]>
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 | 28 | #include "tree.h" |
e73cf9a2 | 29 | #include "memmodel.h" |
c7131fb2 | 30 | #include "gimple.h" |
e11c4407 AM |
31 | #include "cfghooks.h" |
32 | #include "cfgloop.h" | |
c7131fb2 | 33 | #include "df.h" |
e11c4407 AM |
34 | #include "tm_p.h" |
35 | #include "stringpool.h" | |
36 | #include "optabs.h" | |
37 | #include "regs.h" | |
38 | #include "emit-rtl.h" | |
39 | #include "recog.h" | |
40 | #include "diagnostic.h" | |
43e9d192 | 41 | #include "insn-attr.h" |
40e23961 | 42 | #include "alias.h" |
40e23961 | 43 | #include "fold-const.h" |
d8a2d370 DN |
44 | #include "stor-layout.h" |
45 | #include "calls.h" | |
46 | #include "varasm.h" | |
43e9d192 | 47 | #include "output.h" |
36566b39 | 48 | #include "flags.h" |
36566b39 | 49 | #include "explow.h" |
43e9d192 IB |
50 | #include "expr.h" |
51 | #include "reload.h" | |
43e9d192 | 52 | #include "langhooks.h" |
5a2c8331 | 53 | #include "opts.h" |
2d6bc7fa | 54 | #include "params.h" |
45b0be94 | 55 | #include "gimplify.h" |
43e9d192 | 56 | #include "dwarf2.h" |
61d371eb | 57 | #include "gimple-iterator.h" |
8990e73a | 58 | #include "tree-vectorizer.h" |
d1bcc29f | 59 | #include "aarch64-cost-tables.h" |
0ee859b5 | 60 | #include "dumpfile.h" |
9b2b7279 | 61 | #include "builtins.h" |
8baff86e | 62 | #include "rtl-iter.h" |
9bbe08fe | 63 | #include "tm-constrs.h" |
d03f7e44 | 64 | #include "sched-int.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 | }; | |
8dec06f2 | 174 | |
a339a01c | 175 | #define AARCH64_EXTRA_TUNING_OPTION(name, internal_name) \ |
8dec06f2 JG |
176 | { name, AARCH64_EXTRA_TUNE_##internal_name }, |
177 | static const struct aarch64_flag_desc aarch64_tuning_flags[] = | |
178 | { | |
179 | { "none", AARCH64_EXTRA_TUNE_NONE }, | |
180 | #include "aarch64-tuning-flags.def" | |
181 | { "all", AARCH64_EXTRA_TUNE_ALL }, | |
182 | { NULL, AARCH64_EXTRA_TUNE_NONE } | |
183 | }; | |
8dec06f2 | 184 | |
43e9d192 IB |
185 | /* Tuning parameters. */ |
186 | ||
43e9d192 IB |
187 | static const struct cpu_addrcost_table generic_addrcost_table = |
188 | { | |
67747367 | 189 | { |
bd95e655 JG |
190 | 0, /* hi */ |
191 | 0, /* si */ | |
192 | 0, /* di */ | |
193 | 0, /* ti */ | |
67747367 | 194 | }, |
bd95e655 JG |
195 | 0, /* pre_modify */ |
196 | 0, /* post_modify */ | |
197 | 0, /* register_offset */ | |
783879e6 EM |
198 | 0, /* register_sextend */ |
199 | 0, /* register_zextend */ | |
bd95e655 | 200 | 0 /* imm_offset */ |
43e9d192 IB |
201 | }; |
202 | ||
60bff090 JG |
203 | static const struct cpu_addrcost_table cortexa57_addrcost_table = |
204 | { | |
60bff090 | 205 | { |
bd95e655 JG |
206 | 1, /* hi */ |
207 | 0, /* si */ | |
208 | 0, /* di */ | |
209 | 1, /* ti */ | |
60bff090 | 210 | }, |
bd95e655 JG |
211 | 0, /* pre_modify */ |
212 | 0, /* post_modify */ | |
213 | 0, /* register_offset */ | |
783879e6 EM |
214 | 0, /* register_sextend */ |
215 | 0, /* register_zextend */ | |
bd95e655 | 216 | 0, /* imm_offset */ |
60bff090 JG |
217 | }; |
218 | ||
5ec1ae3b EM |
219 | static const struct cpu_addrcost_table exynosm1_addrcost_table = |
220 | { | |
221 | { | |
222 | 0, /* hi */ | |
223 | 0, /* si */ | |
224 | 0, /* di */ | |
225 | 2, /* ti */ | |
226 | }, | |
227 | 0, /* pre_modify */ | |
228 | 0, /* post_modify */ | |
229 | 1, /* register_offset */ | |
230 | 1, /* register_sextend */ | |
231 | 2, /* register_zextend */ | |
232 | 0, /* imm_offset */ | |
233 | }; | |
234 | ||
381e27aa PT |
235 | static const struct cpu_addrcost_table xgene1_addrcost_table = |
236 | { | |
381e27aa | 237 | { |
bd95e655 JG |
238 | 1, /* hi */ |
239 | 0, /* si */ | |
240 | 0, /* di */ | |
241 | 1, /* ti */ | |
381e27aa | 242 | }, |
bd95e655 JG |
243 | 1, /* pre_modify */ |
244 | 0, /* post_modify */ | |
245 | 0, /* register_offset */ | |
783879e6 EM |
246 | 1, /* register_sextend */ |
247 | 1, /* register_zextend */ | |
bd95e655 | 248 | 0, /* imm_offset */ |
381e27aa PT |
249 | }; |
250 | ||
ee446d9f JW |
251 | static const struct cpu_addrcost_table qdf24xx_addrcost_table = |
252 | { | |
253 | { | |
254 | 1, /* hi */ | |
255 | 0, /* si */ | |
256 | 0, /* di */ | |
257 | 1, /* ti */ | |
258 | }, | |
259 | 0, /* pre_modify */ | |
260 | 0, /* post_modify */ | |
261 | 0, /* register_offset */ | |
262 | 0, /* register_sextend */ | |
263 | 0, /* register_zextend */ | |
264 | 0 /* imm_offset */ | |
265 | }; | |
266 | ||
ad611a4c VP |
267 | static const struct cpu_addrcost_table vulcan_addrcost_table = |
268 | { | |
269 | { | |
270 | 0, /* hi */ | |
271 | 0, /* si */ | |
272 | 0, /* di */ | |
273 | 2, /* ti */ | |
274 | }, | |
275 | 0, /* pre_modify */ | |
276 | 0, /* post_modify */ | |
277 | 2, /* register_offset */ | |
278 | 3, /* register_sextend */ | |
279 | 3, /* register_zextend */ | |
280 | 0, /* imm_offset */ | |
281 | }; | |
282 | ||
43e9d192 IB |
283 | static const struct cpu_regmove_cost generic_regmove_cost = |
284 | { | |
bd95e655 | 285 | 1, /* GP2GP */ |
3969c510 WD |
286 | /* Avoid the use of slow int<->fp moves for spilling by setting |
287 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
288 | 5, /* GP2FP */ |
289 | 5, /* FP2GP */ | |
290 | 2 /* FP2FP */ | |
43e9d192 IB |
291 | }; |
292 | ||
e4a9c55a WD |
293 | static const struct cpu_regmove_cost cortexa57_regmove_cost = |
294 | { | |
bd95e655 | 295 | 1, /* GP2GP */ |
e4a9c55a WD |
296 | /* Avoid the use of slow int<->fp moves for spilling by setting |
297 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
298 | 5, /* GP2FP */ |
299 | 5, /* FP2GP */ | |
300 | 2 /* FP2FP */ | |
e4a9c55a WD |
301 | }; |
302 | ||
303 | static const struct cpu_regmove_cost cortexa53_regmove_cost = | |
304 | { | |
bd95e655 | 305 | 1, /* GP2GP */ |
e4a9c55a WD |
306 | /* Avoid the use of slow int<->fp moves for spilling by setting |
307 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
308 | 5, /* GP2FP */ |
309 | 5, /* FP2GP */ | |
310 | 2 /* FP2FP */ | |
e4a9c55a WD |
311 | }; |
312 | ||
5ec1ae3b EM |
313 | static const struct cpu_regmove_cost exynosm1_regmove_cost = |
314 | { | |
315 | 1, /* GP2GP */ | |
316 | /* Avoid the use of slow int<->fp moves for spilling by setting | |
317 | their cost higher than memmov_cost (actual, 4 and 9). */ | |
318 | 9, /* GP2FP */ | |
319 | 9, /* FP2GP */ | |
320 | 1 /* FP2FP */ | |
321 | }; | |
322 | ||
d1bcc29f AP |
323 | static const struct cpu_regmove_cost thunderx_regmove_cost = |
324 | { | |
bd95e655 JG |
325 | 2, /* GP2GP */ |
326 | 2, /* GP2FP */ | |
327 | 6, /* FP2GP */ | |
328 | 4 /* FP2FP */ | |
d1bcc29f AP |
329 | }; |
330 | ||
381e27aa PT |
331 | static const struct cpu_regmove_cost xgene1_regmove_cost = |
332 | { | |
bd95e655 | 333 | 1, /* GP2GP */ |
381e27aa PT |
334 | /* Avoid the use of slow int<->fp moves for spilling by setting |
335 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
336 | 8, /* GP2FP */ |
337 | 8, /* FP2GP */ | |
338 | 2 /* FP2FP */ | |
381e27aa PT |
339 | }; |
340 | ||
ee446d9f JW |
341 | static const struct cpu_regmove_cost qdf24xx_regmove_cost = |
342 | { | |
343 | 2, /* GP2GP */ | |
344 | /* Avoid the use of int<->fp moves for spilling. */ | |
345 | 6, /* GP2FP */ | |
346 | 6, /* FP2GP */ | |
347 | 4 /* FP2FP */ | |
348 | }; | |
349 | ||
ad611a4c VP |
350 | static const struct cpu_regmove_cost vulcan_regmove_cost = |
351 | { | |
352 | 1, /* GP2GP */ | |
353 | /* Avoid the use of int<->fp moves for spilling. */ | |
354 | 8, /* GP2FP */ | |
355 | 8, /* FP2GP */ | |
356 | 4 /* FP2FP */ | |
357 | }; | |
358 | ||
8990e73a | 359 | /* Generic costs for vector insn classes. */ |
8990e73a TB |
360 | static const struct cpu_vector_cost generic_vector_cost = |
361 | { | |
bd95e655 JG |
362 | 1, /* scalar_stmt_cost */ |
363 | 1, /* scalar_load_cost */ | |
364 | 1, /* scalar_store_cost */ | |
365 | 1, /* vec_stmt_cost */ | |
c428f91c | 366 | 2, /* vec_permute_cost */ |
bd95e655 JG |
367 | 1, /* vec_to_scalar_cost */ |
368 | 1, /* scalar_to_vec_cost */ | |
369 | 1, /* vec_align_load_cost */ | |
370 | 1, /* vec_unalign_load_cost */ | |
371 | 1, /* vec_unalign_store_cost */ | |
372 | 1, /* vec_store_cost */ | |
373 | 3, /* cond_taken_branch_cost */ | |
374 | 1 /* cond_not_taken_branch_cost */ | |
8990e73a TB |
375 | }; |
376 | ||
c3f20327 AP |
377 | /* ThunderX costs for vector insn classes. */ |
378 | static const struct cpu_vector_cost thunderx_vector_cost = | |
379 | { | |
380 | 1, /* scalar_stmt_cost */ | |
381 | 3, /* scalar_load_cost */ | |
382 | 1, /* scalar_store_cost */ | |
383 | 4, /* vec_stmt_cost */ | |
384 | 4, /* vec_permute_cost */ | |
385 | 2, /* vec_to_scalar_cost */ | |
386 | 2, /* scalar_to_vec_cost */ | |
387 | 3, /* vec_align_load_cost */ | |
388 | 10, /* vec_unalign_load_cost */ | |
389 | 10, /* vec_unalign_store_cost */ | |
390 | 1, /* vec_store_cost */ | |
391 | 3, /* cond_taken_branch_cost */ | |
392 | 3 /* cond_not_taken_branch_cost */ | |
393 | }; | |
394 | ||
60bff090 | 395 | /* Generic costs for vector insn classes. */ |
60bff090 JG |
396 | static const struct cpu_vector_cost cortexa57_vector_cost = |
397 | { | |
bd95e655 JG |
398 | 1, /* scalar_stmt_cost */ |
399 | 4, /* scalar_load_cost */ | |
400 | 1, /* scalar_store_cost */ | |
db4a1c18 | 401 | 2, /* vec_stmt_cost */ |
c428f91c | 402 | 3, /* vec_permute_cost */ |
bd95e655 JG |
403 | 8, /* vec_to_scalar_cost */ |
404 | 8, /* scalar_to_vec_cost */ | |
db4a1c18 WD |
405 | 4, /* vec_align_load_cost */ |
406 | 4, /* vec_unalign_load_cost */ | |
bd95e655 JG |
407 | 1, /* vec_unalign_store_cost */ |
408 | 1, /* vec_store_cost */ | |
409 | 1, /* cond_taken_branch_cost */ | |
410 | 1 /* cond_not_taken_branch_cost */ | |
60bff090 JG |
411 | }; |
412 | ||
5ec1ae3b EM |
413 | static const struct cpu_vector_cost exynosm1_vector_cost = |
414 | { | |
415 | 1, /* scalar_stmt_cost */ | |
416 | 5, /* scalar_load_cost */ | |
417 | 1, /* scalar_store_cost */ | |
418 | 3, /* vec_stmt_cost */ | |
c428f91c | 419 | 3, /* vec_permute_cost */ |
5ec1ae3b EM |
420 | 3, /* vec_to_scalar_cost */ |
421 | 3, /* scalar_to_vec_cost */ | |
422 | 5, /* vec_align_load_cost */ | |
423 | 5, /* vec_unalign_load_cost */ | |
424 | 1, /* vec_unalign_store_cost */ | |
425 | 1, /* vec_store_cost */ | |
426 | 1, /* cond_taken_branch_cost */ | |
427 | 1 /* cond_not_taken_branch_cost */ | |
428 | }; | |
429 | ||
381e27aa | 430 | /* Generic costs for vector insn classes. */ |
381e27aa PT |
431 | static const struct cpu_vector_cost xgene1_vector_cost = |
432 | { | |
bd95e655 JG |
433 | 1, /* scalar_stmt_cost */ |
434 | 5, /* scalar_load_cost */ | |
435 | 1, /* scalar_store_cost */ | |
436 | 2, /* vec_stmt_cost */ | |
c428f91c | 437 | 2, /* vec_permute_cost */ |
bd95e655 JG |
438 | 4, /* vec_to_scalar_cost */ |
439 | 4, /* scalar_to_vec_cost */ | |
440 | 10, /* vec_align_load_cost */ | |
441 | 10, /* vec_unalign_load_cost */ | |
442 | 2, /* vec_unalign_store_cost */ | |
443 | 2, /* vec_store_cost */ | |
444 | 2, /* cond_taken_branch_cost */ | |
445 | 1 /* cond_not_taken_branch_cost */ | |
381e27aa PT |
446 | }; |
447 | ||
ad611a4c VP |
448 | /* Costs for vector insn classes for Vulcan. */ |
449 | static const struct cpu_vector_cost vulcan_vector_cost = | |
450 | { | |
451 | 6, /* scalar_stmt_cost */ | |
452 | 4, /* scalar_load_cost */ | |
453 | 1, /* scalar_store_cost */ | |
454 | 6, /* vec_stmt_cost */ | |
455 | 3, /* vec_permute_cost */ | |
456 | 6, /* vec_to_scalar_cost */ | |
457 | 5, /* scalar_to_vec_cost */ | |
458 | 8, /* vec_align_load_cost */ | |
459 | 8, /* vec_unalign_load_cost */ | |
460 | 4, /* vec_unalign_store_cost */ | |
461 | 4, /* vec_store_cost */ | |
462 | 2, /* cond_taken_branch_cost */ | |
463 | 1 /* cond_not_taken_branch_cost */ | |
464 | }; | |
465 | ||
b9066f5a MW |
466 | /* Generic costs for branch instructions. */ |
467 | static const struct cpu_branch_cost generic_branch_cost = | |
468 | { | |
469 | 2, /* Predictable. */ | |
470 | 2 /* Unpredictable. */ | |
471 | }; | |
472 | ||
67707f65 JG |
473 | /* Branch costs for Cortex-A57. */ |
474 | static const struct cpu_branch_cost cortexa57_branch_cost = | |
475 | { | |
476 | 1, /* Predictable. */ | |
477 | 3 /* Unpredictable. */ | |
478 | }; | |
479 | ||
ad611a4c VP |
480 | /* Branch costs for Vulcan. */ |
481 | static const struct cpu_branch_cost vulcan_branch_cost = | |
482 | { | |
483 | 1, /* Predictable. */ | |
484 | 3 /* Unpredictable. */ | |
485 | }; | |
486 | ||
9acc9cbe EM |
487 | /* Generic approximation modes. */ |
488 | static const cpu_approx_modes generic_approx_modes = | |
489 | { | |
79a2bc2d | 490 | AARCH64_APPROX_NONE, /* division */ |
98daafa0 | 491 | AARCH64_APPROX_NONE, /* sqrt */ |
9acc9cbe EM |
492 | AARCH64_APPROX_NONE /* recip_sqrt */ |
493 | }; | |
494 | ||
495 | /* Approximation modes for Exynos M1. */ | |
496 | static const cpu_approx_modes exynosm1_approx_modes = | |
497 | { | |
79a2bc2d | 498 | AARCH64_APPROX_NONE, /* division */ |
98daafa0 | 499 | AARCH64_APPROX_ALL, /* sqrt */ |
9acc9cbe EM |
500 | AARCH64_APPROX_ALL /* recip_sqrt */ |
501 | }; | |
502 | ||
503 | /* Approximation modes for X-Gene 1. */ | |
504 | static const cpu_approx_modes xgene1_approx_modes = | |
505 | { | |
79a2bc2d | 506 | AARCH64_APPROX_NONE, /* division */ |
98daafa0 | 507 | AARCH64_APPROX_NONE, /* sqrt */ |
9acc9cbe EM |
508 | AARCH64_APPROX_ALL /* recip_sqrt */ |
509 | }; | |
510 | ||
43e9d192 IB |
511 | static const struct tune_params generic_tunings = |
512 | { | |
4e2cd668 | 513 | &cortexa57_extra_costs, |
43e9d192 IB |
514 | &generic_addrcost_table, |
515 | &generic_regmove_cost, | |
8990e73a | 516 | &generic_vector_cost, |
b9066f5a | 517 | &generic_branch_cost, |
9acc9cbe | 518 | &generic_approx_modes, |
bd95e655 JG |
519 | 4, /* memmov_cost */ |
520 | 2, /* issue_rate */ | |
e9a3a175 | 521 | AARCH64_FUSE_NOTHING, /* fusible_ops */ |
0b82a5a2 WD |
522 | 8, /* function_align. */ |
523 | 8, /* jump_align. */ | |
524 | 4, /* loop_align. */ | |
cee66c68 WD |
525 | 2, /* int_reassoc_width. */ |
526 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
527 | 1, /* vec_reassoc_width. */ |
528 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 529 | 2, /* min_div_recip_mul_df. */ |
50487d79 EM |
530 | 0, /* max_case_values. */ |
531 | 0, /* cache_line_size. */ | |
2d6bc7fa | 532 | tune_params::AUTOPREFETCHER_OFF, /* autoprefetcher_model. */ |
dfba575f | 533 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ |
43e9d192 IB |
534 | }; |
535 | ||
1c72a3ca JG |
536 | static const struct tune_params cortexa35_tunings = |
537 | { | |
538 | &cortexa53_extra_costs, | |
539 | &generic_addrcost_table, | |
540 | &cortexa53_regmove_cost, | |
541 | &generic_vector_cost, | |
0bc24338 | 542 | &cortexa57_branch_cost, |
9acc9cbe | 543 | &generic_approx_modes, |
1c72a3ca JG |
544 | 4, /* memmov_cost */ |
545 | 1, /* issue_rate */ | |
0bc24338 | 546 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD |
1c72a3ca | 547 | | AARCH64_FUSE_MOVK_MOVK | AARCH64_FUSE_ADRP_LDR), /* fusible_ops */ |
d4407370 | 548 | 16, /* function_align. */ |
1c72a3ca | 549 | 8, /* jump_align. */ |
d4407370 | 550 | 8, /* loop_align. */ |
1c72a3ca JG |
551 | 2, /* int_reassoc_width. */ |
552 | 4, /* fp_reassoc_width. */ | |
553 | 1, /* vec_reassoc_width. */ | |
554 | 2, /* min_div_recip_mul_sf. */ | |
555 | 2, /* min_div_recip_mul_df. */ | |
556 | 0, /* max_case_values. */ | |
557 | 0, /* cache_line_size. */ | |
558 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ | |
559 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ | |
560 | }; | |
561 | ||
984239ad KT |
562 | static const struct tune_params cortexa53_tunings = |
563 | { | |
564 | &cortexa53_extra_costs, | |
565 | &generic_addrcost_table, | |
e4a9c55a | 566 | &cortexa53_regmove_cost, |
984239ad | 567 | &generic_vector_cost, |
0bc24338 | 568 | &cortexa57_branch_cost, |
9acc9cbe | 569 | &generic_approx_modes, |
bd95e655 JG |
570 | 4, /* memmov_cost */ |
571 | 2, /* issue_rate */ | |
00a8574a | 572 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD |
e9a3a175 | 573 | | AARCH64_FUSE_MOVK_MOVK | AARCH64_FUSE_ADRP_LDR), /* fusible_ops */ |
d4407370 | 574 | 16, /* function_align. */ |
0b82a5a2 | 575 | 8, /* jump_align. */ |
d4407370 | 576 | 8, /* loop_align. */ |
cee66c68 WD |
577 | 2, /* int_reassoc_width. */ |
578 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
579 | 1, /* vec_reassoc_width. */ |
580 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 581 | 2, /* min_div_recip_mul_df. */ |
50487d79 EM |
582 | 0, /* max_case_values. */ |
583 | 0, /* cache_line_size. */ | |
2d6bc7fa | 584 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
dfba575f | 585 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ |
984239ad KT |
586 | }; |
587 | ||
4fd92af6 KT |
588 | static const struct tune_params cortexa57_tunings = |
589 | { | |
590 | &cortexa57_extra_costs, | |
60bff090 | 591 | &cortexa57_addrcost_table, |
e4a9c55a | 592 | &cortexa57_regmove_cost, |
60bff090 | 593 | &cortexa57_vector_cost, |
67707f65 | 594 | &cortexa57_branch_cost, |
9acc9cbe | 595 | &generic_approx_modes, |
bd95e655 JG |
596 | 4, /* memmov_cost */ |
597 | 3, /* issue_rate */ | |
00a8574a | 598 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD |
e9a3a175 | 599 | | AARCH64_FUSE_MOVK_MOVK), /* fusible_ops */ |
0b82a5a2 WD |
600 | 16, /* function_align. */ |
601 | 8, /* jump_align. */ | |
d4407370 | 602 | 8, /* loop_align. */ |
cee66c68 WD |
603 | 2, /* int_reassoc_width. */ |
604 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
605 | 1, /* vec_reassoc_width. */ |
606 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 607 | 2, /* min_div_recip_mul_df. */ |
50487d79 EM |
608 | 0, /* max_case_values. */ |
609 | 0, /* cache_line_size. */ | |
2d6bc7fa | 610 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
7c175186 | 611 | (AARCH64_EXTRA_TUNE_RENAME_FMA_REGS) /* tune_flags. */ |
dfba575f JG |
612 | }; |
613 | ||
614 | static const struct tune_params cortexa72_tunings = | |
615 | { | |
616 | &cortexa57_extra_costs, | |
617 | &cortexa57_addrcost_table, | |
618 | &cortexa57_regmove_cost, | |
619 | &cortexa57_vector_cost, | |
0bc24338 | 620 | &cortexa57_branch_cost, |
9acc9cbe | 621 | &generic_approx_modes, |
dfba575f JG |
622 | 4, /* memmov_cost */ |
623 | 3, /* issue_rate */ | |
00a8574a | 624 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD |
dfba575f JG |
625 | | AARCH64_FUSE_MOVK_MOVK), /* fusible_ops */ |
626 | 16, /* function_align. */ | |
627 | 8, /* jump_align. */ | |
d4407370 | 628 | 8, /* loop_align. */ |
dfba575f JG |
629 | 2, /* int_reassoc_width. */ |
630 | 4, /* fp_reassoc_width. */ | |
631 | 1, /* vec_reassoc_width. */ | |
632 | 2, /* min_div_recip_mul_sf. */ | |
633 | 2, /* min_div_recip_mul_df. */ | |
50487d79 EM |
634 | 0, /* max_case_values. */ |
635 | 0, /* cache_line_size. */ | |
0bc24338 | 636 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
dfba575f | 637 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ |
4fd92af6 KT |
638 | }; |
639 | ||
4fb570c4 KT |
640 | static const struct tune_params cortexa73_tunings = |
641 | { | |
642 | &cortexa57_extra_costs, | |
643 | &cortexa57_addrcost_table, | |
644 | &cortexa57_regmove_cost, | |
645 | &cortexa57_vector_cost, | |
0bc24338 | 646 | &cortexa57_branch_cost, |
4fb570c4 KT |
647 | &generic_approx_modes, |
648 | 4, /* memmov_cost. */ | |
649 | 2, /* issue_rate. */ | |
650 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD | |
651 | | AARCH64_FUSE_MOVK_MOVK | AARCH64_FUSE_ADRP_LDR), /* fusible_ops */ | |
652 | 16, /* function_align. */ | |
653 | 8, /* jump_align. */ | |
d4407370 | 654 | 8, /* loop_align. */ |
4fb570c4 KT |
655 | 2, /* int_reassoc_width. */ |
656 | 4, /* fp_reassoc_width. */ | |
657 | 1, /* vec_reassoc_width. */ | |
658 | 2, /* min_div_recip_mul_sf. */ | |
659 | 2, /* min_div_recip_mul_df. */ | |
660 | 0, /* max_case_values. */ | |
661 | 0, /* cache_line_size. */ | |
662 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ | |
663 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ | |
664 | }; | |
665 | ||
5ec1ae3b EM |
666 | static const struct tune_params exynosm1_tunings = |
667 | { | |
668 | &exynosm1_extra_costs, | |
669 | &exynosm1_addrcost_table, | |
670 | &exynosm1_regmove_cost, | |
671 | &exynosm1_vector_cost, | |
672 | &generic_branch_cost, | |
9acc9cbe | 673 | &exynosm1_approx_modes, |
5ec1ae3b EM |
674 | 4, /* memmov_cost */ |
675 | 3, /* issue_rate */ | |
25cc2199 | 676 | (AARCH64_FUSE_AES_AESMC), /* fusible_ops */ |
5ec1ae3b EM |
677 | 4, /* function_align. */ |
678 | 4, /* jump_align. */ | |
679 | 4, /* loop_align. */ | |
680 | 2, /* int_reassoc_width. */ | |
681 | 4, /* fp_reassoc_width. */ | |
682 | 1, /* vec_reassoc_width. */ | |
683 | 2, /* min_div_recip_mul_sf. */ | |
684 | 2, /* min_div_recip_mul_df. */ | |
685 | 48, /* max_case_values. */ | |
686 | 64, /* cache_line_size. */ | |
220379df | 687 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
9acc9cbe | 688 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ |
5ec1ae3b EM |
689 | }; |
690 | ||
d1bcc29f AP |
691 | static const struct tune_params thunderx_tunings = |
692 | { | |
693 | &thunderx_extra_costs, | |
694 | &generic_addrcost_table, | |
695 | &thunderx_regmove_cost, | |
c3f20327 | 696 | &thunderx_vector_cost, |
b9066f5a | 697 | &generic_branch_cost, |
9acc9cbe | 698 | &generic_approx_modes, |
bd95e655 JG |
699 | 6, /* memmov_cost */ |
700 | 2, /* issue_rate */ | |
e9a3a175 | 701 | AARCH64_FUSE_CMP_BRANCH, /* fusible_ops */ |
0b82a5a2 WD |
702 | 8, /* function_align. */ |
703 | 8, /* jump_align. */ | |
704 | 8, /* loop_align. */ | |
cee66c68 WD |
705 | 2, /* int_reassoc_width. */ |
706 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
707 | 1, /* vec_reassoc_width. */ |
708 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 709 | 2, /* min_div_recip_mul_df. */ |
50487d79 EM |
710 | 0, /* max_case_values. */ |
711 | 0, /* cache_line_size. */ | |
2d6bc7fa | 712 | tune_params::AUTOPREFETCHER_OFF, /* autoprefetcher_model. */ |
54700e2e | 713 | (AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW) /* tune_flags. */ |
d1bcc29f AP |
714 | }; |
715 | ||
381e27aa PT |
716 | static const struct tune_params xgene1_tunings = |
717 | { | |
718 | &xgene1_extra_costs, | |
719 | &xgene1_addrcost_table, | |
720 | &xgene1_regmove_cost, | |
721 | &xgene1_vector_cost, | |
b9066f5a | 722 | &generic_branch_cost, |
9acc9cbe | 723 | &xgene1_approx_modes, |
bd95e655 JG |
724 | 6, /* memmov_cost */ |
725 | 4, /* issue_rate */ | |
e9a3a175 | 726 | AARCH64_FUSE_NOTHING, /* fusible_ops */ |
381e27aa PT |
727 | 16, /* function_align. */ |
728 | 8, /* jump_align. */ | |
729 | 16, /* loop_align. */ | |
730 | 2, /* int_reassoc_width. */ | |
731 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
732 | 1, /* vec_reassoc_width. */ |
733 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 734 | 2, /* min_div_recip_mul_df. */ |
50487d79 EM |
735 | 0, /* max_case_values. */ |
736 | 0, /* cache_line_size. */ | |
2d6bc7fa | 737 | tune_params::AUTOPREFETCHER_OFF, /* autoprefetcher_model. */ |
9acc9cbe | 738 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ |
381e27aa PT |
739 | }; |
740 | ||
ee446d9f JW |
741 | static const struct tune_params qdf24xx_tunings = |
742 | { | |
743 | &qdf24xx_extra_costs, | |
744 | &qdf24xx_addrcost_table, | |
745 | &qdf24xx_regmove_cost, | |
746 | &generic_vector_cost, | |
747 | &generic_branch_cost, | |
748 | &generic_approx_modes, | |
749 | 4, /* memmov_cost */ | |
750 | 4, /* issue_rate */ | |
751 | (AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD | |
752 | | AARCH64_FUSE_MOVK_MOVK), /* fuseable_ops */ | |
753 | 16, /* function_align. */ | |
754 | 8, /* jump_align. */ | |
755 | 16, /* loop_align. */ | |
756 | 2, /* int_reassoc_width. */ | |
757 | 4, /* fp_reassoc_width. */ | |
758 | 1, /* vec_reassoc_width. */ | |
759 | 2, /* min_div_recip_mul_sf. */ | |
760 | 2, /* min_div_recip_mul_df. */ | |
761 | 0, /* max_case_values. */ | |
762 | 64, /* cache_line_size. */ | |
763 | tune_params::AUTOPREFETCHER_STRONG, /* autoprefetcher_model. */ | |
764 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ | |
765 | }; | |
766 | ||
ad611a4c VP |
767 | static const struct tune_params vulcan_tunings = |
768 | { | |
769 | &vulcan_extra_costs, | |
770 | &vulcan_addrcost_table, | |
771 | &vulcan_regmove_cost, | |
772 | &vulcan_vector_cost, | |
773 | &vulcan_branch_cost, | |
774 | &generic_approx_modes, | |
775 | 4, /* memmov_cost. */ | |
776 | 4, /* issue_rate. */ | |
777 | AARCH64_FUSE_NOTHING, /* fuseable_ops. */ | |
778 | 16, /* function_align. */ | |
779 | 8, /* jump_align. */ | |
780 | 16, /* loop_align. */ | |
781 | 3, /* int_reassoc_width. */ | |
782 | 2, /* fp_reassoc_width. */ | |
783 | 2, /* vec_reassoc_width. */ | |
784 | 2, /* min_div_recip_mul_sf. */ | |
785 | 2, /* min_div_recip_mul_df. */ | |
786 | 0, /* max_case_values. */ | |
b91cd96b | 787 | 64, /* cache_line_size. */ |
ad611a4c VP |
788 | tune_params::AUTOPREFETCHER_OFF, /* autoprefetcher_model. */ |
789 | (AARCH64_EXTRA_TUNE_NONE) /* tune_flags. */ | |
790 | }; | |
791 | ||
8dec06f2 JG |
792 | /* Support for fine-grained override of the tuning structures. */ |
793 | struct aarch64_tuning_override_function | |
794 | { | |
795 | const char* name; | |
796 | void (*parse_override)(const char*, struct tune_params*); | |
797 | }; | |
798 | ||
799 | static void aarch64_parse_fuse_string (const char*, struct tune_params*); | |
800 | static void aarch64_parse_tune_string (const char*, struct tune_params*); | |
801 | ||
802 | static const struct aarch64_tuning_override_function | |
803 | aarch64_tuning_override_functions[] = | |
804 | { | |
805 | { "fuse", aarch64_parse_fuse_string }, | |
806 | { "tune", aarch64_parse_tune_string }, | |
807 | { NULL, NULL } | |
808 | }; | |
809 | ||
43e9d192 IB |
810 | /* A processor implementing AArch64. */ |
811 | struct processor | |
812 | { | |
813 | const char *const name; | |
46806c44 KT |
814 | enum aarch64_processor ident; |
815 | enum aarch64_processor sched_core; | |
393ae126 | 816 | enum aarch64_arch arch; |
0c6caaf8 | 817 | unsigned architecture_version; |
43e9d192 IB |
818 | const unsigned long flags; |
819 | const struct tune_params *const tune; | |
820 | }; | |
821 | ||
393ae126 KT |
822 | /* Architectures implementing AArch64. */ |
823 | static const struct processor all_architectures[] = | |
824 | { | |
825 | #define AARCH64_ARCH(NAME, CORE, ARCH_IDENT, ARCH_REV, FLAGS) \ | |
826 | {NAME, CORE, CORE, AARCH64_ARCH_##ARCH_IDENT, ARCH_REV, FLAGS, NULL}, | |
827 | #include "aarch64-arches.def" | |
393ae126 KT |
828 | {NULL, aarch64_none, aarch64_none, aarch64_no_arch, 0, 0, NULL} |
829 | }; | |
830 | ||
43e9d192 IB |
831 | /* Processor cores implementing AArch64. */ |
832 | static const struct processor all_cores[] = | |
833 | { | |
7e1bcce3 | 834 | #define AARCH64_CORE(NAME, IDENT, SCHED, ARCH, FLAGS, COSTS, IMP, PART) \ |
393ae126 KT |
835 | {NAME, IDENT, SCHED, AARCH64_ARCH_##ARCH, \ |
836 | all_architectures[AARCH64_ARCH_##ARCH].architecture_version, \ | |
837 | FLAGS, &COSTS##_tunings}, | |
43e9d192 | 838 | #include "aarch64-cores.def" |
393ae126 KT |
839 | {"generic", generic, cortexa53, AARCH64_ARCH_8A, 8, |
840 | AARCH64_FL_FOR_ARCH8, &generic_tunings}, | |
841 | {NULL, aarch64_none, aarch64_none, aarch64_no_arch, 0, 0, NULL} | |
43e9d192 IB |
842 | }; |
843 | ||
43e9d192 | 844 | |
361fb3ee KT |
845 | /* Target specification. These are populated by the -march, -mtune, -mcpu |
846 | handling code or by target attributes. */ | |
43e9d192 IB |
847 | static const struct processor *selected_arch; |
848 | static const struct processor *selected_cpu; | |
849 | static const struct processor *selected_tune; | |
850 | ||
b175b679 JG |
851 | /* The current tuning set. */ |
852 | struct tune_params aarch64_tune_params = generic_tunings; | |
853 | ||
43e9d192 IB |
854 | #define AARCH64_CPU_DEFAULT_FLAGS ((selected_cpu) ? selected_cpu->flags : 0) |
855 | ||
856 | /* An ISA extension in the co-processor and main instruction set space. */ | |
857 | struct aarch64_option_extension | |
858 | { | |
859 | const char *const name; | |
860 | const unsigned long flags_on; | |
861 | const unsigned long flags_off; | |
862 | }; | |
863 | ||
43e9d192 IB |
864 | typedef enum aarch64_cond_code |
865 | { | |
866 | AARCH64_EQ = 0, AARCH64_NE, AARCH64_CS, AARCH64_CC, AARCH64_MI, AARCH64_PL, | |
867 | AARCH64_VS, AARCH64_VC, AARCH64_HI, AARCH64_LS, AARCH64_GE, AARCH64_LT, | |
868 | AARCH64_GT, AARCH64_LE, AARCH64_AL, AARCH64_NV | |
869 | } | |
870 | aarch64_cc; | |
871 | ||
872 | #define AARCH64_INVERSE_CONDITION_CODE(X) ((aarch64_cc) (((int) X) ^ 1)) | |
873 | ||
874 | /* The condition codes of the processor, and the inverse function. */ | |
875 | static const char * const aarch64_condition_codes[] = | |
876 | { | |
877 | "eq", "ne", "cs", "cc", "mi", "pl", "vs", "vc", | |
878 | "hi", "ls", "ge", "lt", "gt", "le", "al", "nv" | |
879 | }; | |
880 | ||
973d2e01 TP |
881 | /* Generate code to enable conditional branches in functions over 1 MiB. */ |
882 | const char * | |
883 | aarch64_gen_far_branch (rtx * operands, int pos_label, const char * dest, | |
884 | const char * branch_format) | |
885 | { | |
886 | rtx_code_label * tmp_label = gen_label_rtx (); | |
887 | char label_buf[256]; | |
888 | char buffer[128]; | |
889 | ASM_GENERATE_INTERNAL_LABEL (label_buf, dest, | |
890 | CODE_LABEL_NUMBER (tmp_label)); | |
891 | const char *label_ptr = targetm.strip_name_encoding (label_buf); | |
892 | rtx dest_label = operands[pos_label]; | |
893 | operands[pos_label] = tmp_label; | |
894 | ||
895 | snprintf (buffer, sizeof (buffer), "%s%s", branch_format, label_ptr); | |
896 | output_asm_insn (buffer, operands); | |
897 | ||
898 | snprintf (buffer, sizeof (buffer), "b\t%%l%d\n%s:", pos_label, label_ptr); | |
899 | operands[pos_label] = dest_label; | |
900 | output_asm_insn (buffer, operands); | |
901 | return ""; | |
902 | } | |
903 | ||
261fb553 AL |
904 | void |
905 | aarch64_err_no_fpadvsimd (machine_mode mode, const char *msg) | |
906 | { | |
907 | const char *mc = FLOAT_MODE_P (mode) ? "floating-point" : "vector"; | |
908 | if (TARGET_GENERAL_REGS_ONLY) | |
909 | error ("%qs is incompatible with %s %s", "-mgeneral-regs-only", mc, msg); | |
910 | else | |
911 | error ("%qs feature modifier is incompatible with %s %s", "+nofp", mc, msg); | |
912 | } | |
913 | ||
c64f7d37 WD |
914 | /* Implement TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS. |
915 | The register allocator chooses ALL_REGS if FP_REGS and GENERAL_REGS have | |
31e2b5a3 WD |
916 | the same cost even if ALL_REGS has a much larger cost. ALL_REGS is also |
917 | used if the cost of both FP_REGS and GENERAL_REGS is lower than the memory | |
918 | cost (in this case the best class is the lowest cost one). Using ALL_REGS | |
919 | irrespectively of its cost results in bad allocations with many redundant | |
920 | int<->FP moves which are expensive on various cores. | |
921 | To avoid this we don't allow ALL_REGS as the allocno class, but force a | |
922 | decision between FP_REGS and GENERAL_REGS. We use the allocno class if it | |
923 | isn't ALL_REGS. Similarly, use the best class if it isn't ALL_REGS. | |
924 | Otherwise set the allocno class depending on the mode. | |
925 | The result of this is that it is no longer inefficient to have a higher | |
926 | memory move cost than the register move cost. | |
927 | */ | |
c64f7d37 WD |
928 | |
929 | static reg_class_t | |
31e2b5a3 WD |
930 | aarch64_ira_change_pseudo_allocno_class (int regno, reg_class_t allocno_class, |
931 | reg_class_t best_class) | |
c64f7d37 WD |
932 | { |
933 | enum machine_mode mode; | |
934 | ||
935 | if (allocno_class != ALL_REGS) | |
936 | return allocno_class; | |
937 | ||
31e2b5a3 WD |
938 | if (best_class != ALL_REGS) |
939 | return best_class; | |
940 | ||
c64f7d37 WD |
941 | mode = PSEUDO_REGNO_MODE (regno); |
942 | return FLOAT_MODE_P (mode) || VECTOR_MODE_P (mode) ? FP_REGS : GENERAL_REGS; | |
943 | } | |
944 | ||
26e0ff94 | 945 | static unsigned int |
50093a33 | 946 | aarch64_min_divisions_for_recip_mul (enum machine_mode mode) |
26e0ff94 | 947 | { |
50093a33 | 948 | if (GET_MODE_UNIT_SIZE (mode) == 4) |
b175b679 JG |
949 | return aarch64_tune_params.min_div_recip_mul_sf; |
950 | return aarch64_tune_params.min_div_recip_mul_df; | |
26e0ff94 WD |
951 | } |
952 | ||
cee66c68 WD |
953 | static int |
954 | aarch64_reassociation_width (unsigned opc ATTRIBUTE_UNUSED, | |
955 | enum machine_mode mode) | |
956 | { | |
957 | if (VECTOR_MODE_P (mode)) | |
b175b679 | 958 | return aarch64_tune_params.vec_reassoc_width; |
cee66c68 | 959 | if (INTEGRAL_MODE_P (mode)) |
b175b679 | 960 | return aarch64_tune_params.int_reassoc_width; |
cee66c68 | 961 | if (FLOAT_MODE_P (mode)) |
b175b679 | 962 | return aarch64_tune_params.fp_reassoc_width; |
cee66c68 WD |
963 | return 1; |
964 | } | |
965 | ||
43e9d192 IB |
966 | /* Provide a mapping from gcc register numbers to dwarf register numbers. */ |
967 | unsigned | |
968 | aarch64_dbx_register_number (unsigned regno) | |
969 | { | |
970 | if (GP_REGNUM_P (regno)) | |
971 | return AARCH64_DWARF_R0 + regno - R0_REGNUM; | |
972 | else if (regno == SP_REGNUM) | |
973 | return AARCH64_DWARF_SP; | |
974 | else if (FP_REGNUM_P (regno)) | |
975 | return AARCH64_DWARF_V0 + regno - V0_REGNUM; | |
976 | ||
977 | /* Return values >= DWARF_FRAME_REGISTERS indicate that there is no | |
978 | equivalent DWARF register. */ | |
979 | return DWARF_FRAME_REGISTERS; | |
980 | } | |
981 | ||
982 | /* Return TRUE if MODE is any of the large INT modes. */ | |
983 | static bool | |
ef4bddc2 | 984 | aarch64_vect_struct_mode_p (machine_mode mode) |
43e9d192 IB |
985 | { |
986 | return mode == OImode || mode == CImode || mode == XImode; | |
987 | } | |
988 | ||
989 | /* Return TRUE if MODE is any of the vector modes. */ | |
990 | static bool | |
ef4bddc2 | 991 | aarch64_vector_mode_p (machine_mode mode) |
43e9d192 IB |
992 | { |
993 | return aarch64_vector_mode_supported_p (mode) | |
994 | || aarch64_vect_struct_mode_p (mode); | |
995 | } | |
996 | ||
997 | /* Implement target hook TARGET_ARRAY_MODE_SUPPORTED_P. */ | |
998 | static bool | |
ef4bddc2 | 999 | aarch64_array_mode_supported_p (machine_mode mode, |
43e9d192 IB |
1000 | unsigned HOST_WIDE_INT nelems) |
1001 | { | |
1002 | if (TARGET_SIMD | |
635e66fe AL |
1003 | && (AARCH64_VALID_SIMD_QREG_MODE (mode) |
1004 | || AARCH64_VALID_SIMD_DREG_MODE (mode)) | |
43e9d192 IB |
1005 | && (nelems >= 2 && nelems <= 4)) |
1006 | return true; | |
1007 | ||
1008 | return false; | |
1009 | } | |
1010 | ||
1011 | /* Implement HARD_REGNO_NREGS. */ | |
1012 | ||
1013 | int | |
ef4bddc2 | 1014 | aarch64_hard_regno_nregs (unsigned regno, machine_mode mode) |
43e9d192 IB |
1015 | { |
1016 | switch (aarch64_regno_regclass (regno)) | |
1017 | { | |
1018 | case FP_REGS: | |
1019 | case FP_LO_REGS: | |
1020 | return (GET_MODE_SIZE (mode) + UNITS_PER_VREG - 1) / UNITS_PER_VREG; | |
1021 | default: | |
1022 | return (GET_MODE_SIZE (mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
1023 | } | |
1024 | gcc_unreachable (); | |
1025 | } | |
1026 | ||
1027 | /* Implement HARD_REGNO_MODE_OK. */ | |
1028 | ||
1029 | int | |
ef4bddc2 | 1030 | aarch64_hard_regno_mode_ok (unsigned regno, machine_mode mode) |
43e9d192 IB |
1031 | { |
1032 | if (GET_MODE_CLASS (mode) == MODE_CC) | |
1033 | return regno == CC_REGNUM; | |
1034 | ||
9259db42 YZ |
1035 | if (regno == SP_REGNUM) |
1036 | /* The purpose of comparing with ptr_mode is to support the | |
1037 | global register variable associated with the stack pointer | |
1038 | register via the syntax of asm ("wsp") in ILP32. */ | |
1039 | return mode == Pmode || mode == ptr_mode; | |
1040 | ||
1041 | if (regno == FRAME_POINTER_REGNUM || regno == ARG_POINTER_REGNUM) | |
43e9d192 IB |
1042 | return mode == Pmode; |
1043 | ||
1044 | if (GP_REGNUM_P (regno) && ! aarch64_vect_struct_mode_p (mode)) | |
1045 | return 1; | |
1046 | ||
1047 | if (FP_REGNUM_P (regno)) | |
1048 | { | |
1049 | if (aarch64_vect_struct_mode_p (mode)) | |
1050 | return | |
1051 | (regno + aarch64_hard_regno_nregs (regno, mode) - 1) <= V31_REGNUM; | |
1052 | else | |
1053 | return 1; | |
1054 | } | |
1055 | ||
1056 | return 0; | |
1057 | } | |
1058 | ||
73d9ac6a | 1059 | /* Implement HARD_REGNO_CALLER_SAVE_MODE. */ |
ef4bddc2 | 1060 | machine_mode |
73d9ac6a | 1061 | aarch64_hard_regno_caller_save_mode (unsigned regno, unsigned nregs, |
ef4bddc2 | 1062 | machine_mode mode) |
73d9ac6a IB |
1063 | { |
1064 | /* Handle modes that fit within single registers. */ | |
1065 | if (nregs == 1 && GET_MODE_SIZE (mode) <= 16) | |
1066 | { | |
1067 | if (GET_MODE_SIZE (mode) >= 4) | |
1068 | return mode; | |
1069 | else | |
1070 | return SImode; | |
1071 | } | |
1072 | /* Fall back to generic for multi-reg and very large modes. */ | |
1073 | else | |
1074 | return choose_hard_reg_mode (regno, nregs, false); | |
1075 | } | |
1076 | ||
43e9d192 IB |
1077 | /* Return true if calls to DECL should be treated as |
1078 | long-calls (ie called via a register). */ | |
1079 | static bool | |
1080 | aarch64_decl_is_long_call_p (const_tree decl ATTRIBUTE_UNUSED) | |
1081 | { | |
1082 | return false; | |
1083 | } | |
1084 | ||
1085 | /* Return true if calls to symbol-ref SYM should be treated as | |
1086 | long-calls (ie called via a register). */ | |
1087 | bool | |
1088 | aarch64_is_long_call_p (rtx sym) | |
1089 | { | |
1090 | return aarch64_decl_is_long_call_p (SYMBOL_REF_DECL (sym)); | |
1091 | } | |
1092 | ||
b60d63cb JW |
1093 | /* Return true if calls to symbol-ref SYM should not go through |
1094 | plt stubs. */ | |
1095 | ||
1096 | bool | |
1097 | aarch64_is_noplt_call_p (rtx sym) | |
1098 | { | |
1099 | const_tree decl = SYMBOL_REF_DECL (sym); | |
1100 | ||
1101 | if (flag_pic | |
1102 | && decl | |
1103 | && (!flag_plt | |
1104 | || lookup_attribute ("noplt", DECL_ATTRIBUTES (decl))) | |
1105 | && !targetm.binds_local_p (decl)) | |
1106 | return true; | |
1107 | ||
1108 | return false; | |
1109 | } | |
1110 | ||
43e9d192 IB |
1111 | /* Return true if the offsets to a zero/sign-extract operation |
1112 | represent an expression that matches an extend operation. The | |
1113 | operands represent the paramters from | |
1114 | ||
4745e701 | 1115 | (extract:MODE (mult (reg) (MULT_IMM)) (EXTRACT_IMM) (const_int 0)). */ |
43e9d192 | 1116 | bool |
ef4bddc2 | 1117 | aarch64_is_extend_from_extract (machine_mode mode, rtx mult_imm, |
43e9d192 IB |
1118 | rtx extract_imm) |
1119 | { | |
1120 | HOST_WIDE_INT mult_val, extract_val; | |
1121 | ||
1122 | if (! CONST_INT_P (mult_imm) || ! CONST_INT_P (extract_imm)) | |
1123 | return false; | |
1124 | ||
1125 | mult_val = INTVAL (mult_imm); | |
1126 | extract_val = INTVAL (extract_imm); | |
1127 | ||
1128 | if (extract_val > 8 | |
1129 | && extract_val < GET_MODE_BITSIZE (mode) | |
1130 | && exact_log2 (extract_val & ~7) > 0 | |
1131 | && (extract_val & 7) <= 4 | |
1132 | && mult_val == (1 << (extract_val & 7))) | |
1133 | return true; | |
1134 | ||
1135 | return false; | |
1136 | } | |
1137 | ||
1138 | /* Emit an insn that's a simple single-set. Both the operands must be | |
1139 | known to be valid. */ | |
1140 | inline static rtx | |
1141 | emit_set_insn (rtx x, rtx y) | |
1142 | { | |
f7df4a84 | 1143 | return emit_insn (gen_rtx_SET (x, y)); |
43e9d192 IB |
1144 | } |
1145 | ||
1146 | /* X and Y are two things to compare using CODE. Emit the compare insn and | |
1147 | return the rtx for register 0 in the proper mode. */ | |
1148 | rtx | |
1149 | aarch64_gen_compare_reg (RTX_CODE code, rtx x, rtx y) | |
1150 | { | |
ef4bddc2 | 1151 | machine_mode mode = SELECT_CC_MODE (code, x, y); |
43e9d192 IB |
1152 | rtx cc_reg = gen_rtx_REG (mode, CC_REGNUM); |
1153 | ||
1154 | emit_set_insn (cc_reg, gen_rtx_COMPARE (mode, x, y)); | |
1155 | return cc_reg; | |
1156 | } | |
1157 | ||
1158 | /* Build the SYMBOL_REF for __tls_get_addr. */ | |
1159 | ||
1160 | static GTY(()) rtx tls_get_addr_libfunc; | |
1161 | ||
1162 | rtx | |
1163 | aarch64_tls_get_addr (void) | |
1164 | { | |
1165 | if (!tls_get_addr_libfunc) | |
1166 | tls_get_addr_libfunc = init_one_libfunc ("__tls_get_addr"); | |
1167 | return tls_get_addr_libfunc; | |
1168 | } | |
1169 | ||
1170 | /* Return the TLS model to use for ADDR. */ | |
1171 | ||
1172 | static enum tls_model | |
1173 | tls_symbolic_operand_type (rtx addr) | |
1174 | { | |
1175 | enum tls_model tls_kind = TLS_MODEL_NONE; | |
1176 | rtx sym, addend; | |
1177 | ||
1178 | if (GET_CODE (addr) == CONST) | |
1179 | { | |
1180 | split_const (addr, &sym, &addend); | |
1181 | if (GET_CODE (sym) == SYMBOL_REF) | |
1182 | tls_kind = SYMBOL_REF_TLS_MODEL (sym); | |
1183 | } | |
1184 | else if (GET_CODE (addr) == SYMBOL_REF) | |
1185 | tls_kind = SYMBOL_REF_TLS_MODEL (addr); | |
1186 | ||
1187 | return tls_kind; | |
1188 | } | |
1189 | ||
1190 | /* We'll allow lo_sum's in addresses in our legitimate addresses | |
1191 | so that combine would take care of combining addresses where | |
1192 | necessary, but for generation purposes, we'll generate the address | |
1193 | as : | |
1194 | RTL Absolute | |
1195 | tmp = hi (symbol_ref); adrp x1, foo | |
1196 | dest = lo_sum (tmp, symbol_ref); add dest, x1, :lo_12:foo | |
1197 | nop | |
1198 | ||
1199 | PIC TLS | |
1200 | adrp x1, :got:foo adrp tmp, :tlsgd:foo | |
1201 | ldr x1, [:got_lo12:foo] add dest, tmp, :tlsgd_lo12:foo | |
1202 | bl __tls_get_addr | |
1203 | nop | |
1204 | ||
1205 | Load TLS symbol, depending on TLS mechanism and TLS access model. | |
1206 | ||
1207 | Global Dynamic - Traditional TLS: | |
1208 | adrp tmp, :tlsgd:imm | |
1209 | add dest, tmp, #:tlsgd_lo12:imm | |
1210 | bl __tls_get_addr | |
1211 | ||
1212 | Global Dynamic - TLS Descriptors: | |
1213 | adrp dest, :tlsdesc:imm | |
1214 | ldr tmp, [dest, #:tlsdesc_lo12:imm] | |
1215 | add dest, dest, #:tlsdesc_lo12:imm | |
1216 | blr tmp | |
1217 | mrs tp, tpidr_el0 | |
1218 | add dest, dest, tp | |
1219 | ||
1220 | Initial Exec: | |
1221 | mrs tp, tpidr_el0 | |
1222 | adrp tmp, :gottprel:imm | |
1223 | ldr dest, [tmp, #:gottprel_lo12:imm] | |
1224 | add dest, dest, tp | |
1225 | ||
1226 | Local Exec: | |
1227 | mrs tp, tpidr_el0 | |
0699caae RL |
1228 | add t0, tp, #:tprel_hi12:imm, lsl #12 |
1229 | add t0, t0, #:tprel_lo12_nc:imm | |
43e9d192 IB |
1230 | */ |
1231 | ||
1232 | static void | |
1233 | aarch64_load_symref_appropriately (rtx dest, rtx imm, | |
1234 | enum aarch64_symbol_type type) | |
1235 | { | |
1236 | switch (type) | |
1237 | { | |
1238 | case SYMBOL_SMALL_ABSOLUTE: | |
1239 | { | |
28514dda | 1240 | /* In ILP32, the mode of dest can be either SImode or DImode. */ |
43e9d192 | 1241 | rtx tmp_reg = dest; |
ef4bddc2 | 1242 | machine_mode mode = GET_MODE (dest); |
28514dda YZ |
1243 | |
1244 | gcc_assert (mode == Pmode || mode == ptr_mode); | |
1245 | ||
43e9d192 | 1246 | if (can_create_pseudo_p ()) |
28514dda | 1247 | tmp_reg = gen_reg_rtx (mode); |
43e9d192 | 1248 | |
28514dda | 1249 | emit_move_insn (tmp_reg, gen_rtx_HIGH (mode, imm)); |
43e9d192 IB |
1250 | emit_insn (gen_add_losym (dest, tmp_reg, imm)); |
1251 | return; | |
1252 | } | |
1253 | ||
a5350ddc | 1254 | case SYMBOL_TINY_ABSOLUTE: |
f7df4a84 | 1255 | emit_insn (gen_rtx_SET (dest, imm)); |
a5350ddc CSS |
1256 | return; |
1257 | ||
1b1e81f8 JW |
1258 | case SYMBOL_SMALL_GOT_28K: |
1259 | { | |
1260 | machine_mode mode = GET_MODE (dest); | |
1261 | rtx gp_rtx = pic_offset_table_rtx; | |
53021678 JW |
1262 | rtx insn; |
1263 | rtx mem; | |
1b1e81f8 JW |
1264 | |
1265 | /* NOTE: pic_offset_table_rtx can be NULL_RTX, because we can reach | |
1266 | here before rtl expand. Tree IVOPT will generate rtl pattern to | |
1267 | decide rtx costs, in which case pic_offset_table_rtx is not | |
1268 | initialized. For that case no need to generate the first adrp | |
026c3cfd | 1269 | instruction as the final cost for global variable access is |
1b1e81f8 JW |
1270 | one instruction. */ |
1271 | if (gp_rtx != NULL) | |
1272 | { | |
1273 | /* -fpic for -mcmodel=small allow 32K GOT table size (but we are | |
1274 | using the page base as GOT base, the first page may be wasted, | |
1275 | in the worst scenario, there is only 28K space for GOT). | |
1276 | ||
1277 | The generate instruction sequence for accessing global variable | |
1278 | is: | |
1279 | ||
a3957742 | 1280 | ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym] |
1b1e81f8 JW |
1281 | |
1282 | Only one instruction needed. But we must initialize | |
1283 | pic_offset_table_rtx properly. We generate initialize insn for | |
1284 | every global access, and allow CSE to remove all redundant. | |
1285 | ||
1286 | The final instruction sequences will look like the following | |
1287 | for multiply global variables access. | |
1288 | ||
a3957742 | 1289 | adrp pic_offset_table_rtx, _GLOBAL_OFFSET_TABLE_ |
1b1e81f8 | 1290 | |
a3957742 JW |
1291 | ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym1] |
1292 | ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym2] | |
1293 | ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym3] | |
1294 | ... */ | |
1b1e81f8 JW |
1295 | |
1296 | rtx s = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_"); | |
1297 | crtl->uses_pic_offset_table = 1; | |
1298 | emit_move_insn (gp_rtx, gen_rtx_HIGH (Pmode, s)); | |
1299 | ||
1300 | if (mode != GET_MODE (gp_rtx)) | |
1301 | gp_rtx = simplify_gen_subreg (mode, gp_rtx, GET_MODE (gp_rtx), 0); | |
1302 | } | |
1303 | ||
1304 | if (mode == ptr_mode) | |
1305 | { | |
1306 | if (mode == DImode) | |
53021678 | 1307 | insn = gen_ldr_got_small_28k_di (dest, gp_rtx, imm); |
1b1e81f8 | 1308 | else |
53021678 JW |
1309 | insn = gen_ldr_got_small_28k_si (dest, gp_rtx, imm); |
1310 | ||
1311 | mem = XVECEXP (SET_SRC (insn), 0, 0); | |
1b1e81f8 JW |
1312 | } |
1313 | else | |
1314 | { | |
1315 | gcc_assert (mode == Pmode); | |
53021678 JW |
1316 | |
1317 | insn = gen_ldr_got_small_28k_sidi (dest, gp_rtx, imm); | |
1318 | mem = XVECEXP (XEXP (SET_SRC (insn), 0), 0, 0); | |
1b1e81f8 JW |
1319 | } |
1320 | ||
53021678 JW |
1321 | /* The operand is expected to be MEM. Whenever the related insn |
1322 | pattern changed, above code which calculate mem should be | |
1323 | updated. */ | |
1324 | gcc_assert (GET_CODE (mem) == MEM); | |
1325 | MEM_READONLY_P (mem) = 1; | |
1326 | MEM_NOTRAP_P (mem) = 1; | |
1327 | emit_insn (insn); | |
1b1e81f8 JW |
1328 | return; |
1329 | } | |
1330 | ||
6642bdb4 | 1331 | case SYMBOL_SMALL_GOT_4G: |
43e9d192 | 1332 | { |
28514dda YZ |
1333 | /* In ILP32, the mode of dest can be either SImode or DImode, |
1334 | while the got entry is always of SImode size. The mode of | |
1335 | dest depends on how dest is used: if dest is assigned to a | |
1336 | pointer (e.g. in the memory), it has SImode; it may have | |
1337 | DImode if dest is dereferenced to access the memeory. | |
1338 | This is why we have to handle three different ldr_got_small | |
1339 | patterns here (two patterns for ILP32). */ | |
53021678 JW |
1340 | |
1341 | rtx insn; | |
1342 | rtx mem; | |
43e9d192 | 1343 | rtx tmp_reg = dest; |
ef4bddc2 | 1344 | machine_mode mode = GET_MODE (dest); |
28514dda | 1345 | |
43e9d192 | 1346 | if (can_create_pseudo_p ()) |
28514dda YZ |
1347 | tmp_reg = gen_reg_rtx (mode); |
1348 | ||
1349 | emit_move_insn (tmp_reg, gen_rtx_HIGH (mode, imm)); | |
1350 | if (mode == ptr_mode) | |
1351 | { | |
1352 | if (mode == DImode) | |
53021678 | 1353 | insn = gen_ldr_got_small_di (dest, tmp_reg, imm); |
28514dda | 1354 | else |
53021678 JW |
1355 | insn = gen_ldr_got_small_si (dest, tmp_reg, imm); |
1356 | ||
1357 | mem = XVECEXP (SET_SRC (insn), 0, 0); | |
28514dda YZ |
1358 | } |
1359 | else | |
1360 | { | |
1361 | gcc_assert (mode == Pmode); | |
53021678 JW |
1362 | |
1363 | insn = gen_ldr_got_small_sidi (dest, tmp_reg, imm); | |
1364 | mem = XVECEXP (XEXP (SET_SRC (insn), 0), 0, 0); | |
28514dda YZ |
1365 | } |
1366 | ||
53021678 JW |
1367 | gcc_assert (GET_CODE (mem) == MEM); |
1368 | MEM_READONLY_P (mem) = 1; | |
1369 | MEM_NOTRAP_P (mem) = 1; | |
1370 | emit_insn (insn); | |
43e9d192 IB |
1371 | return; |
1372 | } | |
1373 | ||
1374 | case SYMBOL_SMALL_TLSGD: | |
1375 | { | |
5d8a22a5 | 1376 | rtx_insn *insns; |
43e9d192 IB |
1377 | rtx result = gen_rtx_REG (Pmode, R0_REGNUM); |
1378 | ||
1379 | start_sequence (); | |
78607708 | 1380 | aarch64_emit_call_insn (gen_tlsgd_small (result, imm)); |
43e9d192 IB |
1381 | insns = get_insns (); |
1382 | end_sequence (); | |
1383 | ||
1384 | RTL_CONST_CALL_P (insns) = 1; | |
1385 | emit_libcall_block (insns, dest, result, imm); | |
1386 | return; | |
1387 | } | |
1388 | ||
1389 | case SYMBOL_SMALL_TLSDESC: | |
1390 | { | |
ef4bddc2 | 1391 | machine_mode mode = GET_MODE (dest); |
621ad2de | 1392 | rtx x0 = gen_rtx_REG (mode, R0_REGNUM); |
43e9d192 IB |
1393 | rtx tp; |
1394 | ||
621ad2de AP |
1395 | gcc_assert (mode == Pmode || mode == ptr_mode); |
1396 | ||
2876a13f JW |
1397 | /* In ILP32, the got entry is always of SImode size. Unlike |
1398 | small GOT, the dest is fixed at reg 0. */ | |
1399 | if (TARGET_ILP32) | |
1400 | emit_insn (gen_tlsdesc_small_si (imm)); | |
621ad2de | 1401 | else |
2876a13f | 1402 | emit_insn (gen_tlsdesc_small_di (imm)); |
43e9d192 | 1403 | tp = aarch64_load_tp (NULL); |
621ad2de AP |
1404 | |
1405 | if (mode != Pmode) | |
1406 | tp = gen_lowpart (mode, tp); | |
1407 | ||
2876a13f | 1408 | emit_insn (gen_rtx_SET (dest, gen_rtx_PLUS (mode, tp, x0))); |
43e9d192 IB |
1409 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); |
1410 | return; | |
1411 | } | |
1412 | ||
79496620 | 1413 | case SYMBOL_SMALL_TLSIE: |
43e9d192 | 1414 | { |
621ad2de AP |
1415 | /* In ILP32, the mode of dest can be either SImode or DImode, |
1416 | while the got entry is always of SImode size. The mode of | |
1417 | dest depends on how dest is used: if dest is assigned to a | |
1418 | pointer (e.g. in the memory), it has SImode; it may have | |
1419 | DImode if dest is dereferenced to access the memeory. | |
1420 | This is why we have to handle three different tlsie_small | |
1421 | patterns here (two patterns for ILP32). */ | |
ef4bddc2 | 1422 | machine_mode mode = GET_MODE (dest); |
621ad2de | 1423 | rtx tmp_reg = gen_reg_rtx (mode); |
43e9d192 | 1424 | rtx tp = aarch64_load_tp (NULL); |
621ad2de AP |
1425 | |
1426 | if (mode == ptr_mode) | |
1427 | { | |
1428 | if (mode == DImode) | |
1429 | emit_insn (gen_tlsie_small_di (tmp_reg, imm)); | |
1430 | else | |
1431 | { | |
1432 | emit_insn (gen_tlsie_small_si (tmp_reg, imm)); | |
1433 | tp = gen_lowpart (mode, tp); | |
1434 | } | |
1435 | } | |
1436 | else | |
1437 | { | |
1438 | gcc_assert (mode == Pmode); | |
1439 | emit_insn (gen_tlsie_small_sidi (tmp_reg, imm)); | |
1440 | } | |
1441 | ||
f7df4a84 | 1442 | emit_insn (gen_rtx_SET (dest, gen_rtx_PLUS (mode, tp, tmp_reg))); |
43e9d192 IB |
1443 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); |
1444 | return; | |
1445 | } | |
1446 | ||
cbf5629e | 1447 | case SYMBOL_TLSLE12: |
d18ba284 | 1448 | case SYMBOL_TLSLE24: |
cbf5629e JW |
1449 | case SYMBOL_TLSLE32: |
1450 | case SYMBOL_TLSLE48: | |
43e9d192 | 1451 | { |
cbf5629e | 1452 | machine_mode mode = GET_MODE (dest); |
43e9d192 | 1453 | rtx tp = aarch64_load_tp (NULL); |
e6f7f0e9 | 1454 | |
cbf5629e JW |
1455 | if (mode != Pmode) |
1456 | tp = gen_lowpart (mode, tp); | |
1457 | ||
1458 | switch (type) | |
1459 | { | |
1460 | case SYMBOL_TLSLE12: | |
1461 | emit_insn ((mode == DImode ? gen_tlsle12_di : gen_tlsle12_si) | |
1462 | (dest, tp, imm)); | |
1463 | break; | |
1464 | case SYMBOL_TLSLE24: | |
1465 | emit_insn ((mode == DImode ? gen_tlsle24_di : gen_tlsle24_si) | |
1466 | (dest, tp, imm)); | |
1467 | break; | |
1468 | case SYMBOL_TLSLE32: | |
1469 | emit_insn ((mode == DImode ? gen_tlsle32_di : gen_tlsle32_si) | |
1470 | (dest, imm)); | |
1471 | emit_insn ((mode == DImode ? gen_adddi3 : gen_addsi3) | |
1472 | (dest, dest, tp)); | |
1473 | break; | |
1474 | case SYMBOL_TLSLE48: | |
1475 | emit_insn ((mode == DImode ? gen_tlsle48_di : gen_tlsle48_si) | |
1476 | (dest, imm)); | |
1477 | emit_insn ((mode == DImode ? gen_adddi3 : gen_addsi3) | |
1478 | (dest, dest, tp)); | |
1479 | break; | |
1480 | default: | |
1481 | gcc_unreachable (); | |
1482 | } | |
e6f7f0e9 | 1483 | |
43e9d192 IB |
1484 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); |
1485 | return; | |
1486 | } | |
1487 | ||
87dd8ab0 MS |
1488 | case SYMBOL_TINY_GOT: |
1489 | emit_insn (gen_ldr_got_tiny (dest, imm)); | |
1490 | return; | |
1491 | ||
5ae7caad JW |
1492 | case SYMBOL_TINY_TLSIE: |
1493 | { | |
1494 | machine_mode mode = GET_MODE (dest); | |
1495 | rtx tp = aarch64_load_tp (NULL); | |
1496 | ||
1497 | if (mode == ptr_mode) | |
1498 | { | |
1499 | if (mode == DImode) | |
1500 | emit_insn (gen_tlsie_tiny_di (dest, imm, tp)); | |
1501 | else | |
1502 | { | |
1503 | tp = gen_lowpart (mode, tp); | |
1504 | emit_insn (gen_tlsie_tiny_si (dest, imm, tp)); | |
1505 | } | |
1506 | } | |
1507 | else | |
1508 | { | |
1509 | gcc_assert (mode == Pmode); | |
1510 | emit_insn (gen_tlsie_tiny_sidi (dest, imm, tp)); | |
1511 | } | |
1512 | ||
1513 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); | |
1514 | return; | |
1515 | } | |
1516 | ||
43e9d192 IB |
1517 | default: |
1518 | gcc_unreachable (); | |
1519 | } | |
1520 | } | |
1521 | ||
1522 | /* Emit a move from SRC to DEST. Assume that the move expanders can | |
1523 | handle all moves if !can_create_pseudo_p (). The distinction is | |
1524 | important because, unlike emit_move_insn, the move expanders know | |
1525 | how to force Pmode objects into the constant pool even when the | |
1526 | constant pool address is not itself legitimate. */ | |
1527 | static rtx | |
1528 | aarch64_emit_move (rtx dest, rtx src) | |
1529 | { | |
1530 | return (can_create_pseudo_p () | |
1531 | ? emit_move_insn (dest, src) | |
1532 | : emit_move_insn_1 (dest, src)); | |
1533 | } | |
1534 | ||
030d03b8 RE |
1535 | /* Split a 128-bit move operation into two 64-bit move operations, |
1536 | taking care to handle partial overlap of register to register | |
1537 | copies. Special cases are needed when moving between GP regs and | |
1538 | FP regs. SRC can be a register, constant or memory; DST a register | |
1539 | or memory. If either operand is memory it must not have any side | |
1540 | effects. */ | |
43e9d192 IB |
1541 | void |
1542 | aarch64_split_128bit_move (rtx dst, rtx src) | |
1543 | { | |
030d03b8 RE |
1544 | rtx dst_lo, dst_hi; |
1545 | rtx src_lo, src_hi; | |
43e9d192 | 1546 | |
ef4bddc2 | 1547 | machine_mode mode = GET_MODE (dst); |
12dc6974 | 1548 | |
030d03b8 RE |
1549 | gcc_assert (mode == TImode || mode == TFmode); |
1550 | gcc_assert (!(side_effects_p (src) || side_effects_p (dst))); | |
1551 | gcc_assert (mode == GET_MODE (src) || GET_MODE (src) == VOIDmode); | |
43e9d192 IB |
1552 | |
1553 | if (REG_P (dst) && REG_P (src)) | |
1554 | { | |
030d03b8 RE |
1555 | int src_regno = REGNO (src); |
1556 | int dst_regno = REGNO (dst); | |
43e9d192 | 1557 | |
030d03b8 | 1558 | /* Handle FP <-> GP regs. */ |
43e9d192 IB |
1559 | if (FP_REGNUM_P (dst_regno) && GP_REGNUM_P (src_regno)) |
1560 | { | |
030d03b8 RE |
1561 | src_lo = gen_lowpart (word_mode, src); |
1562 | src_hi = gen_highpart (word_mode, src); | |
1563 | ||
1564 | if (mode == TImode) | |
1565 | { | |
1566 | emit_insn (gen_aarch64_movtilow_di (dst, src_lo)); | |
1567 | emit_insn (gen_aarch64_movtihigh_di (dst, src_hi)); | |
1568 | } | |
1569 | else | |
1570 | { | |
1571 | emit_insn (gen_aarch64_movtflow_di (dst, src_lo)); | |
1572 | emit_insn (gen_aarch64_movtfhigh_di (dst, src_hi)); | |
1573 | } | |
1574 | return; | |
43e9d192 IB |
1575 | } |
1576 | else if (GP_REGNUM_P (dst_regno) && FP_REGNUM_P (src_regno)) | |
1577 | { | |
030d03b8 RE |
1578 | dst_lo = gen_lowpart (word_mode, dst); |
1579 | dst_hi = gen_highpart (word_mode, dst); | |
1580 | ||
1581 | if (mode == TImode) | |
1582 | { | |
1583 | emit_insn (gen_aarch64_movdi_tilow (dst_lo, src)); | |
1584 | emit_insn (gen_aarch64_movdi_tihigh (dst_hi, src)); | |
1585 | } | |
1586 | else | |
1587 | { | |
1588 | emit_insn (gen_aarch64_movdi_tflow (dst_lo, src)); | |
1589 | emit_insn (gen_aarch64_movdi_tfhigh (dst_hi, src)); | |
1590 | } | |
1591 | return; | |
43e9d192 | 1592 | } |
43e9d192 IB |
1593 | } |
1594 | ||
030d03b8 RE |
1595 | dst_lo = gen_lowpart (word_mode, dst); |
1596 | dst_hi = gen_highpart (word_mode, dst); | |
1597 | src_lo = gen_lowpart (word_mode, src); | |
1598 | src_hi = gen_highpart_mode (word_mode, mode, src); | |
1599 | ||
1600 | /* At most one pairing may overlap. */ | |
1601 | if (reg_overlap_mentioned_p (dst_lo, src_hi)) | |
1602 | { | |
1603 | aarch64_emit_move (dst_hi, src_hi); | |
1604 | aarch64_emit_move (dst_lo, src_lo); | |
1605 | } | |
1606 | else | |
1607 | { | |
1608 | aarch64_emit_move (dst_lo, src_lo); | |
1609 | aarch64_emit_move (dst_hi, src_hi); | |
1610 | } | |
43e9d192 IB |
1611 | } |
1612 | ||
1613 | bool | |
1614 | aarch64_split_128bit_move_p (rtx dst, rtx src) | |
1615 | { | |
1616 | return (! REG_P (src) | |
1617 | || ! (FP_REGNUM_P (REGNO (dst)) && FP_REGNUM_P (REGNO (src)))); | |
1618 | } | |
1619 | ||
8b033a8a SN |
1620 | /* Split a complex SIMD combine. */ |
1621 | ||
1622 | void | |
1623 | aarch64_split_simd_combine (rtx dst, rtx src1, rtx src2) | |
1624 | { | |
ef4bddc2 RS |
1625 | machine_mode src_mode = GET_MODE (src1); |
1626 | machine_mode dst_mode = GET_MODE (dst); | |
8b033a8a SN |
1627 | |
1628 | gcc_assert (VECTOR_MODE_P (dst_mode)); | |
1629 | ||
1630 | if (REG_P (dst) && REG_P (src1) && REG_P (src2)) | |
1631 | { | |
1632 | rtx (*gen) (rtx, rtx, rtx); | |
1633 | ||
1634 | switch (src_mode) | |
1635 | { | |
1636 | case V8QImode: | |
1637 | gen = gen_aarch64_simd_combinev8qi; | |
1638 | break; | |
1639 | case V4HImode: | |
1640 | gen = gen_aarch64_simd_combinev4hi; | |
1641 | break; | |
1642 | case V2SImode: | |
1643 | gen = gen_aarch64_simd_combinev2si; | |
1644 | break; | |
7c369485 AL |
1645 | case V4HFmode: |
1646 | gen = gen_aarch64_simd_combinev4hf; | |
1647 | break; | |
8b033a8a SN |
1648 | case V2SFmode: |
1649 | gen = gen_aarch64_simd_combinev2sf; | |
1650 | break; | |
1651 | case DImode: | |
1652 | gen = gen_aarch64_simd_combinedi; | |
1653 | break; | |
1654 | case DFmode: | |
1655 | gen = gen_aarch64_simd_combinedf; | |
1656 | break; | |
1657 | default: | |
1658 | gcc_unreachable (); | |
1659 | } | |
1660 | ||
1661 | emit_insn (gen (dst, src1, src2)); | |
1662 | return; | |
1663 | } | |
1664 | } | |
1665 | ||
fd4842cd SN |
1666 | /* Split a complex SIMD move. */ |
1667 | ||
1668 | void | |
1669 | aarch64_split_simd_move (rtx dst, rtx src) | |
1670 | { | |
ef4bddc2 RS |
1671 | machine_mode src_mode = GET_MODE (src); |
1672 | machine_mode dst_mode = GET_MODE (dst); | |
fd4842cd SN |
1673 | |
1674 | gcc_assert (VECTOR_MODE_P (dst_mode)); | |
1675 | ||
1676 | if (REG_P (dst) && REG_P (src)) | |
1677 | { | |
c59b7e28 SN |
1678 | rtx (*gen) (rtx, rtx); |
1679 | ||
fd4842cd SN |
1680 | gcc_assert (VECTOR_MODE_P (src_mode)); |
1681 | ||
1682 | switch (src_mode) | |
1683 | { | |
1684 | case V16QImode: | |
c59b7e28 | 1685 | gen = gen_aarch64_split_simd_movv16qi; |
fd4842cd SN |
1686 | break; |
1687 | case V8HImode: | |
c59b7e28 | 1688 | gen = gen_aarch64_split_simd_movv8hi; |
fd4842cd SN |
1689 | break; |
1690 | case V4SImode: | |
c59b7e28 | 1691 | gen = gen_aarch64_split_simd_movv4si; |
fd4842cd SN |
1692 | break; |
1693 | case V2DImode: | |
c59b7e28 | 1694 | gen = gen_aarch64_split_simd_movv2di; |
fd4842cd | 1695 | break; |
71a11456 AL |
1696 | case V8HFmode: |
1697 | gen = gen_aarch64_split_simd_movv8hf; | |
1698 | break; | |
fd4842cd | 1699 | case V4SFmode: |
c59b7e28 | 1700 | gen = gen_aarch64_split_simd_movv4sf; |
fd4842cd SN |
1701 | break; |
1702 | case V2DFmode: | |
c59b7e28 | 1703 | gen = gen_aarch64_split_simd_movv2df; |
fd4842cd SN |
1704 | break; |
1705 | default: | |
1706 | gcc_unreachable (); | |
1707 | } | |
c59b7e28 SN |
1708 | |
1709 | emit_insn (gen (dst, src)); | |
fd4842cd SN |
1710 | return; |
1711 | } | |
1712 | } | |
1713 | ||
ef22810a RH |
1714 | bool |
1715 | aarch64_zero_extend_const_eq (machine_mode xmode, rtx x, | |
1716 | machine_mode ymode, rtx y) | |
1717 | { | |
1718 | rtx r = simplify_const_unary_operation (ZERO_EXTEND, xmode, y, ymode); | |
1719 | gcc_assert (r != NULL); | |
1720 | return rtx_equal_p (x, r); | |
1721 | } | |
1722 | ||
1723 | ||
43e9d192 | 1724 | static rtx |
ef4bddc2 | 1725 | aarch64_force_temporary (machine_mode mode, rtx x, rtx value) |
43e9d192 IB |
1726 | { |
1727 | if (can_create_pseudo_p ()) | |
e18b4a81 | 1728 | return force_reg (mode, value); |
43e9d192 IB |
1729 | else |
1730 | { | |
1731 | x = aarch64_emit_move (x, value); | |
1732 | return x; | |
1733 | } | |
1734 | } | |
1735 | ||
1736 | ||
1737 | static rtx | |
ef4bddc2 | 1738 | aarch64_add_offset (machine_mode mode, rtx temp, rtx reg, HOST_WIDE_INT offset) |
43e9d192 | 1739 | { |
9c023bf0 | 1740 | if (!aarch64_plus_immediate (GEN_INT (offset), mode)) |
43e9d192 IB |
1741 | { |
1742 | rtx high; | |
1743 | /* Load the full offset into a register. This | |
1744 | might be improvable in the future. */ | |
1745 | high = GEN_INT (offset); | |
1746 | offset = 0; | |
e18b4a81 YZ |
1747 | high = aarch64_force_temporary (mode, temp, high); |
1748 | reg = aarch64_force_temporary (mode, temp, | |
1749 | gen_rtx_PLUS (mode, high, reg)); | |
43e9d192 IB |
1750 | } |
1751 | return plus_constant (mode, reg, offset); | |
1752 | } | |
1753 | ||
82614948 RR |
1754 | static int |
1755 | aarch64_internal_mov_immediate (rtx dest, rtx imm, bool generate, | |
1756 | machine_mode mode) | |
43e9d192 | 1757 | { |
43e9d192 | 1758 | int i; |
9a4865db WD |
1759 | unsigned HOST_WIDE_INT val, val2, mask; |
1760 | int one_match, zero_match; | |
1761 | int num_insns; | |
43e9d192 | 1762 | |
9a4865db WD |
1763 | val = INTVAL (imm); |
1764 | ||
1765 | if (aarch64_move_imm (val, mode)) | |
43e9d192 | 1766 | { |
82614948 | 1767 | if (generate) |
f7df4a84 | 1768 | emit_insn (gen_rtx_SET (dest, imm)); |
9a4865db | 1769 | return 1; |
43e9d192 IB |
1770 | } |
1771 | ||
9a4865db | 1772 | if ((val >> 32) == 0 || mode == SImode) |
43e9d192 | 1773 | { |
82614948 RR |
1774 | if (generate) |
1775 | { | |
9a4865db WD |
1776 | emit_insn (gen_rtx_SET (dest, GEN_INT (val & 0xffff))); |
1777 | if (mode == SImode) | |
1778 | emit_insn (gen_insv_immsi (dest, GEN_INT (16), | |
1779 | GEN_INT ((val >> 16) & 0xffff))); | |
1780 | else | |
1781 | emit_insn (gen_insv_immdi (dest, GEN_INT (16), | |
1782 | GEN_INT ((val >> 16) & 0xffff))); | |
82614948 | 1783 | } |
9a4865db | 1784 | return 2; |
43e9d192 IB |
1785 | } |
1786 | ||
1787 | /* Remaining cases are all for DImode. */ | |
1788 | ||
43e9d192 | 1789 | mask = 0xffff; |
9a4865db WD |
1790 | zero_match = ((val & mask) == 0) + ((val & (mask << 16)) == 0) + |
1791 | ((val & (mask << 32)) == 0) + ((val & (mask << 48)) == 0); | |
1792 | one_match = ((~val & mask) == 0) + ((~val & (mask << 16)) == 0) + | |
1793 | ((~val & (mask << 32)) == 0) + ((~val & (mask << 48)) == 0); | |
43e9d192 | 1794 | |
62c8d76c | 1795 | if (zero_match != 2 && one_match != 2) |
43e9d192 | 1796 | { |
62c8d76c WD |
1797 | /* Try emitting a bitmask immediate with a movk replacing 16 bits. |
1798 | For a 64-bit bitmask try whether changing 16 bits to all ones or | |
1799 | zeroes creates a valid bitmask. To check any repeated bitmask, | |
1800 | try using 16 bits from the other 32-bit half of val. */ | |
43e9d192 | 1801 | |
62c8d76c | 1802 | for (i = 0; i < 64; i += 16, mask <<= 16) |
43e9d192 | 1803 | { |
62c8d76c WD |
1804 | val2 = val & ~mask; |
1805 | if (val2 != val && aarch64_bitmask_imm (val2, mode)) | |
1806 | break; | |
1807 | val2 = val | mask; | |
1808 | if (val2 != val && aarch64_bitmask_imm (val2, mode)) | |
1809 | break; | |
1810 | val2 = val2 & ~mask; | |
1811 | val2 = val2 | (((val2 >> 32) | (val2 << 32)) & mask); | |
1812 | if (val2 != val && aarch64_bitmask_imm (val2, mode)) | |
1813 | break; | |
43e9d192 | 1814 | } |
62c8d76c | 1815 | if (i != 64) |
43e9d192 | 1816 | { |
62c8d76c | 1817 | if (generate) |
43e9d192 | 1818 | { |
62c8d76c WD |
1819 | emit_insn (gen_rtx_SET (dest, GEN_INT (val2))); |
1820 | emit_insn (gen_insv_immdi (dest, GEN_INT (i), | |
9a4865db | 1821 | GEN_INT ((val >> i) & 0xffff))); |
43e9d192 | 1822 | } |
1312b1ba | 1823 | return 2; |
43e9d192 IB |
1824 | } |
1825 | } | |
1826 | ||
9a4865db WD |
1827 | /* Generate 2-4 instructions, skipping 16 bits of all zeroes or ones which |
1828 | are emitted by the initial mov. If one_match > zero_match, skip set bits, | |
1829 | otherwise skip zero bits. */ | |
2c274197 | 1830 | |
9a4865db | 1831 | num_insns = 1; |
43e9d192 | 1832 | mask = 0xffff; |
9a4865db WD |
1833 | val2 = one_match > zero_match ? ~val : val; |
1834 | i = (val2 & mask) != 0 ? 0 : (val2 & (mask << 16)) != 0 ? 16 : 32; | |
1835 | ||
1836 | if (generate) | |
1837 | emit_insn (gen_rtx_SET (dest, GEN_INT (one_match > zero_match | |
1838 | ? (val | ~(mask << i)) | |
1839 | : (val & (mask << i))))); | |
1840 | for (i += 16; i < 64; i += 16) | |
43e9d192 | 1841 | { |
9a4865db WD |
1842 | if ((val2 & (mask << i)) == 0) |
1843 | continue; | |
1844 | if (generate) | |
1845 | emit_insn (gen_insv_immdi (dest, GEN_INT (i), | |
1846 | GEN_INT ((val >> i) & 0xffff))); | |
1847 | num_insns ++; | |
82614948 RR |
1848 | } |
1849 | ||
1850 | return num_insns; | |
1851 | } | |
1852 | ||
1853 | ||
1854 | void | |
1855 | aarch64_expand_mov_immediate (rtx dest, rtx imm) | |
1856 | { | |
1857 | machine_mode mode = GET_MODE (dest); | |
1858 | ||
1859 | gcc_assert (mode == SImode || mode == DImode); | |
1860 | ||
1861 | /* Check on what type of symbol it is. */ | |
1862 | if (GET_CODE (imm) == SYMBOL_REF | |
1863 | || GET_CODE (imm) == LABEL_REF | |
1864 | || GET_CODE (imm) == CONST) | |
1865 | { | |
1866 | rtx mem, base, offset; | |
1867 | enum aarch64_symbol_type sty; | |
1868 | ||
1869 | /* If we have (const (plus symbol offset)), separate out the offset | |
1870 | before we start classifying the symbol. */ | |
1871 | split_const (imm, &base, &offset); | |
1872 | ||
a6e0bfa7 | 1873 | sty = aarch64_classify_symbol (base, offset); |
82614948 RR |
1874 | switch (sty) |
1875 | { | |
1876 | case SYMBOL_FORCE_TO_MEM: | |
1877 | if (offset != const0_rtx | |
1878 | && targetm.cannot_force_const_mem (mode, imm)) | |
1879 | { | |
1880 | gcc_assert (can_create_pseudo_p ()); | |
1881 | base = aarch64_force_temporary (mode, dest, base); | |
1882 | base = aarch64_add_offset (mode, NULL, base, INTVAL (offset)); | |
1883 | aarch64_emit_move (dest, base); | |
1884 | return; | |
1885 | } | |
b4f50fd4 | 1886 | |
82614948 RR |
1887 | mem = force_const_mem (ptr_mode, imm); |
1888 | gcc_assert (mem); | |
b4f50fd4 RR |
1889 | |
1890 | /* If we aren't generating PC relative literals, then | |
1891 | we need to expand the literal pool access carefully. | |
1892 | This is something that needs to be done in a number | |
1893 | of places, so could well live as a separate function. */ | |
9ee6540a | 1894 | if (!aarch64_pcrelative_literal_loads) |
b4f50fd4 RR |
1895 | { |
1896 | gcc_assert (can_create_pseudo_p ()); | |
1897 | base = gen_reg_rtx (ptr_mode); | |
1898 | aarch64_expand_mov_immediate (base, XEXP (mem, 0)); | |
1899 | mem = gen_rtx_MEM (ptr_mode, base); | |
1900 | } | |
1901 | ||
82614948 RR |
1902 | if (mode != ptr_mode) |
1903 | mem = gen_rtx_ZERO_EXTEND (mode, mem); | |
b4f50fd4 | 1904 | |
f7df4a84 | 1905 | emit_insn (gen_rtx_SET (dest, mem)); |
b4f50fd4 | 1906 | |
82614948 RR |
1907 | return; |
1908 | ||
1909 | case SYMBOL_SMALL_TLSGD: | |
1910 | case SYMBOL_SMALL_TLSDESC: | |
79496620 | 1911 | case SYMBOL_SMALL_TLSIE: |
1b1e81f8 | 1912 | case SYMBOL_SMALL_GOT_28K: |
6642bdb4 | 1913 | case SYMBOL_SMALL_GOT_4G: |
82614948 | 1914 | case SYMBOL_TINY_GOT: |
5ae7caad | 1915 | case SYMBOL_TINY_TLSIE: |
82614948 RR |
1916 | if (offset != const0_rtx) |
1917 | { | |
1918 | gcc_assert(can_create_pseudo_p ()); | |
1919 | base = aarch64_force_temporary (mode, dest, base); | |
1920 | base = aarch64_add_offset (mode, NULL, base, INTVAL (offset)); | |
1921 | aarch64_emit_move (dest, base); | |
1922 | return; | |
1923 | } | |
1924 | /* FALLTHRU */ | |
1925 | ||
82614948 RR |
1926 | case SYMBOL_SMALL_ABSOLUTE: |
1927 | case SYMBOL_TINY_ABSOLUTE: | |
cbf5629e | 1928 | case SYMBOL_TLSLE12: |
d18ba284 | 1929 | case SYMBOL_TLSLE24: |
cbf5629e JW |
1930 | case SYMBOL_TLSLE32: |
1931 | case SYMBOL_TLSLE48: | |
82614948 RR |
1932 | aarch64_load_symref_appropriately (dest, imm, sty); |
1933 | return; | |
1934 | ||
1935 | default: | |
1936 | gcc_unreachable (); | |
1937 | } | |
1938 | } | |
1939 | ||
1940 | if (!CONST_INT_P (imm)) | |
1941 | { | |
1942 | if (GET_CODE (imm) == HIGH) | |
f7df4a84 | 1943 | emit_insn (gen_rtx_SET (dest, imm)); |
82614948 RR |
1944 | else |
1945 | { | |
1946 | rtx mem = force_const_mem (mode, imm); | |
1947 | gcc_assert (mem); | |
f7df4a84 | 1948 | emit_insn (gen_rtx_SET (dest, mem)); |
43e9d192 | 1949 | } |
82614948 RR |
1950 | |
1951 | return; | |
43e9d192 | 1952 | } |
82614948 RR |
1953 | |
1954 | aarch64_internal_mov_immediate (dest, imm, true, GET_MODE (dest)); | |
43e9d192 IB |
1955 | } |
1956 | ||
5be6b295 WD |
1957 | /* Add DELTA to REGNUM in mode MODE. SCRATCHREG can be used to hold a |
1958 | temporary value if necessary. FRAME_RELATED_P should be true if | |
1959 | the RTX_FRAME_RELATED flag should be set and CFA adjustments added | |
1960 | to the generated instructions. If SCRATCHREG is known to hold | |
1961 | abs (delta), EMIT_MOVE_IMM can be set to false to avoid emitting the | |
1962 | immediate again. | |
1963 | ||
1964 | Since this function may be used to adjust the stack pointer, we must | |
1965 | ensure that it cannot cause transient stack deallocation (for example | |
1966 | by first incrementing SP and then decrementing when adjusting by a | |
1967 | large immediate). */ | |
c4ddc43a JW |
1968 | |
1969 | static void | |
5be6b295 WD |
1970 | aarch64_add_constant_internal (machine_mode mode, int regnum, int scratchreg, |
1971 | HOST_WIDE_INT delta, bool frame_related_p, | |
1972 | bool emit_move_imm) | |
c4ddc43a JW |
1973 | { |
1974 | HOST_WIDE_INT mdelta = abs_hwi (delta); | |
1975 | rtx this_rtx = gen_rtx_REG (mode, regnum); | |
37d6a4b7 | 1976 | rtx_insn *insn; |
c4ddc43a | 1977 | |
c4ddc43a JW |
1978 | if (!mdelta) |
1979 | return; | |
1980 | ||
5be6b295 | 1981 | /* Single instruction adjustment. */ |
c4ddc43a JW |
1982 | if (aarch64_uimm12_shift (mdelta)) |
1983 | { | |
37d6a4b7 JW |
1984 | insn = emit_insn (gen_add2_insn (this_rtx, GEN_INT (delta))); |
1985 | RTX_FRAME_RELATED_P (insn) = frame_related_p; | |
c4ddc43a JW |
1986 | return; |
1987 | } | |
1988 | ||
5be6b295 WD |
1989 | /* Emit 2 additions/subtractions if the adjustment is less than 24 bits. |
1990 | Only do this if mdelta is not a 16-bit move as adjusting using a move | |
1991 | is better. */ | |
1992 | if (mdelta < 0x1000000 && !aarch64_move_imm (mdelta, mode)) | |
c4ddc43a JW |
1993 | { |
1994 | HOST_WIDE_INT low_off = mdelta & 0xfff; | |
1995 | ||
1996 | low_off = delta < 0 ? -low_off : low_off; | |
37d6a4b7 JW |
1997 | insn = emit_insn (gen_add2_insn (this_rtx, GEN_INT (low_off))); |
1998 | RTX_FRAME_RELATED_P (insn) = frame_related_p; | |
1999 | insn = emit_insn (gen_add2_insn (this_rtx, GEN_INT (delta - low_off))); | |
2000 | RTX_FRAME_RELATED_P (insn) = frame_related_p; | |
c4ddc43a JW |
2001 | return; |
2002 | } | |
2003 | ||
5be6b295 | 2004 | /* Emit a move immediate if required and an addition/subtraction. */ |
c4ddc43a | 2005 | rtx scratch_rtx = gen_rtx_REG (mode, scratchreg); |
5be6b295 WD |
2006 | if (emit_move_imm) |
2007 | aarch64_internal_mov_immediate (scratch_rtx, GEN_INT (mdelta), true, mode); | |
2008 | insn = emit_insn (delta < 0 ? gen_sub2_insn (this_rtx, scratch_rtx) | |
2009 | : gen_add2_insn (this_rtx, scratch_rtx)); | |
37d6a4b7 JW |
2010 | if (frame_related_p) |
2011 | { | |
2012 | RTX_FRAME_RELATED_P (insn) = frame_related_p; | |
2013 | rtx adj = plus_constant (mode, this_rtx, delta); | |
2014 | add_reg_note (insn , REG_CFA_ADJUST_CFA, gen_rtx_SET (this_rtx, adj)); | |
2015 | } | |
c4ddc43a JW |
2016 | } |
2017 | ||
5be6b295 WD |
2018 | static inline void |
2019 | aarch64_add_constant (machine_mode mode, int regnum, int scratchreg, | |
2020 | HOST_WIDE_INT delta) | |
2021 | { | |
2022 | aarch64_add_constant_internal (mode, regnum, scratchreg, delta, false, true); | |
2023 | } | |
2024 | ||
2025 | static inline void | |
2026 | aarch64_add_sp (int scratchreg, HOST_WIDE_INT delta, bool emit_move_imm) | |
2027 | { | |
2028 | aarch64_add_constant_internal (Pmode, SP_REGNUM, scratchreg, delta, | |
2029 | true, emit_move_imm); | |
2030 | } | |
2031 | ||
2032 | static inline void | |
2033 | aarch64_sub_sp (int scratchreg, HOST_WIDE_INT delta, bool frame_related_p) | |
2034 | { | |
2035 | aarch64_add_constant_internal (Pmode, SP_REGNUM, scratchreg, -delta, | |
2036 | frame_related_p, true); | |
2037 | } | |
2038 | ||
43e9d192 | 2039 | static bool |
fee9ba42 JW |
2040 | aarch64_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED, |
2041 | tree exp ATTRIBUTE_UNUSED) | |
43e9d192 | 2042 | { |
fee9ba42 | 2043 | /* Currently, always true. */ |
43e9d192 IB |
2044 | return true; |
2045 | } | |
2046 | ||
2047 | /* Implement TARGET_PASS_BY_REFERENCE. */ | |
2048 | ||
2049 | static bool | |
2050 | aarch64_pass_by_reference (cumulative_args_t pcum ATTRIBUTE_UNUSED, | |
ef4bddc2 | 2051 | machine_mode mode, |
43e9d192 IB |
2052 | const_tree type, |
2053 | bool named ATTRIBUTE_UNUSED) | |
2054 | { | |
2055 | HOST_WIDE_INT size; | |
ef4bddc2 | 2056 | machine_mode dummymode; |
43e9d192 IB |
2057 | int nregs; |
2058 | ||
2059 | /* GET_MODE_SIZE (BLKmode) is useless since it is 0. */ | |
2060 | size = (mode == BLKmode && type) | |
2061 | ? int_size_in_bytes (type) : (int) GET_MODE_SIZE (mode); | |
2062 | ||
aadc1c43 MHD |
2063 | /* Aggregates are passed by reference based on their size. */ |
2064 | if (type && AGGREGATE_TYPE_P (type)) | |
43e9d192 | 2065 | { |
aadc1c43 | 2066 | size = int_size_in_bytes (type); |
43e9d192 IB |
2067 | } |
2068 | ||
2069 | /* Variable sized arguments are always returned by reference. */ | |
2070 | if (size < 0) | |
2071 | return true; | |
2072 | ||
2073 | /* Can this be a candidate to be passed in fp/simd register(s)? */ | |
2074 | if (aarch64_vfp_is_call_or_return_candidate (mode, type, | |
2075 | &dummymode, &nregs, | |
2076 | NULL)) | |
2077 | return false; | |
2078 | ||
2079 | /* Arguments which are variable sized or larger than 2 registers are | |
2080 | passed by reference unless they are a homogenous floating point | |
2081 | aggregate. */ | |
2082 | return size > 2 * UNITS_PER_WORD; | |
2083 | } | |
2084 | ||
2085 | /* Return TRUE if VALTYPE is padded to its least significant bits. */ | |
2086 | static bool | |
2087 | aarch64_return_in_msb (const_tree valtype) | |
2088 | { | |
ef4bddc2 | 2089 | machine_mode dummy_mode; |
43e9d192 IB |
2090 | int dummy_int; |
2091 | ||
2092 | /* Never happens in little-endian mode. */ | |
2093 | if (!BYTES_BIG_ENDIAN) | |
2094 | return false; | |
2095 | ||
2096 | /* Only composite types smaller than or equal to 16 bytes can | |
2097 | be potentially returned in registers. */ | |
2098 | if (!aarch64_composite_type_p (valtype, TYPE_MODE (valtype)) | |
2099 | || int_size_in_bytes (valtype) <= 0 | |
2100 | || int_size_in_bytes (valtype) > 16) | |
2101 | return false; | |
2102 | ||
2103 | /* But not a composite that is an HFA (Homogeneous Floating-point Aggregate) | |
2104 | or an HVA (Homogeneous Short-Vector Aggregate); such a special composite | |
2105 | is always passed/returned in the least significant bits of fp/simd | |
2106 | register(s). */ | |
2107 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (valtype), valtype, | |
2108 | &dummy_mode, &dummy_int, NULL)) | |
2109 | return false; | |
2110 | ||
2111 | return true; | |
2112 | } | |
2113 | ||
2114 | /* Implement TARGET_FUNCTION_VALUE. | |
2115 | Define how to find the value returned by a function. */ | |
2116 | ||
2117 | static rtx | |
2118 | aarch64_function_value (const_tree type, const_tree func, | |
2119 | bool outgoing ATTRIBUTE_UNUSED) | |
2120 | { | |
ef4bddc2 | 2121 | machine_mode mode; |
43e9d192 IB |
2122 | int unsignedp; |
2123 | int count; | |
ef4bddc2 | 2124 | machine_mode ag_mode; |
43e9d192 IB |
2125 | |
2126 | mode = TYPE_MODE (type); | |
2127 | if (INTEGRAL_TYPE_P (type)) | |
2128 | mode = promote_function_mode (type, mode, &unsignedp, func, 1); | |
2129 | ||
2130 | if (aarch64_return_in_msb (type)) | |
2131 | { | |
2132 | HOST_WIDE_INT size = int_size_in_bytes (type); | |
2133 | ||
2134 | if (size % UNITS_PER_WORD != 0) | |
2135 | { | |
2136 | size += UNITS_PER_WORD - size % UNITS_PER_WORD; | |
2137 | mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0); | |
2138 | } | |
2139 | } | |
2140 | ||
2141 | if (aarch64_vfp_is_call_or_return_candidate (mode, type, | |
2142 | &ag_mode, &count, NULL)) | |
2143 | { | |
2144 | if (!aarch64_composite_type_p (type, mode)) | |
2145 | { | |
2146 | gcc_assert (count == 1 && mode == ag_mode); | |
2147 | return gen_rtx_REG (mode, V0_REGNUM); | |
2148 | } | |
2149 | else | |
2150 | { | |
2151 | int i; | |
2152 | rtx par; | |
2153 | ||
2154 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (count)); | |
2155 | for (i = 0; i < count; i++) | |
2156 | { | |
2157 | rtx tmp = gen_rtx_REG (ag_mode, V0_REGNUM + i); | |
2158 | tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp, | |
2159 | GEN_INT (i * GET_MODE_SIZE (ag_mode))); | |
2160 | XVECEXP (par, 0, i) = tmp; | |
2161 | } | |
2162 | return par; | |
2163 | } | |
2164 | } | |
2165 | else | |
2166 | return gen_rtx_REG (mode, R0_REGNUM); | |
2167 | } | |
2168 | ||
2169 | /* Implements TARGET_FUNCTION_VALUE_REGNO_P. | |
2170 | Return true if REGNO is the number of a hard register in which the values | |
2171 | of called function may come back. */ | |
2172 | ||
2173 | static bool | |
2174 | aarch64_function_value_regno_p (const unsigned int regno) | |
2175 | { | |
2176 | /* Maximum of 16 bytes can be returned in the general registers. Examples | |
2177 | of 16-byte return values are: 128-bit integers and 16-byte small | |
2178 | structures (excluding homogeneous floating-point aggregates). */ | |
2179 | if (regno == R0_REGNUM || regno == R1_REGNUM) | |
2180 | return true; | |
2181 | ||
2182 | /* Up to four fp/simd registers can return a function value, e.g. a | |
2183 | homogeneous floating-point aggregate having four members. */ | |
2184 | if (regno >= V0_REGNUM && regno < V0_REGNUM + HA_MAX_NUM_FLDS) | |
d5726973 | 2185 | return TARGET_FLOAT; |
43e9d192 IB |
2186 | |
2187 | return false; | |
2188 | } | |
2189 | ||
2190 | /* Implement TARGET_RETURN_IN_MEMORY. | |
2191 | ||
2192 | If the type T of the result of a function is such that | |
2193 | void func (T arg) | |
2194 | would require that arg be passed as a value in a register (or set of | |
2195 | registers) according to the parameter passing rules, then the result | |
2196 | is returned in the same registers as would be used for such an | |
2197 | argument. */ | |
2198 | ||
2199 | static bool | |
2200 | aarch64_return_in_memory (const_tree type, const_tree fndecl ATTRIBUTE_UNUSED) | |
2201 | { | |
2202 | HOST_WIDE_INT size; | |
ef4bddc2 | 2203 | machine_mode ag_mode; |
43e9d192 IB |
2204 | int count; |
2205 | ||
2206 | if (!AGGREGATE_TYPE_P (type) | |
2207 | && TREE_CODE (type) != COMPLEX_TYPE | |
2208 | && TREE_CODE (type) != VECTOR_TYPE) | |
2209 | /* Simple scalar types always returned in registers. */ | |
2210 | return false; | |
2211 | ||
2212 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (type), | |
2213 | type, | |
2214 | &ag_mode, | |
2215 | &count, | |
2216 | NULL)) | |
2217 | return false; | |
2218 | ||
2219 | /* Types larger than 2 registers returned in memory. */ | |
2220 | size = int_size_in_bytes (type); | |
2221 | return (size < 0 || size > 2 * UNITS_PER_WORD); | |
2222 | } | |
2223 | ||
2224 | static bool | |
ef4bddc2 | 2225 | aarch64_vfp_is_call_candidate (cumulative_args_t pcum_v, machine_mode mode, |
43e9d192 IB |
2226 | const_tree type, int *nregs) |
2227 | { | |
2228 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
2229 | return aarch64_vfp_is_call_or_return_candidate (mode, | |
2230 | type, | |
2231 | &pcum->aapcs_vfp_rmode, | |
2232 | nregs, | |
2233 | NULL); | |
2234 | } | |
2235 | ||
2236 | /* Given MODE and TYPE of a function argument, return the alignment in | |
2237 | bits. The idea is to suppress any stronger alignment requested by | |
2238 | the user and opt for the natural alignment (specified in AAPCS64 \S 4.1). | |
2239 | This is a helper function for local use only. */ | |
2240 | ||
2241 | static unsigned int | |
ef4bddc2 | 2242 | aarch64_function_arg_alignment (machine_mode mode, const_tree type) |
43e9d192 | 2243 | { |
75d6cc81 AL |
2244 | if (!type) |
2245 | return GET_MODE_ALIGNMENT (mode); | |
2246 | if (integer_zerop (TYPE_SIZE (type))) | |
2247 | return 0; | |
43e9d192 | 2248 | |
75d6cc81 AL |
2249 | gcc_assert (TYPE_MODE (type) == mode); |
2250 | ||
2251 | if (!AGGREGATE_TYPE_P (type)) | |
2252 | return TYPE_ALIGN (TYPE_MAIN_VARIANT (type)); | |
2253 | ||
2254 | if (TREE_CODE (type) == ARRAY_TYPE) | |
2255 | return TYPE_ALIGN (TREE_TYPE (type)); | |
2256 | ||
2257 | unsigned int alignment = 0; | |
2258 | ||
2259 | for (tree field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field)) | |
2260 | alignment = std::max (alignment, DECL_ALIGN (field)); | |
43e9d192 IB |
2261 | |
2262 | return alignment; | |
2263 | } | |
2264 | ||
2265 | /* Layout a function argument according to the AAPCS64 rules. The rule | |
2266 | numbers refer to the rule numbers in the AAPCS64. */ | |
2267 | ||
2268 | static void | |
ef4bddc2 | 2269 | aarch64_layout_arg (cumulative_args_t pcum_v, machine_mode mode, |
43e9d192 IB |
2270 | const_tree type, |
2271 | bool named ATTRIBUTE_UNUSED) | |
2272 | { | |
2273 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
2274 | int ncrn, nvrn, nregs; | |
2275 | bool allocate_ncrn, allocate_nvrn; | |
3abf17cf | 2276 | HOST_WIDE_INT size; |
43e9d192 IB |
2277 | |
2278 | /* We need to do this once per argument. */ | |
2279 | if (pcum->aapcs_arg_processed) | |
2280 | return; | |
2281 | ||
2282 | pcum->aapcs_arg_processed = true; | |
2283 | ||
3abf17cf YZ |
2284 | /* Size in bytes, rounded to the nearest multiple of 8 bytes. */ |
2285 | size | |
4f59f9f2 UB |
2286 | = ROUND_UP (type ? int_size_in_bytes (type) : GET_MODE_SIZE (mode), |
2287 | UNITS_PER_WORD); | |
3abf17cf | 2288 | |
43e9d192 IB |
2289 | allocate_ncrn = (type) ? !(FLOAT_TYPE_P (type)) : !FLOAT_MODE_P (mode); |
2290 | allocate_nvrn = aarch64_vfp_is_call_candidate (pcum_v, | |
2291 | mode, | |
2292 | type, | |
2293 | &nregs); | |
2294 | ||
2295 | /* allocate_ncrn may be false-positive, but allocate_nvrn is quite reliable. | |
2296 | The following code thus handles passing by SIMD/FP registers first. */ | |
2297 | ||
2298 | nvrn = pcum->aapcs_nvrn; | |
2299 | ||
2300 | /* C1 - C5 for floating point, homogenous floating point aggregates (HFA) | |
2301 | and homogenous short-vector aggregates (HVA). */ | |
2302 | if (allocate_nvrn) | |
2303 | { | |
261fb553 AL |
2304 | if (!TARGET_FLOAT) |
2305 | aarch64_err_no_fpadvsimd (mode, "argument"); | |
2306 | ||
43e9d192 IB |
2307 | if (nvrn + nregs <= NUM_FP_ARG_REGS) |
2308 | { | |
2309 | pcum->aapcs_nextnvrn = nvrn + nregs; | |
2310 | if (!aarch64_composite_type_p (type, mode)) | |
2311 | { | |
2312 | gcc_assert (nregs == 1); | |
2313 | pcum->aapcs_reg = gen_rtx_REG (mode, V0_REGNUM + nvrn); | |
2314 | } | |
2315 | else | |
2316 | { | |
2317 | rtx par; | |
2318 | int i; | |
2319 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (nregs)); | |
2320 | for (i = 0; i < nregs; i++) | |
2321 | { | |
2322 | rtx tmp = gen_rtx_REG (pcum->aapcs_vfp_rmode, | |
2323 | V0_REGNUM + nvrn + i); | |
2324 | tmp = gen_rtx_EXPR_LIST | |
2325 | (VOIDmode, tmp, | |
2326 | GEN_INT (i * GET_MODE_SIZE (pcum->aapcs_vfp_rmode))); | |
2327 | XVECEXP (par, 0, i) = tmp; | |
2328 | } | |
2329 | pcum->aapcs_reg = par; | |
2330 | } | |
2331 | return; | |
2332 | } | |
2333 | else | |
2334 | { | |
2335 | /* C.3 NSRN is set to 8. */ | |
2336 | pcum->aapcs_nextnvrn = NUM_FP_ARG_REGS; | |
2337 | goto on_stack; | |
2338 | } | |
2339 | } | |
2340 | ||
2341 | ncrn = pcum->aapcs_ncrn; | |
3abf17cf | 2342 | nregs = size / UNITS_PER_WORD; |
43e9d192 IB |
2343 | |
2344 | /* C6 - C9. though the sign and zero extension semantics are | |
2345 | handled elsewhere. This is the case where the argument fits | |
2346 | entirely general registers. */ | |
2347 | if (allocate_ncrn && (ncrn + nregs <= NUM_ARG_REGS)) | |
2348 | { | |
2349 | unsigned int alignment = aarch64_function_arg_alignment (mode, type); | |
2350 | ||
2351 | gcc_assert (nregs == 0 || nregs == 1 || nregs == 2); | |
2352 | ||
2353 | /* C.8 if the argument has an alignment of 16 then the NGRN is | |
2354 | rounded up to the next even number. */ | |
2355 | if (nregs == 2 && alignment == 16 * BITS_PER_UNIT && ncrn % 2) | |
2356 | { | |
2357 | ++ncrn; | |
2358 | gcc_assert (ncrn + nregs <= NUM_ARG_REGS); | |
2359 | } | |
2360 | /* NREGS can be 0 when e.g. an empty structure is to be passed. | |
2361 | A reg is still generated for it, but the caller should be smart | |
2362 | enough not to use it. */ | |
2363 | if (nregs == 0 || nregs == 1 || GET_MODE_CLASS (mode) == MODE_INT) | |
2364 | { | |
2365 | pcum->aapcs_reg = gen_rtx_REG (mode, R0_REGNUM + ncrn); | |
2366 | } | |
2367 | else | |
2368 | { | |
2369 | rtx par; | |
2370 | int i; | |
2371 | ||
2372 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (nregs)); | |
2373 | for (i = 0; i < nregs; i++) | |
2374 | { | |
2375 | rtx tmp = gen_rtx_REG (word_mode, R0_REGNUM + ncrn + i); | |
2376 | tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp, | |
2377 | GEN_INT (i * UNITS_PER_WORD)); | |
2378 | XVECEXP (par, 0, i) = tmp; | |
2379 | } | |
2380 | pcum->aapcs_reg = par; | |
2381 | } | |
2382 | ||
2383 | pcum->aapcs_nextncrn = ncrn + nregs; | |
2384 | return; | |
2385 | } | |
2386 | ||
2387 | /* C.11 */ | |
2388 | pcum->aapcs_nextncrn = NUM_ARG_REGS; | |
2389 | ||
2390 | /* The argument is passed on stack; record the needed number of words for | |
3abf17cf | 2391 | this argument and align the total size if necessary. */ |
43e9d192 | 2392 | on_stack: |
3abf17cf | 2393 | pcum->aapcs_stack_words = size / UNITS_PER_WORD; |
43e9d192 | 2394 | if (aarch64_function_arg_alignment (mode, type) == 16 * BITS_PER_UNIT) |
4f59f9f2 UB |
2395 | pcum->aapcs_stack_size = ROUND_UP (pcum->aapcs_stack_size, |
2396 | 16 / UNITS_PER_WORD); | |
43e9d192 IB |
2397 | return; |
2398 | } | |
2399 | ||
2400 | /* Implement TARGET_FUNCTION_ARG. */ | |
2401 | ||
2402 | static rtx | |
ef4bddc2 | 2403 | aarch64_function_arg (cumulative_args_t pcum_v, machine_mode mode, |
43e9d192 IB |
2404 | const_tree type, bool named) |
2405 | { | |
2406 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
2407 | gcc_assert (pcum->pcs_variant == ARM_PCS_AAPCS64); | |
2408 | ||
2409 | if (mode == VOIDmode) | |
2410 | return NULL_RTX; | |
2411 | ||
2412 | aarch64_layout_arg (pcum_v, mode, type, named); | |
2413 | return pcum->aapcs_reg; | |
2414 | } | |
2415 | ||
2416 | void | |
2417 | aarch64_init_cumulative_args (CUMULATIVE_ARGS *pcum, | |
2418 | const_tree fntype ATTRIBUTE_UNUSED, | |
2419 | rtx libname ATTRIBUTE_UNUSED, | |
2420 | const_tree fndecl ATTRIBUTE_UNUSED, | |
2421 | unsigned n_named ATTRIBUTE_UNUSED) | |
2422 | { | |
2423 | pcum->aapcs_ncrn = 0; | |
2424 | pcum->aapcs_nvrn = 0; | |
2425 | pcum->aapcs_nextncrn = 0; | |
2426 | pcum->aapcs_nextnvrn = 0; | |
2427 | pcum->pcs_variant = ARM_PCS_AAPCS64; | |
2428 | pcum->aapcs_reg = NULL_RTX; | |
2429 | pcum->aapcs_arg_processed = false; | |
2430 | pcum->aapcs_stack_words = 0; | |
2431 | pcum->aapcs_stack_size = 0; | |
2432 | ||
261fb553 AL |
2433 | if (!TARGET_FLOAT |
2434 | && fndecl && TREE_PUBLIC (fndecl) | |
2435 | && fntype && fntype != error_mark_node) | |
2436 | { | |
2437 | const_tree type = TREE_TYPE (fntype); | |
2438 | machine_mode mode ATTRIBUTE_UNUSED; /* To pass pointer as argument. */ | |
2439 | int nregs ATTRIBUTE_UNUSED; /* Likewise. */ | |
2440 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (type), type, | |
2441 | &mode, &nregs, NULL)) | |
2442 | aarch64_err_no_fpadvsimd (TYPE_MODE (type), "return type"); | |
2443 | } | |
43e9d192 IB |
2444 | return; |
2445 | } | |
2446 | ||
2447 | static void | |
2448 | aarch64_function_arg_advance (cumulative_args_t pcum_v, | |
ef4bddc2 | 2449 | machine_mode mode, |
43e9d192 IB |
2450 | const_tree type, |
2451 | bool named) | |
2452 | { | |
2453 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
2454 | if (pcum->pcs_variant == ARM_PCS_AAPCS64) | |
2455 | { | |
2456 | aarch64_layout_arg (pcum_v, mode, type, named); | |
2457 | gcc_assert ((pcum->aapcs_reg != NULL_RTX) | |
2458 | != (pcum->aapcs_stack_words != 0)); | |
2459 | pcum->aapcs_arg_processed = false; | |
2460 | pcum->aapcs_ncrn = pcum->aapcs_nextncrn; | |
2461 | pcum->aapcs_nvrn = pcum->aapcs_nextnvrn; | |
2462 | pcum->aapcs_stack_size += pcum->aapcs_stack_words; | |
2463 | pcum->aapcs_stack_words = 0; | |
2464 | pcum->aapcs_reg = NULL_RTX; | |
2465 | } | |
2466 | } | |
2467 | ||
2468 | bool | |
2469 | aarch64_function_arg_regno_p (unsigned regno) | |
2470 | { | |
2471 | return ((GP_REGNUM_P (regno) && regno < R0_REGNUM + NUM_ARG_REGS) | |
2472 | || (FP_REGNUM_P (regno) && regno < V0_REGNUM + NUM_FP_ARG_REGS)); | |
2473 | } | |
2474 | ||
2475 | /* Implement FUNCTION_ARG_BOUNDARY. Every parameter gets at least | |
2476 | PARM_BOUNDARY bits of alignment, but will be given anything up | |
2477 | to STACK_BOUNDARY bits if the type requires it. This makes sure | |
2478 | that both before and after the layout of each argument, the Next | |
2479 | Stacked Argument Address (NSAA) will have a minimum alignment of | |
2480 | 8 bytes. */ | |
2481 | ||
2482 | static unsigned int | |
ef4bddc2 | 2483 | aarch64_function_arg_boundary (machine_mode mode, const_tree type) |
43e9d192 IB |
2484 | { |
2485 | unsigned int alignment = aarch64_function_arg_alignment (mode, type); | |
2486 | ||
2487 | if (alignment < PARM_BOUNDARY) | |
2488 | alignment = PARM_BOUNDARY; | |
2489 | if (alignment > STACK_BOUNDARY) | |
2490 | alignment = STACK_BOUNDARY; | |
2491 | return alignment; | |
2492 | } | |
2493 | ||
2494 | /* For use by FUNCTION_ARG_PADDING (MODE, TYPE). | |
2495 | ||
2496 | Return true if an argument passed on the stack should be padded upwards, | |
2497 | i.e. if the least-significant byte of the stack slot has useful data. | |
2498 | ||
2499 | Small aggregate types are placed in the lowest memory address. | |
2500 | ||
2501 | The related parameter passing rules are B.4, C.3, C.5 and C.14. */ | |
2502 | ||
2503 | bool | |
ef4bddc2 | 2504 | aarch64_pad_arg_upward (machine_mode mode, const_tree type) |
43e9d192 IB |
2505 | { |
2506 | /* On little-endian targets, the least significant byte of every stack | |
2507 | argument is passed at the lowest byte address of the stack slot. */ | |
2508 | if (!BYTES_BIG_ENDIAN) | |
2509 | return true; | |
2510 | ||
00edcfbe | 2511 | /* Otherwise, integral, floating-point and pointer types are padded downward: |
43e9d192 IB |
2512 | the least significant byte of a stack argument is passed at the highest |
2513 | byte address of the stack slot. */ | |
2514 | if (type | |
00edcfbe YZ |
2515 | ? (INTEGRAL_TYPE_P (type) || SCALAR_FLOAT_TYPE_P (type) |
2516 | || POINTER_TYPE_P (type)) | |
43e9d192 IB |
2517 | : (SCALAR_INT_MODE_P (mode) || SCALAR_FLOAT_MODE_P (mode))) |
2518 | return false; | |
2519 | ||
2520 | /* Everything else padded upward, i.e. data in first byte of stack slot. */ | |
2521 | return true; | |
2522 | } | |
2523 | ||
2524 | /* Similarly, for use by BLOCK_REG_PADDING (MODE, TYPE, FIRST). | |
2525 | ||
2526 | It specifies padding for the last (may also be the only) | |
2527 | element of a block move between registers and memory. If | |
2528 | assuming the block is in the memory, padding upward means that | |
2529 | the last element is padded after its highest significant byte, | |
2530 | while in downward padding, the last element is padded at the | |
2531 | its least significant byte side. | |
2532 | ||
2533 | Small aggregates and small complex types are always padded | |
2534 | upwards. | |
2535 | ||
2536 | We don't need to worry about homogeneous floating-point or | |
2537 | short-vector aggregates; their move is not affected by the | |
2538 | padding direction determined here. Regardless of endianness, | |
2539 | each element of such an aggregate is put in the least | |
2540 | significant bits of a fp/simd register. | |
2541 | ||
2542 | Return !BYTES_BIG_ENDIAN if the least significant byte of the | |
2543 | register has useful data, and return the opposite if the most | |
2544 | significant byte does. */ | |
2545 | ||
2546 | bool | |
ef4bddc2 | 2547 | aarch64_pad_reg_upward (machine_mode mode, const_tree type, |
43e9d192 IB |
2548 | bool first ATTRIBUTE_UNUSED) |
2549 | { | |
2550 | ||
2551 | /* Small composite types are always padded upward. */ | |
2552 | if (BYTES_BIG_ENDIAN && aarch64_composite_type_p (type, mode)) | |
2553 | { | |
2554 | HOST_WIDE_INT size = (type ? int_size_in_bytes (type) | |
2555 | : GET_MODE_SIZE (mode)); | |
2556 | if (size < 2 * UNITS_PER_WORD) | |
2557 | return true; | |
2558 | } | |
2559 | ||
2560 | /* Otherwise, use the default padding. */ | |
2561 | return !BYTES_BIG_ENDIAN; | |
2562 | } | |
2563 | ||
ef4bddc2 | 2564 | static machine_mode |
43e9d192 IB |
2565 | aarch64_libgcc_cmp_return_mode (void) |
2566 | { | |
2567 | return SImode; | |
2568 | } | |
2569 | ||
a3eb8a52 EB |
2570 | #define PROBE_INTERVAL (1 << STACK_CHECK_PROBE_INTERVAL_EXP) |
2571 | ||
2572 | /* We use the 12-bit shifted immediate arithmetic instructions so values | |
2573 | must be multiple of (1 << 12), i.e. 4096. */ | |
2574 | #define ARITH_FACTOR 4096 | |
2575 | ||
2576 | #if (PROBE_INTERVAL % ARITH_FACTOR) != 0 | |
2577 | #error Cannot use simple address calculation for stack probing | |
2578 | #endif | |
2579 | ||
2580 | /* The pair of scratch registers used for stack probing. */ | |
2581 | #define PROBE_STACK_FIRST_REG 9 | |
2582 | #define PROBE_STACK_SECOND_REG 10 | |
2583 | ||
2584 | /* Emit code to probe a range of stack addresses from FIRST to FIRST+SIZE, | |
2585 | inclusive. These are offsets from the current stack pointer. */ | |
2586 | ||
2587 | static void | |
2588 | aarch64_emit_probe_stack_range (HOST_WIDE_INT first, HOST_WIDE_INT size) | |
2589 | { | |
2590 | rtx reg1 = gen_rtx_REG (ptr_mode, PROBE_STACK_FIRST_REG); | |
2591 | ||
2592 | /* See the same assertion on PROBE_INTERVAL above. */ | |
2593 | gcc_assert ((first % ARITH_FACTOR) == 0); | |
2594 | ||
2595 | /* See if we have a constant small number of probes to generate. If so, | |
2596 | that's the easy case. */ | |
2597 | if (size <= PROBE_INTERVAL) | |
2598 | { | |
2599 | const HOST_WIDE_INT base = ROUND_UP (size, ARITH_FACTOR); | |
2600 | ||
2601 | emit_set_insn (reg1, | |
2602 | plus_constant (ptr_mode, | |
2603 | stack_pointer_rtx, -(first + base))); | |
2604 | emit_stack_probe (plus_constant (ptr_mode, reg1, base - size)); | |
2605 | } | |
2606 | ||
2607 | /* The run-time loop is made up of 8 insns in the generic case while the | |
2608 | compile-time loop is made up of 4+2*(n-2) insns for n # of intervals. */ | |
2609 | else if (size <= 4 * PROBE_INTERVAL) | |
2610 | { | |
2611 | HOST_WIDE_INT i, rem; | |
2612 | ||
2613 | emit_set_insn (reg1, | |
2614 | plus_constant (ptr_mode, | |
2615 | stack_pointer_rtx, | |
2616 | -(first + PROBE_INTERVAL))); | |
2617 | emit_stack_probe (reg1); | |
2618 | ||
2619 | /* Probe at FIRST + N * PROBE_INTERVAL for values of N from 2 until | |
2620 | it exceeds SIZE. If only two probes are needed, this will not | |
2621 | generate any code. Then probe at FIRST + SIZE. */ | |
2622 | for (i = 2 * PROBE_INTERVAL; i < size; i += PROBE_INTERVAL) | |
2623 | { | |
2624 | emit_set_insn (reg1, | |
2625 | plus_constant (ptr_mode, reg1, -PROBE_INTERVAL)); | |
2626 | emit_stack_probe (reg1); | |
2627 | } | |
2628 | ||
2629 | rem = size - (i - PROBE_INTERVAL); | |
2630 | if (rem > 256) | |
2631 | { | |
2632 | const HOST_WIDE_INT base = ROUND_UP (rem, ARITH_FACTOR); | |
2633 | ||
2634 | emit_set_insn (reg1, plus_constant (ptr_mode, reg1, -base)); | |
2635 | emit_stack_probe (plus_constant (ptr_mode, reg1, base - rem)); | |
2636 | } | |
2637 | else | |
2638 | emit_stack_probe (plus_constant (ptr_mode, reg1, -rem)); | |
2639 | } | |
2640 | ||
2641 | /* Otherwise, do the same as above, but in a loop. Note that we must be | |
2642 | extra careful with variables wrapping around because we might be at | |
2643 | the very top (or the very bottom) of the address space and we have | |
2644 | to be able to handle this case properly; in particular, we use an | |
2645 | equality test for the loop condition. */ | |
2646 | else | |
2647 | { | |
2648 | rtx reg2 = gen_rtx_REG (ptr_mode, PROBE_STACK_SECOND_REG); | |
2649 | ||
2650 | /* Step 1: round SIZE to the previous multiple of the interval. */ | |
2651 | ||
2652 | HOST_WIDE_INT rounded_size = size & -PROBE_INTERVAL; | |
2653 | ||
2654 | ||
2655 | /* Step 2: compute initial and final value of the loop counter. */ | |
2656 | ||
2657 | /* TEST_ADDR = SP + FIRST. */ | |
2658 | emit_set_insn (reg1, | |
2659 | plus_constant (ptr_mode, stack_pointer_rtx, -first)); | |
2660 | ||
2661 | /* LAST_ADDR = SP + FIRST + ROUNDED_SIZE. */ | |
2662 | emit_set_insn (reg2, | |
2663 | plus_constant (ptr_mode, stack_pointer_rtx, | |
2664 | -(first + rounded_size))); | |
2665 | ||
2666 | ||
2667 | /* Step 3: the loop | |
2668 | ||
2669 | do | |
2670 | { | |
2671 | TEST_ADDR = TEST_ADDR + PROBE_INTERVAL | |
2672 | probe at TEST_ADDR | |
2673 | } | |
2674 | while (TEST_ADDR != LAST_ADDR) | |
2675 | ||
2676 | probes at FIRST + N * PROBE_INTERVAL for values of N from 1 | |
2677 | until it is equal to ROUNDED_SIZE. */ | |
2678 | ||
2679 | if (ptr_mode == DImode) | |
2680 | emit_insn (gen_probe_stack_range_di (reg1, reg1, reg2)); | |
2681 | else | |
2682 | emit_insn (gen_probe_stack_range_si (reg1, reg1, reg2)); | |
2683 | ||
2684 | ||
2685 | /* Step 4: probe at FIRST + SIZE if we cannot assert at compile-time | |
2686 | that SIZE is equal to ROUNDED_SIZE. */ | |
2687 | ||
2688 | if (size != rounded_size) | |
2689 | { | |
2690 | HOST_WIDE_INT rem = size - rounded_size; | |
2691 | ||
2692 | if (rem > 256) | |
2693 | { | |
2694 | const HOST_WIDE_INT base = ROUND_UP (rem, ARITH_FACTOR); | |
2695 | ||
2696 | emit_set_insn (reg2, plus_constant (ptr_mode, reg2, -base)); | |
2697 | emit_stack_probe (plus_constant (ptr_mode, reg2, base - rem)); | |
2698 | } | |
2699 | else | |
2700 | emit_stack_probe (plus_constant (ptr_mode, reg2, -rem)); | |
2701 | } | |
2702 | } | |
2703 | ||
2704 | /* Make sure nothing is scheduled before we are done. */ | |
2705 | emit_insn (gen_blockage ()); | |
2706 | } | |
2707 | ||
2708 | /* Probe a range of stack addresses from REG1 to REG2 inclusive. These are | |
2709 | absolute addresses. */ | |
2710 | ||
2711 | const char * | |
2712 | aarch64_output_probe_stack_range (rtx reg1, rtx reg2) | |
2713 | { | |
2714 | static int labelno = 0; | |
2715 | char loop_lab[32]; | |
2716 | rtx xops[2]; | |
2717 | ||
2718 | ASM_GENERATE_INTERNAL_LABEL (loop_lab, "LPSRL", labelno++); | |
2719 | ||
2720 | /* Loop. */ | |
2721 | ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, loop_lab); | |
2722 | ||
2723 | /* TEST_ADDR = TEST_ADDR + PROBE_INTERVAL. */ | |
2724 | xops[0] = reg1; | |
2725 | xops[1] = GEN_INT (PROBE_INTERVAL); | |
2726 | output_asm_insn ("sub\t%0, %0, %1", xops); | |
2727 | ||
2728 | /* Probe at TEST_ADDR. */ | |
2729 | output_asm_insn ("str\txzr, [%0]", xops); | |
2730 | ||
2731 | /* Test if TEST_ADDR == LAST_ADDR. */ | |
2732 | xops[1] = reg2; | |
2733 | output_asm_insn ("cmp\t%0, %1", xops); | |
2734 | ||
2735 | /* Branch. */ | |
2736 | fputs ("\tb.ne\t", asm_out_file); | |
2737 | assemble_name_raw (asm_out_file, loop_lab); | |
2738 | fputc ('\n', asm_out_file); | |
2739 | ||
2740 | return ""; | |
2741 | } | |
2742 | ||
43e9d192 IB |
2743 | static bool |
2744 | aarch64_frame_pointer_required (void) | |
2745 | { | |
0b7f8166 MS |
2746 | /* In aarch64_override_options_after_change |
2747 | flag_omit_leaf_frame_pointer turns off the frame pointer by | |
2748 | default. Turn it back on now if we've not got a leaf | |
2749 | function. */ | |
2750 | if (flag_omit_leaf_frame_pointer | |
2751 | && (!crtl->is_leaf || df_regs_ever_live_p (LR_REGNUM))) | |
2752 | return true; | |
43e9d192 | 2753 | |
0b7f8166 | 2754 | return false; |
43e9d192 IB |
2755 | } |
2756 | ||
2757 | /* Mark the registers that need to be saved by the callee and calculate | |
2758 | the size of the callee-saved registers area and frame record (both FP | |
2759 | and LR may be omitted). */ | |
2760 | static void | |
2761 | aarch64_layout_frame (void) | |
2762 | { | |
2763 | HOST_WIDE_INT offset = 0; | |
4b0685d9 | 2764 | int regno, last_fp_reg = INVALID_REGNUM; |
43e9d192 IB |
2765 | |
2766 | if (reload_completed && cfun->machine->frame.laid_out) | |
2767 | return; | |
2768 | ||
97826595 MS |
2769 | #define SLOT_NOT_REQUIRED (-2) |
2770 | #define SLOT_REQUIRED (-1) | |
2771 | ||
71bfb77a WD |
2772 | cfun->machine->frame.wb_candidate1 = INVALID_REGNUM; |
2773 | cfun->machine->frame.wb_candidate2 = INVALID_REGNUM; | |
363ffa50 | 2774 | |
43e9d192 IB |
2775 | /* First mark all the registers that really need to be saved... */ |
2776 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) | |
97826595 | 2777 | cfun->machine->frame.reg_offset[regno] = SLOT_NOT_REQUIRED; |
43e9d192 IB |
2778 | |
2779 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
97826595 | 2780 | cfun->machine->frame.reg_offset[regno] = SLOT_NOT_REQUIRED; |
43e9d192 IB |
2781 | |
2782 | /* ... that includes the eh data registers (if needed)... */ | |
2783 | if (crtl->calls_eh_return) | |
2784 | for (regno = 0; EH_RETURN_DATA_REGNO (regno) != INVALID_REGNUM; regno++) | |
97826595 MS |
2785 | cfun->machine->frame.reg_offset[EH_RETURN_DATA_REGNO (regno)] |
2786 | = SLOT_REQUIRED; | |
43e9d192 IB |
2787 | |
2788 | /* ... and any callee saved register that dataflow says is live. */ | |
2789 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) | |
2790 | if (df_regs_ever_live_p (regno) | |
1c923b60 JW |
2791 | && (regno == R30_REGNUM |
2792 | || !call_used_regs[regno])) | |
97826595 | 2793 | cfun->machine->frame.reg_offset[regno] = SLOT_REQUIRED; |
43e9d192 IB |
2794 | |
2795 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
2796 | if (df_regs_ever_live_p (regno) | |
2797 | && !call_used_regs[regno]) | |
4b0685d9 WD |
2798 | { |
2799 | cfun->machine->frame.reg_offset[regno] = SLOT_REQUIRED; | |
2800 | last_fp_reg = regno; | |
2801 | } | |
43e9d192 IB |
2802 | |
2803 | if (frame_pointer_needed) | |
2804 | { | |
2e1cdae5 | 2805 | /* FP and LR are placed in the linkage record. */ |
43e9d192 | 2806 | cfun->machine->frame.reg_offset[R29_REGNUM] = 0; |
363ffa50 | 2807 | cfun->machine->frame.wb_candidate1 = R29_REGNUM; |
2e1cdae5 | 2808 | cfun->machine->frame.reg_offset[R30_REGNUM] = UNITS_PER_WORD; |
363ffa50 | 2809 | cfun->machine->frame.wb_candidate2 = R30_REGNUM; |
2e1cdae5 | 2810 | offset += 2 * UNITS_PER_WORD; |
43e9d192 IB |
2811 | } |
2812 | ||
2813 | /* Now assign stack slots for them. */ | |
2e1cdae5 | 2814 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) |
97826595 | 2815 | if (cfun->machine->frame.reg_offset[regno] == SLOT_REQUIRED) |
43e9d192 IB |
2816 | { |
2817 | cfun->machine->frame.reg_offset[regno] = offset; | |
71bfb77a | 2818 | if (cfun->machine->frame.wb_candidate1 == INVALID_REGNUM) |
363ffa50 | 2819 | cfun->machine->frame.wb_candidate1 = regno; |
71bfb77a | 2820 | else if (cfun->machine->frame.wb_candidate2 == INVALID_REGNUM) |
363ffa50 | 2821 | cfun->machine->frame.wb_candidate2 = regno; |
43e9d192 IB |
2822 | offset += UNITS_PER_WORD; |
2823 | } | |
2824 | ||
4b0685d9 WD |
2825 | HOST_WIDE_INT max_int_offset = offset; |
2826 | offset = ROUND_UP (offset, STACK_BOUNDARY / BITS_PER_UNIT); | |
2827 | bool has_align_gap = offset != max_int_offset; | |
2828 | ||
43e9d192 | 2829 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) |
97826595 | 2830 | if (cfun->machine->frame.reg_offset[regno] == SLOT_REQUIRED) |
43e9d192 | 2831 | { |
4b0685d9 WD |
2832 | /* If there is an alignment gap between integer and fp callee-saves, |
2833 | allocate the last fp register to it if possible. */ | |
2834 | if (regno == last_fp_reg && has_align_gap && (offset & 8) == 0) | |
2835 | { | |
2836 | cfun->machine->frame.reg_offset[regno] = max_int_offset; | |
2837 | break; | |
2838 | } | |
2839 | ||
43e9d192 | 2840 | cfun->machine->frame.reg_offset[regno] = offset; |
71bfb77a | 2841 | if (cfun->machine->frame.wb_candidate1 == INVALID_REGNUM) |
363ffa50 | 2842 | cfun->machine->frame.wb_candidate1 = regno; |
71bfb77a | 2843 | else if (cfun->machine->frame.wb_candidate2 == INVALID_REGNUM |
363ffa50 JW |
2844 | && cfun->machine->frame.wb_candidate1 >= V0_REGNUM) |
2845 | cfun->machine->frame.wb_candidate2 = regno; | |
43e9d192 IB |
2846 | offset += UNITS_PER_WORD; |
2847 | } | |
2848 | ||
4f59f9f2 | 2849 | offset = ROUND_UP (offset, STACK_BOUNDARY / BITS_PER_UNIT); |
43e9d192 IB |
2850 | |
2851 | cfun->machine->frame.saved_regs_size = offset; | |
1c960e02 | 2852 | |
71bfb77a WD |
2853 | HOST_WIDE_INT varargs_and_saved_regs_size |
2854 | = offset + cfun->machine->frame.saved_varargs_size; | |
2855 | ||
1c960e02 | 2856 | cfun->machine->frame.hard_fp_offset |
71bfb77a | 2857 | = ROUND_UP (varargs_and_saved_regs_size + get_frame_size (), |
4f59f9f2 | 2858 | STACK_BOUNDARY / BITS_PER_UNIT); |
1c960e02 MS |
2859 | |
2860 | cfun->machine->frame.frame_size | |
4f59f9f2 UB |
2861 | = ROUND_UP (cfun->machine->frame.hard_fp_offset |
2862 | + crtl->outgoing_args_size, | |
2863 | STACK_BOUNDARY / BITS_PER_UNIT); | |
1c960e02 | 2864 | |
71bfb77a WD |
2865 | cfun->machine->frame.locals_offset = cfun->machine->frame.saved_varargs_size; |
2866 | ||
2867 | cfun->machine->frame.initial_adjust = 0; | |
2868 | cfun->machine->frame.final_adjust = 0; | |
2869 | cfun->machine->frame.callee_adjust = 0; | |
2870 | cfun->machine->frame.callee_offset = 0; | |
2871 | ||
2872 | HOST_WIDE_INT max_push_offset = 0; | |
2873 | if (cfun->machine->frame.wb_candidate2 != INVALID_REGNUM) | |
2874 | max_push_offset = 512; | |
2875 | else if (cfun->machine->frame.wb_candidate1 != INVALID_REGNUM) | |
2876 | max_push_offset = 256; | |
2877 | ||
2878 | if (cfun->machine->frame.frame_size < max_push_offset | |
2879 | && crtl->outgoing_args_size == 0) | |
2880 | { | |
2881 | /* Simple, small frame with no outgoing arguments: | |
2882 | stp reg1, reg2, [sp, -frame_size]! | |
2883 | stp reg3, reg4, [sp, 16] */ | |
2884 | cfun->machine->frame.callee_adjust = cfun->machine->frame.frame_size; | |
2885 | } | |
2886 | else if ((crtl->outgoing_args_size | |
2887 | + cfun->machine->frame.saved_regs_size < 512) | |
2888 | && !(cfun->calls_alloca | |
2889 | && cfun->machine->frame.hard_fp_offset < max_push_offset)) | |
2890 | { | |
2891 | /* Frame with small outgoing arguments: | |
2892 | sub sp, sp, frame_size | |
2893 | stp reg1, reg2, [sp, outgoing_args_size] | |
2894 | stp reg3, reg4, [sp, outgoing_args_size + 16] */ | |
2895 | cfun->machine->frame.initial_adjust = cfun->machine->frame.frame_size; | |
2896 | cfun->machine->frame.callee_offset | |
2897 | = cfun->machine->frame.frame_size - cfun->machine->frame.hard_fp_offset; | |
2898 | } | |
2899 | else if (cfun->machine->frame.hard_fp_offset < max_push_offset) | |
2900 | { | |
2901 | /* Frame with large outgoing arguments but a small local area: | |
2902 | stp reg1, reg2, [sp, -hard_fp_offset]! | |
2903 | stp reg3, reg4, [sp, 16] | |
2904 | sub sp, sp, outgoing_args_size */ | |
2905 | cfun->machine->frame.callee_adjust = cfun->machine->frame.hard_fp_offset; | |
2906 | cfun->machine->frame.final_adjust | |
2907 | = cfun->machine->frame.frame_size - cfun->machine->frame.callee_adjust; | |
2908 | } | |
2909 | else if (!frame_pointer_needed | |
2910 | && varargs_and_saved_regs_size < max_push_offset) | |
2911 | { | |
2912 | /* Frame with large local area and outgoing arguments (this pushes the | |
2913 | callee-saves first, followed by the locals and outgoing area): | |
2914 | stp reg1, reg2, [sp, -varargs_and_saved_regs_size]! | |
2915 | stp reg3, reg4, [sp, 16] | |
2916 | sub sp, sp, frame_size - varargs_and_saved_regs_size */ | |
2917 | cfun->machine->frame.callee_adjust = varargs_and_saved_regs_size; | |
2918 | cfun->machine->frame.final_adjust | |
2919 | = cfun->machine->frame.frame_size - cfun->machine->frame.callee_adjust; | |
2920 | cfun->machine->frame.hard_fp_offset = cfun->machine->frame.callee_adjust; | |
2921 | cfun->machine->frame.locals_offset = cfun->machine->frame.hard_fp_offset; | |
2922 | } | |
2923 | else | |
2924 | { | |
2925 | /* Frame with large local area and outgoing arguments using frame pointer: | |
2926 | sub sp, sp, hard_fp_offset | |
2927 | stp x29, x30, [sp, 0] | |
2928 | add x29, sp, 0 | |
2929 | stp reg3, reg4, [sp, 16] | |
2930 | sub sp, sp, outgoing_args_size */ | |
2931 | cfun->machine->frame.initial_adjust = cfun->machine->frame.hard_fp_offset; | |
2932 | cfun->machine->frame.final_adjust | |
2933 | = cfun->machine->frame.frame_size - cfun->machine->frame.initial_adjust; | |
2934 | } | |
2935 | ||
43e9d192 IB |
2936 | cfun->machine->frame.laid_out = true; |
2937 | } | |
2938 | ||
04ddfe06 KT |
2939 | /* Return true if the register REGNO is saved on entry to |
2940 | the current function. */ | |
2941 | ||
43e9d192 IB |
2942 | static bool |
2943 | aarch64_register_saved_on_entry (int regno) | |
2944 | { | |
97826595 | 2945 | return cfun->machine->frame.reg_offset[regno] >= 0; |
43e9d192 IB |
2946 | } |
2947 | ||
04ddfe06 KT |
2948 | /* Return the next register up from REGNO up to LIMIT for the callee |
2949 | to save. */ | |
2950 | ||
64dedd72 JW |
2951 | static unsigned |
2952 | aarch64_next_callee_save (unsigned regno, unsigned limit) | |
2953 | { | |
2954 | while (regno <= limit && !aarch64_register_saved_on_entry (regno)) | |
2955 | regno ++; | |
2956 | return regno; | |
2957 | } | |
43e9d192 | 2958 | |
04ddfe06 KT |
2959 | /* Push the register number REGNO of mode MODE to the stack with write-back |
2960 | adjusting the stack by ADJUSTMENT. */ | |
2961 | ||
c5e1f66e | 2962 | static void |
ef4bddc2 | 2963 | aarch64_pushwb_single_reg (machine_mode mode, unsigned regno, |
c5e1f66e JW |
2964 | HOST_WIDE_INT adjustment) |
2965 | { | |
2966 | rtx base_rtx = stack_pointer_rtx; | |
2967 | rtx insn, reg, mem; | |
2968 | ||
2969 | reg = gen_rtx_REG (mode, regno); | |
2970 | mem = gen_rtx_PRE_MODIFY (Pmode, base_rtx, | |
2971 | plus_constant (Pmode, base_rtx, -adjustment)); | |
2972 | mem = gen_rtx_MEM (mode, mem); | |
2973 | ||
2974 | insn = emit_move_insn (mem, reg); | |
2975 | RTX_FRAME_RELATED_P (insn) = 1; | |
2976 | } | |
2977 | ||
04ddfe06 KT |
2978 | /* Generate and return an instruction to store the pair of registers |
2979 | REG and REG2 of mode MODE to location BASE with write-back adjusting | |
2980 | the stack location BASE by ADJUSTMENT. */ | |
2981 | ||
80c11907 | 2982 | static rtx |
ef4bddc2 | 2983 | aarch64_gen_storewb_pair (machine_mode mode, rtx base, rtx reg, rtx reg2, |
80c11907 JW |
2984 | HOST_WIDE_INT adjustment) |
2985 | { | |
2986 | switch (mode) | |
2987 | { | |
2988 | case DImode: | |
2989 | return gen_storewb_pairdi_di (base, base, reg, reg2, | |
2990 | GEN_INT (-adjustment), | |
2991 | GEN_INT (UNITS_PER_WORD - adjustment)); | |
2992 | case DFmode: | |
2993 | return gen_storewb_pairdf_di (base, base, reg, reg2, | |
2994 | GEN_INT (-adjustment), | |
2995 | GEN_INT (UNITS_PER_WORD - adjustment)); | |
2996 | default: | |
2997 | gcc_unreachable (); | |
2998 | } | |
2999 | } | |
3000 | ||
04ddfe06 KT |
3001 | /* Push registers numbered REGNO1 and REGNO2 to the stack, adjusting the |
3002 | stack pointer by ADJUSTMENT. */ | |
3003 | ||
80c11907 | 3004 | static void |
89ac681e | 3005 | aarch64_push_regs (unsigned regno1, unsigned regno2, HOST_WIDE_INT adjustment) |
80c11907 | 3006 | { |
5d8a22a5 | 3007 | rtx_insn *insn; |
89ac681e WD |
3008 | machine_mode mode = (regno1 <= R30_REGNUM) ? DImode : DFmode; |
3009 | ||
71bfb77a | 3010 | if (regno2 == INVALID_REGNUM) |
89ac681e WD |
3011 | return aarch64_pushwb_single_reg (mode, regno1, adjustment); |
3012 | ||
80c11907 JW |
3013 | rtx reg1 = gen_rtx_REG (mode, regno1); |
3014 | rtx reg2 = gen_rtx_REG (mode, regno2); | |
3015 | ||
3016 | insn = emit_insn (aarch64_gen_storewb_pair (mode, stack_pointer_rtx, reg1, | |
3017 | reg2, adjustment)); | |
3018 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 2)) = 1; | |
80c11907 JW |
3019 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; |
3020 | RTX_FRAME_RELATED_P (insn) = 1; | |
3021 | } | |
3022 | ||
04ddfe06 KT |
3023 | /* Load the pair of register REG, REG2 of mode MODE from stack location BASE, |
3024 | adjusting it by ADJUSTMENT afterwards. */ | |
3025 | ||
159313d9 | 3026 | static rtx |
ef4bddc2 | 3027 | aarch64_gen_loadwb_pair (machine_mode mode, rtx base, rtx reg, rtx reg2, |
159313d9 JW |
3028 | HOST_WIDE_INT adjustment) |
3029 | { | |
3030 | switch (mode) | |
3031 | { | |
3032 | case DImode: | |
3033 | return gen_loadwb_pairdi_di (base, base, reg, reg2, GEN_INT (adjustment), | |
3e322b3f | 3034 | GEN_INT (UNITS_PER_WORD)); |
159313d9 JW |
3035 | case DFmode: |
3036 | return gen_loadwb_pairdf_di (base, base, reg, reg2, GEN_INT (adjustment), | |
3e322b3f | 3037 | GEN_INT (UNITS_PER_WORD)); |
159313d9 JW |
3038 | default: |
3039 | gcc_unreachable (); | |
3040 | } | |
3041 | } | |
3042 | ||
04ddfe06 KT |
3043 | /* Pop the two registers numbered REGNO1, REGNO2 from the stack, adjusting it |
3044 | afterwards by ADJUSTMENT and writing the appropriate REG_CFA_RESTORE notes | |
3045 | into CFI_OPS. */ | |
3046 | ||
89ac681e WD |
3047 | static void |
3048 | aarch64_pop_regs (unsigned regno1, unsigned regno2, HOST_WIDE_INT adjustment, | |
3049 | rtx *cfi_ops) | |
3050 | { | |
3051 | machine_mode mode = (regno1 <= R30_REGNUM) ? DImode : DFmode; | |
3052 | rtx reg1 = gen_rtx_REG (mode, regno1); | |
3053 | ||
3054 | *cfi_ops = alloc_reg_note (REG_CFA_RESTORE, reg1, *cfi_ops); | |
3055 | ||
71bfb77a | 3056 | if (regno2 == INVALID_REGNUM) |
89ac681e WD |
3057 | { |
3058 | rtx mem = plus_constant (Pmode, stack_pointer_rtx, adjustment); | |
3059 | mem = gen_rtx_POST_MODIFY (Pmode, stack_pointer_rtx, mem); | |
3060 | emit_move_insn (reg1, gen_rtx_MEM (mode, mem)); | |
3061 | } | |
3062 | else | |
3063 | { | |
3064 | rtx reg2 = gen_rtx_REG (mode, regno2); | |
3065 | *cfi_ops = alloc_reg_note (REG_CFA_RESTORE, reg2, *cfi_ops); | |
3066 | emit_insn (aarch64_gen_loadwb_pair (mode, stack_pointer_rtx, reg1, | |
3067 | reg2, adjustment)); | |
3068 | } | |
3069 | } | |
3070 | ||
04ddfe06 KT |
3071 | /* Generate and return a store pair instruction of mode MODE to store |
3072 | register REG1 to MEM1 and register REG2 to MEM2. */ | |
3073 | ||
72df5c1f | 3074 | static rtx |
ef4bddc2 | 3075 | aarch64_gen_store_pair (machine_mode mode, rtx mem1, rtx reg1, rtx mem2, |
72df5c1f JW |
3076 | rtx reg2) |
3077 | { | |
3078 | switch (mode) | |
3079 | { | |
3080 | case DImode: | |
3081 | return gen_store_pairdi (mem1, reg1, mem2, reg2); | |
3082 | ||
3083 | case DFmode: | |
3084 | return gen_store_pairdf (mem1, reg1, mem2, reg2); | |
3085 | ||
3086 | default: | |
3087 | gcc_unreachable (); | |
3088 | } | |
3089 | } | |
3090 | ||
04ddfe06 KT |
3091 | /* Generate and regurn a load pair isntruction of mode MODE to load register |
3092 | REG1 from MEM1 and register REG2 from MEM2. */ | |
3093 | ||
72df5c1f | 3094 | static rtx |
ef4bddc2 | 3095 | aarch64_gen_load_pair (machine_mode mode, rtx reg1, rtx mem1, rtx reg2, |
72df5c1f JW |
3096 | rtx mem2) |
3097 | { | |
3098 | switch (mode) | |
3099 | { | |
3100 | case DImode: | |
3101 | return gen_load_pairdi (reg1, mem1, reg2, mem2); | |
3102 | ||
3103 | case DFmode: | |
3104 | return gen_load_pairdf (reg1, mem1, reg2, mem2); | |
3105 | ||
3106 | default: | |
3107 | gcc_unreachable (); | |
3108 | } | |
3109 | } | |
3110 | ||
04ddfe06 KT |
3111 | /* Emit code to save the callee-saved registers from register number START |
3112 | to LIMIT to the stack at the location starting at offset START_OFFSET, | |
3113 | skipping any write-back candidates if SKIP_WB is true. */ | |
43e9d192 | 3114 | |
43e9d192 | 3115 | static void |
ef4bddc2 | 3116 | aarch64_save_callee_saves (machine_mode mode, HOST_WIDE_INT start_offset, |
ae13fce3 | 3117 | unsigned start, unsigned limit, bool skip_wb) |
43e9d192 | 3118 | { |
5d8a22a5 | 3119 | rtx_insn *insn; |
ef4bddc2 | 3120 | rtx (*gen_mem_ref) (machine_mode, rtx) = (frame_pointer_needed |
a007a21c | 3121 | ? gen_frame_mem : gen_rtx_MEM); |
43e9d192 IB |
3122 | unsigned regno; |
3123 | unsigned regno2; | |
3124 | ||
0ec74a1e | 3125 | for (regno = aarch64_next_callee_save (start, limit); |
64dedd72 JW |
3126 | regno <= limit; |
3127 | regno = aarch64_next_callee_save (regno + 1, limit)) | |
43e9d192 | 3128 | { |
ae13fce3 JW |
3129 | rtx reg, mem; |
3130 | HOST_WIDE_INT offset; | |
64dedd72 | 3131 | |
ae13fce3 JW |
3132 | if (skip_wb |
3133 | && (regno == cfun->machine->frame.wb_candidate1 | |
3134 | || regno == cfun->machine->frame.wb_candidate2)) | |
3135 | continue; | |
3136 | ||
3137 | reg = gen_rtx_REG (mode, regno); | |
3138 | offset = start_offset + cfun->machine->frame.reg_offset[regno]; | |
0ec74a1e JW |
3139 | mem = gen_mem_ref (mode, plus_constant (Pmode, stack_pointer_rtx, |
3140 | offset)); | |
64dedd72 JW |
3141 | |
3142 | regno2 = aarch64_next_callee_save (regno + 1, limit); | |
3143 | ||
3144 | if (regno2 <= limit | |
3145 | && ((cfun->machine->frame.reg_offset[regno] + UNITS_PER_WORD) | |
3146 | == cfun->machine->frame.reg_offset[regno2])) | |
3147 | ||
43e9d192 | 3148 | { |
0ec74a1e | 3149 | rtx reg2 = gen_rtx_REG (mode, regno2); |
64dedd72 JW |
3150 | rtx mem2; |
3151 | ||
3152 | offset = start_offset + cfun->machine->frame.reg_offset[regno2]; | |
8ed2fc62 JW |
3153 | mem2 = gen_mem_ref (mode, plus_constant (Pmode, stack_pointer_rtx, |
3154 | offset)); | |
3155 | insn = emit_insn (aarch64_gen_store_pair (mode, mem, reg, mem2, | |
3156 | reg2)); | |
0b4a9743 | 3157 | |
64dedd72 JW |
3158 | /* The first part of a frame-related parallel insn is |
3159 | always assumed to be relevant to the frame | |
3160 | calculations; subsequent parts, are only | |
3161 | frame-related if explicitly marked. */ | |
3162 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; | |
3163 | regno = regno2; | |
3164 | } | |
3165 | else | |
8ed2fc62 JW |
3166 | insn = emit_move_insn (mem, reg); |
3167 | ||
3168 | RTX_FRAME_RELATED_P (insn) = 1; | |
3169 | } | |
3170 | } | |
3171 | ||
04ddfe06 KT |
3172 | /* Emit code to restore the callee registers of mode MODE from register |
3173 | number START up to and including LIMIT. Restore from the stack offset | |
3174 | START_OFFSET, skipping any write-back candidates if SKIP_WB is true. | |
3175 | Write the appropriate REG_CFA_RESTORE notes into CFI_OPS. */ | |
3176 | ||
8ed2fc62 | 3177 | static void |
ef4bddc2 | 3178 | aarch64_restore_callee_saves (machine_mode mode, |
8ed2fc62 | 3179 | HOST_WIDE_INT start_offset, unsigned start, |
dd991abb | 3180 | unsigned limit, bool skip_wb, rtx *cfi_ops) |
8ed2fc62 | 3181 | { |
8ed2fc62 | 3182 | rtx base_rtx = stack_pointer_rtx; |
ef4bddc2 | 3183 | rtx (*gen_mem_ref) (machine_mode, rtx) = (frame_pointer_needed |
8ed2fc62 JW |
3184 | ? gen_frame_mem : gen_rtx_MEM); |
3185 | unsigned regno; | |
3186 | unsigned regno2; | |
3187 | HOST_WIDE_INT offset; | |
3188 | ||
3189 | for (regno = aarch64_next_callee_save (start, limit); | |
3190 | regno <= limit; | |
3191 | regno = aarch64_next_callee_save (regno + 1, limit)) | |
3192 | { | |
ae13fce3 | 3193 | rtx reg, mem; |
8ed2fc62 | 3194 | |
ae13fce3 JW |
3195 | if (skip_wb |
3196 | && (regno == cfun->machine->frame.wb_candidate1 | |
3197 | || regno == cfun->machine->frame.wb_candidate2)) | |
3198 | continue; | |
3199 | ||
3200 | reg = gen_rtx_REG (mode, regno); | |
8ed2fc62 JW |
3201 | offset = start_offset + cfun->machine->frame.reg_offset[regno]; |
3202 | mem = gen_mem_ref (mode, plus_constant (Pmode, base_rtx, offset)); | |
3203 | ||
3204 | regno2 = aarch64_next_callee_save (regno + 1, limit); | |
3205 | ||
3206 | if (regno2 <= limit | |
3207 | && ((cfun->machine->frame.reg_offset[regno] + UNITS_PER_WORD) | |
3208 | == cfun->machine->frame.reg_offset[regno2])) | |
64dedd72 | 3209 | { |
8ed2fc62 JW |
3210 | rtx reg2 = gen_rtx_REG (mode, regno2); |
3211 | rtx mem2; | |
3212 | ||
3213 | offset = start_offset + cfun->machine->frame.reg_offset[regno2]; | |
3214 | mem2 = gen_mem_ref (mode, plus_constant (Pmode, base_rtx, offset)); | |
dd991abb | 3215 | emit_insn (aarch64_gen_load_pair (mode, reg, mem, reg2, mem2)); |
8ed2fc62 | 3216 | |
dd991abb | 3217 | *cfi_ops = alloc_reg_note (REG_CFA_RESTORE, reg2, *cfi_ops); |
8ed2fc62 | 3218 | regno = regno2; |
43e9d192 | 3219 | } |
8ed2fc62 | 3220 | else |
dd991abb RH |
3221 | emit_move_insn (reg, mem); |
3222 | *cfi_ops = alloc_reg_note (REG_CFA_RESTORE, reg, *cfi_ops); | |
43e9d192 | 3223 | } |
43e9d192 IB |
3224 | } |
3225 | ||
3226 | /* AArch64 stack frames generated by this compiler look like: | |
3227 | ||
3228 | +-------------------------------+ | |
3229 | | | | |
3230 | | incoming stack arguments | | |
3231 | | | | |
34834420 MS |
3232 | +-------------------------------+ |
3233 | | | <-- incoming stack pointer (aligned) | |
43e9d192 IB |
3234 | | callee-allocated save area | |
3235 | | for register varargs | | |
3236 | | | | |
34834420 MS |
3237 | +-------------------------------+ |
3238 | | local variables | <-- frame_pointer_rtx | |
43e9d192 IB |
3239 | | | |
3240 | +-------------------------------+ | |
454fdba9 RL |
3241 | | padding0 | \ |
3242 | +-------------------------------+ | | |
454fdba9 | 3243 | | callee-saved registers | | frame.saved_regs_size |
454fdba9 RL |
3244 | +-------------------------------+ | |
3245 | | LR' | | | |
3246 | +-------------------------------+ | | |
34834420 MS |
3247 | | FP' | / <- hard_frame_pointer_rtx (aligned) |
3248 | +-------------------------------+ | |
43e9d192 IB |
3249 | | dynamic allocation | |
3250 | +-------------------------------+ | |
34834420 MS |
3251 | | padding | |
3252 | +-------------------------------+ | |
3253 | | outgoing stack arguments | <-- arg_pointer | |
3254 | | | | |
3255 | +-------------------------------+ | |
3256 | | | <-- stack_pointer_rtx (aligned) | |
43e9d192 | 3257 | |
34834420 MS |
3258 | Dynamic stack allocations via alloca() decrease stack_pointer_rtx |
3259 | but leave frame_pointer_rtx and hard_frame_pointer_rtx | |
3260 | unchanged. */ | |
43e9d192 IB |
3261 | |
3262 | /* Generate the prologue instructions for entry into a function. | |
3263 | Establish the stack frame by decreasing the stack pointer with a | |
3264 | properly calculated size and, if necessary, create a frame record | |
3265 | filled with the values of LR and previous frame pointer. The | |
6991c977 | 3266 | current FP is also set up if it is in use. */ |
43e9d192 IB |
3267 | |
3268 | void | |
3269 | aarch64_expand_prologue (void) | |
3270 | { | |
43e9d192 | 3271 | aarch64_layout_frame (); |
43e9d192 | 3272 | |
71bfb77a WD |
3273 | HOST_WIDE_INT frame_size = cfun->machine->frame.frame_size; |
3274 | HOST_WIDE_INT initial_adjust = cfun->machine->frame.initial_adjust; | |
3275 | HOST_WIDE_INT callee_adjust = cfun->machine->frame.callee_adjust; | |
3276 | HOST_WIDE_INT final_adjust = cfun->machine->frame.final_adjust; | |
3277 | HOST_WIDE_INT callee_offset = cfun->machine->frame.callee_offset; | |
3278 | unsigned reg1 = cfun->machine->frame.wb_candidate1; | |
3279 | unsigned reg2 = cfun->machine->frame.wb_candidate2; | |
3280 | rtx_insn *insn; | |
43e9d192 | 3281 | |
dd991abb RH |
3282 | if (flag_stack_usage_info) |
3283 | current_function_static_stack_size = frame_size; | |
43e9d192 | 3284 | |
a3eb8a52 EB |
3285 | if (flag_stack_check == STATIC_BUILTIN_STACK_CHECK) |
3286 | { | |
3287 | if (crtl->is_leaf && !cfun->calls_alloca) | |
3288 | { | |
3289 | if (frame_size > PROBE_INTERVAL && frame_size > STACK_CHECK_PROTECT) | |
3290 | aarch64_emit_probe_stack_range (STACK_CHECK_PROTECT, | |
3291 | frame_size - STACK_CHECK_PROTECT); | |
3292 | } | |
3293 | else if (frame_size > 0) | |
3294 | aarch64_emit_probe_stack_range (STACK_CHECK_PROTECT, frame_size); | |
3295 | } | |
3296 | ||
5be6b295 | 3297 | aarch64_sub_sp (IP0_REGNUM, initial_adjust, true); |
43e9d192 | 3298 | |
71bfb77a WD |
3299 | if (callee_adjust != 0) |
3300 | aarch64_push_regs (reg1, reg2, callee_adjust); | |
43e9d192 | 3301 | |
71bfb77a | 3302 | if (frame_pointer_needed) |
43e9d192 | 3303 | { |
71bfb77a WD |
3304 | if (callee_adjust == 0) |
3305 | aarch64_save_callee_saves (DImode, callee_offset, R29_REGNUM, | |
3306 | R30_REGNUM, false); | |
3307 | insn = emit_insn (gen_add3_insn (hard_frame_pointer_rtx, | |
3308 | stack_pointer_rtx, | |
3309 | GEN_INT (callee_offset))); | |
3310 | RTX_FRAME_RELATED_P (insn) = 1; | |
3311 | emit_insn (gen_stack_tie (stack_pointer_rtx, hard_frame_pointer_rtx)); | |
43e9d192 | 3312 | } |
71bfb77a WD |
3313 | |
3314 | aarch64_save_callee_saves (DImode, callee_offset, R0_REGNUM, R30_REGNUM, | |
3315 | callee_adjust != 0 || frame_pointer_needed); | |
3316 | aarch64_save_callee_saves (DFmode, callee_offset, V0_REGNUM, V31_REGNUM, | |
3317 | callee_adjust != 0 || frame_pointer_needed); | |
5be6b295 | 3318 | aarch64_sub_sp (IP1_REGNUM, final_adjust, !frame_pointer_needed); |
43e9d192 IB |
3319 | } |
3320 | ||
4f942779 RL |
3321 | /* Return TRUE if we can use a simple_return insn. |
3322 | ||
3323 | This function checks whether the callee saved stack is empty, which | |
3324 | means no restore actions are need. The pro_and_epilogue will use | |
3325 | this to check whether shrink-wrapping opt is feasible. */ | |
3326 | ||
3327 | bool | |
3328 | aarch64_use_return_insn_p (void) | |
3329 | { | |
3330 | if (!reload_completed) | |
3331 | return false; | |
3332 | ||
3333 | if (crtl->profile) | |
3334 | return false; | |
3335 | ||
3336 | aarch64_layout_frame (); | |
3337 | ||
3338 | return cfun->machine->frame.frame_size == 0; | |
3339 | } | |
3340 | ||
71bfb77a WD |
3341 | /* Generate the epilogue instructions for returning from a function. |
3342 | This is almost exactly the reverse of the prolog sequence, except | |
3343 | that we need to insert barriers to avoid scheduling loads that read | |
3344 | from a deallocated stack, and we optimize the unwind records by | |
3345 | emitting them all together if possible. */ | |
43e9d192 IB |
3346 | void |
3347 | aarch64_expand_epilogue (bool for_sibcall) | |
3348 | { | |
43e9d192 | 3349 | aarch64_layout_frame (); |
43e9d192 | 3350 | |
71bfb77a WD |
3351 | HOST_WIDE_INT initial_adjust = cfun->machine->frame.initial_adjust; |
3352 | HOST_WIDE_INT callee_adjust = cfun->machine->frame.callee_adjust; | |
3353 | HOST_WIDE_INT final_adjust = cfun->machine->frame.final_adjust; | |
3354 | HOST_WIDE_INT callee_offset = cfun->machine->frame.callee_offset; | |
3355 | unsigned reg1 = cfun->machine->frame.wb_candidate1; | |
3356 | unsigned reg2 = cfun->machine->frame.wb_candidate2; | |
3357 | rtx cfi_ops = NULL; | |
3358 | rtx_insn *insn; | |
44c0e7b9 | 3359 | |
71bfb77a WD |
3360 | /* We need to add memory barrier to prevent read from deallocated stack. */ |
3361 | bool need_barrier_p = (get_frame_size () | |
3362 | + cfun->machine->frame.saved_varargs_size) != 0; | |
43e9d192 | 3363 | |
71bfb77a WD |
3364 | /* Emit a barrier to prevent loads from a deallocated stack. */ |
3365 | if (final_adjust > crtl->outgoing_args_size || cfun->calls_alloca) | |
43e9d192 | 3366 | { |
71bfb77a WD |
3367 | emit_insn (gen_stack_tie (stack_pointer_rtx, stack_pointer_rtx)); |
3368 | need_barrier_p = false; | |
3369 | } | |
7e8c2bd5 | 3370 | |
71bfb77a WD |
3371 | /* Restore the stack pointer from the frame pointer if it may not |
3372 | be the same as the stack pointer. */ | |
3373 | if (frame_pointer_needed && (final_adjust || cfun->calls_alloca)) | |
3374 | { | |
43e9d192 IB |
3375 | insn = emit_insn (gen_add3_insn (stack_pointer_rtx, |
3376 | hard_frame_pointer_rtx, | |
71bfb77a WD |
3377 | GEN_INT (-callee_offset))); |
3378 | /* If writeback is used when restoring callee-saves, the CFA | |
3379 | is restored on the instruction doing the writeback. */ | |
3380 | RTX_FRAME_RELATED_P (insn) = callee_adjust == 0; | |
43e9d192 | 3381 | } |
71bfb77a | 3382 | else |
5be6b295 | 3383 | aarch64_add_sp (IP1_REGNUM, final_adjust, df_regs_ever_live_p (IP1_REGNUM)); |
43e9d192 | 3384 | |
71bfb77a WD |
3385 | aarch64_restore_callee_saves (DImode, callee_offset, R0_REGNUM, R30_REGNUM, |
3386 | callee_adjust != 0, &cfi_ops); | |
3387 | aarch64_restore_callee_saves (DFmode, callee_offset, V0_REGNUM, V31_REGNUM, | |
3388 | callee_adjust != 0, &cfi_ops); | |
43e9d192 | 3389 | |
71bfb77a WD |
3390 | if (need_barrier_p) |
3391 | emit_insn (gen_stack_tie (stack_pointer_rtx, stack_pointer_rtx)); | |
3392 | ||
3393 | if (callee_adjust != 0) | |
3394 | aarch64_pop_regs (reg1, reg2, callee_adjust, &cfi_ops); | |
3395 | ||
3396 | if (callee_adjust != 0 || initial_adjust > 65536) | |
3397 | { | |
3398 | /* Emit delayed restores and set the CFA to be SP + initial_adjust. */ | |
89ac681e | 3399 | insn = get_last_insn (); |
71bfb77a WD |
3400 | rtx new_cfa = plus_constant (Pmode, stack_pointer_rtx, initial_adjust); |
3401 | REG_NOTES (insn) = alloc_reg_note (REG_CFA_DEF_CFA, new_cfa, cfi_ops); | |
43e9d192 | 3402 | RTX_FRAME_RELATED_P (insn) = 1; |
71bfb77a | 3403 | cfi_ops = NULL; |
43e9d192 IB |
3404 | } |
3405 | ||
5be6b295 | 3406 | aarch64_add_sp (IP0_REGNUM, initial_adjust, df_regs_ever_live_p (IP0_REGNUM)); |
7e8c2bd5 | 3407 | |
71bfb77a WD |
3408 | if (cfi_ops) |
3409 | { | |
3410 | /* Emit delayed restores and reset the CFA to be SP. */ | |
3411 | insn = get_last_insn (); | |
3412 | cfi_ops = alloc_reg_note (REG_CFA_DEF_CFA, stack_pointer_rtx, cfi_ops); | |
3413 | REG_NOTES (insn) = cfi_ops; | |
3414 | RTX_FRAME_RELATED_P (insn) = 1; | |
dd991abb RH |
3415 | } |
3416 | ||
3417 | /* Stack adjustment for exception handler. */ | |
3418 | if (crtl->calls_eh_return) | |
3419 | { | |
3420 | /* We need to unwind the stack by the offset computed by | |
3421 | EH_RETURN_STACKADJ_RTX. We have already reset the CFA | |
3422 | to be SP; letting the CFA move during this adjustment | |
3423 | is just as correct as retaining the CFA from the body | |
3424 | of the function. Therefore, do nothing special. */ | |
3425 | emit_insn (gen_add2_insn (stack_pointer_rtx, EH_RETURN_STACKADJ_RTX)); | |
43e9d192 IB |
3426 | } |
3427 | ||
3428 | emit_use (gen_rtx_REG (DImode, LR_REGNUM)); | |
3429 | if (!for_sibcall) | |
3430 | emit_jump_insn (ret_rtx); | |
3431 | } | |
3432 | ||
3433 | /* Return the place to copy the exception unwinding return address to. | |
3434 | This will probably be a stack slot, but could (in theory be the | |
3435 | return register). */ | |
3436 | rtx | |
3437 | aarch64_final_eh_return_addr (void) | |
3438 | { | |
1c960e02 MS |
3439 | HOST_WIDE_INT fp_offset; |
3440 | ||
43e9d192 | 3441 | aarch64_layout_frame (); |
1c960e02 MS |
3442 | |
3443 | fp_offset = cfun->machine->frame.frame_size | |
3444 | - cfun->machine->frame.hard_fp_offset; | |
43e9d192 IB |
3445 | |
3446 | if (cfun->machine->frame.reg_offset[LR_REGNUM] < 0) | |
3447 | return gen_rtx_REG (DImode, LR_REGNUM); | |
3448 | ||
3449 | /* DSE and CSELIB do not detect an alias between sp+k1 and fp+k2. This can | |
3450 | result in a store to save LR introduced by builtin_eh_return () being | |
3451 | incorrectly deleted because the alias is not detected. | |
3452 | So in the calculation of the address to copy the exception unwinding | |
3453 | return address to, we note 2 cases. | |
3454 | If FP is needed and the fp_offset is 0, it means that SP = FP and hence | |
3455 | we return a SP-relative location since all the addresses are SP-relative | |
3456 | in this case. This prevents the store from being optimized away. | |
3457 | If the fp_offset is not 0, then the addresses will be FP-relative and | |
3458 | therefore we return a FP-relative location. */ | |
3459 | ||
3460 | if (frame_pointer_needed) | |
3461 | { | |
3462 | if (fp_offset) | |
3463 | return gen_frame_mem (DImode, | |
3464 | plus_constant (Pmode, hard_frame_pointer_rtx, UNITS_PER_WORD)); | |
3465 | else | |
3466 | return gen_frame_mem (DImode, | |
3467 | plus_constant (Pmode, stack_pointer_rtx, UNITS_PER_WORD)); | |
3468 | } | |
3469 | ||
3470 | /* If FP is not needed, we calculate the location of LR, which would be | |
3471 | at the top of the saved registers block. */ | |
3472 | ||
3473 | return gen_frame_mem (DImode, | |
3474 | plus_constant (Pmode, | |
3475 | stack_pointer_rtx, | |
3476 | fp_offset | |
3477 | + cfun->machine->frame.saved_regs_size | |
3478 | - 2 * UNITS_PER_WORD)); | |
3479 | } | |
3480 | ||
43e9d192 IB |
3481 | /* Output code to add DELTA to the first argument, and then jump |
3482 | to FUNCTION. Used for C++ multiple inheritance. */ | |
3483 | static void | |
3484 | aarch64_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED, | |
3485 | HOST_WIDE_INT delta, | |
3486 | HOST_WIDE_INT vcall_offset, | |
3487 | tree function) | |
3488 | { | |
3489 | /* The this pointer is always in x0. Note that this differs from | |
3490 | Arm where the this pointer maybe bumped to r1 if r0 is required | |
3491 | to return a pointer to an aggregate. On AArch64 a result value | |
3492 | pointer will be in x8. */ | |
3493 | int this_regno = R0_REGNUM; | |
5d8a22a5 DM |
3494 | rtx this_rtx, temp0, temp1, addr, funexp; |
3495 | rtx_insn *insn; | |
43e9d192 | 3496 | |
75f1d6fc SN |
3497 | reload_completed = 1; |
3498 | emit_note (NOTE_INSN_PROLOGUE_END); | |
43e9d192 IB |
3499 | |
3500 | if (vcall_offset == 0) | |
5be6b295 | 3501 | aarch64_add_constant (Pmode, this_regno, IP1_REGNUM, delta); |
43e9d192 IB |
3502 | else |
3503 | { | |
28514dda | 3504 | gcc_assert ((vcall_offset & (POINTER_BYTES - 1)) == 0); |
43e9d192 | 3505 | |
75f1d6fc SN |
3506 | this_rtx = gen_rtx_REG (Pmode, this_regno); |
3507 | temp0 = gen_rtx_REG (Pmode, IP0_REGNUM); | |
3508 | temp1 = gen_rtx_REG (Pmode, IP1_REGNUM); | |
43e9d192 | 3509 | |
75f1d6fc SN |
3510 | addr = this_rtx; |
3511 | if (delta != 0) | |
3512 | { | |
3513 | if (delta >= -256 && delta < 256) | |
3514 | addr = gen_rtx_PRE_MODIFY (Pmode, this_rtx, | |
3515 | plus_constant (Pmode, this_rtx, delta)); | |
3516 | else | |
5be6b295 | 3517 | aarch64_add_constant (Pmode, this_regno, IP1_REGNUM, delta); |
43e9d192 IB |
3518 | } |
3519 | ||
28514dda YZ |
3520 | if (Pmode == ptr_mode) |
3521 | aarch64_emit_move (temp0, gen_rtx_MEM (ptr_mode, addr)); | |
3522 | else | |
3523 | aarch64_emit_move (temp0, | |
3524 | gen_rtx_ZERO_EXTEND (Pmode, | |
3525 | gen_rtx_MEM (ptr_mode, addr))); | |
75f1d6fc | 3526 | |
28514dda | 3527 | if (vcall_offset >= -256 && vcall_offset < 4096 * POINTER_BYTES) |
75f1d6fc | 3528 | addr = plus_constant (Pmode, temp0, vcall_offset); |
43e9d192 IB |
3529 | else |
3530 | { | |
f43657b4 JW |
3531 | aarch64_internal_mov_immediate (temp1, GEN_INT (vcall_offset), true, |
3532 | Pmode); | |
75f1d6fc | 3533 | addr = gen_rtx_PLUS (Pmode, temp0, temp1); |
43e9d192 IB |
3534 | } |
3535 | ||
28514dda YZ |
3536 | if (Pmode == ptr_mode) |
3537 | aarch64_emit_move (temp1, gen_rtx_MEM (ptr_mode,addr)); | |
3538 | else | |
3539 | aarch64_emit_move (temp1, | |
3540 | gen_rtx_SIGN_EXTEND (Pmode, | |
3541 | gen_rtx_MEM (ptr_mode, addr))); | |
3542 | ||
75f1d6fc | 3543 | emit_insn (gen_add2_insn (this_rtx, temp1)); |
43e9d192 IB |
3544 | } |
3545 | ||
75f1d6fc SN |
3546 | /* Generate a tail call to the target function. */ |
3547 | if (!TREE_USED (function)) | |
3548 | { | |
3549 | assemble_external (function); | |
3550 | TREE_USED (function) = 1; | |
3551 | } | |
3552 | funexp = XEXP (DECL_RTL (function), 0); | |
3553 | funexp = gen_rtx_MEM (FUNCTION_MODE, funexp); | |
3554 | insn = emit_call_insn (gen_sibcall (funexp, const0_rtx, NULL_RTX)); | |
3555 | SIBLING_CALL_P (insn) = 1; | |
3556 | ||
3557 | insn = get_insns (); | |
3558 | shorten_branches (insn); | |
3559 | final_start_function (insn, file, 1); | |
3560 | final (insn, file, 1); | |
43e9d192 | 3561 | final_end_function (); |
75f1d6fc SN |
3562 | |
3563 | /* Stop pretending to be a post-reload pass. */ | |
3564 | reload_completed = 0; | |
43e9d192 IB |
3565 | } |
3566 | ||
43e9d192 IB |
3567 | static bool |
3568 | aarch64_tls_referenced_p (rtx x) | |
3569 | { | |
3570 | if (!TARGET_HAVE_TLS) | |
3571 | return false; | |
e7de8563 RS |
3572 | subrtx_iterator::array_type array; |
3573 | FOR_EACH_SUBRTX (iter, array, x, ALL) | |
3574 | { | |
3575 | const_rtx x = *iter; | |
3576 | if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (x) != 0) | |
3577 | return true; | |
3578 | /* Don't recurse into UNSPEC_TLS looking for TLS symbols; these are | |
3579 | TLS offsets, not real symbol references. */ | |
3580 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
3581 | iter.skip_subrtxes (); | |
3582 | } | |
3583 | return false; | |
43e9d192 IB |
3584 | } |
3585 | ||
3586 | ||
43e9d192 IB |
3587 | /* Return true if val can be encoded as a 12-bit unsigned immediate with |
3588 | a left shift of 0 or 12 bits. */ | |
3589 | bool | |
3590 | aarch64_uimm12_shift (HOST_WIDE_INT val) | |
3591 | { | |
3592 | return ((val & (((HOST_WIDE_INT) 0xfff) << 0)) == val | |
3593 | || (val & (((HOST_WIDE_INT) 0xfff) << 12)) == val | |
3594 | ); | |
3595 | } | |
3596 | ||
3597 | ||
3598 | /* Return true if val is an immediate that can be loaded into a | |
3599 | register by a MOVZ instruction. */ | |
3600 | static bool | |
ef4bddc2 | 3601 | aarch64_movw_imm (HOST_WIDE_INT val, machine_mode mode) |
43e9d192 IB |
3602 | { |
3603 | if (GET_MODE_SIZE (mode) > 4) | |
3604 | { | |
3605 | if ((val & (((HOST_WIDE_INT) 0xffff) << 32)) == val | |
3606 | || (val & (((HOST_WIDE_INT) 0xffff) << 48)) == val) | |
3607 | return 1; | |
3608 | } | |
3609 | else | |
3610 | { | |
3611 | /* Ignore sign extension. */ | |
3612 | val &= (HOST_WIDE_INT) 0xffffffff; | |
3613 | } | |
3614 | return ((val & (((HOST_WIDE_INT) 0xffff) << 0)) == val | |
3615 | || (val & (((HOST_WIDE_INT) 0xffff) << 16)) == val); | |
3616 | } | |
3617 | ||
a64c73a2 WD |
3618 | /* Multipliers for repeating bitmasks of width 32, 16, 8, 4, and 2. */ |
3619 | ||
3620 | static const unsigned HOST_WIDE_INT bitmask_imm_mul[] = | |
3621 | { | |
3622 | 0x0000000100000001ull, | |
3623 | 0x0001000100010001ull, | |
3624 | 0x0101010101010101ull, | |
3625 | 0x1111111111111111ull, | |
3626 | 0x5555555555555555ull, | |
3627 | }; | |
3628 | ||
43e9d192 IB |
3629 | |
3630 | /* Return true if val is a valid bitmask immediate. */ | |
a64c73a2 | 3631 | |
43e9d192 | 3632 | bool |
a64c73a2 | 3633 | aarch64_bitmask_imm (HOST_WIDE_INT val_in, machine_mode mode) |
43e9d192 | 3634 | { |
a64c73a2 WD |
3635 | unsigned HOST_WIDE_INT val, tmp, mask, first_one, next_one; |
3636 | int bits; | |
3637 | ||
3638 | /* Check for a single sequence of one bits and return quickly if so. | |
3639 | The special cases of all ones and all zeroes returns false. */ | |
3640 | val = (unsigned HOST_WIDE_INT) val_in; | |
3641 | tmp = val + (val & -val); | |
3642 | ||
3643 | if (tmp == (tmp & -tmp)) | |
3644 | return (val + 1) > 1; | |
3645 | ||
3646 | /* Replicate 32-bit immediates so we can treat them as 64-bit. */ | |
3647 | if (mode == SImode) | |
3648 | val = (val << 32) | (val & 0xffffffff); | |
3649 | ||
3650 | /* Invert if the immediate doesn't start with a zero bit - this means we | |
3651 | only need to search for sequences of one bits. */ | |
3652 | if (val & 1) | |
3653 | val = ~val; | |
3654 | ||
3655 | /* Find the first set bit and set tmp to val with the first sequence of one | |
3656 | bits removed. Return success if there is a single sequence of ones. */ | |
3657 | first_one = val & -val; | |
3658 | tmp = val & (val + first_one); | |
3659 | ||
3660 | if (tmp == 0) | |
3661 | return true; | |
3662 | ||
3663 | /* Find the next set bit and compute the difference in bit position. */ | |
3664 | next_one = tmp & -tmp; | |
3665 | bits = clz_hwi (first_one) - clz_hwi (next_one); | |
3666 | mask = val ^ tmp; | |
3667 | ||
3668 | /* Check the bit position difference is a power of 2, and that the first | |
3669 | sequence of one bits fits within 'bits' bits. */ | |
3670 | if ((mask >> bits) != 0 || bits != (bits & -bits)) | |
3671 | return false; | |
3672 | ||
3673 | /* Check the sequence of one bits is repeated 64/bits times. */ | |
3674 | return val == mask * bitmask_imm_mul[__builtin_clz (bits) - 26]; | |
43e9d192 IB |
3675 | } |
3676 | ||
3677 | ||
3678 | /* Return true if val is an immediate that can be loaded into a | |
3679 | register in a single instruction. */ | |
3680 | bool | |
ef4bddc2 | 3681 | aarch64_move_imm (HOST_WIDE_INT val, machine_mode mode) |
43e9d192 IB |
3682 | { |
3683 | if (aarch64_movw_imm (val, mode) || aarch64_movw_imm (~val, mode)) | |
3684 | return 1; | |
3685 | return aarch64_bitmask_imm (val, mode); | |
3686 | } | |
3687 | ||
3688 | static bool | |
ef4bddc2 | 3689 | aarch64_cannot_force_const_mem (machine_mode mode ATTRIBUTE_UNUSED, rtx x) |
43e9d192 IB |
3690 | { |
3691 | rtx base, offset; | |
7eda14e1 | 3692 | |
43e9d192 IB |
3693 | if (GET_CODE (x) == HIGH) |
3694 | return true; | |
3695 | ||
3696 | split_const (x, &base, &offset); | |
3697 | if (GET_CODE (base) == SYMBOL_REF || GET_CODE (base) == LABEL_REF) | |
28514dda | 3698 | { |
a6e0bfa7 | 3699 | if (aarch64_classify_symbol (base, offset) |
28514dda YZ |
3700 | != SYMBOL_FORCE_TO_MEM) |
3701 | return true; | |
3702 | else | |
3703 | /* Avoid generating a 64-bit relocation in ILP32; leave | |
3704 | to aarch64_expand_mov_immediate to handle it properly. */ | |
3705 | return mode != ptr_mode; | |
3706 | } | |
43e9d192 IB |
3707 | |
3708 | return aarch64_tls_referenced_p (x); | |
3709 | } | |
3710 | ||
e79136e4 WD |
3711 | /* Implement TARGET_CASE_VALUES_THRESHOLD. |
3712 | The expansion for a table switch is quite expensive due to the number | |
3713 | of instructions, the table lookup and hard to predict indirect jump. | |
3714 | When optimizing for speed, and -O3 enabled, use the per-core tuning if | |
3715 | set, otherwise use tables for > 16 cases as a tradeoff between size and | |
3716 | performance. When optimizing for size, use the default setting. */ | |
50487d79 EM |
3717 | |
3718 | static unsigned int | |
3719 | aarch64_case_values_threshold (void) | |
3720 | { | |
3721 | /* Use the specified limit for the number of cases before using jump | |
3722 | tables at higher optimization levels. */ | |
3723 | if (optimize > 2 | |
3724 | && selected_cpu->tune->max_case_values != 0) | |
3725 | return selected_cpu->tune->max_case_values; | |
3726 | else | |
e79136e4 | 3727 | return optimize_size ? default_case_values_threshold () : 17; |
50487d79 EM |
3728 | } |
3729 | ||
43e9d192 IB |
3730 | /* Return true if register REGNO is a valid index register. |
3731 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
3732 | ||
3733 | bool | |
3734 | aarch64_regno_ok_for_index_p (int regno, bool strict_p) | |
3735 | { | |
3736 | if (!HARD_REGISTER_NUM_P (regno)) | |
3737 | { | |
3738 | if (!strict_p) | |
3739 | return true; | |
3740 | ||
3741 | if (!reg_renumber) | |
3742 | return false; | |
3743 | ||
3744 | regno = reg_renumber[regno]; | |
3745 | } | |
3746 | return GP_REGNUM_P (regno); | |
3747 | } | |
3748 | ||
3749 | /* Return true if register REGNO is a valid base register for mode MODE. | |
3750 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
3751 | ||
3752 | bool | |
3753 | aarch64_regno_ok_for_base_p (int regno, bool strict_p) | |
3754 | { | |
3755 | if (!HARD_REGISTER_NUM_P (regno)) | |
3756 | { | |
3757 | if (!strict_p) | |
3758 | return true; | |
3759 | ||
3760 | if (!reg_renumber) | |
3761 | return false; | |
3762 | ||
3763 | regno = reg_renumber[regno]; | |
3764 | } | |
3765 | ||
3766 | /* The fake registers will be eliminated to either the stack or | |
3767 | hard frame pointer, both of which are usually valid base registers. | |
3768 | Reload deals with the cases where the eliminated form isn't valid. */ | |
3769 | return (GP_REGNUM_P (regno) | |
3770 | || regno == SP_REGNUM | |
3771 | || regno == FRAME_POINTER_REGNUM | |
3772 | || regno == ARG_POINTER_REGNUM); | |
3773 | } | |
3774 | ||
3775 | /* Return true if X is a valid base register for mode MODE. | |
3776 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
3777 | ||
3778 | static bool | |
3779 | aarch64_base_register_rtx_p (rtx x, bool strict_p) | |
3780 | { | |
3781 | if (!strict_p && GET_CODE (x) == SUBREG) | |
3782 | x = SUBREG_REG (x); | |
3783 | ||
3784 | return (REG_P (x) && aarch64_regno_ok_for_base_p (REGNO (x), strict_p)); | |
3785 | } | |
3786 | ||
3787 | /* Return true if address offset is a valid index. If it is, fill in INFO | |
3788 | appropriately. STRICT_P is true if REG_OK_STRICT is in effect. */ | |
3789 | ||
3790 | static bool | |
3791 | aarch64_classify_index (struct aarch64_address_info *info, rtx x, | |
ef4bddc2 | 3792 | machine_mode mode, bool strict_p) |
43e9d192 IB |
3793 | { |
3794 | enum aarch64_address_type type; | |
3795 | rtx index; | |
3796 | int shift; | |
3797 | ||
3798 | /* (reg:P) */ | |
3799 | if ((REG_P (x) || GET_CODE (x) == SUBREG) | |
3800 | && GET_MODE (x) == Pmode) | |
3801 | { | |
3802 | type = ADDRESS_REG_REG; | |
3803 | index = x; | |
3804 | shift = 0; | |
3805 | } | |
3806 | /* (sign_extend:DI (reg:SI)) */ | |
3807 | else if ((GET_CODE (x) == SIGN_EXTEND | |
3808 | || GET_CODE (x) == ZERO_EXTEND) | |
3809 | && GET_MODE (x) == DImode | |
3810 | && GET_MODE (XEXP (x, 0)) == SImode) | |
3811 | { | |
3812 | type = (GET_CODE (x) == SIGN_EXTEND) | |
3813 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3814 | index = XEXP (x, 0); | |
3815 | shift = 0; | |
3816 | } | |
3817 | /* (mult:DI (sign_extend:DI (reg:SI)) (const_int scale)) */ | |
3818 | else if (GET_CODE (x) == MULT | |
3819 | && (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
3820 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
3821 | && GET_MODE (XEXP (x, 0)) == DImode | |
3822 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == SImode | |
3823 | && CONST_INT_P (XEXP (x, 1))) | |
3824 | { | |
3825 | type = (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
3826 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3827 | index = XEXP (XEXP (x, 0), 0); | |
3828 | shift = exact_log2 (INTVAL (XEXP (x, 1))); | |
3829 | } | |
3830 | /* (ashift:DI (sign_extend:DI (reg:SI)) (const_int shift)) */ | |
3831 | else if (GET_CODE (x) == ASHIFT | |
3832 | && (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
3833 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
3834 | && GET_MODE (XEXP (x, 0)) == DImode | |
3835 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == SImode | |
3836 | && CONST_INT_P (XEXP (x, 1))) | |
3837 | { | |
3838 | type = (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
3839 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3840 | index = XEXP (XEXP (x, 0), 0); | |
3841 | shift = INTVAL (XEXP (x, 1)); | |
3842 | } | |
3843 | /* (sign_extract:DI (mult:DI (reg:DI) (const_int scale)) 32+shift 0) */ | |
3844 | else if ((GET_CODE (x) == SIGN_EXTRACT | |
3845 | || GET_CODE (x) == ZERO_EXTRACT) | |
3846 | && GET_MODE (x) == DImode | |
3847 | && GET_CODE (XEXP (x, 0)) == MULT | |
3848 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3849 | && CONST_INT_P (XEXP (XEXP (x, 0), 1))) | |
3850 | { | |
3851 | type = (GET_CODE (x) == SIGN_EXTRACT) | |
3852 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3853 | index = XEXP (XEXP (x, 0), 0); | |
3854 | shift = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1))); | |
3855 | if (INTVAL (XEXP (x, 1)) != 32 + shift | |
3856 | || INTVAL (XEXP (x, 2)) != 0) | |
3857 | shift = -1; | |
3858 | } | |
3859 | /* (and:DI (mult:DI (reg:DI) (const_int scale)) | |
3860 | (const_int 0xffffffff<<shift)) */ | |
3861 | else if (GET_CODE (x) == AND | |
3862 | && GET_MODE (x) == DImode | |
3863 | && GET_CODE (XEXP (x, 0)) == MULT | |
3864 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3865 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
3866 | && CONST_INT_P (XEXP (x, 1))) | |
3867 | { | |
3868 | type = ADDRESS_REG_UXTW; | |
3869 | index = XEXP (XEXP (x, 0), 0); | |
3870 | shift = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1))); | |
3871 | if (INTVAL (XEXP (x, 1)) != (HOST_WIDE_INT)0xffffffff << shift) | |
3872 | shift = -1; | |
3873 | } | |
3874 | /* (sign_extract:DI (ashift:DI (reg:DI) (const_int shift)) 32+shift 0) */ | |
3875 | else if ((GET_CODE (x) == SIGN_EXTRACT | |
3876 | || GET_CODE (x) == ZERO_EXTRACT) | |
3877 | && GET_MODE (x) == DImode | |
3878 | && GET_CODE (XEXP (x, 0)) == ASHIFT | |
3879 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3880 | && CONST_INT_P (XEXP (XEXP (x, 0), 1))) | |
3881 | { | |
3882 | type = (GET_CODE (x) == SIGN_EXTRACT) | |
3883 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3884 | index = XEXP (XEXP (x, 0), 0); | |
3885 | shift = INTVAL (XEXP (XEXP (x, 0), 1)); | |
3886 | if (INTVAL (XEXP (x, 1)) != 32 + shift | |
3887 | || INTVAL (XEXP (x, 2)) != 0) | |
3888 | shift = -1; | |
3889 | } | |
3890 | /* (and:DI (ashift:DI (reg:DI) (const_int shift)) | |
3891 | (const_int 0xffffffff<<shift)) */ | |
3892 | else if (GET_CODE (x) == AND | |
3893 | && GET_MODE (x) == DImode | |
3894 | && GET_CODE (XEXP (x, 0)) == ASHIFT | |
3895 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3896 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
3897 | && CONST_INT_P (XEXP (x, 1))) | |
3898 | { | |
3899 | type = ADDRESS_REG_UXTW; | |
3900 | index = XEXP (XEXP (x, 0), 0); | |
3901 | shift = INTVAL (XEXP (XEXP (x, 0), 1)); | |
3902 | if (INTVAL (XEXP (x, 1)) != (HOST_WIDE_INT)0xffffffff << shift) | |
3903 | shift = -1; | |
3904 | } | |
3905 | /* (mult:P (reg:P) (const_int scale)) */ | |
3906 | else if (GET_CODE (x) == MULT | |
3907 | && GET_MODE (x) == Pmode | |
3908 | && GET_MODE (XEXP (x, 0)) == Pmode | |
3909 | && CONST_INT_P (XEXP (x, 1))) | |
3910 | { | |
3911 | type = ADDRESS_REG_REG; | |
3912 | index = XEXP (x, 0); | |
3913 | shift = exact_log2 (INTVAL (XEXP (x, 1))); | |
3914 | } | |
3915 | /* (ashift:P (reg:P) (const_int shift)) */ | |
3916 | else if (GET_CODE (x) == ASHIFT | |
3917 | && GET_MODE (x) == Pmode | |
3918 | && GET_MODE (XEXP (x, 0)) == Pmode | |
3919 | && CONST_INT_P (XEXP (x, 1))) | |
3920 | { | |
3921 | type = ADDRESS_REG_REG; | |
3922 | index = XEXP (x, 0); | |
3923 | shift = INTVAL (XEXP (x, 1)); | |
3924 | } | |
3925 | else | |
3926 | return false; | |
3927 | ||
3928 | if (GET_CODE (index) == SUBREG) | |
3929 | index = SUBREG_REG (index); | |
3930 | ||
3931 | if ((shift == 0 || | |
3932 | (shift > 0 && shift <= 3 | |
3933 | && (1 << shift) == GET_MODE_SIZE (mode))) | |
3934 | && REG_P (index) | |
3935 | && aarch64_regno_ok_for_index_p (REGNO (index), strict_p)) | |
3936 | { | |
3937 | info->type = type; | |
3938 | info->offset = index; | |
3939 | info->shift = shift; | |
3940 | return true; | |
3941 | } | |
3942 | ||
3943 | return false; | |
3944 | } | |
3945 | ||
44707478 | 3946 | bool |
ef4bddc2 | 3947 | aarch64_offset_7bit_signed_scaled_p (machine_mode mode, HOST_WIDE_INT offset) |
43e9d192 IB |
3948 | { |
3949 | return (offset >= -64 * GET_MODE_SIZE (mode) | |
3950 | && offset < 64 * GET_MODE_SIZE (mode) | |
3951 | && offset % GET_MODE_SIZE (mode) == 0); | |
3952 | } | |
3953 | ||
3954 | static inline bool | |
ef4bddc2 | 3955 | offset_9bit_signed_unscaled_p (machine_mode mode ATTRIBUTE_UNUSED, |
43e9d192 IB |
3956 | HOST_WIDE_INT offset) |
3957 | { | |
3958 | return offset >= -256 && offset < 256; | |
3959 | } | |
3960 | ||
3961 | static inline bool | |
ef4bddc2 | 3962 | offset_12bit_unsigned_scaled_p (machine_mode mode, HOST_WIDE_INT offset) |
43e9d192 IB |
3963 | { |
3964 | return (offset >= 0 | |
3965 | && offset < 4096 * GET_MODE_SIZE (mode) | |
3966 | && offset % GET_MODE_SIZE (mode) == 0); | |
3967 | } | |
3968 | ||
abc52318 KT |
3969 | /* Return true if MODE is one of the modes for which we |
3970 | support LDP/STP operations. */ | |
3971 | ||
3972 | static bool | |
3973 | aarch64_mode_valid_for_sched_fusion_p (machine_mode mode) | |
3974 | { | |
3975 | return mode == SImode || mode == DImode | |
3976 | || mode == SFmode || mode == DFmode | |
3977 | || (aarch64_vector_mode_supported_p (mode) | |
3978 | && GET_MODE_SIZE (mode) == 8); | |
3979 | } | |
3980 | ||
9e0218fc RH |
3981 | /* Return true if REGNO is a virtual pointer register, or an eliminable |
3982 | "soft" frame register. Like REGNO_PTR_FRAME_P except that we don't | |
3983 | include stack_pointer or hard_frame_pointer. */ | |
3984 | static bool | |
3985 | virt_or_elim_regno_p (unsigned regno) | |
3986 | { | |
3987 | return ((regno >= FIRST_VIRTUAL_REGISTER | |
3988 | && regno <= LAST_VIRTUAL_POINTER_REGISTER) | |
3989 | || regno == FRAME_POINTER_REGNUM | |
3990 | || regno == ARG_POINTER_REGNUM); | |
3991 | } | |
3992 | ||
43e9d192 IB |
3993 | /* Return true if X is a valid address for machine mode MODE. If it is, |
3994 | fill in INFO appropriately. STRICT_P is true if REG_OK_STRICT is in | |
3995 | effect. OUTER_CODE is PARALLEL for a load/store pair. */ | |
3996 | ||
3997 | static bool | |
3998 | aarch64_classify_address (struct aarch64_address_info *info, | |
ef4bddc2 | 3999 | rtx x, machine_mode mode, |
43e9d192 IB |
4000 | RTX_CODE outer_code, bool strict_p) |
4001 | { | |
4002 | enum rtx_code code = GET_CODE (x); | |
4003 | rtx op0, op1; | |
2d8c6dc1 AH |
4004 | |
4005 | /* On BE, we use load/store pair for all large int mode load/stores. */ | |
4006 | bool load_store_pair_p = (outer_code == PARALLEL | |
4007 | || (BYTES_BIG_ENDIAN | |
4008 | && aarch64_vect_struct_mode_p (mode))); | |
4009 | ||
43e9d192 | 4010 | bool allow_reg_index_p = |
2d8c6dc1 AH |
4011 | !load_store_pair_p |
4012 | && (GET_MODE_SIZE (mode) != 16 || aarch64_vector_mode_supported_p (mode)) | |
4013 | && !aarch64_vect_struct_mode_p (mode); | |
4014 | ||
4015 | /* On LE, for AdvSIMD, don't support anything other than POST_INC or | |
4016 | REG addressing. */ | |
4017 | if (aarch64_vect_struct_mode_p (mode) && !BYTES_BIG_ENDIAN | |
43e9d192 IB |
4018 | && (code != POST_INC && code != REG)) |
4019 | return false; | |
4020 | ||
4021 | switch (code) | |
4022 | { | |
4023 | case REG: | |
4024 | case SUBREG: | |
4025 | info->type = ADDRESS_REG_IMM; | |
4026 | info->base = x; | |
4027 | info->offset = const0_rtx; | |
4028 | return aarch64_base_register_rtx_p (x, strict_p); | |
4029 | ||
4030 | case PLUS: | |
4031 | op0 = XEXP (x, 0); | |
4032 | op1 = XEXP (x, 1); | |
15c0c5c9 JW |
4033 | |
4034 | if (! strict_p | |
4aa81c2e | 4035 | && REG_P (op0) |
9e0218fc | 4036 | && virt_or_elim_regno_p (REGNO (op0)) |
4aa81c2e | 4037 | && CONST_INT_P (op1)) |
15c0c5c9 JW |
4038 | { |
4039 | info->type = ADDRESS_REG_IMM; | |
4040 | info->base = op0; | |
4041 | info->offset = op1; | |
4042 | ||
4043 | return true; | |
4044 | } | |
4045 | ||
43e9d192 IB |
4046 | if (GET_MODE_SIZE (mode) != 0 |
4047 | && CONST_INT_P (op1) | |
4048 | && aarch64_base_register_rtx_p (op0, strict_p)) | |
4049 | { | |
4050 | HOST_WIDE_INT offset = INTVAL (op1); | |
4051 | ||
4052 | info->type = ADDRESS_REG_IMM; | |
4053 | info->base = op0; | |
4054 | info->offset = op1; | |
4055 | ||
4056 | /* TImode and TFmode values are allowed in both pairs of X | |
4057 | registers and individual Q registers. The available | |
4058 | address modes are: | |
4059 | X,X: 7-bit signed scaled offset | |
4060 | Q: 9-bit signed offset | |
4061 | We conservatively require an offset representable in either mode. | |
8ed49fab KT |
4062 | When performing the check for pairs of X registers i.e. LDP/STP |
4063 | pass down DImode since that is the natural size of the LDP/STP | |
4064 | instruction memory accesses. */ | |
43e9d192 | 4065 | if (mode == TImode || mode == TFmode) |
8ed49fab | 4066 | return (aarch64_offset_7bit_signed_scaled_p (DImode, offset) |
43e9d192 IB |
4067 | && offset_9bit_signed_unscaled_p (mode, offset)); |
4068 | ||
2d8c6dc1 AH |
4069 | /* A 7bit offset check because OImode will emit a ldp/stp |
4070 | instruction (only big endian will get here). | |
4071 | For ldp/stp instructions, the offset is scaled for the size of a | |
4072 | single element of the pair. */ | |
4073 | if (mode == OImode) | |
4074 | return aarch64_offset_7bit_signed_scaled_p (TImode, offset); | |
4075 | ||
4076 | /* Three 9/12 bit offsets checks because CImode will emit three | |
4077 | ldr/str instructions (only big endian will get here). */ | |
4078 | if (mode == CImode) | |
4079 | return (aarch64_offset_7bit_signed_scaled_p (TImode, offset) | |
4080 | && (offset_9bit_signed_unscaled_p (V16QImode, offset + 32) | |
4081 | || offset_12bit_unsigned_scaled_p (V16QImode, | |
4082 | offset + 32))); | |
4083 | ||
4084 | /* Two 7bit offsets checks because XImode will emit two ldp/stp | |
4085 | instructions (only big endian will get here). */ | |
4086 | if (mode == XImode) | |
4087 | return (aarch64_offset_7bit_signed_scaled_p (TImode, offset) | |
4088 | && aarch64_offset_7bit_signed_scaled_p (TImode, | |
4089 | offset + 32)); | |
4090 | ||
4091 | if (load_store_pair_p) | |
43e9d192 | 4092 | return ((GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) |
44707478 | 4093 | && aarch64_offset_7bit_signed_scaled_p (mode, offset)); |
43e9d192 IB |
4094 | else |
4095 | return (offset_9bit_signed_unscaled_p (mode, offset) | |
4096 | || offset_12bit_unsigned_scaled_p (mode, offset)); | |
4097 | } | |
4098 | ||
4099 | if (allow_reg_index_p) | |
4100 | { | |
4101 | /* Look for base + (scaled/extended) index register. */ | |
4102 | if (aarch64_base_register_rtx_p (op0, strict_p) | |
4103 | && aarch64_classify_index (info, op1, mode, strict_p)) | |
4104 | { | |
4105 | info->base = op0; | |
4106 | return true; | |
4107 | } | |
4108 | if (aarch64_base_register_rtx_p (op1, strict_p) | |
4109 | && aarch64_classify_index (info, op0, mode, strict_p)) | |
4110 | { | |
4111 | info->base = op1; | |
4112 | return true; | |
4113 | } | |
4114 | } | |
4115 | ||
4116 | return false; | |
4117 | ||
4118 | case POST_INC: | |
4119 | case POST_DEC: | |
4120 | case PRE_INC: | |
4121 | case PRE_DEC: | |
4122 | info->type = ADDRESS_REG_WB; | |
4123 | info->base = XEXP (x, 0); | |
4124 | info->offset = NULL_RTX; | |
4125 | return aarch64_base_register_rtx_p (info->base, strict_p); | |
4126 | ||
4127 | case POST_MODIFY: | |
4128 | case PRE_MODIFY: | |
4129 | info->type = ADDRESS_REG_WB; | |
4130 | info->base = XEXP (x, 0); | |
4131 | if (GET_CODE (XEXP (x, 1)) == PLUS | |
4132 | && CONST_INT_P (XEXP (XEXP (x, 1), 1)) | |
4133 | && rtx_equal_p (XEXP (XEXP (x, 1), 0), info->base) | |
4134 | && aarch64_base_register_rtx_p (info->base, strict_p)) | |
4135 | { | |
4136 | HOST_WIDE_INT offset; | |
4137 | info->offset = XEXP (XEXP (x, 1), 1); | |
4138 | offset = INTVAL (info->offset); | |
4139 | ||
4140 | /* TImode and TFmode values are allowed in both pairs of X | |
4141 | registers and individual Q registers. The available | |
4142 | address modes are: | |
4143 | X,X: 7-bit signed scaled offset | |
4144 | Q: 9-bit signed offset | |
4145 | We conservatively require an offset representable in either mode. | |
4146 | */ | |
4147 | if (mode == TImode || mode == TFmode) | |
44707478 | 4148 | return (aarch64_offset_7bit_signed_scaled_p (mode, offset) |
43e9d192 IB |
4149 | && offset_9bit_signed_unscaled_p (mode, offset)); |
4150 | ||
2d8c6dc1 | 4151 | if (load_store_pair_p) |
43e9d192 | 4152 | return ((GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) |
44707478 | 4153 | && aarch64_offset_7bit_signed_scaled_p (mode, offset)); |
43e9d192 IB |
4154 | else |
4155 | return offset_9bit_signed_unscaled_p (mode, offset); | |
4156 | } | |
4157 | return false; | |
4158 | ||
4159 | case CONST: | |
4160 | case SYMBOL_REF: | |
4161 | case LABEL_REF: | |
79517551 SN |
4162 | /* load literal: pc-relative constant pool entry. Only supported |
4163 | for SI mode or larger. */ | |
43e9d192 | 4164 | info->type = ADDRESS_SYMBOLIC; |
2d8c6dc1 AH |
4165 | |
4166 | if (!load_store_pair_p && GET_MODE_SIZE (mode) >= 4) | |
43e9d192 IB |
4167 | { |
4168 | rtx sym, addend; | |
4169 | ||
4170 | split_const (x, &sym, &addend); | |
b4f50fd4 RR |
4171 | return ((GET_CODE (sym) == LABEL_REF |
4172 | || (GET_CODE (sym) == SYMBOL_REF | |
4173 | && CONSTANT_POOL_ADDRESS_P (sym) | |
9ee6540a | 4174 | && aarch64_pcrelative_literal_loads))); |
43e9d192 IB |
4175 | } |
4176 | return false; | |
4177 | ||
4178 | case LO_SUM: | |
4179 | info->type = ADDRESS_LO_SUM; | |
4180 | info->base = XEXP (x, 0); | |
4181 | info->offset = XEXP (x, 1); | |
4182 | if (allow_reg_index_p | |
4183 | && aarch64_base_register_rtx_p (info->base, strict_p)) | |
4184 | { | |
4185 | rtx sym, offs; | |
4186 | split_const (info->offset, &sym, &offs); | |
4187 | if (GET_CODE (sym) == SYMBOL_REF | |
a6e0bfa7 | 4188 | && (aarch64_classify_symbol (sym, offs) == SYMBOL_SMALL_ABSOLUTE)) |
43e9d192 IB |
4189 | { |
4190 | /* The symbol and offset must be aligned to the access size. */ | |
4191 | unsigned int align; | |
4192 | unsigned int ref_size; | |
4193 | ||
4194 | if (CONSTANT_POOL_ADDRESS_P (sym)) | |
4195 | align = GET_MODE_ALIGNMENT (get_pool_mode (sym)); | |
4196 | else if (TREE_CONSTANT_POOL_ADDRESS_P (sym)) | |
4197 | { | |
4198 | tree exp = SYMBOL_REF_DECL (sym); | |
4199 | align = TYPE_ALIGN (TREE_TYPE (exp)); | |
4200 | align = CONSTANT_ALIGNMENT (exp, align); | |
4201 | } | |
4202 | else if (SYMBOL_REF_DECL (sym)) | |
4203 | align = DECL_ALIGN (SYMBOL_REF_DECL (sym)); | |
6c031d8d KV |
4204 | else if (SYMBOL_REF_HAS_BLOCK_INFO_P (sym) |
4205 | && SYMBOL_REF_BLOCK (sym) != NULL) | |
4206 | align = SYMBOL_REF_BLOCK (sym)->alignment; | |
43e9d192 IB |
4207 | else |
4208 | align = BITS_PER_UNIT; | |
4209 | ||
4210 | ref_size = GET_MODE_SIZE (mode); | |
4211 | if (ref_size == 0) | |
4212 | ref_size = GET_MODE_SIZE (DImode); | |
4213 | ||
4214 | return ((INTVAL (offs) & (ref_size - 1)) == 0 | |
4215 | && ((align / BITS_PER_UNIT) & (ref_size - 1)) == 0); | |
4216 | } | |
4217 | } | |
4218 | return false; | |
4219 | ||
4220 | default: | |
4221 | return false; | |
4222 | } | |
4223 | } | |
4224 | ||
4225 | bool | |
4226 | aarch64_symbolic_address_p (rtx x) | |
4227 | { | |
4228 | rtx offset; | |
4229 | ||
4230 | split_const (x, &x, &offset); | |
4231 | return GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == LABEL_REF; | |
4232 | } | |
4233 | ||
a6e0bfa7 | 4234 | /* Classify the base of symbolic expression X. */ |
da4f13a4 MS |
4235 | |
4236 | enum aarch64_symbol_type | |
a6e0bfa7 | 4237 | aarch64_classify_symbolic_expression (rtx x) |
43e9d192 IB |
4238 | { |
4239 | rtx offset; | |
da4f13a4 | 4240 | |
43e9d192 | 4241 | split_const (x, &x, &offset); |
a6e0bfa7 | 4242 | return aarch64_classify_symbol (x, offset); |
43e9d192 IB |
4243 | } |
4244 | ||
4245 | ||
4246 | /* Return TRUE if X is a legitimate address for accessing memory in | |
4247 | mode MODE. */ | |
4248 | static bool | |
ef4bddc2 | 4249 | aarch64_legitimate_address_hook_p (machine_mode mode, rtx x, bool strict_p) |
43e9d192 IB |
4250 | { |
4251 | struct aarch64_address_info addr; | |
4252 | ||
4253 | return aarch64_classify_address (&addr, x, mode, MEM, strict_p); | |
4254 | } | |
4255 | ||
4256 | /* Return TRUE if X is a legitimate address for accessing memory in | |
4257 | mode MODE. OUTER_CODE will be PARALLEL if this is a load/store | |
4258 | pair operation. */ | |
4259 | bool | |
ef4bddc2 | 4260 | aarch64_legitimate_address_p (machine_mode mode, rtx x, |
aef66c94 | 4261 | RTX_CODE outer_code, bool strict_p) |
43e9d192 IB |
4262 | { |
4263 | struct aarch64_address_info addr; | |
4264 | ||
4265 | return aarch64_classify_address (&addr, x, mode, outer_code, strict_p); | |
4266 | } | |
4267 | ||
491ec060 WD |
4268 | /* Split an out-of-range address displacement into a base and offset. |
4269 | Use 4KB range for 1- and 2-byte accesses and a 16KB range otherwise | |
4270 | to increase opportunities for sharing the base address of different sizes. | |
4271 | For TI/TFmode and unaligned accesses use a 256-byte range. */ | |
4272 | static bool | |
4273 | aarch64_legitimize_address_displacement (rtx *disp, rtx *off, machine_mode mode) | |
4274 | { | |
4275 | HOST_WIDE_INT mask = GET_MODE_SIZE (mode) < 4 ? 0xfff : 0x3fff; | |
4276 | ||
4277 | if (mode == TImode || mode == TFmode || | |
4278 | (INTVAL (*disp) & (GET_MODE_SIZE (mode) - 1)) != 0) | |
4279 | mask = 0xff; | |
4280 | ||
4281 | *off = GEN_INT (INTVAL (*disp) & ~mask); | |
4282 | *disp = GEN_INT (INTVAL (*disp) & mask); | |
4283 | return true; | |
4284 | } | |
4285 | ||
43e9d192 IB |
4286 | /* Return TRUE if rtx X is immediate constant 0.0 */ |
4287 | bool | |
3520f7cc | 4288 | aarch64_float_const_zero_rtx_p (rtx x) |
43e9d192 | 4289 | { |
43e9d192 IB |
4290 | if (GET_MODE (x) == VOIDmode) |
4291 | return false; | |
4292 | ||
34a72c33 | 4293 | if (REAL_VALUE_MINUS_ZERO (*CONST_DOUBLE_REAL_VALUE (x))) |
43e9d192 | 4294 | return !HONOR_SIGNED_ZEROS (GET_MODE (x)); |
34a72c33 | 4295 | return real_equal (CONST_DOUBLE_REAL_VALUE (x), &dconst0); |
43e9d192 IB |
4296 | } |
4297 | ||
70f09188 AP |
4298 | /* Return the fixed registers used for condition codes. */ |
4299 | ||
4300 | static bool | |
4301 | aarch64_fixed_condition_code_regs (unsigned int *p1, unsigned int *p2) | |
4302 | { | |
4303 | *p1 = CC_REGNUM; | |
4304 | *p2 = INVALID_REGNUM; | |
4305 | return true; | |
4306 | } | |
4307 | ||
78607708 TV |
4308 | /* Emit call insn with PAT and do aarch64-specific handling. */ |
4309 | ||
d07a3fed | 4310 | void |
78607708 TV |
4311 | aarch64_emit_call_insn (rtx pat) |
4312 | { | |
4313 | rtx insn = emit_call_insn (pat); | |
4314 | ||
4315 | rtx *fusage = &CALL_INSN_FUNCTION_USAGE (insn); | |
4316 | clobber_reg (fusage, gen_rtx_REG (word_mode, IP0_REGNUM)); | |
4317 | clobber_reg (fusage, gen_rtx_REG (word_mode, IP1_REGNUM)); | |
4318 | } | |
4319 | ||
ef4bddc2 | 4320 | machine_mode |
43e9d192 IB |
4321 | aarch64_select_cc_mode (RTX_CODE code, rtx x, rtx y) |
4322 | { | |
4323 | /* All floating point compares return CCFP if it is an equality | |
4324 | comparison, and CCFPE otherwise. */ | |
4325 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
4326 | { | |
4327 | switch (code) | |
4328 | { | |
4329 | case EQ: | |
4330 | case NE: | |
4331 | case UNORDERED: | |
4332 | case ORDERED: | |
4333 | case UNLT: | |
4334 | case UNLE: | |
4335 | case UNGT: | |
4336 | case UNGE: | |
4337 | case UNEQ: | |
4338 | case LTGT: | |
4339 | return CCFPmode; | |
4340 | ||
4341 | case LT: | |
4342 | case LE: | |
4343 | case GT: | |
4344 | case GE: | |
4345 | return CCFPEmode; | |
4346 | ||
4347 | default: | |
4348 | gcc_unreachable (); | |
4349 | } | |
4350 | } | |
4351 | ||
2b8568fe KT |
4352 | /* Equality comparisons of short modes against zero can be performed |
4353 | using the TST instruction with the appropriate bitmask. */ | |
4354 | if (y == const0_rtx && REG_P (x) | |
4355 | && (code == EQ || code == NE) | |
4356 | && (GET_MODE (x) == HImode || GET_MODE (x) == QImode)) | |
4357 | return CC_NZmode; | |
4358 | ||
b06335f9 KT |
4359 | /* Similarly, comparisons of zero_extends from shorter modes can |
4360 | be performed using an ANDS with an immediate mask. */ | |
4361 | if (y == const0_rtx && GET_CODE (x) == ZERO_EXTEND | |
4362 | && (GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
4363 | && (GET_MODE (XEXP (x, 0)) == HImode || GET_MODE (XEXP (x, 0)) == QImode) | |
4364 | && (code == EQ || code == NE)) | |
4365 | return CC_NZmode; | |
4366 | ||
43e9d192 IB |
4367 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) |
4368 | && y == const0_rtx | |
4369 | && (code == EQ || code == NE || code == LT || code == GE) | |
b056c910 | 4370 | && (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS || GET_CODE (x) == AND |
7325d85a KT |
4371 | || GET_CODE (x) == NEG |
4372 | || (GET_CODE (x) == ZERO_EXTRACT && CONST_INT_P (XEXP (x, 1)) | |
4373 | && CONST_INT_P (XEXP (x, 2))))) | |
43e9d192 IB |
4374 | return CC_NZmode; |
4375 | ||
1c992d1e | 4376 | /* A compare with a shifted operand. Because of canonicalization, |
43e9d192 IB |
4377 | the comparison will have to be swapped when we emit the assembly |
4378 | code. */ | |
4379 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
4aa81c2e | 4380 | && (REG_P (y) || GET_CODE (y) == SUBREG) |
43e9d192 IB |
4381 | && (GET_CODE (x) == ASHIFT || GET_CODE (x) == ASHIFTRT |
4382 | || GET_CODE (x) == LSHIFTRT | |
1c992d1e | 4383 | || GET_CODE (x) == ZERO_EXTEND || GET_CODE (x) == SIGN_EXTEND)) |
43e9d192 IB |
4384 | return CC_SWPmode; |
4385 | ||
1c992d1e RE |
4386 | /* Similarly for a negated operand, but we can only do this for |
4387 | equalities. */ | |
4388 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
4aa81c2e | 4389 | && (REG_P (y) || GET_CODE (y) == SUBREG) |
1c992d1e RE |
4390 | && (code == EQ || code == NE) |
4391 | && GET_CODE (x) == NEG) | |
4392 | return CC_Zmode; | |
4393 | ||
ef22810a RH |
4394 | /* A test for unsigned overflow. */ |
4395 | if ((GET_MODE (x) == DImode || GET_MODE (x) == TImode) | |
4396 | && code == NE | |
4397 | && GET_CODE (x) == PLUS | |
4398 | && GET_CODE (y) == ZERO_EXTEND) | |
4399 | return CC_Cmode; | |
4400 | ||
43e9d192 IB |
4401 | /* For everything else, return CCmode. */ |
4402 | return CCmode; | |
4403 | } | |
4404 | ||
3dfa7055 ZC |
4405 | static int |
4406 | aarch64_get_condition_code_1 (enum machine_mode, enum rtx_code); | |
4407 | ||
cd5660ab | 4408 | int |
43e9d192 IB |
4409 | aarch64_get_condition_code (rtx x) |
4410 | { | |
ef4bddc2 | 4411 | machine_mode mode = GET_MODE (XEXP (x, 0)); |
43e9d192 IB |
4412 | enum rtx_code comp_code = GET_CODE (x); |
4413 | ||
4414 | if (GET_MODE_CLASS (mode) != MODE_CC) | |
4415 | mode = SELECT_CC_MODE (comp_code, XEXP (x, 0), XEXP (x, 1)); | |
3dfa7055 ZC |
4416 | return aarch64_get_condition_code_1 (mode, comp_code); |
4417 | } | |
43e9d192 | 4418 | |
3dfa7055 ZC |
4419 | static int |
4420 | aarch64_get_condition_code_1 (enum machine_mode mode, enum rtx_code comp_code) | |
4421 | { | |
43e9d192 IB |
4422 | switch (mode) |
4423 | { | |
4424 | case CCFPmode: | |
4425 | case CCFPEmode: | |
4426 | switch (comp_code) | |
4427 | { | |
4428 | case GE: return AARCH64_GE; | |
4429 | case GT: return AARCH64_GT; | |
4430 | case LE: return AARCH64_LS; | |
4431 | case LT: return AARCH64_MI; | |
4432 | case NE: return AARCH64_NE; | |
4433 | case EQ: return AARCH64_EQ; | |
4434 | case ORDERED: return AARCH64_VC; | |
4435 | case UNORDERED: return AARCH64_VS; | |
4436 | case UNLT: return AARCH64_LT; | |
4437 | case UNLE: return AARCH64_LE; | |
4438 | case UNGT: return AARCH64_HI; | |
4439 | case UNGE: return AARCH64_PL; | |
cd5660ab | 4440 | default: return -1; |
43e9d192 IB |
4441 | } |
4442 | break; | |
4443 | ||
4444 | case CCmode: | |
4445 | switch (comp_code) | |
4446 | { | |
4447 | case NE: return AARCH64_NE; | |
4448 | case EQ: return AARCH64_EQ; | |
4449 | case GE: return AARCH64_GE; | |
4450 | case GT: return AARCH64_GT; | |
4451 | case LE: return AARCH64_LE; | |
4452 | case LT: return AARCH64_LT; | |
4453 | case GEU: return AARCH64_CS; | |
4454 | case GTU: return AARCH64_HI; | |
4455 | case LEU: return AARCH64_LS; | |
4456 | case LTU: return AARCH64_CC; | |
cd5660ab | 4457 | default: return -1; |
43e9d192 IB |
4458 | } |
4459 | break; | |
4460 | ||
4461 | case CC_SWPmode: | |
43e9d192 IB |
4462 | switch (comp_code) |
4463 | { | |
4464 | case NE: return AARCH64_NE; | |
4465 | case EQ: return AARCH64_EQ; | |
4466 | case GE: return AARCH64_LE; | |
4467 | case GT: return AARCH64_LT; | |
4468 | case LE: return AARCH64_GE; | |
4469 | case LT: return AARCH64_GT; | |
4470 | case GEU: return AARCH64_LS; | |
4471 | case GTU: return AARCH64_CC; | |
4472 | case LEU: return AARCH64_CS; | |
4473 | case LTU: return AARCH64_HI; | |
cd5660ab | 4474 | default: return -1; |
43e9d192 IB |
4475 | } |
4476 | break; | |
4477 | ||
4478 | case CC_NZmode: | |
4479 | switch (comp_code) | |
4480 | { | |
4481 | case NE: return AARCH64_NE; | |
4482 | case EQ: return AARCH64_EQ; | |
4483 | case GE: return AARCH64_PL; | |
4484 | case LT: return AARCH64_MI; | |
cd5660ab | 4485 | default: return -1; |
43e9d192 IB |
4486 | } |
4487 | break; | |
4488 | ||
1c992d1e RE |
4489 | case CC_Zmode: |
4490 | switch (comp_code) | |
4491 | { | |
4492 | case NE: return AARCH64_NE; | |
4493 | case EQ: return AARCH64_EQ; | |
cd5660ab | 4494 | default: return -1; |
1c992d1e RE |
4495 | } |
4496 | break; | |
4497 | ||
ef22810a RH |
4498 | case CC_Cmode: |
4499 | switch (comp_code) | |
4500 | { | |
4501 | case NE: return AARCH64_CS; | |
4502 | case EQ: return AARCH64_CC; | |
4503 | default: return -1; | |
4504 | } | |
4505 | break; | |
4506 | ||
43e9d192 | 4507 | default: |
cd5660ab | 4508 | return -1; |
43e9d192 | 4509 | } |
3dfa7055 | 4510 | |
3dfa7055 | 4511 | return -1; |
43e9d192 IB |
4512 | } |
4513 | ||
ddeabd3e AL |
4514 | bool |
4515 | aarch64_const_vec_all_same_in_range_p (rtx x, | |
4516 | HOST_WIDE_INT minval, | |
4517 | HOST_WIDE_INT maxval) | |
4518 | { | |
4519 | HOST_WIDE_INT firstval; | |
4520 | int count, i; | |
4521 | ||
4522 | if (GET_CODE (x) != CONST_VECTOR | |
4523 | || GET_MODE_CLASS (GET_MODE (x)) != MODE_VECTOR_INT) | |
4524 | return false; | |
4525 | ||
4526 | firstval = INTVAL (CONST_VECTOR_ELT (x, 0)); | |
4527 | if (firstval < minval || firstval > maxval) | |
4528 | return false; | |
4529 | ||
4530 | count = CONST_VECTOR_NUNITS (x); | |
4531 | for (i = 1; i < count; i++) | |
4532 | if (INTVAL (CONST_VECTOR_ELT (x, i)) != firstval) | |
4533 | return false; | |
4534 | ||
4535 | return true; | |
4536 | } | |
4537 | ||
4538 | bool | |
4539 | aarch64_const_vec_all_same_int_p (rtx x, HOST_WIDE_INT val) | |
4540 | { | |
4541 | return aarch64_const_vec_all_same_in_range_p (x, val, val); | |
4542 | } | |
4543 | ||
43e9d192 | 4544 | |
cf670503 ZC |
4545 | /* N Z C V. */ |
4546 | #define AARCH64_CC_V 1 | |
4547 | #define AARCH64_CC_C (1 << 1) | |
4548 | #define AARCH64_CC_Z (1 << 2) | |
4549 | #define AARCH64_CC_N (1 << 3) | |
4550 | ||
c8012fbc WD |
4551 | /* N Z C V flags for ccmp. Indexed by AARCH64_COND_CODE. */ |
4552 | static const int aarch64_nzcv_codes[] = | |
4553 | { | |
4554 | 0, /* EQ, Z == 1. */ | |
4555 | AARCH64_CC_Z, /* NE, Z == 0. */ | |
4556 | 0, /* CS, C == 1. */ | |
4557 | AARCH64_CC_C, /* CC, C == 0. */ | |
4558 | 0, /* MI, N == 1. */ | |
4559 | AARCH64_CC_N, /* PL, N == 0. */ | |
4560 | 0, /* VS, V == 1. */ | |
4561 | AARCH64_CC_V, /* VC, V == 0. */ | |
4562 | 0, /* HI, C ==1 && Z == 0. */ | |
4563 | AARCH64_CC_C, /* LS, !(C == 1 && Z == 0). */ | |
4564 | AARCH64_CC_V, /* GE, N == V. */ | |
4565 | 0, /* LT, N != V. */ | |
4566 | AARCH64_CC_Z, /* GT, Z == 0 && N == V. */ | |
4567 | 0, /* LE, !(Z == 0 && N == V). */ | |
4568 | 0, /* AL, Any. */ | |
4569 | 0 /* NV, Any. */ | |
cf670503 ZC |
4570 | }; |
4571 | ||
cc8ca59e JB |
4572 | static void |
4573 | aarch64_print_operand (FILE *f, rtx x, int code) | |
43e9d192 IB |
4574 | { |
4575 | switch (code) | |
4576 | { | |
f541a481 KT |
4577 | /* An integer or symbol address without a preceding # sign. */ |
4578 | case 'c': | |
4579 | switch (GET_CODE (x)) | |
4580 | { | |
4581 | case CONST_INT: | |
4582 | fprintf (f, HOST_WIDE_INT_PRINT_DEC, INTVAL (x)); | |
4583 | break; | |
4584 | ||
4585 | case SYMBOL_REF: | |
4586 | output_addr_const (f, x); | |
4587 | break; | |
4588 | ||
4589 | case CONST: | |
4590 | if (GET_CODE (XEXP (x, 0)) == PLUS | |
4591 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF) | |
4592 | { | |
4593 | output_addr_const (f, x); | |
4594 | break; | |
4595 | } | |
4596 | /* Fall through. */ | |
4597 | ||
4598 | default: | |
4599 | output_operand_lossage ("Unsupported operand for code '%c'", code); | |
4600 | } | |
4601 | break; | |
4602 | ||
43e9d192 IB |
4603 | case 'e': |
4604 | /* Print the sign/zero-extend size as a character 8->b, 16->h, 32->w. */ | |
4605 | { | |
4606 | int n; | |
4607 | ||
4aa81c2e | 4608 | if (!CONST_INT_P (x) |
43e9d192 IB |
4609 | || (n = exact_log2 (INTVAL (x) & ~7)) <= 0) |
4610 | { | |
4611 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4612 | return; | |
4613 | } | |
4614 | ||
4615 | switch (n) | |
4616 | { | |
4617 | case 3: | |
4618 | fputc ('b', f); | |
4619 | break; | |
4620 | case 4: | |
4621 | fputc ('h', f); | |
4622 | break; | |
4623 | case 5: | |
4624 | fputc ('w', f); | |
4625 | break; | |
4626 | default: | |
4627 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4628 | return; | |
4629 | } | |
4630 | } | |
4631 | break; | |
4632 | ||
4633 | case 'p': | |
4634 | { | |
4635 | int n; | |
4636 | ||
4637 | /* Print N such that 2^N == X. */ | |
4aa81c2e | 4638 | if (!CONST_INT_P (x) || (n = exact_log2 (INTVAL (x))) < 0) |
43e9d192 IB |
4639 | { |
4640 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4641 | return; | |
4642 | } | |
4643 | ||
4644 | asm_fprintf (f, "%d", n); | |
4645 | } | |
4646 | break; | |
4647 | ||
4648 | case 'P': | |
4649 | /* Print the number of non-zero bits in X (a const_int). */ | |
4aa81c2e | 4650 | if (!CONST_INT_P (x)) |
43e9d192 IB |
4651 | { |
4652 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4653 | return; | |
4654 | } | |
4655 | ||
8d55c61b | 4656 | asm_fprintf (f, "%u", popcount_hwi (INTVAL (x))); |
43e9d192 IB |
4657 | break; |
4658 | ||
4659 | case 'H': | |
4660 | /* Print the higher numbered register of a pair (TImode) of regs. */ | |
4aa81c2e | 4661 | if (!REG_P (x) || !GP_REGNUM_P (REGNO (x) + 1)) |
43e9d192 IB |
4662 | { |
4663 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4664 | return; | |
4665 | } | |
4666 | ||
01a3a324 | 4667 | asm_fprintf (f, "%s", reg_names [REGNO (x) + 1]); |
43e9d192 IB |
4668 | break; |
4669 | ||
43e9d192 | 4670 | case 'M': |
c8012fbc | 4671 | case 'm': |
cd5660ab KT |
4672 | { |
4673 | int cond_code; | |
c8012fbc | 4674 | /* Print a condition (eq, ne, etc) or its inverse. */ |
43e9d192 | 4675 | |
c8012fbc WD |
4676 | /* CONST_TRUE_RTX means al/nv (al is the default, don't print it). */ |
4677 | if (x == const_true_rtx) | |
cd5660ab | 4678 | { |
c8012fbc WD |
4679 | if (code == 'M') |
4680 | fputs ("nv", f); | |
cd5660ab KT |
4681 | return; |
4682 | } | |
43e9d192 | 4683 | |
cd5660ab KT |
4684 | if (!COMPARISON_P (x)) |
4685 | { | |
4686 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4687 | return; | |
4688 | } | |
c8012fbc | 4689 | |
cd5660ab KT |
4690 | cond_code = aarch64_get_condition_code (x); |
4691 | gcc_assert (cond_code >= 0); | |
c8012fbc WD |
4692 | if (code == 'M') |
4693 | cond_code = AARCH64_INVERSE_CONDITION_CODE (cond_code); | |
4694 | fputs (aarch64_condition_codes[cond_code], f); | |
cd5660ab | 4695 | } |
43e9d192 IB |
4696 | break; |
4697 | ||
4698 | case 'b': | |
4699 | case 'h': | |
4700 | case 's': | |
4701 | case 'd': | |
4702 | case 'q': | |
4703 | /* Print a scalar FP/SIMD register name. */ | |
4704 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) | |
4705 | { | |
4706 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
4707 | return; | |
4708 | } | |
50ce6f88 | 4709 | asm_fprintf (f, "%c%d", code, REGNO (x) - V0_REGNUM); |
43e9d192 IB |
4710 | break; |
4711 | ||
4712 | case 'S': | |
4713 | case 'T': | |
4714 | case 'U': | |
4715 | case 'V': | |
4716 | /* Print the first FP/SIMD register name in a list. */ | |
4717 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) | |
4718 | { | |
4719 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
4720 | return; | |
4721 | } | |
50ce6f88 | 4722 | asm_fprintf (f, "v%d", REGNO (x) - V0_REGNUM + (code - 'S')); |
43e9d192 IB |
4723 | break; |
4724 | ||
2d8c6dc1 AH |
4725 | case 'R': |
4726 | /* Print a scalar FP/SIMD register name + 1. */ | |
4727 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) | |
4728 | { | |
4729 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
4730 | return; | |
4731 | } | |
4732 | asm_fprintf (f, "q%d", REGNO (x) - V0_REGNUM + 1); | |
4733 | break; | |
4734 | ||
a05c0ddf | 4735 | case 'X': |
50d38551 | 4736 | /* Print bottom 16 bits of integer constant in hex. */ |
4aa81c2e | 4737 | if (!CONST_INT_P (x)) |
a05c0ddf IB |
4738 | { |
4739 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4740 | return; | |
4741 | } | |
50d38551 | 4742 | asm_fprintf (f, "0x%wx", UINTVAL (x) & 0xffff); |
a05c0ddf IB |
4743 | break; |
4744 | ||
43e9d192 IB |
4745 | case 'w': |
4746 | case 'x': | |
4747 | /* Print a general register name or the zero register (32-bit or | |
4748 | 64-bit). */ | |
3520f7cc JG |
4749 | if (x == const0_rtx |
4750 | || (CONST_DOUBLE_P (x) && aarch64_float_const_zero_rtx_p (x))) | |
43e9d192 | 4751 | { |
50ce6f88 | 4752 | asm_fprintf (f, "%czr", code); |
43e9d192 IB |
4753 | break; |
4754 | } | |
4755 | ||
4756 | if (REG_P (x) && GP_REGNUM_P (REGNO (x))) | |
4757 | { | |
50ce6f88 | 4758 | asm_fprintf (f, "%c%d", code, REGNO (x) - R0_REGNUM); |
43e9d192 IB |
4759 | break; |
4760 | } | |
4761 | ||
4762 | if (REG_P (x) && REGNO (x) == SP_REGNUM) | |
4763 | { | |
50ce6f88 | 4764 | asm_fprintf (f, "%ssp", code == 'w' ? "w" : ""); |
43e9d192 IB |
4765 | break; |
4766 | } | |
4767 | ||
4768 | /* Fall through */ | |
4769 | ||
4770 | case 0: | |
4771 | /* Print a normal operand, if it's a general register, then we | |
4772 | assume DImode. */ | |
4773 | if (x == NULL) | |
4774 | { | |
4775 | output_operand_lossage ("missing operand"); | |
4776 | return; | |
4777 | } | |
4778 | ||
4779 | switch (GET_CODE (x)) | |
4780 | { | |
4781 | case REG: | |
01a3a324 | 4782 | asm_fprintf (f, "%s", reg_names [REGNO (x)]); |
43e9d192 IB |
4783 | break; |
4784 | ||
4785 | case MEM: | |
cc8ca59e | 4786 | output_address (GET_MODE (x), XEXP (x, 0)); |
43e9d192 IB |
4787 | break; |
4788 | ||
2af16a7c | 4789 | case CONST: |
43e9d192 IB |
4790 | case LABEL_REF: |
4791 | case SYMBOL_REF: | |
4792 | output_addr_const (asm_out_file, x); | |
4793 | break; | |
4794 | ||
4795 | case CONST_INT: | |
4796 | asm_fprintf (f, "%wd", INTVAL (x)); | |
4797 | break; | |
4798 | ||
4799 | case CONST_VECTOR: | |
3520f7cc JG |
4800 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_VECTOR_INT) |
4801 | { | |
ddeabd3e AL |
4802 | gcc_assert ( |
4803 | aarch64_const_vec_all_same_in_range_p (x, | |
4804 | HOST_WIDE_INT_MIN, | |
4805 | HOST_WIDE_INT_MAX)); | |
3520f7cc JG |
4806 | asm_fprintf (f, "%wd", INTVAL (CONST_VECTOR_ELT (x, 0))); |
4807 | } | |
4808 | else if (aarch64_simd_imm_zero_p (x, GET_MODE (x))) | |
4809 | { | |
4810 | fputc ('0', f); | |
4811 | } | |
4812 | else | |
4813 | gcc_unreachable (); | |
43e9d192 IB |
4814 | break; |
4815 | ||
3520f7cc | 4816 | case CONST_DOUBLE: |
2ca5b430 KT |
4817 | /* Since we define TARGET_SUPPORTS_WIDE_INT we shouldn't ever |
4818 | be getting CONST_DOUBLEs holding integers. */ | |
4819 | gcc_assert (GET_MODE (x) != VOIDmode); | |
4820 | if (aarch64_float_const_zero_rtx_p (x)) | |
3520f7cc JG |
4821 | { |
4822 | fputc ('0', f); | |
4823 | break; | |
4824 | } | |
4825 | else if (aarch64_float_const_representable_p (x)) | |
4826 | { | |
4827 | #define buf_size 20 | |
4828 | char float_buf[buf_size] = {'\0'}; | |
34a72c33 RS |
4829 | real_to_decimal_for_mode (float_buf, |
4830 | CONST_DOUBLE_REAL_VALUE (x), | |
3520f7cc JG |
4831 | buf_size, buf_size, |
4832 | 1, GET_MODE (x)); | |
4833 | asm_fprintf (asm_out_file, "%s", float_buf); | |
4834 | break; | |
4835 | #undef buf_size | |
4836 | } | |
4837 | output_operand_lossage ("invalid constant"); | |
4838 | return; | |
43e9d192 IB |
4839 | default: |
4840 | output_operand_lossage ("invalid operand"); | |
4841 | return; | |
4842 | } | |
4843 | break; | |
4844 | ||
4845 | case 'A': | |
4846 | if (GET_CODE (x) == HIGH) | |
4847 | x = XEXP (x, 0); | |
4848 | ||
a6e0bfa7 | 4849 | switch (aarch64_classify_symbolic_expression (x)) |
43e9d192 | 4850 | { |
6642bdb4 | 4851 | case SYMBOL_SMALL_GOT_4G: |
43e9d192 IB |
4852 | asm_fprintf (asm_out_file, ":got:"); |
4853 | break; | |
4854 | ||
4855 | case SYMBOL_SMALL_TLSGD: | |
4856 | asm_fprintf (asm_out_file, ":tlsgd:"); | |
4857 | break; | |
4858 | ||
4859 | case SYMBOL_SMALL_TLSDESC: | |
4860 | asm_fprintf (asm_out_file, ":tlsdesc:"); | |
4861 | break; | |
4862 | ||
79496620 | 4863 | case SYMBOL_SMALL_TLSIE: |
43e9d192 IB |
4864 | asm_fprintf (asm_out_file, ":gottprel:"); |
4865 | break; | |
4866 | ||
d18ba284 | 4867 | case SYMBOL_TLSLE24: |
43e9d192 IB |
4868 | asm_fprintf (asm_out_file, ":tprel:"); |
4869 | break; | |
4870 | ||
87dd8ab0 MS |
4871 | case SYMBOL_TINY_GOT: |
4872 | gcc_unreachable (); | |
4873 | break; | |
4874 | ||
43e9d192 IB |
4875 | default: |
4876 | break; | |
4877 | } | |
4878 | output_addr_const (asm_out_file, x); | |
4879 | break; | |
4880 | ||
4881 | case 'L': | |
a6e0bfa7 | 4882 | switch (aarch64_classify_symbolic_expression (x)) |
43e9d192 | 4883 | { |
6642bdb4 | 4884 | case SYMBOL_SMALL_GOT_4G: |
43e9d192 IB |
4885 | asm_fprintf (asm_out_file, ":lo12:"); |
4886 | break; | |
4887 | ||
4888 | case SYMBOL_SMALL_TLSGD: | |
4889 | asm_fprintf (asm_out_file, ":tlsgd_lo12:"); | |
4890 | break; | |
4891 | ||
4892 | case SYMBOL_SMALL_TLSDESC: | |
4893 | asm_fprintf (asm_out_file, ":tlsdesc_lo12:"); | |
4894 | break; | |
4895 | ||
79496620 | 4896 | case SYMBOL_SMALL_TLSIE: |
43e9d192 IB |
4897 | asm_fprintf (asm_out_file, ":gottprel_lo12:"); |
4898 | break; | |
4899 | ||
cbf5629e JW |
4900 | case SYMBOL_TLSLE12: |
4901 | asm_fprintf (asm_out_file, ":tprel_lo12:"); | |
4902 | break; | |
4903 | ||
d18ba284 | 4904 | case SYMBOL_TLSLE24: |
43e9d192 IB |
4905 | asm_fprintf (asm_out_file, ":tprel_lo12_nc:"); |
4906 | break; | |
4907 | ||
87dd8ab0 MS |
4908 | case SYMBOL_TINY_GOT: |
4909 | asm_fprintf (asm_out_file, ":got:"); | |
4910 | break; | |
4911 | ||
5ae7caad JW |
4912 | case SYMBOL_TINY_TLSIE: |
4913 | asm_fprintf (asm_out_file, ":gottprel:"); | |
4914 | break; | |
4915 | ||
43e9d192 IB |
4916 | default: |
4917 | break; | |
4918 | } | |
4919 | output_addr_const (asm_out_file, x); | |
4920 | break; | |
4921 | ||
4922 | case 'G': | |
4923 | ||
a6e0bfa7 | 4924 | switch (aarch64_classify_symbolic_expression (x)) |
43e9d192 | 4925 | { |
d18ba284 | 4926 | case SYMBOL_TLSLE24: |
43e9d192 IB |
4927 | asm_fprintf (asm_out_file, ":tprel_hi12:"); |
4928 | break; | |
4929 | default: | |
4930 | break; | |
4931 | } | |
4932 | output_addr_const (asm_out_file, x); | |
4933 | break; | |
4934 | ||
cf670503 ZC |
4935 | case 'k': |
4936 | { | |
c8012fbc | 4937 | HOST_WIDE_INT cond_code; |
cf670503 ZC |
4938 | /* Print nzcv. */ |
4939 | ||
c8012fbc | 4940 | if (!CONST_INT_P (x)) |
cf670503 ZC |
4941 | { |
4942 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4943 | return; | |
4944 | } | |
4945 | ||
c8012fbc WD |
4946 | cond_code = INTVAL (x); |
4947 | gcc_assert (cond_code >= 0 && cond_code <= AARCH64_NV); | |
4948 | asm_fprintf (f, "%d", aarch64_nzcv_codes[cond_code]); | |
cf670503 ZC |
4949 | } |
4950 | break; | |
4951 | ||
43e9d192 IB |
4952 | default: |
4953 | output_operand_lossage ("invalid operand prefix '%%%c'", code); | |
4954 | return; | |
4955 | } | |
4956 | } | |
4957 | ||
cc8ca59e JB |
4958 | static void |
4959 | aarch64_print_operand_address (FILE *f, machine_mode mode, rtx x) | |
43e9d192 IB |
4960 | { |
4961 | struct aarch64_address_info addr; | |
4962 | ||
cc8ca59e | 4963 | if (aarch64_classify_address (&addr, x, mode, MEM, true)) |
43e9d192 IB |
4964 | switch (addr.type) |
4965 | { | |
4966 | case ADDRESS_REG_IMM: | |
4967 | if (addr.offset == const0_rtx) | |
01a3a324 | 4968 | asm_fprintf (f, "[%s]", reg_names [REGNO (addr.base)]); |
43e9d192 | 4969 | else |
16a3246f | 4970 | asm_fprintf (f, "[%s, %wd]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4971 | INTVAL (addr.offset)); |
4972 | return; | |
4973 | ||
4974 | case ADDRESS_REG_REG: | |
4975 | if (addr.shift == 0) | |
16a3246f | 4976 | asm_fprintf (f, "[%s, %s]", reg_names [REGNO (addr.base)], |
01a3a324 | 4977 | reg_names [REGNO (addr.offset)]); |
43e9d192 | 4978 | else |
16a3246f | 4979 | asm_fprintf (f, "[%s, %s, lsl %u]", reg_names [REGNO (addr.base)], |
01a3a324 | 4980 | reg_names [REGNO (addr.offset)], addr.shift); |
43e9d192 IB |
4981 | return; |
4982 | ||
4983 | case ADDRESS_REG_UXTW: | |
4984 | if (addr.shift == 0) | |
16a3246f | 4985 | asm_fprintf (f, "[%s, w%d, uxtw]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4986 | REGNO (addr.offset) - R0_REGNUM); |
4987 | else | |
16a3246f | 4988 | asm_fprintf (f, "[%s, w%d, uxtw %u]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4989 | REGNO (addr.offset) - R0_REGNUM, addr.shift); |
4990 | return; | |
4991 | ||
4992 | case ADDRESS_REG_SXTW: | |
4993 | if (addr.shift == 0) | |
16a3246f | 4994 | asm_fprintf (f, "[%s, w%d, sxtw]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4995 | REGNO (addr.offset) - R0_REGNUM); |
4996 | else | |
16a3246f | 4997 | asm_fprintf (f, "[%s, w%d, sxtw %u]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4998 | REGNO (addr.offset) - R0_REGNUM, addr.shift); |
4999 | return; | |
5000 | ||
5001 | case ADDRESS_REG_WB: | |
5002 | switch (GET_CODE (x)) | |
5003 | { | |
5004 | case PRE_INC: | |
16a3246f | 5005 | asm_fprintf (f, "[%s, %d]!", reg_names [REGNO (addr.base)], |
cc8ca59e | 5006 | GET_MODE_SIZE (mode)); |
43e9d192 IB |
5007 | return; |
5008 | case POST_INC: | |
16a3246f | 5009 | asm_fprintf (f, "[%s], %d", reg_names [REGNO (addr.base)], |
cc8ca59e | 5010 | GET_MODE_SIZE (mode)); |
43e9d192 IB |
5011 | return; |
5012 | case PRE_DEC: | |
16a3246f | 5013 | asm_fprintf (f, "[%s, -%d]!", reg_names [REGNO (addr.base)], |
cc8ca59e | 5014 | GET_MODE_SIZE (mode)); |
43e9d192 IB |
5015 | return; |
5016 | case POST_DEC: | |
16a3246f | 5017 | asm_fprintf (f, "[%s], -%d", reg_names [REGNO (addr.base)], |
cc8ca59e | 5018 | GET_MODE_SIZE (mode)); |
43e9d192 IB |
5019 | return; |
5020 | case PRE_MODIFY: | |
16a3246f | 5021 | asm_fprintf (f, "[%s, %wd]!", reg_names [REGNO (addr.base)], |
43e9d192 IB |
5022 | INTVAL (addr.offset)); |
5023 | return; | |
5024 | case POST_MODIFY: | |
16a3246f | 5025 | asm_fprintf (f, "[%s], %wd", reg_names [REGNO (addr.base)], |
43e9d192 IB |
5026 | INTVAL (addr.offset)); |
5027 | return; | |
5028 | default: | |
5029 | break; | |
5030 | } | |
5031 | break; | |
5032 | ||
5033 | case ADDRESS_LO_SUM: | |
16a3246f | 5034 | asm_fprintf (f, "[%s, #:lo12:", reg_names [REGNO (addr.base)]); |
43e9d192 IB |
5035 | output_addr_const (f, addr.offset); |
5036 | asm_fprintf (f, "]"); | |
5037 | return; | |
5038 | ||
5039 | case ADDRESS_SYMBOLIC: | |
5040 | break; | |
5041 | } | |
5042 | ||
5043 | output_addr_const (f, x); | |
5044 | } | |
5045 | ||
43e9d192 IB |
5046 | bool |
5047 | aarch64_label_mentioned_p (rtx x) | |
5048 | { | |
5049 | const char *fmt; | |
5050 | int i; | |
5051 | ||
5052 | if (GET_CODE (x) == LABEL_REF) | |
5053 | return true; | |
5054 | ||
5055 | /* UNSPEC_TLS entries for a symbol include a LABEL_REF for the | |
5056 | referencing instruction, but they are constant offsets, not | |
5057 | symbols. */ | |
5058 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
5059 | return false; | |
5060 | ||
5061 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
5062 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
5063 | { | |
5064 | if (fmt[i] == 'E') | |
5065 | { | |
5066 | int j; | |
5067 | ||
5068 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
5069 | if (aarch64_label_mentioned_p (XVECEXP (x, i, j))) | |
5070 | return 1; | |
5071 | } | |
5072 | else if (fmt[i] == 'e' && aarch64_label_mentioned_p (XEXP (x, i))) | |
5073 | return 1; | |
5074 | } | |
5075 | ||
5076 | return 0; | |
5077 | } | |
5078 | ||
5079 | /* Implement REGNO_REG_CLASS. */ | |
5080 | ||
5081 | enum reg_class | |
5082 | aarch64_regno_regclass (unsigned regno) | |
5083 | { | |
5084 | if (GP_REGNUM_P (regno)) | |
a4a182c6 | 5085 | return GENERAL_REGS; |
43e9d192 IB |
5086 | |
5087 | if (regno == SP_REGNUM) | |
5088 | return STACK_REG; | |
5089 | ||
5090 | if (regno == FRAME_POINTER_REGNUM | |
5091 | || regno == ARG_POINTER_REGNUM) | |
f24bb080 | 5092 | return POINTER_REGS; |
43e9d192 IB |
5093 | |
5094 | if (FP_REGNUM_P (regno)) | |
5095 | return FP_LO_REGNUM_P (regno) ? FP_LO_REGS : FP_REGS; | |
5096 | ||
5097 | return NO_REGS; | |
5098 | } | |
5099 | ||
0c4ec427 | 5100 | static rtx |
ef4bddc2 | 5101 | aarch64_legitimize_address (rtx x, rtx /* orig_x */, machine_mode mode) |
0c4ec427 RE |
5102 | { |
5103 | /* Try to split X+CONST into Y=X+(CONST & ~mask), Y+(CONST&mask), | |
5104 | where mask is selected by alignment and size of the offset. | |
5105 | We try to pick as large a range for the offset as possible to | |
5106 | maximize the chance of a CSE. However, for aligned addresses | |
5107 | we limit the range to 4k so that structures with different sized | |
e8426e0a BC |
5108 | elements are likely to use the same base. We need to be careful |
5109 | not to split a CONST for some forms of address expression, otherwise | |
5110 | it will generate sub-optimal code. */ | |
0c4ec427 RE |
5111 | |
5112 | if (GET_CODE (x) == PLUS && CONST_INT_P (XEXP (x, 1))) | |
5113 | { | |
9e0218fc | 5114 | rtx base = XEXP (x, 0); |
17d7bdd8 | 5115 | rtx offset_rtx = XEXP (x, 1); |
9e0218fc | 5116 | HOST_WIDE_INT offset = INTVAL (offset_rtx); |
0c4ec427 | 5117 | |
9e0218fc | 5118 | if (GET_CODE (base) == PLUS) |
e8426e0a | 5119 | { |
9e0218fc RH |
5120 | rtx op0 = XEXP (base, 0); |
5121 | rtx op1 = XEXP (base, 1); | |
5122 | ||
5123 | /* Force any scaling into a temp for CSE. */ | |
5124 | op0 = force_reg (Pmode, op0); | |
5125 | op1 = force_reg (Pmode, op1); | |
5126 | ||
5127 | /* Let the pointer register be in op0. */ | |
5128 | if (REG_POINTER (op1)) | |
5129 | std::swap (op0, op1); | |
5130 | ||
5131 | /* If the pointer is virtual or frame related, then we know that | |
5132 | virtual register instantiation or register elimination is going | |
5133 | to apply a second constant. We want the two constants folded | |
5134 | together easily. Therefore, emit as (OP0 + CONST) + OP1. */ | |
5135 | if (virt_or_elim_regno_p (REGNO (op0))) | |
e8426e0a | 5136 | { |
9e0218fc RH |
5137 | base = expand_binop (Pmode, add_optab, op0, offset_rtx, |
5138 | NULL_RTX, true, OPTAB_DIRECT); | |
5139 | return gen_rtx_PLUS (Pmode, base, op1); | |
e8426e0a | 5140 | } |
e8426e0a | 5141 | |
9e0218fc RH |
5142 | /* Otherwise, in order to encourage CSE (and thence loop strength |
5143 | reduce) scaled addresses, emit as (OP0 + OP1) + CONST. */ | |
5144 | base = expand_binop (Pmode, add_optab, op0, op1, | |
5145 | NULL_RTX, true, OPTAB_DIRECT); | |
5146 | x = gen_rtx_PLUS (Pmode, base, offset_rtx); | |
e8426e0a BC |
5147 | } |
5148 | ||
0c4ec427 | 5149 | /* Does it look like we'll need a load/store-pair operation? */ |
9e0218fc | 5150 | HOST_WIDE_INT base_offset; |
0c4ec427 RE |
5151 | if (GET_MODE_SIZE (mode) > 16 |
5152 | || mode == TImode) | |
5153 | base_offset = ((offset + 64 * GET_MODE_SIZE (mode)) | |
5154 | & ~((128 * GET_MODE_SIZE (mode)) - 1)); | |
5155 | /* For offsets aren't a multiple of the access size, the limit is | |
5156 | -256...255. */ | |
5157 | else if (offset & (GET_MODE_SIZE (mode) - 1)) | |
ff0f3f1c WD |
5158 | { |
5159 | base_offset = (offset + 0x100) & ~0x1ff; | |
5160 | ||
5161 | /* BLKmode typically uses LDP of X-registers. */ | |
5162 | if (mode == BLKmode) | |
5163 | base_offset = (offset + 512) & ~0x3ff; | |
5164 | } | |
5165 | /* Small negative offsets are supported. */ | |
5166 | else if (IN_RANGE (offset, -256, 0)) | |
5167 | base_offset = 0; | |
5168 | /* Use 12-bit offset by access size. */ | |
0c4ec427 | 5169 | else |
ff0f3f1c | 5170 | base_offset = offset & (~0xfff * GET_MODE_SIZE (mode)); |
0c4ec427 | 5171 | |
9e0218fc RH |
5172 | if (base_offset != 0) |
5173 | { | |
5174 | base = plus_constant (Pmode, base, base_offset); | |
5175 | base = force_operand (base, NULL_RTX); | |
5176 | return plus_constant (Pmode, base, offset - base_offset); | |
5177 | } | |
0c4ec427 RE |
5178 | } |
5179 | ||
5180 | return x; | |
5181 | } | |
5182 | ||
b4f50fd4 RR |
5183 | /* Return the reload icode required for a constant pool in mode. */ |
5184 | static enum insn_code | |
5185 | aarch64_constant_pool_reload_icode (machine_mode mode) | |
5186 | { | |
5187 | switch (mode) | |
5188 | { | |
5189 | case SFmode: | |
5190 | return CODE_FOR_aarch64_reload_movcpsfdi; | |
5191 | ||
5192 | case DFmode: | |
5193 | return CODE_FOR_aarch64_reload_movcpdfdi; | |
5194 | ||
5195 | case TFmode: | |
5196 | return CODE_FOR_aarch64_reload_movcptfdi; | |
5197 | ||
5198 | case V8QImode: | |
5199 | return CODE_FOR_aarch64_reload_movcpv8qidi; | |
5200 | ||
5201 | case V16QImode: | |
5202 | return CODE_FOR_aarch64_reload_movcpv16qidi; | |
5203 | ||
5204 | case V4HImode: | |
5205 | return CODE_FOR_aarch64_reload_movcpv4hidi; | |
5206 | ||
5207 | case V8HImode: | |
5208 | return CODE_FOR_aarch64_reload_movcpv8hidi; | |
5209 | ||
5210 | case V2SImode: | |
5211 | return CODE_FOR_aarch64_reload_movcpv2sidi; | |
5212 | ||
5213 | case V4SImode: | |
5214 | return CODE_FOR_aarch64_reload_movcpv4sidi; | |
5215 | ||
5216 | case V2DImode: | |
5217 | return CODE_FOR_aarch64_reload_movcpv2didi; | |
5218 | ||
5219 | case V2DFmode: | |
5220 | return CODE_FOR_aarch64_reload_movcpv2dfdi; | |
5221 | ||
5222 | default: | |
5223 | gcc_unreachable (); | |
5224 | } | |
5225 | ||
5226 | gcc_unreachable (); | |
5227 | } | |
43e9d192 IB |
5228 | static reg_class_t |
5229 | aarch64_secondary_reload (bool in_p ATTRIBUTE_UNUSED, rtx x, | |
5230 | reg_class_t rclass, | |
ef4bddc2 | 5231 | machine_mode mode, |
43e9d192 IB |
5232 | secondary_reload_info *sri) |
5233 | { | |
b4f50fd4 RR |
5234 | |
5235 | /* If we have to disable direct literal pool loads and stores because the | |
5236 | function is too big, then we need a scratch register. */ | |
5237 | if (MEM_P (x) && GET_CODE (x) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (x) | |
5238 | && (SCALAR_FLOAT_MODE_P (GET_MODE (x)) | |
5239 | || targetm.vector_mode_supported_p (GET_MODE (x))) | |
9ee6540a | 5240 | && !aarch64_pcrelative_literal_loads) |
b4f50fd4 RR |
5241 | { |
5242 | sri->icode = aarch64_constant_pool_reload_icode (mode); | |
5243 | return NO_REGS; | |
5244 | } | |
5245 | ||
43e9d192 IB |
5246 | /* Without the TARGET_SIMD instructions we cannot move a Q register |
5247 | to a Q register directly. We need a scratch. */ | |
5248 | if (REG_P (x) && (mode == TFmode || mode == TImode) && mode == GET_MODE (x) | |
5249 | && FP_REGNUM_P (REGNO (x)) && !TARGET_SIMD | |
5250 | && reg_class_subset_p (rclass, FP_REGS)) | |
5251 | { | |
5252 | if (mode == TFmode) | |
5253 | sri->icode = CODE_FOR_aarch64_reload_movtf; | |
5254 | else if (mode == TImode) | |
5255 | sri->icode = CODE_FOR_aarch64_reload_movti; | |
5256 | return NO_REGS; | |
5257 | } | |
5258 | ||
5259 | /* A TFmode or TImode memory access should be handled via an FP_REGS | |
5260 | because AArch64 has richer addressing modes for LDR/STR instructions | |
5261 | than LDP/STP instructions. */ | |
d5726973 | 5262 | if (TARGET_FLOAT && rclass == GENERAL_REGS |
43e9d192 IB |
5263 | && GET_MODE_SIZE (mode) == 16 && MEM_P (x)) |
5264 | return FP_REGS; | |
5265 | ||
5266 | if (rclass == FP_REGS && (mode == TImode || mode == TFmode) && CONSTANT_P(x)) | |
a4a182c6 | 5267 | return GENERAL_REGS; |
43e9d192 IB |
5268 | |
5269 | return NO_REGS; | |
5270 | } | |
5271 | ||
5272 | static bool | |
5273 | aarch64_can_eliminate (const int from, const int to) | |
5274 | { | |
5275 | /* If we need a frame pointer, we must eliminate FRAME_POINTER_REGNUM into | |
5276 | HARD_FRAME_POINTER_REGNUM and not into STACK_POINTER_REGNUM. */ | |
5277 | ||
5278 | if (frame_pointer_needed) | |
5279 | { | |
5280 | if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
5281 | return true; | |
5282 | if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) | |
5283 | return false; | |
5284 | if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM | |
5285 | && !cfun->calls_alloca) | |
5286 | return true; | |
5287 | if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
5288 | return true; | |
0b7f8166 MS |
5289 | |
5290 | return false; | |
43e9d192 | 5291 | } |
1c923b60 JW |
5292 | else |
5293 | { | |
5294 | /* If we decided that we didn't need a leaf frame pointer but then used | |
5295 | LR in the function, then we'll want a frame pointer after all, so | |
5296 | prevent this elimination to ensure a frame pointer is used. */ | |
5297 | if (to == STACK_POINTER_REGNUM | |
5298 | && flag_omit_leaf_frame_pointer | |
5299 | && df_regs_ever_live_p (LR_REGNUM)) | |
5300 | return false; | |
5301 | } | |
777e6976 | 5302 | |
43e9d192 IB |
5303 | return true; |
5304 | } | |
5305 | ||
5306 | HOST_WIDE_INT | |
5307 | aarch64_initial_elimination_offset (unsigned from, unsigned to) | |
5308 | { | |
43e9d192 | 5309 | aarch64_layout_frame (); |
78c29983 MS |
5310 | |
5311 | if (to == HARD_FRAME_POINTER_REGNUM) | |
5312 | { | |
5313 | if (from == ARG_POINTER_REGNUM) | |
71bfb77a | 5314 | return cfun->machine->frame.hard_fp_offset; |
78c29983 MS |
5315 | |
5316 | if (from == FRAME_POINTER_REGNUM) | |
71bfb77a WD |
5317 | return cfun->machine->frame.hard_fp_offset |
5318 | - cfun->machine->frame.locals_offset; | |
78c29983 MS |
5319 | } |
5320 | ||
5321 | if (to == STACK_POINTER_REGNUM) | |
5322 | { | |
5323 | if (from == FRAME_POINTER_REGNUM) | |
71bfb77a WD |
5324 | return cfun->machine->frame.frame_size |
5325 | - cfun->machine->frame.locals_offset; | |
78c29983 MS |
5326 | } |
5327 | ||
1c960e02 | 5328 | return cfun->machine->frame.frame_size; |
43e9d192 IB |
5329 | } |
5330 | ||
43e9d192 IB |
5331 | /* Implement RETURN_ADDR_RTX. We do not support moving back to a |
5332 | previous frame. */ | |
5333 | ||
5334 | rtx | |
5335 | aarch64_return_addr (int count, rtx frame ATTRIBUTE_UNUSED) | |
5336 | { | |
5337 | if (count != 0) | |
5338 | return const0_rtx; | |
5339 | return get_hard_reg_initial_val (Pmode, LR_REGNUM); | |
5340 | } | |
5341 | ||
5342 | ||
5343 | static void | |
5344 | aarch64_asm_trampoline_template (FILE *f) | |
5345 | { | |
28514dda YZ |
5346 | if (TARGET_ILP32) |
5347 | { | |
5348 | asm_fprintf (f, "\tldr\tw%d, .+16\n", IP1_REGNUM - R0_REGNUM); | |
5349 | asm_fprintf (f, "\tldr\tw%d, .+16\n", STATIC_CHAIN_REGNUM - R0_REGNUM); | |
5350 | } | |
5351 | else | |
5352 | { | |
5353 | asm_fprintf (f, "\tldr\t%s, .+16\n", reg_names [IP1_REGNUM]); | |
5354 | asm_fprintf (f, "\tldr\t%s, .+20\n", reg_names [STATIC_CHAIN_REGNUM]); | |
5355 | } | |
01a3a324 | 5356 | asm_fprintf (f, "\tbr\t%s\n", reg_names [IP1_REGNUM]); |
43e9d192 | 5357 | assemble_aligned_integer (4, const0_rtx); |
28514dda YZ |
5358 | assemble_aligned_integer (POINTER_BYTES, const0_rtx); |
5359 | assemble_aligned_integer (POINTER_BYTES, const0_rtx); | |
43e9d192 IB |
5360 | } |
5361 | ||
5362 | static void | |
5363 | aarch64_trampoline_init (rtx m_tramp, tree fndecl, rtx chain_value) | |
5364 | { | |
5365 | rtx fnaddr, mem, a_tramp; | |
28514dda | 5366 | const int tramp_code_sz = 16; |
43e9d192 IB |
5367 | |
5368 | /* Don't need to copy the trailing D-words, we fill those in below. */ | |
5369 | emit_block_move (m_tramp, assemble_trampoline_template (), | |
28514dda YZ |
5370 | GEN_INT (tramp_code_sz), BLOCK_OP_NORMAL); |
5371 | mem = adjust_address (m_tramp, ptr_mode, tramp_code_sz); | |
43e9d192 | 5372 | fnaddr = XEXP (DECL_RTL (fndecl), 0); |
28514dda YZ |
5373 | if (GET_MODE (fnaddr) != ptr_mode) |
5374 | fnaddr = convert_memory_address (ptr_mode, fnaddr); | |
43e9d192 IB |
5375 | emit_move_insn (mem, fnaddr); |
5376 | ||
28514dda | 5377 | mem = adjust_address (m_tramp, ptr_mode, tramp_code_sz + POINTER_BYTES); |
43e9d192 IB |
5378 | emit_move_insn (mem, chain_value); |
5379 | ||
5380 | /* XXX We should really define a "clear_cache" pattern and use | |
5381 | gen_clear_cache(). */ | |
5382 | a_tramp = XEXP (m_tramp, 0); | |
5383 | emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__clear_cache"), | |
28514dda YZ |
5384 | LCT_NORMAL, VOIDmode, 2, a_tramp, ptr_mode, |
5385 | plus_constant (ptr_mode, a_tramp, TRAMPOLINE_SIZE), | |
5386 | ptr_mode); | |
43e9d192 IB |
5387 | } |
5388 | ||
5389 | static unsigned char | |
ef4bddc2 | 5390 | aarch64_class_max_nregs (reg_class_t regclass, machine_mode mode) |
43e9d192 IB |
5391 | { |
5392 | switch (regclass) | |
5393 | { | |
fee9ba42 | 5394 | case CALLER_SAVE_REGS: |
43e9d192 IB |
5395 | case POINTER_REGS: |
5396 | case GENERAL_REGS: | |
5397 | case ALL_REGS: | |
5398 | case FP_REGS: | |
5399 | case FP_LO_REGS: | |
5400 | return | |
7bd11911 KT |
5401 | aarch64_vector_mode_p (mode) |
5402 | ? (GET_MODE_SIZE (mode) + UNITS_PER_VREG - 1) / UNITS_PER_VREG | |
5403 | : (GET_MODE_SIZE (mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
43e9d192 IB |
5404 | case STACK_REG: |
5405 | return 1; | |
5406 | ||
5407 | case NO_REGS: | |
5408 | return 0; | |
5409 | ||
5410 | default: | |
5411 | break; | |
5412 | } | |
5413 | gcc_unreachable (); | |
5414 | } | |
5415 | ||
5416 | static reg_class_t | |
78d8b9f0 | 5417 | aarch64_preferred_reload_class (rtx x, reg_class_t regclass) |
43e9d192 | 5418 | { |
51bb310d | 5419 | if (regclass == POINTER_REGS) |
78d8b9f0 IB |
5420 | return GENERAL_REGS; |
5421 | ||
51bb310d MS |
5422 | if (regclass == STACK_REG) |
5423 | { | |
5424 | if (REG_P(x) | |
5425 | && reg_class_subset_p (REGNO_REG_CLASS (REGNO (x)), POINTER_REGS)) | |
5426 | return regclass; | |
5427 | ||
5428 | return NO_REGS; | |
5429 | } | |
5430 | ||
78d8b9f0 IB |
5431 | /* If it's an integer immediate that MOVI can't handle, then |
5432 | FP_REGS is not an option, so we return NO_REGS instead. */ | |
5433 | if (CONST_INT_P (x) && reg_class_subset_p (regclass, FP_REGS) | |
5434 | && !aarch64_simd_imm_scalar_p (x, GET_MODE (x))) | |
5435 | return NO_REGS; | |
5436 | ||
27bd251b IB |
5437 | /* Register eliminiation can result in a request for |
5438 | SP+constant->FP_REGS. We cannot support such operations which | |
5439 | use SP as source and an FP_REG as destination, so reject out | |
5440 | right now. */ | |
5441 | if (! reg_class_subset_p (regclass, GENERAL_REGS) && GET_CODE (x) == PLUS) | |
5442 | { | |
5443 | rtx lhs = XEXP (x, 0); | |
5444 | ||
5445 | /* Look through a possible SUBREG introduced by ILP32. */ | |
5446 | if (GET_CODE (lhs) == SUBREG) | |
5447 | lhs = SUBREG_REG (lhs); | |
5448 | ||
5449 | gcc_assert (REG_P (lhs)); | |
5450 | gcc_assert (reg_class_subset_p (REGNO_REG_CLASS (REGNO (lhs)), | |
5451 | POINTER_REGS)); | |
5452 | return NO_REGS; | |
5453 | } | |
5454 | ||
78d8b9f0 | 5455 | return regclass; |
43e9d192 IB |
5456 | } |
5457 | ||
5458 | void | |
5459 | aarch64_asm_output_labelref (FILE* f, const char *name) | |
5460 | { | |
5461 | asm_fprintf (f, "%U%s", name); | |
5462 | } | |
5463 | ||
5464 | static void | |
5465 | aarch64_elf_asm_constructor (rtx symbol, int priority) | |
5466 | { | |
5467 | if (priority == DEFAULT_INIT_PRIORITY) | |
5468 | default_ctor_section_asm_out_constructor (symbol, priority); | |
5469 | else | |
5470 | { | |
5471 | section *s; | |
5472 | char buf[18]; | |
5473 | snprintf (buf, sizeof (buf), ".init_array.%.5u", priority); | |
5474 | s = get_section (buf, SECTION_WRITE, NULL); | |
5475 | switch_to_section (s); | |
5476 | assemble_align (POINTER_SIZE); | |
28514dda | 5477 | assemble_aligned_integer (POINTER_BYTES, symbol); |
43e9d192 IB |
5478 | } |
5479 | } | |
5480 | ||
5481 | static void | |
5482 | aarch64_elf_asm_destructor (rtx symbol, int priority) | |
5483 | { | |
5484 | if (priority == DEFAULT_INIT_PRIORITY) | |
5485 | default_dtor_section_asm_out_destructor (symbol, priority); | |
5486 | else | |
5487 | { | |
5488 | section *s; | |
5489 | char buf[18]; | |
5490 | snprintf (buf, sizeof (buf), ".fini_array.%.5u", priority); | |
5491 | s = get_section (buf, SECTION_WRITE, NULL); | |
5492 | switch_to_section (s); | |
5493 | assemble_align (POINTER_SIZE); | |
28514dda | 5494 | assemble_aligned_integer (POINTER_BYTES, symbol); |
43e9d192 IB |
5495 | } |
5496 | } | |
5497 | ||
5498 | const char* | |
5499 | aarch64_output_casesi (rtx *operands) | |
5500 | { | |
5501 | char buf[100]; | |
5502 | char label[100]; | |
b32d5189 | 5503 | rtx diff_vec = PATTERN (NEXT_INSN (as_a <rtx_insn *> (operands[2]))); |
43e9d192 IB |
5504 | int index; |
5505 | static const char *const patterns[4][2] = | |
5506 | { | |
5507 | { | |
5508 | "ldrb\t%w3, [%0,%w1,uxtw]", | |
5509 | "add\t%3, %4, %w3, sxtb #2" | |
5510 | }, | |
5511 | { | |
5512 | "ldrh\t%w3, [%0,%w1,uxtw #1]", | |
5513 | "add\t%3, %4, %w3, sxth #2" | |
5514 | }, | |
5515 | { | |
5516 | "ldr\t%w3, [%0,%w1,uxtw #2]", | |
5517 | "add\t%3, %4, %w3, sxtw #2" | |
5518 | }, | |
5519 | /* We assume that DImode is only generated when not optimizing and | |
5520 | that we don't really need 64-bit address offsets. That would | |
5521 | imply an object file with 8GB of code in a single function! */ | |
5522 | { | |
5523 | "ldr\t%w3, [%0,%w1,uxtw #2]", | |
5524 | "add\t%3, %4, %w3, sxtw #2" | |
5525 | } | |
5526 | }; | |
5527 | ||
5528 | gcc_assert (GET_CODE (diff_vec) == ADDR_DIFF_VEC); | |
5529 | ||
5530 | index = exact_log2 (GET_MODE_SIZE (GET_MODE (diff_vec))); | |
5531 | ||
5532 | gcc_assert (index >= 0 && index <= 3); | |
5533 | ||
5534 | /* Need to implement table size reduction, by chaning the code below. */ | |
5535 | output_asm_insn (patterns[index][0], operands); | |
5536 | ASM_GENERATE_INTERNAL_LABEL (label, "Lrtx", CODE_LABEL_NUMBER (operands[2])); | |
5537 | snprintf (buf, sizeof (buf), | |
5538 | "adr\t%%4, %s", targetm.strip_name_encoding (label)); | |
5539 | output_asm_insn (buf, operands); | |
5540 | output_asm_insn (patterns[index][1], operands); | |
5541 | output_asm_insn ("br\t%3", operands); | |
5542 | assemble_label (asm_out_file, label); | |
5543 | return ""; | |
5544 | } | |
5545 | ||
5546 | ||
5547 | /* Return size in bits of an arithmetic operand which is shifted/scaled and | |
5548 | masked such that it is suitable for a UXTB, UXTH, or UXTW extend | |
5549 | operator. */ | |
5550 | ||
5551 | int | |
5552 | aarch64_uxt_size (int shift, HOST_WIDE_INT mask) | |
5553 | { | |
5554 | if (shift >= 0 && shift <= 3) | |
5555 | { | |
5556 | int size; | |
5557 | for (size = 8; size <= 32; size *= 2) | |
5558 | { | |
5559 | HOST_WIDE_INT bits = ((HOST_WIDE_INT)1U << size) - 1; | |
5560 | if (mask == bits << shift) | |
5561 | return size; | |
5562 | } | |
5563 | } | |
5564 | return 0; | |
5565 | } | |
5566 | ||
e78d485e RR |
5567 | /* Constant pools are per function only when PC relative |
5568 | literal loads are true or we are in the large memory | |
5569 | model. */ | |
5570 | ||
5571 | static inline bool | |
5572 | aarch64_can_use_per_function_literal_pools_p (void) | |
5573 | { | |
9ee6540a | 5574 | return (aarch64_pcrelative_literal_loads |
e78d485e RR |
5575 | || aarch64_cmodel == AARCH64_CMODEL_LARGE); |
5576 | } | |
5577 | ||
43e9d192 | 5578 | static bool |
e78d485e | 5579 | aarch64_use_blocks_for_constant_p (machine_mode, const_rtx) |
43e9d192 | 5580 | { |
3eece53d RR |
5581 | /* Fixme:: In an ideal world this would work similar |
5582 | to the logic in aarch64_select_rtx_section but this | |
5583 | breaks bootstrap in gcc go. For now we workaround | |
5584 | this by returning false here. */ | |
5585 | return false; | |
43e9d192 IB |
5586 | } |
5587 | ||
e78d485e RR |
5588 | /* Select appropriate section for constants depending |
5589 | on where we place literal pools. */ | |
5590 | ||
43e9d192 | 5591 | static section * |
e78d485e RR |
5592 | aarch64_select_rtx_section (machine_mode mode, |
5593 | rtx x, | |
5594 | unsigned HOST_WIDE_INT align) | |
43e9d192 | 5595 | { |
e78d485e RR |
5596 | if (aarch64_can_use_per_function_literal_pools_p ()) |
5597 | return function_section (current_function_decl); | |
43e9d192 | 5598 | |
e78d485e RR |
5599 | return default_elf_select_rtx_section (mode, x, align); |
5600 | } | |
43e9d192 | 5601 | |
5fca7b66 RH |
5602 | /* Implement ASM_OUTPUT_POOL_EPILOGUE. */ |
5603 | void | |
5604 | aarch64_asm_output_pool_epilogue (FILE *f, const char *, tree, | |
5605 | HOST_WIDE_INT offset) | |
5606 | { | |
5607 | /* When using per-function literal pools, we must ensure that any code | |
5608 | section is aligned to the minimal instruction length, lest we get | |
5609 | errors from the assembler re "unaligned instructions". */ | |
5610 | if ((offset & 3) && aarch64_can_use_per_function_literal_pools_p ()) | |
5611 | ASM_OUTPUT_ALIGN (f, 2); | |
5612 | } | |
5613 | ||
43e9d192 IB |
5614 | /* Costs. */ |
5615 | ||
5616 | /* Helper function for rtx cost calculation. Strip a shift expression | |
5617 | from X. Returns the inner operand if successful, or the original | |
5618 | expression on failure. */ | |
5619 | static rtx | |
5620 | aarch64_strip_shift (rtx x) | |
5621 | { | |
5622 | rtx op = x; | |
5623 | ||
57b77d46 RE |
5624 | /* We accept both ROTATERT and ROTATE: since the RHS must be a constant |
5625 | we can convert both to ROR during final output. */ | |
43e9d192 IB |
5626 | if ((GET_CODE (op) == ASHIFT |
5627 | || GET_CODE (op) == ASHIFTRT | |
57b77d46 RE |
5628 | || GET_CODE (op) == LSHIFTRT |
5629 | || GET_CODE (op) == ROTATERT | |
5630 | || GET_CODE (op) == ROTATE) | |
43e9d192 IB |
5631 | && CONST_INT_P (XEXP (op, 1))) |
5632 | return XEXP (op, 0); | |
5633 | ||
5634 | if (GET_CODE (op) == MULT | |
5635 | && CONST_INT_P (XEXP (op, 1)) | |
5636 | && ((unsigned) exact_log2 (INTVAL (XEXP (op, 1)))) < 64) | |
5637 | return XEXP (op, 0); | |
5638 | ||
5639 | return x; | |
5640 | } | |
5641 | ||
4745e701 | 5642 | /* Helper function for rtx cost calculation. Strip an extend |
43e9d192 IB |
5643 | expression from X. Returns the inner operand if successful, or the |
5644 | original expression on failure. We deal with a number of possible | |
5645 | canonicalization variations here. */ | |
5646 | static rtx | |
4745e701 | 5647 | aarch64_strip_extend (rtx x) |
43e9d192 IB |
5648 | { |
5649 | rtx op = x; | |
5650 | ||
5651 | /* Zero and sign extraction of a widened value. */ | |
5652 | if ((GET_CODE (op) == ZERO_EXTRACT || GET_CODE (op) == SIGN_EXTRACT) | |
5653 | && XEXP (op, 2) == const0_rtx | |
4745e701 | 5654 | && GET_CODE (XEXP (op, 0)) == MULT |
43e9d192 IB |
5655 | && aarch64_is_extend_from_extract (GET_MODE (op), XEXP (XEXP (op, 0), 1), |
5656 | XEXP (op, 1))) | |
5657 | return XEXP (XEXP (op, 0), 0); | |
5658 | ||
5659 | /* It can also be represented (for zero-extend) as an AND with an | |
5660 | immediate. */ | |
5661 | if (GET_CODE (op) == AND | |
5662 | && GET_CODE (XEXP (op, 0)) == MULT | |
5663 | && CONST_INT_P (XEXP (XEXP (op, 0), 1)) | |
5664 | && CONST_INT_P (XEXP (op, 1)) | |
5665 | && aarch64_uxt_size (exact_log2 (INTVAL (XEXP (XEXP (op, 0), 1))), | |
5666 | INTVAL (XEXP (op, 1))) != 0) | |
5667 | return XEXP (XEXP (op, 0), 0); | |
5668 | ||
5669 | /* Now handle extended register, as this may also have an optional | |
5670 | left shift by 1..4. */ | |
5671 | if (GET_CODE (op) == ASHIFT | |
5672 | && CONST_INT_P (XEXP (op, 1)) | |
5673 | && ((unsigned HOST_WIDE_INT) INTVAL (XEXP (op, 1))) <= 4) | |
5674 | op = XEXP (op, 0); | |
5675 | ||
5676 | if (GET_CODE (op) == ZERO_EXTEND | |
5677 | || GET_CODE (op) == SIGN_EXTEND) | |
5678 | op = XEXP (op, 0); | |
5679 | ||
5680 | if (op != x) | |
5681 | return op; | |
5682 | ||
4745e701 JG |
5683 | return x; |
5684 | } | |
5685 | ||
0a78ebe4 KT |
5686 | /* Return true iff CODE is a shift supported in combination |
5687 | with arithmetic instructions. */ | |
4d1919ed | 5688 | |
0a78ebe4 KT |
5689 | static bool |
5690 | aarch64_shift_p (enum rtx_code code) | |
5691 | { | |
5692 | return code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT; | |
5693 | } | |
5694 | ||
4745e701 | 5695 | /* Helper function for rtx cost calculation. Calculate the cost of |
0a78ebe4 KT |
5696 | a MULT or ASHIFT, which may be part of a compound PLUS/MINUS rtx. |
5697 | Return the calculated cost of the expression, recursing manually in to | |
4745e701 JG |
5698 | operands where needed. */ |
5699 | ||
5700 | static int | |
e548c9df | 5701 | aarch64_rtx_mult_cost (rtx x, enum rtx_code code, int outer, bool speed) |
4745e701 JG |
5702 | { |
5703 | rtx op0, op1; | |
5704 | const struct cpu_cost_table *extra_cost | |
b175b679 | 5705 | = aarch64_tune_params.insn_extra_cost; |
4745e701 | 5706 | int cost = 0; |
0a78ebe4 | 5707 | bool compound_p = (outer == PLUS || outer == MINUS); |
ef4bddc2 | 5708 | machine_mode mode = GET_MODE (x); |
4745e701 JG |
5709 | |
5710 | gcc_checking_assert (code == MULT); | |
5711 | ||
5712 | op0 = XEXP (x, 0); | |
5713 | op1 = XEXP (x, 1); | |
5714 | ||
5715 | if (VECTOR_MODE_P (mode)) | |
5716 | mode = GET_MODE_INNER (mode); | |
5717 | ||
5718 | /* Integer multiply/fma. */ | |
5719 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
5720 | { | |
5721 | /* The multiply will be canonicalized as a shift, cost it as such. */ | |
0a78ebe4 KT |
5722 | if (aarch64_shift_p (GET_CODE (x)) |
5723 | || (CONST_INT_P (op1) | |
5724 | && exact_log2 (INTVAL (op1)) > 0)) | |
4745e701 | 5725 | { |
0a78ebe4 KT |
5726 | bool is_extend = GET_CODE (op0) == ZERO_EXTEND |
5727 | || GET_CODE (op0) == SIGN_EXTEND; | |
4745e701 JG |
5728 | if (speed) |
5729 | { | |
0a78ebe4 KT |
5730 | if (compound_p) |
5731 | { | |
5732 | if (REG_P (op1)) | |
5733 | /* ARITH + shift-by-register. */ | |
5734 | cost += extra_cost->alu.arith_shift_reg; | |
5735 | else if (is_extend) | |
5736 | /* ARITH + extended register. We don't have a cost field | |
5737 | for ARITH+EXTEND+SHIFT, so use extend_arith here. */ | |
5738 | cost += extra_cost->alu.extend_arith; | |
5739 | else | |
5740 | /* ARITH + shift-by-immediate. */ | |
5741 | cost += extra_cost->alu.arith_shift; | |
5742 | } | |
4745e701 JG |
5743 | else |
5744 | /* LSL (immediate). */ | |
0a78ebe4 KT |
5745 | cost += extra_cost->alu.shift; |
5746 | ||
4745e701 | 5747 | } |
0a78ebe4 KT |
5748 | /* Strip extends as we will have costed them in the case above. */ |
5749 | if (is_extend) | |
5750 | op0 = aarch64_strip_extend (op0); | |
4745e701 | 5751 | |
e548c9df | 5752 | cost += rtx_cost (op0, VOIDmode, code, 0, speed); |
4745e701 JG |
5753 | |
5754 | return cost; | |
5755 | } | |
5756 | ||
d2ac256b KT |
5757 | /* MNEG or [US]MNEGL. Extract the NEG operand and indicate that it's a |
5758 | compound and let the below cases handle it. After all, MNEG is a | |
5759 | special-case alias of MSUB. */ | |
5760 | if (GET_CODE (op0) == NEG) | |
5761 | { | |
5762 | op0 = XEXP (op0, 0); | |
5763 | compound_p = true; | |
5764 | } | |
5765 | ||
4745e701 JG |
5766 | /* Integer multiplies or FMAs have zero/sign extending variants. */ |
5767 | if ((GET_CODE (op0) == ZERO_EXTEND | |
5768 | && GET_CODE (op1) == ZERO_EXTEND) | |
5769 | || (GET_CODE (op0) == SIGN_EXTEND | |
5770 | && GET_CODE (op1) == SIGN_EXTEND)) | |
5771 | { | |
e548c9df AM |
5772 | cost += rtx_cost (XEXP (op0, 0), VOIDmode, MULT, 0, speed); |
5773 | cost += rtx_cost (XEXP (op1, 0), VOIDmode, MULT, 1, speed); | |
4745e701 JG |
5774 | |
5775 | if (speed) | |
5776 | { | |
0a78ebe4 | 5777 | if (compound_p) |
d2ac256b | 5778 | /* SMADDL/UMADDL/UMSUBL/SMSUBL. */ |
4745e701 JG |
5779 | cost += extra_cost->mult[0].extend_add; |
5780 | else | |
5781 | /* MUL/SMULL/UMULL. */ | |
5782 | cost += extra_cost->mult[0].extend; | |
5783 | } | |
5784 | ||
5785 | return cost; | |
5786 | } | |
5787 | ||
d2ac256b | 5788 | /* This is either an integer multiply or a MADD. In both cases |
4745e701 | 5789 | we want to recurse and cost the operands. */ |
e548c9df AM |
5790 | cost += rtx_cost (op0, mode, MULT, 0, speed); |
5791 | cost += rtx_cost (op1, mode, MULT, 1, speed); | |
4745e701 JG |
5792 | |
5793 | if (speed) | |
5794 | { | |
0a78ebe4 | 5795 | if (compound_p) |
d2ac256b | 5796 | /* MADD/MSUB. */ |
4745e701 JG |
5797 | cost += extra_cost->mult[mode == DImode].add; |
5798 | else | |
5799 | /* MUL. */ | |
5800 | cost += extra_cost->mult[mode == DImode].simple; | |
5801 | } | |
5802 | ||
5803 | return cost; | |
5804 | } | |
5805 | else | |
5806 | { | |
5807 | if (speed) | |
5808 | { | |
3d840f7d | 5809 | /* Floating-point FMA/FMUL can also support negations of the |
d318517d SN |
5810 | operands, unless the rounding mode is upward or downward in |
5811 | which case FNMUL is different than FMUL with operand negation. */ | |
5812 | bool neg0 = GET_CODE (op0) == NEG; | |
5813 | bool neg1 = GET_CODE (op1) == NEG; | |
5814 | if (compound_p || !flag_rounding_math || (neg0 && neg1)) | |
5815 | { | |
5816 | if (neg0) | |
5817 | op0 = XEXP (op0, 0); | |
5818 | if (neg1) | |
5819 | op1 = XEXP (op1, 0); | |
5820 | } | |
4745e701 | 5821 | |
0a78ebe4 | 5822 | if (compound_p) |
4745e701 JG |
5823 | /* FMADD/FNMADD/FNMSUB/FMSUB. */ |
5824 | cost += extra_cost->fp[mode == DFmode].fma; | |
5825 | else | |
3d840f7d | 5826 | /* FMUL/FNMUL. */ |
4745e701 JG |
5827 | cost += extra_cost->fp[mode == DFmode].mult; |
5828 | } | |
5829 | ||
e548c9df AM |
5830 | cost += rtx_cost (op0, mode, MULT, 0, speed); |
5831 | cost += rtx_cost (op1, mode, MULT, 1, speed); | |
4745e701 JG |
5832 | return cost; |
5833 | } | |
43e9d192 IB |
5834 | } |
5835 | ||
67747367 JG |
5836 | static int |
5837 | aarch64_address_cost (rtx x, | |
ef4bddc2 | 5838 | machine_mode mode, |
67747367 JG |
5839 | addr_space_t as ATTRIBUTE_UNUSED, |
5840 | bool speed) | |
5841 | { | |
5842 | enum rtx_code c = GET_CODE (x); | |
b175b679 | 5843 | const struct cpu_addrcost_table *addr_cost = aarch64_tune_params.addr_cost; |
67747367 JG |
5844 | struct aarch64_address_info info; |
5845 | int cost = 0; | |
5846 | info.shift = 0; | |
5847 | ||
5848 | if (!aarch64_classify_address (&info, x, mode, c, false)) | |
5849 | { | |
5850 | if (GET_CODE (x) == CONST || GET_CODE (x) == SYMBOL_REF) | |
5851 | { | |
5852 | /* This is a CONST or SYMBOL ref which will be split | |
5853 | in a different way depending on the code model in use. | |
5854 | Cost it through the generic infrastructure. */ | |
e548c9df | 5855 | int cost_symbol_ref = rtx_cost (x, Pmode, MEM, 1, speed); |
67747367 JG |
5856 | /* Divide through by the cost of one instruction to |
5857 | bring it to the same units as the address costs. */ | |
5858 | cost_symbol_ref /= COSTS_N_INSNS (1); | |
5859 | /* The cost is then the cost of preparing the address, | |
5860 | followed by an immediate (possibly 0) offset. */ | |
5861 | return cost_symbol_ref + addr_cost->imm_offset; | |
5862 | } | |
5863 | else | |
5864 | { | |
5865 | /* This is most likely a jump table from a case | |
5866 | statement. */ | |
5867 | return addr_cost->register_offset; | |
5868 | } | |
5869 | } | |
5870 | ||
5871 | switch (info.type) | |
5872 | { | |
5873 | case ADDRESS_LO_SUM: | |
5874 | case ADDRESS_SYMBOLIC: | |
5875 | case ADDRESS_REG_IMM: | |
5876 | cost += addr_cost->imm_offset; | |
5877 | break; | |
5878 | ||
5879 | case ADDRESS_REG_WB: | |
5880 | if (c == PRE_INC || c == PRE_DEC || c == PRE_MODIFY) | |
5881 | cost += addr_cost->pre_modify; | |
5882 | else if (c == POST_INC || c == POST_DEC || c == POST_MODIFY) | |
5883 | cost += addr_cost->post_modify; | |
5884 | else | |
5885 | gcc_unreachable (); | |
5886 | ||
5887 | break; | |
5888 | ||
5889 | case ADDRESS_REG_REG: | |
5890 | cost += addr_cost->register_offset; | |
5891 | break; | |
5892 | ||
67747367 | 5893 | case ADDRESS_REG_SXTW: |
783879e6 EM |
5894 | cost += addr_cost->register_sextend; |
5895 | break; | |
5896 | ||
5897 | case ADDRESS_REG_UXTW: | |
5898 | cost += addr_cost->register_zextend; | |
67747367 JG |
5899 | break; |
5900 | ||
5901 | default: | |
5902 | gcc_unreachable (); | |
5903 | } | |
5904 | ||
5905 | ||
5906 | if (info.shift > 0) | |
5907 | { | |
5908 | /* For the sake of calculating the cost of the shifted register | |
5909 | component, we can treat same sized modes in the same way. */ | |
5910 | switch (GET_MODE_BITSIZE (mode)) | |
5911 | { | |
5912 | case 16: | |
5913 | cost += addr_cost->addr_scale_costs.hi; | |
5914 | break; | |
5915 | ||
5916 | case 32: | |
5917 | cost += addr_cost->addr_scale_costs.si; | |
5918 | break; | |
5919 | ||
5920 | case 64: | |
5921 | cost += addr_cost->addr_scale_costs.di; | |
5922 | break; | |
5923 | ||
5924 | /* We can't tell, or this is a 128-bit vector. */ | |
5925 | default: | |
5926 | cost += addr_cost->addr_scale_costs.ti; | |
5927 | break; | |
5928 | } | |
5929 | } | |
5930 | ||
5931 | return cost; | |
5932 | } | |
5933 | ||
b9066f5a MW |
5934 | /* Return the cost of a branch. If SPEED_P is true then the compiler is |
5935 | optimizing for speed. If PREDICTABLE_P is true then the branch is predicted | |
5936 | to be taken. */ | |
5937 | ||
5938 | int | |
5939 | aarch64_branch_cost (bool speed_p, bool predictable_p) | |
5940 | { | |
5941 | /* When optimizing for speed, use the cost of unpredictable branches. */ | |
5942 | const struct cpu_branch_cost *branch_costs = | |
b175b679 | 5943 | aarch64_tune_params.branch_costs; |
b9066f5a MW |
5944 | |
5945 | if (!speed_p || predictable_p) | |
5946 | return branch_costs->predictable; | |
5947 | else | |
5948 | return branch_costs->unpredictable; | |
5949 | } | |
5950 | ||
7cc2145f JG |
5951 | /* Return true if the RTX X in mode MODE is a zero or sign extract |
5952 | usable in an ADD or SUB (extended register) instruction. */ | |
5953 | static bool | |
ef4bddc2 | 5954 | aarch64_rtx_arith_op_extract_p (rtx x, machine_mode mode) |
7cc2145f JG |
5955 | { |
5956 | /* Catch add with a sign extract. | |
5957 | This is add_<optab><mode>_multp2. */ | |
5958 | if (GET_CODE (x) == SIGN_EXTRACT | |
5959 | || GET_CODE (x) == ZERO_EXTRACT) | |
5960 | { | |
5961 | rtx op0 = XEXP (x, 0); | |
5962 | rtx op1 = XEXP (x, 1); | |
5963 | rtx op2 = XEXP (x, 2); | |
5964 | ||
5965 | if (GET_CODE (op0) == MULT | |
5966 | && CONST_INT_P (op1) | |
5967 | && op2 == const0_rtx | |
5968 | && CONST_INT_P (XEXP (op0, 1)) | |
5969 | && aarch64_is_extend_from_extract (mode, | |
5970 | XEXP (op0, 1), | |
5971 | op1)) | |
5972 | { | |
5973 | return true; | |
5974 | } | |
5975 | } | |
e47c4031 KT |
5976 | /* The simple case <ARITH>, XD, XN, XM, [us]xt. |
5977 | No shift. */ | |
5978 | else if (GET_CODE (x) == SIGN_EXTEND | |
5979 | || GET_CODE (x) == ZERO_EXTEND) | |
5980 | return REG_P (XEXP (x, 0)); | |
7cc2145f JG |
5981 | |
5982 | return false; | |
5983 | } | |
5984 | ||
61263118 KT |
5985 | static bool |
5986 | aarch64_frint_unspec_p (unsigned int u) | |
5987 | { | |
5988 | switch (u) | |
5989 | { | |
5990 | case UNSPEC_FRINTZ: | |
5991 | case UNSPEC_FRINTP: | |
5992 | case UNSPEC_FRINTM: | |
5993 | case UNSPEC_FRINTA: | |
5994 | case UNSPEC_FRINTN: | |
5995 | case UNSPEC_FRINTX: | |
5996 | case UNSPEC_FRINTI: | |
5997 | return true; | |
5998 | ||
5999 | default: | |
6000 | return false; | |
6001 | } | |
6002 | } | |
6003 | ||
fb0cb7fa KT |
6004 | /* Return true iff X is an rtx that will match an extr instruction |
6005 | i.e. as described in the *extr<mode>5_insn family of patterns. | |
6006 | OP0 and OP1 will be set to the operands of the shifts involved | |
6007 | on success and will be NULL_RTX otherwise. */ | |
6008 | ||
6009 | static bool | |
6010 | aarch64_extr_rtx_p (rtx x, rtx *res_op0, rtx *res_op1) | |
6011 | { | |
6012 | rtx op0, op1; | |
6013 | machine_mode mode = GET_MODE (x); | |
6014 | ||
6015 | *res_op0 = NULL_RTX; | |
6016 | *res_op1 = NULL_RTX; | |
6017 | ||
6018 | if (GET_CODE (x) != IOR) | |
6019 | return false; | |
6020 | ||
6021 | op0 = XEXP (x, 0); | |
6022 | op1 = XEXP (x, 1); | |
6023 | ||
6024 | if ((GET_CODE (op0) == ASHIFT && GET_CODE (op1) == LSHIFTRT) | |
6025 | || (GET_CODE (op1) == ASHIFT && GET_CODE (op0) == LSHIFTRT)) | |
6026 | { | |
6027 | /* Canonicalise locally to ashift in op0, lshiftrt in op1. */ | |
6028 | if (GET_CODE (op1) == ASHIFT) | |
6029 | std::swap (op0, op1); | |
6030 | ||
6031 | if (!CONST_INT_P (XEXP (op0, 1)) || !CONST_INT_P (XEXP (op1, 1))) | |
6032 | return false; | |
6033 | ||
6034 | unsigned HOST_WIDE_INT shft_amnt_0 = UINTVAL (XEXP (op0, 1)); | |
6035 | unsigned HOST_WIDE_INT shft_amnt_1 = UINTVAL (XEXP (op1, 1)); | |
6036 | ||
6037 | if (shft_amnt_0 < GET_MODE_BITSIZE (mode) | |
6038 | && shft_amnt_0 + shft_amnt_1 == GET_MODE_BITSIZE (mode)) | |
6039 | { | |
6040 | *res_op0 = XEXP (op0, 0); | |
6041 | *res_op1 = XEXP (op1, 0); | |
6042 | return true; | |
6043 | } | |
6044 | } | |
6045 | ||
6046 | return false; | |
6047 | } | |
6048 | ||
2d5ffe46 AP |
6049 | /* Calculate the cost of calculating (if_then_else (OP0) (OP1) (OP2)), |
6050 | storing it in *COST. Result is true if the total cost of the operation | |
6051 | has now been calculated. */ | |
6052 | static bool | |
6053 | aarch64_if_then_else_costs (rtx op0, rtx op1, rtx op2, int *cost, bool speed) | |
6054 | { | |
b9e3afe9 AP |
6055 | rtx inner; |
6056 | rtx comparator; | |
6057 | enum rtx_code cmpcode; | |
6058 | ||
6059 | if (COMPARISON_P (op0)) | |
6060 | { | |
6061 | inner = XEXP (op0, 0); | |
6062 | comparator = XEXP (op0, 1); | |
6063 | cmpcode = GET_CODE (op0); | |
6064 | } | |
6065 | else | |
6066 | { | |
6067 | inner = op0; | |
6068 | comparator = const0_rtx; | |
6069 | cmpcode = NE; | |
6070 | } | |
6071 | ||
2d5ffe46 AP |
6072 | if (GET_CODE (op1) == PC || GET_CODE (op2) == PC) |
6073 | { | |
6074 | /* Conditional branch. */ | |
b9e3afe9 | 6075 | if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_CC) |
2d5ffe46 AP |
6076 | return true; |
6077 | else | |
6078 | { | |
b9e3afe9 | 6079 | if (cmpcode == NE || cmpcode == EQ) |
2d5ffe46 | 6080 | { |
2d5ffe46 AP |
6081 | if (comparator == const0_rtx) |
6082 | { | |
6083 | /* TBZ/TBNZ/CBZ/CBNZ. */ | |
6084 | if (GET_CODE (inner) == ZERO_EXTRACT) | |
6085 | /* TBZ/TBNZ. */ | |
e548c9df AM |
6086 | *cost += rtx_cost (XEXP (inner, 0), VOIDmode, |
6087 | ZERO_EXTRACT, 0, speed); | |
6088 | else | |
6089 | /* CBZ/CBNZ. */ | |
6090 | *cost += rtx_cost (inner, VOIDmode, cmpcode, 0, speed); | |
2d5ffe46 AP |
6091 | |
6092 | return true; | |
6093 | } | |
6094 | } | |
b9e3afe9 | 6095 | else if (cmpcode == LT || cmpcode == GE) |
2d5ffe46 | 6096 | { |
2d5ffe46 AP |
6097 | /* TBZ/TBNZ. */ |
6098 | if (comparator == const0_rtx) | |
6099 | return true; | |
6100 | } | |
6101 | } | |
6102 | } | |
b9e3afe9 | 6103 | else if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_CC) |
2d5ffe46 | 6104 | { |
786298dc | 6105 | /* CCMP. */ |
6dfeb7ce | 6106 | if (GET_CODE (op1) == COMPARE) |
786298dc WD |
6107 | { |
6108 | /* Increase cost of CCMP reg, 0, imm, CC to prefer CMP reg, 0. */ | |
6109 | if (XEXP (op1, 1) == const0_rtx) | |
6110 | *cost += 1; | |
6111 | if (speed) | |
6112 | { | |
6113 | machine_mode mode = GET_MODE (XEXP (op1, 0)); | |
6114 | const struct cpu_cost_table *extra_cost | |
6115 | = aarch64_tune_params.insn_extra_cost; | |
6116 | ||
6117 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
6118 | *cost += extra_cost->alu.arith; | |
6119 | else | |
6120 | *cost += extra_cost->fp[mode == DFmode].compare; | |
6121 | } | |
6122 | return true; | |
6123 | } | |
6124 | ||
2d5ffe46 AP |
6125 | /* It's a conditional operation based on the status flags, |
6126 | so it must be some flavor of CSEL. */ | |
6127 | ||
6128 | /* CSNEG, CSINV, and CSINC are handled for free as part of CSEL. */ | |
6129 | if (GET_CODE (op1) == NEG | |
6130 | || GET_CODE (op1) == NOT | |
6131 | || (GET_CODE (op1) == PLUS && XEXP (op1, 1) == const1_rtx)) | |
6132 | op1 = XEXP (op1, 0); | |
bad00732 KT |
6133 | else if (GET_CODE (op1) == ZERO_EXTEND && GET_CODE (op2) == ZERO_EXTEND) |
6134 | { | |
6135 | /* CSEL with zero-extension (*cmovdi_insn_uxtw). */ | |
6136 | op1 = XEXP (op1, 0); | |
6137 | op2 = XEXP (op2, 0); | |
6138 | } | |
2d5ffe46 | 6139 | |
e548c9df AM |
6140 | *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed); |
6141 | *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed); | |
2d5ffe46 AP |
6142 | return true; |
6143 | } | |
6144 | ||
6145 | /* We don't know what this is, cost all operands. */ | |
6146 | return false; | |
6147 | } | |
6148 | ||
283b6c85 KT |
6149 | /* Check whether X is a bitfield operation of the form shift + extend that |
6150 | maps down to a UBFIZ/SBFIZ/UBFX/SBFX instruction. If so, return the | |
6151 | operand to which the bitfield operation is applied. Otherwise return | |
6152 | NULL_RTX. */ | |
6153 | ||
6154 | static rtx | |
6155 | aarch64_extend_bitfield_pattern_p (rtx x) | |
6156 | { | |
6157 | rtx_code outer_code = GET_CODE (x); | |
6158 | machine_mode outer_mode = GET_MODE (x); | |
6159 | ||
6160 | if (outer_code != ZERO_EXTEND && outer_code != SIGN_EXTEND | |
6161 | && outer_mode != SImode && outer_mode != DImode) | |
6162 | return NULL_RTX; | |
6163 | ||
6164 | rtx inner = XEXP (x, 0); | |
6165 | rtx_code inner_code = GET_CODE (inner); | |
6166 | machine_mode inner_mode = GET_MODE (inner); | |
6167 | rtx op = NULL_RTX; | |
6168 | ||
6169 | switch (inner_code) | |
6170 | { | |
6171 | case ASHIFT: | |
6172 | if (CONST_INT_P (XEXP (inner, 1)) | |
6173 | && (inner_mode == QImode || inner_mode == HImode)) | |
6174 | op = XEXP (inner, 0); | |
6175 | break; | |
6176 | case LSHIFTRT: | |
6177 | if (outer_code == ZERO_EXTEND && CONST_INT_P (XEXP (inner, 1)) | |
6178 | && (inner_mode == QImode || inner_mode == HImode)) | |
6179 | op = XEXP (inner, 0); | |
6180 | break; | |
6181 | case ASHIFTRT: | |
6182 | if (outer_code == SIGN_EXTEND && CONST_INT_P (XEXP (inner, 1)) | |
6183 | && (inner_mode == QImode || inner_mode == HImode)) | |
6184 | op = XEXP (inner, 0); | |
6185 | break; | |
6186 | default: | |
6187 | break; | |
6188 | } | |
6189 | ||
6190 | return op; | |
6191 | } | |
6192 | ||
8c83f71d KT |
6193 | /* Return true if the mask and a shift amount from an RTX of the form |
6194 | (x << SHFT_AMNT) & MASK are valid to combine into a UBFIZ instruction of | |
6195 | mode MODE. See the *andim_ashift<mode>_bfiz pattern. */ | |
6196 | ||
6197 | bool | |
6198 | aarch64_mask_and_shift_for_ubfiz_p (machine_mode mode, rtx mask, rtx shft_amnt) | |
6199 | { | |
6200 | return CONST_INT_P (mask) && CONST_INT_P (shft_amnt) | |
6201 | && INTVAL (shft_amnt) < GET_MODE_BITSIZE (mode) | |
6202 | && exact_log2 ((INTVAL (mask) >> INTVAL (shft_amnt)) + 1) >= 0 | |
6203 | && (INTVAL (mask) & ((1 << INTVAL (shft_amnt)) - 1)) == 0; | |
6204 | } | |
6205 | ||
43e9d192 IB |
6206 | /* Calculate the cost of calculating X, storing it in *COST. Result |
6207 | is true if the total cost of the operation has now been calculated. */ | |
6208 | static bool | |
e548c9df | 6209 | aarch64_rtx_costs (rtx x, machine_mode mode, int outer ATTRIBUTE_UNUSED, |
43e9d192 IB |
6210 | int param ATTRIBUTE_UNUSED, int *cost, bool speed) |
6211 | { | |
a8eecd00 | 6212 | rtx op0, op1, op2; |
73250c4c | 6213 | const struct cpu_cost_table *extra_cost |
b175b679 | 6214 | = aarch64_tune_params.insn_extra_cost; |
e548c9df | 6215 | int code = GET_CODE (x); |
43e9d192 | 6216 | |
7fc5ef02 JG |
6217 | /* By default, assume that everything has equivalent cost to the |
6218 | cheapest instruction. Any additional costs are applied as a delta | |
6219 | above this default. */ | |
6220 | *cost = COSTS_N_INSNS (1); | |
6221 | ||
43e9d192 IB |
6222 | switch (code) |
6223 | { | |
6224 | case SET: | |
ba123b0d JG |
6225 | /* The cost depends entirely on the operands to SET. */ |
6226 | *cost = 0; | |
43e9d192 IB |
6227 | op0 = SET_DEST (x); |
6228 | op1 = SET_SRC (x); | |
6229 | ||
6230 | switch (GET_CODE (op0)) | |
6231 | { | |
6232 | case MEM: | |
6233 | if (speed) | |
2961177e JG |
6234 | { |
6235 | rtx address = XEXP (op0, 0); | |
b6875aac KV |
6236 | if (VECTOR_MODE_P (mode)) |
6237 | *cost += extra_cost->ldst.storev; | |
6238 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
2961177e JG |
6239 | *cost += extra_cost->ldst.store; |
6240 | else if (mode == SFmode) | |
6241 | *cost += extra_cost->ldst.storef; | |
6242 | else if (mode == DFmode) | |
6243 | *cost += extra_cost->ldst.stored; | |
6244 | ||
6245 | *cost += | |
6246 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
6247 | 0, speed)); | |
6248 | } | |
43e9d192 | 6249 | |
e548c9df | 6250 | *cost += rtx_cost (op1, mode, SET, 1, speed); |
43e9d192 IB |
6251 | return true; |
6252 | ||
6253 | case SUBREG: | |
6254 | if (! REG_P (SUBREG_REG (op0))) | |
e548c9df | 6255 | *cost += rtx_cost (SUBREG_REG (op0), VOIDmode, SET, 0, speed); |
ba123b0d | 6256 | |
43e9d192 IB |
6257 | /* Fall through. */ |
6258 | case REG: | |
b6875aac KV |
6259 | /* The cost is one per vector-register copied. */ |
6260 | if (VECTOR_MODE_P (GET_MODE (op0)) && REG_P (op1)) | |
6261 | { | |
6262 | int n_minus_1 = (GET_MODE_SIZE (GET_MODE (op0)) - 1) | |
6263 | / GET_MODE_SIZE (V4SImode); | |
6264 | *cost = COSTS_N_INSNS (n_minus_1 + 1); | |
6265 | } | |
ba123b0d JG |
6266 | /* const0_rtx is in general free, but we will use an |
6267 | instruction to set a register to 0. */ | |
b6875aac KV |
6268 | else if (REG_P (op1) || op1 == const0_rtx) |
6269 | { | |
6270 | /* The cost is 1 per register copied. */ | |
6271 | int n_minus_1 = (GET_MODE_SIZE (GET_MODE (op0)) - 1) | |
ba123b0d | 6272 | / UNITS_PER_WORD; |
b6875aac KV |
6273 | *cost = COSTS_N_INSNS (n_minus_1 + 1); |
6274 | } | |
ba123b0d JG |
6275 | else |
6276 | /* Cost is just the cost of the RHS of the set. */ | |
e548c9df | 6277 | *cost += rtx_cost (op1, mode, SET, 1, speed); |
43e9d192 IB |
6278 | return true; |
6279 | ||
ba123b0d | 6280 | case ZERO_EXTRACT: |
43e9d192 | 6281 | case SIGN_EXTRACT: |
ba123b0d JG |
6282 | /* Bit-field insertion. Strip any redundant widening of |
6283 | the RHS to meet the width of the target. */ | |
43e9d192 IB |
6284 | if (GET_CODE (op1) == SUBREG) |
6285 | op1 = SUBREG_REG (op1); | |
6286 | if ((GET_CODE (op1) == ZERO_EXTEND | |
6287 | || GET_CODE (op1) == SIGN_EXTEND) | |
4aa81c2e | 6288 | && CONST_INT_P (XEXP (op0, 1)) |
43e9d192 IB |
6289 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (op1, 0))) |
6290 | >= INTVAL (XEXP (op0, 1)))) | |
6291 | op1 = XEXP (op1, 0); | |
ba123b0d JG |
6292 | |
6293 | if (CONST_INT_P (op1)) | |
6294 | { | |
6295 | /* MOV immediate is assumed to always be cheap. */ | |
6296 | *cost = COSTS_N_INSNS (1); | |
6297 | } | |
6298 | else | |
6299 | { | |
6300 | /* BFM. */ | |
6301 | if (speed) | |
6302 | *cost += extra_cost->alu.bfi; | |
e548c9df | 6303 | *cost += rtx_cost (op1, VOIDmode, (enum rtx_code) code, 1, speed); |
ba123b0d JG |
6304 | } |
6305 | ||
43e9d192 IB |
6306 | return true; |
6307 | ||
6308 | default: | |
ba123b0d JG |
6309 | /* We can't make sense of this, assume default cost. */ |
6310 | *cost = COSTS_N_INSNS (1); | |
61263118 | 6311 | return false; |
43e9d192 IB |
6312 | } |
6313 | return false; | |
6314 | ||
9dfc162c JG |
6315 | case CONST_INT: |
6316 | /* If an instruction can incorporate a constant within the | |
6317 | instruction, the instruction's expression avoids calling | |
6318 | rtx_cost() on the constant. If rtx_cost() is called on a | |
6319 | constant, then it is usually because the constant must be | |
6320 | moved into a register by one or more instructions. | |
6321 | ||
6322 | The exception is constant 0, which can be expressed | |
6323 | as XZR/WZR and is therefore free. The exception to this is | |
6324 | if we have (set (reg) (const0_rtx)) in which case we must cost | |
6325 | the move. However, we can catch that when we cost the SET, so | |
6326 | we don't need to consider that here. */ | |
6327 | if (x == const0_rtx) | |
6328 | *cost = 0; | |
6329 | else | |
6330 | { | |
6331 | /* To an approximation, building any other constant is | |
6332 | proportionally expensive to the number of instructions | |
6333 | required to build that constant. This is true whether we | |
6334 | are compiling for SPEED or otherwise. */ | |
82614948 RR |
6335 | *cost = COSTS_N_INSNS (aarch64_internal_mov_immediate |
6336 | (NULL_RTX, x, false, mode)); | |
9dfc162c JG |
6337 | } |
6338 | return true; | |
6339 | ||
6340 | case CONST_DOUBLE: | |
6341 | if (speed) | |
6342 | { | |
6343 | /* mov[df,sf]_aarch64. */ | |
6344 | if (aarch64_float_const_representable_p (x)) | |
6345 | /* FMOV (scalar immediate). */ | |
6346 | *cost += extra_cost->fp[mode == DFmode].fpconst; | |
6347 | else if (!aarch64_float_const_zero_rtx_p (x)) | |
6348 | { | |
6349 | /* This will be a load from memory. */ | |
6350 | if (mode == DFmode) | |
6351 | *cost += extra_cost->ldst.loadd; | |
6352 | else | |
6353 | *cost += extra_cost->ldst.loadf; | |
6354 | } | |
6355 | else | |
6356 | /* Otherwise this is +0.0. We get this using MOVI d0, #0 | |
6357 | or MOV v0.s[0], wzr - neither of which are modeled by the | |
6358 | cost tables. Just use the default cost. */ | |
6359 | { | |
6360 | } | |
6361 | } | |
6362 | ||
6363 | return true; | |
6364 | ||
43e9d192 IB |
6365 | case MEM: |
6366 | if (speed) | |
2961177e JG |
6367 | { |
6368 | /* For loads we want the base cost of a load, plus an | |
6369 | approximation for the additional cost of the addressing | |
6370 | mode. */ | |
6371 | rtx address = XEXP (x, 0); | |
b6875aac KV |
6372 | if (VECTOR_MODE_P (mode)) |
6373 | *cost += extra_cost->ldst.loadv; | |
6374 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
2961177e JG |
6375 | *cost += extra_cost->ldst.load; |
6376 | else if (mode == SFmode) | |
6377 | *cost += extra_cost->ldst.loadf; | |
6378 | else if (mode == DFmode) | |
6379 | *cost += extra_cost->ldst.loadd; | |
6380 | ||
6381 | *cost += | |
6382 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
6383 | 0, speed)); | |
6384 | } | |
43e9d192 IB |
6385 | |
6386 | return true; | |
6387 | ||
6388 | case NEG: | |
4745e701 JG |
6389 | op0 = XEXP (x, 0); |
6390 | ||
b6875aac KV |
6391 | if (VECTOR_MODE_P (mode)) |
6392 | { | |
6393 | if (speed) | |
6394 | { | |
6395 | /* FNEG. */ | |
6396 | *cost += extra_cost->vect.alu; | |
6397 | } | |
6398 | return false; | |
6399 | } | |
6400 | ||
e548c9df AM |
6401 | if (GET_MODE_CLASS (mode) == MODE_INT) |
6402 | { | |
4745e701 JG |
6403 | if (GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMPARE |
6404 | || GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMM_COMPARE) | |
6405 | { | |
6406 | /* CSETM. */ | |
e548c9df | 6407 | *cost += rtx_cost (XEXP (op0, 0), VOIDmode, NEG, 0, speed); |
4745e701 JG |
6408 | return true; |
6409 | } | |
6410 | ||
6411 | /* Cost this as SUB wzr, X. */ | |
e548c9df | 6412 | op0 = CONST0_RTX (mode); |
4745e701 JG |
6413 | op1 = XEXP (x, 0); |
6414 | goto cost_minus; | |
6415 | } | |
6416 | ||
e548c9df | 6417 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
4745e701 JG |
6418 | { |
6419 | /* Support (neg(fma...)) as a single instruction only if | |
6420 | sign of zeros is unimportant. This matches the decision | |
6421 | making in aarch64.md. */ | |
6422 | if (GET_CODE (op0) == FMA && !HONOR_SIGNED_ZEROS (GET_MODE (op0))) | |
6423 | { | |
6424 | /* FNMADD. */ | |
e548c9df | 6425 | *cost = rtx_cost (op0, mode, NEG, 0, speed); |
4745e701 JG |
6426 | return true; |
6427 | } | |
d318517d SN |
6428 | if (GET_CODE (op0) == MULT) |
6429 | { | |
6430 | /* FNMUL. */ | |
6431 | *cost = rtx_cost (op0, mode, NEG, 0, speed); | |
6432 | return true; | |
6433 | } | |
4745e701 JG |
6434 | if (speed) |
6435 | /* FNEG. */ | |
6436 | *cost += extra_cost->fp[mode == DFmode].neg; | |
6437 | return false; | |
6438 | } | |
6439 | ||
6440 | return false; | |
43e9d192 | 6441 | |
781aeb73 KT |
6442 | case CLRSB: |
6443 | case CLZ: | |
6444 | if (speed) | |
b6875aac KV |
6445 | { |
6446 | if (VECTOR_MODE_P (mode)) | |
6447 | *cost += extra_cost->vect.alu; | |
6448 | else | |
6449 | *cost += extra_cost->alu.clz; | |
6450 | } | |
781aeb73 KT |
6451 | |
6452 | return false; | |
6453 | ||
43e9d192 IB |
6454 | case COMPARE: |
6455 | op0 = XEXP (x, 0); | |
6456 | op1 = XEXP (x, 1); | |
6457 | ||
6458 | if (op1 == const0_rtx | |
6459 | && GET_CODE (op0) == AND) | |
6460 | { | |
6461 | x = op0; | |
e548c9df | 6462 | mode = GET_MODE (op0); |
43e9d192 IB |
6463 | goto cost_logic; |
6464 | } | |
6465 | ||
a8eecd00 JG |
6466 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT) |
6467 | { | |
6468 | /* TODO: A write to the CC flags possibly costs extra, this | |
6469 | needs encoding in the cost tables. */ | |
6470 | ||
e548c9df | 6471 | mode = GET_MODE (op0); |
a8eecd00 JG |
6472 | /* ANDS. */ |
6473 | if (GET_CODE (op0) == AND) | |
6474 | { | |
6475 | x = op0; | |
6476 | goto cost_logic; | |
6477 | } | |
6478 | ||
6479 | if (GET_CODE (op0) == PLUS) | |
6480 | { | |
6481 | /* ADDS (and CMN alias). */ | |
6482 | x = op0; | |
6483 | goto cost_plus; | |
6484 | } | |
6485 | ||
6486 | if (GET_CODE (op0) == MINUS) | |
6487 | { | |
6488 | /* SUBS. */ | |
6489 | x = op0; | |
6490 | goto cost_minus; | |
6491 | } | |
6492 | ||
345854d8 KT |
6493 | if (GET_CODE (op0) == ZERO_EXTRACT && op1 == const0_rtx |
6494 | && GET_MODE (x) == CC_NZmode && CONST_INT_P (XEXP (op0, 1)) | |
6495 | && CONST_INT_P (XEXP (op0, 2))) | |
6496 | { | |
6497 | /* COMPARE of ZERO_EXTRACT form of TST-immediate. | |
6498 | Handle it here directly rather than going to cost_logic | |
6499 | since we know the immediate generated for the TST is valid | |
6500 | so we can avoid creating an intermediate rtx for it only | |
6501 | for costing purposes. */ | |
6502 | if (speed) | |
6503 | *cost += extra_cost->alu.logical; | |
6504 | ||
6505 | *cost += rtx_cost (XEXP (op0, 0), GET_MODE (op0), | |
6506 | ZERO_EXTRACT, 0, speed); | |
6507 | return true; | |
6508 | } | |
6509 | ||
a8eecd00 JG |
6510 | if (GET_CODE (op1) == NEG) |
6511 | { | |
6512 | /* CMN. */ | |
6513 | if (speed) | |
6514 | *cost += extra_cost->alu.arith; | |
6515 | ||
e548c9df AM |
6516 | *cost += rtx_cost (op0, mode, COMPARE, 0, speed); |
6517 | *cost += rtx_cost (XEXP (op1, 0), mode, NEG, 1, speed); | |
a8eecd00 JG |
6518 | return true; |
6519 | } | |
6520 | ||
6521 | /* CMP. | |
6522 | ||
6523 | Compare can freely swap the order of operands, and | |
6524 | canonicalization puts the more complex operation first. | |
6525 | But the integer MINUS logic expects the shift/extend | |
6526 | operation in op1. */ | |
6527 | if (! (REG_P (op0) | |
6528 | || (GET_CODE (op0) == SUBREG && REG_P (SUBREG_REG (op0))))) | |
6529 | { | |
6530 | op0 = XEXP (x, 1); | |
6531 | op1 = XEXP (x, 0); | |
6532 | } | |
6533 | goto cost_minus; | |
6534 | } | |
6535 | ||
6536 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT) | |
6537 | { | |
6538 | /* FCMP. */ | |
6539 | if (speed) | |
6540 | *cost += extra_cost->fp[mode == DFmode].compare; | |
6541 | ||
6542 | if (CONST_DOUBLE_P (op1) && aarch64_float_const_zero_rtx_p (op1)) | |
6543 | { | |
e548c9df | 6544 | *cost += rtx_cost (op0, VOIDmode, COMPARE, 0, speed); |
a8eecd00 JG |
6545 | /* FCMP supports constant 0.0 for no extra cost. */ |
6546 | return true; | |
6547 | } | |
6548 | return false; | |
6549 | } | |
6550 | ||
b6875aac KV |
6551 | if (VECTOR_MODE_P (mode)) |
6552 | { | |
6553 | /* Vector compare. */ | |
6554 | if (speed) | |
6555 | *cost += extra_cost->vect.alu; | |
6556 | ||
6557 | if (aarch64_float_const_zero_rtx_p (op1)) | |
6558 | { | |
6559 | /* Vector cm (eq|ge|gt|lt|le) supports constant 0.0 for no extra | |
6560 | cost. */ | |
6561 | return true; | |
6562 | } | |
6563 | return false; | |
6564 | } | |
a8eecd00 | 6565 | return false; |
43e9d192 IB |
6566 | |
6567 | case MINUS: | |
4745e701 JG |
6568 | { |
6569 | op0 = XEXP (x, 0); | |
6570 | op1 = XEXP (x, 1); | |
6571 | ||
6572 | cost_minus: | |
e548c9df | 6573 | *cost += rtx_cost (op0, mode, MINUS, 0, speed); |
23cb6618 | 6574 | |
4745e701 JG |
6575 | /* Detect valid immediates. */ |
6576 | if ((GET_MODE_CLASS (mode) == MODE_INT | |
6577 | || (GET_MODE_CLASS (mode) == MODE_CC | |
6578 | && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT)) | |
6579 | && CONST_INT_P (op1) | |
6580 | && aarch64_uimm12_shift (INTVAL (op1))) | |
6581 | { | |
4745e701 JG |
6582 | if (speed) |
6583 | /* SUB(S) (immediate). */ | |
6584 | *cost += extra_cost->alu.arith; | |
6585 | return true; | |
4745e701 JG |
6586 | } |
6587 | ||
7cc2145f JG |
6588 | /* Look for SUB (extended register). */ |
6589 | if (aarch64_rtx_arith_op_extract_p (op1, mode)) | |
6590 | { | |
6591 | if (speed) | |
2533c820 | 6592 | *cost += extra_cost->alu.extend_arith; |
7cc2145f | 6593 | |
e47c4031 KT |
6594 | op1 = aarch64_strip_extend (op1); |
6595 | *cost += rtx_cost (op1, VOIDmode, | |
e548c9df | 6596 | (enum rtx_code) GET_CODE (op1), 0, speed); |
7cc2145f JG |
6597 | return true; |
6598 | } | |
6599 | ||
4745e701 JG |
6600 | rtx new_op1 = aarch64_strip_extend (op1); |
6601 | ||
6602 | /* Cost this as an FMA-alike operation. */ | |
6603 | if ((GET_CODE (new_op1) == MULT | |
0a78ebe4 | 6604 | || aarch64_shift_p (GET_CODE (new_op1))) |
4745e701 JG |
6605 | && code != COMPARE) |
6606 | { | |
6607 | *cost += aarch64_rtx_mult_cost (new_op1, MULT, | |
6608 | (enum rtx_code) code, | |
6609 | speed); | |
4745e701 JG |
6610 | return true; |
6611 | } | |
43e9d192 | 6612 | |
e548c9df | 6613 | *cost += rtx_cost (new_op1, VOIDmode, MINUS, 1, speed); |
43e9d192 | 6614 | |
4745e701 JG |
6615 | if (speed) |
6616 | { | |
b6875aac KV |
6617 | if (VECTOR_MODE_P (mode)) |
6618 | { | |
6619 | /* Vector SUB. */ | |
6620 | *cost += extra_cost->vect.alu; | |
6621 | } | |
6622 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
6623 | { | |
6624 | /* SUB(S). */ | |
6625 | *cost += extra_cost->alu.arith; | |
6626 | } | |
4745e701 | 6627 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
b6875aac KV |
6628 | { |
6629 | /* FSUB. */ | |
6630 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
6631 | } | |
4745e701 JG |
6632 | } |
6633 | return true; | |
6634 | } | |
43e9d192 IB |
6635 | |
6636 | case PLUS: | |
4745e701 JG |
6637 | { |
6638 | rtx new_op0; | |
43e9d192 | 6639 | |
4745e701 JG |
6640 | op0 = XEXP (x, 0); |
6641 | op1 = XEXP (x, 1); | |
43e9d192 | 6642 | |
a8eecd00 | 6643 | cost_plus: |
4745e701 JG |
6644 | if (GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMPARE |
6645 | || GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMM_COMPARE) | |
6646 | { | |
6647 | /* CSINC. */ | |
e548c9df AM |
6648 | *cost += rtx_cost (XEXP (op0, 0), mode, PLUS, 0, speed); |
6649 | *cost += rtx_cost (op1, mode, PLUS, 1, speed); | |
4745e701 JG |
6650 | return true; |
6651 | } | |
43e9d192 | 6652 | |
4745e701 JG |
6653 | if (GET_MODE_CLASS (mode) == MODE_INT |
6654 | && CONST_INT_P (op1) | |
6655 | && aarch64_uimm12_shift (INTVAL (op1))) | |
6656 | { | |
e548c9df | 6657 | *cost += rtx_cost (op0, mode, PLUS, 0, speed); |
43e9d192 | 6658 | |
4745e701 JG |
6659 | if (speed) |
6660 | /* ADD (immediate). */ | |
6661 | *cost += extra_cost->alu.arith; | |
6662 | return true; | |
6663 | } | |
6664 | ||
e548c9df | 6665 | *cost += rtx_cost (op1, mode, PLUS, 1, speed); |
23cb6618 | 6666 | |
7cc2145f JG |
6667 | /* Look for ADD (extended register). */ |
6668 | if (aarch64_rtx_arith_op_extract_p (op0, mode)) | |
6669 | { | |
6670 | if (speed) | |
2533c820 | 6671 | *cost += extra_cost->alu.extend_arith; |
7cc2145f | 6672 | |
e47c4031 KT |
6673 | op0 = aarch64_strip_extend (op0); |
6674 | *cost += rtx_cost (op0, VOIDmode, | |
e548c9df | 6675 | (enum rtx_code) GET_CODE (op0), 0, speed); |
7cc2145f JG |
6676 | return true; |
6677 | } | |
6678 | ||
4745e701 JG |
6679 | /* Strip any extend, leave shifts behind as we will |
6680 | cost them through mult_cost. */ | |
6681 | new_op0 = aarch64_strip_extend (op0); | |
6682 | ||
6683 | if (GET_CODE (new_op0) == MULT | |
0a78ebe4 | 6684 | || aarch64_shift_p (GET_CODE (new_op0))) |
4745e701 JG |
6685 | { |
6686 | *cost += aarch64_rtx_mult_cost (new_op0, MULT, PLUS, | |
6687 | speed); | |
4745e701 JG |
6688 | return true; |
6689 | } | |
6690 | ||
e548c9df | 6691 | *cost += rtx_cost (new_op0, VOIDmode, PLUS, 0, speed); |
4745e701 JG |
6692 | |
6693 | if (speed) | |
6694 | { | |
b6875aac KV |
6695 | if (VECTOR_MODE_P (mode)) |
6696 | { | |
6697 | /* Vector ADD. */ | |
6698 | *cost += extra_cost->vect.alu; | |
6699 | } | |
6700 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
6701 | { | |
6702 | /* ADD. */ | |
6703 | *cost += extra_cost->alu.arith; | |
6704 | } | |
4745e701 | 6705 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
b6875aac KV |
6706 | { |
6707 | /* FADD. */ | |
6708 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
6709 | } | |
4745e701 JG |
6710 | } |
6711 | return true; | |
6712 | } | |
43e9d192 | 6713 | |
18b42b2a KT |
6714 | case BSWAP: |
6715 | *cost = COSTS_N_INSNS (1); | |
6716 | ||
6717 | if (speed) | |
b6875aac KV |
6718 | { |
6719 | if (VECTOR_MODE_P (mode)) | |
6720 | *cost += extra_cost->vect.alu; | |
6721 | else | |
6722 | *cost += extra_cost->alu.rev; | |
6723 | } | |
18b42b2a KT |
6724 | return false; |
6725 | ||
43e9d192 | 6726 | case IOR: |
f7d5cf8d KT |
6727 | if (aarch_rev16_p (x)) |
6728 | { | |
6729 | *cost = COSTS_N_INSNS (1); | |
6730 | ||
b6875aac KV |
6731 | if (speed) |
6732 | { | |
6733 | if (VECTOR_MODE_P (mode)) | |
6734 | *cost += extra_cost->vect.alu; | |
6735 | else | |
6736 | *cost += extra_cost->alu.rev; | |
6737 | } | |
6738 | return true; | |
f7d5cf8d | 6739 | } |
fb0cb7fa KT |
6740 | |
6741 | if (aarch64_extr_rtx_p (x, &op0, &op1)) | |
6742 | { | |
e548c9df AM |
6743 | *cost += rtx_cost (op0, mode, IOR, 0, speed); |
6744 | *cost += rtx_cost (op1, mode, IOR, 1, speed); | |
fb0cb7fa KT |
6745 | if (speed) |
6746 | *cost += extra_cost->alu.shift; | |
6747 | ||
6748 | return true; | |
6749 | } | |
f7d5cf8d | 6750 | /* Fall through. */ |
43e9d192 IB |
6751 | case XOR: |
6752 | case AND: | |
6753 | cost_logic: | |
6754 | op0 = XEXP (x, 0); | |
6755 | op1 = XEXP (x, 1); | |
6756 | ||
b6875aac KV |
6757 | if (VECTOR_MODE_P (mode)) |
6758 | { | |
6759 | if (speed) | |
6760 | *cost += extra_cost->vect.alu; | |
6761 | return true; | |
6762 | } | |
6763 | ||
268c3b47 JG |
6764 | if (code == AND |
6765 | && GET_CODE (op0) == MULT | |
6766 | && CONST_INT_P (XEXP (op0, 1)) | |
6767 | && CONST_INT_P (op1) | |
6768 | && aarch64_uxt_size (exact_log2 (INTVAL (XEXP (op0, 1))), | |
6769 | INTVAL (op1)) != 0) | |
6770 | { | |
6771 | /* This is a UBFM/SBFM. */ | |
e548c9df | 6772 | *cost += rtx_cost (XEXP (op0, 0), mode, ZERO_EXTRACT, 0, speed); |
268c3b47 JG |
6773 | if (speed) |
6774 | *cost += extra_cost->alu.bfx; | |
6775 | return true; | |
6776 | } | |
6777 | ||
e548c9df | 6778 | if (GET_MODE_CLASS (mode) == MODE_INT) |
43e9d192 | 6779 | { |
8c83f71d | 6780 | if (CONST_INT_P (op1)) |
43e9d192 | 6781 | { |
8c83f71d KT |
6782 | /* We have a mask + shift version of a UBFIZ |
6783 | i.e. the *andim_ashift<mode>_bfiz pattern. */ | |
6784 | if (GET_CODE (op0) == ASHIFT | |
6785 | && aarch64_mask_and_shift_for_ubfiz_p (mode, op1, | |
6786 | XEXP (op0, 1))) | |
6787 | { | |
6788 | *cost += rtx_cost (XEXP (op0, 0), mode, | |
6789 | (enum rtx_code) code, 0, speed); | |
6790 | if (speed) | |
6791 | *cost += extra_cost->alu.bfx; | |
268c3b47 | 6792 | |
8c83f71d KT |
6793 | return true; |
6794 | } | |
6795 | else if (aarch64_bitmask_imm (INTVAL (op1), mode)) | |
6796 | { | |
6797 | /* We possibly get the immediate for free, this is not | |
6798 | modelled. */ | |
6799 | *cost += rtx_cost (op0, mode, (enum rtx_code) code, 0, speed); | |
6800 | if (speed) | |
6801 | *cost += extra_cost->alu.logical; | |
268c3b47 | 6802 | |
8c83f71d KT |
6803 | return true; |
6804 | } | |
43e9d192 IB |
6805 | } |
6806 | else | |
6807 | { | |
268c3b47 JG |
6808 | rtx new_op0 = op0; |
6809 | ||
6810 | /* Handle ORN, EON, or BIC. */ | |
43e9d192 IB |
6811 | if (GET_CODE (op0) == NOT) |
6812 | op0 = XEXP (op0, 0); | |
268c3b47 JG |
6813 | |
6814 | new_op0 = aarch64_strip_shift (op0); | |
6815 | ||
6816 | /* If we had a shift on op0 then this is a logical-shift- | |
6817 | by-register/immediate operation. Otherwise, this is just | |
6818 | a logical operation. */ | |
6819 | if (speed) | |
6820 | { | |
6821 | if (new_op0 != op0) | |
6822 | { | |
6823 | /* Shift by immediate. */ | |
6824 | if (CONST_INT_P (XEXP (op0, 1))) | |
6825 | *cost += extra_cost->alu.log_shift; | |
6826 | else | |
6827 | *cost += extra_cost->alu.log_shift_reg; | |
6828 | } | |
6829 | else | |
6830 | *cost += extra_cost->alu.logical; | |
6831 | } | |
6832 | ||
6833 | /* In both cases we want to cost both operands. */ | |
e548c9df AM |
6834 | *cost += rtx_cost (new_op0, mode, (enum rtx_code) code, 0, speed); |
6835 | *cost += rtx_cost (op1, mode, (enum rtx_code) code, 1, speed); | |
268c3b47 JG |
6836 | |
6837 | return true; | |
43e9d192 | 6838 | } |
43e9d192 IB |
6839 | } |
6840 | return false; | |
6841 | ||
268c3b47 | 6842 | case NOT: |
6365da9e KT |
6843 | x = XEXP (x, 0); |
6844 | op0 = aarch64_strip_shift (x); | |
6845 | ||
b6875aac KV |
6846 | if (VECTOR_MODE_P (mode)) |
6847 | { | |
6848 | /* Vector NOT. */ | |
6849 | *cost += extra_cost->vect.alu; | |
6850 | return false; | |
6851 | } | |
6852 | ||
6365da9e KT |
6853 | /* MVN-shifted-reg. */ |
6854 | if (op0 != x) | |
6855 | { | |
e548c9df | 6856 | *cost += rtx_cost (op0, mode, (enum rtx_code) code, 0, speed); |
6365da9e KT |
6857 | |
6858 | if (speed) | |
6859 | *cost += extra_cost->alu.log_shift; | |
6860 | ||
6861 | return true; | |
6862 | } | |
6863 | /* EON can have two forms: (xor (not a) b) but also (not (xor a b)). | |
6864 | Handle the second form here taking care that 'a' in the above can | |
6865 | be a shift. */ | |
6866 | else if (GET_CODE (op0) == XOR) | |
6867 | { | |
6868 | rtx newop0 = XEXP (op0, 0); | |
6869 | rtx newop1 = XEXP (op0, 1); | |
6870 | rtx op0_stripped = aarch64_strip_shift (newop0); | |
6871 | ||
e548c9df AM |
6872 | *cost += rtx_cost (newop1, mode, (enum rtx_code) code, 1, speed); |
6873 | *cost += rtx_cost (op0_stripped, mode, XOR, 0, speed); | |
6365da9e KT |
6874 | |
6875 | if (speed) | |
6876 | { | |
6877 | if (op0_stripped != newop0) | |
6878 | *cost += extra_cost->alu.log_shift; | |
6879 | else | |
6880 | *cost += extra_cost->alu.logical; | |
6881 | } | |
6882 | ||
6883 | return true; | |
6884 | } | |
268c3b47 JG |
6885 | /* MVN. */ |
6886 | if (speed) | |
6887 | *cost += extra_cost->alu.logical; | |
6888 | ||
268c3b47 JG |
6889 | return false; |
6890 | ||
43e9d192 | 6891 | case ZERO_EXTEND: |
b1685e62 JG |
6892 | |
6893 | op0 = XEXP (x, 0); | |
6894 | /* If a value is written in SI mode, then zero extended to DI | |
6895 | mode, the operation will in general be free as a write to | |
6896 | a 'w' register implicitly zeroes the upper bits of an 'x' | |
6897 | register. However, if this is | |
6898 | ||
6899 | (set (reg) (zero_extend (reg))) | |
6900 | ||
6901 | we must cost the explicit register move. */ | |
6902 | if (mode == DImode | |
6903 | && GET_MODE (op0) == SImode | |
6904 | && outer == SET) | |
6905 | { | |
e548c9df | 6906 | int op_cost = rtx_cost (op0, VOIDmode, ZERO_EXTEND, 0, speed); |
b1685e62 | 6907 | |
dde23f43 KM |
6908 | /* If OP_COST is non-zero, then the cost of the zero extend |
6909 | is effectively the cost of the inner operation. Otherwise | |
6910 | we have a MOV instruction and we take the cost from the MOV | |
6911 | itself. This is true independently of whether we are | |
6912 | optimizing for space or time. */ | |
6913 | if (op_cost) | |
b1685e62 JG |
6914 | *cost = op_cost; |
6915 | ||
6916 | return true; | |
6917 | } | |
e548c9df | 6918 | else if (MEM_P (op0)) |
43e9d192 | 6919 | { |
b1685e62 | 6920 | /* All loads can zero extend to any size for free. */ |
e548c9df | 6921 | *cost = rtx_cost (op0, VOIDmode, ZERO_EXTEND, param, speed); |
43e9d192 IB |
6922 | return true; |
6923 | } | |
b1685e62 | 6924 | |
283b6c85 KT |
6925 | op0 = aarch64_extend_bitfield_pattern_p (x); |
6926 | if (op0) | |
6927 | { | |
6928 | *cost += rtx_cost (op0, mode, ZERO_EXTEND, 0, speed); | |
6929 | if (speed) | |
6930 | *cost += extra_cost->alu.bfx; | |
6931 | return true; | |
6932 | } | |
6933 | ||
b1685e62 | 6934 | if (speed) |
b6875aac KV |
6935 | { |
6936 | if (VECTOR_MODE_P (mode)) | |
6937 | { | |
6938 | /* UMOV. */ | |
6939 | *cost += extra_cost->vect.alu; | |
6940 | } | |
6941 | else | |
6942 | { | |
63715e5e WD |
6943 | /* We generate an AND instead of UXTB/UXTH. */ |
6944 | *cost += extra_cost->alu.logical; | |
b6875aac KV |
6945 | } |
6946 | } | |
43e9d192 IB |
6947 | return false; |
6948 | ||
6949 | case SIGN_EXTEND: | |
b1685e62 | 6950 | if (MEM_P (XEXP (x, 0))) |
43e9d192 | 6951 | { |
b1685e62 JG |
6952 | /* LDRSH. */ |
6953 | if (speed) | |
6954 | { | |
6955 | rtx address = XEXP (XEXP (x, 0), 0); | |
6956 | *cost += extra_cost->ldst.load_sign_extend; | |
6957 | ||
6958 | *cost += | |
6959 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
6960 | 0, speed)); | |
6961 | } | |
43e9d192 IB |
6962 | return true; |
6963 | } | |
b1685e62 | 6964 | |
283b6c85 KT |
6965 | op0 = aarch64_extend_bitfield_pattern_p (x); |
6966 | if (op0) | |
6967 | { | |
6968 | *cost += rtx_cost (op0, mode, SIGN_EXTEND, 0, speed); | |
6969 | if (speed) | |
6970 | *cost += extra_cost->alu.bfx; | |
6971 | return true; | |
6972 | } | |
6973 | ||
b1685e62 | 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.extend; | |
6980 | } | |
43e9d192 IB |
6981 | return false; |
6982 | ||
ba0cfa17 JG |
6983 | case ASHIFT: |
6984 | op0 = XEXP (x, 0); | |
6985 | op1 = XEXP (x, 1); | |
6986 | ||
6987 | if (CONST_INT_P (op1)) | |
6988 | { | |
ba0cfa17 | 6989 | if (speed) |
b6875aac KV |
6990 | { |
6991 | if (VECTOR_MODE_P (mode)) | |
6992 | { | |
6993 | /* Vector shift (immediate). */ | |
6994 | *cost += extra_cost->vect.alu; | |
6995 | } | |
6996 | else | |
6997 | { | |
6998 | /* LSL (immediate), UBMF, UBFIZ and friends. These are all | |
6999 | aliases. */ | |
7000 | *cost += extra_cost->alu.shift; | |
7001 | } | |
7002 | } | |
ba0cfa17 JG |
7003 | |
7004 | /* We can incorporate zero/sign extend for free. */ | |
7005 | if (GET_CODE (op0) == ZERO_EXTEND | |
7006 | || GET_CODE (op0) == SIGN_EXTEND) | |
7007 | op0 = XEXP (op0, 0); | |
7008 | ||
e548c9df | 7009 | *cost += rtx_cost (op0, VOIDmode, ASHIFT, 0, speed); |
ba0cfa17 JG |
7010 | return true; |
7011 | } | |
7012 | else | |
7013 | { | |
ba0cfa17 | 7014 | if (speed) |
b6875aac KV |
7015 | { |
7016 | if (VECTOR_MODE_P (mode)) | |
7017 | { | |
7018 | /* Vector shift (register). */ | |
7019 | *cost += extra_cost->vect.alu; | |
7020 | } | |
7021 | else | |
7022 | { | |
7023 | /* LSLV. */ | |
7024 | *cost += extra_cost->alu.shift_reg; | |
7025 | } | |
7026 | } | |
ba0cfa17 JG |
7027 | return false; /* All arguments need to be in registers. */ |
7028 | } | |
7029 | ||
43e9d192 | 7030 | case ROTATE: |
43e9d192 IB |
7031 | case ROTATERT: |
7032 | case LSHIFTRT: | |
43e9d192 | 7033 | case ASHIFTRT: |
ba0cfa17 JG |
7034 | op0 = XEXP (x, 0); |
7035 | op1 = XEXP (x, 1); | |
43e9d192 | 7036 | |
ba0cfa17 JG |
7037 | if (CONST_INT_P (op1)) |
7038 | { | |
7039 | /* ASR (immediate) and friends. */ | |
7040 | if (speed) | |
b6875aac KV |
7041 | { |
7042 | if (VECTOR_MODE_P (mode)) | |
7043 | *cost += extra_cost->vect.alu; | |
7044 | else | |
7045 | *cost += extra_cost->alu.shift; | |
7046 | } | |
43e9d192 | 7047 | |
e548c9df | 7048 | *cost += rtx_cost (op0, mode, (enum rtx_code) code, 0, speed); |
ba0cfa17 JG |
7049 | return true; |
7050 | } | |
7051 | else | |
7052 | { | |
7053 | ||
7054 | /* ASR (register) and friends. */ | |
7055 | if (speed) | |
b6875aac KV |
7056 | { |
7057 | if (VECTOR_MODE_P (mode)) | |
7058 | *cost += extra_cost->vect.alu; | |
7059 | else | |
7060 | *cost += extra_cost->alu.shift_reg; | |
7061 | } | |
ba0cfa17 JG |
7062 | return false; /* All arguments need to be in registers. */ |
7063 | } | |
43e9d192 | 7064 | |
909734be JG |
7065 | case SYMBOL_REF: |
7066 | ||
1b1e81f8 JW |
7067 | if (aarch64_cmodel == AARCH64_CMODEL_LARGE |
7068 | || aarch64_cmodel == AARCH64_CMODEL_SMALL_SPIC) | |
909734be JG |
7069 | { |
7070 | /* LDR. */ | |
7071 | if (speed) | |
7072 | *cost += extra_cost->ldst.load; | |
7073 | } | |
7074 | else if (aarch64_cmodel == AARCH64_CMODEL_SMALL | |
7075 | || aarch64_cmodel == AARCH64_CMODEL_SMALL_PIC) | |
7076 | { | |
7077 | /* ADRP, followed by ADD. */ | |
7078 | *cost += COSTS_N_INSNS (1); | |
7079 | if (speed) | |
7080 | *cost += 2 * extra_cost->alu.arith; | |
7081 | } | |
7082 | else if (aarch64_cmodel == AARCH64_CMODEL_TINY | |
7083 | || aarch64_cmodel == AARCH64_CMODEL_TINY_PIC) | |
7084 | { | |
7085 | /* ADR. */ | |
7086 | if (speed) | |
7087 | *cost += extra_cost->alu.arith; | |
7088 | } | |
7089 | ||
7090 | if (flag_pic) | |
7091 | { | |
7092 | /* One extra load instruction, after accessing the GOT. */ | |
7093 | *cost += COSTS_N_INSNS (1); | |
7094 | if (speed) | |
7095 | *cost += extra_cost->ldst.load; | |
7096 | } | |
43e9d192 IB |
7097 | return true; |
7098 | ||
909734be | 7099 | case HIGH: |
43e9d192 | 7100 | case LO_SUM: |
909734be JG |
7101 | /* ADRP/ADD (immediate). */ |
7102 | if (speed) | |
7103 | *cost += extra_cost->alu.arith; | |
43e9d192 IB |
7104 | return true; |
7105 | ||
7106 | case ZERO_EXTRACT: | |
7107 | case SIGN_EXTRACT: | |
7cc2145f JG |
7108 | /* UBFX/SBFX. */ |
7109 | if (speed) | |
b6875aac KV |
7110 | { |
7111 | if (VECTOR_MODE_P (mode)) | |
7112 | *cost += extra_cost->vect.alu; | |
7113 | else | |
7114 | *cost += extra_cost->alu.bfx; | |
7115 | } | |
7cc2145f JG |
7116 | |
7117 | /* We can trust that the immediates used will be correct (there | |
7118 | are no by-register forms), so we need only cost op0. */ | |
e548c9df | 7119 | *cost += rtx_cost (XEXP (x, 0), VOIDmode, (enum rtx_code) code, 0, speed); |
43e9d192 IB |
7120 | return true; |
7121 | ||
7122 | case MULT: | |
4745e701 JG |
7123 | *cost += aarch64_rtx_mult_cost (x, MULT, 0, speed); |
7124 | /* aarch64_rtx_mult_cost always handles recursion to its | |
7125 | operands. */ | |
7126 | return true; | |
43e9d192 IB |
7127 | |
7128 | case MOD: | |
4f58fe36 KT |
7129 | /* We can expand signed mod by power of 2 using a NEGS, two parallel |
7130 | ANDs and a CSNEG. Assume here that CSNEG is the same as the cost of | |
7131 | an unconditional negate. This case should only ever be reached through | |
7132 | the set_smod_pow2_cheap check in expmed.c. */ | |
7133 | if (CONST_INT_P (XEXP (x, 1)) | |
7134 | && exact_log2 (INTVAL (XEXP (x, 1))) > 0 | |
7135 | && (mode == SImode || mode == DImode)) | |
7136 | { | |
7137 | /* We expand to 4 instructions. Reset the baseline. */ | |
7138 | *cost = COSTS_N_INSNS (4); | |
7139 | ||
7140 | if (speed) | |
7141 | *cost += 2 * extra_cost->alu.logical | |
7142 | + 2 * extra_cost->alu.arith; | |
7143 | ||
7144 | return true; | |
7145 | } | |
7146 | ||
7147 | /* Fall-through. */ | |
43e9d192 | 7148 | case UMOD: |
43e9d192 IB |
7149 | if (speed) |
7150 | { | |
b6875aac KV |
7151 | if (VECTOR_MODE_P (mode)) |
7152 | *cost += extra_cost->vect.alu; | |
e548c9df AM |
7153 | else if (GET_MODE_CLASS (mode) == MODE_INT) |
7154 | *cost += (extra_cost->mult[mode == DImode].add | |
7155 | + extra_cost->mult[mode == DImode].idiv); | |
7156 | else if (mode == DFmode) | |
73250c4c KT |
7157 | *cost += (extra_cost->fp[1].mult |
7158 | + extra_cost->fp[1].div); | |
e548c9df | 7159 | else if (mode == SFmode) |
73250c4c KT |
7160 | *cost += (extra_cost->fp[0].mult |
7161 | + extra_cost->fp[0].div); | |
43e9d192 IB |
7162 | } |
7163 | return false; /* All arguments need to be in registers. */ | |
7164 | ||
7165 | case DIV: | |
7166 | case UDIV: | |
4105fe38 | 7167 | case SQRT: |
43e9d192 IB |
7168 | if (speed) |
7169 | { | |
b6875aac KV |
7170 | if (VECTOR_MODE_P (mode)) |
7171 | *cost += extra_cost->vect.alu; | |
7172 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
4105fe38 JG |
7173 | /* There is no integer SQRT, so only DIV and UDIV can get |
7174 | here. */ | |
7175 | *cost += extra_cost->mult[mode == DImode].idiv; | |
7176 | else | |
7177 | *cost += extra_cost->fp[mode == DFmode].div; | |
43e9d192 IB |
7178 | } |
7179 | return false; /* All arguments need to be in registers. */ | |
7180 | ||
a8eecd00 | 7181 | case IF_THEN_ELSE: |
2d5ffe46 AP |
7182 | return aarch64_if_then_else_costs (XEXP (x, 0), XEXP (x, 1), |
7183 | XEXP (x, 2), cost, speed); | |
a8eecd00 JG |
7184 | |
7185 | case EQ: | |
7186 | case NE: | |
7187 | case GT: | |
7188 | case GTU: | |
7189 | case LT: | |
7190 | case LTU: | |
7191 | case GE: | |
7192 | case GEU: | |
7193 | case LE: | |
7194 | case LEU: | |
7195 | ||
7196 | return false; /* All arguments must be in registers. */ | |
7197 | ||
b292109f JG |
7198 | case FMA: |
7199 | op0 = XEXP (x, 0); | |
7200 | op1 = XEXP (x, 1); | |
7201 | op2 = XEXP (x, 2); | |
7202 | ||
7203 | if (speed) | |
b6875aac KV |
7204 | { |
7205 | if (VECTOR_MODE_P (mode)) | |
7206 | *cost += extra_cost->vect.alu; | |
7207 | else | |
7208 | *cost += extra_cost->fp[mode == DFmode].fma; | |
7209 | } | |
b292109f JG |
7210 | |
7211 | /* FMSUB, FNMADD, and FNMSUB are free. */ | |
7212 | if (GET_CODE (op0) == NEG) | |
7213 | op0 = XEXP (op0, 0); | |
7214 | ||
7215 | if (GET_CODE (op2) == NEG) | |
7216 | op2 = XEXP (op2, 0); | |
7217 | ||
7218 | /* aarch64_fnma4_elt_to_64v2df has the NEG as operand 1, | |
7219 | and the by-element operand as operand 0. */ | |
7220 | if (GET_CODE (op1) == NEG) | |
7221 | op1 = XEXP (op1, 0); | |
7222 | ||
7223 | /* Catch vector-by-element operations. The by-element operand can | |
7224 | either be (vec_duplicate (vec_select (x))) or just | |
7225 | (vec_select (x)), depending on whether we are multiplying by | |
7226 | a vector or a scalar. | |
7227 | ||
7228 | Canonicalization is not very good in these cases, FMA4 will put the | |
7229 | by-element operand as operand 0, FNMA4 will have it as operand 1. */ | |
7230 | if (GET_CODE (op0) == VEC_DUPLICATE) | |
7231 | op0 = XEXP (op0, 0); | |
7232 | else if (GET_CODE (op1) == VEC_DUPLICATE) | |
7233 | op1 = XEXP (op1, 0); | |
7234 | ||
7235 | if (GET_CODE (op0) == VEC_SELECT) | |
7236 | op0 = XEXP (op0, 0); | |
7237 | else if (GET_CODE (op1) == VEC_SELECT) | |
7238 | op1 = XEXP (op1, 0); | |
7239 | ||
7240 | /* If the remaining parameters are not registers, | |
7241 | get the cost to put them into registers. */ | |
e548c9df AM |
7242 | *cost += rtx_cost (op0, mode, FMA, 0, speed); |
7243 | *cost += rtx_cost (op1, mode, FMA, 1, speed); | |
7244 | *cost += rtx_cost (op2, mode, FMA, 2, speed); | |
b292109f JG |
7245 | return true; |
7246 | ||
5e2a765b KT |
7247 | case FLOAT: |
7248 | case UNSIGNED_FLOAT: | |
7249 | if (speed) | |
7250 | *cost += extra_cost->fp[mode == DFmode].fromint; | |
7251 | return false; | |
7252 | ||
b292109f JG |
7253 | case FLOAT_EXTEND: |
7254 | if (speed) | |
b6875aac KV |
7255 | { |
7256 | if (VECTOR_MODE_P (mode)) | |
7257 | { | |
7258 | /*Vector truncate. */ | |
7259 | *cost += extra_cost->vect.alu; | |
7260 | } | |
7261 | else | |
7262 | *cost += extra_cost->fp[mode == DFmode].widen; | |
7263 | } | |
b292109f JG |
7264 | return false; |
7265 | ||
7266 | case FLOAT_TRUNCATE: | |
7267 | if (speed) | |
b6875aac KV |
7268 | { |
7269 | if (VECTOR_MODE_P (mode)) | |
7270 | { | |
7271 | /*Vector conversion. */ | |
7272 | *cost += extra_cost->vect.alu; | |
7273 | } | |
7274 | else | |
7275 | *cost += extra_cost->fp[mode == DFmode].narrow; | |
7276 | } | |
b292109f JG |
7277 | return false; |
7278 | ||
61263118 KT |
7279 | case FIX: |
7280 | case UNSIGNED_FIX: | |
7281 | x = XEXP (x, 0); | |
7282 | /* Strip the rounding part. They will all be implemented | |
7283 | by the fcvt* family of instructions anyway. */ | |
7284 | if (GET_CODE (x) == UNSPEC) | |
7285 | { | |
7286 | unsigned int uns_code = XINT (x, 1); | |
7287 | ||
7288 | if (uns_code == UNSPEC_FRINTA | |
7289 | || uns_code == UNSPEC_FRINTM | |
7290 | || uns_code == UNSPEC_FRINTN | |
7291 | || uns_code == UNSPEC_FRINTP | |
7292 | || uns_code == UNSPEC_FRINTZ) | |
7293 | x = XVECEXP (x, 0, 0); | |
7294 | } | |
7295 | ||
7296 | if (speed) | |
b6875aac KV |
7297 | { |
7298 | if (VECTOR_MODE_P (mode)) | |
7299 | *cost += extra_cost->vect.alu; | |
7300 | else | |
7301 | *cost += extra_cost->fp[GET_MODE (x) == DFmode].toint; | |
7302 | } | |
39252973 KT |
7303 | |
7304 | /* We can combine fmul by a power of 2 followed by a fcvt into a single | |
7305 | fixed-point fcvt. */ | |
7306 | if (GET_CODE (x) == MULT | |
7307 | && ((VECTOR_MODE_P (mode) | |
7308 | && aarch64_vec_fpconst_pow_of_2 (XEXP (x, 1)) > 0) | |
7309 | || aarch64_fpconst_pow_of_2 (XEXP (x, 1)) > 0)) | |
7310 | { | |
7311 | *cost += rtx_cost (XEXP (x, 0), VOIDmode, (rtx_code) code, | |
7312 | 0, speed); | |
7313 | return true; | |
7314 | } | |
7315 | ||
e548c9df | 7316 | *cost += rtx_cost (x, VOIDmode, (enum rtx_code) code, 0, speed); |
61263118 KT |
7317 | return true; |
7318 | ||
b292109f | 7319 | case ABS: |
b6875aac KV |
7320 | if (VECTOR_MODE_P (mode)) |
7321 | { | |
7322 | /* ABS (vector). */ | |
7323 | if (speed) | |
7324 | *cost += extra_cost->vect.alu; | |
7325 | } | |
7326 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
b292109f | 7327 | { |
19261b99 KT |
7328 | op0 = XEXP (x, 0); |
7329 | ||
7330 | /* FABD, which is analogous to FADD. */ | |
7331 | if (GET_CODE (op0) == MINUS) | |
7332 | { | |
e548c9df AM |
7333 | *cost += rtx_cost (XEXP (op0, 0), mode, MINUS, 0, speed); |
7334 | *cost += rtx_cost (XEXP (op0, 1), mode, MINUS, 1, speed); | |
19261b99 KT |
7335 | if (speed) |
7336 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
7337 | ||
7338 | return true; | |
7339 | } | |
7340 | /* Simple FABS is analogous to FNEG. */ | |
b292109f JG |
7341 | if (speed) |
7342 | *cost += extra_cost->fp[mode == DFmode].neg; | |
7343 | } | |
7344 | else | |
7345 | { | |
7346 | /* Integer ABS will either be split to | |
7347 | two arithmetic instructions, or will be an ABS | |
7348 | (scalar), which we don't model. */ | |
7349 | *cost = COSTS_N_INSNS (2); | |
7350 | if (speed) | |
7351 | *cost += 2 * extra_cost->alu.arith; | |
7352 | } | |
7353 | return false; | |
7354 | ||
7355 | case SMAX: | |
7356 | case SMIN: | |
7357 | if (speed) | |
7358 | { | |
b6875aac KV |
7359 | if (VECTOR_MODE_P (mode)) |
7360 | *cost += extra_cost->vect.alu; | |
7361 | else | |
7362 | { | |
7363 | /* FMAXNM/FMINNM/FMAX/FMIN. | |
7364 | TODO: This may not be accurate for all implementations, but | |
7365 | we do not model this in the cost tables. */ | |
7366 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
7367 | } | |
b292109f JG |
7368 | } |
7369 | return false; | |
7370 | ||
61263118 KT |
7371 | case UNSPEC: |
7372 | /* The floating point round to integer frint* instructions. */ | |
7373 | if (aarch64_frint_unspec_p (XINT (x, 1))) | |
7374 | { | |
7375 | if (speed) | |
7376 | *cost += extra_cost->fp[mode == DFmode].roundint; | |
7377 | ||
7378 | return false; | |
7379 | } | |
781aeb73 KT |
7380 | |
7381 | if (XINT (x, 1) == UNSPEC_RBIT) | |
7382 | { | |
7383 | if (speed) | |
7384 | *cost += extra_cost->alu.rev; | |
7385 | ||
7386 | return false; | |
7387 | } | |
61263118 KT |
7388 | break; |
7389 | ||
fb620c4a JG |
7390 | case TRUNCATE: |
7391 | ||
7392 | /* Decompose <su>muldi3_highpart. */ | |
7393 | if (/* (truncate:DI */ | |
7394 | mode == DImode | |
7395 | /* (lshiftrt:TI */ | |
7396 | && GET_MODE (XEXP (x, 0)) == TImode | |
7397 | && GET_CODE (XEXP (x, 0)) == LSHIFTRT | |
7398 | /* (mult:TI */ | |
7399 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT | |
7400 | /* (ANY_EXTEND:TI (reg:DI)) | |
7401 | (ANY_EXTEND:TI (reg:DI))) */ | |
7402 | && ((GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == ZERO_EXTEND | |
7403 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == ZERO_EXTEND) | |
7404 | || (GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == SIGN_EXTEND | |
7405 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == SIGN_EXTEND)) | |
7406 | && GET_MODE (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 0)) == DImode | |
7407 | && GET_MODE (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 1), 0)) == DImode | |
7408 | /* (const_int 64) */ | |
7409 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
7410 | && UINTVAL (XEXP (XEXP (x, 0), 1)) == 64) | |
7411 | { | |
7412 | /* UMULH/SMULH. */ | |
7413 | if (speed) | |
7414 | *cost += extra_cost->mult[mode == DImode].extend; | |
e548c9df AM |
7415 | *cost += rtx_cost (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 0), |
7416 | mode, MULT, 0, speed); | |
7417 | *cost += rtx_cost (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 1), 0), | |
7418 | mode, MULT, 1, speed); | |
fb620c4a JG |
7419 | return true; |
7420 | } | |
7421 | ||
7422 | /* Fall through. */ | |
43e9d192 | 7423 | default: |
61263118 | 7424 | break; |
43e9d192 | 7425 | } |
61263118 | 7426 | |
c10e3d7f AP |
7427 | if (dump_file |
7428 | && flag_aarch64_verbose_cost) | |
61263118 KT |
7429 | fprintf (dump_file, |
7430 | "\nFailed to cost RTX. Assuming default cost.\n"); | |
7431 | ||
7432 | return true; | |
43e9d192 IB |
7433 | } |
7434 | ||
0ee859b5 JG |
7435 | /* Wrapper around aarch64_rtx_costs, dumps the partial, or total cost |
7436 | calculated for X. This cost is stored in *COST. Returns true | |
7437 | if the total cost of X was calculated. */ | |
7438 | static bool | |
e548c9df | 7439 | aarch64_rtx_costs_wrapper (rtx x, machine_mode mode, int outer, |
0ee859b5 JG |
7440 | int param, int *cost, bool speed) |
7441 | { | |
e548c9df | 7442 | bool result = aarch64_rtx_costs (x, mode, outer, param, cost, speed); |
0ee859b5 | 7443 | |
c10e3d7f AP |
7444 | if (dump_file |
7445 | && flag_aarch64_verbose_cost) | |
0ee859b5 JG |
7446 | { |
7447 | print_rtl_single (dump_file, x); | |
7448 | fprintf (dump_file, "\n%s cost: %d (%s)\n", | |
7449 | speed ? "Hot" : "Cold", | |
7450 | *cost, result ? "final" : "partial"); | |
7451 | } | |
7452 | ||
7453 | return result; | |
7454 | } | |
7455 | ||
43e9d192 | 7456 | static int |
ef4bddc2 | 7457 | aarch64_register_move_cost (machine_mode mode, |
8a3a7e67 | 7458 | reg_class_t from_i, reg_class_t to_i) |
43e9d192 | 7459 | { |
8a3a7e67 RH |
7460 | enum reg_class from = (enum reg_class) from_i; |
7461 | enum reg_class to = (enum reg_class) to_i; | |
43e9d192 | 7462 | const struct cpu_regmove_cost *regmove_cost |
b175b679 | 7463 | = aarch64_tune_params.regmove_cost; |
43e9d192 | 7464 | |
3be07662 | 7465 | /* Caller save and pointer regs are equivalent to GENERAL_REGS. */ |
2876a13f | 7466 | if (to == CALLER_SAVE_REGS || to == POINTER_REGS) |
3be07662 WD |
7467 | to = GENERAL_REGS; |
7468 | ||
2876a13f | 7469 | if (from == CALLER_SAVE_REGS || from == POINTER_REGS) |
3be07662 WD |
7470 | from = GENERAL_REGS; |
7471 | ||
6ee70f81 AP |
7472 | /* Moving between GPR and stack cost is the same as GP2GP. */ |
7473 | if ((from == GENERAL_REGS && to == STACK_REG) | |
7474 | || (to == GENERAL_REGS && from == STACK_REG)) | |
7475 | return regmove_cost->GP2GP; | |
7476 | ||
7477 | /* To/From the stack register, we move via the gprs. */ | |
7478 | if (to == STACK_REG || from == STACK_REG) | |
7479 | return aarch64_register_move_cost (mode, from, GENERAL_REGS) | |
7480 | + aarch64_register_move_cost (mode, GENERAL_REGS, to); | |
7481 | ||
8919453c WD |
7482 | if (GET_MODE_SIZE (mode) == 16) |
7483 | { | |
7484 | /* 128-bit operations on general registers require 2 instructions. */ | |
7485 | if (from == GENERAL_REGS && to == GENERAL_REGS) | |
7486 | return regmove_cost->GP2GP * 2; | |
7487 | else if (from == GENERAL_REGS) | |
7488 | return regmove_cost->GP2FP * 2; | |
7489 | else if (to == GENERAL_REGS) | |
7490 | return regmove_cost->FP2GP * 2; | |
7491 | ||
7492 | /* When AdvSIMD instructions are disabled it is not possible to move | |
7493 | a 128-bit value directly between Q registers. This is handled in | |
7494 | secondary reload. A general register is used as a scratch to move | |
7495 | the upper DI value and the lower DI value is moved directly, | |
7496 | hence the cost is the sum of three moves. */ | |
7497 | if (! TARGET_SIMD) | |
7498 | return regmove_cost->GP2FP + regmove_cost->FP2GP + regmove_cost->FP2FP; | |
7499 | ||
7500 | return regmove_cost->FP2FP; | |
7501 | } | |
7502 | ||
43e9d192 IB |
7503 | if (from == GENERAL_REGS && to == GENERAL_REGS) |
7504 | return regmove_cost->GP2GP; | |
7505 | else if (from == GENERAL_REGS) | |
7506 | return regmove_cost->GP2FP; | |
7507 | else if (to == GENERAL_REGS) | |
7508 | return regmove_cost->FP2GP; | |
7509 | ||
43e9d192 IB |
7510 | return regmove_cost->FP2FP; |
7511 | } | |
7512 | ||
7513 | static int | |
ef4bddc2 | 7514 | aarch64_memory_move_cost (machine_mode mode ATTRIBUTE_UNUSED, |
43e9d192 IB |
7515 | reg_class_t rclass ATTRIBUTE_UNUSED, |
7516 | bool in ATTRIBUTE_UNUSED) | |
7517 | { | |
b175b679 | 7518 | return aarch64_tune_params.memmov_cost; |
43e9d192 IB |
7519 | } |
7520 | ||
0c30e0f3 EM |
7521 | /* Return true if it is safe and beneficial to use the approximate rsqrt optabs |
7522 | to optimize 1.0/sqrt. */ | |
ee62a5a6 RS |
7523 | |
7524 | static bool | |
9acc9cbe | 7525 | use_rsqrt_p (machine_mode mode) |
ee62a5a6 RS |
7526 | { |
7527 | return (!flag_trapping_math | |
7528 | && flag_unsafe_math_optimizations | |
9acc9cbe EM |
7529 | && ((aarch64_tune_params.approx_modes->recip_sqrt |
7530 | & AARCH64_APPROX_MODE (mode)) | |
1a33079e | 7531 | || flag_mrecip_low_precision_sqrt)); |
ee62a5a6 RS |
7532 | } |
7533 | ||
0c30e0f3 EM |
7534 | /* Function to decide when to use the approximate reciprocal square root |
7535 | builtin. */ | |
a6fc00da BH |
7536 | |
7537 | static tree | |
ee62a5a6 | 7538 | aarch64_builtin_reciprocal (tree fndecl) |
a6fc00da | 7539 | { |
9acc9cbe EM |
7540 | machine_mode mode = TYPE_MODE (TREE_TYPE (fndecl)); |
7541 | ||
7542 | if (!use_rsqrt_p (mode)) | |
a6fc00da | 7543 | return NULL_TREE; |
ee62a5a6 | 7544 | return aarch64_builtin_rsqrt (DECL_FUNCTION_CODE (fndecl)); |
a6fc00da BH |
7545 | } |
7546 | ||
7547 | typedef rtx (*rsqrte_type) (rtx, rtx); | |
7548 | ||
98daafa0 EM |
7549 | /* Select reciprocal square root initial estimate insn depending on machine |
7550 | mode. */ | |
a6fc00da | 7551 | |
98daafa0 | 7552 | static rsqrte_type |
a6fc00da BH |
7553 | get_rsqrte_type (machine_mode mode) |
7554 | { | |
7555 | switch (mode) | |
7556 | { | |
2a823433 JW |
7557 | case DFmode: return gen_aarch64_rsqrtedf; |
7558 | case SFmode: return gen_aarch64_rsqrtesf; | |
7559 | case V2DFmode: return gen_aarch64_rsqrtev2df; | |
7560 | case V2SFmode: return gen_aarch64_rsqrtev2sf; | |
7561 | case V4SFmode: return gen_aarch64_rsqrtev4sf; | |
a6fc00da BH |
7562 | default: gcc_unreachable (); |
7563 | } | |
7564 | } | |
7565 | ||
7566 | typedef rtx (*rsqrts_type) (rtx, rtx, rtx); | |
7567 | ||
98daafa0 | 7568 | /* Select reciprocal square root series step insn depending on machine mode. */ |
a6fc00da | 7569 | |
98daafa0 | 7570 | static rsqrts_type |
a6fc00da BH |
7571 | get_rsqrts_type (machine_mode mode) |
7572 | { | |
7573 | switch (mode) | |
7574 | { | |
00ea75d4 JW |
7575 | case DFmode: return gen_aarch64_rsqrtsdf; |
7576 | case SFmode: return gen_aarch64_rsqrtssf; | |
7577 | case V2DFmode: return gen_aarch64_rsqrtsv2df; | |
7578 | case V2SFmode: return gen_aarch64_rsqrtsv2sf; | |
7579 | case V4SFmode: return gen_aarch64_rsqrtsv4sf; | |
a6fc00da BH |
7580 | default: gcc_unreachable (); |
7581 | } | |
7582 | } | |
7583 | ||
98daafa0 EM |
7584 | /* Emit instruction sequence to compute either the approximate square root |
7585 | or its approximate reciprocal, depending on the flag RECP, and return | |
7586 | whether the sequence was emitted or not. */ | |
a6fc00da | 7587 | |
98daafa0 EM |
7588 | bool |
7589 | aarch64_emit_approx_sqrt (rtx dst, rtx src, bool recp) | |
a6fc00da | 7590 | { |
98daafa0 | 7591 | machine_mode mode = GET_MODE (dst); |
daef0a8c JW |
7592 | |
7593 | if (GET_MODE_INNER (mode) == HFmode) | |
7594 | return false; | |
7595 | ||
98daafa0 EM |
7596 | machine_mode mmsk = mode_for_vector |
7597 | (int_mode_for_mode (GET_MODE_INNER (mode)), | |
7598 | GET_MODE_NUNITS (mode)); | |
7599 | bool use_approx_sqrt_p = (!recp | |
7600 | && (flag_mlow_precision_sqrt | |
7601 | || (aarch64_tune_params.approx_modes->sqrt | |
7602 | & AARCH64_APPROX_MODE (mode)))); | |
7603 | bool use_approx_rsqrt_p = (recp | |
7604 | && (flag_mrecip_low_precision_sqrt | |
7605 | || (aarch64_tune_params.approx_modes->recip_sqrt | |
7606 | & AARCH64_APPROX_MODE (mode)))); | |
7607 | ||
7608 | if (!flag_finite_math_only | |
7609 | || flag_trapping_math | |
7610 | || !flag_unsafe_math_optimizations | |
7611 | || !(use_approx_sqrt_p || use_approx_rsqrt_p) | |
7612 | || optimize_function_for_size_p (cfun)) | |
7613 | return false; | |
a6fc00da | 7614 | |
98daafa0 EM |
7615 | rtx xmsk = gen_reg_rtx (mmsk); |
7616 | if (!recp) | |
7617 | /* When calculating the approximate square root, compare the argument with | |
7618 | 0.0 and create a mask. */ | |
7619 | emit_insn (gen_rtx_SET (xmsk, gen_rtx_NEG (mmsk, gen_rtx_EQ (mmsk, src, | |
7620 | CONST0_RTX (mode))))); | |
a6fc00da | 7621 | |
98daafa0 EM |
7622 | /* Estimate the approximate reciprocal square root. */ |
7623 | rtx xdst = gen_reg_rtx (mode); | |
7624 | emit_insn ((*get_rsqrte_type (mode)) (xdst, src)); | |
a6fc00da | 7625 | |
98daafa0 EM |
7626 | /* Iterate over the series twice for SF and thrice for DF. */ |
7627 | int iterations = (GET_MODE_INNER (mode) == DFmode) ? 3 : 2; | |
a6fc00da | 7628 | |
98daafa0 EM |
7629 | /* Optionally iterate over the series once less for faster performance |
7630 | while sacrificing the accuracy. */ | |
7631 | if ((recp && flag_mrecip_low_precision_sqrt) | |
7632 | || (!recp && flag_mlow_precision_sqrt)) | |
a6fc00da BH |
7633 | iterations--; |
7634 | ||
98daafa0 EM |
7635 | /* Iterate over the series to calculate the approximate reciprocal square |
7636 | root. */ | |
7637 | rtx x1 = gen_reg_rtx (mode); | |
7638 | while (iterations--) | |
a6fc00da | 7639 | { |
a6fc00da | 7640 | rtx x2 = gen_reg_rtx (mode); |
98daafa0 EM |
7641 | emit_set_insn (x2, gen_rtx_MULT (mode, xdst, xdst)); |
7642 | ||
7643 | emit_insn ((*get_rsqrts_type (mode)) (x1, src, x2)); | |
a6fc00da | 7644 | |
98daafa0 EM |
7645 | if (iterations > 0) |
7646 | emit_set_insn (xdst, gen_rtx_MULT (mode, xdst, x1)); | |
7647 | } | |
7648 | ||
7649 | if (!recp) | |
7650 | { | |
7651 | /* Qualify the approximate reciprocal square root when the argument is | |
7652 | 0.0 by squashing the intermediary result to 0.0. */ | |
7653 | rtx xtmp = gen_reg_rtx (mmsk); | |
7654 | emit_set_insn (xtmp, gen_rtx_AND (mmsk, gen_rtx_NOT (mmsk, xmsk), | |
7655 | gen_rtx_SUBREG (mmsk, xdst, 0))); | |
7656 | emit_move_insn (xdst, gen_rtx_SUBREG (mode, xtmp, 0)); | |
a6fc00da | 7657 | |
98daafa0 EM |
7658 | /* Calculate the approximate square root. */ |
7659 | emit_set_insn (xdst, gen_rtx_MULT (mode, xdst, src)); | |
a6fc00da BH |
7660 | } |
7661 | ||
98daafa0 EM |
7662 | /* Finalize the approximation. */ |
7663 | emit_set_insn (dst, gen_rtx_MULT (mode, xdst, x1)); | |
7664 | ||
7665 | return true; | |
a6fc00da BH |
7666 | } |
7667 | ||
79a2bc2d EM |
7668 | typedef rtx (*recpe_type) (rtx, rtx); |
7669 | ||
7670 | /* Select reciprocal initial estimate insn depending on machine mode. */ | |
7671 | ||
7672 | static recpe_type | |
7673 | get_recpe_type (machine_mode mode) | |
7674 | { | |
7675 | switch (mode) | |
7676 | { | |
7677 | case SFmode: return (gen_aarch64_frecpesf); | |
7678 | case V2SFmode: return (gen_aarch64_frecpev2sf); | |
7679 | case V4SFmode: return (gen_aarch64_frecpev4sf); | |
7680 | case DFmode: return (gen_aarch64_frecpedf); | |
7681 | case V2DFmode: return (gen_aarch64_frecpev2df); | |
7682 | default: gcc_unreachable (); | |
7683 | } | |
7684 | } | |
7685 | ||
7686 | typedef rtx (*recps_type) (rtx, rtx, rtx); | |
7687 | ||
7688 | /* Select reciprocal series step insn depending on machine mode. */ | |
7689 | ||
7690 | static recps_type | |
7691 | get_recps_type (machine_mode mode) | |
7692 | { | |
7693 | switch (mode) | |
7694 | { | |
7695 | case SFmode: return (gen_aarch64_frecpssf); | |
7696 | case V2SFmode: return (gen_aarch64_frecpsv2sf); | |
7697 | case V4SFmode: return (gen_aarch64_frecpsv4sf); | |
7698 | case DFmode: return (gen_aarch64_frecpsdf); | |
7699 | case V2DFmode: return (gen_aarch64_frecpsv2df); | |
7700 | default: gcc_unreachable (); | |
7701 | } | |
7702 | } | |
7703 | ||
7704 | /* Emit the instruction sequence to compute the approximation for the division | |
7705 | of NUM by DEN in QUO and return whether the sequence was emitted or not. */ | |
7706 | ||
7707 | bool | |
7708 | aarch64_emit_approx_div (rtx quo, rtx num, rtx den) | |
7709 | { | |
7710 | machine_mode mode = GET_MODE (quo); | |
33d72b63 JW |
7711 | |
7712 | if (GET_MODE_INNER (mode) == HFmode) | |
7713 | return false; | |
7714 | ||
79a2bc2d EM |
7715 | bool use_approx_division_p = (flag_mlow_precision_div |
7716 | || (aarch64_tune_params.approx_modes->division | |
7717 | & AARCH64_APPROX_MODE (mode))); | |
7718 | ||
7719 | if (!flag_finite_math_only | |
7720 | || flag_trapping_math | |
7721 | || !flag_unsafe_math_optimizations | |
7722 | || optimize_function_for_size_p (cfun) | |
7723 | || !use_approx_division_p) | |
7724 | return false; | |
7725 | ||
7726 | /* Estimate the approximate reciprocal. */ | |
7727 | rtx xrcp = gen_reg_rtx (mode); | |
7728 | emit_insn ((*get_recpe_type (mode)) (xrcp, den)); | |
7729 | ||
7730 | /* Iterate over the series twice for SF and thrice for DF. */ | |
7731 | int iterations = (GET_MODE_INNER (mode) == DFmode) ? 3 : 2; | |
7732 | ||
7733 | /* Optionally iterate over the series once less for faster performance, | |
7734 | while sacrificing the accuracy. */ | |
7735 | if (flag_mlow_precision_div) | |
7736 | iterations--; | |
7737 | ||
7738 | /* Iterate over the series to calculate the approximate reciprocal. */ | |
7739 | rtx xtmp = gen_reg_rtx (mode); | |
7740 | while (iterations--) | |
7741 | { | |
7742 | emit_insn ((*get_recps_type (mode)) (xtmp, xrcp, den)); | |
7743 | ||
7744 | if (iterations > 0) | |
7745 | emit_set_insn (xrcp, gen_rtx_MULT (mode, xrcp, xtmp)); | |
7746 | } | |
7747 | ||
7748 | if (num != CONST1_RTX (mode)) | |
7749 | { | |
7750 | /* As the approximate reciprocal of DEN is already calculated, only | |
7751 | calculate the approximate division when NUM is not 1.0. */ | |
7752 | rtx xnum = force_reg (mode, num); | |
7753 | emit_set_insn (xrcp, gen_rtx_MULT (mode, xrcp, xnum)); | |
7754 | } | |
7755 | ||
7756 | /* Finalize the approximation. */ | |
7757 | emit_set_insn (quo, gen_rtx_MULT (mode, xrcp, xtmp)); | |
7758 | return true; | |
7759 | } | |
7760 | ||
d126a4ae AP |
7761 | /* Return the number of instructions that can be issued per cycle. */ |
7762 | static int | |
7763 | aarch64_sched_issue_rate (void) | |
7764 | { | |
b175b679 | 7765 | return aarch64_tune_params.issue_rate; |
d126a4ae AP |
7766 | } |
7767 | ||
d03f7e44 MK |
7768 | static int |
7769 | aarch64_sched_first_cycle_multipass_dfa_lookahead (void) | |
7770 | { | |
7771 | int issue_rate = aarch64_sched_issue_rate (); | |
7772 | ||
7773 | return issue_rate > 1 && !sched_fusion ? issue_rate : 0; | |
7774 | } | |
7775 | ||
2d6bc7fa KT |
7776 | |
7777 | /* Implement TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD as | |
7778 | autopref_multipass_dfa_lookahead_guard from haifa-sched.c. It only | |
7779 | has an effect if PARAM_SCHED_AUTOPREF_QUEUE_DEPTH > 0. */ | |
7780 | ||
7781 | static int | |
7782 | aarch64_first_cycle_multipass_dfa_lookahead_guard (rtx_insn *insn, | |
7783 | int ready_index) | |
7784 | { | |
7785 | return autopref_multipass_dfa_lookahead_guard (insn, ready_index); | |
7786 | } | |
7787 | ||
7788 | ||
8990e73a TB |
7789 | /* Vectorizer cost model target hooks. */ |
7790 | ||
7791 | /* Implement targetm.vectorize.builtin_vectorization_cost. */ | |
7792 | static int | |
7793 | aarch64_builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost, | |
7794 | tree vectype, | |
7795 | int misalign ATTRIBUTE_UNUSED) | |
7796 | { | |
7797 | unsigned elements; | |
7798 | ||
7799 | switch (type_of_cost) | |
7800 | { | |
7801 | case scalar_stmt: | |
b175b679 | 7802 | return aarch64_tune_params.vec_costs->scalar_stmt_cost; |
8990e73a TB |
7803 | |
7804 | case scalar_load: | |
b175b679 | 7805 | return aarch64_tune_params.vec_costs->scalar_load_cost; |
8990e73a TB |
7806 | |
7807 | case scalar_store: | |
b175b679 | 7808 | return aarch64_tune_params.vec_costs->scalar_store_cost; |
8990e73a TB |
7809 | |
7810 | case vector_stmt: | |
b175b679 | 7811 | return aarch64_tune_params.vec_costs->vec_stmt_cost; |
8990e73a TB |
7812 | |
7813 | case vector_load: | |
b175b679 | 7814 | return aarch64_tune_params.vec_costs->vec_align_load_cost; |
8990e73a TB |
7815 | |
7816 | case vector_store: | |
b175b679 | 7817 | return aarch64_tune_params.vec_costs->vec_store_cost; |
8990e73a TB |
7818 | |
7819 | case vec_to_scalar: | |
b175b679 | 7820 | return aarch64_tune_params.vec_costs->vec_to_scalar_cost; |
8990e73a TB |
7821 | |
7822 | case scalar_to_vec: | |
b175b679 | 7823 | return aarch64_tune_params.vec_costs->scalar_to_vec_cost; |
8990e73a TB |
7824 | |
7825 | case unaligned_load: | |
b175b679 | 7826 | return aarch64_tune_params.vec_costs->vec_unalign_load_cost; |
8990e73a TB |
7827 | |
7828 | case unaligned_store: | |
b175b679 | 7829 | return aarch64_tune_params.vec_costs->vec_unalign_store_cost; |
8990e73a TB |
7830 | |
7831 | case cond_branch_taken: | |
b175b679 | 7832 | return aarch64_tune_params.vec_costs->cond_taken_branch_cost; |
8990e73a TB |
7833 | |
7834 | case cond_branch_not_taken: | |
b175b679 | 7835 | return aarch64_tune_params.vec_costs->cond_not_taken_branch_cost; |
8990e73a TB |
7836 | |
7837 | case vec_perm: | |
c428f91c WD |
7838 | return aarch64_tune_params.vec_costs->vec_permute_cost; |
7839 | ||
8990e73a | 7840 | case vec_promote_demote: |
b175b679 | 7841 | return aarch64_tune_params.vec_costs->vec_stmt_cost; |
8990e73a TB |
7842 | |
7843 | case vec_construct: | |
7844 | elements = TYPE_VECTOR_SUBPARTS (vectype); | |
7845 | return elements / 2 + 1; | |
7846 | ||
7847 | default: | |
7848 | gcc_unreachable (); | |
7849 | } | |
7850 | } | |
7851 | ||
7852 | /* Implement targetm.vectorize.add_stmt_cost. */ | |
7853 | static unsigned | |
7854 | aarch64_add_stmt_cost (void *data, int count, enum vect_cost_for_stmt kind, | |
7855 | struct _stmt_vec_info *stmt_info, int misalign, | |
7856 | enum vect_cost_model_location where) | |
7857 | { | |
7858 | unsigned *cost = (unsigned *) data; | |
7859 | unsigned retval = 0; | |
7860 | ||
7861 | if (flag_vect_cost_model) | |
7862 | { | |
7863 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; | |
7864 | int stmt_cost = | |
7865 | aarch64_builtin_vectorization_cost (kind, vectype, misalign); | |
7866 | ||
7867 | /* Statements in an inner loop relative to the loop being | |
7868 | vectorized are weighted more heavily. The value here is | |
058e4c71 | 7869 | arbitrary and could potentially be improved with analysis. */ |
8990e73a | 7870 | if (where == vect_body && stmt_info && stmt_in_inner_loop_p (stmt_info)) |
058e4c71 | 7871 | count *= 50; /* FIXME */ |
8990e73a TB |
7872 | |
7873 | retval = (unsigned) (count * stmt_cost); | |
7874 | cost[where] += retval; | |
7875 | } | |
7876 | ||
7877 | return retval; | |
7878 | } | |
7879 | ||
0cfff2a1 | 7880 | static void initialize_aarch64_code_model (struct gcc_options *); |
43e9d192 | 7881 | |
0cfff2a1 KT |
7882 | /* Parse the TO_PARSE string and put the architecture struct that it |
7883 | selects into RES and the architectural features into ISA_FLAGS. | |
7884 | Return an aarch64_parse_opt_result describing the parse result. | |
7885 | If there is an error parsing, RES and ISA_FLAGS are left unchanged. */ | |
43e9d192 | 7886 | |
0cfff2a1 KT |
7887 | static enum aarch64_parse_opt_result |
7888 | aarch64_parse_arch (const char *to_parse, const struct processor **res, | |
7889 | unsigned long *isa_flags) | |
43e9d192 IB |
7890 | { |
7891 | char *ext; | |
7892 | const struct processor *arch; | |
0cfff2a1 | 7893 | char *str = (char *) alloca (strlen (to_parse) + 1); |
43e9d192 IB |
7894 | size_t len; |
7895 | ||
0cfff2a1 | 7896 | strcpy (str, to_parse); |
43e9d192 IB |
7897 | |
7898 | ext = strchr (str, '+'); | |
7899 | ||
7900 | if (ext != NULL) | |
7901 | len = ext - str; | |
7902 | else | |
7903 | len = strlen (str); | |
7904 | ||
7905 | if (len == 0) | |
0cfff2a1 KT |
7906 | return AARCH64_PARSE_MISSING_ARG; |
7907 | ||
43e9d192 | 7908 | |
0cfff2a1 | 7909 | /* Loop through the list of supported ARCHes to find a match. */ |
43e9d192 IB |
7910 | for (arch = all_architectures; arch->name != NULL; arch++) |
7911 | { | |
7912 | if (strlen (arch->name) == len && strncmp (arch->name, str, len) == 0) | |
7913 | { | |
0cfff2a1 | 7914 | unsigned long isa_temp = arch->flags; |
43e9d192 IB |
7915 | |
7916 | if (ext != NULL) | |
7917 | { | |
0cfff2a1 KT |
7918 | /* TO_PARSE string contains at least one extension. */ |
7919 | enum aarch64_parse_opt_result ext_res | |
7920 | = aarch64_parse_extension (ext, &isa_temp); | |
43e9d192 | 7921 | |
0cfff2a1 KT |
7922 | if (ext_res != AARCH64_PARSE_OK) |
7923 | return ext_res; | |
ffee7aa9 | 7924 | } |
0cfff2a1 KT |
7925 | /* Extension parsing was successful. Confirm the result |
7926 | arch and ISA flags. */ | |
7927 | *res = arch; | |
7928 | *isa_flags = isa_temp; | |
7929 | return AARCH64_PARSE_OK; | |
43e9d192 IB |
7930 | } |
7931 | } | |
7932 | ||
7933 | /* ARCH name not found in list. */ | |
0cfff2a1 | 7934 | return AARCH64_PARSE_INVALID_ARG; |
43e9d192 IB |
7935 | } |
7936 | ||
0cfff2a1 KT |
7937 | /* Parse the TO_PARSE string and put the result tuning in RES and the |
7938 | architecture flags in ISA_FLAGS. Return an aarch64_parse_opt_result | |
7939 | describing the parse result. If there is an error parsing, RES and | |
7940 | ISA_FLAGS are left unchanged. */ | |
43e9d192 | 7941 | |
0cfff2a1 KT |
7942 | static enum aarch64_parse_opt_result |
7943 | aarch64_parse_cpu (const char *to_parse, const struct processor **res, | |
7944 | unsigned long *isa_flags) | |
43e9d192 IB |
7945 | { |
7946 | char *ext; | |
7947 | const struct processor *cpu; | |
0cfff2a1 | 7948 | char *str = (char *) alloca (strlen (to_parse) + 1); |
43e9d192 IB |
7949 | size_t len; |
7950 | ||
0cfff2a1 | 7951 | strcpy (str, to_parse); |
43e9d192 IB |
7952 | |
7953 | ext = strchr (str, '+'); | |
7954 | ||
7955 | if (ext != NULL) | |
7956 | len = ext - str; | |
7957 | else | |
7958 | len = strlen (str); | |
7959 | ||
7960 | if (len == 0) | |
0cfff2a1 KT |
7961 | return AARCH64_PARSE_MISSING_ARG; |
7962 | ||
43e9d192 IB |
7963 | |
7964 | /* Loop through the list of supported CPUs to find a match. */ | |
7965 | for (cpu = all_cores; cpu->name != NULL; cpu++) | |
7966 | { | |
7967 | if (strlen (cpu->name) == len && strncmp (cpu->name, str, len) == 0) | |
7968 | { | |
0cfff2a1 KT |
7969 | unsigned long isa_temp = cpu->flags; |
7970 | ||
43e9d192 IB |
7971 | |
7972 | if (ext != NULL) | |
7973 | { | |
0cfff2a1 KT |
7974 | /* TO_PARSE string contains at least one extension. */ |
7975 | enum aarch64_parse_opt_result ext_res | |
7976 | = aarch64_parse_extension (ext, &isa_temp); | |
43e9d192 | 7977 | |
0cfff2a1 KT |
7978 | if (ext_res != AARCH64_PARSE_OK) |
7979 | return ext_res; | |
7980 | } | |
7981 | /* Extension parsing was successfull. Confirm the result | |
7982 | cpu and ISA flags. */ | |
7983 | *res = cpu; | |
7984 | *isa_flags = isa_temp; | |
7985 | return AARCH64_PARSE_OK; | |
43e9d192 IB |
7986 | } |
7987 | } | |
7988 | ||
7989 | /* CPU name not found in list. */ | |
0cfff2a1 | 7990 | return AARCH64_PARSE_INVALID_ARG; |
43e9d192 IB |
7991 | } |
7992 | ||
0cfff2a1 KT |
7993 | /* Parse the TO_PARSE string and put the cpu it selects into RES. |
7994 | Return an aarch64_parse_opt_result describing the parse result. | |
7995 | If the parsing fails the RES does not change. */ | |
43e9d192 | 7996 | |
0cfff2a1 KT |
7997 | static enum aarch64_parse_opt_result |
7998 | aarch64_parse_tune (const char *to_parse, const struct processor **res) | |
43e9d192 IB |
7999 | { |
8000 | const struct processor *cpu; | |
0cfff2a1 KT |
8001 | char *str = (char *) alloca (strlen (to_parse) + 1); |
8002 | ||
8003 | strcpy (str, to_parse); | |
43e9d192 IB |
8004 | |
8005 | /* Loop through the list of supported CPUs to find a match. */ | |
8006 | for (cpu = all_cores; cpu->name != NULL; cpu++) | |
8007 | { | |
8008 | if (strcmp (cpu->name, str) == 0) | |
8009 | { | |
0cfff2a1 KT |
8010 | *res = cpu; |
8011 | return AARCH64_PARSE_OK; | |
43e9d192 IB |
8012 | } |
8013 | } | |
8014 | ||
8015 | /* CPU name not found in list. */ | |
0cfff2a1 | 8016 | return AARCH64_PARSE_INVALID_ARG; |
43e9d192 IB |
8017 | } |
8018 | ||
8dec06f2 JG |
8019 | /* Parse TOKEN, which has length LENGTH to see if it is an option |
8020 | described in FLAG. If it is, return the index bit for that fusion type. | |
8021 | If not, error (printing OPTION_NAME) and return zero. */ | |
8022 | ||
8023 | static unsigned int | |
8024 | aarch64_parse_one_option_token (const char *token, | |
8025 | size_t length, | |
8026 | const struct aarch64_flag_desc *flag, | |
8027 | const char *option_name) | |
8028 | { | |
8029 | for (; flag->name != NULL; flag++) | |
8030 | { | |
8031 | if (length == strlen (flag->name) | |
8032 | && !strncmp (flag->name, token, length)) | |
8033 | return flag->flag; | |
8034 | } | |
8035 | ||
8036 | error ("unknown flag passed in -moverride=%s (%s)", option_name, token); | |
8037 | return 0; | |
8038 | } | |
8039 | ||
8040 | /* Parse OPTION which is a comma-separated list of flags to enable. | |
8041 | FLAGS gives the list of flags we understand, INITIAL_STATE gives any | |
8042 | default state we inherit from the CPU tuning structures. OPTION_NAME | |
8043 | gives the top-level option we are parsing in the -moverride string, | |
8044 | for use in error messages. */ | |
8045 | ||
8046 | static unsigned int | |
8047 | aarch64_parse_boolean_options (const char *option, | |
8048 | const struct aarch64_flag_desc *flags, | |
8049 | unsigned int initial_state, | |
8050 | const char *option_name) | |
8051 | { | |
8052 | const char separator = '.'; | |
8053 | const char* specs = option; | |
8054 | const char* ntoken = option; | |
8055 | unsigned int found_flags = initial_state; | |
8056 | ||
8057 | while ((ntoken = strchr (specs, separator))) | |
8058 | { | |
8059 | size_t token_length = ntoken - specs; | |
8060 | unsigned token_ops = aarch64_parse_one_option_token (specs, | |
8061 | token_length, | |
8062 | flags, | |
8063 | option_name); | |
8064 | /* If we find "none" (or, for simplicity's sake, an error) anywhere | |
8065 | in the token stream, reset the supported operations. So: | |
8066 | ||
8067 | adrp+add.cmp+branch.none.adrp+add | |
8068 | ||
8069 | would have the result of turning on only adrp+add fusion. */ | |
8070 | if (!token_ops) | |
8071 | found_flags = 0; | |
8072 | ||
8073 | found_flags |= token_ops; | |
8074 | specs = ++ntoken; | |
8075 | } | |
8076 | ||
8077 | /* We ended with a comma, print something. */ | |
8078 | if (!(*specs)) | |
8079 | { | |
8080 | error ("%s string ill-formed\n", option_name); | |
8081 | return 0; | |
8082 | } | |
8083 | ||
8084 | /* We still have one more token to parse. */ | |
8085 | size_t token_length = strlen (specs); | |
8086 | unsigned token_ops = aarch64_parse_one_option_token (specs, | |
8087 | token_length, | |
8088 | flags, | |
8089 | option_name); | |
8090 | if (!token_ops) | |
8091 | found_flags = 0; | |
8092 | ||
8093 | found_flags |= token_ops; | |
8094 | return found_flags; | |
8095 | } | |
8096 | ||
8097 | /* Support for overriding instruction fusion. */ | |
8098 | ||
8099 | static void | |
8100 | aarch64_parse_fuse_string (const char *fuse_string, | |
8101 | struct tune_params *tune) | |
8102 | { | |
8103 | tune->fusible_ops = aarch64_parse_boolean_options (fuse_string, | |
8104 | aarch64_fusible_pairs, | |
8105 | tune->fusible_ops, | |
8106 | "fuse="); | |
8107 | } | |
8108 | ||
8109 | /* Support for overriding other tuning flags. */ | |
8110 | ||
8111 | static void | |
8112 | aarch64_parse_tune_string (const char *tune_string, | |
8113 | struct tune_params *tune) | |
8114 | { | |
8115 | tune->extra_tuning_flags | |
8116 | = aarch64_parse_boolean_options (tune_string, | |
8117 | aarch64_tuning_flags, | |
8118 | tune->extra_tuning_flags, | |
8119 | "tune="); | |
8120 | } | |
8121 | ||
8122 | /* Parse TOKEN, which has length LENGTH to see if it is a tuning option | |
8123 | we understand. If it is, extract the option string and handoff to | |
8124 | the appropriate function. */ | |
8125 | ||
8126 | void | |
8127 | aarch64_parse_one_override_token (const char* token, | |
8128 | size_t length, | |
8129 | struct tune_params *tune) | |
8130 | { | |
8131 | const struct aarch64_tuning_override_function *fn | |
8132 | = aarch64_tuning_override_functions; | |
8133 | ||
8134 | const char *option_part = strchr (token, '='); | |
8135 | if (!option_part) | |
8136 | { | |
8137 | error ("tuning string missing in option (%s)", token); | |
8138 | return; | |
8139 | } | |
8140 | ||
8141 | /* Get the length of the option name. */ | |
8142 | length = option_part - token; | |
8143 | /* Skip the '=' to get to the option string. */ | |
8144 | option_part++; | |
8145 | ||
8146 | for (; fn->name != NULL; fn++) | |
8147 | { | |
8148 | if (!strncmp (fn->name, token, length)) | |
8149 | { | |
8150 | fn->parse_override (option_part, tune); | |
8151 | return; | |
8152 | } | |
8153 | } | |
8154 | ||
8155 | error ("unknown tuning option (%s)",token); | |
8156 | return; | |
8157 | } | |
8158 | ||
5eee3c34 JW |
8159 | /* A checking mechanism for the implementation of the tls size. */ |
8160 | ||
8161 | static void | |
8162 | initialize_aarch64_tls_size (struct gcc_options *opts) | |
8163 | { | |
8164 | if (aarch64_tls_size == 0) | |
8165 | aarch64_tls_size = 24; | |
8166 | ||
8167 | switch (opts->x_aarch64_cmodel_var) | |
8168 | { | |
8169 | case AARCH64_CMODEL_TINY: | |
8170 | /* Both the default and maximum TLS size allowed under tiny is 1M which | |
8171 | needs two instructions to address, so we clamp the size to 24. */ | |
8172 | if (aarch64_tls_size > 24) | |
8173 | aarch64_tls_size = 24; | |
8174 | break; | |
8175 | case AARCH64_CMODEL_SMALL: | |
8176 | /* The maximum TLS size allowed under small is 4G. */ | |
8177 | if (aarch64_tls_size > 32) | |
8178 | aarch64_tls_size = 32; | |
8179 | break; | |
8180 | case AARCH64_CMODEL_LARGE: | |
8181 | /* The maximum TLS size allowed under large is 16E. | |
8182 | FIXME: 16E should be 64bit, we only support 48bit offset now. */ | |
8183 | if (aarch64_tls_size > 48) | |
8184 | aarch64_tls_size = 48; | |
8185 | break; | |
8186 | default: | |
8187 | gcc_unreachable (); | |
8188 | } | |
8189 | ||
8190 | return; | |
8191 | } | |
8192 | ||
8dec06f2 JG |
8193 | /* Parse STRING looking for options in the format: |
8194 | string :: option:string | |
8195 | option :: name=substring | |
8196 | name :: {a-z} | |
8197 | substring :: defined by option. */ | |
8198 | ||
8199 | static void | |
8200 | aarch64_parse_override_string (const char* input_string, | |
8201 | struct tune_params* tune) | |
8202 | { | |
8203 | const char separator = ':'; | |
8204 | size_t string_length = strlen (input_string) + 1; | |
8205 | char *string_root = (char *) xmalloc (sizeof (*string_root) * string_length); | |
8206 | char *string = string_root; | |
8207 | strncpy (string, input_string, string_length); | |
8208 | string[string_length - 1] = '\0'; | |
8209 | ||
8210 | char* ntoken = string; | |
8211 | ||
8212 | while ((ntoken = strchr (string, separator))) | |
8213 | { | |
8214 | size_t token_length = ntoken - string; | |
8215 | /* Make this substring look like a string. */ | |
8216 | *ntoken = '\0'; | |
8217 | aarch64_parse_one_override_token (string, token_length, tune); | |
8218 | string = ++ntoken; | |
8219 | } | |
8220 | ||
8221 | /* One last option to parse. */ | |
8222 | aarch64_parse_one_override_token (string, strlen (string), tune); | |
8223 | free (string_root); | |
8224 | } | |
43e9d192 | 8225 | |
43e9d192 IB |
8226 | |
8227 | static void | |
0cfff2a1 | 8228 | aarch64_override_options_after_change_1 (struct gcc_options *opts) |
43e9d192 | 8229 | { |
a3dc8760 NC |
8230 | /* The logic here is that if we are disabling all frame pointer generation |
8231 | then we do not need to disable leaf frame pointer generation as a | |
8232 | separate operation. But if we are *only* disabling leaf frame pointer | |
8233 | generation then we set flag_omit_frame_pointer to true, but in | |
8234 | aarch64_frame_pointer_required we return false only for leaf functions. | |
8235 | ||
8236 | PR 70044: We have to be careful about being called multiple times for the | |
8237 | same function. Once we have decided to set flag_omit_frame_pointer just | |
8238 | so that we can omit leaf frame pointers, we must then not interpret a | |
8239 | second call as meaning that all frame pointer generation should be | |
8240 | omitted. We do this by setting flag_omit_frame_pointer to a special, | |
8241 | non-zero value. */ | |
8242 | if (opts->x_flag_omit_frame_pointer == 2) | |
8243 | opts->x_flag_omit_frame_pointer = 0; | |
8244 | ||
0cfff2a1 KT |
8245 | if (opts->x_flag_omit_frame_pointer) |
8246 | opts->x_flag_omit_leaf_frame_pointer = false; | |
8247 | else if (opts->x_flag_omit_leaf_frame_pointer) | |
a3dc8760 | 8248 | opts->x_flag_omit_frame_pointer = 2; |
43e9d192 | 8249 | |
1be34295 | 8250 | /* If not optimizing for size, set the default |
0cfff2a1 KT |
8251 | alignment to what the target wants. */ |
8252 | if (!opts->x_optimize_size) | |
43e9d192 | 8253 | { |
0cfff2a1 KT |
8254 | if (opts->x_align_loops <= 0) |
8255 | opts->x_align_loops = aarch64_tune_params.loop_align; | |
8256 | if (opts->x_align_jumps <= 0) | |
8257 | opts->x_align_jumps = aarch64_tune_params.jump_align; | |
8258 | if (opts->x_align_functions <= 0) | |
8259 | opts->x_align_functions = aarch64_tune_params.function_align; | |
43e9d192 | 8260 | } |
b4f50fd4 | 8261 | |
9ee6540a WD |
8262 | /* We default to no pc-relative literal loads. */ |
8263 | ||
8264 | aarch64_pcrelative_literal_loads = false; | |
8265 | ||
8266 | /* If -mpc-relative-literal-loads is set on the command line, this | |
b4f50fd4 | 8267 | implies that the user asked for PC relative literal loads. */ |
9ee6540a WD |
8268 | if (opts->x_pcrelative_literal_loads == 1) |
8269 | aarch64_pcrelative_literal_loads = true; | |
b4f50fd4 | 8270 | |
48bb1a55 CL |
8271 | /* This is PR70113. When building the Linux kernel with |
8272 | CONFIG_ARM64_ERRATUM_843419, support for relocations | |
8273 | R_AARCH64_ADR_PREL_PG_HI21 and R_AARCH64_ADR_PREL_PG_HI21_NC is | |
8274 | removed from the kernel to avoid loading objects with possibly | |
9ee6540a | 8275 | offending sequences. Without -mpc-relative-literal-loads we would |
48bb1a55 CL |
8276 | generate such relocations, preventing the kernel build from |
8277 | succeeding. */ | |
9ee6540a WD |
8278 | if (opts->x_pcrelative_literal_loads == 2 |
8279 | && TARGET_FIX_ERR_A53_843419) | |
8280 | aarch64_pcrelative_literal_loads = true; | |
8281 | ||
8282 | /* In the tiny memory model it makes no sense to disallow PC relative | |
8283 | literal pool loads. */ | |
8284 | if (aarch64_cmodel == AARCH64_CMODEL_TINY | |
8285 | || aarch64_cmodel == AARCH64_CMODEL_TINY_PIC) | |
8286 | aarch64_pcrelative_literal_loads = true; | |
98daafa0 EM |
8287 | |
8288 | /* When enabling the lower precision Newton series for the square root, also | |
8289 | enable it for the reciprocal square root, since the latter is an | |
8290 | intermediary step for the former. */ | |
8291 | if (flag_mlow_precision_sqrt) | |
8292 | flag_mrecip_low_precision_sqrt = true; | |
0cfff2a1 | 8293 | } |
43e9d192 | 8294 | |
0cfff2a1 KT |
8295 | /* 'Unpack' up the internal tuning structs and update the options |
8296 | in OPTS. The caller must have set up selected_tune and selected_arch | |
8297 | as all the other target-specific codegen decisions are | |
8298 | derived from them. */ | |
8299 | ||
e4ea20c8 | 8300 | void |
0cfff2a1 KT |
8301 | aarch64_override_options_internal (struct gcc_options *opts) |
8302 | { | |
8303 | aarch64_tune_flags = selected_tune->flags; | |
8304 | aarch64_tune = selected_tune->sched_core; | |
8305 | /* Make a copy of the tuning parameters attached to the core, which | |
8306 | we may later overwrite. */ | |
8307 | aarch64_tune_params = *(selected_tune->tune); | |
8308 | aarch64_architecture_version = selected_arch->architecture_version; | |
8309 | ||
8310 | if (opts->x_aarch64_override_tune_string) | |
8311 | aarch64_parse_override_string (opts->x_aarch64_override_tune_string, | |
8312 | &aarch64_tune_params); | |
8313 | ||
8314 | /* This target defaults to strict volatile bitfields. */ | |
8315 | if (opts->x_flag_strict_volatile_bitfields < 0 && abi_version_at_least (2)) | |
8316 | opts->x_flag_strict_volatile_bitfields = 1; | |
8317 | ||
0cfff2a1 | 8318 | initialize_aarch64_code_model (opts); |
5eee3c34 | 8319 | initialize_aarch64_tls_size (opts); |
63892fa2 | 8320 | |
2d6bc7fa KT |
8321 | int queue_depth = 0; |
8322 | switch (aarch64_tune_params.autoprefetcher_model) | |
8323 | { | |
8324 | case tune_params::AUTOPREFETCHER_OFF: | |
8325 | queue_depth = -1; | |
8326 | break; | |
8327 | case tune_params::AUTOPREFETCHER_WEAK: | |
8328 | queue_depth = 0; | |
8329 | break; | |
8330 | case tune_params::AUTOPREFETCHER_STRONG: | |
8331 | queue_depth = max_insn_queue_index + 1; | |
8332 | break; | |
8333 | default: | |
8334 | gcc_unreachable (); | |
8335 | } | |
8336 | ||
8337 | /* We don't mind passing in global_options_set here as we don't use | |
8338 | the *options_set structs anyway. */ | |
8339 | maybe_set_param_value (PARAM_SCHED_AUTOPREF_QUEUE_DEPTH, | |
8340 | queue_depth, | |
8341 | opts->x_param_values, | |
8342 | global_options_set.x_param_values); | |
8343 | ||
50487d79 EM |
8344 | /* Set the L1 cache line size. */ |
8345 | if (selected_cpu->tune->cache_line_size != 0) | |
8346 | maybe_set_param_value (PARAM_L1_CACHE_LINE_SIZE, | |
8347 | selected_cpu->tune->cache_line_size, | |
8348 | opts->x_param_values, | |
8349 | global_options_set.x_param_values); | |
8350 | ||
0cfff2a1 KT |
8351 | aarch64_override_options_after_change_1 (opts); |
8352 | } | |
43e9d192 | 8353 | |
01f44038 KT |
8354 | /* Print a hint with a suggestion for a core or architecture name that |
8355 | most closely resembles what the user passed in STR. ARCH is true if | |
8356 | the user is asking for an architecture name. ARCH is false if the user | |
8357 | is asking for a core name. */ | |
8358 | ||
8359 | static void | |
8360 | aarch64_print_hint_for_core_or_arch (const char *str, bool arch) | |
8361 | { | |
8362 | auto_vec<const char *> candidates; | |
8363 | const struct processor *entry = arch ? all_architectures : all_cores; | |
8364 | for (; entry->name != NULL; entry++) | |
8365 | candidates.safe_push (entry->name); | |
8366 | char *s; | |
8367 | const char *hint = candidates_list_and_hint (str, s, candidates); | |
8368 | if (hint) | |
8369 | inform (input_location, "valid arguments are: %s;" | |
8370 | " did you mean %qs?", s, hint); | |
8371 | XDELETEVEC (s); | |
8372 | } | |
8373 | ||
8374 | /* Print a hint with a suggestion for a core name that most closely resembles | |
8375 | what the user passed in STR. */ | |
8376 | ||
8377 | inline static void | |
8378 | aarch64_print_hint_for_core (const char *str) | |
8379 | { | |
8380 | aarch64_print_hint_for_core_or_arch (str, false); | |
8381 | } | |
8382 | ||
8383 | /* Print a hint with a suggestion for an architecture name that most closely | |
8384 | resembles what the user passed in STR. */ | |
8385 | ||
8386 | inline static void | |
8387 | aarch64_print_hint_for_arch (const char *str) | |
8388 | { | |
8389 | aarch64_print_hint_for_core_or_arch (str, true); | |
8390 | } | |
8391 | ||
0cfff2a1 KT |
8392 | /* Validate a command-line -mcpu option. Parse the cpu and extensions (if any) |
8393 | specified in STR and throw errors if appropriate. Put the results if | |
361fb3ee KT |
8394 | they are valid in RES and ISA_FLAGS. Return whether the option is |
8395 | valid. */ | |
43e9d192 | 8396 | |
361fb3ee | 8397 | static bool |
0cfff2a1 KT |
8398 | aarch64_validate_mcpu (const char *str, const struct processor **res, |
8399 | unsigned long *isa_flags) | |
8400 | { | |
8401 | enum aarch64_parse_opt_result parse_res | |
8402 | = aarch64_parse_cpu (str, res, isa_flags); | |
8403 | ||
8404 | if (parse_res == AARCH64_PARSE_OK) | |
361fb3ee | 8405 | return true; |
0cfff2a1 KT |
8406 | |
8407 | switch (parse_res) | |
8408 | { | |
8409 | case AARCH64_PARSE_MISSING_ARG: | |
8410 | error ("missing cpu name in -mcpu=%qs", str); | |
8411 | break; | |
8412 | case AARCH64_PARSE_INVALID_ARG: | |
8413 | error ("unknown value %qs for -mcpu", str); | |
01f44038 | 8414 | aarch64_print_hint_for_core (str); |
0cfff2a1 KT |
8415 | break; |
8416 | case AARCH64_PARSE_INVALID_FEATURE: | |
8417 | error ("invalid feature modifier in -mcpu=%qs", str); | |
8418 | break; | |
8419 | default: | |
8420 | gcc_unreachable (); | |
8421 | } | |
361fb3ee KT |
8422 | |
8423 | return false; | |
0cfff2a1 KT |
8424 | } |
8425 | ||
8426 | /* Validate a command-line -march option. Parse the arch and extensions | |
8427 | (if any) specified in STR and throw errors if appropriate. Put the | |
361fb3ee KT |
8428 | results, if they are valid, in RES and ISA_FLAGS. Return whether the |
8429 | option is valid. */ | |
0cfff2a1 | 8430 | |
361fb3ee | 8431 | static bool |
0cfff2a1 | 8432 | aarch64_validate_march (const char *str, const struct processor **res, |
01f44038 | 8433 | unsigned long *isa_flags) |
0cfff2a1 KT |
8434 | { |
8435 | enum aarch64_parse_opt_result parse_res | |
8436 | = aarch64_parse_arch (str, res, isa_flags); | |
8437 | ||
8438 | if (parse_res == AARCH64_PARSE_OK) | |
361fb3ee | 8439 | return true; |
0cfff2a1 KT |
8440 | |
8441 | switch (parse_res) | |
8442 | { | |
8443 | case AARCH64_PARSE_MISSING_ARG: | |
8444 | error ("missing arch name in -march=%qs", str); | |
8445 | break; | |
8446 | case AARCH64_PARSE_INVALID_ARG: | |
8447 | error ("unknown value %qs for -march", str); | |
01f44038 | 8448 | aarch64_print_hint_for_arch (str); |
0cfff2a1 KT |
8449 | break; |
8450 | case AARCH64_PARSE_INVALID_FEATURE: | |
8451 | error ("invalid feature modifier in -march=%qs", str); | |
8452 | break; | |
8453 | default: | |
8454 | gcc_unreachable (); | |
8455 | } | |
361fb3ee KT |
8456 | |
8457 | return false; | |
0cfff2a1 KT |
8458 | } |
8459 | ||
8460 | /* Validate a command-line -mtune option. Parse the cpu | |
8461 | specified in STR and throw errors if appropriate. Put the | |
361fb3ee KT |
8462 | result, if it is valid, in RES. Return whether the option is |
8463 | valid. */ | |
0cfff2a1 | 8464 | |
361fb3ee | 8465 | static bool |
0cfff2a1 KT |
8466 | aarch64_validate_mtune (const char *str, const struct processor **res) |
8467 | { | |
8468 | enum aarch64_parse_opt_result parse_res | |
8469 | = aarch64_parse_tune (str, res); | |
8470 | ||
8471 | if (parse_res == AARCH64_PARSE_OK) | |
361fb3ee | 8472 | return true; |
0cfff2a1 KT |
8473 | |
8474 | switch (parse_res) | |
8475 | { | |
8476 | case AARCH64_PARSE_MISSING_ARG: | |
8477 | error ("missing cpu name in -mtune=%qs", str); | |
8478 | break; | |
8479 | case AARCH64_PARSE_INVALID_ARG: | |
8480 | error ("unknown value %qs for -mtune", str); | |
01f44038 | 8481 | aarch64_print_hint_for_core (str); |
0cfff2a1 KT |
8482 | break; |
8483 | default: | |
8484 | gcc_unreachable (); | |
8485 | } | |
361fb3ee KT |
8486 | return false; |
8487 | } | |
8488 | ||
8489 | /* Return the CPU corresponding to the enum CPU. | |
8490 | If it doesn't specify a cpu, return the default. */ | |
8491 | ||
8492 | static const struct processor * | |
8493 | aarch64_get_tune_cpu (enum aarch64_processor cpu) | |
8494 | { | |
8495 | if (cpu != aarch64_none) | |
8496 | return &all_cores[cpu]; | |
8497 | ||
8498 | /* The & 0x3f is to extract the bottom 6 bits that encode the | |
8499 | default cpu as selected by the --with-cpu GCC configure option | |
8500 | in config.gcc. | |
8501 | ???: The whole TARGET_CPU_DEFAULT and AARCH64_CPU_DEFAULT_FLAGS | |
8502 | flags mechanism should be reworked to make it more sane. */ | |
8503 | return &all_cores[TARGET_CPU_DEFAULT & 0x3f]; | |
8504 | } | |
8505 | ||
8506 | /* Return the architecture corresponding to the enum ARCH. | |
8507 | If it doesn't specify a valid architecture, return the default. */ | |
8508 | ||
8509 | static const struct processor * | |
8510 | aarch64_get_arch (enum aarch64_arch arch) | |
8511 | { | |
8512 | if (arch != aarch64_no_arch) | |
8513 | return &all_architectures[arch]; | |
8514 | ||
8515 | const struct processor *cpu = &all_cores[TARGET_CPU_DEFAULT & 0x3f]; | |
8516 | ||
8517 | return &all_architectures[cpu->arch]; | |
0cfff2a1 KT |
8518 | } |
8519 | ||
8520 | /* Implement TARGET_OPTION_OVERRIDE. This is called once in the beginning | |
8521 | and is used to parse the -m{cpu,tune,arch} strings and setup the initial | |
8522 | tuning structs. In particular it must set selected_tune and | |
8523 | aarch64_isa_flags that define the available ISA features and tuning | |
8524 | decisions. It must also set selected_arch as this will be used to | |
8525 | output the .arch asm tags for each function. */ | |
8526 | ||
8527 | static void | |
8528 | aarch64_override_options (void) | |
8529 | { | |
8530 | unsigned long cpu_isa = 0; | |
8531 | unsigned long arch_isa = 0; | |
8532 | aarch64_isa_flags = 0; | |
8533 | ||
361fb3ee KT |
8534 | bool valid_cpu = true; |
8535 | bool valid_tune = true; | |
8536 | bool valid_arch = true; | |
8537 | ||
0cfff2a1 KT |
8538 | selected_cpu = NULL; |
8539 | selected_arch = NULL; | |
8540 | selected_tune = NULL; | |
8541 | ||
8542 | /* -mcpu=CPU is shorthand for -march=ARCH_FOR_CPU, -mtune=CPU. | |
8543 | If either of -march or -mtune is given, they override their | |
8544 | respective component of -mcpu. */ | |
8545 | if (aarch64_cpu_string) | |
361fb3ee KT |
8546 | valid_cpu = aarch64_validate_mcpu (aarch64_cpu_string, &selected_cpu, |
8547 | &cpu_isa); | |
0cfff2a1 KT |
8548 | |
8549 | if (aarch64_arch_string) | |
361fb3ee KT |
8550 | valid_arch = aarch64_validate_march (aarch64_arch_string, &selected_arch, |
8551 | &arch_isa); | |
0cfff2a1 KT |
8552 | |
8553 | if (aarch64_tune_string) | |
361fb3ee | 8554 | valid_tune = aarch64_validate_mtune (aarch64_tune_string, &selected_tune); |
43e9d192 IB |
8555 | |
8556 | /* If the user did not specify a processor, choose the default | |
8557 | one for them. This will be the CPU set during configuration using | |
a3cd0246 | 8558 | --with-cpu, otherwise it is "generic". */ |
43e9d192 IB |
8559 | if (!selected_cpu) |
8560 | { | |
0cfff2a1 KT |
8561 | if (selected_arch) |
8562 | { | |
8563 | selected_cpu = &all_cores[selected_arch->ident]; | |
8564 | aarch64_isa_flags = arch_isa; | |
361fb3ee | 8565 | explicit_arch = selected_arch->arch; |
0cfff2a1 KT |
8566 | } |
8567 | else | |
8568 | { | |
361fb3ee KT |
8569 | /* Get default configure-time CPU. */ |
8570 | selected_cpu = aarch64_get_tune_cpu (aarch64_none); | |
0cfff2a1 KT |
8571 | aarch64_isa_flags = TARGET_CPU_DEFAULT >> 6; |
8572 | } | |
361fb3ee KT |
8573 | |
8574 | if (selected_tune) | |
8575 | explicit_tune_core = selected_tune->ident; | |
0cfff2a1 KT |
8576 | } |
8577 | /* If both -mcpu and -march are specified check that they are architecturally | |
8578 | compatible, warn if they're not and prefer the -march ISA flags. */ | |
8579 | else if (selected_arch) | |
8580 | { | |
8581 | if (selected_arch->arch != selected_cpu->arch) | |
8582 | { | |
8583 | warning (0, "switch -mcpu=%s conflicts with -march=%s switch", | |
8584 | all_architectures[selected_cpu->arch].name, | |
8585 | selected_arch->name); | |
8586 | } | |
8587 | aarch64_isa_flags = arch_isa; | |
361fb3ee KT |
8588 | explicit_arch = selected_arch->arch; |
8589 | explicit_tune_core = selected_tune ? selected_tune->ident | |
8590 | : selected_cpu->ident; | |
0cfff2a1 KT |
8591 | } |
8592 | else | |
8593 | { | |
8594 | /* -mcpu but no -march. */ | |
8595 | aarch64_isa_flags = cpu_isa; | |
361fb3ee KT |
8596 | explicit_tune_core = selected_tune ? selected_tune->ident |
8597 | : selected_cpu->ident; | |
8598 | gcc_assert (selected_cpu); | |
8599 | selected_arch = &all_architectures[selected_cpu->arch]; | |
8600 | explicit_arch = selected_arch->arch; | |
43e9d192 IB |
8601 | } |
8602 | ||
0cfff2a1 KT |
8603 | /* Set the arch as well as we will need it when outputing |
8604 | the .arch directive in assembly. */ | |
8605 | if (!selected_arch) | |
8606 | { | |
8607 | gcc_assert (selected_cpu); | |
8608 | selected_arch = &all_architectures[selected_cpu->arch]; | |
8609 | } | |
43e9d192 | 8610 | |
43e9d192 | 8611 | if (!selected_tune) |
3edaf26d | 8612 | selected_tune = selected_cpu; |
43e9d192 | 8613 | |
0cfff2a1 KT |
8614 | #ifndef HAVE_AS_MABI_OPTION |
8615 | /* The compiler may have been configured with 2.23.* binutils, which does | |
8616 | not have support for ILP32. */ | |
8617 | if (TARGET_ILP32) | |
8618 | error ("Assembler does not support -mabi=ilp32"); | |
8619 | #endif | |
43e9d192 | 8620 | |
361fb3ee KT |
8621 | /* Make sure we properly set up the explicit options. */ |
8622 | if ((aarch64_cpu_string && valid_cpu) | |
8623 | || (aarch64_tune_string && valid_tune)) | |
8624 | gcc_assert (explicit_tune_core != aarch64_none); | |
8625 | ||
8626 | if ((aarch64_cpu_string && valid_cpu) | |
8627 | || (aarch64_arch_string && valid_arch)) | |
8628 | gcc_assert (explicit_arch != aarch64_no_arch); | |
8629 | ||
0cfff2a1 KT |
8630 | aarch64_override_options_internal (&global_options); |
8631 | ||
8632 | /* Save these options as the default ones in case we push and pop them later | |
8633 | while processing functions with potential target attributes. */ | |
8634 | target_option_default_node = target_option_current_node | |
8635 | = build_target_option_node (&global_options); | |
43e9d192 IB |
8636 | } |
8637 | ||
8638 | /* Implement targetm.override_options_after_change. */ | |
8639 | ||
8640 | static void | |
8641 | aarch64_override_options_after_change (void) | |
8642 | { | |
0cfff2a1 | 8643 | aarch64_override_options_after_change_1 (&global_options); |
43e9d192 IB |
8644 | } |
8645 | ||
8646 | static struct machine_function * | |
8647 | aarch64_init_machine_status (void) | |
8648 | { | |
8649 | struct machine_function *machine; | |
766090c2 | 8650 | machine = ggc_cleared_alloc<machine_function> (); |
43e9d192 IB |
8651 | return machine; |
8652 | } | |
8653 | ||
8654 | void | |
8655 | aarch64_init_expanders (void) | |
8656 | { | |
8657 | init_machine_status = aarch64_init_machine_status; | |
8658 | } | |
8659 | ||
8660 | /* A checking mechanism for the implementation of the various code models. */ | |
8661 | static void | |
0cfff2a1 | 8662 | initialize_aarch64_code_model (struct gcc_options *opts) |
43e9d192 | 8663 | { |
0cfff2a1 | 8664 | if (opts->x_flag_pic) |
43e9d192 | 8665 | { |
0cfff2a1 | 8666 | switch (opts->x_aarch64_cmodel_var) |
43e9d192 IB |
8667 | { |
8668 | case AARCH64_CMODEL_TINY: | |
8669 | aarch64_cmodel = AARCH64_CMODEL_TINY_PIC; | |
8670 | break; | |
8671 | case AARCH64_CMODEL_SMALL: | |
34ecdb0f | 8672 | #ifdef HAVE_AS_SMALL_PIC_RELOCS |
1b1e81f8 JW |
8673 | aarch64_cmodel = (flag_pic == 2 |
8674 | ? AARCH64_CMODEL_SMALL_PIC | |
8675 | : AARCH64_CMODEL_SMALL_SPIC); | |
34ecdb0f JW |
8676 | #else |
8677 | aarch64_cmodel = AARCH64_CMODEL_SMALL_PIC; | |
8678 | #endif | |
43e9d192 IB |
8679 | break; |
8680 | case AARCH64_CMODEL_LARGE: | |
8681 | sorry ("code model %qs with -f%s", "large", | |
0cfff2a1 | 8682 | opts->x_flag_pic > 1 ? "PIC" : "pic"); |
1c652781 | 8683 | break; |
43e9d192 IB |
8684 | default: |
8685 | gcc_unreachable (); | |
8686 | } | |
8687 | } | |
8688 | else | |
0cfff2a1 | 8689 | aarch64_cmodel = opts->x_aarch64_cmodel_var; |
43e9d192 IB |
8690 | } |
8691 | ||
361fb3ee KT |
8692 | /* Implement TARGET_OPTION_SAVE. */ |
8693 | ||
8694 | static void | |
8695 | aarch64_option_save (struct cl_target_option *ptr, struct gcc_options *opts) | |
8696 | { | |
8697 | ptr->x_aarch64_override_tune_string = opts->x_aarch64_override_tune_string; | |
8698 | } | |
8699 | ||
8700 | /* Implements TARGET_OPTION_RESTORE. Restore the backend codegen decisions | |
8701 | using the information saved in PTR. */ | |
8702 | ||
8703 | static void | |
8704 | aarch64_option_restore (struct gcc_options *opts, struct cl_target_option *ptr) | |
8705 | { | |
8706 | opts->x_explicit_tune_core = ptr->x_explicit_tune_core; | |
8707 | selected_tune = aarch64_get_tune_cpu (ptr->x_explicit_tune_core); | |
8708 | opts->x_explicit_arch = ptr->x_explicit_arch; | |
8709 | selected_arch = aarch64_get_arch (ptr->x_explicit_arch); | |
8710 | opts->x_aarch64_override_tune_string = ptr->x_aarch64_override_tune_string; | |
8711 | ||
8712 | aarch64_override_options_internal (opts); | |
8713 | } | |
8714 | ||
8715 | /* Implement TARGET_OPTION_PRINT. */ | |
8716 | ||
8717 | static void | |
8718 | aarch64_option_print (FILE *file, int indent, struct cl_target_option *ptr) | |
8719 | { | |
8720 | const struct processor *cpu | |
8721 | = aarch64_get_tune_cpu (ptr->x_explicit_tune_core); | |
8722 | unsigned long isa_flags = ptr->x_aarch64_isa_flags; | |
8723 | const struct processor *arch = aarch64_get_arch (ptr->x_explicit_arch); | |
054b4005 | 8724 | std::string extension |
04a99ebe | 8725 | = aarch64_get_extension_string_for_isa_flags (isa_flags, arch->flags); |
361fb3ee KT |
8726 | |
8727 | fprintf (file, "%*sselected tune = %s\n", indent, "", cpu->name); | |
054b4005 JG |
8728 | fprintf (file, "%*sselected arch = %s%s\n", indent, "", |
8729 | arch->name, extension.c_str ()); | |
361fb3ee KT |
8730 | } |
8731 | ||
d78006d9 KT |
8732 | static GTY(()) tree aarch64_previous_fndecl; |
8733 | ||
e4ea20c8 KT |
8734 | void |
8735 | aarch64_reset_previous_fndecl (void) | |
8736 | { | |
8737 | aarch64_previous_fndecl = NULL; | |
8738 | } | |
8739 | ||
acfc1ac1 KT |
8740 | /* Restore or save the TREE_TARGET_GLOBALS from or to NEW_TREE. |
8741 | Used by aarch64_set_current_function and aarch64_pragma_target_parse to | |
8742 | make sure optab availability predicates are recomputed when necessary. */ | |
8743 | ||
8744 | void | |
8745 | aarch64_save_restore_target_globals (tree new_tree) | |
8746 | { | |
8747 | if (TREE_TARGET_GLOBALS (new_tree)) | |
8748 | restore_target_globals (TREE_TARGET_GLOBALS (new_tree)); | |
8749 | else if (new_tree == target_option_default_node) | |
8750 | restore_target_globals (&default_target_globals); | |
8751 | else | |
8752 | TREE_TARGET_GLOBALS (new_tree) = save_target_globals_default_opts (); | |
8753 | } | |
8754 | ||
d78006d9 KT |
8755 | /* Implement TARGET_SET_CURRENT_FUNCTION. Unpack the codegen decisions |
8756 | like tuning and ISA features from the DECL_FUNCTION_SPECIFIC_TARGET | |
8757 | of the function, if such exists. This function may be called multiple | |
8758 | times on a single function so use aarch64_previous_fndecl to avoid | |
8759 | setting up identical state. */ | |
8760 | ||
8761 | static void | |
8762 | aarch64_set_current_function (tree fndecl) | |
8763 | { | |
acfc1ac1 KT |
8764 | if (!fndecl || fndecl == aarch64_previous_fndecl) |
8765 | return; | |
8766 | ||
d78006d9 KT |
8767 | tree old_tree = (aarch64_previous_fndecl |
8768 | ? DECL_FUNCTION_SPECIFIC_TARGET (aarch64_previous_fndecl) | |
8769 | : NULL_TREE); | |
8770 | ||
acfc1ac1 | 8771 | tree new_tree = DECL_FUNCTION_SPECIFIC_TARGET (fndecl); |
d78006d9 | 8772 | |
acfc1ac1 KT |
8773 | /* If current function has no attributes but the previous one did, |
8774 | use the default node. */ | |
8775 | if (!new_tree && old_tree) | |
8776 | new_tree = target_option_default_node; | |
d78006d9 | 8777 | |
acfc1ac1 KT |
8778 | /* If nothing to do, return. #pragma GCC reset or #pragma GCC pop to |
8779 | the default have been handled by aarch64_save_restore_target_globals from | |
8780 | aarch64_pragma_target_parse. */ | |
8781 | if (old_tree == new_tree) | |
8782 | return; | |
d78006d9 | 8783 | |
acfc1ac1 | 8784 | aarch64_previous_fndecl = fndecl; |
6e17a23b | 8785 | |
acfc1ac1 KT |
8786 | /* First set the target options. */ |
8787 | cl_target_option_restore (&global_options, TREE_TARGET_OPTION (new_tree)); | |
6e17a23b | 8788 | |
acfc1ac1 | 8789 | aarch64_save_restore_target_globals (new_tree); |
d78006d9 | 8790 | } |
361fb3ee | 8791 | |
5a2c8331 KT |
8792 | /* Enum describing the various ways we can handle attributes. |
8793 | In many cases we can reuse the generic option handling machinery. */ | |
8794 | ||
8795 | enum aarch64_attr_opt_type | |
8796 | { | |
8797 | aarch64_attr_mask, /* Attribute should set a bit in target_flags. */ | |
8798 | aarch64_attr_bool, /* Attribute sets or unsets a boolean variable. */ | |
8799 | aarch64_attr_enum, /* Attribute sets an enum variable. */ | |
8800 | aarch64_attr_custom /* Attribute requires a custom handling function. */ | |
8801 | }; | |
8802 | ||
8803 | /* All the information needed to handle a target attribute. | |
8804 | NAME is the name of the attribute. | |
9c582551 | 8805 | ATTR_TYPE specifies the type of behavior of the attribute as described |
5a2c8331 KT |
8806 | in the definition of enum aarch64_attr_opt_type. |
8807 | ALLOW_NEG is true if the attribute supports a "no-" form. | |
8808 | HANDLER is the function that takes the attribute string and whether | |
8809 | it is a pragma or attribute and handles the option. It is needed only | |
8810 | when the ATTR_TYPE is aarch64_attr_custom. | |
8811 | OPT_NUM is the enum specifying the option that the attribute modifies. | |
9c582551 | 8812 | This is needed for attributes that mirror the behavior of a command-line |
5a2c8331 KT |
8813 | option, that is it has ATTR_TYPE aarch64_attr_mask, aarch64_attr_bool or |
8814 | aarch64_attr_enum. */ | |
8815 | ||
8816 | struct aarch64_attribute_info | |
8817 | { | |
8818 | const char *name; | |
8819 | enum aarch64_attr_opt_type attr_type; | |
8820 | bool allow_neg; | |
8821 | bool (*handler) (const char *, const char *); | |
8822 | enum opt_code opt_num; | |
8823 | }; | |
8824 | ||
8825 | /* Handle the ARCH_STR argument to the arch= target attribute. | |
8826 | PRAGMA_OR_ATTR is used in potential error messages. */ | |
8827 | ||
8828 | static bool | |
8829 | aarch64_handle_attr_arch (const char *str, const char *pragma_or_attr) | |
8830 | { | |
8831 | const struct processor *tmp_arch = NULL; | |
8832 | enum aarch64_parse_opt_result parse_res | |
8833 | = aarch64_parse_arch (str, &tmp_arch, &aarch64_isa_flags); | |
8834 | ||
8835 | if (parse_res == AARCH64_PARSE_OK) | |
8836 | { | |
8837 | gcc_assert (tmp_arch); | |
8838 | selected_arch = tmp_arch; | |
8839 | explicit_arch = selected_arch->arch; | |
8840 | return true; | |
8841 | } | |
8842 | ||
8843 | switch (parse_res) | |
8844 | { | |
8845 | case AARCH64_PARSE_MISSING_ARG: | |
8846 | error ("missing architecture name in 'arch' target %s", pragma_or_attr); | |
8847 | break; | |
8848 | case AARCH64_PARSE_INVALID_ARG: | |
8849 | error ("unknown value %qs for 'arch' target %s", str, pragma_or_attr); | |
01f44038 | 8850 | aarch64_print_hint_for_arch (str); |
5a2c8331 KT |
8851 | break; |
8852 | case AARCH64_PARSE_INVALID_FEATURE: | |
8853 | error ("invalid feature modifier %qs for 'arch' target %s", | |
8854 | str, pragma_or_attr); | |
8855 | break; | |
8856 | default: | |
8857 | gcc_unreachable (); | |
8858 | } | |
8859 | ||
8860 | return false; | |
8861 | } | |
8862 | ||
8863 | /* Handle the argument CPU_STR to the cpu= target attribute. | |
8864 | PRAGMA_OR_ATTR is used in potential error messages. */ | |
8865 | ||
8866 | static bool | |
8867 | aarch64_handle_attr_cpu (const char *str, const char *pragma_or_attr) | |
8868 | { | |
8869 | const struct processor *tmp_cpu = NULL; | |
8870 | enum aarch64_parse_opt_result parse_res | |
8871 | = aarch64_parse_cpu (str, &tmp_cpu, &aarch64_isa_flags); | |
8872 | ||
8873 | if (parse_res == AARCH64_PARSE_OK) | |
8874 | { | |
8875 | gcc_assert (tmp_cpu); | |
8876 | selected_tune = tmp_cpu; | |
8877 | explicit_tune_core = selected_tune->ident; | |
8878 | ||
8879 | selected_arch = &all_architectures[tmp_cpu->arch]; | |
8880 | explicit_arch = selected_arch->arch; | |
8881 | return true; | |
8882 | } | |
8883 | ||
8884 | switch (parse_res) | |
8885 | { | |
8886 | case AARCH64_PARSE_MISSING_ARG: | |
8887 | error ("missing cpu name in 'cpu' target %s", pragma_or_attr); | |
8888 | break; | |
8889 | case AARCH64_PARSE_INVALID_ARG: | |
8890 | error ("unknown value %qs for 'cpu' target %s", str, pragma_or_attr); | |
01f44038 | 8891 | aarch64_print_hint_for_core (str); |
5a2c8331 KT |
8892 | break; |
8893 | case AARCH64_PARSE_INVALID_FEATURE: | |
8894 | error ("invalid feature modifier %qs for 'cpu' target %s", | |
8895 | str, pragma_or_attr); | |
8896 | break; | |
8897 | default: | |
8898 | gcc_unreachable (); | |
8899 | } | |
8900 | ||
8901 | return false; | |
8902 | } | |
8903 | ||
8904 | /* Handle the argument STR to the tune= target attribute. | |
8905 | PRAGMA_OR_ATTR is used in potential error messages. */ | |
8906 | ||
8907 | static bool | |
8908 | aarch64_handle_attr_tune (const char *str, const char *pragma_or_attr) | |
8909 | { | |
8910 | const struct processor *tmp_tune = NULL; | |
8911 | enum aarch64_parse_opt_result parse_res | |
8912 | = aarch64_parse_tune (str, &tmp_tune); | |
8913 | ||
8914 | if (parse_res == AARCH64_PARSE_OK) | |
8915 | { | |
8916 | gcc_assert (tmp_tune); | |
8917 | selected_tune = tmp_tune; | |
8918 | explicit_tune_core = selected_tune->ident; | |
8919 | return true; | |
8920 | } | |
8921 | ||
8922 | switch (parse_res) | |
8923 | { | |
8924 | case AARCH64_PARSE_INVALID_ARG: | |
8925 | error ("unknown value %qs for 'tune' target %s", str, pragma_or_attr); | |
01f44038 | 8926 | aarch64_print_hint_for_core (str); |
5a2c8331 KT |
8927 | break; |
8928 | default: | |
8929 | gcc_unreachable (); | |
8930 | } | |
8931 | ||
8932 | return false; | |
8933 | } | |
8934 | ||
8935 | /* Parse an architecture extensions target attribute string specified in STR. | |
8936 | For example "+fp+nosimd". Show any errors if needed. Return TRUE | |
8937 | if successful. Update aarch64_isa_flags to reflect the ISA features | |
8938 | modified. | |
8939 | PRAGMA_OR_ATTR is used in potential error messages. */ | |
8940 | ||
8941 | static bool | |
8942 | aarch64_handle_attr_isa_flags (char *str, const char *pragma_or_attr) | |
8943 | { | |
8944 | enum aarch64_parse_opt_result parse_res; | |
8945 | unsigned long isa_flags = aarch64_isa_flags; | |
8946 | ||
e4ea20c8 KT |
8947 | /* We allow "+nothing" in the beginning to clear out all architectural |
8948 | features if the user wants to handpick specific features. */ | |
8949 | if (strncmp ("+nothing", str, 8) == 0) | |
8950 | { | |
8951 | isa_flags = 0; | |
8952 | str += 8; | |
8953 | } | |
8954 | ||
5a2c8331 KT |
8955 | parse_res = aarch64_parse_extension (str, &isa_flags); |
8956 | ||
8957 | if (parse_res == AARCH64_PARSE_OK) | |
8958 | { | |
8959 | aarch64_isa_flags = isa_flags; | |
8960 | return true; | |
8961 | } | |
8962 | ||
8963 | switch (parse_res) | |
8964 | { | |
8965 | case AARCH64_PARSE_MISSING_ARG: | |
8966 | error ("missing feature modifier in target %s %qs", | |
8967 | pragma_or_attr, str); | |
8968 | break; | |
8969 | ||
8970 | case AARCH64_PARSE_INVALID_FEATURE: | |
8971 | error ("invalid feature modifier in target %s %qs", | |
8972 | pragma_or_attr, str); | |
8973 | break; | |
8974 | ||
8975 | default: | |
8976 | gcc_unreachable (); | |
8977 | } | |
8978 | ||
8979 | return false; | |
8980 | } | |
8981 | ||
8982 | /* The target attributes that we support. On top of these we also support just | |
8983 | ISA extensions, like __attribute__ ((target ("+crc"))), but that case is | |
8984 | handled explicitly in aarch64_process_one_target_attr. */ | |
8985 | ||
8986 | static const struct aarch64_attribute_info aarch64_attributes[] = | |
8987 | { | |
8988 | { "general-regs-only", aarch64_attr_mask, false, NULL, | |
8989 | OPT_mgeneral_regs_only }, | |
8990 | { "fix-cortex-a53-835769", aarch64_attr_bool, true, NULL, | |
8991 | OPT_mfix_cortex_a53_835769 }, | |
48bb1a55 CL |
8992 | { "fix-cortex-a53-843419", aarch64_attr_bool, true, NULL, |
8993 | OPT_mfix_cortex_a53_843419 }, | |
5a2c8331 KT |
8994 | { "cmodel", aarch64_attr_enum, false, NULL, OPT_mcmodel_ }, |
8995 | { "strict-align", aarch64_attr_mask, false, NULL, OPT_mstrict_align }, | |
8996 | { "omit-leaf-frame-pointer", aarch64_attr_bool, true, NULL, | |
8997 | OPT_momit_leaf_frame_pointer }, | |
8998 | { "tls-dialect", aarch64_attr_enum, false, NULL, OPT_mtls_dialect_ }, | |
8999 | { "arch", aarch64_attr_custom, false, aarch64_handle_attr_arch, | |
9000 | OPT_march_ }, | |
9001 | { "cpu", aarch64_attr_custom, false, aarch64_handle_attr_cpu, OPT_mcpu_ }, | |
9002 | { "tune", aarch64_attr_custom, false, aarch64_handle_attr_tune, | |
9003 | OPT_mtune_ }, | |
9004 | { NULL, aarch64_attr_custom, false, NULL, OPT____ } | |
9005 | }; | |
9006 | ||
9007 | /* Parse ARG_STR which contains the definition of one target attribute. | |
9008 | Show appropriate errors if any or return true if the attribute is valid. | |
9009 | PRAGMA_OR_ATTR holds the string to use in error messages about whether | |
9010 | we're processing a target attribute or pragma. */ | |
9011 | ||
9012 | static bool | |
9013 | aarch64_process_one_target_attr (char *arg_str, const char* pragma_or_attr) | |
9014 | { | |
9015 | bool invert = false; | |
9016 | ||
9017 | size_t len = strlen (arg_str); | |
9018 | ||
9019 | if (len == 0) | |
9020 | { | |
9021 | error ("malformed target %s", pragma_or_attr); | |
9022 | return false; | |
9023 | } | |
9024 | ||
9025 | char *str_to_check = (char *) alloca (len + 1); | |
9026 | strcpy (str_to_check, arg_str); | |
9027 | ||
9028 | /* Skip leading whitespace. */ | |
9029 | while (*str_to_check == ' ' || *str_to_check == '\t') | |
9030 | str_to_check++; | |
9031 | ||
9032 | /* We have something like __attribute__ ((target ("+fp+nosimd"))). | |
9033 | It is easier to detect and handle it explicitly here rather than going | |
9034 | through the machinery for the rest of the target attributes in this | |
9035 | function. */ | |
9036 | if (*str_to_check == '+') | |
9037 | return aarch64_handle_attr_isa_flags (str_to_check, pragma_or_attr); | |
9038 | ||
9039 | if (len > 3 && strncmp (str_to_check, "no-", 3) == 0) | |
9040 | { | |
9041 | invert = true; | |
9042 | str_to_check += 3; | |
9043 | } | |
9044 | char *arg = strchr (str_to_check, '='); | |
9045 | ||
9046 | /* If we found opt=foo then terminate STR_TO_CHECK at the '=' | |
9047 | and point ARG to "foo". */ | |
9048 | if (arg) | |
9049 | { | |
9050 | *arg = '\0'; | |
9051 | arg++; | |
9052 | } | |
9053 | const struct aarch64_attribute_info *p_attr; | |
16d12992 | 9054 | bool found = false; |
5a2c8331 KT |
9055 | for (p_attr = aarch64_attributes; p_attr->name; p_attr++) |
9056 | { | |
9057 | /* If the names don't match up, or the user has given an argument | |
9058 | to an attribute that doesn't accept one, or didn't give an argument | |
9059 | to an attribute that expects one, fail to match. */ | |
9060 | if (strcmp (str_to_check, p_attr->name) != 0) | |
9061 | continue; | |
9062 | ||
16d12992 | 9063 | found = true; |
5a2c8331 KT |
9064 | bool attr_need_arg_p = p_attr->attr_type == aarch64_attr_custom |
9065 | || p_attr->attr_type == aarch64_attr_enum; | |
9066 | ||
9067 | if (attr_need_arg_p ^ (arg != NULL)) | |
9068 | { | |
9069 | error ("target %s %qs does not accept an argument", | |
9070 | pragma_or_attr, str_to_check); | |
9071 | return false; | |
9072 | } | |
9073 | ||
9074 | /* If the name matches but the attribute does not allow "no-" versions | |
9075 | then we can't match. */ | |
9076 | if (invert && !p_attr->allow_neg) | |
9077 | { | |
9078 | error ("target %s %qs does not allow a negated form", | |
9079 | pragma_or_attr, str_to_check); | |
9080 | return false; | |
9081 | } | |
9082 | ||
9083 | switch (p_attr->attr_type) | |
9084 | { | |
9085 | /* Has a custom handler registered. | |
9086 | For example, cpu=, arch=, tune=. */ | |
9087 | case aarch64_attr_custom: | |
9088 | gcc_assert (p_attr->handler); | |
9089 | if (!p_attr->handler (arg, pragma_or_attr)) | |
9090 | return false; | |
9091 | break; | |
9092 | ||
9093 | /* Either set or unset a boolean option. */ | |
9094 | case aarch64_attr_bool: | |
9095 | { | |
9096 | struct cl_decoded_option decoded; | |
9097 | ||
9098 | generate_option (p_attr->opt_num, NULL, !invert, | |
9099 | CL_TARGET, &decoded); | |
9100 | aarch64_handle_option (&global_options, &global_options_set, | |
9101 | &decoded, input_location); | |
9102 | break; | |
9103 | } | |
9104 | /* Set or unset a bit in the target_flags. aarch64_handle_option | |
9105 | should know what mask to apply given the option number. */ | |
9106 | case aarch64_attr_mask: | |
9107 | { | |
9108 | struct cl_decoded_option decoded; | |
9109 | /* We only need to specify the option number. | |
9110 | aarch64_handle_option will know which mask to apply. */ | |
9111 | decoded.opt_index = p_attr->opt_num; | |
9112 | decoded.value = !invert; | |
9113 | aarch64_handle_option (&global_options, &global_options_set, | |
9114 | &decoded, input_location); | |
9115 | break; | |
9116 | } | |
9117 | /* Use the option setting machinery to set an option to an enum. */ | |
9118 | case aarch64_attr_enum: | |
9119 | { | |
9120 | gcc_assert (arg); | |
9121 | bool valid; | |
9122 | int value; | |
9123 | valid = opt_enum_arg_to_value (p_attr->opt_num, arg, | |
9124 | &value, CL_TARGET); | |
9125 | if (valid) | |
9126 | { | |
9127 | set_option (&global_options, NULL, p_attr->opt_num, value, | |
9128 | NULL, DK_UNSPECIFIED, input_location, | |
9129 | global_dc); | |
9130 | } | |
9131 | else | |
9132 | { | |
9133 | error ("target %s %s=%s is not valid", | |
9134 | pragma_or_attr, str_to_check, arg); | |
9135 | } | |
9136 | break; | |
9137 | } | |
9138 | default: | |
9139 | gcc_unreachable (); | |
9140 | } | |
9141 | } | |
9142 | ||
16d12992 KT |
9143 | /* If we reached here we either have found an attribute and validated |
9144 | it or didn't match any. If we matched an attribute but its arguments | |
9145 | were malformed we will have returned false already. */ | |
9146 | return found; | |
5a2c8331 KT |
9147 | } |
9148 | ||
9149 | /* Count how many times the character C appears in | |
9150 | NULL-terminated string STR. */ | |
9151 | ||
9152 | static unsigned int | |
9153 | num_occurences_in_str (char c, char *str) | |
9154 | { | |
9155 | unsigned int res = 0; | |
9156 | while (*str != '\0') | |
9157 | { | |
9158 | if (*str == c) | |
9159 | res++; | |
9160 | ||
9161 | str++; | |
9162 | } | |
9163 | ||
9164 | return res; | |
9165 | } | |
9166 | ||
9167 | /* Parse the tree in ARGS that contains the target attribute information | |
9168 | and update the global target options space. PRAGMA_OR_ATTR is a string | |
9169 | to be used in error messages, specifying whether this is processing | |
9170 | a target attribute or a target pragma. */ | |
9171 | ||
9172 | bool | |
9173 | aarch64_process_target_attr (tree args, const char* pragma_or_attr) | |
9174 | { | |
9175 | if (TREE_CODE (args) == TREE_LIST) | |
9176 | { | |
9177 | do | |
9178 | { | |
9179 | tree head = TREE_VALUE (args); | |
9180 | if (head) | |
9181 | { | |
9182 | if (!aarch64_process_target_attr (head, pragma_or_attr)) | |
9183 | return false; | |
9184 | } | |
9185 | args = TREE_CHAIN (args); | |
9186 | } while (args); | |
9187 | ||
9188 | return true; | |
9189 | } | |
9190 | /* We expect to find a string to parse. */ | |
9191 | gcc_assert (TREE_CODE (args) == STRING_CST); | |
9192 | ||
9193 | size_t len = strlen (TREE_STRING_POINTER (args)); | |
9194 | char *str_to_check = (char *) alloca (len + 1); | |
9195 | strcpy (str_to_check, TREE_STRING_POINTER (args)); | |
9196 | ||
9197 | if (len == 0) | |
9198 | { | |
9199 | error ("malformed target %s value", pragma_or_attr); | |
9200 | return false; | |
9201 | } | |
9202 | ||
9203 | /* Used to catch empty spaces between commas i.e. | |
9204 | attribute ((target ("attr1,,attr2"))). */ | |
9205 | unsigned int num_commas = num_occurences_in_str (',', str_to_check); | |
9206 | ||
9207 | /* Handle multiple target attributes separated by ','. */ | |
9208 | char *token = strtok (str_to_check, ","); | |
9209 | ||
9210 | unsigned int num_attrs = 0; | |
9211 | while (token) | |
9212 | { | |
9213 | num_attrs++; | |
9214 | if (!aarch64_process_one_target_attr (token, pragma_or_attr)) | |
9215 | { | |
9216 | error ("target %s %qs is invalid", pragma_or_attr, token); | |
9217 | return false; | |
9218 | } | |
9219 | ||
9220 | token = strtok (NULL, ","); | |
9221 | } | |
9222 | ||
9223 | if (num_attrs != num_commas + 1) | |
9224 | { | |
9225 | error ("malformed target %s list %qs", | |
9226 | pragma_or_attr, TREE_STRING_POINTER (args)); | |
9227 | return false; | |
9228 | } | |
9229 | ||
9230 | return true; | |
9231 | } | |
9232 | ||
9233 | /* Implement TARGET_OPTION_VALID_ATTRIBUTE_P. This is used to | |
9234 | process attribute ((target ("..."))). */ | |
9235 | ||
9236 | static bool | |
9237 | aarch64_option_valid_attribute_p (tree fndecl, tree, tree args, int) | |
9238 | { | |
9239 | struct cl_target_option cur_target; | |
9240 | bool ret; | |
9241 | tree old_optimize; | |
9242 | tree new_target, new_optimize; | |
9243 | tree existing_target = DECL_FUNCTION_SPECIFIC_TARGET (fndecl); | |
91d0e8de KT |
9244 | |
9245 | /* If what we're processing is the current pragma string then the | |
9246 | target option node is already stored in target_option_current_node | |
9247 | by aarch64_pragma_target_parse in aarch64-c.c. Use that to avoid | |
9248 | having to re-parse the string. This is especially useful to keep | |
9249 | arm_neon.h compile times down since that header contains a lot | |
9250 | of intrinsics enclosed in pragmas. */ | |
9251 | if (!existing_target && args == current_target_pragma) | |
9252 | { | |
9253 | DECL_FUNCTION_SPECIFIC_TARGET (fndecl) = target_option_current_node; | |
9254 | return true; | |
9255 | } | |
5a2c8331 KT |
9256 | tree func_optimize = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl); |
9257 | ||
9258 | old_optimize = build_optimization_node (&global_options); | |
9259 | func_optimize = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl); | |
9260 | ||
9261 | /* If the function changed the optimization levels as well as setting | |
9262 | target options, start with the optimizations specified. */ | |
9263 | if (func_optimize && func_optimize != old_optimize) | |
9264 | cl_optimization_restore (&global_options, | |
9265 | TREE_OPTIMIZATION (func_optimize)); | |
9266 | ||
9267 | /* Save the current target options to restore at the end. */ | |
9268 | cl_target_option_save (&cur_target, &global_options); | |
9269 | ||
9270 | /* If fndecl already has some target attributes applied to it, unpack | |
9271 | them so that we add this attribute on top of them, rather than | |
9272 | overwriting them. */ | |
9273 | if (existing_target) | |
9274 | { | |
9275 | struct cl_target_option *existing_options | |
9276 | = TREE_TARGET_OPTION (existing_target); | |
9277 | ||
9278 | if (existing_options) | |
9279 | cl_target_option_restore (&global_options, existing_options); | |
9280 | } | |
9281 | else | |
9282 | cl_target_option_restore (&global_options, | |
9283 | TREE_TARGET_OPTION (target_option_current_node)); | |
9284 | ||
9285 | ||
9286 | ret = aarch64_process_target_attr (args, "attribute"); | |
9287 | ||
9288 | /* Set up any additional state. */ | |
9289 | if (ret) | |
9290 | { | |
9291 | aarch64_override_options_internal (&global_options); | |
e95a988a KT |
9292 | /* Initialize SIMD builtins if we haven't already. |
9293 | Set current_target_pragma to NULL for the duration so that | |
9294 | the builtin initialization code doesn't try to tag the functions | |
9295 | being built with the attributes specified by any current pragma, thus | |
9296 | going into an infinite recursion. */ | |
9297 | if (TARGET_SIMD) | |
9298 | { | |
9299 | tree saved_current_target_pragma = current_target_pragma; | |
9300 | current_target_pragma = NULL; | |
9301 | aarch64_init_simd_builtins (); | |
9302 | current_target_pragma = saved_current_target_pragma; | |
9303 | } | |
5a2c8331 KT |
9304 | new_target = build_target_option_node (&global_options); |
9305 | } | |
9306 | else | |
9307 | new_target = NULL; | |
9308 | ||
9309 | new_optimize = build_optimization_node (&global_options); | |
9310 | ||
9311 | if (fndecl && ret) | |
9312 | { | |
9313 | DECL_FUNCTION_SPECIFIC_TARGET (fndecl) = new_target; | |
9314 | ||
9315 | if (old_optimize != new_optimize) | |
9316 | DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl) = new_optimize; | |
9317 | } | |
9318 | ||
9319 | cl_target_option_restore (&global_options, &cur_target); | |
9320 | ||
9321 | if (old_optimize != new_optimize) | |
9322 | cl_optimization_restore (&global_options, | |
9323 | TREE_OPTIMIZATION (old_optimize)); | |
9324 | return ret; | |
9325 | } | |
9326 | ||
1fd8d40c KT |
9327 | /* Helper for aarch64_can_inline_p. In the case where CALLER and CALLEE are |
9328 | tri-bool options (yes, no, don't care) and the default value is | |
9329 | DEF, determine whether to reject inlining. */ | |
9330 | ||
9331 | static bool | |
9332 | aarch64_tribools_ok_for_inlining_p (int caller, int callee, | |
9333 | int dont_care, int def) | |
9334 | { | |
9335 | /* If the callee doesn't care, always allow inlining. */ | |
9336 | if (callee == dont_care) | |
9337 | return true; | |
9338 | ||
9339 | /* If the caller doesn't care, always allow inlining. */ | |
9340 | if (caller == dont_care) | |
9341 | return true; | |
9342 | ||
9343 | /* Otherwise, allow inlining if either the callee and caller values | |
9344 | agree, or if the callee is using the default value. */ | |
9345 | return (callee == caller || callee == def); | |
9346 | } | |
9347 | ||
9348 | /* Implement TARGET_CAN_INLINE_P. Decide whether it is valid | |
9349 | to inline CALLEE into CALLER based on target-specific info. | |
9350 | Make sure that the caller and callee have compatible architectural | |
9351 | features. Then go through the other possible target attributes | |
9352 | and see if they can block inlining. Try not to reject always_inline | |
9353 | callees unless they are incompatible architecturally. */ | |
9354 | ||
9355 | static bool | |
9356 | aarch64_can_inline_p (tree caller, tree callee) | |
9357 | { | |
9358 | tree caller_tree = DECL_FUNCTION_SPECIFIC_TARGET (caller); | |
9359 | tree callee_tree = DECL_FUNCTION_SPECIFIC_TARGET (callee); | |
9360 | ||
9361 | /* If callee has no option attributes, then it is ok to inline. */ | |
9362 | if (!callee_tree) | |
9363 | return true; | |
9364 | ||
9365 | struct cl_target_option *caller_opts | |
9366 | = TREE_TARGET_OPTION (caller_tree ? caller_tree | |
9367 | : target_option_default_node); | |
9368 | ||
9369 | struct cl_target_option *callee_opts = TREE_TARGET_OPTION (callee_tree); | |
9370 | ||
9371 | ||
9372 | /* Callee's ISA flags should be a subset of the caller's. */ | |
9373 | if ((caller_opts->x_aarch64_isa_flags & callee_opts->x_aarch64_isa_flags) | |
9374 | != callee_opts->x_aarch64_isa_flags) | |
9375 | return false; | |
9376 | ||
9377 | /* Allow non-strict aligned functions inlining into strict | |
9378 | aligned ones. */ | |
9379 | if ((TARGET_STRICT_ALIGN_P (caller_opts->x_target_flags) | |
9380 | != TARGET_STRICT_ALIGN_P (callee_opts->x_target_flags)) | |
9381 | && !(!TARGET_STRICT_ALIGN_P (callee_opts->x_target_flags) | |
9382 | && TARGET_STRICT_ALIGN_P (caller_opts->x_target_flags))) | |
9383 | return false; | |
9384 | ||
9385 | bool always_inline = lookup_attribute ("always_inline", | |
9386 | DECL_ATTRIBUTES (callee)); | |
9387 | ||
9388 | /* If the architectural features match up and the callee is always_inline | |
9389 | then the other attributes don't matter. */ | |
9390 | if (always_inline) | |
9391 | return true; | |
9392 | ||
9393 | if (caller_opts->x_aarch64_cmodel_var | |
9394 | != callee_opts->x_aarch64_cmodel_var) | |
9395 | return false; | |
9396 | ||
9397 | if (caller_opts->x_aarch64_tls_dialect | |
9398 | != callee_opts->x_aarch64_tls_dialect) | |
9399 | return false; | |
9400 | ||
9401 | /* Honour explicit requests to workaround errata. */ | |
9402 | if (!aarch64_tribools_ok_for_inlining_p ( | |
9403 | caller_opts->x_aarch64_fix_a53_err835769, | |
9404 | callee_opts->x_aarch64_fix_a53_err835769, | |
9405 | 2, TARGET_FIX_ERR_A53_835769_DEFAULT)) | |
9406 | return false; | |
9407 | ||
48bb1a55 CL |
9408 | if (!aarch64_tribools_ok_for_inlining_p ( |
9409 | caller_opts->x_aarch64_fix_a53_err843419, | |
9410 | callee_opts->x_aarch64_fix_a53_err843419, | |
9411 | 2, TARGET_FIX_ERR_A53_843419)) | |
9412 | return false; | |
9413 | ||
1fd8d40c KT |
9414 | /* If the user explicitly specified -momit-leaf-frame-pointer for the |
9415 | caller and calle and they don't match up, reject inlining. */ | |
9416 | if (!aarch64_tribools_ok_for_inlining_p ( | |
9417 | caller_opts->x_flag_omit_leaf_frame_pointer, | |
9418 | callee_opts->x_flag_omit_leaf_frame_pointer, | |
9419 | 2, 1)) | |
9420 | return false; | |
9421 | ||
9422 | /* If the callee has specific tuning overrides, respect them. */ | |
9423 | if (callee_opts->x_aarch64_override_tune_string != NULL | |
9424 | && caller_opts->x_aarch64_override_tune_string == NULL) | |
9425 | return false; | |
9426 | ||
9427 | /* If the user specified tuning override strings for the | |
9428 | caller and callee and they don't match up, reject inlining. | |
9429 | We just do a string compare here, we don't analyze the meaning | |
9430 | of the string, as it would be too costly for little gain. */ | |
9431 | if (callee_opts->x_aarch64_override_tune_string | |
9432 | && caller_opts->x_aarch64_override_tune_string | |
9433 | && (strcmp (callee_opts->x_aarch64_override_tune_string, | |
9434 | caller_opts->x_aarch64_override_tune_string) != 0)) | |
9435 | return false; | |
9436 | ||
9437 | return true; | |
9438 | } | |
9439 | ||
43e9d192 IB |
9440 | /* Return true if SYMBOL_REF X binds locally. */ |
9441 | ||
9442 | static bool | |
9443 | aarch64_symbol_binds_local_p (const_rtx x) | |
9444 | { | |
9445 | return (SYMBOL_REF_DECL (x) | |
9446 | ? targetm.binds_local_p (SYMBOL_REF_DECL (x)) | |
9447 | : SYMBOL_REF_LOCAL_P (x)); | |
9448 | } | |
9449 | ||
9450 | /* Return true if SYMBOL_REF X is thread local */ | |
9451 | static bool | |
9452 | aarch64_tls_symbol_p (rtx x) | |
9453 | { | |
9454 | if (! TARGET_HAVE_TLS) | |
9455 | return false; | |
9456 | ||
9457 | if (GET_CODE (x) != SYMBOL_REF) | |
9458 | return false; | |
9459 | ||
9460 | return SYMBOL_REF_TLS_MODEL (x) != 0; | |
9461 | } | |
9462 | ||
9463 | /* Classify a TLS symbol into one of the TLS kinds. */ | |
9464 | enum aarch64_symbol_type | |
9465 | aarch64_classify_tls_symbol (rtx x) | |
9466 | { | |
9467 | enum tls_model tls_kind = tls_symbolic_operand_type (x); | |
9468 | ||
9469 | switch (tls_kind) | |
9470 | { | |
9471 | case TLS_MODEL_GLOBAL_DYNAMIC: | |
9472 | case TLS_MODEL_LOCAL_DYNAMIC: | |
9473 | return TARGET_TLS_DESC ? SYMBOL_SMALL_TLSDESC : SYMBOL_SMALL_TLSGD; | |
9474 | ||
9475 | case TLS_MODEL_INITIAL_EXEC: | |
5ae7caad JW |
9476 | switch (aarch64_cmodel) |
9477 | { | |
9478 | case AARCH64_CMODEL_TINY: | |
9479 | case AARCH64_CMODEL_TINY_PIC: | |
9480 | return SYMBOL_TINY_TLSIE; | |
9481 | default: | |
79496620 | 9482 | return SYMBOL_SMALL_TLSIE; |
5ae7caad | 9483 | } |
43e9d192 IB |
9484 | |
9485 | case TLS_MODEL_LOCAL_EXEC: | |
cbf5629e JW |
9486 | if (aarch64_tls_size == 12) |
9487 | return SYMBOL_TLSLE12; | |
9488 | else if (aarch64_tls_size == 24) | |
9489 | return SYMBOL_TLSLE24; | |
9490 | else if (aarch64_tls_size == 32) | |
9491 | return SYMBOL_TLSLE32; | |
9492 | else if (aarch64_tls_size == 48) | |
9493 | return SYMBOL_TLSLE48; | |
9494 | else | |
9495 | gcc_unreachable (); | |
43e9d192 IB |
9496 | |
9497 | case TLS_MODEL_EMULATED: | |
9498 | case TLS_MODEL_NONE: | |
9499 | return SYMBOL_FORCE_TO_MEM; | |
9500 | ||
9501 | default: | |
9502 | gcc_unreachable (); | |
9503 | } | |
9504 | } | |
9505 | ||
9506 | /* Return the method that should be used to access SYMBOL_REF or | |
a6e0bfa7 | 9507 | LABEL_REF X. */ |
17f4d4bf | 9508 | |
43e9d192 | 9509 | enum aarch64_symbol_type |
a6e0bfa7 | 9510 | aarch64_classify_symbol (rtx x, rtx offset) |
43e9d192 IB |
9511 | { |
9512 | if (GET_CODE (x) == LABEL_REF) | |
9513 | { | |
9514 | switch (aarch64_cmodel) | |
9515 | { | |
9516 | case AARCH64_CMODEL_LARGE: | |
9517 | return SYMBOL_FORCE_TO_MEM; | |
9518 | ||
9519 | case AARCH64_CMODEL_TINY_PIC: | |
9520 | case AARCH64_CMODEL_TINY: | |
a5350ddc CSS |
9521 | return SYMBOL_TINY_ABSOLUTE; |
9522 | ||
1b1e81f8 | 9523 | case AARCH64_CMODEL_SMALL_SPIC: |
43e9d192 IB |
9524 | case AARCH64_CMODEL_SMALL_PIC: |
9525 | case AARCH64_CMODEL_SMALL: | |
9526 | return SYMBOL_SMALL_ABSOLUTE; | |
9527 | ||
9528 | default: | |
9529 | gcc_unreachable (); | |
9530 | } | |
9531 | } | |
9532 | ||
17f4d4bf | 9533 | if (GET_CODE (x) == SYMBOL_REF) |
43e9d192 | 9534 | { |
43e9d192 IB |
9535 | if (aarch64_tls_symbol_p (x)) |
9536 | return aarch64_classify_tls_symbol (x); | |
9537 | ||
17f4d4bf CSS |
9538 | switch (aarch64_cmodel) |
9539 | { | |
9540 | case AARCH64_CMODEL_TINY: | |
15f6e0da | 9541 | /* When we retrieve symbol + offset address, we have to make sure |
f8b756b7 TB |
9542 | the offset does not cause overflow of the final address. But |
9543 | we have no way of knowing the address of symbol at compile time | |
9544 | so we can't accurately say if the distance between the PC and | |
9545 | symbol + offset is outside the addressible range of +/-1M in the | |
9546 | TINY code model. So we rely on images not being greater than | |
9547 | 1M and cap the offset at 1M and anything beyond 1M will have to | |
15f6e0da RR |
9548 | be loaded using an alternative mechanism. Furthermore if the |
9549 | symbol is a weak reference to something that isn't known to | |
9550 | resolve to a symbol in this module, then force to memory. */ | |
9551 | if ((SYMBOL_REF_WEAK (x) | |
9552 | && !aarch64_symbol_binds_local_p (x)) | |
f8b756b7 | 9553 | || INTVAL (offset) < -1048575 || INTVAL (offset) > 1048575) |
a5350ddc CSS |
9554 | return SYMBOL_FORCE_TO_MEM; |
9555 | return SYMBOL_TINY_ABSOLUTE; | |
9556 | ||
17f4d4bf | 9557 | case AARCH64_CMODEL_SMALL: |
f8b756b7 TB |
9558 | /* Same reasoning as the tiny code model, but the offset cap here is |
9559 | 4G. */ | |
15f6e0da RR |
9560 | if ((SYMBOL_REF_WEAK (x) |
9561 | && !aarch64_symbol_binds_local_p (x)) | |
3ff5d1f0 TB |
9562 | || !IN_RANGE (INTVAL (offset), HOST_WIDE_INT_C (-4294967263), |
9563 | HOST_WIDE_INT_C (4294967264))) | |
17f4d4bf CSS |
9564 | return SYMBOL_FORCE_TO_MEM; |
9565 | return SYMBOL_SMALL_ABSOLUTE; | |
43e9d192 | 9566 | |
17f4d4bf | 9567 | case AARCH64_CMODEL_TINY_PIC: |
38e6c9a6 | 9568 | if (!aarch64_symbol_binds_local_p (x)) |
87dd8ab0 | 9569 | return SYMBOL_TINY_GOT; |
38e6c9a6 MS |
9570 | return SYMBOL_TINY_ABSOLUTE; |
9571 | ||
1b1e81f8 | 9572 | case AARCH64_CMODEL_SMALL_SPIC: |
17f4d4bf CSS |
9573 | case AARCH64_CMODEL_SMALL_PIC: |
9574 | if (!aarch64_symbol_binds_local_p (x)) | |
1b1e81f8 JW |
9575 | return (aarch64_cmodel == AARCH64_CMODEL_SMALL_SPIC |
9576 | ? SYMBOL_SMALL_GOT_28K : SYMBOL_SMALL_GOT_4G); | |
17f4d4bf | 9577 | return SYMBOL_SMALL_ABSOLUTE; |
43e9d192 | 9578 | |
9ee6540a WD |
9579 | case AARCH64_CMODEL_LARGE: |
9580 | /* This is alright even in PIC code as the constant | |
9581 | pool reference is always PC relative and within | |
9582 | the same translation unit. */ | |
9583 | if (CONSTANT_POOL_ADDRESS_P (x)) | |
9584 | return SYMBOL_SMALL_ABSOLUTE; | |
9585 | else | |
9586 | return SYMBOL_FORCE_TO_MEM; | |
9587 | ||
17f4d4bf CSS |
9588 | default: |
9589 | gcc_unreachable (); | |
9590 | } | |
43e9d192 | 9591 | } |
17f4d4bf | 9592 | |
43e9d192 IB |
9593 | /* By default push everything into the constant pool. */ |
9594 | return SYMBOL_FORCE_TO_MEM; | |
9595 | } | |
9596 | ||
43e9d192 IB |
9597 | bool |
9598 | aarch64_constant_address_p (rtx x) | |
9599 | { | |
9600 | return (CONSTANT_P (x) && memory_address_p (DImode, x)); | |
9601 | } | |
9602 | ||
9603 | bool | |
9604 | aarch64_legitimate_pic_operand_p (rtx x) | |
9605 | { | |
9606 | if (GET_CODE (x) == SYMBOL_REF | |
9607 | || (GET_CODE (x) == CONST | |
9608 | && GET_CODE (XEXP (x, 0)) == PLUS | |
9609 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF)) | |
9610 | return false; | |
9611 | ||
9612 | return true; | |
9613 | } | |
9614 | ||
3520f7cc JG |
9615 | /* Return true if X holds either a quarter-precision or |
9616 | floating-point +0.0 constant. */ | |
9617 | static bool | |
ef4bddc2 | 9618 | aarch64_valid_floating_const (machine_mode mode, rtx x) |
3520f7cc JG |
9619 | { |
9620 | if (!CONST_DOUBLE_P (x)) | |
9621 | return false; | |
9622 | ||
6a0f8c01 JW |
9623 | if (aarch64_float_const_zero_rtx_p (x)) |
9624 | return true; | |
9625 | ||
9626 | /* We only handle moving 0.0 to a TFmode register. */ | |
3520f7cc JG |
9627 | if (!(mode == SFmode || mode == DFmode)) |
9628 | return false; | |
9629 | ||
3520f7cc JG |
9630 | return aarch64_float_const_representable_p (x); |
9631 | } | |
9632 | ||
43e9d192 | 9633 | static bool |
ef4bddc2 | 9634 | aarch64_legitimate_constant_p (machine_mode mode, rtx x) |
43e9d192 IB |
9635 | { |
9636 | /* Do not allow vector struct mode constants. We could support | |
9637 | 0 and -1 easily, but they need support in aarch64-simd.md. */ | |
9638 | if (TARGET_SIMD && aarch64_vect_struct_mode_p (mode)) | |
9639 | return false; | |
9640 | ||
9641 | /* This could probably go away because | |
9642 | we now decompose CONST_INTs according to expand_mov_immediate. */ | |
9643 | if ((GET_CODE (x) == CONST_VECTOR | |
48063b9d | 9644 | && aarch64_simd_valid_immediate (x, mode, false, NULL)) |
3520f7cc JG |
9645 | || CONST_INT_P (x) || aarch64_valid_floating_const (mode, x)) |
9646 | return !targetm.cannot_force_const_mem (mode, x); | |
43e9d192 IB |
9647 | |
9648 | if (GET_CODE (x) == HIGH | |
9649 | && aarch64_valid_symref (XEXP (x, 0), GET_MODE (XEXP (x, 0)))) | |
9650 | return true; | |
9651 | ||
9652 | return aarch64_constant_address_p (x); | |
9653 | } | |
9654 | ||
a5bc806c | 9655 | rtx |
43e9d192 IB |
9656 | aarch64_load_tp (rtx target) |
9657 | { | |
9658 | if (!target | |
9659 | || GET_MODE (target) != Pmode | |
9660 | || !register_operand (target, Pmode)) | |
9661 | target = gen_reg_rtx (Pmode); | |
9662 | ||
9663 | /* Can return in any reg. */ | |
9664 | emit_insn (gen_aarch64_load_tp_hard (target)); | |
9665 | return target; | |
9666 | } | |
9667 | ||
43e9d192 IB |
9668 | /* On AAPCS systems, this is the "struct __va_list". */ |
9669 | static GTY(()) tree va_list_type; | |
9670 | ||
9671 | /* Implement TARGET_BUILD_BUILTIN_VA_LIST. | |
9672 | Return the type to use as __builtin_va_list. | |
9673 | ||
9674 | AAPCS64 \S 7.1.4 requires that va_list be a typedef for a type defined as: | |
9675 | ||
9676 | struct __va_list | |
9677 | { | |
9678 | void *__stack; | |
9679 | void *__gr_top; | |
9680 | void *__vr_top; | |
9681 | int __gr_offs; | |
9682 | int __vr_offs; | |
9683 | }; */ | |
9684 | ||
9685 | static tree | |
9686 | aarch64_build_builtin_va_list (void) | |
9687 | { | |
9688 | tree va_list_name; | |
9689 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
9690 | ||
9691 | /* Create the type. */ | |
9692 | va_list_type = lang_hooks.types.make_type (RECORD_TYPE); | |
9693 | /* Give it the required name. */ | |
9694 | va_list_name = build_decl (BUILTINS_LOCATION, | |
9695 | TYPE_DECL, | |
9696 | get_identifier ("__va_list"), | |
9697 | va_list_type); | |
9698 | DECL_ARTIFICIAL (va_list_name) = 1; | |
9699 | TYPE_NAME (va_list_type) = va_list_name; | |
665c56c6 | 9700 | TYPE_STUB_DECL (va_list_type) = va_list_name; |
43e9d192 IB |
9701 | |
9702 | /* Create the fields. */ | |
9703 | f_stack = build_decl (BUILTINS_LOCATION, | |
9704 | FIELD_DECL, get_identifier ("__stack"), | |
9705 | ptr_type_node); | |
9706 | f_grtop = build_decl (BUILTINS_LOCATION, | |
9707 | FIELD_DECL, get_identifier ("__gr_top"), | |
9708 | ptr_type_node); | |
9709 | f_vrtop = build_decl (BUILTINS_LOCATION, | |
9710 | FIELD_DECL, get_identifier ("__vr_top"), | |
9711 | ptr_type_node); | |
9712 | f_groff = build_decl (BUILTINS_LOCATION, | |
9713 | FIELD_DECL, get_identifier ("__gr_offs"), | |
9714 | integer_type_node); | |
9715 | f_vroff = build_decl (BUILTINS_LOCATION, | |
9716 | FIELD_DECL, get_identifier ("__vr_offs"), | |
9717 | integer_type_node); | |
9718 | ||
88e3bdd1 | 9719 | /* Tell tree-stdarg pass about our internal offset fields. |
3fd6b9cc JW |
9720 | NOTE: va_list_gpr/fpr_counter_field are only used for tree comparision |
9721 | purpose to identify whether the code is updating va_list internal | |
9722 | offset fields through irregular way. */ | |
9723 | va_list_gpr_counter_field = f_groff; | |
9724 | va_list_fpr_counter_field = f_vroff; | |
9725 | ||
43e9d192 IB |
9726 | DECL_ARTIFICIAL (f_stack) = 1; |
9727 | DECL_ARTIFICIAL (f_grtop) = 1; | |
9728 | DECL_ARTIFICIAL (f_vrtop) = 1; | |
9729 | DECL_ARTIFICIAL (f_groff) = 1; | |
9730 | DECL_ARTIFICIAL (f_vroff) = 1; | |
9731 | ||
9732 | DECL_FIELD_CONTEXT (f_stack) = va_list_type; | |
9733 | DECL_FIELD_CONTEXT (f_grtop) = va_list_type; | |
9734 | DECL_FIELD_CONTEXT (f_vrtop) = va_list_type; | |
9735 | DECL_FIELD_CONTEXT (f_groff) = va_list_type; | |
9736 | DECL_FIELD_CONTEXT (f_vroff) = va_list_type; | |
9737 | ||
9738 | TYPE_FIELDS (va_list_type) = f_stack; | |
9739 | DECL_CHAIN (f_stack) = f_grtop; | |
9740 | DECL_CHAIN (f_grtop) = f_vrtop; | |
9741 | DECL_CHAIN (f_vrtop) = f_groff; | |
9742 | DECL_CHAIN (f_groff) = f_vroff; | |
9743 | ||
9744 | /* Compute its layout. */ | |
9745 | layout_type (va_list_type); | |
9746 | ||
9747 | return va_list_type; | |
9748 | } | |
9749 | ||
9750 | /* Implement TARGET_EXPAND_BUILTIN_VA_START. */ | |
9751 | static void | |
9752 | aarch64_expand_builtin_va_start (tree valist, rtx nextarg ATTRIBUTE_UNUSED) | |
9753 | { | |
9754 | const CUMULATIVE_ARGS *cum; | |
9755 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
9756 | tree stack, grtop, vrtop, groff, vroff; | |
9757 | tree t; | |
88e3bdd1 JW |
9758 | int gr_save_area_size = cfun->va_list_gpr_size; |
9759 | int vr_save_area_size = cfun->va_list_fpr_size; | |
43e9d192 IB |
9760 | int vr_offset; |
9761 | ||
9762 | cum = &crtl->args.info; | |
88e3bdd1 JW |
9763 | if (cfun->va_list_gpr_size) |
9764 | gr_save_area_size = MIN ((NUM_ARG_REGS - cum->aapcs_ncrn) * UNITS_PER_WORD, | |
9765 | cfun->va_list_gpr_size); | |
9766 | if (cfun->va_list_fpr_size) | |
9767 | vr_save_area_size = MIN ((NUM_FP_ARG_REGS - cum->aapcs_nvrn) | |
9768 | * UNITS_PER_VREG, cfun->va_list_fpr_size); | |
43e9d192 | 9769 | |
d5726973 | 9770 | if (!TARGET_FLOAT) |
43e9d192 | 9771 | { |
261fb553 | 9772 | gcc_assert (cum->aapcs_nvrn == 0); |
43e9d192 IB |
9773 | vr_save_area_size = 0; |
9774 | } | |
9775 | ||
9776 | f_stack = TYPE_FIELDS (va_list_type_node); | |
9777 | f_grtop = DECL_CHAIN (f_stack); | |
9778 | f_vrtop = DECL_CHAIN (f_grtop); | |
9779 | f_groff = DECL_CHAIN (f_vrtop); | |
9780 | f_vroff = DECL_CHAIN (f_groff); | |
9781 | ||
9782 | stack = build3 (COMPONENT_REF, TREE_TYPE (f_stack), valist, f_stack, | |
9783 | NULL_TREE); | |
9784 | grtop = build3 (COMPONENT_REF, TREE_TYPE (f_grtop), valist, f_grtop, | |
9785 | NULL_TREE); | |
9786 | vrtop = build3 (COMPONENT_REF, TREE_TYPE (f_vrtop), valist, f_vrtop, | |
9787 | NULL_TREE); | |
9788 | groff = build3 (COMPONENT_REF, TREE_TYPE (f_groff), valist, f_groff, | |
9789 | NULL_TREE); | |
9790 | vroff = build3 (COMPONENT_REF, TREE_TYPE (f_vroff), valist, f_vroff, | |
9791 | NULL_TREE); | |
9792 | ||
9793 | /* Emit code to initialize STACK, which points to the next varargs stack | |
9794 | argument. CUM->AAPCS_STACK_SIZE gives the number of stack words used | |
9795 | by named arguments. STACK is 8-byte aligned. */ | |
9796 | t = make_tree (TREE_TYPE (stack), virtual_incoming_args_rtx); | |
9797 | if (cum->aapcs_stack_size > 0) | |
9798 | t = fold_build_pointer_plus_hwi (t, cum->aapcs_stack_size * UNITS_PER_WORD); | |
9799 | t = build2 (MODIFY_EXPR, TREE_TYPE (stack), stack, t); | |
9800 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
9801 | ||
9802 | /* Emit code to initialize GRTOP, the top of the GR save area. | |
9803 | virtual_incoming_args_rtx should have been 16 byte aligned. */ | |
9804 | t = make_tree (TREE_TYPE (grtop), virtual_incoming_args_rtx); | |
9805 | t = build2 (MODIFY_EXPR, TREE_TYPE (grtop), grtop, t); | |
9806 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
9807 | ||
9808 | /* Emit code to initialize VRTOP, the top of the VR save area. | |
9809 | This address is gr_save_area_bytes below GRTOP, rounded | |
9810 | down to the next 16-byte boundary. */ | |
9811 | t = make_tree (TREE_TYPE (vrtop), virtual_incoming_args_rtx); | |
4f59f9f2 UB |
9812 | vr_offset = ROUND_UP (gr_save_area_size, |
9813 | STACK_BOUNDARY / BITS_PER_UNIT); | |
43e9d192 IB |
9814 | |
9815 | if (vr_offset) | |
9816 | t = fold_build_pointer_plus_hwi (t, -vr_offset); | |
9817 | t = build2 (MODIFY_EXPR, TREE_TYPE (vrtop), vrtop, t); | |
9818 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
9819 | ||
9820 | /* Emit code to initialize GROFF, the offset from GRTOP of the | |
9821 | next GPR argument. */ | |
9822 | t = build2 (MODIFY_EXPR, TREE_TYPE (groff), groff, | |
9823 | build_int_cst (TREE_TYPE (groff), -gr_save_area_size)); | |
9824 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
9825 | ||
9826 | /* Likewise emit code to initialize VROFF, the offset from FTOP | |
9827 | of the next VR argument. */ | |
9828 | t = build2 (MODIFY_EXPR, TREE_TYPE (vroff), vroff, | |
9829 | build_int_cst (TREE_TYPE (vroff), -vr_save_area_size)); | |
9830 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
9831 | } | |
9832 | ||
9833 | /* Implement TARGET_GIMPLIFY_VA_ARG_EXPR. */ | |
9834 | ||
9835 | static tree | |
9836 | aarch64_gimplify_va_arg_expr (tree valist, tree type, gimple_seq *pre_p, | |
9837 | gimple_seq *post_p ATTRIBUTE_UNUSED) | |
9838 | { | |
9839 | tree addr; | |
9840 | bool indirect_p; | |
9841 | bool is_ha; /* is HFA or HVA. */ | |
9842 | bool dw_align; /* double-word align. */ | |
ef4bddc2 | 9843 | machine_mode ag_mode = VOIDmode; |
43e9d192 | 9844 | int nregs; |
ef4bddc2 | 9845 | machine_mode mode; |
43e9d192 IB |
9846 | |
9847 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
9848 | tree stack, f_top, f_off, off, arg, roundup, on_stack; | |
9849 | HOST_WIDE_INT size, rsize, adjust, align; | |
9850 | tree t, u, cond1, cond2; | |
9851 | ||
9852 | indirect_p = pass_by_reference (NULL, TYPE_MODE (type), type, false); | |
9853 | if (indirect_p) | |
9854 | type = build_pointer_type (type); | |
9855 | ||
9856 | mode = TYPE_MODE (type); | |
9857 | ||
9858 | f_stack = TYPE_FIELDS (va_list_type_node); | |
9859 | f_grtop = DECL_CHAIN (f_stack); | |
9860 | f_vrtop = DECL_CHAIN (f_grtop); | |
9861 | f_groff = DECL_CHAIN (f_vrtop); | |
9862 | f_vroff = DECL_CHAIN (f_groff); | |
9863 | ||
9864 | stack = build3 (COMPONENT_REF, TREE_TYPE (f_stack), unshare_expr (valist), | |
9865 | f_stack, NULL_TREE); | |
9866 | size = int_size_in_bytes (type); | |
9867 | align = aarch64_function_arg_alignment (mode, type) / BITS_PER_UNIT; | |
9868 | ||
9869 | dw_align = false; | |
9870 | adjust = 0; | |
9871 | if (aarch64_vfp_is_call_or_return_candidate (mode, | |
9872 | type, | |
9873 | &ag_mode, | |
9874 | &nregs, | |
9875 | &is_ha)) | |
9876 | { | |
9877 | /* TYPE passed in fp/simd registers. */ | |
d5726973 | 9878 | if (!TARGET_FLOAT) |
261fb553 | 9879 | aarch64_err_no_fpadvsimd (mode, "varargs"); |
43e9d192 IB |
9880 | |
9881 | f_top = build3 (COMPONENT_REF, TREE_TYPE (f_vrtop), | |
9882 | unshare_expr (valist), f_vrtop, NULL_TREE); | |
9883 | f_off = build3 (COMPONENT_REF, TREE_TYPE (f_vroff), | |
9884 | unshare_expr (valist), f_vroff, NULL_TREE); | |
9885 | ||
9886 | rsize = nregs * UNITS_PER_VREG; | |
9887 | ||
9888 | if (is_ha) | |
9889 | { | |
9890 | if (BYTES_BIG_ENDIAN && GET_MODE_SIZE (ag_mode) < UNITS_PER_VREG) | |
9891 | adjust = UNITS_PER_VREG - GET_MODE_SIZE (ag_mode); | |
9892 | } | |
9893 | else if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
9894 | && size < UNITS_PER_VREG) | |
9895 | { | |
9896 | adjust = UNITS_PER_VREG - size; | |
9897 | } | |
9898 | } | |
9899 | else | |
9900 | { | |
9901 | /* TYPE passed in general registers. */ | |
9902 | f_top = build3 (COMPONENT_REF, TREE_TYPE (f_grtop), | |
9903 | unshare_expr (valist), f_grtop, NULL_TREE); | |
9904 | f_off = build3 (COMPONENT_REF, TREE_TYPE (f_groff), | |
9905 | unshare_expr (valist), f_groff, NULL_TREE); | |
4f59f9f2 | 9906 | rsize = ROUND_UP (size, UNITS_PER_WORD); |
43e9d192 IB |
9907 | nregs = rsize / UNITS_PER_WORD; |
9908 | ||
9909 | if (align > 8) | |
9910 | dw_align = true; | |
9911 | ||
9912 | if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
9913 | && size < UNITS_PER_WORD) | |
9914 | { | |
9915 | adjust = UNITS_PER_WORD - size; | |
9916 | } | |
9917 | } | |
9918 | ||
9919 | /* Get a local temporary for the field value. */ | |
9920 | off = get_initialized_tmp_var (f_off, pre_p, NULL); | |
9921 | ||
9922 | /* Emit code to branch if off >= 0. */ | |
9923 | t = build2 (GE_EXPR, boolean_type_node, off, | |
9924 | build_int_cst (TREE_TYPE (off), 0)); | |
9925 | cond1 = build3 (COND_EXPR, ptr_type_node, t, NULL_TREE, NULL_TREE); | |
9926 | ||
9927 | if (dw_align) | |
9928 | { | |
9929 | /* Emit: offs = (offs + 15) & -16. */ | |
9930 | t = build2 (PLUS_EXPR, TREE_TYPE (off), off, | |
9931 | build_int_cst (TREE_TYPE (off), 15)); | |
9932 | t = build2 (BIT_AND_EXPR, TREE_TYPE (off), t, | |
9933 | build_int_cst (TREE_TYPE (off), -16)); | |
9934 | roundup = build2 (MODIFY_EXPR, TREE_TYPE (off), off, t); | |
9935 | } | |
9936 | else | |
9937 | roundup = NULL; | |
9938 | ||
9939 | /* Update ap.__[g|v]r_offs */ | |
9940 | t = build2 (PLUS_EXPR, TREE_TYPE (off), off, | |
9941 | build_int_cst (TREE_TYPE (off), rsize)); | |
9942 | t = build2 (MODIFY_EXPR, TREE_TYPE (f_off), unshare_expr (f_off), t); | |
9943 | ||
9944 | /* String up. */ | |
9945 | if (roundup) | |
9946 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), roundup, t); | |
9947 | ||
9948 | /* [cond2] if (ap.__[g|v]r_offs > 0) */ | |
9949 | u = build2 (GT_EXPR, boolean_type_node, unshare_expr (f_off), | |
9950 | build_int_cst (TREE_TYPE (f_off), 0)); | |
9951 | cond2 = build3 (COND_EXPR, ptr_type_node, u, NULL_TREE, NULL_TREE); | |
9952 | ||
9953 | /* String up: make sure the assignment happens before the use. */ | |
9954 | t = build2 (COMPOUND_EXPR, TREE_TYPE (cond2), t, cond2); | |
9955 | COND_EXPR_ELSE (cond1) = t; | |
9956 | ||
9957 | /* Prepare the trees handling the argument that is passed on the stack; | |
9958 | the top level node will store in ON_STACK. */ | |
9959 | arg = get_initialized_tmp_var (stack, pre_p, NULL); | |
9960 | if (align > 8) | |
9961 | { | |
9962 | /* if (alignof(type) > 8) (arg = arg + 15) & -16; */ | |
9963 | t = fold_convert (intDI_type_node, arg); | |
9964 | t = build2 (PLUS_EXPR, TREE_TYPE (t), t, | |
9965 | build_int_cst (TREE_TYPE (t), 15)); | |
9966 | t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t, | |
9967 | build_int_cst (TREE_TYPE (t), -16)); | |
9968 | t = fold_convert (TREE_TYPE (arg), t); | |
9969 | roundup = build2 (MODIFY_EXPR, TREE_TYPE (arg), arg, t); | |
9970 | } | |
9971 | else | |
9972 | roundup = NULL; | |
9973 | /* Advance ap.__stack */ | |
9974 | t = fold_convert (intDI_type_node, arg); | |
9975 | t = build2 (PLUS_EXPR, TREE_TYPE (t), t, | |
9976 | build_int_cst (TREE_TYPE (t), size + 7)); | |
9977 | t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t, | |
9978 | build_int_cst (TREE_TYPE (t), -8)); | |
9979 | t = fold_convert (TREE_TYPE (arg), t); | |
9980 | t = build2 (MODIFY_EXPR, TREE_TYPE (stack), unshare_expr (stack), t); | |
9981 | /* String up roundup and advance. */ | |
9982 | if (roundup) | |
9983 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), roundup, t); | |
9984 | /* String up with arg */ | |
9985 | on_stack = build2 (COMPOUND_EXPR, TREE_TYPE (arg), t, arg); | |
9986 | /* Big-endianness related address adjustment. */ | |
9987 | if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
9988 | && size < UNITS_PER_WORD) | |
9989 | { | |
9990 | t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (arg), arg, | |
9991 | size_int (UNITS_PER_WORD - size)); | |
9992 | on_stack = build2 (COMPOUND_EXPR, TREE_TYPE (arg), on_stack, t); | |
9993 | } | |
9994 | ||
9995 | COND_EXPR_THEN (cond1) = unshare_expr (on_stack); | |
9996 | COND_EXPR_THEN (cond2) = unshare_expr (on_stack); | |
9997 | ||
9998 | /* Adjustment to OFFSET in the case of BIG_ENDIAN. */ | |
9999 | t = off; | |
10000 | if (adjust) | |
10001 | t = build2 (PREINCREMENT_EXPR, TREE_TYPE (off), off, | |
10002 | build_int_cst (TREE_TYPE (off), adjust)); | |
10003 | ||
10004 | t = fold_convert (sizetype, t); | |
10005 | t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (f_top), f_top, t); | |
10006 | ||
10007 | if (is_ha) | |
10008 | { | |
10009 | /* type ha; // treat as "struct {ftype field[n];}" | |
10010 | ... [computing offs] | |
10011 | for (i = 0; i <nregs; ++i, offs += 16) | |
10012 | ha.field[i] = *((ftype *)(ap.__vr_top + offs)); | |
10013 | return ha; */ | |
10014 | int i; | |
10015 | tree tmp_ha, field_t, field_ptr_t; | |
10016 | ||
10017 | /* Declare a local variable. */ | |
10018 | tmp_ha = create_tmp_var_raw (type, "ha"); | |
10019 | gimple_add_tmp_var (tmp_ha); | |
10020 | ||
10021 | /* Establish the base type. */ | |
10022 | switch (ag_mode) | |
10023 | { | |
10024 | case SFmode: | |
10025 | field_t = float_type_node; | |
10026 | field_ptr_t = float_ptr_type_node; | |
10027 | break; | |
10028 | case DFmode: | |
10029 | field_t = double_type_node; | |
10030 | field_ptr_t = double_ptr_type_node; | |
10031 | break; | |
10032 | case TFmode: | |
10033 | field_t = long_double_type_node; | |
10034 | field_ptr_t = long_double_ptr_type_node; | |
10035 | break; | |
43e9d192 | 10036 | case HFmode: |
1b62ed4f JG |
10037 | field_t = aarch64_fp16_type_node; |
10038 | field_ptr_t = aarch64_fp16_ptr_type_node; | |
43e9d192 | 10039 | break; |
43e9d192 IB |
10040 | case V2SImode: |
10041 | case V4SImode: | |
10042 | { | |
10043 | tree innertype = make_signed_type (GET_MODE_PRECISION (SImode)); | |
10044 | field_t = build_vector_type_for_mode (innertype, ag_mode); | |
10045 | field_ptr_t = build_pointer_type (field_t); | |
10046 | } | |
10047 | break; | |
10048 | default: | |
10049 | gcc_assert (0); | |
10050 | } | |
10051 | ||
10052 | /* *(field_ptr_t)&ha = *((field_ptr_t)vr_saved_area */ | |
10053 | tmp_ha = build1 (ADDR_EXPR, field_ptr_t, tmp_ha); | |
10054 | addr = t; | |
10055 | t = fold_convert (field_ptr_t, addr); | |
10056 | t = build2 (MODIFY_EXPR, field_t, | |
10057 | build1 (INDIRECT_REF, field_t, tmp_ha), | |
10058 | build1 (INDIRECT_REF, field_t, t)); | |
10059 | ||
10060 | /* ha.field[i] = *((field_ptr_t)vr_saved_area + i) */ | |
10061 | for (i = 1; i < nregs; ++i) | |
10062 | { | |
10063 | addr = fold_build_pointer_plus_hwi (addr, UNITS_PER_VREG); | |
10064 | u = fold_convert (field_ptr_t, addr); | |
10065 | u = build2 (MODIFY_EXPR, field_t, | |
10066 | build2 (MEM_REF, field_t, tmp_ha, | |
10067 | build_int_cst (field_ptr_t, | |
10068 | (i * | |
10069 | int_size_in_bytes (field_t)))), | |
10070 | build1 (INDIRECT_REF, field_t, u)); | |
10071 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), t, u); | |
10072 | } | |
10073 | ||
10074 | u = fold_convert (TREE_TYPE (f_top), tmp_ha); | |
10075 | t = build2 (COMPOUND_EXPR, TREE_TYPE (f_top), t, u); | |
10076 | } | |
10077 | ||
10078 | COND_EXPR_ELSE (cond2) = t; | |
10079 | addr = fold_convert (build_pointer_type (type), cond1); | |
10080 | addr = build_va_arg_indirect_ref (addr); | |
10081 | ||
10082 | if (indirect_p) | |
10083 | addr = build_va_arg_indirect_ref (addr); | |
10084 | ||
10085 | return addr; | |
10086 | } | |
10087 | ||
10088 | /* Implement TARGET_SETUP_INCOMING_VARARGS. */ | |
10089 | ||
10090 | static void | |
ef4bddc2 | 10091 | aarch64_setup_incoming_varargs (cumulative_args_t cum_v, machine_mode mode, |
43e9d192 IB |
10092 | tree type, int *pretend_size ATTRIBUTE_UNUSED, |
10093 | int no_rtl) | |
10094 | { | |
10095 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
10096 | CUMULATIVE_ARGS local_cum; | |
88e3bdd1 JW |
10097 | int gr_saved = cfun->va_list_gpr_size; |
10098 | int vr_saved = cfun->va_list_fpr_size; | |
43e9d192 IB |
10099 | |
10100 | /* The caller has advanced CUM up to, but not beyond, the last named | |
10101 | argument. Advance a local copy of CUM past the last "real" named | |
10102 | argument, to find out how many registers are left over. */ | |
10103 | local_cum = *cum; | |
10104 | aarch64_function_arg_advance (pack_cumulative_args(&local_cum), mode, type, true); | |
10105 | ||
88e3bdd1 JW |
10106 | /* Found out how many registers we need to save. |
10107 | Honor tree-stdvar analysis results. */ | |
10108 | if (cfun->va_list_gpr_size) | |
10109 | gr_saved = MIN (NUM_ARG_REGS - local_cum.aapcs_ncrn, | |
10110 | cfun->va_list_gpr_size / UNITS_PER_WORD); | |
10111 | if (cfun->va_list_fpr_size) | |
10112 | vr_saved = MIN (NUM_FP_ARG_REGS - local_cum.aapcs_nvrn, | |
10113 | cfun->va_list_fpr_size / UNITS_PER_VREG); | |
43e9d192 | 10114 | |
d5726973 | 10115 | if (!TARGET_FLOAT) |
43e9d192 | 10116 | { |
261fb553 | 10117 | gcc_assert (local_cum.aapcs_nvrn == 0); |
43e9d192 IB |
10118 | vr_saved = 0; |
10119 | } | |
10120 | ||
10121 | if (!no_rtl) | |
10122 | { | |
10123 | if (gr_saved > 0) | |
10124 | { | |
10125 | rtx ptr, mem; | |
10126 | ||
10127 | /* virtual_incoming_args_rtx should have been 16-byte aligned. */ | |
10128 | ptr = plus_constant (Pmode, virtual_incoming_args_rtx, | |
10129 | - gr_saved * UNITS_PER_WORD); | |
10130 | mem = gen_frame_mem (BLKmode, ptr); | |
10131 | set_mem_alias_set (mem, get_varargs_alias_set ()); | |
10132 | ||
10133 | move_block_from_reg (local_cum.aapcs_ncrn + R0_REGNUM, | |
10134 | mem, gr_saved); | |
10135 | } | |
10136 | if (vr_saved > 0) | |
10137 | { | |
10138 | /* We can't use move_block_from_reg, because it will use | |
10139 | the wrong mode, storing D regs only. */ | |
ef4bddc2 | 10140 | machine_mode mode = TImode; |
88e3bdd1 | 10141 | int off, i, vr_start; |
43e9d192 IB |
10142 | |
10143 | /* Set OFF to the offset from virtual_incoming_args_rtx of | |
10144 | the first vector register. The VR save area lies below | |
10145 | the GR one, and is aligned to 16 bytes. */ | |
4f59f9f2 UB |
10146 | off = -ROUND_UP (gr_saved * UNITS_PER_WORD, |
10147 | STACK_BOUNDARY / BITS_PER_UNIT); | |
43e9d192 IB |
10148 | off -= vr_saved * UNITS_PER_VREG; |
10149 | ||
88e3bdd1 JW |
10150 | vr_start = V0_REGNUM + local_cum.aapcs_nvrn; |
10151 | for (i = 0; i < vr_saved; ++i) | |
43e9d192 IB |
10152 | { |
10153 | rtx ptr, mem; | |
10154 | ||
10155 | ptr = plus_constant (Pmode, virtual_incoming_args_rtx, off); | |
10156 | mem = gen_frame_mem (mode, ptr); | |
10157 | set_mem_alias_set (mem, get_varargs_alias_set ()); | |
88e3bdd1 | 10158 | aarch64_emit_move (mem, gen_rtx_REG (mode, vr_start + i)); |
43e9d192 IB |
10159 | off += UNITS_PER_VREG; |
10160 | } | |
10161 | } | |
10162 | } | |
10163 | ||
10164 | /* We don't save the size into *PRETEND_SIZE because we want to avoid | |
10165 | any complication of having crtl->args.pretend_args_size changed. */ | |
8799637a | 10166 | cfun->machine->frame.saved_varargs_size |
4f59f9f2 UB |
10167 | = (ROUND_UP (gr_saved * UNITS_PER_WORD, |
10168 | STACK_BOUNDARY / BITS_PER_UNIT) | |
43e9d192 IB |
10169 | + vr_saved * UNITS_PER_VREG); |
10170 | } | |
10171 | ||
10172 | static void | |
10173 | aarch64_conditional_register_usage (void) | |
10174 | { | |
10175 | int i; | |
10176 | if (!TARGET_FLOAT) | |
10177 | { | |
10178 | for (i = V0_REGNUM; i <= V31_REGNUM; i++) | |
10179 | { | |
10180 | fixed_regs[i] = 1; | |
10181 | call_used_regs[i] = 1; | |
10182 | } | |
10183 | } | |
10184 | } | |
10185 | ||
10186 | /* Walk down the type tree of TYPE counting consecutive base elements. | |
10187 | If *MODEP is VOIDmode, then set it to the first valid floating point | |
10188 | type. If a non-floating point type is found, or if a floating point | |
10189 | type that doesn't match a non-VOIDmode *MODEP is found, then return -1, | |
10190 | otherwise return the count in the sub-tree. */ | |
10191 | static int | |
ef4bddc2 | 10192 | aapcs_vfp_sub_candidate (const_tree type, machine_mode *modep) |
43e9d192 | 10193 | { |
ef4bddc2 | 10194 | machine_mode mode; |
43e9d192 IB |
10195 | HOST_WIDE_INT size; |
10196 | ||
10197 | switch (TREE_CODE (type)) | |
10198 | { | |
10199 | case REAL_TYPE: | |
10200 | mode = TYPE_MODE (type); | |
1b62ed4f JG |
10201 | if (mode != DFmode && mode != SFmode |
10202 | && mode != TFmode && mode != HFmode) | |
43e9d192 IB |
10203 | return -1; |
10204 | ||
10205 | if (*modep == VOIDmode) | |
10206 | *modep = mode; | |
10207 | ||
10208 | if (*modep == mode) | |
10209 | return 1; | |
10210 | ||
10211 | break; | |
10212 | ||
10213 | case COMPLEX_TYPE: | |
10214 | mode = TYPE_MODE (TREE_TYPE (type)); | |
1b62ed4f JG |
10215 | if (mode != DFmode && mode != SFmode |
10216 | && mode != TFmode && mode != HFmode) | |
43e9d192 IB |
10217 | return -1; |
10218 | ||
10219 | if (*modep == VOIDmode) | |
10220 | *modep = mode; | |
10221 | ||
10222 | if (*modep == mode) | |
10223 | return 2; | |
10224 | ||
10225 | break; | |
10226 | ||
10227 | case VECTOR_TYPE: | |
10228 | /* Use V2SImode and V4SImode as representatives of all 64-bit | |
10229 | and 128-bit vector types. */ | |
10230 | size = int_size_in_bytes (type); | |
10231 | switch (size) | |
10232 | { | |
10233 | case 8: | |
10234 | mode = V2SImode; | |
10235 | break; | |
10236 | case 16: | |
10237 | mode = V4SImode; | |
10238 | break; | |
10239 | default: | |
10240 | return -1; | |
10241 | } | |
10242 | ||
10243 | if (*modep == VOIDmode) | |
10244 | *modep = mode; | |
10245 | ||
10246 | /* Vector modes are considered to be opaque: two vectors are | |
10247 | equivalent for the purposes of being homogeneous aggregates | |
10248 | if they are the same size. */ | |
10249 | if (*modep == mode) | |
10250 | return 1; | |
10251 | ||
10252 | break; | |
10253 | ||
10254 | case ARRAY_TYPE: | |
10255 | { | |
10256 | int count; | |
10257 | tree index = TYPE_DOMAIN (type); | |
10258 | ||
807e902e KZ |
10259 | /* Can't handle incomplete types nor sizes that are not |
10260 | fixed. */ | |
10261 | if (!COMPLETE_TYPE_P (type) | |
10262 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
43e9d192 IB |
10263 | return -1; |
10264 | ||
10265 | count = aapcs_vfp_sub_candidate (TREE_TYPE (type), modep); | |
10266 | if (count == -1 | |
10267 | || !index | |
10268 | || !TYPE_MAX_VALUE (index) | |
cc269bb6 | 10269 | || !tree_fits_uhwi_p (TYPE_MAX_VALUE (index)) |
43e9d192 | 10270 | || !TYPE_MIN_VALUE (index) |
cc269bb6 | 10271 | || !tree_fits_uhwi_p (TYPE_MIN_VALUE (index)) |
43e9d192 IB |
10272 | || count < 0) |
10273 | return -1; | |
10274 | ||
ae7e9ddd RS |
10275 | count *= (1 + tree_to_uhwi (TYPE_MAX_VALUE (index)) |
10276 | - tree_to_uhwi (TYPE_MIN_VALUE (index))); | |
43e9d192 IB |
10277 | |
10278 | /* There must be no padding. */ | |
807e902e | 10279 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) |
43e9d192 IB |
10280 | return -1; |
10281 | ||
10282 | return count; | |
10283 | } | |
10284 | ||
10285 | case RECORD_TYPE: | |
10286 | { | |
10287 | int count = 0; | |
10288 | int sub_count; | |
10289 | tree field; | |
10290 | ||
807e902e KZ |
10291 | /* Can't handle incomplete types nor sizes that are not |
10292 | fixed. */ | |
10293 | if (!COMPLETE_TYPE_P (type) | |
10294 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
43e9d192 IB |
10295 | return -1; |
10296 | ||
10297 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
10298 | { | |
10299 | if (TREE_CODE (field) != FIELD_DECL) | |
10300 | continue; | |
10301 | ||
10302 | sub_count = aapcs_vfp_sub_candidate (TREE_TYPE (field), modep); | |
10303 | if (sub_count < 0) | |
10304 | return -1; | |
10305 | count += sub_count; | |
10306 | } | |
10307 | ||
10308 | /* There must be no padding. */ | |
807e902e | 10309 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) |
43e9d192 IB |
10310 | return -1; |
10311 | ||
10312 | return count; | |
10313 | } | |
10314 | ||
10315 | case UNION_TYPE: | |
10316 | case QUAL_UNION_TYPE: | |
10317 | { | |
10318 | /* These aren't very interesting except in a degenerate case. */ | |
10319 | int count = 0; | |
10320 | int sub_count; | |
10321 | tree field; | |
10322 | ||
807e902e KZ |
10323 | /* Can't handle incomplete types nor sizes that are not |
10324 | fixed. */ | |
10325 | if (!COMPLETE_TYPE_P (type) | |
10326 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
43e9d192 IB |
10327 | return -1; |
10328 | ||
10329 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
10330 | { | |
10331 | if (TREE_CODE (field) != FIELD_DECL) | |
10332 | continue; | |
10333 | ||
10334 | sub_count = aapcs_vfp_sub_candidate (TREE_TYPE (field), modep); | |
10335 | if (sub_count < 0) | |
10336 | return -1; | |
10337 | count = count > sub_count ? count : sub_count; | |
10338 | } | |
10339 | ||
10340 | /* There must be no padding. */ | |
807e902e | 10341 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) |
43e9d192 IB |
10342 | return -1; |
10343 | ||
10344 | return count; | |
10345 | } | |
10346 | ||
10347 | default: | |
10348 | break; | |
10349 | } | |
10350 | ||
10351 | return -1; | |
10352 | } | |
10353 | ||
b6ec6215 KT |
10354 | /* Return TRUE if the type, as described by TYPE and MODE, is a short vector |
10355 | type as described in AAPCS64 \S 4.1.2. | |
10356 | ||
10357 | See the comment above aarch64_composite_type_p for the notes on MODE. */ | |
10358 | ||
10359 | static bool | |
10360 | aarch64_short_vector_p (const_tree type, | |
10361 | machine_mode mode) | |
10362 | { | |
10363 | HOST_WIDE_INT size = -1; | |
10364 | ||
10365 | if (type && TREE_CODE (type) == VECTOR_TYPE) | |
10366 | size = int_size_in_bytes (type); | |
10367 | else if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT | |
10368 | || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) | |
10369 | size = GET_MODE_SIZE (mode); | |
10370 | ||
10371 | return (size == 8 || size == 16); | |
10372 | } | |
10373 | ||
43e9d192 IB |
10374 | /* Return TRUE if the type, as described by TYPE and MODE, is a composite |
10375 | type as described in AAPCS64 \S 4.3. This includes aggregate, union and | |
10376 | array types. The C99 floating-point complex types are also considered | |
10377 | as composite types, according to AAPCS64 \S 7.1.1. The complex integer | |
10378 | types, which are GCC extensions and out of the scope of AAPCS64, are | |
10379 | treated as composite types here as well. | |
10380 | ||
10381 | Note that MODE itself is not sufficient in determining whether a type | |
10382 | is such a composite type or not. This is because | |
10383 | stor-layout.c:compute_record_mode may have already changed the MODE | |
10384 | (BLKmode) of a RECORD_TYPE TYPE to some other mode. For example, a | |
10385 | structure with only one field may have its MODE set to the mode of the | |
10386 | field. Also an integer mode whose size matches the size of the | |
10387 | RECORD_TYPE type may be used to substitute the original mode | |
10388 | (i.e. BLKmode) in certain circumstances. In other words, MODE cannot be | |
10389 | solely relied on. */ | |
10390 | ||
10391 | static bool | |
10392 | aarch64_composite_type_p (const_tree type, | |
ef4bddc2 | 10393 | machine_mode mode) |
43e9d192 | 10394 | { |
b6ec6215 KT |
10395 | if (aarch64_short_vector_p (type, mode)) |
10396 | return false; | |
10397 | ||
43e9d192 IB |
10398 | if (type && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == COMPLEX_TYPE)) |
10399 | return true; | |
10400 | ||
10401 | if (mode == BLKmode | |
10402 | || GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT | |
10403 | || GET_MODE_CLASS (mode) == MODE_COMPLEX_INT) | |
10404 | return true; | |
10405 | ||
10406 | return false; | |
10407 | } | |
10408 | ||
43e9d192 IB |
10409 | /* Return TRUE if an argument, whose type is described by TYPE and MODE, |
10410 | shall be passed or returned in simd/fp register(s) (providing these | |
10411 | parameter passing registers are available). | |
10412 | ||
10413 | Upon successful return, *COUNT returns the number of needed registers, | |
10414 | *BASE_MODE returns the mode of the individual register and when IS_HAF | |
10415 | is not NULL, *IS_HA indicates whether or not the argument is a homogeneous | |
10416 | floating-point aggregate or a homogeneous short-vector aggregate. */ | |
10417 | ||
10418 | static bool | |
ef4bddc2 | 10419 | aarch64_vfp_is_call_or_return_candidate (machine_mode mode, |
43e9d192 | 10420 | const_tree type, |
ef4bddc2 | 10421 | machine_mode *base_mode, |
43e9d192 IB |
10422 | int *count, |
10423 | bool *is_ha) | |
10424 | { | |
ef4bddc2 | 10425 | machine_mode new_mode = VOIDmode; |
43e9d192 IB |
10426 | bool composite_p = aarch64_composite_type_p (type, mode); |
10427 | ||
10428 | if (is_ha != NULL) *is_ha = false; | |
10429 | ||
10430 | if ((!composite_p && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
10431 | || aarch64_short_vector_p (type, mode)) | |
10432 | { | |
10433 | *count = 1; | |
10434 | new_mode = mode; | |
10435 | } | |
10436 | else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT) | |
10437 | { | |
10438 | if (is_ha != NULL) *is_ha = true; | |
10439 | *count = 2; | |
10440 | new_mode = GET_MODE_INNER (mode); | |
10441 | } | |
10442 | else if (type && composite_p) | |
10443 | { | |
10444 | int ag_count = aapcs_vfp_sub_candidate (type, &new_mode); | |
10445 | ||
10446 | if (ag_count > 0 && ag_count <= HA_MAX_NUM_FLDS) | |
10447 | { | |
10448 | if (is_ha != NULL) *is_ha = true; | |
10449 | *count = ag_count; | |
10450 | } | |
10451 | else | |
10452 | return false; | |
10453 | } | |
10454 | else | |
10455 | return false; | |
10456 | ||
10457 | *base_mode = new_mode; | |
10458 | return true; | |
10459 | } | |
10460 | ||
10461 | /* Implement TARGET_STRUCT_VALUE_RTX. */ | |
10462 | ||
10463 | static rtx | |
10464 | aarch64_struct_value_rtx (tree fndecl ATTRIBUTE_UNUSED, | |
10465 | int incoming ATTRIBUTE_UNUSED) | |
10466 | { | |
10467 | return gen_rtx_REG (Pmode, AARCH64_STRUCT_VALUE_REGNUM); | |
10468 | } | |
10469 | ||
10470 | /* Implements target hook vector_mode_supported_p. */ | |
10471 | static bool | |
ef4bddc2 | 10472 | aarch64_vector_mode_supported_p (machine_mode mode) |
43e9d192 IB |
10473 | { |
10474 | if (TARGET_SIMD | |
10475 | && (mode == V4SImode || mode == V8HImode | |
10476 | || mode == V16QImode || mode == V2DImode | |
10477 | || mode == V2SImode || mode == V4HImode | |
10478 | || mode == V8QImode || mode == V2SFmode | |
ad7d90cc | 10479 | || mode == V4SFmode || mode == V2DFmode |
71a11456 | 10480 | || mode == V4HFmode || mode == V8HFmode |
ad7d90cc | 10481 | || mode == V1DFmode)) |
43e9d192 IB |
10482 | return true; |
10483 | ||
10484 | return false; | |
10485 | } | |
10486 | ||
b7342d25 IB |
10487 | /* Return appropriate SIMD container |
10488 | for MODE within a vector of WIDTH bits. */ | |
ef4bddc2 RS |
10489 | static machine_mode |
10490 | aarch64_simd_container_mode (machine_mode mode, unsigned width) | |
43e9d192 | 10491 | { |
b7342d25 | 10492 | gcc_assert (width == 64 || width == 128); |
43e9d192 | 10493 | if (TARGET_SIMD) |
b7342d25 IB |
10494 | { |
10495 | if (width == 128) | |
10496 | switch (mode) | |
10497 | { | |
10498 | case DFmode: | |
10499 | return V2DFmode; | |
10500 | case SFmode: | |
10501 | return V4SFmode; | |
10502 | case SImode: | |
10503 | return V4SImode; | |
10504 | case HImode: | |
10505 | return V8HImode; | |
10506 | case QImode: | |
10507 | return V16QImode; | |
10508 | case DImode: | |
10509 | return V2DImode; | |
10510 | default: | |
10511 | break; | |
10512 | } | |
10513 | else | |
10514 | switch (mode) | |
10515 | { | |
10516 | case SFmode: | |
10517 | return V2SFmode; | |
10518 | case SImode: | |
10519 | return V2SImode; | |
10520 | case HImode: | |
10521 | return V4HImode; | |
10522 | case QImode: | |
10523 | return V8QImode; | |
10524 | default: | |
10525 | break; | |
10526 | } | |
10527 | } | |
43e9d192 IB |
10528 | return word_mode; |
10529 | } | |
10530 | ||
b7342d25 | 10531 | /* Return 128-bit container as the preferred SIMD mode for MODE. */ |
ef4bddc2 RS |
10532 | static machine_mode |
10533 | aarch64_preferred_simd_mode (machine_mode mode) | |
b7342d25 IB |
10534 | { |
10535 | return aarch64_simd_container_mode (mode, 128); | |
10536 | } | |
10537 | ||
3b357264 JG |
10538 | /* Return the bitmask of possible vector sizes for the vectorizer |
10539 | to iterate over. */ | |
10540 | static unsigned int | |
10541 | aarch64_autovectorize_vector_sizes (void) | |
10542 | { | |
10543 | return (16 | 8); | |
10544 | } | |
10545 | ||
ac2b960f YZ |
10546 | /* Implement TARGET_MANGLE_TYPE. */ |
10547 | ||
6f549691 | 10548 | static const char * |
ac2b960f YZ |
10549 | aarch64_mangle_type (const_tree type) |
10550 | { | |
10551 | /* The AArch64 ABI documents say that "__va_list" has to be | |
10552 | managled as if it is in the "std" namespace. */ | |
10553 | if (lang_hooks.types_compatible_p (CONST_CAST_TREE (type), va_list_type)) | |
10554 | return "St9__va_list"; | |
10555 | ||
c2ec330c AL |
10556 | /* Half-precision float. */ |
10557 | if (TREE_CODE (type) == REAL_TYPE && TYPE_PRECISION (type) == 16) | |
10558 | return "Dh"; | |
10559 | ||
f9d53c27 TB |
10560 | /* Mangle AArch64-specific internal types. TYPE_NAME is non-NULL_TREE for |
10561 | builtin types. */ | |
10562 | if (TYPE_NAME (type) != NULL) | |
10563 | return aarch64_mangle_builtin_type (type); | |
c6fc9e43 | 10564 | |
ac2b960f YZ |
10565 | /* Use the default mangling. */ |
10566 | return NULL; | |
10567 | } | |
10568 | ||
8baff86e KT |
10569 | |
10570 | /* Return true if the rtx_insn contains a MEM RTX somewhere | |
10571 | in it. */ | |
75cf1494 KT |
10572 | |
10573 | static bool | |
8baff86e | 10574 | has_memory_op (rtx_insn *mem_insn) |
75cf1494 | 10575 | { |
8baff86e KT |
10576 | subrtx_iterator::array_type array; |
10577 | FOR_EACH_SUBRTX (iter, array, PATTERN (mem_insn), ALL) | |
10578 | if (MEM_P (*iter)) | |
10579 | return true; | |
10580 | ||
10581 | return false; | |
75cf1494 KT |
10582 | } |
10583 | ||
10584 | /* Find the first rtx_insn before insn that will generate an assembly | |
10585 | instruction. */ | |
10586 | ||
10587 | static rtx_insn * | |
10588 | aarch64_prev_real_insn (rtx_insn *insn) | |
10589 | { | |
10590 | if (!insn) | |
10591 | return NULL; | |
10592 | ||
10593 | do | |
10594 | { | |
10595 | insn = prev_real_insn (insn); | |
10596 | } | |
10597 | while (insn && recog_memoized (insn) < 0); | |
10598 | ||
10599 | return insn; | |
10600 | } | |
10601 | ||
10602 | static bool | |
10603 | is_madd_op (enum attr_type t1) | |
10604 | { | |
10605 | unsigned int i; | |
10606 | /* A number of these may be AArch32 only. */ | |
10607 | enum attr_type mlatypes[] = { | |
10608 | TYPE_MLA, TYPE_MLAS, TYPE_SMLAD, TYPE_SMLADX, TYPE_SMLAL, TYPE_SMLALD, | |
10609 | TYPE_SMLALS, TYPE_SMLALXY, TYPE_SMLAWX, TYPE_SMLAWY, TYPE_SMLAXY, | |
10610 | TYPE_SMMLA, TYPE_UMLAL, TYPE_UMLALS,TYPE_SMLSD, TYPE_SMLSDX, TYPE_SMLSLD | |
10611 | }; | |
10612 | ||
10613 | for (i = 0; i < sizeof (mlatypes) / sizeof (enum attr_type); i++) | |
10614 | { | |
10615 | if (t1 == mlatypes[i]) | |
10616 | return true; | |
10617 | } | |
10618 | ||
10619 | return false; | |
10620 | } | |
10621 | ||
10622 | /* Check if there is a register dependency between a load and the insn | |
10623 | for which we hold recog_data. */ | |
10624 | ||
10625 | static bool | |
10626 | dep_between_memop_and_curr (rtx memop) | |
10627 | { | |
10628 | rtx load_reg; | |
10629 | int opno; | |
10630 | ||
8baff86e | 10631 | gcc_assert (GET_CODE (memop) == SET); |
75cf1494 KT |
10632 | |
10633 | if (!REG_P (SET_DEST (memop))) | |
10634 | return false; | |
10635 | ||
10636 | load_reg = SET_DEST (memop); | |
8baff86e | 10637 | for (opno = 1; opno < recog_data.n_operands; opno++) |
75cf1494 KT |
10638 | { |
10639 | rtx operand = recog_data.operand[opno]; | |
10640 | if (REG_P (operand) | |
10641 | && reg_overlap_mentioned_p (load_reg, operand)) | |
10642 | return true; | |
10643 | ||
10644 | } | |
10645 | return false; | |
10646 | } | |
10647 | ||
8baff86e KT |
10648 | |
10649 | /* When working around the Cortex-A53 erratum 835769, | |
10650 | given rtx_insn INSN, return true if it is a 64-bit multiply-accumulate | |
10651 | instruction and has a preceding memory instruction such that a NOP | |
10652 | should be inserted between them. */ | |
10653 | ||
75cf1494 KT |
10654 | bool |
10655 | aarch64_madd_needs_nop (rtx_insn* insn) | |
10656 | { | |
10657 | enum attr_type attr_type; | |
10658 | rtx_insn *prev; | |
10659 | rtx body; | |
10660 | ||
b32c1043 | 10661 | if (!TARGET_FIX_ERR_A53_835769) |
75cf1494 KT |
10662 | return false; |
10663 | ||
e322d6e3 | 10664 | if (!INSN_P (insn) || recog_memoized (insn) < 0) |
75cf1494 KT |
10665 | return false; |
10666 | ||
10667 | attr_type = get_attr_type (insn); | |
10668 | if (!is_madd_op (attr_type)) | |
10669 | return false; | |
10670 | ||
10671 | prev = aarch64_prev_real_insn (insn); | |
3fea1a75 KT |
10672 | /* aarch64_prev_real_insn can call recog_memoized on insns other than INSN. |
10673 | Restore recog state to INSN to avoid state corruption. */ | |
10674 | extract_constrain_insn_cached (insn); | |
10675 | ||
8baff86e | 10676 | if (!prev || !has_memory_op (prev)) |
75cf1494 KT |
10677 | return false; |
10678 | ||
10679 | body = single_set (prev); | |
10680 | ||
10681 | /* If the previous insn is a memory op and there is no dependency between | |
8baff86e KT |
10682 | it and the DImode madd, emit a NOP between them. If body is NULL then we |
10683 | have a complex memory operation, probably a load/store pair. | |
10684 | Be conservative for now and emit a NOP. */ | |
10685 | if (GET_MODE (recog_data.operand[0]) == DImode | |
10686 | && (!body || !dep_between_memop_and_curr (body))) | |
75cf1494 KT |
10687 | return true; |
10688 | ||
10689 | return false; | |
10690 | ||
10691 | } | |
10692 | ||
8baff86e KT |
10693 | |
10694 | /* Implement FINAL_PRESCAN_INSN. */ | |
10695 | ||
75cf1494 KT |
10696 | void |
10697 | aarch64_final_prescan_insn (rtx_insn *insn) | |
10698 | { | |
10699 | if (aarch64_madd_needs_nop (insn)) | |
10700 | fprintf (asm_out_file, "\tnop // between mem op and mult-accumulate\n"); | |
10701 | } | |
10702 | ||
10703 | ||
43e9d192 | 10704 | /* Return the equivalent letter for size. */ |
81c2dfb9 | 10705 | static char |
43e9d192 IB |
10706 | sizetochar (int size) |
10707 | { | |
10708 | switch (size) | |
10709 | { | |
10710 | case 64: return 'd'; | |
10711 | case 32: return 's'; | |
10712 | case 16: return 'h'; | |
10713 | case 8 : return 'b'; | |
10714 | default: gcc_unreachable (); | |
10715 | } | |
10716 | } | |
10717 | ||
3520f7cc JG |
10718 | /* Return true iff x is a uniform vector of floating-point |
10719 | constants, and the constant can be represented in | |
10720 | quarter-precision form. Note, as aarch64_float_const_representable | |
10721 | rejects both +0.0 and -0.0, we will also reject +0.0 and -0.0. */ | |
10722 | static bool | |
10723 | aarch64_vect_float_const_representable_p (rtx x) | |
10724 | { | |
92695fbb RS |
10725 | rtx elt; |
10726 | return (GET_MODE_CLASS (GET_MODE (x)) == MODE_VECTOR_FLOAT | |
10727 | && const_vec_duplicate_p (x, &elt) | |
10728 | && aarch64_float_const_representable_p (elt)); | |
3520f7cc JG |
10729 | } |
10730 | ||
d8edd899 | 10731 | /* Return true for valid and false for invalid. */ |
3ea63f60 | 10732 | bool |
ef4bddc2 | 10733 | aarch64_simd_valid_immediate (rtx op, machine_mode mode, bool inverse, |
48063b9d | 10734 | struct simd_immediate_info *info) |
43e9d192 IB |
10735 | { |
10736 | #define CHECK(STRIDE, ELSIZE, CLASS, TEST, SHIFT, NEG) \ | |
10737 | matches = 1; \ | |
10738 | for (i = 0; i < idx; i += (STRIDE)) \ | |
10739 | if (!(TEST)) \ | |
10740 | matches = 0; \ | |
10741 | if (matches) \ | |
10742 | { \ | |
10743 | immtype = (CLASS); \ | |
10744 | elsize = (ELSIZE); \ | |
43e9d192 IB |
10745 | eshift = (SHIFT); \ |
10746 | emvn = (NEG); \ | |
10747 | break; \ | |
10748 | } | |
10749 | ||
10750 | unsigned int i, elsize = 0, idx = 0, n_elts = CONST_VECTOR_NUNITS (op); | |
cb5ca315 | 10751 | unsigned int innersize = GET_MODE_UNIT_SIZE (mode); |
43e9d192 | 10752 | unsigned char bytes[16]; |
43e9d192 IB |
10753 | int immtype = -1, matches; |
10754 | unsigned int invmask = inverse ? 0xff : 0; | |
10755 | int eshift, emvn; | |
10756 | ||
43e9d192 | 10757 | if (GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) |
3520f7cc | 10758 | { |
81c2dfb9 IB |
10759 | if (! (aarch64_simd_imm_zero_p (op, mode) |
10760 | || aarch64_vect_float_const_representable_p (op))) | |
d8edd899 | 10761 | return false; |
3520f7cc | 10762 | |
48063b9d IB |
10763 | if (info) |
10764 | { | |
10765 | info->value = CONST_VECTOR_ELT (op, 0); | |
81c2dfb9 | 10766 | info->element_width = GET_MODE_BITSIZE (GET_MODE (info->value)); |
48063b9d IB |
10767 | info->mvn = false; |
10768 | info->shift = 0; | |
10769 | } | |
3520f7cc | 10770 | |
d8edd899 | 10771 | return true; |
3520f7cc | 10772 | } |
43e9d192 IB |
10773 | |
10774 | /* Splat vector constant out into a byte vector. */ | |
10775 | for (i = 0; i < n_elts; i++) | |
10776 | { | |
4b1e108c AL |
10777 | /* The vector is provided in gcc endian-neutral fashion. For aarch64_be, |
10778 | it must be laid out in the vector register in reverse order. */ | |
10779 | rtx el = CONST_VECTOR_ELT (op, BYTES_BIG_ENDIAN ? (n_elts - 1 - i) : i); | |
43e9d192 | 10780 | unsigned HOST_WIDE_INT elpart; |
43e9d192 | 10781 | |
ee78df47 KT |
10782 | gcc_assert (CONST_INT_P (el)); |
10783 | elpart = INTVAL (el); | |
10784 | ||
10785 | for (unsigned int byte = 0; byte < innersize; byte++) | |
10786 | { | |
10787 | bytes[idx++] = (elpart & 0xff) ^ invmask; | |
10788 | elpart >>= BITS_PER_UNIT; | |
10789 | } | |
43e9d192 | 10790 | |
43e9d192 IB |
10791 | } |
10792 | ||
10793 | /* Sanity check. */ | |
10794 | gcc_assert (idx == GET_MODE_SIZE (mode)); | |
10795 | ||
10796 | do | |
10797 | { | |
10798 | CHECK (4, 32, 0, bytes[i] == bytes[0] && bytes[i + 1] == 0 | |
10799 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 0, 0); | |
10800 | ||
10801 | CHECK (4, 32, 1, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
10802 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 8, 0); | |
10803 | ||
10804 | CHECK (4, 32, 2, bytes[i] == 0 && bytes[i + 1] == 0 | |
10805 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0, 16, 0); | |
10806 | ||
10807 | CHECK (4, 32, 3, bytes[i] == 0 && bytes[i + 1] == 0 | |
10808 | && bytes[i + 2] == 0 && bytes[i + 3] == bytes[3], 24, 0); | |
10809 | ||
10810 | CHECK (2, 16, 4, bytes[i] == bytes[0] && bytes[i + 1] == 0, 0, 0); | |
10811 | ||
10812 | CHECK (2, 16, 5, bytes[i] == 0 && bytes[i + 1] == bytes[1], 8, 0); | |
10813 | ||
10814 | CHECK (4, 32, 6, bytes[i] == bytes[0] && bytes[i + 1] == 0xff | |
10815 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 0, 1); | |
10816 | ||
10817 | CHECK (4, 32, 7, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
10818 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 8, 1); | |
10819 | ||
10820 | CHECK (4, 32, 8, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
10821 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff, 16, 1); | |
10822 | ||
10823 | CHECK (4, 32, 9, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
10824 | && bytes[i + 2] == 0xff && bytes[i + 3] == bytes[3], 24, 1); | |
10825 | ||
10826 | CHECK (2, 16, 10, bytes[i] == bytes[0] && bytes[i + 1] == 0xff, 0, 1); | |
10827 | ||
10828 | CHECK (2, 16, 11, bytes[i] == 0xff && bytes[i + 1] == bytes[1], 8, 1); | |
10829 | ||
10830 | CHECK (4, 32, 12, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
e4f0f84d | 10831 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 8, 0); |
43e9d192 IB |
10832 | |
10833 | CHECK (4, 32, 13, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
e4f0f84d | 10834 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 8, 1); |
43e9d192 IB |
10835 | |
10836 | CHECK (4, 32, 14, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
e4f0f84d | 10837 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0, 16, 0); |
43e9d192 IB |
10838 | |
10839 | CHECK (4, 32, 15, bytes[i] == 0 && bytes[i + 1] == 0 | |
e4f0f84d | 10840 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff, 16, 1); |
43e9d192 IB |
10841 | |
10842 | CHECK (1, 8, 16, bytes[i] == bytes[0], 0, 0); | |
10843 | ||
10844 | CHECK (1, 64, 17, (bytes[i] == 0 || bytes[i] == 0xff) | |
10845 | && bytes[i] == bytes[(i + 8) % idx], 0, 0); | |
10846 | } | |
10847 | while (0); | |
10848 | ||
e4f0f84d | 10849 | if (immtype == -1) |
d8edd899 | 10850 | return false; |
43e9d192 | 10851 | |
48063b9d | 10852 | if (info) |
43e9d192 | 10853 | { |
48063b9d | 10854 | info->element_width = elsize; |
48063b9d IB |
10855 | info->mvn = emvn != 0; |
10856 | info->shift = eshift; | |
10857 | ||
43e9d192 IB |
10858 | unsigned HOST_WIDE_INT imm = 0; |
10859 | ||
e4f0f84d TB |
10860 | if (immtype >= 12 && immtype <= 15) |
10861 | info->msl = true; | |
10862 | ||
43e9d192 IB |
10863 | /* Un-invert bytes of recognized vector, if necessary. */ |
10864 | if (invmask != 0) | |
10865 | for (i = 0; i < idx; i++) | |
10866 | bytes[i] ^= invmask; | |
10867 | ||
10868 | if (immtype == 17) | |
10869 | { | |
10870 | /* FIXME: Broken on 32-bit H_W_I hosts. */ | |
10871 | gcc_assert (sizeof (HOST_WIDE_INT) == 8); | |
10872 | ||
10873 | for (i = 0; i < 8; i++) | |
10874 | imm |= (unsigned HOST_WIDE_INT) (bytes[i] ? 0xff : 0) | |
10875 | << (i * BITS_PER_UNIT); | |
10876 | ||
43e9d192 | 10877 | |
48063b9d IB |
10878 | info->value = GEN_INT (imm); |
10879 | } | |
10880 | else | |
10881 | { | |
10882 | for (i = 0; i < elsize / BITS_PER_UNIT; i++) | |
10883 | imm |= (unsigned HOST_WIDE_INT) bytes[i] << (i * BITS_PER_UNIT); | |
43e9d192 IB |
10884 | |
10885 | /* Construct 'abcdefgh' because the assembler cannot handle | |
48063b9d IB |
10886 | generic constants. */ |
10887 | if (info->mvn) | |
43e9d192 | 10888 | imm = ~imm; |
48063b9d IB |
10889 | imm = (imm >> info->shift) & 0xff; |
10890 | info->value = GEN_INT (imm); | |
10891 | } | |
43e9d192 IB |
10892 | } |
10893 | ||
48063b9d | 10894 | return true; |
43e9d192 IB |
10895 | #undef CHECK |
10896 | } | |
10897 | ||
43e9d192 IB |
10898 | /* Check of immediate shift constants are within range. */ |
10899 | bool | |
ef4bddc2 | 10900 | aarch64_simd_shift_imm_p (rtx x, machine_mode mode, bool left) |
43e9d192 IB |
10901 | { |
10902 | int bit_width = GET_MODE_UNIT_SIZE (mode) * BITS_PER_UNIT; | |
10903 | if (left) | |
ddeabd3e | 10904 | return aarch64_const_vec_all_same_in_range_p (x, 0, bit_width - 1); |
43e9d192 | 10905 | else |
ddeabd3e | 10906 | return aarch64_const_vec_all_same_in_range_p (x, 1, bit_width); |
43e9d192 IB |
10907 | } |
10908 | ||
3520f7cc JG |
10909 | /* Return true if X is a uniform vector where all elements |
10910 | are either the floating-point constant 0.0 or the | |
10911 | integer constant 0. */ | |
43e9d192 | 10912 | bool |
ef4bddc2 | 10913 | aarch64_simd_imm_zero_p (rtx x, machine_mode mode) |
43e9d192 | 10914 | { |
3520f7cc | 10915 | return x == CONST0_RTX (mode); |
43e9d192 IB |
10916 | } |
10917 | ||
7325d85a KT |
10918 | |
10919 | /* Return the bitmask CONST_INT to select the bits required by a zero extract | |
10920 | operation of width WIDTH at bit position POS. */ | |
10921 | ||
10922 | rtx | |
10923 | aarch64_mask_from_zextract_ops (rtx width, rtx pos) | |
10924 | { | |
10925 | gcc_assert (CONST_INT_P (width)); | |
10926 | gcc_assert (CONST_INT_P (pos)); | |
10927 | ||
10928 | unsigned HOST_WIDE_INT mask | |
10929 | = ((unsigned HOST_WIDE_INT) 1 << UINTVAL (width)) - 1; | |
10930 | return GEN_INT (mask << UINTVAL (pos)); | |
10931 | } | |
10932 | ||
43e9d192 | 10933 | bool |
ef4bddc2 | 10934 | aarch64_simd_imm_scalar_p (rtx x, machine_mode mode ATTRIBUTE_UNUSED) |
43e9d192 IB |
10935 | { |
10936 | HOST_WIDE_INT imm = INTVAL (x); | |
10937 | int i; | |
10938 | ||
10939 | for (i = 0; i < 8; i++) | |
10940 | { | |
10941 | unsigned int byte = imm & 0xff; | |
10942 | if (byte != 0xff && byte != 0) | |
10943 | return false; | |
10944 | imm >>= 8; | |
10945 | } | |
10946 | ||
10947 | return true; | |
10948 | } | |
10949 | ||
83f8c414 | 10950 | bool |
a6e0bfa7 | 10951 | aarch64_mov_operand_p (rtx x, machine_mode mode) |
83f8c414 | 10952 | { |
83f8c414 CSS |
10953 | if (GET_CODE (x) == HIGH |
10954 | && aarch64_valid_symref (XEXP (x, 0), GET_MODE (XEXP (x, 0)))) | |
10955 | return true; | |
10956 | ||
82614948 | 10957 | if (CONST_INT_P (x)) |
83f8c414 CSS |
10958 | return true; |
10959 | ||
10960 | if (GET_CODE (x) == SYMBOL_REF && mode == DImode && CONSTANT_ADDRESS_P (x)) | |
10961 | return true; | |
10962 | ||
a6e0bfa7 | 10963 | return aarch64_classify_symbolic_expression (x) |
a5350ddc | 10964 | == SYMBOL_TINY_ABSOLUTE; |
83f8c414 CSS |
10965 | } |
10966 | ||
43e9d192 IB |
10967 | /* Return a const_int vector of VAL. */ |
10968 | rtx | |
ef4bddc2 | 10969 | aarch64_simd_gen_const_vector_dup (machine_mode mode, int val) |
43e9d192 IB |
10970 | { |
10971 | int nunits = GET_MODE_NUNITS (mode); | |
10972 | rtvec v = rtvec_alloc (nunits); | |
10973 | int i; | |
10974 | ||
10975 | for (i=0; i < nunits; i++) | |
10976 | RTVEC_ELT (v, i) = GEN_INT (val); | |
10977 | ||
10978 | return gen_rtx_CONST_VECTOR (mode, v); | |
10979 | } | |
10980 | ||
051d0e2f SN |
10981 | /* Check OP is a legal scalar immediate for the MOVI instruction. */ |
10982 | ||
10983 | bool | |
ef4bddc2 | 10984 | aarch64_simd_scalar_immediate_valid_for_move (rtx op, machine_mode mode) |
051d0e2f | 10985 | { |
ef4bddc2 | 10986 | machine_mode vmode; |
051d0e2f SN |
10987 | |
10988 | gcc_assert (!VECTOR_MODE_P (mode)); | |
10989 | vmode = aarch64_preferred_simd_mode (mode); | |
10990 | rtx op_v = aarch64_simd_gen_const_vector_dup (vmode, INTVAL (op)); | |
48063b9d | 10991 | return aarch64_simd_valid_immediate (op_v, vmode, false, NULL); |
051d0e2f SN |
10992 | } |
10993 | ||
988fa693 JG |
10994 | /* Construct and return a PARALLEL RTX vector with elements numbering the |
10995 | lanes of either the high (HIGH == TRUE) or low (HIGH == FALSE) half of | |
10996 | the vector - from the perspective of the architecture. This does not | |
10997 | line up with GCC's perspective on lane numbers, so we end up with | |
10998 | different masks depending on our target endian-ness. The diagram | |
10999 | below may help. We must draw the distinction when building masks | |
11000 | which select one half of the vector. An instruction selecting | |
11001 | architectural low-lanes for a big-endian target, must be described using | |
11002 | a mask selecting GCC high-lanes. | |
11003 | ||
11004 | Big-Endian Little-Endian | |
11005 | ||
11006 | GCC 0 1 2 3 3 2 1 0 | |
11007 | | x | x | x | x | | x | x | x | x | | |
11008 | Architecture 3 2 1 0 3 2 1 0 | |
11009 | ||
11010 | Low Mask: { 2, 3 } { 0, 1 } | |
11011 | High Mask: { 0, 1 } { 2, 3 } | |
11012 | */ | |
11013 | ||
43e9d192 | 11014 | rtx |
ef4bddc2 | 11015 | aarch64_simd_vect_par_cnst_half (machine_mode mode, bool high) |
43e9d192 IB |
11016 | { |
11017 | int nunits = GET_MODE_NUNITS (mode); | |
11018 | rtvec v = rtvec_alloc (nunits / 2); | |
988fa693 JG |
11019 | int high_base = nunits / 2; |
11020 | int low_base = 0; | |
11021 | int base; | |
43e9d192 IB |
11022 | rtx t1; |
11023 | int i; | |
11024 | ||
988fa693 JG |
11025 | if (BYTES_BIG_ENDIAN) |
11026 | base = high ? low_base : high_base; | |
11027 | else | |
11028 | base = high ? high_base : low_base; | |
11029 | ||
11030 | for (i = 0; i < nunits / 2; i++) | |
43e9d192 IB |
11031 | RTVEC_ELT (v, i) = GEN_INT (base + i); |
11032 | ||
11033 | t1 = gen_rtx_PARALLEL (mode, v); | |
11034 | return t1; | |
11035 | } | |
11036 | ||
988fa693 JG |
11037 | /* Check OP for validity as a PARALLEL RTX vector with elements |
11038 | numbering the lanes of either the high (HIGH == TRUE) or low lanes, | |
11039 | from the perspective of the architecture. See the diagram above | |
11040 | aarch64_simd_vect_par_cnst_half for more details. */ | |
11041 | ||
11042 | bool | |
ef4bddc2 | 11043 | aarch64_simd_check_vect_par_cnst_half (rtx op, machine_mode mode, |
988fa693 JG |
11044 | bool high) |
11045 | { | |
11046 | rtx ideal = aarch64_simd_vect_par_cnst_half (mode, high); | |
11047 | HOST_WIDE_INT count_op = XVECLEN (op, 0); | |
11048 | HOST_WIDE_INT count_ideal = XVECLEN (ideal, 0); | |
11049 | int i = 0; | |
11050 | ||
11051 | if (!VECTOR_MODE_P (mode)) | |
11052 | return false; | |
11053 | ||
11054 | if (count_op != count_ideal) | |
11055 | return false; | |
11056 | ||
11057 | for (i = 0; i < count_ideal; i++) | |
11058 | { | |
11059 | rtx elt_op = XVECEXP (op, 0, i); | |
11060 | rtx elt_ideal = XVECEXP (ideal, 0, i); | |
11061 | ||
4aa81c2e | 11062 | if (!CONST_INT_P (elt_op) |
988fa693 JG |
11063 | || INTVAL (elt_ideal) != INTVAL (elt_op)) |
11064 | return false; | |
11065 | } | |
11066 | return true; | |
11067 | } | |
11068 | ||
43e9d192 IB |
11069 | /* Bounds-check lanes. Ensure OPERAND lies between LOW (inclusive) and |
11070 | HIGH (exclusive). */ | |
11071 | void | |
46ed6024 CB |
11072 | aarch64_simd_lane_bounds (rtx operand, HOST_WIDE_INT low, HOST_WIDE_INT high, |
11073 | const_tree exp) | |
43e9d192 IB |
11074 | { |
11075 | HOST_WIDE_INT lane; | |
4aa81c2e | 11076 | gcc_assert (CONST_INT_P (operand)); |
43e9d192 IB |
11077 | lane = INTVAL (operand); |
11078 | ||
11079 | if (lane < low || lane >= high) | |
46ed6024 CB |
11080 | { |
11081 | if (exp) | |
cf0c27ef | 11082 | error ("%Klane %wd out of range %wd - %wd", exp, lane, low, high - 1); |
46ed6024 | 11083 | else |
cf0c27ef | 11084 | error ("lane %wd out of range %wd - %wd", lane, low, high - 1); |
46ed6024 | 11085 | } |
43e9d192 IB |
11086 | } |
11087 | ||
43e9d192 IB |
11088 | /* Return TRUE if OP is a valid vector addressing mode. */ |
11089 | bool | |
11090 | aarch64_simd_mem_operand_p (rtx op) | |
11091 | { | |
11092 | return MEM_P (op) && (GET_CODE (XEXP (op, 0)) == POST_INC | |
4aa81c2e | 11093 | || REG_P (XEXP (op, 0))); |
43e9d192 IB |
11094 | } |
11095 | ||
2d8c6dc1 AH |
11096 | /* Emit a register copy from operand to operand, taking care not to |
11097 | early-clobber source registers in the process. | |
43e9d192 | 11098 | |
2d8c6dc1 AH |
11099 | COUNT is the number of components into which the copy needs to be |
11100 | decomposed. */ | |
43e9d192 | 11101 | void |
2d8c6dc1 AH |
11102 | aarch64_simd_emit_reg_reg_move (rtx *operands, enum machine_mode mode, |
11103 | unsigned int count) | |
43e9d192 IB |
11104 | { |
11105 | unsigned int i; | |
2d8c6dc1 AH |
11106 | int rdest = REGNO (operands[0]); |
11107 | int rsrc = REGNO (operands[1]); | |
43e9d192 IB |
11108 | |
11109 | if (!reg_overlap_mentioned_p (operands[0], operands[1]) | |
2d8c6dc1 AH |
11110 | || rdest < rsrc) |
11111 | for (i = 0; i < count; i++) | |
11112 | emit_move_insn (gen_rtx_REG (mode, rdest + i), | |
11113 | gen_rtx_REG (mode, rsrc + i)); | |
43e9d192 | 11114 | else |
2d8c6dc1 AH |
11115 | for (i = 0; i < count; i++) |
11116 | emit_move_insn (gen_rtx_REG (mode, rdest + count - i - 1), | |
11117 | gen_rtx_REG (mode, rsrc + count - i - 1)); | |
43e9d192 IB |
11118 | } |
11119 | ||
668046d1 | 11120 | /* Compute and return the length of aarch64_simd_reglist<mode>, where <mode> is |
6ec0e5b9 | 11121 | one of VSTRUCT modes: OI, CI, or XI. */ |
668046d1 DS |
11122 | int |
11123 | aarch64_simd_attr_length_rglist (enum machine_mode mode) | |
11124 | { | |
11125 | return (GET_MODE_SIZE (mode) / UNITS_PER_VREG) * 4; | |
11126 | } | |
11127 | ||
db0253a4 TB |
11128 | /* Implement target hook TARGET_VECTOR_ALIGNMENT. The AAPCS64 sets the maximum |
11129 | alignment of a vector to 128 bits. */ | |
11130 | static HOST_WIDE_INT | |
11131 | aarch64_simd_vector_alignment (const_tree type) | |
11132 | { | |
9439e9a1 | 11133 | HOST_WIDE_INT align = tree_to_shwi (TYPE_SIZE (type)); |
db0253a4 TB |
11134 | return MIN (align, 128); |
11135 | } | |
11136 | ||
11137 | /* Implement target hook TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE. */ | |
11138 | static bool | |
11139 | aarch64_simd_vector_alignment_reachable (const_tree type, bool is_packed) | |
11140 | { | |
11141 | if (is_packed) | |
11142 | return false; | |
11143 | ||
11144 | /* We guarantee alignment for vectors up to 128-bits. */ | |
11145 | if (tree_int_cst_compare (TYPE_SIZE (type), | |
11146 | bitsize_int (BIGGEST_ALIGNMENT)) > 0) | |
11147 | return false; | |
11148 | ||
11149 | /* Vectors whose size is <= BIGGEST_ALIGNMENT are naturally aligned. */ | |
11150 | return true; | |
11151 | } | |
11152 | ||
4369c11e TB |
11153 | /* If VALS is a vector constant that can be loaded into a register |
11154 | using DUP, generate instructions to do so and return an RTX to | |
11155 | assign to the register. Otherwise return NULL_RTX. */ | |
11156 | static rtx | |
11157 | aarch64_simd_dup_constant (rtx vals) | |
11158 | { | |
ef4bddc2 RS |
11159 | machine_mode mode = GET_MODE (vals); |
11160 | machine_mode inner_mode = GET_MODE_INNER (mode); | |
4369c11e | 11161 | rtx x; |
4369c11e | 11162 | |
92695fbb | 11163 | if (!const_vec_duplicate_p (vals, &x)) |
4369c11e TB |
11164 | return NULL_RTX; |
11165 | ||
11166 | /* We can load this constant by using DUP and a constant in a | |
11167 | single ARM register. This will be cheaper than a vector | |
11168 | load. */ | |
92695fbb | 11169 | x = copy_to_mode_reg (inner_mode, x); |
4369c11e TB |
11170 | return gen_rtx_VEC_DUPLICATE (mode, x); |
11171 | } | |
11172 | ||
11173 | ||
11174 | /* Generate code to load VALS, which is a PARALLEL containing only | |
11175 | constants (for vec_init) or CONST_VECTOR, efficiently into a | |
11176 | register. Returns an RTX to copy into the register, or NULL_RTX | |
11177 | for a PARALLEL that can not be converted into a CONST_VECTOR. */ | |
1df3f464 | 11178 | static rtx |
4369c11e TB |
11179 | aarch64_simd_make_constant (rtx vals) |
11180 | { | |
ef4bddc2 | 11181 | machine_mode mode = GET_MODE (vals); |
4369c11e TB |
11182 | rtx const_dup; |
11183 | rtx const_vec = NULL_RTX; | |
11184 | int n_elts = GET_MODE_NUNITS (mode); | |
11185 | int n_const = 0; | |
11186 | int i; | |
11187 | ||
11188 | if (GET_CODE (vals) == CONST_VECTOR) | |
11189 | const_vec = vals; | |
11190 | else if (GET_CODE (vals) == PARALLEL) | |
11191 | { | |
11192 | /* A CONST_VECTOR must contain only CONST_INTs and | |
11193 | CONST_DOUBLEs, but CONSTANT_P allows more (e.g. SYMBOL_REF). | |
11194 | Only store valid constants in a CONST_VECTOR. */ | |
11195 | for (i = 0; i < n_elts; ++i) | |
11196 | { | |
11197 | rtx x = XVECEXP (vals, 0, i); | |
11198 | if (CONST_INT_P (x) || CONST_DOUBLE_P (x)) | |
11199 | n_const++; | |
11200 | } | |
11201 | if (n_const == n_elts) | |
11202 | const_vec = gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0)); | |
11203 | } | |
11204 | else | |
11205 | gcc_unreachable (); | |
11206 | ||
11207 | if (const_vec != NULL_RTX | |
48063b9d | 11208 | && aarch64_simd_valid_immediate (const_vec, mode, false, NULL)) |
4369c11e TB |
11209 | /* Load using MOVI/MVNI. */ |
11210 | return const_vec; | |
11211 | else if ((const_dup = aarch64_simd_dup_constant (vals)) != NULL_RTX) | |
11212 | /* Loaded using DUP. */ | |
11213 | return const_dup; | |
11214 | else if (const_vec != NULL_RTX) | |
11215 | /* Load from constant pool. We can not take advantage of single-cycle | |
11216 | LD1 because we need a PC-relative addressing mode. */ | |
11217 | return const_vec; | |
11218 | else | |
11219 | /* A PARALLEL containing something not valid inside CONST_VECTOR. | |
11220 | We can not construct an initializer. */ | |
11221 | return NULL_RTX; | |
11222 | } | |
11223 | ||
35a093b6 JG |
11224 | /* Expand a vector initialisation sequence, such that TARGET is |
11225 | initialised to contain VALS. */ | |
11226 | ||
4369c11e TB |
11227 | void |
11228 | aarch64_expand_vector_init (rtx target, rtx vals) | |
11229 | { | |
ef4bddc2 RS |
11230 | machine_mode mode = GET_MODE (target); |
11231 | machine_mode inner_mode = GET_MODE_INNER (mode); | |
35a093b6 | 11232 | /* The number of vector elements. */ |
4369c11e | 11233 | int n_elts = GET_MODE_NUNITS (mode); |
35a093b6 | 11234 | /* The number of vector elements which are not constant. */ |
8b66a2d4 AL |
11235 | int n_var = 0; |
11236 | rtx any_const = NULL_RTX; | |
35a093b6 JG |
11237 | /* The first element of vals. */ |
11238 | rtx v0 = XVECEXP (vals, 0, 0); | |
4369c11e | 11239 | bool all_same = true; |
4369c11e | 11240 | |
35a093b6 | 11241 | /* Count the number of variable elements to initialise. */ |
8b66a2d4 | 11242 | for (int i = 0; i < n_elts; ++i) |
4369c11e | 11243 | { |
8b66a2d4 | 11244 | rtx x = XVECEXP (vals, 0, i); |
35a093b6 | 11245 | if (!(CONST_INT_P (x) || CONST_DOUBLE_P (x))) |
8b66a2d4 AL |
11246 | ++n_var; |
11247 | else | |
11248 | any_const = x; | |
4369c11e | 11249 | |
35a093b6 | 11250 | all_same &= rtx_equal_p (x, v0); |
4369c11e TB |
11251 | } |
11252 | ||
35a093b6 JG |
11253 | /* No variable elements, hand off to aarch64_simd_make_constant which knows |
11254 | how best to handle this. */ | |
4369c11e TB |
11255 | if (n_var == 0) |
11256 | { | |
11257 | rtx constant = aarch64_simd_make_constant (vals); | |
11258 | if (constant != NULL_RTX) | |
11259 | { | |
11260 | emit_move_insn (target, constant); | |
11261 | return; | |
11262 | } | |
11263 | } | |
11264 | ||
11265 | /* Splat a single non-constant element if we can. */ | |
11266 | if (all_same) | |
11267 | { | |
35a093b6 | 11268 | rtx x = copy_to_mode_reg (inner_mode, v0); |
4369c11e TB |
11269 | aarch64_emit_move (target, gen_rtx_VEC_DUPLICATE (mode, x)); |
11270 | return; | |
11271 | } | |
11272 | ||
35a093b6 JG |
11273 | /* Initialise a vector which is part-variable. We want to first try |
11274 | to build those lanes which are constant in the most efficient way we | |
11275 | can. */ | |
11276 | if (n_var != n_elts) | |
4369c11e TB |
11277 | { |
11278 | rtx copy = copy_rtx (vals); | |
4369c11e | 11279 | |
8b66a2d4 AL |
11280 | /* Load constant part of vector. We really don't care what goes into the |
11281 | parts we will overwrite, but we're more likely to be able to load the | |
11282 | constant efficiently if it has fewer, larger, repeating parts | |
11283 | (see aarch64_simd_valid_immediate). */ | |
11284 | for (int i = 0; i < n_elts; i++) | |
11285 | { | |
11286 | rtx x = XVECEXP (vals, 0, i); | |
11287 | if (CONST_INT_P (x) || CONST_DOUBLE_P (x)) | |
11288 | continue; | |
11289 | rtx subst = any_const; | |
11290 | for (int bit = n_elts / 2; bit > 0; bit /= 2) | |
11291 | { | |
11292 | /* Look in the copied vector, as more elements are const. */ | |
11293 | rtx test = XVECEXP (copy, 0, i ^ bit); | |
11294 | if (CONST_INT_P (test) || CONST_DOUBLE_P (test)) | |
11295 | { | |
11296 | subst = test; | |
11297 | break; | |
11298 | } | |
11299 | } | |
11300 | XVECEXP (copy, 0, i) = subst; | |
11301 | } | |
4369c11e | 11302 | aarch64_expand_vector_init (target, copy); |
35a093b6 | 11303 | } |
4369c11e | 11304 | |
35a093b6 | 11305 | /* Insert the variable lanes directly. */ |
8b66a2d4 | 11306 | |
35a093b6 JG |
11307 | enum insn_code icode = optab_handler (vec_set_optab, mode); |
11308 | gcc_assert (icode != CODE_FOR_nothing); | |
4369c11e | 11309 | |
8b66a2d4 | 11310 | for (int i = 0; i < n_elts; i++) |
35a093b6 JG |
11311 | { |
11312 | rtx x = XVECEXP (vals, 0, i); | |
11313 | if (CONST_INT_P (x) || CONST_DOUBLE_P (x)) | |
11314 | continue; | |
11315 | x = copy_to_mode_reg (inner_mode, x); | |
11316 | emit_insn (GEN_FCN (icode) (target, x, GEN_INT (i))); | |
11317 | } | |
4369c11e TB |
11318 | } |
11319 | ||
43e9d192 | 11320 | static unsigned HOST_WIDE_INT |
ef4bddc2 | 11321 | aarch64_shift_truncation_mask (machine_mode mode) |
43e9d192 IB |
11322 | { |
11323 | return | |
ac59ad4e KT |
11324 | (!SHIFT_COUNT_TRUNCATED |
11325 | || aarch64_vector_mode_supported_p (mode) | |
43e9d192 IB |
11326 | || aarch64_vect_struct_mode_p (mode)) ? 0 : (GET_MODE_BITSIZE (mode) - 1); |
11327 | } | |
11328 | ||
43e9d192 IB |
11329 | /* Select a format to encode pointers in exception handling data. */ |
11330 | int | |
11331 | aarch64_asm_preferred_eh_data_format (int code ATTRIBUTE_UNUSED, int global) | |
11332 | { | |
11333 | int type; | |
11334 | switch (aarch64_cmodel) | |
11335 | { | |
11336 | case AARCH64_CMODEL_TINY: | |
11337 | case AARCH64_CMODEL_TINY_PIC: | |
11338 | case AARCH64_CMODEL_SMALL: | |
11339 | case AARCH64_CMODEL_SMALL_PIC: | |
1b1e81f8 | 11340 | case AARCH64_CMODEL_SMALL_SPIC: |
43e9d192 IB |
11341 | /* text+got+data < 4Gb. 4-byte signed relocs are sufficient |
11342 | for everything. */ | |
11343 | type = DW_EH_PE_sdata4; | |
11344 | break; | |
11345 | default: | |
11346 | /* No assumptions here. 8-byte relocs required. */ | |
11347 | type = DW_EH_PE_sdata8; | |
11348 | break; | |
11349 | } | |
11350 | return (global ? DW_EH_PE_indirect : 0) | DW_EH_PE_pcrel | type; | |
11351 | } | |
11352 | ||
e1c1ecb0 KT |
11353 | /* The last .arch and .tune assembly strings that we printed. */ |
11354 | static std::string aarch64_last_printed_arch_string; | |
11355 | static std::string aarch64_last_printed_tune_string; | |
11356 | ||
361fb3ee KT |
11357 | /* Implement ASM_DECLARE_FUNCTION_NAME. Output the ISA features used |
11358 | by the function fndecl. */ | |
11359 | ||
11360 | void | |
11361 | aarch64_declare_function_name (FILE *stream, const char* name, | |
11362 | tree fndecl) | |
11363 | { | |
11364 | tree target_parts = DECL_FUNCTION_SPECIFIC_TARGET (fndecl); | |
11365 | ||
11366 | struct cl_target_option *targ_options; | |
11367 | if (target_parts) | |
11368 | targ_options = TREE_TARGET_OPTION (target_parts); | |
11369 | else | |
11370 | targ_options = TREE_TARGET_OPTION (target_option_current_node); | |
11371 | gcc_assert (targ_options); | |
11372 | ||
11373 | const struct processor *this_arch | |
11374 | = aarch64_get_arch (targ_options->x_explicit_arch); | |
11375 | ||
054b4005 JG |
11376 | unsigned long isa_flags = targ_options->x_aarch64_isa_flags; |
11377 | std::string extension | |
04a99ebe JG |
11378 | = aarch64_get_extension_string_for_isa_flags (isa_flags, |
11379 | this_arch->flags); | |
e1c1ecb0 KT |
11380 | /* Only update the assembler .arch string if it is distinct from the last |
11381 | such string we printed. */ | |
11382 | std::string to_print = this_arch->name + extension; | |
11383 | if (to_print != aarch64_last_printed_arch_string) | |
11384 | { | |
11385 | asm_fprintf (asm_out_file, "\t.arch %s\n", to_print.c_str ()); | |
11386 | aarch64_last_printed_arch_string = to_print; | |
11387 | } | |
361fb3ee KT |
11388 | |
11389 | /* Print the cpu name we're tuning for in the comments, might be | |
e1c1ecb0 KT |
11390 | useful to readers of the generated asm. Do it only when it changes |
11391 | from function to function and verbose assembly is requested. */ | |
361fb3ee KT |
11392 | const struct processor *this_tune |
11393 | = aarch64_get_tune_cpu (targ_options->x_explicit_tune_core); | |
11394 | ||
e1c1ecb0 KT |
11395 | if (flag_debug_asm && aarch64_last_printed_tune_string != this_tune->name) |
11396 | { | |
11397 | asm_fprintf (asm_out_file, "\t" ASM_COMMENT_START ".tune %s\n", | |
11398 | this_tune->name); | |
11399 | aarch64_last_printed_tune_string = this_tune->name; | |
11400 | } | |
361fb3ee KT |
11401 | |
11402 | /* Don't forget the type directive for ELF. */ | |
11403 | ASM_OUTPUT_TYPE_DIRECTIVE (stream, name, "function"); | |
11404 | ASM_OUTPUT_LABEL (stream, name); | |
11405 | } | |
11406 | ||
e1c1ecb0 KT |
11407 | /* Implements TARGET_ASM_FILE_START. Output the assembly header. */ |
11408 | ||
11409 | static void | |
11410 | aarch64_start_file (void) | |
11411 | { | |
11412 | struct cl_target_option *default_options | |
11413 | = TREE_TARGET_OPTION (target_option_default_node); | |
11414 | ||
11415 | const struct processor *default_arch | |
11416 | = aarch64_get_arch (default_options->x_explicit_arch); | |
11417 | unsigned long default_isa_flags = default_options->x_aarch64_isa_flags; | |
11418 | std::string extension | |
04a99ebe JG |
11419 | = aarch64_get_extension_string_for_isa_flags (default_isa_flags, |
11420 | default_arch->flags); | |
e1c1ecb0 KT |
11421 | |
11422 | aarch64_last_printed_arch_string = default_arch->name + extension; | |
11423 | aarch64_last_printed_tune_string = ""; | |
11424 | asm_fprintf (asm_out_file, "\t.arch %s\n", | |
11425 | aarch64_last_printed_arch_string.c_str ()); | |
11426 | ||
11427 | default_file_start (); | |
11428 | } | |
11429 | ||
0462169c SN |
11430 | /* Emit load exclusive. */ |
11431 | ||
11432 | static void | |
ef4bddc2 | 11433 | aarch64_emit_load_exclusive (machine_mode mode, rtx rval, |
0462169c SN |
11434 | rtx mem, rtx model_rtx) |
11435 | { | |
11436 | rtx (*gen) (rtx, rtx, rtx); | |
11437 | ||
11438 | switch (mode) | |
11439 | { | |
11440 | case QImode: gen = gen_aarch64_load_exclusiveqi; break; | |
11441 | case HImode: gen = gen_aarch64_load_exclusivehi; break; | |
11442 | case SImode: gen = gen_aarch64_load_exclusivesi; break; | |
11443 | case DImode: gen = gen_aarch64_load_exclusivedi; break; | |
11444 | default: | |
11445 | gcc_unreachable (); | |
11446 | } | |
11447 | ||
11448 | emit_insn (gen (rval, mem, model_rtx)); | |
11449 | } | |
11450 | ||
11451 | /* Emit store exclusive. */ | |
11452 | ||
11453 | static void | |
ef4bddc2 | 11454 | aarch64_emit_store_exclusive (machine_mode mode, rtx bval, |
0462169c SN |
11455 | rtx rval, rtx mem, rtx model_rtx) |
11456 | { | |
11457 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
11458 | ||
11459 | switch (mode) | |
11460 | { | |
11461 | case QImode: gen = gen_aarch64_store_exclusiveqi; break; | |
11462 | case HImode: gen = gen_aarch64_store_exclusivehi; break; | |
11463 | case SImode: gen = gen_aarch64_store_exclusivesi; break; | |
11464 | case DImode: gen = gen_aarch64_store_exclusivedi; break; | |
11465 | default: | |
11466 | gcc_unreachable (); | |
11467 | } | |
11468 | ||
11469 | emit_insn (gen (bval, rval, mem, model_rtx)); | |
11470 | } | |
11471 | ||
11472 | /* Mark the previous jump instruction as unlikely. */ | |
11473 | ||
11474 | static void | |
11475 | aarch64_emit_unlikely_jump (rtx insn) | |
11476 | { | |
e5af9ddd | 11477 | int very_unlikely = REG_BR_PROB_BASE / 100 - 1; |
0462169c | 11478 | |
f370536c TS |
11479 | rtx_insn *jump = emit_jump_insn (insn); |
11480 | add_int_reg_note (jump, REG_BR_PROB, very_unlikely); | |
0462169c SN |
11481 | } |
11482 | ||
11483 | /* Expand a compare and swap pattern. */ | |
11484 | ||
11485 | void | |
11486 | aarch64_expand_compare_and_swap (rtx operands[]) | |
11487 | { | |
11488 | rtx bval, rval, mem, oldval, newval, is_weak, mod_s, mod_f, x; | |
ef4bddc2 | 11489 | machine_mode mode, cmp_mode; |
b0770c0f MW |
11490 | typedef rtx (*gen_cas_fn) (rtx, rtx, rtx, rtx, rtx, rtx, rtx); |
11491 | int idx; | |
11492 | gen_cas_fn gen; | |
11493 | const gen_cas_fn split_cas[] = | |
11494 | { | |
11495 | gen_aarch64_compare_and_swapqi, | |
11496 | gen_aarch64_compare_and_swaphi, | |
11497 | gen_aarch64_compare_and_swapsi, | |
11498 | gen_aarch64_compare_and_swapdi | |
11499 | }; | |
11500 | const gen_cas_fn atomic_cas[] = | |
11501 | { | |
11502 | gen_aarch64_compare_and_swapqi_lse, | |
11503 | gen_aarch64_compare_and_swaphi_lse, | |
11504 | gen_aarch64_compare_and_swapsi_lse, | |
11505 | gen_aarch64_compare_and_swapdi_lse | |
11506 | }; | |
0462169c SN |
11507 | |
11508 | bval = operands[0]; | |
11509 | rval = operands[1]; | |
11510 | mem = operands[2]; | |
11511 | oldval = operands[3]; | |
11512 | newval = operands[4]; | |
11513 | is_weak = operands[5]; | |
11514 | mod_s = operands[6]; | |
11515 | mod_f = operands[7]; | |
11516 | mode = GET_MODE (mem); | |
11517 | cmp_mode = mode; | |
11518 | ||
11519 | /* Normally the succ memory model must be stronger than fail, but in the | |
11520 | unlikely event of fail being ACQUIRE and succ being RELEASE we need to | |
11521 | promote succ to ACQ_REL so that we don't lose the acquire semantics. */ | |
11522 | ||
46b35980 AM |
11523 | if (is_mm_acquire (memmodel_from_int (INTVAL (mod_f))) |
11524 | && is_mm_release (memmodel_from_int (INTVAL (mod_s)))) | |
0462169c SN |
11525 | mod_s = GEN_INT (MEMMODEL_ACQ_REL); |
11526 | ||
11527 | switch (mode) | |
11528 | { | |
11529 | case QImode: | |
11530 | case HImode: | |
11531 | /* For short modes, we're going to perform the comparison in SImode, | |
11532 | so do the zero-extension now. */ | |
11533 | cmp_mode = SImode; | |
11534 | rval = gen_reg_rtx (SImode); | |
11535 | oldval = convert_modes (SImode, mode, oldval, true); | |
11536 | /* Fall through. */ | |
11537 | ||
11538 | case SImode: | |
11539 | case DImode: | |
11540 | /* Force the value into a register if needed. */ | |
11541 | if (!aarch64_plus_operand (oldval, mode)) | |
11542 | oldval = force_reg (cmp_mode, oldval); | |
11543 | break; | |
11544 | ||
11545 | default: | |
11546 | gcc_unreachable (); | |
11547 | } | |
11548 | ||
11549 | switch (mode) | |
11550 | { | |
b0770c0f MW |
11551 | case QImode: idx = 0; break; |
11552 | case HImode: idx = 1; break; | |
11553 | case SImode: idx = 2; break; | |
11554 | case DImode: idx = 3; break; | |
0462169c SN |
11555 | default: |
11556 | gcc_unreachable (); | |
11557 | } | |
b0770c0f MW |
11558 | if (TARGET_LSE) |
11559 | gen = atomic_cas[idx]; | |
11560 | else | |
11561 | gen = split_cas[idx]; | |
0462169c SN |
11562 | |
11563 | emit_insn (gen (rval, mem, oldval, newval, is_weak, mod_s, mod_f)); | |
11564 | ||
11565 | if (mode == QImode || mode == HImode) | |
11566 | emit_move_insn (operands[1], gen_lowpart (mode, rval)); | |
11567 | ||
11568 | x = gen_rtx_REG (CCmode, CC_REGNUM); | |
11569 | x = gen_rtx_EQ (SImode, x, const0_rtx); | |
f7df4a84 | 11570 | emit_insn (gen_rtx_SET (bval, x)); |
0462169c SN |
11571 | } |
11572 | ||
641c2f8b MW |
11573 | /* Test whether the target supports using a atomic load-operate instruction. |
11574 | CODE is the operation and AFTER is TRUE if the data in memory after the | |
11575 | operation should be returned and FALSE if the data before the operation | |
11576 | should be returned. Returns FALSE if the operation isn't supported by the | |
11577 | architecture. */ | |
11578 | ||
11579 | bool | |
11580 | aarch64_atomic_ldop_supported_p (enum rtx_code code) | |
11581 | { | |
11582 | if (!TARGET_LSE) | |
11583 | return false; | |
11584 | ||
11585 | switch (code) | |
11586 | { | |
11587 | case SET: | |
11588 | case AND: | |
11589 | case IOR: | |
11590 | case XOR: | |
11591 | case MINUS: | |
11592 | case PLUS: | |
11593 | return true; | |
11594 | default: | |
11595 | return false; | |
11596 | } | |
11597 | } | |
11598 | ||
f70fb3b6 MW |
11599 | /* Emit a barrier, that is appropriate for memory model MODEL, at the end of a |
11600 | sequence implementing an atomic operation. */ | |
11601 | ||
11602 | static void | |
11603 | aarch64_emit_post_barrier (enum memmodel model) | |
11604 | { | |
11605 | const enum memmodel base_model = memmodel_base (model); | |
11606 | ||
11607 | if (is_mm_sync (model) | |
11608 | && (base_model == MEMMODEL_ACQUIRE | |
11609 | || base_model == MEMMODEL_ACQ_REL | |
11610 | || base_model == MEMMODEL_SEQ_CST)) | |
11611 | { | |
11612 | emit_insn (gen_mem_thread_fence (GEN_INT (MEMMODEL_SEQ_CST))); | |
11613 | } | |
11614 | } | |
11615 | ||
b0770c0f MW |
11616 | /* Emit an atomic compare-and-swap operation. RVAL is the destination register |
11617 | for the data in memory. EXPECTED is the value expected to be in memory. | |
11618 | DESIRED is the value to store to memory. MEM is the memory location. MODEL | |
11619 | is the memory ordering to use. */ | |
11620 | ||
11621 | void | |
11622 | aarch64_gen_atomic_cas (rtx rval, rtx mem, | |
11623 | rtx expected, rtx desired, | |
11624 | rtx model) | |
11625 | { | |
11626 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
11627 | machine_mode mode; | |
11628 | ||
11629 | mode = GET_MODE (mem); | |
11630 | ||
11631 | switch (mode) | |
11632 | { | |
11633 | case QImode: gen = gen_aarch64_atomic_casqi; break; | |
11634 | case HImode: gen = gen_aarch64_atomic_cashi; break; | |
11635 | case SImode: gen = gen_aarch64_atomic_cassi; break; | |
11636 | case DImode: gen = gen_aarch64_atomic_casdi; break; | |
11637 | default: | |
11638 | gcc_unreachable (); | |
11639 | } | |
11640 | ||
11641 | /* Move the expected value into the CAS destination register. */ | |
11642 | emit_insn (gen_rtx_SET (rval, expected)); | |
11643 | ||
11644 | /* Emit the CAS. */ | |
11645 | emit_insn (gen (rval, mem, desired, model)); | |
11646 | ||
11647 | /* Compare the expected value with the value loaded by the CAS, to establish | |
11648 | whether the swap was made. */ | |
11649 | aarch64_gen_compare_reg (EQ, rval, expected); | |
11650 | } | |
11651 | ||
0462169c SN |
11652 | /* Split a compare and swap pattern. */ |
11653 | ||
11654 | void | |
11655 | aarch64_split_compare_and_swap (rtx operands[]) | |
11656 | { | |
11657 | rtx rval, mem, oldval, newval, scratch; | |
ef4bddc2 | 11658 | machine_mode mode; |
0462169c | 11659 | bool is_weak; |
5d8a22a5 DM |
11660 | rtx_code_label *label1, *label2; |
11661 | rtx x, cond; | |
ab876106 MW |
11662 | enum memmodel model; |
11663 | rtx model_rtx; | |
0462169c SN |
11664 | |
11665 | rval = operands[0]; | |
11666 | mem = operands[1]; | |
11667 | oldval = operands[2]; | |
11668 | newval = operands[3]; | |
11669 | is_weak = (operands[4] != const0_rtx); | |
ab876106 | 11670 | model_rtx = operands[5]; |
0462169c SN |
11671 | scratch = operands[7]; |
11672 | mode = GET_MODE (mem); | |
ab876106 | 11673 | model = memmodel_from_int (INTVAL (model_rtx)); |
0462169c | 11674 | |
5d8a22a5 | 11675 | label1 = NULL; |
0462169c SN |
11676 | if (!is_weak) |
11677 | { | |
11678 | label1 = gen_label_rtx (); | |
11679 | emit_label (label1); | |
11680 | } | |
11681 | label2 = gen_label_rtx (); | |
11682 | ||
ab876106 MW |
11683 | /* The initial load can be relaxed for a __sync operation since a final |
11684 | barrier will be emitted to stop code hoisting. */ | |
11685 | if (is_mm_sync (model)) | |
11686 | aarch64_emit_load_exclusive (mode, rval, mem, | |
11687 | GEN_INT (MEMMODEL_RELAXED)); | |
11688 | else | |
11689 | aarch64_emit_load_exclusive (mode, rval, mem, model_rtx); | |
0462169c SN |
11690 | |
11691 | cond = aarch64_gen_compare_reg (NE, rval, oldval); | |
11692 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
11693 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
11694 | gen_rtx_LABEL_REF (Pmode, label2), pc_rtx); | |
f7df4a84 | 11695 | aarch64_emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); |
0462169c | 11696 | |
ab876106 | 11697 | aarch64_emit_store_exclusive (mode, scratch, mem, newval, model_rtx); |
0462169c SN |
11698 | |
11699 | if (!is_weak) | |
11700 | { | |
11701 | x = gen_rtx_NE (VOIDmode, scratch, const0_rtx); | |
11702 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
11703 | gen_rtx_LABEL_REF (Pmode, label1), pc_rtx); | |
f7df4a84 | 11704 | aarch64_emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); |
0462169c SN |
11705 | } |
11706 | else | |
11707 | { | |
11708 | cond = gen_rtx_REG (CCmode, CC_REGNUM); | |
11709 | x = gen_rtx_COMPARE (CCmode, scratch, const0_rtx); | |
f7df4a84 | 11710 | emit_insn (gen_rtx_SET (cond, x)); |
0462169c SN |
11711 | } |
11712 | ||
11713 | emit_label (label2); | |
ab876106 MW |
11714 | |
11715 | /* Emit any final barrier needed for a __sync operation. */ | |
11716 | if (is_mm_sync (model)) | |
11717 | aarch64_emit_post_barrier (model); | |
0462169c SN |
11718 | } |
11719 | ||
68729b06 MW |
11720 | /* Emit a BIC instruction. */ |
11721 | ||
11722 | static void | |
11723 | aarch64_emit_bic (machine_mode mode, rtx dst, rtx s1, rtx s2, int shift) | |
11724 | { | |
11725 | rtx shift_rtx = GEN_INT (shift); | |
11726 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
11727 | ||
11728 | switch (mode) | |
11729 | { | |
11730 | case SImode: gen = gen_and_one_cmpl_lshrsi3; break; | |
11731 | case DImode: gen = gen_and_one_cmpl_lshrdi3; break; | |
11732 | default: | |
11733 | gcc_unreachable (); | |
11734 | } | |
11735 | ||
11736 | emit_insn (gen (dst, s2, shift_rtx, s1)); | |
11737 | } | |
11738 | ||
9cd7b720 MW |
11739 | /* Emit an atomic swap. */ |
11740 | ||
11741 | static void | |
11742 | aarch64_emit_atomic_swap (machine_mode mode, rtx dst, rtx value, | |
11743 | rtx mem, rtx model) | |
11744 | { | |
11745 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
11746 | ||
11747 | switch (mode) | |
11748 | { | |
11749 | case QImode: gen = gen_aarch64_atomic_swpqi; break; | |
11750 | case HImode: gen = gen_aarch64_atomic_swphi; break; | |
11751 | case SImode: gen = gen_aarch64_atomic_swpsi; break; | |
11752 | case DImode: gen = gen_aarch64_atomic_swpdi; break; | |
11753 | default: | |
11754 | gcc_unreachable (); | |
11755 | } | |
11756 | ||
11757 | emit_insn (gen (dst, mem, value, model)); | |
11758 | } | |
11759 | ||
641c2f8b MW |
11760 | /* Operations supported by aarch64_emit_atomic_load_op. */ |
11761 | ||
11762 | enum aarch64_atomic_load_op_code | |
11763 | { | |
11764 | AARCH64_LDOP_PLUS, /* A + B */ | |
11765 | AARCH64_LDOP_XOR, /* A ^ B */ | |
11766 | AARCH64_LDOP_OR, /* A | B */ | |
11767 | AARCH64_LDOP_BIC /* A & ~B */ | |
11768 | }; | |
11769 | ||
11770 | /* Emit an atomic load-operate. */ | |
11771 | ||
11772 | static void | |
11773 | aarch64_emit_atomic_load_op (enum aarch64_atomic_load_op_code code, | |
11774 | machine_mode mode, rtx dst, rtx src, | |
11775 | rtx mem, rtx model) | |
11776 | { | |
11777 | typedef rtx (*aarch64_atomic_load_op_fn) (rtx, rtx, rtx, rtx); | |
11778 | const aarch64_atomic_load_op_fn plus[] = | |
11779 | { | |
11780 | gen_aarch64_atomic_loadaddqi, | |
11781 | gen_aarch64_atomic_loadaddhi, | |
11782 | gen_aarch64_atomic_loadaddsi, | |
11783 | gen_aarch64_atomic_loadadddi | |
11784 | }; | |
11785 | const aarch64_atomic_load_op_fn eor[] = | |
11786 | { | |
11787 | gen_aarch64_atomic_loadeorqi, | |
11788 | gen_aarch64_atomic_loadeorhi, | |
11789 | gen_aarch64_atomic_loadeorsi, | |
11790 | gen_aarch64_atomic_loadeordi | |
11791 | }; | |
11792 | const aarch64_atomic_load_op_fn ior[] = | |
11793 | { | |
11794 | gen_aarch64_atomic_loadsetqi, | |
11795 | gen_aarch64_atomic_loadsethi, | |
11796 | gen_aarch64_atomic_loadsetsi, | |
11797 | gen_aarch64_atomic_loadsetdi | |
11798 | }; | |
11799 | const aarch64_atomic_load_op_fn bic[] = | |
11800 | { | |
11801 | gen_aarch64_atomic_loadclrqi, | |
11802 | gen_aarch64_atomic_loadclrhi, | |
11803 | gen_aarch64_atomic_loadclrsi, | |
11804 | gen_aarch64_atomic_loadclrdi | |
11805 | }; | |
11806 | aarch64_atomic_load_op_fn gen; | |
11807 | int idx = 0; | |
11808 | ||
11809 | switch (mode) | |
11810 | { | |
11811 | case QImode: idx = 0; break; | |
11812 | case HImode: idx = 1; break; | |
11813 | case SImode: idx = 2; break; | |
11814 | case DImode: idx = 3; break; | |
11815 | default: | |
11816 | gcc_unreachable (); | |
11817 | } | |
11818 | ||
11819 | switch (code) | |
11820 | { | |
11821 | case AARCH64_LDOP_PLUS: gen = plus[idx]; break; | |
11822 | case AARCH64_LDOP_XOR: gen = eor[idx]; break; | |
11823 | case AARCH64_LDOP_OR: gen = ior[idx]; break; | |
11824 | case AARCH64_LDOP_BIC: gen = bic[idx]; break; | |
11825 | default: | |
11826 | gcc_unreachable (); | |
11827 | } | |
11828 | ||
11829 | emit_insn (gen (dst, mem, src, model)); | |
11830 | } | |
11831 | ||
11832 | /* Emit an atomic load+operate. CODE is the operation. OUT_DATA is the | |
68729b06 MW |
11833 | location to store the data read from memory. OUT_RESULT is the location to |
11834 | store the result of the operation. MEM is the memory location to read and | |
11835 | modify. MODEL_RTX is the memory ordering to use. VALUE is the second | |
11836 | operand for the operation. Either OUT_DATA or OUT_RESULT, but not both, can | |
11837 | be NULL. */ | |
9cd7b720 MW |
11838 | |
11839 | void | |
68729b06 | 11840 | aarch64_gen_atomic_ldop (enum rtx_code code, rtx out_data, rtx out_result, |
9cd7b720 MW |
11841 | rtx mem, rtx value, rtx model_rtx) |
11842 | { | |
11843 | machine_mode mode = GET_MODE (mem); | |
641c2f8b MW |
11844 | machine_mode wmode = (mode == DImode ? DImode : SImode); |
11845 | const bool short_mode = (mode < SImode); | |
11846 | aarch64_atomic_load_op_code ldop_code; | |
11847 | rtx src; | |
11848 | rtx x; | |
11849 | ||
11850 | if (out_data) | |
11851 | out_data = gen_lowpart (mode, out_data); | |
9cd7b720 | 11852 | |
68729b06 MW |
11853 | if (out_result) |
11854 | out_result = gen_lowpart (mode, out_result); | |
11855 | ||
641c2f8b MW |
11856 | /* Make sure the value is in a register, putting it into a destination |
11857 | register if it needs to be manipulated. */ | |
11858 | if (!register_operand (value, mode) | |
11859 | || code == AND || code == MINUS) | |
11860 | { | |
68729b06 | 11861 | src = out_result ? out_result : out_data; |
641c2f8b MW |
11862 | emit_move_insn (src, gen_lowpart (mode, value)); |
11863 | } | |
11864 | else | |
11865 | src = value; | |
11866 | gcc_assert (register_operand (src, mode)); | |
9cd7b720 | 11867 | |
641c2f8b MW |
11868 | /* Preprocess the data for the operation as necessary. If the operation is |
11869 | a SET then emit a swap instruction and finish. */ | |
9cd7b720 MW |
11870 | switch (code) |
11871 | { | |
11872 | case SET: | |
641c2f8b | 11873 | aarch64_emit_atomic_swap (mode, out_data, src, mem, model_rtx); |
9cd7b720 MW |
11874 | return; |
11875 | ||
641c2f8b MW |
11876 | case MINUS: |
11877 | /* Negate the value and treat it as a PLUS. */ | |
11878 | { | |
11879 | rtx neg_src; | |
11880 | ||
11881 | /* Resize the value if necessary. */ | |
11882 | if (short_mode) | |
11883 | src = gen_lowpart (wmode, src); | |
11884 | ||
11885 | neg_src = gen_rtx_NEG (wmode, src); | |
11886 | emit_insn (gen_rtx_SET (src, neg_src)); | |
11887 | ||
11888 | if (short_mode) | |
11889 | src = gen_lowpart (mode, src); | |
11890 | } | |
11891 | /* Fall-through. */ | |
11892 | case PLUS: | |
11893 | ldop_code = AARCH64_LDOP_PLUS; | |
11894 | break; | |
11895 | ||
11896 | case IOR: | |
11897 | ldop_code = AARCH64_LDOP_OR; | |
11898 | break; | |
11899 | ||
11900 | case XOR: | |
11901 | ldop_code = AARCH64_LDOP_XOR; | |
11902 | break; | |
11903 | ||
11904 | case AND: | |
11905 | { | |
11906 | rtx not_src; | |
11907 | ||
11908 | /* Resize the value if necessary. */ | |
11909 | if (short_mode) | |
11910 | src = gen_lowpart (wmode, src); | |
11911 | ||
11912 | not_src = gen_rtx_NOT (wmode, src); | |
11913 | emit_insn (gen_rtx_SET (src, not_src)); | |
11914 | ||
11915 | if (short_mode) | |
11916 | src = gen_lowpart (mode, src); | |
11917 | } | |
11918 | ldop_code = AARCH64_LDOP_BIC; | |
11919 | break; | |
11920 | ||
9cd7b720 MW |
11921 | default: |
11922 | /* The operation can't be done with atomic instructions. */ | |
11923 | gcc_unreachable (); | |
11924 | } | |
641c2f8b MW |
11925 | |
11926 | aarch64_emit_atomic_load_op (ldop_code, mode, out_data, src, mem, model_rtx); | |
68729b06 MW |
11927 | |
11928 | /* If necessary, calculate the data in memory after the update by redoing the | |
11929 | operation from values in registers. */ | |
11930 | if (!out_result) | |
11931 | return; | |
11932 | ||
11933 | if (short_mode) | |
11934 | { | |
11935 | src = gen_lowpart (wmode, src); | |
11936 | out_data = gen_lowpart (wmode, out_data); | |
11937 | out_result = gen_lowpart (wmode, out_result); | |
11938 | } | |
11939 | ||
11940 | x = NULL_RTX; | |
11941 | ||
11942 | switch (code) | |
11943 | { | |
11944 | case MINUS: | |
11945 | case PLUS: | |
11946 | x = gen_rtx_PLUS (wmode, out_data, src); | |
11947 | break; | |
11948 | case IOR: | |
11949 | x = gen_rtx_IOR (wmode, out_data, src); | |
11950 | break; | |
11951 | case XOR: | |
11952 | x = gen_rtx_XOR (wmode, out_data, src); | |
11953 | break; | |
11954 | case AND: | |
11955 | aarch64_emit_bic (wmode, out_result, out_data, src, 0); | |
11956 | return; | |
11957 | default: | |
11958 | gcc_unreachable (); | |
11959 | } | |
11960 | ||
11961 | emit_set_insn (out_result, x); | |
11962 | ||
11963 | return; | |
9cd7b720 MW |
11964 | } |
11965 | ||
0462169c SN |
11966 | /* Split an atomic operation. */ |
11967 | ||
11968 | void | |
11969 | aarch64_split_atomic_op (enum rtx_code code, rtx old_out, rtx new_out, rtx mem, | |
9cd7b720 | 11970 | rtx value, rtx model_rtx, rtx cond) |
0462169c | 11971 | { |
ef4bddc2 RS |
11972 | machine_mode mode = GET_MODE (mem); |
11973 | machine_mode wmode = (mode == DImode ? DImode : SImode); | |
f70fb3b6 MW |
11974 | const enum memmodel model = memmodel_from_int (INTVAL (model_rtx)); |
11975 | const bool is_sync = is_mm_sync (model); | |
5d8a22a5 DM |
11976 | rtx_code_label *label; |
11977 | rtx x; | |
0462169c | 11978 | |
9cd7b720 | 11979 | /* Split the atomic operation into a sequence. */ |
0462169c SN |
11980 | label = gen_label_rtx (); |
11981 | emit_label (label); | |
11982 | ||
11983 | if (new_out) | |
11984 | new_out = gen_lowpart (wmode, new_out); | |
11985 | if (old_out) | |
11986 | old_out = gen_lowpart (wmode, old_out); | |
11987 | else | |
11988 | old_out = new_out; | |
11989 | value = simplify_gen_subreg (wmode, value, mode, 0); | |
11990 | ||
f70fb3b6 MW |
11991 | /* The initial load can be relaxed for a __sync operation since a final |
11992 | barrier will be emitted to stop code hoisting. */ | |
11993 | if (is_sync) | |
11994 | aarch64_emit_load_exclusive (mode, old_out, mem, | |
11995 | GEN_INT (MEMMODEL_RELAXED)); | |
11996 | else | |
11997 | aarch64_emit_load_exclusive (mode, old_out, mem, model_rtx); | |
0462169c SN |
11998 | |
11999 | switch (code) | |
12000 | { | |
12001 | case SET: | |
12002 | new_out = value; | |
12003 | break; | |
12004 | ||
12005 | case NOT: | |
12006 | x = gen_rtx_AND (wmode, old_out, value); | |
f7df4a84 | 12007 | emit_insn (gen_rtx_SET (new_out, x)); |
0462169c | 12008 | x = gen_rtx_NOT (wmode, new_out); |
f7df4a84 | 12009 | emit_insn (gen_rtx_SET (new_out, x)); |
0462169c SN |
12010 | break; |
12011 | ||
12012 | case MINUS: | |
12013 | if (CONST_INT_P (value)) | |
12014 | { | |
12015 | value = GEN_INT (-INTVAL (value)); | |
12016 | code = PLUS; | |
12017 | } | |
12018 | /* Fall through. */ | |
12019 | ||
12020 | default: | |
12021 | x = gen_rtx_fmt_ee (code, wmode, old_out, value); | |
f7df4a84 | 12022 | emit_insn (gen_rtx_SET (new_out, x)); |
0462169c SN |
12023 | break; |
12024 | } | |
12025 | ||
12026 | aarch64_emit_store_exclusive (mode, cond, mem, | |
12027 | gen_lowpart (mode, new_out), model_rtx); | |
12028 | ||
12029 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
12030 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
12031 | gen_rtx_LABEL_REF (Pmode, label), pc_rtx); | |
f7df4a84 | 12032 | aarch64_emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); |
f70fb3b6 MW |
12033 | |
12034 | /* Emit any final barrier needed for a __sync operation. */ | |
12035 | if (is_sync) | |
12036 | aarch64_emit_post_barrier (model); | |
0462169c SN |
12037 | } |
12038 | ||
c2ec330c AL |
12039 | static void |
12040 | aarch64_init_libfuncs (void) | |
12041 | { | |
12042 | /* Half-precision float operations. The compiler handles all operations | |
12043 | with NULL libfuncs by converting to SFmode. */ | |
12044 | ||
12045 | /* Conversions. */ | |
12046 | set_conv_libfunc (trunc_optab, HFmode, SFmode, "__gnu_f2h_ieee"); | |
12047 | set_conv_libfunc (sext_optab, SFmode, HFmode, "__gnu_h2f_ieee"); | |
12048 | ||
12049 | /* Arithmetic. */ | |
12050 | set_optab_libfunc (add_optab, HFmode, NULL); | |
12051 | set_optab_libfunc (sdiv_optab, HFmode, NULL); | |
12052 | set_optab_libfunc (smul_optab, HFmode, NULL); | |
12053 | set_optab_libfunc (neg_optab, HFmode, NULL); | |
12054 | set_optab_libfunc (sub_optab, HFmode, NULL); | |
12055 | ||
12056 | /* Comparisons. */ | |
12057 | set_optab_libfunc (eq_optab, HFmode, NULL); | |
12058 | set_optab_libfunc (ne_optab, HFmode, NULL); | |
12059 | set_optab_libfunc (lt_optab, HFmode, NULL); | |
12060 | set_optab_libfunc (le_optab, HFmode, NULL); | |
12061 | set_optab_libfunc (ge_optab, HFmode, NULL); | |
12062 | set_optab_libfunc (gt_optab, HFmode, NULL); | |
12063 | set_optab_libfunc (unord_optab, HFmode, NULL); | |
12064 | } | |
12065 | ||
43e9d192 | 12066 | /* Target hook for c_mode_for_suffix. */ |
ef4bddc2 | 12067 | static machine_mode |
43e9d192 IB |
12068 | aarch64_c_mode_for_suffix (char suffix) |
12069 | { | |
12070 | if (suffix == 'q') | |
12071 | return TFmode; | |
12072 | ||
12073 | return VOIDmode; | |
12074 | } | |
12075 | ||
3520f7cc JG |
12076 | /* We can only represent floating point constants which will fit in |
12077 | "quarter-precision" values. These values are characterised by | |
12078 | a sign bit, a 4-bit mantissa and a 3-bit exponent. And are given | |
12079 | by: | |
12080 | ||
12081 | (-1)^s * (n/16) * 2^r | |
12082 | ||
12083 | Where: | |
12084 | 's' is the sign bit. | |
12085 | 'n' is an integer in the range 16 <= n <= 31. | |
12086 | 'r' is an integer in the range -3 <= r <= 4. */ | |
12087 | ||
12088 | /* Return true iff X can be represented by a quarter-precision | |
12089 | floating point immediate operand X. Note, we cannot represent 0.0. */ | |
12090 | bool | |
12091 | aarch64_float_const_representable_p (rtx x) | |
12092 | { | |
12093 | /* This represents our current view of how many bits | |
12094 | make up the mantissa. */ | |
12095 | int point_pos = 2 * HOST_BITS_PER_WIDE_INT - 1; | |
ba96cdfb | 12096 | int exponent; |
3520f7cc | 12097 | unsigned HOST_WIDE_INT mantissa, mask; |
3520f7cc | 12098 | REAL_VALUE_TYPE r, m; |
807e902e | 12099 | bool fail; |
3520f7cc JG |
12100 | |
12101 | if (!CONST_DOUBLE_P (x)) | |
12102 | return false; | |
12103 | ||
c2ec330c AL |
12104 | /* We don't support HFmode constants yet. */ |
12105 | if (GET_MODE (x) == VOIDmode || GET_MODE (x) == HFmode) | |
94bfa2da TV |
12106 | return false; |
12107 | ||
34a72c33 | 12108 | r = *CONST_DOUBLE_REAL_VALUE (x); |
3520f7cc JG |
12109 | |
12110 | /* We cannot represent infinities, NaNs or +/-zero. We won't | |
12111 | know if we have +zero until we analyse the mantissa, but we | |
12112 | can reject the other invalid values. */ | |
12113 | if (REAL_VALUE_ISINF (r) || REAL_VALUE_ISNAN (r) | |
12114 | || REAL_VALUE_MINUS_ZERO (r)) | |
12115 | return false; | |
12116 | ||
ba96cdfb | 12117 | /* Extract exponent. */ |
3520f7cc JG |
12118 | r = real_value_abs (&r); |
12119 | exponent = REAL_EXP (&r); | |
12120 | ||
12121 | /* For the mantissa, we expand into two HOST_WIDE_INTS, apart from the | |
12122 | highest (sign) bit, with a fixed binary point at bit point_pos. | |
12123 | m1 holds the low part of the mantissa, m2 the high part. | |
12124 | WARNING: If we ever have a representation using more than 2 * H_W_I - 1 | |
12125 | bits for the mantissa, this can fail (low bits will be lost). */ | |
12126 | real_ldexp (&m, &r, point_pos - exponent); | |
807e902e | 12127 | wide_int w = real_to_integer (&m, &fail, HOST_BITS_PER_WIDE_INT * 2); |
3520f7cc JG |
12128 | |
12129 | /* If the low part of the mantissa has bits set we cannot represent | |
12130 | the value. */ | |
807e902e | 12131 | if (w.elt (0) != 0) |
3520f7cc JG |
12132 | return false; |
12133 | /* We have rejected the lower HOST_WIDE_INT, so update our | |
12134 | understanding of how many bits lie in the mantissa and | |
12135 | look only at the high HOST_WIDE_INT. */ | |
807e902e | 12136 | mantissa = w.elt (1); |
3520f7cc JG |
12137 | point_pos -= HOST_BITS_PER_WIDE_INT; |
12138 | ||
12139 | /* We can only represent values with a mantissa of the form 1.xxxx. */ | |
12140 | mask = ((unsigned HOST_WIDE_INT)1 << (point_pos - 5)) - 1; | |
12141 | if ((mantissa & mask) != 0) | |
12142 | return false; | |
12143 | ||
12144 | /* Having filtered unrepresentable values, we may now remove all | |
12145 | but the highest 5 bits. */ | |
12146 | mantissa >>= point_pos - 5; | |
12147 | ||
12148 | /* We cannot represent the value 0.0, so reject it. This is handled | |
12149 | elsewhere. */ | |
12150 | if (mantissa == 0) | |
12151 | return false; | |
12152 | ||
12153 | /* Then, as bit 4 is always set, we can mask it off, leaving | |
12154 | the mantissa in the range [0, 15]. */ | |
12155 | mantissa &= ~(1 << 4); | |
12156 | gcc_assert (mantissa <= 15); | |
12157 | ||
12158 | /* GCC internally does not use IEEE754-like encoding (where normalized | |
12159 | significands are in the range [1, 2). GCC uses [0.5, 1) (see real.c). | |
12160 | Our mantissa values are shifted 4 places to the left relative to | |
12161 | normalized IEEE754 so we must modify the exponent returned by REAL_EXP | |
12162 | by 5 places to correct for GCC's representation. */ | |
12163 | exponent = 5 - exponent; | |
12164 | ||
12165 | return (exponent >= 0 && exponent <= 7); | |
12166 | } | |
12167 | ||
12168 | char* | |
81c2dfb9 | 12169 | aarch64_output_simd_mov_immediate (rtx const_vector, |
ef4bddc2 | 12170 | machine_mode mode, |
3520f7cc JG |
12171 | unsigned width) |
12172 | { | |
3ea63f60 | 12173 | bool is_valid; |
3520f7cc | 12174 | static char templ[40]; |
3520f7cc | 12175 | const char *mnemonic; |
e4f0f84d | 12176 | const char *shift_op; |
3520f7cc | 12177 | unsigned int lane_count = 0; |
81c2dfb9 | 12178 | char element_char; |
3520f7cc | 12179 | |
e4f0f84d | 12180 | struct simd_immediate_info info = { NULL_RTX, 0, 0, false, false }; |
48063b9d IB |
12181 | |
12182 | /* This will return true to show const_vector is legal for use as either | |
12183 | a AdvSIMD MOVI instruction (or, implicitly, MVNI) immediate. It will | |
12184 | also update INFO to show how the immediate should be generated. */ | |
81c2dfb9 | 12185 | is_valid = aarch64_simd_valid_immediate (const_vector, mode, false, &info); |
3520f7cc JG |
12186 | gcc_assert (is_valid); |
12187 | ||
81c2dfb9 | 12188 | element_char = sizetochar (info.element_width); |
48063b9d IB |
12189 | lane_count = width / info.element_width; |
12190 | ||
3520f7cc | 12191 | mode = GET_MODE_INNER (mode); |
0d8e1702 | 12192 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
3520f7cc | 12193 | { |
48063b9d | 12194 | gcc_assert (info.shift == 0 && ! info.mvn); |
0d8e1702 KT |
12195 | /* For FP zero change it to a CONST_INT 0 and use the integer SIMD |
12196 | move immediate path. */ | |
48063b9d IB |
12197 | if (aarch64_float_const_zero_rtx_p (info.value)) |
12198 | info.value = GEN_INT (0); | |
12199 | else | |
12200 | { | |
83faf7d0 | 12201 | const unsigned int buf_size = 20; |
48063b9d | 12202 | char float_buf[buf_size] = {'\0'}; |
34a72c33 RS |
12203 | real_to_decimal_for_mode (float_buf, |
12204 | CONST_DOUBLE_REAL_VALUE (info.value), | |
12205 | buf_size, buf_size, 1, mode); | |
48063b9d IB |
12206 | |
12207 | if (lane_count == 1) | |
12208 | snprintf (templ, sizeof (templ), "fmov\t%%d0, %s", float_buf); | |
12209 | else | |
12210 | snprintf (templ, sizeof (templ), "fmov\t%%0.%d%c, %s", | |
81c2dfb9 | 12211 | lane_count, element_char, float_buf); |
48063b9d IB |
12212 | return templ; |
12213 | } | |
3520f7cc | 12214 | } |
3520f7cc | 12215 | |
48063b9d | 12216 | mnemonic = info.mvn ? "mvni" : "movi"; |
e4f0f84d | 12217 | shift_op = info.msl ? "msl" : "lsl"; |
3520f7cc | 12218 | |
0d8e1702 | 12219 | gcc_assert (CONST_INT_P (info.value)); |
3520f7cc | 12220 | if (lane_count == 1) |
48063b9d IB |
12221 | snprintf (templ, sizeof (templ), "%s\t%%d0, " HOST_WIDE_INT_PRINT_HEX, |
12222 | mnemonic, UINTVAL (info.value)); | |
12223 | else if (info.shift) | |
12224 | snprintf (templ, sizeof (templ), "%s\t%%0.%d%c, " HOST_WIDE_INT_PRINT_HEX | |
e4f0f84d TB |
12225 | ", %s %d", mnemonic, lane_count, element_char, |
12226 | UINTVAL (info.value), shift_op, info.shift); | |
3520f7cc | 12227 | else |
48063b9d | 12228 | snprintf (templ, sizeof (templ), "%s\t%%0.%d%c, " HOST_WIDE_INT_PRINT_HEX, |
81c2dfb9 | 12229 | mnemonic, lane_count, element_char, UINTVAL (info.value)); |
3520f7cc JG |
12230 | return templ; |
12231 | } | |
12232 | ||
b7342d25 IB |
12233 | char* |
12234 | aarch64_output_scalar_simd_mov_immediate (rtx immediate, | |
ef4bddc2 | 12235 | machine_mode mode) |
b7342d25 | 12236 | { |
ef4bddc2 | 12237 | machine_mode vmode; |
b7342d25 IB |
12238 | |
12239 | gcc_assert (!VECTOR_MODE_P (mode)); | |
12240 | vmode = aarch64_simd_container_mode (mode, 64); | |
12241 | rtx v_op = aarch64_simd_gen_const_vector_dup (vmode, INTVAL (immediate)); | |
12242 | return aarch64_output_simd_mov_immediate (v_op, vmode, 64); | |
12243 | } | |
12244 | ||
88b08073 JG |
12245 | /* Split operands into moves from op[1] + op[2] into op[0]. */ |
12246 | ||
12247 | void | |
12248 | aarch64_split_combinev16qi (rtx operands[3]) | |
12249 | { | |
12250 | unsigned int dest = REGNO (operands[0]); | |
12251 | unsigned int src1 = REGNO (operands[1]); | |
12252 | unsigned int src2 = REGNO (operands[2]); | |
ef4bddc2 | 12253 | machine_mode halfmode = GET_MODE (operands[1]); |
88b08073 JG |
12254 | unsigned int halfregs = HARD_REGNO_NREGS (src1, halfmode); |
12255 | rtx destlo, desthi; | |
12256 | ||
12257 | gcc_assert (halfmode == V16QImode); | |
12258 | ||
12259 | if (src1 == dest && src2 == dest + halfregs) | |
12260 | { | |
12261 | /* No-op move. Can't split to nothing; emit something. */ | |
12262 | emit_note (NOTE_INSN_DELETED); | |
12263 | return; | |
12264 | } | |
12265 | ||
12266 | /* Preserve register attributes for variable tracking. */ | |
12267 | destlo = gen_rtx_REG_offset (operands[0], halfmode, dest, 0); | |
12268 | desthi = gen_rtx_REG_offset (operands[0], halfmode, dest + halfregs, | |
12269 | GET_MODE_SIZE (halfmode)); | |
12270 | ||
12271 | /* Special case of reversed high/low parts. */ | |
12272 | if (reg_overlap_mentioned_p (operands[2], destlo) | |
12273 | && reg_overlap_mentioned_p (operands[1], desthi)) | |
12274 | { | |
12275 | emit_insn (gen_xorv16qi3 (operands[1], operands[1], operands[2])); | |
12276 | emit_insn (gen_xorv16qi3 (operands[2], operands[1], operands[2])); | |
12277 | emit_insn (gen_xorv16qi3 (operands[1], operands[1], operands[2])); | |
12278 | } | |
12279 | else if (!reg_overlap_mentioned_p (operands[2], destlo)) | |
12280 | { | |
12281 | /* Try to avoid unnecessary moves if part of the result | |
12282 | is in the right place already. */ | |
12283 | if (src1 != dest) | |
12284 | emit_move_insn (destlo, operands[1]); | |
12285 | if (src2 != dest + halfregs) | |
12286 | emit_move_insn (desthi, operands[2]); | |
12287 | } | |
12288 | else | |
12289 | { | |
12290 | if (src2 != dest + halfregs) | |
12291 | emit_move_insn (desthi, operands[2]); | |
12292 | if (src1 != dest) | |
12293 | emit_move_insn (destlo, operands[1]); | |
12294 | } | |
12295 | } | |
12296 | ||
12297 | /* vec_perm support. */ | |
12298 | ||
12299 | #define MAX_VECT_LEN 16 | |
12300 | ||
12301 | struct expand_vec_perm_d | |
12302 | { | |
12303 | rtx target, op0, op1; | |
12304 | unsigned char perm[MAX_VECT_LEN]; | |
ef4bddc2 | 12305 | machine_mode vmode; |
88b08073 JG |
12306 | unsigned char nelt; |
12307 | bool one_vector_p; | |
12308 | bool testing_p; | |
12309 | }; | |
12310 | ||
12311 | /* Generate a variable permutation. */ | |
12312 | ||
12313 | static void | |
12314 | aarch64_expand_vec_perm_1 (rtx target, rtx op0, rtx op1, rtx sel) | |
12315 | { | |
ef4bddc2 | 12316 | machine_mode vmode = GET_MODE (target); |
88b08073 JG |
12317 | bool one_vector_p = rtx_equal_p (op0, op1); |
12318 | ||
12319 | gcc_checking_assert (vmode == V8QImode || vmode == V16QImode); | |
12320 | gcc_checking_assert (GET_MODE (op0) == vmode); | |
12321 | gcc_checking_assert (GET_MODE (op1) == vmode); | |
12322 | gcc_checking_assert (GET_MODE (sel) == vmode); | |
12323 | gcc_checking_assert (TARGET_SIMD); | |
12324 | ||
12325 | if (one_vector_p) | |
12326 | { | |
12327 | if (vmode == V8QImode) | |
12328 | { | |
12329 | /* Expand the argument to a V16QI mode by duplicating it. */ | |
12330 | rtx pair = gen_reg_rtx (V16QImode); | |
12331 | emit_insn (gen_aarch64_combinev8qi (pair, op0, op0)); | |
12332 | emit_insn (gen_aarch64_tbl1v8qi (target, pair, sel)); | |
12333 | } | |
12334 | else | |
12335 | { | |
12336 | emit_insn (gen_aarch64_tbl1v16qi (target, op0, sel)); | |
12337 | } | |
12338 | } | |
12339 | else | |
12340 | { | |
12341 | rtx pair; | |
12342 | ||
12343 | if (vmode == V8QImode) | |
12344 | { | |
12345 | pair = gen_reg_rtx (V16QImode); | |
12346 | emit_insn (gen_aarch64_combinev8qi (pair, op0, op1)); | |
12347 | emit_insn (gen_aarch64_tbl1v8qi (target, pair, sel)); | |
12348 | } | |
12349 | else | |
12350 | { | |
12351 | pair = gen_reg_rtx (OImode); | |
12352 | emit_insn (gen_aarch64_combinev16qi (pair, op0, op1)); | |
12353 | emit_insn (gen_aarch64_tbl2v16qi (target, pair, sel)); | |
12354 | } | |
12355 | } | |
12356 | } | |
12357 | ||
12358 | void | |
12359 | aarch64_expand_vec_perm (rtx target, rtx op0, rtx op1, rtx sel) | |
12360 | { | |
ef4bddc2 | 12361 | machine_mode vmode = GET_MODE (target); |
c9d1a16a | 12362 | unsigned int nelt = GET_MODE_NUNITS (vmode); |
88b08073 | 12363 | bool one_vector_p = rtx_equal_p (op0, op1); |
f7c4e5b8 | 12364 | rtx mask; |
88b08073 JG |
12365 | |
12366 | /* The TBL instruction does not use a modulo index, so we must take care | |
12367 | of that ourselves. */ | |
f7c4e5b8 AL |
12368 | mask = aarch64_simd_gen_const_vector_dup (vmode, |
12369 | one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
88b08073 JG |
12370 | sel = expand_simple_binop (vmode, AND, sel, mask, NULL, 0, OPTAB_LIB_WIDEN); |
12371 | ||
f7c4e5b8 AL |
12372 | /* For big-endian, we also need to reverse the index within the vector |
12373 | (but not which vector). */ | |
12374 | if (BYTES_BIG_ENDIAN) | |
12375 | { | |
12376 | /* If one_vector_p, mask is a vector of (nelt - 1)'s already. */ | |
12377 | if (!one_vector_p) | |
12378 | mask = aarch64_simd_gen_const_vector_dup (vmode, nelt - 1); | |
12379 | sel = expand_simple_binop (vmode, XOR, sel, mask, | |
12380 | NULL, 0, OPTAB_LIB_WIDEN); | |
12381 | } | |
88b08073 JG |
12382 | aarch64_expand_vec_perm_1 (target, op0, op1, sel); |
12383 | } | |
12384 | ||
cc4d934f JG |
12385 | /* Recognize patterns suitable for the TRN instructions. */ |
12386 | static bool | |
12387 | aarch64_evpc_trn (struct expand_vec_perm_d *d) | |
12388 | { | |
12389 | unsigned int i, odd, mask, nelt = d->nelt; | |
12390 | rtx out, in0, in1, x; | |
12391 | rtx (*gen) (rtx, rtx, rtx); | |
ef4bddc2 | 12392 | machine_mode vmode = d->vmode; |
cc4d934f JG |
12393 | |
12394 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
12395 | return false; | |
12396 | ||
12397 | /* Note that these are little-endian tests. | |
12398 | We correct for big-endian later. */ | |
12399 | if (d->perm[0] == 0) | |
12400 | odd = 0; | |
12401 | else if (d->perm[0] == 1) | |
12402 | odd = 1; | |
12403 | else | |
12404 | return false; | |
12405 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
12406 | ||
12407 | for (i = 0; i < nelt; i += 2) | |
12408 | { | |
12409 | if (d->perm[i] != i + odd) | |
12410 | return false; | |
12411 | if (d->perm[i + 1] != ((i + nelt + odd) & mask)) | |
12412 | return false; | |
12413 | } | |
12414 | ||
12415 | /* Success! */ | |
12416 | if (d->testing_p) | |
12417 | return true; | |
12418 | ||
12419 | in0 = d->op0; | |
12420 | in1 = d->op1; | |
12421 | if (BYTES_BIG_ENDIAN) | |
12422 | { | |
12423 | x = in0, in0 = in1, in1 = x; | |
12424 | odd = !odd; | |
12425 | } | |
12426 | out = d->target; | |
12427 | ||
12428 | if (odd) | |
12429 | { | |
12430 | switch (vmode) | |
12431 | { | |
12432 | case V16QImode: gen = gen_aarch64_trn2v16qi; break; | |
12433 | case V8QImode: gen = gen_aarch64_trn2v8qi; break; | |
12434 | case V8HImode: gen = gen_aarch64_trn2v8hi; break; | |
12435 | case V4HImode: gen = gen_aarch64_trn2v4hi; break; | |
12436 | case V4SImode: gen = gen_aarch64_trn2v4si; break; | |
12437 | case V2SImode: gen = gen_aarch64_trn2v2si; break; | |
12438 | case V2DImode: gen = gen_aarch64_trn2v2di; break; | |
358decd5 JW |
12439 | case V4HFmode: gen = gen_aarch64_trn2v4hf; break; |
12440 | case V8HFmode: gen = gen_aarch64_trn2v8hf; break; | |
cc4d934f JG |
12441 | case V4SFmode: gen = gen_aarch64_trn2v4sf; break; |
12442 | case V2SFmode: gen = gen_aarch64_trn2v2sf; break; | |
12443 | case V2DFmode: gen = gen_aarch64_trn2v2df; break; | |
12444 | default: | |
12445 | return false; | |
12446 | } | |
12447 | } | |
12448 | else | |
12449 | { | |
12450 | switch (vmode) | |
12451 | { | |
12452 | case V16QImode: gen = gen_aarch64_trn1v16qi; break; | |
12453 | case V8QImode: gen = gen_aarch64_trn1v8qi; break; | |
12454 | case V8HImode: gen = gen_aarch64_trn1v8hi; break; | |
12455 | case V4HImode: gen = gen_aarch64_trn1v4hi; break; | |
12456 | case V4SImode: gen = gen_aarch64_trn1v4si; break; | |
12457 | case V2SImode: gen = gen_aarch64_trn1v2si; break; | |
12458 | case V2DImode: gen = gen_aarch64_trn1v2di; break; | |
358decd5 JW |
12459 | case V4HFmode: gen = gen_aarch64_trn1v4hf; break; |
12460 | case V8HFmode: gen = gen_aarch64_trn1v8hf; break; | |
cc4d934f JG |
12461 | case V4SFmode: gen = gen_aarch64_trn1v4sf; break; |
12462 | case V2SFmode: gen = gen_aarch64_trn1v2sf; break; | |
12463 | case V2DFmode: gen = gen_aarch64_trn1v2df; break; | |
12464 | default: | |
12465 | return false; | |
12466 | } | |
12467 | } | |
12468 | ||
12469 | emit_insn (gen (out, in0, in1)); | |
12470 | return true; | |
12471 | } | |
12472 | ||
12473 | /* Recognize patterns suitable for the UZP instructions. */ | |
12474 | static bool | |
12475 | aarch64_evpc_uzp (struct expand_vec_perm_d *d) | |
12476 | { | |
12477 | unsigned int i, odd, mask, nelt = d->nelt; | |
12478 | rtx out, in0, in1, x; | |
12479 | rtx (*gen) (rtx, rtx, rtx); | |
ef4bddc2 | 12480 | machine_mode vmode = d->vmode; |
cc4d934f JG |
12481 | |
12482 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
12483 | return false; | |
12484 | ||
12485 | /* Note that these are little-endian tests. | |
12486 | We correct for big-endian later. */ | |
12487 | if (d->perm[0] == 0) | |
12488 | odd = 0; | |
12489 | else if (d->perm[0] == 1) | |
12490 | odd = 1; | |
12491 | else | |
12492 | return false; | |
12493 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
12494 | ||
12495 | for (i = 0; i < nelt; i++) | |
12496 | { | |
12497 | unsigned elt = (i * 2 + odd) & mask; | |
12498 | if (d->perm[i] != elt) | |
12499 | return false; | |
12500 | } | |
12501 | ||
12502 | /* Success! */ | |
12503 | if (d->testing_p) | |
12504 | return true; | |
12505 | ||
12506 | in0 = d->op0; | |
12507 | in1 = d->op1; | |
12508 | if (BYTES_BIG_ENDIAN) | |
12509 | { | |
12510 | x = in0, in0 = in1, in1 = x; | |
12511 | odd = !odd; | |
12512 | } | |
12513 | out = d->target; | |
12514 | ||
12515 | if (odd) | |
12516 | { | |
12517 | switch (vmode) | |
12518 | { | |
12519 | case V16QImode: gen = gen_aarch64_uzp2v16qi; break; | |
12520 | case V8QImode: gen = gen_aarch64_uzp2v8qi; break; | |
12521 | case V8HImode: gen = gen_aarch64_uzp2v8hi; break; | |
12522 | case V4HImode: gen = gen_aarch64_uzp2v4hi; break; | |
12523 | case V4SImode: gen = gen_aarch64_uzp2v4si; break; | |
12524 | case V2SImode: gen = gen_aarch64_uzp2v2si; break; | |
12525 | case V2DImode: gen = gen_aarch64_uzp2v2di; break; | |
358decd5 JW |
12526 | case V4HFmode: gen = gen_aarch64_uzp2v4hf; break; |
12527 | case V8HFmode: gen = gen_aarch64_uzp2v8hf; break; | |
cc4d934f JG |
12528 | case V4SFmode: gen = gen_aarch64_uzp2v4sf; break; |
12529 | case V2SFmode: gen = gen_aarch64_uzp2v2sf; break; | |
12530 | case V2DFmode: gen = gen_aarch64_uzp2v2df; break; | |
12531 | default: | |
12532 | return false; | |
12533 | } | |
12534 | } | |
12535 | else | |
12536 | { | |
12537 | switch (vmode) | |
12538 | { | |
12539 | case V16QImode: gen = gen_aarch64_uzp1v16qi; break; | |
12540 | case V8QImode: gen = gen_aarch64_uzp1v8qi; break; | |
12541 | case V8HImode: gen = gen_aarch64_uzp1v8hi; break; | |
12542 | case V4HImode: gen = gen_aarch64_uzp1v4hi; break; | |
12543 | case V4SImode: gen = gen_aarch64_uzp1v4si; break; | |
12544 | case V2SImode: gen = gen_aarch64_uzp1v2si; break; | |
12545 | case V2DImode: gen = gen_aarch64_uzp1v2di; break; | |
358decd5 JW |
12546 | case V4HFmode: gen = gen_aarch64_uzp1v4hf; break; |
12547 | case V8HFmode: gen = gen_aarch64_uzp1v8hf; break; | |
cc4d934f JG |
12548 | case V4SFmode: gen = gen_aarch64_uzp1v4sf; break; |
12549 | case V2SFmode: gen = gen_aarch64_uzp1v2sf; break; | |
12550 | case V2DFmode: gen = gen_aarch64_uzp1v2df; break; | |
12551 | default: | |
12552 | return false; | |
12553 | } | |
12554 | } | |
12555 | ||
12556 | emit_insn (gen (out, in0, in1)); | |
12557 | return true; | |
12558 | } | |
12559 | ||
12560 | /* Recognize patterns suitable for the ZIP instructions. */ | |
12561 | static bool | |
12562 | aarch64_evpc_zip (struct expand_vec_perm_d *d) | |
12563 | { | |
12564 | unsigned int i, high, mask, nelt = d->nelt; | |
12565 | rtx out, in0, in1, x; | |
12566 | rtx (*gen) (rtx, rtx, rtx); | |
ef4bddc2 | 12567 | machine_mode vmode = d->vmode; |
cc4d934f JG |
12568 | |
12569 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
12570 | return false; | |
12571 | ||
12572 | /* Note that these are little-endian tests. | |
12573 | We correct for big-endian later. */ | |
12574 | high = nelt / 2; | |
12575 | if (d->perm[0] == high) | |
12576 | /* Do Nothing. */ | |
12577 | ; | |
12578 | else if (d->perm[0] == 0) | |
12579 | high = 0; | |
12580 | else | |
12581 | return false; | |
12582 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
12583 | ||
12584 | for (i = 0; i < nelt / 2; i++) | |
12585 | { | |
12586 | unsigned elt = (i + high) & mask; | |
12587 | if (d->perm[i * 2] != elt) | |
12588 | return false; | |
12589 | elt = (elt + nelt) & mask; | |
12590 | if (d->perm[i * 2 + 1] != elt) | |
12591 | return false; | |
12592 | } | |
12593 | ||
12594 | /* Success! */ | |
12595 | if (d->testing_p) | |
12596 | return true; | |
12597 | ||
12598 | in0 = d->op0; | |
12599 | in1 = d->op1; | |
12600 | if (BYTES_BIG_ENDIAN) | |
12601 | { | |
12602 | x = in0, in0 = in1, in1 = x; | |
12603 | high = !high; | |
12604 | } | |
12605 | out = d->target; | |
12606 | ||
12607 | if (high) | |
12608 | { | |
12609 | switch (vmode) | |
12610 | { | |
12611 | case V16QImode: gen = gen_aarch64_zip2v16qi; break; | |
12612 | case V8QImode: gen = gen_aarch64_zip2v8qi; break; | |
12613 | case V8HImode: gen = gen_aarch64_zip2v8hi; break; | |
12614 | case V4HImode: gen = gen_aarch64_zip2v4hi; break; | |
12615 | case V4SImode: gen = gen_aarch64_zip2v4si; break; | |
12616 | case V2SImode: gen = gen_aarch64_zip2v2si; break; | |
12617 | case V2DImode: gen = gen_aarch64_zip2v2di; break; | |
358decd5 JW |
12618 | case V4HFmode: gen = gen_aarch64_zip2v4hf; break; |
12619 | case V8HFmode: gen = gen_aarch64_zip2v8hf; break; | |
cc4d934f JG |
12620 | case V4SFmode: gen = gen_aarch64_zip2v4sf; break; |
12621 | case V2SFmode: gen = gen_aarch64_zip2v2sf; break; | |
12622 | case V2DFmode: gen = gen_aarch64_zip2v2df; break; | |
12623 | default: | |
12624 | return false; | |
12625 | } | |
12626 | } | |
12627 | else | |
12628 | { | |
12629 | switch (vmode) | |
12630 | { | |
12631 | case V16QImode: gen = gen_aarch64_zip1v16qi; break; | |
12632 | case V8QImode: gen = gen_aarch64_zip1v8qi; break; | |
12633 | case V8HImode: gen = gen_aarch64_zip1v8hi; break; | |
12634 | case V4HImode: gen = gen_aarch64_zip1v4hi; break; | |
12635 | case V4SImode: gen = gen_aarch64_zip1v4si; break; | |
12636 | case V2SImode: gen = gen_aarch64_zip1v2si; break; | |
12637 | case V2DImode: gen = gen_aarch64_zip1v2di; break; | |
358decd5 JW |
12638 | case V4HFmode: gen = gen_aarch64_zip1v4hf; break; |
12639 | case V8HFmode: gen = gen_aarch64_zip1v8hf; break; | |
cc4d934f JG |
12640 | case V4SFmode: gen = gen_aarch64_zip1v4sf; break; |
12641 | case V2SFmode: gen = gen_aarch64_zip1v2sf; break; | |
12642 | case V2DFmode: gen = gen_aarch64_zip1v2df; break; | |
12643 | default: | |
12644 | return false; | |
12645 | } | |
12646 | } | |
12647 | ||
12648 | emit_insn (gen (out, in0, in1)); | |
12649 | return true; | |
12650 | } | |
12651 | ||
ae0533da AL |
12652 | /* Recognize patterns for the EXT insn. */ |
12653 | ||
12654 | static bool | |
12655 | aarch64_evpc_ext (struct expand_vec_perm_d *d) | |
12656 | { | |
12657 | unsigned int i, nelt = d->nelt; | |
12658 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
12659 | rtx offset; | |
12660 | ||
12661 | unsigned int location = d->perm[0]; /* Always < nelt. */ | |
12662 | ||
12663 | /* Check if the extracted indices are increasing by one. */ | |
12664 | for (i = 1; i < nelt; i++) | |
12665 | { | |
12666 | unsigned int required = location + i; | |
12667 | if (d->one_vector_p) | |
12668 | { | |
12669 | /* We'll pass the same vector in twice, so allow indices to wrap. */ | |
12670 | required &= (nelt - 1); | |
12671 | } | |
12672 | if (d->perm[i] != required) | |
12673 | return false; | |
12674 | } | |
12675 | ||
ae0533da AL |
12676 | switch (d->vmode) |
12677 | { | |
12678 | case V16QImode: gen = gen_aarch64_extv16qi; break; | |
12679 | case V8QImode: gen = gen_aarch64_extv8qi; break; | |
12680 | case V4HImode: gen = gen_aarch64_extv4hi; break; | |
12681 | case V8HImode: gen = gen_aarch64_extv8hi; break; | |
12682 | case V2SImode: gen = gen_aarch64_extv2si; break; | |
12683 | case V4SImode: gen = gen_aarch64_extv4si; break; | |
358decd5 JW |
12684 | case V4HFmode: gen = gen_aarch64_extv4hf; break; |
12685 | case V8HFmode: gen = gen_aarch64_extv8hf; break; | |
ae0533da AL |
12686 | case V2SFmode: gen = gen_aarch64_extv2sf; break; |
12687 | case V4SFmode: gen = gen_aarch64_extv4sf; break; | |
12688 | case V2DImode: gen = gen_aarch64_extv2di; break; | |
12689 | case V2DFmode: gen = gen_aarch64_extv2df; break; | |
12690 | default: | |
12691 | return false; | |
12692 | } | |
12693 | ||
12694 | /* Success! */ | |
12695 | if (d->testing_p) | |
12696 | return true; | |
12697 | ||
b31e65bb AL |
12698 | /* The case where (location == 0) is a no-op for both big- and little-endian, |
12699 | and is removed by the mid-end at optimization levels -O1 and higher. */ | |
12700 | ||
12701 | if (BYTES_BIG_ENDIAN && (location != 0)) | |
ae0533da AL |
12702 | { |
12703 | /* After setup, we want the high elements of the first vector (stored | |
12704 | at the LSB end of the register), and the low elements of the second | |
12705 | vector (stored at the MSB end of the register). So swap. */ | |
cb5c6c29 | 12706 | std::swap (d->op0, d->op1); |
ae0533da AL |
12707 | /* location != 0 (above), so safe to assume (nelt - location) < nelt. */ |
12708 | location = nelt - location; | |
12709 | } | |
12710 | ||
12711 | offset = GEN_INT (location); | |
12712 | emit_insn (gen (d->target, d->op0, d->op1, offset)); | |
12713 | return true; | |
12714 | } | |
12715 | ||
923fcec3 AL |
12716 | /* Recognize patterns for the REV insns. */ |
12717 | ||
12718 | static bool | |
12719 | aarch64_evpc_rev (struct expand_vec_perm_d *d) | |
12720 | { | |
12721 | unsigned int i, j, diff, nelt = d->nelt; | |
12722 | rtx (*gen) (rtx, rtx); | |
12723 | ||
12724 | if (!d->one_vector_p) | |
12725 | return false; | |
12726 | ||
12727 | diff = d->perm[0]; | |
12728 | switch (diff) | |
12729 | { | |
12730 | case 7: | |
12731 | switch (d->vmode) | |
12732 | { | |
12733 | case V16QImode: gen = gen_aarch64_rev64v16qi; break; | |
12734 | case V8QImode: gen = gen_aarch64_rev64v8qi; break; | |
12735 | default: | |
12736 | return false; | |
12737 | } | |
12738 | break; | |
12739 | case 3: | |
12740 | switch (d->vmode) | |
12741 | { | |
12742 | case V16QImode: gen = gen_aarch64_rev32v16qi; break; | |
12743 | case V8QImode: gen = gen_aarch64_rev32v8qi; break; | |
12744 | case V8HImode: gen = gen_aarch64_rev64v8hi; break; | |
12745 | case V4HImode: gen = gen_aarch64_rev64v4hi; break; | |
12746 | default: | |
12747 | return false; | |
12748 | } | |
12749 | break; | |
12750 | case 1: | |
12751 | switch (d->vmode) | |
12752 | { | |
12753 | case V16QImode: gen = gen_aarch64_rev16v16qi; break; | |
12754 | case V8QImode: gen = gen_aarch64_rev16v8qi; break; | |
12755 | case V8HImode: gen = gen_aarch64_rev32v8hi; break; | |
12756 | case V4HImode: gen = gen_aarch64_rev32v4hi; break; | |
12757 | case V4SImode: gen = gen_aarch64_rev64v4si; break; | |
12758 | case V2SImode: gen = gen_aarch64_rev64v2si; break; | |
12759 | case V4SFmode: gen = gen_aarch64_rev64v4sf; break; | |
12760 | case V2SFmode: gen = gen_aarch64_rev64v2sf; break; | |
358decd5 JW |
12761 | case V8HFmode: gen = gen_aarch64_rev64v8hf; break; |
12762 | case V4HFmode: gen = gen_aarch64_rev64v4hf; break; | |
923fcec3 AL |
12763 | default: |
12764 | return false; | |
12765 | } | |
12766 | break; | |
12767 | default: | |
12768 | return false; | |
12769 | } | |
12770 | ||
12771 | for (i = 0; i < nelt ; i += diff + 1) | |
12772 | for (j = 0; j <= diff; j += 1) | |
12773 | { | |
12774 | /* This is guaranteed to be true as the value of diff | |
12775 | is 7, 3, 1 and we should have enough elements in the | |
12776 | queue to generate this. Getting a vector mask with a | |
12777 | value of diff other than these values implies that | |
12778 | something is wrong by the time we get here. */ | |
12779 | gcc_assert (i + j < nelt); | |
12780 | if (d->perm[i + j] != i + diff - j) | |
12781 | return false; | |
12782 | } | |
12783 | ||
12784 | /* Success! */ | |
12785 | if (d->testing_p) | |
12786 | return true; | |
12787 | ||
12788 | emit_insn (gen (d->target, d->op0)); | |
12789 | return true; | |
12790 | } | |
12791 | ||
91bd4114 JG |
12792 | static bool |
12793 | aarch64_evpc_dup (struct expand_vec_perm_d *d) | |
12794 | { | |
12795 | rtx (*gen) (rtx, rtx, rtx); | |
12796 | rtx out = d->target; | |
12797 | rtx in0; | |
ef4bddc2 | 12798 | machine_mode vmode = d->vmode; |
91bd4114 JG |
12799 | unsigned int i, elt, nelt = d->nelt; |
12800 | rtx lane; | |
12801 | ||
91bd4114 JG |
12802 | elt = d->perm[0]; |
12803 | for (i = 1; i < nelt; i++) | |
12804 | { | |
12805 | if (elt != d->perm[i]) | |
12806 | return false; | |
12807 | } | |
12808 | ||
12809 | /* The generic preparation in aarch64_expand_vec_perm_const_1 | |
12810 | swaps the operand order and the permute indices if it finds | |
12811 | d->perm[0] to be in the second operand. Thus, we can always | |
12812 | use d->op0 and need not do any extra arithmetic to get the | |
12813 | correct lane number. */ | |
12814 | in0 = d->op0; | |
f901401e | 12815 | lane = GEN_INT (elt); /* The pattern corrects for big-endian. */ |
91bd4114 JG |
12816 | |
12817 | switch (vmode) | |
12818 | { | |
12819 | case V16QImode: gen = gen_aarch64_dup_lanev16qi; break; | |
12820 | case V8QImode: gen = gen_aarch64_dup_lanev8qi; break; | |
12821 | case V8HImode: gen = gen_aarch64_dup_lanev8hi; break; | |
12822 | case V4HImode: gen = gen_aarch64_dup_lanev4hi; break; | |
12823 | case V4SImode: gen = gen_aarch64_dup_lanev4si; break; | |
12824 | case V2SImode: gen = gen_aarch64_dup_lanev2si; break; | |
12825 | case V2DImode: gen = gen_aarch64_dup_lanev2di; break; | |
862abc04 AL |
12826 | case V8HFmode: gen = gen_aarch64_dup_lanev8hf; break; |
12827 | case V4HFmode: gen = gen_aarch64_dup_lanev4hf; break; | |
91bd4114 JG |
12828 | case V4SFmode: gen = gen_aarch64_dup_lanev4sf; break; |
12829 | case V2SFmode: gen = gen_aarch64_dup_lanev2sf; break; | |
12830 | case V2DFmode: gen = gen_aarch64_dup_lanev2df; break; | |
12831 | default: | |
12832 | return false; | |
12833 | } | |
12834 | ||
12835 | emit_insn (gen (out, in0, lane)); | |
12836 | return true; | |
12837 | } | |
12838 | ||
88b08073 JG |
12839 | static bool |
12840 | aarch64_evpc_tbl (struct expand_vec_perm_d *d) | |
12841 | { | |
12842 | rtx rperm[MAX_VECT_LEN], sel; | |
ef4bddc2 | 12843 | machine_mode vmode = d->vmode; |
88b08073 JG |
12844 | unsigned int i, nelt = d->nelt; |
12845 | ||
88b08073 JG |
12846 | if (d->testing_p) |
12847 | return true; | |
12848 | ||
12849 | /* Generic code will try constant permutation twice. Once with the | |
12850 | original mode and again with the elements lowered to QImode. | |
12851 | So wait and don't do the selector expansion ourselves. */ | |
12852 | if (vmode != V8QImode && vmode != V16QImode) | |
12853 | return false; | |
12854 | ||
12855 | for (i = 0; i < nelt; ++i) | |
bbcc9c00 TB |
12856 | { |
12857 | int nunits = GET_MODE_NUNITS (vmode); | |
12858 | ||
12859 | /* If big-endian and two vectors we end up with a weird mixed-endian | |
12860 | mode on NEON. Reverse the index within each word but not the word | |
12861 | itself. */ | |
12862 | rperm[i] = GEN_INT (BYTES_BIG_ENDIAN ? d->perm[i] ^ (nunits - 1) | |
12863 | : d->perm[i]); | |
12864 | } | |
88b08073 JG |
12865 | sel = gen_rtx_CONST_VECTOR (vmode, gen_rtvec_v (nelt, rperm)); |
12866 | sel = force_reg (vmode, sel); | |
12867 | ||
12868 | aarch64_expand_vec_perm_1 (d->target, d->op0, d->op1, sel); | |
12869 | return true; | |
12870 | } | |
12871 | ||
12872 | static bool | |
12873 | aarch64_expand_vec_perm_const_1 (struct expand_vec_perm_d *d) | |
12874 | { | |
12875 | /* The pattern matching functions above are written to look for a small | |
12876 | number to begin the sequence (0, 1, N/2). If we begin with an index | |
12877 | from the second operand, we can swap the operands. */ | |
12878 | if (d->perm[0] >= d->nelt) | |
12879 | { | |
12880 | unsigned i, nelt = d->nelt; | |
88b08073 | 12881 | |
0696116a | 12882 | gcc_assert (nelt == (nelt & -nelt)); |
88b08073 | 12883 | for (i = 0; i < nelt; ++i) |
0696116a | 12884 | d->perm[i] ^= nelt; /* Keep the same index, but in the other vector. */ |
88b08073 | 12885 | |
cb5c6c29 | 12886 | std::swap (d->op0, d->op1); |
88b08073 JG |
12887 | } |
12888 | ||
12889 | if (TARGET_SIMD) | |
cc4d934f | 12890 | { |
923fcec3 AL |
12891 | if (aarch64_evpc_rev (d)) |
12892 | return true; | |
12893 | else if (aarch64_evpc_ext (d)) | |
ae0533da | 12894 | return true; |
f901401e AL |
12895 | else if (aarch64_evpc_dup (d)) |
12896 | return true; | |
ae0533da | 12897 | else if (aarch64_evpc_zip (d)) |
cc4d934f JG |
12898 | return true; |
12899 | else if (aarch64_evpc_uzp (d)) | |
12900 | return true; | |
12901 | else if (aarch64_evpc_trn (d)) | |
12902 | return true; | |
12903 | return aarch64_evpc_tbl (d); | |
12904 | } | |
88b08073 JG |
12905 | return false; |
12906 | } | |
12907 | ||
12908 | /* Expand a vec_perm_const pattern. */ | |
12909 | ||
12910 | bool | |
12911 | aarch64_expand_vec_perm_const (rtx target, rtx op0, rtx op1, rtx sel) | |
12912 | { | |
12913 | struct expand_vec_perm_d d; | |
12914 | int i, nelt, which; | |
12915 | ||
12916 | d.target = target; | |
12917 | d.op0 = op0; | |
12918 | d.op1 = op1; | |
12919 | ||
12920 | d.vmode = GET_MODE (target); | |
12921 | gcc_assert (VECTOR_MODE_P (d.vmode)); | |
12922 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
12923 | d.testing_p = false; | |
12924 | ||
12925 | for (i = which = 0; i < nelt; ++i) | |
12926 | { | |
12927 | rtx e = XVECEXP (sel, 0, i); | |
12928 | int ei = INTVAL (e) & (2 * nelt - 1); | |
12929 | which |= (ei < nelt ? 1 : 2); | |
12930 | d.perm[i] = ei; | |
12931 | } | |
12932 | ||
12933 | switch (which) | |
12934 | { | |
12935 | default: | |
12936 | gcc_unreachable (); | |
12937 | ||
12938 | case 3: | |
12939 | d.one_vector_p = false; | |
12940 | if (!rtx_equal_p (op0, op1)) | |
12941 | break; | |
12942 | ||
12943 | /* The elements of PERM do not suggest that only the first operand | |
12944 | is used, but both operands are identical. Allow easier matching | |
12945 | of the permutation by folding the permutation into the single | |
12946 | input vector. */ | |
12947 | /* Fall Through. */ | |
12948 | case 2: | |
12949 | for (i = 0; i < nelt; ++i) | |
12950 | d.perm[i] &= nelt - 1; | |
12951 | d.op0 = op1; | |
12952 | d.one_vector_p = true; | |
12953 | break; | |
12954 | ||
12955 | case 1: | |
12956 | d.op1 = op0; | |
12957 | d.one_vector_p = true; | |
12958 | break; | |
12959 | } | |
12960 | ||
12961 | return aarch64_expand_vec_perm_const_1 (&d); | |
12962 | } | |
12963 | ||
12964 | static bool | |
ef4bddc2 | 12965 | aarch64_vectorize_vec_perm_const_ok (machine_mode vmode, |
88b08073 JG |
12966 | const unsigned char *sel) |
12967 | { | |
12968 | struct expand_vec_perm_d d; | |
12969 | unsigned int i, nelt, which; | |
12970 | bool ret; | |
12971 | ||
12972 | d.vmode = vmode; | |
12973 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
12974 | d.testing_p = true; | |
12975 | memcpy (d.perm, sel, nelt); | |
12976 | ||
12977 | /* Calculate whether all elements are in one vector. */ | |
12978 | for (i = which = 0; i < nelt; ++i) | |
12979 | { | |
12980 | unsigned char e = d.perm[i]; | |
12981 | gcc_assert (e < 2 * nelt); | |
12982 | which |= (e < nelt ? 1 : 2); | |
12983 | } | |
12984 | ||
12985 | /* If all elements are from the second vector, reindex as if from the | |
12986 | first vector. */ | |
12987 | if (which == 2) | |
12988 | for (i = 0; i < nelt; ++i) | |
12989 | d.perm[i] -= nelt; | |
12990 | ||
12991 | /* Check whether the mask can be applied to a single vector. */ | |
12992 | d.one_vector_p = (which != 3); | |
12993 | ||
12994 | d.target = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 1); | |
12995 | d.op1 = d.op0 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 2); | |
12996 | if (!d.one_vector_p) | |
12997 | d.op1 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 3); | |
12998 | ||
12999 | start_sequence (); | |
13000 | ret = aarch64_expand_vec_perm_const_1 (&d); | |
13001 | end_sequence (); | |
13002 | ||
13003 | return ret; | |
13004 | } | |
13005 | ||
668046d1 DS |
13006 | rtx |
13007 | aarch64_reverse_mask (enum machine_mode mode) | |
13008 | { | |
13009 | /* We have to reverse each vector because we dont have | |
13010 | a permuted load that can reverse-load according to ABI rules. */ | |
13011 | rtx mask; | |
13012 | rtvec v = rtvec_alloc (16); | |
13013 | int i, j; | |
13014 | int nunits = GET_MODE_NUNITS (mode); | |
13015 | int usize = GET_MODE_UNIT_SIZE (mode); | |
13016 | ||
13017 | gcc_assert (BYTES_BIG_ENDIAN); | |
13018 | gcc_assert (AARCH64_VALID_SIMD_QREG_MODE (mode)); | |
13019 | ||
13020 | for (i = 0; i < nunits; i++) | |
13021 | for (j = 0; j < usize; j++) | |
13022 | RTVEC_ELT (v, i * usize + j) = GEN_INT ((i + 1) * usize - 1 - j); | |
13023 | mask = gen_rtx_CONST_VECTOR (V16QImode, v); | |
13024 | return force_reg (V16QImode, mask); | |
13025 | } | |
13026 | ||
61f17a5c WD |
13027 | /* Implement MODES_TIEABLE_P. In principle we should always return true. |
13028 | However due to issues with register allocation it is preferable to avoid | |
13029 | tieing integer scalar and FP scalar modes. Executing integer operations | |
13030 | in general registers is better than treating them as scalar vector | |
13031 | operations. This reduces latency and avoids redundant int<->FP moves. | |
13032 | So tie modes if they are either the same class, or vector modes with | |
13033 | other vector modes, vector structs or any scalar mode. | |
13034 | */ | |
97e1ad78 JG |
13035 | |
13036 | bool | |
ef4bddc2 | 13037 | aarch64_modes_tieable_p (machine_mode mode1, machine_mode mode2) |
97e1ad78 JG |
13038 | { |
13039 | if (GET_MODE_CLASS (mode1) == GET_MODE_CLASS (mode2)) | |
13040 | return true; | |
13041 | ||
13042 | /* We specifically want to allow elements of "structure" modes to | |
13043 | be tieable to the structure. This more general condition allows | |
13044 | other rarer situations too. */ | |
61f17a5c WD |
13045 | if (aarch64_vector_mode_p (mode1) && aarch64_vector_mode_p (mode2)) |
13046 | return true; | |
13047 | ||
13048 | /* Also allow any scalar modes with vectors. */ | |
13049 | if (aarch64_vector_mode_supported_p (mode1) | |
13050 | || aarch64_vector_mode_supported_p (mode2)) | |
97e1ad78 JG |
13051 | return true; |
13052 | ||
13053 | return false; | |
13054 | } | |
13055 | ||
e2c75eea JG |
13056 | /* Return a new RTX holding the result of moving POINTER forward by |
13057 | AMOUNT bytes. */ | |
13058 | ||
13059 | static rtx | |
13060 | aarch64_move_pointer (rtx pointer, int amount) | |
13061 | { | |
13062 | rtx next = plus_constant (Pmode, XEXP (pointer, 0), amount); | |
13063 | ||
13064 | return adjust_automodify_address (pointer, GET_MODE (pointer), | |
13065 | next, amount); | |
13066 | } | |
13067 | ||
13068 | /* Return a new RTX holding the result of moving POINTER forward by the | |
13069 | size of the mode it points to. */ | |
13070 | ||
13071 | static rtx | |
13072 | aarch64_progress_pointer (rtx pointer) | |
13073 | { | |
13074 | HOST_WIDE_INT amount = GET_MODE_SIZE (GET_MODE (pointer)); | |
13075 | ||
13076 | return aarch64_move_pointer (pointer, amount); | |
13077 | } | |
13078 | ||
13079 | /* Copy one MODE sized block from SRC to DST, then progress SRC and DST by | |
13080 | MODE bytes. */ | |
13081 | ||
13082 | static void | |
13083 | aarch64_copy_one_block_and_progress_pointers (rtx *src, rtx *dst, | |
ef4bddc2 | 13084 | machine_mode mode) |
e2c75eea JG |
13085 | { |
13086 | rtx reg = gen_reg_rtx (mode); | |
13087 | ||
13088 | /* "Cast" the pointers to the correct mode. */ | |
13089 | *src = adjust_address (*src, mode, 0); | |
13090 | *dst = adjust_address (*dst, mode, 0); | |
13091 | /* Emit the memcpy. */ | |
13092 | emit_move_insn (reg, *src); | |
13093 | emit_move_insn (*dst, reg); | |
13094 | /* Move the pointers forward. */ | |
13095 | *src = aarch64_progress_pointer (*src); | |
13096 | *dst = aarch64_progress_pointer (*dst); | |
13097 | } | |
13098 | ||
13099 | /* Expand movmem, as if from a __builtin_memcpy. Return true if | |
13100 | we succeed, otherwise return false. */ | |
13101 | ||
13102 | bool | |
13103 | aarch64_expand_movmem (rtx *operands) | |
13104 | { | |
13105 | unsigned int n; | |
13106 | rtx dst = operands[0]; | |
13107 | rtx src = operands[1]; | |
13108 | rtx base; | |
13109 | bool speed_p = !optimize_function_for_size_p (cfun); | |
13110 | ||
13111 | /* When optimizing for size, give a better estimate of the length of a | |
13112 | memcpy call, but use the default otherwise. */ | |
13113 | unsigned int max_instructions = (speed_p ? 15 : AARCH64_CALL_RATIO) / 2; | |
13114 | ||
13115 | /* We can't do anything smart if the amount to copy is not constant. */ | |
13116 | if (!CONST_INT_P (operands[2])) | |
13117 | return false; | |
13118 | ||
13119 | n = UINTVAL (operands[2]); | |
13120 | ||
13121 | /* Try to keep the number of instructions low. For cases below 16 bytes we | |
13122 | need to make at most two moves. For cases above 16 bytes it will be one | |
13123 | move for each 16 byte chunk, then at most two additional moves. */ | |
13124 | if (((n / 16) + (n % 16 ? 2 : 0)) > max_instructions) | |
13125 | return false; | |
13126 | ||
13127 | base = copy_to_mode_reg (Pmode, XEXP (dst, 0)); | |
13128 | dst = adjust_automodify_address (dst, VOIDmode, base, 0); | |
13129 | ||
13130 | base = copy_to_mode_reg (Pmode, XEXP (src, 0)); | |
13131 | src = adjust_automodify_address (src, VOIDmode, base, 0); | |
13132 | ||
13133 | /* Simple cases. Copy 0-3 bytes, as (if applicable) a 2-byte, then a | |
13134 | 1-byte chunk. */ | |
13135 | if (n < 4) | |
13136 | { | |
13137 | if (n >= 2) | |
13138 | { | |
13139 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, HImode); | |
13140 | n -= 2; | |
13141 | } | |
13142 | ||
13143 | if (n == 1) | |
13144 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, QImode); | |
13145 | ||
13146 | return true; | |
13147 | } | |
13148 | ||
13149 | /* Copy 4-8 bytes. First a 4-byte chunk, then (if applicable) a second | |
13150 | 4-byte chunk, partially overlapping with the previously copied chunk. */ | |
13151 | if (n < 8) | |
13152 | { | |
13153 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
13154 | n -= 4; | |
13155 | if (n > 0) | |
13156 | { | |
13157 | int move = n - 4; | |
13158 | ||
13159 | src = aarch64_move_pointer (src, move); | |
13160 | dst = aarch64_move_pointer (dst, move); | |
13161 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
13162 | } | |
13163 | return true; | |
13164 | } | |
13165 | ||
13166 | /* Copy more than 8 bytes. Copy chunks of 16 bytes until we run out of | |
13167 | them, then (if applicable) an 8-byte chunk. */ | |
13168 | while (n >= 8) | |
13169 | { | |
13170 | if (n / 16) | |
13171 | { | |
13172 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, TImode); | |
13173 | n -= 16; | |
13174 | } | |
13175 | else | |
13176 | { | |
13177 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, DImode); | |
13178 | n -= 8; | |
13179 | } | |
13180 | } | |
13181 | ||
13182 | /* Finish the final bytes of the copy. We can always do this in one | |
13183 | instruction. We either copy the exact amount we need, or partially | |
13184 | overlap with the previous chunk we copied and copy 8-bytes. */ | |
13185 | if (n == 0) | |
13186 | return true; | |
13187 | else if (n == 1) | |
13188 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, QImode); | |
13189 | else if (n == 2) | |
13190 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, HImode); | |
13191 | else if (n == 4) | |
13192 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
13193 | else | |
13194 | { | |
13195 | if (n == 3) | |
13196 | { | |
13197 | src = aarch64_move_pointer (src, -1); | |
13198 | dst = aarch64_move_pointer (dst, -1); | |
13199 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
13200 | } | |
13201 | else | |
13202 | { | |
13203 | int move = n - 8; | |
13204 | ||
13205 | src = aarch64_move_pointer (src, move); | |
13206 | dst = aarch64_move_pointer (dst, move); | |
13207 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, DImode); | |
13208 | } | |
13209 | } | |
13210 | ||
13211 | return true; | |
13212 | } | |
13213 | ||
141a3ccf KT |
13214 | /* Split a DImode store of a CONST_INT SRC to MEM DST as two |
13215 | SImode stores. Handle the case when the constant has identical | |
13216 | bottom and top halves. This is beneficial when the two stores can be | |
13217 | merged into an STP and we avoid synthesising potentially expensive | |
13218 | immediates twice. Return true if such a split is possible. */ | |
13219 | ||
13220 | bool | |
13221 | aarch64_split_dimode_const_store (rtx dst, rtx src) | |
13222 | { | |
13223 | rtx lo = gen_lowpart (SImode, src); | |
13224 | rtx hi = gen_highpart_mode (SImode, DImode, src); | |
13225 | ||
13226 | bool size_p = optimize_function_for_size_p (cfun); | |
13227 | ||
13228 | if (!rtx_equal_p (lo, hi)) | |
13229 | return false; | |
13230 | ||
13231 | unsigned int orig_cost | |
13232 | = aarch64_internal_mov_immediate (NULL_RTX, src, false, DImode); | |
13233 | unsigned int lo_cost | |
13234 | = aarch64_internal_mov_immediate (NULL_RTX, lo, false, SImode); | |
13235 | ||
13236 | /* We want to transform: | |
13237 | MOV x1, 49370 | |
13238 | MOVK x1, 0x140, lsl 16 | |
13239 | MOVK x1, 0xc0da, lsl 32 | |
13240 | MOVK x1, 0x140, lsl 48 | |
13241 | STR x1, [x0] | |
13242 | into: | |
13243 | MOV w1, 49370 | |
13244 | MOVK w1, 0x140, lsl 16 | |
13245 | STP w1, w1, [x0] | |
13246 | So we want to perform this only when we save two instructions | |
13247 | or more. When optimizing for size, however, accept any code size | |
13248 | savings we can. */ | |
13249 | if (size_p && orig_cost <= lo_cost) | |
13250 | return false; | |
13251 | ||
13252 | if (!size_p | |
13253 | && (orig_cost <= lo_cost + 1)) | |
13254 | return false; | |
13255 | ||
13256 | rtx mem_lo = adjust_address (dst, SImode, 0); | |
13257 | if (!aarch64_mem_pair_operand (mem_lo, SImode)) | |
13258 | return false; | |
13259 | ||
13260 | rtx tmp_reg = gen_reg_rtx (SImode); | |
13261 | aarch64_expand_mov_immediate (tmp_reg, lo); | |
13262 | rtx mem_hi = aarch64_move_pointer (mem_lo, GET_MODE_SIZE (SImode)); | |
13263 | /* Don't emit an explicit store pair as this may not be always profitable. | |
13264 | Let the sched-fusion logic decide whether to merge them. */ | |
13265 | emit_move_insn (mem_lo, tmp_reg); | |
13266 | emit_move_insn (mem_hi, tmp_reg); | |
13267 | ||
13268 | return true; | |
13269 | } | |
13270 | ||
a3125fc2 CL |
13271 | /* Implement the TARGET_ASAN_SHADOW_OFFSET hook. */ |
13272 | ||
13273 | static unsigned HOST_WIDE_INT | |
13274 | aarch64_asan_shadow_offset (void) | |
13275 | { | |
13276 | return (HOST_WIDE_INT_1 << 36); | |
13277 | } | |
13278 | ||
d3006da6 | 13279 | static bool |
445d7826 | 13280 | aarch64_use_by_pieces_infrastructure_p (unsigned HOST_WIDE_INT size, |
d3006da6 JG |
13281 | unsigned int align, |
13282 | enum by_pieces_operation op, | |
13283 | bool speed_p) | |
13284 | { | |
13285 | /* STORE_BY_PIECES can be used when copying a constant string, but | |
13286 | in that case each 64-bit chunk takes 5 insns instead of 2 (LDR/STR). | |
13287 | For now we always fail this and let the move_by_pieces code copy | |
13288 | the string from read-only memory. */ | |
13289 | if (op == STORE_BY_PIECES) | |
13290 | return false; | |
13291 | ||
13292 | return default_use_by_pieces_infrastructure_p (size, align, op, speed_p); | |
13293 | } | |
13294 | ||
5f3bc026 | 13295 | static rtx |
cb4347e8 | 13296 | aarch64_gen_ccmp_first (rtx_insn **prep_seq, rtx_insn **gen_seq, |
5f3bc026 ZC |
13297 | int code, tree treeop0, tree treeop1) |
13298 | { | |
c8012fbc WD |
13299 | machine_mode op_mode, cmp_mode, cc_mode = CCmode; |
13300 | rtx op0, op1; | |
5f3bc026 | 13301 | int unsignedp = TYPE_UNSIGNED (TREE_TYPE (treeop0)); |
c8012fbc | 13302 | insn_code icode; |
5f3bc026 ZC |
13303 | struct expand_operand ops[4]; |
13304 | ||
5f3bc026 ZC |
13305 | start_sequence (); |
13306 | expand_operands (treeop0, treeop1, NULL_RTX, &op0, &op1, EXPAND_NORMAL); | |
13307 | ||
13308 | op_mode = GET_MODE (op0); | |
13309 | if (op_mode == VOIDmode) | |
13310 | op_mode = GET_MODE (op1); | |
13311 | ||
13312 | switch (op_mode) | |
13313 | { | |
13314 | case QImode: | |
13315 | case HImode: | |
13316 | case SImode: | |
13317 | cmp_mode = SImode; | |
13318 | icode = CODE_FOR_cmpsi; | |
13319 | break; | |
13320 | ||
13321 | case DImode: | |
13322 | cmp_mode = DImode; | |
13323 | icode = CODE_FOR_cmpdi; | |
13324 | break; | |
13325 | ||
786e3c06 WD |
13326 | case SFmode: |
13327 | cmp_mode = SFmode; | |
13328 | cc_mode = aarch64_select_cc_mode ((rtx_code) code, op0, op1); | |
13329 | icode = cc_mode == CCFPEmode ? CODE_FOR_fcmpesf : CODE_FOR_fcmpsf; | |
13330 | break; | |
13331 | ||
13332 | case DFmode: | |
13333 | cmp_mode = DFmode; | |
13334 | cc_mode = aarch64_select_cc_mode ((rtx_code) code, op0, op1); | |
13335 | icode = cc_mode == CCFPEmode ? CODE_FOR_fcmpedf : CODE_FOR_fcmpdf; | |
13336 | break; | |
13337 | ||
5f3bc026 ZC |
13338 | default: |
13339 | end_sequence (); | |
13340 | return NULL_RTX; | |
13341 | } | |
13342 | ||
c8012fbc WD |
13343 | op0 = prepare_operand (icode, op0, 0, op_mode, cmp_mode, unsignedp); |
13344 | op1 = prepare_operand (icode, op1, 1, op_mode, cmp_mode, unsignedp); | |
5f3bc026 ZC |
13345 | if (!op0 || !op1) |
13346 | { | |
13347 | end_sequence (); | |
13348 | return NULL_RTX; | |
13349 | } | |
13350 | *prep_seq = get_insns (); | |
13351 | end_sequence (); | |
13352 | ||
c8012fbc WD |
13353 | create_fixed_operand (&ops[0], op0); |
13354 | create_fixed_operand (&ops[1], op1); | |
5f3bc026 ZC |
13355 | |
13356 | start_sequence (); | |
c8012fbc | 13357 | if (!maybe_expand_insn (icode, 2, ops)) |
5f3bc026 ZC |
13358 | { |
13359 | end_sequence (); | |
13360 | return NULL_RTX; | |
13361 | } | |
13362 | *gen_seq = get_insns (); | |
13363 | end_sequence (); | |
13364 | ||
c8012fbc WD |
13365 | return gen_rtx_fmt_ee ((rtx_code) code, cc_mode, |
13366 | gen_rtx_REG (cc_mode, CC_REGNUM), const0_rtx); | |
5f3bc026 ZC |
13367 | } |
13368 | ||
13369 | static rtx | |
cb4347e8 TS |
13370 | aarch64_gen_ccmp_next (rtx_insn **prep_seq, rtx_insn **gen_seq, rtx prev, |
13371 | int cmp_code, tree treeop0, tree treeop1, int bit_code) | |
5f3bc026 | 13372 | { |
c8012fbc WD |
13373 | rtx op0, op1, target; |
13374 | machine_mode op_mode, cmp_mode, cc_mode = CCmode; | |
5f3bc026 | 13375 | int unsignedp = TYPE_UNSIGNED (TREE_TYPE (treeop0)); |
c8012fbc | 13376 | insn_code icode; |
5f3bc026 | 13377 | struct expand_operand ops[6]; |
c8012fbc | 13378 | int aarch64_cond; |
5f3bc026 | 13379 | |
cb4347e8 | 13380 | push_to_sequence (*prep_seq); |
5f3bc026 ZC |
13381 | expand_operands (treeop0, treeop1, NULL_RTX, &op0, &op1, EXPAND_NORMAL); |
13382 | ||
13383 | op_mode = GET_MODE (op0); | |
13384 | if (op_mode == VOIDmode) | |
13385 | op_mode = GET_MODE (op1); | |
13386 | ||
13387 | switch (op_mode) | |
13388 | { | |
13389 | case QImode: | |
13390 | case HImode: | |
13391 | case SImode: | |
13392 | cmp_mode = SImode; | |
c8012fbc | 13393 | icode = CODE_FOR_ccmpsi; |
5f3bc026 ZC |
13394 | break; |
13395 | ||
13396 | case DImode: | |
13397 | cmp_mode = DImode; | |
c8012fbc | 13398 | icode = CODE_FOR_ccmpdi; |
5f3bc026 ZC |
13399 | break; |
13400 | ||
786e3c06 WD |
13401 | case SFmode: |
13402 | cmp_mode = SFmode; | |
13403 | cc_mode = aarch64_select_cc_mode ((rtx_code) cmp_code, op0, op1); | |
13404 | icode = cc_mode == CCFPEmode ? CODE_FOR_fccmpesf : CODE_FOR_fccmpsf; | |
13405 | break; | |
13406 | ||
13407 | case DFmode: | |
13408 | cmp_mode = DFmode; | |
13409 | cc_mode = aarch64_select_cc_mode ((rtx_code) cmp_code, op0, op1); | |
13410 | icode = cc_mode == CCFPEmode ? CODE_FOR_fccmpedf : CODE_FOR_fccmpdf; | |
13411 | break; | |
13412 | ||
5f3bc026 ZC |
13413 | default: |
13414 | end_sequence (); | |
13415 | return NULL_RTX; | |
13416 | } | |
13417 | ||
13418 | op0 = prepare_operand (icode, op0, 2, op_mode, cmp_mode, unsignedp); | |
13419 | op1 = prepare_operand (icode, op1, 3, op_mode, cmp_mode, unsignedp); | |
13420 | if (!op0 || !op1) | |
13421 | { | |
13422 | end_sequence (); | |
13423 | return NULL_RTX; | |
13424 | } | |
13425 | *prep_seq = get_insns (); | |
13426 | end_sequence (); | |
13427 | ||
13428 | target = gen_rtx_REG (cc_mode, CC_REGNUM); | |
c8012fbc | 13429 | aarch64_cond = aarch64_get_condition_code_1 (cc_mode, (rtx_code) cmp_code); |
5f3bc026 | 13430 | |
c8012fbc WD |
13431 | if (bit_code != AND) |
13432 | { | |
13433 | prev = gen_rtx_fmt_ee (REVERSE_CONDITION (GET_CODE (prev), | |
13434 | GET_MODE (XEXP (prev, 0))), | |
13435 | VOIDmode, XEXP (prev, 0), const0_rtx); | |
13436 | aarch64_cond = AARCH64_INVERSE_CONDITION_CODE (aarch64_cond); | |
13437 | } | |
13438 | ||
13439 | create_fixed_operand (&ops[0], XEXP (prev, 0)); | |
5f3bc026 ZC |
13440 | create_fixed_operand (&ops[1], target); |
13441 | create_fixed_operand (&ops[2], op0); | |
13442 | create_fixed_operand (&ops[3], op1); | |
c8012fbc WD |
13443 | create_fixed_operand (&ops[4], prev); |
13444 | create_fixed_operand (&ops[5], GEN_INT (aarch64_cond)); | |
5f3bc026 | 13445 | |
cb4347e8 | 13446 | push_to_sequence (*gen_seq); |
5f3bc026 ZC |
13447 | if (!maybe_expand_insn (icode, 6, ops)) |
13448 | { | |
13449 | end_sequence (); | |
13450 | return NULL_RTX; | |
13451 | } | |
13452 | ||
13453 | *gen_seq = get_insns (); | |
13454 | end_sequence (); | |
13455 | ||
c8012fbc | 13456 | return gen_rtx_fmt_ee ((rtx_code) cmp_code, VOIDmode, target, const0_rtx); |
5f3bc026 ZC |
13457 | } |
13458 | ||
13459 | #undef TARGET_GEN_CCMP_FIRST | |
13460 | #define TARGET_GEN_CCMP_FIRST aarch64_gen_ccmp_first | |
13461 | ||
13462 | #undef TARGET_GEN_CCMP_NEXT | |
13463 | #define TARGET_GEN_CCMP_NEXT aarch64_gen_ccmp_next | |
13464 | ||
6a569cdd KT |
13465 | /* Implement TARGET_SCHED_MACRO_FUSION_P. Return true if target supports |
13466 | instruction fusion of some sort. */ | |
13467 | ||
13468 | static bool | |
13469 | aarch64_macro_fusion_p (void) | |
13470 | { | |
b175b679 | 13471 | return aarch64_tune_params.fusible_ops != AARCH64_FUSE_NOTHING; |
6a569cdd KT |
13472 | } |
13473 | ||
13474 | ||
13475 | /* Implement TARGET_SCHED_MACRO_FUSION_PAIR_P. Return true if PREV and CURR | |
13476 | should be kept together during scheduling. */ | |
13477 | ||
13478 | static bool | |
13479 | aarch_macro_fusion_pair_p (rtx_insn *prev, rtx_insn *curr) | |
13480 | { | |
13481 | rtx set_dest; | |
13482 | rtx prev_set = single_set (prev); | |
13483 | rtx curr_set = single_set (curr); | |
13484 | /* prev and curr are simple SET insns i.e. no flag setting or branching. */ | |
13485 | bool simple_sets_p = prev_set && curr_set && !any_condjump_p (curr); | |
13486 | ||
13487 | if (!aarch64_macro_fusion_p ()) | |
13488 | return false; | |
13489 | ||
d7b03373 | 13490 | if (simple_sets_p && aarch64_fusion_enabled_p (AARCH64_FUSE_MOV_MOVK)) |
6a569cdd KT |
13491 | { |
13492 | /* We are trying to match: | |
13493 | prev (mov) == (set (reg r0) (const_int imm16)) | |
13494 | curr (movk) == (set (zero_extract (reg r0) | |
13495 | (const_int 16) | |
13496 | (const_int 16)) | |
13497 | (const_int imm16_1)) */ | |
13498 | ||
13499 | set_dest = SET_DEST (curr_set); | |
13500 | ||
13501 | if (GET_CODE (set_dest) == ZERO_EXTRACT | |
13502 | && CONST_INT_P (SET_SRC (curr_set)) | |
13503 | && CONST_INT_P (SET_SRC (prev_set)) | |
13504 | && CONST_INT_P (XEXP (set_dest, 2)) | |
13505 | && INTVAL (XEXP (set_dest, 2)) == 16 | |
13506 | && REG_P (XEXP (set_dest, 0)) | |
13507 | && REG_P (SET_DEST (prev_set)) | |
13508 | && REGNO (XEXP (set_dest, 0)) == REGNO (SET_DEST (prev_set))) | |
13509 | { | |
13510 | return true; | |
13511 | } | |
13512 | } | |
13513 | ||
d7b03373 | 13514 | if (simple_sets_p && aarch64_fusion_enabled_p (AARCH64_FUSE_ADRP_ADD)) |
9bbe08fe KT |
13515 | { |
13516 | ||
13517 | /* We're trying to match: | |
13518 | prev (adrp) == (set (reg r1) | |
13519 | (high (symbol_ref ("SYM")))) | |
13520 | curr (add) == (set (reg r0) | |
13521 | (lo_sum (reg r1) | |
13522 | (symbol_ref ("SYM")))) | |
13523 | Note that r0 need not necessarily be the same as r1, especially | |
13524 | during pre-regalloc scheduling. */ | |
13525 | ||
13526 | if (satisfies_constraint_Ush (SET_SRC (prev_set)) | |
13527 | && REG_P (SET_DEST (prev_set)) && REG_P (SET_DEST (curr_set))) | |
13528 | { | |
13529 | if (GET_CODE (SET_SRC (curr_set)) == LO_SUM | |
13530 | && REG_P (XEXP (SET_SRC (curr_set), 0)) | |
13531 | && REGNO (XEXP (SET_SRC (curr_set), 0)) | |
13532 | == REGNO (SET_DEST (prev_set)) | |
13533 | && rtx_equal_p (XEXP (SET_SRC (prev_set), 0), | |
13534 | XEXP (SET_SRC (curr_set), 1))) | |
13535 | return true; | |
13536 | } | |
13537 | } | |
13538 | ||
d7b03373 | 13539 | if (simple_sets_p && aarch64_fusion_enabled_p (AARCH64_FUSE_MOVK_MOVK)) |
cd0cb232 KT |
13540 | { |
13541 | ||
13542 | /* We're trying to match: | |
13543 | prev (movk) == (set (zero_extract (reg r0) | |
13544 | (const_int 16) | |
13545 | (const_int 32)) | |
13546 | (const_int imm16_1)) | |
13547 | curr (movk) == (set (zero_extract (reg r0) | |
13548 | (const_int 16) | |
13549 | (const_int 48)) | |
13550 | (const_int imm16_2)) */ | |
13551 | ||
13552 | if (GET_CODE (SET_DEST (prev_set)) == ZERO_EXTRACT | |
13553 | && GET_CODE (SET_DEST (curr_set)) == ZERO_EXTRACT | |
13554 | && REG_P (XEXP (SET_DEST (prev_set), 0)) | |
13555 | && REG_P (XEXP (SET_DEST (curr_set), 0)) | |
13556 | && REGNO (XEXP (SET_DEST (prev_set), 0)) | |
13557 | == REGNO (XEXP (SET_DEST (curr_set), 0)) | |
13558 | && CONST_INT_P (XEXP (SET_DEST (prev_set), 2)) | |
13559 | && CONST_INT_P (XEXP (SET_DEST (curr_set), 2)) | |
13560 | && INTVAL (XEXP (SET_DEST (prev_set), 2)) == 32 | |
13561 | && INTVAL (XEXP (SET_DEST (curr_set), 2)) == 48 | |
13562 | && CONST_INT_P (SET_SRC (prev_set)) | |
13563 | && CONST_INT_P (SET_SRC (curr_set))) | |
13564 | return true; | |
13565 | ||
13566 | } | |
d7b03373 | 13567 | if (simple_sets_p && aarch64_fusion_enabled_p (AARCH64_FUSE_ADRP_LDR)) |
d8354ad7 KT |
13568 | { |
13569 | /* We're trying to match: | |
13570 | prev (adrp) == (set (reg r0) | |
13571 | (high (symbol_ref ("SYM")))) | |
13572 | curr (ldr) == (set (reg r1) | |
13573 | (mem (lo_sum (reg r0) | |
13574 | (symbol_ref ("SYM"))))) | |
13575 | or | |
13576 | curr (ldr) == (set (reg r1) | |
13577 | (zero_extend (mem | |
13578 | (lo_sum (reg r0) | |
13579 | (symbol_ref ("SYM")))))) */ | |
13580 | if (satisfies_constraint_Ush (SET_SRC (prev_set)) | |
13581 | && REG_P (SET_DEST (prev_set)) && REG_P (SET_DEST (curr_set))) | |
13582 | { | |
13583 | rtx curr_src = SET_SRC (curr_set); | |
13584 | ||
13585 | if (GET_CODE (curr_src) == ZERO_EXTEND) | |
13586 | curr_src = XEXP (curr_src, 0); | |
13587 | ||
13588 | if (MEM_P (curr_src) && GET_CODE (XEXP (curr_src, 0)) == LO_SUM | |
13589 | && REG_P (XEXP (XEXP (curr_src, 0), 0)) | |
13590 | && REGNO (XEXP (XEXP (curr_src, 0), 0)) | |
13591 | == REGNO (SET_DEST (prev_set)) | |
13592 | && rtx_equal_p (XEXP (XEXP (curr_src, 0), 1), | |
13593 | XEXP (SET_SRC (prev_set), 0))) | |
13594 | return true; | |
13595 | } | |
13596 | } | |
cd0cb232 | 13597 | |
d7b03373 | 13598 | if (aarch64_fusion_enabled_p (AARCH64_FUSE_AES_AESMC) |
00a8574a WD |
13599 | && aarch_crypto_can_dual_issue (prev, curr)) |
13600 | return true; | |
13601 | ||
d7b03373 | 13602 | if (aarch64_fusion_enabled_p (AARCH64_FUSE_CMP_BRANCH) |
3759108f AP |
13603 | && any_condjump_p (curr)) |
13604 | { | |
13605 | enum attr_type prev_type = get_attr_type (prev); | |
13606 | ||
13607 | /* FIXME: this misses some which is considered simple arthematic | |
13608 | instructions for ThunderX. Simple shifts are missed here. */ | |
13609 | if (prev_type == TYPE_ALUS_SREG | |
13610 | || prev_type == TYPE_ALUS_IMM | |
13611 | || prev_type == TYPE_LOGICS_REG | |
13612 | || prev_type == TYPE_LOGICS_IMM) | |
13613 | return true; | |
13614 | } | |
13615 | ||
6a569cdd KT |
13616 | return false; |
13617 | } | |
13618 | ||
f2879a90 KT |
13619 | /* Return true iff the instruction fusion described by OP is enabled. */ |
13620 | ||
13621 | bool | |
13622 | aarch64_fusion_enabled_p (enum aarch64_fusion_pairs op) | |
13623 | { | |
13624 | return (aarch64_tune_params.fusible_ops & op) != 0; | |
13625 | } | |
13626 | ||
350013bc BC |
13627 | /* If MEM is in the form of [base+offset], extract the two parts |
13628 | of address and set to BASE and OFFSET, otherwise return false | |
13629 | after clearing BASE and OFFSET. */ | |
13630 | ||
13631 | bool | |
13632 | extract_base_offset_in_addr (rtx mem, rtx *base, rtx *offset) | |
13633 | { | |
13634 | rtx addr; | |
13635 | ||
13636 | gcc_assert (MEM_P (mem)); | |
13637 | ||
13638 | addr = XEXP (mem, 0); | |
13639 | ||
13640 | if (REG_P (addr)) | |
13641 | { | |
13642 | *base = addr; | |
13643 | *offset = const0_rtx; | |
13644 | return true; | |
13645 | } | |
13646 | ||
13647 | if (GET_CODE (addr) == PLUS | |
13648 | && REG_P (XEXP (addr, 0)) && CONST_INT_P (XEXP (addr, 1))) | |
13649 | { | |
13650 | *base = XEXP (addr, 0); | |
13651 | *offset = XEXP (addr, 1); | |
13652 | return true; | |
13653 | } | |
13654 | ||
13655 | *base = NULL_RTX; | |
13656 | *offset = NULL_RTX; | |
13657 | ||
13658 | return false; | |
13659 | } | |
13660 | ||
13661 | /* Types for scheduling fusion. */ | |
13662 | enum sched_fusion_type | |
13663 | { | |
13664 | SCHED_FUSION_NONE = 0, | |
13665 | SCHED_FUSION_LD_SIGN_EXTEND, | |
13666 | SCHED_FUSION_LD_ZERO_EXTEND, | |
13667 | SCHED_FUSION_LD, | |
13668 | SCHED_FUSION_ST, | |
13669 | SCHED_FUSION_NUM | |
13670 | }; | |
13671 | ||
13672 | /* If INSN is a load or store of address in the form of [base+offset], | |
13673 | extract the two parts and set to BASE and OFFSET. Return scheduling | |
13674 | fusion type this INSN is. */ | |
13675 | ||
13676 | static enum sched_fusion_type | |
13677 | fusion_load_store (rtx_insn *insn, rtx *base, rtx *offset) | |
13678 | { | |
13679 | rtx x, dest, src; | |
13680 | enum sched_fusion_type fusion = SCHED_FUSION_LD; | |
13681 | ||
13682 | gcc_assert (INSN_P (insn)); | |
13683 | x = PATTERN (insn); | |
13684 | if (GET_CODE (x) != SET) | |
13685 | return SCHED_FUSION_NONE; | |
13686 | ||
13687 | src = SET_SRC (x); | |
13688 | dest = SET_DEST (x); | |
13689 | ||
abc52318 KT |
13690 | machine_mode dest_mode = GET_MODE (dest); |
13691 | ||
13692 | if (!aarch64_mode_valid_for_sched_fusion_p (dest_mode)) | |
350013bc BC |
13693 | return SCHED_FUSION_NONE; |
13694 | ||
13695 | if (GET_CODE (src) == SIGN_EXTEND) | |
13696 | { | |
13697 | fusion = SCHED_FUSION_LD_SIGN_EXTEND; | |
13698 | src = XEXP (src, 0); | |
13699 | if (GET_CODE (src) != MEM || GET_MODE (src) != SImode) | |
13700 | return SCHED_FUSION_NONE; | |
13701 | } | |
13702 | else if (GET_CODE (src) == ZERO_EXTEND) | |
13703 | { | |
13704 | fusion = SCHED_FUSION_LD_ZERO_EXTEND; | |
13705 | src = XEXP (src, 0); | |
13706 | if (GET_CODE (src) != MEM || GET_MODE (src) != SImode) | |
13707 | return SCHED_FUSION_NONE; | |
13708 | } | |
13709 | ||
13710 | if (GET_CODE (src) == MEM && REG_P (dest)) | |
13711 | extract_base_offset_in_addr (src, base, offset); | |
13712 | else if (GET_CODE (dest) == MEM && (REG_P (src) || src == const0_rtx)) | |
13713 | { | |
13714 | fusion = SCHED_FUSION_ST; | |
13715 | extract_base_offset_in_addr (dest, base, offset); | |
13716 | } | |
13717 | else | |
13718 | return SCHED_FUSION_NONE; | |
13719 | ||
13720 | if (*base == NULL_RTX || *offset == NULL_RTX) | |
13721 | fusion = SCHED_FUSION_NONE; | |
13722 | ||
13723 | return fusion; | |
13724 | } | |
13725 | ||
13726 | /* Implement the TARGET_SCHED_FUSION_PRIORITY hook. | |
13727 | ||
13728 | Currently we only support to fuse ldr or str instructions, so FUSION_PRI | |
13729 | and PRI are only calculated for these instructions. For other instruction, | |
13730 | FUSION_PRI and PRI are simply set to MAX_PRI - 1. In the future, other | |
13731 | type instruction fusion can be added by returning different priorities. | |
13732 | ||
13733 | It's important that irrelevant instructions get the largest FUSION_PRI. */ | |
13734 | ||
13735 | static void | |
13736 | aarch64_sched_fusion_priority (rtx_insn *insn, int max_pri, | |
13737 | int *fusion_pri, int *pri) | |
13738 | { | |
13739 | int tmp, off_val; | |
13740 | rtx base, offset; | |
13741 | enum sched_fusion_type fusion; | |
13742 | ||
13743 | gcc_assert (INSN_P (insn)); | |
13744 | ||
13745 | tmp = max_pri - 1; | |
13746 | fusion = fusion_load_store (insn, &base, &offset); | |
13747 | if (fusion == SCHED_FUSION_NONE) | |
13748 | { | |
13749 | *pri = tmp; | |
13750 | *fusion_pri = tmp; | |
13751 | return; | |
13752 | } | |
13753 | ||
13754 | /* Set FUSION_PRI according to fusion type and base register. */ | |
13755 | *fusion_pri = tmp - fusion * FIRST_PSEUDO_REGISTER - REGNO (base); | |
13756 | ||
13757 | /* Calculate PRI. */ | |
13758 | tmp /= 2; | |
13759 | ||
13760 | /* INSN with smaller offset goes first. */ | |
13761 | off_val = (int)(INTVAL (offset)); | |
13762 | if (off_val >= 0) | |
13763 | tmp -= (off_val & 0xfffff); | |
13764 | else | |
13765 | tmp += ((- off_val) & 0xfffff); | |
13766 | ||
13767 | *pri = tmp; | |
13768 | return; | |
13769 | } | |
13770 | ||
13771 | /* Given OPERANDS of consecutive load/store, check if we can merge | |
13772 | them into ldp/stp. LOAD is true if they are load instructions. | |
13773 | MODE is the mode of memory operands. */ | |
13774 | ||
13775 | bool | |
13776 | aarch64_operands_ok_for_ldpstp (rtx *operands, bool load, | |
13777 | enum machine_mode mode) | |
13778 | { | |
13779 | HOST_WIDE_INT offval_1, offval_2, msize; | |
13780 | enum reg_class rclass_1, rclass_2; | |
13781 | rtx mem_1, mem_2, reg_1, reg_2, base_1, base_2, offset_1, offset_2; | |
13782 | ||
13783 | if (load) | |
13784 | { | |
13785 | mem_1 = operands[1]; | |
13786 | mem_2 = operands[3]; | |
13787 | reg_1 = operands[0]; | |
13788 | reg_2 = operands[2]; | |
13789 | gcc_assert (REG_P (reg_1) && REG_P (reg_2)); | |
13790 | if (REGNO (reg_1) == REGNO (reg_2)) | |
13791 | return false; | |
13792 | } | |
13793 | else | |
13794 | { | |
13795 | mem_1 = operands[0]; | |
13796 | mem_2 = operands[2]; | |
13797 | reg_1 = operands[1]; | |
13798 | reg_2 = operands[3]; | |
13799 | } | |
13800 | ||
bf84ac44 AP |
13801 | /* The mems cannot be volatile. */ |
13802 | if (MEM_VOLATILE_P (mem_1) || MEM_VOLATILE_P (mem_2)) | |
13803 | return false; | |
13804 | ||
54700e2e AP |
13805 | /* If we have SImode and slow unaligned ldp, |
13806 | check the alignment to be at least 8 byte. */ | |
13807 | if (mode == SImode | |
13808 | && (aarch64_tune_params.extra_tuning_flags | |
13809 | & AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW) | |
13810 | && !optimize_size | |
13811 | && MEM_ALIGN (mem_1) < 8 * BITS_PER_UNIT) | |
13812 | return false; | |
13813 | ||
350013bc BC |
13814 | /* Check if the addresses are in the form of [base+offset]. */ |
13815 | extract_base_offset_in_addr (mem_1, &base_1, &offset_1); | |
13816 | if (base_1 == NULL_RTX || offset_1 == NULL_RTX) | |
13817 | return false; | |
13818 | extract_base_offset_in_addr (mem_2, &base_2, &offset_2); | |
13819 | if (base_2 == NULL_RTX || offset_2 == NULL_RTX) | |
13820 | return false; | |
13821 | ||
13822 | /* Check if the bases are same. */ | |
13823 | if (!rtx_equal_p (base_1, base_2)) | |
13824 | return false; | |
13825 | ||
13826 | offval_1 = INTVAL (offset_1); | |
13827 | offval_2 = INTVAL (offset_2); | |
13828 | msize = GET_MODE_SIZE (mode); | |
13829 | /* Check if the offsets are consecutive. */ | |
13830 | if (offval_1 != (offval_2 + msize) && offval_2 != (offval_1 + msize)) | |
13831 | return false; | |
13832 | ||
13833 | /* Check if the addresses are clobbered by load. */ | |
13834 | if (load) | |
13835 | { | |
13836 | if (reg_mentioned_p (reg_1, mem_1)) | |
13837 | return false; | |
13838 | ||
13839 | /* In increasing order, the last load can clobber the address. */ | |
13840 | if (offval_1 > offval_2 && reg_mentioned_p (reg_2, mem_2)) | |
13841 | return false; | |
13842 | } | |
13843 | ||
13844 | if (REG_P (reg_1) && FP_REGNUM_P (REGNO (reg_1))) | |
13845 | rclass_1 = FP_REGS; | |
13846 | else | |
13847 | rclass_1 = GENERAL_REGS; | |
13848 | ||
13849 | if (REG_P (reg_2) && FP_REGNUM_P (REGNO (reg_2))) | |
13850 | rclass_2 = FP_REGS; | |
13851 | else | |
13852 | rclass_2 = GENERAL_REGS; | |
13853 | ||
13854 | /* Check if the registers are of same class. */ | |
13855 | if (rclass_1 != rclass_2) | |
13856 | return false; | |
13857 | ||
13858 | return true; | |
13859 | } | |
13860 | ||
13861 | /* Given OPERANDS of consecutive load/store, check if we can merge | |
13862 | them into ldp/stp by adjusting the offset. LOAD is true if they | |
13863 | are load instructions. MODE is the mode of memory operands. | |
13864 | ||
13865 | Given below consecutive stores: | |
13866 | ||
13867 | str w1, [xb, 0x100] | |
13868 | str w1, [xb, 0x104] | |
13869 | str w1, [xb, 0x108] | |
13870 | str w1, [xb, 0x10c] | |
13871 | ||
13872 | Though the offsets are out of the range supported by stp, we can | |
13873 | still pair them after adjusting the offset, like: | |
13874 | ||
13875 | add scratch, xb, 0x100 | |
13876 | stp w1, w1, [scratch] | |
13877 | stp w1, w1, [scratch, 0x8] | |
13878 | ||
13879 | The peephole patterns detecting this opportunity should guarantee | |
13880 | the scratch register is avaliable. */ | |
13881 | ||
13882 | bool | |
13883 | aarch64_operands_adjust_ok_for_ldpstp (rtx *operands, bool load, | |
13884 | enum machine_mode mode) | |
13885 | { | |
13886 | enum reg_class rclass_1, rclass_2, rclass_3, rclass_4; | |
13887 | HOST_WIDE_INT offval_1, offval_2, offval_3, offval_4, msize; | |
13888 | rtx mem_1, mem_2, mem_3, mem_4, reg_1, reg_2, reg_3, reg_4; | |
13889 | rtx base_1, base_2, base_3, base_4, offset_1, offset_2, offset_3, offset_4; | |
13890 | ||
13891 | if (load) | |
13892 | { | |
13893 | reg_1 = operands[0]; | |
13894 | mem_1 = operands[1]; | |
13895 | reg_2 = operands[2]; | |
13896 | mem_2 = operands[3]; | |
13897 | reg_3 = operands[4]; | |
13898 | mem_3 = operands[5]; | |
13899 | reg_4 = operands[6]; | |
13900 | mem_4 = operands[7]; | |
13901 | gcc_assert (REG_P (reg_1) && REG_P (reg_2) | |
13902 | && REG_P (reg_3) && REG_P (reg_4)); | |
13903 | if (REGNO (reg_1) == REGNO (reg_2) || REGNO (reg_3) == REGNO (reg_4)) | |
13904 | return false; | |
13905 | } | |
13906 | else | |
13907 | { | |
13908 | mem_1 = operands[0]; | |
13909 | reg_1 = operands[1]; | |
13910 | mem_2 = operands[2]; | |
13911 | reg_2 = operands[3]; | |
13912 | mem_3 = operands[4]; | |
13913 | reg_3 = operands[5]; | |
13914 | mem_4 = operands[6]; | |
13915 | reg_4 = operands[7]; | |
13916 | } | |
13917 | /* Skip if memory operand is by itslef valid for ldp/stp. */ | |
13918 | if (!MEM_P (mem_1) || aarch64_mem_pair_operand (mem_1, mode)) | |
13919 | return false; | |
13920 | ||
bf84ac44 AP |
13921 | /* The mems cannot be volatile. */ |
13922 | if (MEM_VOLATILE_P (mem_1) || MEM_VOLATILE_P (mem_2) | |
13923 | || MEM_VOLATILE_P (mem_3) ||MEM_VOLATILE_P (mem_4)) | |
13924 | return false; | |
13925 | ||
350013bc BC |
13926 | /* Check if the addresses are in the form of [base+offset]. */ |
13927 | extract_base_offset_in_addr (mem_1, &base_1, &offset_1); | |
13928 | if (base_1 == NULL_RTX || offset_1 == NULL_RTX) | |
13929 | return false; | |
13930 | extract_base_offset_in_addr (mem_2, &base_2, &offset_2); | |
13931 | if (base_2 == NULL_RTX || offset_2 == NULL_RTX) | |
13932 | return false; | |
13933 | extract_base_offset_in_addr (mem_3, &base_3, &offset_3); | |
13934 | if (base_3 == NULL_RTX || offset_3 == NULL_RTX) | |
13935 | return false; | |
13936 | extract_base_offset_in_addr (mem_4, &base_4, &offset_4); | |
13937 | if (base_4 == NULL_RTX || offset_4 == NULL_RTX) | |
13938 | return false; | |
13939 | ||
13940 | /* Check if the bases are same. */ | |
13941 | if (!rtx_equal_p (base_1, base_2) | |
13942 | || !rtx_equal_p (base_2, base_3) | |
13943 | || !rtx_equal_p (base_3, base_4)) | |
13944 | return false; | |
13945 | ||
13946 | offval_1 = INTVAL (offset_1); | |
13947 | offval_2 = INTVAL (offset_2); | |
13948 | offval_3 = INTVAL (offset_3); | |
13949 | offval_4 = INTVAL (offset_4); | |
13950 | msize = GET_MODE_SIZE (mode); | |
13951 | /* Check if the offsets are consecutive. */ | |
13952 | if ((offval_1 != (offval_2 + msize) | |
13953 | || offval_1 != (offval_3 + msize * 2) | |
13954 | || offval_1 != (offval_4 + msize * 3)) | |
13955 | && (offval_4 != (offval_3 + msize) | |
13956 | || offval_4 != (offval_2 + msize * 2) | |
13957 | || offval_4 != (offval_1 + msize * 3))) | |
13958 | return false; | |
13959 | ||
13960 | /* Check if the addresses are clobbered by load. */ | |
13961 | if (load) | |
13962 | { | |
13963 | if (reg_mentioned_p (reg_1, mem_1) | |
13964 | || reg_mentioned_p (reg_2, mem_2) | |
13965 | || reg_mentioned_p (reg_3, mem_3)) | |
13966 | return false; | |
13967 | ||
13968 | /* In increasing order, the last load can clobber the address. */ | |
13969 | if (offval_1 > offval_2 && reg_mentioned_p (reg_4, mem_4)) | |
13970 | return false; | |
13971 | } | |
13972 | ||
54700e2e AP |
13973 | /* If we have SImode and slow unaligned ldp, |
13974 | check the alignment to be at least 8 byte. */ | |
13975 | if (mode == SImode | |
13976 | && (aarch64_tune_params.extra_tuning_flags | |
13977 | & AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW) | |
13978 | && !optimize_size | |
13979 | && MEM_ALIGN (mem_1) < 8 * BITS_PER_UNIT) | |
13980 | return false; | |
13981 | ||
350013bc BC |
13982 | if (REG_P (reg_1) && FP_REGNUM_P (REGNO (reg_1))) |
13983 | rclass_1 = FP_REGS; | |
13984 | else | |
13985 | rclass_1 = GENERAL_REGS; | |
13986 | ||
13987 | if (REG_P (reg_2) && FP_REGNUM_P (REGNO (reg_2))) | |
13988 | rclass_2 = FP_REGS; | |
13989 | else | |
13990 | rclass_2 = GENERAL_REGS; | |
13991 | ||
13992 | if (REG_P (reg_3) && FP_REGNUM_P (REGNO (reg_3))) | |
13993 | rclass_3 = FP_REGS; | |
13994 | else | |
13995 | rclass_3 = GENERAL_REGS; | |
13996 | ||
13997 | if (REG_P (reg_4) && FP_REGNUM_P (REGNO (reg_4))) | |
13998 | rclass_4 = FP_REGS; | |
13999 | else | |
14000 | rclass_4 = GENERAL_REGS; | |
14001 | ||
14002 | /* Check if the registers are of same class. */ | |
14003 | if (rclass_1 != rclass_2 || rclass_2 != rclass_3 || rclass_3 != rclass_4) | |
14004 | return false; | |
14005 | ||
14006 | return true; | |
14007 | } | |
14008 | ||
14009 | /* Given OPERANDS of consecutive load/store, this function pairs them | |
14010 | into ldp/stp after adjusting the offset. It depends on the fact | |
14011 | that addresses of load/store instructions are in increasing order. | |
14012 | MODE is the mode of memory operands. CODE is the rtl operator | |
14013 | which should be applied to all memory operands, it's SIGN_EXTEND, | |
14014 | ZERO_EXTEND or UNKNOWN. */ | |
14015 | ||
14016 | bool | |
14017 | aarch64_gen_adjusted_ldpstp (rtx *operands, bool load, | |
14018 | enum machine_mode mode, RTX_CODE code) | |
14019 | { | |
14020 | rtx base, offset, t1, t2; | |
14021 | rtx mem_1, mem_2, mem_3, mem_4; | |
14022 | HOST_WIDE_INT off_val, abs_off, adj_off, new_off, stp_off_limit, msize; | |
14023 | ||
14024 | if (load) | |
14025 | { | |
14026 | mem_1 = operands[1]; | |
14027 | mem_2 = operands[3]; | |
14028 | mem_3 = operands[5]; | |
14029 | mem_4 = operands[7]; | |
14030 | } | |
14031 | else | |
14032 | { | |
14033 | mem_1 = operands[0]; | |
14034 | mem_2 = operands[2]; | |
14035 | mem_3 = operands[4]; | |
14036 | mem_4 = operands[6]; | |
14037 | gcc_assert (code == UNKNOWN); | |
14038 | } | |
14039 | ||
14040 | extract_base_offset_in_addr (mem_1, &base, &offset); | |
14041 | gcc_assert (base != NULL_RTX && offset != NULL_RTX); | |
14042 | ||
14043 | /* Adjust offset thus it can fit in ldp/stp instruction. */ | |
14044 | msize = GET_MODE_SIZE (mode); | |
14045 | stp_off_limit = msize * 0x40; | |
14046 | off_val = INTVAL (offset); | |
14047 | abs_off = (off_val < 0) ? -off_val : off_val; | |
14048 | new_off = abs_off % stp_off_limit; | |
14049 | adj_off = abs_off - new_off; | |
14050 | ||
14051 | /* Further adjust to make sure all offsets are OK. */ | |
14052 | if ((new_off + msize * 2) >= stp_off_limit) | |
14053 | { | |
14054 | adj_off += stp_off_limit; | |
14055 | new_off -= stp_off_limit; | |
14056 | } | |
14057 | ||
14058 | /* Make sure the adjustment can be done with ADD/SUB instructions. */ | |
14059 | if (adj_off >= 0x1000) | |
14060 | return false; | |
14061 | ||
14062 | if (off_val < 0) | |
14063 | { | |
14064 | adj_off = -adj_off; | |
14065 | new_off = -new_off; | |
14066 | } | |
14067 | ||
14068 | /* Create new memory references. */ | |
14069 | mem_1 = change_address (mem_1, VOIDmode, | |
14070 | plus_constant (DImode, operands[8], new_off)); | |
14071 | ||
14072 | /* Check if the adjusted address is OK for ldp/stp. */ | |
14073 | if (!aarch64_mem_pair_operand (mem_1, mode)) | |
14074 | return false; | |
14075 | ||
14076 | msize = GET_MODE_SIZE (mode); | |
14077 | mem_2 = change_address (mem_2, VOIDmode, | |
14078 | plus_constant (DImode, | |
14079 | operands[8], | |
14080 | new_off + msize)); | |
14081 | mem_3 = change_address (mem_3, VOIDmode, | |
14082 | plus_constant (DImode, | |
14083 | operands[8], | |
14084 | new_off + msize * 2)); | |
14085 | mem_4 = change_address (mem_4, VOIDmode, | |
14086 | plus_constant (DImode, | |
14087 | operands[8], | |
14088 | new_off + msize * 3)); | |
14089 | ||
14090 | if (code == ZERO_EXTEND) | |
14091 | { | |
14092 | mem_1 = gen_rtx_ZERO_EXTEND (DImode, mem_1); | |
14093 | mem_2 = gen_rtx_ZERO_EXTEND (DImode, mem_2); | |
14094 | mem_3 = gen_rtx_ZERO_EXTEND (DImode, mem_3); | |
14095 | mem_4 = gen_rtx_ZERO_EXTEND (DImode, mem_4); | |
14096 | } | |
14097 | else if (code == SIGN_EXTEND) | |
14098 | { | |
14099 | mem_1 = gen_rtx_SIGN_EXTEND (DImode, mem_1); | |
14100 | mem_2 = gen_rtx_SIGN_EXTEND (DImode, mem_2); | |
14101 | mem_3 = gen_rtx_SIGN_EXTEND (DImode, mem_3); | |
14102 | mem_4 = gen_rtx_SIGN_EXTEND (DImode, mem_4); | |
14103 | } | |
14104 | ||
14105 | if (load) | |
14106 | { | |
14107 | operands[1] = mem_1; | |
14108 | operands[3] = mem_2; | |
14109 | operands[5] = mem_3; | |
14110 | operands[7] = mem_4; | |
14111 | } | |
14112 | else | |
14113 | { | |
14114 | operands[0] = mem_1; | |
14115 | operands[2] = mem_2; | |
14116 | operands[4] = mem_3; | |
14117 | operands[6] = mem_4; | |
14118 | } | |
14119 | ||
14120 | /* Emit adjusting instruction. */ | |
f7df4a84 | 14121 | emit_insn (gen_rtx_SET (operands[8], plus_constant (DImode, base, adj_off))); |
350013bc | 14122 | /* Emit ldp/stp instructions. */ |
f7df4a84 RS |
14123 | t1 = gen_rtx_SET (operands[0], operands[1]); |
14124 | t2 = gen_rtx_SET (operands[2], operands[3]); | |
350013bc | 14125 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, t1, t2))); |
f7df4a84 RS |
14126 | t1 = gen_rtx_SET (operands[4], operands[5]); |
14127 | t2 = gen_rtx_SET (operands[6], operands[7]); | |
350013bc BC |
14128 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, t1, t2))); |
14129 | return true; | |
14130 | } | |
14131 | ||
1b1e81f8 JW |
14132 | /* Return 1 if pseudo register should be created and used to hold |
14133 | GOT address for PIC code. */ | |
14134 | ||
14135 | bool | |
14136 | aarch64_use_pseudo_pic_reg (void) | |
14137 | { | |
14138 | return aarch64_cmodel == AARCH64_CMODEL_SMALL_SPIC; | |
14139 | } | |
14140 | ||
7b841a12 JW |
14141 | /* Implement TARGET_UNSPEC_MAY_TRAP_P. */ |
14142 | ||
14143 | static int | |
14144 | aarch64_unspec_may_trap_p (const_rtx x, unsigned flags) | |
14145 | { | |
14146 | switch (XINT (x, 1)) | |
14147 | { | |
14148 | case UNSPEC_GOTSMALLPIC: | |
14149 | case UNSPEC_GOTSMALLPIC28K: | |
14150 | case UNSPEC_GOTTINYPIC: | |
14151 | return 0; | |
14152 | default: | |
14153 | break; | |
14154 | } | |
14155 | ||
14156 | return default_unspec_may_trap_p (x, flags); | |
14157 | } | |
14158 | ||
39252973 KT |
14159 | |
14160 | /* If X is a positive CONST_DOUBLE with a value that is a power of 2 | |
14161 | return the log2 of that value. Otherwise return -1. */ | |
14162 | ||
14163 | int | |
14164 | aarch64_fpconst_pow_of_2 (rtx x) | |
14165 | { | |
14166 | const REAL_VALUE_TYPE *r; | |
14167 | ||
14168 | if (!CONST_DOUBLE_P (x)) | |
14169 | return -1; | |
14170 | ||
14171 | r = CONST_DOUBLE_REAL_VALUE (x); | |
14172 | ||
14173 | if (REAL_VALUE_NEGATIVE (*r) | |
14174 | || REAL_VALUE_ISNAN (*r) | |
14175 | || REAL_VALUE_ISINF (*r) | |
14176 | || !real_isinteger (r, DFmode)) | |
14177 | return -1; | |
14178 | ||
14179 | return exact_log2 (real_to_integer (r)); | |
14180 | } | |
14181 | ||
14182 | /* If X is a vector of equal CONST_DOUBLE values and that value is | |
14183 | Y, return the aarch64_fpconst_pow_of_2 of Y. Otherwise return -1. */ | |
14184 | ||
14185 | int | |
14186 | aarch64_vec_fpconst_pow_of_2 (rtx x) | |
14187 | { | |
14188 | if (GET_CODE (x) != CONST_VECTOR) | |
14189 | return -1; | |
14190 | ||
14191 | if (GET_MODE_CLASS (GET_MODE (x)) != MODE_VECTOR_FLOAT) | |
14192 | return -1; | |
14193 | ||
14194 | int firstval = aarch64_fpconst_pow_of_2 (CONST_VECTOR_ELT (x, 0)); | |
14195 | if (firstval <= 0) | |
14196 | return -1; | |
14197 | ||
14198 | for (int i = 1; i < CONST_VECTOR_NUNITS (x); i++) | |
14199 | if (aarch64_fpconst_pow_of_2 (CONST_VECTOR_ELT (x, i)) != firstval) | |
14200 | return -1; | |
14201 | ||
14202 | return firstval; | |
14203 | } | |
14204 | ||
c2ec330c AL |
14205 | /* Implement TARGET_PROMOTED_TYPE to promote __fp16 to float. */ |
14206 | static tree | |
14207 | aarch64_promoted_type (const_tree t) | |
14208 | { | |
14209 | if (SCALAR_FLOAT_TYPE_P (t) && TYPE_PRECISION (t) == 16) | |
14210 | return float_type_node; | |
14211 | return NULL_TREE; | |
14212 | } | |
ee62a5a6 RS |
14213 | |
14214 | /* Implement the TARGET_OPTAB_SUPPORTED_P hook. */ | |
14215 | ||
14216 | static bool | |
9acc9cbe | 14217 | aarch64_optab_supported_p (int op, machine_mode mode1, machine_mode, |
ee62a5a6 RS |
14218 | optimization_type opt_type) |
14219 | { | |
14220 | switch (op) | |
14221 | { | |
14222 | case rsqrt_optab: | |
9acc9cbe | 14223 | return opt_type == OPTIMIZE_FOR_SPEED && use_rsqrt_p (mode1); |
ee62a5a6 RS |
14224 | |
14225 | default: | |
14226 | return true; | |
14227 | } | |
14228 | } | |
14229 | ||
43e9d192 IB |
14230 | #undef TARGET_ADDRESS_COST |
14231 | #define TARGET_ADDRESS_COST aarch64_address_cost | |
14232 | ||
14233 | /* This hook will determines whether unnamed bitfields affect the alignment | |
14234 | of the containing structure. The hook returns true if the structure | |
14235 | should inherit the alignment requirements of an unnamed bitfield's | |
14236 | type. */ | |
14237 | #undef TARGET_ALIGN_ANON_BITFIELD | |
14238 | #define TARGET_ALIGN_ANON_BITFIELD hook_bool_void_true | |
14239 | ||
14240 | #undef TARGET_ASM_ALIGNED_DI_OP | |
14241 | #define TARGET_ASM_ALIGNED_DI_OP "\t.xword\t" | |
14242 | ||
14243 | #undef TARGET_ASM_ALIGNED_HI_OP | |
14244 | #define TARGET_ASM_ALIGNED_HI_OP "\t.hword\t" | |
14245 | ||
14246 | #undef TARGET_ASM_ALIGNED_SI_OP | |
14247 | #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t" | |
14248 | ||
14249 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
14250 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK \ | |
14251 | hook_bool_const_tree_hwi_hwi_const_tree_true | |
14252 | ||
e1c1ecb0 KT |
14253 | #undef TARGET_ASM_FILE_START |
14254 | #define TARGET_ASM_FILE_START aarch64_start_file | |
14255 | ||
43e9d192 IB |
14256 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
14257 | #define TARGET_ASM_OUTPUT_MI_THUNK aarch64_output_mi_thunk | |
14258 | ||
14259 | #undef TARGET_ASM_SELECT_RTX_SECTION | |
14260 | #define TARGET_ASM_SELECT_RTX_SECTION aarch64_select_rtx_section | |
14261 | ||
14262 | #undef TARGET_ASM_TRAMPOLINE_TEMPLATE | |
14263 | #define TARGET_ASM_TRAMPOLINE_TEMPLATE aarch64_asm_trampoline_template | |
14264 | ||
14265 | #undef TARGET_BUILD_BUILTIN_VA_LIST | |
14266 | #define TARGET_BUILD_BUILTIN_VA_LIST aarch64_build_builtin_va_list | |
14267 | ||
14268 | #undef TARGET_CALLEE_COPIES | |
14269 | #define TARGET_CALLEE_COPIES hook_bool_CUMULATIVE_ARGS_mode_tree_bool_false | |
14270 | ||
14271 | #undef TARGET_CAN_ELIMINATE | |
14272 | #define TARGET_CAN_ELIMINATE aarch64_can_eliminate | |
14273 | ||
1fd8d40c KT |
14274 | #undef TARGET_CAN_INLINE_P |
14275 | #define TARGET_CAN_INLINE_P aarch64_can_inline_p | |
14276 | ||
43e9d192 IB |
14277 | #undef TARGET_CANNOT_FORCE_CONST_MEM |
14278 | #define TARGET_CANNOT_FORCE_CONST_MEM aarch64_cannot_force_const_mem | |
14279 | ||
50487d79 EM |
14280 | #undef TARGET_CASE_VALUES_THRESHOLD |
14281 | #define TARGET_CASE_VALUES_THRESHOLD aarch64_case_values_threshold | |
14282 | ||
43e9d192 IB |
14283 | #undef TARGET_CONDITIONAL_REGISTER_USAGE |
14284 | #define TARGET_CONDITIONAL_REGISTER_USAGE aarch64_conditional_register_usage | |
14285 | ||
14286 | /* Only the least significant bit is used for initialization guard | |
14287 | variables. */ | |
14288 | #undef TARGET_CXX_GUARD_MASK_BIT | |
14289 | #define TARGET_CXX_GUARD_MASK_BIT hook_bool_void_true | |
14290 | ||
14291 | #undef TARGET_C_MODE_FOR_SUFFIX | |
14292 | #define TARGET_C_MODE_FOR_SUFFIX aarch64_c_mode_for_suffix | |
14293 | ||
14294 | #ifdef TARGET_BIG_ENDIAN_DEFAULT | |
14295 | #undef TARGET_DEFAULT_TARGET_FLAGS | |
14296 | #define TARGET_DEFAULT_TARGET_FLAGS (MASK_BIG_END) | |
14297 | #endif | |
14298 | ||
14299 | #undef TARGET_CLASS_MAX_NREGS | |
14300 | #define TARGET_CLASS_MAX_NREGS aarch64_class_max_nregs | |
14301 | ||
119103ca JG |
14302 | #undef TARGET_BUILTIN_DECL |
14303 | #define TARGET_BUILTIN_DECL aarch64_builtin_decl | |
14304 | ||
a6fc00da BH |
14305 | #undef TARGET_BUILTIN_RECIPROCAL |
14306 | #define TARGET_BUILTIN_RECIPROCAL aarch64_builtin_reciprocal | |
14307 | ||
43e9d192 IB |
14308 | #undef TARGET_EXPAND_BUILTIN |
14309 | #define TARGET_EXPAND_BUILTIN aarch64_expand_builtin | |
14310 | ||
14311 | #undef TARGET_EXPAND_BUILTIN_VA_START | |
14312 | #define TARGET_EXPAND_BUILTIN_VA_START aarch64_expand_builtin_va_start | |
14313 | ||
9697e620 JG |
14314 | #undef TARGET_FOLD_BUILTIN |
14315 | #define TARGET_FOLD_BUILTIN aarch64_fold_builtin | |
14316 | ||
43e9d192 IB |
14317 | #undef TARGET_FUNCTION_ARG |
14318 | #define TARGET_FUNCTION_ARG aarch64_function_arg | |
14319 | ||
14320 | #undef TARGET_FUNCTION_ARG_ADVANCE | |
14321 | #define TARGET_FUNCTION_ARG_ADVANCE aarch64_function_arg_advance | |
14322 | ||
14323 | #undef TARGET_FUNCTION_ARG_BOUNDARY | |
14324 | #define TARGET_FUNCTION_ARG_BOUNDARY aarch64_function_arg_boundary | |
14325 | ||
14326 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL | |
14327 | #define TARGET_FUNCTION_OK_FOR_SIBCALL aarch64_function_ok_for_sibcall | |
14328 | ||
14329 | #undef TARGET_FUNCTION_VALUE | |
14330 | #define TARGET_FUNCTION_VALUE aarch64_function_value | |
14331 | ||
14332 | #undef TARGET_FUNCTION_VALUE_REGNO_P | |
14333 | #define TARGET_FUNCTION_VALUE_REGNO_P aarch64_function_value_regno_p | |
14334 | ||
14335 | #undef TARGET_FRAME_POINTER_REQUIRED | |
14336 | #define TARGET_FRAME_POINTER_REQUIRED aarch64_frame_pointer_required | |
14337 | ||
fc72cba7 AL |
14338 | #undef TARGET_GIMPLE_FOLD_BUILTIN |
14339 | #define TARGET_GIMPLE_FOLD_BUILTIN aarch64_gimple_fold_builtin | |
0ac198d3 | 14340 | |
43e9d192 IB |
14341 | #undef TARGET_GIMPLIFY_VA_ARG_EXPR |
14342 | #define TARGET_GIMPLIFY_VA_ARG_EXPR aarch64_gimplify_va_arg_expr | |
14343 | ||
14344 | #undef TARGET_INIT_BUILTINS | |
14345 | #define TARGET_INIT_BUILTINS aarch64_init_builtins | |
14346 | ||
c64f7d37 WD |
14347 | #undef TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS |
14348 | #define TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS \ | |
14349 | aarch64_ira_change_pseudo_allocno_class | |
14350 | ||
43e9d192 IB |
14351 | #undef TARGET_LEGITIMATE_ADDRESS_P |
14352 | #define TARGET_LEGITIMATE_ADDRESS_P aarch64_legitimate_address_hook_p | |
14353 | ||
14354 | #undef TARGET_LEGITIMATE_CONSTANT_P | |
14355 | #define TARGET_LEGITIMATE_CONSTANT_P aarch64_legitimate_constant_p | |
14356 | ||
491ec060 WD |
14357 | #undef TARGET_LEGITIMIZE_ADDRESS_DISPLACEMENT |
14358 | #define TARGET_LEGITIMIZE_ADDRESS_DISPLACEMENT \ | |
14359 | aarch64_legitimize_address_displacement | |
14360 | ||
43e9d192 IB |
14361 | #undef TARGET_LIBGCC_CMP_RETURN_MODE |
14362 | #define TARGET_LIBGCC_CMP_RETURN_MODE aarch64_libgcc_cmp_return_mode | |
14363 | ||
ac2b960f YZ |
14364 | #undef TARGET_MANGLE_TYPE |
14365 | #define TARGET_MANGLE_TYPE aarch64_mangle_type | |
14366 | ||
43e9d192 IB |
14367 | #undef TARGET_MEMORY_MOVE_COST |
14368 | #define TARGET_MEMORY_MOVE_COST aarch64_memory_move_cost | |
14369 | ||
26e0ff94 WD |
14370 | #undef TARGET_MIN_DIVISIONS_FOR_RECIP_MUL |
14371 | #define TARGET_MIN_DIVISIONS_FOR_RECIP_MUL aarch64_min_divisions_for_recip_mul | |
14372 | ||
43e9d192 IB |
14373 | #undef TARGET_MUST_PASS_IN_STACK |
14374 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
14375 | ||
14376 | /* This target hook should return true if accesses to volatile bitfields | |
14377 | should use the narrowest mode possible. It should return false if these | |
14378 | accesses should use the bitfield container type. */ | |
14379 | #undef TARGET_NARROW_VOLATILE_BITFIELD | |
14380 | #define TARGET_NARROW_VOLATILE_BITFIELD hook_bool_void_false | |
14381 | ||
14382 | #undef TARGET_OPTION_OVERRIDE | |
14383 | #define TARGET_OPTION_OVERRIDE aarch64_override_options | |
14384 | ||
14385 | #undef TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE | |
14386 | #define TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE \ | |
14387 | aarch64_override_options_after_change | |
14388 | ||
361fb3ee KT |
14389 | #undef TARGET_OPTION_SAVE |
14390 | #define TARGET_OPTION_SAVE aarch64_option_save | |
14391 | ||
14392 | #undef TARGET_OPTION_RESTORE | |
14393 | #define TARGET_OPTION_RESTORE aarch64_option_restore | |
14394 | ||
14395 | #undef TARGET_OPTION_PRINT | |
14396 | #define TARGET_OPTION_PRINT aarch64_option_print | |
14397 | ||
5a2c8331 KT |
14398 | #undef TARGET_OPTION_VALID_ATTRIBUTE_P |
14399 | #define TARGET_OPTION_VALID_ATTRIBUTE_P aarch64_option_valid_attribute_p | |
14400 | ||
d78006d9 KT |
14401 | #undef TARGET_SET_CURRENT_FUNCTION |
14402 | #define TARGET_SET_CURRENT_FUNCTION aarch64_set_current_function | |
14403 | ||
43e9d192 IB |
14404 | #undef TARGET_PASS_BY_REFERENCE |
14405 | #define TARGET_PASS_BY_REFERENCE aarch64_pass_by_reference | |
14406 | ||
14407 | #undef TARGET_PREFERRED_RELOAD_CLASS | |
14408 | #define TARGET_PREFERRED_RELOAD_CLASS aarch64_preferred_reload_class | |
14409 | ||
cee66c68 WD |
14410 | #undef TARGET_SCHED_REASSOCIATION_WIDTH |
14411 | #define TARGET_SCHED_REASSOCIATION_WIDTH aarch64_reassociation_width | |
14412 | ||
c2ec330c AL |
14413 | #undef TARGET_PROMOTED_TYPE |
14414 | #define TARGET_PROMOTED_TYPE aarch64_promoted_type | |
14415 | ||
43e9d192 IB |
14416 | #undef TARGET_SECONDARY_RELOAD |
14417 | #define TARGET_SECONDARY_RELOAD aarch64_secondary_reload | |
14418 | ||
14419 | #undef TARGET_SHIFT_TRUNCATION_MASK | |
14420 | #define TARGET_SHIFT_TRUNCATION_MASK aarch64_shift_truncation_mask | |
14421 | ||
14422 | #undef TARGET_SETUP_INCOMING_VARARGS | |
14423 | #define TARGET_SETUP_INCOMING_VARARGS aarch64_setup_incoming_varargs | |
14424 | ||
14425 | #undef TARGET_STRUCT_VALUE_RTX | |
14426 | #define TARGET_STRUCT_VALUE_RTX aarch64_struct_value_rtx | |
14427 | ||
14428 | #undef TARGET_REGISTER_MOVE_COST | |
14429 | #define TARGET_REGISTER_MOVE_COST aarch64_register_move_cost | |
14430 | ||
14431 | #undef TARGET_RETURN_IN_MEMORY | |
14432 | #define TARGET_RETURN_IN_MEMORY aarch64_return_in_memory | |
14433 | ||
14434 | #undef TARGET_RETURN_IN_MSB | |
14435 | #define TARGET_RETURN_IN_MSB aarch64_return_in_msb | |
14436 | ||
14437 | #undef TARGET_RTX_COSTS | |
7cc2145f | 14438 | #define TARGET_RTX_COSTS aarch64_rtx_costs_wrapper |
43e9d192 | 14439 | |
d126a4ae AP |
14440 | #undef TARGET_SCHED_ISSUE_RATE |
14441 | #define TARGET_SCHED_ISSUE_RATE aarch64_sched_issue_rate | |
14442 | ||
d03f7e44 MK |
14443 | #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD |
14444 | #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD \ | |
14445 | aarch64_sched_first_cycle_multipass_dfa_lookahead | |
14446 | ||
2d6bc7fa KT |
14447 | #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD |
14448 | #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD \ | |
14449 | aarch64_first_cycle_multipass_dfa_lookahead_guard | |
14450 | ||
43e9d192 IB |
14451 | #undef TARGET_TRAMPOLINE_INIT |
14452 | #define TARGET_TRAMPOLINE_INIT aarch64_trampoline_init | |
14453 | ||
14454 | #undef TARGET_USE_BLOCKS_FOR_CONSTANT_P | |
14455 | #define TARGET_USE_BLOCKS_FOR_CONSTANT_P aarch64_use_blocks_for_constant_p | |
14456 | ||
14457 | #undef TARGET_VECTOR_MODE_SUPPORTED_P | |
14458 | #define TARGET_VECTOR_MODE_SUPPORTED_P aarch64_vector_mode_supported_p | |
14459 | ||
14460 | #undef TARGET_ARRAY_MODE_SUPPORTED_P | |
14461 | #define TARGET_ARRAY_MODE_SUPPORTED_P aarch64_array_mode_supported_p | |
14462 | ||
8990e73a TB |
14463 | #undef TARGET_VECTORIZE_ADD_STMT_COST |
14464 | #define TARGET_VECTORIZE_ADD_STMT_COST aarch64_add_stmt_cost | |
14465 | ||
14466 | #undef TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST | |
14467 | #define TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST \ | |
14468 | aarch64_builtin_vectorization_cost | |
14469 | ||
43e9d192 IB |
14470 | #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE |
14471 | #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE aarch64_preferred_simd_mode | |
14472 | ||
42fc9a7f JG |
14473 | #undef TARGET_VECTORIZE_BUILTINS |
14474 | #define TARGET_VECTORIZE_BUILTINS | |
14475 | ||
14476 | #undef TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION | |
14477 | #define TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION \ | |
14478 | aarch64_builtin_vectorized_function | |
14479 | ||
3b357264 JG |
14480 | #undef TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES |
14481 | #define TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES \ | |
14482 | aarch64_autovectorize_vector_sizes | |
14483 | ||
aa87aced KV |
14484 | #undef TARGET_ATOMIC_ASSIGN_EXPAND_FENV |
14485 | #define TARGET_ATOMIC_ASSIGN_EXPAND_FENV \ | |
14486 | aarch64_atomic_assign_expand_fenv | |
14487 | ||
43e9d192 IB |
14488 | /* Section anchor support. */ |
14489 | ||
14490 | #undef TARGET_MIN_ANCHOR_OFFSET | |
14491 | #define TARGET_MIN_ANCHOR_OFFSET -256 | |
14492 | ||
14493 | /* Limit the maximum anchor offset to 4k-1, since that's the limit for a | |
14494 | byte offset; we can do much more for larger data types, but have no way | |
14495 | to determine the size of the access. We assume accesses are aligned. */ | |
14496 | #undef TARGET_MAX_ANCHOR_OFFSET | |
14497 | #define TARGET_MAX_ANCHOR_OFFSET 4095 | |
14498 | ||
db0253a4 TB |
14499 | #undef TARGET_VECTOR_ALIGNMENT |
14500 | #define TARGET_VECTOR_ALIGNMENT aarch64_simd_vector_alignment | |
14501 | ||
14502 | #undef TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE | |
14503 | #define TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE \ | |
14504 | aarch64_simd_vector_alignment_reachable | |
14505 | ||
88b08073 JG |
14506 | /* vec_perm support. */ |
14507 | ||
14508 | #undef TARGET_VECTORIZE_VEC_PERM_CONST_OK | |
14509 | #define TARGET_VECTORIZE_VEC_PERM_CONST_OK \ | |
14510 | aarch64_vectorize_vec_perm_const_ok | |
14511 | ||
c2ec330c AL |
14512 | #undef TARGET_INIT_LIBFUNCS |
14513 | #define TARGET_INIT_LIBFUNCS aarch64_init_libfuncs | |
70f09188 | 14514 | |
706b2314 | 14515 | #undef TARGET_FIXED_CONDITION_CODE_REGS |
70f09188 AP |
14516 | #define TARGET_FIXED_CONDITION_CODE_REGS aarch64_fixed_condition_code_regs |
14517 | ||
5cb74e90 RR |
14518 | #undef TARGET_FLAGS_REGNUM |
14519 | #define TARGET_FLAGS_REGNUM CC_REGNUM | |
14520 | ||
78607708 TV |
14521 | #undef TARGET_CALL_FUSAGE_CONTAINS_NON_CALLEE_CLOBBERS |
14522 | #define TARGET_CALL_FUSAGE_CONTAINS_NON_CALLEE_CLOBBERS true | |
14523 | ||
a3125fc2 CL |
14524 | #undef TARGET_ASAN_SHADOW_OFFSET |
14525 | #define TARGET_ASAN_SHADOW_OFFSET aarch64_asan_shadow_offset | |
14526 | ||
0c4ec427 RE |
14527 | #undef TARGET_LEGITIMIZE_ADDRESS |
14528 | #define TARGET_LEGITIMIZE_ADDRESS aarch64_legitimize_address | |
14529 | ||
d3006da6 JG |
14530 | #undef TARGET_USE_BY_PIECES_INFRASTRUCTURE_P |
14531 | #define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \ | |
14532 | aarch64_use_by_pieces_infrastructure_p | |
14533 | ||
594bdd53 FY |
14534 | #undef TARGET_CAN_USE_DOLOOP_P |
14535 | #define TARGET_CAN_USE_DOLOOP_P can_use_doloop_if_innermost | |
14536 | ||
6a569cdd KT |
14537 | #undef TARGET_SCHED_MACRO_FUSION_P |
14538 | #define TARGET_SCHED_MACRO_FUSION_P aarch64_macro_fusion_p | |
14539 | ||
14540 | #undef TARGET_SCHED_MACRO_FUSION_PAIR_P | |
14541 | #define TARGET_SCHED_MACRO_FUSION_PAIR_P aarch_macro_fusion_pair_p | |
14542 | ||
350013bc BC |
14543 | #undef TARGET_SCHED_FUSION_PRIORITY |
14544 | #define TARGET_SCHED_FUSION_PRIORITY aarch64_sched_fusion_priority | |
14545 | ||
7b841a12 JW |
14546 | #undef TARGET_UNSPEC_MAY_TRAP_P |
14547 | #define TARGET_UNSPEC_MAY_TRAP_P aarch64_unspec_may_trap_p | |
14548 | ||
1b1e81f8 JW |
14549 | #undef TARGET_USE_PSEUDO_PIC_REG |
14550 | #define TARGET_USE_PSEUDO_PIC_REG aarch64_use_pseudo_pic_reg | |
14551 | ||
cc8ca59e JB |
14552 | #undef TARGET_PRINT_OPERAND |
14553 | #define TARGET_PRINT_OPERAND aarch64_print_operand | |
14554 | ||
14555 | #undef TARGET_PRINT_OPERAND_ADDRESS | |
14556 | #define TARGET_PRINT_OPERAND_ADDRESS aarch64_print_operand_address | |
14557 | ||
ee62a5a6 RS |
14558 | #undef TARGET_OPTAB_SUPPORTED_P |
14559 | #define TARGET_OPTAB_SUPPORTED_P aarch64_optab_supported_p | |
14560 | ||
43203dea RR |
14561 | #undef TARGET_OMIT_STRUCT_RETURN_REG |
14562 | #define TARGET_OMIT_STRUCT_RETURN_REG true | |
14563 | ||
43e9d192 IB |
14564 | struct gcc_target targetm = TARGET_INITIALIZER; |
14565 | ||
14566 | #include "gt-aarch64.h" |