b175b679 | 7395 | = aarch64_tune_params.regmove_cost; |
43e9d192 | 7396 | |
3be07662 | 7397 | /* Caller save and pointer regs are equivalent to GENERAL_REGS. */ |
2876a13f | 7398 | if (to == CALLER_SAVE_REGS || to == POINTER_REGS) |
3be07662 WD |
7399 | to = GENERAL_REGS; |
7400 | ||
2876a13f | 7401 | if (from == CALLER_SAVE_REGS || from == POINTER_REGS) |
3be07662 WD |
7402 | from = GENERAL_REGS; |
7403 | ||
6ee70f81 AP |
7404 | /* Moving between GPR and stack cost is the same as GP2GP. */ |
7405 | if ((from == GENERAL_REGS && to == STACK_REG) | |
7406 | || (to == GENERAL_REGS && from == STACK_REG)) | |
7407 | return regmove_cost->GP2GP; | |
7408 | ||
7409 | /* To/From the stack register, we move via the gprs. */ | |
7410 | if (to == STACK_REG || from == STACK_REG) | |
7411 | return aarch64_register_move_cost (mode, from, GENERAL_REGS) | |
7412 | + aarch64_register_move_cost (mode, GENERAL_REGS, to); | |
7413 | ||
8919453c WD |
7414 | if (GET_MODE_SIZE (mode) == 16) |
7415 | { | |
7416 | /* 128-bit operations on general registers require 2 instructions. */ | |
7417 | if (from == GENERAL_REGS && to == GENERAL_REGS) | |
7418 | return regmove_cost->GP2GP * 2; | |
7419 | else if (from == GENERAL_REGS) | |
7420 | return regmove_cost->GP2FP * 2; | |
7421 | else if (to == GENERAL_REGS) | |
7422 | return regmove_cost->FP2GP * 2; | |
7423 | ||
7424 | /* When AdvSIMD instructions are disabled it is not possible to move | |
7425 | a 128-bit value directly between Q registers. This is handled in | |
7426 | secondary reload. A general register is used as a scratch to move | |
7427 | the upper DI value and the lower DI value is moved directly, | |
7428 | hence the cost is the sum of three moves. */ | |
7429 | if (! TARGET_SIMD) | |
7430 | return regmove_cost->GP2FP + regmove_cost->FP2GP + regmove_cost->FP2FP; | |
7431 | ||
7432 | return regmove_cost->FP2FP; | |
7433 | } | |
7434 | ||
43e9d192 IB |
7435 | if (from == GENERAL_REGS && to == GENERAL_REGS) |
7436 | return regmove_cost->GP2GP; | |
7437 | else if (from == GENERAL_REGS) | |
7438 | return regmove_cost->GP2FP; | |
7439 | else if (to == GENERAL_REGS) | |
7440 | return regmove_cost->FP2GP; | |
7441 | ||
43e9d192 IB |
7442 | return regmove_cost->FP2FP; |
7443 | } | |
7444 | ||
7445 | static int | |
ef4bddc2 | 7446 | aarch64_memory_move_cost (machine_mode mode ATTRIBUTE_UNUSED, |
43e9d192 IB |
7447 | reg_class_t rclass ATTRIBUTE_UNUSED, |
7448 | bool in ATTRIBUTE_UNUSED) | |
7449 | { | |
b175b679 | 7450 | return aarch64_tune_params.memmov_cost; |
43e9d192 IB |
7451 | } |
7452 | ||
0c30e0f3 EM |
7453 | /* Return true if it is safe and beneficial to use the approximate rsqrt optabs |
7454 | to optimize 1.0/sqrt. */ | |
ee62a5a6 RS |
7455 | |
7456 | static bool | |
9acc9cbe | 7457 | use_rsqrt_p (machine_mode mode) |
ee62a5a6 RS |
7458 | { |
7459 | return (!flag_trapping_math | |
7460 | && flag_unsafe_math_optimizations | |
9acc9cbe EM |
7461 | && ((aarch64_tune_params.approx_modes->recip_sqrt |
7462 | & AARCH64_APPROX_MODE (mode)) | |
1a33079e | 7463 | || flag_mrecip_low_precision_sqrt)); |
ee62a5a6 RS |
7464 | } |
7465 | ||
0c30e0f3 EM |
7466 | /* Function to decide when to use the approximate reciprocal square root |
7467 | builtin. */ | |
a6fc00da BH |
7468 | |
7469 | static tree | |
ee62a5a6 | 7470 | aarch64_builtin_reciprocal (tree fndecl) |
a6fc00da | 7471 | { |
9acc9cbe EM |
7472 | machine_mode mode = TYPE_MODE (TREE_TYPE (fndecl)); |
7473 | ||
7474 | if (!use_rsqrt_p (mode)) | |
a6fc00da | 7475 | return NULL_TREE; |
ee62a5a6 | 7476 | return aarch64_builtin_rsqrt (DECL_FUNCTION_CODE (fndecl)); |
a6fc00da BH |
7477 | } |
7478 | ||
7479 | typedef rtx (*rsqrte_type) (rtx, rtx); | |
7480 | ||
98daafa0 EM |
7481 | /* Select reciprocal square root initial estimate insn depending on machine |
7482 | mode. */ | |
a6fc00da | 7483 | |
98daafa0 | 7484 | static rsqrte_type |
a6fc00da BH |
7485 | get_rsqrte_type (machine_mode mode) |
7486 | { | |
7487 | switch (mode) | |
7488 | { | |
2a823433 JW |
7489 | case DFmode: return gen_aarch64_rsqrtedf; |
7490 | case SFmode: return gen_aarch64_rsqrtesf; | |
7491 | case V2DFmode: return gen_aarch64_rsqrtev2df; | |
7492 | case V2SFmode: return gen_aarch64_rsqrtev2sf; | |
7493 | case V4SFmode: return gen_aarch64_rsqrtev4sf; | |
a6fc00da BH |
7494 | default: gcc_unreachable (); |
7495 | } | |
7496 | } | |
7497 | ||
7498 | typedef rtx (*rsqrts_type) (rtx, rtx, rtx); | |
7499 | ||
98daafa0 | 7500 | /* Select reciprocal square root series step insn depending on machine mode. */ |
a6fc00da | 7501 | |
98daafa0 | 7502 | static rsqrts_type |
a6fc00da BH |
7503 | get_rsqrts_type (machine_mode mode) |
7504 | { | |
7505 | switch (mode) | |
7506 | { | |
00ea75d4 JW |
7507 | case DFmode: return gen_aarch64_rsqrtsdf; |
7508 | case SFmode: return gen_aarch64_rsqrtssf; | |
7509 | case V2DFmode: return gen_aarch64_rsqrtsv2df; | |
7510 | case V2SFmode: return gen_aarch64_rsqrtsv2sf; | |
7511 | case V4SFmode: return gen_aarch64_rsqrtsv4sf; | |
a6fc00da BH |
7512 | default: gcc_unreachable (); |
7513 | } | |
7514 | } | |
7515 | ||
98daafa0 EM |
7516 | /* Emit instruction sequence to compute either the approximate square root |
7517 | or its approximate reciprocal, depending on the flag RECP, and return | |
7518 | whether the sequence was emitted or not. */ | |
a6fc00da | 7519 | |
98daafa0 EM |
7520 | bool |
7521 | aarch64_emit_approx_sqrt (rtx dst, rtx src, bool recp) | |
a6fc00da | 7522 | { |
98daafa0 EM |
7523 | machine_mode mode = GET_MODE (dst); |
7524 | machine_mode mmsk = mode_for_vector | |
7525 | (int_mode_for_mode (GET_MODE_INNER (mode)), | |
7526 | GET_MODE_NUNITS (mode)); | |
7527 | bool use_approx_sqrt_p = (!recp | |
7528 | && (flag_mlow_precision_sqrt | |
7529 | || (aarch64_tune_params.approx_modes->sqrt | |
7530 | & AARCH64_APPROX_MODE (mode)))); | |
7531 | bool use_approx_rsqrt_p = (recp | |
7532 | && (flag_mrecip_low_precision_sqrt | |
7533 | || (aarch64_tune_params.approx_modes->recip_sqrt | |
7534 | & AARCH64_APPROX_MODE (mode)))); | |
7535 | ||
7536 | if (!flag_finite_math_only | |
7537 | || flag_trapping_math | |
7538 | || !flag_unsafe_math_optimizations | |
7539 | || !(use_approx_sqrt_p || use_approx_rsqrt_p) | |
7540 | || optimize_function_for_size_p (cfun)) | |
7541 | return false; | |
a6fc00da | 7542 | |
98daafa0 EM |
7543 | rtx xmsk = gen_reg_rtx (mmsk); |
7544 | if (!recp) | |
7545 | /* When calculating the approximate square root, compare the argument with | |
7546 | 0.0 and create a mask. */ | |
7547 | emit_insn (gen_rtx_SET (xmsk, gen_rtx_NEG (mmsk, gen_rtx_EQ (mmsk, src, | |
7548 | CONST0_RTX (mode))))); | |
a6fc00da | 7549 | |
98daafa0 EM |
7550 | /* Estimate the approximate reciprocal square root. */ |
7551 | rtx xdst = gen_reg_rtx (mode); | |
7552 | emit_insn ((*get_rsqrte_type (mode)) (xdst, src)); | |
a6fc00da | 7553 | |
98daafa0 EM |
7554 | /* Iterate over the series twice for SF and thrice for DF. */ |
7555 | int iterations = (GET_MODE_INNER (mode) == DFmode) ? 3 : 2; | |
a6fc00da | 7556 | |
98daafa0 EM |
7557 | /* Optionally iterate over the series once less for faster performance |
7558 | while sacrificing the accuracy. */ | |
7559 | if ((recp && flag_mrecip_low_precision_sqrt) | |
7560 | || (!recp && flag_mlow_precision_sqrt)) | |
a6fc00da BH |
7561 | iterations--; |
7562 | ||
98daafa0 EM |
7563 | /* Iterate over the series to calculate the approximate reciprocal square |
7564 | root. */ | |
7565 | rtx x1 = gen_reg_rtx (mode); | |
7566 | while (iterations--) | |
a6fc00da | 7567 | { |
a6fc00da | 7568 | rtx x2 = gen_reg_rtx (mode); |
98daafa0 EM |
7569 | emit_set_insn (x2, gen_rtx_MULT (mode, xdst, xdst)); |
7570 | ||
7571 | emit_insn ((*get_rsqrts_type (mode)) (x1, src, x2)); | |
a6fc00da | 7572 | |
98daafa0 EM |
7573 | if (iterations > 0) |
7574 | emit_set_insn (xdst, gen_rtx_MULT (mode, xdst, x1)); | |
7575 | } | |
7576 | ||
7577 | if (!recp) | |
7578 | { | |
7579 | /* Qualify the approximate reciprocal square root when the argument is | |
7580 | 0.0 by squashing the intermediary result to 0.0. */ | |
7581 | rtx xtmp = gen_reg_rtx (mmsk); | |
7582 | emit_set_insn (xtmp, gen_rtx_AND (mmsk, gen_rtx_NOT (mmsk, xmsk), | |
7583 | gen_rtx_SUBREG (mmsk, xdst, 0))); | |
7584 | emit_move_insn (xdst, gen_rtx_SUBREG (mode, xtmp, 0)); | |
a6fc00da | 7585 | |
98daafa0 EM |
7586 | /* Calculate the approximate square root. */ |
7587 | emit_set_insn (xdst, gen_rtx_MULT (mode, xdst, src)); | |
a6fc00da BH |
7588 | } |
7589 | ||
98daafa0 EM |
7590 | /* Finalize the approximation. */ |
7591 | emit_set_insn (dst, gen_rtx_MULT (mode, xdst, x1)); | |
7592 | ||
7593 | return true; | |
a6fc00da BH |
7594 | } |
7595 | ||
79a2bc2d EM |
7596 | typedef rtx (*recpe_type) (rtx, rtx); |
7597 | ||
7598 | /* Select reciprocal initial estimate insn depending on machine mode. */ | |
7599 | ||
7600 | static recpe_type | |
7601 | get_recpe_type (machine_mode mode) | |
7602 | { | |
7603 | switch (mode) | |
7604 | { | |
7605 | case SFmode: return (gen_aarch64_frecpesf); | |
7606 | case V2SFmode: return (gen_aarch64_frecpev2sf); | |
7607 | case V4SFmode: return (gen_aarch64_frecpev4sf); | |
7608 | case DFmode: return (gen_aarch64_frecpedf); | |
7609 | case V2DFmode: return (gen_aarch64_frecpev2df); | |
7610 | default: gcc_unreachable (); | |
7611 | } | |
7612 | } | |
7613 | ||
7614 | typedef rtx (*recps_type) (rtx, rtx, rtx); | |
7615 | ||
7616 | /* Select reciprocal series step insn depending on machine mode. */ | |
7617 | ||
7618 | static recps_type | |
7619 | get_recps_type (machine_mode mode) | |
7620 | { | |
7621 | switch (mode) | |
7622 | { | |
7623 | case SFmode: return (gen_aarch64_frecpssf); | |
7624 | case V2SFmode: return (gen_aarch64_frecpsv2sf); | |
7625 | case V4SFmode: return (gen_aarch64_frecpsv4sf); | |
7626 | case DFmode: return (gen_aarch64_frecpsdf); | |
7627 | case V2DFmode: return (gen_aarch64_frecpsv2df); | |
7628 | default: gcc_unreachable (); | |
7629 | } | |
7630 | } | |
7631 | ||
7632 | /* Emit the instruction sequence to compute the approximation for the division | |
7633 | of NUM by DEN in QUO and return whether the sequence was emitted or not. */ | |
7634 | ||
7635 | bool | |
7636 | aarch64_emit_approx_div (rtx quo, rtx num, rtx den) | |
7637 | { | |
7638 | machine_mode mode = GET_MODE (quo); | |
7639 | bool use_approx_division_p = (flag_mlow_precision_div | |
7640 | || (aarch64_tune_params.approx_modes->division | |
7641 | & AARCH64_APPROX_MODE (mode))); | |
7642 | ||
7643 | if (!flag_finite_math_only | |
7644 | || flag_trapping_math | |
7645 | || !flag_unsafe_math_optimizations | |
7646 | || optimize_function_for_size_p (cfun) | |
7647 | || !use_approx_division_p) | |
7648 | return false; | |
7649 | ||
7650 | /* Estimate the approximate reciprocal. */ | |
7651 | rtx xrcp = gen_reg_rtx (mode); | |
7652 | emit_insn ((*get_recpe_type (mode)) (xrcp, den)); | |
7653 | ||
7654 | /* Iterate over the series twice for SF and thrice for DF. */ | |
7655 | int iterations = (GET_MODE_INNER (mode) == DFmode) ? 3 : 2; | |
7656 | ||
7657 | /* Optionally iterate over the series once less for faster performance, | |
7658 | while sacrificing the accuracy. */ | |
7659 | if (flag_mlow_precision_div) | |
7660 | iterations--; | |
7661 | ||
7662 | /* Iterate over the series to calculate the approximate reciprocal. */ | |
7663 | rtx xtmp = gen_reg_rtx (mode); | |
7664 | while (iterations--) | |
7665 | { | |
7666 | emit_insn ((*get_recps_type (mode)) (xtmp, xrcp, den)); | |
7667 | ||
7668 | if (iterations > 0) | |
7669 | emit_set_insn (xrcp, gen_rtx_MULT (mode, xrcp, xtmp)); | |
7670 | } | |
7671 | ||
7672 | if (num != CONST1_RTX (mode)) | |
7673 | { | |
7674 | /* As the approximate reciprocal of DEN is already calculated, only | |
7675 | calculate the approximate division when NUM is not 1.0. */ | |
7676 | rtx xnum = force_reg (mode, num); | |
7677 | emit_set_insn (xrcp, gen_rtx_MULT (mode, xrcp, xnum)); | |
7678 | } | |
7679 | ||
7680 | /* Finalize the approximation. */ | |
7681 | emit_set_insn (quo, gen_rtx_MULT (mode, xrcp, xtmp)); | |
7682 | return true; | |
7683 | } | |
7684 | ||
d126a4ae AP |
7685 | /* Return the number of instructions that can be issued per cycle. */ |
7686 | static int | |
7687 | aarch64_sched_issue_rate (void) | |
7688 | { | |
b175b679 | 7689 | return aarch64_tune_params.issue_rate; |
d126a4ae AP |
7690 | } |
7691 | ||
d03f7e44 MK |
7692 | static int |
7693 | aarch64_sched_first_cycle_multipass_dfa_lookahead (void) | |
7694 | { | |
7695 | int issue_rate = aarch64_sched_issue_rate (); | |
7696 | ||
7697 | return issue_rate > 1 && !sched_fusion ? issue_rate : 0; | |
7698 | } | |
7699 | ||
2d6bc7fa KT |
7700 | |
7701 | /* Implement TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD as | |
7702 | autopref_multipass_dfa_lookahead_guard from haifa-sched.c. It only | |
7703 | has an effect if PARAM_SCHED_AUTOPREF_QUEUE_DEPTH > 0. */ | |
7704 | ||
7705 | static int | |
7706 | aarch64_first_cycle_multipass_dfa_lookahead_guard (rtx_insn *insn, | |
7707 | int ready_index) | |
7708 | { | |
7709 | return autopref_multipass_dfa_lookahead_guard (insn, ready_index); | |
7710 | } | |
7711 | ||
7712 | ||
8990e73a TB |
7713 | /* Vectorizer cost model target hooks. */ |
7714 | ||
7715 | /* Implement targetm.vectorize.builtin_vectorization_cost. */ | |
7716 | static int | |
7717 | aarch64_builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost, | |
7718 | tree vectype, | |
7719 | int misalign ATTRIBUTE_UNUSED) | |
7720 | { | |
7721 | unsigned elements; | |
7722 | ||
7723 | switch (type_of_cost) | |
7724 | { | |
7725 | case scalar_stmt: | |
b175b679 | 7726 | return aarch64_tune_params.vec_costs->scalar_stmt_cost; |
8990e73a TB |
7727 | |
7728 | case scalar_load: | |
b175b679 | 7729 | return aarch64_tune_params.vec_costs->scalar_load_cost; |
8990e73a TB |
7730 | |
7731 | case scalar_store: | |
b175b679 | 7732 | return aarch64_tune_params.vec_costs->scalar_store_cost; |
8990e73a TB |
7733 | |
7734 | case vector_stmt: | |
b175b679 | 7735 | return aarch64_tune_params.vec_costs->vec_stmt_cost; |
8990e73a TB |
7736 | |
7737 | case vector_load: | |
b175b679 | 7738 | return aarch64_tune_params.vec_costs->vec_align_load_cost; |
8990e73a TB |
7739 | |
7740 | case vector_store: | |
b175b679 | 7741 | return aarch64_tune_params.vec_costs->vec_store_cost; |
8990e73a TB |
7742 | |
7743 | case vec_to_scalar: | |
b175b679 | 7744 | return aarch64_tune_params.vec_costs->vec_to_scalar_cost; |
8990e73a TB |
7745 | |
7746 | case scalar_to_vec: | |
b175b679 | 7747 | return aarch64_tune_params.vec_costs->scalar_to_vec_cost; |
8990e73a TB |
7748 | |
7749 | case unaligned_load: | |
b175b679 | 7750 | return aarch64_tune_params.vec_costs->vec_unalign_load_cost; |
8990e73a TB |
7751 | |
7752 | case unaligned_store: | |
b175b679 | 7753 | return aarch64_tune_params.vec_costs->vec_unalign_store_cost; |
8990e73a TB |
7754 | |
7755 | case cond_branch_taken: | |
b175b679 | 7756 | return aarch64_tune_params.vec_costs->cond_taken_branch_cost; |
8990e73a TB |
7757 | |
7758 | case cond_branch_not_taken: | |
b175b679 | 7759 | return aarch64_tune_params.vec_costs->cond_not_taken_branch_cost; |
8990e73a TB |
7760 | |
7761 | case vec_perm: | |
c428f91c WD |
7762 | return aarch64_tune_params.vec_costs->vec_permute_cost; |
7763 | ||
8990e73a | 7764 | case vec_promote_demote: |
b175b679 | 7765 | return aarch64_tune_params.vec_costs->vec_stmt_cost; |
8990e73a TB |
7766 | |
7767 | case vec_construct: | |
7768 | elements = TYPE_VECTOR_SUBPARTS (vectype); | |
7769 | return elements / 2 + 1; | |
7770 | ||
7771 | default: | |
7772 | gcc_unreachable (); | |
7773 | } | |
7774 | } | |
7775 | ||
7776 | /* Implement targetm.vectorize.add_stmt_cost. */ | |
7777 | static unsigned | |
7778 | aarch64_add_stmt_cost (void *data, int count, enum vect_cost_for_stmt kind, | |
7779 | struct _stmt_vec_info *stmt_info, int misalign, | |
7780 | enum vect_cost_model_location where) | |
7781 | { | |
7782 | unsigned *cost = (unsigned *) data; | |
7783 | unsigned retval = 0; | |
7784 | ||
7785 | if (flag_vect_cost_model) | |
7786 | { | |
7787 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; | |
7788 | int stmt_cost = | |
7789 | aarch64_builtin_vectorization_cost (kind, vectype, misalign); | |
7790 | ||
7791 | /* Statements in an inner loop relative to the loop being | |
7792 | vectorized are weighted more heavily. The value here is | |
058e4c71 | 7793 | arbitrary and could potentially be improved with analysis. */ |
8990e73a | 7794 | if (where == vect_body && stmt_info && stmt_in_inner_loop_p (stmt_info)) |
058e4c71 | 7795 | count *= 50; /* FIXME */ |
8990e73a TB |
7796 | |
7797 | retval = (unsigned) (count * stmt_cost); | |
7798 | cost[where] += retval; | |
7799 | } | |
7800 | ||
7801 | return retval; | |
7802 | } | |
7803 | ||
0cfff2a1 | 7804 | static void initialize_aarch64_code_model (struct gcc_options *); |
43e9d192 | 7805 | |
0cfff2a1 KT |
7806 | /* Parse the TO_PARSE string and put the architecture struct that it |
7807 | selects into RES and the architectural features into ISA_FLAGS. | |
7808 | Return an aarch64_parse_opt_result describing the parse result. | |
7809 | If there is an error parsing, RES and ISA_FLAGS are left unchanged. */ | |
43e9d192 | 7810 | |
0cfff2a1 KT |
7811 | static enum aarch64_parse_opt_result |
7812 | aarch64_parse_arch (const char *to_parse, const struct processor **res, | |
7813 | unsigned long *isa_flags) | |
43e9d192 IB |
7814 | { |
7815 | char *ext; | |
7816 | const struct processor *arch; | |
0cfff2a1 | 7817 | char *str = (char *) alloca (strlen (to_parse) + 1); |
43e9d192 IB |
7818 | size_t len; |
7819 | ||
0cfff2a1 | 7820 | strcpy (str, to_parse); |
43e9d192 IB |
7821 | |
7822 | ext = strchr (str, '+'); | |
7823 | ||
7824 | if (ext != NULL) | |
7825 | len = ext - str; | |
7826 | else | |
7827 | len = strlen (str); | |
7828 | ||
7829 | if (len == 0) | |
0cfff2a1 KT |
7830 | return AARCH64_PARSE_MISSING_ARG; |
7831 | ||
43e9d192 | 7832 | |
0cfff2a1 | 7833 | /* Loop through the list of supported ARCHes to find a match. */ |
43e9d192 IB |
7834 | for (arch = all_architectures; arch->name != NULL; arch++) |
7835 | { | |
7836 | if (strlen (arch->name) == len && strncmp (arch->name, str, len) == 0) | |
7837 | { | |
0cfff2a1 | 7838 | unsigned long isa_temp = arch->flags; |
43e9d192 IB |
7839 | |
7840 | if (ext != NULL) | |
7841 | { | |
0cfff2a1 KT |
7842 | /* TO_PARSE string contains at least one extension. */ |
7843 | enum aarch64_parse_opt_result ext_res | |
7844 | = aarch64_parse_extension (ext, &isa_temp); | |
43e9d192 | 7845 | |
0cfff2a1 KT |
7846 | if (ext_res != AARCH64_PARSE_OK) |
7847 | return ext_res; | |
ffee7aa9 | 7848 | } |
0cfff2a1 KT |
7849 | /* Extension parsing was successful. Confirm the result |
7850 | arch and ISA flags. */ | |
7851 | *res = arch; | |
7852 | *isa_flags = isa_temp; | |
7853 | return AARCH64_PARSE_OK; | |
43e9d192 IB |
7854 | } |
7855 | } | |
7856 | ||
7857 | /* ARCH name not found in list. */ | |
0cfff2a1 | 7858 | return AARCH64_PARSE_INVALID_ARG; |
43e9d192 IB |
7859 | } |
7860 | ||
0cfff2a1 KT |
7861 | /* Parse the TO_PARSE string and put the result tuning in RES and the |
7862 | architecture flags in ISA_FLAGS. Return an aarch64_parse_opt_result | |
7863 | describing the parse result. If there is an error parsing, RES and | |
7864 | ISA_FLAGS are left unchanged. */ | |
43e9d192 | 7865 | |
0cfff2a1 KT |
7866 | static enum aarch64_parse_opt_result |
7867 | aarch64_parse_cpu (const char *to_parse, const struct processor **res, | |
7868 | unsigned long *isa_flags) | |
43e9d192 IB |
7869 | { |
7870 | char *ext; | |
7871 | const struct processor *cpu; | |
0cfff2a1 | 7872 | char *str = (char *) alloca (strlen (to_parse) + 1); |
43e9d192 IB |
7873 | size_t len; |
7874 | ||
0cfff2a1 | 7875 | strcpy (str, to_parse); |
43e9d192 IB |
7876 | |
7877 | ext = strchr (str, '+'); | |
7878 | ||
7879 | if (ext != NULL) | |
7880 | len = ext - str; | |
7881 | else | |
7882 | len = strlen (str); | |
7883 | ||
7884 | if (len == 0) | |
0cfff2a1 KT |
7885 | return AARCH64_PARSE_MISSING_ARG; |
7886 | ||
43e9d192 IB |
7887 | |
7888 | /* Loop through the list of supported CPUs to find a match. */ | |
7889 | for (cpu = all_cores; cpu->name != NULL; cpu++) | |
7890 | { | |
7891 | if (strlen (cpu->name) == len && strncmp (cpu->name, str, len) == 0) | |
7892 | { | |
0cfff2a1 KT |
7893 | unsigned long isa_temp = cpu->flags; |
7894 | ||
43e9d192 IB |
7895 | |
7896 | if (ext != NULL) | |
7897 | { | |
0cfff2a1 KT |
7898 | /* TO_PARSE string contains at least one extension. */ |
7899 | enum aarch64_parse_opt_result ext_res | |
7900 | = aarch64_parse_extension (ext, &isa_temp); | |
43e9d192 | 7901 | |
0cfff2a1 KT |
7902 | if (ext_res != AARCH64_PARSE_OK) |
7903 | return ext_res; | |
7904 | } | |
7905 | /* Extension parsing was successfull. Confirm the result | |
7906 | cpu and ISA flags. */ | |
7907 | *res = cpu; | |
7908 | *isa_flags = isa_temp; | |
7909 | return AARCH64_PARSE_OK; | |
43e9d192 IB |
7910 | } |
7911 | } | |
7912 | ||
7913 | /* CPU name not found in list. */ | |
0cfff2a1 | 7914 | return AARCH64_PARSE_INVALID_ARG; |
43e9d192 IB |
7915 | } |
7916 | ||
0cfff2a1 KT |
7917 | /* Parse the TO_PARSE string and put the cpu it selects into RES. |
7918 | Return an aarch64_parse_opt_result describing the parse result. | |
7919 | If the parsing fails the RES does not change. */ | |
43e9d192 | 7920 | |
0cfff2a1 KT |
7921 | static enum aarch64_parse_opt_result |
7922 | aarch64_parse_tune (const char *to_parse, const struct processor **res) | |
43e9d192 IB |
7923 | { |
7924 | const struct processor *cpu; | |
0cfff2a1 KT |
7925 | char *str = (char *) alloca (strlen (to_parse) + 1); |
7926 | ||
7927 | strcpy (str, to_parse); | |
43e9d192 IB |
7928 | |
7929 | /* Loop through the list of supported CPUs to find a match. */ | |
7930 | for (cpu = all_cores; cpu->name != NULL; cpu++) | |
7931 | { | |
7932 | if (strcmp (cpu->name, str) == 0) | |
7933 | { | |
0cfff2a1 KT |
7934 | *res = cpu; |
7935 | return AARCH64_PARSE_OK; | |
43e9d192 IB |
7936 | } |
7937 | } | |
7938 | ||
7939 | /* CPU name not found in list. */ | |
0cfff2a1 | 7940 | return AARCH64_PARSE_INVALID_ARG; |
43e9d192 IB |
7941 | } |
7942 | ||
8dec06f2 JG |
7943 | /* Parse TOKEN, which has length LENGTH to see if it is an option |
7944 | described in FLAG. If it is, return the index bit for that fusion type. | |
7945 | If not, error (printing OPTION_NAME) and return zero. */ | |
7946 | ||
7947 | static unsigned int | |
7948 | aarch64_parse_one_option_token (const char *token, | |
7949 | size_t length, | |
7950 | const struct aarch64_flag_desc *flag, | |
7951 | const char *option_name) | |
7952 | { | |
7953 | for (; flag->name != NULL; flag++) | |
7954 | { | |
7955 | if (length == strlen (flag->name) | |
7956 | && !strncmp (flag->name, token, length)) | |
7957 | return flag->flag; | |
7958 | } | |
7959 | ||
7960 | error ("unknown flag passed in -moverride=%s (%s)", option_name, token); | |
7961 | return 0; | |
7962 | } | |
7963 | ||
7964 | /* Parse OPTION which is a comma-separated list of flags to enable. | |
7965 | FLAGS gives the list of flags we understand, INITIAL_STATE gives any | |
7966 | default state we inherit from the CPU tuning structures. OPTION_NAME | |
7967 | gives the top-level option we are parsing in the -moverride string, | |
7968 | for use in error messages. */ | |
7969 | ||
7970 | static unsigned int | |
7971 | aarch64_parse_boolean_options (const char *option, | |
7972 | const struct aarch64_flag_desc *flags, | |
7973 | unsigned int initial_state, | |
7974 | const char *option_name) | |
7975 | { | |
7976 | const char separator = '.'; | |
7977 | const char* specs = option; | |
7978 | const char* ntoken = option; | |
7979 | unsigned int found_flags = initial_state; | |
7980 | ||
7981 | while ((ntoken = strchr (specs, separator))) | |
7982 | { | |
7983 | size_t token_length = ntoken - specs; | |
7984 | unsigned token_ops = aarch64_parse_one_option_token (specs, | |
7985 | token_length, | |
7986 | flags, | |
7987 | option_name); | |
7988 | /* If we find "none" (or, for simplicity's sake, an error) anywhere | |
7989 | in the token stream, reset the supported operations. So: | |
7990 | ||
7991 | adrp+add.cmp+branch.none.adrp+add | |
7992 | ||
7993 | would have the result of turning on only adrp+add fusion. */ | |
7994 | if (!token_ops) | |
7995 | found_flags = 0; | |
7996 | ||
7997 | found_flags |= token_ops; | |
7998 | specs = ++ntoken; | |
7999 | } | |
8000 | ||
8001 | /* We ended with a comma, print something. */ | |
8002 | if (!(*specs)) | |
8003 | { | |
8004 | error ("%s string ill-formed\n", option_name); | |
8005 | return 0; | |
8006 | } | |
8007 | ||
8008 | /* We still have one more token to parse. */ | |
8009 | size_t token_length = strlen (specs); | |
8010 | unsigned token_ops = aarch64_parse_one_option_token (specs, | |
8011 | token_length, | |
8012 | flags, | |
8013 | option_name); | |
8014 | if (!token_ops) | |
8015 | found_flags = 0; | |
8016 | ||
8017 | found_flags |= token_ops; | |
8018 | return found_flags; | |
8019 | } | |
8020 | ||
8021 | /* Support for overriding instruction fusion. */ | |
8022 | ||
8023 | static void | |
8024 | aarch64_parse_fuse_string (const char *fuse_string, | |
8025 | struct tune_params *tune) | |
8026 | { | |
8027 | tune->fusible_ops = aarch64_parse_boolean_options (fuse_string, | |
8028 | aarch64_fusible_pairs, | |
8029 | tune->fusible_ops, | |
8030 | "fuse="); | |
8031 | } | |
8032 | ||
8033 | /* Support for overriding other tuning flags. */ | |
8034 | ||
8035 | static void | |
8036 | aarch64_parse_tune_string (const char *tune_string, | |
8037 | struct tune_params *tune) | |
8038 | { | |
8039 | tune->extra_tuning_flags | |
8040 | = aarch64_parse_boolean_options (tune_string, | |
8041 | aarch64_tuning_flags, | |
8042 | tune->extra_tuning_flags, | |
8043 | "tune="); | |
8044 | } | |
8045 | ||
8046 | /* Parse TOKEN, which has length LENGTH to see if it is a tuning option | |
8047 | we understand. If it is, extract the option string and handoff to | |
8048 | the appropriate function. */ | |
8049 | ||
8050 | void | |
8051 | aarch64_parse_one_override_token (const char* token, | |
8052 | size_t length, | |
8053 | struct tune_params *tune) | |
8054 | { | |
8055 | const struct aarch64_tuning_override_function *fn | |
8056 | = aarch64_tuning_override_functions; | |
8057 | ||
8058 | const char *option_part = strchr (token, '='); | |
8059 | if (!option_part) | |
8060 | { | |
8061 | error ("tuning string missing in option (%s)", token); | |
8062 | return; | |
8063 | } | |
8064 | ||
8065 | /* Get the length of the option name. */ | |
8066 | length = option_part - token; | |
8067 | /* Skip the '=' to get to the option string. */ | |
8068 | option_part++; | |
8069 | ||
8070 | for (; fn->name != NULL; fn++) | |
8071 | { | |
8072 | if (!strncmp (fn->name, token, length)) | |
8073 | { | |
8074 | fn->parse_override (option_part, tune); | |
8075 | return; | |
8076 | } | |
8077 | } | |
8078 | ||
8079 | error ("unknown tuning option (%s)",token); | |
8080 | return; | |
8081 | } | |
8082 | ||
5eee3c34 JW |
8083 | /* A checking mechanism for the implementation of the tls size. */ |
8084 | ||
8085 | static void | |
8086 | initialize_aarch64_tls_size (struct gcc_options *opts) | |
8087 | { | |
8088 | if (aarch64_tls_size == 0) | |
8089 | aarch64_tls_size = 24; | |
8090 | ||
8091 | switch (opts->x_aarch64_cmodel_var) | |
8092 | { | |
8093 | case AARCH64_CMODEL_TINY: | |
8094 | /* Both the default and maximum TLS size allowed under tiny is 1M which | |
8095 | needs two instructions to address, so we clamp the size to 24. */ | |
8096 | if (aarch64_tls_size > 24) | |
8097 | aarch64_tls_size = 24; | |
8098 | break; | |
8099 | case AARCH64_CMODEL_SMALL: | |
8100 | /* The maximum TLS size allowed under small is 4G. */ | |
8101 | if (aarch64_tls_size > 32) | |
8102 | aarch64_tls_size = 32; | |
8103 | break; | |
8104 | case AARCH64_CMODEL_LARGE: | |
8105 | /* The maximum TLS size allowed under large is 16E. | |
8106 | FIXME: 16E should be 64bit, we only support 48bit offset now. */ | |
8107 | if (aarch64_tls_size > 48) | |
8108 | aarch64_tls_size = 48; | |
8109 | break; | |
8110 | default: | |
8111 | gcc_unreachable (); | |
8112 | } | |
8113 | ||
8114 | return; | |
8115 | } | |
8116 | ||
8dec06f2 JG |
8117 | /* Parse STRING looking for options in the format: |
8118 | string :: option:string | |
8119 | option :: name=substring | |
8120 | name :: {a-z} | |
8121 | substring :: defined by option. */ | |
8122 | ||
8123 | static void | |
8124 | aarch64_parse_override_string (const char* input_string, | |
8125 | struct tune_params* tune) | |
8126 | { | |
8127 | const char separator = ':'; | |
8128 | size_t string_length = strlen (input_string) + 1; | |
8129 | char *string_root = (char *) xmalloc (sizeof (*string_root) * string_length); | |
8130 | char *string = string_root; | |
8131 | strncpy (string, input_string, string_length); | |
8132 | string[string_length - 1] = '\0'; | |
8133 | ||
8134 | char* ntoken = string; | |
8135 | ||
8136 | while ((ntoken = strchr (string, separator))) | |
8137 | { | |
8138 | size_t token_length = ntoken - string; | |
8139 | /* Make this substring look like a string. */ | |
8140 | *ntoken = '\0'; | |
8141 | aarch64_parse_one_override_token (string, token_length, tune); | |
8142 | string = ++ntoken; | |
8143 | } | |
8144 | ||
8145 | /* One last option to parse. */ | |
8146 | aarch64_parse_one_override_token (string, strlen (string), tune); | |
8147 | free (string_root); | |
8148 | } | |
43e9d192 | 8149 | |
43e9d192 IB |
8150 | |
8151 | static void | |
0cfff2a1 | 8152 | aarch64_override_options_after_change_1 (struct gcc_options *opts) |
43e9d192 | 8153 | { |
a3dc8760 NC |
8154 | /* The logic here is that if we are disabling all frame pointer generation |
8155 | then we do not need to disable leaf frame pointer generation as a | |
8156 | separate operation. But if we are *only* disabling leaf frame pointer | |
8157 | generation then we set flag_omit_frame_pointer to true, but in | |
8158 | aarch64_frame_pointer_required we return false only for leaf functions. | |
8159 | ||
8160 | PR 70044: We have to be careful about being called multiple times for the | |
8161 | same function. Once we have decided to set flag_omit_frame_pointer just | |
8162 | so that we can omit leaf frame pointers, we must then not interpret a | |
8163 | second call as meaning that all frame pointer generation should be | |
8164 | omitted. We do this by setting flag_omit_frame_pointer to a special, | |
8165 | non-zero value. */ | |
8166 | if (opts->x_flag_omit_frame_pointer == 2) | |
8167 | opts->x_flag_omit_frame_pointer = 0; | |
8168 | ||
0cfff2a1 KT |
8169 | if (opts->x_flag_omit_frame_pointer) |
8170 | opts->x_flag_omit_leaf_frame_pointer = false; | |
8171 | else if (opts->x_flag_omit_leaf_frame_pointer) | |
a3dc8760 | 8172 | opts->x_flag_omit_frame_pointer = 2; |
43e9d192 | 8173 | |
1be34295 | 8174 | /* If not optimizing for size, set the default |
0cfff2a1 KT |
8175 | alignment to what the target wants. */ |
8176 | if (!opts->x_optimize_size) | |
43e9d192 | 8177 | { |
0cfff2a1 KT |
8178 | if (opts->x_align_loops <= 0) |
8179 | opts->x_align_loops = aarch64_tune_params.loop_align; | |
8180 | if (opts->x_align_jumps <= 0) | |
8181 | opts->x_align_jumps = aarch64_tune_params.jump_align; | |
8182 | if (opts->x_align_functions <= 0) | |
8183 | opts->x_align_functions = aarch64_tune_params.function_align; | |
43e9d192 | 8184 | } |
b4f50fd4 | 8185 | |
9ee6540a WD |
8186 | /* We default to no pc-relative literal loads. */ |
8187 | ||
8188 | aarch64_pcrelative_literal_loads = false; | |
8189 | ||
8190 | /* If -mpc-relative-literal-loads is set on the command line, this | |
b4f50fd4 | 8191 | implies that the user asked for PC relative literal loads. */ |
9ee6540a WD |
8192 | if (opts->x_pcrelative_literal_loads == 1) |
8193 | aarch64_pcrelative_literal_loads = true; | |
b4f50fd4 | 8194 | |
48bb1a55 CL |
8195 | /* This is PR70113. When building the Linux kernel with |
8196 | CONFIG_ARM64_ERRATUM_843419, support for relocations | |
8197 | R_AARCH64_ADR_PREL_PG_HI21 and R_AARCH64_ADR_PREL_PG_HI21_NC is | |
8198 | removed from the kernel to avoid loading objects with possibly | |
9ee6540a | 8199 | offending sequences. Without -mpc-relative-literal-loads we would |
48bb1a55 CL |
8200 | generate such relocations, preventing the kernel build from |
8201 | succeeding. */ | |
9ee6540a WD |
8202 | if (opts->x_pcrelative_literal_loads == 2 |
8203 | && TARGET_FIX_ERR_A53_843419) | |
8204 | aarch64_pcrelative_literal_loads = true; | |
8205 | ||
8206 | /* In the tiny memory model it makes no sense to disallow PC relative | |
8207 | literal pool loads. */ | |
8208 | if (aarch64_cmodel == AARCH64_CMODEL_TINY | |
8209 | || aarch64_cmodel == AARCH64_CMODEL_TINY_PIC) | |
8210 | aarch64_pcrelative_literal_loads = true; | |
98daafa0 EM |
8211 | |
8212 | /* When enabling the lower precision Newton series for the square root, also | |
8213 | enable it for the reciprocal square root, since the latter is an | |
8214 | intermediary step for the former. */ | |
8215 | if (flag_mlow_precision_sqrt) | |
8216 | flag_mrecip_low_precision_sqrt = true; | |
0cfff2a1 | 8217 | } |
43e9d192 | 8218 | |
0cfff2a1 KT |
8219 | /* 'Unpack' up the internal tuning structs and update the options |
8220 | in OPTS. The caller must have set up selected_tune and selected_arch | |
8221 | as all the other target-specific codegen decisions are | |
8222 | derived from them. */ | |
8223 | ||
e4ea20c8 | 8224 | void |
0cfff2a1 KT |
8225 | aarch64_override_options_internal (struct gcc_options *opts) |
8226 | { | |
8227 | aarch64_tune_flags = selected_tune->flags; | |
8228 | aarch64_tune = selected_tune->sched_core; | |
8229 | /* Make a copy of the tuning parameters attached to the core, which | |
8230 | we may later overwrite. */ | |
8231 | aarch64_tune_params = *(selected_tune->tune); | |
8232 | aarch64_architecture_version = selected_arch->architecture_version; | |
8233 | ||
8234 | if (opts->x_aarch64_override_tune_string) | |
8235 | aarch64_parse_override_string (opts->x_aarch64_override_tune_string, | |
8236 | &aarch64_tune_params); | |
8237 | ||
8238 | /* This target defaults to strict volatile bitfields. */ | |
8239 | if (opts->x_flag_strict_volatile_bitfields < 0 && abi_version_at_least (2)) | |
8240 | opts->x_flag_strict_volatile_bitfields = 1; | |
8241 | ||
0cfff2a1 | 8242 | initialize_aarch64_code_model (opts); |
5eee3c34 | 8243 | initialize_aarch64_tls_size (opts); |
63892fa2 | 8244 | |
2d6bc7fa KT |
8245 | int queue_depth = 0; |
8246 | switch (aarch64_tune_params.autoprefetcher_model) | |
8247 | { | |
8248 | case tune_params::AUTOPREFETCHER_OFF: | |
8249 | queue_depth = -1; | |
8250 | break; | |
8251 | case tune_params::AUTOPREFETCHER_WEAK: | |
8252 | queue_depth = 0; | |
8253 | break; | |
8254 | case tune_params::AUTOPREFETCHER_STRONG: | |
8255 | queue_depth = max_insn_queue_index + 1; | |
8256 | break; | |
8257 | default: | |
8258 | gcc_unreachable (); | |
8259 | } | |
8260 | ||
8261 | /* We don't mind passing in global_options_set here as we don't use | |
8262 | the *options_set structs anyway. */ | |
8263 | maybe_set_param_value (PARAM_SCHED_AUTOPREF_QUEUE_DEPTH, | |
8264 | queue_depth, | |
8265 | opts->x_param_values, | |
8266 | global_options_set.x_param_values); | |
8267 | ||
50487d79 EM |
8268 | /* Set the L1 cache line size. */ |
8269 | if (selected_cpu->tune->cache_line_size != 0) | |
8270 | maybe_set_param_value (PARAM_L1_CACHE_LINE_SIZE, | |
8271 | selected_cpu->tune->cache_line_size, | |
8272 | opts->x_param_values, | |
8273 | global_options_set.x_param_values); | |
8274 | ||
0cfff2a1 KT |
8275 | aarch64_override_options_after_change_1 (opts); |
8276 | } | |
43e9d192 | 8277 | |
0cfff2a1 KT |
8278 | /* Validate a command-line -mcpu option. Parse the cpu and extensions (if any) |
8279 | specified in STR and throw errors if appropriate. Put the results if | |
361fb3ee KT |
8280 | they are valid in RES and ISA_FLAGS. Return whether the option is |
8281 | valid. */ | |
43e9d192 | 8282 | |
361fb3ee | 8283 | static bool |
0cfff2a1 KT |
8284 | aarch64_validate_mcpu (const char *str, const struct processor **res, |
8285 | unsigned long *isa_flags) | |
8286 | { | |
8287 | enum aarch64_parse_opt_result parse_res | |
8288 | = aarch64_parse_cpu (str, res, isa_flags); | |
8289 | ||
8290 | if (parse_res == AARCH64_PARSE_OK) | |
361fb3ee | 8291 | return true; |
0cfff2a1 KT |
8292 | |
8293 | switch (parse_res) | |
8294 | { | |
8295 | case AARCH64_PARSE_MISSING_ARG: | |
8296 | error ("missing cpu name in -mcpu=%qs", str); | |
8297 | break; | |
8298 | case AARCH64_PARSE_INVALID_ARG: | |
8299 | error ("unknown value %qs for -mcpu", str); | |
8300 | break; | |
8301 | case AARCH64_PARSE_INVALID_FEATURE: | |
8302 | error ("invalid feature modifier in -mcpu=%qs", str); | |
8303 | break; | |
8304 | default: | |
8305 | gcc_unreachable (); | |
8306 | } | |
361fb3ee KT |
8307 | |
8308 | return false; | |
0cfff2a1 KT |
8309 | } |
8310 | ||
8311 | /* Validate a command-line -march option. Parse the arch and extensions | |
8312 | (if any) specified in STR and throw errors if appropriate. Put the | |
361fb3ee KT |
8313 | results, if they are valid, in RES and ISA_FLAGS. Return whether the |
8314 | option is valid. */ | |
0cfff2a1 | 8315 | |
361fb3ee | 8316 | static bool |
0cfff2a1 KT |
8317 | aarch64_validate_march (const char *str, const struct processor **res, |
8318 | unsigned long *isa_flags) | |
8319 | { | |
8320 | enum aarch64_parse_opt_result parse_res | |
8321 | = aarch64_parse_arch (str, res, isa_flags); | |
8322 | ||
8323 | if (parse_res == AARCH64_PARSE_OK) | |
361fb3ee | 8324 | return true; |
0cfff2a1 KT |
8325 | |
8326 | switch (parse_res) | |
8327 | { | |
8328 | case AARCH64_PARSE_MISSING_ARG: | |
8329 | error ("missing arch name in -march=%qs", str); | |
8330 | break; | |
8331 | case AARCH64_PARSE_INVALID_ARG: | |
8332 | error ("unknown value %qs for -march", str); | |
8333 | break; | |
8334 | case AARCH64_PARSE_INVALID_FEATURE: | |
8335 | error ("invalid feature modifier in -march=%qs", str); | |
8336 | break; | |
8337 | default: | |
8338 | gcc_unreachable (); | |
8339 | } | |
361fb3ee KT |
8340 | |
8341 | return false; | |
0cfff2a1 KT |
8342 | } |
8343 | ||
8344 | /* Validate a command-line -mtune option. Parse the cpu | |
8345 | specified in STR and throw errors if appropriate. Put the | |
361fb3ee KT |
8346 | result, if it is valid, in RES. Return whether the option is |
8347 | valid. */ | |
0cfff2a1 | 8348 | |
361fb3ee | 8349 | static bool |
0cfff2a1 KT |
8350 | aarch64_validate_mtune (const char *str, const struct processor **res) |
8351 | { | |
8352 | enum aarch64_parse_opt_result parse_res | |
8353 | = aarch64_parse_tune (str, res); | |
8354 | ||
8355 | if (parse_res == AARCH64_PARSE_OK) | |
361fb3ee | 8356 | return true; |
0cfff2a1 KT |
8357 | |
8358 | switch (parse_res) | |
8359 | { | |
8360 | case AARCH64_PARSE_MISSING_ARG: | |
8361 | error ("missing cpu name in -mtune=%qs", str); | |
8362 | break; | |
8363 | case AARCH64_PARSE_INVALID_ARG: | |
8364 | error ("unknown value %qs for -mtune", str); | |
8365 | break; | |
8366 | default: | |
8367 | gcc_unreachable (); | |
8368 | } | |
361fb3ee KT |
8369 | return false; |
8370 | } | |
8371 | ||
8372 | /* Return the CPU corresponding to the enum CPU. | |
8373 | If it doesn't specify a cpu, return the default. */ | |
8374 | ||
8375 | static const struct processor * | |
8376 | aarch64_get_tune_cpu (enum aarch64_processor cpu) | |
8377 | { | |
8378 | if (cpu != aarch64_none) | |
8379 | return &all_cores[cpu]; | |
8380 | ||
8381 | /* The & 0x3f is to extract the bottom 6 bits that encode the | |
8382 | default cpu as selected by the --with-cpu GCC configure option | |
8383 | in config.gcc. | |
8384 | ???: The whole TARGET_CPU_DEFAULT and AARCH64_CPU_DEFAULT_FLAGS | |
8385 | flags mechanism should be reworked to make it more sane. */ | |
8386 | return &all_cores[TARGET_CPU_DEFAULT & 0x3f]; | |
8387 | } | |
8388 | ||
8389 | /* Return the architecture corresponding to the enum ARCH. | |
8390 | If it doesn't specify a valid architecture, return the default. */ | |
8391 | ||
8392 | static const struct processor * | |
8393 | aarch64_get_arch (enum aarch64_arch arch) | |
8394 | { | |
8395 | if (arch != aarch64_no_arch) | |
8396 | return &all_architectures[arch]; | |
8397 | ||
8398 | const struct processor *cpu = &all_cores[TARGET_CPU_DEFAULT & 0x3f]; | |
8399 | ||
8400 | return &all_architectures[cpu->arch]; | |
0cfff2a1 KT |
8401 | } |
8402 | ||
8403 | /* Implement TARGET_OPTION_OVERRIDE. This is called once in the beginning | |
8404 | and is used to parse the -m{cpu,tune,arch} strings and setup the initial | |
8405 | tuning structs. In particular it must set selected_tune and | |
8406 | aarch64_isa_flags that define the available ISA features and tuning | |
8407 | decisions. It must also set selected_arch as this will be used to | |
8408 | output the .arch asm tags for each function. */ | |
8409 | ||
8410 | static void | |
8411 | aarch64_override_options (void) | |
8412 | { | |
8413 | unsigned long cpu_isa = 0; | |
8414 | unsigned long arch_isa = 0; | |
8415 | aarch64_isa_flags = 0; | |
8416 | ||
361fb3ee KT |
8417 | bool valid_cpu = true; |
8418 | bool valid_tune = true; | |
8419 | bool valid_arch = true; | |
8420 | ||
0cfff2a1 KT |
8421 | selected_cpu = NULL; |
8422 | selected_arch = NULL; | |
8423 | selected_tune = NULL; | |
8424 | ||
8425 | /* -mcpu=CPU is shorthand for -march=ARCH_FOR_CPU, -mtune=CPU. | |
8426 | If either of -march or -mtune is given, they override their | |
8427 | respective component of -mcpu. */ | |
8428 | if (aarch64_cpu_string) | |
361fb3ee KT |
8429 | valid_cpu = aarch64_validate_mcpu (aarch64_cpu_string, &selected_cpu, |
8430 | &cpu_isa); | |
0cfff2a1 KT |
8431 | |
8432 | if (aarch64_arch_string) | |
361fb3ee KT |
8433 | valid_arch = aarch64_validate_march (aarch64_arch_string, &selected_arch, |
8434 | &arch_isa); | |
0cfff2a1 KT |
8435 | |
8436 | if (aarch64_tune_string) | |
361fb3ee | 8437 | valid_tune = aarch64_validate_mtune (aarch64_tune_string, &selected_tune); |
43e9d192 IB |
8438 | |
8439 | /* If the user did not specify a processor, choose the default | |
8440 | one for them. This will be the CPU set during configuration using | |
a3cd0246 | 8441 | --with-cpu, otherwise it is "generic". */ |
43e9d192 IB |
8442 | if (!selected_cpu) |
8443 | { | |
0cfff2a1 KT |
8444 | if (selected_arch) |
8445 | { | |
8446 | selected_cpu = &all_cores[selected_arch->ident]; | |
8447 | aarch64_isa_flags = arch_isa; | |
361fb3ee | 8448 | explicit_arch = selected_arch->arch; |
0cfff2a1 KT |
8449 | } |
8450 | else | |
8451 | { | |
361fb3ee KT |
8452 | /* Get default configure-time CPU. */ |
8453 | selected_cpu = aarch64_get_tune_cpu (aarch64_none); | |
0cfff2a1 KT |
8454 | aarch64_isa_flags = TARGET_CPU_DEFAULT >> 6; |
8455 | } | |
361fb3ee KT |
8456 | |
8457 | if (selected_tune) | |
8458 | explicit_tune_core = selected_tune->ident; | |
0cfff2a1 KT |
8459 | } |
8460 | /* If both -mcpu and -march are specified check that they are architecturally | |
8461 | compatible, warn if they're not and prefer the -march ISA flags. */ | |
8462 | else if (selected_arch) | |
8463 | { | |
8464 | if (selected_arch->arch != selected_cpu->arch) | |
8465 | { | |
8466 | warning (0, "switch -mcpu=%s conflicts with -march=%s switch", | |
8467 | all_architectures[selected_cpu->arch].name, | |
8468 | selected_arch->name); | |
8469 | } | |
8470 | aarch64_isa_flags = arch_isa; | |
361fb3ee KT |
8471 | explicit_arch = selected_arch->arch; |
8472 | explicit_tune_core = selected_tune ? selected_tune->ident | |
8473 | : selected_cpu->ident; | |
0cfff2a1 KT |
8474 | } |
8475 | else | |
8476 | { | |
8477 | /* -mcpu but no -march. */ | |
8478 | aarch64_isa_flags = cpu_isa; | |
361fb3ee KT |
8479 | explicit_tune_core = selected_tune ? selected_tune->ident |
8480 | : selected_cpu->ident; | |
8481 | gcc_assert (selected_cpu); | |
8482 | selected_arch = &all_architectures[selected_cpu->arch]; | |
8483 | explicit_arch = selected_arch->arch; | |
43e9d192 IB |
8484 | } |
8485 | ||
0cfff2a1 KT |
8486 | /* Set the arch as well as we will need it when outputing |
8487 | the .arch directive in assembly. */ | |
8488 | if (!selected_arch) | |
8489 | { | |
8490 | gcc_assert (selected_cpu); | |
8491 | selected_arch = &all_architectures[selected_cpu->arch]; | |
8492 | } | |
43e9d192 | 8493 | |
43e9d192 | 8494 | if (!selected_tune) |
3edaf26d | 8495 | selected_tune = selected_cpu; |
43e9d192 | 8496 | |
0cfff2a1 KT |
8497 | #ifndef HAVE_AS_MABI_OPTION |
8498 | /* The compiler may have been configured with 2.23.* binutils, which does | |
8499 | not have support for ILP32. */ | |
8500 | if (TARGET_ILP32) | |
8501 | error ("Assembler does not support -mabi=ilp32"); | |
8502 | #endif | |
43e9d192 | 8503 | |
361fb3ee KT |
8504 | /* Make sure we properly set up the explicit options. */ |
8505 | if ((aarch64_cpu_string && valid_cpu) | |
8506 | || (aarch64_tune_string && valid_tune)) | |
8507 | gcc_assert (explicit_tune_core != aarch64_none); | |
8508 | ||
8509 | if ((aarch64_cpu_string && valid_cpu) | |
8510 | || (aarch64_arch_string && valid_arch)) | |
8511 | gcc_assert (explicit_arch != aarch64_no_arch); | |
8512 | ||
0cfff2a1 KT |
8513 | aarch64_override_options_internal (&global_options); |
8514 | ||
8515 | /* Save these options as the default ones in case we push and pop them later | |
8516 | while processing functions with potential target attributes. */ | |
8517 | target_option_default_node = target_option_current_node | |
8518 | = build_target_option_node (&global_options); | |
5e396da6 | 8519 | |
e2fc7193 | 8520 | aarch64_register_fma_steering (); |
fde9b31b | 8521 | |
43e9d192 IB |
8522 | } |
8523 | ||
8524 | /* Implement targetm.override_options_after_change. */ | |
8525 | ||
8526 | static void | |
8527 | aarch64_override_options_after_change (void) | |
8528 | { | |
0cfff2a1 | 8529 | aarch64_override_options_after_change_1 (&global_options); |
43e9d192 IB |
8530 | } |
8531 | ||
8532 | static struct machine_function * | |
8533 | aarch64_init_machine_status (void) | |
8534 | { | |
8535 | struct machine_function *machine; | |
766090c2 | 8536 | machine = ggc_cleared_alloc<machine_function> (); |
43e9d192 IB |
8537 | return machine; |
8538 | } | |
8539 | ||
8540 | void | |
8541 | aarch64_init_expanders (void) | |
8542 | { | |
8543 | init_machine_status = aarch64_init_machine_status; | |
8544 | } | |
8545 | ||
8546 | /* A checking mechanism for the implementation of the various code models. */ | |
8547 | static void | |
0cfff2a1 | 8548 | initialize_aarch64_code_model (struct gcc_options *opts) |
43e9d192 | 8549 | { |
0cfff2a1 | 8550 | if (opts->x_flag_pic) |
43e9d192 | 8551 | { |
0cfff2a1 | 8552 | switch (opts->x_aarch64_cmodel_var) |
43e9d192 IB |
8553 | { |
8554 | case AARCH64_CMODEL_TINY: | |
8555 | aarch64_cmodel = AARCH64_CMODEL_TINY_PIC; | |
8556 | break; | |
8557 | case AARCH64_CMODEL_SMALL: | |
34ecdb0f | 8558 | #ifdef HAVE_AS_SMALL_PIC_RELOCS |
1b1e81f8 JW |
8559 | aarch64_cmodel = (flag_pic == 2 |
8560 | ? AARCH64_CMODEL_SMALL_PIC | |
8561 | : AARCH64_CMODEL_SMALL_SPIC); | |
34ecdb0f JW |
8562 | #else |
8563 | aarch64_cmodel = AARCH64_CMODEL_SMALL_PIC; | |
8564 | #endif | |
43e9d192 IB |
8565 | break; |
8566 | case AARCH64_CMODEL_LARGE: | |
8567 | sorry ("code model %qs with -f%s", "large", | |
0cfff2a1 | 8568 | opts->x_flag_pic > 1 ? "PIC" : "pic"); |
1c652781 | 8569 | break; |
43e9d192 IB |
8570 | default: |
8571 | gcc_unreachable (); | |
8572 | } | |
8573 | } | |
8574 | else | |
0cfff2a1 | 8575 | aarch64_cmodel = opts->x_aarch64_cmodel_var; |
43e9d192 IB |
8576 | } |
8577 | ||
361fb3ee KT |
8578 | /* Implement TARGET_OPTION_SAVE. */ |
8579 | ||
8580 | static void | |
8581 | aarch64_option_save (struct cl_target_option *ptr, struct gcc_options *opts) | |
8582 | { | |
8583 | ptr->x_aarch64_override_tune_string = opts->x_aarch64_override_tune_string; | |
8584 | } | |
8585 | ||
8586 | /* Implements TARGET_OPTION_RESTORE. Restore the backend codegen decisions | |
8587 | using the information saved in PTR. */ | |
8588 | ||
8589 | static void | |
8590 | aarch64_option_restore (struct gcc_options *opts, struct cl_target_option *ptr) | |
8591 | { | |
8592 | opts->x_explicit_tune_core = ptr->x_explicit_tune_core; | |
8593 | selected_tune = aarch64_get_tune_cpu (ptr->x_explicit_tune_core); | |
8594 | opts->x_explicit_arch = ptr->x_explicit_arch; | |
8595 | selected_arch = aarch64_get_arch (ptr->x_explicit_arch); | |
8596 | opts->x_aarch64_override_tune_string = ptr->x_aarch64_override_tune_string; | |
8597 | ||
8598 | aarch64_override_options_internal (opts); | |
8599 | } | |
8600 | ||
8601 | /* Implement TARGET_OPTION_PRINT. */ | |
8602 | ||
8603 | static void | |
8604 | aarch64_option_print (FILE *file, int indent, struct cl_target_option *ptr) | |
8605 | { | |
8606 | const struct processor *cpu | |
8607 | = aarch64_get_tune_cpu (ptr->x_explicit_tune_core); | |
8608 | unsigned long isa_flags = ptr->x_aarch64_isa_flags; | |
8609 | const struct processor *arch = aarch64_get_arch (ptr->x_explicit_arch); | |
054b4005 | 8610 | std::string extension |
04a99ebe | 8611 | = aarch64_get_extension_string_for_isa_flags (isa_flags, arch->flags); |
361fb3ee KT |
8612 | |
8613 | fprintf (file, "%*sselected tune = %s\n", indent, "", cpu->name); | |
054b4005 JG |
8614 | fprintf (file, "%*sselected arch = %s%s\n", indent, "", |
8615 | arch->name, extension.c_str ()); | |
361fb3ee KT |
8616 | } |
8617 | ||
d78006d9 KT |
8618 | static GTY(()) tree aarch64_previous_fndecl; |
8619 | ||
e4ea20c8 KT |
8620 | void |
8621 | aarch64_reset_previous_fndecl (void) | |
8622 | { | |
8623 | aarch64_previous_fndecl = NULL; | |
8624 | } | |
8625 | ||
acfc1ac1 KT |
8626 | /* Restore or save the TREE_TARGET_GLOBALS from or to NEW_TREE. |
8627 | Used by aarch64_set_current_function and aarch64_pragma_target_parse to | |
8628 | make sure optab availability predicates are recomputed when necessary. */ | |
8629 | ||
8630 | void | |
8631 | aarch64_save_restore_target_globals (tree new_tree) | |
8632 | { | |
8633 | if (TREE_TARGET_GLOBALS (new_tree)) | |
8634 | restore_target_globals (TREE_TARGET_GLOBALS (new_tree)); | |
8635 | else if (new_tree == target_option_default_node) | |
8636 | restore_target_globals (&default_target_globals); | |
8637 | else | |
8638 | TREE_TARGET_GLOBALS (new_tree) = save_target_globals_default_opts (); | |
8639 | } | |
8640 | ||
d78006d9 KT |
8641 | /* Implement TARGET_SET_CURRENT_FUNCTION. Unpack the codegen decisions |
8642 | like tuning and ISA features from the DECL_FUNCTION_SPECIFIC_TARGET | |
8643 | of the function, if such exists. This function may be called multiple | |
8644 | times on a single function so use aarch64_previous_fndecl to avoid | |
8645 | setting up identical state. */ | |
8646 | ||
8647 | static void | |
8648 | aarch64_set_current_function (tree fndecl) | |
8649 | { | |
acfc1ac1 KT |
8650 | if (!fndecl || fndecl == aarch64_previous_fndecl) |
8651 | return; | |
8652 | ||
d78006d9 KT |
8653 | tree old_tree = (aarch64_previous_fndecl |
8654 | ? DECL_FUNCTION_SPECIFIC_TARGET (aarch64_previous_fndecl) | |
8655 | : NULL_TREE); | |
8656 | ||
acfc1ac1 | 8657 | tree new_tree = DECL_FUNCTION_SPECIFIC_TARGET (fndecl); |
d78006d9 | 8658 | |
acfc1ac1 KT |
8659 | /* If current function has no attributes but the previous one did, |
8660 | use the default node. */ | |
8661 | if (!new_tree && old_tree) | |
8662 | new_tree = target_option_default_node; | |
d78006d9 | 8663 | |
acfc1ac1 KT |
8664 | /* If nothing to do, return. #pragma GCC reset or #pragma GCC pop to |
8665 | the default have been handled by aarch64_save_restore_target_globals from | |
8666 | aarch64_pragma_target_parse. */ | |
8667 | if (old_tree == new_tree) | |
8668 | return; | |
d78006d9 | 8669 | |
acfc1ac1 | 8670 | aarch64_previous_fndecl = fndecl; |
6e17a23b | 8671 | |
acfc1ac1 KT |
8672 | /* First set the target options. */ |
8673 | cl_target_option_restore (&global_options, TREE_TARGET_OPTION (new_tree)); | |
6e17a23b | 8674 | |
acfc1ac1 | 8675 | aarch64_save_restore_target_globals (new_tree); |
d78006d9 | 8676 | } |
361fb3ee | 8677 | |
5a2c8331 KT |
8678 | /* Enum describing the various ways we can handle attributes. |
8679 | In many cases we can reuse the generic option handling machinery. */ | |
8680 | ||
8681 | enum aarch64_attr_opt_type | |
8682 | { | |
8683 | aarch64_attr_mask, /* Attribute should set a bit in target_flags. */ | |
8684 | aarch64_attr_bool, /* Attribute sets or unsets a boolean variable. */ | |
8685 | aarch64_attr_enum, /* Attribute sets an enum variable. */ | |
8686 | aarch64_attr_custom /* Attribute requires a custom handling function. */ | |
8687 | }; | |
8688 | ||
8689 | /* All the information needed to handle a target attribute. | |
8690 | NAME is the name of the attribute. | |
9c582551 | 8691 | ATTR_TYPE specifies the type of behavior of the attribute as described |
5a2c8331 KT |
8692 | in the definition of enum aarch64_attr_opt_type. |
8693 | ALLOW_NEG is true if the attribute supports a "no-" form. | |
8694 | HANDLER is the function that takes the attribute string and whether | |
8695 | it is a pragma or attribute and handles the option. It is needed only | |
8696 | when the ATTR_TYPE is aarch64_attr_custom. | |
8697 | OPT_NUM is the enum specifying the option that the attribute modifies. | |
9c582551 | 8698 | This is needed for attributes that mirror the behavior of a command-line |
5a2c8331 KT |
8699 | option, that is it has ATTR_TYPE aarch64_attr_mask, aarch64_attr_bool or |
8700 | aarch64_attr_enum. */ | |
8701 | ||
8702 | struct aarch64_attribute_info | |
8703 | { | |
8704 | const char *name; | |
8705 | enum aarch64_attr_opt_type attr_type; | |
8706 | bool allow_neg; | |
8707 | bool (*handler) (const char *, const char *); | |
8708 | enum opt_code opt_num; | |
8709 | }; | |
8710 | ||
8711 | /* Handle the ARCH_STR argument to the arch= target attribute. | |
8712 | PRAGMA_OR_ATTR is used in potential error messages. */ | |
8713 | ||
8714 | static bool | |
8715 | aarch64_handle_attr_arch (const char *str, const char *pragma_or_attr) | |
8716 | { | |
8717 | const struct processor *tmp_arch = NULL; | |
8718 | enum aarch64_parse_opt_result parse_res | |
8719 | = aarch64_parse_arch (str, &tmp_arch, &aarch64_isa_flags); | |
8720 | ||
8721 | if (parse_res == AARCH64_PARSE_OK) | |
8722 | { | |
8723 | gcc_assert (tmp_arch); | |
8724 | selected_arch = tmp_arch; | |
8725 | explicit_arch = selected_arch->arch; | |
8726 | return true; | |
8727 | } | |
8728 | ||
8729 | switch (parse_res) | |
8730 | { | |
8731 | case AARCH64_PARSE_MISSING_ARG: | |
8732 | error ("missing architecture name in 'arch' target %s", pragma_or_attr); | |
8733 | break; | |
8734 | case AARCH64_PARSE_INVALID_ARG: | |
8735 | error ("unknown value %qs for 'arch' target %s", str, pragma_or_attr); | |
8736 | break; | |
8737 | case AARCH64_PARSE_INVALID_FEATURE: | |
8738 | error ("invalid feature modifier %qs for 'arch' target %s", | |
8739 | str, pragma_or_attr); | |
8740 | break; | |
8741 | default: | |
8742 | gcc_unreachable (); | |
8743 | } | |
8744 | ||
8745 | return false; | |
8746 | } | |
8747 | ||
8748 | /* Handle the argument CPU_STR to the cpu= target attribute. | |
8749 | PRAGMA_OR_ATTR is used in potential error messages. */ | |
8750 | ||
8751 | static bool | |
8752 | aarch64_handle_attr_cpu (const char *str, const char *pragma_or_attr) | |
8753 | { | |
8754 | const struct processor *tmp_cpu = NULL; | |
8755 | enum aarch64_parse_opt_result parse_res | |
8756 | = aarch64_parse_cpu (str, &tmp_cpu, &aarch64_isa_flags); | |
8757 | ||
8758 | if (parse_res == AARCH64_PARSE_OK) | |
8759 | { | |
8760 | gcc_assert (tmp_cpu); | |
8761 | selected_tune = tmp_cpu; | |
8762 | explicit_tune_core = selected_tune->ident; | |
8763 | ||
8764 | selected_arch = &all_architectures[tmp_cpu->arch]; | |
8765 | explicit_arch = selected_arch->arch; | |
8766 | return true; | |
8767 | } | |
8768 | ||
8769 | switch (parse_res) | |
8770 | { | |
8771 | case AARCH64_PARSE_MISSING_ARG: | |
8772 | error ("missing cpu name in 'cpu' target %s", pragma_or_attr); | |
8773 | break; | |
8774 | case AARCH64_PARSE_INVALID_ARG: | |
8775 | error ("unknown value %qs for 'cpu' target %s", str, pragma_or_attr); | |
8776 | break; | |
8777 | case AARCH64_PARSE_INVALID_FEATURE: | |
8778 | error ("invalid feature modifier %qs for 'cpu' target %s", | |
8779 | str, pragma_or_attr); | |
8780 | break; | |
8781 | default: | |
8782 | gcc_unreachable (); | |
8783 | } | |
8784 | ||
8785 | return false; | |
8786 | } | |
8787 | ||
8788 | /* Handle the argument STR to the tune= target attribute. | |
8789 | PRAGMA_OR_ATTR is used in potential error messages. */ | |
8790 | ||
8791 | static bool | |
8792 | aarch64_handle_attr_tune (const char *str, const char *pragma_or_attr) | |
8793 | { | |
8794 | const struct processor *tmp_tune = NULL; | |
8795 | enum aarch64_parse_opt_result parse_res | |
8796 | = aarch64_parse_tune (str, &tmp_tune); | |
8797 | ||
8798 | if (parse_res == AARCH64_PARSE_OK) | |
8799 | { | |
8800 | gcc_assert (tmp_tune); | |
8801 | selected_tune = tmp_tune; | |
8802 | explicit_tune_core = selected_tune->ident; | |
8803 | return true; | |
8804 | } | |
8805 | ||
8806 | switch (parse_res) | |
8807 | { | |
8808 | case AARCH64_PARSE_INVALID_ARG: | |
8809 | error ("unknown value %qs for 'tune' target %s", str, pragma_or_attr); | |
8810 | break; | |
8811 | default: | |
8812 | gcc_unreachable (); | |
8813 | } | |
8814 | ||
8815 | return false; | |
8816 | } | |
8817 | ||
8818 | /* Parse an architecture extensions target attribute string specified in STR. | |
8819 | For example "+fp+nosimd". Show any errors if needed. Return TRUE | |
8820 | if successful. Update aarch64_isa_flags to reflect the ISA features | |
8821 | modified. | |
8822 | PRAGMA_OR_ATTR is used in potential error messages. */ | |
8823 | ||
8824 | static bool | |
8825 | aarch64_handle_attr_isa_flags (char *str, const char *pragma_or_attr) | |
8826 | { | |
8827 | enum aarch64_parse_opt_result parse_res; | |
8828 | unsigned long isa_flags = aarch64_isa_flags; | |
8829 | ||
e4ea20c8 KT |
8830 | /* We allow "+nothing" in the beginning to clear out all architectural |
8831 | features if the user wants to handpick specific features. */ | |
8832 | if (strncmp ("+nothing", str, 8) == 0) | |
8833 | { | |
8834 | isa_flags = 0; | |
8835 | str += 8; | |
8836 | } | |
8837 | ||
5a2c8331 KT |
8838 | parse_res = aarch64_parse_extension (str, &isa_flags); |
8839 | ||
8840 | if (parse_res == AARCH64_PARSE_OK) | |
8841 | { | |
8842 | aarch64_isa_flags = isa_flags; | |
8843 | return true; | |
8844 | } | |
8845 | ||
8846 | switch (parse_res) | |
8847 | { | |
8848 | case AARCH64_PARSE_MISSING_ARG: | |
8849 | error ("missing feature modifier in target %s %qs", | |
8850 | pragma_or_attr, str); | |
8851 | break; | |
8852 | ||
8853 | case AARCH64_PARSE_INVALID_FEATURE: | |
8854 | error ("invalid feature modifier in target %s %qs", | |
8855 | pragma_or_attr, str); | |
8856 | break; | |
8857 | ||
8858 | default: | |
8859 | gcc_unreachable (); | |
8860 | } | |
8861 | ||
8862 | return false; | |
8863 | } | |
8864 | ||
8865 | /* The target attributes that we support. On top of these we also support just | |
8866 | ISA extensions, like __attribute__ ((target ("+crc"))), but that case is | |
8867 | handled explicitly in aarch64_process_one_target_attr. */ | |
8868 | ||
8869 | static const struct aarch64_attribute_info aarch64_attributes[] = | |
8870 | { | |
8871 | { "general-regs-only", aarch64_attr_mask, false, NULL, | |
8872 | OPT_mgeneral_regs_only }, | |
8873 | { "fix-cortex-a53-835769", aarch64_attr_bool, true, NULL, | |
8874 | OPT_mfix_cortex_a53_835769 }, | |
48bb1a55 CL |
8875 | { "fix-cortex-a53-843419", aarch64_attr_bool, true, NULL, |
8876 | OPT_mfix_cortex_a53_843419 }, | |
5a2c8331 KT |
8877 | { "cmodel", aarch64_attr_enum, false, NULL, OPT_mcmodel_ }, |
8878 | { "strict-align", aarch64_attr_mask, false, NULL, OPT_mstrict_align }, | |
8879 | { "omit-leaf-frame-pointer", aarch64_attr_bool, true, NULL, | |
8880 | OPT_momit_leaf_frame_pointer }, | |
8881 | { "tls-dialect", aarch64_attr_enum, false, NULL, OPT_mtls_dialect_ }, | |
8882 | { "arch", aarch64_attr_custom, false, aarch64_handle_attr_arch, | |
8883 | OPT_march_ }, | |
8884 | { "cpu", aarch64_attr_custom, false, aarch64_handle_attr_cpu, OPT_mcpu_ }, | |
8885 | { "tune", aarch64_attr_custom, false, aarch64_handle_attr_tune, | |
8886 | OPT_mtune_ }, | |
8887 | { NULL, aarch64_attr_custom, false, NULL, OPT____ } | |
8888 | }; | |
8889 | ||
8890 | /* Parse ARG_STR which contains the definition of one target attribute. | |
8891 | Show appropriate errors if any or return true if the attribute is valid. | |
8892 | PRAGMA_OR_ATTR holds the string to use in error messages about whether | |
8893 | we're processing a target attribute or pragma. */ | |
8894 | ||
8895 | static bool | |
8896 | aarch64_process_one_target_attr (char *arg_str, const char* pragma_or_attr) | |
8897 | { | |
8898 | bool invert = false; | |
8899 | ||
8900 | size_t len = strlen (arg_str); | |
8901 | ||
8902 | if (len == 0) | |
8903 | { | |
8904 | error ("malformed target %s", pragma_or_attr); | |
8905 | return false; | |
8906 | } | |
8907 | ||
8908 | char *str_to_check = (char *) alloca (len + 1); | |
8909 | strcpy (str_to_check, arg_str); | |
8910 | ||
8911 | /* Skip leading whitespace. */ | |
8912 | while (*str_to_check == ' ' || *str_to_check == '\t') | |
8913 | str_to_check++; | |
8914 | ||
8915 | /* We have something like __attribute__ ((target ("+fp+nosimd"))). | |
8916 | It is easier to detect and handle it explicitly here rather than going | |
8917 | through the machinery for the rest of the target attributes in this | |
8918 | function. */ | |
8919 | if (*str_to_check == '+') | |
8920 | return aarch64_handle_attr_isa_flags (str_to_check, pragma_or_attr); | |
8921 | ||
8922 | if (len > 3 && strncmp (str_to_check, "no-", 3) == 0) | |
8923 | { | |
8924 | invert = true; | |
8925 | str_to_check += 3; | |
8926 | } | |
8927 | char *arg = strchr (str_to_check, '='); | |
8928 | ||
8929 | /* If we found opt=foo then terminate STR_TO_CHECK at the '=' | |
8930 | and point ARG to "foo". */ | |
8931 | if (arg) | |
8932 | { | |
8933 | *arg = '\0'; | |
8934 | arg++; | |
8935 | } | |
8936 | const struct aarch64_attribute_info *p_attr; | |
16d12992 | 8937 | bool found = false; |
5a2c8331 KT |
8938 | for (p_attr = aarch64_attributes; p_attr->name; p_attr++) |
8939 | { | |
8940 | /* If the names don't match up, or the user has given an argument | |
8941 | to an attribute that doesn't accept one, or didn't give an argument | |
8942 | to an attribute that expects one, fail to match. */ | |
8943 | if (strcmp (str_to_check, p_attr->name) != 0) | |
8944 | continue; | |
8945 | ||
16d12992 | 8946 | found = true; |
5a2c8331 KT |
8947 | bool attr_need_arg_p = p_attr->attr_type == aarch64_attr_custom |
8948 | || p_attr->attr_type == aarch64_attr_enum; | |
8949 | ||
8950 | if (attr_need_arg_p ^ (arg != NULL)) | |
8951 | { | |
8952 | error ("target %s %qs does not accept an argument", | |
8953 | pragma_or_attr, str_to_check); | |
8954 | return false; | |
8955 | } | |
8956 | ||
8957 | /* If the name matches but the attribute does not allow "no-" versions | |
8958 | then we can't match. */ | |
8959 | if (invert && !p_attr->allow_neg) | |
8960 | { | |
8961 | error ("target %s %qs does not allow a negated form", | |
8962 | pragma_or_attr, str_to_check); | |
8963 | return false; | |
8964 | } | |
8965 | ||
8966 | switch (p_attr->attr_type) | |
8967 | { | |
8968 | /* Has a custom handler registered. | |
8969 | For example, cpu=, arch=, tune=. */ | |
8970 | case aarch64_attr_custom: | |
8971 | gcc_assert (p_attr->handler); | |
8972 | if (!p_attr->handler (arg, pragma_or_attr)) | |
8973 | return false; | |
8974 | break; | |
8975 | ||
8976 | /* Either set or unset a boolean option. */ | |
8977 | case aarch64_attr_bool: | |
8978 | { | |
8979 | struct cl_decoded_option decoded; | |
8980 | ||
8981 | generate_option (p_attr->opt_num, NULL, !invert, | |
8982 | CL_TARGET, &decoded); | |
8983 | aarch64_handle_option (&global_options, &global_options_set, | |
8984 | &decoded, input_location); | |
8985 | break; | |
8986 | } | |
8987 | /* Set or unset a bit in the target_flags. aarch64_handle_option | |
8988 | should know what mask to apply given the option number. */ | |
8989 | case aarch64_attr_mask: | |
8990 | { | |
8991 | struct cl_decoded_option decoded; | |
8992 | /* We only need to specify the option number. | |
8993 | aarch64_handle_option will know which mask to apply. */ | |
8994 | decoded.opt_index = p_attr->opt_num; | |
8995 | decoded.value = !invert; | |
8996 | aarch64_handle_option (&global_options, &global_options_set, | |
8997 | &decoded, input_location); | |
8998 | break; | |
8999 | } | |
9000 | /* Use the option setting machinery to set an option to an enum. */ | |
9001 | case aarch64_attr_enum: | |
9002 | { | |
9003 | gcc_assert (arg); | |
9004 | bool valid; | |
9005 | int value; | |
9006 | valid = opt_enum_arg_to_value (p_attr->opt_num, arg, | |
9007 | &value, CL_TARGET); | |
9008 | if (valid) | |
9009 | { | |
9010 | set_option (&global_options, NULL, p_attr->opt_num, value, | |
9011 | NULL, DK_UNSPECIFIED, input_location, | |
9012 | global_dc); | |
9013 | } | |
9014 | else | |
9015 | { | |
9016 | error ("target %s %s=%s is not valid", | |
9017 | pragma_or_attr, str_to_check, arg); | |
9018 | } | |
9019 | break; | |
9020 | } | |
9021 | default: | |
9022 | gcc_unreachable (); | |
9023 | } | |
9024 | } | |
9025 | ||
16d12992 KT |
9026 | /* If we reached here we either have found an attribute and validated |
9027 | it or didn't match any. If we matched an attribute but its arguments | |
9028 | were malformed we will have returned false already. */ | |
9029 | return found; | |
5a2c8331 KT |
9030 | } |
9031 | ||
9032 | /* Count how many times the character C appears in | |
9033 | NULL-terminated string STR. */ | |
9034 | ||
9035 | static unsigned int | |
9036 | num_occurences_in_str (char c, char *str) | |
9037 | { | |
9038 | unsigned int res = 0; | |
9039 | while (*str != '\0') | |
9040 | { | |
9041 | if (*str == c) | |
9042 | res++; | |
9043 | ||
9044 | str++; | |
9045 | } | |
9046 | ||
9047 | return res; | |
9048 | } | |
9049 | ||
9050 | /* Parse the tree in ARGS that contains the target attribute information | |
9051 | and update the global target options space. PRAGMA_OR_ATTR is a string | |
9052 | to be used in error messages, specifying whether this is processing | |
9053 | a target attribute or a target pragma. */ | |
9054 | ||
9055 | bool | |
9056 | aarch64_process_target_attr (tree args, const char* pragma_or_attr) | |
9057 | { | |
9058 | if (TREE_CODE (args) == TREE_LIST) | |
9059 | { | |
9060 | do | |
9061 | { | |
9062 | tree head = TREE_VALUE (args); | |
9063 | if (head) | |
9064 | { | |
9065 | if (!aarch64_process_target_attr (head, pragma_or_attr)) | |
9066 | return false; | |
9067 | } | |
9068 | args = TREE_CHAIN (args); | |
9069 | } while (args); | |
9070 | ||
9071 | return true; | |
9072 | } | |
9073 | /* We expect to find a string to parse. */ | |
9074 | gcc_assert (TREE_CODE (args) == STRING_CST); | |
9075 | ||
9076 | size_t len = strlen (TREE_STRING_POINTER (args)); | |
9077 | char *str_to_check = (char *) alloca (len + 1); | |
9078 | strcpy (str_to_check, TREE_STRING_POINTER (args)); | |
9079 | ||
9080 | if (len == 0) | |
9081 | { | |
9082 | error ("malformed target %s value", pragma_or_attr); | |
9083 | return false; | |
9084 | } | |
9085 | ||
9086 | /* Used to catch empty spaces between commas i.e. | |
9087 | attribute ((target ("attr1,,attr2"))). */ | |
9088 | unsigned int num_commas = num_occurences_in_str (',', str_to_check); | |
9089 | ||
9090 | /* Handle multiple target attributes separated by ','. */ | |
9091 | char *token = strtok (str_to_check, ","); | |
9092 | ||
9093 | unsigned int num_attrs = 0; | |
9094 | while (token) | |
9095 | { | |
9096 | num_attrs++; | |
9097 | if (!aarch64_process_one_target_attr (token, pragma_or_attr)) | |
9098 | { | |
9099 | error ("target %s %qs is invalid", pragma_or_attr, token); | |
9100 | return false; | |
9101 | } | |
9102 | ||
9103 | token = strtok (NULL, ","); | |
9104 | } | |
9105 | ||
9106 | if (num_attrs != num_commas + 1) | |
9107 | { | |
9108 | error ("malformed target %s list %qs", | |
9109 | pragma_or_attr, TREE_STRING_POINTER (args)); | |
9110 | return false; | |
9111 | } | |
9112 | ||
9113 | return true; | |
9114 | } | |
9115 | ||
9116 | /* Implement TARGET_OPTION_VALID_ATTRIBUTE_P. This is used to | |
9117 | process attribute ((target ("..."))). */ | |
9118 | ||
9119 | static bool | |
9120 | aarch64_option_valid_attribute_p (tree fndecl, tree, tree args, int) | |
9121 | { | |
9122 | struct cl_target_option cur_target; | |
9123 | bool ret; | |
9124 | tree old_optimize; | |
9125 | tree new_target, new_optimize; | |
9126 | tree existing_target = DECL_FUNCTION_SPECIFIC_TARGET (fndecl); | |
91d0e8de KT |
9127 | |
9128 | /* If what we're processing is the current pragma string then the | |
9129 | target option node is already stored in target_option_current_node | |
9130 | by aarch64_pragma_target_parse in aarch64-c.c. Use that to avoid | |
9131 | having to re-parse the string. This is especially useful to keep | |
9132 | arm_neon.h compile times down since that header contains a lot | |
9133 | of intrinsics enclosed in pragmas. */ | |
9134 | if (!existing_target && args == current_target_pragma) | |
9135 | { | |
9136 | DECL_FUNCTION_SPECIFIC_TARGET (fndecl) = target_option_current_node; | |
9137 | return true; | |
9138 | } | |
5a2c8331 KT |
9139 | tree func_optimize = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl); |
9140 | ||
9141 | old_optimize = build_optimization_node (&global_options); | |
9142 | func_optimize = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl); | |
9143 | ||
9144 | /* If the function changed the optimization levels as well as setting | |
9145 | target options, start with the optimizations specified. */ | |
9146 | if (func_optimize && func_optimize != old_optimize) | |
9147 | cl_optimization_restore (&global_options, | |
9148 | TREE_OPTIMIZATION (func_optimize)); | |
9149 | ||
9150 | /* Save the current target options to restore at the end. */ | |
9151 | cl_target_option_save (&cur_target, &global_options); | |
9152 | ||
9153 | /* If fndecl already has some target attributes applied to it, unpack | |
9154 | them so that we add this attribute on top of them, rather than | |
9155 | overwriting them. */ | |
9156 | if (existing_target) | |
9157 | { | |
9158 | struct cl_target_option *existing_options | |
9159 | = TREE_TARGET_OPTION (existing_target); | |
9160 | ||
9161 | if (existing_options) | |
9162 | cl_target_option_restore (&global_options, existing_options); | |
9163 | } | |
9164 | else | |
9165 | cl_target_option_restore (&global_options, | |
9166 | TREE_TARGET_OPTION (target_option_current_node)); | |
9167 | ||
9168 | ||
9169 | ret = aarch64_process_target_attr (args, "attribute"); | |
9170 | ||
9171 | /* Set up any additional state. */ | |
9172 | if (ret) | |
9173 | { | |
9174 | aarch64_override_options_internal (&global_options); | |
e95a988a KT |
9175 | /* Initialize SIMD builtins if we haven't already. |
9176 | Set current_target_pragma to NULL for the duration so that | |
9177 | the builtin initialization code doesn't try to tag the functions | |
9178 | being built with the attributes specified by any current pragma, thus | |
9179 | going into an infinite recursion. */ | |
9180 | if (TARGET_SIMD) | |
9181 | { | |
9182 | tree saved_current_target_pragma = current_target_pragma; | |
9183 | current_target_pragma = NULL; | |
9184 | aarch64_init_simd_builtins (); | |
9185 | current_target_pragma = saved_current_target_pragma; | |
9186 | } | |
5a2c8331 KT |
9187 | new_target = build_target_option_node (&global_options); |
9188 | } | |
9189 | else | |
9190 | new_target = NULL; | |
9191 | ||
9192 | new_optimize = build_optimization_node (&global_options); | |
9193 | ||
9194 | if (fndecl && ret) | |
9195 | { | |
9196 | DECL_FUNCTION_SPECIFIC_TARGET (fndecl) = new_target; | |
9197 | ||
9198 | if (old_optimize != new_optimize) | |
9199 | DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl) = new_optimize; | |
9200 | } | |
9201 | ||
9202 | cl_target_option_restore (&global_options, &cur_target); | |
9203 | ||
9204 | if (old_optimize != new_optimize) | |
9205 | cl_optimization_restore (&global_options, | |
9206 | TREE_OPTIMIZATION (old_optimize)); | |
9207 | return ret; | |
9208 | } | |
9209 | ||
1fd8d40c KT |
9210 | /* Helper for aarch64_can_inline_p. In the case where CALLER and CALLEE are |
9211 | tri-bool options (yes, no, don't care) and the default value is | |
9212 | DEF, determine whether to reject inlining. */ | |
9213 | ||
9214 | static bool | |
9215 | aarch64_tribools_ok_for_inlining_p (int caller, int callee, | |
9216 | int dont_care, int def) | |
9217 | { | |
9218 | /* If the callee doesn't care, always allow inlining. */ | |
9219 | if (callee == dont_care) | |
9220 | return true; | |
9221 | ||
9222 | /* If the caller doesn't care, always allow inlining. */ | |
9223 | if (caller == dont_care) | |
9224 | return true; | |
9225 | ||
9226 | /* Otherwise, allow inlining if either the callee and caller values | |
9227 | agree, or if the callee is using the default value. */ | |
9228 | return (callee == caller || callee == def); | |
9229 | } | |
9230 | ||
9231 | /* Implement TARGET_CAN_INLINE_P. Decide whether it is valid | |
9232 | to inline CALLEE into CALLER based on target-specific info. | |
9233 | Make sure that the caller and callee have compatible architectural | |
9234 | features. Then go through the other possible target attributes | |
9235 | and see if they can block inlining. Try not to reject always_inline | |
9236 | callees unless they are incompatible architecturally. */ | |
9237 | ||
9238 | static bool | |
9239 | aarch64_can_inline_p (tree caller, tree callee) | |
9240 | { | |
9241 | tree caller_tree = DECL_FUNCTION_SPECIFIC_TARGET (caller); | |
9242 | tree callee_tree = DECL_FUNCTION_SPECIFIC_TARGET (callee); | |
9243 | ||
9244 | /* If callee has no option attributes, then it is ok to inline. */ | |
9245 | if (!callee_tree) | |
9246 | return true; | |
9247 | ||
9248 | struct cl_target_option *caller_opts | |
9249 | = TREE_TARGET_OPTION (caller_tree ? caller_tree | |
9250 | : target_option_default_node); | |
9251 | ||
9252 | struct cl_target_option *callee_opts = TREE_TARGET_OPTION (callee_tree); | |
9253 | ||
9254 | ||
9255 | /* Callee's ISA flags should be a subset of the caller's. */ | |
9256 | if ((caller_opts->x_aarch64_isa_flags & callee_opts->x_aarch64_isa_flags) | |
9257 | != callee_opts->x_aarch64_isa_flags) | |
9258 | return false; | |
9259 | ||
9260 | /* Allow non-strict aligned functions inlining into strict | |
9261 | aligned ones. */ | |
9262 | if ((TARGET_STRICT_ALIGN_P (caller_opts->x_target_flags) | |
9263 | != TARGET_STRICT_ALIGN_P (callee_opts->x_target_flags)) | |
9264 | && !(!TARGET_STRICT_ALIGN_P (callee_opts->x_target_flags) | |
9265 | && TARGET_STRICT_ALIGN_P (caller_opts->x_target_flags))) | |
9266 | return false; | |
9267 | ||
9268 | bool always_inline = lookup_attribute ("always_inline", | |
9269 | DECL_ATTRIBUTES (callee)); | |
9270 | ||
9271 | /* If the architectural features match up and the callee is always_inline | |
9272 | then the other attributes don't matter. */ | |
9273 | if (always_inline) | |
9274 | return true; | |
9275 | ||
9276 | if (caller_opts->x_aarch64_cmodel_var | |
9277 | != callee_opts->x_aarch64_cmodel_var) | |
9278 | return false; | |
9279 | ||
9280 | if (caller_opts->x_aarch64_tls_dialect | |
9281 | != callee_opts->x_aarch64_tls_dialect) | |
9282 | return false; | |
9283 | ||
9284 | /* Honour explicit requests to workaround errata. */ | |
9285 | if (!aarch64_tribools_ok_for_inlining_p ( | |
9286 | caller_opts->x_aarch64_fix_a53_err835769, | |
9287 | callee_opts->x_aarch64_fix_a53_err835769, | |
9288 | 2, TARGET_FIX_ERR_A53_835769_DEFAULT)) | |
9289 | return false; | |
9290 | ||
48bb1a55 CL |
9291 | if (!aarch64_tribools_ok_for_inlining_p ( |
9292 | caller_opts->x_aarch64_fix_a53_err843419, | |
9293 | callee_opts->x_aarch64_fix_a53_err843419, | |
9294 | 2, TARGET_FIX_ERR_A53_843419)) | |
9295 | return false; | |
9296 | ||
1fd8d40c KT |
9297 | /* If the user explicitly specified -momit-leaf-frame-pointer for the |
9298 | caller and calle and they don't match up, reject inlining. */ | |
9299 | if (!aarch64_tribools_ok_for_inlining_p ( | |
9300 | caller_opts->x_flag_omit_leaf_frame_pointer, | |
9301 | callee_opts->x_flag_omit_leaf_frame_pointer, | |
9302 | 2, 1)) | |
9303 | return false; | |
9304 | ||
9305 | /* If the callee has specific tuning overrides, respect them. */ | |
9306 | if (callee_opts->x_aarch64_override_tune_string != NULL | |
9307 | && caller_opts->x_aarch64_override_tune_string == NULL) | |
9308 | return false; | |
9309 | ||
9310 | /* If the user specified tuning override strings for the | |
9311 | caller and callee and they don't match up, reject inlining. | |
9312 | We just do a string compare here, we don't analyze the meaning | |
9313 | of the string, as it would be too costly for little gain. */ | |
9314 | if (callee_opts->x_aarch64_override_tune_string | |
9315 | && caller_opts->x_aarch64_override_tune_string | |
9316 | && (strcmp (callee_opts->x_aarch64_override_tune_string, | |
9317 | caller_opts->x_aarch64_override_tune_string) != 0)) | |
9318 | return false; | |
9319 | ||
9320 | return true; | |
9321 | } | |
9322 | ||
43e9d192 IB |
9323 | /* Return true if SYMBOL_REF X binds locally. */ |
9324 | ||
9325 | static bool | |
9326 | aarch64_symbol_binds_local_p (const_rtx x) | |
9327 | { | |
9328 | return (SYMBOL_REF_DECL (x) | |
9329 | ? targetm.binds_local_p (SYMBOL_REF_DECL (x)) | |
9330 | : SYMBOL_REF_LOCAL_P (x)); | |
9331 | } | |
9332 | ||
9333 | /* Return true if SYMBOL_REF X is thread local */ | |
9334 | static bool | |
9335 | aarch64_tls_symbol_p (rtx x) | |
9336 | { | |
9337 | if (! TARGET_HAVE_TLS) | |
9338 | return false; | |
9339 | ||
9340 | if (GET_CODE (x) != SYMBOL_REF) | |
9341 | return false; | |
9342 | ||
9343 | return SYMBOL_REF_TLS_MODEL (x) != 0; | |
9344 | } | |
9345 | ||
9346 | /* Classify a TLS symbol into one of the TLS kinds. */ | |
9347 | enum aarch64_symbol_type | |
9348 | aarch64_classify_tls_symbol (rtx x) | |
9349 | { | |
9350 | enum tls_model tls_kind = tls_symbolic_operand_type (x); | |
9351 | ||
9352 | switch (tls_kind) | |
9353 | { | |
9354 | case TLS_MODEL_GLOBAL_DYNAMIC: | |
9355 | case TLS_MODEL_LOCAL_DYNAMIC: | |
9356 | return TARGET_TLS_DESC ? SYMBOL_SMALL_TLSDESC : SYMBOL_SMALL_TLSGD; | |
9357 | ||
9358 | case TLS_MODEL_INITIAL_EXEC: | |
5ae7caad JW |
9359 | switch (aarch64_cmodel) |
9360 | { | |
9361 | case AARCH64_CMODEL_TINY: | |
9362 | case AARCH64_CMODEL_TINY_PIC: | |
9363 | return SYMBOL_TINY_TLSIE; | |
9364 | default: | |
79496620 | 9365 | return SYMBOL_SMALL_TLSIE; |
5ae7caad | 9366 | } |
43e9d192 IB |
9367 | |
9368 | case TLS_MODEL_LOCAL_EXEC: | |
cbf5629e JW |
9369 | if (aarch64_tls_size == 12) |
9370 | return SYMBOL_TLSLE12; | |
9371 | else if (aarch64_tls_size == 24) | |
9372 | return SYMBOL_TLSLE24; | |
9373 | else if (aarch64_tls_size == 32) | |
9374 | return SYMBOL_TLSLE32; | |
9375 | else if (aarch64_tls_size == 48) | |
9376 | return SYMBOL_TLSLE48; | |
9377 | else | |
9378 | gcc_unreachable (); | |
43e9d192 IB |
9379 | |
9380 | case TLS_MODEL_EMULATED: | |
9381 | case TLS_MODEL_NONE: | |
9382 | return SYMBOL_FORCE_TO_MEM; | |
9383 | ||
9384 | default: | |
9385 | gcc_unreachable (); | |
9386 | } | |
9387 | } | |
9388 | ||
9389 | /* Return the method that should be used to access SYMBOL_REF or | |
a6e0bfa7 | 9390 | LABEL_REF X. */ |
17f4d4bf | 9391 | |
43e9d192 | 9392 | enum aarch64_symbol_type |
a6e0bfa7 | 9393 | aarch64_classify_symbol (rtx x, rtx offset) |
43e9d192 IB |
9394 | { |
9395 | if (GET_CODE (x) == LABEL_REF) | |
9396 | { | |
9397 | switch (aarch64_cmodel) | |
9398 | { | |
9399 | case AARCH64_CMODEL_LARGE: | |
9400 | return SYMBOL_FORCE_TO_MEM; | |
9401 | ||
9402 | case AARCH64_CMODEL_TINY_PIC: | |
9403 | case AARCH64_CMODEL_TINY: | |
a5350ddc CSS |
9404 | return SYMBOL_TINY_ABSOLUTE; |
9405 | ||
1b1e81f8 | 9406 | case AARCH64_CMODEL_SMALL_SPIC: |
43e9d192 IB |
9407 | case AARCH64_CMODEL_SMALL_PIC: |
9408 | case AARCH64_CMODEL_SMALL: | |
9409 | return SYMBOL_SMALL_ABSOLUTE; | |
9410 | ||
9411 | default: | |
9412 | gcc_unreachable (); | |
9413 | } | |
9414 | } | |
9415 | ||
17f4d4bf | 9416 | if (GET_CODE (x) == SYMBOL_REF) |
43e9d192 | 9417 | { |
43e9d192 IB |
9418 | if (aarch64_tls_symbol_p (x)) |
9419 | return aarch64_classify_tls_symbol (x); | |
9420 | ||
17f4d4bf CSS |
9421 | switch (aarch64_cmodel) |
9422 | { | |
9423 | case AARCH64_CMODEL_TINY: | |
15f6e0da | 9424 | /* When we retrieve symbol + offset address, we have to make sure |
f8b756b7 TB |
9425 | the offset does not cause overflow of the final address. But |
9426 | we have no way of knowing the address of symbol at compile time | |
9427 | so we can't accurately say if the distance between the PC and | |
9428 | symbol + offset is outside the addressible range of +/-1M in the | |
9429 | TINY code model. So we rely on images not being greater than | |
9430 | 1M and cap the offset at 1M and anything beyond 1M will have to | |
15f6e0da RR |
9431 | be loaded using an alternative mechanism. Furthermore if the |
9432 | symbol is a weak reference to something that isn't known to | |
9433 | resolve to a symbol in this module, then force to memory. */ | |
9434 | if ((SYMBOL_REF_WEAK (x) | |
9435 | && !aarch64_symbol_binds_local_p (x)) | |
f8b756b7 | 9436 | || INTVAL (offset) < -1048575 || INTVAL (offset) > 1048575) |
a5350ddc CSS |
9437 | return SYMBOL_FORCE_TO_MEM; |
9438 | return SYMBOL_TINY_ABSOLUTE; | |
9439 | ||
17f4d4bf | 9440 | case AARCH64_CMODEL_SMALL: |
f8b756b7 TB |
9441 | /* Same reasoning as the tiny code model, but the offset cap here is |
9442 | 4G. */ | |
15f6e0da RR |
9443 | if ((SYMBOL_REF_WEAK (x) |
9444 | && !aarch64_symbol_binds_local_p (x)) | |
3ff5d1f0 TB |
9445 | || !IN_RANGE (INTVAL (offset), HOST_WIDE_INT_C (-4294967263), |
9446 | HOST_WIDE_INT_C (4294967264))) | |
17f4d4bf CSS |
9447 | return SYMBOL_FORCE_TO_MEM; |
9448 | return SYMBOL_SMALL_ABSOLUTE; | |
43e9d192 | 9449 | |
17f4d4bf | 9450 | case AARCH64_CMODEL_TINY_PIC: |
38e6c9a6 | 9451 | if (!aarch64_symbol_binds_local_p (x)) |
87dd8ab0 | 9452 | return SYMBOL_TINY_GOT; |
38e6c9a6 MS |
9453 | return SYMBOL_TINY_ABSOLUTE; |
9454 | ||
1b1e81f8 | 9455 | case AARCH64_CMODEL_SMALL_SPIC: |
17f4d4bf CSS |
9456 | case AARCH64_CMODEL_SMALL_PIC: |
9457 | if (!aarch64_symbol_binds_local_p (x)) | |
1b1e81f8 JW |
9458 | return (aarch64_cmodel == AARCH64_CMODEL_SMALL_SPIC |
9459 | ? SYMBOL_SMALL_GOT_28K : SYMBOL_SMALL_GOT_4G); | |
17f4d4bf | 9460 | return SYMBOL_SMALL_ABSOLUTE; |
43e9d192 | 9461 | |
9ee6540a WD |
9462 | case AARCH64_CMODEL_LARGE: |
9463 | /* This is alright even in PIC code as the constant | |
9464 | pool reference is always PC relative and within | |
9465 | the same translation unit. */ | |
9466 | if (CONSTANT_POOL_ADDRESS_P (x)) | |
9467 | return SYMBOL_SMALL_ABSOLUTE; | |
9468 | else | |
9469 | return SYMBOL_FORCE_TO_MEM; | |
9470 | ||
17f4d4bf CSS |
9471 | default: |
9472 | gcc_unreachable (); | |
9473 | } | |
43e9d192 | 9474 | } |
17f4d4bf | 9475 | |
43e9d192 IB |
9476 | /* By default push everything into the constant pool. */ |
9477 | return SYMBOL_FORCE_TO_MEM; | |
9478 | } | |
9479 | ||
43e9d192 IB |
9480 | bool |
9481 | aarch64_constant_address_p (rtx x) | |
9482 | { | |
9483 | return (CONSTANT_P (x) && memory_address_p (DImode, x)); | |
9484 | } | |
9485 | ||
9486 | bool | |
9487 | aarch64_legitimate_pic_operand_p (rtx x) | |
9488 | { | |
9489 | if (GET_CODE (x) == SYMBOL_REF | |
9490 | || (GET_CODE (x) == CONST | |
9491 | && GET_CODE (XEXP (x, 0)) == PLUS | |
9492 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF)) | |
9493 | return false; | |
9494 | ||
9495 | return true; | |
9496 | } | |
9497 | ||
3520f7cc JG |
9498 | /* Return true if X holds either a quarter-precision or |
9499 | floating-point +0.0 constant. */ | |
9500 | static bool | |
ef4bddc2 | 9501 | aarch64_valid_floating_const (machine_mode mode, rtx x) |
3520f7cc JG |
9502 | { |
9503 | if (!CONST_DOUBLE_P (x)) | |
9504 | return false; | |
9505 | ||
6a0f8c01 JW |
9506 | if (aarch64_float_const_zero_rtx_p (x)) |
9507 | return true; | |
9508 | ||
9509 | /* We only handle moving 0.0 to a TFmode register. */ | |
3520f7cc JG |
9510 | if (!(mode == SFmode || mode == DFmode)) |
9511 | return false; | |
9512 | ||
3520f7cc JG |
9513 | return aarch64_float_const_representable_p (x); |
9514 | } | |
9515 | ||
43e9d192 | 9516 | static bool |
ef4bddc2 | 9517 | aarch64_legitimate_constant_p (machine_mode mode, rtx x) |
43e9d192 IB |
9518 | { |
9519 | /* Do not allow vector struct mode constants. We could support | |
9520 | 0 and -1 easily, but they need support in aarch64-simd.md. */ | |
9521 | if (TARGET_SIMD && aarch64_vect_struct_mode_p (mode)) | |
9522 | return false; | |
9523 | ||
9524 | /* This could probably go away because | |
9525 | we now decompose CONST_INTs according to expand_mov_immediate. */ | |
9526 | if ((GET_CODE (x) == CONST_VECTOR | |
48063b9d | 9527 | && aarch64_simd_valid_immediate (x, mode, false, NULL)) |
3520f7cc JG |
9528 | || CONST_INT_P (x) || aarch64_valid_floating_const (mode, x)) |
9529 | return !targetm.cannot_force_const_mem (mode, x); | |
43e9d192 IB |
9530 | |
9531 | if (GET_CODE (x) == HIGH | |
9532 | && aarch64_valid_symref (XEXP (x, 0), GET_MODE (XEXP (x, 0)))) | |
9533 | return true; | |
9534 | ||
9535 | return aarch64_constant_address_p (x); | |
9536 | } | |
9537 | ||
a5bc806c | 9538 | rtx |
43e9d192 IB |
9539 | aarch64_load_tp (rtx target) |
9540 | { | |
9541 | if (!target | |
9542 | || GET_MODE (target) != Pmode | |
9543 | || !register_operand (target, Pmode)) | |
9544 | target = gen_reg_rtx (Pmode); | |
9545 | ||
9546 | /* Can return in any reg. */ | |
9547 | emit_insn (gen_aarch64_load_tp_hard (target)); | |
9548 | return target; | |
9549 | } | |
9550 | ||
43e9d192 IB |
9551 | /* On AAPCS systems, this is the "struct __va_list". */ |
9552 | static GTY(()) tree va_list_type; | |
9553 | ||
9554 | /* Implement TARGET_BUILD_BUILTIN_VA_LIST. | |
9555 | Return the type to use as __builtin_va_list. | |
9556 | ||
9557 | AAPCS64 \S 7.1.4 requires that va_list be a typedef for a type defined as: | |
9558 | ||
9559 | struct __va_list | |
9560 | { | |
9561 | void *__stack; | |
9562 | void *__gr_top; | |
9563 | void *__vr_top; | |
9564 | int __gr_offs; | |
9565 | int __vr_offs; | |
9566 | }; */ | |
9567 | ||
9568 | static tree | |
9569 | aarch64_build_builtin_va_list (void) | |
9570 | { | |
9571 | tree va_list_name; | |
9572 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
9573 | ||
9574 | /* Create the type. */ | |
9575 | va_list_type = lang_hooks.types.make_type (RECORD_TYPE); | |
9576 | /* Give it the required name. */ | |
9577 | va_list_name = build_decl (BUILTINS_LOCATION, | |
9578 | TYPE_DECL, | |
9579 | get_identifier ("__va_list"), | |
9580 | va_list_type); | |
9581 | DECL_ARTIFICIAL (va_list_name) = 1; | |
9582 | TYPE_NAME (va_list_type) = va_list_name; | |
665c56c6 | 9583 | TYPE_STUB_DECL (va_list_type) = va_list_name; |
43e9d192 IB |
9584 | |
9585 | /* Create the fields. */ | |
9586 | f_stack = build_decl (BUILTINS_LOCATION, | |
9587 | FIELD_DECL, get_identifier ("__stack"), | |
9588 | ptr_type_node); | |
9589 | f_grtop = build_decl (BUILTINS_LOCATION, | |
9590 | FIELD_DECL, get_identifier ("__gr_top"), | |
9591 | ptr_type_node); | |
9592 | f_vrtop = build_decl (BUILTINS_LOCATION, | |
9593 | FIELD_DECL, get_identifier ("__vr_top"), | |
9594 | ptr_type_node); | |
9595 | f_groff = build_decl (BUILTINS_LOCATION, | |
9596 | FIELD_DECL, get_identifier ("__gr_offs"), | |
9597 | integer_type_node); | |
9598 | f_vroff = build_decl (BUILTINS_LOCATION, | |
9599 | FIELD_DECL, get_identifier ("__vr_offs"), | |
9600 | integer_type_node); | |
9601 | ||
88e3bdd1 | 9602 | /* Tell tree-stdarg pass about our internal offset fields. |
3fd6b9cc JW |
9603 | NOTE: va_list_gpr/fpr_counter_field are only used for tree comparision |
9604 | purpose to identify whether the code is updating va_list internal | |
9605 | offset fields through irregular way. */ | |
9606 | va_list_gpr_counter_field = f_groff; | |
9607 | va_list_fpr_counter_field = f_vroff; | |
9608 | ||
43e9d192 IB |
9609 | DECL_ARTIFICIAL (f_stack) = 1; |
9610 | DECL_ARTIFICIAL (f_grtop) = 1; | |
9611 | DECL_ARTIFICIAL (f_vrtop) = 1; | |
9612 | DECL_ARTIFICIAL (f_groff) = 1; | |
9613 | DECL_ARTIFICIAL (f_vroff) = 1; | |
9614 | ||
9615 | DECL_FIELD_CONTEXT (f_stack) = va_list_type; | |
9616 | DECL_FIELD_CONTEXT (f_grtop) = va_list_type; | |
9617 | DECL_FIELD_CONTEXT (f_vrtop) = va_list_type; | |
9618 | DECL_FIELD_CONTEXT (f_groff) = va_list_type; | |
9619 | DECL_FIELD_CONTEXT (f_vroff) = va_list_type; | |
9620 | ||
9621 | TYPE_FIELDS (va_list_type) = f_stack; | |
9622 | DECL_CHAIN (f_stack) = f_grtop; | |
9623 | DECL_CHAIN (f_grtop) = f_vrtop; | |
9624 | DECL_CHAIN (f_vrtop) = f_groff; | |
9625 | DECL_CHAIN (f_groff) = f_vroff; | |
9626 | ||
9627 | /* Compute its layout. */ | |
9628 | layout_type (va_list_type); | |
9629 | ||
9630 | return va_list_type; | |
9631 | } | |
9632 | ||
9633 | /* Implement TARGET_EXPAND_BUILTIN_VA_START. */ | |
9634 | static void | |
9635 | aarch64_expand_builtin_va_start (tree valist, rtx nextarg ATTRIBUTE_UNUSED) | |
9636 | { | |
9637 | const CUMULATIVE_ARGS *cum; | |
9638 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
9639 | tree stack, grtop, vrtop, groff, vroff; | |
9640 | tree t; | |
88e3bdd1 JW |
9641 | int gr_save_area_size = cfun->va_list_gpr_size; |
9642 | int vr_save_area_size = cfun->va_list_fpr_size; | |
43e9d192 IB |
9643 | int vr_offset; |
9644 | ||
9645 | cum = &crtl->args.info; | |
88e3bdd1 JW |
9646 | if (cfun->va_list_gpr_size) |
9647 | gr_save_area_size = MIN ((NUM_ARG_REGS - cum->aapcs_ncrn) * UNITS_PER_WORD, | |
9648 | cfun->va_list_gpr_size); | |
9649 | if (cfun->va_list_fpr_size) | |
9650 | vr_save_area_size = MIN ((NUM_FP_ARG_REGS - cum->aapcs_nvrn) | |
9651 | * UNITS_PER_VREG, cfun->va_list_fpr_size); | |
43e9d192 | 9652 | |
d5726973 | 9653 | if (!TARGET_FLOAT) |
43e9d192 | 9654 | { |
261fb553 | 9655 | gcc_assert (cum->aapcs_nvrn == 0); |
43e9d192 IB |
9656 | vr_save_area_size = 0; |
9657 | } | |
9658 | ||
9659 | f_stack = TYPE_FIELDS (va_list_type_node); | |
9660 | f_grtop = DECL_CHAIN (f_stack); | |
9661 | f_vrtop = DECL_CHAIN (f_grtop); | |
9662 | f_groff = DECL_CHAIN (f_vrtop); | |
9663 | f_vroff = DECL_CHAIN (f_groff); | |
9664 | ||
9665 | stack = build3 (COMPONENT_REF, TREE_TYPE (f_stack), valist, f_stack, | |
9666 | NULL_TREE); | |
9667 | grtop = build3 (COMPONENT_REF, TREE_TYPE (f_grtop), valist, f_grtop, | |
9668 | NULL_TREE); | |
9669 | vrtop = build3 (COMPONENT_REF, TREE_TYPE (f_vrtop), valist, f_vrtop, | |
9670 | NULL_TREE); | |
9671 | groff = build3 (COMPONENT_REF, TREE_TYPE (f_groff), valist, f_groff, | |
9672 | NULL_TREE); | |
9673 | vroff = build3 (COMPONENT_REF, TREE_TYPE (f_vroff), valist, f_vroff, | |
9674 | NULL_TREE); | |
9675 | ||
9676 | /* Emit code to initialize STACK, which points to the next varargs stack | |
9677 | argument. CUM->AAPCS_STACK_SIZE gives the number of stack words used | |
9678 | by named arguments. STACK is 8-byte aligned. */ | |
9679 | t = make_tree (TREE_TYPE (stack), virtual_incoming_args_rtx); | |
9680 | if (cum->aapcs_stack_size > 0) | |
9681 | t = fold_build_pointer_plus_hwi (t, cum->aapcs_stack_size * UNITS_PER_WORD); | |
9682 | t = build2 (MODIFY_EXPR, TREE_TYPE (stack), stack, t); | |
9683 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
9684 | ||
9685 | /* Emit code to initialize GRTOP, the top of the GR save area. | |
9686 | virtual_incoming_args_rtx should have been 16 byte aligned. */ | |
9687 | t = make_tree (TREE_TYPE (grtop), virtual_incoming_args_rtx); | |
9688 | t = build2 (MODIFY_EXPR, TREE_TYPE (grtop), grtop, t); | |
9689 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
9690 | ||
9691 | /* Emit code to initialize VRTOP, the top of the VR save area. | |
9692 | This address is gr_save_area_bytes below GRTOP, rounded | |
9693 | down to the next 16-byte boundary. */ | |
9694 | t = make_tree (TREE_TYPE (vrtop), virtual_incoming_args_rtx); | |
4f59f9f2 UB |
9695 | vr_offset = ROUND_UP (gr_save_area_size, |
9696 | STACK_BOUNDARY / BITS_PER_UNIT); | |
43e9d192 IB |
9697 | |
9698 | if (vr_offset) | |
9699 | t = fold_build_pointer_plus_hwi (t, -vr_offset); | |
9700 | t = build2 (MODIFY_EXPR, TREE_TYPE (vrtop), vrtop, t); | |
9701 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
9702 | ||
9703 | /* Emit code to initialize GROFF, the offset from GRTOP of the | |
9704 | next GPR argument. */ | |
9705 | t = build2 (MODIFY_EXPR, TREE_TYPE (groff), groff, | |
9706 | build_int_cst (TREE_TYPE (groff), -gr_save_area_size)); | |
9707 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
9708 | ||
9709 | /* Likewise emit code to initialize VROFF, the offset from FTOP | |
9710 | of the next VR argument. */ | |
9711 | t = build2 (MODIFY_EXPR, TREE_TYPE (vroff), vroff, | |
9712 | build_int_cst (TREE_TYPE (vroff), -vr_save_area_size)); | |
9713 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
9714 | } | |
9715 | ||
9716 | /* Implement TARGET_GIMPLIFY_VA_ARG_EXPR. */ | |
9717 | ||
9718 | static tree | |
9719 | aarch64_gimplify_va_arg_expr (tree valist, tree type, gimple_seq *pre_p, | |
9720 | gimple_seq *post_p ATTRIBUTE_UNUSED) | |
9721 | { | |
9722 | tree addr; | |
9723 | bool indirect_p; | |
9724 | bool is_ha; /* is HFA or HVA. */ | |
9725 | bool dw_align; /* double-word align. */ | |
ef4bddc2 | 9726 | machine_mode ag_mode = VOIDmode; |
43e9d192 | 9727 | int nregs; |
ef4bddc2 | 9728 | machine_mode mode; |
43e9d192 IB |
9729 | |
9730 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
9731 | tree stack, f_top, f_off, off, arg, roundup, on_stack; | |
9732 | HOST_WIDE_INT size, rsize, adjust, align; | |
9733 | tree t, u, cond1, cond2; | |
9734 | ||
9735 | indirect_p = pass_by_reference (NULL, TYPE_MODE (type), type, false); | |
9736 | if (indirect_p) | |
9737 | type = build_pointer_type (type); | |
9738 | ||
9739 | mode = TYPE_MODE (type); | |
9740 | ||
9741 | f_stack = TYPE_FIELDS (va_list_type_node); | |
9742 | f_grtop = DECL_CHAIN (f_stack); | |
9743 | f_vrtop = DECL_CHAIN (f_grtop); | |
9744 | f_groff = DECL_CHAIN (f_vrtop); | |
9745 | f_vroff = DECL_CHAIN (f_groff); | |
9746 | ||
9747 | stack = build3 (COMPONENT_REF, TREE_TYPE (f_stack), unshare_expr (valist), | |
9748 | f_stack, NULL_TREE); | |
9749 | size = int_size_in_bytes (type); | |
9750 | align = aarch64_function_arg_alignment (mode, type) / BITS_PER_UNIT; | |
9751 | ||
9752 | dw_align = false; | |
9753 | adjust = 0; | |
9754 | if (aarch64_vfp_is_call_or_return_candidate (mode, | |
9755 | type, | |
9756 | &ag_mode, | |
9757 | &nregs, | |
9758 | &is_ha)) | |
9759 | { | |
9760 | /* TYPE passed in fp/simd registers. */ | |
d5726973 | 9761 | if (!TARGET_FLOAT) |
261fb553 | 9762 | aarch64_err_no_fpadvsimd (mode, "varargs"); |
43e9d192 IB |
9763 | |
9764 | f_top = build3 (COMPONENT_REF, TREE_TYPE (f_vrtop), | |
9765 | unshare_expr (valist), f_vrtop, NULL_TREE); | |
9766 | f_off = build3 (COMPONENT_REF, TREE_TYPE (f_vroff), | |
9767 | unshare_expr (valist), f_vroff, NULL_TREE); | |
9768 | ||
9769 | rsize = nregs * UNITS_PER_VREG; | |
9770 | ||
9771 | if (is_ha) | |
9772 | { | |
9773 | if (BYTES_BIG_ENDIAN && GET_MODE_SIZE (ag_mode) < UNITS_PER_VREG) | |
9774 | adjust = UNITS_PER_VREG - GET_MODE_SIZE (ag_mode); | |
9775 | } | |
9776 | else if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
9777 | && size < UNITS_PER_VREG) | |
9778 | { | |
9779 | adjust = UNITS_PER_VREG - size; | |
9780 | } | |
9781 | } | |
9782 | else | |
9783 | { | |
9784 | /* TYPE passed in general registers. */ | |
9785 | f_top = build3 (COMPONENT_REF, TREE_TYPE (f_grtop), | |
9786 | unshare_expr (valist), f_grtop, NULL_TREE); | |
9787 | f_off = build3 (COMPONENT_REF, TREE_TYPE (f_groff), | |
9788 | unshare_expr (valist), f_groff, NULL_TREE); | |
4f59f9f2 | 9789 | rsize = ROUND_UP (size, UNITS_PER_WORD); |
43e9d192 IB |
9790 | nregs = rsize / UNITS_PER_WORD; |
9791 | ||
9792 | if (align > 8) | |
9793 | dw_align = true; | |
9794 | ||
9795 | if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
9796 | && size < UNITS_PER_WORD) | |
9797 | { | |
9798 | adjust = UNITS_PER_WORD - size; | |
9799 | } | |
9800 | } | |
9801 | ||
9802 | /* Get a local temporary for the field value. */ | |
9803 | off = get_initialized_tmp_var (f_off, pre_p, NULL); | |
9804 | ||
9805 | /* Emit code to branch if off >= 0. */ | |
9806 | t = build2 (GE_EXPR, boolean_type_node, off, | |
9807 | build_int_cst (TREE_TYPE (off), 0)); | |
9808 | cond1 = build3 (COND_EXPR, ptr_type_node, t, NULL_TREE, NULL_TREE); | |
9809 | ||
9810 | if (dw_align) | |
9811 | { | |
9812 | /* Emit: offs = (offs + 15) & -16. */ | |
9813 | t = build2 (PLUS_EXPR, TREE_TYPE (off), off, | |
9814 | build_int_cst (TREE_TYPE (off), 15)); | |
9815 | t = build2 (BIT_AND_EXPR, TREE_TYPE (off), t, | |
9816 | build_int_cst (TREE_TYPE (off), -16)); | |
9817 | roundup = build2 (MODIFY_EXPR, TREE_TYPE (off), off, t); | |
9818 | } | |
9819 | else | |
9820 | roundup = NULL; | |
9821 | ||
9822 | /* Update ap.__[g|v]r_offs */ | |
9823 | t = build2 (PLUS_EXPR, TREE_TYPE (off), off, | |
9824 | build_int_cst (TREE_TYPE (off), rsize)); | |
9825 | t = build2 (MODIFY_EXPR, TREE_TYPE (f_off), unshare_expr (f_off), t); | |
9826 | ||
9827 | /* String up. */ | |
9828 | if (roundup) | |
9829 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), roundup, t); | |
9830 | ||
9831 | /* [cond2] if (ap.__[g|v]r_offs > 0) */ | |
9832 | u = build2 (GT_EXPR, boolean_type_node, unshare_expr (f_off), | |
9833 | build_int_cst (TREE_TYPE (f_off), 0)); | |
9834 | cond2 = build3 (COND_EXPR, ptr_type_node, u, NULL_TREE, NULL_TREE); | |
9835 | ||
9836 | /* String up: make sure the assignment happens before the use. */ | |
9837 | t = build2 (COMPOUND_EXPR, TREE_TYPE (cond2), t, cond2); | |
9838 | COND_EXPR_ELSE (cond1) = t; | |
9839 | ||
9840 | /* Prepare the trees handling the argument that is passed on the stack; | |
9841 | the top level node will store in ON_STACK. */ | |
9842 | arg = get_initialized_tmp_var (stack, pre_p, NULL); | |
9843 | if (align > 8) | |
9844 | { | |
9845 | /* if (alignof(type) > 8) (arg = arg + 15) & -16; */ | |
9846 | t = fold_convert (intDI_type_node, arg); | |
9847 | t = build2 (PLUS_EXPR, TREE_TYPE (t), t, | |
9848 | build_int_cst (TREE_TYPE (t), 15)); | |
9849 | t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t, | |
9850 | build_int_cst (TREE_TYPE (t), -16)); | |
9851 | t = fold_convert (TREE_TYPE (arg), t); | |
9852 | roundup = build2 (MODIFY_EXPR, TREE_TYPE (arg), arg, t); | |
9853 | } | |
9854 | else | |
9855 | roundup = NULL; | |
9856 | /* Advance ap.__stack */ | |
9857 | t = fold_convert (intDI_type_node, arg); | |
9858 | t = build2 (PLUS_EXPR, TREE_TYPE (t), t, | |
9859 | build_int_cst (TREE_TYPE (t), size + 7)); | |
9860 | t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t, | |
9861 | build_int_cst (TREE_TYPE (t), -8)); | |
9862 | t = fold_convert (TREE_TYPE (arg), t); | |
9863 | t = build2 (MODIFY_EXPR, TREE_TYPE (stack), unshare_expr (stack), t); | |
9864 | /* String up roundup and advance. */ | |
9865 | if (roundup) | |
9866 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), roundup, t); | |
9867 | /* String up with arg */ | |
9868 | on_stack = build2 (COMPOUND_EXPR, TREE_TYPE (arg), t, arg); | |
9869 | /* Big-endianness related address adjustment. */ | |
9870 | if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
9871 | && size < UNITS_PER_WORD) | |
9872 | { | |
9873 | t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (arg), arg, | |
9874 | size_int (UNITS_PER_WORD - size)); | |
9875 | on_stack = build2 (COMPOUND_EXPR, TREE_TYPE (arg), on_stack, t); | |
9876 | } | |
9877 | ||
9878 | COND_EXPR_THEN (cond1) = unshare_expr (on_stack); | |
9879 | COND_EXPR_THEN (cond2) = unshare_expr (on_stack); | |
9880 | ||
9881 | /* Adjustment to OFFSET in the case of BIG_ENDIAN. */ | |
9882 | t = off; | |
9883 | if (adjust) | |
9884 | t = build2 (PREINCREMENT_EXPR, TREE_TYPE (off), off, | |
9885 | build_int_cst (TREE_TYPE (off), adjust)); | |
9886 | ||
9887 | t = fold_convert (sizetype, t); | |
9888 | t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (f_top), f_top, t); | |
9889 | ||
9890 | if (is_ha) | |
9891 | { | |
9892 | /* type ha; // treat as "struct {ftype field[n];}" | |
9893 | ... [computing offs] | |
9894 | for (i = 0; i <nregs; ++i, offs += 16) | |
9895 | ha.field[i] = *((ftype *)(ap.__vr_top + offs)); | |
9896 | return ha; */ | |
9897 | int i; | |
9898 | tree tmp_ha, field_t, field_ptr_t; | |
9899 | ||
9900 | /* Declare a local variable. */ | |
9901 | tmp_ha = create_tmp_var_raw (type, "ha"); | |
9902 | gimple_add_tmp_var (tmp_ha); | |
9903 | ||
9904 | /* Establish the base type. */ | |
9905 | switch (ag_mode) | |
9906 | { | |
9907 | case SFmode: | |
9908 | field_t = float_type_node; | |
9909 | field_ptr_t = float_ptr_type_node; | |
9910 | break; | |
9911 | case DFmode: | |
9912 | field_t = double_type_node; | |
9913 | field_ptr_t = double_ptr_type_node; | |
9914 | break; | |
9915 | case TFmode: | |
9916 | field_t = long_double_type_node; | |
9917 | field_ptr_t = long_double_ptr_type_node; | |
9918 | break; | |
9919 | /* The half precision and quad precision are not fully supported yet. Enable | |
9920 | the following code after the support is complete. Need to find the correct | |
9921 | type node for __fp16 *. */ | |
9922 | #if 0 | |
9923 | case HFmode: | |
9924 | field_t = float_type_node; | |
9925 | field_ptr_t = float_ptr_type_node; | |
9926 | break; | |
9927 | #endif | |
9928 | case V2SImode: | |
9929 | case V4SImode: | |
9930 | { | |
9931 | tree innertype = make_signed_type (GET_MODE_PRECISION (SImode)); | |
9932 | field_t = build_vector_type_for_mode (innertype, ag_mode); | |
9933 | field_ptr_t = build_pointer_type (field_t); | |
9934 | } | |
9935 | break; | |
9936 | default: | |
9937 | gcc_assert (0); | |
9938 | } | |
9939 | ||
9940 | /* *(field_ptr_t)&ha = *((field_ptr_t)vr_saved_area */ | |
9941 | tmp_ha = build1 (ADDR_EXPR, field_ptr_t, tmp_ha); | |
9942 | addr = t; | |
9943 | t = fold_convert (field_ptr_t, addr); | |
9944 | t = build2 (MODIFY_EXPR, field_t, | |
9945 | build1 (INDIRECT_REF, field_t, tmp_ha), | |
9946 | build1 (INDIRECT_REF, field_t, t)); | |
9947 | ||
9948 | /* ha.field[i] = *((field_ptr_t)vr_saved_area + i) */ | |
9949 | for (i = 1; i < nregs; ++i) | |
9950 | { | |
9951 | addr = fold_build_pointer_plus_hwi (addr, UNITS_PER_VREG); | |
9952 | u = fold_convert (field_ptr_t, addr); | |
9953 | u = build2 (MODIFY_EXPR, field_t, | |
9954 | build2 (MEM_REF, field_t, tmp_ha, | |
9955 | build_int_cst (field_ptr_t, | |
9956 | (i * | |
9957 | int_size_in_bytes (field_t)))), | |
9958 | build1 (INDIRECT_REF, field_t, u)); | |
9959 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), t, u); | |
9960 | } | |
9961 | ||
9962 | u = fold_convert (TREE_TYPE (f_top), tmp_ha); | |
9963 | t = build2 (COMPOUND_EXPR, TREE_TYPE (f_top), t, u); | |
9964 | } | |
9965 | ||
9966 | COND_EXPR_ELSE (cond2) = t; | |
9967 | addr = fold_convert (build_pointer_type (type), cond1); | |
9968 | addr = build_va_arg_indirect_ref (addr); | |
9969 | ||
9970 | if (indirect_p) | |
9971 | addr = build_va_arg_indirect_ref (addr); | |
9972 | ||
9973 | return addr; | |
9974 | } | |
9975 | ||
9976 | /* Implement TARGET_SETUP_INCOMING_VARARGS. */ | |
9977 | ||
9978 | static void | |
ef4bddc2 | 9979 | aarch64_setup_incoming_varargs (cumulative_args_t cum_v, machine_mode mode, |
43e9d192 IB |
9980 | tree type, int *pretend_size ATTRIBUTE_UNUSED, |
9981 | int no_rtl) | |
9982 | { | |
9983 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
9984 | CUMULATIVE_ARGS local_cum; | |
88e3bdd1 JW |
9985 | int gr_saved = cfun->va_list_gpr_size; |
9986 | int vr_saved = cfun->va_list_fpr_size; | |
43e9d192 IB |
9987 | |
9988 | /* The caller has advanced CUM up to, but not beyond, the last named | |
9989 | argument. Advance a local copy of CUM past the last "real" named | |
9990 | argument, to find out how many registers are left over. */ | |
9991 | local_cum = *cum; | |
9992 | aarch64_function_arg_advance (pack_cumulative_args(&local_cum), mode, type, true); | |
9993 | ||
88e3bdd1 JW |
9994 | /* Found out how many registers we need to save. |
9995 | Honor tree-stdvar analysis results. */ | |
9996 | if (cfun->va_list_gpr_size) | |
9997 | gr_saved = MIN (NUM_ARG_REGS - local_cum.aapcs_ncrn, | |
9998 | cfun->va_list_gpr_size / UNITS_PER_WORD); | |
9999 | if (cfun->va_list_fpr_size) | |
10000 | vr_saved = MIN (NUM_FP_ARG_REGS - local_cum.aapcs_nvrn, | |
10001 | cfun->va_list_fpr_size / UNITS_PER_VREG); | |
43e9d192 | 10002 | |
d5726973 | 10003 | if (!TARGET_FLOAT) |
43e9d192 | 10004 | { |
261fb553 | 10005 | gcc_assert (local_cum.aapcs_nvrn == 0); |
43e9d192 IB |
10006 | vr_saved = 0; |
10007 | } | |
10008 | ||
10009 | if (!no_rtl) | |
10010 | { | |
10011 | if (gr_saved > 0) | |
10012 | { | |
10013 | rtx ptr, mem; | |
10014 | ||
10015 | /* virtual_incoming_args_rtx should have been 16-byte aligned. */ | |
10016 | ptr = plus_constant (Pmode, virtual_incoming_args_rtx, | |
10017 | - gr_saved * UNITS_PER_WORD); | |
10018 | mem = gen_frame_mem (BLKmode, ptr); | |
10019 | set_mem_alias_set (mem, get_varargs_alias_set ()); | |
10020 | ||
10021 | move_block_from_reg (local_cum.aapcs_ncrn + R0_REGNUM, | |
10022 | mem, gr_saved); | |
10023 | } | |
10024 | if (vr_saved > 0) | |
10025 | { | |
10026 | /* We can't use move_block_from_reg, because it will use | |
10027 | the wrong mode, storing D regs only. */ | |
ef4bddc2 | 10028 | machine_mode mode = TImode; |
88e3bdd1 | 10029 | int off, i, vr_start; |
43e9d192 IB |
10030 | |
10031 | /* Set OFF to the offset from virtual_incoming_args_rtx of | |
10032 | the first vector register. The VR save area lies below | |
10033 | the GR one, and is aligned to 16 bytes. */ | |
4f59f9f2 UB |
10034 | off = -ROUND_UP (gr_saved * UNITS_PER_WORD, |
10035 | STACK_BOUNDARY / BITS_PER_UNIT); | |
43e9d192 IB |
10036 | off -= vr_saved * UNITS_PER_VREG; |
10037 | ||
88e3bdd1 JW |
10038 | vr_start = V0_REGNUM + local_cum.aapcs_nvrn; |
10039 | for (i = 0; i < vr_saved; ++i) | |
43e9d192 IB |
10040 | { |
10041 | rtx ptr, mem; | |
10042 | ||
10043 | ptr = plus_constant (Pmode, virtual_incoming_args_rtx, off); | |
10044 | mem = gen_frame_mem (mode, ptr); | |
10045 | set_mem_alias_set (mem, get_varargs_alias_set ()); | |
88e3bdd1 | 10046 | aarch64_emit_move (mem, gen_rtx_REG (mode, vr_start + i)); |
43e9d192 IB |
10047 | off += UNITS_PER_VREG; |
10048 | } | |
10049 | } | |
10050 | } | |
10051 | ||
10052 | /* We don't save the size into *PRETEND_SIZE because we want to avoid | |
10053 | any complication of having crtl->args.pretend_args_size changed. */ | |
8799637a | 10054 | cfun->machine->frame.saved_varargs_size |
4f59f9f2 UB |
10055 | = (ROUND_UP (gr_saved * UNITS_PER_WORD, |
10056 | STACK_BOUNDARY / BITS_PER_UNIT) | |
43e9d192 IB |
10057 | + vr_saved * UNITS_PER_VREG); |
10058 | } | |
10059 | ||
10060 | static void | |
10061 | aarch64_conditional_register_usage (void) | |
10062 | { | |
10063 | int i; | |
10064 | if (!TARGET_FLOAT) | |
10065 | { | |
10066 | for (i = V0_REGNUM; i <= V31_REGNUM; i++) | |
10067 | { | |
10068 | fixed_regs[i] = 1; | |
10069 | call_used_regs[i] = 1; | |
10070 | } | |
10071 | } | |
10072 | } | |
10073 | ||
10074 | /* Walk down the type tree of TYPE counting consecutive base elements. | |
10075 | If *MODEP is VOIDmode, then set it to the first valid floating point | |
10076 | type. If a non-floating point type is found, or if a floating point | |
10077 | type that doesn't match a non-VOIDmode *MODEP is found, then return -1, | |
10078 | otherwise return the count in the sub-tree. */ | |
10079 | static int | |
ef4bddc2 | 10080 | aapcs_vfp_sub_candidate (const_tree type, machine_mode *modep) |
43e9d192 | 10081 | { |
ef4bddc2 | 10082 | machine_mode mode; |
43e9d192 IB |
10083 | HOST_WIDE_INT size; |
10084 | ||
10085 | switch (TREE_CODE (type)) | |
10086 | { | |
10087 | case REAL_TYPE: | |
10088 | mode = TYPE_MODE (type); | |
10089 | if (mode != DFmode && mode != SFmode && mode != TFmode) | |
10090 | return -1; | |
10091 | ||
10092 | if (*modep == VOIDmode) | |
10093 | *modep = mode; | |
10094 | ||
10095 | if (*modep == mode) | |
10096 | return 1; | |
10097 | ||
10098 | break; | |
10099 | ||
10100 | case COMPLEX_TYPE: | |
10101 | mode = TYPE_MODE (TREE_TYPE (type)); | |
10102 | if (mode != DFmode && mode != SFmode && mode != TFmode) | |
10103 | return -1; | |
10104 | ||
10105 | if (*modep == VOIDmode) | |
10106 | *modep = mode; | |
10107 | ||
10108 | if (*modep == mode) | |
10109 | return 2; | |
10110 | ||
10111 | break; | |
10112 | ||
10113 | case VECTOR_TYPE: | |
10114 | /* Use V2SImode and V4SImode as representatives of all 64-bit | |
10115 | and 128-bit vector types. */ | |
10116 | size = int_size_in_bytes (type); | |
10117 | switch (size) | |
10118 | { | |
10119 | case 8: | |
10120 | mode = V2SImode; | |
10121 | break; | |
10122 | case 16: | |
10123 | mode = V4SImode; | |
10124 | break; | |
10125 | default: | |
10126 | return -1; | |
10127 | } | |
10128 | ||
10129 | if (*modep == VOIDmode) | |
10130 | *modep = mode; | |
10131 | ||
10132 | /* Vector modes are considered to be opaque: two vectors are | |
10133 | equivalent for the purposes of being homogeneous aggregates | |
10134 | if they are the same size. */ | |
10135 | if (*modep == mode) | |
10136 | return 1; | |
10137 | ||
10138 | break; | |
10139 | ||
10140 | case ARRAY_TYPE: | |
10141 | { | |
10142 | int count; | |
10143 | tree index = TYPE_DOMAIN (type); | |
10144 | ||
807e902e KZ |
10145 | /* Can't handle incomplete types nor sizes that are not |
10146 | fixed. */ | |
10147 | if (!COMPLETE_TYPE_P (type) | |
10148 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
43e9d192 IB |
10149 | return -1; |
10150 | ||
10151 | count = aapcs_vfp_sub_candidate (TREE_TYPE (type), modep); | |
10152 | if (count == -1 | |
10153 | || !index | |
10154 | || !TYPE_MAX_VALUE (index) | |
cc269bb6 | 10155 | || !tree_fits_uhwi_p (TYPE_MAX_VALUE (index)) |
43e9d192 | 10156 | || !TYPE_MIN_VALUE (index) |
cc269bb6 | 10157 | || !tree_fits_uhwi_p (TYPE_MIN_VALUE (index)) |
43e9d192 IB |
10158 | || count < 0) |
10159 | return -1; | |
10160 | ||
ae7e9ddd RS |
10161 | count *= (1 + tree_to_uhwi (TYPE_MAX_VALUE (index)) |
10162 | - tree_to_uhwi (TYPE_MIN_VALUE (index))); | |
43e9d192 IB |
10163 | |
10164 | /* There must be no padding. */ | |
807e902e | 10165 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) |
43e9d192 IB |
10166 | return -1; |
10167 | ||
10168 | return count; | |
10169 | } | |
10170 | ||
10171 | case RECORD_TYPE: | |
10172 | { | |
10173 | int count = 0; | |
10174 | int sub_count; | |
10175 | tree field; | |
10176 | ||
807e902e KZ |
10177 | /* Can't handle incomplete types nor sizes that are not |
10178 | fixed. */ | |
10179 | if (!COMPLETE_TYPE_P (type) | |
10180 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
43e9d192 IB |
10181 | return -1; |
10182 | ||
10183 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
10184 | { | |
10185 | if (TREE_CODE (field) != FIELD_DECL) | |
10186 | continue; | |
10187 | ||
10188 | sub_count = aapcs_vfp_sub_candidate (TREE_TYPE (field), modep); | |
10189 | if (sub_count < 0) | |
10190 | return -1; | |
10191 | count += sub_count; | |
10192 | } | |
10193 | ||
10194 | /* There must be no padding. */ | |
807e902e | 10195 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) |
43e9d192 IB |
10196 | return -1; |
10197 | ||
10198 | return count; | |
10199 | } | |
10200 | ||
10201 | case UNION_TYPE: | |
10202 | case QUAL_UNION_TYPE: | |
10203 | { | |
10204 | /* These aren't very interesting except in a degenerate case. */ | |
10205 | int count = 0; | |
10206 | int sub_count; | |
10207 | tree field; | |
10208 | ||
807e902e KZ |
10209 | /* Can't handle incomplete types nor sizes that are not |
10210 | fixed. */ | |
10211 | if (!COMPLETE_TYPE_P (type) | |
10212 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
43e9d192 IB |
10213 | return -1; |
10214 | ||
10215 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
10216 | { | |
10217 | if (TREE_CODE (field) != FIELD_DECL) | |
10218 | continue; | |
10219 | ||
10220 | sub_count = aapcs_vfp_sub_candidate (TREE_TYPE (field), modep); | |
10221 | if (sub_count < 0) | |
10222 | return -1; | |
10223 | count = count > sub_count ? count : sub_count; | |
10224 | } | |
10225 | ||
10226 | /* There must be no padding. */ | |
807e902e | 10227 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) |
43e9d192 IB |
10228 | return -1; |
10229 | ||
10230 | return count; | |
10231 | } | |
10232 | ||
10233 | default: | |
10234 | break; | |
10235 | } | |
10236 | ||
10237 | return -1; | |
10238 | } | |
10239 | ||
b6ec6215 KT |
10240 | /* Return TRUE if the type, as described by TYPE and MODE, is a short vector |
10241 | type as described in AAPCS64 \S 4.1.2. | |
10242 | ||
10243 | See the comment above aarch64_composite_type_p for the notes on MODE. */ | |
10244 | ||
10245 | static bool | |
10246 | aarch64_short_vector_p (const_tree type, | |
10247 | machine_mode mode) | |
10248 | { | |
10249 | HOST_WIDE_INT size = -1; | |
10250 | ||
10251 | if (type && TREE_CODE (type) == VECTOR_TYPE) | |
10252 | size = int_size_in_bytes (type); | |
10253 | else if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT | |
10254 | || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) | |
10255 | size = GET_MODE_SIZE (mode); | |
10256 | ||
10257 | return (size == 8 || size == 16); | |
10258 | } | |
10259 | ||
43e9d192 IB |
10260 | /* Return TRUE if the type, as described by TYPE and MODE, is a composite |
10261 | type as described in AAPCS64 \S 4.3. This includes aggregate, union and | |
10262 | array types. The C99 floating-point complex types are also considered | |
10263 | as composite types, according to AAPCS64 \S 7.1.1. The complex integer | |
10264 | types, which are GCC extensions and out of the scope of AAPCS64, are | |
10265 | treated as composite types here as well. | |
10266 | ||
10267 | Note that MODE itself is not sufficient in determining whether a type | |
10268 | is such a composite type or not. This is because | |
10269 | stor-layout.c:compute_record_mode may have already changed the MODE | |
10270 | (BLKmode) of a RECORD_TYPE TYPE to some other mode. For example, a | |
10271 | structure with only one field may have its MODE set to the mode of the | |
10272 | field. Also an integer mode whose size matches the size of the | |
10273 | RECORD_TYPE type may be used to substitute the original mode | |
10274 | (i.e. BLKmode) in certain circumstances. In other words, MODE cannot be | |
10275 | solely relied on. */ | |
10276 | ||
10277 | static bool | |
10278 | aarch64_composite_type_p (const_tree type, | |
ef4bddc2 | 10279 | machine_mode mode) |
43e9d192 | 10280 | { |
b6ec6215 KT |
10281 | if (aarch64_short_vector_p (type, mode)) |
10282 | return false; | |
10283 | ||
43e9d192 IB |
10284 | if (type && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == COMPLEX_TYPE)) |
10285 | return true; | |
10286 | ||
10287 | if (mode == BLKmode | |
10288 | || GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT | |
10289 | || GET_MODE_CLASS (mode) == MODE_COMPLEX_INT) | |
10290 | return true; | |
10291 | ||
10292 | return false; | |
10293 | } | |
10294 | ||
43e9d192 IB |
10295 | /* Return TRUE if an argument, whose type is described by TYPE and MODE, |
10296 | shall be passed or returned in simd/fp register(s) (providing these | |
10297 | parameter passing registers are available). | |
10298 | ||
10299 | Upon successful return, *COUNT returns the number of needed registers, | |
10300 | *BASE_MODE returns the mode of the individual register and when IS_HAF | |
10301 | is not NULL, *IS_HA indicates whether or not the argument is a homogeneous | |
10302 | floating-point aggregate or a homogeneous short-vector aggregate. */ | |
10303 | ||
10304 | static bool | |
ef4bddc2 | 10305 | aarch64_vfp_is_call_or_return_candidate (machine_mode mode, |
43e9d192 | 10306 | const_tree type, |
ef4bddc2 | 10307 | machine_mode *base_mode, |
43e9d192 IB |
10308 | int *count, |
10309 | bool *is_ha) | |
10310 | { | |
ef4bddc2 | 10311 | machine_mode new_mode = VOIDmode; |
43e9d192 IB |
10312 | bool composite_p = aarch64_composite_type_p (type, mode); |
10313 | ||
10314 | if (is_ha != NULL) *is_ha = false; | |
10315 | ||
10316 | if ((!composite_p && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
10317 | || aarch64_short_vector_p (type, mode)) | |
10318 | { | |
10319 | *count = 1; | |
10320 | new_mode = mode; | |
10321 | } | |
10322 | else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT) | |
10323 | { | |
10324 | if (is_ha != NULL) *is_ha = true; | |
10325 | *count = 2; | |
10326 | new_mode = GET_MODE_INNER (mode); | |
10327 | } | |
10328 | else if (type && composite_p) | |
10329 | { | |
10330 | int ag_count = aapcs_vfp_sub_candidate (type, &new_mode); | |
10331 | ||
10332 | if (ag_count > 0 && ag_count <= HA_MAX_NUM_FLDS) | |
10333 | { | |
10334 | if (is_ha != NULL) *is_ha = true; | |
10335 | *count = ag_count; | |
10336 | } | |
10337 | else | |
10338 | return false; | |
10339 | } | |
10340 | else | |
10341 | return false; | |
10342 | ||
10343 | *base_mode = new_mode; | |
10344 | return true; | |
10345 | } | |
10346 | ||
10347 | /* Implement TARGET_STRUCT_VALUE_RTX. */ | |
10348 | ||
10349 | static rtx | |
10350 | aarch64_struct_value_rtx (tree fndecl ATTRIBUTE_UNUSED, | |
10351 | int incoming ATTRIBUTE_UNUSED) | |
10352 | { | |
10353 | return gen_rtx_REG (Pmode, AARCH64_STRUCT_VALUE_REGNUM); | |
10354 | } | |
10355 | ||
10356 | /* Implements target hook vector_mode_supported_p. */ | |
10357 | static bool | |
ef4bddc2 | 10358 | aarch64_vector_mode_supported_p (machine_mode mode) |
43e9d192 IB |
10359 | { |
10360 | if (TARGET_SIMD | |
10361 | && (mode == V4SImode || mode == V8HImode | |
10362 | || mode == V16QImode || mode == V2DImode | |
10363 | || mode == V2SImode || mode == V4HImode | |
10364 | || mode == V8QImode || mode == V2SFmode | |
ad7d90cc | 10365 | || mode == V4SFmode || mode == V2DFmode |
71a11456 | 10366 | || mode == V4HFmode || mode == V8HFmode |
ad7d90cc | 10367 | || mode == V1DFmode)) |
43e9d192 IB |
10368 | return true; |
10369 | ||
10370 | return false; | |
10371 | } | |
10372 | ||
b7342d25 IB |
10373 | /* Return appropriate SIMD container |
10374 | for MODE within a vector of WIDTH bits. */ | |
ef4bddc2 RS |
10375 | static machine_mode |
10376 | aarch64_simd_container_mode (machine_mode mode, unsigned width) | |
43e9d192 | 10377 | { |
b7342d25 | 10378 | gcc_assert (width == 64 || width == 128); |
43e9d192 | 10379 | if (TARGET_SIMD) |
b7342d25 IB |
10380 | { |
10381 | if (width == 128) | |
10382 | switch (mode) | |
10383 | { | |
10384 | case DFmode: | |
10385 | return V2DFmode; | |
10386 | case SFmode: | |
10387 | return V4SFmode; | |
10388 | case SImode: | |
10389 | return V4SImode; | |
10390 | case HImode: | |
10391 | return V8HImode; | |
10392 | case QImode: | |
10393 | return V16QImode; | |
10394 | case DImode: | |
10395 | return V2DImode; | |
10396 | default: | |
10397 | break; | |
10398 | } | |
10399 | else | |
10400 | switch (mode) | |
10401 | { | |
10402 | case SFmode: | |
10403 | return V2SFmode; | |
10404 | case SImode: | |
10405 | return V2SImode; | |
10406 | case HImode: | |
10407 | return V4HImode; | |
10408 | case QImode: | |
10409 | return V8QImode; | |
10410 | default: | |
10411 | break; | |
10412 | } | |
10413 | } | |
43e9d192 IB |
10414 | return word_mode; |
10415 | } | |
10416 | ||
b7342d25 | 10417 | /* Return 128-bit container as the preferred SIMD mode for MODE. */ |
ef4bddc2 RS |
10418 | static machine_mode |
10419 | aarch64_preferred_simd_mode (machine_mode mode) | |
b7342d25 IB |
10420 | { |
10421 | return aarch64_simd_container_mode (mode, 128); | |
10422 | } | |
10423 | ||
3b357264 JG |
10424 | /* Return the bitmask of possible vector sizes for the vectorizer |
10425 | to iterate over. */ | |
10426 | static unsigned int | |
10427 | aarch64_autovectorize_vector_sizes (void) | |
10428 | { | |
10429 | return (16 | 8); | |
10430 | } | |
10431 | ||
ac2b960f YZ |
10432 | /* Implement TARGET_MANGLE_TYPE. */ |
10433 | ||
6f549691 | 10434 | static const char * |
ac2b960f YZ |
10435 | aarch64_mangle_type (const_tree type) |
10436 | { | |
10437 | /* The AArch64 ABI documents say that "__va_list" has to be | |
10438 | managled as if it is in the "std" namespace. */ | |
10439 | if (lang_hooks.types_compatible_p (CONST_CAST_TREE (type), va_list_type)) | |
10440 | return "St9__va_list"; | |
10441 | ||
c2ec330c AL |
10442 | /* Half-precision float. */ |
10443 | if (TREE_CODE (type) == REAL_TYPE && TYPE_PRECISION (type) == 16) | |
10444 | return "Dh"; | |
10445 | ||
f9d53c27 TB |
10446 | /* Mangle AArch64-specific internal types. TYPE_NAME is non-NULL_TREE for |
10447 | builtin types. */ | |
10448 | if (TYPE_NAME (type) != NULL) | |
10449 | return aarch64_mangle_builtin_type (type); | |
c6fc9e43 | 10450 | |
ac2b960f YZ |
10451 | /* Use the default mangling. */ |
10452 | return NULL; | |
10453 | } | |
10454 | ||
8baff86e KT |
10455 | |
10456 | /* Return true if the rtx_insn contains a MEM RTX somewhere | |
10457 | in it. */ | |
75cf1494 KT |
10458 | |
10459 | static bool | |
8baff86e | 10460 | has_memory_op (rtx_insn *mem_insn) |
75cf1494 | 10461 | { |
8baff86e KT |
10462 | subrtx_iterator::array_type array; |
10463 | FOR_EACH_SUBRTX (iter, array, PATTERN (mem_insn), ALL) | |
10464 | if (MEM_P (*iter)) | |
10465 | return true; | |
10466 | ||
10467 | return false; | |
75cf1494 KT |
10468 | } |
10469 | ||
10470 | /* Find the first rtx_insn before insn that will generate an assembly | |
10471 | instruction. */ | |
10472 | ||
10473 | static rtx_insn * | |
10474 | aarch64_prev_real_insn (rtx_insn *insn) | |
10475 | { | |
10476 | if (!insn) | |
10477 | return NULL; | |
10478 | ||
10479 | do | |
10480 | { | |
10481 | insn = prev_real_insn (insn); | |
10482 | } | |
10483 | while (insn && recog_memoized (insn) < 0); | |
10484 | ||
10485 | return insn; | |
10486 | } | |
10487 | ||
10488 | static bool | |
10489 | is_madd_op (enum attr_type t1) | |
10490 | { | |
10491 | unsigned int i; | |
10492 | /* A number of these may be AArch32 only. */ | |
10493 | enum attr_type mlatypes[] = { | |
10494 | TYPE_MLA, TYPE_MLAS, TYPE_SMLAD, TYPE_SMLADX, TYPE_SMLAL, TYPE_SMLALD, | |
10495 | TYPE_SMLALS, TYPE_SMLALXY, TYPE_SMLAWX, TYPE_SMLAWY, TYPE_SMLAXY, | |
10496 | TYPE_SMMLA, TYPE_UMLAL, TYPE_UMLALS,TYPE_SMLSD, TYPE_SMLSDX, TYPE_SMLSLD | |
10497 | }; | |
10498 | ||
10499 | for (i = 0; i < sizeof (mlatypes) / sizeof (enum attr_type); i++) | |
10500 | { | |
10501 | if (t1 == mlatypes[i]) | |
10502 | return true; | |
10503 | } | |
10504 | ||
10505 | return false; | |
10506 | } | |
10507 | ||
10508 | /* Check if there is a register dependency between a load and the insn | |
10509 | for which we hold recog_data. */ | |
10510 | ||
10511 | static bool | |
10512 | dep_between_memop_and_curr (rtx memop) | |
10513 | { | |
10514 | rtx load_reg; | |
10515 | int opno; | |
10516 | ||
8baff86e | 10517 | gcc_assert (GET_CODE (memop) == SET); |
75cf1494 KT |
10518 | |
10519 | if (!REG_P (SET_DEST (memop))) | |
10520 | return false; | |
10521 | ||
10522 | load_reg = SET_DEST (memop); | |
8baff86e | 10523 | for (opno = 1; opno < recog_data.n_operands; opno++) |
75cf1494 KT |
10524 | { |
10525 | rtx operand = recog_data.operand[opno]; | |
10526 | if (REG_P (operand) | |
10527 | && reg_overlap_mentioned_p (load_reg, operand)) | |
10528 | return true; | |
10529 | ||
10530 | } | |
10531 | return false; | |
10532 | } | |
10533 | ||
8baff86e KT |
10534 | |
10535 | /* When working around the Cortex-A53 erratum 835769, | |
10536 | given rtx_insn INSN, return true if it is a 64-bit multiply-accumulate | |
10537 | instruction and has a preceding memory instruction such that a NOP | |
10538 | should be inserted between them. */ | |
10539 | ||
75cf1494 KT |
10540 | bool |
10541 | aarch64_madd_needs_nop (rtx_insn* insn) | |
10542 | { | |
10543 | enum attr_type attr_type; | |
10544 | rtx_insn *prev; | |
10545 | rtx body; | |
10546 | ||
b32c1043 | 10547 | if (!TARGET_FIX_ERR_A53_835769) |
75cf1494 KT |
10548 | return false; |
10549 | ||
e322d6e3 | 10550 | if (!INSN_P (insn) || recog_memoized (insn) < 0) |
75cf1494 KT |
10551 | return false; |
10552 | ||
10553 | attr_type = get_attr_type (insn); | |
10554 | if (!is_madd_op (attr_type)) | |
10555 | return false; | |
10556 | ||
10557 | prev = aarch64_prev_real_insn (insn); | |
3fea1a75 KT |
10558 | /* aarch64_prev_real_insn can call recog_memoized on insns other than INSN. |
10559 | Restore recog state to INSN to avoid state corruption. */ | |
10560 | extract_constrain_insn_cached (insn); | |
10561 | ||
8baff86e | 10562 | if (!prev || !has_memory_op (prev)) |
75cf1494 KT |
10563 | return false; |
10564 | ||
10565 | body = single_set (prev); | |
10566 | ||
10567 | /* If the previous insn is a memory op and there is no dependency between | |
8baff86e KT |
10568 | it and the DImode madd, emit a NOP between them. If body is NULL then we |
10569 | have a complex memory operation, probably a load/store pair. | |
10570 | Be conservative for now and emit a NOP. */ | |
10571 | if (GET_MODE (recog_data.operand[0]) == DImode | |
10572 | && (!body || !dep_between_memop_and_curr (body))) | |
75cf1494 KT |
10573 | return true; |
10574 | ||
10575 | return false; | |
10576 | ||
10577 | } | |
10578 | ||
8baff86e KT |
10579 | |
10580 | /* Implement FINAL_PRESCAN_INSN. */ | |
10581 | ||
75cf1494 KT |
10582 | void |
10583 | aarch64_final_prescan_insn (rtx_insn *insn) | |
10584 | { | |
10585 | if (aarch64_madd_needs_nop (insn)) | |
10586 | fprintf (asm_out_file, "\tnop // between mem op and mult-accumulate\n"); | |
10587 | } | |
10588 | ||
10589 | ||
43e9d192 | 10590 | /* Return the equivalent letter for size. */ |
81c2dfb9 | 10591 | static char |
43e9d192 IB |
10592 | sizetochar (int size) |
10593 | { | |
10594 | switch (size) | |
10595 | { | |
10596 | case 64: return 'd'; | |
10597 | case 32: return 's'; | |
10598 | case 16: return 'h'; | |
10599 | case 8 : return 'b'; | |
10600 | default: gcc_unreachable (); | |
10601 | } | |
10602 | } | |
10603 | ||
3520f7cc JG |
10604 | /* Return true iff x is a uniform vector of floating-point |
10605 | constants, and the constant can be represented in | |
10606 | quarter-precision form. Note, as aarch64_float_const_representable | |
10607 | rejects both +0.0 and -0.0, we will also reject +0.0 and -0.0. */ | |
10608 | static bool | |
10609 | aarch64_vect_float_const_representable_p (rtx x) | |
10610 | { | |
92695fbb RS |
10611 | rtx elt; |
10612 | return (GET_MODE_CLASS (GET_MODE (x)) == MODE_VECTOR_FLOAT | |
10613 | && const_vec_duplicate_p (x, &elt) | |
10614 | && aarch64_float_const_representable_p (elt)); | |
3520f7cc JG |
10615 | } |
10616 | ||
d8edd899 | 10617 | /* Return true for valid and false for invalid. */ |
3ea63f60 | 10618 | bool |
ef4bddc2 | 10619 | aarch64_simd_valid_immediate (rtx op, machine_mode mode, bool inverse, |
48063b9d | 10620 | struct simd_immediate_info *info) |
43e9d192 IB |
10621 | { |
10622 | #define CHECK(STRIDE, ELSIZE, CLASS, TEST, SHIFT, NEG) \ | |
10623 | matches = 1; \ | |
10624 | for (i = 0; i < idx; i += (STRIDE)) \ | |
10625 | if (!(TEST)) \ | |
10626 | matches = 0; \ | |
10627 | if (matches) \ | |
10628 | { \ | |
10629 | immtype = (CLASS); \ | |
10630 | elsize = (ELSIZE); \ | |
43e9d192 IB |
10631 | eshift = (SHIFT); \ |
10632 | emvn = (NEG); \ | |
10633 | break; \ | |
10634 | } | |
10635 | ||
10636 | unsigned int i, elsize = 0, idx = 0, n_elts = CONST_VECTOR_NUNITS (op); | |
cb5ca315 | 10637 | unsigned int innersize = GET_MODE_UNIT_SIZE (mode); |
43e9d192 | 10638 | unsigned char bytes[16]; |
43e9d192 IB |
10639 | int immtype = -1, matches; |
10640 | unsigned int invmask = inverse ? 0xff : 0; | |
10641 | int eshift, emvn; | |
10642 | ||
43e9d192 | 10643 | if (GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) |
3520f7cc | 10644 | { |
81c2dfb9 IB |
10645 | if (! (aarch64_simd_imm_zero_p (op, mode) |
10646 | || aarch64_vect_float_const_representable_p (op))) | |
d8edd899 | 10647 | return false; |
3520f7cc | 10648 | |
48063b9d IB |
10649 | if (info) |
10650 | { | |
10651 | info->value = CONST_VECTOR_ELT (op, 0); | |
81c2dfb9 | 10652 | info->element_width = GET_MODE_BITSIZE (GET_MODE (info->value)); |
48063b9d IB |
10653 | info->mvn = false; |
10654 | info->shift = 0; | |
10655 | } | |
3520f7cc | 10656 | |
d8edd899 | 10657 | return true; |
3520f7cc | 10658 | } |
43e9d192 IB |
10659 | |
10660 | /* Splat vector constant out into a byte vector. */ | |
10661 | for (i = 0; i < n_elts; i++) | |
10662 | { | |
4b1e108c AL |
10663 | /* The vector is provided in gcc endian-neutral fashion. For aarch64_be, |
10664 | it must be laid out in the vector register in reverse order. */ | |
10665 | rtx el = CONST_VECTOR_ELT (op, BYTES_BIG_ENDIAN ? (n_elts - 1 - i) : i); | |
43e9d192 | 10666 | unsigned HOST_WIDE_INT elpart; |
43e9d192 | 10667 | |
ee78df47 KT |
10668 | gcc_assert (CONST_INT_P (el)); |
10669 | elpart = INTVAL (el); | |
10670 | ||
10671 | for (unsigned int byte = 0; byte < innersize; byte++) | |
10672 | { | |
10673 | bytes[idx++] = (elpart & 0xff) ^ invmask; | |
10674 | elpart >>= BITS_PER_UNIT; | |
10675 | } | |
43e9d192 | 10676 | |
43e9d192 IB |
10677 | } |
10678 | ||
10679 | /* Sanity check. */ | |
10680 | gcc_assert (idx == GET_MODE_SIZE (mode)); | |
10681 | ||
10682 | do | |
10683 | { | |
10684 | CHECK (4, 32, 0, bytes[i] == bytes[0] && bytes[i + 1] == 0 | |
10685 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 0, 0); | |
10686 | ||
10687 | CHECK (4, 32, 1, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
10688 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 8, 0); | |
10689 | ||
10690 | CHECK (4, 32, 2, bytes[i] == 0 && bytes[i + 1] == 0 | |
10691 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0, 16, 0); | |
10692 | ||
10693 | CHECK (4, 32, 3, bytes[i] == 0 && bytes[i + 1] == 0 | |
10694 | && bytes[i + 2] == 0 && bytes[i + 3] == bytes[3], 24, 0); | |
10695 | ||
10696 | CHECK (2, 16, 4, bytes[i] == bytes[0] && bytes[i + 1] == 0, 0, 0); | |
10697 | ||
10698 | CHECK (2, 16, 5, bytes[i] == 0 && bytes[i + 1] == bytes[1], 8, 0); | |
10699 | ||
10700 | CHECK (4, 32, 6, bytes[i] == bytes[0] && bytes[i + 1] == 0xff | |
10701 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 0, 1); | |
10702 | ||
10703 | CHECK (4, 32, 7, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
10704 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 8, 1); | |
10705 | ||
10706 | CHECK (4, 32, 8, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
10707 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff, 16, 1); | |
10708 | ||
10709 | CHECK (4, 32, 9, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
10710 | && bytes[i + 2] == 0xff && bytes[i + 3] == bytes[3], 24, 1); | |
10711 | ||
10712 | CHECK (2, 16, 10, bytes[i] == bytes[0] && bytes[i + 1] == 0xff, 0, 1); | |
10713 | ||
10714 | CHECK (2, 16, 11, bytes[i] == 0xff && bytes[i + 1] == bytes[1], 8, 1); | |
10715 | ||
10716 | CHECK (4, 32, 12, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
e4f0f84d | 10717 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 8, 0); |
43e9d192 IB |
10718 | |
10719 | CHECK (4, 32, 13, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
e4f0f84d | 10720 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 8, 1); |
43e9d192 IB |
10721 | |
10722 | CHECK (4, 32, 14, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
e4f0f84d | 10723 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0, 16, 0); |
43e9d192 IB |
10724 | |
10725 | CHECK (4, 32, 15, bytes[i] == 0 && bytes[i + 1] == 0 | |
e4f0f84d | 10726 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff, 16, 1); |
43e9d192 IB |
10727 | |
10728 | CHECK (1, 8, 16, bytes[i] == bytes[0], 0, 0); | |
10729 | ||
10730 | CHECK (1, 64, 17, (bytes[i] == 0 || bytes[i] == 0xff) | |
10731 | && bytes[i] == bytes[(i + 8) % idx], 0, 0); | |
10732 | } | |
10733 | while (0); | |
10734 | ||
e4f0f84d | 10735 | if (immtype == -1) |
d8edd899 | 10736 | return false; |
43e9d192 | 10737 | |
48063b9d | 10738 | if (info) |
43e9d192 | 10739 | { |
48063b9d | 10740 | info->element_width = elsize; |
48063b9d IB |
10741 | info->mvn = emvn != 0; |
10742 | info->shift = eshift; | |
10743 | ||
43e9d192 IB |
10744 | unsigned HOST_WIDE_INT imm = 0; |
10745 | ||
e4f0f84d TB |
10746 | if (immtype >= 12 && immtype <= 15) |
10747 | info->msl = true; | |
10748 | ||
43e9d192 IB |
10749 | /* Un-invert bytes of recognized vector, if necessary. */ |
10750 | if (invmask != 0) | |
10751 | for (i = 0; i < idx; i++) | |
10752 | bytes[i] ^= invmask; | |
10753 | ||
10754 | if (immtype == 17) | |
10755 | { | |
10756 | /* FIXME: Broken on 32-bit H_W_I hosts. */ | |
10757 | gcc_assert (sizeof (HOST_WIDE_INT) == 8); | |
10758 | ||
10759 | for (i = 0; i < 8; i++) | |
10760 | imm |= (unsigned HOST_WIDE_INT) (bytes[i] ? 0xff : 0) | |
10761 | << (i * BITS_PER_UNIT); | |
10762 | ||
43e9d192 | 10763 | |
48063b9d IB |
10764 | info->value = GEN_INT (imm); |
10765 | } | |
10766 | else | |
10767 | { | |
10768 | for (i = 0; i < elsize / BITS_PER_UNIT; i++) | |
10769 | imm |= (unsigned HOST_WIDE_INT) bytes[i] << (i * BITS_PER_UNIT); | |
43e9d192 IB |
10770 | |
10771 | /* Construct 'abcdefgh' because the assembler cannot handle | |
48063b9d IB |
10772 | generic constants. */ |
10773 | if (info->mvn) | |
43e9d192 | 10774 | imm = ~imm; |
48063b9d IB |
10775 | imm = (imm >> info->shift) & 0xff; |
10776 | info->value = GEN_INT (imm); | |
10777 | } | |
43e9d192 IB |
10778 | } |
10779 | ||
48063b9d | 10780 | return true; |
43e9d192 IB |
10781 | #undef CHECK |
10782 | } | |
10783 | ||
43e9d192 IB |
10784 | /* Check of immediate shift constants are within range. */ |
10785 | bool | |
ef4bddc2 | 10786 | aarch64_simd_shift_imm_p (rtx x, machine_mode mode, bool left) |
43e9d192 IB |
10787 | { |
10788 | int bit_width = GET_MODE_UNIT_SIZE (mode) * BITS_PER_UNIT; | |
10789 | if (left) | |
ddeabd3e | 10790 | return aarch64_const_vec_all_same_in_range_p (x, 0, bit_width - 1); |
43e9d192 | 10791 | else |
ddeabd3e | 10792 | return aarch64_const_vec_all_same_in_range_p (x, 1, bit_width); |
43e9d192 IB |
10793 | } |
10794 | ||
3520f7cc JG |
10795 | /* Return true if X is a uniform vector where all elements |
10796 | are either the floating-point constant 0.0 or the | |
10797 | integer constant 0. */ | |
43e9d192 | 10798 | bool |
ef4bddc2 | 10799 | aarch64_simd_imm_zero_p (rtx x, machine_mode mode) |
43e9d192 | 10800 | { |
3520f7cc | 10801 | return x == CONST0_RTX (mode); |
43e9d192 IB |
10802 | } |
10803 | ||
7325d85a KT |
10804 | |
10805 | /* Return the bitmask CONST_INT to select the bits required by a zero extract | |
10806 | operation of width WIDTH at bit position POS. */ | |
10807 | ||
10808 | rtx | |
10809 | aarch64_mask_from_zextract_ops (rtx width, rtx pos) | |
10810 | { | |
10811 | gcc_assert (CONST_INT_P (width)); | |
10812 | gcc_assert (CONST_INT_P (pos)); | |
10813 | ||
10814 | unsigned HOST_WIDE_INT mask | |
10815 | = ((unsigned HOST_WIDE_INT) 1 << UINTVAL (width)) - 1; | |
10816 | return GEN_INT (mask << UINTVAL (pos)); | |
10817 | } | |
10818 | ||
43e9d192 | 10819 | bool |
ef4bddc2 | 10820 | aarch64_simd_imm_scalar_p (rtx x, machine_mode mode ATTRIBUTE_UNUSED) |
43e9d192 IB |
10821 | { |
10822 | HOST_WIDE_INT imm = INTVAL (x); | |
10823 | int i; | |
10824 | ||
10825 | for (i = 0; i < 8; i++) | |
10826 | { | |
10827 | unsigned int byte = imm & 0xff; | |
10828 | if (byte != 0xff && byte != 0) | |
10829 | return false; | |
10830 | imm >>= 8; | |
10831 | } | |
10832 | ||
10833 | return true; | |
10834 | } | |
10835 | ||
83f8c414 | 10836 | bool |
a6e0bfa7 | 10837 | aarch64_mov_operand_p (rtx x, machine_mode mode) |
83f8c414 | 10838 | { |
83f8c414 CSS |
10839 | if (GET_CODE (x) == HIGH |
10840 | && aarch64_valid_symref (XEXP (x, 0), GET_MODE (XEXP (x, 0)))) | |
10841 | return true; | |
10842 | ||
82614948 | 10843 | if (CONST_INT_P (x)) |
83f8c414 CSS |
10844 | return true; |
10845 | ||
10846 | if (GET_CODE (x) == SYMBOL_REF && mode == DImode && CONSTANT_ADDRESS_P (x)) | |
10847 | return true; | |
10848 | ||
a6e0bfa7 | 10849 | return aarch64_classify_symbolic_expression (x) |
a5350ddc | 10850 | == SYMBOL_TINY_ABSOLUTE; |
83f8c414 CSS |
10851 | } |
10852 | ||
43e9d192 IB |
10853 | /* Return a const_int vector of VAL. */ |
10854 | rtx | |
ef4bddc2 | 10855 | aarch64_simd_gen_const_vector_dup (machine_mode mode, int val) |
43e9d192 IB |
10856 | { |
10857 | int nunits = GET_MODE_NUNITS (mode); | |
10858 | rtvec v = rtvec_alloc (nunits); | |
10859 | int i; | |
10860 | ||
10861 | for (i=0; i < nunits; i++) | |
10862 | RTVEC_ELT (v, i) = GEN_INT (val); | |
10863 | ||
10864 | return gen_rtx_CONST_VECTOR (mode, v); | |
10865 | } | |
10866 | ||
051d0e2f SN |
10867 | /* Check OP is a legal scalar immediate for the MOVI instruction. */ |
10868 | ||
10869 | bool | |
ef4bddc2 | 10870 | aarch64_simd_scalar_immediate_valid_for_move (rtx op, machine_mode mode) |
051d0e2f | 10871 | { |
ef4bddc2 | 10872 | machine_mode vmode; |
051d0e2f SN |
10873 | |
10874 | gcc_assert (!VECTOR_MODE_P (mode)); | |
10875 | vmode = aarch64_preferred_simd_mode (mode); | |
10876 | rtx op_v = aarch64_simd_gen_const_vector_dup (vmode, INTVAL (op)); | |
48063b9d | 10877 | return aarch64_simd_valid_immediate (op_v, vmode, false, NULL); |
051d0e2f SN |
10878 | } |
10879 | ||
988fa693 JG |
10880 | /* Construct and return a PARALLEL RTX vector with elements numbering the |
10881 | lanes of either the high (HIGH == TRUE) or low (HIGH == FALSE) half of | |
10882 | the vector - from the perspective of the architecture. This does not | |
10883 | line up with GCC's perspective on lane numbers, so we end up with | |
10884 | different masks depending on our target endian-ness. The diagram | |
10885 | below may help. We must draw the distinction when building masks | |
10886 | which select one half of the vector. An instruction selecting | |
10887 | architectural low-lanes for a big-endian target, must be described using | |
10888 | a mask selecting GCC high-lanes. | |
10889 | ||
10890 | Big-Endian Little-Endian | |
10891 | ||
10892 | GCC 0 1 2 3 3 2 1 0 | |
10893 | | x | x | x | x | | x | x | x | x | | |
10894 | Architecture 3 2 1 0 3 2 1 0 | |
10895 | ||
10896 | Low Mask: { 2, 3 } { 0, 1 } | |
10897 | High Mask: { 0, 1 } { 2, 3 } | |
10898 | */ | |
10899 | ||
43e9d192 | 10900 | rtx |
ef4bddc2 | 10901 | aarch64_simd_vect_par_cnst_half (machine_mode mode, bool high) |
43e9d192 IB |
10902 | { |
10903 | int nunits = GET_MODE_NUNITS (mode); | |
10904 | rtvec v = rtvec_alloc (nunits / 2); | |
988fa693 JG |
10905 | int high_base = nunits / 2; |
10906 | int low_base = 0; | |
10907 | int base; | |
43e9d192 IB |
10908 | rtx t1; |
10909 | int i; | |
10910 | ||
988fa693 JG |
10911 | if (BYTES_BIG_ENDIAN) |
10912 | base = high ? low_base : high_base; | |
10913 | else | |
10914 | base = high ? high_base : low_base; | |
10915 | ||
10916 | for (i = 0; i < nunits / 2; i++) | |
43e9d192 IB |
10917 | RTVEC_ELT (v, i) = GEN_INT (base + i); |
10918 | ||
10919 | t1 = gen_rtx_PARALLEL (mode, v); | |
10920 | return t1; | |
10921 | } | |
10922 | ||
988fa693 JG |
10923 | /* Check OP for validity as a PARALLEL RTX vector with elements |
10924 | numbering the lanes of either the high (HIGH == TRUE) or low lanes, | |
10925 | from the perspective of the architecture. See the diagram above | |
10926 | aarch64_simd_vect_par_cnst_half for more details. */ | |
10927 | ||
10928 | bool | |
ef4bddc2 | 10929 | aarch64_simd_check_vect_par_cnst_half (rtx op, machine_mode mode, |
988fa693 JG |
10930 | bool high) |
10931 | { | |
10932 | rtx ideal = aarch64_simd_vect_par_cnst_half (mode, high); | |
10933 | HOST_WIDE_INT count_op = XVECLEN (op, 0); | |
10934 | HOST_WIDE_INT count_ideal = XVECLEN (ideal, 0); | |
10935 | int i = 0; | |
10936 | ||
10937 | if (!VECTOR_MODE_P (mode)) | |
10938 | return false; | |
10939 | ||
10940 | if (count_op != count_ideal) | |
10941 | return false; | |
10942 | ||
10943 | for (i = 0; i < count_ideal; i++) | |
10944 | { | |
10945 | rtx elt_op = XVECEXP (op, 0, i); | |
10946 | rtx elt_ideal = XVECEXP (ideal, 0, i); | |
10947 | ||
4aa81c2e | 10948 | if (!CONST_INT_P (elt_op) |
988fa693 JG |
10949 | || INTVAL (elt_ideal) != INTVAL (elt_op)) |
10950 | return false; | |
10951 | } | |
10952 | return true; | |
10953 | } | |
10954 | ||
43e9d192 IB |
10955 | /* Bounds-check lanes. Ensure OPERAND lies between LOW (inclusive) and |
10956 | HIGH (exclusive). */ | |
10957 | void | |
46ed6024 CB |
10958 | aarch64_simd_lane_bounds (rtx operand, HOST_WIDE_INT low, HOST_WIDE_INT high, |
10959 | const_tree exp) | |
43e9d192 IB |
10960 | { |
10961 | HOST_WIDE_INT lane; | |
4aa81c2e | 10962 | gcc_assert (CONST_INT_P (operand)); |
43e9d192 IB |
10963 | lane = INTVAL (operand); |
10964 | ||
10965 | if (lane < low || lane >= high) | |
46ed6024 CB |
10966 | { |
10967 | if (exp) | |
cf0c27ef | 10968 | error ("%Klane %wd out of range %wd - %wd", exp, lane, low, high - 1); |
46ed6024 | 10969 | else |
cf0c27ef | 10970 | error ("lane %wd out of range %wd - %wd", lane, low, high - 1); |
46ed6024 | 10971 | } |
43e9d192 IB |
10972 | } |
10973 | ||
43e9d192 IB |
10974 | /* Return TRUE if OP is a valid vector addressing mode. */ |
10975 | bool | |
10976 | aarch64_simd_mem_operand_p (rtx op) | |
10977 | { | |
10978 | return MEM_P (op) && (GET_CODE (XEXP (op, 0)) == POST_INC | |
4aa81c2e | 10979 | || REG_P (XEXP (op, 0))); |
43e9d192 IB |
10980 | } |
10981 | ||
2d8c6dc1 AH |
10982 | /* Emit a register copy from operand to operand, taking care not to |
10983 | early-clobber source registers in the process. | |
43e9d192 | 10984 | |
2d8c6dc1 AH |
10985 | COUNT is the number of components into which the copy needs to be |
10986 | decomposed. */ | |
43e9d192 | 10987 | void |
2d8c6dc1 AH |
10988 | aarch64_simd_emit_reg_reg_move (rtx *operands, enum machine_mode mode, |
10989 | unsigned int count) | |
43e9d192 IB |
10990 | { |
10991 | unsigned int i; | |
2d8c6dc1 AH |
10992 | int rdest = REGNO (operands[0]); |
10993 | int rsrc = REGNO (operands[1]); | |
43e9d192 IB |
10994 | |
10995 | if (!reg_overlap_mentioned_p (operands[0], operands[1]) | |
2d8c6dc1 AH |
10996 | || rdest < rsrc) |
10997 | for (i = 0; i < count; i++) | |
10998 | emit_move_insn (gen_rtx_REG (mode, rdest + i), | |
10999 | gen_rtx_REG (mode, rsrc + i)); | |
43e9d192 | 11000 | else |
2d8c6dc1 AH |
11001 | for (i = 0; i < count; i++) |
11002 | emit_move_insn (gen_rtx_REG (mode, rdest + count - i - 1), | |
11003 | gen_rtx_REG (mode, rsrc + count - i - 1)); | |
43e9d192 IB |
11004 | } |
11005 | ||
668046d1 | 11006 | /* Compute and return the length of aarch64_simd_reglist<mode>, where <mode> is |
6ec0e5b9 | 11007 | one of VSTRUCT modes: OI, CI, or XI. */ |
668046d1 DS |
11008 | int |
11009 | aarch64_simd_attr_length_rglist (enum machine_mode mode) | |
11010 | { | |
11011 | return (GET_MODE_SIZE (mode) / UNITS_PER_VREG) * 4; | |
11012 | } | |
11013 | ||
db0253a4 TB |
11014 | /* Implement target hook TARGET_VECTOR_ALIGNMENT. The AAPCS64 sets the maximum |
11015 | alignment of a vector to 128 bits. */ | |
11016 | static HOST_WIDE_INT | |
11017 | aarch64_simd_vector_alignment (const_tree type) | |
11018 | { | |
9439e9a1 | 11019 | HOST_WIDE_INT align = tree_to_shwi (TYPE_SIZE (type)); |
db0253a4 TB |
11020 | return MIN (align, 128); |
11021 | } | |
11022 | ||
11023 | /* Implement target hook TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE. */ | |
11024 | static bool | |
11025 | aarch64_simd_vector_alignment_reachable (const_tree type, bool is_packed) | |
11026 | { | |
11027 | if (is_packed) | |
11028 | return false; | |
11029 | ||
11030 | /* We guarantee alignment for vectors up to 128-bits. */ | |
11031 | if (tree_int_cst_compare (TYPE_SIZE (type), | |
11032 | bitsize_int (BIGGEST_ALIGNMENT)) > 0) | |
11033 | return false; | |
11034 | ||
11035 | /* Vectors whose size is <= BIGGEST_ALIGNMENT are naturally aligned. */ | |
11036 | return true; | |
11037 | } | |
11038 | ||
4369c11e TB |
11039 | /* If VALS is a vector constant that can be loaded into a register |
11040 | using DUP, generate instructions to do so and return an RTX to | |
11041 | assign to the register. Otherwise return NULL_RTX. */ | |
11042 | static rtx | |
11043 | aarch64_simd_dup_constant (rtx vals) | |
11044 | { | |
ef4bddc2 RS |
11045 | machine_mode mode = GET_MODE (vals); |
11046 | machine_mode inner_mode = GET_MODE_INNER (mode); | |
4369c11e | 11047 | rtx x; |
4369c11e | 11048 | |
92695fbb | 11049 | if (!const_vec_duplicate_p (vals, &x)) |
4369c11e TB |
11050 | return NULL_RTX; |
11051 | ||
11052 | /* We can load this constant by using DUP and a constant in a | |
11053 | single ARM register. This will be cheaper than a vector | |
11054 | load. */ | |
92695fbb | 11055 | x = copy_to_mode_reg (inner_mode, x); |
4369c11e TB |
11056 | return gen_rtx_VEC_DUPLICATE (mode, x); |
11057 | } | |
11058 | ||
11059 | ||
11060 | /* Generate code to load VALS, which is a PARALLEL containing only | |
11061 | constants (for vec_init) or CONST_VECTOR, efficiently into a | |
11062 | register. Returns an RTX to copy into the register, or NULL_RTX | |
11063 | for a PARALLEL that can not be converted into a CONST_VECTOR. */ | |
1df3f464 | 11064 | static rtx |
4369c11e TB |
11065 | aarch64_simd_make_constant (rtx vals) |
11066 | { | |
ef4bddc2 | 11067 | machine_mode mode = GET_MODE (vals); |
4369c11e TB |
11068 | rtx const_dup; |
11069 | rtx const_vec = NULL_RTX; | |
11070 | int n_elts = GET_MODE_NUNITS (mode); | |
11071 | int n_const = 0; | |
11072 | int i; | |
11073 | ||
11074 | if (GET_CODE (vals) == CONST_VECTOR) | |
11075 | const_vec = vals; | |
11076 | else if (GET_CODE (vals) == PARALLEL) | |
11077 | { | |
11078 | /* A CONST_VECTOR must contain only CONST_INTs and | |
11079 | CONST_DOUBLEs, but CONSTANT_P allows more (e.g. SYMBOL_REF). | |
11080 | Only store valid constants in a CONST_VECTOR. */ | |
11081 | for (i = 0; i < n_elts; ++i) | |
11082 | { | |
11083 | rtx x = XVECEXP (vals, 0, i); | |
11084 | if (CONST_INT_P (x) || CONST_DOUBLE_P (x)) | |
11085 | n_const++; | |
11086 | } | |
11087 | if (n_const == n_elts) | |
11088 | const_vec = gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0)); | |
11089 | } | |
11090 | else | |
11091 | gcc_unreachable (); | |
11092 | ||
11093 | if (const_vec != NULL_RTX | |
48063b9d | 11094 | && aarch64_simd_valid_immediate (const_vec, mode, false, NULL)) |
4369c11e TB |
11095 | /* Load using MOVI/MVNI. */ |
11096 | return const_vec; | |
11097 | else if ((const_dup = aarch64_simd_dup_constant (vals)) != NULL_RTX) | |
11098 | /* Loaded using DUP. */ | |
11099 | return const_dup; | |
11100 | else if (const_vec != NULL_RTX) | |
11101 | /* Load from constant pool. We can not take advantage of single-cycle | |
11102 | LD1 because we need a PC-relative addressing mode. */ | |
11103 | return const_vec; | |
11104 | else | |
11105 | /* A PARALLEL containing something not valid inside CONST_VECTOR. | |
11106 | We can not construct an initializer. */ | |
11107 | return NULL_RTX; | |
11108 | } | |
11109 | ||
35a093b6 JG |
11110 | /* Expand a vector initialisation sequence, such that TARGET is |
11111 | initialised to contain VALS. */ | |
11112 | ||
4369c11e TB |
11113 | void |
11114 | aarch64_expand_vector_init (rtx target, rtx vals) | |
11115 | { | |
ef4bddc2 RS |
11116 | machine_mode mode = GET_MODE (target); |
11117 | machine_mode inner_mode = GET_MODE_INNER (mode); | |
35a093b6 | 11118 | /* The number of vector elements. */ |
4369c11e | 11119 | int n_elts = GET_MODE_NUNITS (mode); |
35a093b6 | 11120 | /* The number of vector elements which are not constant. */ |
8b66a2d4 AL |
11121 | int n_var = 0; |
11122 | rtx any_const = NULL_RTX; | |
35a093b6 JG |
11123 | /* The first element of vals. */ |
11124 | rtx v0 = XVECEXP (vals, 0, 0); | |
4369c11e | 11125 | bool all_same = true; |
4369c11e | 11126 | |
35a093b6 | 11127 | /* Count the number of variable elements to initialise. */ |
8b66a2d4 | 11128 | for (int i = 0; i < n_elts; ++i) |
4369c11e | 11129 | { |
8b66a2d4 | 11130 | rtx x = XVECEXP (vals, 0, i); |
35a093b6 | 11131 | if (!(CONST_INT_P (x) || CONST_DOUBLE_P (x))) |
8b66a2d4 AL |
11132 | ++n_var; |
11133 | else | |
11134 | any_const = x; | |
4369c11e | 11135 | |
35a093b6 | 11136 | all_same &= rtx_equal_p (x, v0); |
4369c11e TB |
11137 | } |
11138 | ||
35a093b6 JG |
11139 | /* No variable elements, hand off to aarch64_simd_make_constant which knows |
11140 | how best to handle this. */ | |
4369c11e TB |
11141 | if (n_var == 0) |
11142 | { | |
11143 | rtx constant = aarch64_simd_make_constant (vals); | |
11144 | if (constant != NULL_RTX) | |
11145 | { | |
11146 | emit_move_insn (target, constant); | |
11147 | return; | |
11148 | } | |
11149 | } | |
11150 | ||
11151 | /* Splat a single non-constant element if we can. */ | |
11152 | if (all_same) | |
11153 | { | |
35a093b6 | 11154 | rtx x = copy_to_mode_reg (inner_mode, v0); |
4369c11e TB |
11155 | aarch64_emit_move (target, gen_rtx_VEC_DUPLICATE (mode, x)); |
11156 | return; | |
11157 | } | |
11158 | ||
35a093b6 JG |
11159 | /* Initialise a vector which is part-variable. We want to first try |
11160 | to build those lanes which are constant in the most efficient way we | |
11161 | can. */ | |
11162 | if (n_var != n_elts) | |
4369c11e TB |
11163 | { |
11164 | rtx copy = copy_rtx (vals); | |
4369c11e | 11165 | |
8b66a2d4 AL |
11166 | /* Load constant part of vector. We really don't care what goes into the |
11167 | parts we will overwrite, but we're more likely to be able to load the | |
11168 | constant efficiently if it has fewer, larger, repeating parts | |
11169 | (see aarch64_simd_valid_immediate). */ | |
11170 | for (int i = 0; i < n_elts; i++) | |
11171 | { | |
11172 | rtx x = XVECEXP (vals, 0, i); | |
11173 | if (CONST_INT_P (x) || CONST_DOUBLE_P (x)) | |
11174 | continue; | |
11175 | rtx subst = any_const; | |
11176 | for (int bit = n_elts / 2; bit > 0; bit /= 2) | |
11177 | { | |
11178 | /* Look in the copied vector, as more elements are const. */ | |
11179 | rtx test = XVECEXP (copy, 0, i ^ bit); | |
11180 | if (CONST_INT_P (test) || CONST_DOUBLE_P (test)) | |
11181 | { | |
11182 | subst = test; | |
11183 | break; | |
11184 | } | |
11185 | } | |
11186 | XVECEXP (copy, 0, i) = subst; | |
11187 | } | |
4369c11e | 11188 | aarch64_expand_vector_init (target, copy); |
35a093b6 | 11189 | } |
4369c11e | 11190 | |
35a093b6 | 11191 | /* Insert the variable lanes directly. */ |
8b66a2d4 | 11192 | |
35a093b6 JG |
11193 | enum insn_code icode = optab_handler (vec_set_optab, mode); |
11194 | gcc_assert (icode != CODE_FOR_nothing); | |
4369c11e | 11195 | |
8b66a2d4 | 11196 | for (int i = 0; i < n_elts; i++) |
35a093b6 JG |
11197 | { |
11198 | rtx x = XVECEXP (vals, 0, i); | |
11199 | if (CONST_INT_P (x) || CONST_DOUBLE_P (x)) | |
11200 | continue; | |
11201 | x = copy_to_mode_reg (inner_mode, x); | |
11202 | emit_insn (GEN_FCN (icode) (target, x, GEN_INT (i))); | |
11203 | } | |
4369c11e TB |
11204 | } |
11205 | ||
43e9d192 | 11206 | static unsigned HOST_WIDE_INT |
ef4bddc2 | 11207 | aarch64_shift_truncation_mask (machine_mode mode) |
43e9d192 IB |
11208 | { |
11209 | return | |
ac59ad4e KT |
11210 | (!SHIFT_COUNT_TRUNCATED |
11211 | || aarch64_vector_mode_supported_p (mode) | |
43e9d192 IB |
11212 | || aarch64_vect_struct_mode_p (mode)) ? 0 : (GET_MODE_BITSIZE (mode) - 1); |
11213 | } | |
11214 | ||
43e9d192 IB |
11215 | /* Select a format to encode pointers in exception handling data. */ |
11216 | int | |
11217 | aarch64_asm_preferred_eh_data_format (int code ATTRIBUTE_UNUSED, int global) | |
11218 | { | |
11219 | int type; | |
11220 | switch (aarch64_cmodel) | |
11221 | { | |
11222 | case AARCH64_CMODEL_TINY: | |
11223 | case AARCH64_CMODEL_TINY_PIC: | |
11224 | case AARCH64_CMODEL_SMALL: | |
11225 | case AARCH64_CMODEL_SMALL_PIC: | |
1b1e81f8 | 11226 | case AARCH64_CMODEL_SMALL_SPIC: |
43e9d192 IB |
11227 | /* text+got+data < 4Gb. 4-byte signed relocs are sufficient |
11228 | for everything. */ | |
11229 | type = DW_EH_PE_sdata4; | |
11230 | break; | |
11231 | default: | |
11232 | /* No assumptions here. 8-byte relocs required. */ | |
11233 | type = DW_EH_PE_sdata8; | |
11234 | break; | |
11235 | } | |
11236 | return (global ? DW_EH_PE_indirect : 0) | DW_EH_PE_pcrel | type; | |
11237 | } | |
11238 | ||
e1c1ecb0 KT |
11239 | /* The last .arch and .tune assembly strings that we printed. */ |
11240 | static std::string aarch64_last_printed_arch_string; | |
11241 | static std::string aarch64_last_printed_tune_string; | |
11242 | ||
361fb3ee KT |
11243 | /* Implement ASM_DECLARE_FUNCTION_NAME. Output the ISA features used |
11244 | by the function fndecl. */ | |
11245 | ||
11246 | void | |
11247 | aarch64_declare_function_name (FILE *stream, const char* name, | |
11248 | tree fndecl) | |
11249 | { | |
11250 | tree target_parts = DECL_FUNCTION_SPECIFIC_TARGET (fndecl); | |
11251 | ||
11252 | struct cl_target_option *targ_options; | |
11253 | if (target_parts) | |
11254 | targ_options = TREE_TARGET_OPTION (target_parts); | |
11255 | else | |
11256 | targ_options = TREE_TARGET_OPTION (target_option_current_node); | |
11257 | gcc_assert (targ_options); | |
11258 | ||
11259 | const struct processor *this_arch | |
11260 | = aarch64_get_arch (targ_options->x_explicit_arch); | |
11261 | ||
054b4005 JG |
11262 | unsigned long isa_flags = targ_options->x_aarch64_isa_flags; |
11263 | std::string extension | |
04a99ebe JG |
11264 | = aarch64_get_extension_string_for_isa_flags (isa_flags, |
11265 | this_arch->flags); | |
e1c1ecb0 KT |
11266 | /* Only update the assembler .arch string if it is distinct from the last |
11267 | such string we printed. */ | |
11268 | std::string to_print = this_arch->name + extension; | |
11269 | if (to_print != aarch64_last_printed_arch_string) | |
11270 | { | |
11271 | asm_fprintf (asm_out_file, "\t.arch %s\n", to_print.c_str ()); | |
11272 | aarch64_last_printed_arch_string = to_print; | |
11273 | } | |
361fb3ee KT |
11274 | |
11275 | /* Print the cpu name we're tuning for in the comments, might be | |
e1c1ecb0 KT |
11276 | useful to readers of the generated asm. Do it only when it changes |
11277 | from function to function and verbose assembly is requested. */ | |
361fb3ee KT |
11278 | const struct processor *this_tune |
11279 | = aarch64_get_tune_cpu (targ_options->x_explicit_tune_core); | |
11280 | ||
e1c1ecb0 KT |
11281 | if (flag_debug_asm && aarch64_last_printed_tune_string != this_tune->name) |
11282 | { | |
11283 | asm_fprintf (asm_out_file, "\t" ASM_COMMENT_START ".tune %s\n", | |
11284 | this_tune->name); | |
11285 | aarch64_last_printed_tune_string = this_tune->name; | |
11286 | } | |
361fb3ee KT |
11287 | |
11288 | /* Don't forget the type directive for ELF. */ | |
11289 | ASM_OUTPUT_TYPE_DIRECTIVE (stream, name, "function"); | |
11290 | ASM_OUTPUT_LABEL (stream, name); | |
11291 | } | |
11292 | ||
e1c1ecb0 KT |
11293 | /* Implements TARGET_ASM_FILE_START. Output the assembly header. */ |
11294 | ||
11295 | static void | |
11296 | aarch64_start_file (void) | |
11297 | { | |
11298 | struct cl_target_option *default_options | |
11299 | = TREE_TARGET_OPTION (target_option_default_node); | |
11300 | ||
11301 | const struct processor *default_arch | |
11302 | = aarch64_get_arch (default_options->x_explicit_arch); | |
11303 | unsigned long default_isa_flags = default_options->x_aarch64_isa_flags; | |
11304 | std::string extension | |
04a99ebe JG |
11305 | = aarch64_get_extension_string_for_isa_flags (default_isa_flags, |
11306 | default_arch->flags); | |
e1c1ecb0 KT |
11307 | |
11308 | aarch64_last_printed_arch_string = default_arch->name + extension; | |
11309 | aarch64_last_printed_tune_string = ""; | |
11310 | asm_fprintf (asm_out_file, "\t.arch %s\n", | |
11311 | aarch64_last_printed_arch_string.c_str ()); | |
11312 | ||
11313 | default_file_start (); | |
11314 | } | |
11315 | ||
0462169c SN |
11316 | /* Emit load exclusive. */ |
11317 | ||
11318 | static void | |
ef4bddc2 | 11319 | aarch64_emit_load_exclusive (machine_mode mode, rtx rval, |
0462169c SN |
11320 | rtx mem, rtx model_rtx) |
11321 | { | |
11322 | rtx (*gen) (rtx, rtx, rtx); | |
11323 | ||
11324 | switch (mode) | |
11325 | { | |
11326 | case QImode: gen = gen_aarch64_load_exclusiveqi; break; | |
11327 | case HImode: gen = gen_aarch64_load_exclusivehi; break; | |
11328 | case SImode: gen = gen_aarch64_load_exclusivesi; break; | |
11329 | case DImode: gen = gen_aarch64_load_exclusivedi; break; | |
11330 | default: | |
11331 | gcc_unreachable (); | |
11332 | } | |
11333 | ||
11334 | emit_insn (gen (rval, mem, model_rtx)); | |
11335 | } | |
11336 | ||
11337 | /* Emit store exclusive. */ | |
11338 | ||
11339 | static void | |
ef4bddc2 | 11340 | aarch64_emit_store_exclusive (machine_mode mode, rtx bval, |
0462169c SN |
11341 | rtx rval, rtx mem, rtx model_rtx) |
11342 | { | |
11343 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
11344 | ||
11345 | switch (mode) | |
11346 | { | |
11347 | case QImode: gen = gen_aarch64_store_exclusiveqi; break; | |
11348 | case HImode: gen = gen_aarch64_store_exclusivehi; break; | |
11349 | case SImode: gen = gen_aarch64_store_exclusivesi; break; | |
11350 | case DImode: gen = gen_aarch64_store_exclusivedi; break; | |
11351 | default: | |
11352 | gcc_unreachable (); | |
11353 | } | |
11354 | ||
11355 | emit_insn (gen (bval, rval, mem, model_rtx)); | |
11356 | } | |
11357 | ||
11358 | /* Mark the previous jump instruction as unlikely. */ | |
11359 | ||
11360 | static void | |
11361 | aarch64_emit_unlikely_jump (rtx insn) | |
11362 | { | |
e5af9ddd | 11363 | int very_unlikely = REG_BR_PROB_BASE / 100 - 1; |
0462169c SN |
11364 | |
11365 | insn = emit_jump_insn (insn); | |
e5af9ddd | 11366 | add_int_reg_note (insn, REG_BR_PROB, very_unlikely); |
0462169c SN |
11367 | } |
11368 | ||
11369 | /* Expand a compare and swap pattern. */ | |
11370 | ||
11371 | void | |
11372 | aarch64_expand_compare_and_swap (rtx operands[]) | |
11373 | { | |
11374 | rtx bval, rval, mem, oldval, newval, is_weak, mod_s, mod_f, x; | |
ef4bddc2 | 11375 | machine_mode mode, cmp_mode; |
b0770c0f MW |
11376 | typedef rtx (*gen_cas_fn) (rtx, rtx, rtx, rtx, rtx, rtx, rtx); |
11377 | int idx; | |
11378 | gen_cas_fn gen; | |
11379 | const gen_cas_fn split_cas[] = | |
11380 | { | |
11381 | gen_aarch64_compare_and_swapqi, | |
11382 | gen_aarch64_compare_and_swaphi, | |
11383 | gen_aarch64_compare_and_swapsi, | |
11384 | gen_aarch64_compare_and_swapdi | |
11385 | }; | |
11386 | const gen_cas_fn atomic_cas[] = | |
11387 | { | |
11388 | gen_aarch64_compare_and_swapqi_lse, | |
11389 | gen_aarch64_compare_and_swaphi_lse, | |
11390 | gen_aarch64_compare_and_swapsi_lse, | |
11391 | gen_aarch64_compare_and_swapdi_lse | |
11392 | }; | |
0462169c SN |
11393 | |
11394 | bval = operands[0]; | |
11395 | rval = operands[1]; | |
11396 | mem = operands[2]; | |
11397 | oldval = operands[3]; | |
11398 | newval = operands[4]; | |
11399 | is_weak = operands[5]; | |
11400 | mod_s = operands[6]; | |
11401 | mod_f = operands[7]; | |
11402 | mode = GET_MODE (mem); | |
11403 | cmp_mode = mode; | |
11404 | ||
11405 | /* Normally the succ memory model must be stronger than fail, but in the | |
11406 | unlikely event of fail being ACQUIRE and succ being RELEASE we need to | |
11407 | promote succ to ACQ_REL so that we don't lose the acquire semantics. */ | |
11408 | ||
46b35980 AM |
11409 | if (is_mm_acquire (memmodel_from_int (INTVAL (mod_f))) |
11410 | && is_mm_release (memmodel_from_int (INTVAL (mod_s)))) | |
0462169c SN |
11411 | mod_s = GEN_INT (MEMMODEL_ACQ_REL); |
11412 | ||
11413 | switch (mode) | |
11414 | { | |
11415 | case QImode: | |
11416 | case HImode: | |
11417 | /* For short modes, we're going to perform the comparison in SImode, | |
11418 | so do the zero-extension now. */ | |
11419 | cmp_mode = SImode; | |
11420 | rval = gen_reg_rtx (SImode); | |
11421 | oldval = convert_modes (SImode, mode, oldval, true); | |
11422 | /* Fall through. */ | |
11423 | ||
11424 | case SImode: | |
11425 | case DImode: | |
11426 | /* Force the value into a register if needed. */ | |
11427 | if (!aarch64_plus_operand (oldval, mode)) | |
11428 | oldval = force_reg (cmp_mode, oldval); | |
11429 | break; | |
11430 | ||
11431 | default: | |
11432 | gcc_unreachable (); | |
11433 | } | |
11434 | ||
11435 | switch (mode) | |
11436 | { | |
b0770c0f MW |
11437 | case QImode: idx = 0; break; |
11438 | case HImode: idx = 1; break; | |
11439 | case SImode: idx = 2; break; | |
11440 | case DImode: idx = 3; break; | |
0462169c SN |
11441 | default: |
11442 | gcc_unreachable (); | |
11443 | } | |
b0770c0f MW |
11444 | if (TARGET_LSE) |
11445 | gen = atomic_cas[idx]; | |
11446 | else | |
11447 | gen = split_cas[idx]; | |
0462169c SN |
11448 | |
11449 | emit_insn (gen (rval, mem, oldval, newval, is_weak, mod_s, mod_f)); | |
11450 | ||
11451 | if (mode == QImode || mode == HImode) | |
11452 | emit_move_insn (operands[1], gen_lowpart (mode, rval)); | |
11453 | ||
11454 | x = gen_rtx_REG (CCmode, CC_REGNUM); | |
11455 | x = gen_rtx_EQ (SImode, x, const0_rtx); | |
f7df4a84 | 11456 | emit_insn (gen_rtx_SET (bval, x)); |
0462169c SN |
11457 | } |
11458 | ||
641c2f8b MW |
11459 | /* Test whether the target supports using a atomic load-operate instruction. |
11460 | CODE is the operation and AFTER is TRUE if the data in memory after the | |
11461 | operation should be returned and FALSE if the data before the operation | |
11462 | should be returned. Returns FALSE if the operation isn't supported by the | |
11463 | architecture. */ | |
11464 | ||
11465 | bool | |
11466 | aarch64_atomic_ldop_supported_p (enum rtx_code code) | |
11467 | { | |
11468 | if (!TARGET_LSE) | |
11469 | return false; | |
11470 | ||
11471 | switch (code) | |
11472 | { | |
11473 | case SET: | |
11474 | case AND: | |
11475 | case IOR: | |
11476 | case XOR: | |
11477 | case MINUS: | |
11478 | case PLUS: | |
11479 | return true; | |
11480 | default: | |
11481 | return false; | |
11482 | } | |
11483 | } | |
11484 | ||
f70fb3b6 MW |
11485 | /* Emit a barrier, that is appropriate for memory model MODEL, at the end of a |
11486 | sequence implementing an atomic operation. */ | |
11487 | ||
11488 | static void | |
11489 | aarch64_emit_post_barrier (enum memmodel model) | |
11490 | { | |
11491 | const enum memmodel base_model = memmodel_base (model); | |
11492 | ||
11493 | if (is_mm_sync (model) | |
11494 | && (base_model == MEMMODEL_ACQUIRE | |
11495 | || base_model == MEMMODEL_ACQ_REL | |
11496 | || base_model == MEMMODEL_SEQ_CST)) | |
11497 | { | |
11498 | emit_insn (gen_mem_thread_fence (GEN_INT (MEMMODEL_SEQ_CST))); | |
11499 | } | |
11500 | } | |
11501 | ||
b0770c0f MW |
11502 | /* Emit an atomic compare-and-swap operation. RVAL is the destination register |
11503 | for the data in memory. EXPECTED is the value expected to be in memory. | |
11504 | DESIRED is the value to store to memory. MEM is the memory location. MODEL | |
11505 | is the memory ordering to use. */ | |
11506 | ||
11507 | void | |
11508 | aarch64_gen_atomic_cas (rtx rval, rtx mem, | |
11509 | rtx expected, rtx desired, | |
11510 | rtx model) | |
11511 | { | |
11512 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
11513 | machine_mode mode; | |
11514 | ||
11515 | mode = GET_MODE (mem); | |
11516 | ||
11517 | switch (mode) | |
11518 | { | |
11519 | case QImode: gen = gen_aarch64_atomic_casqi; break; | |
11520 | case HImode: gen = gen_aarch64_atomic_cashi; break; | |
11521 | case SImode: gen = gen_aarch64_atomic_cassi; break; | |
11522 | case DImode: gen = gen_aarch64_atomic_casdi; break; | |
11523 | default: | |
11524 | gcc_unreachable (); | |
11525 | } | |
11526 | ||
11527 | /* Move the expected value into the CAS destination register. */ | |
11528 | emit_insn (gen_rtx_SET (rval, expected)); | |
11529 | ||
11530 | /* Emit the CAS. */ | |
11531 | emit_insn (gen (rval, mem, desired, model)); | |
11532 | ||
11533 | /* Compare the expected value with the value loaded by the CAS, to establish | |
11534 | whether the swap was made. */ | |
11535 | aarch64_gen_compare_reg (EQ, rval, expected); | |
11536 | } | |
11537 | ||
0462169c SN |
11538 | /* Split a compare and swap pattern. */ |
11539 | ||
11540 | void | |
11541 | aarch64_split_compare_and_swap (rtx operands[]) | |
11542 | { | |
11543 | rtx rval, mem, oldval, newval, scratch; | |
ef4bddc2 | 11544 | machine_mode mode; |
0462169c | 11545 | bool is_weak; |
5d8a22a5 DM |
11546 | rtx_code_label *label1, *label2; |
11547 | rtx x, cond; | |
ab876106 MW |
11548 | enum memmodel model; |
11549 | rtx model_rtx; | |
0462169c SN |
11550 | |
11551 | rval = operands[0]; | |
11552 | mem = operands[1]; | |
11553 | oldval = operands[2]; | |
11554 | newval = operands[3]; | |
11555 | is_weak = (operands[4] != const0_rtx); | |
ab876106 | 11556 | model_rtx = operands[5]; |
0462169c SN |
11557 | scratch = operands[7]; |
11558 | mode = GET_MODE (mem); | |
ab876106 | 11559 | model = memmodel_from_int (INTVAL (model_rtx)); |
0462169c | 11560 | |
5d8a22a5 | 11561 | label1 = NULL; |
0462169c SN |
11562 | if (!is_weak) |
11563 | { | |
11564 | label1 = gen_label_rtx (); | |
11565 | emit_label (label1); | |
11566 | } | |
11567 | label2 = gen_label_rtx (); | |
11568 | ||
ab876106 MW |
11569 | /* The initial load can be relaxed for a __sync operation since a final |
11570 | barrier will be emitted to stop code hoisting. */ | |
11571 | if (is_mm_sync (model)) | |
11572 | aarch64_emit_load_exclusive (mode, rval, mem, | |
11573 | GEN_INT (MEMMODEL_RELAXED)); | |
11574 | else | |
11575 | aarch64_emit_load_exclusive (mode, rval, mem, model_rtx); | |
0462169c SN |
11576 | |
11577 | cond = aarch64_gen_compare_reg (NE, rval, oldval); | |
11578 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
11579 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
11580 | gen_rtx_LABEL_REF (Pmode, label2), pc_rtx); | |
f7df4a84 | 11581 | aarch64_emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); |
0462169c | 11582 | |
ab876106 | 11583 | aarch64_emit_store_exclusive (mode, scratch, mem, newval, model_rtx); |
0462169c SN |
11584 | |
11585 | if (!is_weak) | |
11586 | { | |
11587 | x = gen_rtx_NE (VOIDmode, scratch, const0_rtx); | |
11588 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
11589 | gen_rtx_LABEL_REF (Pmode, label1), pc_rtx); | |
f7df4a84 | 11590 | aarch64_emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); |
0462169c SN |
11591 | } |
11592 | else | |
11593 | { | |
11594 | cond = gen_rtx_REG (CCmode, CC_REGNUM); | |
11595 | x = gen_rtx_COMPARE (CCmode, scratch, const0_rtx); | |
f7df4a84 | 11596 | emit_insn (gen_rtx_SET (cond, x)); |
0462169c SN |
11597 | } |
11598 | ||
11599 | emit_label (label2); | |
ab876106 MW |
11600 | |
11601 | /* Emit any final barrier needed for a __sync operation. */ | |
11602 | if (is_mm_sync (model)) | |
11603 | aarch64_emit_post_barrier (model); | |
0462169c SN |
11604 | } |
11605 | ||
68729b06 MW |
11606 | /* Emit a BIC instruction. */ |
11607 | ||
11608 | static void | |
11609 | aarch64_emit_bic (machine_mode mode, rtx dst, rtx s1, rtx s2, int shift) | |
11610 | { | |
11611 | rtx shift_rtx = GEN_INT (shift); | |
11612 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
11613 | ||
11614 | switch (mode) | |
11615 | { | |
11616 | case SImode: gen = gen_and_one_cmpl_lshrsi3; break; | |
11617 | case DImode: gen = gen_and_one_cmpl_lshrdi3; break; | |
11618 | default: | |
11619 | gcc_unreachable (); | |
11620 | } | |
11621 | ||
11622 | emit_insn (gen (dst, s2, shift_rtx, s1)); | |
11623 | } | |
11624 | ||
9cd7b720 MW |
11625 | /* Emit an atomic swap. */ |
11626 | ||
11627 | static void | |
11628 | aarch64_emit_atomic_swap (machine_mode mode, rtx dst, rtx value, | |
11629 | rtx mem, rtx model) | |
11630 | { | |
11631 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
11632 | ||
11633 | switch (mode) | |
11634 | { | |
11635 | case QImode: gen = gen_aarch64_atomic_swpqi; break; | |
11636 | case HImode: gen = gen_aarch64_atomic_swphi; break; | |
11637 | case SImode: gen = gen_aarch64_atomic_swpsi; break; | |
11638 | case DImode: gen = gen_aarch64_atomic_swpdi; break; | |
11639 | default: | |
11640 | gcc_unreachable (); | |
11641 | } | |
11642 | ||
11643 | emit_insn (gen (dst, mem, value, model)); | |
11644 | } | |
11645 | ||
641c2f8b MW |
11646 | /* Operations supported by aarch64_emit_atomic_load_op. */ |
11647 | ||
11648 | enum aarch64_atomic_load_op_code | |
11649 | { | |
11650 | AARCH64_LDOP_PLUS, /* A + B */ | |
11651 | AARCH64_LDOP_XOR, /* A ^ B */ | |
11652 | AARCH64_LDOP_OR, /* A | B */ | |
11653 | AARCH64_LDOP_BIC /* A & ~B */ | |
11654 | }; | |
11655 | ||
11656 | /* Emit an atomic load-operate. */ | |
11657 | ||
11658 | static void | |
11659 | aarch64_emit_atomic_load_op (enum aarch64_atomic_load_op_code code, | |
11660 | machine_mode mode, rtx dst, rtx src, | |
11661 | rtx mem, rtx model) | |
11662 | { | |
11663 | typedef rtx (*aarch64_atomic_load_op_fn) (rtx, rtx, rtx, rtx); | |
11664 | const aarch64_atomic_load_op_fn plus[] = | |
11665 | { | |
11666 | gen_aarch64_atomic_loadaddqi, | |
11667 | gen_aarch64_atomic_loadaddhi, | |
11668 | gen_aarch64_atomic_loadaddsi, | |
11669 | gen_aarch64_atomic_loadadddi | |
11670 | }; | |
11671 | const aarch64_atomic_load_op_fn eor[] = | |
11672 | { | |
11673 | gen_aarch64_atomic_loadeorqi, | |
11674 | gen_aarch64_atomic_loadeorhi, | |
11675 | gen_aarch64_atomic_loadeorsi, | |
11676 | gen_aarch64_atomic_loadeordi | |
11677 | }; | |
11678 | const aarch64_atomic_load_op_fn ior[] = | |
11679 | { | |
11680 | gen_aarch64_atomic_loadsetqi, | |
11681 | gen_aarch64_atomic_loadsethi, | |
11682 | gen_aarch64_atomic_loadsetsi, | |
11683 | gen_aarch64_atomic_loadsetdi | |
11684 | }; | |
11685 | const aarch64_atomic_load_op_fn bic[] = | |
11686 | { | |
11687 | gen_aarch64_atomic_loadclrqi, | |
11688 | gen_aarch64_atomic_loadclrhi, | |
11689 | gen_aarch64_atomic_loadclrsi, | |
11690 | gen_aarch64_atomic_loadclrdi | |
11691 | }; | |
11692 | aarch64_atomic_load_op_fn gen; | |
11693 | int idx = 0; | |
11694 | ||
11695 | switch (mode) | |
11696 | { | |
11697 | case QImode: idx = 0; break; | |
11698 | case HImode: idx = 1; break; | |
11699 | case SImode: idx = 2; break; | |
11700 | case DImode: idx = 3; break; | |
11701 | default: | |
11702 | gcc_unreachable (); | |
11703 | } | |
11704 | ||
11705 | switch (code) | |
11706 | { | |
11707 | case AARCH64_LDOP_PLUS: gen = plus[idx]; break; | |
11708 | case AARCH64_LDOP_XOR: gen = eor[idx]; break; | |
11709 | case AARCH64_LDOP_OR: gen = ior[idx]; break; | |
11710 | case AARCH64_LDOP_BIC: gen = bic[idx]; break; | |
11711 | default: | |
11712 | gcc_unreachable (); | |
11713 | } | |
11714 | ||
11715 | emit_insn (gen (dst, mem, src, model)); | |
11716 | } | |
11717 | ||
11718 | /* Emit an atomic load+operate. CODE is the operation. OUT_DATA is the | |
68729b06 MW |
11719 | location to store the data read from memory. OUT_RESULT is the location to |
11720 | store the result of the operation. MEM is the memory location to read and | |
11721 | modify. MODEL_RTX is the memory ordering to use. VALUE is the second | |
11722 | operand for the operation. Either OUT_DATA or OUT_RESULT, but not both, can | |
11723 | be NULL. */ | |
9cd7b720 MW |
11724 | |
11725 | void | |
68729b06 | 11726 | aarch64_gen_atomic_ldop (enum rtx_code code, rtx out_data, rtx out_result, |
9cd7b720 MW |
11727 | rtx mem, rtx value, rtx model_rtx) |
11728 | { | |
11729 | machine_mode mode = GET_MODE (mem); | |
641c2f8b MW |
11730 | machine_mode wmode = (mode == DImode ? DImode : SImode); |
11731 | const bool short_mode = (mode < SImode); | |
11732 | aarch64_atomic_load_op_code ldop_code; | |
11733 | rtx src; | |
11734 | rtx x; | |
11735 | ||
11736 | if (out_data) | |
11737 | out_data = gen_lowpart (mode, out_data); | |
9cd7b720 | 11738 | |
68729b06 MW |
11739 | if (out_result) |
11740 | out_result = gen_lowpart (mode, out_result); | |
11741 | ||
641c2f8b MW |
11742 | /* Make sure the value is in a register, putting it into a destination |
11743 | register if it needs to be manipulated. */ | |
11744 | if (!register_operand (value, mode) | |
11745 | || code == AND || code == MINUS) | |
11746 | { | |
68729b06 | 11747 | src = out_result ? out_result : out_data; |
641c2f8b MW |
11748 | emit_move_insn (src, gen_lowpart (mode, value)); |
11749 | } | |
11750 | else | |
11751 | src = value; | |
11752 | gcc_assert (register_operand (src, mode)); | |
9cd7b720 | 11753 | |
641c2f8b MW |
11754 | /* Preprocess the data for the operation as necessary. If the operation is |
11755 | a SET then emit a swap instruction and finish. */ | |
9cd7b720 MW |
11756 | switch (code) |
11757 | { | |
11758 | case SET: | |
641c2f8b | 11759 | aarch64_emit_atomic_swap (mode, out_data, src, mem, model_rtx); |
9cd7b720 MW |
11760 | return; |
11761 | ||
641c2f8b MW |
11762 | case MINUS: |
11763 | /* Negate the value and treat it as a PLUS. */ | |
11764 | { | |
11765 | rtx neg_src; | |
11766 | ||
11767 | /* Resize the value if necessary. */ | |
11768 | if (short_mode) | |
11769 | src = gen_lowpart (wmode, src); | |
11770 | ||
11771 | neg_src = gen_rtx_NEG (wmode, src); | |
11772 | emit_insn (gen_rtx_SET (src, neg_src)); | |
11773 | ||
11774 | if (short_mode) | |
11775 | src = gen_lowpart (mode, src); | |
11776 | } | |
11777 | /* Fall-through. */ | |
11778 | case PLUS: | |
11779 | ldop_code = AARCH64_LDOP_PLUS; | |
11780 | break; | |
11781 | ||
11782 | case IOR: | |
11783 | ldop_code = AARCH64_LDOP_OR; | |
11784 | break; | |
11785 | ||
11786 | case XOR: | |
11787 | ldop_code = AARCH64_LDOP_XOR; | |
11788 | break; | |
11789 | ||
11790 | case AND: | |
11791 | { | |
11792 | rtx not_src; | |
11793 | ||
11794 | /* Resize the value if necessary. */ | |
11795 | if (short_mode) | |
11796 | src = gen_lowpart (wmode, src); | |
11797 | ||
11798 | not_src = gen_rtx_NOT (wmode, src); | |
11799 | emit_insn (gen_rtx_SET (src, not_src)); | |
11800 | ||
11801 | if (short_mode) | |
11802 | src = gen_lowpart (mode, src); | |
11803 | } | |
11804 | ldop_code = AARCH64_LDOP_BIC; | |
11805 | break; | |
11806 | ||
9cd7b720 MW |
11807 | default: |
11808 | /* The operation can't be done with atomic instructions. */ | |
11809 | gcc_unreachable (); | |
11810 | } | |
641c2f8b MW |
11811 | |
11812 | aarch64_emit_atomic_load_op (ldop_code, mode, out_data, src, mem, model_rtx); | |
68729b06 MW |
11813 | |
11814 | /* If necessary, calculate the data in memory after the update by redoing the | |
11815 | operation from values in registers. */ | |
11816 | if (!out_result) | |
11817 | return; | |
11818 | ||
11819 | if (short_mode) | |
11820 | { | |
11821 | src = gen_lowpart (wmode, src); | |
11822 | out_data = gen_lowpart (wmode, out_data); | |
11823 | out_result = gen_lowpart (wmode, out_result); | |
11824 | } | |
11825 | ||
11826 | x = NULL_RTX; | |
11827 | ||
11828 | switch (code) | |
11829 | { | |
11830 | case MINUS: | |
11831 | case PLUS: | |
11832 | x = gen_rtx_PLUS (wmode, out_data, src); | |
11833 | break; | |
11834 | case IOR: | |
11835 | x = gen_rtx_IOR (wmode, out_data, src); | |
11836 | break; | |
11837 | case XOR: | |
11838 | x = gen_rtx_XOR (wmode, out_data, src); | |
11839 | break; | |
11840 | case AND: | |
11841 | aarch64_emit_bic (wmode, out_result, out_data, src, 0); | |
11842 | return; | |
11843 | default: | |
11844 | gcc_unreachable (); | |
11845 | } | |
11846 | ||
11847 | emit_set_insn (out_result, x); | |
11848 | ||
11849 | return; | |
9cd7b720 MW |
11850 | } |
11851 | ||
0462169c SN |
11852 | /* Split an atomic operation. */ |
11853 | ||
11854 | void | |
11855 | aarch64_split_atomic_op (enum rtx_code code, rtx old_out, rtx new_out, rtx mem, | |
9cd7b720 | 11856 | rtx value, rtx model_rtx, rtx cond) |
0462169c | 11857 | { |
ef4bddc2 RS |
11858 | machine_mode mode = GET_MODE (mem); |
11859 | machine_mode wmode = (mode == DImode ? DImode : SImode); | |
f70fb3b6 MW |
11860 | const enum memmodel model = memmodel_from_int (INTVAL (model_rtx)); |
11861 | const bool is_sync = is_mm_sync (model); | |
5d8a22a5 DM |
11862 | rtx_code_label *label; |
11863 | rtx x; | |
0462169c | 11864 | |
9cd7b720 | 11865 | /* Split the atomic operation into a sequence. */ |
0462169c SN |
11866 | label = gen_label_rtx (); |
11867 | emit_label (label); | |
11868 | ||
11869 | if (new_out) | |
11870 | new_out = gen_lowpart (wmode, new_out); | |
11871 | if (old_out) | |
11872 | old_out = gen_lowpart (wmode, old_out); | |
11873 | else | |
11874 | old_out = new_out; | |
11875 | value = simplify_gen_subreg (wmode, value, mode, 0); | |
11876 | ||
f70fb3b6 MW |
11877 | /* The initial load can be relaxed for a __sync operation since a final |
11878 | barrier will be emitted to stop code hoisting. */ | |
11879 | if (is_sync) | |
11880 | aarch64_emit_load_exclusive (mode, old_out, mem, | |
11881 | GEN_INT (MEMMODEL_RELAXED)); | |
11882 | else | |
11883 | aarch64_emit_load_exclusive (mode, old_out, mem, model_rtx); | |
0462169c SN |
11884 | |
11885 | switch (code) | |
11886 | { | |
11887 | case SET: | |
11888 | new_out = value; | |
11889 | break; | |
11890 | ||
11891 | case NOT: | |
11892 | x = gen_rtx_AND (wmode, old_out, value); | |
f7df4a84 | 11893 | emit_insn (gen_rtx_SET (new_out, x)); |
0462169c | 11894 | x = gen_rtx_NOT (wmode, new_out); |
f7df4a84 | 11895 | emit_insn (gen_rtx_SET (new_out, x)); |
0462169c SN |
11896 | break; |
11897 | ||
11898 | case MINUS: | |
11899 | if (CONST_INT_P (value)) | |
11900 | { | |
11901 | value = GEN_INT (-INTVAL (value)); | |
11902 | code = PLUS; | |
11903 | } | |
11904 | /* Fall through. */ | |
11905 | ||
11906 | default: | |
11907 | x = gen_rtx_fmt_ee (code, wmode, old_out, value); | |
f7df4a84 | 11908 | emit_insn (gen_rtx_SET (new_out, x)); |
0462169c SN |
11909 | break; |
11910 | } | |
11911 | ||
11912 | aarch64_emit_store_exclusive (mode, cond, mem, | |
11913 | gen_lowpart (mode, new_out), model_rtx); | |
11914 | ||
11915 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
11916 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
11917 | gen_rtx_LABEL_REF (Pmode, label), pc_rtx); | |
f7df4a84 | 11918 | aarch64_emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); |
f70fb3b6 MW |
11919 | |
11920 | /* Emit any final barrier needed for a __sync operation. */ | |
11921 | if (is_sync) | |
11922 | aarch64_emit_post_barrier (model); | |
0462169c SN |
11923 | } |
11924 | ||
c2ec330c AL |
11925 | static void |
11926 | aarch64_init_libfuncs (void) | |
11927 | { | |
11928 | /* Half-precision float operations. The compiler handles all operations | |
11929 | with NULL libfuncs by converting to SFmode. */ | |
11930 | ||
11931 | /* Conversions. */ | |
11932 | set_conv_libfunc (trunc_optab, HFmode, SFmode, "__gnu_f2h_ieee"); | |
11933 | set_conv_libfunc (sext_optab, SFmode, HFmode, "__gnu_h2f_ieee"); | |
11934 | ||
11935 | /* Arithmetic. */ | |
11936 | set_optab_libfunc (add_optab, HFmode, NULL); | |
11937 | set_optab_libfunc (sdiv_optab, HFmode, NULL); | |
11938 | set_optab_libfunc (smul_optab, HFmode, NULL); | |
11939 | set_optab_libfunc (neg_optab, HFmode, NULL); | |
11940 | set_optab_libfunc (sub_optab, HFmode, NULL); | |
11941 | ||
11942 | /* Comparisons. */ | |
11943 | set_optab_libfunc (eq_optab, HFmode, NULL); | |
11944 | set_optab_libfunc (ne_optab, HFmode, NULL); | |
11945 | set_optab_libfunc (lt_optab, HFmode, NULL); | |
11946 | set_optab_libfunc (le_optab, HFmode, NULL); | |
11947 | set_optab_libfunc (ge_optab, HFmode, NULL); | |
11948 | set_optab_libfunc (gt_optab, HFmode, NULL); | |
11949 | set_optab_libfunc (unord_optab, HFmode, NULL); | |
11950 | } | |
11951 | ||
43e9d192 | 11952 | /* Target hook for c_mode_for_suffix. */ |
ef4bddc2 | 11953 | static machine_mode |
43e9d192 IB |
11954 | aarch64_c_mode_for_suffix (char suffix) |
11955 | { | |
11956 | if (suffix == 'q') | |
11957 | return TFmode; | |
11958 | ||
11959 | return VOIDmode; | |
11960 | } | |
11961 | ||
3520f7cc JG |
11962 | /* We can only represent floating point constants which will fit in |
11963 | "quarter-precision" values. These values are characterised by | |
11964 | a sign bit, a 4-bit mantissa and a 3-bit exponent. And are given | |
11965 | by: | |
11966 | ||
11967 | (-1)^s * (n/16) * 2^r | |
11968 | ||
11969 | Where: | |
11970 | 's' is the sign bit. | |
11971 | 'n' is an integer in the range 16 <= n <= 31. | |
11972 | 'r' is an integer in the range -3 <= r <= 4. */ | |
11973 | ||
11974 | /* Return true iff X can be represented by a quarter-precision | |
11975 | floating point immediate operand X. Note, we cannot represent 0.0. */ | |
11976 | bool | |
11977 | aarch64_float_const_representable_p (rtx x) | |
11978 | { | |
11979 | /* This represents our current view of how many bits | |
11980 | make up the mantissa. */ | |
11981 | int point_pos = 2 * HOST_BITS_PER_WIDE_INT - 1; | |
ba96cdfb | 11982 | int exponent; |
3520f7cc | 11983 | unsigned HOST_WIDE_INT mantissa, mask; |
3520f7cc | 11984 | REAL_VALUE_TYPE r, m; |
807e902e | 11985 | bool fail; |
3520f7cc JG |
11986 | |
11987 | if (!CONST_DOUBLE_P (x)) | |
11988 | return false; | |
11989 | ||
c2ec330c AL |
11990 | /* We don't support HFmode constants yet. */ |
11991 | if (GET_MODE (x) == VOIDmode || GET_MODE (x) == HFmode) | |
94bfa2da TV |
11992 | return false; |
11993 | ||
34a72c33 | 11994 | r = *CONST_DOUBLE_REAL_VALUE (x); |
3520f7cc JG |
11995 | |
11996 | /* We cannot represent infinities, NaNs or +/-zero. We won't | |
11997 | know if we have +zero until we analyse the mantissa, but we | |
11998 | can reject the other invalid values. */ | |
11999 | if (REAL_VALUE_ISINF (r) || REAL_VALUE_ISNAN (r) | |
12000 | || REAL_VALUE_MINUS_ZERO (r)) | |
12001 | return false; | |
12002 | ||
ba96cdfb | 12003 | /* Extract exponent. */ |
3520f7cc JG |
12004 | r = real_value_abs (&r); |
12005 | exponent = REAL_EXP (&r); | |
12006 | ||
12007 | /* For the mantissa, we expand into two HOST_WIDE_INTS, apart from the | |
12008 | highest (sign) bit, with a fixed binary point at bit point_pos. | |
12009 | m1 holds the low part of the mantissa, m2 the high part. | |
12010 | WARNING: If we ever have a representation using more than 2 * H_W_I - 1 | |
12011 | bits for the mantissa, this can fail (low bits will be lost). */ | |
12012 | real_ldexp (&m, &r, point_pos - exponent); | |
807e902e | 12013 | wide_int w = real_to_integer (&m, &fail, HOST_BITS_PER_WIDE_INT * 2); |
3520f7cc JG |
12014 | |
12015 | /* If the low part of the mantissa has bits set we cannot represent | |
12016 | the value. */ | |
807e902e | 12017 | if (w.elt (0) != 0) |
3520f7cc JG |
12018 | return false; |
12019 | /* We have rejected the lower HOST_WIDE_INT, so update our | |
12020 | understanding of how many bits lie in the mantissa and | |
12021 | look only at the high HOST_WIDE_INT. */ | |
807e902e | 12022 | mantissa = w.elt (1); |
3520f7cc JG |
12023 | point_pos -= HOST_BITS_PER_WIDE_INT; |
12024 | ||
12025 | /* We can only represent values with a mantissa of the form 1.xxxx. */ | |
12026 | mask = ((unsigned HOST_WIDE_INT)1 << (point_pos - 5)) - 1; | |
12027 | if ((mantissa & mask) != 0) | |
12028 | return false; | |
12029 | ||
12030 | /* Having filtered unrepresentable values, we may now remove all | |
12031 | but the highest 5 bits. */ | |
12032 | mantissa >>= point_pos - 5; | |
12033 | ||
12034 | /* We cannot represent the value 0.0, so reject it. This is handled | |
12035 | elsewhere. */ | |
12036 | if (mantissa == 0) | |
12037 | return false; | |
12038 | ||
12039 | /* Then, as bit 4 is always set, we can mask it off, leaving | |
12040 | the mantissa in the range [0, 15]. */ | |
12041 | mantissa &= ~(1 << 4); | |
12042 | gcc_assert (mantissa <= 15); | |
12043 | ||
12044 | /* GCC internally does not use IEEE754-like encoding (where normalized | |
12045 | significands are in the range [1, 2). GCC uses [0.5, 1) (see real.c). | |
12046 | Our mantissa values are shifted 4 places to the left relative to | |
12047 | normalized IEEE754 so we must modify the exponent returned by REAL_EXP | |
12048 | by 5 places to correct for GCC's representation. */ | |
12049 | exponent = 5 - exponent; | |
12050 | ||
12051 | return (exponent >= 0 && exponent <= 7); | |
12052 | } | |
12053 | ||
12054 | char* | |
81c2dfb9 | 12055 | aarch64_output_simd_mov_immediate (rtx const_vector, |
ef4bddc2 | 12056 | machine_mode mode, |
3520f7cc JG |
12057 | unsigned width) |
12058 | { | |
3ea63f60 | 12059 | bool is_valid; |
3520f7cc | 12060 | static char templ[40]; |
3520f7cc | 12061 | const char *mnemonic; |
e4f0f84d | 12062 | const char *shift_op; |
3520f7cc | 12063 | unsigned int lane_count = 0; |
81c2dfb9 | 12064 | char element_char; |
3520f7cc | 12065 | |
e4f0f84d | 12066 | struct simd_immediate_info info = { NULL_RTX, 0, 0, false, false }; |
48063b9d IB |
12067 | |
12068 | /* This will return true to show const_vector is legal for use as either | |
12069 | a AdvSIMD MOVI instruction (or, implicitly, MVNI) immediate. It will | |
12070 | also update INFO to show how the immediate should be generated. */ | |
81c2dfb9 | 12071 | is_valid = aarch64_simd_valid_immediate (const_vector, mode, false, &info); |
3520f7cc JG |
12072 | gcc_assert (is_valid); |
12073 | ||
81c2dfb9 | 12074 | element_char = sizetochar (info.element_width); |
48063b9d IB |
12075 | lane_count = width / info.element_width; |
12076 | ||
3520f7cc | 12077 | mode = GET_MODE_INNER (mode); |
0d8e1702 | 12078 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
3520f7cc | 12079 | { |
48063b9d | 12080 | gcc_assert (info.shift == 0 && ! info.mvn); |
0d8e1702 KT |
12081 | /* For FP zero change it to a CONST_INT 0 and use the integer SIMD |
12082 | move immediate path. */ | |
48063b9d IB |
12083 | if (aarch64_float_const_zero_rtx_p (info.value)) |
12084 | info.value = GEN_INT (0); | |
12085 | else | |
12086 | { | |
83faf7d0 | 12087 | const unsigned int buf_size = 20; |
48063b9d | 12088 | char float_buf[buf_size] = {'\0'}; |
34a72c33 RS |
12089 | real_to_decimal_for_mode (float_buf, |
12090 | CONST_DOUBLE_REAL_VALUE (info.value), | |
12091 | buf_size, buf_size, 1, mode); | |
48063b9d IB |
12092 | |
12093 | if (lane_count == 1) | |
12094 | snprintf (templ, sizeof (templ), "fmov\t%%d0, %s", float_buf); | |
12095 | else | |
12096 | snprintf (templ, sizeof (templ), "fmov\t%%0.%d%c, %s", | |
81c2dfb9 | 12097 | lane_count, element_char, float_buf); |
48063b9d IB |
12098 | return templ; |
12099 | } | |
3520f7cc | 12100 | } |
3520f7cc | 12101 | |
48063b9d | 12102 | mnemonic = info.mvn ? "mvni" : "movi"; |
e4f0f84d | 12103 | shift_op = info.msl ? "msl" : "lsl"; |
3520f7cc | 12104 | |
0d8e1702 | 12105 | gcc_assert (CONST_INT_P (info.value)); |
3520f7cc | 12106 | if (lane_count == 1) |
48063b9d IB |
12107 | snprintf (templ, sizeof (templ), "%s\t%%d0, " HOST_WIDE_INT_PRINT_HEX, |
12108 | mnemonic, UINTVAL (info.value)); | |
12109 | else if (info.shift) | |
12110 | snprintf (templ, sizeof (templ), "%s\t%%0.%d%c, " HOST_WIDE_INT_PRINT_HEX | |
e4f0f84d TB |
12111 | ", %s %d", mnemonic, lane_count, element_char, |
12112 | UINTVAL (info.value), shift_op, info.shift); | |
3520f7cc | 12113 | else |
48063b9d | 12114 | snprintf (templ, sizeof (templ), "%s\t%%0.%d%c, " HOST_WIDE_INT_PRINT_HEX, |
81c2dfb9 | 12115 | mnemonic, lane_count, element_char, UINTVAL (info.value)); |
3520f7cc JG |
12116 | return templ; |
12117 | } | |
12118 | ||
b7342d25 IB |
12119 | char* |
12120 | aarch64_output_scalar_simd_mov_immediate (rtx immediate, | |
ef4bddc2 | 12121 | machine_mode mode) |
b7342d25 | 12122 | { |
ef4bddc2 | 12123 | machine_mode vmode; |
b7342d25 IB |
12124 | |
12125 | gcc_assert (!VECTOR_MODE_P (mode)); | |
12126 | vmode = aarch64_simd_container_mode (mode, 64); | |
12127 | rtx v_op = aarch64_simd_gen_const_vector_dup (vmode, INTVAL (immediate)); | |
12128 | return aarch64_output_simd_mov_immediate (v_op, vmode, 64); | |
12129 | } | |
12130 | ||
88b08073 JG |
12131 | /* Split operands into moves from op[1] + op[2] into op[0]. */ |
12132 | ||
12133 | void | |
12134 | aarch64_split_combinev16qi (rtx operands[3]) | |
12135 | { | |
12136 | unsigned int dest = REGNO (operands[0]); | |
12137 | unsigned int src1 = REGNO (operands[1]); | |
12138 | unsigned int src2 = REGNO (operands[2]); | |
ef4bddc2 | 12139 | machine_mode halfmode = GET_MODE (operands[1]); |
88b08073 JG |
12140 | unsigned int halfregs = HARD_REGNO_NREGS (src1, halfmode); |
12141 | rtx destlo, desthi; | |
12142 | ||
12143 | gcc_assert (halfmode == V16QImode); | |
12144 | ||
12145 | if (src1 == dest && src2 == dest + halfregs) | |
12146 | { | |
12147 | /* No-op move. Can't split to nothing; emit something. */ | |
12148 | emit_note (NOTE_INSN_DELETED); | |
12149 | return; | |
12150 | } | |
12151 | ||
12152 | /* Preserve register attributes for variable tracking. */ | |
12153 | destlo = gen_rtx_REG_offset (operands[0], halfmode, dest, 0); | |
12154 | desthi = gen_rtx_REG_offset (operands[0], halfmode, dest + halfregs, | |
12155 | GET_MODE_SIZE (halfmode)); | |
12156 | ||
12157 | /* Special case of reversed high/low parts. */ | |
12158 | if (reg_overlap_mentioned_p (operands[2], destlo) | |
12159 | && reg_overlap_mentioned_p (operands[1], desthi)) | |
12160 | { | |
12161 | emit_insn (gen_xorv16qi3 (operands[1], operands[1], operands[2])); | |
12162 | emit_insn (gen_xorv16qi3 (operands[2], operands[1], operands[2])); | |
12163 | emit_insn (gen_xorv16qi3 (operands[1], operands[1], operands[2])); | |
12164 | } | |
12165 | else if (!reg_overlap_mentioned_p (operands[2], destlo)) | |
12166 | { | |
12167 | /* Try to avoid unnecessary moves if part of the result | |
12168 | is in the right place already. */ | |
12169 | if (src1 != dest) | |
12170 | emit_move_insn (destlo, operands[1]); | |
12171 | if (src2 != dest + halfregs) | |
12172 | emit_move_insn (desthi, operands[2]); | |
12173 | } | |
12174 | else | |
12175 | { | |
12176 | if (src2 != dest + halfregs) | |
12177 | emit_move_insn (desthi, operands[2]); | |
12178 | if (src1 != dest) | |
12179 | emit_move_insn (destlo, operands[1]); | |
12180 | } | |
12181 | } | |
12182 | ||
12183 | /* vec_perm support. */ | |
12184 | ||
12185 | #define MAX_VECT_LEN 16 | |
12186 | ||
12187 | struct expand_vec_perm_d | |
12188 | { | |
12189 | rtx target, op0, op1; | |
12190 | unsigned char perm[MAX_VECT_LEN]; | |
ef4bddc2 | 12191 | machine_mode vmode; |
88b08073 JG |
12192 | unsigned char nelt; |
12193 | bool one_vector_p; | |
12194 | bool testing_p; | |
12195 | }; | |
12196 | ||
12197 | /* Generate a variable permutation. */ | |
12198 | ||
12199 | static void | |
12200 | aarch64_expand_vec_perm_1 (rtx target, rtx op0, rtx op1, rtx sel) | |
12201 | { | |
ef4bddc2 | 12202 | machine_mode vmode = GET_MODE (target); |
88b08073 JG |
12203 | bool one_vector_p = rtx_equal_p (op0, op1); |
12204 | ||
12205 | gcc_checking_assert (vmode == V8QImode || vmode == V16QImode); | |
12206 | gcc_checking_assert (GET_MODE (op0) == vmode); | |
12207 | gcc_checking_assert (GET_MODE (op1) == vmode); | |
12208 | gcc_checking_assert (GET_MODE (sel) == vmode); | |
12209 | gcc_checking_assert (TARGET_SIMD); | |
12210 | ||
12211 | if (one_vector_p) | |
12212 | { | |
12213 | if (vmode == V8QImode) | |
12214 | { | |
12215 | /* Expand the argument to a V16QI mode by duplicating it. */ | |
12216 | rtx pair = gen_reg_rtx (V16QImode); | |
12217 | emit_insn (gen_aarch64_combinev8qi (pair, op0, op0)); | |
12218 | emit_insn (gen_aarch64_tbl1v8qi (target, pair, sel)); | |
12219 | } | |
12220 | else | |
12221 | { | |
12222 | emit_insn (gen_aarch64_tbl1v16qi (target, op0, sel)); | |
12223 | } | |
12224 | } | |
12225 | else | |
12226 | { | |
12227 | rtx pair; | |
12228 | ||
12229 | if (vmode == V8QImode) | |
12230 | { | |
12231 | pair = gen_reg_rtx (V16QImode); | |
12232 | emit_insn (gen_aarch64_combinev8qi (pair, op0, op1)); | |
12233 | emit_insn (gen_aarch64_tbl1v8qi (target, pair, sel)); | |
12234 | } | |
12235 | else | |
12236 | { | |
12237 | pair = gen_reg_rtx (OImode); | |
12238 | emit_insn (gen_aarch64_combinev16qi (pair, op0, op1)); | |
12239 | emit_insn (gen_aarch64_tbl2v16qi (target, pair, sel)); | |
12240 | } | |
12241 | } | |
12242 | } | |
12243 | ||
12244 | void | |
12245 | aarch64_expand_vec_perm (rtx target, rtx op0, rtx op1, rtx sel) | |
12246 | { | |
ef4bddc2 | 12247 | machine_mode vmode = GET_MODE (target); |
c9d1a16a | 12248 | unsigned int nelt = GET_MODE_NUNITS (vmode); |
88b08073 | 12249 | bool one_vector_p = rtx_equal_p (op0, op1); |
f7c4e5b8 | 12250 | rtx mask; |
88b08073 JG |
12251 | |
12252 | /* The TBL instruction does not use a modulo index, so we must take care | |
12253 | of that ourselves. */ | |
f7c4e5b8 AL |
12254 | mask = aarch64_simd_gen_const_vector_dup (vmode, |
12255 | one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
88b08073 JG |
12256 | sel = expand_simple_binop (vmode, AND, sel, mask, NULL, 0, OPTAB_LIB_WIDEN); |
12257 | ||
f7c4e5b8 AL |
12258 | /* For big-endian, we also need to reverse the index within the vector |
12259 | (but not which vector). */ | |
12260 | if (BYTES_BIG_ENDIAN) | |
12261 | { | |
12262 | /* If one_vector_p, mask is a vector of (nelt - 1)'s already. */ | |
12263 | if (!one_vector_p) | |
12264 | mask = aarch64_simd_gen_const_vector_dup (vmode, nelt - 1); | |
12265 | sel = expand_simple_binop (vmode, XOR, sel, mask, | |
12266 | NULL, 0, OPTAB_LIB_WIDEN); | |
12267 | } | |
88b08073 JG |
12268 | aarch64_expand_vec_perm_1 (target, op0, op1, sel); |
12269 | } | |
12270 | ||
cc4d934f JG |
12271 | /* Recognize patterns suitable for the TRN instructions. */ |
12272 | static bool | |
12273 | aarch64_evpc_trn (struct expand_vec_perm_d *d) | |
12274 | { | |
12275 | unsigned int i, odd, mask, nelt = d->nelt; | |
12276 | rtx out, in0, in1, x; | |
12277 | rtx (*gen) (rtx, rtx, rtx); | |
ef4bddc2 | 12278 | machine_mode vmode = d->vmode; |
cc4d934f JG |
12279 | |
12280 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
12281 | return false; | |
12282 | ||
12283 | /* Note that these are little-endian tests. | |
12284 | We correct for big-endian later. */ | |
12285 | if (d->perm[0] == 0) | |
12286 | odd = 0; | |
12287 | else if (d->perm[0] == 1) | |
12288 | odd = 1; | |
12289 | else | |
12290 | return false; | |
12291 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
12292 | ||
12293 | for (i = 0; i < nelt; i += 2) | |
12294 | { | |
12295 | if (d->perm[i] != i + odd) | |
12296 | return false; | |
12297 | if (d->perm[i + 1] != ((i + nelt + odd) & mask)) | |
12298 | return false; | |
12299 | } | |
12300 | ||
12301 | /* Success! */ | |
12302 | if (d->testing_p) | |
12303 | return true; | |
12304 | ||
12305 | in0 = d->op0; | |
12306 | in1 = d->op1; | |
12307 | if (BYTES_BIG_ENDIAN) | |
12308 | { | |
12309 | x = in0, in0 = in1, in1 = x; | |
12310 | odd = !odd; | |
12311 | } | |
12312 | out = d->target; | |
12313 | ||
12314 | if (odd) | |
12315 | { | |
12316 | switch (vmode) | |
12317 | { | |
12318 | case V16QImode: gen = gen_aarch64_trn2v16qi; break; | |
12319 | case V8QImode: gen = gen_aarch64_trn2v8qi; break; | |
12320 | case V8HImode: gen = gen_aarch64_trn2v8hi; break; | |
12321 | case V4HImode: gen = gen_aarch64_trn2v4hi; break; | |
12322 | case V4SImode: gen = gen_aarch64_trn2v4si; break; | |
12323 | case V2SImode: gen = gen_aarch64_trn2v2si; break; | |
12324 | case V2DImode: gen = gen_aarch64_trn2v2di; break; | |
12325 | case V4SFmode: gen = gen_aarch64_trn2v4sf; break; | |
12326 | case V2SFmode: gen = gen_aarch64_trn2v2sf; break; | |
12327 | case V2DFmode: gen = gen_aarch64_trn2v2df; break; | |
12328 | default: | |
12329 | return false; | |
12330 | } | |
12331 | } | |
12332 | else | |
12333 | { | |
12334 | switch (vmode) | |
12335 | { | |
12336 | case V16QImode: gen = gen_aarch64_trn1v16qi; break; | |
12337 | case V8QImode: gen = gen_aarch64_trn1v8qi; break; | |
12338 | case V8HImode: gen = gen_aarch64_trn1v8hi; break; | |
12339 | case V4HImode: gen = gen_aarch64_trn1v4hi; break; | |
12340 | case V4SImode: gen = gen_aarch64_trn1v4si; break; | |
12341 | case V2SImode: gen = gen_aarch64_trn1v2si; break; | |
12342 | case V2DImode: gen = gen_aarch64_trn1v2di; break; | |
12343 | case V4SFmode: gen = gen_aarch64_trn1v4sf; break; | |
12344 | case V2SFmode: gen = gen_aarch64_trn1v2sf; break; | |
12345 | case V2DFmode: gen = gen_aarch64_trn1v2df; break; | |
12346 | default: | |
12347 | return false; | |
12348 | } | |
12349 | } | |
12350 | ||
12351 | emit_insn (gen (out, in0, in1)); | |
12352 | return true; | |
12353 | } | |
12354 | ||
12355 | /* Recognize patterns suitable for the UZP instructions. */ | |
12356 | static bool | |
12357 | aarch64_evpc_uzp (struct expand_vec_perm_d *d) | |
12358 | { | |
12359 | unsigned int i, odd, mask, nelt = d->nelt; | |
12360 | rtx out, in0, in1, x; | |
12361 | rtx (*gen) (rtx, rtx, rtx); | |
ef4bddc2 | 12362 | machine_mode vmode = d->vmode; |
cc4d934f JG |
12363 | |
12364 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
12365 | return false; | |
12366 | ||
12367 | /* Note that these are little-endian tests. | |
12368 | We correct for big-endian later. */ | |
12369 | if (d->perm[0] == 0) | |
12370 | odd = 0; | |
12371 | else if (d->perm[0] == 1) | |
12372 | odd = 1; | |
12373 | else | |
12374 | return false; | |
12375 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
12376 | ||
12377 | for (i = 0; i < nelt; i++) | |
12378 | { | |
12379 | unsigned elt = (i * 2 + odd) & mask; | |
12380 | if (d->perm[i] != elt) | |
12381 | return false; | |
12382 | } | |
12383 | ||
12384 | /* Success! */ | |
12385 | if (d->testing_p) | |
12386 | return true; | |
12387 | ||
12388 | in0 = d->op0; | |
12389 | in1 = d->op1; | |
12390 | if (BYTES_BIG_ENDIAN) | |
12391 | { | |
12392 | x = in0, in0 = in1, in1 = x; | |
12393 | odd = !odd; | |
12394 | } | |
12395 | out = d->target; | |
12396 | ||
12397 | if (odd) | |
12398 | { | |
12399 | switch (vmode) | |
12400 | { | |
12401 | case V16QImode: gen = gen_aarch64_uzp2v16qi; break; | |
12402 | case V8QImode: gen = gen_aarch64_uzp2v8qi; break; | |
12403 | case V8HImode: gen = gen_aarch64_uzp2v8hi; break; | |
12404 | case V4HImode: gen = gen_aarch64_uzp2v4hi; break; | |
12405 | case V4SImode: gen = gen_aarch64_uzp2v4si; break; | |
12406 | case V2SImode: gen = gen_aarch64_uzp2v2si; break; | |
12407 | case V2DImode: gen = gen_aarch64_uzp2v2di; break; | |
12408 | case V4SFmode: gen = gen_aarch64_uzp2v4sf; break; | |
12409 | case V2SFmode: gen = gen_aarch64_uzp2v2sf; break; | |
12410 | case V2DFmode: gen = gen_aarch64_uzp2v2df; break; | |
12411 | default: | |
12412 | return false; | |
12413 | } | |
12414 | } | |
12415 | else | |
12416 | { | |
12417 | switch (vmode) | |
12418 | { | |
12419 | case V16QImode: gen = gen_aarch64_uzp1v16qi; break; | |
12420 | case V8QImode: gen = gen_aarch64_uzp1v8qi; break; | |
12421 | case V8HImode: gen = gen_aarch64_uzp1v8hi; break; | |
12422 | case V4HImode: gen = gen_aarch64_uzp1v4hi; break; | |
12423 | case V4SImode: gen = gen_aarch64_uzp1v4si; break; | |
12424 | case V2SImode: gen = gen_aarch64_uzp1v2si; break; | |
12425 | case V2DImode: gen = gen_aarch64_uzp1v2di; break; | |
12426 | case V4SFmode: gen = gen_aarch64_uzp1v4sf; break; | |
12427 | case V2SFmode: gen = gen_aarch64_uzp1v2sf; break; | |
12428 | case V2DFmode: gen = gen_aarch64_uzp1v2df; break; | |
12429 | default: | |
12430 | return false; | |
12431 | } | |
12432 | } | |
12433 | ||
12434 | emit_insn (gen (out, in0, in1)); | |
12435 | return true; | |
12436 | } | |
12437 | ||
12438 | /* Recognize patterns suitable for the ZIP instructions. */ | |
12439 | static bool | |
12440 | aarch64_evpc_zip (struct expand_vec_perm_d *d) | |
12441 | { | |
12442 | unsigned int i, high, mask, nelt = d->nelt; | |
12443 | rtx out, in0, in1, x; | |
12444 | rtx (*gen) (rtx, rtx, rtx); | |
ef4bddc2 | 12445 | machine_mode vmode = d->vmode; |
cc4d934f JG |
12446 | |
12447 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
12448 | return false; | |
12449 | ||
12450 | /* Note that these are little-endian tests. | |
12451 | We correct for big-endian later. */ | |
12452 | high = nelt / 2; | |
12453 | if (d->perm[0] == high) | |
12454 | /* Do Nothing. */ | |
12455 | ; | |
12456 | else if (d->perm[0] == 0) | |
12457 | high = 0; | |
12458 | else | |
12459 | return false; | |
12460 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
12461 | ||
12462 | for (i = 0; i < nelt / 2; i++) | |
12463 | { | |
12464 | unsigned elt = (i + high) & mask; | |
12465 | if (d->perm[i * 2] != elt) | |
12466 | return false; | |
12467 | elt = (elt + nelt) & mask; | |
12468 | if (d->perm[i * 2 + 1] != elt) | |
12469 | return false; | |
12470 | } | |
12471 | ||
12472 | /* Success! */ | |
12473 | if (d->testing_p) | |
12474 | return true; | |
12475 | ||
12476 | in0 = d->op0; | |
12477 | in1 = d->op1; | |
12478 | if (BYTES_BIG_ENDIAN) | |
12479 | { | |
12480 | x = in0, in0 = in1, in1 = x; | |
12481 | high = !high; | |
12482 | } | |
12483 | out = d->target; | |
12484 | ||
12485 | if (high) | |
12486 | { | |
12487 | switch (vmode) | |
12488 | { | |
12489 | case V16QImode: gen = gen_aarch64_zip2v16qi; break; | |
12490 | case V8QImode: gen = gen_aarch64_zip2v8qi; break; | |
12491 | case V8HImode: gen = gen_aarch64_zip2v8hi; break; | |
12492 | case V4HImode: gen = gen_aarch64_zip2v4hi; break; | |
12493 | case V4SImode: gen = gen_aarch64_zip2v4si; break; | |
12494 | case V2SImode: gen = gen_aarch64_zip2v2si; break; | |
12495 | case V2DImode: gen = gen_aarch64_zip2v2di; break; | |
12496 | case V4SFmode: gen = gen_aarch64_zip2v4sf; break; | |
12497 | case V2SFmode: gen = gen_aarch64_zip2v2sf; break; | |
12498 | case V2DFmode: gen = gen_aarch64_zip2v2df; break; | |
12499 | default: | |
12500 | return false; | |
12501 | } | |
12502 | } | |
12503 | else | |
12504 | { | |
12505 | switch (vmode) | |
12506 | { | |
12507 | case V16QImode: gen = gen_aarch64_zip1v16qi; break; | |
12508 | case V8QImode: gen = gen_aarch64_zip1v8qi; break; | |
12509 | case V8HImode: gen = gen_aarch64_zip1v8hi; break; | |
12510 | case V4HImode: gen = gen_aarch64_zip1v4hi; break; | |
12511 | case V4SImode: gen = gen_aarch64_zip1v4si; break; | |
12512 | case V2SImode: gen = gen_aarch64_zip1v2si; break; | |
12513 | case V2DImode: gen = gen_aarch64_zip1v2di; break; | |
12514 | case V4SFmode: gen = gen_aarch64_zip1v4sf; break; | |
12515 | case V2SFmode: gen = gen_aarch64_zip1v2sf; break; | |
12516 | case V2DFmode: gen = gen_aarch64_zip1v2df; break; | |
12517 | default: | |
12518 | return false; | |
12519 | } | |
12520 | } | |
12521 | ||
12522 | emit_insn (gen (out, in0, in1)); | |
12523 | return true; | |
12524 | } | |
12525 | ||
ae0533da AL |
12526 | /* Recognize patterns for the EXT insn. */ |
12527 | ||
12528 | static bool | |
12529 | aarch64_evpc_ext (struct expand_vec_perm_d *d) | |
12530 | { | |
12531 | unsigned int i, nelt = d->nelt; | |
12532 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
12533 | rtx offset; | |
12534 | ||
12535 | unsigned int location = d->perm[0]; /* Always < nelt. */ | |
12536 | ||
12537 | /* Check if the extracted indices are increasing by one. */ | |
12538 | for (i = 1; i < nelt; i++) | |
12539 | { | |
12540 | unsigned int required = location + i; | |
12541 | if (d->one_vector_p) | |
12542 | { | |
12543 | /* We'll pass the same vector in twice, so allow indices to wrap. */ | |
12544 | required &= (nelt - 1); | |
12545 | } | |
12546 | if (d->perm[i] != required) | |
12547 | return false; | |
12548 | } | |
12549 | ||
ae0533da AL |
12550 | switch (d->vmode) |
12551 | { | |
12552 | case V16QImode: gen = gen_aarch64_extv16qi; break; | |
12553 | case V8QImode: gen = gen_aarch64_extv8qi; break; | |
12554 | case V4HImode: gen = gen_aarch64_extv4hi; break; | |
12555 | case V8HImode: gen = gen_aarch64_extv8hi; break; | |
12556 | case V2SImode: gen = gen_aarch64_extv2si; break; | |
12557 | case V4SImode: gen = gen_aarch64_extv4si; break; | |
12558 | case V2SFmode: gen = gen_aarch64_extv2sf; break; | |
12559 | case V4SFmode: gen = gen_aarch64_extv4sf; break; | |
12560 | case V2DImode: gen = gen_aarch64_extv2di; break; | |
12561 | case V2DFmode: gen = gen_aarch64_extv2df; break; | |
12562 | default: | |
12563 | return false; | |
12564 | } | |
12565 | ||
12566 | /* Success! */ | |
12567 | if (d->testing_p) | |
12568 | return true; | |
12569 | ||
b31e65bb AL |
12570 | /* The case where (location == 0) is a no-op for both big- and little-endian, |
12571 | and is removed by the mid-end at optimization levels -O1 and higher. */ | |
12572 | ||
12573 | if (BYTES_BIG_ENDIAN && (location != 0)) | |
ae0533da AL |
12574 | { |
12575 | /* After setup, we want the high elements of the first vector (stored | |
12576 | at the LSB end of the register), and the low elements of the second | |
12577 | vector (stored at the MSB end of the register). So swap. */ | |
cb5c6c29 | 12578 | std::swap (d->op0, d->op1); |
ae0533da AL |
12579 | /* location != 0 (above), so safe to assume (nelt - location) < nelt. */ |
12580 | location = nelt - location; | |
12581 | } | |
12582 | ||
12583 | offset = GEN_INT (location); | |
12584 | emit_insn (gen (d->target, d->op0, d->op1, offset)); | |
12585 | return true; | |
12586 | } | |
12587 | ||
923fcec3 AL |
12588 | /* Recognize patterns for the REV insns. */ |
12589 | ||
12590 | static bool | |
12591 | aarch64_evpc_rev (struct expand_vec_perm_d *d) | |
12592 | { | |
12593 | unsigned int i, j, diff, nelt = d->nelt; | |
12594 | rtx (*gen) (rtx, rtx); | |
12595 | ||
12596 | if (!d->one_vector_p) | |
12597 | return false; | |
12598 | ||
12599 | diff = d->perm[0]; | |
12600 | switch (diff) | |
12601 | { | |
12602 | case 7: | |
12603 | switch (d->vmode) | |
12604 | { | |
12605 | case V16QImode: gen = gen_aarch64_rev64v16qi; break; | |
12606 | case V8QImode: gen = gen_aarch64_rev64v8qi; break; | |
12607 | default: | |
12608 | return false; | |
12609 | } | |
12610 | break; | |
12611 | case 3: | |
12612 | switch (d->vmode) | |
12613 | { | |
12614 | case V16QImode: gen = gen_aarch64_rev32v16qi; break; | |
12615 | case V8QImode: gen = gen_aarch64_rev32v8qi; break; | |
12616 | case V8HImode: gen = gen_aarch64_rev64v8hi; break; | |
12617 | case V4HImode: gen = gen_aarch64_rev64v4hi; break; | |
12618 | default: | |
12619 | return false; | |
12620 | } | |
12621 | break; | |
12622 | case 1: | |
12623 | switch (d->vmode) | |
12624 | { | |
12625 | case V16QImode: gen = gen_aarch64_rev16v16qi; break; | |
12626 | case V8QImode: gen = gen_aarch64_rev16v8qi; break; | |
12627 | case V8HImode: gen = gen_aarch64_rev32v8hi; break; | |
12628 | case V4HImode: gen = gen_aarch64_rev32v4hi; break; | |
12629 | case V4SImode: gen = gen_aarch64_rev64v4si; break; | |
12630 | case V2SImode: gen = gen_aarch64_rev64v2si; break; | |
12631 | case V4SFmode: gen = gen_aarch64_rev64v4sf; break; | |
12632 | case V2SFmode: gen = gen_aarch64_rev64v2sf; break; | |
12633 | default: | |
12634 | return false; | |
12635 | } | |
12636 | break; | |
12637 | default: | |
12638 | return false; | |
12639 | } | |
12640 | ||
12641 | for (i = 0; i < nelt ; i += diff + 1) | |
12642 | for (j = 0; j <= diff; j += 1) | |
12643 | { | |
12644 | /* This is guaranteed to be true as the value of diff | |
12645 | is 7, 3, 1 and we should have enough elements in the | |
12646 | queue to generate this. Getting a vector mask with a | |
12647 | value of diff other than these values implies that | |
12648 | something is wrong by the time we get here. */ | |
12649 | gcc_assert (i + j < nelt); | |
12650 | if (d->perm[i + j] != i + diff - j) | |
12651 | return false; | |
12652 | } | |
12653 | ||
12654 | /* Success! */ | |
12655 | if (d->testing_p) | |
12656 | return true; | |
12657 | ||
12658 | emit_insn (gen (d->target, d->op0)); | |
12659 | return true; | |
12660 | } | |
12661 | ||
91bd4114 JG |
12662 | static bool |
12663 | aarch64_evpc_dup (struct expand_vec_perm_d *d) | |
12664 | { | |
12665 | rtx (*gen) (rtx, rtx, rtx); | |
12666 | rtx out = d->target; | |
12667 | rtx in0; | |
ef4bddc2 | 12668 | machine_mode vmode = d->vmode; |
91bd4114 JG |
12669 | unsigned int i, elt, nelt = d->nelt; |
12670 | rtx lane; | |
12671 | ||
91bd4114 JG |
12672 | elt = d->perm[0]; |
12673 | for (i = 1; i < nelt; i++) | |
12674 | { | |
12675 | if (elt != d->perm[i]) | |
12676 | return false; | |
12677 | } | |
12678 | ||
12679 | /* The generic preparation in aarch64_expand_vec_perm_const_1 | |
12680 | swaps the operand order and the permute indices if it finds | |
12681 | d->perm[0] to be in the second operand. Thus, we can always | |
12682 | use d->op0 and need not do any extra arithmetic to get the | |
12683 | correct lane number. */ | |
12684 | in0 = d->op0; | |
f901401e | 12685 | lane = GEN_INT (elt); /* The pattern corrects for big-endian. */ |
91bd4114 JG |
12686 | |
12687 | switch (vmode) | |
12688 | { | |
12689 | case V16QImode: gen = gen_aarch64_dup_lanev16qi; break; | |
12690 | case V8QImode: gen = gen_aarch64_dup_lanev8qi; break; | |
12691 | case V8HImode: gen = gen_aarch64_dup_lanev8hi; break; | |
12692 | case V4HImode: gen = gen_aarch64_dup_lanev4hi; break; | |
12693 | case V4SImode: gen = gen_aarch64_dup_lanev4si; break; | |
12694 | case V2SImode: gen = gen_aarch64_dup_lanev2si; break; | |
12695 | case V2DImode: gen = gen_aarch64_dup_lanev2di; break; | |
862abc04 AL |
12696 | case V8HFmode: gen = gen_aarch64_dup_lanev8hf; break; |
12697 | case V4HFmode: gen = gen_aarch64_dup_lanev4hf; break; | |
91bd4114 JG |
12698 | case V4SFmode: gen = gen_aarch64_dup_lanev4sf; break; |
12699 | case V2SFmode: gen = gen_aarch64_dup_lanev2sf; break; | |
12700 | case V2DFmode: gen = gen_aarch64_dup_lanev2df; break; | |
12701 | default: | |
12702 | return false; | |
12703 | } | |
12704 | ||
12705 | emit_insn (gen (out, in0, lane)); | |
12706 | return true; | |
12707 | } | |
12708 | ||
88b08073 JG |
12709 | static bool |
12710 | aarch64_evpc_tbl (struct expand_vec_perm_d *d) | |
12711 | { | |
12712 | rtx rperm[MAX_VECT_LEN], sel; | |
ef4bddc2 | 12713 | machine_mode vmode = d->vmode; |
88b08073 JG |
12714 | unsigned int i, nelt = d->nelt; |
12715 | ||
88b08073 JG |
12716 | if (d->testing_p) |
12717 | return true; | |
12718 | ||
12719 | /* Generic code will try constant permutation twice. Once with the | |
12720 | original mode and again with the elements lowered to QImode. | |
12721 | So wait and don't do the selector expansion ourselves. */ | |
12722 | if (vmode != V8QImode && vmode != V16QImode) | |
12723 | return false; | |
12724 | ||
12725 | for (i = 0; i < nelt; ++i) | |
bbcc9c00 TB |
12726 | { |
12727 | int nunits = GET_MODE_NUNITS (vmode); | |
12728 | ||
12729 | /* If big-endian and two vectors we end up with a weird mixed-endian | |
12730 | mode on NEON. Reverse the index within each word but not the word | |
12731 | itself. */ | |
12732 | rperm[i] = GEN_INT (BYTES_BIG_ENDIAN ? d->perm[i] ^ (nunits - 1) | |
12733 | : d->perm[i]); | |
12734 | } | |
88b08073 JG |
12735 | sel = gen_rtx_CONST_VECTOR (vmode, gen_rtvec_v (nelt, rperm)); |
12736 | sel = force_reg (vmode, sel); | |
12737 | ||
12738 | aarch64_expand_vec_perm_1 (d->target, d->op0, d->op1, sel); | |
12739 | return true; | |
12740 | } | |
12741 | ||
12742 | static bool | |
12743 | aarch64_expand_vec_perm_const_1 (struct expand_vec_perm_d *d) | |
12744 | { | |
12745 | /* The pattern matching functions above are written to look for a small | |
12746 | number to begin the sequence (0, 1, N/2). If we begin with an index | |
12747 | from the second operand, we can swap the operands. */ | |
12748 | if (d->perm[0] >= d->nelt) | |
12749 | { | |
12750 | unsigned i, nelt = d->nelt; | |
88b08073 | 12751 | |
0696116a | 12752 | gcc_assert (nelt == (nelt & -nelt)); |
88b08073 | 12753 | for (i = 0; i < nelt; ++i) |
0696116a | 12754 | d->perm[i] ^= nelt; /* Keep the same index, but in the other vector. */ |
88b08073 | 12755 | |
cb5c6c29 | 12756 | std::swap (d->op0, d->op1); |
88b08073 JG |
12757 | } |
12758 | ||
12759 | if (TARGET_SIMD) | |
cc4d934f | 12760 | { |
923fcec3 AL |
12761 | if (aarch64_evpc_rev (d)) |
12762 | return true; | |
12763 | else if (aarch64_evpc_ext (d)) | |
ae0533da | 12764 | return true; |
f901401e AL |
12765 | else if (aarch64_evpc_dup (d)) |
12766 | return true; | |
ae0533da | 12767 | else if (aarch64_evpc_zip (d)) |
cc4d934f JG |
12768 | return true; |
12769 | else if (aarch64_evpc_uzp (d)) | |
12770 | return true; | |
12771 | else if (aarch64_evpc_trn (d)) | |
12772 | return true; | |
12773 | return aarch64_evpc_tbl (d); | |
12774 | } | |
88b08073 JG |
12775 | return false; |
12776 | } | |
12777 | ||
12778 | /* Expand a vec_perm_const pattern. */ | |
12779 | ||
12780 | bool | |
12781 | aarch64_expand_vec_perm_const (rtx target, rtx op0, rtx op1, rtx sel) | |
12782 | { | |
12783 | struct expand_vec_perm_d d; | |
12784 | int i, nelt, which; | |
12785 | ||
12786 | d.target = target; | |
12787 | d.op0 = op0; | |
12788 | d.op1 = op1; | |
12789 | ||
12790 | d.vmode = GET_MODE (target); | |
12791 | gcc_assert (VECTOR_MODE_P (d.vmode)); | |
12792 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
12793 | d.testing_p = false; | |
12794 | ||
12795 | for (i = which = 0; i < nelt; ++i) | |
12796 | { | |
12797 | rtx e = XVECEXP (sel, 0, i); | |
12798 | int ei = INTVAL (e) & (2 * nelt - 1); | |
12799 | which |= (ei < nelt ? 1 : 2); | |
12800 | d.perm[i] = ei; | |
12801 | } | |
12802 | ||
12803 | switch (which) | |
12804 | { | |
12805 | default: | |
12806 | gcc_unreachable (); | |
12807 | ||
12808 | case 3: | |
12809 | d.one_vector_p = false; | |
12810 | if (!rtx_equal_p (op0, op1)) | |
12811 | break; | |
12812 | ||
12813 | /* The elements of PERM do not suggest that only the first operand | |
12814 | is used, but both operands are identical. Allow easier matching | |
12815 | of the permutation by folding the permutation into the single | |
12816 | input vector. */ | |
12817 | /* Fall Through. */ | |
12818 | case 2: | |
12819 | for (i = 0; i < nelt; ++i) | |
12820 | d.perm[i] &= nelt - 1; | |
12821 | d.op0 = op1; | |
12822 | d.one_vector_p = true; | |
12823 | break; | |
12824 | ||
12825 | case 1: | |
12826 | d.op1 = op0; | |
12827 | d.one_vector_p = true; | |
12828 | break; | |
12829 | } | |
12830 | ||
12831 | return aarch64_expand_vec_perm_const_1 (&d); | |
12832 | } | |
12833 | ||
12834 | static bool | |
ef4bddc2 | 12835 | aarch64_vectorize_vec_perm_const_ok (machine_mode vmode, |
88b08073 JG |
12836 | const unsigned char *sel) |
12837 | { | |
12838 | struct expand_vec_perm_d d; | |
12839 | unsigned int i, nelt, which; | |
12840 | bool ret; | |
12841 | ||
12842 | d.vmode = vmode; | |
12843 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
12844 | d.testing_p = true; | |
12845 | memcpy (d.perm, sel, nelt); | |
12846 | ||
12847 | /* Calculate whether all elements are in one vector. */ | |
12848 | for (i = which = 0; i < nelt; ++i) | |
12849 | { | |
12850 | unsigned char e = d.perm[i]; | |
12851 | gcc_assert (e < 2 * nelt); | |
12852 | which |= (e < nelt ? 1 : 2); | |
12853 | } | |
12854 | ||
12855 | /* If all elements are from the second vector, reindex as if from the | |
12856 | first vector. */ | |
12857 | if (which == 2) | |
12858 | for (i = 0; i < nelt; ++i) | |
12859 | d.perm[i] -= nelt; | |
12860 | ||
12861 | /* Check whether the mask can be applied to a single vector. */ | |
12862 | d.one_vector_p = (which != 3); | |
12863 | ||
12864 | d.target = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 1); | |
12865 | d.op1 = d.op0 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 2); | |
12866 | if (!d.one_vector_p) | |
12867 | d.op1 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 3); | |
12868 | ||
12869 | start_sequence (); | |
12870 | ret = aarch64_expand_vec_perm_const_1 (&d); | |
12871 | end_sequence (); | |
12872 | ||
12873 | return ret; | |
12874 | } | |
12875 | ||
668046d1 DS |
12876 | rtx |
12877 | aarch64_reverse_mask (enum machine_mode mode) | |
12878 | { | |
12879 | /* We have to reverse each vector because we dont have | |
12880 | a permuted load that can reverse-load according to ABI rules. */ | |
12881 | rtx mask; | |
12882 | rtvec v = rtvec_alloc (16); | |
12883 | int i, j; | |
12884 | int nunits = GET_MODE_NUNITS (mode); | |
12885 | int usize = GET_MODE_UNIT_SIZE (mode); | |
12886 | ||
12887 | gcc_assert (BYTES_BIG_ENDIAN); | |
12888 | gcc_assert (AARCH64_VALID_SIMD_QREG_MODE (mode)); | |
12889 | ||
12890 | for (i = 0; i < nunits; i++) | |
12891 | for (j = 0; j < usize; j++) | |
12892 | RTVEC_ELT (v, i * usize + j) = GEN_INT ((i + 1) * usize - 1 - j); | |
12893 | mask = gen_rtx_CONST_VECTOR (V16QImode, v); | |
12894 | return force_reg (V16QImode, mask); | |
12895 | } | |
12896 | ||
61f17a5c WD |
12897 | /* Implement MODES_TIEABLE_P. In principle we should always return true. |
12898 | However due to issues with register allocation it is preferable to avoid | |
12899 | tieing integer scalar and FP scalar modes. Executing integer operations | |
12900 | in general registers is better than treating them as scalar vector | |
12901 | operations. This reduces latency and avoids redundant int<->FP moves. | |
12902 | So tie modes if they are either the same class, or vector modes with | |
12903 | other vector modes, vector structs or any scalar mode. | |
12904 | */ | |
97e1ad78 JG |
12905 | |
12906 | bool | |
ef4bddc2 | 12907 | aarch64_modes_tieable_p (machine_mode mode1, machine_mode mode2) |
97e1ad78 JG |
12908 | { |
12909 | if (GET_MODE_CLASS (mode1) == GET_MODE_CLASS (mode2)) | |
12910 | return true; | |
12911 | ||
12912 | /* We specifically want to allow elements of "structure" modes to | |
12913 | be tieable to the structure. This more general condition allows | |
12914 | other rarer situations too. */ | |
61f17a5c WD |
12915 | if (aarch64_vector_mode_p (mode1) && aarch64_vector_mode_p (mode2)) |
12916 | return true; | |
12917 | ||
12918 | /* Also allow any scalar modes with vectors. */ | |
12919 | if (aarch64_vector_mode_supported_p (mode1) | |
12920 | || aarch64_vector_mode_supported_p (mode2)) | |
97e1ad78 JG |
12921 | return true; |
12922 | ||
12923 | return false; | |
12924 | } | |
12925 | ||
e2c75eea JG |
12926 | /* Return a new RTX holding the result of moving POINTER forward by |
12927 | AMOUNT bytes. */ | |
12928 | ||
12929 | static rtx | |
12930 | aarch64_move_pointer (rtx pointer, int amount) | |
12931 | { | |
12932 | rtx next = plus_constant (Pmode, XEXP (pointer, 0), amount); | |
12933 | ||
12934 | return adjust_automodify_address (pointer, GET_MODE (pointer), | |
12935 | next, amount); | |
12936 | } | |
12937 | ||
12938 | /* Return a new RTX holding the result of moving POINTER forward by the | |
12939 | size of the mode it points to. */ | |
12940 | ||
12941 | static rtx | |
12942 | aarch64_progress_pointer (rtx pointer) | |
12943 | { | |
12944 | HOST_WIDE_INT amount = GET_MODE_SIZE (GET_MODE (pointer)); | |
12945 | ||
12946 | return aarch64_move_pointer (pointer, amount); | |
12947 | } | |
12948 | ||
12949 | /* Copy one MODE sized block from SRC to DST, then progress SRC and DST by | |
12950 | MODE bytes. */ | |
12951 | ||
12952 | static void | |
12953 | aarch64_copy_one_block_and_progress_pointers (rtx *src, rtx *dst, | |
ef4bddc2 | 12954 | machine_mode mode) |
e2c75eea JG |
12955 | { |
12956 | rtx reg = gen_reg_rtx (mode); | |
12957 | ||
12958 | /* "Cast" the pointers to the correct mode. */ | |
12959 | *src = adjust_address (*src, mode, 0); | |
12960 | *dst = adjust_address (*dst, mode, 0); | |
12961 | /* Emit the memcpy. */ | |
12962 | emit_move_insn (reg, *src); | |
12963 | emit_move_insn (*dst, reg); | |
12964 | /* Move the pointers forward. */ | |
12965 | *src = aarch64_progress_pointer (*src); | |
12966 | *dst = aarch64_progress_pointer (*dst); | |
12967 | } | |
12968 | ||
12969 | /* Expand movmem, as if from a __builtin_memcpy. Return true if | |
12970 | we succeed, otherwise return false. */ | |
12971 | ||
12972 | bool | |
12973 | aarch64_expand_movmem (rtx *operands) | |
12974 | { | |
12975 | unsigned int n; | |
12976 | rtx dst = operands[0]; | |
12977 | rtx src = operands[1]; | |
12978 | rtx base; | |
12979 | bool speed_p = !optimize_function_for_size_p (cfun); | |
12980 | ||
12981 | /* When optimizing for size, give a better estimate of the length of a | |
12982 | memcpy call, but use the default otherwise. */ | |
12983 | unsigned int max_instructions = (speed_p ? 15 : AARCH64_CALL_RATIO) / 2; | |
12984 | ||
12985 | /* We can't do anything smart if the amount to copy is not constant. */ | |
12986 | if (!CONST_INT_P (operands[2])) | |
12987 | return false; | |
12988 | ||
12989 | n = UINTVAL (operands[2]); | |
12990 | ||
12991 | /* Try to keep the number of instructions low. For cases below 16 bytes we | |
12992 | need to make at most two moves. For cases above 16 bytes it will be one | |
12993 | move for each 16 byte chunk, then at most two additional moves. */ | |
12994 | if (((n / 16) + (n % 16 ? 2 : 0)) > max_instructions) | |
12995 | return false; | |
12996 | ||
12997 | base = copy_to_mode_reg (Pmode, XEXP (dst, 0)); | |
12998 | dst = adjust_automodify_address (dst, VOIDmode, base, 0); | |
12999 | ||
13000 | base = copy_to_mode_reg (Pmode, XEXP (src, 0)); | |
13001 | src = adjust_automodify_address (src, VOIDmode, base, 0); | |
13002 | ||
13003 | /* Simple cases. Copy 0-3 bytes, as (if applicable) a 2-byte, then a | |
13004 | 1-byte chunk. */ | |
13005 | if (n < 4) | |
13006 | { | |
13007 | if (n >= 2) | |
13008 | { | |
13009 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, HImode); | |
13010 | n -= 2; | |
13011 | } | |
13012 | ||
13013 | if (n == 1) | |
13014 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, QImode); | |
13015 | ||
13016 | return true; | |
13017 | } | |
13018 | ||
13019 | /* Copy 4-8 bytes. First a 4-byte chunk, then (if applicable) a second | |
13020 | 4-byte chunk, partially overlapping with the previously copied chunk. */ | |
13021 | if (n < 8) | |
13022 | { | |
13023 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
13024 | n -= 4; | |
13025 | if (n > 0) | |
13026 | { | |
13027 | int move = n - 4; | |
13028 | ||
13029 | src = aarch64_move_pointer (src, move); | |
13030 | dst = aarch64_move_pointer (dst, move); | |
13031 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
13032 | } | |
13033 | return true; | |
13034 | } | |
13035 | ||
13036 | /* Copy more than 8 bytes. Copy chunks of 16 bytes until we run out of | |
13037 | them, then (if applicable) an 8-byte chunk. */ | |
13038 | while (n >= 8) | |
13039 | { | |
13040 | if (n / 16) | |
13041 | { | |
13042 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, TImode); | |
13043 | n -= 16; | |
13044 | } | |
13045 | else | |
13046 | { | |
13047 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, DImode); | |
13048 | n -= 8; | |
13049 | } | |
13050 | } | |
13051 | ||
13052 | /* Finish the final bytes of the copy. We can always do this in one | |
13053 | instruction. We either copy the exact amount we need, or partially | |
13054 | overlap with the previous chunk we copied and copy 8-bytes. */ | |
13055 | if (n == 0) | |
13056 | return true; | |
13057 | else if (n == 1) | |
13058 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, QImode); | |
13059 | else if (n == 2) | |
13060 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, HImode); | |
13061 | else if (n == 4) | |
13062 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
13063 | else | |
13064 | { | |
13065 | if (n == 3) | |
13066 | { | |
13067 | src = aarch64_move_pointer (src, -1); | |
13068 | dst = aarch64_move_pointer (dst, -1); | |
13069 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
13070 | } | |
13071 | else | |
13072 | { | |
13073 | int move = n - 8; | |
13074 | ||
13075 | src = aarch64_move_pointer (src, move); | |
13076 | dst = aarch64_move_pointer (dst, move); | |
13077 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, DImode); | |
13078 | } | |
13079 | } | |
13080 | ||
13081 | return true; | |
13082 | } | |
13083 | ||
a3125fc2 CL |
13084 | /* Implement the TARGET_ASAN_SHADOW_OFFSET hook. */ |
13085 | ||
13086 | static unsigned HOST_WIDE_INT | |
13087 | aarch64_asan_shadow_offset (void) | |
13088 | { | |
13089 | return (HOST_WIDE_INT_1 << 36); | |
13090 | } | |
13091 | ||
d3006da6 | 13092 | static bool |
445d7826 | 13093 | aarch64_use_by_pieces_infrastructure_p (unsigned HOST_WIDE_INT size, |
d3006da6 JG |
13094 | unsigned int align, |
13095 | enum by_pieces_operation op, | |
13096 | bool speed_p) | |
13097 | { | |
13098 | /* STORE_BY_PIECES can be used when copying a constant string, but | |
13099 | in that case each 64-bit chunk takes 5 insns instead of 2 (LDR/STR). | |
13100 | For now we always fail this and let the move_by_pieces code copy | |
13101 | the string from read-only memory. */ | |
13102 | if (op == STORE_BY_PIECES) | |
13103 | return false; | |
13104 | ||
13105 | return default_use_by_pieces_infrastructure_p (size, align, op, speed_p); | |
13106 | } | |
13107 | ||
5f3bc026 ZC |
13108 | static rtx |
13109 | aarch64_gen_ccmp_first (rtx *prep_seq, rtx *gen_seq, | |
13110 | int code, tree treeop0, tree treeop1) | |
13111 | { | |
c8012fbc WD |
13112 | machine_mode op_mode, cmp_mode, cc_mode = CCmode; |
13113 | rtx op0, op1; | |
5f3bc026 | 13114 | int unsignedp = TYPE_UNSIGNED (TREE_TYPE (treeop0)); |
c8012fbc | 13115 | insn_code icode; |
5f3bc026 ZC |
13116 | struct expand_operand ops[4]; |
13117 | ||
5f3bc026 ZC |
13118 | start_sequence (); |
13119 | expand_operands (treeop0, treeop1, NULL_RTX, &op0, &op1, EXPAND_NORMAL); | |
13120 | ||
13121 | op_mode = GET_MODE (op0); | |
13122 | if (op_mode == VOIDmode) | |
13123 | op_mode = GET_MODE (op1); | |
13124 | ||
13125 | switch (op_mode) | |
13126 | { | |
13127 | case QImode: | |
13128 | case HImode: | |
13129 | case SImode: | |
13130 | cmp_mode = SImode; | |
13131 | icode = CODE_FOR_cmpsi; | |
13132 | break; | |
13133 | ||
13134 | case DImode: | |
13135 | cmp_mode = DImode; | |
13136 | icode = CODE_FOR_cmpdi; | |
13137 | break; | |
13138 | ||
786e3c06 WD |
13139 | case SFmode: |
13140 | cmp_mode = SFmode; | |
13141 | cc_mode = aarch64_select_cc_mode ((rtx_code) code, op0, op1); | |
13142 | icode = cc_mode == CCFPEmode ? CODE_FOR_fcmpesf : CODE_FOR_fcmpsf; | |
13143 | break; | |
13144 | ||
13145 | case DFmode: | |
13146 | cmp_mode = DFmode; | |
13147 | cc_mode = aarch64_select_cc_mode ((rtx_code) code, op0, op1); | |
13148 | icode = cc_mode == CCFPEmode ? CODE_FOR_fcmpedf : CODE_FOR_fcmpdf; | |
13149 | break; | |
13150 | ||
5f3bc026 ZC |
13151 | default: |
13152 | end_sequence (); | |
13153 | return NULL_RTX; | |
13154 | } | |
13155 | ||
c8012fbc WD |
13156 | op0 = prepare_operand (icode, op0, 0, op_mode, cmp_mode, unsignedp); |
13157 | op1 = prepare_operand (icode, op1, 1, op_mode, cmp_mode, unsignedp); | |
5f3bc026 ZC |
13158 | if (!op0 || !op1) |
13159 | { | |
13160 | end_sequence (); | |
13161 | return NULL_RTX; | |
13162 | } | |
13163 | *prep_seq = get_insns (); | |
13164 | end_sequence (); | |
13165 | ||
c8012fbc WD |
13166 | create_fixed_operand (&ops[0], op0); |
13167 | create_fixed_operand (&ops[1], op1); | |
5f3bc026 ZC |
13168 | |
13169 | start_sequence (); | |
c8012fbc | 13170 | if (!maybe_expand_insn (icode, 2, ops)) |
5f3bc026 ZC |
13171 | { |
13172 | end_sequence (); | |
13173 | return NULL_RTX; | |
13174 | } | |
13175 | *gen_seq = get_insns (); | |
13176 | end_sequence (); | |
13177 | ||
c8012fbc WD |
13178 | return gen_rtx_fmt_ee ((rtx_code) code, cc_mode, |
13179 | gen_rtx_REG (cc_mode, CC_REGNUM), const0_rtx); | |
5f3bc026 ZC |
13180 | } |
13181 | ||
13182 | static rtx | |
13183 | aarch64_gen_ccmp_next (rtx *prep_seq, rtx *gen_seq, rtx prev, int cmp_code, | |
13184 | tree treeop0, tree treeop1, int bit_code) | |
13185 | { | |
c8012fbc WD |
13186 | rtx op0, op1, target; |
13187 | machine_mode op_mode, cmp_mode, cc_mode = CCmode; | |
5f3bc026 | 13188 | int unsignedp = TYPE_UNSIGNED (TREE_TYPE (treeop0)); |
c8012fbc | 13189 | insn_code icode; |
5f3bc026 | 13190 | struct expand_operand ops[6]; |
c8012fbc | 13191 | int aarch64_cond; |
5f3bc026 ZC |
13192 | |
13193 | push_to_sequence ((rtx_insn*) *prep_seq); | |
13194 | expand_operands (treeop0, treeop1, NULL_RTX, &op0, &op1, EXPAND_NORMAL); | |
13195 | ||
13196 | op_mode = GET_MODE (op0); | |
13197 | if (op_mode == VOIDmode) | |
13198 | op_mode = GET_MODE (op1); | |
13199 | ||
13200 | switch (op_mode) | |
13201 | { | |
13202 | case QImode: | |
13203 | case HImode: | |
13204 | case SImode: | |
13205 | cmp_mode = SImode; | |
c8012fbc | 13206 | icode = CODE_FOR_ccmpsi; |
5f3bc026 ZC |
13207 | break; |
13208 | ||
13209 | case DImode: | |
13210 | cmp_mode = DImode; | |
c8012fbc | 13211 | icode = CODE_FOR_ccmpdi; |
5f3bc026 ZC |
13212 | break; |
13213 | ||
786e3c06 WD |
13214 | case SFmode: |
13215 | cmp_mode = SFmode; | |
13216 | cc_mode = aarch64_select_cc_mode ((rtx_code) cmp_code, op0, op1); | |
13217 | icode = cc_mode == CCFPEmode ? CODE_FOR_fccmpesf : CODE_FOR_fccmpsf; | |
13218 | break; | |
13219 | ||
13220 | case DFmode: | |
13221 | cmp_mode = DFmode; | |
13222 | cc_mode = aarch64_select_cc_mode ((rtx_code) cmp_code, op0, op1); | |
13223 | icode = cc_mode == CCFPEmode ? CODE_FOR_fccmpedf : CODE_FOR_fccmpdf; | |
13224 | break; | |
13225 | ||
5f3bc026 ZC |
13226 | default: |
13227 | end_sequence (); | |
13228 | return NULL_RTX; | |
13229 | } | |
13230 | ||
13231 | op0 = prepare_operand (icode, op0, 2, op_mode, cmp_mode, unsignedp); | |
13232 | op1 = prepare_operand (icode, op1, 3, op_mode, cmp_mode, unsignedp); | |
13233 | if (!op0 || !op1) | |
13234 | { | |
13235 | end_sequence (); | |
13236 | return NULL_RTX; | |
13237 | } | |
13238 | *prep_seq = get_insns (); | |
13239 | end_sequence (); | |
13240 | ||
13241 | target = gen_rtx_REG (cc_mode, CC_REGNUM); | |
c8012fbc | 13242 | aarch64_cond = aarch64_get_condition_code_1 (cc_mode, (rtx_code) cmp_code); |
5f3bc026 | 13243 | |
c8012fbc WD |
13244 | if (bit_code != AND) |
13245 | { | |
13246 | prev = gen_rtx_fmt_ee (REVERSE_CONDITION (GET_CODE (prev), | |
13247 | GET_MODE (XEXP (prev, 0))), | |
13248 | VOIDmode, XEXP (prev, 0), const0_rtx); | |
13249 | aarch64_cond = AARCH64_INVERSE_CONDITION_CODE (aarch64_cond); | |
13250 | } | |
13251 | ||
13252 | create_fixed_operand (&ops[0], XEXP (prev, 0)); | |
5f3bc026 ZC |
13253 | create_fixed_operand (&ops[1], target); |
13254 | create_fixed_operand (&ops[2], op0); | |
13255 | create_fixed_operand (&ops[3], op1); | |
c8012fbc WD |
13256 | create_fixed_operand (&ops[4], prev); |
13257 | create_fixed_operand (&ops[5], GEN_INT (aarch64_cond)); | |
5f3bc026 ZC |
13258 | |
13259 | push_to_sequence ((rtx_insn*) *gen_seq); | |
13260 | if (!maybe_expand_insn (icode, 6, ops)) | |
13261 | { | |
13262 | end_sequence (); | |
13263 | return NULL_RTX; | |
13264 | } | |
13265 | ||
13266 | *gen_seq = get_insns (); | |
13267 | end_sequence (); | |
13268 | ||
c8012fbc | 13269 | return gen_rtx_fmt_ee ((rtx_code) cmp_code, VOIDmode, target, const0_rtx); |
5f3bc026 ZC |
13270 | } |
13271 | ||
13272 | #undef TARGET_GEN_CCMP_FIRST | |
13273 | #define TARGET_GEN_CCMP_FIRST aarch64_gen_ccmp_first | |
13274 | ||
13275 | #undef TARGET_GEN_CCMP_NEXT | |
13276 | #define TARGET_GEN_CCMP_NEXT aarch64_gen_ccmp_next | |
13277 | ||
6a569cdd KT |
13278 | /* Implement TARGET_SCHED_MACRO_FUSION_P. Return true if target supports |
13279 | instruction fusion of some sort. */ | |
13280 | ||
13281 | static bool | |
13282 | aarch64_macro_fusion_p (void) | |
13283 | { | |
b175b679 | 13284 | return aarch64_tune_params.fusible_ops != AARCH64_FUSE_NOTHING; |
6a569cdd KT |
13285 | } |
13286 | ||
13287 | ||
13288 | /* Implement TARGET_SCHED_MACRO_FUSION_PAIR_P. Return true if PREV and CURR | |
13289 | should be kept together during scheduling. */ | |
13290 | ||
13291 | static bool | |
13292 | aarch_macro_fusion_pair_p (rtx_insn *prev, rtx_insn *curr) | |
13293 | { | |
13294 | rtx set_dest; | |
13295 | rtx prev_set = single_set (prev); | |
13296 | rtx curr_set = single_set (curr); | |
13297 | /* prev and curr are simple SET insns i.e. no flag setting or branching. */ | |
13298 | bool simple_sets_p = prev_set && curr_set && !any_condjump_p (curr); | |
13299 | ||
13300 | if (!aarch64_macro_fusion_p ()) | |
13301 | return false; | |
13302 | ||
d7b03373 | 13303 | if (simple_sets_p && aarch64_fusion_enabled_p (AARCH64_FUSE_MOV_MOVK)) |
6a569cdd KT |
13304 | { |
13305 | /* We are trying to match: | |
13306 | prev (mov) == (set (reg r0) (const_int imm16)) | |
13307 | curr (movk) == (set (zero_extract (reg r0) | |
13308 | (const_int 16) | |
13309 | (const_int 16)) | |
13310 | (const_int imm16_1)) */ | |
13311 | ||
13312 | set_dest = SET_DEST (curr_set); | |
13313 | ||
13314 | if (GET_CODE (set_dest) == ZERO_EXTRACT | |
13315 | && CONST_INT_P (SET_SRC (curr_set)) | |
13316 | && CONST_INT_P (SET_SRC (prev_set)) | |
13317 | && CONST_INT_P (XEXP (set_dest, 2)) | |
13318 | && INTVAL (XEXP (set_dest, 2)) == 16 | |
13319 | && REG_P (XEXP (set_dest, 0)) | |
13320 | && REG_P (SET_DEST (prev_set)) | |
13321 | && REGNO (XEXP (set_dest, 0)) == REGNO (SET_DEST (prev_set))) | |
13322 | { | |
13323 | return true; | |
13324 | } | |
13325 | } | |
13326 | ||
d7b03373 | 13327 | if (simple_sets_p && aarch64_fusion_enabled_p (AARCH64_FUSE_ADRP_ADD)) |
9bbe08fe KT |
13328 | { |
13329 | ||
13330 | /* We're trying to match: | |
13331 | prev (adrp) == (set (reg r1) | |
13332 | (high (symbol_ref ("SYM")))) | |
13333 | curr (add) == (set (reg r0) | |
13334 | (lo_sum (reg r1) | |
13335 | (symbol_ref ("SYM")))) | |
13336 | Note that r0 need not necessarily be the same as r1, especially | |
13337 | during pre-regalloc scheduling. */ | |
13338 | ||
13339 | if (satisfies_constraint_Ush (SET_SRC (prev_set)) | |
13340 | && REG_P (SET_DEST (prev_set)) && REG_P (SET_DEST (curr_set))) | |
13341 | { | |
13342 | if (GET_CODE (SET_SRC (curr_set)) == LO_SUM | |
13343 | && REG_P (XEXP (SET_SRC (curr_set), 0)) | |
13344 | && REGNO (XEXP (SET_SRC (curr_set), 0)) | |
13345 | == REGNO (SET_DEST (prev_set)) | |
13346 | && rtx_equal_p (XEXP (SET_SRC (prev_set), 0), | |
13347 | XEXP (SET_SRC (curr_set), 1))) | |
13348 | return true; | |
13349 | } | |
13350 | } | |
13351 | ||
d7b03373 | 13352 | if (simple_sets_p && aarch64_fusion_enabled_p (AARCH64_FUSE_MOVK_MOVK)) |
cd0cb232 KT |
13353 | { |
13354 | ||
13355 | /* We're trying to match: | |
13356 | prev (movk) == (set (zero_extract (reg r0) | |
13357 | (const_int 16) | |
13358 | (const_int 32)) | |
13359 | (const_int imm16_1)) | |
13360 | curr (movk) == (set (zero_extract (reg r0) | |
13361 | (const_int 16) | |
13362 | (const_int 48)) | |
13363 | (const_int imm16_2)) */ | |
13364 | ||
13365 | if (GET_CODE (SET_DEST (prev_set)) == ZERO_EXTRACT | |
13366 | && GET_CODE (SET_DEST (curr_set)) == ZERO_EXTRACT | |
13367 | && REG_P (XEXP (SET_DEST (prev_set), 0)) | |
13368 | && REG_P (XEXP (SET_DEST (curr_set), 0)) | |
13369 | && REGNO (XEXP (SET_DEST (prev_set), 0)) | |
13370 | == REGNO (XEXP (SET_DEST (curr_set), 0)) | |
13371 | && CONST_INT_P (XEXP (SET_DEST (prev_set), 2)) | |
13372 | && CONST_INT_P (XEXP (SET_DEST (curr_set), 2)) | |
13373 | && INTVAL (XEXP (SET_DEST (prev_set), 2)) == 32 | |
13374 | && INTVAL (XEXP (SET_DEST (curr_set), 2)) == 48 | |
13375 | && CONST_INT_P (SET_SRC (prev_set)) | |
13376 | && CONST_INT_P (SET_SRC (curr_set))) | |
13377 | return true; | |
13378 | ||
13379 | } | |
d7b03373 | 13380 | if (simple_sets_p && aarch64_fusion_enabled_p (AARCH64_FUSE_ADRP_LDR)) |
d8354ad7 KT |
13381 | { |
13382 | /* We're trying to match: | |
13383 | prev (adrp) == (set (reg r0) | |
13384 | (high (symbol_ref ("SYM")))) | |
13385 | curr (ldr) == (set (reg r1) | |
13386 | (mem (lo_sum (reg r0) | |
13387 | (symbol_ref ("SYM"))))) | |
13388 | or | |
13389 | curr (ldr) == (set (reg r1) | |
13390 | (zero_extend (mem | |
13391 | (lo_sum (reg r0) | |
13392 | (symbol_ref ("SYM")))))) */ | |
13393 | if (satisfies_constraint_Ush (SET_SRC (prev_set)) | |
13394 | && REG_P (SET_DEST (prev_set)) && REG_P (SET_DEST (curr_set))) | |
13395 | { | |
13396 | rtx curr_src = SET_SRC (curr_set); | |
13397 | ||
13398 | if (GET_CODE (curr_src) == ZERO_EXTEND) | |
13399 | curr_src = XEXP (curr_src, 0); | |
13400 | ||
13401 | if (MEM_P (curr_src) && GET_CODE (XEXP (curr_src, 0)) == LO_SUM | |
13402 | && REG_P (XEXP (XEXP (curr_src, 0), 0)) | |
13403 | && REGNO (XEXP (XEXP (curr_src, 0), 0)) | |
13404 | == REGNO (SET_DEST (prev_set)) | |
13405 | && rtx_equal_p (XEXP (XEXP (curr_src, 0), 1), | |
13406 | XEXP (SET_SRC (prev_set), 0))) | |
13407 | return true; | |
13408 | } | |
13409 | } | |
cd0cb232 | 13410 | |
d7b03373 | 13411 | if (aarch64_fusion_enabled_p (AARCH64_FUSE_AES_AESMC) |
00a8574a WD |
13412 | && aarch_crypto_can_dual_issue (prev, curr)) |
13413 | return true; | |
13414 | ||
d7b03373 | 13415 | if (aarch64_fusion_enabled_p (AARCH64_FUSE_CMP_BRANCH) |
3759108f AP |
13416 | && any_condjump_p (curr)) |
13417 | { | |
13418 | enum attr_type prev_type = get_attr_type (prev); | |
13419 | ||
13420 | /* FIXME: this misses some which is considered simple arthematic | |
13421 | instructions for ThunderX. Simple shifts are missed here. */ | |
13422 | if (prev_type == TYPE_ALUS_SREG | |
13423 | || prev_type == TYPE_ALUS_IMM | |
13424 | || prev_type == TYPE_LOGICS_REG | |
13425 | || prev_type == TYPE_LOGICS_IMM) | |
13426 | return true; | |
13427 | } | |
13428 | ||
6a569cdd KT |
13429 | return false; |
13430 | } | |
13431 | ||
f2879a90 KT |
13432 | /* Return true iff the instruction fusion described by OP is enabled. */ |
13433 | ||
13434 | bool | |
13435 | aarch64_fusion_enabled_p (enum aarch64_fusion_pairs op) | |
13436 | { | |
13437 | return (aarch64_tune_params.fusible_ops & op) != 0; | |
13438 | } | |
13439 | ||
350013bc BC |
13440 | /* If MEM is in the form of [base+offset], extract the two parts |
13441 | of address and set to BASE and OFFSET, otherwise return false | |
13442 | after clearing BASE and OFFSET. */ | |
13443 | ||
13444 | bool | |
13445 | extract_base_offset_in_addr (rtx mem, rtx *base, rtx *offset) | |
13446 | { | |
13447 | rtx addr; | |
13448 | ||
13449 | gcc_assert (MEM_P (mem)); | |
13450 | ||
13451 | addr = XEXP (mem, 0); | |
13452 | ||
13453 | if (REG_P (addr)) | |
13454 | { | |
13455 | *base = addr; | |
13456 | *offset = const0_rtx; | |
13457 | return true; | |
13458 | } | |
13459 | ||
13460 | if (GET_CODE (addr) == PLUS | |
13461 | && REG_P (XEXP (addr, 0)) && CONST_INT_P (XEXP (addr, 1))) | |
13462 | { | |
13463 | *base = XEXP (addr, 0); | |
13464 | *offset = XEXP (addr, 1); | |
13465 | return true; | |
13466 | } | |
13467 | ||
13468 | *base = NULL_RTX; | |
13469 | *offset = NULL_RTX; | |
13470 | ||
13471 | return false; | |
13472 | } | |
13473 | ||
13474 | /* Types for scheduling fusion. */ | |
13475 | enum sched_fusion_type | |
13476 | { | |
13477 | SCHED_FUSION_NONE = 0, | |
13478 | SCHED_FUSION_LD_SIGN_EXTEND, | |
13479 | SCHED_FUSION_LD_ZERO_EXTEND, | |
13480 | SCHED_FUSION_LD, | |
13481 | SCHED_FUSION_ST, | |
13482 | SCHED_FUSION_NUM | |
13483 | }; | |
13484 | ||
13485 | /* If INSN is a load or store of address in the form of [base+offset], | |
13486 | extract the two parts and set to BASE and OFFSET. Return scheduling | |
13487 | fusion type this INSN is. */ | |
13488 | ||
13489 | static enum sched_fusion_type | |
13490 | fusion_load_store (rtx_insn *insn, rtx *base, rtx *offset) | |
13491 | { | |
13492 | rtx x, dest, src; | |
13493 | enum sched_fusion_type fusion = SCHED_FUSION_LD; | |
13494 | ||
13495 | gcc_assert (INSN_P (insn)); | |
13496 | x = PATTERN (insn); | |
13497 | if (GET_CODE (x) != SET) | |
13498 | return SCHED_FUSION_NONE; | |
13499 | ||
13500 | src = SET_SRC (x); | |
13501 | dest = SET_DEST (x); | |
13502 | ||
abc52318 KT |
13503 | machine_mode dest_mode = GET_MODE (dest); |
13504 | ||
13505 | if (!aarch64_mode_valid_for_sched_fusion_p (dest_mode)) | |
350013bc BC |
13506 | return SCHED_FUSION_NONE; |
13507 | ||
13508 | if (GET_CODE (src) == SIGN_EXTEND) | |
13509 | { | |
13510 | fusion = SCHED_FUSION_LD_SIGN_EXTEND; | |
13511 | src = XEXP (src, 0); | |
13512 | if (GET_CODE (src) != MEM || GET_MODE (src) != SImode) | |
13513 | return SCHED_FUSION_NONE; | |
13514 | } | |
13515 | else if (GET_CODE (src) == ZERO_EXTEND) | |
13516 | { | |
13517 | fusion = SCHED_FUSION_LD_ZERO_EXTEND; | |
13518 | src = XEXP (src, 0); | |
13519 | if (GET_CODE (src) != MEM || GET_MODE (src) != SImode) | |
13520 | return SCHED_FUSION_NONE; | |
13521 | } | |
13522 | ||
13523 | if (GET_CODE (src) == MEM && REG_P (dest)) | |
13524 | extract_base_offset_in_addr (src, base, offset); | |
13525 | else if (GET_CODE (dest) == MEM && (REG_P (src) || src == const0_rtx)) | |
13526 | { | |
13527 | fusion = SCHED_FUSION_ST; | |
13528 | extract_base_offset_in_addr (dest, base, offset); | |
13529 | } | |
13530 | else | |
13531 | return SCHED_FUSION_NONE; | |
13532 | ||
13533 | if (*base == NULL_RTX || *offset == NULL_RTX) | |
13534 | fusion = SCHED_FUSION_NONE; | |
13535 | ||
13536 | return fusion; | |
13537 | } | |
13538 | ||
13539 | /* Implement the TARGET_SCHED_FUSION_PRIORITY hook. | |
13540 | ||
13541 | Currently we only support to fuse ldr or str instructions, so FUSION_PRI | |
13542 | and PRI are only calculated for these instructions. For other instruction, | |
13543 | FUSION_PRI and PRI are simply set to MAX_PRI - 1. In the future, other | |
13544 | type instruction fusion can be added by returning different priorities. | |
13545 | ||
13546 | It's important that irrelevant instructions get the largest FUSION_PRI. */ | |
13547 | ||
13548 | static void | |
13549 | aarch64_sched_fusion_priority (rtx_insn *insn, int max_pri, | |
13550 | int *fusion_pri, int *pri) | |
13551 | { | |
13552 | int tmp, off_val; | |
13553 | rtx base, offset; | |
13554 | enum sched_fusion_type fusion; | |
13555 | ||
13556 | gcc_assert (INSN_P (insn)); | |
13557 | ||
13558 | tmp = max_pri - 1; | |
13559 | fusion = fusion_load_store (insn, &base, &offset); | |
13560 | if (fusion == SCHED_FUSION_NONE) | |
13561 | { | |
13562 | *pri = tmp; | |
13563 | *fusion_pri = tmp; | |
13564 | return; | |
13565 | } | |
13566 | ||
13567 | /* Set FUSION_PRI according to fusion type and base register. */ | |
13568 | *fusion_pri = tmp - fusion * FIRST_PSEUDO_REGISTER - REGNO (base); | |
13569 | ||
13570 | /* Calculate PRI. */ | |
13571 | tmp /= 2; | |
13572 | ||
13573 | /* INSN with smaller offset goes first. */ | |
13574 | off_val = (int)(INTVAL (offset)); | |
13575 | if (off_val >= 0) | |
13576 | tmp -= (off_val & 0xfffff); | |
13577 | else | |
13578 | tmp += ((- off_val) & 0xfffff); | |
13579 | ||
13580 | *pri = tmp; | |
13581 | return; | |
13582 | } | |
13583 | ||
13584 | /* Given OPERANDS of consecutive load/store, check if we can merge | |
13585 | them into ldp/stp. LOAD is true if they are load instructions. | |
13586 | MODE is the mode of memory operands. */ | |
13587 | ||
13588 | bool | |
13589 | aarch64_operands_ok_for_ldpstp (rtx *operands, bool load, | |
13590 | enum machine_mode mode) | |
13591 | { | |
13592 | HOST_WIDE_INT offval_1, offval_2, msize; | |
13593 | enum reg_class rclass_1, rclass_2; | |
13594 | rtx mem_1, mem_2, reg_1, reg_2, base_1, base_2, offset_1, offset_2; | |
13595 | ||
13596 | if (load) | |
13597 | { | |
13598 | mem_1 = operands[1]; | |
13599 | mem_2 = operands[3]; | |
13600 | reg_1 = operands[0]; | |
13601 | reg_2 = operands[2]; | |
13602 | gcc_assert (REG_P (reg_1) && REG_P (reg_2)); | |
13603 | if (REGNO (reg_1) == REGNO (reg_2)) | |
13604 | return false; | |
13605 | } | |
13606 | else | |
13607 | { | |
13608 | mem_1 = operands[0]; | |
13609 | mem_2 = operands[2]; | |
13610 | reg_1 = operands[1]; | |
13611 | reg_2 = operands[3]; | |
13612 | } | |
13613 | ||
bf84ac44 AP |
13614 | /* The mems cannot be volatile. */ |
13615 | if (MEM_VOLATILE_P (mem_1) || MEM_VOLATILE_P (mem_2)) | |
13616 | return false; | |
13617 | ||
350013bc BC |
13618 | /* Check if the addresses are in the form of [base+offset]. */ |
13619 | extract_base_offset_in_addr (mem_1, &base_1, &offset_1); | |
13620 | if (base_1 == NULL_RTX || offset_1 == NULL_RTX) | |
13621 | return false; | |
13622 | extract_base_offset_in_addr (mem_2, &base_2, &offset_2); | |
13623 | if (base_2 == NULL_RTX || offset_2 == NULL_RTX) | |
13624 | return false; | |
13625 | ||
13626 | /* Check if the bases are same. */ | |
13627 | if (!rtx_equal_p (base_1, base_2)) | |
13628 | return false; | |
13629 | ||
13630 | offval_1 = INTVAL (offset_1); | |
13631 | offval_2 = INTVAL (offset_2); | |
13632 | msize = GET_MODE_SIZE (mode); | |
13633 | /* Check if the offsets are consecutive. */ | |
13634 | if (offval_1 != (offval_2 + msize) && offval_2 != (offval_1 + msize)) | |
13635 | return false; | |
13636 | ||
13637 | /* Check if the addresses are clobbered by load. */ | |
13638 | if (load) | |
13639 | { | |
13640 | if (reg_mentioned_p (reg_1, mem_1)) | |
13641 | return false; | |
13642 | ||
13643 | /* In increasing order, the last load can clobber the address. */ | |
13644 | if (offval_1 > offval_2 && reg_mentioned_p (reg_2, mem_2)) | |
13645 | return false; | |
13646 | } | |
13647 | ||
13648 | if (REG_P (reg_1) && FP_REGNUM_P (REGNO (reg_1))) | |
13649 | rclass_1 = FP_REGS; | |
13650 | else | |
13651 | rclass_1 = GENERAL_REGS; | |
13652 | ||
13653 | if (REG_P (reg_2) && FP_REGNUM_P (REGNO (reg_2))) | |
13654 | rclass_2 = FP_REGS; | |
13655 | else | |
13656 | rclass_2 = GENERAL_REGS; | |
13657 | ||
13658 | /* Check if the registers are of same class. */ | |
13659 | if (rclass_1 != rclass_2) | |
13660 | return false; | |
13661 | ||
13662 | return true; | |
13663 | } | |
13664 | ||
13665 | /* Given OPERANDS of consecutive load/store, check if we can merge | |
13666 | them into ldp/stp by adjusting the offset. LOAD is true if they | |
13667 | are load instructions. MODE is the mode of memory operands. | |
13668 | ||
13669 | Given below consecutive stores: | |
13670 | ||
13671 | str w1, [xb, 0x100] | |
13672 | str w1, [xb, 0x104] | |
13673 | str w1, [xb, 0x108] | |
13674 | str w1, [xb, 0x10c] | |
13675 | ||
13676 | Though the offsets are out of the range supported by stp, we can | |
13677 | still pair them after adjusting the offset, like: | |
13678 | ||
13679 | add scratch, xb, 0x100 | |
13680 | stp w1, w1, [scratch] | |
13681 | stp w1, w1, [scratch, 0x8] | |
13682 | ||
13683 | The peephole patterns detecting this opportunity should guarantee | |
13684 | the scratch register is avaliable. */ | |
13685 | ||
13686 | bool | |
13687 | aarch64_operands_adjust_ok_for_ldpstp (rtx *operands, bool load, | |
13688 | enum machine_mode mode) | |
13689 | { | |
13690 | enum reg_class rclass_1, rclass_2, rclass_3, rclass_4; | |
13691 | HOST_WIDE_INT offval_1, offval_2, offval_3, offval_4, msize; | |
13692 | rtx mem_1, mem_2, mem_3, mem_4, reg_1, reg_2, reg_3, reg_4; | |
13693 | rtx base_1, base_2, base_3, base_4, offset_1, offset_2, offset_3, offset_4; | |
13694 | ||
13695 | if (load) | |
13696 | { | |
13697 | reg_1 = operands[0]; | |
13698 | mem_1 = operands[1]; | |
13699 | reg_2 = operands[2]; | |
13700 | mem_2 = operands[3]; | |
13701 | reg_3 = operands[4]; | |
13702 | mem_3 = operands[5]; | |
13703 | reg_4 = operands[6]; | |
13704 | mem_4 = operands[7]; | |
13705 | gcc_assert (REG_P (reg_1) && REG_P (reg_2) | |
13706 | && REG_P (reg_3) && REG_P (reg_4)); | |
13707 | if (REGNO (reg_1) == REGNO (reg_2) || REGNO (reg_3) == REGNO (reg_4)) | |
13708 | return false; | |
13709 | } | |
13710 | else | |
13711 | { | |
13712 | mem_1 = operands[0]; | |
13713 | reg_1 = operands[1]; | |
13714 | mem_2 = operands[2]; | |
13715 | reg_2 = operands[3]; | |
13716 | mem_3 = operands[4]; | |
13717 | reg_3 = operands[5]; | |
13718 | mem_4 = operands[6]; | |
13719 | reg_4 = operands[7]; | |
13720 | } | |
13721 | /* Skip if memory operand is by itslef valid for ldp/stp. */ | |
13722 | if (!MEM_P (mem_1) || aarch64_mem_pair_operand (mem_1, mode)) | |
13723 | return false; | |
13724 | ||
bf84ac44 AP |
13725 | /* The mems cannot be volatile. */ |
13726 | if (MEM_VOLATILE_P (mem_1) || MEM_VOLATILE_P (mem_2) | |
13727 | || MEM_VOLATILE_P (mem_3) ||MEM_VOLATILE_P (mem_4)) | |
13728 | return false; | |
13729 | ||
350013bc BC |
13730 | /* Check if the addresses are in the form of [base+offset]. */ |
13731 | extract_base_offset_in_addr (mem_1, &base_1, &offset_1); | |
13732 | if (base_1 == NULL_RTX || offset_1 == NULL_RTX) | |
13733 | return false; | |
13734 | extract_base_offset_in_addr (mem_2, &base_2, &offset_2); | |
13735 | if (base_2 == NULL_RTX || offset_2 == NULL_RTX) | |
13736 | return false; | |
13737 | extract_base_offset_in_addr (mem_3, &base_3, &offset_3); | |
13738 | if (base_3 == NULL_RTX || offset_3 == NULL_RTX) | |
13739 | return false; | |
13740 | extract_base_offset_in_addr (mem_4, &base_4, &offset_4); | |
13741 | if (base_4 == NULL_RTX || offset_4 == NULL_RTX) | |
13742 | return false; | |
13743 | ||
13744 | /* Check if the bases are same. */ | |
13745 | if (!rtx_equal_p (base_1, base_2) | |
13746 | || !rtx_equal_p (base_2, base_3) | |
13747 | || !rtx_equal_p (base_3, base_4)) | |
13748 | return false; | |
13749 | ||
13750 | offval_1 = INTVAL (offset_1); | |
13751 | offval_2 = INTVAL (offset_2); | |
13752 | offval_3 = INTVAL (offset_3); | |
13753 | offval_4 = INTVAL (offset_4); | |
13754 | msize = GET_MODE_SIZE (mode); | |
13755 | /* Check if the offsets are consecutive. */ | |
13756 | if ((offval_1 != (offval_2 + msize) | |
13757 | || offval_1 != (offval_3 + msize * 2) | |
13758 | || offval_1 != (offval_4 + msize * 3)) | |
13759 | && (offval_4 != (offval_3 + msize) | |
13760 | || offval_4 != (offval_2 + msize * 2) | |
13761 | || offval_4 != (offval_1 + msize * 3))) | |
13762 | return false; | |
13763 | ||
13764 | /* Check if the addresses are clobbered by load. */ | |
13765 | if (load) | |
13766 | { | |
13767 | if (reg_mentioned_p (reg_1, mem_1) | |
13768 | || reg_mentioned_p (reg_2, mem_2) | |
13769 | || reg_mentioned_p (reg_3, mem_3)) | |
13770 | return false; | |
13771 | ||
13772 | /* In increasing order, the last load can clobber the address. */ | |
13773 | if (offval_1 > offval_2 && reg_mentioned_p (reg_4, mem_4)) | |
13774 | return false; | |
13775 | } | |
13776 | ||
13777 | if (REG_P (reg_1) && FP_REGNUM_P (REGNO (reg_1))) | |
13778 | rclass_1 = FP_REGS; | |
13779 | else | |
13780 | rclass_1 = GENERAL_REGS; | |
13781 | ||
13782 | if (REG_P (reg_2) && FP_REGNUM_P (REGNO (reg_2))) | |
13783 | rclass_2 = FP_REGS; | |
13784 | else | |
13785 | rclass_2 = GENERAL_REGS; | |
13786 | ||
13787 | if (REG_P (reg_3) && FP_REGNUM_P (REGNO (reg_3))) | |
13788 | rclass_3 = FP_REGS; | |
13789 | else | |
13790 | rclass_3 = GENERAL_REGS; | |
13791 | ||
13792 | if (REG_P (reg_4) && FP_REGNUM_P (REGNO (reg_4))) | |
13793 | rclass_4 = FP_REGS; | |
13794 | else | |
13795 | rclass_4 = GENERAL_REGS; | |
13796 | ||
13797 | /* Check if the registers are of same class. */ | |
13798 | if (rclass_1 != rclass_2 || rclass_2 != rclass_3 || rclass_3 != rclass_4) | |
13799 | return false; | |
13800 | ||
13801 | return true; | |
13802 | } | |
13803 | ||
13804 | /* Given OPERANDS of consecutive load/store, this function pairs them | |
13805 | into ldp/stp after adjusting the offset. It depends on the fact | |
13806 | that addresses of load/store instructions are in increasing order. | |
13807 | MODE is the mode of memory operands. CODE is the rtl operator | |
13808 | which should be applied to all memory operands, it's SIGN_EXTEND, | |
13809 | ZERO_EXTEND or UNKNOWN. */ | |
13810 | ||
13811 | bool | |
13812 | aarch64_gen_adjusted_ldpstp (rtx *operands, bool load, | |
13813 | enum machine_mode mode, RTX_CODE code) | |
13814 | { | |
13815 | rtx base, offset, t1, t2; | |
13816 | rtx mem_1, mem_2, mem_3, mem_4; | |
13817 | HOST_WIDE_INT off_val, abs_off, adj_off, new_off, stp_off_limit, msize; | |
13818 | ||
13819 | if (load) | |
13820 | { | |
13821 | mem_1 = operands[1]; | |
13822 | mem_2 = operands[3]; | |
13823 | mem_3 = operands[5]; | |
13824 | mem_4 = operands[7]; | |
13825 | } | |
13826 | else | |
13827 | { | |
13828 | mem_1 = operands[0]; | |
13829 | mem_2 = operands[2]; | |
13830 | mem_3 = operands[4]; | |
13831 | mem_4 = operands[6]; | |
13832 | gcc_assert (code == UNKNOWN); | |
13833 | } | |
13834 | ||
13835 | extract_base_offset_in_addr (mem_1, &base, &offset); | |
13836 | gcc_assert (base != NULL_RTX && offset != NULL_RTX); | |
13837 | ||
13838 | /* Adjust offset thus it can fit in ldp/stp instruction. */ | |
13839 | msize = GET_MODE_SIZE (mode); | |
13840 | stp_off_limit = msize * 0x40; | |
13841 | off_val = INTVAL (offset); | |
13842 | abs_off = (off_val < 0) ? -off_val : off_val; | |
13843 | new_off = abs_off % stp_off_limit; | |
13844 | adj_off = abs_off - new_off; | |
13845 | ||
13846 | /* Further adjust to make sure all offsets are OK. */ | |
13847 | if ((new_off + msize * 2) >= stp_off_limit) | |
13848 | { | |
13849 | adj_off += stp_off_limit; | |
13850 | new_off -= stp_off_limit; | |
13851 | } | |
13852 | ||
13853 | /* Make sure the adjustment can be done with ADD/SUB instructions. */ | |
13854 | if (adj_off >= 0x1000) | |
13855 | return false; | |
13856 | ||
13857 | if (off_val < 0) | |
13858 | { | |
13859 | adj_off = -adj_off; | |
13860 | new_off = -new_off; | |
13861 | } | |
13862 | ||
13863 | /* Create new memory references. */ | |
13864 | mem_1 = change_address (mem_1, VOIDmode, | |
13865 | plus_constant (DImode, operands[8], new_off)); | |
13866 | ||
13867 | /* Check if the adjusted address is OK for ldp/stp. */ | |
13868 | if (!aarch64_mem_pair_operand (mem_1, mode)) | |
13869 | return false; | |
13870 | ||
13871 | msize = GET_MODE_SIZE (mode); | |
13872 | mem_2 = change_address (mem_2, VOIDmode, | |
13873 | plus_constant (DImode, | |
13874 | operands[8], | |
13875 | new_off + msize)); | |
13876 | mem_3 = change_address (mem_3, VOIDmode, | |
13877 | plus_constant (DImode, | |
13878 | operands[8], | |
13879 | new_off + msize * 2)); | |
13880 | mem_4 = change_address (mem_4, VOIDmode, | |
13881 | plus_constant (DImode, | |
13882 | operands[8], | |
13883 | new_off + msize * 3)); | |
13884 | ||
13885 | if (code == ZERO_EXTEND) | |
13886 | { | |
13887 | mem_1 = gen_rtx_ZERO_EXTEND (DImode, mem_1); | |
13888 | mem_2 = gen_rtx_ZERO_EXTEND (DImode, mem_2); | |
13889 | mem_3 = gen_rtx_ZERO_EXTEND (DImode, mem_3); | |
13890 | mem_4 = gen_rtx_ZERO_EXTEND (DImode, mem_4); | |
13891 | } | |
13892 | else if (code == SIGN_EXTEND) | |
13893 | { | |
13894 | mem_1 = gen_rtx_SIGN_EXTEND (DImode, mem_1); | |
13895 | mem_2 = gen_rtx_SIGN_EXTEND (DImode, mem_2); | |
13896 | mem_3 = gen_rtx_SIGN_EXTEND (DImode, mem_3); | |
13897 | mem_4 = gen_rtx_SIGN_EXTEND (DImode, mem_4); | |
13898 | } | |
13899 | ||
13900 | if (load) | |
13901 | { | |
13902 | operands[1] = mem_1; | |
13903 | operands[3] = mem_2; | |
13904 | operands[5] = mem_3; | |
13905 | operands[7] = mem_4; | |
13906 | } | |
13907 | else | |
13908 | { | |
13909 | operands[0] = mem_1; | |
13910 | operands[2] = mem_2; | |
13911 | operands[4] = mem_3; | |
13912 | operands[6] = mem_4; | |
13913 | } | |
13914 | ||
13915 | /* Emit adjusting instruction. */ | |
f7df4a84 | 13916 | emit_insn (gen_rtx_SET (operands[8], plus_constant (DImode, base, adj_off))); |
350013bc | 13917 | /* Emit ldp/stp instructions. */ |
f7df4a84 RS |
13918 | t1 = gen_rtx_SET (operands[0], operands[1]); |
13919 | t2 = gen_rtx_SET (operands[2], operands[3]); | |
350013bc | 13920 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, t1, t2))); |
f7df4a84 RS |
13921 | t1 = gen_rtx_SET (operands[4], operands[5]); |
13922 | t2 = gen_rtx_SET (operands[6], operands[7]); | |
350013bc BC |
13923 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, t1, t2))); |
13924 | return true; | |
13925 | } | |
13926 | ||
1b1e81f8 JW |
13927 | /* Return 1 if pseudo register should be created and used to hold |
13928 | GOT address for PIC code. */ | |
13929 | ||
13930 | bool | |
13931 | aarch64_use_pseudo_pic_reg (void) | |
13932 | { | |
13933 | return aarch64_cmodel == AARCH64_CMODEL_SMALL_SPIC; | |
13934 | } | |
13935 | ||
7b841a12 JW |
13936 | /* Implement TARGET_UNSPEC_MAY_TRAP_P. */ |
13937 | ||
13938 | static int | |
13939 | aarch64_unspec_may_trap_p (const_rtx x, unsigned flags) | |
13940 | { | |
13941 | switch (XINT (x, 1)) | |
13942 | { | |
13943 | case UNSPEC_GOTSMALLPIC: | |
13944 | case UNSPEC_GOTSMALLPIC28K: | |
13945 | case UNSPEC_GOTTINYPIC: | |
13946 | return 0; | |
13947 | default: | |
13948 | break; | |
13949 | } | |
13950 | ||
13951 | return default_unspec_may_trap_p (x, flags); | |
13952 | } | |
13953 | ||
39252973 KT |
13954 | |
13955 | /* If X is a positive CONST_DOUBLE with a value that is a power of 2 | |
13956 | return the log2 of that value. Otherwise return -1. */ | |
13957 | ||
13958 | int | |
13959 | aarch64_fpconst_pow_of_2 (rtx x) | |
13960 | { | |
13961 | const REAL_VALUE_TYPE *r; | |
13962 | ||
13963 | if (!CONST_DOUBLE_P (x)) | |
13964 | return -1; | |
13965 | ||
13966 | r = CONST_DOUBLE_REAL_VALUE (x); | |
13967 | ||
13968 | if (REAL_VALUE_NEGATIVE (*r) | |
13969 | || REAL_VALUE_ISNAN (*r) | |
13970 | || REAL_VALUE_ISINF (*r) | |
13971 | || !real_isinteger (r, DFmode)) | |
13972 | return -1; | |
13973 | ||
13974 | return exact_log2 (real_to_integer (r)); | |
13975 | } | |
13976 | ||
13977 | /* If X is a vector of equal CONST_DOUBLE values and that value is | |
13978 | Y, return the aarch64_fpconst_pow_of_2 of Y. Otherwise return -1. */ | |
13979 | ||
13980 | int | |
13981 | aarch64_vec_fpconst_pow_of_2 (rtx x) | |
13982 | { | |
13983 | if (GET_CODE (x) != CONST_VECTOR) | |
13984 | return -1; | |
13985 | ||
13986 | if (GET_MODE_CLASS (GET_MODE (x)) != MODE_VECTOR_FLOAT) | |
13987 | return -1; | |
13988 | ||
13989 | int firstval = aarch64_fpconst_pow_of_2 (CONST_VECTOR_ELT (x, 0)); | |
13990 | if (firstval <= 0) | |
13991 | return -1; | |
13992 | ||
13993 | for (int i = 1; i < CONST_VECTOR_NUNITS (x); i++) | |
13994 | if (aarch64_fpconst_pow_of_2 (CONST_VECTOR_ELT (x, i)) != firstval) | |
13995 | return -1; | |
13996 | ||
13997 | return firstval; | |
13998 | } | |
13999 | ||
c2ec330c AL |
14000 | /* Implement TARGET_PROMOTED_TYPE to promote __fp16 to float. */ |
14001 | static tree | |
14002 | aarch64_promoted_type (const_tree t) | |
14003 | { | |
14004 | if (SCALAR_FLOAT_TYPE_P (t) && TYPE_PRECISION (t) == 16) | |
14005 | return float_type_node; | |
14006 | return NULL_TREE; | |
14007 | } | |
ee62a5a6 RS |
14008 | |
14009 | /* Implement the TARGET_OPTAB_SUPPORTED_P hook. */ | |
14010 | ||
14011 | static bool | |
9acc9cbe | 14012 | aarch64_optab_supported_p (int op, machine_mode mode1, machine_mode, |
ee62a5a6 RS |
14013 | optimization_type opt_type) |
14014 | { | |
14015 | switch (op) | |
14016 | { | |
14017 | case rsqrt_optab: | |
9acc9cbe | 14018 | return opt_type == OPTIMIZE_FOR_SPEED && use_rsqrt_p (mode1); |
ee62a5a6 RS |
14019 | |
14020 | default: | |
14021 | return true; | |
14022 | } | |
14023 | } | |
14024 | ||
43e9d192 IB |
14025 | #undef TARGET_ADDRESS_COST |
14026 | #define TARGET_ADDRESS_COST aarch64_address_cost | |
14027 | ||
14028 | /* This hook will determines whether unnamed bitfields affect the alignment | |
14029 | of the containing structure. The hook returns true if the structure | |
14030 | should inherit the alignment requirements of an unnamed bitfield's | |
14031 | type. */ | |
14032 | #undef TARGET_ALIGN_ANON_BITFIELD | |
14033 | #define TARGET_ALIGN_ANON_BITFIELD hook_bool_void_true | |
14034 | ||
14035 | #undef TARGET_ASM_ALIGNED_DI_OP | |
14036 | #define TARGET_ASM_ALIGNED_DI_OP "\t.xword\t" | |
14037 | ||
14038 | #undef TARGET_ASM_ALIGNED_HI_OP | |
14039 | #define TARGET_ASM_ALIGNED_HI_OP "\t.hword\t" | |
14040 | ||
14041 | #undef TARGET_ASM_ALIGNED_SI_OP | |
14042 | #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t" | |
14043 | ||
14044 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
14045 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK \ | |
14046 | hook_bool_const_tree_hwi_hwi_const_tree_true | |
14047 | ||
e1c1ecb0 KT |
14048 | #undef TARGET_ASM_FILE_START |
14049 | #define TARGET_ASM_FILE_START aarch64_start_file | |
14050 | ||
43e9d192 IB |
14051 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
14052 | #define TARGET_ASM_OUTPUT_MI_THUNK aarch64_output_mi_thunk | |
14053 | ||
14054 | #undef TARGET_ASM_SELECT_RTX_SECTION | |
14055 | #define TARGET_ASM_SELECT_RTX_SECTION aarch64_select_rtx_section | |
14056 | ||
14057 | #undef TARGET_ASM_TRAMPOLINE_TEMPLATE | |
14058 | #define TARGET_ASM_TRAMPOLINE_TEMPLATE aarch64_asm_trampoline_template | |
14059 | ||
14060 | #undef TARGET_BUILD_BUILTIN_VA_LIST | |
14061 | #define TARGET_BUILD_BUILTIN_VA_LIST aarch64_build_builtin_va_list | |
14062 | ||
14063 | #undef TARGET_CALLEE_COPIES | |
14064 | #define TARGET_CALLEE_COPIES hook_bool_CUMULATIVE_ARGS_mode_tree_bool_false | |
14065 | ||
14066 | #undef TARGET_CAN_ELIMINATE | |
14067 | #define TARGET_CAN_ELIMINATE aarch64_can_eliminate | |
14068 | ||
1fd8d40c KT |
14069 | #undef TARGET_CAN_INLINE_P |
14070 | #define TARGET_CAN_INLINE_P aarch64_can_inline_p | |
14071 | ||
43e9d192 IB |
14072 | #undef TARGET_CANNOT_FORCE_CONST_MEM |
14073 | #define TARGET_CANNOT_FORCE_CONST_MEM aarch64_cannot_force_const_mem | |
14074 | ||
50487d79 EM |
14075 | #undef TARGET_CASE_VALUES_THRESHOLD |
14076 | #define TARGET_CASE_VALUES_THRESHOLD aarch64_case_values_threshold | |
14077 | ||
43e9d192 IB |
14078 | #undef TARGET_CONDITIONAL_REGISTER_USAGE |
14079 | #define TARGET_CONDITIONAL_REGISTER_USAGE aarch64_conditional_register_usage | |
14080 | ||
14081 | /* Only the least significant bit is used for initialization guard | |
14082 | variables. */ | |
14083 | #undef TARGET_CXX_GUARD_MASK_BIT | |
14084 | #define TARGET_CXX_GUARD_MASK_BIT hook_bool_void_true | |
14085 | ||
14086 | #undef TARGET_C_MODE_FOR_SUFFIX | |
14087 | #define TARGET_C_MODE_FOR_SUFFIX aarch64_c_mode_for_suffix | |
14088 | ||
14089 | #ifdef TARGET_BIG_ENDIAN_DEFAULT | |
14090 | #undef TARGET_DEFAULT_TARGET_FLAGS | |
14091 | #define TARGET_DEFAULT_TARGET_FLAGS (MASK_BIG_END) | |
14092 | #endif | |
14093 | ||
14094 | #undef TARGET_CLASS_MAX_NREGS | |
14095 | #define TARGET_CLASS_MAX_NREGS aarch64_class_max_nregs | |
14096 | ||
119103ca JG |
14097 | #undef TARGET_BUILTIN_DECL |
14098 | #define TARGET_BUILTIN_DECL aarch64_builtin_decl | |
14099 | ||
a6fc00da BH |
14100 | #undef TARGET_BUILTIN_RECIPROCAL |
14101 | #define TARGET_BUILTIN_RECIPROCAL aarch64_builtin_reciprocal | |
14102 | ||
43e9d192 IB |
14103 | #undef TARGET_EXPAND_BUILTIN |
14104 | #define TARGET_EXPAND_BUILTIN aarch64_expand_builtin | |
14105 | ||
14106 | #undef TARGET_EXPAND_BUILTIN_VA_START | |
14107 | #define TARGET_EXPAND_BUILTIN_VA_START aarch64_expand_builtin_va_start | |
14108 | ||
9697e620 JG |
14109 | #undef TARGET_FOLD_BUILTIN |
14110 | #define TARGET_FOLD_BUILTIN aarch64_fold_builtin | |
14111 | ||
43e9d192 IB |
14112 | #undef TARGET_FUNCTION_ARG |
14113 | #define TARGET_FUNCTION_ARG aarch64_function_arg | |
14114 | ||
14115 | #undef TARGET_FUNCTION_ARG_ADVANCE | |
14116 | #define TARGET_FUNCTION_ARG_ADVANCE aarch64_function_arg_advance | |
14117 | ||
14118 | #undef TARGET_FUNCTION_ARG_BOUNDARY | |
14119 | #define TARGET_FUNCTION_ARG_BOUNDARY aarch64_function_arg_boundary | |
14120 | ||
14121 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL | |
14122 | #define TARGET_FUNCTION_OK_FOR_SIBCALL aarch64_function_ok_for_sibcall | |
14123 | ||
14124 | #undef TARGET_FUNCTION_VALUE | |
14125 | #define TARGET_FUNCTION_VALUE aarch64_function_value | |
14126 | ||
14127 | #undef TARGET_FUNCTION_VALUE_REGNO_P | |
14128 | #define TARGET_FUNCTION_VALUE_REGNO_P aarch64_function_value_regno_p | |
14129 | ||
14130 | #undef TARGET_FRAME_POINTER_REQUIRED | |
14131 | #define TARGET_FRAME_POINTER_REQUIRED aarch64_frame_pointer_required | |
14132 | ||
fc72cba7 AL |
14133 | #undef TARGET_GIMPLE_FOLD_BUILTIN |
14134 | #define TARGET_GIMPLE_FOLD_BUILTIN aarch64_gimple_fold_builtin | |
0ac198d3 | 14135 | |
43e9d192 IB |
14136 | #undef TARGET_GIMPLIFY_VA_ARG_EXPR |
14137 | #define TARGET_GIMPLIFY_VA_ARG_EXPR aarch64_gimplify_va_arg_expr | |
14138 | ||
14139 | #undef TARGET_INIT_BUILTINS | |
14140 | #define TARGET_INIT_BUILTINS aarch64_init_builtins | |
14141 | ||
c64f7d37 WD |
14142 | #undef TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS |
14143 | #define TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS \ | |
14144 | aarch64_ira_change_pseudo_allocno_class | |
14145 | ||
43e9d192 IB |
14146 | #undef TARGET_LEGITIMATE_ADDRESS_P |
14147 | #define TARGET_LEGITIMATE_ADDRESS_P aarch64_legitimate_address_hook_p | |
14148 | ||
14149 | #undef TARGET_LEGITIMATE_CONSTANT_P | |
14150 | #define TARGET_LEGITIMATE_CONSTANT_P aarch64_legitimate_constant_p | |
14151 | ||
14152 | #undef TARGET_LIBGCC_CMP_RETURN_MODE | |
14153 | #define TARGET_LIBGCC_CMP_RETURN_MODE aarch64_libgcc_cmp_return_mode | |
14154 | ||
38e8f663 | 14155 | #undef TARGET_LRA_P |
98d404be | 14156 | #define TARGET_LRA_P hook_bool_void_true |
38e8f663 | 14157 | |
ac2b960f YZ |
14158 | #undef TARGET_MANGLE_TYPE |
14159 | #define TARGET_MANGLE_TYPE aarch64_mangle_type | |
14160 | ||
43e9d192 IB |
14161 | #undef TARGET_MEMORY_MOVE_COST |
14162 | #define TARGET_MEMORY_MOVE_COST aarch64_memory_move_cost | |
14163 | ||
26e0ff94 WD |
14164 | #undef TARGET_MIN_DIVISIONS_FOR_RECIP_MUL |
14165 | #define TARGET_MIN_DIVISIONS_FOR_RECIP_MUL aarch64_min_divisions_for_recip_mul | |
14166 | ||
43e9d192 IB |
14167 | #undef TARGET_MUST_PASS_IN_STACK |
14168 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
14169 | ||
14170 | /* This target hook should return true if accesses to volatile bitfields | |
14171 | should use the narrowest mode possible. It should return false if these | |
14172 | accesses should use the bitfield container type. */ | |
14173 | #undef TARGET_NARROW_VOLATILE_BITFIELD | |
14174 | #define TARGET_NARROW_VOLATILE_BITFIELD hook_bool_void_false | |
14175 | ||
14176 | #undef TARGET_OPTION_OVERRIDE | |
14177 | #define TARGET_OPTION_OVERRIDE aarch64_override_options | |
14178 | ||
14179 | #undef TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE | |
14180 | #define TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE \ | |
14181 | aarch64_override_options_after_change | |
14182 | ||
361fb3ee KT |
14183 | #undef TARGET_OPTION_SAVE |
14184 | #define TARGET_OPTION_SAVE aarch64_option_save | |
14185 | ||
14186 | #undef TARGET_OPTION_RESTORE | |
14187 | #define TARGET_OPTION_RESTORE aarch64_option_restore | |
14188 | ||
14189 | #undef TARGET_OPTION_PRINT | |
14190 | #define TARGET_OPTION_PRINT aarch64_option_print | |
14191 | ||
5a2c8331 KT |
14192 | #undef TARGET_OPTION_VALID_ATTRIBUTE_P |
14193 | #define TARGET_OPTION_VALID_ATTRIBUTE_P aarch64_option_valid_attribute_p | |
14194 | ||
d78006d9 KT |
14195 | #undef TARGET_SET_CURRENT_FUNCTION |
14196 | #define TARGET_SET_CURRENT_FUNCTION aarch64_set_current_function | |
14197 | ||
43e9d192 IB |
14198 | #undef TARGET_PASS_BY_REFERENCE |
14199 | #define TARGET_PASS_BY_REFERENCE aarch64_pass_by_reference | |
14200 | ||
14201 | #undef TARGET_PREFERRED_RELOAD_CLASS | |
14202 | #define TARGET_PREFERRED_RELOAD_CLASS aarch64_preferred_reload_class | |
14203 | ||
cee66c68 WD |
14204 | #undef TARGET_SCHED_REASSOCIATION_WIDTH |
14205 | #define TARGET_SCHED_REASSOCIATION_WIDTH aarch64_reassociation_width | |
14206 | ||
c2ec330c AL |
14207 | #undef TARGET_PROMOTED_TYPE |
14208 | #define TARGET_PROMOTED_TYPE aarch64_promoted_type | |
14209 | ||
43e9d192 IB |
14210 | #undef TARGET_SECONDARY_RELOAD |
14211 | #define TARGET_SECONDARY_RELOAD aarch64_secondary_reload | |
14212 | ||
14213 | #undef TARGET_SHIFT_TRUNCATION_MASK | |
14214 | #define TARGET_SHIFT_TRUNCATION_MASK aarch64_shift_truncation_mask | |
14215 | ||
14216 | #undef TARGET_SETUP_INCOMING_VARARGS | |
14217 | #define TARGET_SETUP_INCOMING_VARARGS aarch64_setup_incoming_varargs | |
14218 | ||
14219 | #undef TARGET_STRUCT_VALUE_RTX | |
14220 | #define TARGET_STRUCT_VALUE_RTX aarch64_struct_value_rtx | |
14221 | ||
14222 | #undef TARGET_REGISTER_MOVE_COST | |
14223 | #define TARGET_REGISTER_MOVE_COST aarch64_register_move_cost | |
14224 | ||
14225 | #undef TARGET_RETURN_IN_MEMORY | |
14226 | #define TARGET_RETURN_IN_MEMORY aarch64_return_in_memory | |
14227 | ||
14228 | #undef TARGET_RETURN_IN_MSB | |
14229 | #define TARGET_RETURN_IN_MSB aarch64_return_in_msb | |
14230 | ||
14231 | #undef TARGET_RTX_COSTS | |
7cc2145f | 14232 | #define TARGET_RTX_COSTS aarch64_rtx_costs_wrapper |
43e9d192 | 14233 | |
d126a4ae AP |
14234 | #undef TARGET_SCHED_ISSUE_RATE |
14235 | #define TARGET_SCHED_ISSUE_RATE aarch64_sched_issue_rate | |
14236 | ||
d03f7e44 MK |
14237 | #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD |
14238 | #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD \ | |
14239 | aarch64_sched_first_cycle_multipass_dfa_lookahead | |
14240 | ||
2d6bc7fa KT |
14241 | #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD |
14242 | #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD \ | |
14243 | aarch64_first_cycle_multipass_dfa_lookahead_guard | |
14244 | ||
43e9d192 IB |
14245 | #undef TARGET_TRAMPOLINE_INIT |
14246 | #define TARGET_TRAMPOLINE_INIT aarch64_trampoline_init | |
14247 | ||
14248 | #undef TARGET_USE_BLOCKS_FOR_CONSTANT_P | |
14249 | #define TARGET_USE_BLOCKS_FOR_CONSTANT_P aarch64_use_blocks_for_constant_p | |
14250 | ||
14251 | #undef TARGET_VECTOR_MODE_SUPPORTED_P | |
14252 | #define TARGET_VECTOR_MODE_SUPPORTED_P aarch64_vector_mode_supported_p | |
14253 | ||
14254 | #undef TARGET_ARRAY_MODE_SUPPORTED_P | |
14255 | #define TARGET_ARRAY_MODE_SUPPORTED_P aarch64_array_mode_supported_p | |
14256 | ||
8990e73a TB |
14257 | #undef TARGET_VECTORIZE_ADD_STMT_COST |
14258 | #define TARGET_VECTORIZE_ADD_STMT_COST aarch64_add_stmt_cost | |
14259 | ||
14260 | #undef TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST | |
14261 | #define TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST \ | |
14262 | aarch64_builtin_vectorization_cost | |
14263 | ||
43e9d192 IB |
14264 | #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE |
14265 | #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE aarch64_preferred_simd_mode | |
14266 | ||
42fc9a7f JG |
14267 | #undef TARGET_VECTORIZE_BUILTINS |
14268 | #define TARGET_VECTORIZE_BUILTINS | |
14269 | ||
14270 | #undef TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION | |
14271 | #define TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION \ | |
14272 | aarch64_builtin_vectorized_function | |
14273 | ||
3b357264 JG |
14274 | #undef TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES |
14275 | #define TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES \ | |
14276 | aarch64_autovectorize_vector_sizes | |
14277 | ||
aa87aced KV |
14278 | #undef TARGET_ATOMIC_ASSIGN_EXPAND_FENV |
14279 | #define TARGET_ATOMIC_ASSIGN_EXPAND_FENV \ | |
14280 | aarch64_atomic_assign_expand_fenv | |
14281 | ||
43e9d192 IB |
14282 | /* Section anchor support. */ |
14283 | ||
14284 | #undef TARGET_MIN_ANCHOR_OFFSET | |
14285 | #define TARGET_MIN_ANCHOR_OFFSET -256 | |
14286 | ||
14287 | /* Limit the maximum anchor offset to 4k-1, since that's the limit for a | |
14288 | byte offset; we can do much more for larger data types, but have no way | |
14289 | to determine the size of the access. We assume accesses are aligned. */ | |
14290 | #undef TARGET_MAX_ANCHOR_OFFSET | |
14291 | #define TARGET_MAX_ANCHOR_OFFSET 4095 | |
14292 | ||
db0253a4 TB |
14293 | #undef TARGET_VECTOR_ALIGNMENT |
14294 | #define TARGET_VECTOR_ALIGNMENT aarch64_simd_vector_alignment | |
14295 | ||
14296 | #undef TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE | |
14297 | #define TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE \ | |
14298 | aarch64_simd_vector_alignment_reachable | |
14299 | ||
88b08073 JG |
14300 | /* vec_perm support. */ |
14301 | ||
14302 | #undef TARGET_VECTORIZE_VEC_PERM_CONST_OK | |
14303 | #define TARGET_VECTORIZE_VEC_PERM_CONST_OK \ | |
14304 | aarch64_vectorize_vec_perm_const_ok | |
14305 | ||
c2ec330c AL |
14306 | #undef TARGET_INIT_LIBFUNCS |
14307 | #define TARGET_INIT_LIBFUNCS aarch64_init_libfuncs | |
70f09188 | 14308 | |
706b2314 | 14309 | #undef TARGET_FIXED_CONDITION_CODE_REGS |
70f09188 AP |
14310 | #define TARGET_FIXED_CONDITION_CODE_REGS aarch64_fixed_condition_code_regs |
14311 | ||
5cb74e90 RR |
14312 | #undef TARGET_FLAGS_REGNUM |
14313 | #define TARGET_FLAGS_REGNUM CC_REGNUM | |
14314 | ||
78607708 TV |
14315 | #undef TARGET_CALL_FUSAGE_CONTAINS_NON_CALLEE_CLOBBERS |
14316 | #define TARGET_CALL_FUSAGE_CONTAINS_NON_CALLEE_CLOBBERS true | |
14317 | ||
a3125fc2 CL |
14318 | #undef TARGET_ASAN_SHADOW_OFFSET |
14319 | #define TARGET_ASAN_SHADOW_OFFSET aarch64_asan_shadow_offset | |
14320 | ||
0c4ec427 RE |
14321 | #undef TARGET_LEGITIMIZE_ADDRESS |
14322 | #define TARGET_LEGITIMIZE_ADDRESS aarch64_legitimize_address | |
14323 | ||
d3006da6 JG |
14324 | #undef TARGET_USE_BY_PIECES_INFRASTRUCTURE_P |
14325 | #define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \ | |
14326 | aarch64_use_by_pieces_infrastructure_p | |
14327 | ||
594bdd53 FY |
14328 | #undef TARGET_CAN_USE_DOLOOP_P |
14329 | #define TARGET_CAN_USE_DOLOOP_P can_use_doloop_if_innermost | |
14330 | ||
6a569cdd KT |
14331 | #undef TARGET_SCHED_MACRO_FUSION_P |
14332 | #define TARGET_SCHED_MACRO_FUSION_P aarch64_macro_fusion_p | |
14333 | ||
14334 | #undef TARGET_SCHED_MACRO_FUSION_PAIR_P | |
14335 | #define TARGET_SCHED_MACRO_FUSION_PAIR_P aarch_macro_fusion_pair_p | |
14336 | ||
350013bc BC |
14337 | #undef TARGET_SCHED_FUSION_PRIORITY |
14338 | #define TARGET_SCHED_FUSION_PRIORITY aarch64_sched_fusion_priority | |
14339 | ||
7b841a12 JW |
14340 | #undef TARGET_UNSPEC_MAY_TRAP_P |
14341 | #define TARGET_UNSPEC_MAY_TRAP_P aarch64_unspec_may_trap_p | |
14342 | ||
1b1e81f8 JW |
14343 | #undef TARGET_USE_PSEUDO_PIC_REG |
14344 | #define TARGET_USE_PSEUDO_PIC_REG aarch64_use_pseudo_pic_reg | |
14345 | ||
cc8ca59e JB |
14346 | #undef TARGET_PRINT_OPERAND |
14347 | #define TARGET_PRINT_OPERAND aarch64_print_operand | |
14348 | ||
14349 | #undef TARGET_PRINT_OPERAND_ADDRESS | |
14350 | #define TARGET_PRINT_OPERAND_ADDRESS aarch64_print_operand_address | |
14351 | ||
ee62a5a6 RS |
14352 | #undef TARGET_OPTAB_SUPPORTED_P |
14353 | #define TARGET_OPTAB_SUPPORTED_P aarch64_optab_supported_p | |
14354 | ||
43203dea RR |
14355 | #undef TARGET_OMIT_STRUCT_RETURN_REG |
14356 | #define TARGET_OMIT_STRUCT_RETURN_REG true | |
14357 | ||
43e9d192 IB |
14358 | struct gcc_target targetm = TARGET_INITIALIZER; |
14359 | ||
14360 | #include "gt-aarch64.h" |