]>
Commit | Line | Data |
---|---|---|
bdb7bf8a | 1 | /* Machine description for AArch64 architecture. |
cbe34bb5 | 2 | Copyright (C) 2009-2017 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" | |
314e6352 | 36 | #include "attribs.h" |
e11c4407 AM |
37 | #include "optabs.h" |
38 | #include "regs.h" | |
39 | #include "emit-rtl.h" | |
40 | #include "recog.h" | |
41 | #include "diagnostic.h" | |
43e9d192 | 42 | #include "insn-attr.h" |
40e23961 | 43 | #include "alias.h" |
40e23961 | 44 | #include "fold-const.h" |
d8a2d370 DN |
45 | #include "stor-layout.h" |
46 | #include "calls.h" | |
47 | #include "varasm.h" | |
43e9d192 | 48 | #include "output.h" |
36566b39 | 49 | #include "flags.h" |
36566b39 | 50 | #include "explow.h" |
43e9d192 IB |
51 | #include "expr.h" |
52 | #include "reload.h" | |
43e9d192 | 53 | #include "langhooks.h" |
5a2c8331 | 54 | #include "opts.h" |
2d6bc7fa | 55 | #include "params.h" |
45b0be94 | 56 | #include "gimplify.h" |
43e9d192 | 57 | #include "dwarf2.h" |
61d371eb | 58 | #include "gimple-iterator.h" |
8990e73a | 59 | #include "tree-vectorizer.h" |
d1bcc29f | 60 | #include "aarch64-cost-tables.h" |
0ee859b5 | 61 | #include "dumpfile.h" |
9b2b7279 | 62 | #include "builtins.h" |
8baff86e | 63 | #include "rtl-iter.h" |
9bbe08fe | 64 | #include "tm-constrs.h" |
d03f7e44 | 65 | #include "sched-int.h" |
d78006d9 | 66 | #include "target-globals.h" |
a3eb8a52 | 67 | #include "common/common-target.h" |
51b86113 DM |
68 | #include "selftest.h" |
69 | #include "selftest-rtl.h" | |
43e9d192 | 70 | |
994c5d85 | 71 | /* This file should be included last. */ |
d58627a0 RS |
72 | #include "target-def.h" |
73 | ||
28514dda YZ |
74 | /* Defined for convenience. */ |
75 | #define POINTER_BYTES (POINTER_SIZE / BITS_PER_UNIT) | |
76 | ||
43e9d192 IB |
77 | /* Classifies an address. |
78 | ||
79 | ADDRESS_REG_IMM | |
80 | A simple base register plus immediate offset. | |
81 | ||
82 | ADDRESS_REG_WB | |
83 | A base register indexed by immediate offset with writeback. | |
84 | ||
85 | ADDRESS_REG_REG | |
86 | A base register indexed by (optionally scaled) register. | |
87 | ||
88 | ADDRESS_REG_UXTW | |
89 | A base register indexed by (optionally scaled) zero-extended register. | |
90 | ||
91 | ADDRESS_REG_SXTW | |
92 | A base register indexed by (optionally scaled) sign-extended register. | |
93 | ||
94 | ADDRESS_LO_SUM | |
95 | A LO_SUM rtx with a base register and "LO12" symbol relocation. | |
96 | ||
97 | ADDRESS_SYMBOLIC: | |
98 | A constant symbolic address, in pc-relative literal pool. */ | |
99 | ||
100 | enum aarch64_address_type { | |
101 | ADDRESS_REG_IMM, | |
102 | ADDRESS_REG_WB, | |
103 | ADDRESS_REG_REG, | |
104 | ADDRESS_REG_UXTW, | |
105 | ADDRESS_REG_SXTW, | |
106 | ADDRESS_LO_SUM, | |
107 | ADDRESS_SYMBOLIC | |
108 | }; | |
109 | ||
110 | struct aarch64_address_info { | |
111 | enum aarch64_address_type type; | |
112 | rtx base; | |
113 | rtx offset; | |
114 | int shift; | |
115 | enum aarch64_symbol_type symbol_type; | |
116 | }; | |
117 | ||
48063b9d IB |
118 | struct simd_immediate_info |
119 | { | |
120 | rtx value; | |
121 | int shift; | |
122 | int element_width; | |
48063b9d | 123 | bool mvn; |
e4f0f84d | 124 | bool msl; |
48063b9d IB |
125 | }; |
126 | ||
43e9d192 IB |
127 | /* The current code model. */ |
128 | enum aarch64_code_model aarch64_cmodel; | |
129 | ||
130 | #ifdef HAVE_AS_TLS | |
131 | #undef TARGET_HAVE_TLS | |
132 | #define TARGET_HAVE_TLS 1 | |
133 | #endif | |
134 | ||
ef4bddc2 RS |
135 | static bool aarch64_composite_type_p (const_tree, machine_mode); |
136 | static bool aarch64_vfp_is_call_or_return_candidate (machine_mode, | |
43e9d192 | 137 | const_tree, |
ef4bddc2 | 138 | machine_mode *, int *, |
43e9d192 IB |
139 | bool *); |
140 | static void aarch64_elf_asm_constructor (rtx, int) ATTRIBUTE_UNUSED; | |
141 | static void aarch64_elf_asm_destructor (rtx, int) ATTRIBUTE_UNUSED; | |
43e9d192 | 142 | static void aarch64_override_options_after_change (void); |
ef4bddc2 | 143 | static bool aarch64_vector_mode_supported_p (machine_mode); |
ef4bddc2 | 144 | static bool aarch64_vectorize_vec_perm_const_ok (machine_mode vmode, |
88b08073 | 145 | const unsigned char *sel); |
ef4bddc2 | 146 | static int aarch64_address_cost (rtx, machine_mode, addr_space_t, bool); |
7df76747 N |
147 | static bool aarch64_builtin_support_vector_misalignment (machine_mode mode, |
148 | const_tree type, | |
149 | int misalignment, | |
150 | bool is_packed); | |
a2170965 TC |
151 | static machine_mode |
152 | aarch64_simd_container_mode (machine_mode mode, unsigned width); | |
88b08073 | 153 | |
0c6caaf8 RL |
154 | /* Major revision number of the ARM Architecture implemented by the target. */ |
155 | unsigned aarch64_architecture_version; | |
156 | ||
43e9d192 | 157 | /* The processor for which instructions should be scheduled. */ |
02fdbd5b | 158 | enum aarch64_processor aarch64_tune = cortexa53; |
43e9d192 | 159 | |
43e9d192 IB |
160 | /* Mask to specify which instruction scheduling options should be used. */ |
161 | unsigned long aarch64_tune_flags = 0; | |
162 | ||
1be34295 | 163 | /* Global flag for PC relative loads. */ |
9ee6540a | 164 | bool aarch64_pcrelative_literal_loads; |
1be34295 | 165 | |
8dec06f2 JG |
166 | /* Support for command line parsing of boolean flags in the tuning |
167 | structures. */ | |
168 | struct aarch64_flag_desc | |
169 | { | |
170 | const char* name; | |
171 | unsigned int flag; | |
172 | }; | |
173 | ||
ed9fa8d2 | 174 | #define AARCH64_FUSION_PAIR(name, internal_name) \ |
8dec06f2 JG |
175 | { name, AARCH64_FUSE_##internal_name }, |
176 | static const struct aarch64_flag_desc aarch64_fusible_pairs[] = | |
177 | { | |
178 | { "none", AARCH64_FUSE_NOTHING }, | |
179 | #include "aarch64-fusion-pairs.def" | |
180 | { "all", AARCH64_FUSE_ALL }, | |
181 | { NULL, AARCH64_FUSE_NOTHING } | |
182 | }; | |
8dec06f2 | 183 | |
a339a01c | 184 | #define AARCH64_EXTRA_TUNING_OPTION(name, internal_name) \ |
8dec06f2 JG |
185 | { name, AARCH64_EXTRA_TUNE_##internal_name }, |
186 | static const struct aarch64_flag_desc aarch64_tuning_flags[] = | |
187 | { | |
188 | { "none", AARCH64_EXTRA_TUNE_NONE }, | |
189 | #include "aarch64-tuning-flags.def" | |
190 | { "all", AARCH64_EXTRA_TUNE_ALL }, | |
191 | { NULL, AARCH64_EXTRA_TUNE_NONE } | |
192 | }; | |
8dec06f2 | 193 | |
43e9d192 IB |
194 | /* Tuning parameters. */ |
195 | ||
43e9d192 IB |
196 | static const struct cpu_addrcost_table generic_addrcost_table = |
197 | { | |
67747367 | 198 | { |
2fae724a | 199 | 1, /* hi */ |
bd95e655 JG |
200 | 0, /* si */ |
201 | 0, /* di */ | |
2fae724a | 202 | 1, /* ti */ |
67747367 | 203 | }, |
bd95e655 JG |
204 | 0, /* pre_modify */ |
205 | 0, /* post_modify */ | |
206 | 0, /* register_offset */ | |
783879e6 EM |
207 | 0, /* register_sextend */ |
208 | 0, /* register_zextend */ | |
bd95e655 | 209 | 0 /* imm_offset */ |
43e9d192 IB |
210 | }; |
211 | ||
5ec1ae3b EM |
212 | static const struct cpu_addrcost_table exynosm1_addrcost_table = |
213 | { | |
214 | { | |
215 | 0, /* hi */ | |
216 | 0, /* si */ | |
217 | 0, /* di */ | |
218 | 2, /* ti */ | |
219 | }, | |
220 | 0, /* pre_modify */ | |
221 | 0, /* post_modify */ | |
222 | 1, /* register_offset */ | |
223 | 1, /* register_sextend */ | |
224 | 2, /* register_zextend */ | |
225 | 0, /* imm_offset */ | |
226 | }; | |
227 | ||
381e27aa PT |
228 | static const struct cpu_addrcost_table xgene1_addrcost_table = |
229 | { | |
381e27aa | 230 | { |
bd95e655 JG |
231 | 1, /* hi */ |
232 | 0, /* si */ | |
233 | 0, /* di */ | |
234 | 1, /* ti */ | |
381e27aa | 235 | }, |
bd95e655 JG |
236 | 1, /* pre_modify */ |
237 | 0, /* post_modify */ | |
238 | 0, /* register_offset */ | |
783879e6 EM |
239 | 1, /* register_sextend */ |
240 | 1, /* register_zextend */ | |
bd95e655 | 241 | 0, /* imm_offset */ |
381e27aa PT |
242 | }; |
243 | ||
d1261ac6 | 244 | static const struct cpu_addrcost_table thunderx2t99_addrcost_table = |
ad611a4c VP |
245 | { |
246 | { | |
5f407e57 AP |
247 | 1, /* hi */ |
248 | 1, /* si */ | |
249 | 1, /* di */ | |
ad611a4c VP |
250 | 2, /* ti */ |
251 | }, | |
252 | 0, /* pre_modify */ | |
253 | 0, /* post_modify */ | |
254 | 2, /* register_offset */ | |
255 | 3, /* register_sextend */ | |
256 | 3, /* register_zextend */ | |
257 | 0, /* imm_offset */ | |
258 | }; | |
259 | ||
43e9d192 IB |
260 | static const struct cpu_regmove_cost generic_regmove_cost = |
261 | { | |
bd95e655 | 262 | 1, /* GP2GP */ |
3969c510 WD |
263 | /* Avoid the use of slow int<->fp moves for spilling by setting |
264 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
265 | 5, /* GP2FP */ |
266 | 5, /* FP2GP */ | |
267 | 2 /* FP2FP */ | |
43e9d192 IB |
268 | }; |
269 | ||
e4a9c55a WD |
270 | static const struct cpu_regmove_cost cortexa57_regmove_cost = |
271 | { | |
bd95e655 | 272 | 1, /* GP2GP */ |
e4a9c55a WD |
273 | /* Avoid the use of slow int<->fp moves for spilling by setting |
274 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
275 | 5, /* GP2FP */ |
276 | 5, /* FP2GP */ | |
277 | 2 /* FP2FP */ | |
e4a9c55a WD |
278 | }; |
279 | ||
280 | static const struct cpu_regmove_cost cortexa53_regmove_cost = | |
281 | { | |
bd95e655 | 282 | 1, /* GP2GP */ |
e4a9c55a WD |
283 | /* Avoid the use of slow int<->fp moves for spilling by setting |
284 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
285 | 5, /* GP2FP */ |
286 | 5, /* FP2GP */ | |
287 | 2 /* FP2FP */ | |
e4a9c55a WD |
288 | }; |
289 | ||
5ec1ae3b EM |
290 | static const struct cpu_regmove_cost exynosm1_regmove_cost = |
291 | { | |
292 | 1, /* GP2GP */ | |
293 | /* Avoid the use of slow int<->fp moves for spilling by setting | |
294 | their cost higher than memmov_cost (actual, 4 and 9). */ | |
295 | 9, /* GP2FP */ | |
296 | 9, /* FP2GP */ | |
297 | 1 /* FP2FP */ | |
298 | }; | |
299 | ||
d1bcc29f AP |
300 | static const struct cpu_regmove_cost thunderx_regmove_cost = |
301 | { | |
bd95e655 JG |
302 | 2, /* GP2GP */ |
303 | 2, /* GP2FP */ | |
304 | 6, /* FP2GP */ | |
305 | 4 /* FP2FP */ | |
d1bcc29f AP |
306 | }; |
307 | ||
381e27aa PT |
308 | static const struct cpu_regmove_cost xgene1_regmove_cost = |
309 | { | |
bd95e655 | 310 | 1, /* GP2GP */ |
381e27aa PT |
311 | /* Avoid the use of slow int<->fp moves for spilling by setting |
312 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
313 | 8, /* GP2FP */ |
314 | 8, /* FP2GP */ | |
315 | 2 /* FP2FP */ | |
381e27aa PT |
316 | }; |
317 | ||
ee446d9f JW |
318 | static const struct cpu_regmove_cost qdf24xx_regmove_cost = |
319 | { | |
320 | 2, /* GP2GP */ | |
321 | /* Avoid the use of int<->fp moves for spilling. */ | |
322 | 6, /* GP2FP */ | |
323 | 6, /* FP2GP */ | |
324 | 4 /* FP2FP */ | |
325 | }; | |
326 | ||
d1261ac6 | 327 | static const struct cpu_regmove_cost thunderx2t99_regmove_cost = |
ad611a4c VP |
328 | { |
329 | 1, /* GP2GP */ | |
330 | /* Avoid the use of int<->fp moves for spilling. */ | |
331 | 8, /* GP2FP */ | |
332 | 8, /* FP2GP */ | |
333 | 4 /* FP2FP */ | |
334 | }; | |
335 | ||
8990e73a | 336 | /* Generic costs for vector insn classes. */ |
8990e73a TB |
337 | static const struct cpu_vector_cost generic_vector_cost = |
338 | { | |
cd8ae5ed AP |
339 | 1, /* scalar_int_stmt_cost */ |
340 | 1, /* scalar_fp_stmt_cost */ | |
bd95e655 JG |
341 | 1, /* scalar_load_cost */ |
342 | 1, /* scalar_store_cost */ | |
cd8ae5ed AP |
343 | 1, /* vec_int_stmt_cost */ |
344 | 1, /* vec_fp_stmt_cost */ | |
c428f91c | 345 | 2, /* vec_permute_cost */ |
bd95e655 JG |
346 | 1, /* vec_to_scalar_cost */ |
347 | 1, /* scalar_to_vec_cost */ | |
348 | 1, /* vec_align_load_cost */ | |
349 | 1, /* vec_unalign_load_cost */ | |
350 | 1, /* vec_unalign_store_cost */ | |
351 | 1, /* vec_store_cost */ | |
352 | 3, /* cond_taken_branch_cost */ | |
353 | 1 /* cond_not_taken_branch_cost */ | |
8990e73a TB |
354 | }; |
355 | ||
c3f20327 AP |
356 | /* ThunderX costs for vector insn classes. */ |
357 | static const struct cpu_vector_cost thunderx_vector_cost = | |
358 | { | |
cd8ae5ed AP |
359 | 1, /* scalar_int_stmt_cost */ |
360 | 1, /* scalar_fp_stmt_cost */ | |
c3f20327 AP |
361 | 3, /* scalar_load_cost */ |
362 | 1, /* scalar_store_cost */ | |
cd8ae5ed | 363 | 4, /* vec_int_stmt_cost */ |
b29d7591 | 364 | 1, /* vec_fp_stmt_cost */ |
c3f20327 AP |
365 | 4, /* vec_permute_cost */ |
366 | 2, /* vec_to_scalar_cost */ | |
367 | 2, /* scalar_to_vec_cost */ | |
368 | 3, /* vec_align_load_cost */ | |
7e87a3d9 AP |
369 | 5, /* vec_unalign_load_cost */ |
370 | 5, /* vec_unalign_store_cost */ | |
c3f20327 AP |
371 | 1, /* vec_store_cost */ |
372 | 3, /* cond_taken_branch_cost */ | |
373 | 3 /* cond_not_taken_branch_cost */ | |
374 | }; | |
375 | ||
60bff090 | 376 | /* Generic costs for vector insn classes. */ |
60bff090 JG |
377 | static const struct cpu_vector_cost cortexa57_vector_cost = |
378 | { | |
cd8ae5ed AP |
379 | 1, /* scalar_int_stmt_cost */ |
380 | 1, /* scalar_fp_stmt_cost */ | |
bd95e655 JG |
381 | 4, /* scalar_load_cost */ |
382 | 1, /* scalar_store_cost */ | |
cd8ae5ed AP |
383 | 2, /* vec_int_stmt_cost */ |
384 | 2, /* vec_fp_stmt_cost */ | |
c428f91c | 385 | 3, /* vec_permute_cost */ |
bd95e655 JG |
386 | 8, /* vec_to_scalar_cost */ |
387 | 8, /* scalar_to_vec_cost */ | |
db4a1c18 WD |
388 | 4, /* vec_align_load_cost */ |
389 | 4, /* vec_unalign_load_cost */ | |
bd95e655 JG |
390 | 1, /* vec_unalign_store_cost */ |
391 | 1, /* vec_store_cost */ | |
392 | 1, /* cond_taken_branch_cost */ | |
393 | 1 /* cond_not_taken_branch_cost */ | |
60bff090 JG |
394 | }; |
395 | ||
5ec1ae3b EM |
396 | static const struct cpu_vector_cost exynosm1_vector_cost = |
397 | { | |
cd8ae5ed AP |
398 | 1, /* scalar_int_stmt_cost */ |
399 | 1, /* scalar_fp_stmt_cost */ | |
5ec1ae3b EM |
400 | 5, /* scalar_load_cost */ |
401 | 1, /* scalar_store_cost */ | |
cd8ae5ed AP |
402 | 3, /* vec_int_stmt_cost */ |
403 | 3, /* vec_fp_stmt_cost */ | |
c428f91c | 404 | 3, /* vec_permute_cost */ |
5ec1ae3b EM |
405 | 3, /* vec_to_scalar_cost */ |
406 | 3, /* scalar_to_vec_cost */ | |
407 | 5, /* vec_align_load_cost */ | |
408 | 5, /* vec_unalign_load_cost */ | |
409 | 1, /* vec_unalign_store_cost */ | |
410 | 1, /* vec_store_cost */ | |
411 | 1, /* cond_taken_branch_cost */ | |
412 | 1 /* cond_not_taken_branch_cost */ | |
413 | }; | |
414 | ||
381e27aa | 415 | /* Generic costs for vector insn classes. */ |
381e27aa PT |
416 | static const struct cpu_vector_cost xgene1_vector_cost = |
417 | { | |
cd8ae5ed AP |
418 | 1, /* scalar_int_stmt_cost */ |
419 | 1, /* scalar_fp_stmt_cost */ | |
bd95e655 JG |
420 | 5, /* scalar_load_cost */ |
421 | 1, /* scalar_store_cost */ | |
cd8ae5ed AP |
422 | 2, /* vec_int_stmt_cost */ |
423 | 2, /* vec_fp_stmt_cost */ | |
c428f91c | 424 | 2, /* vec_permute_cost */ |
bd95e655 JG |
425 | 4, /* vec_to_scalar_cost */ |
426 | 4, /* scalar_to_vec_cost */ | |
427 | 10, /* vec_align_load_cost */ | |
428 | 10, /* vec_unalign_load_cost */ | |
429 | 2, /* vec_unalign_store_cost */ | |
430 | 2, /* vec_store_cost */ | |
431 | 2, /* cond_taken_branch_cost */ | |
432 | 1 /* cond_not_taken_branch_cost */ | |
381e27aa PT |
433 | }; |
434 | ||
ad611a4c | 435 | /* Costs for vector insn classes for Vulcan. */ |
d1261ac6 | 436 | static const struct cpu_vector_cost thunderx2t99_vector_cost = |
ad611a4c | 437 | { |
cd8ae5ed AP |
438 | 1, /* scalar_int_stmt_cost */ |
439 | 6, /* scalar_fp_stmt_cost */ | |
ad611a4c VP |
440 | 4, /* scalar_load_cost */ |
441 | 1, /* scalar_store_cost */ | |
cd8ae5ed AP |
442 | 5, /* vec_int_stmt_cost */ |
443 | 6, /* vec_fp_stmt_cost */ | |
ad611a4c VP |
444 | 3, /* vec_permute_cost */ |
445 | 6, /* vec_to_scalar_cost */ | |
446 | 5, /* scalar_to_vec_cost */ | |
447 | 8, /* vec_align_load_cost */ | |
448 | 8, /* vec_unalign_load_cost */ | |
449 | 4, /* vec_unalign_store_cost */ | |
450 | 4, /* vec_store_cost */ | |
451 | 2, /* cond_taken_branch_cost */ | |
452 | 1 /* cond_not_taken_branch_cost */ | |
453 | }; | |
454 | ||
b9066f5a MW |
455 | /* Generic costs for branch instructions. */ |
456 | static const struct cpu_branch_cost generic_branch_cost = | |
457 | { | |
9094d4a4 WD |
458 | 1, /* Predictable. */ |
459 | 3 /* Unpredictable. */ | |
b9066f5a MW |
460 | }; |
461 | ||
9acc9cbe EM |
462 | /* Generic approximation modes. */ |
463 | static const cpu_approx_modes generic_approx_modes = | |
464 | { | |
79a2bc2d | 465 | AARCH64_APPROX_NONE, /* division */ |
98daafa0 | 466 | AARCH64_APPROX_NONE, /* sqrt */ |
9acc9cbe EM |
467 | AARCH64_APPROX_NONE /* recip_sqrt */ |
468 | }; | |
469 | ||
470 | /* Approximation modes for Exynos M1. */ | |
471 | static const cpu_approx_modes exynosm1_approx_modes = | |
472 | { | |
79a2bc2d | 473 | AARCH64_APPROX_NONE, /* division */ |
98daafa0 | 474 | AARCH64_APPROX_ALL, /* sqrt */ |
9acc9cbe EM |
475 | AARCH64_APPROX_ALL /* recip_sqrt */ |
476 | }; | |
477 | ||
478 | /* Approximation modes for X-Gene 1. */ | |
479 | static const cpu_approx_modes xgene1_approx_modes = | |
480 | { | |
79a2bc2d | 481 | AARCH64_APPROX_NONE, /* division */ |
98daafa0 | 482 | AARCH64_APPROX_NONE, /* sqrt */ |
9acc9cbe EM |
483 | AARCH64_APPROX_ALL /* recip_sqrt */ |
484 | }; | |
485 | ||
9d2c6e2e MK |
486 | /* Generic prefetch settings (which disable prefetch). */ |
487 | static const cpu_prefetch_tune generic_prefetch_tune = | |
488 | { | |
489 | 0, /* num_slots */ | |
490 | -1, /* l1_cache_size */ | |
491 | -1, /* l1_cache_line_size */ | |
16b2cafd MK |
492 | -1, /* l2_cache_size */ |
493 | -1 /* default_opt_level */ | |
9d2c6e2e MK |
494 | }; |
495 | ||
496 | static const cpu_prefetch_tune exynosm1_prefetch_tune = | |
497 | { | |
498 | 0, /* num_slots */ | |
499 | -1, /* l1_cache_size */ | |
500 | 64, /* l1_cache_line_size */ | |
16b2cafd MK |
501 | -1, /* l2_cache_size */ |
502 | -1 /* default_opt_level */ | |
9d2c6e2e MK |
503 | }; |
504 | ||
505 | static const cpu_prefetch_tune qdf24xx_prefetch_tune = | |
506 | { | |
70c51b58 MK |
507 | 4, /* num_slots */ |
508 | 32, /* l1_cache_size */ | |
9d2c6e2e | 509 | 64, /* l1_cache_line_size */ |
70c51b58 MK |
510 | 1024, /* l2_cache_size */ |
511 | 3 /* default_opt_level */ | |
9d2c6e2e MK |
512 | }; |
513 | ||
f1e247d0 AP |
514 | static const cpu_prefetch_tune thunderxt88_prefetch_tune = |
515 | { | |
516 | 8, /* num_slots */ | |
517 | 32, /* l1_cache_size */ | |
518 | 128, /* l1_cache_line_size */ | |
519 | 16*1024, /* l2_cache_size */ | |
520 | 3 /* default_opt_level */ | |
521 | }; | |
522 | ||
523 | static const cpu_prefetch_tune thunderx_prefetch_tune = | |
524 | { | |
525 | 8, /* num_slots */ | |
526 | 32, /* l1_cache_size */ | |
527 | 128, /* l1_cache_line_size */ | |
528 | -1, /* l2_cache_size */ | |
529 | -1 /* default_opt_level */ | |
530 | }; | |
531 | ||
9d2c6e2e MK |
532 | static const cpu_prefetch_tune thunderx2t99_prefetch_tune = |
533 | { | |
f1e247d0 AP |
534 | 8, /* num_slots */ |
535 | 32, /* l1_cache_size */ | |
9d2c6e2e | 536 | 64, /* l1_cache_line_size */ |
f1e247d0 | 537 | 256, /* l2_cache_size */ |
16b2cafd | 538 | -1 /* default_opt_level */ |
9d2c6e2e MK |
539 | }; |
540 | ||
43e9d192 IB |
541 | static const struct tune_params generic_tunings = |
542 | { | |
4e2cd668 | 543 | &cortexa57_extra_costs, |
43e9d192 IB |
544 | &generic_addrcost_table, |
545 | &generic_regmove_cost, | |
8990e73a | 546 | &generic_vector_cost, |
b9066f5a | 547 | &generic_branch_cost, |
9acc9cbe | 548 | &generic_approx_modes, |
bd95e655 JG |
549 | 4, /* memmov_cost */ |
550 | 2, /* issue_rate */ | |
e0701ef0 | 551 | (AARCH64_FUSE_AES_AESMC), /* fusible_ops */ |
0b82a5a2 | 552 | 8, /* function_align. */ |
6b13482b WD |
553 | 4, /* jump_align. */ |
554 | 8, /* loop_align. */ | |
cee66c68 WD |
555 | 2, /* int_reassoc_width. */ |
556 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
557 | 1, /* vec_reassoc_width. */ |
558 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 559 | 2, /* min_div_recip_mul_df. */ |
50487d79 | 560 | 0, /* max_case_values. */ |
3b4c0f7e | 561 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
9d2c6e2e MK |
562 | (AARCH64_EXTRA_TUNE_NONE), /* tune_flags. */ |
563 | &generic_prefetch_tune | |
43e9d192 IB |
564 | }; |
565 | ||
1c72a3ca JG |
566 | static const struct tune_params cortexa35_tunings = |
567 | { | |
568 | &cortexa53_extra_costs, | |
569 | &generic_addrcost_table, | |
570 | &cortexa53_regmove_cost, | |
571 | &generic_vector_cost, | |
aca97ef8 | 572 | &generic_branch_cost, |
9acc9cbe | 573 | &generic_approx_modes, |
1c72a3ca JG |
574 | 4, /* memmov_cost */ |
575 | 1, /* issue_rate */ | |
0bc24338 | 576 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD |
1c72a3ca | 577 | | AARCH64_FUSE_MOVK_MOVK | AARCH64_FUSE_ADRP_LDR), /* fusible_ops */ |
d4407370 | 578 | 16, /* function_align. */ |
9779b2e8 | 579 | 4, /* jump_align. */ |
d4407370 | 580 | 8, /* loop_align. */ |
1c72a3ca JG |
581 | 2, /* int_reassoc_width. */ |
582 | 4, /* fp_reassoc_width. */ | |
583 | 1, /* vec_reassoc_width. */ | |
584 | 2, /* min_div_recip_mul_sf. */ | |
585 | 2, /* min_div_recip_mul_df. */ | |
586 | 0, /* max_case_values. */ | |
1c72a3ca | 587 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
9d2c6e2e MK |
588 | (AARCH64_EXTRA_TUNE_NONE), /* tune_flags. */ |
589 | &generic_prefetch_tune | |
1c72a3ca JG |
590 | }; |
591 | ||
984239ad KT |
592 | static const struct tune_params cortexa53_tunings = |
593 | { | |
594 | &cortexa53_extra_costs, | |
595 | &generic_addrcost_table, | |
e4a9c55a | 596 | &cortexa53_regmove_cost, |
984239ad | 597 | &generic_vector_cost, |
aca97ef8 | 598 | &generic_branch_cost, |
9acc9cbe | 599 | &generic_approx_modes, |
bd95e655 JG |
600 | 4, /* memmov_cost */ |
601 | 2, /* issue_rate */ | |
00a8574a | 602 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD |
e9a3a175 | 603 | | AARCH64_FUSE_MOVK_MOVK | AARCH64_FUSE_ADRP_LDR), /* fusible_ops */ |
d4407370 | 604 | 16, /* function_align. */ |
9779b2e8 | 605 | 4, /* jump_align. */ |
d4407370 | 606 | 8, /* loop_align. */ |
cee66c68 WD |
607 | 2, /* int_reassoc_width. */ |
608 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
609 | 1, /* vec_reassoc_width. */ |
610 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 611 | 2, /* min_div_recip_mul_df. */ |
50487d79 | 612 | 0, /* max_case_values. */ |
2d6bc7fa | 613 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
9d2c6e2e MK |
614 | (AARCH64_EXTRA_TUNE_NONE), /* tune_flags. */ |
615 | &generic_prefetch_tune | |
984239ad KT |
616 | }; |
617 | ||
4fd92af6 KT |
618 | static const struct tune_params cortexa57_tunings = |
619 | { | |
620 | &cortexa57_extra_costs, | |
a39d4348 | 621 | &generic_addrcost_table, |
e4a9c55a | 622 | &cortexa57_regmove_cost, |
60bff090 | 623 | &cortexa57_vector_cost, |
aca97ef8 | 624 | &generic_branch_cost, |
9acc9cbe | 625 | &generic_approx_modes, |
bd95e655 JG |
626 | 4, /* memmov_cost */ |
627 | 3, /* issue_rate */ | |
00a8574a | 628 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD |
e9a3a175 | 629 | | AARCH64_FUSE_MOVK_MOVK), /* fusible_ops */ |
0b82a5a2 | 630 | 16, /* function_align. */ |
9779b2e8 | 631 | 4, /* jump_align. */ |
d4407370 | 632 | 8, /* loop_align. */ |
cee66c68 WD |
633 | 2, /* int_reassoc_width. */ |
634 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
635 | 1, /* vec_reassoc_width. */ |
636 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 637 | 2, /* min_div_recip_mul_df. */ |
50487d79 | 638 | 0, /* max_case_values. */ |
2d6bc7fa | 639 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
9d2c6e2e MK |
640 | (AARCH64_EXTRA_TUNE_RENAME_FMA_REGS), /* tune_flags. */ |
641 | &generic_prefetch_tune | |
dfba575f JG |
642 | }; |
643 | ||
644 | static const struct tune_params cortexa72_tunings = | |
645 | { | |
646 | &cortexa57_extra_costs, | |
a39d4348 | 647 | &generic_addrcost_table, |
dfba575f JG |
648 | &cortexa57_regmove_cost, |
649 | &cortexa57_vector_cost, | |
aca97ef8 | 650 | &generic_branch_cost, |
9acc9cbe | 651 | &generic_approx_modes, |
dfba575f JG |
652 | 4, /* memmov_cost */ |
653 | 3, /* issue_rate */ | |
00a8574a | 654 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD |
dfba575f JG |
655 | | AARCH64_FUSE_MOVK_MOVK), /* fusible_ops */ |
656 | 16, /* function_align. */ | |
9779b2e8 | 657 | 4, /* jump_align. */ |
d4407370 | 658 | 8, /* loop_align. */ |
dfba575f JG |
659 | 2, /* int_reassoc_width. */ |
660 | 4, /* fp_reassoc_width. */ | |
661 | 1, /* vec_reassoc_width. */ | |
662 | 2, /* min_div_recip_mul_sf. */ | |
663 | 2, /* min_div_recip_mul_df. */ | |
50487d79 | 664 | 0, /* max_case_values. */ |
0bc24338 | 665 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
9d2c6e2e MK |
666 | (AARCH64_EXTRA_TUNE_NONE), /* tune_flags. */ |
667 | &generic_prefetch_tune | |
4fd92af6 KT |
668 | }; |
669 | ||
4fb570c4 KT |
670 | static const struct tune_params cortexa73_tunings = |
671 | { | |
672 | &cortexa57_extra_costs, | |
a39d4348 | 673 | &generic_addrcost_table, |
4fb570c4 KT |
674 | &cortexa57_regmove_cost, |
675 | &cortexa57_vector_cost, | |
aca97ef8 | 676 | &generic_branch_cost, |
4fb570c4 KT |
677 | &generic_approx_modes, |
678 | 4, /* memmov_cost. */ | |
679 | 2, /* issue_rate. */ | |
680 | (AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD | |
681 | | AARCH64_FUSE_MOVK_MOVK | AARCH64_FUSE_ADRP_LDR), /* fusible_ops */ | |
682 | 16, /* function_align. */ | |
9779b2e8 | 683 | 4, /* jump_align. */ |
d4407370 | 684 | 8, /* loop_align. */ |
4fb570c4 KT |
685 | 2, /* int_reassoc_width. */ |
686 | 4, /* fp_reassoc_width. */ | |
687 | 1, /* vec_reassoc_width. */ | |
688 | 2, /* min_div_recip_mul_sf. */ | |
689 | 2, /* min_div_recip_mul_df. */ | |
690 | 0, /* max_case_values. */ | |
4fb570c4 | 691 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
9d2c6e2e MK |
692 | (AARCH64_EXTRA_TUNE_NONE), /* tune_flags. */ |
693 | &generic_prefetch_tune | |
4fb570c4 KT |
694 | }; |
695 | ||
9d2c6e2e MK |
696 | |
697 | ||
5ec1ae3b EM |
698 | static const struct tune_params exynosm1_tunings = |
699 | { | |
700 | &exynosm1_extra_costs, | |
701 | &exynosm1_addrcost_table, | |
702 | &exynosm1_regmove_cost, | |
703 | &exynosm1_vector_cost, | |
704 | &generic_branch_cost, | |
9acc9cbe | 705 | &exynosm1_approx_modes, |
5ec1ae3b EM |
706 | 4, /* memmov_cost */ |
707 | 3, /* issue_rate */ | |
25cc2199 | 708 | (AARCH64_FUSE_AES_AESMC), /* fusible_ops */ |
5ec1ae3b EM |
709 | 4, /* function_align. */ |
710 | 4, /* jump_align. */ | |
711 | 4, /* loop_align. */ | |
712 | 2, /* int_reassoc_width. */ | |
713 | 4, /* fp_reassoc_width. */ | |
714 | 1, /* vec_reassoc_width. */ | |
715 | 2, /* min_div_recip_mul_sf. */ | |
716 | 2, /* min_div_recip_mul_df. */ | |
717 | 48, /* max_case_values. */ | |
220379df | 718 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
9d2c6e2e MK |
719 | (AARCH64_EXTRA_TUNE_NONE), /* tune_flags. */ |
720 | &exynosm1_prefetch_tune | |
5ec1ae3b EM |
721 | }; |
722 | ||
f1e247d0 AP |
723 | static const struct tune_params thunderxt88_tunings = |
724 | { | |
725 | &thunderx_extra_costs, | |
726 | &generic_addrcost_table, | |
727 | &thunderx_regmove_cost, | |
728 | &thunderx_vector_cost, | |
729 | &generic_branch_cost, | |
730 | &generic_approx_modes, | |
731 | 6, /* memmov_cost */ | |
732 | 2, /* issue_rate */ | |
733 | AARCH64_FUSE_CMP_BRANCH, /* fusible_ops */ | |
734 | 8, /* function_align. */ | |
735 | 8, /* jump_align. */ | |
736 | 8, /* loop_align. */ | |
737 | 2, /* int_reassoc_width. */ | |
738 | 4, /* fp_reassoc_width. */ | |
739 | 1, /* vec_reassoc_width. */ | |
740 | 2, /* min_div_recip_mul_sf. */ | |
741 | 2, /* min_div_recip_mul_df. */ | |
742 | 0, /* max_case_values. */ | |
743 | tune_params::AUTOPREFETCHER_OFF, /* autoprefetcher_model. */ | |
744 | (AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW), /* tune_flags. */ | |
745 | &thunderxt88_prefetch_tune | |
746 | }; | |
747 | ||
d1bcc29f AP |
748 | static const struct tune_params thunderx_tunings = |
749 | { | |
750 | &thunderx_extra_costs, | |
751 | &generic_addrcost_table, | |
752 | &thunderx_regmove_cost, | |
c3f20327 | 753 | &thunderx_vector_cost, |
b9066f5a | 754 | &generic_branch_cost, |
9acc9cbe | 755 | &generic_approx_modes, |
bd95e655 JG |
756 | 6, /* memmov_cost */ |
757 | 2, /* issue_rate */ | |
e9a3a175 | 758 | AARCH64_FUSE_CMP_BRANCH, /* fusible_ops */ |
0b82a5a2 WD |
759 | 8, /* function_align. */ |
760 | 8, /* jump_align. */ | |
761 | 8, /* loop_align. */ | |
cee66c68 WD |
762 | 2, /* int_reassoc_width. */ |
763 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
764 | 1, /* vec_reassoc_width. */ |
765 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 766 | 2, /* min_div_recip_mul_df. */ |
50487d79 | 767 | 0, /* max_case_values. */ |
2d6bc7fa | 768 | tune_params::AUTOPREFETCHER_OFF, /* autoprefetcher_model. */ |
b10f1009 AP |
769 | (AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW |
770 | | AARCH64_EXTRA_TUNE_CHEAP_SHIFT_EXTEND), /* tune_flags. */ | |
f1e247d0 | 771 | &thunderx_prefetch_tune |
d1bcc29f AP |
772 | }; |
773 | ||
381e27aa PT |
774 | static const struct tune_params xgene1_tunings = |
775 | { | |
776 | &xgene1_extra_costs, | |
777 | &xgene1_addrcost_table, | |
778 | &xgene1_regmove_cost, | |
779 | &xgene1_vector_cost, | |
b9066f5a | 780 | &generic_branch_cost, |
9acc9cbe | 781 | &xgene1_approx_modes, |
bd95e655 JG |
782 | 6, /* memmov_cost */ |
783 | 4, /* issue_rate */ | |
e9a3a175 | 784 | AARCH64_FUSE_NOTHING, /* fusible_ops */ |
381e27aa PT |
785 | 16, /* function_align. */ |
786 | 8, /* jump_align. */ | |
787 | 16, /* loop_align. */ | |
788 | 2, /* int_reassoc_width. */ | |
789 | 4, /* fp_reassoc_width. */ | |
50093a33 WD |
790 | 1, /* vec_reassoc_width. */ |
791 | 2, /* min_div_recip_mul_sf. */ | |
dfba575f | 792 | 2, /* min_div_recip_mul_df. */ |
50487d79 | 793 | 0, /* max_case_values. */ |
2d6bc7fa | 794 | tune_params::AUTOPREFETCHER_OFF, /* autoprefetcher_model. */ |
9d2c6e2e MK |
795 | (AARCH64_EXTRA_TUNE_NONE), /* tune_flags. */ |
796 | &generic_prefetch_tune | |
381e27aa PT |
797 | }; |
798 | ||
ee446d9f JW |
799 | static const struct tune_params qdf24xx_tunings = |
800 | { | |
801 | &qdf24xx_extra_costs, | |
a39d4348 | 802 | &generic_addrcost_table, |
ee446d9f JW |
803 | &qdf24xx_regmove_cost, |
804 | &generic_vector_cost, | |
805 | &generic_branch_cost, | |
806 | &generic_approx_modes, | |
807 | 4, /* memmov_cost */ | |
808 | 4, /* issue_rate */ | |
809 | (AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD | |
810 | | AARCH64_FUSE_MOVK_MOVK), /* fuseable_ops */ | |
811 | 16, /* function_align. */ | |
812 | 8, /* jump_align. */ | |
813 | 16, /* loop_align. */ | |
814 | 2, /* int_reassoc_width. */ | |
815 | 4, /* fp_reassoc_width. */ | |
816 | 1, /* vec_reassoc_width. */ | |
817 | 2, /* min_div_recip_mul_sf. */ | |
818 | 2, /* min_div_recip_mul_df. */ | |
819 | 0, /* max_case_values. */ | |
ee446d9f | 820 | tune_params::AUTOPREFETCHER_STRONG, /* autoprefetcher_model. */ |
9d2c6e2e MK |
821 | (AARCH64_EXTRA_TUNE_NONE), /* tune_flags. */ |
822 | &qdf24xx_prefetch_tune | |
ee446d9f JW |
823 | }; |
824 | ||
d1261ac6 | 825 | static const struct tune_params thunderx2t99_tunings = |
ad611a4c | 826 | { |
d1261ac6 AP |
827 | &thunderx2t99_extra_costs, |
828 | &thunderx2t99_addrcost_table, | |
829 | &thunderx2t99_regmove_cost, | |
830 | &thunderx2t99_vector_cost, | |
aca97ef8 | 831 | &generic_branch_cost, |
ad611a4c VP |
832 | &generic_approx_modes, |
833 | 4, /* memmov_cost. */ | |
834 | 4, /* issue_rate. */ | |
00c7c57f JB |
835 | (AARCH64_FUSE_CMP_BRANCH | AARCH64_FUSE_AES_AESMC |
836 | | AARCH64_FUSE_ALU_BRANCH), /* fusible_ops */ | |
ad611a4c VP |
837 | 16, /* function_align. */ |
838 | 8, /* jump_align. */ | |
839 | 16, /* loop_align. */ | |
840 | 3, /* int_reassoc_width. */ | |
841 | 2, /* fp_reassoc_width. */ | |
842 | 2, /* vec_reassoc_width. */ | |
843 | 2, /* min_div_recip_mul_sf. */ | |
844 | 2, /* min_div_recip_mul_df. */ | |
845 | 0, /* max_case_values. */ | |
f1e247d0 | 846 | tune_params::AUTOPREFETCHER_WEAK, /* autoprefetcher_model. */ |
9d2c6e2e MK |
847 | (AARCH64_EXTRA_TUNE_NONE), /* tune_flags. */ |
848 | &thunderx2t99_prefetch_tune | |
ad611a4c VP |
849 | }; |
850 | ||
8dec06f2 JG |
851 | /* Support for fine-grained override of the tuning structures. */ |
852 | struct aarch64_tuning_override_function | |
853 | { | |
854 | const char* name; | |
855 | void (*parse_override)(const char*, struct tune_params*); | |
856 | }; | |
857 | ||
858 | static void aarch64_parse_fuse_string (const char*, struct tune_params*); | |
859 | static void aarch64_parse_tune_string (const char*, struct tune_params*); | |
860 | ||
861 | static const struct aarch64_tuning_override_function | |
862 | aarch64_tuning_override_functions[] = | |
863 | { | |
864 | { "fuse", aarch64_parse_fuse_string }, | |
865 | { "tune", aarch64_parse_tune_string }, | |
866 | { NULL, NULL } | |
867 | }; | |
868 | ||
43e9d192 IB |
869 | /* A processor implementing AArch64. */ |
870 | struct processor | |
871 | { | |
872 | const char *const name; | |
46806c44 KT |
873 | enum aarch64_processor ident; |
874 | enum aarch64_processor sched_core; | |
393ae126 | 875 | enum aarch64_arch arch; |
0c6caaf8 | 876 | unsigned architecture_version; |
43e9d192 IB |
877 | const unsigned long flags; |
878 | const struct tune_params *const tune; | |
879 | }; | |
880 | ||
393ae126 KT |
881 | /* Architectures implementing AArch64. */ |
882 | static const struct processor all_architectures[] = | |
883 | { | |
884 | #define AARCH64_ARCH(NAME, CORE, ARCH_IDENT, ARCH_REV, FLAGS) \ | |
885 | {NAME, CORE, CORE, AARCH64_ARCH_##ARCH_IDENT, ARCH_REV, FLAGS, NULL}, | |
886 | #include "aarch64-arches.def" | |
393ae126 KT |
887 | {NULL, aarch64_none, aarch64_none, aarch64_no_arch, 0, 0, NULL} |
888 | }; | |
889 | ||
43e9d192 IB |
890 | /* Processor cores implementing AArch64. */ |
891 | static const struct processor all_cores[] = | |
892 | { | |
e8fcc9fa | 893 | #define AARCH64_CORE(NAME, IDENT, SCHED, ARCH, FLAGS, COSTS, IMP, PART, VARIANT) \ |
393ae126 KT |
894 | {NAME, IDENT, SCHED, AARCH64_ARCH_##ARCH, \ |
895 | all_architectures[AARCH64_ARCH_##ARCH].architecture_version, \ | |
896 | FLAGS, &COSTS##_tunings}, | |
43e9d192 | 897 | #include "aarch64-cores.def" |
393ae126 KT |
898 | {"generic", generic, cortexa53, AARCH64_ARCH_8A, 8, |
899 | AARCH64_FL_FOR_ARCH8, &generic_tunings}, | |
900 | {NULL, aarch64_none, aarch64_none, aarch64_no_arch, 0, 0, NULL} | |
43e9d192 IB |
901 | }; |
902 | ||
43e9d192 | 903 | |
361fb3ee KT |
904 | /* Target specification. These are populated by the -march, -mtune, -mcpu |
905 | handling code or by target attributes. */ | |
43e9d192 IB |
906 | static const struct processor *selected_arch; |
907 | static const struct processor *selected_cpu; | |
908 | static const struct processor *selected_tune; | |
909 | ||
b175b679 JG |
910 | /* The current tuning set. */ |
911 | struct tune_params aarch64_tune_params = generic_tunings; | |
912 | ||
43e9d192 IB |
913 | #define AARCH64_CPU_DEFAULT_FLAGS ((selected_cpu) ? selected_cpu->flags : 0) |
914 | ||
915 | /* An ISA extension in the co-processor and main instruction set space. */ | |
916 | struct aarch64_option_extension | |
917 | { | |
918 | const char *const name; | |
919 | const unsigned long flags_on; | |
920 | const unsigned long flags_off; | |
921 | }; | |
922 | ||
43e9d192 IB |
923 | typedef enum aarch64_cond_code |
924 | { | |
925 | AARCH64_EQ = 0, AARCH64_NE, AARCH64_CS, AARCH64_CC, AARCH64_MI, AARCH64_PL, | |
926 | AARCH64_VS, AARCH64_VC, AARCH64_HI, AARCH64_LS, AARCH64_GE, AARCH64_LT, | |
927 | AARCH64_GT, AARCH64_LE, AARCH64_AL, AARCH64_NV | |
928 | } | |
929 | aarch64_cc; | |
930 | ||
931 | #define AARCH64_INVERSE_CONDITION_CODE(X) ((aarch64_cc) (((int) X) ^ 1)) | |
932 | ||
933 | /* The condition codes of the processor, and the inverse function. */ | |
934 | static const char * const aarch64_condition_codes[] = | |
935 | { | |
936 | "eq", "ne", "cs", "cc", "mi", "pl", "vs", "vc", | |
937 | "hi", "ls", "ge", "lt", "gt", "le", "al", "nv" | |
938 | }; | |
939 | ||
973d2e01 TP |
940 | /* Generate code to enable conditional branches in functions over 1 MiB. */ |
941 | const char * | |
942 | aarch64_gen_far_branch (rtx * operands, int pos_label, const char * dest, | |
943 | const char * branch_format) | |
944 | { | |
945 | rtx_code_label * tmp_label = gen_label_rtx (); | |
946 | char label_buf[256]; | |
947 | char buffer[128]; | |
948 | ASM_GENERATE_INTERNAL_LABEL (label_buf, dest, | |
949 | CODE_LABEL_NUMBER (tmp_label)); | |
950 | const char *label_ptr = targetm.strip_name_encoding (label_buf); | |
951 | rtx dest_label = operands[pos_label]; | |
952 | operands[pos_label] = tmp_label; | |
953 | ||
954 | snprintf (buffer, sizeof (buffer), "%s%s", branch_format, label_ptr); | |
955 | output_asm_insn (buffer, operands); | |
956 | ||
957 | snprintf (buffer, sizeof (buffer), "b\t%%l%d\n%s:", pos_label, label_ptr); | |
958 | operands[pos_label] = dest_label; | |
959 | output_asm_insn (buffer, operands); | |
960 | return ""; | |
961 | } | |
962 | ||
261fb553 AL |
963 | void |
964 | aarch64_err_no_fpadvsimd (machine_mode mode, const char *msg) | |
965 | { | |
966 | const char *mc = FLOAT_MODE_P (mode) ? "floating-point" : "vector"; | |
967 | if (TARGET_GENERAL_REGS_ONLY) | |
968 | error ("%qs is incompatible with %s %s", "-mgeneral-regs-only", mc, msg); | |
969 | else | |
970 | error ("%qs feature modifier is incompatible with %s %s", "+nofp", mc, msg); | |
971 | } | |
972 | ||
c64f7d37 WD |
973 | /* Implement TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS. |
974 | The register allocator chooses ALL_REGS if FP_REGS and GENERAL_REGS have | |
31e2b5a3 WD |
975 | the same cost even if ALL_REGS has a much larger cost. ALL_REGS is also |
976 | used if the cost of both FP_REGS and GENERAL_REGS is lower than the memory | |
977 | cost (in this case the best class is the lowest cost one). Using ALL_REGS | |
978 | irrespectively of its cost results in bad allocations with many redundant | |
979 | int<->FP moves which are expensive on various cores. | |
980 | To avoid this we don't allow ALL_REGS as the allocno class, but force a | |
981 | decision between FP_REGS and GENERAL_REGS. We use the allocno class if it | |
982 | isn't ALL_REGS. Similarly, use the best class if it isn't ALL_REGS. | |
983 | Otherwise set the allocno class depending on the mode. | |
984 | The result of this is that it is no longer inefficient to have a higher | |
985 | memory move cost than the register move cost. | |
986 | */ | |
c64f7d37 WD |
987 | |
988 | static reg_class_t | |
31e2b5a3 WD |
989 | aarch64_ira_change_pseudo_allocno_class (int regno, reg_class_t allocno_class, |
990 | reg_class_t best_class) | |
c64f7d37 | 991 | { |
b8506a8a | 992 | machine_mode mode; |
c64f7d37 WD |
993 | |
994 | if (allocno_class != ALL_REGS) | |
995 | return allocno_class; | |
996 | ||
31e2b5a3 WD |
997 | if (best_class != ALL_REGS) |
998 | return best_class; | |
999 | ||
c64f7d37 WD |
1000 | mode = PSEUDO_REGNO_MODE (regno); |
1001 | return FLOAT_MODE_P (mode) || VECTOR_MODE_P (mode) ? FP_REGS : GENERAL_REGS; | |
1002 | } | |
1003 | ||
26e0ff94 | 1004 | static unsigned int |
b8506a8a | 1005 | aarch64_min_divisions_for_recip_mul (machine_mode mode) |
26e0ff94 | 1006 | { |
50093a33 | 1007 | if (GET_MODE_UNIT_SIZE (mode) == 4) |
b175b679 JG |
1008 | return aarch64_tune_params.min_div_recip_mul_sf; |
1009 | return aarch64_tune_params.min_div_recip_mul_df; | |
26e0ff94 WD |
1010 | } |
1011 | ||
cee66c68 WD |
1012 | static int |
1013 | aarch64_reassociation_width (unsigned opc ATTRIBUTE_UNUSED, | |
b8506a8a | 1014 | machine_mode mode) |
cee66c68 WD |
1015 | { |
1016 | if (VECTOR_MODE_P (mode)) | |
b175b679 | 1017 | return aarch64_tune_params.vec_reassoc_width; |
cee66c68 | 1018 | if (INTEGRAL_MODE_P (mode)) |
b175b679 | 1019 | return aarch64_tune_params.int_reassoc_width; |
cee66c68 | 1020 | if (FLOAT_MODE_P (mode)) |
b175b679 | 1021 | return aarch64_tune_params.fp_reassoc_width; |
cee66c68 WD |
1022 | return 1; |
1023 | } | |
1024 | ||
43e9d192 IB |
1025 | /* Provide a mapping from gcc register numbers to dwarf register numbers. */ |
1026 | unsigned | |
1027 | aarch64_dbx_register_number (unsigned regno) | |
1028 | { | |
1029 | if (GP_REGNUM_P (regno)) | |
1030 | return AARCH64_DWARF_R0 + regno - R0_REGNUM; | |
1031 | else if (regno == SP_REGNUM) | |
1032 | return AARCH64_DWARF_SP; | |
1033 | else if (FP_REGNUM_P (regno)) | |
1034 | return AARCH64_DWARF_V0 + regno - V0_REGNUM; | |
1035 | ||
1036 | /* Return values >= DWARF_FRAME_REGISTERS indicate that there is no | |
1037 | equivalent DWARF register. */ | |
1038 | return DWARF_FRAME_REGISTERS; | |
1039 | } | |
1040 | ||
1041 | /* Return TRUE if MODE is any of the large INT modes. */ | |
1042 | static bool | |
ef4bddc2 | 1043 | aarch64_vect_struct_mode_p (machine_mode mode) |
43e9d192 IB |
1044 | { |
1045 | return mode == OImode || mode == CImode || mode == XImode; | |
1046 | } | |
1047 | ||
1048 | /* Return TRUE if MODE is any of the vector modes. */ | |
1049 | static bool | |
ef4bddc2 | 1050 | aarch64_vector_mode_p (machine_mode mode) |
43e9d192 IB |
1051 | { |
1052 | return aarch64_vector_mode_supported_p (mode) | |
1053 | || aarch64_vect_struct_mode_p (mode); | |
1054 | } | |
1055 | ||
1056 | /* Implement target hook TARGET_ARRAY_MODE_SUPPORTED_P. */ | |
1057 | static bool | |
ef4bddc2 | 1058 | aarch64_array_mode_supported_p (machine_mode mode, |
43e9d192 IB |
1059 | unsigned HOST_WIDE_INT nelems) |
1060 | { | |
1061 | if (TARGET_SIMD | |
635e66fe AL |
1062 | && (AARCH64_VALID_SIMD_QREG_MODE (mode) |
1063 | || AARCH64_VALID_SIMD_DREG_MODE (mode)) | |
43e9d192 IB |
1064 | && (nelems >= 2 && nelems <= 4)) |
1065 | return true; | |
1066 | ||
1067 | return false; | |
1068 | } | |
1069 | ||
1070 | /* Implement HARD_REGNO_NREGS. */ | |
1071 | ||
1072 | int | |
ef4bddc2 | 1073 | aarch64_hard_regno_nregs (unsigned regno, machine_mode mode) |
43e9d192 IB |
1074 | { |
1075 | switch (aarch64_regno_regclass (regno)) | |
1076 | { | |
1077 | case FP_REGS: | |
1078 | case FP_LO_REGS: | |
1079 | return (GET_MODE_SIZE (mode) + UNITS_PER_VREG - 1) / UNITS_PER_VREG; | |
1080 | default: | |
1081 | return (GET_MODE_SIZE (mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
1082 | } | |
1083 | gcc_unreachable (); | |
1084 | } | |
1085 | ||
1086 | /* Implement HARD_REGNO_MODE_OK. */ | |
1087 | ||
1088 | int | |
ef4bddc2 | 1089 | aarch64_hard_regno_mode_ok (unsigned regno, machine_mode mode) |
43e9d192 IB |
1090 | { |
1091 | if (GET_MODE_CLASS (mode) == MODE_CC) | |
1092 | return regno == CC_REGNUM; | |
1093 | ||
9259db42 YZ |
1094 | if (regno == SP_REGNUM) |
1095 | /* The purpose of comparing with ptr_mode is to support the | |
1096 | global register variable associated with the stack pointer | |
1097 | register via the syntax of asm ("wsp") in ILP32. */ | |
1098 | return mode == Pmode || mode == ptr_mode; | |
1099 | ||
1100 | if (regno == FRAME_POINTER_REGNUM || regno == ARG_POINTER_REGNUM) | |
43e9d192 IB |
1101 | return mode == Pmode; |
1102 | ||
1103 | if (GP_REGNUM_P (regno) && ! aarch64_vect_struct_mode_p (mode)) | |
1104 | return 1; | |
1105 | ||
1106 | if (FP_REGNUM_P (regno)) | |
1107 | { | |
1108 | if (aarch64_vect_struct_mode_p (mode)) | |
1109 | return | |
1110 | (regno + aarch64_hard_regno_nregs (regno, mode) - 1) <= V31_REGNUM; | |
1111 | else | |
1112 | return 1; | |
1113 | } | |
1114 | ||
1115 | return 0; | |
1116 | } | |
1117 | ||
73d9ac6a | 1118 | /* Implement HARD_REGNO_CALLER_SAVE_MODE. */ |
ef4bddc2 | 1119 | machine_mode |
73d9ac6a | 1120 | aarch64_hard_regno_caller_save_mode (unsigned regno, unsigned nregs, |
ef4bddc2 | 1121 | machine_mode mode) |
73d9ac6a IB |
1122 | { |
1123 | /* Handle modes that fit within single registers. */ | |
1124 | if (nregs == 1 && GET_MODE_SIZE (mode) <= 16) | |
1125 | { | |
1126 | if (GET_MODE_SIZE (mode) >= 4) | |
1127 | return mode; | |
1128 | else | |
1129 | return SImode; | |
1130 | } | |
1131 | /* Fall back to generic for multi-reg and very large modes. */ | |
1132 | else | |
1133 | return choose_hard_reg_mode (regno, nregs, false); | |
1134 | } | |
1135 | ||
43e9d192 IB |
1136 | /* Return true if calls to DECL should be treated as |
1137 | long-calls (ie called via a register). */ | |
1138 | static bool | |
1139 | aarch64_decl_is_long_call_p (const_tree decl ATTRIBUTE_UNUSED) | |
1140 | { | |
1141 | return false; | |
1142 | } | |
1143 | ||
1144 | /* Return true if calls to symbol-ref SYM should be treated as | |
1145 | long-calls (ie called via a register). */ | |
1146 | bool | |
1147 | aarch64_is_long_call_p (rtx sym) | |
1148 | { | |
1149 | return aarch64_decl_is_long_call_p (SYMBOL_REF_DECL (sym)); | |
1150 | } | |
1151 | ||
b60d63cb JW |
1152 | /* Return true if calls to symbol-ref SYM should not go through |
1153 | plt stubs. */ | |
1154 | ||
1155 | bool | |
1156 | aarch64_is_noplt_call_p (rtx sym) | |
1157 | { | |
1158 | const_tree decl = SYMBOL_REF_DECL (sym); | |
1159 | ||
1160 | if (flag_pic | |
1161 | && decl | |
1162 | && (!flag_plt | |
1163 | || lookup_attribute ("noplt", DECL_ATTRIBUTES (decl))) | |
1164 | && !targetm.binds_local_p (decl)) | |
1165 | return true; | |
1166 | ||
1167 | return false; | |
1168 | } | |
1169 | ||
43e9d192 IB |
1170 | /* Return true if the offsets to a zero/sign-extract operation |
1171 | represent an expression that matches an extend operation. The | |
1172 | operands represent the paramters from | |
1173 | ||
4745e701 | 1174 | (extract:MODE (mult (reg) (MULT_IMM)) (EXTRACT_IMM) (const_int 0)). */ |
43e9d192 | 1175 | bool |
ef4bddc2 | 1176 | aarch64_is_extend_from_extract (machine_mode mode, rtx mult_imm, |
43e9d192 IB |
1177 | rtx extract_imm) |
1178 | { | |
1179 | HOST_WIDE_INT mult_val, extract_val; | |
1180 | ||
1181 | if (! CONST_INT_P (mult_imm) || ! CONST_INT_P (extract_imm)) | |
1182 | return false; | |
1183 | ||
1184 | mult_val = INTVAL (mult_imm); | |
1185 | extract_val = INTVAL (extract_imm); | |
1186 | ||
1187 | if (extract_val > 8 | |
1188 | && extract_val < GET_MODE_BITSIZE (mode) | |
1189 | && exact_log2 (extract_val & ~7) > 0 | |
1190 | && (extract_val & 7) <= 4 | |
1191 | && mult_val == (1 << (extract_val & 7))) | |
1192 | return true; | |
1193 | ||
1194 | return false; | |
1195 | } | |
1196 | ||
1197 | /* Emit an insn that's a simple single-set. Both the operands must be | |
1198 | known to be valid. */ | |
827ab47a | 1199 | inline static rtx_insn * |
43e9d192 IB |
1200 | emit_set_insn (rtx x, rtx y) |
1201 | { | |
f7df4a84 | 1202 | return emit_insn (gen_rtx_SET (x, y)); |
43e9d192 IB |
1203 | } |
1204 | ||
1205 | /* X and Y are two things to compare using CODE. Emit the compare insn and | |
1206 | return the rtx for register 0 in the proper mode. */ | |
1207 | rtx | |
1208 | aarch64_gen_compare_reg (RTX_CODE code, rtx x, rtx y) | |
1209 | { | |
ef4bddc2 | 1210 | machine_mode mode = SELECT_CC_MODE (code, x, y); |
43e9d192 IB |
1211 | rtx cc_reg = gen_rtx_REG (mode, CC_REGNUM); |
1212 | ||
1213 | emit_set_insn (cc_reg, gen_rtx_COMPARE (mode, x, y)); | |
1214 | return cc_reg; | |
1215 | } | |
1216 | ||
1217 | /* Build the SYMBOL_REF for __tls_get_addr. */ | |
1218 | ||
1219 | static GTY(()) rtx tls_get_addr_libfunc; | |
1220 | ||
1221 | rtx | |
1222 | aarch64_tls_get_addr (void) | |
1223 | { | |
1224 | if (!tls_get_addr_libfunc) | |
1225 | tls_get_addr_libfunc = init_one_libfunc ("__tls_get_addr"); | |
1226 | return tls_get_addr_libfunc; | |
1227 | } | |
1228 | ||
1229 | /* Return the TLS model to use for ADDR. */ | |
1230 | ||
1231 | static enum tls_model | |
1232 | tls_symbolic_operand_type (rtx addr) | |
1233 | { | |
1234 | enum tls_model tls_kind = TLS_MODEL_NONE; | |
1235 | rtx sym, addend; | |
1236 | ||
1237 | if (GET_CODE (addr) == CONST) | |
1238 | { | |
1239 | split_const (addr, &sym, &addend); | |
1240 | if (GET_CODE (sym) == SYMBOL_REF) | |
1241 | tls_kind = SYMBOL_REF_TLS_MODEL (sym); | |
1242 | } | |
1243 | else if (GET_CODE (addr) == SYMBOL_REF) | |
1244 | tls_kind = SYMBOL_REF_TLS_MODEL (addr); | |
1245 | ||
1246 | return tls_kind; | |
1247 | } | |
1248 | ||
1249 | /* We'll allow lo_sum's in addresses in our legitimate addresses | |
1250 | so that combine would take care of combining addresses where | |
1251 | necessary, but for generation purposes, we'll generate the address | |
1252 | as : | |
1253 | RTL Absolute | |
1254 | tmp = hi (symbol_ref); adrp x1, foo | |
1255 | dest = lo_sum (tmp, symbol_ref); add dest, x1, :lo_12:foo | |
1256 | nop | |
1257 | ||
1258 | PIC TLS | |
1259 | adrp x1, :got:foo adrp tmp, :tlsgd:foo | |
1260 | ldr x1, [:got_lo12:foo] add dest, tmp, :tlsgd_lo12:foo | |
1261 | bl __tls_get_addr | |
1262 | nop | |
1263 | ||
1264 | Load TLS symbol, depending on TLS mechanism and TLS access model. | |
1265 | ||
1266 | Global Dynamic - Traditional TLS: | |
1267 | adrp tmp, :tlsgd:imm | |
1268 | add dest, tmp, #:tlsgd_lo12:imm | |
1269 | bl __tls_get_addr | |
1270 | ||
1271 | Global Dynamic - TLS Descriptors: | |
1272 | adrp dest, :tlsdesc:imm | |
1273 | ldr tmp, [dest, #:tlsdesc_lo12:imm] | |
1274 | add dest, dest, #:tlsdesc_lo12:imm | |
1275 | blr tmp | |
1276 | mrs tp, tpidr_el0 | |
1277 | add dest, dest, tp | |
1278 | ||
1279 | Initial Exec: | |
1280 | mrs tp, tpidr_el0 | |
1281 | adrp tmp, :gottprel:imm | |
1282 | ldr dest, [tmp, #:gottprel_lo12:imm] | |
1283 | add dest, dest, tp | |
1284 | ||
1285 | Local Exec: | |
1286 | mrs tp, tpidr_el0 | |
0699caae RL |
1287 | add t0, tp, #:tprel_hi12:imm, lsl #12 |
1288 | add t0, t0, #:tprel_lo12_nc:imm | |
43e9d192 IB |
1289 | */ |
1290 | ||
1291 | static void | |
1292 | aarch64_load_symref_appropriately (rtx dest, rtx imm, | |
1293 | enum aarch64_symbol_type type) | |
1294 | { | |
1295 | switch (type) | |
1296 | { | |
1297 | case SYMBOL_SMALL_ABSOLUTE: | |
1298 | { | |
28514dda | 1299 | /* In ILP32, the mode of dest can be either SImode or DImode. */ |
43e9d192 | 1300 | rtx tmp_reg = dest; |
ef4bddc2 | 1301 | machine_mode mode = GET_MODE (dest); |
28514dda YZ |
1302 | |
1303 | gcc_assert (mode == Pmode || mode == ptr_mode); | |
1304 | ||
43e9d192 | 1305 | if (can_create_pseudo_p ()) |
28514dda | 1306 | tmp_reg = gen_reg_rtx (mode); |
43e9d192 | 1307 | |
28514dda | 1308 | emit_move_insn (tmp_reg, gen_rtx_HIGH (mode, imm)); |
43e9d192 IB |
1309 | emit_insn (gen_add_losym (dest, tmp_reg, imm)); |
1310 | return; | |
1311 | } | |
1312 | ||
a5350ddc | 1313 | case SYMBOL_TINY_ABSOLUTE: |
f7df4a84 | 1314 | emit_insn (gen_rtx_SET (dest, imm)); |
a5350ddc CSS |
1315 | return; |
1316 | ||
1b1e81f8 JW |
1317 | case SYMBOL_SMALL_GOT_28K: |
1318 | { | |
1319 | machine_mode mode = GET_MODE (dest); | |
1320 | rtx gp_rtx = pic_offset_table_rtx; | |
53021678 JW |
1321 | rtx insn; |
1322 | rtx mem; | |
1b1e81f8 JW |
1323 | |
1324 | /* NOTE: pic_offset_table_rtx can be NULL_RTX, because we can reach | |
1325 | here before rtl expand. Tree IVOPT will generate rtl pattern to | |
1326 | decide rtx costs, in which case pic_offset_table_rtx is not | |
1327 | initialized. For that case no need to generate the first adrp | |
026c3cfd | 1328 | instruction as the final cost for global variable access is |
1b1e81f8 JW |
1329 | one instruction. */ |
1330 | if (gp_rtx != NULL) | |
1331 | { | |
1332 | /* -fpic for -mcmodel=small allow 32K GOT table size (but we are | |
1333 | using the page base as GOT base, the first page may be wasted, | |
1334 | in the worst scenario, there is only 28K space for GOT). | |
1335 | ||
1336 | The generate instruction sequence for accessing global variable | |
1337 | is: | |
1338 | ||
a3957742 | 1339 | ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym] |
1b1e81f8 JW |
1340 | |
1341 | Only one instruction needed. But we must initialize | |
1342 | pic_offset_table_rtx properly. We generate initialize insn for | |
1343 | every global access, and allow CSE to remove all redundant. | |
1344 | ||
1345 | The final instruction sequences will look like the following | |
1346 | for multiply global variables access. | |
1347 | ||
a3957742 | 1348 | adrp pic_offset_table_rtx, _GLOBAL_OFFSET_TABLE_ |
1b1e81f8 | 1349 | |
a3957742 JW |
1350 | ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym1] |
1351 | ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym2] | |
1352 | ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym3] | |
1353 | ... */ | |
1b1e81f8 JW |
1354 | |
1355 | rtx s = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_"); | |
1356 | crtl->uses_pic_offset_table = 1; | |
1357 | emit_move_insn (gp_rtx, gen_rtx_HIGH (Pmode, s)); | |
1358 | ||
1359 | if (mode != GET_MODE (gp_rtx)) | |
4ba8f0a3 AP |
1360 | gp_rtx = gen_lowpart (mode, gp_rtx); |
1361 | ||
1b1e81f8 JW |
1362 | } |
1363 | ||
1364 | if (mode == ptr_mode) | |
1365 | { | |
1366 | if (mode == DImode) | |
53021678 | 1367 | insn = gen_ldr_got_small_28k_di (dest, gp_rtx, imm); |
1b1e81f8 | 1368 | else |
53021678 JW |
1369 | insn = gen_ldr_got_small_28k_si (dest, gp_rtx, imm); |
1370 | ||
1371 | mem = XVECEXP (SET_SRC (insn), 0, 0); | |
1b1e81f8 JW |
1372 | } |
1373 | else | |
1374 | { | |
1375 | gcc_assert (mode == Pmode); | |
53021678 JW |
1376 | |
1377 | insn = gen_ldr_got_small_28k_sidi (dest, gp_rtx, imm); | |
1378 | mem = XVECEXP (XEXP (SET_SRC (insn), 0), 0, 0); | |
1b1e81f8 JW |
1379 | } |
1380 | ||
53021678 JW |
1381 | /* The operand is expected to be MEM. Whenever the related insn |
1382 | pattern changed, above code which calculate mem should be | |
1383 | updated. */ | |
1384 | gcc_assert (GET_CODE (mem) == MEM); | |
1385 | MEM_READONLY_P (mem) = 1; | |
1386 | MEM_NOTRAP_P (mem) = 1; | |
1387 | emit_insn (insn); | |
1b1e81f8 JW |
1388 | return; |
1389 | } | |
1390 | ||
6642bdb4 | 1391 | case SYMBOL_SMALL_GOT_4G: |
43e9d192 | 1392 | { |
28514dda YZ |
1393 | /* In ILP32, the mode of dest can be either SImode or DImode, |
1394 | while the got entry is always of SImode size. The mode of | |
1395 | dest depends on how dest is used: if dest is assigned to a | |
1396 | pointer (e.g. in the memory), it has SImode; it may have | |
1397 | DImode if dest is dereferenced to access the memeory. | |
1398 | This is why we have to handle three different ldr_got_small | |
1399 | patterns here (two patterns for ILP32). */ | |
53021678 JW |
1400 | |
1401 | rtx insn; | |
1402 | rtx mem; | |
43e9d192 | 1403 | rtx tmp_reg = dest; |
ef4bddc2 | 1404 | machine_mode mode = GET_MODE (dest); |
28514dda | 1405 | |
43e9d192 | 1406 | if (can_create_pseudo_p ()) |
28514dda YZ |
1407 | tmp_reg = gen_reg_rtx (mode); |
1408 | ||
1409 | emit_move_insn (tmp_reg, gen_rtx_HIGH (mode, imm)); | |
1410 | if (mode == ptr_mode) | |
1411 | { | |
1412 | if (mode == DImode) | |
53021678 | 1413 | insn = gen_ldr_got_small_di (dest, tmp_reg, imm); |
28514dda | 1414 | else |
53021678 JW |
1415 | insn = gen_ldr_got_small_si (dest, tmp_reg, imm); |
1416 | ||
1417 | mem = XVECEXP (SET_SRC (insn), 0, 0); | |
28514dda YZ |
1418 | } |
1419 | else | |
1420 | { | |
1421 | gcc_assert (mode == Pmode); | |
53021678 JW |
1422 | |
1423 | insn = gen_ldr_got_small_sidi (dest, tmp_reg, imm); | |
1424 | mem = XVECEXP (XEXP (SET_SRC (insn), 0), 0, 0); | |
28514dda YZ |
1425 | } |
1426 | ||
53021678 JW |
1427 | gcc_assert (GET_CODE (mem) == MEM); |
1428 | MEM_READONLY_P (mem) = 1; | |
1429 | MEM_NOTRAP_P (mem) = 1; | |
1430 | emit_insn (insn); | |
43e9d192 IB |
1431 | return; |
1432 | } | |
1433 | ||
1434 | case SYMBOL_SMALL_TLSGD: | |
1435 | { | |
5d8a22a5 | 1436 | rtx_insn *insns; |
23b88fda N |
1437 | machine_mode mode = GET_MODE (dest); |
1438 | rtx result = gen_rtx_REG (mode, R0_REGNUM); | |
43e9d192 IB |
1439 | |
1440 | start_sequence (); | |
23b88fda N |
1441 | if (TARGET_ILP32) |
1442 | aarch64_emit_call_insn (gen_tlsgd_small_si (result, imm)); | |
1443 | else | |
1444 | aarch64_emit_call_insn (gen_tlsgd_small_di (result, imm)); | |
43e9d192 IB |
1445 | insns = get_insns (); |
1446 | end_sequence (); | |
1447 | ||
1448 | RTL_CONST_CALL_P (insns) = 1; | |
1449 | emit_libcall_block (insns, dest, result, imm); | |
1450 | return; | |
1451 | } | |
1452 | ||
1453 | case SYMBOL_SMALL_TLSDESC: | |
1454 | { | |
ef4bddc2 | 1455 | machine_mode mode = GET_MODE (dest); |
621ad2de | 1456 | rtx x0 = gen_rtx_REG (mode, R0_REGNUM); |
43e9d192 IB |
1457 | rtx tp; |
1458 | ||
621ad2de AP |
1459 | gcc_assert (mode == Pmode || mode == ptr_mode); |
1460 | ||
2876a13f JW |
1461 | /* In ILP32, the got entry is always of SImode size. Unlike |
1462 | small GOT, the dest is fixed at reg 0. */ | |
1463 | if (TARGET_ILP32) | |
1464 | emit_insn (gen_tlsdesc_small_si (imm)); | |
621ad2de | 1465 | else |
2876a13f | 1466 | emit_insn (gen_tlsdesc_small_di (imm)); |
43e9d192 | 1467 | tp = aarch64_load_tp (NULL); |
621ad2de AP |
1468 | |
1469 | if (mode != Pmode) | |
1470 | tp = gen_lowpart (mode, tp); | |
1471 | ||
2876a13f | 1472 | emit_insn (gen_rtx_SET (dest, gen_rtx_PLUS (mode, tp, x0))); |
43e9d192 IB |
1473 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); |
1474 | return; | |
1475 | } | |
1476 | ||
79496620 | 1477 | case SYMBOL_SMALL_TLSIE: |
43e9d192 | 1478 | { |
621ad2de AP |
1479 | /* In ILP32, the mode of dest can be either SImode or DImode, |
1480 | while the got entry is always of SImode size. The mode of | |
1481 | dest depends on how dest is used: if dest is assigned to a | |
1482 | pointer (e.g. in the memory), it has SImode; it may have | |
1483 | DImode if dest is dereferenced to access the memeory. | |
1484 | This is why we have to handle three different tlsie_small | |
1485 | patterns here (two patterns for ILP32). */ | |
ef4bddc2 | 1486 | machine_mode mode = GET_MODE (dest); |
621ad2de | 1487 | rtx tmp_reg = gen_reg_rtx (mode); |
43e9d192 | 1488 | rtx tp = aarch64_load_tp (NULL); |
621ad2de AP |
1489 | |
1490 | if (mode == ptr_mode) | |
1491 | { | |
1492 | if (mode == DImode) | |
1493 | emit_insn (gen_tlsie_small_di (tmp_reg, imm)); | |
1494 | else | |
1495 | { | |
1496 | emit_insn (gen_tlsie_small_si (tmp_reg, imm)); | |
1497 | tp = gen_lowpart (mode, tp); | |
1498 | } | |
1499 | } | |
1500 | else | |
1501 | { | |
1502 | gcc_assert (mode == Pmode); | |
1503 | emit_insn (gen_tlsie_small_sidi (tmp_reg, imm)); | |
1504 | } | |
1505 | ||
f7df4a84 | 1506 | emit_insn (gen_rtx_SET (dest, gen_rtx_PLUS (mode, tp, tmp_reg))); |
43e9d192 IB |
1507 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); |
1508 | return; | |
1509 | } | |
1510 | ||
cbf5629e | 1511 | case SYMBOL_TLSLE12: |
d18ba284 | 1512 | case SYMBOL_TLSLE24: |
cbf5629e JW |
1513 | case SYMBOL_TLSLE32: |
1514 | case SYMBOL_TLSLE48: | |
43e9d192 | 1515 | { |
cbf5629e | 1516 | machine_mode mode = GET_MODE (dest); |
43e9d192 | 1517 | rtx tp = aarch64_load_tp (NULL); |
e6f7f0e9 | 1518 | |
cbf5629e JW |
1519 | if (mode != Pmode) |
1520 | tp = gen_lowpart (mode, tp); | |
1521 | ||
1522 | switch (type) | |
1523 | { | |
1524 | case SYMBOL_TLSLE12: | |
1525 | emit_insn ((mode == DImode ? gen_tlsle12_di : gen_tlsle12_si) | |
1526 | (dest, tp, imm)); | |
1527 | break; | |
1528 | case SYMBOL_TLSLE24: | |
1529 | emit_insn ((mode == DImode ? gen_tlsle24_di : gen_tlsle24_si) | |
1530 | (dest, tp, imm)); | |
1531 | break; | |
1532 | case SYMBOL_TLSLE32: | |
1533 | emit_insn ((mode == DImode ? gen_tlsle32_di : gen_tlsle32_si) | |
1534 | (dest, imm)); | |
1535 | emit_insn ((mode == DImode ? gen_adddi3 : gen_addsi3) | |
1536 | (dest, dest, tp)); | |
1537 | break; | |
1538 | case SYMBOL_TLSLE48: | |
1539 | emit_insn ((mode == DImode ? gen_tlsle48_di : gen_tlsle48_si) | |
1540 | (dest, imm)); | |
1541 | emit_insn ((mode == DImode ? gen_adddi3 : gen_addsi3) | |
1542 | (dest, dest, tp)); | |
1543 | break; | |
1544 | default: | |
1545 | gcc_unreachable (); | |
1546 | } | |
e6f7f0e9 | 1547 | |
43e9d192 IB |
1548 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); |
1549 | return; | |
1550 | } | |
1551 | ||
87dd8ab0 MS |
1552 | case SYMBOL_TINY_GOT: |
1553 | emit_insn (gen_ldr_got_tiny (dest, imm)); | |
1554 | return; | |
1555 | ||
5ae7caad JW |
1556 | case SYMBOL_TINY_TLSIE: |
1557 | { | |
1558 | machine_mode mode = GET_MODE (dest); | |
1559 | rtx tp = aarch64_load_tp (NULL); | |
1560 | ||
1561 | if (mode == ptr_mode) | |
1562 | { | |
1563 | if (mode == DImode) | |
1564 | emit_insn (gen_tlsie_tiny_di (dest, imm, tp)); | |
1565 | else | |
1566 | { | |
1567 | tp = gen_lowpart (mode, tp); | |
1568 | emit_insn (gen_tlsie_tiny_si (dest, imm, tp)); | |
1569 | } | |
1570 | } | |
1571 | else | |
1572 | { | |
1573 | gcc_assert (mode == Pmode); | |
1574 | emit_insn (gen_tlsie_tiny_sidi (dest, imm, tp)); | |
1575 | } | |
1576 | ||
1577 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); | |
1578 | return; | |
1579 | } | |
1580 | ||
43e9d192 IB |
1581 | default: |
1582 | gcc_unreachable (); | |
1583 | } | |
1584 | } | |
1585 | ||
1586 | /* Emit a move from SRC to DEST. Assume that the move expanders can | |
1587 | handle all moves if !can_create_pseudo_p (). The distinction is | |
1588 | important because, unlike emit_move_insn, the move expanders know | |
1589 | how to force Pmode objects into the constant pool even when the | |
1590 | constant pool address is not itself legitimate. */ | |
1591 | static rtx | |
1592 | aarch64_emit_move (rtx dest, rtx src) | |
1593 | { | |
1594 | return (can_create_pseudo_p () | |
1595 | ? emit_move_insn (dest, src) | |
1596 | : emit_move_insn_1 (dest, src)); | |
1597 | } | |
1598 | ||
030d03b8 RE |
1599 | /* Split a 128-bit move operation into two 64-bit move operations, |
1600 | taking care to handle partial overlap of register to register | |
1601 | copies. Special cases are needed when moving between GP regs and | |
1602 | FP regs. SRC can be a register, constant or memory; DST a register | |
1603 | or memory. If either operand is memory it must not have any side | |
1604 | effects. */ | |
43e9d192 IB |
1605 | void |
1606 | aarch64_split_128bit_move (rtx dst, rtx src) | |
1607 | { | |
030d03b8 RE |
1608 | rtx dst_lo, dst_hi; |
1609 | rtx src_lo, src_hi; | |
43e9d192 | 1610 | |
ef4bddc2 | 1611 | machine_mode mode = GET_MODE (dst); |
12dc6974 | 1612 | |
030d03b8 RE |
1613 | gcc_assert (mode == TImode || mode == TFmode); |
1614 | gcc_assert (!(side_effects_p (src) || side_effects_p (dst))); | |
1615 | gcc_assert (mode == GET_MODE (src) || GET_MODE (src) == VOIDmode); | |
43e9d192 IB |
1616 | |
1617 | if (REG_P (dst) && REG_P (src)) | |
1618 | { | |
030d03b8 RE |
1619 | int src_regno = REGNO (src); |
1620 | int dst_regno = REGNO (dst); | |
43e9d192 | 1621 | |
030d03b8 | 1622 | /* Handle FP <-> GP regs. */ |
43e9d192 IB |
1623 | if (FP_REGNUM_P (dst_regno) && GP_REGNUM_P (src_regno)) |
1624 | { | |
030d03b8 RE |
1625 | src_lo = gen_lowpart (word_mode, src); |
1626 | src_hi = gen_highpart (word_mode, src); | |
1627 | ||
1628 | if (mode == TImode) | |
1629 | { | |
1630 | emit_insn (gen_aarch64_movtilow_di (dst, src_lo)); | |
1631 | emit_insn (gen_aarch64_movtihigh_di (dst, src_hi)); | |
1632 | } | |
1633 | else | |
1634 | { | |
1635 | emit_insn (gen_aarch64_movtflow_di (dst, src_lo)); | |
1636 | emit_insn (gen_aarch64_movtfhigh_di (dst, src_hi)); | |
1637 | } | |
1638 | return; | |
43e9d192 IB |
1639 | } |
1640 | else if (GP_REGNUM_P (dst_regno) && FP_REGNUM_P (src_regno)) | |
1641 | { | |
030d03b8 RE |
1642 | dst_lo = gen_lowpart (word_mode, dst); |
1643 | dst_hi = gen_highpart (word_mode, dst); | |
1644 | ||
1645 | if (mode == TImode) | |
1646 | { | |
1647 | emit_insn (gen_aarch64_movdi_tilow (dst_lo, src)); | |
1648 | emit_insn (gen_aarch64_movdi_tihigh (dst_hi, src)); | |
1649 | } | |
1650 | else | |
1651 | { | |
1652 | emit_insn (gen_aarch64_movdi_tflow (dst_lo, src)); | |
1653 | emit_insn (gen_aarch64_movdi_tfhigh (dst_hi, src)); | |
1654 | } | |
1655 | return; | |
43e9d192 | 1656 | } |
43e9d192 IB |
1657 | } |
1658 | ||
030d03b8 RE |
1659 | dst_lo = gen_lowpart (word_mode, dst); |
1660 | dst_hi = gen_highpart (word_mode, dst); | |
1661 | src_lo = gen_lowpart (word_mode, src); | |
1662 | src_hi = gen_highpart_mode (word_mode, mode, src); | |
1663 | ||
1664 | /* At most one pairing may overlap. */ | |
1665 | if (reg_overlap_mentioned_p (dst_lo, src_hi)) | |
1666 | { | |
1667 | aarch64_emit_move (dst_hi, src_hi); | |
1668 | aarch64_emit_move (dst_lo, src_lo); | |
1669 | } | |
1670 | else | |
1671 | { | |
1672 | aarch64_emit_move (dst_lo, src_lo); | |
1673 | aarch64_emit_move (dst_hi, src_hi); | |
1674 | } | |
43e9d192 IB |
1675 | } |
1676 | ||
1677 | bool | |
1678 | aarch64_split_128bit_move_p (rtx dst, rtx src) | |
1679 | { | |
1680 | return (! REG_P (src) | |
1681 | || ! (FP_REGNUM_P (REGNO (dst)) && FP_REGNUM_P (REGNO (src)))); | |
1682 | } | |
1683 | ||
8b033a8a SN |
1684 | /* Split a complex SIMD combine. */ |
1685 | ||
1686 | void | |
1687 | aarch64_split_simd_combine (rtx dst, rtx src1, rtx src2) | |
1688 | { | |
ef4bddc2 RS |
1689 | machine_mode src_mode = GET_MODE (src1); |
1690 | machine_mode dst_mode = GET_MODE (dst); | |
8b033a8a SN |
1691 | |
1692 | gcc_assert (VECTOR_MODE_P (dst_mode)); | |
a977dc0c MC |
1693 | gcc_assert (register_operand (dst, dst_mode) |
1694 | && register_operand (src1, src_mode) | |
1695 | && register_operand (src2, src_mode)); | |
8b033a8a | 1696 | |
a977dc0c | 1697 | rtx (*gen) (rtx, rtx, rtx); |
8b033a8a | 1698 | |
a977dc0c MC |
1699 | switch (src_mode) |
1700 | { | |
4e10a5a7 | 1701 | case E_V8QImode: |
a977dc0c MC |
1702 | gen = gen_aarch64_simd_combinev8qi; |
1703 | break; | |
4e10a5a7 | 1704 | case E_V4HImode: |
a977dc0c MC |
1705 | gen = gen_aarch64_simd_combinev4hi; |
1706 | break; | |
4e10a5a7 | 1707 | case E_V2SImode: |
a977dc0c MC |
1708 | gen = gen_aarch64_simd_combinev2si; |
1709 | break; | |
4e10a5a7 | 1710 | case E_V4HFmode: |
a977dc0c MC |
1711 | gen = gen_aarch64_simd_combinev4hf; |
1712 | break; | |
4e10a5a7 | 1713 | case E_V2SFmode: |
a977dc0c MC |
1714 | gen = gen_aarch64_simd_combinev2sf; |
1715 | break; | |
4e10a5a7 | 1716 | case E_DImode: |
a977dc0c MC |
1717 | gen = gen_aarch64_simd_combinedi; |
1718 | break; | |
4e10a5a7 | 1719 | case E_DFmode: |
a977dc0c MC |
1720 | gen = gen_aarch64_simd_combinedf; |
1721 | break; | |
1722 | default: | |
1723 | gcc_unreachable (); | |
8b033a8a | 1724 | } |
a977dc0c MC |
1725 | |
1726 | emit_insn (gen (dst, src1, src2)); | |
1727 | return; | |
8b033a8a SN |
1728 | } |
1729 | ||
fd4842cd SN |
1730 | /* Split a complex SIMD move. */ |
1731 | ||
1732 | void | |
1733 | aarch64_split_simd_move (rtx dst, rtx src) | |
1734 | { | |
ef4bddc2 RS |
1735 | machine_mode src_mode = GET_MODE (src); |
1736 | machine_mode dst_mode = GET_MODE (dst); | |
fd4842cd SN |
1737 | |
1738 | gcc_assert (VECTOR_MODE_P (dst_mode)); | |
1739 | ||
1740 | if (REG_P (dst) && REG_P (src)) | |
1741 | { | |
c59b7e28 SN |
1742 | rtx (*gen) (rtx, rtx); |
1743 | ||
fd4842cd SN |
1744 | gcc_assert (VECTOR_MODE_P (src_mode)); |
1745 | ||
1746 | switch (src_mode) | |
1747 | { | |
4e10a5a7 | 1748 | case E_V16QImode: |
c59b7e28 | 1749 | gen = gen_aarch64_split_simd_movv16qi; |
fd4842cd | 1750 | break; |
4e10a5a7 | 1751 | case E_V8HImode: |
c59b7e28 | 1752 | gen = gen_aarch64_split_simd_movv8hi; |
fd4842cd | 1753 | break; |
4e10a5a7 | 1754 | case E_V4SImode: |
c59b7e28 | 1755 | gen = gen_aarch64_split_simd_movv4si; |
fd4842cd | 1756 | break; |
4e10a5a7 | 1757 | case E_V2DImode: |
c59b7e28 | 1758 | gen = gen_aarch64_split_simd_movv2di; |
fd4842cd | 1759 | break; |
4e10a5a7 | 1760 | case E_V8HFmode: |
71a11456 AL |
1761 | gen = gen_aarch64_split_simd_movv8hf; |
1762 | break; | |
4e10a5a7 | 1763 | case E_V4SFmode: |
c59b7e28 | 1764 | gen = gen_aarch64_split_simd_movv4sf; |
fd4842cd | 1765 | break; |
4e10a5a7 | 1766 | case E_V2DFmode: |
c59b7e28 | 1767 | gen = gen_aarch64_split_simd_movv2df; |
fd4842cd SN |
1768 | break; |
1769 | default: | |
1770 | gcc_unreachable (); | |
1771 | } | |
c59b7e28 SN |
1772 | |
1773 | emit_insn (gen (dst, src)); | |
fd4842cd SN |
1774 | return; |
1775 | } | |
1776 | } | |
1777 | ||
ef22810a RH |
1778 | bool |
1779 | aarch64_zero_extend_const_eq (machine_mode xmode, rtx x, | |
1780 | machine_mode ymode, rtx y) | |
1781 | { | |
1782 | rtx r = simplify_const_unary_operation (ZERO_EXTEND, xmode, y, ymode); | |
1783 | gcc_assert (r != NULL); | |
1784 | return rtx_equal_p (x, r); | |
1785 | } | |
1786 | ||
1787 | ||
43e9d192 | 1788 | static rtx |
ef4bddc2 | 1789 | aarch64_force_temporary (machine_mode mode, rtx x, rtx value) |
43e9d192 IB |
1790 | { |
1791 | if (can_create_pseudo_p ()) | |
e18b4a81 | 1792 | return force_reg (mode, value); |
43e9d192 IB |
1793 | else |
1794 | { | |
1795 | x = aarch64_emit_move (x, value); | |
1796 | return x; | |
1797 | } | |
1798 | } | |
1799 | ||
1800 | ||
1801 | static rtx | |
ef4bddc2 | 1802 | aarch64_add_offset (machine_mode mode, rtx temp, rtx reg, HOST_WIDE_INT offset) |
43e9d192 | 1803 | { |
9c023bf0 | 1804 | if (!aarch64_plus_immediate (GEN_INT (offset), mode)) |
43e9d192 IB |
1805 | { |
1806 | rtx high; | |
1807 | /* Load the full offset into a register. This | |
1808 | might be improvable in the future. */ | |
1809 | high = GEN_INT (offset); | |
1810 | offset = 0; | |
e18b4a81 YZ |
1811 | high = aarch64_force_temporary (mode, temp, high); |
1812 | reg = aarch64_force_temporary (mode, temp, | |
1813 | gen_rtx_PLUS (mode, high, reg)); | |
43e9d192 IB |
1814 | } |
1815 | return plus_constant (mode, reg, offset); | |
1816 | } | |
1817 | ||
82614948 RR |
1818 | static int |
1819 | aarch64_internal_mov_immediate (rtx dest, rtx imm, bool generate, | |
1820 | machine_mode mode) | |
43e9d192 | 1821 | { |
43e9d192 | 1822 | int i; |
9a4865db WD |
1823 | unsigned HOST_WIDE_INT val, val2, mask; |
1824 | int one_match, zero_match; | |
1825 | int num_insns; | |
43e9d192 | 1826 | |
9a4865db WD |
1827 | val = INTVAL (imm); |
1828 | ||
1829 | if (aarch64_move_imm (val, mode)) | |
43e9d192 | 1830 | { |
82614948 | 1831 | if (generate) |
f7df4a84 | 1832 | emit_insn (gen_rtx_SET (dest, imm)); |
9a4865db | 1833 | return 1; |
43e9d192 IB |
1834 | } |
1835 | ||
9de00935 TC |
1836 | /* Check to see if the low 32 bits are either 0xffffXXXX or 0xXXXXffff |
1837 | (with XXXX non-zero). In that case check to see if the move can be done in | |
1838 | a smaller mode. */ | |
1839 | val2 = val & 0xffffffff; | |
1840 | if (mode == DImode | |
1841 | && aarch64_move_imm (val2, SImode) | |
1842 | && (((val >> 32) & 0xffff) == 0 || (val >> 48) == 0)) | |
1843 | { | |
1844 | if (generate) | |
1845 | emit_insn (gen_rtx_SET (dest, GEN_INT (val2))); | |
1846 | ||
1847 | /* Check if we have to emit a second instruction by checking to see | |
1848 | if any of the upper 32 bits of the original DI mode value is set. */ | |
1849 | if (val == val2) | |
1850 | return 1; | |
1851 | ||
1852 | i = (val >> 48) ? 48 : 32; | |
1853 | ||
1854 | if (generate) | |
1855 | emit_insn (gen_insv_immdi (dest, GEN_INT (i), | |
1856 | GEN_INT ((val >> i) & 0xffff))); | |
1857 | ||
1858 | return 2; | |
1859 | } | |
1860 | ||
9a4865db | 1861 | if ((val >> 32) == 0 || mode == SImode) |
43e9d192 | 1862 | { |
82614948 RR |
1863 | if (generate) |
1864 | { | |
9a4865db WD |
1865 | emit_insn (gen_rtx_SET (dest, GEN_INT (val & 0xffff))); |
1866 | if (mode == SImode) | |
1867 | emit_insn (gen_insv_immsi (dest, GEN_INT (16), | |
1868 | GEN_INT ((val >> 16) & 0xffff))); | |
1869 | else | |
1870 | emit_insn (gen_insv_immdi (dest, GEN_INT (16), | |
1871 | GEN_INT ((val >> 16) & 0xffff))); | |
82614948 | 1872 | } |
9a4865db | 1873 | return 2; |
43e9d192 IB |
1874 | } |
1875 | ||
1876 | /* Remaining cases are all for DImode. */ | |
1877 | ||
43e9d192 | 1878 | mask = 0xffff; |
9a4865db WD |
1879 | zero_match = ((val & mask) == 0) + ((val & (mask << 16)) == 0) + |
1880 | ((val & (mask << 32)) == 0) + ((val & (mask << 48)) == 0); | |
1881 | one_match = ((~val & mask) == 0) + ((~val & (mask << 16)) == 0) + | |
1882 | ((~val & (mask << 32)) == 0) + ((~val & (mask << 48)) == 0); | |
43e9d192 | 1883 | |
62c8d76c | 1884 | if (zero_match != 2 && one_match != 2) |
43e9d192 | 1885 | { |
62c8d76c WD |
1886 | /* Try emitting a bitmask immediate with a movk replacing 16 bits. |
1887 | For a 64-bit bitmask try whether changing 16 bits to all ones or | |
1888 | zeroes creates a valid bitmask. To check any repeated bitmask, | |
1889 | try using 16 bits from the other 32-bit half of val. */ | |
43e9d192 | 1890 | |
62c8d76c | 1891 | for (i = 0; i < 64; i += 16, mask <<= 16) |
43e9d192 | 1892 | { |
62c8d76c WD |
1893 | val2 = val & ~mask; |
1894 | if (val2 != val && aarch64_bitmask_imm (val2, mode)) | |
1895 | break; | |
1896 | val2 = val | mask; | |
1897 | if (val2 != val && aarch64_bitmask_imm (val2, mode)) | |
1898 | break; | |
1899 | val2 = val2 & ~mask; | |
1900 | val2 = val2 | (((val2 >> 32) | (val2 << 32)) & mask); | |
1901 | if (val2 != val && aarch64_bitmask_imm (val2, mode)) | |
1902 | break; | |
43e9d192 | 1903 | } |
62c8d76c | 1904 | if (i != 64) |
43e9d192 | 1905 | { |
62c8d76c | 1906 | if (generate) |
43e9d192 | 1907 | { |
62c8d76c WD |
1908 | emit_insn (gen_rtx_SET (dest, GEN_INT (val2))); |
1909 | emit_insn (gen_insv_immdi (dest, GEN_INT (i), | |
9a4865db | 1910 | GEN_INT ((val >> i) & 0xffff))); |
43e9d192 | 1911 | } |
1312b1ba | 1912 | return 2; |
43e9d192 IB |
1913 | } |
1914 | } | |
1915 | ||
9a4865db WD |
1916 | /* Generate 2-4 instructions, skipping 16 bits of all zeroes or ones which |
1917 | are emitted by the initial mov. If one_match > zero_match, skip set bits, | |
1918 | otherwise skip zero bits. */ | |
2c274197 | 1919 | |
9a4865db | 1920 | num_insns = 1; |
43e9d192 | 1921 | mask = 0xffff; |
9a4865db WD |
1922 | val2 = one_match > zero_match ? ~val : val; |
1923 | i = (val2 & mask) != 0 ? 0 : (val2 & (mask << 16)) != 0 ? 16 : 32; | |
1924 | ||
1925 | if (generate) | |
1926 | emit_insn (gen_rtx_SET (dest, GEN_INT (one_match > zero_match | |
1927 | ? (val | ~(mask << i)) | |
1928 | : (val & (mask << i))))); | |
1929 | for (i += 16; i < 64; i += 16) | |
43e9d192 | 1930 | { |
9a4865db WD |
1931 | if ((val2 & (mask << i)) == 0) |
1932 | continue; | |
1933 | if (generate) | |
1934 | emit_insn (gen_insv_immdi (dest, GEN_INT (i), | |
1935 | GEN_INT ((val >> i) & 0xffff))); | |
1936 | num_insns ++; | |
82614948 RR |
1937 | } |
1938 | ||
1939 | return num_insns; | |
1940 | } | |
1941 | ||
1942 | ||
1943 | void | |
1944 | aarch64_expand_mov_immediate (rtx dest, rtx imm) | |
1945 | { | |
1946 | machine_mode mode = GET_MODE (dest); | |
1947 | ||
1948 | gcc_assert (mode == SImode || mode == DImode); | |
1949 | ||
1950 | /* Check on what type of symbol it is. */ | |
1951 | if (GET_CODE (imm) == SYMBOL_REF | |
1952 | || GET_CODE (imm) == LABEL_REF | |
1953 | || GET_CODE (imm) == CONST) | |
1954 | { | |
1955 | rtx mem, base, offset; | |
1956 | enum aarch64_symbol_type sty; | |
1957 | ||
1958 | /* If we have (const (plus symbol offset)), separate out the offset | |
1959 | before we start classifying the symbol. */ | |
1960 | split_const (imm, &base, &offset); | |
1961 | ||
a6e0bfa7 | 1962 | sty = aarch64_classify_symbol (base, offset); |
82614948 RR |
1963 | switch (sty) |
1964 | { | |
1965 | case SYMBOL_FORCE_TO_MEM: | |
1966 | if (offset != const0_rtx | |
1967 | && targetm.cannot_force_const_mem (mode, imm)) | |
1968 | { | |
1969 | gcc_assert (can_create_pseudo_p ()); | |
1970 | base = aarch64_force_temporary (mode, dest, base); | |
1971 | base = aarch64_add_offset (mode, NULL, base, INTVAL (offset)); | |
1972 | aarch64_emit_move (dest, base); | |
1973 | return; | |
1974 | } | |
b4f50fd4 | 1975 | |
82614948 RR |
1976 | mem = force_const_mem (ptr_mode, imm); |
1977 | gcc_assert (mem); | |
b4f50fd4 RR |
1978 | |
1979 | /* If we aren't generating PC relative literals, then | |
1980 | we need to expand the literal pool access carefully. | |
1981 | This is something that needs to be done in a number | |
1982 | of places, so could well live as a separate function. */ | |
9ee6540a | 1983 | if (!aarch64_pcrelative_literal_loads) |
b4f50fd4 RR |
1984 | { |
1985 | gcc_assert (can_create_pseudo_p ()); | |
1986 | base = gen_reg_rtx (ptr_mode); | |
1987 | aarch64_expand_mov_immediate (base, XEXP (mem, 0)); | |
00eee3fa WD |
1988 | if (ptr_mode != Pmode) |
1989 | base = convert_memory_address (Pmode, base); | |
b4f50fd4 RR |
1990 | mem = gen_rtx_MEM (ptr_mode, base); |
1991 | } | |
1992 | ||
82614948 RR |
1993 | if (mode != ptr_mode) |
1994 | mem = gen_rtx_ZERO_EXTEND (mode, mem); | |
b4f50fd4 | 1995 | |
f7df4a84 | 1996 | emit_insn (gen_rtx_SET (dest, mem)); |
b4f50fd4 | 1997 | |
82614948 RR |
1998 | return; |
1999 | ||
2000 | case SYMBOL_SMALL_TLSGD: | |
2001 | case SYMBOL_SMALL_TLSDESC: | |
79496620 | 2002 | case SYMBOL_SMALL_TLSIE: |
1b1e81f8 | 2003 | case SYMBOL_SMALL_GOT_28K: |
6642bdb4 | 2004 | case SYMBOL_SMALL_GOT_4G: |
82614948 | 2005 | case SYMBOL_TINY_GOT: |
5ae7caad | 2006 | case SYMBOL_TINY_TLSIE: |
82614948 RR |
2007 | if (offset != const0_rtx) |
2008 | { | |
2009 | gcc_assert(can_create_pseudo_p ()); | |
2010 | base = aarch64_force_temporary (mode, dest, base); | |
2011 | base = aarch64_add_offset (mode, NULL, base, INTVAL (offset)); | |
2012 | aarch64_emit_move (dest, base); | |
2013 | return; | |
2014 | } | |
2015 | /* FALLTHRU */ | |
2016 | ||
82614948 RR |
2017 | case SYMBOL_SMALL_ABSOLUTE: |
2018 | case SYMBOL_TINY_ABSOLUTE: | |
cbf5629e | 2019 | case SYMBOL_TLSLE12: |
d18ba284 | 2020 | case SYMBOL_TLSLE24: |
cbf5629e JW |
2021 | case SYMBOL_TLSLE32: |
2022 | case SYMBOL_TLSLE48: | |
82614948 RR |
2023 | aarch64_load_symref_appropriately (dest, imm, sty); |
2024 | return; | |
2025 | ||
2026 | default: | |
2027 | gcc_unreachable (); | |
2028 | } | |
2029 | } | |
2030 | ||
2031 | if (!CONST_INT_P (imm)) | |
2032 | { | |
2033 | if (GET_CODE (imm) == HIGH) | |
f7df4a84 | 2034 | emit_insn (gen_rtx_SET (dest, imm)); |
82614948 RR |
2035 | else |
2036 | { | |
2037 | rtx mem = force_const_mem (mode, imm); | |
2038 | gcc_assert (mem); | |
f7df4a84 | 2039 | emit_insn (gen_rtx_SET (dest, mem)); |
43e9d192 | 2040 | } |
82614948 RR |
2041 | |
2042 | return; | |
43e9d192 | 2043 | } |
82614948 RR |
2044 | |
2045 | aarch64_internal_mov_immediate (dest, imm, true, GET_MODE (dest)); | |
43e9d192 IB |
2046 | } |
2047 | ||
5be6b295 WD |
2048 | /* Add DELTA to REGNUM in mode MODE. SCRATCHREG can be used to hold a |
2049 | temporary value if necessary. FRAME_RELATED_P should be true if | |
2050 | the RTX_FRAME_RELATED flag should be set and CFA adjustments added | |
2051 | to the generated instructions. If SCRATCHREG is known to hold | |
2052 | abs (delta), EMIT_MOVE_IMM can be set to false to avoid emitting the | |
2053 | immediate again. | |
2054 | ||
2055 | Since this function may be used to adjust the stack pointer, we must | |
2056 | ensure that it cannot cause transient stack deallocation (for example | |
2057 | by first incrementing SP and then decrementing when adjusting by a | |
2058 | large immediate). */ | |
c4ddc43a JW |
2059 | |
2060 | static void | |
5be6b295 WD |
2061 | aarch64_add_constant_internal (machine_mode mode, int regnum, int scratchreg, |
2062 | HOST_WIDE_INT delta, bool frame_related_p, | |
2063 | bool emit_move_imm) | |
c4ddc43a JW |
2064 | { |
2065 | HOST_WIDE_INT mdelta = abs_hwi (delta); | |
2066 | rtx this_rtx = gen_rtx_REG (mode, regnum); | |
37d6a4b7 | 2067 | rtx_insn *insn; |
c4ddc43a | 2068 | |
c4ddc43a JW |
2069 | if (!mdelta) |
2070 | return; | |
2071 | ||
5be6b295 | 2072 | /* Single instruction adjustment. */ |
c4ddc43a JW |
2073 | if (aarch64_uimm12_shift (mdelta)) |
2074 | { | |
37d6a4b7 JW |
2075 | insn = emit_insn (gen_add2_insn (this_rtx, GEN_INT (delta))); |
2076 | RTX_FRAME_RELATED_P (insn) = frame_related_p; | |
c4ddc43a JW |
2077 | return; |
2078 | } | |
2079 | ||
5be6b295 WD |
2080 | /* Emit 2 additions/subtractions if the adjustment is less than 24 bits. |
2081 | Only do this if mdelta is not a 16-bit move as adjusting using a move | |
2082 | is better. */ | |
2083 | if (mdelta < 0x1000000 && !aarch64_move_imm (mdelta, mode)) | |
c4ddc43a JW |
2084 | { |
2085 | HOST_WIDE_INT low_off = mdelta & 0xfff; | |
2086 | ||
2087 | low_off = delta < 0 ? -low_off : low_off; | |
37d6a4b7 JW |
2088 | insn = emit_insn (gen_add2_insn (this_rtx, GEN_INT (low_off))); |
2089 | RTX_FRAME_RELATED_P (insn) = frame_related_p; | |
2090 | insn = emit_insn (gen_add2_insn (this_rtx, GEN_INT (delta - low_off))); | |
2091 | RTX_FRAME_RELATED_P (insn) = frame_related_p; | |
c4ddc43a JW |
2092 | return; |
2093 | } | |
2094 | ||
5be6b295 | 2095 | /* Emit a move immediate if required and an addition/subtraction. */ |
c4ddc43a | 2096 | rtx scratch_rtx = gen_rtx_REG (mode, scratchreg); |
5be6b295 WD |
2097 | if (emit_move_imm) |
2098 | aarch64_internal_mov_immediate (scratch_rtx, GEN_INT (mdelta), true, mode); | |
2099 | insn = emit_insn (delta < 0 ? gen_sub2_insn (this_rtx, scratch_rtx) | |
2100 | : gen_add2_insn (this_rtx, scratch_rtx)); | |
37d6a4b7 JW |
2101 | if (frame_related_p) |
2102 | { | |
2103 | RTX_FRAME_RELATED_P (insn) = frame_related_p; | |
2104 | rtx adj = plus_constant (mode, this_rtx, delta); | |
2105 | add_reg_note (insn , REG_CFA_ADJUST_CFA, gen_rtx_SET (this_rtx, adj)); | |
2106 | } | |
c4ddc43a JW |
2107 | } |
2108 | ||
5be6b295 WD |
2109 | static inline void |
2110 | aarch64_add_constant (machine_mode mode, int regnum, int scratchreg, | |
2111 | HOST_WIDE_INT delta) | |
2112 | { | |
2113 | aarch64_add_constant_internal (mode, regnum, scratchreg, delta, false, true); | |
2114 | } | |
2115 | ||
2116 | static inline void | |
2117 | aarch64_add_sp (int scratchreg, HOST_WIDE_INT delta, bool emit_move_imm) | |
2118 | { | |
2119 | aarch64_add_constant_internal (Pmode, SP_REGNUM, scratchreg, delta, | |
2120 | true, emit_move_imm); | |
2121 | } | |
2122 | ||
2123 | static inline void | |
2124 | aarch64_sub_sp (int scratchreg, HOST_WIDE_INT delta, bool frame_related_p) | |
2125 | { | |
2126 | aarch64_add_constant_internal (Pmode, SP_REGNUM, scratchreg, -delta, | |
2127 | frame_related_p, true); | |
2128 | } | |
2129 | ||
43e9d192 | 2130 | static bool |
fee9ba42 JW |
2131 | aarch64_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED, |
2132 | tree exp ATTRIBUTE_UNUSED) | |
43e9d192 | 2133 | { |
fee9ba42 | 2134 | /* Currently, always true. */ |
43e9d192 IB |
2135 | return true; |
2136 | } | |
2137 | ||
2138 | /* Implement TARGET_PASS_BY_REFERENCE. */ | |
2139 | ||
2140 | static bool | |
2141 | aarch64_pass_by_reference (cumulative_args_t pcum ATTRIBUTE_UNUSED, | |
ef4bddc2 | 2142 | machine_mode mode, |
43e9d192 IB |
2143 | const_tree type, |
2144 | bool named ATTRIBUTE_UNUSED) | |
2145 | { | |
2146 | HOST_WIDE_INT size; | |
ef4bddc2 | 2147 | machine_mode dummymode; |
43e9d192 IB |
2148 | int nregs; |
2149 | ||
2150 | /* GET_MODE_SIZE (BLKmode) is useless since it is 0. */ | |
2151 | size = (mode == BLKmode && type) | |
2152 | ? int_size_in_bytes (type) : (int) GET_MODE_SIZE (mode); | |
2153 | ||
aadc1c43 MHD |
2154 | /* Aggregates are passed by reference based on their size. */ |
2155 | if (type && AGGREGATE_TYPE_P (type)) | |
43e9d192 | 2156 | { |
aadc1c43 | 2157 | size = int_size_in_bytes (type); |
43e9d192 IB |
2158 | } |
2159 | ||
2160 | /* Variable sized arguments are always returned by reference. */ | |
2161 | if (size < 0) | |
2162 | return true; | |
2163 | ||
2164 | /* Can this be a candidate to be passed in fp/simd register(s)? */ | |
2165 | if (aarch64_vfp_is_call_or_return_candidate (mode, type, | |
2166 | &dummymode, &nregs, | |
2167 | NULL)) | |
2168 | return false; | |
2169 | ||
2170 | /* Arguments which are variable sized or larger than 2 registers are | |
2171 | passed by reference unless they are a homogenous floating point | |
2172 | aggregate. */ | |
2173 | return size > 2 * UNITS_PER_WORD; | |
2174 | } | |
2175 | ||
2176 | /* Return TRUE if VALTYPE is padded to its least significant bits. */ | |
2177 | static bool | |
2178 | aarch64_return_in_msb (const_tree valtype) | |
2179 | { | |
ef4bddc2 | 2180 | machine_mode dummy_mode; |
43e9d192 IB |
2181 | int dummy_int; |
2182 | ||
2183 | /* Never happens in little-endian mode. */ | |
2184 | if (!BYTES_BIG_ENDIAN) | |
2185 | return false; | |
2186 | ||
2187 | /* Only composite types smaller than or equal to 16 bytes can | |
2188 | be potentially returned in registers. */ | |
2189 | if (!aarch64_composite_type_p (valtype, TYPE_MODE (valtype)) | |
2190 | || int_size_in_bytes (valtype) <= 0 | |
2191 | || int_size_in_bytes (valtype) > 16) | |
2192 | return false; | |
2193 | ||
2194 | /* But not a composite that is an HFA (Homogeneous Floating-point Aggregate) | |
2195 | or an HVA (Homogeneous Short-Vector Aggregate); such a special composite | |
2196 | is always passed/returned in the least significant bits of fp/simd | |
2197 | register(s). */ | |
2198 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (valtype), valtype, | |
2199 | &dummy_mode, &dummy_int, NULL)) | |
2200 | return false; | |
2201 | ||
2202 | return true; | |
2203 | } | |
2204 | ||
2205 | /* Implement TARGET_FUNCTION_VALUE. | |
2206 | Define how to find the value returned by a function. */ | |
2207 | ||
2208 | static rtx | |
2209 | aarch64_function_value (const_tree type, const_tree func, | |
2210 | bool outgoing ATTRIBUTE_UNUSED) | |
2211 | { | |
ef4bddc2 | 2212 | machine_mode mode; |
43e9d192 IB |
2213 | int unsignedp; |
2214 | int count; | |
ef4bddc2 | 2215 | machine_mode ag_mode; |
43e9d192 IB |
2216 | |
2217 | mode = TYPE_MODE (type); | |
2218 | if (INTEGRAL_TYPE_P (type)) | |
2219 | mode = promote_function_mode (type, mode, &unsignedp, func, 1); | |
2220 | ||
2221 | if (aarch64_return_in_msb (type)) | |
2222 | { | |
2223 | HOST_WIDE_INT size = int_size_in_bytes (type); | |
2224 | ||
2225 | if (size % UNITS_PER_WORD != 0) | |
2226 | { | |
2227 | size += UNITS_PER_WORD - size % UNITS_PER_WORD; | |
2228 | mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0); | |
2229 | } | |
2230 | } | |
2231 | ||
2232 | if (aarch64_vfp_is_call_or_return_candidate (mode, type, | |
2233 | &ag_mode, &count, NULL)) | |
2234 | { | |
2235 | if (!aarch64_composite_type_p (type, mode)) | |
2236 | { | |
2237 | gcc_assert (count == 1 && mode == ag_mode); | |
2238 | return gen_rtx_REG (mode, V0_REGNUM); | |
2239 | } | |
2240 | else | |
2241 | { | |
2242 | int i; | |
2243 | rtx par; | |
2244 | ||
2245 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (count)); | |
2246 | for (i = 0; i < count; i++) | |
2247 | { | |
2248 | rtx tmp = gen_rtx_REG (ag_mode, V0_REGNUM + i); | |
2249 | tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp, | |
2250 | GEN_INT (i * GET_MODE_SIZE (ag_mode))); | |
2251 | XVECEXP (par, 0, i) = tmp; | |
2252 | } | |
2253 | return par; | |
2254 | } | |
2255 | } | |
2256 | else | |
2257 | return gen_rtx_REG (mode, R0_REGNUM); | |
2258 | } | |
2259 | ||
2260 | /* Implements TARGET_FUNCTION_VALUE_REGNO_P. | |
2261 | Return true if REGNO is the number of a hard register in which the values | |
2262 | of called function may come back. */ | |
2263 | ||
2264 | static bool | |
2265 | aarch64_function_value_regno_p (const unsigned int regno) | |
2266 | { | |
2267 | /* Maximum of 16 bytes can be returned in the general registers. Examples | |
2268 | of 16-byte return values are: 128-bit integers and 16-byte small | |
2269 | structures (excluding homogeneous floating-point aggregates). */ | |
2270 | if (regno == R0_REGNUM || regno == R1_REGNUM) | |
2271 | return true; | |
2272 | ||
2273 | /* Up to four fp/simd registers can return a function value, e.g. a | |
2274 | homogeneous floating-point aggregate having four members. */ | |
2275 | if (regno >= V0_REGNUM && regno < V0_REGNUM + HA_MAX_NUM_FLDS) | |
d5726973 | 2276 | return TARGET_FLOAT; |
43e9d192 IB |
2277 | |
2278 | return false; | |
2279 | } | |
2280 | ||
2281 | /* Implement TARGET_RETURN_IN_MEMORY. | |
2282 | ||
2283 | If the type T of the result of a function is such that | |
2284 | void func (T arg) | |
2285 | would require that arg be passed as a value in a register (or set of | |
2286 | registers) according to the parameter passing rules, then the result | |
2287 | is returned in the same registers as would be used for such an | |
2288 | argument. */ | |
2289 | ||
2290 | static bool | |
2291 | aarch64_return_in_memory (const_tree type, const_tree fndecl ATTRIBUTE_UNUSED) | |
2292 | { | |
2293 | HOST_WIDE_INT size; | |
ef4bddc2 | 2294 | machine_mode ag_mode; |
43e9d192 IB |
2295 | int count; |
2296 | ||
2297 | if (!AGGREGATE_TYPE_P (type) | |
2298 | && TREE_CODE (type) != COMPLEX_TYPE | |
2299 | && TREE_CODE (type) != VECTOR_TYPE) | |
2300 | /* Simple scalar types always returned in registers. */ | |
2301 | return false; | |
2302 | ||
2303 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (type), | |
2304 | type, | |
2305 | &ag_mode, | |
2306 | &count, | |
2307 | NULL)) | |
2308 | return false; | |
2309 | ||
2310 | /* Types larger than 2 registers returned in memory. */ | |
2311 | size = int_size_in_bytes (type); | |
2312 | return (size < 0 || size > 2 * UNITS_PER_WORD); | |
2313 | } | |
2314 | ||
2315 | static bool | |
ef4bddc2 | 2316 | aarch64_vfp_is_call_candidate (cumulative_args_t pcum_v, machine_mode mode, |
43e9d192 IB |
2317 | const_tree type, int *nregs) |
2318 | { | |
2319 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
2320 | return aarch64_vfp_is_call_or_return_candidate (mode, | |
2321 | type, | |
2322 | &pcum->aapcs_vfp_rmode, | |
2323 | nregs, | |
2324 | NULL); | |
2325 | } | |
2326 | ||
985b8393 | 2327 | /* Given MODE and TYPE of a function argument, return the alignment in |
43e9d192 IB |
2328 | bits. The idea is to suppress any stronger alignment requested by |
2329 | the user and opt for the natural alignment (specified in AAPCS64 \S 4.1). | |
2330 | This is a helper function for local use only. */ | |
2331 | ||
985b8393 | 2332 | static unsigned int |
ef4bddc2 | 2333 | aarch64_function_arg_alignment (machine_mode mode, const_tree type) |
43e9d192 | 2334 | { |
75d6cc81 | 2335 | if (!type) |
985b8393 | 2336 | return GET_MODE_ALIGNMENT (mode); |
2ec07fa6 | 2337 | |
75d6cc81 | 2338 | if (integer_zerop (TYPE_SIZE (type))) |
985b8393 | 2339 | return 0; |
43e9d192 | 2340 | |
75d6cc81 AL |
2341 | gcc_assert (TYPE_MODE (type) == mode); |
2342 | ||
2343 | if (!AGGREGATE_TYPE_P (type)) | |
985b8393 | 2344 | return TYPE_ALIGN (TYPE_MAIN_VARIANT (type)); |
75d6cc81 AL |
2345 | |
2346 | if (TREE_CODE (type) == ARRAY_TYPE) | |
985b8393 | 2347 | return TYPE_ALIGN (TREE_TYPE (type)); |
75d6cc81 | 2348 | |
985b8393 | 2349 | unsigned int alignment = 0; |
75d6cc81 | 2350 | for (tree field = TYPE_FIELDS (type); field; field = DECL_CHAIN (field)) |
985b8393 JJ |
2351 | if (TREE_CODE (field) == FIELD_DECL) |
2352 | alignment = std::max (alignment, DECL_ALIGN (field)); | |
43e9d192 | 2353 | |
985b8393 | 2354 | return alignment; |
43e9d192 IB |
2355 | } |
2356 | ||
2357 | /* Layout a function argument according to the AAPCS64 rules. The rule | |
2358 | numbers refer to the rule numbers in the AAPCS64. */ | |
2359 | ||
2360 | static void | |
ef4bddc2 | 2361 | aarch64_layout_arg (cumulative_args_t pcum_v, machine_mode mode, |
43e9d192 IB |
2362 | const_tree type, |
2363 | bool named ATTRIBUTE_UNUSED) | |
2364 | { | |
2365 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
2366 | int ncrn, nvrn, nregs; | |
2367 | bool allocate_ncrn, allocate_nvrn; | |
3abf17cf | 2368 | HOST_WIDE_INT size; |
43e9d192 IB |
2369 | |
2370 | /* We need to do this once per argument. */ | |
2371 | if (pcum->aapcs_arg_processed) | |
2372 | return; | |
2373 | ||
2374 | pcum->aapcs_arg_processed = true; | |
2375 | ||
3abf17cf YZ |
2376 | /* Size in bytes, rounded to the nearest multiple of 8 bytes. */ |
2377 | size | |
4f59f9f2 UB |
2378 | = ROUND_UP (type ? int_size_in_bytes (type) : GET_MODE_SIZE (mode), |
2379 | UNITS_PER_WORD); | |
3abf17cf | 2380 | |
43e9d192 IB |
2381 | allocate_ncrn = (type) ? !(FLOAT_TYPE_P (type)) : !FLOAT_MODE_P (mode); |
2382 | allocate_nvrn = aarch64_vfp_is_call_candidate (pcum_v, | |
2383 | mode, | |
2384 | type, | |
2385 | &nregs); | |
2386 | ||
2387 | /* allocate_ncrn may be false-positive, but allocate_nvrn is quite reliable. | |
2388 | The following code thus handles passing by SIMD/FP registers first. */ | |
2389 | ||
2390 | nvrn = pcum->aapcs_nvrn; | |
2391 | ||
2392 | /* C1 - C5 for floating point, homogenous floating point aggregates (HFA) | |
2393 | and homogenous short-vector aggregates (HVA). */ | |
2394 | if (allocate_nvrn) | |
2395 | { | |
261fb553 AL |
2396 | if (!TARGET_FLOAT) |
2397 | aarch64_err_no_fpadvsimd (mode, "argument"); | |
2398 | ||
43e9d192 IB |
2399 | if (nvrn + nregs <= NUM_FP_ARG_REGS) |
2400 | { | |
2401 | pcum->aapcs_nextnvrn = nvrn + nregs; | |
2402 | if (!aarch64_composite_type_p (type, mode)) | |
2403 | { | |
2404 | gcc_assert (nregs == 1); | |
2405 | pcum->aapcs_reg = gen_rtx_REG (mode, V0_REGNUM + nvrn); | |
2406 | } | |
2407 | else | |
2408 | { | |
2409 | rtx par; | |
2410 | int i; | |
2411 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (nregs)); | |
2412 | for (i = 0; i < nregs; i++) | |
2413 | { | |
2414 | rtx tmp = gen_rtx_REG (pcum->aapcs_vfp_rmode, | |
2415 | V0_REGNUM + nvrn + i); | |
2416 | tmp = gen_rtx_EXPR_LIST | |
2417 | (VOIDmode, tmp, | |
2418 | GEN_INT (i * GET_MODE_SIZE (pcum->aapcs_vfp_rmode))); | |
2419 | XVECEXP (par, 0, i) = tmp; | |
2420 | } | |
2421 | pcum->aapcs_reg = par; | |
2422 | } | |
2423 | return; | |
2424 | } | |
2425 | else | |
2426 | { | |
2427 | /* C.3 NSRN is set to 8. */ | |
2428 | pcum->aapcs_nextnvrn = NUM_FP_ARG_REGS; | |
2429 | goto on_stack; | |
2430 | } | |
2431 | } | |
2432 | ||
2433 | ncrn = pcum->aapcs_ncrn; | |
3abf17cf | 2434 | nregs = size / UNITS_PER_WORD; |
43e9d192 IB |
2435 | |
2436 | /* C6 - C9. though the sign and zero extension semantics are | |
2437 | handled elsewhere. This is the case where the argument fits | |
2438 | entirely general registers. */ | |
2439 | if (allocate_ncrn && (ncrn + nregs <= NUM_ARG_REGS)) | |
2440 | { | |
43e9d192 IB |
2441 | |
2442 | gcc_assert (nregs == 0 || nregs == 1 || nregs == 2); | |
2443 | ||
2444 | /* C.8 if the argument has an alignment of 16 then the NGRN is | |
2445 | rounded up to the next even number. */ | |
985b8393 JJ |
2446 | if (nregs == 2 |
2447 | && ncrn % 2 | |
2ec07fa6 | 2448 | /* The == 16 * BITS_PER_UNIT instead of >= 16 * BITS_PER_UNIT |
985b8393 | 2449 | comparison is there because for > 16 * BITS_PER_UNIT |
2ec07fa6 RR |
2450 | alignment nregs should be > 2 and therefore it should be |
2451 | passed by reference rather than value. */ | |
985b8393 JJ |
2452 | && aarch64_function_arg_alignment (mode, type) == 16 * BITS_PER_UNIT) |
2453 | { | |
2454 | ++ncrn; | |
2455 | gcc_assert (ncrn + nregs <= NUM_ARG_REGS); | |
43e9d192 | 2456 | } |
2ec07fa6 | 2457 | |
43e9d192 IB |
2458 | /* NREGS can be 0 when e.g. an empty structure is to be passed. |
2459 | A reg is still generated for it, but the caller should be smart | |
2460 | enough not to use it. */ | |
2461 | if (nregs == 0 || nregs == 1 || GET_MODE_CLASS (mode) == MODE_INT) | |
2ec07fa6 | 2462 | pcum->aapcs_reg = gen_rtx_REG (mode, R0_REGNUM + ncrn); |
43e9d192 IB |
2463 | else |
2464 | { | |
2465 | rtx par; | |
2466 | int i; | |
2467 | ||
2468 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (nregs)); | |
2469 | for (i = 0; i < nregs; i++) | |
2470 | { | |
2471 | rtx tmp = gen_rtx_REG (word_mode, R0_REGNUM + ncrn + i); | |
2472 | tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp, | |
2473 | GEN_INT (i * UNITS_PER_WORD)); | |
2474 | XVECEXP (par, 0, i) = tmp; | |
2475 | } | |
2476 | pcum->aapcs_reg = par; | |
2477 | } | |
2478 | ||
2479 | pcum->aapcs_nextncrn = ncrn + nregs; | |
2480 | return; | |
2481 | } | |
2482 | ||
2483 | /* C.11 */ | |
2484 | pcum->aapcs_nextncrn = NUM_ARG_REGS; | |
2485 | ||
2486 | /* The argument is passed on stack; record the needed number of words for | |
3abf17cf | 2487 | this argument and align the total size if necessary. */ |
43e9d192 | 2488 | on_stack: |
3abf17cf | 2489 | pcum->aapcs_stack_words = size / UNITS_PER_WORD; |
2ec07fa6 | 2490 | |
985b8393 | 2491 | if (aarch64_function_arg_alignment (mode, type) == 16 * BITS_PER_UNIT) |
4f59f9f2 UB |
2492 | pcum->aapcs_stack_size = ROUND_UP (pcum->aapcs_stack_size, |
2493 | 16 / UNITS_PER_WORD); | |
43e9d192 IB |
2494 | return; |
2495 | } | |
2496 | ||
2497 | /* Implement TARGET_FUNCTION_ARG. */ | |
2498 | ||
2499 | static rtx | |
ef4bddc2 | 2500 | aarch64_function_arg (cumulative_args_t pcum_v, machine_mode mode, |
43e9d192 IB |
2501 | const_tree type, bool named) |
2502 | { | |
2503 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
2504 | gcc_assert (pcum->pcs_variant == ARM_PCS_AAPCS64); | |
2505 | ||
2506 | if (mode == VOIDmode) | |
2507 | return NULL_RTX; | |
2508 | ||
2509 | aarch64_layout_arg (pcum_v, mode, type, named); | |
2510 | return pcum->aapcs_reg; | |
2511 | } | |
2512 | ||
2513 | void | |
2514 | aarch64_init_cumulative_args (CUMULATIVE_ARGS *pcum, | |
2515 | const_tree fntype ATTRIBUTE_UNUSED, | |
2516 | rtx libname ATTRIBUTE_UNUSED, | |
2517 | const_tree fndecl ATTRIBUTE_UNUSED, | |
2518 | unsigned n_named ATTRIBUTE_UNUSED) | |
2519 | { | |
2520 | pcum->aapcs_ncrn = 0; | |
2521 | pcum->aapcs_nvrn = 0; | |
2522 | pcum->aapcs_nextncrn = 0; | |
2523 | pcum->aapcs_nextnvrn = 0; | |
2524 | pcum->pcs_variant = ARM_PCS_AAPCS64; | |
2525 | pcum->aapcs_reg = NULL_RTX; | |
2526 | pcum->aapcs_arg_processed = false; | |
2527 | pcum->aapcs_stack_words = 0; | |
2528 | pcum->aapcs_stack_size = 0; | |
2529 | ||
261fb553 AL |
2530 | if (!TARGET_FLOAT |
2531 | && fndecl && TREE_PUBLIC (fndecl) | |
2532 | && fntype && fntype != error_mark_node) | |
2533 | { | |
2534 | const_tree type = TREE_TYPE (fntype); | |
2535 | machine_mode mode ATTRIBUTE_UNUSED; /* To pass pointer as argument. */ | |
2536 | int nregs ATTRIBUTE_UNUSED; /* Likewise. */ | |
2537 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (type), type, | |
2538 | &mode, &nregs, NULL)) | |
2539 | aarch64_err_no_fpadvsimd (TYPE_MODE (type), "return type"); | |
2540 | } | |
43e9d192 IB |
2541 | return; |
2542 | } | |
2543 | ||
2544 | static void | |
2545 | aarch64_function_arg_advance (cumulative_args_t pcum_v, | |
ef4bddc2 | 2546 | machine_mode mode, |
43e9d192 IB |
2547 | const_tree type, |
2548 | bool named) | |
2549 | { | |
2550 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
2551 | if (pcum->pcs_variant == ARM_PCS_AAPCS64) | |
2552 | { | |
2553 | aarch64_layout_arg (pcum_v, mode, type, named); | |
2554 | gcc_assert ((pcum->aapcs_reg != NULL_RTX) | |
2555 | != (pcum->aapcs_stack_words != 0)); | |
2556 | pcum->aapcs_arg_processed = false; | |
2557 | pcum->aapcs_ncrn = pcum->aapcs_nextncrn; | |
2558 | pcum->aapcs_nvrn = pcum->aapcs_nextnvrn; | |
2559 | pcum->aapcs_stack_size += pcum->aapcs_stack_words; | |
2560 | pcum->aapcs_stack_words = 0; | |
2561 | pcum->aapcs_reg = NULL_RTX; | |
2562 | } | |
2563 | } | |
2564 | ||
2565 | bool | |
2566 | aarch64_function_arg_regno_p (unsigned regno) | |
2567 | { | |
2568 | return ((GP_REGNUM_P (regno) && regno < R0_REGNUM + NUM_ARG_REGS) | |
2569 | || (FP_REGNUM_P (regno) && regno < V0_REGNUM + NUM_FP_ARG_REGS)); | |
2570 | } | |
2571 | ||
2572 | /* Implement FUNCTION_ARG_BOUNDARY. Every parameter gets at least | |
2573 | PARM_BOUNDARY bits of alignment, but will be given anything up | |
2574 | to STACK_BOUNDARY bits if the type requires it. This makes sure | |
2575 | that both before and after the layout of each argument, the Next | |
2576 | Stacked Argument Address (NSAA) will have a minimum alignment of | |
2577 | 8 bytes. */ | |
2578 | ||
2579 | static unsigned int | |
ef4bddc2 | 2580 | aarch64_function_arg_boundary (machine_mode mode, const_tree type) |
43e9d192 | 2581 | { |
985b8393 JJ |
2582 | unsigned int alignment = aarch64_function_arg_alignment (mode, type); |
2583 | return MIN (MAX (alignment, PARM_BOUNDARY), STACK_BOUNDARY); | |
43e9d192 IB |
2584 | } |
2585 | ||
2586 | /* For use by FUNCTION_ARG_PADDING (MODE, TYPE). | |
2587 | ||
2588 | Return true if an argument passed on the stack should be padded upwards, | |
2589 | i.e. if the least-significant byte of the stack slot has useful data. | |
2590 | ||
2591 | Small aggregate types are placed in the lowest memory address. | |
2592 | ||
2593 | The related parameter passing rules are B.4, C.3, C.5 and C.14. */ | |
2594 | ||
2595 | bool | |
ef4bddc2 | 2596 | aarch64_pad_arg_upward (machine_mode mode, const_tree type) |
43e9d192 IB |
2597 | { |
2598 | /* On little-endian targets, the least significant byte of every stack | |
2599 | argument is passed at the lowest byte address of the stack slot. */ | |
2600 | if (!BYTES_BIG_ENDIAN) | |
2601 | return true; | |
2602 | ||
00edcfbe | 2603 | /* Otherwise, integral, floating-point and pointer types are padded downward: |
43e9d192 IB |
2604 | the least significant byte of a stack argument is passed at the highest |
2605 | byte address of the stack slot. */ | |
2606 | if (type | |
00edcfbe YZ |
2607 | ? (INTEGRAL_TYPE_P (type) || SCALAR_FLOAT_TYPE_P (type) |
2608 | || POINTER_TYPE_P (type)) | |
43e9d192 IB |
2609 | : (SCALAR_INT_MODE_P (mode) || SCALAR_FLOAT_MODE_P (mode))) |
2610 | return false; | |
2611 | ||
2612 | /* Everything else padded upward, i.e. data in first byte of stack slot. */ | |
2613 | return true; | |
2614 | } | |
2615 | ||
2616 | /* Similarly, for use by BLOCK_REG_PADDING (MODE, TYPE, FIRST). | |
2617 | ||
2618 | It specifies padding for the last (may also be the only) | |
2619 | element of a block move between registers and memory. If | |
2620 | assuming the block is in the memory, padding upward means that | |
2621 | the last element is padded after its highest significant byte, | |
2622 | while in downward padding, the last element is padded at the | |
2623 | its least significant byte side. | |
2624 | ||
2625 | Small aggregates and small complex types are always padded | |
2626 | upwards. | |
2627 | ||
2628 | We don't need to worry about homogeneous floating-point or | |
2629 | short-vector aggregates; their move is not affected by the | |
2630 | padding direction determined here. Regardless of endianness, | |
2631 | each element of such an aggregate is put in the least | |
2632 | significant bits of a fp/simd register. | |
2633 | ||
2634 | Return !BYTES_BIG_ENDIAN if the least significant byte of the | |
2635 | register has useful data, and return the opposite if the most | |
2636 | significant byte does. */ | |
2637 | ||
2638 | bool | |
ef4bddc2 | 2639 | aarch64_pad_reg_upward (machine_mode mode, const_tree type, |
43e9d192 IB |
2640 | bool first ATTRIBUTE_UNUSED) |
2641 | { | |
2642 | ||
2643 | /* Small composite types are always padded upward. */ | |
2644 | if (BYTES_BIG_ENDIAN && aarch64_composite_type_p (type, mode)) | |
2645 | { | |
2646 | HOST_WIDE_INT size = (type ? int_size_in_bytes (type) | |
2647 | : GET_MODE_SIZE (mode)); | |
2648 | if (size < 2 * UNITS_PER_WORD) | |
2649 | return true; | |
2650 | } | |
2651 | ||
2652 | /* Otherwise, use the default padding. */ | |
2653 | return !BYTES_BIG_ENDIAN; | |
2654 | } | |
2655 | ||
095a2d76 | 2656 | static scalar_int_mode |
43e9d192 IB |
2657 | aarch64_libgcc_cmp_return_mode (void) |
2658 | { | |
2659 | return SImode; | |
2660 | } | |
2661 | ||
a3eb8a52 EB |
2662 | #define PROBE_INTERVAL (1 << STACK_CHECK_PROBE_INTERVAL_EXP) |
2663 | ||
2664 | /* We use the 12-bit shifted immediate arithmetic instructions so values | |
2665 | must be multiple of (1 << 12), i.e. 4096. */ | |
2666 | #define ARITH_FACTOR 4096 | |
2667 | ||
2668 | #if (PROBE_INTERVAL % ARITH_FACTOR) != 0 | |
2669 | #error Cannot use simple address calculation for stack probing | |
2670 | #endif | |
2671 | ||
2672 | /* The pair of scratch registers used for stack probing. */ | |
2673 | #define PROBE_STACK_FIRST_REG 9 | |
2674 | #define PROBE_STACK_SECOND_REG 10 | |
2675 | ||
2676 | /* Emit code to probe a range of stack addresses from FIRST to FIRST+SIZE, | |
2677 | inclusive. These are offsets from the current stack pointer. */ | |
2678 | ||
2679 | static void | |
2680 | aarch64_emit_probe_stack_range (HOST_WIDE_INT first, HOST_WIDE_INT size) | |
2681 | { | |
5f5c5e0f | 2682 | rtx reg1 = gen_rtx_REG (Pmode, PROBE_STACK_FIRST_REG); |
a3eb8a52 EB |
2683 | |
2684 | /* See the same assertion on PROBE_INTERVAL above. */ | |
2685 | gcc_assert ((first % ARITH_FACTOR) == 0); | |
2686 | ||
2687 | /* See if we have a constant small number of probes to generate. If so, | |
2688 | that's the easy case. */ | |
2689 | if (size <= PROBE_INTERVAL) | |
2690 | { | |
2691 | const HOST_WIDE_INT base = ROUND_UP (size, ARITH_FACTOR); | |
2692 | ||
2693 | emit_set_insn (reg1, | |
5f5c5e0f | 2694 | plus_constant (Pmode, |
a3eb8a52 | 2695 | stack_pointer_rtx, -(first + base))); |
5f5c5e0f | 2696 | emit_stack_probe (plus_constant (Pmode, reg1, base - size)); |
a3eb8a52 EB |
2697 | } |
2698 | ||
2699 | /* The run-time loop is made up of 8 insns in the generic case while the | |
2700 | compile-time loop is made up of 4+2*(n-2) insns for n # of intervals. */ | |
2701 | else if (size <= 4 * PROBE_INTERVAL) | |
2702 | { | |
2703 | HOST_WIDE_INT i, rem; | |
2704 | ||
2705 | emit_set_insn (reg1, | |
5f5c5e0f | 2706 | plus_constant (Pmode, |
a3eb8a52 EB |
2707 | stack_pointer_rtx, |
2708 | -(first + PROBE_INTERVAL))); | |
2709 | emit_stack_probe (reg1); | |
2710 | ||
2711 | /* Probe at FIRST + N * PROBE_INTERVAL for values of N from 2 until | |
2712 | it exceeds SIZE. If only two probes are needed, this will not | |
2713 | generate any code. Then probe at FIRST + SIZE. */ | |
2714 | for (i = 2 * PROBE_INTERVAL; i < size; i += PROBE_INTERVAL) | |
2715 | { | |
2716 | emit_set_insn (reg1, | |
5f5c5e0f | 2717 | plus_constant (Pmode, reg1, -PROBE_INTERVAL)); |
a3eb8a52 EB |
2718 | emit_stack_probe (reg1); |
2719 | } | |
2720 | ||
2721 | rem = size - (i - PROBE_INTERVAL); | |
2722 | if (rem > 256) | |
2723 | { | |
2724 | const HOST_WIDE_INT base = ROUND_UP (rem, ARITH_FACTOR); | |
2725 | ||
5f5c5e0f EB |
2726 | emit_set_insn (reg1, plus_constant (Pmode, reg1, -base)); |
2727 | emit_stack_probe (plus_constant (Pmode, reg1, base - rem)); | |
a3eb8a52 EB |
2728 | } |
2729 | else | |
5f5c5e0f | 2730 | emit_stack_probe (plus_constant (Pmode, reg1, -rem)); |
a3eb8a52 EB |
2731 | } |
2732 | ||
2733 | /* Otherwise, do the same as above, but in a loop. Note that we must be | |
2734 | extra careful with variables wrapping around because we might be at | |
2735 | the very top (or the very bottom) of the address space and we have | |
2736 | to be able to handle this case properly; in particular, we use an | |
2737 | equality test for the loop condition. */ | |
2738 | else | |
2739 | { | |
5f5c5e0f | 2740 | rtx reg2 = gen_rtx_REG (Pmode, PROBE_STACK_SECOND_REG); |
a3eb8a52 EB |
2741 | |
2742 | /* Step 1: round SIZE to the previous multiple of the interval. */ | |
2743 | ||
2744 | HOST_WIDE_INT rounded_size = size & -PROBE_INTERVAL; | |
2745 | ||
2746 | ||
2747 | /* Step 2: compute initial and final value of the loop counter. */ | |
2748 | ||
2749 | /* TEST_ADDR = SP + FIRST. */ | |
2750 | emit_set_insn (reg1, | |
5f5c5e0f | 2751 | plus_constant (Pmode, stack_pointer_rtx, -first)); |
a3eb8a52 EB |
2752 | |
2753 | /* LAST_ADDR = SP + FIRST + ROUNDED_SIZE. */ | |
13f752b2 JL |
2754 | HOST_WIDE_INT adjustment = - (first + rounded_size); |
2755 | if (! aarch64_uimm12_shift (adjustment)) | |
2756 | { | |
2757 | aarch64_internal_mov_immediate (reg2, GEN_INT (adjustment), | |
2758 | true, Pmode); | |
2759 | emit_set_insn (reg2, gen_rtx_PLUS (Pmode, stack_pointer_rtx, reg2)); | |
2760 | } | |
2761 | else | |
2762 | { | |
2763 | emit_set_insn (reg2, | |
2764 | plus_constant (Pmode, stack_pointer_rtx, adjustment)); | |
2765 | } | |
2766 | ||
a3eb8a52 EB |
2767 | /* Step 3: the loop |
2768 | ||
2769 | do | |
2770 | { | |
2771 | TEST_ADDR = TEST_ADDR + PROBE_INTERVAL | |
2772 | probe at TEST_ADDR | |
2773 | } | |
2774 | while (TEST_ADDR != LAST_ADDR) | |
2775 | ||
2776 | probes at FIRST + N * PROBE_INTERVAL for values of N from 1 | |
2777 | until it is equal to ROUNDED_SIZE. */ | |
2778 | ||
5f5c5e0f | 2779 | emit_insn (gen_probe_stack_range (reg1, reg1, reg2)); |
a3eb8a52 EB |
2780 | |
2781 | ||
2782 | /* Step 4: probe at FIRST + SIZE if we cannot assert at compile-time | |
2783 | that SIZE is equal to ROUNDED_SIZE. */ | |
2784 | ||
2785 | if (size != rounded_size) | |
2786 | { | |
2787 | HOST_WIDE_INT rem = size - rounded_size; | |
2788 | ||
2789 | if (rem > 256) | |
2790 | { | |
2791 | const HOST_WIDE_INT base = ROUND_UP (rem, ARITH_FACTOR); | |
2792 | ||
5f5c5e0f EB |
2793 | emit_set_insn (reg2, plus_constant (Pmode, reg2, -base)); |
2794 | emit_stack_probe (plus_constant (Pmode, reg2, base - rem)); | |
a3eb8a52 EB |
2795 | } |
2796 | else | |
5f5c5e0f | 2797 | emit_stack_probe (plus_constant (Pmode, reg2, -rem)); |
a3eb8a52 EB |
2798 | } |
2799 | } | |
2800 | ||
2801 | /* Make sure nothing is scheduled before we are done. */ | |
2802 | emit_insn (gen_blockage ()); | |
2803 | } | |
2804 | ||
2805 | /* Probe a range of stack addresses from REG1 to REG2 inclusive. These are | |
2806 | absolute addresses. */ | |
2807 | ||
2808 | const char * | |
2809 | aarch64_output_probe_stack_range (rtx reg1, rtx reg2) | |
2810 | { | |
2811 | static int labelno = 0; | |
2812 | char loop_lab[32]; | |
2813 | rtx xops[2]; | |
2814 | ||
2815 | ASM_GENERATE_INTERNAL_LABEL (loop_lab, "LPSRL", labelno++); | |
2816 | ||
2817 | /* Loop. */ | |
2818 | ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, loop_lab); | |
2819 | ||
2820 | /* TEST_ADDR = TEST_ADDR + PROBE_INTERVAL. */ | |
2821 | xops[0] = reg1; | |
2822 | xops[1] = GEN_INT (PROBE_INTERVAL); | |
2823 | output_asm_insn ("sub\t%0, %0, %1", xops); | |
2824 | ||
2825 | /* Probe at TEST_ADDR. */ | |
2826 | output_asm_insn ("str\txzr, [%0]", xops); | |
2827 | ||
2828 | /* Test if TEST_ADDR == LAST_ADDR. */ | |
2829 | xops[1] = reg2; | |
2830 | output_asm_insn ("cmp\t%0, %1", xops); | |
2831 | ||
2832 | /* Branch. */ | |
2833 | fputs ("\tb.ne\t", asm_out_file); | |
2834 | assemble_name_raw (asm_out_file, loop_lab); | |
2835 | fputc ('\n', asm_out_file); | |
2836 | ||
2837 | return ""; | |
2838 | } | |
2839 | ||
43e9d192 IB |
2840 | static bool |
2841 | aarch64_frame_pointer_required (void) | |
2842 | { | |
0b7f8166 MS |
2843 | /* In aarch64_override_options_after_change |
2844 | flag_omit_leaf_frame_pointer turns off the frame pointer by | |
2845 | default. Turn it back on now if we've not got a leaf | |
2846 | function. */ | |
2847 | if (flag_omit_leaf_frame_pointer | |
2848 | && (!crtl->is_leaf || df_regs_ever_live_p (LR_REGNUM))) | |
2849 | return true; | |
43e9d192 | 2850 | |
8144a493 WD |
2851 | /* Force a frame pointer for EH returns so the return address is at FP+8. */ |
2852 | if (crtl->calls_eh_return) | |
2853 | return true; | |
2854 | ||
0b7f8166 | 2855 | return false; |
43e9d192 IB |
2856 | } |
2857 | ||
2858 | /* Mark the registers that need to be saved by the callee and calculate | |
2859 | the size of the callee-saved registers area and frame record (both FP | |
2860 | and LR may be omitted). */ | |
2861 | static void | |
2862 | aarch64_layout_frame (void) | |
2863 | { | |
2864 | HOST_WIDE_INT offset = 0; | |
4b0685d9 | 2865 | int regno, last_fp_reg = INVALID_REGNUM; |
43e9d192 IB |
2866 | |
2867 | if (reload_completed && cfun->machine->frame.laid_out) | |
2868 | return; | |
2869 | ||
97826595 MS |
2870 | #define SLOT_NOT_REQUIRED (-2) |
2871 | #define SLOT_REQUIRED (-1) | |
2872 | ||
71bfb77a WD |
2873 | cfun->machine->frame.wb_candidate1 = INVALID_REGNUM; |
2874 | cfun->machine->frame.wb_candidate2 = INVALID_REGNUM; | |
363ffa50 | 2875 | |
43e9d192 IB |
2876 | /* First mark all the registers that really need to be saved... */ |
2877 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) | |
97826595 | 2878 | cfun->machine->frame.reg_offset[regno] = SLOT_NOT_REQUIRED; |
43e9d192 IB |
2879 | |
2880 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
97826595 | 2881 | cfun->machine->frame.reg_offset[regno] = SLOT_NOT_REQUIRED; |
43e9d192 IB |
2882 | |
2883 | /* ... that includes the eh data registers (if needed)... */ | |
2884 | if (crtl->calls_eh_return) | |
2885 | for (regno = 0; EH_RETURN_DATA_REGNO (regno) != INVALID_REGNUM; regno++) | |
97826595 MS |
2886 | cfun->machine->frame.reg_offset[EH_RETURN_DATA_REGNO (regno)] |
2887 | = SLOT_REQUIRED; | |
43e9d192 IB |
2888 | |
2889 | /* ... and any callee saved register that dataflow says is live. */ | |
2890 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) | |
2891 | if (df_regs_ever_live_p (regno) | |
1c923b60 JW |
2892 | && (regno == R30_REGNUM |
2893 | || !call_used_regs[regno])) | |
97826595 | 2894 | cfun->machine->frame.reg_offset[regno] = SLOT_REQUIRED; |
43e9d192 IB |
2895 | |
2896 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
2897 | if (df_regs_ever_live_p (regno) | |
2898 | && !call_used_regs[regno]) | |
4b0685d9 WD |
2899 | { |
2900 | cfun->machine->frame.reg_offset[regno] = SLOT_REQUIRED; | |
2901 | last_fp_reg = regno; | |
2902 | } | |
43e9d192 IB |
2903 | |
2904 | if (frame_pointer_needed) | |
2905 | { | |
2e1cdae5 | 2906 | /* FP and LR are placed in the linkage record. */ |
43e9d192 | 2907 | cfun->machine->frame.reg_offset[R29_REGNUM] = 0; |
363ffa50 | 2908 | cfun->machine->frame.wb_candidate1 = R29_REGNUM; |
2e1cdae5 | 2909 | cfun->machine->frame.reg_offset[R30_REGNUM] = UNITS_PER_WORD; |
363ffa50 | 2910 | cfun->machine->frame.wb_candidate2 = R30_REGNUM; |
2e1cdae5 | 2911 | offset += 2 * UNITS_PER_WORD; |
43e9d192 IB |
2912 | } |
2913 | ||
2914 | /* Now assign stack slots for them. */ | |
2e1cdae5 | 2915 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) |
97826595 | 2916 | if (cfun->machine->frame.reg_offset[regno] == SLOT_REQUIRED) |
43e9d192 IB |
2917 | { |
2918 | cfun->machine->frame.reg_offset[regno] = offset; | |
71bfb77a | 2919 | if (cfun->machine->frame.wb_candidate1 == INVALID_REGNUM) |
363ffa50 | 2920 | cfun->machine->frame.wb_candidate1 = regno; |
71bfb77a | 2921 | else if (cfun->machine->frame.wb_candidate2 == INVALID_REGNUM) |
363ffa50 | 2922 | cfun->machine->frame.wb_candidate2 = regno; |
43e9d192 IB |
2923 | offset += UNITS_PER_WORD; |
2924 | } | |
2925 | ||
4b0685d9 WD |
2926 | HOST_WIDE_INT max_int_offset = offset; |
2927 | offset = ROUND_UP (offset, STACK_BOUNDARY / BITS_PER_UNIT); | |
2928 | bool has_align_gap = offset != max_int_offset; | |
2929 | ||
43e9d192 | 2930 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) |
97826595 | 2931 | if (cfun->machine->frame.reg_offset[regno] == SLOT_REQUIRED) |
43e9d192 | 2932 | { |
4b0685d9 WD |
2933 | /* If there is an alignment gap between integer and fp callee-saves, |
2934 | allocate the last fp register to it if possible. */ | |
2935 | if (regno == last_fp_reg && has_align_gap && (offset & 8) == 0) | |
2936 | { | |
2937 | cfun->machine->frame.reg_offset[regno] = max_int_offset; | |
2938 | break; | |
2939 | } | |
2940 | ||
43e9d192 | 2941 | cfun->machine->frame.reg_offset[regno] = offset; |
71bfb77a | 2942 | if (cfun->machine->frame.wb_candidate1 == INVALID_REGNUM) |
363ffa50 | 2943 | cfun->machine->frame.wb_candidate1 = regno; |
71bfb77a | 2944 | else if (cfun->machine->frame.wb_candidate2 == INVALID_REGNUM |
363ffa50 JW |
2945 | && cfun->machine->frame.wb_candidate1 >= V0_REGNUM) |
2946 | cfun->machine->frame.wb_candidate2 = regno; | |
43e9d192 IB |
2947 | offset += UNITS_PER_WORD; |
2948 | } | |
2949 | ||
4f59f9f2 | 2950 | offset = ROUND_UP (offset, STACK_BOUNDARY / BITS_PER_UNIT); |
43e9d192 IB |
2951 | |
2952 | cfun->machine->frame.saved_regs_size = offset; | |
1c960e02 | 2953 | |
71bfb77a WD |
2954 | HOST_WIDE_INT varargs_and_saved_regs_size |
2955 | = offset + cfun->machine->frame.saved_varargs_size; | |
2956 | ||
1c960e02 | 2957 | cfun->machine->frame.hard_fp_offset |
71bfb77a | 2958 | = ROUND_UP (varargs_and_saved_regs_size + get_frame_size (), |
4f59f9f2 | 2959 | STACK_BOUNDARY / BITS_PER_UNIT); |
1c960e02 MS |
2960 | |
2961 | cfun->machine->frame.frame_size | |
4f59f9f2 UB |
2962 | = ROUND_UP (cfun->machine->frame.hard_fp_offset |
2963 | + crtl->outgoing_args_size, | |
2964 | STACK_BOUNDARY / BITS_PER_UNIT); | |
1c960e02 | 2965 | |
71bfb77a WD |
2966 | cfun->machine->frame.locals_offset = cfun->machine->frame.saved_varargs_size; |
2967 | ||
2968 | cfun->machine->frame.initial_adjust = 0; | |
2969 | cfun->machine->frame.final_adjust = 0; | |
2970 | cfun->machine->frame.callee_adjust = 0; | |
2971 | cfun->machine->frame.callee_offset = 0; | |
2972 | ||
2973 | HOST_WIDE_INT max_push_offset = 0; | |
2974 | if (cfun->machine->frame.wb_candidate2 != INVALID_REGNUM) | |
2975 | max_push_offset = 512; | |
2976 | else if (cfun->machine->frame.wb_candidate1 != INVALID_REGNUM) | |
2977 | max_push_offset = 256; | |
2978 | ||
2979 | if (cfun->machine->frame.frame_size < max_push_offset | |
2980 | && crtl->outgoing_args_size == 0) | |
2981 | { | |
2982 | /* Simple, small frame with no outgoing arguments: | |
2983 | stp reg1, reg2, [sp, -frame_size]! | |
2984 | stp reg3, reg4, [sp, 16] */ | |
2985 | cfun->machine->frame.callee_adjust = cfun->machine->frame.frame_size; | |
2986 | } | |
2987 | else if ((crtl->outgoing_args_size | |
2988 | + cfun->machine->frame.saved_regs_size < 512) | |
2989 | && !(cfun->calls_alloca | |
2990 | && cfun->machine->frame.hard_fp_offset < max_push_offset)) | |
2991 | { | |
2992 | /* Frame with small outgoing arguments: | |
2993 | sub sp, sp, frame_size | |
2994 | stp reg1, reg2, [sp, outgoing_args_size] | |
2995 | stp reg3, reg4, [sp, outgoing_args_size + 16] */ | |
2996 | cfun->machine->frame.initial_adjust = cfun->machine->frame.frame_size; | |
2997 | cfun->machine->frame.callee_offset | |
2998 | = cfun->machine->frame.frame_size - cfun->machine->frame.hard_fp_offset; | |
2999 | } | |
3000 | else if (cfun->machine->frame.hard_fp_offset < max_push_offset) | |
3001 | { | |
3002 | /* Frame with large outgoing arguments but a small local area: | |
3003 | stp reg1, reg2, [sp, -hard_fp_offset]! | |
3004 | stp reg3, reg4, [sp, 16] | |
3005 | sub sp, sp, outgoing_args_size */ | |
3006 | cfun->machine->frame.callee_adjust = cfun->machine->frame.hard_fp_offset; | |
3007 | cfun->machine->frame.final_adjust | |
3008 | = cfun->machine->frame.frame_size - cfun->machine->frame.callee_adjust; | |
3009 | } | |
3010 | else if (!frame_pointer_needed | |
3011 | && varargs_and_saved_regs_size < max_push_offset) | |
3012 | { | |
3013 | /* Frame with large local area and outgoing arguments (this pushes the | |
3014 | callee-saves first, followed by the locals and outgoing area): | |
3015 | stp reg1, reg2, [sp, -varargs_and_saved_regs_size]! | |
3016 | stp reg3, reg4, [sp, 16] | |
3017 | sub sp, sp, frame_size - varargs_and_saved_regs_size */ | |
3018 | cfun->machine->frame.callee_adjust = varargs_and_saved_regs_size; | |
3019 | cfun->machine->frame.final_adjust | |
3020 | = cfun->machine->frame.frame_size - cfun->machine->frame.callee_adjust; | |
3021 | cfun->machine->frame.hard_fp_offset = cfun->machine->frame.callee_adjust; | |
3022 | cfun->machine->frame.locals_offset = cfun->machine->frame.hard_fp_offset; | |
3023 | } | |
3024 | else | |
3025 | { | |
3026 | /* Frame with large local area and outgoing arguments using frame pointer: | |
3027 | sub sp, sp, hard_fp_offset | |
3028 | stp x29, x30, [sp, 0] | |
3029 | add x29, sp, 0 | |
3030 | stp reg3, reg4, [sp, 16] | |
3031 | sub sp, sp, outgoing_args_size */ | |
3032 | cfun->machine->frame.initial_adjust = cfun->machine->frame.hard_fp_offset; | |
3033 | cfun->machine->frame.final_adjust | |
3034 | = cfun->machine->frame.frame_size - cfun->machine->frame.initial_adjust; | |
3035 | } | |
3036 | ||
43e9d192 IB |
3037 | cfun->machine->frame.laid_out = true; |
3038 | } | |
3039 | ||
04ddfe06 KT |
3040 | /* Return true if the register REGNO is saved on entry to |
3041 | the current function. */ | |
3042 | ||
43e9d192 IB |
3043 | static bool |
3044 | aarch64_register_saved_on_entry (int regno) | |
3045 | { | |
97826595 | 3046 | return cfun->machine->frame.reg_offset[regno] >= 0; |
43e9d192 IB |
3047 | } |
3048 | ||
04ddfe06 KT |
3049 | /* Return the next register up from REGNO up to LIMIT for the callee |
3050 | to save. */ | |
3051 | ||
64dedd72 JW |
3052 | static unsigned |
3053 | aarch64_next_callee_save (unsigned regno, unsigned limit) | |
3054 | { | |
3055 | while (regno <= limit && !aarch64_register_saved_on_entry (regno)) | |
3056 | regno ++; | |
3057 | return regno; | |
3058 | } | |
43e9d192 | 3059 | |
04ddfe06 KT |
3060 | /* Push the register number REGNO of mode MODE to the stack with write-back |
3061 | adjusting the stack by ADJUSTMENT. */ | |
3062 | ||
c5e1f66e | 3063 | static void |
ef4bddc2 | 3064 | aarch64_pushwb_single_reg (machine_mode mode, unsigned regno, |
c5e1f66e JW |
3065 | HOST_WIDE_INT adjustment) |
3066 | { | |
3067 | rtx base_rtx = stack_pointer_rtx; | |
3068 | rtx insn, reg, mem; | |
3069 | ||
3070 | reg = gen_rtx_REG (mode, regno); | |
3071 | mem = gen_rtx_PRE_MODIFY (Pmode, base_rtx, | |
3072 | plus_constant (Pmode, base_rtx, -adjustment)); | |
30079dde | 3073 | mem = gen_frame_mem (mode, mem); |
c5e1f66e JW |
3074 | |
3075 | insn = emit_move_insn (mem, reg); | |
3076 | RTX_FRAME_RELATED_P (insn) = 1; | |
3077 | } | |
3078 | ||
04ddfe06 KT |
3079 | /* Generate and return an instruction to store the pair of registers |
3080 | REG and REG2 of mode MODE to location BASE with write-back adjusting | |
3081 | the stack location BASE by ADJUSTMENT. */ | |
3082 | ||
80c11907 | 3083 | static rtx |
ef4bddc2 | 3084 | aarch64_gen_storewb_pair (machine_mode mode, rtx base, rtx reg, rtx reg2, |
80c11907 JW |
3085 | HOST_WIDE_INT adjustment) |
3086 | { | |
3087 | switch (mode) | |
3088 | { | |
4e10a5a7 | 3089 | case E_DImode: |
80c11907 JW |
3090 | return gen_storewb_pairdi_di (base, base, reg, reg2, |
3091 | GEN_INT (-adjustment), | |
3092 | GEN_INT (UNITS_PER_WORD - adjustment)); | |
4e10a5a7 | 3093 | case E_DFmode: |
80c11907 JW |
3094 | return gen_storewb_pairdf_di (base, base, reg, reg2, |
3095 | GEN_INT (-adjustment), | |
3096 | GEN_INT (UNITS_PER_WORD - adjustment)); | |
3097 | default: | |
3098 | gcc_unreachable (); | |
3099 | } | |
3100 | } | |
3101 | ||
04ddfe06 KT |
3102 | /* Push registers numbered REGNO1 and REGNO2 to the stack, adjusting the |
3103 | stack pointer by ADJUSTMENT. */ | |
3104 | ||
80c11907 | 3105 | static void |
89ac681e | 3106 | aarch64_push_regs (unsigned regno1, unsigned regno2, HOST_WIDE_INT adjustment) |
80c11907 | 3107 | { |
5d8a22a5 | 3108 | rtx_insn *insn; |
0d4a1197 | 3109 | machine_mode mode = (regno1 <= R30_REGNUM) ? E_DImode : E_DFmode; |
89ac681e | 3110 | |
71bfb77a | 3111 | if (regno2 == INVALID_REGNUM) |
89ac681e WD |
3112 | return aarch64_pushwb_single_reg (mode, regno1, adjustment); |
3113 | ||
80c11907 JW |
3114 | rtx reg1 = gen_rtx_REG (mode, regno1); |
3115 | rtx reg2 = gen_rtx_REG (mode, regno2); | |
3116 | ||
3117 | insn = emit_insn (aarch64_gen_storewb_pair (mode, stack_pointer_rtx, reg1, | |
3118 | reg2, adjustment)); | |
3119 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 2)) = 1; | |
80c11907 JW |
3120 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; |
3121 | RTX_FRAME_RELATED_P (insn) = 1; | |
3122 | } | |
3123 | ||
04ddfe06 KT |
3124 | /* Load the pair of register REG, REG2 of mode MODE from stack location BASE, |
3125 | adjusting it by ADJUSTMENT afterwards. */ | |
3126 | ||
159313d9 | 3127 | static rtx |
ef4bddc2 | 3128 | aarch64_gen_loadwb_pair (machine_mode mode, rtx base, rtx reg, rtx reg2, |
159313d9 JW |
3129 | HOST_WIDE_INT adjustment) |
3130 | { | |
3131 | switch (mode) | |
3132 | { | |
4e10a5a7 | 3133 | case E_DImode: |
159313d9 | 3134 | return gen_loadwb_pairdi_di (base, base, reg, reg2, GEN_INT (adjustment), |
3e322b3f | 3135 | GEN_INT (UNITS_PER_WORD)); |
4e10a5a7 | 3136 | case E_DFmode: |
159313d9 | 3137 | return gen_loadwb_pairdf_di (base, base, reg, reg2, GEN_INT (adjustment), |
3e322b3f | 3138 | GEN_INT (UNITS_PER_WORD)); |
159313d9 JW |
3139 | default: |
3140 | gcc_unreachable (); | |
3141 | } | |
3142 | } | |
3143 | ||
04ddfe06 KT |
3144 | /* Pop the two registers numbered REGNO1, REGNO2 from the stack, adjusting it |
3145 | afterwards by ADJUSTMENT and writing the appropriate REG_CFA_RESTORE notes | |
3146 | into CFI_OPS. */ | |
3147 | ||
89ac681e WD |
3148 | static void |
3149 | aarch64_pop_regs (unsigned regno1, unsigned regno2, HOST_WIDE_INT adjustment, | |
3150 | rtx *cfi_ops) | |
3151 | { | |
0d4a1197 | 3152 | machine_mode mode = (regno1 <= R30_REGNUM) ? E_DImode : E_DFmode; |
89ac681e WD |
3153 | rtx reg1 = gen_rtx_REG (mode, regno1); |
3154 | ||
3155 | *cfi_ops = alloc_reg_note (REG_CFA_RESTORE, reg1, *cfi_ops); | |
3156 | ||
71bfb77a | 3157 | if (regno2 == INVALID_REGNUM) |
89ac681e WD |
3158 | { |
3159 | rtx mem = plus_constant (Pmode, stack_pointer_rtx, adjustment); | |
3160 | mem = gen_rtx_POST_MODIFY (Pmode, stack_pointer_rtx, mem); | |
30079dde | 3161 | emit_move_insn (reg1, gen_frame_mem (mode, mem)); |
89ac681e WD |
3162 | } |
3163 | else | |
3164 | { | |
3165 | rtx reg2 = gen_rtx_REG (mode, regno2); | |
3166 | *cfi_ops = alloc_reg_note (REG_CFA_RESTORE, reg2, *cfi_ops); | |
3167 | emit_insn (aarch64_gen_loadwb_pair (mode, stack_pointer_rtx, reg1, | |
3168 | reg2, adjustment)); | |
3169 | } | |
3170 | } | |
3171 | ||
04ddfe06 KT |
3172 | /* Generate and return a store pair instruction of mode MODE to store |
3173 | register REG1 to MEM1 and register REG2 to MEM2. */ | |
3174 | ||
72df5c1f | 3175 | static rtx |
ef4bddc2 | 3176 | aarch64_gen_store_pair (machine_mode mode, rtx mem1, rtx reg1, rtx mem2, |
72df5c1f JW |
3177 | rtx reg2) |
3178 | { | |
3179 | switch (mode) | |
3180 | { | |
4e10a5a7 | 3181 | case E_DImode: |
72df5c1f JW |
3182 | return gen_store_pairdi (mem1, reg1, mem2, reg2); |
3183 | ||
4e10a5a7 | 3184 | case E_DFmode: |
72df5c1f JW |
3185 | return gen_store_pairdf (mem1, reg1, mem2, reg2); |
3186 | ||
3187 | default: | |
3188 | gcc_unreachable (); | |
3189 | } | |
3190 | } | |
3191 | ||
04ddfe06 KT |
3192 | /* Generate and regurn a load pair isntruction of mode MODE to load register |
3193 | REG1 from MEM1 and register REG2 from MEM2. */ | |
3194 | ||
72df5c1f | 3195 | static rtx |
ef4bddc2 | 3196 | aarch64_gen_load_pair (machine_mode mode, rtx reg1, rtx mem1, rtx reg2, |
72df5c1f JW |
3197 | rtx mem2) |
3198 | { | |
3199 | switch (mode) | |
3200 | { | |
4e10a5a7 | 3201 | case E_DImode: |
72df5c1f JW |
3202 | return gen_load_pairdi (reg1, mem1, reg2, mem2); |
3203 | ||
4e10a5a7 | 3204 | case E_DFmode: |
72df5c1f JW |
3205 | return gen_load_pairdf (reg1, mem1, reg2, mem2); |
3206 | ||
3207 | default: | |
3208 | gcc_unreachable (); | |
3209 | } | |
3210 | } | |
3211 | ||
db58fd89 JW |
3212 | /* Return TRUE if return address signing should be enabled for the current |
3213 | function, otherwise return FALSE. */ | |
3214 | ||
3215 | bool | |
3216 | aarch64_return_address_signing_enabled (void) | |
3217 | { | |
3218 | /* This function should only be called after frame laid out. */ | |
3219 | gcc_assert (cfun->machine->frame.laid_out); | |
3220 | ||
3221 | /* If signing scope is AARCH64_FUNCTION_NON_LEAF, we only sign a leaf function | |
3222 | if it's LR is pushed onto stack. */ | |
3223 | return (aarch64_ra_sign_scope == AARCH64_FUNCTION_ALL | |
3224 | || (aarch64_ra_sign_scope == AARCH64_FUNCTION_NON_LEAF | |
3225 | && cfun->machine->frame.reg_offset[LR_REGNUM] >= 0)); | |
3226 | } | |
3227 | ||
04ddfe06 KT |
3228 | /* Emit code to save the callee-saved registers from register number START |
3229 | to LIMIT to the stack at the location starting at offset START_OFFSET, | |
3230 | skipping any write-back candidates if SKIP_WB is true. */ | |
43e9d192 | 3231 | |
43e9d192 | 3232 | static void |
ef4bddc2 | 3233 | aarch64_save_callee_saves (machine_mode mode, HOST_WIDE_INT start_offset, |
ae13fce3 | 3234 | unsigned start, unsigned limit, bool skip_wb) |
43e9d192 | 3235 | { |
5d8a22a5 | 3236 | rtx_insn *insn; |
43e9d192 IB |
3237 | unsigned regno; |
3238 | unsigned regno2; | |
3239 | ||
0ec74a1e | 3240 | for (regno = aarch64_next_callee_save (start, limit); |
64dedd72 JW |
3241 | regno <= limit; |
3242 | regno = aarch64_next_callee_save (regno + 1, limit)) | |
43e9d192 | 3243 | { |
ae13fce3 JW |
3244 | rtx reg, mem; |
3245 | HOST_WIDE_INT offset; | |
64dedd72 | 3246 | |
ae13fce3 JW |
3247 | if (skip_wb |
3248 | && (regno == cfun->machine->frame.wb_candidate1 | |
3249 | || regno == cfun->machine->frame.wb_candidate2)) | |
3250 | continue; | |
3251 | ||
827ab47a KT |
3252 | if (cfun->machine->reg_is_wrapped_separately[regno]) |
3253 | continue; | |
3254 | ||
ae13fce3 JW |
3255 | reg = gen_rtx_REG (mode, regno); |
3256 | offset = start_offset + cfun->machine->frame.reg_offset[regno]; | |
30079dde WD |
3257 | mem = gen_frame_mem (mode, plus_constant (Pmode, stack_pointer_rtx, |
3258 | offset)); | |
64dedd72 JW |
3259 | |
3260 | regno2 = aarch64_next_callee_save (regno + 1, limit); | |
3261 | ||
3262 | if (regno2 <= limit | |
827ab47a | 3263 | && !cfun->machine->reg_is_wrapped_separately[regno2] |
64dedd72 JW |
3264 | && ((cfun->machine->frame.reg_offset[regno] + UNITS_PER_WORD) |
3265 | == cfun->machine->frame.reg_offset[regno2])) | |
3266 | ||
43e9d192 | 3267 | { |
0ec74a1e | 3268 | rtx reg2 = gen_rtx_REG (mode, regno2); |
64dedd72 JW |
3269 | rtx mem2; |
3270 | ||
3271 | offset = start_offset + cfun->machine->frame.reg_offset[regno2]; | |
30079dde WD |
3272 | mem2 = gen_frame_mem (mode, plus_constant (Pmode, stack_pointer_rtx, |
3273 | offset)); | |
8ed2fc62 JW |
3274 | insn = emit_insn (aarch64_gen_store_pair (mode, mem, reg, mem2, |
3275 | reg2)); | |
0b4a9743 | 3276 | |
64dedd72 JW |
3277 | /* The first part of a frame-related parallel insn is |
3278 | always assumed to be relevant to the frame | |
3279 | calculations; subsequent parts, are only | |
3280 | frame-related if explicitly marked. */ | |
3281 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; | |
3282 | regno = regno2; | |
3283 | } | |
3284 | else | |
8ed2fc62 JW |
3285 | insn = emit_move_insn (mem, reg); |
3286 | ||
3287 | RTX_FRAME_RELATED_P (insn) = 1; | |
3288 | } | |
3289 | } | |
3290 | ||
04ddfe06 KT |
3291 | /* Emit code to restore the callee registers of mode MODE from register |
3292 | number START up to and including LIMIT. Restore from the stack offset | |
3293 | START_OFFSET, skipping any write-back candidates if SKIP_WB is true. | |
3294 | Write the appropriate REG_CFA_RESTORE notes into CFI_OPS. */ | |
3295 | ||
8ed2fc62 | 3296 | static void |
ef4bddc2 | 3297 | aarch64_restore_callee_saves (machine_mode mode, |
8ed2fc62 | 3298 | HOST_WIDE_INT start_offset, unsigned start, |
dd991abb | 3299 | unsigned limit, bool skip_wb, rtx *cfi_ops) |
8ed2fc62 | 3300 | { |
8ed2fc62 | 3301 | rtx base_rtx = stack_pointer_rtx; |
8ed2fc62 JW |
3302 | unsigned regno; |
3303 | unsigned regno2; | |
3304 | HOST_WIDE_INT offset; | |
3305 | ||
3306 | for (regno = aarch64_next_callee_save (start, limit); | |
3307 | regno <= limit; | |
3308 | regno = aarch64_next_callee_save (regno + 1, limit)) | |
3309 | { | |
827ab47a KT |
3310 | if (cfun->machine->reg_is_wrapped_separately[regno]) |
3311 | continue; | |
3312 | ||
ae13fce3 | 3313 | rtx reg, mem; |
8ed2fc62 | 3314 | |
ae13fce3 JW |
3315 | if (skip_wb |
3316 | && (regno == cfun->machine->frame.wb_candidate1 | |
3317 | || regno == cfun->machine->frame.wb_candidate2)) | |
3318 | continue; | |
3319 | ||
3320 | reg = gen_rtx_REG (mode, regno); | |
8ed2fc62 | 3321 | offset = start_offset + cfun->machine->frame.reg_offset[regno]; |
30079dde | 3322 | mem = gen_frame_mem (mode, plus_constant (Pmode, base_rtx, offset)); |
8ed2fc62 JW |
3323 | |
3324 | regno2 = aarch64_next_callee_save (regno + 1, limit); | |
3325 | ||
3326 | if (regno2 <= limit | |
827ab47a | 3327 | && !cfun->machine->reg_is_wrapped_separately[regno2] |
8ed2fc62 JW |
3328 | && ((cfun->machine->frame.reg_offset[regno] + UNITS_PER_WORD) |
3329 | == cfun->machine->frame.reg_offset[regno2])) | |
64dedd72 | 3330 | { |
8ed2fc62 JW |
3331 | rtx reg2 = gen_rtx_REG (mode, regno2); |
3332 | rtx mem2; | |
3333 | ||
3334 | offset = start_offset + cfun->machine->frame.reg_offset[regno2]; | |
30079dde | 3335 | mem2 = gen_frame_mem (mode, plus_constant (Pmode, base_rtx, offset)); |
dd991abb | 3336 | emit_insn (aarch64_gen_load_pair (mode, reg, mem, reg2, mem2)); |
8ed2fc62 | 3337 | |
dd991abb | 3338 | *cfi_ops = alloc_reg_note (REG_CFA_RESTORE, reg2, *cfi_ops); |
8ed2fc62 | 3339 | regno = regno2; |
43e9d192 | 3340 | } |
8ed2fc62 | 3341 | else |
dd991abb RH |
3342 | emit_move_insn (reg, mem); |
3343 | *cfi_ops = alloc_reg_note (REG_CFA_RESTORE, reg, *cfi_ops); | |
43e9d192 | 3344 | } |
43e9d192 IB |
3345 | } |
3346 | ||
827ab47a KT |
3347 | static inline bool |
3348 | offset_9bit_signed_unscaled_p (machine_mode mode ATTRIBUTE_UNUSED, | |
3349 | HOST_WIDE_INT offset) | |
3350 | { | |
3351 | return offset >= -256 && offset < 256; | |
3352 | } | |
3353 | ||
3354 | static inline bool | |
3355 | offset_12bit_unsigned_scaled_p (machine_mode mode, HOST_WIDE_INT offset) | |
3356 | { | |
3357 | return (offset >= 0 | |
3358 | && offset < 4096 * GET_MODE_SIZE (mode) | |
3359 | && offset % GET_MODE_SIZE (mode) == 0); | |
3360 | } | |
3361 | ||
3362 | bool | |
3363 | aarch64_offset_7bit_signed_scaled_p (machine_mode mode, HOST_WIDE_INT offset) | |
3364 | { | |
3365 | return (offset >= -64 * GET_MODE_SIZE (mode) | |
3366 | && offset < 64 * GET_MODE_SIZE (mode) | |
3367 | && offset % GET_MODE_SIZE (mode) == 0); | |
3368 | } | |
3369 | ||
3370 | /* Implement TARGET_SHRINK_WRAP_GET_SEPARATE_COMPONENTS. */ | |
3371 | ||
3372 | static sbitmap | |
3373 | aarch64_get_separate_components (void) | |
3374 | { | |
3375 | aarch64_layout_frame (); | |
3376 | ||
3377 | sbitmap components = sbitmap_alloc (LAST_SAVED_REGNUM + 1); | |
3378 | bitmap_clear (components); | |
3379 | ||
3380 | /* The registers we need saved to the frame. */ | |
3381 | for (unsigned regno = 0; regno <= LAST_SAVED_REGNUM; regno++) | |
3382 | if (aarch64_register_saved_on_entry (regno)) | |
3383 | { | |
3384 | HOST_WIDE_INT offset = cfun->machine->frame.reg_offset[regno]; | |
3385 | if (!frame_pointer_needed) | |
3386 | offset += cfun->machine->frame.frame_size | |
3387 | - cfun->machine->frame.hard_fp_offset; | |
3388 | /* Check that we can access the stack slot of the register with one | |
3389 | direct load with no adjustments needed. */ | |
3390 | if (offset_12bit_unsigned_scaled_p (DImode, offset)) | |
3391 | bitmap_set_bit (components, regno); | |
3392 | } | |
3393 | ||
3394 | /* Don't mess with the hard frame pointer. */ | |
3395 | if (frame_pointer_needed) | |
3396 | bitmap_clear_bit (components, HARD_FRAME_POINTER_REGNUM); | |
3397 | ||
3398 | unsigned reg1 = cfun->machine->frame.wb_candidate1; | |
3399 | unsigned reg2 = cfun->machine->frame.wb_candidate2; | |
3400 | /* If aarch64_layout_frame has chosen registers to store/restore with | |
3401 | writeback don't interfere with them to avoid having to output explicit | |
3402 | stack adjustment instructions. */ | |
3403 | if (reg2 != INVALID_REGNUM) | |
3404 | bitmap_clear_bit (components, reg2); | |
3405 | if (reg1 != INVALID_REGNUM) | |
3406 | bitmap_clear_bit (components, reg1); | |
3407 | ||
3408 | bitmap_clear_bit (components, LR_REGNUM); | |
3409 | bitmap_clear_bit (components, SP_REGNUM); | |
3410 | ||
3411 | return components; | |
3412 | } | |
3413 | ||
3414 | /* Implement TARGET_SHRINK_WRAP_COMPONENTS_FOR_BB. */ | |
3415 | ||
3416 | static sbitmap | |
3417 | aarch64_components_for_bb (basic_block bb) | |
3418 | { | |
3419 | bitmap in = DF_LIVE_IN (bb); | |
3420 | bitmap gen = &DF_LIVE_BB_INFO (bb)->gen; | |
3421 | bitmap kill = &DF_LIVE_BB_INFO (bb)->kill; | |
3422 | ||
3423 | sbitmap components = sbitmap_alloc (LAST_SAVED_REGNUM + 1); | |
3424 | bitmap_clear (components); | |
3425 | ||
3426 | /* GPRs are used in a bb if they are in the IN, GEN, or KILL sets. */ | |
3427 | for (unsigned regno = 0; regno <= LAST_SAVED_REGNUM; regno++) | |
3428 | if ((!call_used_regs[regno]) | |
3429 | && (bitmap_bit_p (in, regno) | |
3430 | || bitmap_bit_p (gen, regno) | |
3431 | || bitmap_bit_p (kill, regno))) | |
3432 | bitmap_set_bit (components, regno); | |
3433 | ||
3434 | return components; | |
3435 | } | |
3436 | ||
3437 | /* Implement TARGET_SHRINK_WRAP_DISQUALIFY_COMPONENTS. | |
3438 | Nothing to do for aarch64. */ | |
3439 | ||
3440 | static void | |
3441 | aarch64_disqualify_components (sbitmap, edge, sbitmap, bool) | |
3442 | { | |
3443 | } | |
3444 | ||
3445 | /* Return the next set bit in BMP from START onwards. Return the total number | |
3446 | of bits in BMP if no set bit is found at or after START. */ | |
3447 | ||
3448 | static unsigned int | |
3449 | aarch64_get_next_set_bit (sbitmap bmp, unsigned int start) | |
3450 | { | |
3451 | unsigned int nbits = SBITMAP_SIZE (bmp); | |
3452 | if (start == nbits) | |
3453 | return start; | |
3454 | ||
3455 | gcc_assert (start < nbits); | |
3456 | for (unsigned int i = start; i < nbits; i++) | |
3457 | if (bitmap_bit_p (bmp, i)) | |
3458 | return i; | |
3459 | ||
3460 | return nbits; | |
3461 | } | |
3462 | ||
3463 | /* Do the work for aarch64_emit_prologue_components and | |
3464 | aarch64_emit_epilogue_components. COMPONENTS is the bitmap of registers | |
3465 | to save/restore, PROLOGUE_P indicates whether to emit the prologue sequence | |
3466 | for these components or the epilogue sequence. That is, it determines | |
3467 | whether we should emit stores or loads and what kind of CFA notes to attach | |
3468 | to the insns. Otherwise the logic for the two sequences is very | |
3469 | similar. */ | |
3470 | ||
3471 | static void | |
3472 | aarch64_process_components (sbitmap components, bool prologue_p) | |
3473 | { | |
3474 | rtx ptr_reg = gen_rtx_REG (Pmode, frame_pointer_needed | |
3475 | ? HARD_FRAME_POINTER_REGNUM | |
3476 | : STACK_POINTER_REGNUM); | |
3477 | ||
3478 | unsigned last_regno = SBITMAP_SIZE (components); | |
3479 | unsigned regno = aarch64_get_next_set_bit (components, R0_REGNUM); | |
3480 | rtx_insn *insn = NULL; | |
3481 | ||
3482 | while (regno != last_regno) | |
3483 | { | |
3484 | /* AAPCS64 section 5.1.2 requires only the bottom 64 bits to be saved | |
3485 | so DFmode for the vector registers is enough. */ | |
0d4a1197 | 3486 | machine_mode mode = GP_REGNUM_P (regno) ? E_DImode : E_DFmode; |
827ab47a KT |
3487 | rtx reg = gen_rtx_REG (mode, regno); |
3488 | HOST_WIDE_INT offset = cfun->machine->frame.reg_offset[regno]; | |
3489 | if (!frame_pointer_needed) | |
3490 | offset += cfun->machine->frame.frame_size | |
3491 | - cfun->machine->frame.hard_fp_offset; | |
3492 | rtx addr = plus_constant (Pmode, ptr_reg, offset); | |
3493 | rtx mem = gen_frame_mem (mode, addr); | |
3494 | ||
3495 | rtx set = prologue_p ? gen_rtx_SET (mem, reg) : gen_rtx_SET (reg, mem); | |
3496 | unsigned regno2 = aarch64_get_next_set_bit (components, regno + 1); | |
3497 | /* No more registers to handle after REGNO. | |
3498 | Emit a single save/restore and exit. */ | |
3499 | if (regno2 == last_regno) | |
3500 | { | |
3501 | insn = emit_insn (set); | |
3502 | RTX_FRAME_RELATED_P (insn) = 1; | |
3503 | if (prologue_p) | |
3504 | add_reg_note (insn, REG_CFA_OFFSET, copy_rtx (set)); | |
3505 | else | |
3506 | add_reg_note (insn, REG_CFA_RESTORE, reg); | |
3507 | break; | |
3508 | } | |
3509 | ||
3510 | HOST_WIDE_INT offset2 = cfun->machine->frame.reg_offset[regno2]; | |
3511 | /* The next register is not of the same class or its offset is not | |
3512 | mergeable with the current one into a pair. */ | |
3513 | if (!satisfies_constraint_Ump (mem) | |
3514 | || GP_REGNUM_P (regno) != GP_REGNUM_P (regno2) | |
3515 | || (offset2 - cfun->machine->frame.reg_offset[regno]) | |
3516 | != GET_MODE_SIZE (mode)) | |
3517 | { | |
3518 | insn = emit_insn (set); | |
3519 | RTX_FRAME_RELATED_P (insn) = 1; | |
3520 | if (prologue_p) | |
3521 | add_reg_note (insn, REG_CFA_OFFSET, copy_rtx (set)); | |
3522 | else | |
3523 | add_reg_note (insn, REG_CFA_RESTORE, reg); | |
3524 | ||
3525 | regno = regno2; | |
3526 | continue; | |
3527 | } | |
3528 | ||
3529 | /* REGNO2 can be saved/restored in a pair with REGNO. */ | |
3530 | rtx reg2 = gen_rtx_REG (mode, regno2); | |
3531 | if (!frame_pointer_needed) | |
3532 | offset2 += cfun->machine->frame.frame_size | |
3533 | - cfun->machine->frame.hard_fp_offset; | |
3534 | rtx addr2 = plus_constant (Pmode, ptr_reg, offset2); | |
3535 | rtx mem2 = gen_frame_mem (mode, addr2); | |
3536 | rtx set2 = prologue_p ? gen_rtx_SET (mem2, reg2) | |
3537 | : gen_rtx_SET (reg2, mem2); | |
3538 | ||
3539 | if (prologue_p) | |
3540 | insn = emit_insn (aarch64_gen_store_pair (mode, mem, reg, mem2, reg2)); | |
3541 | else | |
3542 | insn = emit_insn (aarch64_gen_load_pair (mode, reg, mem, reg2, mem2)); | |
3543 | ||
3544 | RTX_FRAME_RELATED_P (insn) = 1; | |
3545 | if (prologue_p) | |
3546 | { | |
3547 | add_reg_note (insn, REG_CFA_OFFSET, set); | |
3548 | add_reg_note (insn, REG_CFA_OFFSET, set2); | |
3549 | } | |
3550 | else | |
3551 | { | |
3552 | add_reg_note (insn, REG_CFA_RESTORE, reg); | |
3553 | add_reg_note (insn, REG_CFA_RESTORE, reg2); | |
3554 | } | |
3555 | ||
3556 | regno = aarch64_get_next_set_bit (components, regno2 + 1); | |
3557 | } | |
3558 | } | |
3559 | ||
3560 | /* Implement TARGET_SHRINK_WRAP_EMIT_PROLOGUE_COMPONENTS. */ | |
3561 | ||
3562 | static void | |
3563 | aarch64_emit_prologue_components (sbitmap components) | |
3564 | { | |
3565 | aarch64_process_components (components, true); | |
3566 | } | |
3567 | ||
3568 | /* Implement TARGET_SHRINK_WRAP_EMIT_EPILOGUE_COMPONENTS. */ | |
3569 | ||
3570 | static void | |
3571 | aarch64_emit_epilogue_components (sbitmap components) | |
3572 | { | |
3573 | aarch64_process_components (components, false); | |
3574 | } | |
3575 | ||
3576 | /* Implement TARGET_SHRINK_WRAP_SET_HANDLED_COMPONENTS. */ | |
3577 | ||
3578 | static void | |
3579 | aarch64_set_handled_components (sbitmap components) | |
3580 | { | |
3581 | for (unsigned regno = 0; regno <= LAST_SAVED_REGNUM; regno++) | |
3582 | if (bitmap_bit_p (components, regno)) | |
3583 | cfun->machine->reg_is_wrapped_separately[regno] = true; | |
3584 | } | |
3585 | ||
43e9d192 IB |
3586 | /* AArch64 stack frames generated by this compiler look like: |
3587 | ||
3588 | +-------------------------------+ | |
3589 | | | | |
3590 | | incoming stack arguments | | |
3591 | | | | |
34834420 MS |
3592 | +-------------------------------+ |
3593 | | | <-- incoming stack pointer (aligned) | |
43e9d192 IB |
3594 | | callee-allocated save area | |
3595 | | for register varargs | | |
3596 | | | | |
34834420 MS |
3597 | +-------------------------------+ |
3598 | | local variables | <-- frame_pointer_rtx | |
43e9d192 IB |
3599 | | | |
3600 | +-------------------------------+ | |
454fdba9 RL |
3601 | | padding0 | \ |
3602 | +-------------------------------+ | | |
454fdba9 | 3603 | | callee-saved registers | | frame.saved_regs_size |
454fdba9 RL |
3604 | +-------------------------------+ | |
3605 | | LR' | | | |
3606 | +-------------------------------+ | | |
34834420 MS |
3607 | | FP' | / <- hard_frame_pointer_rtx (aligned) |
3608 | +-------------------------------+ | |
43e9d192 IB |
3609 | | dynamic allocation | |
3610 | +-------------------------------+ | |
34834420 MS |
3611 | | padding | |
3612 | +-------------------------------+ | |
3613 | | outgoing stack arguments | <-- arg_pointer | |
3614 | | | | |
3615 | +-------------------------------+ | |
3616 | | | <-- stack_pointer_rtx (aligned) | |
43e9d192 | 3617 | |
34834420 MS |
3618 | Dynamic stack allocations via alloca() decrease stack_pointer_rtx |
3619 | but leave frame_pointer_rtx and hard_frame_pointer_rtx | |
3620 | unchanged. */ | |
43e9d192 IB |
3621 | |
3622 | /* Generate the prologue instructions for entry into a function. | |
3623 | Establish the stack frame by decreasing the stack pointer with a | |
3624 | properly calculated size and, if necessary, create a frame record | |
3625 | filled with the values of LR and previous frame pointer. The | |
6991c977 | 3626 | current FP is also set up if it is in use. */ |
43e9d192 IB |
3627 | |
3628 | void | |
3629 | aarch64_expand_prologue (void) | |
3630 | { | |
43e9d192 | 3631 | aarch64_layout_frame (); |
43e9d192 | 3632 | |
71bfb77a WD |
3633 | HOST_WIDE_INT frame_size = cfun->machine->frame.frame_size; |
3634 | HOST_WIDE_INT initial_adjust = cfun->machine->frame.initial_adjust; | |
3635 | HOST_WIDE_INT callee_adjust = cfun->machine->frame.callee_adjust; | |
3636 | HOST_WIDE_INT final_adjust = cfun->machine->frame.final_adjust; | |
3637 | HOST_WIDE_INT callee_offset = cfun->machine->frame.callee_offset; | |
3638 | unsigned reg1 = cfun->machine->frame.wb_candidate1; | |
3639 | unsigned reg2 = cfun->machine->frame.wb_candidate2; | |
3640 | rtx_insn *insn; | |
43e9d192 | 3641 | |
db58fd89 JW |
3642 | /* Sign return address for functions. */ |
3643 | if (aarch64_return_address_signing_enabled ()) | |
27169e45 JW |
3644 | { |
3645 | insn = emit_insn (gen_pacisp ()); | |
3646 | add_reg_note (insn, REG_CFA_TOGGLE_RA_MANGLE, const0_rtx); | |
3647 | RTX_FRAME_RELATED_P (insn) = 1; | |
3648 | } | |
db58fd89 | 3649 | |
dd991abb RH |
3650 | if (flag_stack_usage_info) |
3651 | current_function_static_stack_size = frame_size; | |
43e9d192 | 3652 | |
a3eb8a52 EB |
3653 | if (flag_stack_check == STATIC_BUILTIN_STACK_CHECK) |
3654 | { | |
3655 | if (crtl->is_leaf && !cfun->calls_alloca) | |
3656 | { | |
3657 | if (frame_size > PROBE_INTERVAL && frame_size > STACK_CHECK_PROTECT) | |
3658 | aarch64_emit_probe_stack_range (STACK_CHECK_PROTECT, | |
3659 | frame_size - STACK_CHECK_PROTECT); | |
3660 | } | |
3661 | else if (frame_size > 0) | |
3662 | aarch64_emit_probe_stack_range (STACK_CHECK_PROTECT, frame_size); | |
3663 | } | |
3664 | ||
5be6b295 | 3665 | aarch64_sub_sp (IP0_REGNUM, initial_adjust, true); |
43e9d192 | 3666 | |
71bfb77a WD |
3667 | if (callee_adjust != 0) |
3668 | aarch64_push_regs (reg1, reg2, callee_adjust); | |
43e9d192 | 3669 | |
71bfb77a | 3670 | if (frame_pointer_needed) |
43e9d192 | 3671 | { |
71bfb77a WD |
3672 | if (callee_adjust == 0) |
3673 | aarch64_save_callee_saves (DImode, callee_offset, R29_REGNUM, | |
3674 | R30_REGNUM, false); | |
3675 | insn = emit_insn (gen_add3_insn (hard_frame_pointer_rtx, | |
3676 | stack_pointer_rtx, | |
3677 | GEN_INT (callee_offset))); | |
3678 | RTX_FRAME_RELATED_P (insn) = 1; | |
3679 | emit_insn (gen_stack_tie (stack_pointer_rtx, hard_frame_pointer_rtx)); | |
43e9d192 | 3680 | } |
71bfb77a WD |
3681 | |
3682 | aarch64_save_callee_saves (DImode, callee_offset, R0_REGNUM, R30_REGNUM, | |
3683 | callee_adjust != 0 || frame_pointer_needed); | |
3684 | aarch64_save_callee_saves (DFmode, callee_offset, V0_REGNUM, V31_REGNUM, | |
3685 | callee_adjust != 0 || frame_pointer_needed); | |
5be6b295 | 3686 | aarch64_sub_sp (IP1_REGNUM, final_adjust, !frame_pointer_needed); |
43e9d192 IB |
3687 | } |
3688 | ||
4f942779 RL |
3689 | /* Return TRUE if we can use a simple_return insn. |
3690 | ||
3691 | This function checks whether the callee saved stack is empty, which | |
3692 | means no restore actions are need. The pro_and_epilogue will use | |
3693 | this to check whether shrink-wrapping opt is feasible. */ | |
3694 | ||
3695 | bool | |
3696 | aarch64_use_return_insn_p (void) | |
3697 | { | |
3698 | if (!reload_completed) | |
3699 | return false; | |
3700 | ||
3701 | if (crtl->profile) | |
3702 | return false; | |
3703 | ||
3704 | aarch64_layout_frame (); | |
3705 | ||
3706 | return cfun->machine->frame.frame_size == 0; | |
3707 | } | |
3708 | ||
71bfb77a WD |
3709 | /* Generate the epilogue instructions for returning from a function. |
3710 | This is almost exactly the reverse of the prolog sequence, except | |
3711 | that we need to insert barriers to avoid scheduling loads that read | |
3712 | from a deallocated stack, and we optimize the unwind records by | |
3713 | emitting them all together if possible. */ | |
43e9d192 IB |
3714 | void |
3715 | aarch64_expand_epilogue (bool for_sibcall) | |
3716 | { | |
43e9d192 | 3717 | aarch64_layout_frame (); |
43e9d192 | 3718 | |
71bfb77a WD |
3719 | HOST_WIDE_INT initial_adjust = cfun->machine->frame.initial_adjust; |
3720 | HOST_WIDE_INT callee_adjust = cfun->machine->frame.callee_adjust; | |
3721 | HOST_WIDE_INT final_adjust = cfun->machine->frame.final_adjust; | |
3722 | HOST_WIDE_INT callee_offset = cfun->machine->frame.callee_offset; | |
3723 | unsigned reg1 = cfun->machine->frame.wb_candidate1; | |
3724 | unsigned reg2 = cfun->machine->frame.wb_candidate2; | |
3725 | rtx cfi_ops = NULL; | |
3726 | rtx_insn *insn; | |
44c0e7b9 | 3727 | |
71bfb77a WD |
3728 | /* We need to add memory barrier to prevent read from deallocated stack. */ |
3729 | bool need_barrier_p = (get_frame_size () | |
3730 | + cfun->machine->frame.saved_varargs_size) != 0; | |
43e9d192 | 3731 | |
71bfb77a | 3732 | /* Emit a barrier to prevent loads from a deallocated stack. */ |
8144a493 WD |
3733 | if (final_adjust > crtl->outgoing_args_size || cfun->calls_alloca |
3734 | || crtl->calls_eh_return) | |
43e9d192 | 3735 | { |
71bfb77a WD |
3736 | emit_insn (gen_stack_tie (stack_pointer_rtx, stack_pointer_rtx)); |
3737 | need_barrier_p = false; | |
3738 | } | |
7e8c2bd5 | 3739 | |
71bfb77a WD |
3740 | /* Restore the stack pointer from the frame pointer if it may not |
3741 | be the same as the stack pointer. */ | |
3742 | if (frame_pointer_needed && (final_adjust || cfun->calls_alloca)) | |
3743 | { | |
43e9d192 IB |
3744 | insn = emit_insn (gen_add3_insn (stack_pointer_rtx, |
3745 | hard_frame_pointer_rtx, | |
71bfb77a WD |
3746 | GEN_INT (-callee_offset))); |
3747 | /* If writeback is used when restoring callee-saves, the CFA | |
3748 | is restored on the instruction doing the writeback. */ | |
3749 | RTX_FRAME_RELATED_P (insn) = callee_adjust == 0; | |
43e9d192 | 3750 | } |
71bfb77a | 3751 | else |
5be6b295 | 3752 | aarch64_add_sp (IP1_REGNUM, final_adjust, df_regs_ever_live_p (IP1_REGNUM)); |
43e9d192 | 3753 | |
71bfb77a WD |
3754 | aarch64_restore_callee_saves (DImode, callee_offset, R0_REGNUM, R30_REGNUM, |
3755 | callee_adjust != 0, &cfi_ops); | |
3756 | aarch64_restore_callee_saves (DFmode, callee_offset, V0_REGNUM, V31_REGNUM, | |
3757 | callee_adjust != 0, &cfi_ops); | |
43e9d192 | 3758 | |
71bfb77a WD |
3759 | if (need_barrier_p) |
3760 | emit_insn (gen_stack_tie (stack_pointer_rtx, stack_pointer_rtx)); | |
3761 | ||
3762 | if (callee_adjust != 0) | |
3763 | aarch64_pop_regs (reg1, reg2, callee_adjust, &cfi_ops); | |
3764 | ||
3765 | if (callee_adjust != 0 || initial_adjust > 65536) | |
3766 | { | |
3767 | /* Emit delayed restores and set the CFA to be SP + initial_adjust. */ | |
89ac681e | 3768 | insn = get_last_insn (); |
71bfb77a WD |
3769 | rtx new_cfa = plus_constant (Pmode, stack_pointer_rtx, initial_adjust); |
3770 | REG_NOTES (insn) = alloc_reg_note (REG_CFA_DEF_CFA, new_cfa, cfi_ops); | |
43e9d192 | 3771 | RTX_FRAME_RELATED_P (insn) = 1; |
71bfb77a | 3772 | cfi_ops = NULL; |
43e9d192 IB |
3773 | } |
3774 | ||
5be6b295 | 3775 | aarch64_add_sp (IP0_REGNUM, initial_adjust, df_regs_ever_live_p (IP0_REGNUM)); |
7e8c2bd5 | 3776 | |
71bfb77a WD |
3777 | if (cfi_ops) |
3778 | { | |
3779 | /* Emit delayed restores and reset the CFA to be SP. */ | |
3780 | insn = get_last_insn (); | |
3781 | cfi_ops = alloc_reg_note (REG_CFA_DEF_CFA, stack_pointer_rtx, cfi_ops); | |
3782 | REG_NOTES (insn) = cfi_ops; | |
3783 | RTX_FRAME_RELATED_P (insn) = 1; | |
dd991abb RH |
3784 | } |
3785 | ||
db58fd89 JW |
3786 | /* We prefer to emit the combined return/authenticate instruction RETAA, |
3787 | however there are three cases in which we must instead emit an explicit | |
3788 | authentication instruction. | |
3789 | ||
3790 | 1) Sibcalls don't return in a normal way, so if we're about to call one | |
3791 | we must authenticate. | |
3792 | ||
3793 | 2) The RETAA instruction is not available before ARMv8.3-A, so if we are | |
3794 | generating code for !TARGET_ARMV8_3 we can't use it and must | |
3795 | explicitly authenticate. | |
3796 | ||
3797 | 3) On an eh_return path we make extra stack adjustments to update the | |
3798 | canonical frame address to be the exception handler's CFA. We want | |
3799 | to authenticate using the CFA of the function which calls eh_return. | |
3800 | */ | |
3801 | if (aarch64_return_address_signing_enabled () | |
3802 | && (for_sibcall || !TARGET_ARMV8_3 || crtl->calls_eh_return)) | |
27169e45 JW |
3803 | { |
3804 | insn = emit_insn (gen_autisp ()); | |
3805 | add_reg_note (insn, REG_CFA_TOGGLE_RA_MANGLE, const0_rtx); | |
3806 | RTX_FRAME_RELATED_P (insn) = 1; | |
3807 | } | |
db58fd89 | 3808 | |
dd991abb RH |
3809 | /* Stack adjustment for exception handler. */ |
3810 | if (crtl->calls_eh_return) | |
3811 | { | |
3812 | /* We need to unwind the stack by the offset computed by | |
3813 | EH_RETURN_STACKADJ_RTX. We have already reset the CFA | |
3814 | to be SP; letting the CFA move during this adjustment | |
3815 | is just as correct as retaining the CFA from the body | |
3816 | of the function. Therefore, do nothing special. */ | |
3817 | emit_insn (gen_add2_insn (stack_pointer_rtx, EH_RETURN_STACKADJ_RTX)); | |
43e9d192 IB |
3818 | } |
3819 | ||
3820 | emit_use (gen_rtx_REG (DImode, LR_REGNUM)); | |
3821 | if (!for_sibcall) | |
3822 | emit_jump_insn (ret_rtx); | |
3823 | } | |
3824 | ||
8144a493 WD |
3825 | /* Implement EH_RETURN_HANDLER_RTX. EH returns need to either return |
3826 | normally or return to a previous frame after unwinding. | |
1c960e02 | 3827 | |
8144a493 WD |
3828 | An EH return uses a single shared return sequence. The epilogue is |
3829 | exactly like a normal epilogue except that it has an extra input | |
3830 | register (EH_RETURN_STACKADJ_RTX) which contains the stack adjustment | |
3831 | that must be applied after the frame has been destroyed. An extra label | |
3832 | is inserted before the epilogue which initializes this register to zero, | |
3833 | and this is the entry point for a normal return. | |
43e9d192 | 3834 | |
8144a493 WD |
3835 | An actual EH return updates the return address, initializes the stack |
3836 | adjustment and jumps directly into the epilogue (bypassing the zeroing | |
3837 | of the adjustment). Since the return address is typically saved on the | |
3838 | stack when a function makes a call, the saved LR must be updated outside | |
3839 | the epilogue. | |
43e9d192 | 3840 | |
8144a493 WD |
3841 | This poses problems as the store is generated well before the epilogue, |
3842 | so the offset of LR is not known yet. Also optimizations will remove the | |
3843 | store as it appears dead, even after the epilogue is generated (as the | |
3844 | base or offset for loading LR is different in many cases). | |
43e9d192 | 3845 | |
8144a493 WD |
3846 | To avoid these problems this implementation forces the frame pointer |
3847 | in eh_return functions so that the location of LR is fixed and known early. | |
3848 | It also marks the store volatile, so no optimization is permitted to | |
3849 | remove the store. */ | |
3850 | rtx | |
3851 | aarch64_eh_return_handler_rtx (void) | |
3852 | { | |
3853 | rtx tmp = gen_frame_mem (Pmode, | |
3854 | plus_constant (Pmode, hard_frame_pointer_rtx, UNITS_PER_WORD)); | |
43e9d192 | 3855 | |
8144a493 WD |
3856 | /* Mark the store volatile, so no optimization is permitted to remove it. */ |
3857 | MEM_VOLATILE_P (tmp) = true; | |
3858 | return tmp; | |
43e9d192 IB |
3859 | } |
3860 | ||
43e9d192 IB |
3861 | /* Output code to add DELTA to the first argument, and then jump |
3862 | to FUNCTION. Used for C++ multiple inheritance. */ | |
3863 | static void | |
3864 | aarch64_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED, | |
3865 | HOST_WIDE_INT delta, | |
3866 | HOST_WIDE_INT vcall_offset, | |
3867 | tree function) | |
3868 | { | |
3869 | /* The this pointer is always in x0. Note that this differs from | |
3870 | Arm where the this pointer maybe bumped to r1 if r0 is required | |
3871 | to return a pointer to an aggregate. On AArch64 a result value | |
3872 | pointer will be in x8. */ | |
3873 | int this_regno = R0_REGNUM; | |
5d8a22a5 DM |
3874 | rtx this_rtx, temp0, temp1, addr, funexp; |
3875 | rtx_insn *insn; | |
43e9d192 | 3876 | |
75f1d6fc SN |
3877 | reload_completed = 1; |
3878 | emit_note (NOTE_INSN_PROLOGUE_END); | |
43e9d192 IB |
3879 | |
3880 | if (vcall_offset == 0) | |
5be6b295 | 3881 | aarch64_add_constant (Pmode, this_regno, IP1_REGNUM, delta); |
43e9d192 IB |
3882 | else |
3883 | { | |
28514dda | 3884 | gcc_assert ((vcall_offset & (POINTER_BYTES - 1)) == 0); |
43e9d192 | 3885 | |
75f1d6fc SN |
3886 | this_rtx = gen_rtx_REG (Pmode, this_regno); |
3887 | temp0 = gen_rtx_REG (Pmode, IP0_REGNUM); | |
3888 | temp1 = gen_rtx_REG (Pmode, IP1_REGNUM); | |
43e9d192 | 3889 | |
75f1d6fc SN |
3890 | addr = this_rtx; |
3891 | if (delta != 0) | |
3892 | { | |
3893 | if (delta >= -256 && delta < 256) | |
3894 | addr = gen_rtx_PRE_MODIFY (Pmode, this_rtx, | |
3895 | plus_constant (Pmode, this_rtx, delta)); | |
3896 | else | |
5be6b295 | 3897 | aarch64_add_constant (Pmode, this_regno, IP1_REGNUM, delta); |
43e9d192 IB |
3898 | } |
3899 | ||
28514dda YZ |
3900 | if (Pmode == ptr_mode) |
3901 | aarch64_emit_move (temp0, gen_rtx_MEM (ptr_mode, addr)); | |
3902 | else | |
3903 | aarch64_emit_move (temp0, | |
3904 | gen_rtx_ZERO_EXTEND (Pmode, | |
3905 | gen_rtx_MEM (ptr_mode, addr))); | |
75f1d6fc | 3906 | |
28514dda | 3907 | if (vcall_offset >= -256 && vcall_offset < 4096 * POINTER_BYTES) |
75f1d6fc | 3908 | addr = plus_constant (Pmode, temp0, vcall_offset); |
43e9d192 IB |
3909 | else |
3910 | { | |
f43657b4 JW |
3911 | aarch64_internal_mov_immediate (temp1, GEN_INT (vcall_offset), true, |
3912 | Pmode); | |
75f1d6fc | 3913 | addr = gen_rtx_PLUS (Pmode, temp0, temp1); |
43e9d192 IB |
3914 | } |
3915 | ||
28514dda YZ |
3916 | if (Pmode == ptr_mode) |
3917 | aarch64_emit_move (temp1, gen_rtx_MEM (ptr_mode,addr)); | |
3918 | else | |
3919 | aarch64_emit_move (temp1, | |
3920 | gen_rtx_SIGN_EXTEND (Pmode, | |
3921 | gen_rtx_MEM (ptr_mode, addr))); | |
3922 | ||
75f1d6fc | 3923 | emit_insn (gen_add2_insn (this_rtx, temp1)); |
43e9d192 IB |
3924 | } |
3925 | ||
75f1d6fc SN |
3926 | /* Generate a tail call to the target function. */ |
3927 | if (!TREE_USED (function)) | |
3928 | { | |
3929 | assemble_external (function); | |
3930 | TREE_USED (function) = 1; | |
3931 | } | |
3932 | funexp = XEXP (DECL_RTL (function), 0); | |
3933 | funexp = gen_rtx_MEM (FUNCTION_MODE, funexp); | |
3934 | insn = emit_call_insn (gen_sibcall (funexp, const0_rtx, NULL_RTX)); | |
3935 | SIBLING_CALL_P (insn) = 1; | |
3936 | ||
3937 | insn = get_insns (); | |
3938 | shorten_branches (insn); | |
3939 | final_start_function (insn, file, 1); | |
3940 | final (insn, file, 1); | |
43e9d192 | 3941 | final_end_function (); |
75f1d6fc SN |
3942 | |
3943 | /* Stop pretending to be a post-reload pass. */ | |
3944 | reload_completed = 0; | |
43e9d192 IB |
3945 | } |
3946 | ||
43e9d192 IB |
3947 | static bool |
3948 | aarch64_tls_referenced_p (rtx x) | |
3949 | { | |
3950 | if (!TARGET_HAVE_TLS) | |
3951 | return false; | |
e7de8563 RS |
3952 | subrtx_iterator::array_type array; |
3953 | FOR_EACH_SUBRTX (iter, array, x, ALL) | |
3954 | { | |
3955 | const_rtx x = *iter; | |
3956 | if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (x) != 0) | |
3957 | return true; | |
3958 | /* Don't recurse into UNSPEC_TLS looking for TLS symbols; these are | |
3959 | TLS offsets, not real symbol references. */ | |
3960 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
3961 | iter.skip_subrtxes (); | |
3962 | } | |
3963 | return false; | |
43e9d192 IB |
3964 | } |
3965 | ||
3966 | ||
43e9d192 IB |
3967 | /* Return true if val can be encoded as a 12-bit unsigned immediate with |
3968 | a left shift of 0 or 12 bits. */ | |
3969 | bool | |
3970 | aarch64_uimm12_shift (HOST_WIDE_INT val) | |
3971 | { | |
3972 | return ((val & (((HOST_WIDE_INT) 0xfff) << 0)) == val | |
3973 | || (val & (((HOST_WIDE_INT) 0xfff) << 12)) == val | |
3974 | ); | |
3975 | } | |
3976 | ||
3977 | ||
3978 | /* Return true if val is an immediate that can be loaded into a | |
3979 | register by a MOVZ instruction. */ | |
3980 | static bool | |
ef4bddc2 | 3981 | aarch64_movw_imm (HOST_WIDE_INT val, machine_mode mode) |
43e9d192 IB |
3982 | { |
3983 | if (GET_MODE_SIZE (mode) > 4) | |
3984 | { | |
3985 | if ((val & (((HOST_WIDE_INT) 0xffff) << 32)) == val | |
3986 | || (val & (((HOST_WIDE_INT) 0xffff) << 48)) == val) | |
3987 | return 1; | |
3988 | } | |
3989 | else | |
3990 | { | |
3991 | /* Ignore sign extension. */ | |
3992 | val &= (HOST_WIDE_INT) 0xffffffff; | |
3993 | } | |
3994 | return ((val & (((HOST_WIDE_INT) 0xffff) << 0)) == val | |
3995 | || (val & (((HOST_WIDE_INT) 0xffff) << 16)) == val); | |
3996 | } | |
3997 | ||
a64c73a2 WD |
3998 | /* Multipliers for repeating bitmasks of width 32, 16, 8, 4, and 2. */ |
3999 | ||
4000 | static const unsigned HOST_WIDE_INT bitmask_imm_mul[] = | |
4001 | { | |
4002 | 0x0000000100000001ull, | |
4003 | 0x0001000100010001ull, | |
4004 | 0x0101010101010101ull, | |
4005 | 0x1111111111111111ull, | |
4006 | 0x5555555555555555ull, | |
4007 | }; | |
4008 | ||
43e9d192 IB |
4009 | |
4010 | /* Return true if val is a valid bitmask immediate. */ | |
a64c73a2 | 4011 | |
43e9d192 | 4012 | bool |
a64c73a2 | 4013 | aarch64_bitmask_imm (HOST_WIDE_INT val_in, machine_mode mode) |
43e9d192 | 4014 | { |
a64c73a2 WD |
4015 | unsigned HOST_WIDE_INT val, tmp, mask, first_one, next_one; |
4016 | int bits; | |
4017 | ||
4018 | /* Check for a single sequence of one bits and return quickly if so. | |
4019 | The special cases of all ones and all zeroes returns false. */ | |
4020 | val = (unsigned HOST_WIDE_INT) val_in; | |
4021 | tmp = val + (val & -val); | |
4022 | ||
4023 | if (tmp == (tmp & -tmp)) | |
4024 | return (val + 1) > 1; | |
4025 | ||
4026 | /* Replicate 32-bit immediates so we can treat them as 64-bit. */ | |
4027 | if (mode == SImode) | |
4028 | val = (val << 32) | (val & 0xffffffff); | |
4029 | ||
4030 | /* Invert if the immediate doesn't start with a zero bit - this means we | |
4031 | only need to search for sequences of one bits. */ | |
4032 | if (val & 1) | |
4033 | val = ~val; | |
4034 | ||
4035 | /* Find the first set bit and set tmp to val with the first sequence of one | |
4036 | bits removed. Return success if there is a single sequence of ones. */ | |
4037 | first_one = val & -val; | |
4038 | tmp = val & (val + first_one); | |
4039 | ||
4040 | if (tmp == 0) | |
4041 | return true; | |
4042 | ||
4043 | /* Find the next set bit and compute the difference in bit position. */ | |
4044 | next_one = tmp & -tmp; | |
4045 | bits = clz_hwi (first_one) - clz_hwi (next_one); | |
4046 | mask = val ^ tmp; | |
4047 | ||
4048 | /* Check the bit position difference is a power of 2, and that the first | |
4049 | sequence of one bits fits within 'bits' bits. */ | |
4050 | if ((mask >> bits) != 0 || bits != (bits & -bits)) | |
4051 | return false; | |
4052 | ||
4053 | /* Check the sequence of one bits is repeated 64/bits times. */ | |
4054 | return val == mask * bitmask_imm_mul[__builtin_clz (bits) - 26]; | |
43e9d192 IB |
4055 | } |
4056 | ||
43fd192f MC |
4057 | /* Create mask of ones, covering the lowest to highest bits set in VAL_IN. |
4058 | Assumed precondition: VAL_IN Is not zero. */ | |
4059 | ||
4060 | unsigned HOST_WIDE_INT | |
4061 | aarch64_and_split_imm1 (HOST_WIDE_INT val_in) | |
4062 | { | |
4063 | int lowest_bit_set = ctz_hwi (val_in); | |
4064 | int highest_bit_set = floor_log2 (val_in); | |
4065 | gcc_assert (val_in != 0); | |
4066 | ||
4067 | return ((HOST_WIDE_INT_UC (2) << highest_bit_set) - | |
4068 | (HOST_WIDE_INT_1U << lowest_bit_set)); | |
4069 | } | |
4070 | ||
4071 | /* Create constant where bits outside of lowest bit set to highest bit set | |
4072 | are set to 1. */ | |
4073 | ||
4074 | unsigned HOST_WIDE_INT | |
4075 | aarch64_and_split_imm2 (HOST_WIDE_INT val_in) | |
4076 | { | |
4077 | return val_in | ~aarch64_and_split_imm1 (val_in); | |
4078 | } | |
4079 | ||
4080 | /* Return true if VAL_IN is a valid 'and' bitmask immediate. */ | |
4081 | ||
4082 | bool | |
4083 | aarch64_and_bitmask_imm (unsigned HOST_WIDE_INT val_in, machine_mode mode) | |
4084 | { | |
4085 | if (aarch64_bitmask_imm (val_in, mode)) | |
4086 | return false; | |
4087 | ||
4088 | if (aarch64_move_imm (val_in, mode)) | |
4089 | return false; | |
4090 | ||
4091 | unsigned HOST_WIDE_INT imm2 = aarch64_and_split_imm2 (val_in); | |
4092 | ||
4093 | return aarch64_bitmask_imm (imm2, mode); | |
4094 | } | |
43e9d192 IB |
4095 | |
4096 | /* Return true if val is an immediate that can be loaded into a | |
4097 | register in a single instruction. */ | |
4098 | bool | |
ef4bddc2 | 4099 | aarch64_move_imm (HOST_WIDE_INT val, machine_mode mode) |
43e9d192 IB |
4100 | { |
4101 | if (aarch64_movw_imm (val, mode) || aarch64_movw_imm (~val, mode)) | |
4102 | return 1; | |
4103 | return aarch64_bitmask_imm (val, mode); | |
4104 | } | |
4105 | ||
4106 | static bool | |
ef4bddc2 | 4107 | aarch64_cannot_force_const_mem (machine_mode mode ATTRIBUTE_UNUSED, rtx x) |
43e9d192 IB |
4108 | { |
4109 | rtx base, offset; | |
7eda14e1 | 4110 | |
43e9d192 IB |
4111 | if (GET_CODE (x) == HIGH) |
4112 | return true; | |
4113 | ||
4114 | split_const (x, &base, &offset); | |
4115 | if (GET_CODE (base) == SYMBOL_REF || GET_CODE (base) == LABEL_REF) | |
28514dda | 4116 | { |
a6e0bfa7 | 4117 | if (aarch64_classify_symbol (base, offset) |
28514dda YZ |
4118 | != SYMBOL_FORCE_TO_MEM) |
4119 | return true; | |
4120 | else | |
4121 | /* Avoid generating a 64-bit relocation in ILP32; leave | |
4122 | to aarch64_expand_mov_immediate to handle it properly. */ | |
4123 | return mode != ptr_mode; | |
4124 | } | |
43e9d192 IB |
4125 | |
4126 | return aarch64_tls_referenced_p (x); | |
4127 | } | |
4128 | ||
e79136e4 WD |
4129 | /* Implement TARGET_CASE_VALUES_THRESHOLD. |
4130 | The expansion for a table switch is quite expensive due to the number | |
4131 | of instructions, the table lookup and hard to predict indirect jump. | |
4132 | When optimizing for speed, and -O3 enabled, use the per-core tuning if | |
4133 | set, otherwise use tables for > 16 cases as a tradeoff between size and | |
4134 | performance. When optimizing for size, use the default setting. */ | |
50487d79 EM |
4135 | |
4136 | static unsigned int | |
4137 | aarch64_case_values_threshold (void) | |
4138 | { | |
4139 | /* Use the specified limit for the number of cases before using jump | |
4140 | tables at higher optimization levels. */ | |
4141 | if (optimize > 2 | |
4142 | && selected_cpu->tune->max_case_values != 0) | |
4143 | return selected_cpu->tune->max_case_values; | |
4144 | else | |
e79136e4 | 4145 | return optimize_size ? default_case_values_threshold () : 17; |
50487d79 EM |
4146 | } |
4147 | ||
43e9d192 IB |
4148 | /* Return true if register REGNO is a valid index register. |
4149 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
4150 | ||
4151 | bool | |
4152 | aarch64_regno_ok_for_index_p (int regno, bool strict_p) | |
4153 | { | |
4154 | if (!HARD_REGISTER_NUM_P (regno)) | |
4155 | { | |
4156 | if (!strict_p) | |
4157 | return true; | |
4158 | ||
4159 | if (!reg_renumber) | |
4160 | return false; | |
4161 | ||
4162 | regno = reg_renumber[regno]; | |
4163 | } | |
4164 | return GP_REGNUM_P (regno); | |
4165 | } | |
4166 | ||
4167 | /* Return true if register REGNO is a valid base register for mode MODE. | |
4168 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
4169 | ||
4170 | bool | |
4171 | aarch64_regno_ok_for_base_p (int regno, bool strict_p) | |
4172 | { | |
4173 | if (!HARD_REGISTER_NUM_P (regno)) | |
4174 | { | |
4175 | if (!strict_p) | |
4176 | return true; | |
4177 | ||
4178 | if (!reg_renumber) | |
4179 | return false; | |
4180 | ||
4181 | regno = reg_renumber[regno]; | |
4182 | } | |
4183 | ||
4184 | /* The fake registers will be eliminated to either the stack or | |
4185 | hard frame pointer, both of which are usually valid base registers. | |
4186 | Reload deals with the cases where the eliminated form isn't valid. */ | |
4187 | return (GP_REGNUM_P (regno) | |
4188 | || regno == SP_REGNUM | |
4189 | || regno == FRAME_POINTER_REGNUM | |
4190 | || regno == ARG_POINTER_REGNUM); | |
4191 | } | |
4192 | ||
4193 | /* Return true if X is a valid base register for mode MODE. | |
4194 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
4195 | ||
4196 | static bool | |
4197 | aarch64_base_register_rtx_p (rtx x, bool strict_p) | |
4198 | { | |
4199 | if (!strict_p && GET_CODE (x) == SUBREG) | |
4200 | x = SUBREG_REG (x); | |
4201 | ||
4202 | return (REG_P (x) && aarch64_regno_ok_for_base_p (REGNO (x), strict_p)); | |
4203 | } | |
4204 | ||
4205 | /* Return true if address offset is a valid index. If it is, fill in INFO | |
4206 | appropriately. STRICT_P is true if REG_OK_STRICT is in effect. */ | |
4207 | ||
4208 | static bool | |
4209 | aarch64_classify_index (struct aarch64_address_info *info, rtx x, | |
ef4bddc2 | 4210 | machine_mode mode, bool strict_p) |
43e9d192 IB |
4211 | { |
4212 | enum aarch64_address_type type; | |
4213 | rtx index; | |
4214 | int shift; | |
4215 | ||
4216 | /* (reg:P) */ | |
4217 | if ((REG_P (x) || GET_CODE (x) == SUBREG) | |
4218 | && GET_MODE (x) == Pmode) | |
4219 | { | |
4220 | type = ADDRESS_REG_REG; | |
4221 | index = x; | |
4222 | shift = 0; | |
4223 | } | |
4224 | /* (sign_extend:DI (reg:SI)) */ | |
4225 | else if ((GET_CODE (x) == SIGN_EXTEND | |
4226 | || GET_CODE (x) == ZERO_EXTEND) | |
4227 | && GET_MODE (x) == DImode | |
4228 | && GET_MODE (XEXP (x, 0)) == SImode) | |
4229 | { | |
4230 | type = (GET_CODE (x) == SIGN_EXTEND) | |
4231 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
4232 | index = XEXP (x, 0); | |
4233 | shift = 0; | |
4234 | } | |
4235 | /* (mult:DI (sign_extend:DI (reg:SI)) (const_int scale)) */ | |
4236 | else if (GET_CODE (x) == MULT | |
4237 | && (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
4238 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
4239 | && GET_MODE (XEXP (x, 0)) == DImode | |
4240 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == SImode | |
4241 | && CONST_INT_P (XEXP (x, 1))) | |
4242 | { | |
4243 | type = (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
4244 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
4245 | index = XEXP (XEXP (x, 0), 0); | |
4246 | shift = exact_log2 (INTVAL (XEXP (x, 1))); | |
4247 | } | |
4248 | /* (ashift:DI (sign_extend:DI (reg:SI)) (const_int shift)) */ | |
4249 | else if (GET_CODE (x) == ASHIFT | |
4250 | && (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
4251 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
4252 | && GET_MODE (XEXP (x, 0)) == DImode | |
4253 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == SImode | |
4254 | && CONST_INT_P (XEXP (x, 1))) | |
4255 | { | |
4256 | type = (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
4257 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
4258 | index = XEXP (XEXP (x, 0), 0); | |
4259 | shift = INTVAL (XEXP (x, 1)); | |
4260 | } | |
4261 | /* (sign_extract:DI (mult:DI (reg:DI) (const_int scale)) 32+shift 0) */ | |
4262 | else if ((GET_CODE (x) == SIGN_EXTRACT | |
4263 | || GET_CODE (x) == ZERO_EXTRACT) | |
4264 | && GET_MODE (x) == DImode | |
4265 | && GET_CODE (XEXP (x, 0)) == MULT | |
4266 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
4267 | && CONST_INT_P (XEXP (XEXP (x, 0), 1))) | |
4268 | { | |
4269 | type = (GET_CODE (x) == SIGN_EXTRACT) | |
4270 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
4271 | index = XEXP (XEXP (x, 0), 0); | |
4272 | shift = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1))); | |
4273 | if (INTVAL (XEXP (x, 1)) != 32 + shift | |
4274 | || INTVAL (XEXP (x, 2)) != 0) | |
4275 | shift = -1; | |
4276 | } | |
4277 | /* (and:DI (mult:DI (reg:DI) (const_int scale)) | |
4278 | (const_int 0xffffffff<<shift)) */ | |
4279 | else if (GET_CODE (x) == AND | |
4280 | && GET_MODE (x) == DImode | |
4281 | && GET_CODE (XEXP (x, 0)) == MULT | |
4282 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
4283 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
4284 | && CONST_INT_P (XEXP (x, 1))) | |
4285 | { | |
4286 | type = ADDRESS_REG_UXTW; | |
4287 | index = XEXP (XEXP (x, 0), 0); | |
4288 | shift = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1))); | |
4289 | if (INTVAL (XEXP (x, 1)) != (HOST_WIDE_INT)0xffffffff << shift) | |
4290 | shift = -1; | |
4291 | } | |
4292 | /* (sign_extract:DI (ashift:DI (reg:DI) (const_int shift)) 32+shift 0) */ | |
4293 | else if ((GET_CODE (x) == SIGN_EXTRACT | |
4294 | || GET_CODE (x) == ZERO_EXTRACT) | |
4295 | && GET_MODE (x) == DImode | |
4296 | && GET_CODE (XEXP (x, 0)) == ASHIFT | |
4297 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
4298 | && CONST_INT_P (XEXP (XEXP (x, 0), 1))) | |
4299 | { | |
4300 | type = (GET_CODE (x) == SIGN_EXTRACT) | |
4301 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
4302 | index = XEXP (XEXP (x, 0), 0); | |
4303 | shift = INTVAL (XEXP (XEXP (x, 0), 1)); | |
4304 | if (INTVAL (XEXP (x, 1)) != 32 + shift | |
4305 | || INTVAL (XEXP (x, 2)) != 0) | |
4306 | shift = -1; | |
4307 | } | |
4308 | /* (and:DI (ashift:DI (reg:DI) (const_int shift)) | |
4309 | (const_int 0xffffffff<<shift)) */ | |
4310 | else if (GET_CODE (x) == AND | |
4311 | && GET_MODE (x) == DImode | |
4312 | && GET_CODE (XEXP (x, 0)) == ASHIFT | |
4313 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
4314 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
4315 | && CONST_INT_P (XEXP (x, 1))) | |
4316 | { | |
4317 | type = ADDRESS_REG_UXTW; | |
4318 | index = XEXP (XEXP (x, 0), 0); | |
4319 | shift = INTVAL (XEXP (XEXP (x, 0), 1)); | |
4320 | if (INTVAL (XEXP (x, 1)) != (HOST_WIDE_INT)0xffffffff << shift) | |
4321 | shift = -1; | |
4322 | } | |
4323 | /* (mult:P (reg:P) (const_int scale)) */ | |
4324 | else if (GET_CODE (x) == MULT | |
4325 | && GET_MODE (x) == Pmode | |
4326 | && GET_MODE (XEXP (x, 0)) == Pmode | |
4327 | && CONST_INT_P (XEXP (x, 1))) | |
4328 | { | |
4329 | type = ADDRESS_REG_REG; | |
4330 | index = XEXP (x, 0); | |
4331 | shift = exact_log2 (INTVAL (XEXP (x, 1))); | |
4332 | } | |
4333 | /* (ashift:P (reg:P) (const_int shift)) */ | |
4334 | else if (GET_CODE (x) == ASHIFT | |
4335 | && GET_MODE (x) == Pmode | |
4336 | && GET_MODE (XEXP (x, 0)) == Pmode | |
4337 | && CONST_INT_P (XEXP (x, 1))) | |
4338 | { | |
4339 | type = ADDRESS_REG_REG; | |
4340 | index = XEXP (x, 0); | |
4341 | shift = INTVAL (XEXP (x, 1)); | |
4342 | } | |
4343 | else | |
4344 | return false; | |
4345 | ||
4346 | if (GET_CODE (index) == SUBREG) | |
4347 | index = SUBREG_REG (index); | |
4348 | ||
4349 | if ((shift == 0 || | |
4350 | (shift > 0 && shift <= 3 | |
4351 | && (1 << shift) == GET_MODE_SIZE (mode))) | |
4352 | && REG_P (index) | |
4353 | && aarch64_regno_ok_for_index_p (REGNO (index), strict_p)) | |
4354 | { | |
4355 | info->type = type; | |
4356 | info->offset = index; | |
4357 | info->shift = shift; | |
4358 | return true; | |
4359 | } | |
4360 | ||
4361 | return false; | |
4362 | } | |
4363 | ||
abc52318 KT |
4364 | /* Return true if MODE is one of the modes for which we |
4365 | support LDP/STP operations. */ | |
4366 | ||
4367 | static bool | |
4368 | aarch64_mode_valid_for_sched_fusion_p (machine_mode mode) | |
4369 | { | |
4370 | return mode == SImode || mode == DImode | |
4371 | || mode == SFmode || mode == DFmode | |
4372 | || (aarch64_vector_mode_supported_p (mode) | |
4373 | && GET_MODE_SIZE (mode) == 8); | |
4374 | } | |
4375 | ||
9e0218fc RH |
4376 | /* Return true if REGNO is a virtual pointer register, or an eliminable |
4377 | "soft" frame register. Like REGNO_PTR_FRAME_P except that we don't | |
4378 | include stack_pointer or hard_frame_pointer. */ | |
4379 | static bool | |
4380 | virt_or_elim_regno_p (unsigned regno) | |
4381 | { | |
4382 | return ((regno >= FIRST_VIRTUAL_REGISTER | |
4383 | && regno <= LAST_VIRTUAL_POINTER_REGISTER) | |
4384 | || regno == FRAME_POINTER_REGNUM | |
4385 | || regno == ARG_POINTER_REGNUM); | |
4386 | } | |
4387 | ||
43e9d192 IB |
4388 | /* Return true if X is a valid address for machine mode MODE. If it is, |
4389 | fill in INFO appropriately. STRICT_P is true if REG_OK_STRICT is in | |
4390 | effect. OUTER_CODE is PARALLEL for a load/store pair. */ | |
4391 | ||
4392 | static bool | |
4393 | aarch64_classify_address (struct aarch64_address_info *info, | |
ef4bddc2 | 4394 | rtx x, machine_mode mode, |
43e9d192 IB |
4395 | RTX_CODE outer_code, bool strict_p) |
4396 | { | |
4397 | enum rtx_code code = GET_CODE (x); | |
4398 | rtx op0, op1; | |
2d8c6dc1 | 4399 | |
80d43579 WD |
4400 | /* On BE, we use load/store pair for all large int mode load/stores. |
4401 | TI/TFmode may also use a load/store pair. */ | |
2d8c6dc1 | 4402 | bool load_store_pair_p = (outer_code == PARALLEL |
80d43579 WD |
4403 | || mode == TImode |
4404 | || mode == TFmode | |
2d8c6dc1 AH |
4405 | || (BYTES_BIG_ENDIAN |
4406 | && aarch64_vect_struct_mode_p (mode))); | |
4407 | ||
43e9d192 | 4408 | bool allow_reg_index_p = |
2d8c6dc1 AH |
4409 | !load_store_pair_p |
4410 | && (GET_MODE_SIZE (mode) != 16 || aarch64_vector_mode_supported_p (mode)) | |
4411 | && !aarch64_vect_struct_mode_p (mode); | |
4412 | ||
4413 | /* On LE, for AdvSIMD, don't support anything other than POST_INC or | |
4414 | REG addressing. */ | |
4415 | if (aarch64_vect_struct_mode_p (mode) && !BYTES_BIG_ENDIAN | |
43e9d192 IB |
4416 | && (code != POST_INC && code != REG)) |
4417 | return false; | |
4418 | ||
4419 | switch (code) | |
4420 | { | |
4421 | case REG: | |
4422 | case SUBREG: | |
4423 | info->type = ADDRESS_REG_IMM; | |
4424 | info->base = x; | |
4425 | info->offset = const0_rtx; | |
4426 | return aarch64_base_register_rtx_p (x, strict_p); | |
4427 | ||
4428 | case PLUS: | |
4429 | op0 = XEXP (x, 0); | |
4430 | op1 = XEXP (x, 1); | |
15c0c5c9 JW |
4431 | |
4432 | if (! strict_p | |
4aa81c2e | 4433 | && REG_P (op0) |
9e0218fc | 4434 | && virt_or_elim_regno_p (REGNO (op0)) |
4aa81c2e | 4435 | && CONST_INT_P (op1)) |
15c0c5c9 JW |
4436 | { |
4437 | info->type = ADDRESS_REG_IMM; | |
4438 | info->base = op0; | |
4439 | info->offset = op1; | |
4440 | ||
4441 | return true; | |
4442 | } | |
4443 | ||
43e9d192 IB |
4444 | if (GET_MODE_SIZE (mode) != 0 |
4445 | && CONST_INT_P (op1) | |
4446 | && aarch64_base_register_rtx_p (op0, strict_p)) | |
4447 | { | |
4448 | HOST_WIDE_INT offset = INTVAL (op1); | |
4449 | ||
4450 | info->type = ADDRESS_REG_IMM; | |
4451 | info->base = op0; | |
4452 | info->offset = op1; | |
4453 | ||
4454 | /* TImode and TFmode values are allowed in both pairs of X | |
4455 | registers and individual Q registers. The available | |
4456 | address modes are: | |
4457 | X,X: 7-bit signed scaled offset | |
4458 | Q: 9-bit signed offset | |
4459 | We conservatively require an offset representable in either mode. | |
8ed49fab KT |
4460 | When performing the check for pairs of X registers i.e. LDP/STP |
4461 | pass down DImode since that is the natural size of the LDP/STP | |
4462 | instruction memory accesses. */ | |
43e9d192 | 4463 | if (mode == TImode || mode == TFmode) |
8ed49fab | 4464 | return (aarch64_offset_7bit_signed_scaled_p (DImode, offset) |
8734dfac WD |
4465 | && (offset_9bit_signed_unscaled_p (mode, offset) |
4466 | || offset_12bit_unsigned_scaled_p (mode, offset))); | |
43e9d192 | 4467 | |
2d8c6dc1 AH |
4468 | /* A 7bit offset check because OImode will emit a ldp/stp |
4469 | instruction (only big endian will get here). | |
4470 | For ldp/stp instructions, the offset is scaled for the size of a | |
4471 | single element of the pair. */ | |
4472 | if (mode == OImode) | |
4473 | return aarch64_offset_7bit_signed_scaled_p (TImode, offset); | |
4474 | ||
4475 | /* Three 9/12 bit offsets checks because CImode will emit three | |
4476 | ldr/str instructions (only big endian will get here). */ | |
4477 | if (mode == CImode) | |
4478 | return (aarch64_offset_7bit_signed_scaled_p (TImode, offset) | |
4479 | && (offset_9bit_signed_unscaled_p (V16QImode, offset + 32) | |
4480 | || offset_12bit_unsigned_scaled_p (V16QImode, | |
4481 | offset + 32))); | |
4482 | ||
4483 | /* Two 7bit offsets checks because XImode will emit two ldp/stp | |
4484 | instructions (only big endian will get here). */ | |
4485 | if (mode == XImode) | |
4486 | return (aarch64_offset_7bit_signed_scaled_p (TImode, offset) | |
4487 | && aarch64_offset_7bit_signed_scaled_p (TImode, | |
4488 | offset + 32)); | |
4489 | ||
4490 | if (load_store_pair_p) | |
43e9d192 | 4491 | return ((GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) |
44707478 | 4492 | && aarch64_offset_7bit_signed_scaled_p (mode, offset)); |
43e9d192 IB |
4493 | else |
4494 | return (offset_9bit_signed_unscaled_p (mode, offset) | |
4495 | || offset_12bit_unsigned_scaled_p (mode, offset)); | |
4496 | } | |
4497 | ||
4498 | if (allow_reg_index_p) | |
4499 | { | |
4500 | /* Look for base + (scaled/extended) index register. */ | |
4501 | if (aarch64_base_register_rtx_p (op0, strict_p) | |
4502 | && aarch64_classify_index (info, op1, mode, strict_p)) | |
4503 | { | |
4504 | info->base = op0; | |
4505 | return true; | |
4506 | } | |
4507 | if (aarch64_base_register_rtx_p (op1, strict_p) | |
4508 | && aarch64_classify_index (info, op0, mode, strict_p)) | |
4509 | { | |
4510 | info->base = op1; | |
4511 | return true; | |
4512 | } | |
4513 | } | |
4514 | ||
4515 | return false; | |
4516 | ||
4517 | case POST_INC: | |
4518 | case POST_DEC: | |
4519 | case PRE_INC: | |
4520 | case PRE_DEC: | |
4521 | info->type = ADDRESS_REG_WB; | |
4522 | info->base = XEXP (x, 0); | |
4523 | info->offset = NULL_RTX; | |
4524 | return aarch64_base_register_rtx_p (info->base, strict_p); | |
4525 | ||
4526 | case POST_MODIFY: | |
4527 | case PRE_MODIFY: | |
4528 | info->type = ADDRESS_REG_WB; | |
4529 | info->base = XEXP (x, 0); | |
4530 | if (GET_CODE (XEXP (x, 1)) == PLUS | |
4531 | && CONST_INT_P (XEXP (XEXP (x, 1), 1)) | |
4532 | && rtx_equal_p (XEXP (XEXP (x, 1), 0), info->base) | |
4533 | && aarch64_base_register_rtx_p (info->base, strict_p)) | |
4534 | { | |
4535 | HOST_WIDE_INT offset; | |
4536 | info->offset = XEXP (XEXP (x, 1), 1); | |
4537 | offset = INTVAL (info->offset); | |
4538 | ||
4539 | /* TImode and TFmode values are allowed in both pairs of X | |
4540 | registers and individual Q registers. The available | |
4541 | address modes are: | |
4542 | X,X: 7-bit signed scaled offset | |
4543 | Q: 9-bit signed offset | |
4544 | We conservatively require an offset representable in either mode. | |
4545 | */ | |
4546 | if (mode == TImode || mode == TFmode) | |
44707478 | 4547 | return (aarch64_offset_7bit_signed_scaled_p (mode, offset) |
43e9d192 IB |
4548 | && offset_9bit_signed_unscaled_p (mode, offset)); |
4549 | ||
2d8c6dc1 | 4550 | if (load_store_pair_p) |
43e9d192 | 4551 | return ((GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) |
44707478 | 4552 | && aarch64_offset_7bit_signed_scaled_p (mode, offset)); |
43e9d192 IB |
4553 | else |
4554 | return offset_9bit_signed_unscaled_p (mode, offset); | |
4555 | } | |
4556 | return false; | |
4557 | ||
4558 | case CONST: | |
4559 | case SYMBOL_REF: | |
4560 | case LABEL_REF: | |
79517551 SN |
4561 | /* load literal: pc-relative constant pool entry. Only supported |
4562 | for SI mode or larger. */ | |
43e9d192 | 4563 | info->type = ADDRESS_SYMBOLIC; |
2d8c6dc1 AH |
4564 | |
4565 | if (!load_store_pair_p && GET_MODE_SIZE (mode) >= 4) | |
43e9d192 IB |
4566 | { |
4567 | rtx sym, addend; | |
4568 | ||
4569 | split_const (x, &sym, &addend); | |
b4f50fd4 RR |
4570 | return ((GET_CODE (sym) == LABEL_REF |
4571 | || (GET_CODE (sym) == SYMBOL_REF | |
4572 | && CONSTANT_POOL_ADDRESS_P (sym) | |
9ee6540a | 4573 | && aarch64_pcrelative_literal_loads))); |
43e9d192 IB |
4574 | } |
4575 | return false; | |
4576 | ||
4577 | case LO_SUM: | |
4578 | info->type = ADDRESS_LO_SUM; | |
4579 | info->base = XEXP (x, 0); | |
4580 | info->offset = XEXP (x, 1); | |
4581 | if (allow_reg_index_p | |
4582 | && aarch64_base_register_rtx_p (info->base, strict_p)) | |
4583 | { | |
4584 | rtx sym, offs; | |
4585 | split_const (info->offset, &sym, &offs); | |
4586 | if (GET_CODE (sym) == SYMBOL_REF | |
a6e0bfa7 | 4587 | && (aarch64_classify_symbol (sym, offs) == SYMBOL_SMALL_ABSOLUTE)) |
43e9d192 IB |
4588 | { |
4589 | /* The symbol and offset must be aligned to the access size. */ | |
4590 | unsigned int align; | |
4591 | unsigned int ref_size; | |
4592 | ||
4593 | if (CONSTANT_POOL_ADDRESS_P (sym)) | |
4594 | align = GET_MODE_ALIGNMENT (get_pool_mode (sym)); | |
4595 | else if (TREE_CONSTANT_POOL_ADDRESS_P (sym)) | |
4596 | { | |
4597 | tree exp = SYMBOL_REF_DECL (sym); | |
4598 | align = TYPE_ALIGN (TREE_TYPE (exp)); | |
4599 | align = CONSTANT_ALIGNMENT (exp, align); | |
4600 | } | |
4601 | else if (SYMBOL_REF_DECL (sym)) | |
4602 | align = DECL_ALIGN (SYMBOL_REF_DECL (sym)); | |
6c031d8d KV |
4603 | else if (SYMBOL_REF_HAS_BLOCK_INFO_P (sym) |
4604 | && SYMBOL_REF_BLOCK (sym) != NULL) | |
4605 | align = SYMBOL_REF_BLOCK (sym)->alignment; | |
43e9d192 IB |
4606 | else |
4607 | align = BITS_PER_UNIT; | |
4608 | ||
4609 | ref_size = GET_MODE_SIZE (mode); | |
4610 | if (ref_size == 0) | |
4611 | ref_size = GET_MODE_SIZE (DImode); | |
4612 | ||
4613 | return ((INTVAL (offs) & (ref_size - 1)) == 0 | |
4614 | && ((align / BITS_PER_UNIT) & (ref_size - 1)) == 0); | |
4615 | } | |
4616 | } | |
4617 | return false; | |
4618 | ||
4619 | default: | |
4620 | return false; | |
4621 | } | |
4622 | } | |
4623 | ||
9bf2f779 KT |
4624 | /* Return true if the address X is valid for a PRFM instruction. |
4625 | STRICT_P is true if we should do strict checking with | |
4626 | aarch64_classify_address. */ | |
4627 | ||
4628 | bool | |
4629 | aarch64_address_valid_for_prefetch_p (rtx x, bool strict_p) | |
4630 | { | |
4631 | struct aarch64_address_info addr; | |
4632 | ||
4633 | /* PRFM accepts the same addresses as DImode... */ | |
4634 | bool res = aarch64_classify_address (&addr, x, DImode, MEM, strict_p); | |
4635 | if (!res) | |
4636 | return false; | |
4637 | ||
4638 | /* ... except writeback forms. */ | |
4639 | return addr.type != ADDRESS_REG_WB; | |
4640 | } | |
4641 | ||
43e9d192 IB |
4642 | bool |
4643 | aarch64_symbolic_address_p (rtx x) | |
4644 | { | |
4645 | rtx offset; | |
4646 | ||
4647 | split_const (x, &x, &offset); | |
4648 | return GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == LABEL_REF; | |
4649 | } | |
4650 | ||
a6e0bfa7 | 4651 | /* Classify the base of symbolic expression X. */ |
da4f13a4 MS |
4652 | |
4653 | enum aarch64_symbol_type | |
a6e0bfa7 | 4654 | aarch64_classify_symbolic_expression (rtx x) |
43e9d192 IB |
4655 | { |
4656 | rtx offset; | |
da4f13a4 | 4657 | |
43e9d192 | 4658 | split_const (x, &x, &offset); |
a6e0bfa7 | 4659 | return aarch64_classify_symbol (x, offset); |
43e9d192 IB |
4660 | } |
4661 | ||
4662 | ||
4663 | /* Return TRUE if X is a legitimate address for accessing memory in | |
4664 | mode MODE. */ | |
4665 | static bool | |
ef4bddc2 | 4666 | aarch64_legitimate_address_hook_p (machine_mode mode, rtx x, bool strict_p) |
43e9d192 IB |
4667 | { |
4668 | struct aarch64_address_info addr; | |
4669 | ||
4670 | return aarch64_classify_address (&addr, x, mode, MEM, strict_p); | |
4671 | } | |
4672 | ||
4673 | /* Return TRUE if X is a legitimate address for accessing memory in | |
4674 | mode MODE. OUTER_CODE will be PARALLEL if this is a load/store | |
4675 | pair operation. */ | |
4676 | bool | |
ef4bddc2 | 4677 | aarch64_legitimate_address_p (machine_mode mode, rtx x, |
aef66c94 | 4678 | RTX_CODE outer_code, bool strict_p) |
43e9d192 IB |
4679 | { |
4680 | struct aarch64_address_info addr; | |
4681 | ||
4682 | return aarch64_classify_address (&addr, x, mode, outer_code, strict_p); | |
4683 | } | |
4684 | ||
491ec060 WD |
4685 | /* Split an out-of-range address displacement into a base and offset. |
4686 | Use 4KB range for 1- and 2-byte accesses and a 16KB range otherwise | |
4687 | to increase opportunities for sharing the base address of different sizes. | |
8734dfac | 4688 | For unaligned accesses and TI/TF mode use the signed 9-bit range. */ |
491ec060 WD |
4689 | static bool |
4690 | aarch64_legitimize_address_displacement (rtx *disp, rtx *off, machine_mode mode) | |
4691 | { | |
8734dfac WD |
4692 | HOST_WIDE_INT offset = INTVAL (*disp); |
4693 | HOST_WIDE_INT base = offset & ~(GET_MODE_SIZE (mode) < 4 ? 0xfff : 0x3ffc); | |
491ec060 | 4694 | |
8734dfac WD |
4695 | if (mode == TImode || mode == TFmode |
4696 | || (offset & (GET_MODE_SIZE (mode) - 1)) != 0) | |
4697 | base = (offset + 0x100) & ~0x1ff; | |
491ec060 | 4698 | |
8734dfac WD |
4699 | *off = GEN_INT (base); |
4700 | *disp = GEN_INT (offset - base); | |
491ec060 WD |
4701 | return true; |
4702 | } | |
4703 | ||
a2170965 TC |
4704 | /* Return the binary representation of floating point constant VALUE in INTVAL. |
4705 | If the value cannot be converted, return false without setting INTVAL. | |
4706 | The conversion is done in the given MODE. */ | |
4707 | bool | |
4708 | aarch64_reinterpret_float_as_int (rtx value, unsigned HOST_WIDE_INT *intval) | |
4709 | { | |
4710 | ||
4711 | /* We make a general exception for 0. */ | |
4712 | if (aarch64_float_const_zero_rtx_p (value)) | |
4713 | { | |
4714 | *intval = 0; | |
4715 | return true; | |
4716 | } | |
4717 | ||
4718 | machine_mode mode = GET_MODE (value); | |
4719 | if (GET_CODE (value) != CONST_DOUBLE | |
4720 | || !SCALAR_FLOAT_MODE_P (mode) | |
4721 | || GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT | |
4722 | /* Only support up to DF mode. */ | |
4723 | || GET_MODE_BITSIZE (mode) > GET_MODE_BITSIZE (DFmode)) | |
4724 | return false; | |
4725 | ||
4726 | unsigned HOST_WIDE_INT ival = 0; | |
4727 | ||
4728 | long res[2]; | |
4729 | real_to_target (res, | |
4730 | CONST_DOUBLE_REAL_VALUE (value), | |
4731 | REAL_MODE_FORMAT (mode)); | |
4732 | ||
5c22bb48 TC |
4733 | if (mode == DFmode) |
4734 | { | |
4735 | int order = BYTES_BIG_ENDIAN ? 1 : 0; | |
4736 | ival = zext_hwi (res[order], 32); | |
4737 | ival |= (zext_hwi (res[1 - order], 32) << 32); | |
4738 | } | |
4739 | else | |
4740 | ival = zext_hwi (res[0], 32); | |
a2170965 TC |
4741 | |
4742 | *intval = ival; | |
4743 | return true; | |
4744 | } | |
4745 | ||
4746 | /* Return TRUE if rtx X is an immediate constant that can be moved using a | |
4747 | single MOV(+MOVK) followed by an FMOV. */ | |
4748 | bool | |
4749 | aarch64_float_const_rtx_p (rtx x) | |
4750 | { | |
4751 | machine_mode mode = GET_MODE (x); | |
4752 | if (mode == VOIDmode) | |
4753 | return false; | |
4754 | ||
4755 | /* Determine whether it's cheaper to write float constants as | |
4756 | mov/movk pairs over ldr/adrp pairs. */ | |
4757 | unsigned HOST_WIDE_INT ival; | |
4758 | ||
4759 | if (GET_CODE (x) == CONST_DOUBLE | |
4760 | && SCALAR_FLOAT_MODE_P (mode) | |
4761 | && aarch64_reinterpret_float_as_int (x, &ival)) | |
4762 | { | |
304b9962 RS |
4763 | machine_mode imode = (mode == HFmode |
4764 | ? SImode | |
4765 | : int_mode_for_mode (mode).require ()); | |
a2170965 TC |
4766 | int num_instr = aarch64_internal_mov_immediate |
4767 | (NULL_RTX, gen_int_mode (ival, imode), false, imode); | |
4768 | return num_instr < 3; | |
4769 | } | |
4770 | ||
4771 | return false; | |
4772 | } | |
4773 | ||
43e9d192 IB |
4774 | /* Return TRUE if rtx X is immediate constant 0.0 */ |
4775 | bool | |
3520f7cc | 4776 | aarch64_float_const_zero_rtx_p (rtx x) |
43e9d192 | 4777 | { |
43e9d192 IB |
4778 | if (GET_MODE (x) == VOIDmode) |
4779 | return false; | |
4780 | ||
34a72c33 | 4781 | if (REAL_VALUE_MINUS_ZERO (*CONST_DOUBLE_REAL_VALUE (x))) |
43e9d192 | 4782 | return !HONOR_SIGNED_ZEROS (GET_MODE (x)); |
34a72c33 | 4783 | return real_equal (CONST_DOUBLE_REAL_VALUE (x), &dconst0); |
43e9d192 IB |
4784 | } |
4785 | ||
a2170965 TC |
4786 | /* Return TRUE if rtx X is immediate constant that fits in a single |
4787 | MOVI immediate operation. */ | |
4788 | bool | |
4789 | aarch64_can_const_movi_rtx_p (rtx x, machine_mode mode) | |
4790 | { | |
4791 | if (!TARGET_SIMD) | |
4792 | return false; | |
4793 | ||
a2170965 TC |
4794 | machine_mode vmode, imode; |
4795 | unsigned HOST_WIDE_INT ival; | |
4796 | ||
4797 | if (GET_CODE (x) == CONST_DOUBLE | |
4798 | && SCALAR_FLOAT_MODE_P (mode)) | |
4799 | { | |
4800 | if (!aarch64_reinterpret_float_as_int (x, &ival)) | |
4801 | return false; | |
4802 | ||
35c38fa6 TC |
4803 | /* We make a general exception for 0. */ |
4804 | if (aarch64_float_const_zero_rtx_p (x)) | |
4805 | return true; | |
4806 | ||
304b9962 | 4807 | imode = int_mode_for_mode (mode).require (); |
a2170965 TC |
4808 | } |
4809 | else if (GET_CODE (x) == CONST_INT | |
4810 | && SCALAR_INT_MODE_P (mode)) | |
4811 | { | |
4812 | imode = mode; | |
4813 | ival = INTVAL (x); | |
4814 | } | |
4815 | else | |
4816 | return false; | |
4817 | ||
4818 | /* use a 64 bit mode for everything except for DI/DF mode, where we use | |
4819 | a 128 bit vector mode. */ | |
4820 | int width = GET_MODE_BITSIZE (mode) == 64 ? 128 : 64; | |
4821 | ||
4822 | vmode = aarch64_simd_container_mode (imode, width); | |
4823 | rtx v_op = aarch64_simd_gen_const_vector_dup (vmode, ival); | |
4824 | ||
4825 | return aarch64_simd_valid_immediate (v_op, vmode, false, NULL); | |
4826 | } | |
4827 | ||
4828 | ||
70f09188 AP |
4829 | /* Return the fixed registers used for condition codes. */ |
4830 | ||
4831 | static bool | |
4832 | aarch64_fixed_condition_code_regs (unsigned int *p1, unsigned int *p2) | |
4833 | { | |
4834 | *p1 = CC_REGNUM; | |
4835 | *p2 = INVALID_REGNUM; | |
4836 | return true; | |
4837 | } | |
4838 | ||
47210a04 RL |
4839 | /* This function is used by the call expanders of the machine description. |
4840 | RESULT is the register in which the result is returned. It's NULL for | |
4841 | "call" and "sibcall". | |
4842 | MEM is the location of the function call. | |
4843 | SIBCALL indicates whether this function call is normal call or sibling call. | |
4844 | It will generate different pattern accordingly. */ | |
4845 | ||
4846 | void | |
4847 | aarch64_expand_call (rtx result, rtx mem, bool sibcall) | |
4848 | { | |
4849 | rtx call, callee, tmp; | |
4850 | rtvec vec; | |
4851 | machine_mode mode; | |
4852 | ||
4853 | gcc_assert (MEM_P (mem)); | |
4854 | callee = XEXP (mem, 0); | |
4855 | mode = GET_MODE (callee); | |
4856 | gcc_assert (mode == Pmode); | |
4857 | ||
4858 | /* Decide if we should generate indirect calls by loading the | |
4859 | address of the callee into a register before performing | |
4860 | the branch-and-link. */ | |
4861 | if (SYMBOL_REF_P (callee) | |
4862 | ? (aarch64_is_long_call_p (callee) | |
4863 | || aarch64_is_noplt_call_p (callee)) | |
4864 | : !REG_P (callee)) | |
4865 | XEXP (mem, 0) = force_reg (mode, callee); | |
4866 | ||
4867 | call = gen_rtx_CALL (VOIDmode, mem, const0_rtx); | |
4868 | ||
4869 | if (result != NULL_RTX) | |
4870 | call = gen_rtx_SET (result, call); | |
4871 | ||
4872 | if (sibcall) | |
4873 | tmp = ret_rtx; | |
4874 | else | |
4875 | tmp = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (Pmode, LR_REGNUM)); | |
4876 | ||
4877 | vec = gen_rtvec (2, call, tmp); | |
4878 | call = gen_rtx_PARALLEL (VOIDmode, vec); | |
4879 | ||
4880 | aarch64_emit_call_insn (call); | |
4881 | } | |
4882 | ||
78607708 TV |
4883 | /* Emit call insn with PAT and do aarch64-specific handling. */ |
4884 | ||
d07a3fed | 4885 | void |
78607708 TV |
4886 | aarch64_emit_call_insn (rtx pat) |
4887 | { | |
4888 | rtx insn = emit_call_insn (pat); | |
4889 | ||
4890 | rtx *fusage = &CALL_INSN_FUNCTION_USAGE (insn); | |
4891 | clobber_reg (fusage, gen_rtx_REG (word_mode, IP0_REGNUM)); | |
4892 | clobber_reg (fusage, gen_rtx_REG (word_mode, IP1_REGNUM)); | |
4893 | } | |
4894 | ||
ef4bddc2 | 4895 | machine_mode |
43e9d192 IB |
4896 | aarch64_select_cc_mode (RTX_CODE code, rtx x, rtx y) |
4897 | { | |
4898 | /* All floating point compares return CCFP if it is an equality | |
4899 | comparison, and CCFPE otherwise. */ | |
4900 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
4901 | { | |
4902 | switch (code) | |
4903 | { | |
4904 | case EQ: | |
4905 | case NE: | |
4906 | case UNORDERED: | |
4907 | case ORDERED: | |
4908 | case UNLT: | |
4909 | case UNLE: | |
4910 | case UNGT: | |
4911 | case UNGE: | |
4912 | case UNEQ: | |
4913 | case LTGT: | |
4914 | return CCFPmode; | |
4915 | ||
4916 | case LT: | |
4917 | case LE: | |
4918 | case GT: | |
4919 | case GE: | |
4920 | return CCFPEmode; | |
4921 | ||
4922 | default: | |
4923 | gcc_unreachable (); | |
4924 | } | |
4925 | } | |
4926 | ||
2b8568fe KT |
4927 | /* Equality comparisons of short modes against zero can be performed |
4928 | using the TST instruction with the appropriate bitmask. */ | |
4929 | if (y == const0_rtx && REG_P (x) | |
4930 | && (code == EQ || code == NE) | |
4931 | && (GET_MODE (x) == HImode || GET_MODE (x) == QImode)) | |
4932 | return CC_NZmode; | |
4933 | ||
b06335f9 KT |
4934 | /* Similarly, comparisons of zero_extends from shorter modes can |
4935 | be performed using an ANDS with an immediate mask. */ | |
4936 | if (y == const0_rtx && GET_CODE (x) == ZERO_EXTEND | |
4937 | && (GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
4938 | && (GET_MODE (XEXP (x, 0)) == HImode || GET_MODE (XEXP (x, 0)) == QImode) | |
4939 | && (code == EQ || code == NE)) | |
4940 | return CC_NZmode; | |
4941 | ||
43e9d192 IB |
4942 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) |
4943 | && y == const0_rtx | |
4944 | && (code == EQ || code == NE || code == LT || code == GE) | |
b056c910 | 4945 | && (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS || GET_CODE (x) == AND |
7325d85a KT |
4946 | || GET_CODE (x) == NEG |
4947 | || (GET_CODE (x) == ZERO_EXTRACT && CONST_INT_P (XEXP (x, 1)) | |
4948 | && CONST_INT_P (XEXP (x, 2))))) | |
43e9d192 IB |
4949 | return CC_NZmode; |
4950 | ||
1c992d1e | 4951 | /* A compare with a shifted operand. Because of canonicalization, |
43e9d192 IB |
4952 | the comparison will have to be swapped when we emit the assembly |
4953 | code. */ | |
4954 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
ffa8a921 | 4955 | && (REG_P (y) || GET_CODE (y) == SUBREG || y == const0_rtx) |
43e9d192 IB |
4956 | && (GET_CODE (x) == ASHIFT || GET_CODE (x) == ASHIFTRT |
4957 | || GET_CODE (x) == LSHIFTRT | |
1c992d1e | 4958 | || GET_CODE (x) == ZERO_EXTEND || GET_CODE (x) == SIGN_EXTEND)) |
43e9d192 IB |
4959 | return CC_SWPmode; |
4960 | ||
1c992d1e RE |
4961 | /* Similarly for a negated operand, but we can only do this for |
4962 | equalities. */ | |
4963 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
4aa81c2e | 4964 | && (REG_P (y) || GET_CODE (y) == SUBREG) |
1c992d1e RE |
4965 | && (code == EQ || code == NE) |
4966 | && GET_CODE (x) == NEG) | |
4967 | return CC_Zmode; | |
4968 | ||
ef22810a RH |
4969 | /* A test for unsigned overflow. */ |
4970 | if ((GET_MODE (x) == DImode || GET_MODE (x) == TImode) | |
4971 | && code == NE | |
4972 | && GET_CODE (x) == PLUS | |
4973 | && GET_CODE (y) == ZERO_EXTEND) | |
4974 | return CC_Cmode; | |
4975 | ||
43e9d192 IB |
4976 | /* For everything else, return CCmode. */ |
4977 | return CCmode; | |
4978 | } | |
4979 | ||
3dfa7055 | 4980 | static int |
b8506a8a | 4981 | aarch64_get_condition_code_1 (machine_mode, enum rtx_code); |
3dfa7055 | 4982 | |
cd5660ab | 4983 | int |
43e9d192 IB |
4984 | aarch64_get_condition_code (rtx x) |
4985 | { | |
ef4bddc2 | 4986 | machine_mode mode = GET_MODE (XEXP (x, 0)); |
43e9d192 IB |
4987 | enum rtx_code comp_code = GET_CODE (x); |
4988 | ||
4989 | if (GET_MODE_CLASS (mode) != MODE_CC) | |
4990 | mode = SELECT_CC_MODE (comp_code, XEXP (x, 0), XEXP (x, 1)); | |
3dfa7055 ZC |
4991 | return aarch64_get_condition_code_1 (mode, comp_code); |
4992 | } | |
43e9d192 | 4993 | |
3dfa7055 | 4994 | static int |
b8506a8a | 4995 | aarch64_get_condition_code_1 (machine_mode mode, enum rtx_code comp_code) |
3dfa7055 | 4996 | { |
43e9d192 IB |
4997 | switch (mode) |
4998 | { | |
4e10a5a7 RS |
4999 | case E_CCFPmode: |
5000 | case E_CCFPEmode: | |
43e9d192 IB |
5001 | switch (comp_code) |
5002 | { | |
5003 | case GE: return AARCH64_GE; | |
5004 | case GT: return AARCH64_GT; | |
5005 | case LE: return AARCH64_LS; | |
5006 | case LT: return AARCH64_MI; | |
5007 | case NE: return AARCH64_NE; | |
5008 | case EQ: return AARCH64_EQ; | |
5009 | case ORDERED: return AARCH64_VC; | |
5010 | case UNORDERED: return AARCH64_VS; | |
5011 | case UNLT: return AARCH64_LT; | |
5012 | case UNLE: return AARCH64_LE; | |
5013 | case UNGT: return AARCH64_HI; | |
5014 | case UNGE: return AARCH64_PL; | |
cd5660ab | 5015 | default: return -1; |
43e9d192 IB |
5016 | } |
5017 | break; | |
5018 | ||
4e10a5a7 | 5019 | case E_CCmode: |
43e9d192 IB |
5020 | switch (comp_code) |
5021 | { | |
5022 | case NE: return AARCH64_NE; | |
5023 | case EQ: return AARCH64_EQ; | |
5024 | case GE: return AARCH64_GE; | |
5025 | case GT: return AARCH64_GT; | |
5026 | case LE: return AARCH64_LE; | |
5027 | case LT: return AARCH64_LT; | |
5028 | case GEU: return AARCH64_CS; | |
5029 | case GTU: return AARCH64_HI; | |
5030 | case LEU: return AARCH64_LS; | |
5031 | case LTU: return AARCH64_CC; | |
cd5660ab | 5032 | default: return -1; |
43e9d192 IB |
5033 | } |
5034 | break; | |
5035 | ||
4e10a5a7 | 5036 | case E_CC_SWPmode: |
43e9d192 IB |
5037 | switch (comp_code) |
5038 | { | |
5039 | case NE: return AARCH64_NE; | |
5040 | case EQ: return AARCH64_EQ; | |
5041 | case GE: return AARCH64_LE; | |
5042 | case GT: return AARCH64_LT; | |
5043 | case LE: return AARCH64_GE; | |
5044 | case LT: return AARCH64_GT; | |
5045 | case GEU: return AARCH64_LS; | |
5046 | case GTU: return AARCH64_CC; | |
5047 | case LEU: return AARCH64_CS; | |
5048 | case LTU: return AARCH64_HI; | |
cd5660ab | 5049 | default: return -1; |
43e9d192 IB |
5050 | } |
5051 | break; | |
5052 | ||
4e10a5a7 | 5053 | case E_CC_NZmode: |
43e9d192 IB |
5054 | switch (comp_code) |
5055 | { | |
5056 | case NE: return AARCH64_NE; | |
5057 | case EQ: return AARCH64_EQ; | |
5058 | case GE: return AARCH64_PL; | |
5059 | case LT: return AARCH64_MI; | |
cd5660ab | 5060 | default: return -1; |
43e9d192 IB |
5061 | } |
5062 | break; | |
5063 | ||
4e10a5a7 | 5064 | case E_CC_Zmode: |
1c992d1e RE |
5065 | switch (comp_code) |
5066 | { | |
5067 | case NE: return AARCH64_NE; | |
5068 | case EQ: return AARCH64_EQ; | |
cd5660ab | 5069 | default: return -1; |
1c992d1e RE |
5070 | } |
5071 | break; | |
5072 | ||
4e10a5a7 | 5073 | case E_CC_Cmode: |
ef22810a RH |
5074 | switch (comp_code) |
5075 | { | |
5076 | case NE: return AARCH64_CS; | |
5077 | case EQ: return AARCH64_CC; | |
5078 | default: return -1; | |
5079 | } | |
5080 | break; | |
5081 | ||
43e9d192 | 5082 | default: |
cd5660ab | 5083 | return -1; |
43e9d192 | 5084 | } |
3dfa7055 | 5085 | |
3dfa7055 | 5086 | return -1; |
43e9d192 IB |
5087 | } |
5088 | ||
ddeabd3e AL |
5089 | bool |
5090 | aarch64_const_vec_all_same_in_range_p (rtx x, | |
5091 | HOST_WIDE_INT minval, | |
5092 | HOST_WIDE_INT maxval) | |
5093 | { | |
5094 | HOST_WIDE_INT firstval; | |
5095 | int count, i; | |
5096 | ||
5097 | if (GET_CODE (x) != CONST_VECTOR | |
5098 | || GET_MODE_CLASS (GET_MODE (x)) != MODE_VECTOR_INT) | |
5099 | return false; | |
5100 | ||
5101 | firstval = INTVAL (CONST_VECTOR_ELT (x, 0)); | |
5102 | if (firstval < minval || firstval > maxval) | |
5103 | return false; | |
5104 | ||
5105 | count = CONST_VECTOR_NUNITS (x); | |
5106 | for (i = 1; i < count; i++) | |
5107 | if (INTVAL (CONST_VECTOR_ELT (x, i)) != firstval) | |
5108 | return false; | |
5109 | ||
5110 | return true; | |
5111 | } | |
5112 | ||
5113 | bool | |
5114 | aarch64_const_vec_all_same_int_p (rtx x, HOST_WIDE_INT val) | |
5115 | { | |
5116 | return aarch64_const_vec_all_same_in_range_p (x, val, val); | |
5117 | } | |
5118 | ||
43e9d192 | 5119 | |
cf670503 ZC |
5120 | /* N Z C V. */ |
5121 | #define AARCH64_CC_V 1 | |
5122 | #define AARCH64_CC_C (1 << 1) | |
5123 | #define AARCH64_CC_Z (1 << 2) | |
5124 | #define AARCH64_CC_N (1 << 3) | |
5125 | ||
c8012fbc WD |
5126 | /* N Z C V flags for ccmp. Indexed by AARCH64_COND_CODE. */ |
5127 | static const int aarch64_nzcv_codes[] = | |
5128 | { | |
5129 | 0, /* EQ, Z == 1. */ | |
5130 | AARCH64_CC_Z, /* NE, Z == 0. */ | |
5131 | 0, /* CS, C == 1. */ | |
5132 | AARCH64_CC_C, /* CC, C == 0. */ | |
5133 | 0, /* MI, N == 1. */ | |
5134 | AARCH64_CC_N, /* PL, N == 0. */ | |
5135 | 0, /* VS, V == 1. */ | |
5136 | AARCH64_CC_V, /* VC, V == 0. */ | |
5137 | 0, /* HI, C ==1 && Z == 0. */ | |
5138 | AARCH64_CC_C, /* LS, !(C == 1 && Z == 0). */ | |
5139 | AARCH64_CC_V, /* GE, N == V. */ | |
5140 | 0, /* LT, N != V. */ | |
5141 | AARCH64_CC_Z, /* GT, Z == 0 && N == V. */ | |
5142 | 0, /* LE, !(Z == 0 && N == V). */ | |
5143 | 0, /* AL, Any. */ | |
5144 | 0 /* NV, Any. */ | |
cf670503 ZC |
5145 | }; |
5146 | ||
bcf19844 JW |
5147 | /* Print operand X to file F in a target specific manner according to CODE. |
5148 | The acceptable formatting commands given by CODE are: | |
5149 | 'c': An integer or symbol address without a preceding # | |
5150 | sign. | |
5151 | 'e': Print the sign/zero-extend size as a character 8->b, | |
5152 | 16->h, 32->w. | |
5153 | 'p': Prints N such that 2^N == X (X must be power of 2 and | |
5154 | const int). | |
5155 | 'P': Print the number of non-zero bits in X (a const_int). | |
5156 | 'H': Print the higher numbered register of a pair (TImode) | |
5157 | of regs. | |
5158 | 'm': Print a condition (eq, ne, etc). | |
5159 | 'M': Same as 'm', but invert condition. | |
5160 | 'b/h/s/d/q': Print a scalar FP/SIMD register name. | |
5161 | 'S/T/U/V': Print a FP/SIMD register name for a register list. | |
5162 | The register printed is the FP/SIMD register name | |
5163 | of X + 0/1/2/3 for S/T/U/V. | |
5164 | 'R': Print a scalar FP/SIMD register name + 1. | |
5165 | 'X': Print bottom 16 bits of integer constant in hex. | |
5166 | 'w/x': Print a general register name or the zero register | |
5167 | (32-bit or 64-bit). | |
5168 | '0': Print a normal operand, if it's a general register, | |
5169 | then we assume DImode. | |
5170 | 'k': Print NZCV for conditional compare instructions. | |
5171 | 'A': Output address constant representing the first | |
5172 | argument of X, specifying a relocation offset | |
5173 | if appropriate. | |
5174 | 'L': Output constant address specified by X | |
5175 | with a relocation offset if appropriate. | |
5176 | 'G': Prints address of X, specifying a PC relative | |
5177 | relocation mode if appropriate. */ | |
5178 | ||
cc8ca59e JB |
5179 | static void |
5180 | aarch64_print_operand (FILE *f, rtx x, int code) | |
43e9d192 IB |
5181 | { |
5182 | switch (code) | |
5183 | { | |
f541a481 KT |
5184 | case 'c': |
5185 | switch (GET_CODE (x)) | |
5186 | { | |
5187 | case CONST_INT: | |
5188 | fprintf (f, HOST_WIDE_INT_PRINT_DEC, INTVAL (x)); | |
5189 | break; | |
5190 | ||
5191 | case SYMBOL_REF: | |
5192 | output_addr_const (f, x); | |
5193 | break; | |
5194 | ||
5195 | case CONST: | |
5196 | if (GET_CODE (XEXP (x, 0)) == PLUS | |
5197 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF) | |
5198 | { | |
5199 | output_addr_const (f, x); | |
5200 | break; | |
5201 | } | |
5202 | /* Fall through. */ | |
5203 | ||
5204 | default: | |
5205 | output_operand_lossage ("Unsupported operand for code '%c'", code); | |
5206 | } | |
5207 | break; | |
5208 | ||
43e9d192 | 5209 | case 'e': |
43e9d192 IB |
5210 | { |
5211 | int n; | |
5212 | ||
4aa81c2e | 5213 | if (!CONST_INT_P (x) |
43e9d192 IB |
5214 | || (n = exact_log2 (INTVAL (x) & ~7)) <= 0) |
5215 | { | |
5216 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
5217 | return; | |
5218 | } | |
5219 | ||
5220 | switch (n) | |
5221 | { | |
5222 | case 3: | |
5223 | fputc ('b', f); | |
5224 | break; | |
5225 | case 4: | |
5226 | fputc ('h', f); | |
5227 | break; | |
5228 | case 5: | |
5229 | fputc ('w', f); | |
5230 | break; | |
5231 | default: | |
5232 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
5233 | return; | |
5234 | } | |
5235 | } | |
5236 | break; | |
5237 | ||
5238 | case 'p': | |
5239 | { | |
5240 | int n; | |
5241 | ||
4aa81c2e | 5242 | if (!CONST_INT_P (x) || (n = exact_log2 (INTVAL (x))) < 0) |
43e9d192 IB |
5243 | { |
5244 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
5245 | return; | |
5246 | } | |
5247 | ||
5248 | asm_fprintf (f, "%d", n); | |
5249 | } | |
5250 | break; | |
5251 | ||
5252 | case 'P': | |
4aa81c2e | 5253 | if (!CONST_INT_P (x)) |
43e9d192 IB |
5254 | { |
5255 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
5256 | return; | |
5257 | } | |
5258 | ||
8d55c61b | 5259 | asm_fprintf (f, "%u", popcount_hwi (INTVAL (x))); |
43e9d192 IB |
5260 | break; |
5261 | ||
5262 | case 'H': | |
4aa81c2e | 5263 | if (!REG_P (x) || !GP_REGNUM_P (REGNO (x) + 1)) |
43e9d192 IB |
5264 | { |
5265 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
5266 | return; | |
5267 | } | |
5268 | ||
01a3a324 | 5269 | asm_fprintf (f, "%s", reg_names [REGNO (x) + 1]); |
43e9d192 IB |
5270 | break; |
5271 | ||
43e9d192 | 5272 | case 'M': |
c8012fbc | 5273 | case 'm': |
cd5660ab KT |
5274 | { |
5275 | int cond_code; | |
c8012fbc WD |
5276 | /* CONST_TRUE_RTX means al/nv (al is the default, don't print it). */ |
5277 | if (x == const_true_rtx) | |
cd5660ab | 5278 | { |
c8012fbc WD |
5279 | if (code == 'M') |
5280 | fputs ("nv", f); | |
cd5660ab KT |
5281 | return; |
5282 | } | |
43e9d192 | 5283 | |
cd5660ab KT |
5284 | if (!COMPARISON_P (x)) |
5285 | { | |
5286 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
5287 | return; | |
5288 | } | |
c8012fbc | 5289 | |
cd5660ab KT |
5290 | cond_code = aarch64_get_condition_code (x); |
5291 | gcc_assert (cond_code >= 0); | |
c8012fbc WD |
5292 | if (code == 'M') |
5293 | cond_code = AARCH64_INVERSE_CONDITION_CODE (cond_code); | |
5294 | fputs (aarch64_condition_codes[cond_code], f); | |
cd5660ab | 5295 | } |
43e9d192 IB |
5296 | break; |
5297 | ||
5298 | case 'b': | |
5299 | case 'h': | |
5300 | case 's': | |
5301 | case 'd': | |
5302 | case 'q': | |
43e9d192 IB |
5303 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) |
5304 | { | |
5305 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
5306 | return; | |
5307 | } | |
50ce6f88 | 5308 | asm_fprintf (f, "%c%d", code, REGNO (x) - V0_REGNUM); |
43e9d192 IB |
5309 | break; |
5310 | ||
5311 | case 'S': | |
5312 | case 'T': | |
5313 | case 'U': | |
5314 | case 'V': | |
43e9d192 IB |
5315 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) |
5316 | { | |
5317 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
5318 | return; | |
5319 | } | |
50ce6f88 | 5320 | asm_fprintf (f, "v%d", REGNO (x) - V0_REGNUM + (code - 'S')); |
43e9d192 IB |
5321 | break; |
5322 | ||
2d8c6dc1 | 5323 | case 'R': |
2d8c6dc1 AH |
5324 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) |
5325 | { | |
5326 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
5327 | return; | |
5328 | } | |
5329 | asm_fprintf (f, "q%d", REGNO (x) - V0_REGNUM + 1); | |
5330 | break; | |
5331 | ||
a05c0ddf | 5332 | case 'X': |
4aa81c2e | 5333 | if (!CONST_INT_P (x)) |
a05c0ddf IB |
5334 | { |
5335 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
5336 | return; | |
5337 | } | |
50d38551 | 5338 | asm_fprintf (f, "0x%wx", UINTVAL (x) & 0xffff); |
a05c0ddf IB |
5339 | break; |
5340 | ||
43e9d192 IB |
5341 | case 'w': |
5342 | case 'x': | |
3520f7cc JG |
5343 | if (x == const0_rtx |
5344 | || (CONST_DOUBLE_P (x) && aarch64_float_const_zero_rtx_p (x))) | |
43e9d192 | 5345 | { |
50ce6f88 | 5346 | asm_fprintf (f, "%czr", code); |
43e9d192 IB |
5347 | break; |
5348 | } | |
5349 | ||
5350 | if (REG_P (x) && GP_REGNUM_P (REGNO (x))) | |
5351 | { | |
50ce6f88 | 5352 | asm_fprintf (f, "%c%d", code, REGNO (x) - R0_REGNUM); |
43e9d192 IB |
5353 | break; |
5354 | } | |
5355 | ||
5356 | if (REG_P (x) && REGNO (x) == SP_REGNUM) | |
5357 | { | |
50ce6f88 | 5358 | asm_fprintf (f, "%ssp", code == 'w' ? "w" : ""); |
43e9d192 IB |
5359 | break; |
5360 | } | |
5361 | ||
5362 | /* Fall through */ | |
5363 | ||
5364 | case 0: | |
43e9d192 IB |
5365 | if (x == NULL) |
5366 | { | |
5367 | output_operand_lossage ("missing operand"); | |
5368 | return; | |
5369 | } | |
5370 | ||
5371 | switch (GET_CODE (x)) | |
5372 | { | |
5373 | case REG: | |
01a3a324 | 5374 | asm_fprintf (f, "%s", reg_names [REGNO (x)]); |
43e9d192 IB |
5375 | break; |
5376 | ||
5377 | case MEM: | |
cc8ca59e | 5378 | output_address (GET_MODE (x), XEXP (x, 0)); |
00eee3fa WD |
5379 | /* Check all memory references are Pmode - even with ILP32. */ |
5380 | gcc_assert (GET_MODE (XEXP (x, 0)) == Pmode); | |
43e9d192 IB |
5381 | break; |
5382 | ||
2af16a7c | 5383 | case CONST: |
43e9d192 IB |
5384 | case LABEL_REF: |
5385 | case SYMBOL_REF: | |
5386 | output_addr_const (asm_out_file, x); | |
5387 | break; | |
5388 | ||
5389 | case CONST_INT: | |
5390 | asm_fprintf (f, "%wd", INTVAL (x)); | |
5391 | break; | |
5392 | ||
5393 | case CONST_VECTOR: | |
3520f7cc JG |
5394 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_VECTOR_INT) |
5395 | { | |
ddeabd3e AL |
5396 | gcc_assert ( |
5397 | aarch64_const_vec_all_same_in_range_p (x, | |
5398 | HOST_WIDE_INT_MIN, | |
5399 | HOST_WIDE_INT_MAX)); | |
3520f7cc JG |
5400 | asm_fprintf (f, "%wd", INTVAL (CONST_VECTOR_ELT (x, 0))); |
5401 | } | |
5402 | else if (aarch64_simd_imm_zero_p (x, GET_MODE (x))) | |
5403 | { | |
5404 | fputc ('0', f); | |
5405 | } | |
5406 | else | |
5407 | gcc_unreachable (); | |
43e9d192 IB |
5408 | break; |
5409 | ||
3520f7cc | 5410 | case CONST_DOUBLE: |
2ca5b430 KT |
5411 | /* Since we define TARGET_SUPPORTS_WIDE_INT we shouldn't ever |
5412 | be getting CONST_DOUBLEs holding integers. */ | |
5413 | gcc_assert (GET_MODE (x) != VOIDmode); | |
5414 | if (aarch64_float_const_zero_rtx_p (x)) | |
3520f7cc JG |
5415 | { |
5416 | fputc ('0', f); | |
5417 | break; | |
5418 | } | |
5419 | else if (aarch64_float_const_representable_p (x)) | |
5420 | { | |
5421 | #define buf_size 20 | |
5422 | char float_buf[buf_size] = {'\0'}; | |
34a72c33 RS |
5423 | real_to_decimal_for_mode (float_buf, |
5424 | CONST_DOUBLE_REAL_VALUE (x), | |
3520f7cc JG |
5425 | buf_size, buf_size, |
5426 | 1, GET_MODE (x)); | |
5427 | asm_fprintf (asm_out_file, "%s", float_buf); | |
5428 | break; | |
5429 | #undef buf_size | |
5430 | } | |
5431 | output_operand_lossage ("invalid constant"); | |
5432 | return; | |
43e9d192 IB |
5433 | default: |
5434 | output_operand_lossage ("invalid operand"); | |
5435 | return; | |
5436 | } | |
5437 | break; | |
5438 | ||
5439 | case 'A': | |
5440 | if (GET_CODE (x) == HIGH) | |
5441 | x = XEXP (x, 0); | |
5442 | ||
a6e0bfa7 | 5443 | switch (aarch64_classify_symbolic_expression (x)) |
43e9d192 | 5444 | { |
6642bdb4 | 5445 | case SYMBOL_SMALL_GOT_4G: |
43e9d192 IB |
5446 | asm_fprintf (asm_out_file, ":got:"); |
5447 | break; | |
5448 | ||
5449 | case SYMBOL_SMALL_TLSGD: | |
5450 | asm_fprintf (asm_out_file, ":tlsgd:"); | |
5451 | break; | |
5452 | ||
5453 | case SYMBOL_SMALL_TLSDESC: | |
5454 | asm_fprintf (asm_out_file, ":tlsdesc:"); | |
5455 | break; | |
5456 | ||
79496620 | 5457 | case SYMBOL_SMALL_TLSIE: |
43e9d192 IB |
5458 | asm_fprintf (asm_out_file, ":gottprel:"); |
5459 | break; | |
5460 | ||
d18ba284 | 5461 | case SYMBOL_TLSLE24: |
43e9d192 IB |
5462 | asm_fprintf (asm_out_file, ":tprel:"); |
5463 | break; | |
5464 | ||
87dd8ab0 MS |
5465 | case SYMBOL_TINY_GOT: |
5466 | gcc_unreachable (); | |
5467 | break; | |
5468 | ||
43e9d192 IB |
5469 | default: |
5470 | break; | |
5471 | } | |
5472 | output_addr_const (asm_out_file, x); | |
5473 | break; | |
5474 | ||
5475 | case 'L': | |
a6e0bfa7 | 5476 | switch (aarch64_classify_symbolic_expression (x)) |
43e9d192 | 5477 | { |
6642bdb4 | 5478 | case SYMBOL_SMALL_GOT_4G: |
43e9d192 IB |
5479 | asm_fprintf (asm_out_file, ":lo12:"); |
5480 | break; | |
5481 | ||
5482 | case SYMBOL_SMALL_TLSGD: | |
5483 | asm_fprintf (asm_out_file, ":tlsgd_lo12:"); | |
5484 | break; | |
5485 | ||
5486 | case SYMBOL_SMALL_TLSDESC: | |
5487 | asm_fprintf (asm_out_file, ":tlsdesc_lo12:"); | |
5488 | break; | |
5489 | ||
79496620 | 5490 | case SYMBOL_SMALL_TLSIE: |
43e9d192 IB |
5491 | asm_fprintf (asm_out_file, ":gottprel_lo12:"); |
5492 | break; | |
5493 | ||
cbf5629e JW |
5494 | case SYMBOL_TLSLE12: |
5495 | asm_fprintf (asm_out_file, ":tprel_lo12:"); | |
5496 | break; | |
5497 | ||
d18ba284 | 5498 | case SYMBOL_TLSLE24: |
43e9d192 IB |
5499 | asm_fprintf (asm_out_file, ":tprel_lo12_nc:"); |
5500 | break; | |
5501 | ||
87dd8ab0 MS |
5502 | case SYMBOL_TINY_GOT: |
5503 | asm_fprintf (asm_out_file, ":got:"); | |
5504 | break; | |
5505 | ||
5ae7caad JW |
5506 | case SYMBOL_TINY_TLSIE: |
5507 | asm_fprintf (asm_out_file, ":gottprel:"); | |
5508 | break; | |
5509 | ||
43e9d192 IB |
5510 | default: |
5511 | break; | |
5512 | } | |
5513 | output_addr_const (asm_out_file, x); | |
5514 | break; | |
5515 | ||
5516 | case 'G': | |
a6e0bfa7 | 5517 | switch (aarch64_classify_symbolic_expression (x)) |
43e9d192 | 5518 | { |
d18ba284 | 5519 | case SYMBOL_TLSLE24: |
43e9d192 IB |
5520 | asm_fprintf (asm_out_file, ":tprel_hi12:"); |
5521 | break; | |
5522 | default: | |
5523 | break; | |
5524 | } | |
5525 | output_addr_const (asm_out_file, x); | |
5526 | break; | |
5527 | ||
cf670503 ZC |
5528 | case 'k': |
5529 | { | |
c8012fbc | 5530 | HOST_WIDE_INT cond_code; |
cf670503 | 5531 | |
c8012fbc | 5532 | if (!CONST_INT_P (x)) |
cf670503 ZC |
5533 | { |
5534 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
5535 | return; | |
5536 | } | |
5537 | ||
c8012fbc WD |
5538 | cond_code = INTVAL (x); |
5539 | gcc_assert (cond_code >= 0 && cond_code <= AARCH64_NV); | |
5540 | asm_fprintf (f, "%d", aarch64_nzcv_codes[cond_code]); | |
cf670503 ZC |
5541 | } |
5542 | break; | |
5543 | ||
43e9d192 IB |
5544 | default: |
5545 | output_operand_lossage ("invalid operand prefix '%%%c'", code); | |
5546 | return; | |
5547 | } | |
5548 | } | |
5549 | ||
cc8ca59e JB |
5550 | static void |
5551 | aarch64_print_operand_address (FILE *f, machine_mode mode, rtx x) | |
43e9d192 IB |
5552 | { |
5553 | struct aarch64_address_info addr; | |
5554 | ||
cc8ca59e | 5555 | if (aarch64_classify_address (&addr, x, mode, MEM, true)) |
43e9d192 IB |
5556 | switch (addr.type) |
5557 | { | |
5558 | case ADDRESS_REG_IMM: | |
5559 | if (addr.offset == const0_rtx) | |
01a3a324 | 5560 | asm_fprintf (f, "[%s]", reg_names [REGNO (addr.base)]); |
43e9d192 | 5561 | else |
16a3246f | 5562 | asm_fprintf (f, "[%s, %wd]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
5563 | INTVAL (addr.offset)); |
5564 | return; | |
5565 | ||
5566 | case ADDRESS_REG_REG: | |
5567 | if (addr.shift == 0) | |
16a3246f | 5568 | asm_fprintf (f, "[%s, %s]", reg_names [REGNO (addr.base)], |
01a3a324 | 5569 | reg_names [REGNO (addr.offset)]); |
43e9d192 | 5570 | else |
16a3246f | 5571 | asm_fprintf (f, "[%s, %s, lsl %u]", reg_names [REGNO (addr.base)], |
01a3a324 | 5572 | reg_names [REGNO (addr.offset)], addr.shift); |
43e9d192 IB |
5573 | return; |
5574 | ||
5575 | case ADDRESS_REG_UXTW: | |
5576 | if (addr.shift == 0) | |
16a3246f | 5577 | asm_fprintf (f, "[%s, w%d, uxtw]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
5578 | REGNO (addr.offset) - R0_REGNUM); |
5579 | else | |
16a3246f | 5580 | asm_fprintf (f, "[%s, w%d, uxtw %u]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
5581 | REGNO (addr.offset) - R0_REGNUM, addr.shift); |
5582 | return; | |
5583 | ||
5584 | case ADDRESS_REG_SXTW: | |
5585 | if (addr.shift == 0) | |
16a3246f | 5586 | asm_fprintf (f, "[%s, w%d, sxtw]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
5587 | REGNO (addr.offset) - R0_REGNUM); |
5588 | else | |
16a3246f | 5589 | asm_fprintf (f, "[%s, w%d, sxtw %u]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
5590 | REGNO (addr.offset) - R0_REGNUM, addr.shift); |
5591 | return; | |
5592 | ||
5593 | case ADDRESS_REG_WB: | |
5594 | switch (GET_CODE (x)) | |
5595 | { | |
5596 | case PRE_INC: | |
16a3246f | 5597 | asm_fprintf (f, "[%s, %d]!", reg_names [REGNO (addr.base)], |
cc8ca59e | 5598 | GET_MODE_SIZE (mode)); |
43e9d192 IB |
5599 | return; |
5600 | case POST_INC: | |
16a3246f | 5601 | asm_fprintf (f, "[%s], %d", reg_names [REGNO (addr.base)], |
cc8ca59e | 5602 | GET_MODE_SIZE (mode)); |
43e9d192 IB |
5603 | return; |
5604 | case PRE_DEC: | |
16a3246f | 5605 | asm_fprintf (f, "[%s, -%d]!", reg_names [REGNO (addr.base)], |
cc8ca59e | 5606 | GET_MODE_SIZE (mode)); |
43e9d192 IB |
5607 | return; |
5608 | case POST_DEC: | |
16a3246f | 5609 | asm_fprintf (f, "[%s], -%d", reg_names [REGNO (addr.base)], |
cc8ca59e | 5610 | GET_MODE_SIZE (mode)); |
43e9d192 IB |
5611 | return; |
5612 | case PRE_MODIFY: | |
16a3246f | 5613 | asm_fprintf (f, "[%s, %wd]!", reg_names [REGNO (addr.base)], |
43e9d192 IB |
5614 | INTVAL (addr.offset)); |
5615 | return; | |
5616 | case POST_MODIFY: | |
16a3246f | 5617 | asm_fprintf (f, "[%s], %wd", reg_names [REGNO (addr.base)], |
43e9d192 IB |
5618 | INTVAL (addr.offset)); |
5619 | return; | |
5620 | default: | |
5621 | break; | |
5622 | } | |
5623 | break; | |
5624 | ||
5625 | case ADDRESS_LO_SUM: | |
16a3246f | 5626 | asm_fprintf (f, "[%s, #:lo12:", reg_names [REGNO (addr.base)]); |
43e9d192 IB |
5627 | output_addr_const (f, addr.offset); |
5628 | asm_fprintf (f, "]"); | |
5629 | return; | |
5630 | ||
5631 | case ADDRESS_SYMBOLIC: | |
5632 | break; | |
5633 | } | |
5634 | ||
5635 | output_addr_const (f, x); | |
5636 | } | |
5637 | ||
43e9d192 IB |
5638 | bool |
5639 | aarch64_label_mentioned_p (rtx x) | |
5640 | { | |
5641 | const char *fmt; | |
5642 | int i; | |
5643 | ||
5644 | if (GET_CODE (x) == LABEL_REF) | |
5645 | return true; | |
5646 | ||
5647 | /* UNSPEC_TLS entries for a symbol include a LABEL_REF for the | |
5648 | referencing instruction, but they are constant offsets, not | |
5649 | symbols. */ | |
5650 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
5651 | return false; | |
5652 | ||
5653 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
5654 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
5655 | { | |
5656 | if (fmt[i] == 'E') | |
5657 | { | |
5658 | int j; | |
5659 | ||
5660 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
5661 | if (aarch64_label_mentioned_p (XVECEXP (x, i, j))) | |
5662 | return 1; | |
5663 | } | |
5664 | else if (fmt[i] == 'e' && aarch64_label_mentioned_p (XEXP (x, i))) | |
5665 | return 1; | |
5666 | } | |
5667 | ||
5668 | return 0; | |
5669 | } | |
5670 | ||
5671 | /* Implement REGNO_REG_CLASS. */ | |
5672 | ||
5673 | enum reg_class | |
5674 | aarch64_regno_regclass (unsigned regno) | |
5675 | { | |
5676 | if (GP_REGNUM_P (regno)) | |
a4a182c6 | 5677 | return GENERAL_REGS; |
43e9d192 IB |
5678 | |
5679 | if (regno == SP_REGNUM) | |
5680 | return STACK_REG; | |
5681 | ||
5682 | if (regno == FRAME_POINTER_REGNUM | |
5683 | || regno == ARG_POINTER_REGNUM) | |
f24bb080 | 5684 | return POINTER_REGS; |
43e9d192 IB |
5685 | |
5686 | if (FP_REGNUM_P (regno)) | |
5687 | return FP_LO_REGNUM_P (regno) ? FP_LO_REGS : FP_REGS; | |
5688 | ||
5689 | return NO_REGS; | |
5690 | } | |
5691 | ||
0c4ec427 | 5692 | static rtx |
ef4bddc2 | 5693 | aarch64_legitimize_address (rtx x, rtx /* orig_x */, machine_mode mode) |
0c4ec427 RE |
5694 | { |
5695 | /* Try to split X+CONST into Y=X+(CONST & ~mask), Y+(CONST&mask), | |
5696 | where mask is selected by alignment and size of the offset. | |
5697 | We try to pick as large a range for the offset as possible to | |
5698 | maximize the chance of a CSE. However, for aligned addresses | |
5699 | we limit the range to 4k so that structures with different sized | |
e8426e0a BC |
5700 | elements are likely to use the same base. We need to be careful |
5701 | not to split a CONST for some forms of address expression, otherwise | |
5702 | it will generate sub-optimal code. */ | |
0c4ec427 RE |
5703 | |
5704 | if (GET_CODE (x) == PLUS && CONST_INT_P (XEXP (x, 1))) | |
5705 | { | |
9e0218fc | 5706 | rtx base = XEXP (x, 0); |
17d7bdd8 | 5707 | rtx offset_rtx = XEXP (x, 1); |
9e0218fc | 5708 | HOST_WIDE_INT offset = INTVAL (offset_rtx); |
0c4ec427 | 5709 | |
9e0218fc | 5710 | if (GET_CODE (base) == PLUS) |
e8426e0a | 5711 | { |
9e0218fc RH |
5712 | rtx op0 = XEXP (base, 0); |
5713 | rtx op1 = XEXP (base, 1); | |
5714 | ||
5715 | /* Force any scaling into a temp for CSE. */ | |
5716 | op0 = force_reg (Pmode, op0); | |
5717 | op1 = force_reg (Pmode, op1); | |
5718 | ||
5719 | /* Let the pointer register be in op0. */ | |
5720 | if (REG_POINTER (op1)) | |
5721 | std::swap (op0, op1); | |
5722 | ||
5723 | /* If the pointer is virtual or frame related, then we know that | |
5724 | virtual register instantiation or register elimination is going | |
5725 | to apply a second constant. We want the two constants folded | |
5726 | together easily. Therefore, emit as (OP0 + CONST) + OP1. */ | |
5727 | if (virt_or_elim_regno_p (REGNO (op0))) | |
e8426e0a | 5728 | { |
9e0218fc RH |
5729 | base = expand_binop (Pmode, add_optab, op0, offset_rtx, |
5730 | NULL_RTX, true, OPTAB_DIRECT); | |
5731 | return gen_rtx_PLUS (Pmode, base, op1); | |
e8426e0a | 5732 | } |
e8426e0a | 5733 | |
9e0218fc RH |
5734 | /* Otherwise, in order to encourage CSE (and thence loop strength |
5735 | reduce) scaled addresses, emit as (OP0 + OP1) + CONST. */ | |
5736 | base = expand_binop (Pmode, add_optab, op0, op1, | |
5737 | NULL_RTX, true, OPTAB_DIRECT); | |
5738 | x = gen_rtx_PLUS (Pmode, base, offset_rtx); | |
e8426e0a BC |
5739 | } |
5740 | ||
8734dfac | 5741 | /* Does it look like we'll need a 16-byte load/store-pair operation? */ |
9e0218fc | 5742 | HOST_WIDE_INT base_offset; |
8734dfac WD |
5743 | if (GET_MODE_SIZE (mode) > 16) |
5744 | base_offset = (offset + 0x400) & ~0x7f0; | |
0c4ec427 RE |
5745 | /* For offsets aren't a multiple of the access size, the limit is |
5746 | -256...255. */ | |
5747 | else if (offset & (GET_MODE_SIZE (mode) - 1)) | |
ff0f3f1c WD |
5748 | { |
5749 | base_offset = (offset + 0x100) & ~0x1ff; | |
5750 | ||
5751 | /* BLKmode typically uses LDP of X-registers. */ | |
5752 | if (mode == BLKmode) | |
5753 | base_offset = (offset + 512) & ~0x3ff; | |
5754 | } | |
5755 | /* Small negative offsets are supported. */ | |
5756 | else if (IN_RANGE (offset, -256, 0)) | |
5757 | base_offset = 0; | |
8734dfac WD |
5758 | else if (mode == TImode || mode == TFmode) |
5759 | base_offset = (offset + 0x100) & ~0x1ff; | |
ff0f3f1c | 5760 | /* Use 12-bit offset by access size. */ |
0c4ec427 | 5761 | else |
ff0f3f1c | 5762 | base_offset = offset & (~0xfff * GET_MODE_SIZE (mode)); |
0c4ec427 | 5763 | |
9e0218fc RH |
5764 | if (base_offset != 0) |
5765 | { | |
5766 | base = plus_constant (Pmode, base, base_offset); | |
5767 | base = force_operand (base, NULL_RTX); | |
5768 | return plus_constant (Pmode, base, offset - base_offset); | |
5769 | } | |
0c4ec427 RE |
5770 | } |
5771 | ||
5772 | return x; | |
5773 | } | |
5774 | ||
b4f50fd4 RR |
5775 | /* Return the reload icode required for a constant pool in mode. */ |
5776 | static enum insn_code | |
5777 | aarch64_constant_pool_reload_icode (machine_mode mode) | |
5778 | { | |
5779 | switch (mode) | |
5780 | { | |
4e10a5a7 | 5781 | case E_SFmode: |
b4f50fd4 RR |
5782 | return CODE_FOR_aarch64_reload_movcpsfdi; |
5783 | ||
4e10a5a7 | 5784 | case E_DFmode: |
b4f50fd4 RR |
5785 | return CODE_FOR_aarch64_reload_movcpdfdi; |
5786 | ||
4e10a5a7 | 5787 | case E_TFmode: |
b4f50fd4 RR |
5788 | return CODE_FOR_aarch64_reload_movcptfdi; |
5789 | ||
4e10a5a7 | 5790 | case E_V8QImode: |
b4f50fd4 RR |
5791 | return CODE_FOR_aarch64_reload_movcpv8qidi; |
5792 | ||
4e10a5a7 | 5793 | case E_V16QImode: |
b4f50fd4 RR |
5794 | return CODE_FOR_aarch64_reload_movcpv16qidi; |
5795 | ||
4e10a5a7 | 5796 | case E_V4HImode: |
b4f50fd4 RR |
5797 | return CODE_FOR_aarch64_reload_movcpv4hidi; |
5798 | ||
4e10a5a7 | 5799 | case E_V8HImode: |
b4f50fd4 RR |
5800 | return CODE_FOR_aarch64_reload_movcpv8hidi; |
5801 | ||
4e10a5a7 | 5802 | case E_V2SImode: |
b4f50fd4 RR |
5803 | return CODE_FOR_aarch64_reload_movcpv2sidi; |
5804 | ||
4e10a5a7 | 5805 | case E_V4SImode: |
b4f50fd4 RR |
5806 | return CODE_FOR_aarch64_reload_movcpv4sidi; |
5807 | ||
4e10a5a7 | 5808 | case E_V2DImode: |
b4f50fd4 RR |
5809 | return CODE_FOR_aarch64_reload_movcpv2didi; |
5810 | ||
4e10a5a7 | 5811 | case E_V2DFmode: |
b4f50fd4 RR |
5812 | return CODE_FOR_aarch64_reload_movcpv2dfdi; |
5813 | ||
5814 | default: | |
5815 | gcc_unreachable (); | |
5816 | } | |
5817 | ||
5818 | gcc_unreachable (); | |
5819 | } | |
43e9d192 IB |
5820 | static reg_class_t |
5821 | aarch64_secondary_reload (bool in_p ATTRIBUTE_UNUSED, rtx x, | |
5822 | reg_class_t rclass, | |
ef4bddc2 | 5823 | machine_mode mode, |
43e9d192 IB |
5824 | secondary_reload_info *sri) |
5825 | { | |
b4f50fd4 RR |
5826 | |
5827 | /* If we have to disable direct literal pool loads and stores because the | |
5828 | function is too big, then we need a scratch register. */ | |
5829 | if (MEM_P (x) && GET_CODE (x) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (x) | |
5830 | && (SCALAR_FLOAT_MODE_P (GET_MODE (x)) | |
5831 | || targetm.vector_mode_supported_p (GET_MODE (x))) | |
9ee6540a | 5832 | && !aarch64_pcrelative_literal_loads) |
b4f50fd4 RR |
5833 | { |
5834 | sri->icode = aarch64_constant_pool_reload_icode (mode); | |
5835 | return NO_REGS; | |
5836 | } | |
5837 | ||
43e9d192 IB |
5838 | /* Without the TARGET_SIMD instructions we cannot move a Q register |
5839 | to a Q register directly. We need a scratch. */ | |
5840 | if (REG_P (x) && (mode == TFmode || mode == TImode) && mode == GET_MODE (x) | |
5841 | && FP_REGNUM_P (REGNO (x)) && !TARGET_SIMD | |
5842 | && reg_class_subset_p (rclass, FP_REGS)) | |
5843 | { | |
5844 | if (mode == TFmode) | |
5845 | sri->icode = CODE_FOR_aarch64_reload_movtf; | |
5846 | else if (mode == TImode) | |
5847 | sri->icode = CODE_FOR_aarch64_reload_movti; | |
5848 | return NO_REGS; | |
5849 | } | |
5850 | ||
5851 | /* A TFmode or TImode memory access should be handled via an FP_REGS | |
5852 | because AArch64 has richer addressing modes for LDR/STR instructions | |
5853 | than LDP/STP instructions. */ | |
d5726973 | 5854 | if (TARGET_FLOAT && rclass == GENERAL_REGS |
43e9d192 IB |
5855 | && GET_MODE_SIZE (mode) == 16 && MEM_P (x)) |
5856 | return FP_REGS; | |
5857 | ||
5858 | if (rclass == FP_REGS && (mode == TImode || mode == TFmode) && CONSTANT_P(x)) | |
a4a182c6 | 5859 | return GENERAL_REGS; |
43e9d192 IB |
5860 | |
5861 | return NO_REGS; | |
5862 | } | |
5863 | ||
5864 | static bool | |
5865 | aarch64_can_eliminate (const int from, const int to) | |
5866 | { | |
5867 | /* If we need a frame pointer, we must eliminate FRAME_POINTER_REGNUM into | |
5868 | HARD_FRAME_POINTER_REGNUM and not into STACK_POINTER_REGNUM. */ | |
5869 | ||
5870 | if (frame_pointer_needed) | |
5871 | { | |
5872 | if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
5873 | return true; | |
5874 | if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) | |
5875 | return false; | |
5876 | if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM | |
5877 | && !cfun->calls_alloca) | |
5878 | return true; | |
5879 | if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
5880 | return true; | |
0b7f8166 MS |
5881 | |
5882 | return false; | |
43e9d192 | 5883 | } |
1c923b60 JW |
5884 | else |
5885 | { | |
5886 | /* If we decided that we didn't need a leaf frame pointer but then used | |
5887 | LR in the function, then we'll want a frame pointer after all, so | |
5888 | prevent this elimination to ensure a frame pointer is used. */ | |
5889 | if (to == STACK_POINTER_REGNUM | |
5890 | && flag_omit_leaf_frame_pointer | |
5891 | && df_regs_ever_live_p (LR_REGNUM)) | |
5892 | return false; | |
5893 | } | |
777e6976 | 5894 | |
43e9d192 IB |
5895 | return true; |
5896 | } | |
5897 | ||
5898 | HOST_WIDE_INT | |
5899 | aarch64_initial_elimination_offset (unsigned from, unsigned to) | |
5900 | { | |
43e9d192 | 5901 | aarch64_layout_frame (); |
78c29983 MS |
5902 | |
5903 | if (to == HARD_FRAME_POINTER_REGNUM) | |
5904 | { | |
5905 | if (from == ARG_POINTER_REGNUM) | |
71bfb77a | 5906 | return cfun->machine->frame.hard_fp_offset; |
78c29983 MS |
5907 | |
5908 | if (from == FRAME_POINTER_REGNUM) | |
71bfb77a WD |
5909 | return cfun->machine->frame.hard_fp_offset |
5910 | - cfun->machine->frame.locals_offset; | |
78c29983 MS |
5911 | } |
5912 | ||
5913 | if (to == STACK_POINTER_REGNUM) | |
5914 | { | |
5915 | if (from == FRAME_POINTER_REGNUM) | |
71bfb77a WD |
5916 | return cfun->machine->frame.frame_size |
5917 | - cfun->machine->frame.locals_offset; | |
78c29983 MS |
5918 | } |
5919 | ||
1c960e02 | 5920 | return cfun->machine->frame.frame_size; |
43e9d192 IB |
5921 | } |
5922 | ||
43e9d192 IB |
5923 | /* Implement RETURN_ADDR_RTX. We do not support moving back to a |
5924 | previous frame. */ | |
5925 | ||
5926 | rtx | |
5927 | aarch64_return_addr (int count, rtx frame ATTRIBUTE_UNUSED) | |
5928 | { | |
5929 | if (count != 0) | |
5930 | return const0_rtx; | |
5931 | return get_hard_reg_initial_val (Pmode, LR_REGNUM); | |
5932 | } | |
5933 | ||
5934 | ||
5935 | static void | |
5936 | aarch64_asm_trampoline_template (FILE *f) | |
5937 | { | |
28514dda YZ |
5938 | if (TARGET_ILP32) |
5939 | { | |
5940 | asm_fprintf (f, "\tldr\tw%d, .+16\n", IP1_REGNUM - R0_REGNUM); | |
5941 | asm_fprintf (f, "\tldr\tw%d, .+16\n", STATIC_CHAIN_REGNUM - R0_REGNUM); | |
5942 | } | |
5943 | else | |
5944 | { | |
5945 | asm_fprintf (f, "\tldr\t%s, .+16\n", reg_names [IP1_REGNUM]); | |
5946 | asm_fprintf (f, "\tldr\t%s, .+20\n", reg_names [STATIC_CHAIN_REGNUM]); | |
5947 | } | |
01a3a324 | 5948 | asm_fprintf (f, "\tbr\t%s\n", reg_names [IP1_REGNUM]); |
43e9d192 | 5949 | assemble_aligned_integer (4, const0_rtx); |
28514dda YZ |
5950 | assemble_aligned_integer (POINTER_BYTES, const0_rtx); |
5951 | assemble_aligned_integer (POINTER_BYTES, const0_rtx); | |
43e9d192 IB |
5952 | } |
5953 | ||
5954 | static void | |
5955 | aarch64_trampoline_init (rtx m_tramp, tree fndecl, rtx chain_value) | |
5956 | { | |
5957 | rtx fnaddr, mem, a_tramp; | |
28514dda | 5958 | const int tramp_code_sz = 16; |
43e9d192 IB |
5959 | |
5960 | /* Don't need to copy the trailing D-words, we fill those in below. */ | |
5961 | emit_block_move (m_tramp, assemble_trampoline_template (), | |
28514dda YZ |
5962 | GEN_INT (tramp_code_sz), BLOCK_OP_NORMAL); |
5963 | mem = adjust_address (m_tramp, ptr_mode, tramp_code_sz); | |
43e9d192 | 5964 | fnaddr = XEXP (DECL_RTL (fndecl), 0); |
28514dda YZ |
5965 | if (GET_MODE (fnaddr) != ptr_mode) |
5966 | fnaddr = convert_memory_address (ptr_mode, fnaddr); | |
43e9d192 IB |
5967 | emit_move_insn (mem, fnaddr); |
5968 | ||
28514dda | 5969 | mem = adjust_address (m_tramp, ptr_mode, tramp_code_sz + POINTER_BYTES); |
43e9d192 IB |
5970 | emit_move_insn (mem, chain_value); |
5971 | ||
5972 | /* XXX We should really define a "clear_cache" pattern and use | |
5973 | gen_clear_cache(). */ | |
5974 | a_tramp = XEXP (m_tramp, 0); | |
5975 | emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__clear_cache"), | |
28514dda YZ |
5976 | LCT_NORMAL, VOIDmode, 2, a_tramp, ptr_mode, |
5977 | plus_constant (ptr_mode, a_tramp, TRAMPOLINE_SIZE), | |
5978 | ptr_mode); | |
43e9d192 IB |
5979 | } |
5980 | ||
5981 | static unsigned char | |
ef4bddc2 | 5982 | aarch64_class_max_nregs (reg_class_t regclass, machine_mode mode) |
43e9d192 IB |
5983 | { |
5984 | switch (regclass) | |
5985 | { | |
fee9ba42 | 5986 | case CALLER_SAVE_REGS: |
43e9d192 IB |
5987 | case POINTER_REGS: |
5988 | case GENERAL_REGS: | |
5989 | case ALL_REGS: | |
5990 | case FP_REGS: | |
5991 | case FP_LO_REGS: | |
5992 | return | |
7bd11911 KT |
5993 | aarch64_vector_mode_p (mode) |
5994 | ? (GET_MODE_SIZE (mode) + UNITS_PER_VREG - 1) / UNITS_PER_VREG | |
5995 | : (GET_MODE_SIZE (mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
43e9d192 IB |
5996 | case STACK_REG: |
5997 | return 1; | |
5998 | ||
5999 | case NO_REGS: | |
6000 | return 0; | |
6001 | ||
6002 | default: | |
6003 | break; | |
6004 | } | |
6005 | gcc_unreachable (); | |
6006 | } | |
6007 | ||
6008 | static reg_class_t | |
78d8b9f0 | 6009 | aarch64_preferred_reload_class (rtx x, reg_class_t regclass) |
43e9d192 | 6010 | { |
51bb310d | 6011 | if (regclass == POINTER_REGS) |
78d8b9f0 IB |
6012 | return GENERAL_REGS; |
6013 | ||
51bb310d MS |
6014 | if (regclass == STACK_REG) |
6015 | { | |
6016 | if (REG_P(x) | |
6017 | && reg_class_subset_p (REGNO_REG_CLASS (REGNO (x)), POINTER_REGS)) | |
6018 | return regclass; | |
6019 | ||
6020 | return NO_REGS; | |
6021 | } | |
6022 | ||
27bd251b IB |
6023 | /* Register eliminiation can result in a request for |
6024 | SP+constant->FP_REGS. We cannot support such operations which | |
6025 | use SP as source and an FP_REG as destination, so reject out | |
6026 | right now. */ | |
6027 | if (! reg_class_subset_p (regclass, GENERAL_REGS) && GET_CODE (x) == PLUS) | |
6028 | { | |
6029 | rtx lhs = XEXP (x, 0); | |
6030 | ||
6031 | /* Look through a possible SUBREG introduced by ILP32. */ | |
6032 | if (GET_CODE (lhs) == SUBREG) | |
6033 | lhs = SUBREG_REG (lhs); | |
6034 | ||
6035 | gcc_assert (REG_P (lhs)); | |
6036 | gcc_assert (reg_class_subset_p (REGNO_REG_CLASS (REGNO (lhs)), | |
6037 | POINTER_REGS)); | |
6038 | return NO_REGS; | |
6039 | } | |
6040 | ||
78d8b9f0 | 6041 | return regclass; |
43e9d192 IB |
6042 | } |
6043 | ||
6044 | void | |
6045 | aarch64_asm_output_labelref (FILE* f, const char *name) | |
6046 | { | |
6047 | asm_fprintf (f, "%U%s", name); | |
6048 | } | |
6049 | ||
6050 | static void | |
6051 | aarch64_elf_asm_constructor (rtx symbol, int priority) | |
6052 | { | |
6053 | if (priority == DEFAULT_INIT_PRIORITY) | |
6054 | default_ctor_section_asm_out_constructor (symbol, priority); | |
6055 | else | |
6056 | { | |
6057 | section *s; | |
53d190c1 AT |
6058 | /* While priority is known to be in range [0, 65535], so 18 bytes |
6059 | would be enough, the compiler might not know that. To avoid | |
6060 | -Wformat-truncation false positive, use a larger size. */ | |
6061 | char buf[23]; | |
43e9d192 IB |
6062 | snprintf (buf, sizeof (buf), ".init_array.%.5u", priority); |
6063 | s = get_section (buf, SECTION_WRITE, NULL); | |
6064 | switch_to_section (s); | |
6065 | assemble_align (POINTER_SIZE); | |
28514dda | 6066 | assemble_aligned_integer (POINTER_BYTES, symbol); |
43e9d192 IB |
6067 | } |
6068 | } | |
6069 | ||
6070 | static void | |
6071 | aarch64_elf_asm_destructor (rtx symbol, int priority) | |
6072 | { | |
6073 | if (priority == DEFAULT_INIT_PRIORITY) | |
6074 | default_dtor_section_asm_out_destructor (symbol, priority); | |
6075 | else | |
6076 | { | |
6077 | section *s; | |
53d190c1 AT |
6078 | /* While priority is known to be in range [0, 65535], so 18 bytes |
6079 | would be enough, the compiler might not know that. To avoid | |
6080 | -Wformat-truncation false positive, use a larger size. */ | |
6081 | char buf[23]; | |
43e9d192 IB |
6082 | snprintf (buf, sizeof (buf), ".fini_array.%.5u", priority); |
6083 | s = get_section (buf, SECTION_WRITE, NULL); | |
6084 | switch_to_section (s); | |
6085 | assemble_align (POINTER_SIZE); | |
28514dda | 6086 | assemble_aligned_integer (POINTER_BYTES, symbol); |
43e9d192 IB |
6087 | } |
6088 | } | |
6089 | ||
6090 | const char* | |
6091 | aarch64_output_casesi (rtx *operands) | |
6092 | { | |
6093 | char buf[100]; | |
6094 | char label[100]; | |
b32d5189 | 6095 | rtx diff_vec = PATTERN (NEXT_INSN (as_a <rtx_insn *> (operands[2]))); |
43e9d192 IB |
6096 | int index; |
6097 | static const char *const patterns[4][2] = | |
6098 | { | |
6099 | { | |
6100 | "ldrb\t%w3, [%0,%w1,uxtw]", | |
6101 | "add\t%3, %4, %w3, sxtb #2" | |
6102 | }, | |
6103 | { | |
6104 | "ldrh\t%w3, [%0,%w1,uxtw #1]", | |
6105 | "add\t%3, %4, %w3, sxth #2" | |
6106 | }, | |
6107 | { | |
6108 | "ldr\t%w3, [%0,%w1,uxtw #2]", | |
6109 | "add\t%3, %4, %w3, sxtw #2" | |
6110 | }, | |
6111 | /* We assume that DImode is only generated when not optimizing and | |
6112 | that we don't really need 64-bit address offsets. That would | |
6113 | imply an object file with 8GB of code in a single function! */ | |
6114 | { | |
6115 | "ldr\t%w3, [%0,%w1,uxtw #2]", | |
6116 | "add\t%3, %4, %w3, sxtw #2" | |
6117 | } | |
6118 | }; | |
6119 | ||
6120 | gcc_assert (GET_CODE (diff_vec) == ADDR_DIFF_VEC); | |
6121 | ||
6122 | index = exact_log2 (GET_MODE_SIZE (GET_MODE (diff_vec))); | |
6123 | ||
6124 | gcc_assert (index >= 0 && index <= 3); | |
6125 | ||
6126 | /* Need to implement table size reduction, by chaning the code below. */ | |
6127 | output_asm_insn (patterns[index][0], operands); | |
6128 | ASM_GENERATE_INTERNAL_LABEL (label, "Lrtx", CODE_LABEL_NUMBER (operands[2])); | |
6129 | snprintf (buf, sizeof (buf), | |
6130 | "adr\t%%4, %s", targetm.strip_name_encoding (label)); | |
6131 | output_asm_insn (buf, operands); | |
6132 | output_asm_insn (patterns[index][1], operands); | |
6133 | output_asm_insn ("br\t%3", operands); | |
6134 | assemble_label (asm_out_file, label); | |
6135 | return ""; | |
6136 | } | |
6137 | ||
6138 | ||
6139 | /* Return size in bits of an arithmetic operand which is shifted/scaled and | |
6140 | masked such that it is suitable for a UXTB, UXTH, or UXTW extend | |
6141 | operator. */ | |
6142 | ||
6143 | int | |
6144 | aarch64_uxt_size (int shift, HOST_WIDE_INT mask) | |
6145 | { | |
6146 | if (shift >= 0 && shift <= 3) | |
6147 | { | |
6148 | int size; | |
6149 | for (size = 8; size <= 32; size *= 2) | |
6150 | { | |
6151 | HOST_WIDE_INT bits = ((HOST_WIDE_INT)1U << size) - 1; | |
6152 | if (mask == bits << shift) | |
6153 | return size; | |
6154 | } | |
6155 | } | |
6156 | return 0; | |
6157 | } | |
6158 | ||
e78d485e RR |
6159 | /* Constant pools are per function only when PC relative |
6160 | literal loads are true or we are in the large memory | |
6161 | model. */ | |
6162 | ||
6163 | static inline bool | |
6164 | aarch64_can_use_per_function_literal_pools_p (void) | |
6165 | { | |
9ee6540a | 6166 | return (aarch64_pcrelative_literal_loads |
e78d485e RR |
6167 | || aarch64_cmodel == AARCH64_CMODEL_LARGE); |
6168 | } | |
6169 | ||
43e9d192 | 6170 | static bool |
e78d485e | 6171 | aarch64_use_blocks_for_constant_p (machine_mode, const_rtx) |
43e9d192 | 6172 | { |
3eece53d RR |
6173 | /* Fixme:: In an ideal world this would work similar |
6174 | to the logic in aarch64_select_rtx_section but this | |
6175 | breaks bootstrap in gcc go. For now we workaround | |
6176 | this by returning false here. */ | |
6177 | return false; | |
43e9d192 IB |
6178 | } |
6179 | ||
e78d485e RR |
6180 | /* Select appropriate section for constants depending |
6181 | on where we place literal pools. */ | |
6182 | ||
43e9d192 | 6183 | static section * |
e78d485e RR |
6184 | aarch64_select_rtx_section (machine_mode mode, |
6185 | rtx x, | |
6186 | unsigned HOST_WIDE_INT align) | |
43e9d192 | 6187 | { |
e78d485e RR |
6188 | if (aarch64_can_use_per_function_literal_pools_p ()) |
6189 | return function_section (current_function_decl); | |
43e9d192 | 6190 | |
e78d485e RR |
6191 | return default_elf_select_rtx_section (mode, x, align); |
6192 | } | |
43e9d192 | 6193 | |
5fca7b66 RH |
6194 | /* Implement ASM_OUTPUT_POOL_EPILOGUE. */ |
6195 | void | |
6196 | aarch64_asm_output_pool_epilogue (FILE *f, const char *, tree, | |
6197 | HOST_WIDE_INT offset) | |
6198 | { | |
6199 | /* When using per-function literal pools, we must ensure that any code | |
6200 | section is aligned to the minimal instruction length, lest we get | |
6201 | errors from the assembler re "unaligned instructions". */ | |
6202 | if ((offset & 3) && aarch64_can_use_per_function_literal_pools_p ()) | |
6203 | ASM_OUTPUT_ALIGN (f, 2); | |
6204 | } | |
6205 | ||
43e9d192 IB |
6206 | /* Costs. */ |
6207 | ||
6208 | /* Helper function for rtx cost calculation. Strip a shift expression | |
6209 | from X. Returns the inner operand if successful, or the original | |
6210 | expression on failure. */ | |
6211 | static rtx | |
6212 | aarch64_strip_shift (rtx x) | |
6213 | { | |
6214 | rtx op = x; | |
6215 | ||
57b77d46 RE |
6216 | /* We accept both ROTATERT and ROTATE: since the RHS must be a constant |
6217 | we can convert both to ROR during final output. */ | |
43e9d192 IB |
6218 | if ((GET_CODE (op) == ASHIFT |
6219 | || GET_CODE (op) == ASHIFTRT | |
57b77d46 RE |
6220 | || GET_CODE (op) == LSHIFTRT |
6221 | || GET_CODE (op) == ROTATERT | |
6222 | || GET_CODE (op) == ROTATE) | |
43e9d192 IB |
6223 | && CONST_INT_P (XEXP (op, 1))) |
6224 | return XEXP (op, 0); | |
6225 | ||
6226 | if (GET_CODE (op) == MULT | |
6227 | && CONST_INT_P (XEXP (op, 1)) | |
6228 | && ((unsigned) exact_log2 (INTVAL (XEXP (op, 1)))) < 64) | |
6229 | return XEXP (op, 0); | |
6230 | ||
6231 | return x; | |
6232 | } | |
6233 | ||
4745e701 | 6234 | /* Helper function for rtx cost calculation. Strip an extend |
43e9d192 IB |
6235 | expression from X. Returns the inner operand if successful, or the |
6236 | original expression on failure. We deal with a number of possible | |
b10f1009 AP |
6237 | canonicalization variations here. If STRIP_SHIFT is true, then |
6238 | we can strip off a shift also. */ | |
43e9d192 | 6239 | static rtx |
b10f1009 | 6240 | aarch64_strip_extend (rtx x, bool strip_shift) |
43e9d192 IB |
6241 | { |
6242 | rtx op = x; | |
6243 | ||
6244 | /* Zero and sign extraction of a widened value. */ | |
6245 | if ((GET_CODE (op) == ZERO_EXTRACT || GET_CODE (op) == SIGN_EXTRACT) | |
6246 | && XEXP (op, 2) == const0_rtx | |
4745e701 | 6247 | && GET_CODE (XEXP (op, 0)) == MULT |
43e9d192 IB |
6248 | && aarch64_is_extend_from_extract (GET_MODE (op), XEXP (XEXP (op, 0), 1), |
6249 | XEXP (op, 1))) | |
6250 | return XEXP (XEXP (op, 0), 0); | |
6251 | ||
6252 | /* It can also be represented (for zero-extend) as an AND with an | |
6253 | immediate. */ | |
6254 | if (GET_CODE (op) == AND | |
6255 | && GET_CODE (XEXP (op, 0)) == MULT | |
6256 | && CONST_INT_P (XEXP (XEXP (op, 0), 1)) | |
6257 | && CONST_INT_P (XEXP (op, 1)) | |
6258 | && aarch64_uxt_size (exact_log2 (INTVAL (XEXP (XEXP (op, 0), 1))), | |
6259 | INTVAL (XEXP (op, 1))) != 0) | |
6260 | return XEXP (XEXP (op, 0), 0); | |
6261 | ||
6262 | /* Now handle extended register, as this may also have an optional | |
6263 | left shift by 1..4. */ | |
b10f1009 AP |
6264 | if (strip_shift |
6265 | && GET_CODE (op) == ASHIFT | |
43e9d192 IB |
6266 | && CONST_INT_P (XEXP (op, 1)) |
6267 | && ((unsigned HOST_WIDE_INT) INTVAL (XEXP (op, 1))) <= 4) | |
6268 | op = XEXP (op, 0); | |
6269 | ||
6270 | if (GET_CODE (op) == ZERO_EXTEND | |
6271 | || GET_CODE (op) == SIGN_EXTEND) | |
6272 | op = XEXP (op, 0); | |
6273 | ||
6274 | if (op != x) | |
6275 | return op; | |
6276 | ||
4745e701 JG |
6277 | return x; |
6278 | } | |
6279 | ||
0a78ebe4 KT |
6280 | /* Return true iff CODE is a shift supported in combination |
6281 | with arithmetic instructions. */ | |
4d1919ed | 6282 | |
0a78ebe4 KT |
6283 | static bool |
6284 | aarch64_shift_p (enum rtx_code code) | |
6285 | { | |
6286 | return code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT; | |
6287 | } | |
6288 | ||
b10f1009 AP |
6289 | |
6290 | /* Return true iff X is a cheap shift without a sign extend. */ | |
6291 | ||
6292 | static bool | |
6293 | aarch64_cheap_mult_shift_p (rtx x) | |
6294 | { | |
6295 | rtx op0, op1; | |
6296 | ||
6297 | op0 = XEXP (x, 0); | |
6298 | op1 = XEXP (x, 1); | |
6299 | ||
6300 | if (!(aarch64_tune_params.extra_tuning_flags | |
6301 | & AARCH64_EXTRA_TUNE_CHEAP_SHIFT_EXTEND)) | |
6302 | return false; | |
6303 | ||
6304 | if (GET_CODE (op0) == SIGN_EXTEND) | |
6305 | return false; | |
6306 | ||
6307 | if (GET_CODE (x) == ASHIFT && CONST_INT_P (op1) | |
6308 | && UINTVAL (op1) <= 4) | |
6309 | return true; | |
6310 | ||
6311 | if (GET_CODE (x) != MULT || !CONST_INT_P (op1)) | |
6312 | return false; | |
6313 | ||
6314 | HOST_WIDE_INT l2 = exact_log2 (INTVAL (op1)); | |
6315 | ||
6316 | if (l2 > 0 && l2 <= 4) | |
6317 | return true; | |
6318 | ||
6319 | return false; | |
6320 | } | |
6321 | ||
4745e701 | 6322 | /* Helper function for rtx cost calculation. Calculate the cost of |
0a78ebe4 KT |
6323 | a MULT or ASHIFT, which may be part of a compound PLUS/MINUS rtx. |
6324 | Return the calculated cost of the expression, recursing manually in to | |
4745e701 JG |
6325 | operands where needed. */ |
6326 | ||
6327 | static int | |
e548c9df | 6328 | aarch64_rtx_mult_cost (rtx x, enum rtx_code code, int outer, bool speed) |
4745e701 JG |
6329 | { |
6330 | rtx op0, op1; | |
6331 | const struct cpu_cost_table *extra_cost | |
b175b679 | 6332 | = aarch64_tune_params.insn_extra_cost; |
4745e701 | 6333 | int cost = 0; |
0a78ebe4 | 6334 | bool compound_p = (outer == PLUS || outer == MINUS); |
ef4bddc2 | 6335 | machine_mode mode = GET_MODE (x); |
4745e701 JG |
6336 | |
6337 | gcc_checking_assert (code == MULT); | |
6338 | ||
6339 | op0 = XEXP (x, 0); | |
6340 | op1 = XEXP (x, 1); | |
6341 | ||
6342 | if (VECTOR_MODE_P (mode)) | |
6343 | mode = GET_MODE_INNER (mode); | |
6344 | ||
6345 | /* Integer multiply/fma. */ | |
6346 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
6347 | { | |
6348 | /* The multiply will be canonicalized as a shift, cost it as such. */ | |
0a78ebe4 KT |
6349 | if (aarch64_shift_p (GET_CODE (x)) |
6350 | || (CONST_INT_P (op1) | |
6351 | && exact_log2 (INTVAL (op1)) > 0)) | |
4745e701 | 6352 | { |
0a78ebe4 KT |
6353 | bool is_extend = GET_CODE (op0) == ZERO_EXTEND |
6354 | || GET_CODE (op0) == SIGN_EXTEND; | |
4745e701 JG |
6355 | if (speed) |
6356 | { | |
0a78ebe4 KT |
6357 | if (compound_p) |
6358 | { | |
b10f1009 AP |
6359 | /* If the shift is considered cheap, |
6360 | then don't add any cost. */ | |
6361 | if (aarch64_cheap_mult_shift_p (x)) | |
6362 | ; | |
6363 | else if (REG_P (op1)) | |
0a78ebe4 KT |
6364 | /* ARITH + shift-by-register. */ |
6365 | cost += extra_cost->alu.arith_shift_reg; | |
6366 | else if (is_extend) | |
6367 | /* ARITH + extended register. We don't have a cost field | |
6368 | for ARITH+EXTEND+SHIFT, so use extend_arith here. */ | |
6369 | cost += extra_cost->alu.extend_arith; | |
6370 | else | |
6371 | /* ARITH + shift-by-immediate. */ | |
6372 | cost += extra_cost->alu.arith_shift; | |
6373 | } | |
4745e701 JG |
6374 | else |
6375 | /* LSL (immediate). */ | |
0a78ebe4 KT |
6376 | cost += extra_cost->alu.shift; |
6377 | ||
4745e701 | 6378 | } |
0a78ebe4 KT |
6379 | /* Strip extends as we will have costed them in the case above. */ |
6380 | if (is_extend) | |
b10f1009 | 6381 | op0 = aarch64_strip_extend (op0, true); |
4745e701 | 6382 | |
e548c9df | 6383 | cost += rtx_cost (op0, VOIDmode, code, 0, speed); |
4745e701 JG |
6384 | |
6385 | return cost; | |
6386 | } | |
6387 | ||
d2ac256b KT |
6388 | /* MNEG or [US]MNEGL. Extract the NEG operand and indicate that it's a |
6389 | compound and let the below cases handle it. After all, MNEG is a | |
6390 | special-case alias of MSUB. */ | |
6391 | if (GET_CODE (op0) == NEG) | |
6392 | { | |
6393 | op0 = XEXP (op0, 0); | |
6394 | compound_p = true; | |
6395 | } | |
6396 | ||
4745e701 JG |
6397 | /* Integer multiplies or FMAs have zero/sign extending variants. */ |
6398 | if ((GET_CODE (op0) == ZERO_EXTEND | |
6399 | && GET_CODE (op1) == ZERO_EXTEND) | |
6400 | || (GET_CODE (op0) == SIGN_EXTEND | |
6401 | && GET_CODE (op1) == SIGN_EXTEND)) | |
6402 | { | |
e548c9df AM |
6403 | cost += rtx_cost (XEXP (op0, 0), VOIDmode, MULT, 0, speed); |
6404 | cost += rtx_cost (XEXP (op1, 0), VOIDmode, MULT, 1, speed); | |
4745e701 JG |
6405 | |
6406 | if (speed) | |
6407 | { | |
0a78ebe4 | 6408 | if (compound_p) |
d2ac256b | 6409 | /* SMADDL/UMADDL/UMSUBL/SMSUBL. */ |
4745e701 JG |
6410 | cost += extra_cost->mult[0].extend_add; |
6411 | else | |
6412 | /* MUL/SMULL/UMULL. */ | |
6413 | cost += extra_cost->mult[0].extend; | |
6414 | } | |
6415 | ||
6416 | return cost; | |
6417 | } | |
6418 | ||
d2ac256b | 6419 | /* This is either an integer multiply or a MADD. In both cases |
4745e701 | 6420 | we want to recurse and cost the operands. */ |
e548c9df AM |
6421 | cost += rtx_cost (op0, mode, MULT, 0, speed); |
6422 | cost += rtx_cost (op1, mode, MULT, 1, speed); | |
4745e701 JG |
6423 | |
6424 | if (speed) | |
6425 | { | |
0a78ebe4 | 6426 | if (compound_p) |
d2ac256b | 6427 | /* MADD/MSUB. */ |
4745e701 JG |
6428 | cost += extra_cost->mult[mode == DImode].add; |
6429 | else | |
6430 | /* MUL. */ | |
6431 | cost += extra_cost->mult[mode == DImode].simple; | |
6432 | } | |
6433 | ||
6434 | return cost; | |
6435 | } | |
6436 | else | |
6437 | { | |
6438 | if (speed) | |
6439 | { | |
3d840f7d | 6440 | /* Floating-point FMA/FMUL can also support negations of the |
d318517d SN |
6441 | operands, unless the rounding mode is upward or downward in |
6442 | which case FNMUL is different than FMUL with operand negation. */ | |
6443 | bool neg0 = GET_CODE (op0) == NEG; | |
6444 | bool neg1 = GET_CODE (op1) == NEG; | |
6445 | if (compound_p || !flag_rounding_math || (neg0 && neg1)) | |
6446 | { | |
6447 | if (neg0) | |
6448 | op0 = XEXP (op0, 0); | |
6449 | if (neg1) | |
6450 | op1 = XEXP (op1, 0); | |
6451 | } | |
4745e701 | 6452 | |
0a78ebe4 | 6453 | if (compound_p) |
4745e701 JG |
6454 | /* FMADD/FNMADD/FNMSUB/FMSUB. */ |
6455 | cost += extra_cost->fp[mode == DFmode].fma; | |
6456 | else | |
3d840f7d | 6457 | /* FMUL/FNMUL. */ |
4745e701 JG |
6458 | cost += extra_cost->fp[mode == DFmode].mult; |
6459 | } | |
6460 | ||
e548c9df AM |
6461 | cost += rtx_cost (op0, mode, MULT, 0, speed); |
6462 | cost += rtx_cost (op1, mode, MULT, 1, speed); | |
4745e701 JG |
6463 | return cost; |
6464 | } | |
43e9d192 IB |
6465 | } |
6466 | ||
67747367 JG |
6467 | static int |
6468 | aarch64_address_cost (rtx x, | |
ef4bddc2 | 6469 | machine_mode mode, |
67747367 JG |
6470 | addr_space_t as ATTRIBUTE_UNUSED, |
6471 | bool speed) | |
6472 | { | |
6473 | enum rtx_code c = GET_CODE (x); | |
b175b679 | 6474 | const struct cpu_addrcost_table *addr_cost = aarch64_tune_params.addr_cost; |
67747367 JG |
6475 | struct aarch64_address_info info; |
6476 | int cost = 0; | |
6477 | info.shift = 0; | |
6478 | ||
6479 | if (!aarch64_classify_address (&info, x, mode, c, false)) | |
6480 | { | |
6481 | if (GET_CODE (x) == CONST || GET_CODE (x) == SYMBOL_REF) | |
6482 | { | |
6483 | /* This is a CONST or SYMBOL ref which will be split | |
6484 | in a different way depending on the code model in use. | |
6485 | Cost it through the generic infrastructure. */ | |
e548c9df | 6486 | int cost_symbol_ref = rtx_cost (x, Pmode, MEM, 1, speed); |
67747367 JG |
6487 | /* Divide through by the cost of one instruction to |
6488 | bring it to the same units as the address costs. */ | |
6489 | cost_symbol_ref /= COSTS_N_INSNS (1); | |
6490 | /* The cost is then the cost of preparing the address, | |
6491 | followed by an immediate (possibly 0) offset. */ | |
6492 | return cost_symbol_ref + addr_cost->imm_offset; | |
6493 | } | |
6494 | else | |
6495 | { | |
6496 | /* This is most likely a jump table from a case | |
6497 | statement. */ | |
6498 | return addr_cost->register_offset; | |
6499 | } | |
6500 | } | |
6501 | ||
6502 | switch (info.type) | |
6503 | { | |
6504 | case ADDRESS_LO_SUM: | |
6505 | case ADDRESS_SYMBOLIC: | |
6506 | case ADDRESS_REG_IMM: | |
6507 | cost += addr_cost->imm_offset; | |
6508 | break; | |
6509 | ||
6510 | case ADDRESS_REG_WB: | |
6511 | if (c == PRE_INC || c == PRE_DEC || c == PRE_MODIFY) | |
6512 | cost += addr_cost->pre_modify; | |
6513 | else if (c == POST_INC || c == POST_DEC || c == POST_MODIFY) | |
6514 | cost += addr_cost->post_modify; | |
6515 | else | |
6516 | gcc_unreachable (); | |
6517 | ||
6518 | break; | |
6519 | ||
6520 | case ADDRESS_REG_REG: | |
6521 | cost += addr_cost->register_offset; | |
6522 | break; | |
6523 | ||
67747367 | 6524 | case ADDRESS_REG_SXTW: |
783879e6 EM |
6525 | cost += addr_cost->register_sextend; |
6526 | break; | |
6527 | ||
6528 | case ADDRESS_REG_UXTW: | |
6529 | cost += addr_cost->register_zextend; | |
67747367 JG |
6530 | break; |
6531 | ||
6532 | default: | |
6533 | gcc_unreachable (); | |
6534 | } | |
6535 | ||
6536 | ||
6537 | if (info.shift > 0) | |
6538 | { | |
6539 | /* For the sake of calculating the cost of the shifted register | |
6540 | component, we can treat same sized modes in the same way. */ | |
6541 | switch (GET_MODE_BITSIZE (mode)) | |
6542 | { | |
6543 | case 16: | |
6544 | cost += addr_cost->addr_scale_costs.hi; | |
6545 | break; | |
6546 | ||
6547 | case 32: | |
6548 | cost += addr_cost->addr_scale_costs.si; | |
6549 | break; | |
6550 | ||
6551 | case 64: | |
6552 | cost += addr_cost->addr_scale_costs.di; | |
6553 | break; | |
6554 | ||
6555 | /* We can't tell, or this is a 128-bit vector. */ | |
6556 | default: | |
6557 | cost += addr_cost->addr_scale_costs.ti; | |
6558 | break; | |
6559 | } | |
6560 | } | |
6561 | ||
6562 | return cost; | |
6563 | } | |
6564 | ||
b9066f5a MW |
6565 | /* Return the cost of a branch. If SPEED_P is true then the compiler is |
6566 | optimizing for speed. If PREDICTABLE_P is true then the branch is predicted | |
6567 | to be taken. */ | |
6568 | ||
6569 | int | |
6570 | aarch64_branch_cost (bool speed_p, bool predictable_p) | |
6571 | { | |
6572 | /* When optimizing for speed, use the cost of unpredictable branches. */ | |
6573 | const struct cpu_branch_cost *branch_costs = | |
b175b679 | 6574 | aarch64_tune_params.branch_costs; |
b9066f5a MW |
6575 | |
6576 | if (!speed_p || predictable_p) | |
6577 | return branch_costs->predictable; | |
6578 | else | |
6579 | return branch_costs->unpredictable; | |
6580 | } | |
6581 | ||
7cc2145f JG |
6582 | /* Return true if the RTX X in mode MODE is a zero or sign extract |
6583 | usable in an ADD or SUB (extended register) instruction. */ | |
6584 | static bool | |
ef4bddc2 | 6585 | aarch64_rtx_arith_op_extract_p (rtx x, machine_mode mode) |
7cc2145f JG |
6586 | { |
6587 | /* Catch add with a sign extract. | |
6588 | This is add_<optab><mode>_multp2. */ | |
6589 | if (GET_CODE (x) == SIGN_EXTRACT | |
6590 | || GET_CODE (x) == ZERO_EXTRACT) | |
6591 | { | |
6592 | rtx op0 = XEXP (x, 0); | |
6593 | rtx op1 = XEXP (x, 1); | |
6594 | rtx op2 = XEXP (x, 2); | |
6595 | ||
6596 | if (GET_CODE (op0) == MULT | |
6597 | && CONST_INT_P (op1) | |
6598 | && op2 == const0_rtx | |
6599 | && CONST_INT_P (XEXP (op0, 1)) | |
6600 | && aarch64_is_extend_from_extract (mode, | |
6601 | XEXP (op0, 1), | |
6602 | op1)) | |
6603 | { | |
6604 | return true; | |
6605 | } | |
6606 | } | |
e47c4031 KT |
6607 | /* The simple case <ARITH>, XD, XN, XM, [us]xt. |
6608 | No shift. */ | |
6609 | else if (GET_CODE (x) == SIGN_EXTEND | |
6610 | || GET_CODE (x) == ZERO_EXTEND) | |
6611 | return REG_P (XEXP (x, 0)); | |
7cc2145f JG |
6612 | |
6613 | return false; | |
6614 | } | |
6615 | ||
61263118 KT |
6616 | static bool |
6617 | aarch64_frint_unspec_p (unsigned int u) | |
6618 | { | |
6619 | switch (u) | |
6620 | { | |
6621 | case UNSPEC_FRINTZ: | |
6622 | case UNSPEC_FRINTP: | |
6623 | case UNSPEC_FRINTM: | |
6624 | case UNSPEC_FRINTA: | |
6625 | case UNSPEC_FRINTN: | |
6626 | case UNSPEC_FRINTX: | |
6627 | case UNSPEC_FRINTI: | |
6628 | return true; | |
6629 | ||
6630 | default: | |
6631 | return false; | |
6632 | } | |
6633 | } | |
6634 | ||
fb0cb7fa KT |
6635 | /* Return true iff X is an rtx that will match an extr instruction |
6636 | i.e. as described in the *extr<mode>5_insn family of patterns. | |
6637 | OP0 and OP1 will be set to the operands of the shifts involved | |
6638 | on success and will be NULL_RTX otherwise. */ | |
6639 | ||
6640 | static bool | |
6641 | aarch64_extr_rtx_p (rtx x, rtx *res_op0, rtx *res_op1) | |
6642 | { | |
6643 | rtx op0, op1; | |
6644 | machine_mode mode = GET_MODE (x); | |
6645 | ||
6646 | *res_op0 = NULL_RTX; | |
6647 | *res_op1 = NULL_RTX; | |
6648 | ||
6649 | if (GET_CODE (x) != IOR) | |
6650 | return false; | |
6651 | ||
6652 | op0 = XEXP (x, 0); | |
6653 | op1 = XEXP (x, 1); | |
6654 | ||
6655 | if ((GET_CODE (op0) == ASHIFT && GET_CODE (op1) == LSHIFTRT) | |
6656 | || (GET_CODE (op1) == ASHIFT && GET_CODE (op0) == LSHIFTRT)) | |
6657 | { | |
6658 | /* Canonicalise locally to ashift in op0, lshiftrt in op1. */ | |
6659 | if (GET_CODE (op1) == ASHIFT) | |
6660 | std::swap (op0, op1); | |
6661 | ||
6662 | if (!CONST_INT_P (XEXP (op0, 1)) || !CONST_INT_P (XEXP (op1, 1))) | |
6663 | return false; | |
6664 | ||
6665 | unsigned HOST_WIDE_INT shft_amnt_0 = UINTVAL (XEXP (op0, 1)); | |
6666 | unsigned HOST_WIDE_INT shft_amnt_1 = UINTVAL (XEXP (op1, 1)); | |
6667 | ||
6668 | if (shft_amnt_0 < GET_MODE_BITSIZE (mode) | |
6669 | && shft_amnt_0 + shft_amnt_1 == GET_MODE_BITSIZE (mode)) | |
6670 | { | |
6671 | *res_op0 = XEXP (op0, 0); | |
6672 | *res_op1 = XEXP (op1, 0); | |
6673 | return true; | |
6674 | } | |
6675 | } | |
6676 | ||
6677 | return false; | |
6678 | } | |
6679 | ||
2d5ffe46 AP |
6680 | /* Calculate the cost of calculating (if_then_else (OP0) (OP1) (OP2)), |
6681 | storing it in *COST. Result is true if the total cost of the operation | |
6682 | has now been calculated. */ | |
6683 | static bool | |
6684 | aarch64_if_then_else_costs (rtx op0, rtx op1, rtx op2, int *cost, bool speed) | |
6685 | { | |
b9e3afe9 AP |
6686 | rtx inner; |
6687 | rtx comparator; | |
6688 | enum rtx_code cmpcode; | |
6689 | ||
6690 | if (COMPARISON_P (op0)) | |
6691 | { | |
6692 | inner = XEXP (op0, 0); | |
6693 | comparator = XEXP (op0, 1); | |
6694 | cmpcode = GET_CODE (op0); | |
6695 | } | |
6696 | else | |
6697 | { | |
6698 | inner = op0; | |
6699 | comparator = const0_rtx; | |
6700 | cmpcode = NE; | |
6701 | } | |
6702 | ||
2d5ffe46 AP |
6703 | if (GET_CODE (op1) == PC || GET_CODE (op2) == PC) |
6704 | { | |
6705 | /* Conditional branch. */ | |
b9e3afe9 | 6706 | if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_CC) |
2d5ffe46 AP |
6707 | return true; |
6708 | else | |
6709 | { | |
b9e3afe9 | 6710 | if (cmpcode == NE || cmpcode == EQ) |
2d5ffe46 | 6711 | { |
2d5ffe46 AP |
6712 | if (comparator == const0_rtx) |
6713 | { | |
6714 | /* TBZ/TBNZ/CBZ/CBNZ. */ | |
6715 | if (GET_CODE (inner) == ZERO_EXTRACT) | |
6716 | /* TBZ/TBNZ. */ | |
e548c9df AM |
6717 | *cost += rtx_cost (XEXP (inner, 0), VOIDmode, |
6718 | ZERO_EXTRACT, 0, speed); | |
6719 | else | |
6720 | /* CBZ/CBNZ. */ | |
6721 | *cost += rtx_cost (inner, VOIDmode, cmpcode, 0, speed); | |
2d5ffe46 AP |
6722 | |
6723 | return true; | |
6724 | } | |
6725 | } | |
b9e3afe9 | 6726 | else if (cmpcode == LT || cmpcode == GE) |
2d5ffe46 | 6727 | { |
2d5ffe46 AP |
6728 | /* TBZ/TBNZ. */ |
6729 | if (comparator == const0_rtx) | |
6730 | return true; | |
6731 | } | |
6732 | } | |
6733 | } | |
b9e3afe9 | 6734 | else if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_CC) |
2d5ffe46 | 6735 | { |
786298dc | 6736 | /* CCMP. */ |
6dfeb7ce | 6737 | if (GET_CODE (op1) == COMPARE) |
786298dc WD |
6738 | { |
6739 | /* Increase cost of CCMP reg, 0, imm, CC to prefer CMP reg, 0. */ | |
6740 | if (XEXP (op1, 1) == const0_rtx) | |
6741 | *cost += 1; | |
6742 | if (speed) | |
6743 | { | |
6744 | machine_mode mode = GET_MODE (XEXP (op1, 0)); | |
6745 | const struct cpu_cost_table *extra_cost | |
6746 | = aarch64_tune_params.insn_extra_cost; | |
6747 | ||
6748 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
6749 | *cost += extra_cost->alu.arith; | |
6750 | else | |
6751 | *cost += extra_cost->fp[mode == DFmode].compare; | |
6752 | } | |
6753 | return true; | |
6754 | } | |
6755 | ||
2d5ffe46 AP |
6756 | /* It's a conditional operation based on the status flags, |
6757 | so it must be some flavor of CSEL. */ | |
6758 | ||
6759 | /* CSNEG, CSINV, and CSINC are handled for free as part of CSEL. */ | |
6760 | if (GET_CODE (op1) == NEG | |
6761 | || GET_CODE (op1) == NOT | |
6762 | || (GET_CODE (op1) == PLUS && XEXP (op1, 1) == const1_rtx)) | |
6763 | op1 = XEXP (op1, 0); | |
bad00732 KT |
6764 | else if (GET_CODE (op1) == ZERO_EXTEND && GET_CODE (op2) == ZERO_EXTEND) |
6765 | { | |
6766 | /* CSEL with zero-extension (*cmovdi_insn_uxtw). */ | |
6767 | op1 = XEXP (op1, 0); | |
6768 | op2 = XEXP (op2, 0); | |
6769 | } | |
2d5ffe46 | 6770 | |
e548c9df AM |
6771 | *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed); |
6772 | *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed); | |
2d5ffe46 AP |
6773 | return true; |
6774 | } | |
6775 | ||
6776 | /* We don't know what this is, cost all operands. */ | |
6777 | return false; | |
6778 | } | |
6779 | ||
283b6c85 KT |
6780 | /* Check whether X is a bitfield operation of the form shift + extend that |
6781 | maps down to a UBFIZ/SBFIZ/UBFX/SBFX instruction. If so, return the | |
6782 | operand to which the bitfield operation is applied. Otherwise return | |
6783 | NULL_RTX. */ | |
6784 | ||
6785 | static rtx | |
6786 | aarch64_extend_bitfield_pattern_p (rtx x) | |
6787 | { | |
6788 | rtx_code outer_code = GET_CODE (x); | |
6789 | machine_mode outer_mode = GET_MODE (x); | |
6790 | ||
6791 | if (outer_code != ZERO_EXTEND && outer_code != SIGN_EXTEND | |
6792 | && outer_mode != SImode && outer_mode != DImode) | |
6793 | return NULL_RTX; | |
6794 | ||
6795 | rtx inner = XEXP (x, 0); | |
6796 | rtx_code inner_code = GET_CODE (inner); | |
6797 | machine_mode inner_mode = GET_MODE (inner); | |
6798 | rtx op = NULL_RTX; | |
6799 | ||
6800 | switch (inner_code) | |
6801 | { | |
6802 | case ASHIFT: | |
6803 | if (CONST_INT_P (XEXP (inner, 1)) | |
6804 | && (inner_mode == QImode || inner_mode == HImode)) | |
6805 | op = XEXP (inner, 0); | |
6806 | break; | |
6807 | case LSHIFTRT: | |
6808 | if (outer_code == ZERO_EXTEND && CONST_INT_P (XEXP (inner, 1)) | |
6809 | && (inner_mode == QImode || inner_mode == HImode)) | |
6810 | op = XEXP (inner, 0); | |
6811 | break; | |
6812 | case ASHIFTRT: | |
6813 | if (outer_code == SIGN_EXTEND && CONST_INT_P (XEXP (inner, 1)) | |
6814 | && (inner_mode == QImode || inner_mode == HImode)) | |
6815 | op = XEXP (inner, 0); | |
6816 | break; | |
6817 | default: | |
6818 | break; | |
6819 | } | |
6820 | ||
6821 | return op; | |
6822 | } | |
6823 | ||
8c83f71d KT |
6824 | /* Return true if the mask and a shift amount from an RTX of the form |
6825 | (x << SHFT_AMNT) & MASK are valid to combine into a UBFIZ instruction of | |
6826 | mode MODE. See the *andim_ashift<mode>_bfiz pattern. */ | |
6827 | ||
6828 | bool | |
6829 | aarch64_mask_and_shift_for_ubfiz_p (machine_mode mode, rtx mask, rtx shft_amnt) | |
6830 | { | |
6831 | return CONST_INT_P (mask) && CONST_INT_P (shft_amnt) | |
6832 | && INTVAL (shft_amnt) < GET_MODE_BITSIZE (mode) | |
6833 | && exact_log2 ((INTVAL (mask) >> INTVAL (shft_amnt)) + 1) >= 0 | |
6834 | && (INTVAL (mask) & ((1 << INTVAL (shft_amnt)) - 1)) == 0; | |
6835 | } | |
6836 | ||
43e9d192 IB |
6837 | /* Calculate the cost of calculating X, storing it in *COST. Result |
6838 | is true if the total cost of the operation has now been calculated. */ | |
6839 | static bool | |
e548c9df | 6840 | aarch64_rtx_costs (rtx x, machine_mode mode, int outer ATTRIBUTE_UNUSED, |
43e9d192 IB |
6841 | int param ATTRIBUTE_UNUSED, int *cost, bool speed) |
6842 | { | |
a8eecd00 | 6843 | rtx op0, op1, op2; |
73250c4c | 6844 | const struct cpu_cost_table *extra_cost |
b175b679 | 6845 | = aarch64_tune_params.insn_extra_cost; |
e548c9df | 6846 | int code = GET_CODE (x); |
b4206259 | 6847 | scalar_int_mode int_mode; |
43e9d192 | 6848 | |
7fc5ef02 JG |
6849 | /* By default, assume that everything has equivalent cost to the |
6850 | cheapest instruction. Any additional costs are applied as a delta | |
6851 | above this default. */ | |
6852 | *cost = COSTS_N_INSNS (1); | |
6853 | ||
43e9d192 IB |
6854 | switch (code) |
6855 | { | |
6856 | case SET: | |
ba123b0d JG |
6857 | /* The cost depends entirely on the operands to SET. */ |
6858 | *cost = 0; | |
43e9d192 IB |
6859 | op0 = SET_DEST (x); |
6860 | op1 = SET_SRC (x); | |
6861 | ||
6862 | switch (GET_CODE (op0)) | |
6863 | { | |
6864 | case MEM: | |
6865 | if (speed) | |
2961177e JG |
6866 | { |
6867 | rtx address = XEXP (op0, 0); | |
b6875aac KV |
6868 | if (VECTOR_MODE_P (mode)) |
6869 | *cost += extra_cost->ldst.storev; | |
6870 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
2961177e JG |
6871 | *cost += extra_cost->ldst.store; |
6872 | else if (mode == SFmode) | |
6873 | *cost += extra_cost->ldst.storef; | |
6874 | else if (mode == DFmode) | |
6875 | *cost += extra_cost->ldst.stored; | |
6876 | ||
6877 | *cost += | |
6878 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
6879 | 0, speed)); | |
6880 | } | |
43e9d192 | 6881 | |
e548c9df | 6882 | *cost += rtx_cost (op1, mode, SET, 1, speed); |
43e9d192 IB |
6883 | return true; |
6884 | ||
6885 | case SUBREG: | |
6886 | if (! REG_P (SUBREG_REG (op0))) | |
e548c9df | 6887 | *cost += rtx_cost (SUBREG_REG (op0), VOIDmode, SET, 0, speed); |
ba123b0d | 6888 | |
43e9d192 IB |
6889 | /* Fall through. */ |
6890 | case REG: | |
b6875aac KV |
6891 | /* The cost is one per vector-register copied. */ |
6892 | if (VECTOR_MODE_P (GET_MODE (op0)) && REG_P (op1)) | |
6893 | { | |
6894 | int n_minus_1 = (GET_MODE_SIZE (GET_MODE (op0)) - 1) | |
6895 | / GET_MODE_SIZE (V4SImode); | |
6896 | *cost = COSTS_N_INSNS (n_minus_1 + 1); | |
6897 | } | |
ba123b0d JG |
6898 | /* const0_rtx is in general free, but we will use an |
6899 | instruction to set a register to 0. */ | |
b6875aac KV |
6900 | else if (REG_P (op1) || op1 == const0_rtx) |
6901 | { | |
6902 | /* The cost is 1 per register copied. */ | |
6903 | int n_minus_1 = (GET_MODE_SIZE (GET_MODE (op0)) - 1) | |
ba123b0d | 6904 | / UNITS_PER_WORD; |
b6875aac KV |
6905 | *cost = COSTS_N_INSNS (n_minus_1 + 1); |
6906 | } | |
ba123b0d JG |
6907 | else |
6908 | /* Cost is just the cost of the RHS of the set. */ | |
e548c9df | 6909 | *cost += rtx_cost (op1, mode, SET, 1, speed); |
43e9d192 IB |
6910 | return true; |
6911 | ||
ba123b0d | 6912 | case ZERO_EXTRACT: |
43e9d192 | 6913 | case SIGN_EXTRACT: |
ba123b0d JG |
6914 | /* Bit-field insertion. Strip any redundant widening of |
6915 | the RHS to meet the width of the target. */ | |
43e9d192 IB |
6916 | if (GET_CODE (op1) == SUBREG) |
6917 | op1 = SUBREG_REG (op1); | |
6918 | if ((GET_CODE (op1) == ZERO_EXTEND | |
6919 | || GET_CODE (op1) == SIGN_EXTEND) | |
4aa81c2e | 6920 | && CONST_INT_P (XEXP (op0, 1)) |
43e9d192 IB |
6921 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (op1, 0))) |
6922 | >= INTVAL (XEXP (op0, 1)))) | |
6923 | op1 = XEXP (op1, 0); | |
ba123b0d JG |
6924 | |
6925 | if (CONST_INT_P (op1)) | |
6926 | { | |
6927 | /* MOV immediate is assumed to always be cheap. */ | |
6928 | *cost = COSTS_N_INSNS (1); | |
6929 | } | |
6930 | else | |
6931 | { | |
6932 | /* BFM. */ | |
6933 | if (speed) | |
6934 | *cost += extra_cost->alu.bfi; | |
e548c9df | 6935 | *cost += rtx_cost (op1, VOIDmode, (enum rtx_code) code, 1, speed); |
ba123b0d JG |
6936 | } |
6937 | ||
43e9d192 IB |
6938 | return true; |
6939 | ||
6940 | default: | |
ba123b0d JG |
6941 | /* We can't make sense of this, assume default cost. */ |
6942 | *cost = COSTS_N_INSNS (1); | |
61263118 | 6943 | return false; |
43e9d192 IB |
6944 | } |
6945 | return false; | |
6946 | ||
9dfc162c JG |
6947 | case CONST_INT: |
6948 | /* If an instruction can incorporate a constant within the | |
6949 | instruction, the instruction's expression avoids calling | |
6950 | rtx_cost() on the constant. If rtx_cost() is called on a | |
6951 | constant, then it is usually because the constant must be | |
6952 | moved into a register by one or more instructions. | |
6953 | ||
6954 | The exception is constant 0, which can be expressed | |
6955 | as XZR/WZR and is therefore free. The exception to this is | |
6956 | if we have (set (reg) (const0_rtx)) in which case we must cost | |
6957 | the move. However, we can catch that when we cost the SET, so | |
6958 | we don't need to consider that here. */ | |
6959 | if (x == const0_rtx) | |
6960 | *cost = 0; | |
6961 | else | |
6962 | { | |
6963 | /* To an approximation, building any other constant is | |
6964 | proportionally expensive to the number of instructions | |
6965 | required to build that constant. This is true whether we | |
6966 | are compiling for SPEED or otherwise. */ | |
82614948 RR |
6967 | *cost = COSTS_N_INSNS (aarch64_internal_mov_immediate |
6968 | (NULL_RTX, x, false, mode)); | |
9dfc162c JG |
6969 | } |
6970 | return true; | |
6971 | ||
6972 | case CONST_DOUBLE: | |
a2170965 TC |
6973 | |
6974 | /* First determine number of instructions to do the move | |
6975 | as an integer constant. */ | |
6976 | if (!aarch64_float_const_representable_p (x) | |
6977 | && !aarch64_can_const_movi_rtx_p (x, mode) | |
6978 | && aarch64_float_const_rtx_p (x)) | |
6979 | { | |
6980 | unsigned HOST_WIDE_INT ival; | |
6981 | bool succeed = aarch64_reinterpret_float_as_int (x, &ival); | |
6982 | gcc_assert (succeed); | |
6983 | ||
304b9962 RS |
6984 | machine_mode imode = (mode == HFmode |
6985 | ? SImode | |
6986 | : int_mode_for_mode (mode).require ()); | |
a2170965 TC |
6987 | int ncost = aarch64_internal_mov_immediate |
6988 | (NULL_RTX, gen_int_mode (ival, imode), false, imode); | |
6989 | *cost += COSTS_N_INSNS (ncost); | |
6990 | return true; | |
6991 | } | |
6992 | ||
9dfc162c JG |
6993 | if (speed) |
6994 | { | |
6995 | /* mov[df,sf]_aarch64. */ | |
6996 | if (aarch64_float_const_representable_p (x)) | |
6997 | /* FMOV (scalar immediate). */ | |
6998 | *cost += extra_cost->fp[mode == DFmode].fpconst; | |
6999 | else if (!aarch64_float_const_zero_rtx_p (x)) | |
7000 | { | |
7001 | /* This will be a load from memory. */ | |
7002 | if (mode == DFmode) | |
7003 | *cost += extra_cost->ldst.loadd; | |
7004 | else | |
7005 | *cost += extra_cost->ldst.loadf; | |
7006 | } | |
7007 | else | |
7008 | /* Otherwise this is +0.0. We get this using MOVI d0, #0 | |
7009 | or MOV v0.s[0], wzr - neither of which are modeled by the | |
7010 | cost tables. Just use the default cost. */ | |
7011 | { | |
7012 | } | |
7013 | } | |
7014 | ||
7015 | return true; | |
7016 | ||
43e9d192 IB |
7017 | case MEM: |
7018 | if (speed) | |
2961177e JG |
7019 | { |
7020 | /* For loads we want the base cost of a load, plus an | |
7021 | approximation for the additional cost of the addressing | |
7022 | mode. */ | |
7023 | rtx address = XEXP (x, 0); | |
b6875aac KV |
7024 | if (VECTOR_MODE_P (mode)) |
7025 | *cost += extra_cost->ldst.loadv; | |
7026 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
2961177e JG |
7027 | *cost += extra_cost->ldst.load; |
7028 | else if (mode == SFmode) | |
7029 | *cost += extra_cost->ldst.loadf; | |
7030 | else if (mode == DFmode) | |
7031 | *cost += extra_cost->ldst.loadd; | |
7032 | ||
7033 | *cost += | |
7034 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
7035 | 0, speed)); | |
7036 | } | |
43e9d192 IB |
7037 | |
7038 | return true; | |
7039 | ||
7040 | case NEG: | |
4745e701 JG |
7041 | op0 = XEXP (x, 0); |
7042 | ||
b6875aac KV |
7043 | if (VECTOR_MODE_P (mode)) |
7044 | { | |
7045 | if (speed) | |
7046 | { | |
7047 | /* FNEG. */ | |
7048 | *cost += extra_cost->vect.alu; | |
7049 | } | |
7050 | return false; | |
7051 | } | |
7052 | ||
e548c9df AM |
7053 | if (GET_MODE_CLASS (mode) == MODE_INT) |
7054 | { | |
4745e701 JG |
7055 | if (GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMPARE |
7056 | || GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMM_COMPARE) | |
7057 | { | |
7058 | /* CSETM. */ | |
e548c9df | 7059 | *cost += rtx_cost (XEXP (op0, 0), VOIDmode, NEG, 0, speed); |
4745e701 JG |
7060 | return true; |
7061 | } | |
7062 | ||
7063 | /* Cost this as SUB wzr, X. */ | |
e548c9df | 7064 | op0 = CONST0_RTX (mode); |
4745e701 JG |
7065 | op1 = XEXP (x, 0); |
7066 | goto cost_minus; | |
7067 | } | |
7068 | ||
e548c9df | 7069 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
4745e701 JG |
7070 | { |
7071 | /* Support (neg(fma...)) as a single instruction only if | |
7072 | sign of zeros is unimportant. This matches the decision | |
7073 | making in aarch64.md. */ | |
7074 | if (GET_CODE (op0) == FMA && !HONOR_SIGNED_ZEROS (GET_MODE (op0))) | |
7075 | { | |
7076 | /* FNMADD. */ | |
e548c9df | 7077 | *cost = rtx_cost (op0, mode, NEG, 0, speed); |
4745e701 JG |
7078 | return true; |
7079 | } | |
d318517d SN |
7080 | if (GET_CODE (op0) == MULT) |
7081 | { | |
7082 | /* FNMUL. */ | |
7083 | *cost = rtx_cost (op0, mode, NEG, 0, speed); | |
7084 | return true; | |
7085 | } | |
4745e701 JG |
7086 | if (speed) |
7087 | /* FNEG. */ | |
7088 | *cost += extra_cost->fp[mode == DFmode].neg; | |
7089 | return false; | |
7090 | } | |
7091 | ||
7092 | return false; | |
43e9d192 | 7093 | |
781aeb73 KT |
7094 | case CLRSB: |
7095 | case CLZ: | |
7096 | if (speed) | |
b6875aac KV |
7097 | { |
7098 | if (VECTOR_MODE_P (mode)) | |
7099 | *cost += extra_cost->vect.alu; | |
7100 | else | |
7101 | *cost += extra_cost->alu.clz; | |
7102 | } | |
781aeb73 KT |
7103 | |
7104 | return false; | |
7105 | ||
43e9d192 IB |
7106 | case COMPARE: |
7107 | op0 = XEXP (x, 0); | |
7108 | op1 = XEXP (x, 1); | |
7109 | ||
7110 | if (op1 == const0_rtx | |
7111 | && GET_CODE (op0) == AND) | |
7112 | { | |
7113 | x = op0; | |
e548c9df | 7114 | mode = GET_MODE (op0); |
43e9d192 IB |
7115 | goto cost_logic; |
7116 | } | |
7117 | ||
a8eecd00 JG |
7118 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT) |
7119 | { | |
7120 | /* TODO: A write to the CC flags possibly costs extra, this | |
7121 | needs encoding in the cost tables. */ | |
7122 | ||
e548c9df | 7123 | mode = GET_MODE (op0); |
a8eecd00 JG |
7124 | /* ANDS. */ |
7125 | if (GET_CODE (op0) == AND) | |
7126 | { | |
7127 | x = op0; | |
7128 | goto cost_logic; | |
7129 | } | |
7130 | ||
7131 | if (GET_CODE (op0) == PLUS) | |
7132 | { | |
7133 | /* ADDS (and CMN alias). */ | |
7134 | x = op0; | |
7135 | goto cost_plus; | |
7136 | } | |
7137 | ||
7138 | if (GET_CODE (op0) == MINUS) | |
7139 | { | |
7140 | /* SUBS. */ | |
7141 | x = op0; | |
7142 | goto cost_minus; | |
7143 | } | |
7144 | ||
345854d8 KT |
7145 | if (GET_CODE (op0) == ZERO_EXTRACT && op1 == const0_rtx |
7146 | && GET_MODE (x) == CC_NZmode && CONST_INT_P (XEXP (op0, 1)) | |
7147 | && CONST_INT_P (XEXP (op0, 2))) | |
7148 | { | |
7149 | /* COMPARE of ZERO_EXTRACT form of TST-immediate. | |
7150 | Handle it here directly rather than going to cost_logic | |
7151 | since we know the immediate generated for the TST is valid | |
7152 | so we can avoid creating an intermediate rtx for it only | |
7153 | for costing purposes. */ | |
7154 | if (speed) | |
7155 | *cost += extra_cost->alu.logical; | |
7156 | ||
7157 | *cost += rtx_cost (XEXP (op0, 0), GET_MODE (op0), | |
7158 | ZERO_EXTRACT, 0, speed); | |
7159 | return true; | |
7160 | } | |
7161 | ||
a8eecd00 JG |
7162 | if (GET_CODE (op1) == NEG) |
7163 | { | |
7164 | /* CMN. */ | |
7165 | if (speed) | |
7166 | *cost += extra_cost->alu.arith; | |
7167 | ||
e548c9df AM |
7168 | *cost += rtx_cost (op0, mode, COMPARE, 0, speed); |
7169 | *cost += rtx_cost (XEXP (op1, 0), mode, NEG, 1, speed); | |
a8eecd00 JG |
7170 | return true; |
7171 | } | |
7172 | ||
7173 | /* CMP. | |
7174 | ||
7175 | Compare can freely swap the order of operands, and | |
7176 | canonicalization puts the more complex operation first. | |
7177 | But the integer MINUS logic expects the shift/extend | |
7178 | operation in op1. */ | |
7179 | if (! (REG_P (op0) | |
7180 | || (GET_CODE (op0) == SUBREG && REG_P (SUBREG_REG (op0))))) | |
7181 | { | |
7182 | op0 = XEXP (x, 1); | |
7183 | op1 = XEXP (x, 0); | |
7184 | } | |
7185 | goto cost_minus; | |
7186 | } | |
7187 | ||
7188 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT) | |
7189 | { | |
7190 | /* FCMP. */ | |
7191 | if (speed) | |
7192 | *cost += extra_cost->fp[mode == DFmode].compare; | |
7193 | ||
7194 | if (CONST_DOUBLE_P (op1) && aarch64_float_const_zero_rtx_p (op1)) | |
7195 | { | |
e548c9df | 7196 | *cost += rtx_cost (op0, VOIDmode, COMPARE, 0, speed); |
a8eecd00 JG |
7197 | /* FCMP supports constant 0.0 for no extra cost. */ |
7198 | return true; | |
7199 | } | |
7200 | return false; | |
7201 | } | |
7202 | ||
b6875aac KV |
7203 | if (VECTOR_MODE_P (mode)) |
7204 | { | |
7205 | /* Vector compare. */ | |
7206 | if (speed) | |
7207 | *cost += extra_cost->vect.alu; | |
7208 | ||
7209 | if (aarch64_float_const_zero_rtx_p (op1)) | |
7210 | { | |
7211 | /* Vector cm (eq|ge|gt|lt|le) supports constant 0.0 for no extra | |
7212 | cost. */ | |
7213 | return true; | |
7214 | } | |
7215 | return false; | |
7216 | } | |
a8eecd00 | 7217 | return false; |
43e9d192 IB |
7218 | |
7219 | case MINUS: | |
4745e701 JG |
7220 | { |
7221 | op0 = XEXP (x, 0); | |
7222 | op1 = XEXP (x, 1); | |
7223 | ||
7224 | cost_minus: | |
e548c9df | 7225 | *cost += rtx_cost (op0, mode, MINUS, 0, speed); |
23cb6618 | 7226 | |
4745e701 JG |
7227 | /* Detect valid immediates. */ |
7228 | if ((GET_MODE_CLASS (mode) == MODE_INT | |
7229 | || (GET_MODE_CLASS (mode) == MODE_CC | |
7230 | && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT)) | |
7231 | && CONST_INT_P (op1) | |
7232 | && aarch64_uimm12_shift (INTVAL (op1))) | |
7233 | { | |
4745e701 JG |
7234 | if (speed) |
7235 | /* SUB(S) (immediate). */ | |
7236 | *cost += extra_cost->alu.arith; | |
7237 | return true; | |
4745e701 JG |
7238 | } |
7239 | ||
7cc2145f JG |
7240 | /* Look for SUB (extended register). */ |
7241 | if (aarch64_rtx_arith_op_extract_p (op1, mode)) | |
7242 | { | |
7243 | if (speed) | |
2533c820 | 7244 | *cost += extra_cost->alu.extend_arith; |
7cc2145f | 7245 | |
b10f1009 | 7246 | op1 = aarch64_strip_extend (op1, true); |
e47c4031 | 7247 | *cost += rtx_cost (op1, VOIDmode, |
e548c9df | 7248 | (enum rtx_code) GET_CODE (op1), 0, speed); |
7cc2145f JG |
7249 | return true; |
7250 | } | |
7251 | ||
b10f1009 | 7252 | rtx new_op1 = aarch64_strip_extend (op1, false); |
4745e701 JG |
7253 | |
7254 | /* Cost this as an FMA-alike operation. */ | |
7255 | if ((GET_CODE (new_op1) == MULT | |
0a78ebe4 | 7256 | || aarch64_shift_p (GET_CODE (new_op1))) |
4745e701 JG |
7257 | && code != COMPARE) |
7258 | { | |
7259 | *cost += aarch64_rtx_mult_cost (new_op1, MULT, | |
7260 | (enum rtx_code) code, | |
7261 | speed); | |
4745e701 JG |
7262 | return true; |
7263 | } | |
43e9d192 | 7264 | |
e548c9df | 7265 | *cost += rtx_cost (new_op1, VOIDmode, MINUS, 1, speed); |
43e9d192 | 7266 | |
4745e701 JG |
7267 | if (speed) |
7268 | { | |
b6875aac KV |
7269 | if (VECTOR_MODE_P (mode)) |
7270 | { | |
7271 | /* Vector SUB. */ | |
7272 | *cost += extra_cost->vect.alu; | |
7273 | } | |
7274 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
7275 | { | |
7276 | /* SUB(S). */ | |
7277 | *cost += extra_cost->alu.arith; | |
7278 | } | |
4745e701 | 7279 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
b6875aac KV |
7280 | { |
7281 | /* FSUB. */ | |
7282 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
7283 | } | |
4745e701 JG |
7284 | } |
7285 | return true; | |
7286 | } | |
43e9d192 IB |
7287 | |
7288 | case PLUS: | |
4745e701 JG |
7289 | { |
7290 | rtx new_op0; | |
43e9d192 | 7291 | |
4745e701 JG |
7292 | op0 = XEXP (x, 0); |
7293 | op1 = XEXP (x, 1); | |
43e9d192 | 7294 | |
a8eecd00 | 7295 | cost_plus: |
4745e701 JG |
7296 | if (GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMPARE |
7297 | || GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMM_COMPARE) | |
7298 | { | |
7299 | /* CSINC. */ | |
e548c9df AM |
7300 | *cost += rtx_cost (XEXP (op0, 0), mode, PLUS, 0, speed); |
7301 | *cost += rtx_cost (op1, mode, PLUS, 1, speed); | |
4745e701 JG |
7302 | return true; |
7303 | } | |
43e9d192 | 7304 | |
4745e701 JG |
7305 | if (GET_MODE_CLASS (mode) == MODE_INT |
7306 | && CONST_INT_P (op1) | |
7307 | && aarch64_uimm12_shift (INTVAL (op1))) | |
7308 | { | |
e548c9df | 7309 | *cost += rtx_cost (op0, mode, PLUS, 0, speed); |
43e9d192 | 7310 | |
4745e701 JG |
7311 | if (speed) |
7312 | /* ADD (immediate). */ | |
7313 | *cost += extra_cost->alu.arith; | |
7314 | return true; | |
7315 | } | |
7316 | ||
e548c9df | 7317 | *cost += rtx_cost (op1, mode, PLUS, 1, speed); |
23cb6618 | 7318 | |
7cc2145f JG |
7319 | /* Look for ADD (extended register). */ |
7320 | if (aarch64_rtx_arith_op_extract_p (op0, mode)) | |
7321 | { | |
7322 | if (speed) | |
2533c820 | 7323 | *cost += extra_cost->alu.extend_arith; |
7cc2145f | 7324 | |
b10f1009 | 7325 | op0 = aarch64_strip_extend (op0, true); |
e47c4031 | 7326 | *cost += rtx_cost (op0, VOIDmode, |
e548c9df | 7327 | (enum rtx_code) GET_CODE (op0), 0, speed); |
7cc2145f JG |
7328 | return true; |
7329 | } | |
7330 | ||
4745e701 JG |
7331 | /* Strip any extend, leave shifts behind as we will |
7332 | cost them through mult_cost. */ | |
b10f1009 | 7333 | new_op0 = aarch64_strip_extend (op0, false); |
4745e701 JG |
7334 | |
7335 | if (GET_CODE (new_op0) == MULT | |
0a78ebe4 | 7336 | || aarch64_shift_p (GET_CODE (new_op0))) |
4745e701 JG |
7337 | { |
7338 | *cost += aarch64_rtx_mult_cost (new_op0, MULT, PLUS, | |
7339 | speed); | |
4745e701 JG |
7340 | return true; |
7341 | } | |
7342 | ||
e548c9df | 7343 | *cost += rtx_cost (new_op0, VOIDmode, PLUS, 0, speed); |
4745e701 JG |
7344 | |
7345 | if (speed) | |
7346 | { | |
b6875aac KV |
7347 | if (VECTOR_MODE_P (mode)) |
7348 | { | |
7349 | /* Vector ADD. */ | |
7350 | *cost += extra_cost->vect.alu; | |
7351 | } | |
7352 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
7353 | { | |
7354 | /* ADD. */ | |
7355 | *cost += extra_cost->alu.arith; | |
7356 | } | |
4745e701 | 7357 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
b6875aac KV |
7358 | { |
7359 | /* FADD. */ | |
7360 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
7361 | } | |
4745e701 JG |
7362 | } |
7363 | return true; | |
7364 | } | |
43e9d192 | 7365 | |
18b42b2a KT |
7366 | case BSWAP: |
7367 | *cost = COSTS_N_INSNS (1); | |
7368 | ||
7369 | if (speed) | |
b6875aac KV |
7370 | { |
7371 | if (VECTOR_MODE_P (mode)) | |
7372 | *cost += extra_cost->vect.alu; | |
7373 | else | |
7374 | *cost += extra_cost->alu.rev; | |
7375 | } | |
18b42b2a KT |
7376 | return false; |
7377 | ||
43e9d192 | 7378 | case IOR: |
f7d5cf8d KT |
7379 | if (aarch_rev16_p (x)) |
7380 | { | |
7381 | *cost = COSTS_N_INSNS (1); | |
7382 | ||
b6875aac KV |
7383 | if (speed) |
7384 | { | |
7385 | if (VECTOR_MODE_P (mode)) | |
7386 | *cost += extra_cost->vect.alu; | |
7387 | else | |
7388 | *cost += extra_cost->alu.rev; | |
7389 | } | |
7390 | return true; | |
f7d5cf8d | 7391 | } |
fb0cb7fa KT |
7392 | |
7393 | if (aarch64_extr_rtx_p (x, &op0, &op1)) | |
7394 | { | |
e548c9df AM |
7395 | *cost += rtx_cost (op0, mode, IOR, 0, speed); |
7396 | *cost += rtx_cost (op1, mode, IOR, 1, speed); | |
fb0cb7fa KT |
7397 | if (speed) |
7398 | *cost += extra_cost->alu.shift; | |
7399 | ||
7400 | return true; | |
7401 | } | |
f7d5cf8d | 7402 | /* Fall through. */ |
43e9d192 IB |
7403 | case XOR: |
7404 | case AND: | |
7405 | cost_logic: | |
7406 | op0 = XEXP (x, 0); | |
7407 | op1 = XEXP (x, 1); | |
7408 | ||
b6875aac KV |
7409 | if (VECTOR_MODE_P (mode)) |
7410 | { | |
7411 | if (speed) | |
7412 | *cost += extra_cost->vect.alu; | |
7413 | return true; | |
7414 | } | |
7415 | ||
268c3b47 JG |
7416 | if (code == AND |
7417 | && GET_CODE (op0) == MULT | |
7418 | && CONST_INT_P (XEXP (op0, 1)) | |
7419 | && CONST_INT_P (op1) | |
7420 | && aarch64_uxt_size (exact_log2 (INTVAL (XEXP (op0, 1))), | |
7421 | INTVAL (op1)) != 0) | |
7422 | { | |
7423 | /* This is a UBFM/SBFM. */ | |
e548c9df | 7424 | *cost += rtx_cost (XEXP (op0, 0), mode, ZERO_EXTRACT, 0, speed); |
268c3b47 JG |
7425 | if (speed) |
7426 | *cost += extra_cost->alu.bfx; | |
7427 | return true; | |
7428 | } | |
7429 | ||
b4206259 | 7430 | if (is_int_mode (mode, &int_mode)) |
43e9d192 | 7431 | { |
8c83f71d | 7432 | if (CONST_INT_P (op1)) |
43e9d192 | 7433 | { |
8c83f71d KT |
7434 | /* We have a mask + shift version of a UBFIZ |
7435 | i.e. the *andim_ashift<mode>_bfiz pattern. */ | |
7436 | if (GET_CODE (op0) == ASHIFT | |
b4206259 RS |
7437 | && aarch64_mask_and_shift_for_ubfiz_p (int_mode, op1, |
7438 | XEXP (op0, 1))) | |
8c83f71d | 7439 | { |
b4206259 | 7440 | *cost += rtx_cost (XEXP (op0, 0), int_mode, |
8c83f71d KT |
7441 | (enum rtx_code) code, 0, speed); |
7442 | if (speed) | |
7443 | *cost += extra_cost->alu.bfx; | |
268c3b47 | 7444 | |
8c83f71d KT |
7445 | return true; |
7446 | } | |
b4206259 | 7447 | else if (aarch64_bitmask_imm (INTVAL (op1), int_mode)) |
8c83f71d KT |
7448 | { |
7449 | /* We possibly get the immediate for free, this is not | |
7450 | modelled. */ | |
b4206259 RS |
7451 | *cost += rtx_cost (op0, int_mode, |
7452 | (enum rtx_code) code, 0, speed); | |
8c83f71d KT |
7453 | if (speed) |
7454 | *cost += extra_cost->alu.logical; | |
268c3b47 | 7455 | |
8c83f71d KT |
7456 | return true; |
7457 | } | |
43e9d192 IB |
7458 | } |
7459 | else | |
7460 | { | |
268c3b47 JG |
7461 | rtx new_op0 = op0; |
7462 | ||
7463 | /* Handle ORN, EON, or BIC. */ | |
43e9d192 IB |
7464 | if (GET_CODE (op0) == NOT) |
7465 | op0 = XEXP (op0, 0); | |
268c3b47 JG |
7466 | |
7467 | new_op0 = aarch64_strip_shift (op0); | |
7468 | ||
7469 | /* If we had a shift on op0 then this is a logical-shift- | |
7470 | by-register/immediate operation. Otherwise, this is just | |
7471 | a logical operation. */ | |
7472 | if (speed) | |
7473 | { | |
7474 | if (new_op0 != op0) | |
7475 | { | |
7476 | /* Shift by immediate. */ | |
7477 | if (CONST_INT_P (XEXP (op0, 1))) | |
7478 | *cost += extra_cost->alu.log_shift; | |
7479 | else | |
7480 | *cost += extra_cost->alu.log_shift_reg; | |
7481 | } | |
7482 | else | |
7483 | *cost += extra_cost->alu.logical; | |
7484 | } | |
7485 | ||
7486 | /* In both cases we want to cost both operands. */ | |
b4206259 RS |
7487 | *cost += rtx_cost (new_op0, int_mode, (enum rtx_code) code, |
7488 | 0, speed); | |
7489 | *cost += rtx_cost (op1, int_mode, (enum rtx_code) code, | |
7490 | 1, speed); | |
268c3b47 JG |
7491 | |
7492 | return true; | |
43e9d192 | 7493 | } |
43e9d192 IB |
7494 | } |
7495 | return false; | |
7496 | ||
268c3b47 | 7497 | case NOT: |
6365da9e KT |
7498 | x = XEXP (x, 0); |
7499 | op0 = aarch64_strip_shift (x); | |
7500 | ||
b6875aac KV |
7501 | if (VECTOR_MODE_P (mode)) |
7502 | { | |
7503 | /* Vector NOT. */ | |
7504 | *cost += extra_cost->vect.alu; | |
7505 | return false; | |
7506 | } | |
7507 | ||
6365da9e KT |
7508 | /* MVN-shifted-reg. */ |
7509 | if (op0 != x) | |
7510 | { | |
e548c9df | 7511 | *cost += rtx_cost (op0, mode, (enum rtx_code) code, 0, speed); |
6365da9e KT |
7512 | |
7513 | if (speed) | |
7514 | *cost += extra_cost->alu.log_shift; | |
7515 | ||
7516 | return true; | |
7517 | } | |
7518 | /* EON can have two forms: (xor (not a) b) but also (not (xor a b)). | |
7519 | Handle the second form here taking care that 'a' in the above can | |
7520 | be a shift. */ | |
7521 | else if (GET_CODE (op0) == XOR) | |
7522 | { | |
7523 | rtx newop0 = XEXP (op0, 0); | |
7524 | rtx newop1 = XEXP (op0, 1); | |
7525 | rtx op0_stripped = aarch64_strip_shift (newop0); | |
7526 | ||
e548c9df AM |
7527 | *cost += rtx_cost (newop1, mode, (enum rtx_code) code, 1, speed); |
7528 | *cost += rtx_cost (op0_stripped, mode, XOR, 0, speed); | |
6365da9e KT |
7529 | |
7530 | if (speed) | |
7531 | { | |
7532 | if (op0_stripped != newop0) | |
7533 | *cost += extra_cost->alu.log_shift; | |
7534 | else | |
7535 | *cost += extra_cost->alu.logical; | |
7536 | } | |
7537 | ||
7538 | return true; | |
7539 | } | |
268c3b47 JG |
7540 | /* MVN. */ |
7541 | if (speed) | |
7542 | *cost += extra_cost->alu.logical; | |
7543 | ||
268c3b47 JG |
7544 | return false; |
7545 | ||
43e9d192 | 7546 | case ZERO_EXTEND: |
b1685e62 JG |
7547 | |
7548 | op0 = XEXP (x, 0); | |
7549 | /* If a value is written in SI mode, then zero extended to DI | |
7550 | mode, the operation will in general be free as a write to | |
7551 | a 'w' register implicitly zeroes the upper bits of an 'x' | |
7552 | register. However, if this is | |
7553 | ||
7554 | (set (reg) (zero_extend (reg))) | |
7555 | ||
7556 | we must cost the explicit register move. */ | |
7557 | if (mode == DImode | |
7558 | && GET_MODE (op0) == SImode | |
7559 | && outer == SET) | |
7560 | { | |
e548c9df | 7561 | int op_cost = rtx_cost (op0, VOIDmode, ZERO_EXTEND, 0, speed); |
b1685e62 | 7562 | |
dde23f43 KM |
7563 | /* If OP_COST is non-zero, then the cost of the zero extend |
7564 | is effectively the cost of the inner operation. Otherwise | |
7565 | we have a MOV instruction and we take the cost from the MOV | |
7566 | itself. This is true independently of whether we are | |
7567 | optimizing for space or time. */ | |
7568 | if (op_cost) | |
b1685e62 JG |
7569 | *cost = op_cost; |
7570 | ||
7571 | return true; | |
7572 | } | |
e548c9df | 7573 | else if (MEM_P (op0)) |
43e9d192 | 7574 | { |
b1685e62 | 7575 | /* All loads can zero extend to any size for free. */ |
e548c9df | 7576 | *cost = rtx_cost (op0, VOIDmode, ZERO_EXTEND, param, speed); |
43e9d192 IB |
7577 | return true; |
7578 | } | |
b1685e62 | 7579 | |
283b6c85 KT |
7580 | op0 = aarch64_extend_bitfield_pattern_p (x); |
7581 | if (op0) | |
7582 | { | |
7583 | *cost += rtx_cost (op0, mode, ZERO_EXTEND, 0, speed); | |
7584 | if (speed) | |
7585 | *cost += extra_cost->alu.bfx; | |
7586 | return true; | |
7587 | } | |
7588 | ||
b1685e62 | 7589 | if (speed) |
b6875aac KV |
7590 | { |
7591 | if (VECTOR_MODE_P (mode)) | |
7592 | { | |
7593 | /* UMOV. */ | |
7594 | *cost += extra_cost->vect.alu; | |
7595 | } | |
7596 | else | |
7597 | { | |
63715e5e WD |
7598 | /* We generate an AND instead of UXTB/UXTH. */ |
7599 | *cost += extra_cost->alu.logical; | |
b6875aac KV |
7600 | } |
7601 | } | |
43e9d192 IB |
7602 | return false; |
7603 | ||
7604 | case SIGN_EXTEND: | |
b1685e62 | 7605 | if (MEM_P (XEXP (x, 0))) |
43e9d192 | 7606 | { |
b1685e62 JG |
7607 | /* LDRSH. */ |
7608 | if (speed) | |
7609 | { | |
7610 | rtx address = XEXP (XEXP (x, 0), 0); | |
7611 | *cost += extra_cost->ldst.load_sign_extend; | |
7612 | ||
7613 | *cost += | |
7614 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
7615 | 0, speed)); | |
7616 | } | |
43e9d192 IB |
7617 | return true; |
7618 | } | |
b1685e62 | 7619 | |
283b6c85 KT |
7620 | op0 = aarch64_extend_bitfield_pattern_p (x); |
7621 | if (op0) | |
7622 | { | |
7623 | *cost += rtx_cost (op0, mode, SIGN_EXTEND, 0, speed); | |
7624 | if (speed) | |
7625 | *cost += extra_cost->alu.bfx; | |
7626 | return true; | |
7627 | } | |
7628 | ||
b1685e62 | 7629 | if (speed) |
b6875aac KV |
7630 | { |
7631 | if (VECTOR_MODE_P (mode)) | |
7632 | *cost += extra_cost->vect.alu; | |
7633 | else | |
7634 | *cost += extra_cost->alu.extend; | |
7635 | } | |
43e9d192 IB |
7636 | return false; |
7637 | ||
ba0cfa17 JG |
7638 | case ASHIFT: |
7639 | op0 = XEXP (x, 0); | |
7640 | op1 = XEXP (x, 1); | |
7641 | ||
7642 | if (CONST_INT_P (op1)) | |
7643 | { | |
ba0cfa17 | 7644 | if (speed) |
b6875aac KV |
7645 | { |
7646 | if (VECTOR_MODE_P (mode)) | |
7647 | { | |
7648 | /* Vector shift (immediate). */ | |
7649 | *cost += extra_cost->vect.alu; | |
7650 | } | |
7651 | else | |
7652 | { | |
7653 | /* LSL (immediate), UBMF, UBFIZ and friends. These are all | |
7654 | aliases. */ | |
7655 | *cost += extra_cost->alu.shift; | |
7656 | } | |
7657 | } | |
ba0cfa17 JG |
7658 | |
7659 | /* We can incorporate zero/sign extend for free. */ | |
7660 | if (GET_CODE (op0) == ZERO_EXTEND | |
7661 | || GET_CODE (op0) == SIGN_EXTEND) | |
7662 | op0 = XEXP (op0, 0); | |
7663 | ||
e548c9df | 7664 | *cost += rtx_cost (op0, VOIDmode, ASHIFT, 0, speed); |
ba0cfa17 JG |
7665 | return true; |
7666 | } | |
7667 | else | |
7668 | { | |
7813b280 | 7669 | if (VECTOR_MODE_P (mode)) |
b6875aac | 7670 | { |
7813b280 KT |
7671 | if (speed) |
7672 | /* Vector shift (register). */ | |
7673 | *cost += extra_cost->vect.alu; | |
7674 | } | |
7675 | else | |
7676 | { | |
7677 | if (speed) | |
7678 | /* LSLV. */ | |
7679 | *cost += extra_cost->alu.shift_reg; | |
7680 | ||
7681 | if (GET_CODE (op1) == AND && REG_P (XEXP (op1, 0)) | |
7682 | && CONST_INT_P (XEXP (op1, 1)) | |
7683 | && INTVAL (XEXP (op1, 1)) == GET_MODE_BITSIZE (mode) - 1) | |
b6875aac | 7684 | { |
7813b280 KT |
7685 | *cost += rtx_cost (op0, mode, (rtx_code) code, 0, speed); |
7686 | /* We already demanded XEXP (op1, 0) to be REG_P, so | |
7687 | don't recurse into it. */ | |
7688 | return true; | |
b6875aac KV |
7689 | } |
7690 | } | |
ba0cfa17 JG |
7691 | return false; /* All arguments need to be in registers. */ |
7692 | } | |
7693 | ||
43e9d192 | 7694 | case ROTATE: |
43e9d192 IB |
7695 | case ROTATERT: |
7696 | case LSHIFTRT: | |
43e9d192 | 7697 | case ASHIFTRT: |
ba0cfa17 JG |
7698 | op0 = XEXP (x, 0); |
7699 | op1 = XEXP (x, 1); | |
43e9d192 | 7700 | |
ba0cfa17 JG |
7701 | if (CONST_INT_P (op1)) |
7702 | { | |
7703 | /* ASR (immediate) and friends. */ | |
7704 | if (speed) | |
b6875aac KV |
7705 | { |
7706 | if (VECTOR_MODE_P (mode)) | |
7707 | *cost += extra_cost->vect.alu; | |
7708 | else | |
7709 | *cost += extra_cost->alu.shift; | |
7710 | } | |
43e9d192 | 7711 | |
e548c9df | 7712 | *cost += rtx_cost (op0, mode, (enum rtx_code) code, 0, speed); |
ba0cfa17 JG |
7713 | return true; |
7714 | } | |
7715 | else | |
7716 | { | |
7813b280 | 7717 | if (VECTOR_MODE_P (mode)) |
b6875aac | 7718 | { |
7813b280 KT |
7719 | if (speed) |
7720 | /* Vector shift (register). */ | |
b6875aac | 7721 | *cost += extra_cost->vect.alu; |
7813b280 KT |
7722 | } |
7723 | else | |
7724 | { | |
7725 | if (speed) | |
7726 | /* ASR (register) and friends. */ | |
b6875aac | 7727 | *cost += extra_cost->alu.shift_reg; |
7813b280 KT |
7728 | |
7729 | if (GET_CODE (op1) == AND && REG_P (XEXP (op1, 0)) | |
7730 | && CONST_INT_P (XEXP (op1, 1)) | |
7731 | && INTVAL (XEXP (op1, 1)) == GET_MODE_BITSIZE (mode) - 1) | |
7732 | { | |
7733 | *cost += rtx_cost (op0, mode, (rtx_code) code, 0, speed); | |
7734 | /* We already demanded XEXP (op1, 0) to be REG_P, so | |
7735 | don't recurse into it. */ | |
7736 | return true; | |
7737 | } | |
b6875aac | 7738 | } |
ba0cfa17 JG |
7739 | return false; /* All arguments need to be in registers. */ |
7740 | } | |
43e9d192 | 7741 | |
909734be JG |
7742 | case SYMBOL_REF: |
7743 | ||
1b1e81f8 JW |
7744 | if (aarch64_cmodel == AARCH64_CMODEL_LARGE |
7745 | || aarch64_cmodel == AARCH64_CMODEL_SMALL_SPIC) | |
909734be JG |
7746 | { |
7747 | /* LDR. */ | |
7748 | if (speed) | |
7749 | *cost += extra_cost->ldst.load; | |
7750 | } | |
7751 | else if (aarch64_cmodel == AARCH64_CMODEL_SMALL | |
7752 | || aarch64_cmodel == AARCH64_CMODEL_SMALL_PIC) | |
7753 | { | |
7754 | /* ADRP, followed by ADD. */ | |
7755 | *cost += COSTS_N_INSNS (1); | |
7756 | if (speed) | |
7757 | *cost += 2 * extra_cost->alu.arith; | |
7758 | } | |
7759 | else if (aarch64_cmodel == AARCH64_CMODEL_TINY | |
7760 | || aarch64_cmodel == AARCH64_CMODEL_TINY_PIC) | |
7761 | { | |
7762 | /* ADR. */ | |
7763 | if (speed) | |
7764 | *cost += extra_cost->alu.arith; | |
7765 | } | |
7766 | ||
7767 | if (flag_pic) | |
7768 | { | |
7769 | /* One extra load instruction, after accessing the GOT. */ | |
7770 | *cost += COSTS_N_INSNS (1); | |
7771 | if (speed) | |
7772 | *cost += extra_cost->ldst.load; | |
7773 | } | |
43e9d192 IB |
7774 | return true; |
7775 | ||
909734be | 7776 | case HIGH: |
43e9d192 | 7777 | case LO_SUM: |
909734be JG |
7778 | /* ADRP/ADD (immediate). */ |
7779 | if (speed) | |
7780 | *cost += extra_cost->alu.arith; | |
43e9d192 IB |
7781 | return true; |
7782 | ||
7783 | case ZERO_EXTRACT: | |
7784 | case SIGN_EXTRACT: | |
7cc2145f JG |
7785 | /* UBFX/SBFX. */ |
7786 | if (speed) | |
b6875aac KV |
7787 | { |
7788 | if (VECTOR_MODE_P (mode)) | |
7789 | *cost += extra_cost->vect.alu; | |
7790 | else | |
7791 | *cost += extra_cost->alu.bfx; | |
7792 | } | |
7cc2145f JG |
7793 | |
7794 | /* We can trust that the immediates used will be correct (there | |
7795 | are no by-register forms), so we need only cost op0. */ | |
e548c9df | 7796 | *cost += rtx_cost (XEXP (x, 0), VOIDmode, (enum rtx_code) code, 0, speed); |
43e9d192 IB |
7797 | return true; |
7798 | ||
7799 | case MULT: | |
4745e701 JG |
7800 | *cost += aarch64_rtx_mult_cost (x, MULT, 0, speed); |
7801 | /* aarch64_rtx_mult_cost always handles recursion to its | |
7802 | operands. */ | |
7803 | return true; | |
43e9d192 IB |
7804 | |
7805 | case MOD: | |
4f58fe36 KT |
7806 | /* We can expand signed mod by power of 2 using a NEGS, two parallel |
7807 | ANDs and a CSNEG. Assume here that CSNEG is the same as the cost of | |
7808 | an unconditional negate. This case should only ever be reached through | |
7809 | the set_smod_pow2_cheap check in expmed.c. */ | |
7810 | if (CONST_INT_P (XEXP (x, 1)) | |
7811 | && exact_log2 (INTVAL (XEXP (x, 1))) > 0 | |
7812 | && (mode == SImode || mode == DImode)) | |
7813 | { | |
7814 | /* We expand to 4 instructions. Reset the baseline. */ | |
7815 | *cost = COSTS_N_INSNS (4); | |
7816 | ||
7817 | if (speed) | |
7818 | *cost += 2 * extra_cost->alu.logical | |
7819 | + 2 * extra_cost->alu.arith; | |
7820 | ||
7821 | return true; | |
7822 | } | |
7823 | ||
7824 | /* Fall-through. */ | |
43e9d192 | 7825 | case UMOD: |
43e9d192 IB |
7826 | if (speed) |
7827 | { | |
cb9ac430 | 7828 | /* Slighly prefer UMOD over SMOD. */ |
b6875aac KV |
7829 | if (VECTOR_MODE_P (mode)) |
7830 | *cost += extra_cost->vect.alu; | |
e548c9df AM |
7831 | else if (GET_MODE_CLASS (mode) == MODE_INT) |
7832 | *cost += (extra_cost->mult[mode == DImode].add | |
cb9ac430 TC |
7833 | + extra_cost->mult[mode == DImode].idiv |
7834 | + (code == MOD ? 1 : 0)); | |
43e9d192 IB |
7835 | } |
7836 | return false; /* All arguments need to be in registers. */ | |
7837 | ||
7838 | case DIV: | |
7839 | case UDIV: | |
4105fe38 | 7840 | case SQRT: |
43e9d192 IB |
7841 | if (speed) |
7842 | { | |
b6875aac KV |
7843 | if (VECTOR_MODE_P (mode)) |
7844 | *cost += extra_cost->vect.alu; | |
7845 | else if (GET_MODE_CLASS (mode) == MODE_INT) | |
4105fe38 JG |
7846 | /* There is no integer SQRT, so only DIV and UDIV can get |
7847 | here. */ | |
cb9ac430 TC |
7848 | *cost += (extra_cost->mult[mode == DImode].idiv |
7849 | /* Slighly prefer UDIV over SDIV. */ | |
7850 | + (code == DIV ? 1 : 0)); | |
4105fe38 JG |
7851 | else |
7852 | *cost += extra_cost->fp[mode == DFmode].div; | |
43e9d192 IB |
7853 | } |
7854 | return false; /* All arguments need to be in registers. */ | |
7855 | ||
a8eecd00 | 7856 | case IF_THEN_ELSE: |
2d5ffe46 AP |
7857 | return aarch64_if_then_else_costs (XEXP (x, 0), XEXP (x, 1), |
7858 | XEXP (x, 2), cost, speed); | |
a8eecd00 JG |
7859 | |
7860 | case EQ: | |
7861 | case NE: | |
7862 | case GT: | |
7863 | case GTU: | |
7864 | case LT: | |
7865 | case LTU: | |
7866 | case GE: | |
7867 | case GEU: | |
7868 | case LE: | |
7869 | case LEU: | |
7870 | ||
7871 | return false; /* All arguments must be in registers. */ | |
7872 | ||
b292109f JG |
7873 | case FMA: |
7874 | op0 = XEXP (x, 0); | |
7875 | op1 = XEXP (x, 1); | |
7876 | op2 = XEXP (x, 2); | |
7877 | ||
7878 | if (speed) | |
b6875aac KV |
7879 | { |
7880 | if (VECTOR_MODE_P (mode)) | |
7881 | *cost += extra_cost->vect.alu; | |
7882 | else | |
7883 | *cost += extra_cost->fp[mode == DFmode].fma; | |
7884 | } | |
b292109f JG |
7885 | |
7886 | /* FMSUB, FNMADD, and FNMSUB are free. */ | |
7887 | if (GET_CODE (op0) == NEG) | |
7888 | op0 = XEXP (op0, 0); | |
7889 | ||
7890 | if (GET_CODE (op2) == NEG) | |
7891 | op2 = XEXP (op2, 0); | |
7892 | ||
7893 | /* aarch64_fnma4_elt_to_64v2df has the NEG as operand 1, | |
7894 | and the by-element operand as operand 0. */ | |
7895 | if (GET_CODE (op1) == NEG) | |
7896 | op1 = XEXP (op1, 0); | |
7897 | ||
7898 | /* Catch vector-by-element operations. The by-element operand can | |
7899 | either be (vec_duplicate (vec_select (x))) or just | |
7900 | (vec_select (x)), depending on whether we are multiplying by | |
7901 | a vector or a scalar. | |
7902 | ||
7903 | Canonicalization is not very good in these cases, FMA4 will put the | |
7904 | by-element operand as operand 0, FNMA4 will have it as operand 1. */ | |
7905 | if (GET_CODE (op0) == VEC_DUPLICATE) | |
7906 | op0 = XEXP (op0, 0); | |
7907 | else if (GET_CODE (op1) == VEC_DUPLICATE) | |
7908 | op1 = XEXP (op1, 0); | |
7909 | ||
7910 | if (GET_CODE (op0) == VEC_SELECT) | |
7911 | op0 = XEXP (op0, 0); | |
7912 | else if (GET_CODE (op1) == VEC_SELECT) | |
7913 | op1 = XEXP (op1, 0); | |
7914 | ||
7915 | /* If the remaining parameters are not registers, | |
7916 | get the cost to put them into registers. */ | |
e548c9df AM |
7917 | *cost += rtx_cost (op0, mode, FMA, 0, speed); |
7918 | *cost += rtx_cost (op1, mode, FMA, 1, speed); | |
7919 | *cost += rtx_cost (op2, mode, FMA, 2, speed); | |
b292109f JG |
7920 | return true; |
7921 | ||
5e2a765b KT |
7922 | case FLOAT: |
7923 | case UNSIGNED_FLOAT: | |
7924 | if (speed) | |
7925 | *cost += extra_cost->fp[mode == DFmode].fromint; | |
7926 | return false; | |
7927 | ||
b292109f JG |
7928 | case FLOAT_EXTEND: |
7929 | if (speed) | |
b6875aac KV |
7930 | { |
7931 | if (VECTOR_MODE_P (mode)) | |
7932 | { | |
7933 | /*Vector truncate. */ | |
7934 | *cost += extra_cost->vect.alu; | |
7935 | } | |
7936 | else | |
7937 | *cost += extra_cost->fp[mode == DFmode].widen; | |
7938 | } | |
b292109f JG |
7939 | return false; |
7940 | ||
7941 | case FLOAT_TRUNCATE: | |
7942 | if (speed) | |
b6875aac KV |
7943 | { |
7944 | if (VECTOR_MODE_P (mode)) | |
7945 | { | |
7946 | /*Vector conversion. */ | |
7947 | *cost += extra_cost->vect.alu; | |
7948 | } | |
7949 | else | |
7950 | *cost += extra_cost->fp[mode == DFmode].narrow; | |
7951 | } | |
b292109f JG |
7952 | return false; |
7953 | ||
61263118 KT |
7954 | case FIX: |
7955 | case UNSIGNED_FIX: | |
7956 | x = XEXP (x, 0); | |
7957 | /* Strip the rounding part. They will all be implemented | |
7958 | by the fcvt* family of instructions anyway. */ | |
7959 | if (GET_CODE (x) == UNSPEC) | |
7960 | { | |
7961 | unsigned int uns_code = XINT (x, 1); | |
7962 | ||
7963 | if (uns_code == UNSPEC_FRINTA | |
7964 | || uns_code == UNSPEC_FRINTM | |
7965 | || uns_code == UNSPEC_FRINTN | |
7966 | || uns_code == UNSPEC_FRINTP | |
7967 | || uns_code == UNSPEC_FRINTZ) | |
7968 | x = XVECEXP (x, 0, 0); | |
7969 | } | |
7970 | ||
7971 | if (speed) | |
b6875aac KV |
7972 | { |
7973 | if (VECTOR_MODE_P (mode)) | |
7974 | *cost += extra_cost->vect.alu; | |
7975 | else | |
7976 | *cost += extra_cost->fp[GET_MODE (x) == DFmode].toint; | |
7977 | } | |
39252973 KT |
7978 | |
7979 | /* We can combine fmul by a power of 2 followed by a fcvt into a single | |
7980 | fixed-point fcvt. */ | |
7981 | if (GET_CODE (x) == MULT | |
7982 | && ((VECTOR_MODE_P (mode) | |
7983 | && aarch64_vec_fpconst_pow_of_2 (XEXP (x, 1)) > 0) | |
7984 | || aarch64_fpconst_pow_of_2 (XEXP (x, 1)) > 0)) | |
7985 | { | |
7986 | *cost += rtx_cost (XEXP (x, 0), VOIDmode, (rtx_code) code, | |
7987 | 0, speed); | |
7988 | return true; | |
7989 | } | |
7990 | ||
e548c9df | 7991 | *cost += rtx_cost (x, VOIDmode, (enum rtx_code) code, 0, speed); |
61263118 KT |
7992 | return true; |
7993 | ||
b292109f | 7994 | case ABS: |
b6875aac KV |
7995 | if (VECTOR_MODE_P (mode)) |
7996 | { | |
7997 | /* ABS (vector). */ | |
7998 | if (speed) | |
7999 | *cost += extra_cost->vect.alu; | |
8000 | } | |
8001 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
b292109f | 8002 | { |
19261b99 KT |
8003 | op0 = XEXP (x, 0); |
8004 | ||
8005 | /* FABD, which is analogous to FADD. */ | |
8006 | if (GET_CODE (op0) == MINUS) | |
8007 | { | |
e548c9df AM |
8008 | *cost += rtx_cost (XEXP (op0, 0), mode, MINUS, 0, speed); |
8009 | *cost += rtx_cost (XEXP (op0, 1), mode, MINUS, 1, speed); | |
19261b99 KT |
8010 | if (speed) |
8011 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
8012 | ||
8013 | return true; | |
8014 | } | |
8015 | /* Simple FABS is analogous to FNEG. */ | |
b292109f JG |
8016 | if (speed) |
8017 | *cost += extra_cost->fp[mode == DFmode].neg; | |
8018 | } | |
8019 | else | |
8020 | { | |
8021 | /* Integer ABS will either be split to | |
8022 | two arithmetic instructions, or will be an ABS | |
8023 | (scalar), which we don't model. */ | |
8024 | *cost = COSTS_N_INSNS (2); | |
8025 | if (speed) | |
8026 | *cost += 2 * extra_cost->alu.arith; | |
8027 | } | |
8028 | return false; | |
8029 | ||
8030 | case SMAX: | |
8031 | case SMIN: | |
8032 | if (speed) | |
8033 | { | |
b6875aac KV |
8034 | if (VECTOR_MODE_P (mode)) |
8035 | *cost += extra_cost->vect.alu; | |
8036 | else | |
8037 | { | |
8038 | /* FMAXNM/FMINNM/FMAX/FMIN. | |
8039 | TODO: This may not be accurate for all implementations, but | |
8040 | we do not model this in the cost tables. */ | |
8041 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
8042 | } | |
b292109f JG |
8043 | } |
8044 | return false; | |
8045 | ||
61263118 KT |
8046 | case UNSPEC: |
8047 | /* The floating point round to integer frint* instructions. */ | |
8048 | if (aarch64_frint_unspec_p (XINT (x, 1))) | |
8049 | { | |
8050 | if (speed) | |
8051 | *cost += extra_cost->fp[mode == DFmode].roundint; | |
8052 | ||
8053 | return false; | |
8054 | } | |
781aeb73 KT |
8055 | |
8056 | if (XINT (x, 1) == UNSPEC_RBIT) | |
8057 | { | |
8058 | if (speed) | |
8059 | *cost += extra_cost->alu.rev; | |
8060 | ||
8061 | return false; | |
8062 | } | |
61263118 KT |
8063 | break; |
8064 | ||
fb620c4a JG |
8065 | case TRUNCATE: |
8066 | ||
8067 | /* Decompose <su>muldi3_highpart. */ | |
8068 | if (/* (truncate:DI */ | |
8069 | mode == DImode | |
8070 | /* (lshiftrt:TI */ | |
8071 | && GET_MODE (XEXP (x, 0)) == TImode | |
8072 | && GET_CODE (XEXP (x, 0)) == LSHIFTRT | |
8073 | /* (mult:TI */ | |
8074 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT | |
8075 | /* (ANY_EXTEND:TI (reg:DI)) | |
8076 | (ANY_EXTEND:TI (reg:DI))) */ | |
8077 | && ((GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == ZERO_EXTEND | |
8078 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == ZERO_EXTEND) | |
8079 | || (GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == SIGN_EXTEND | |
8080 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == SIGN_EXTEND)) | |
8081 | && GET_MODE (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 0)) == DImode | |
8082 | && GET_MODE (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 1), 0)) == DImode | |
8083 | /* (const_int 64) */ | |
8084 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
8085 | && UINTVAL (XEXP (XEXP (x, 0), 1)) == 64) | |
8086 | { | |
8087 | /* UMULH/SMULH. */ | |
8088 | if (speed) | |
8089 | *cost += extra_cost->mult[mode == DImode].extend; | |
e548c9df AM |
8090 | *cost += rtx_cost (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 0), |
8091 | mode, MULT, 0, speed); | |
8092 | *cost += rtx_cost (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 1), 0), | |
8093 | mode, MULT, 1, speed); | |
fb620c4a JG |
8094 | return true; |
8095 | } | |
8096 | ||
8097 | /* Fall through. */ | |
43e9d192 | 8098 | default: |
61263118 | 8099 | break; |
43e9d192 | 8100 | } |
61263118 | 8101 | |
c10e3d7f AP |
8102 | if (dump_file |
8103 | && flag_aarch64_verbose_cost) | |
61263118 KT |
8104 | fprintf (dump_file, |
8105 | "\nFailed to cost RTX. Assuming default cost.\n"); | |
8106 | ||
8107 | return true; | |
43e9d192 IB |
8108 | } |
8109 | ||
0ee859b5 JG |
8110 | /* Wrapper around aarch64_rtx_costs, dumps the partial, or total cost |
8111 | calculated for X. This cost is stored in *COST. Returns true | |
8112 | if the total cost of X was calculated. */ | |
8113 | static bool | |
e548c9df | 8114 | aarch64_rtx_costs_wrapper (rtx x, machine_mode mode, int outer, |
0ee859b5 JG |
8115 | int param, int *cost, bool speed) |
8116 | { | |
e548c9df | 8117 | bool result = aarch64_rtx_costs (x, mode, outer, param, cost, speed); |
0ee859b5 | 8118 | |
c10e3d7f AP |
8119 | if (dump_file |
8120 | && flag_aarch64_verbose_cost) | |
0ee859b5 JG |
8121 | { |
8122 | print_rtl_single (dump_file, x); | |
8123 | fprintf (dump_file, "\n%s cost: %d (%s)\n", | |
8124 | speed ? "Hot" : "Cold", | |
8125 | *cost, result ? "final" : "partial"); | |
8126 | } | |
8127 | ||
8128 | return result; | |
8129 | } | |
8130 | ||
43e9d192 | 8131 | static int |
ef4bddc2 | 8132 | aarch64_register_move_cost (machine_mode mode, |
8a3a7e67 | 8133 | reg_class_t from_i, reg_class_t to_i) |
43e9d192 | 8134 | { |
8a3a7e67 RH |
8135 | enum reg_class from = (enum reg_class) from_i; |
8136 | enum reg_class to = (enum reg_class) to_i; | |
43e9d192 | 8137 | const struct cpu_regmove_cost *regmove_cost |
b175b679 | 8138 | = aarch64_tune_params.regmove_cost; |
43e9d192 | 8139 | |
3be07662 | 8140 | /* Caller save and pointer regs are equivalent to GENERAL_REGS. */ |
2876a13f | 8141 | if (to == CALLER_SAVE_REGS || to == POINTER_REGS) |
3be07662 WD |
8142 | to = GENERAL_REGS; |
8143 | ||
2876a13f | 8144 | if (from == CALLER_SAVE_REGS || from == POINTER_REGS) |
3be07662 WD |
8145 | from = GENERAL_REGS; |
8146 | ||
6ee70f81 AP |
8147 | /* Moving between GPR and stack cost is the same as GP2GP. */ |
8148 | if ((from == GENERAL_REGS && to == STACK_REG) | |
8149 | || (to == GENERAL_REGS && from == STACK_REG)) | |
8150 | return regmove_cost->GP2GP; | |
8151 | ||
8152 | /* To/From the stack register, we move via the gprs. */ | |
8153 | if (to == STACK_REG || from == STACK_REG) | |
8154 | return aarch64_register_move_cost (mode, from, GENERAL_REGS) | |
8155 | + aarch64_register_move_cost (mode, GENERAL_REGS, to); | |
8156 | ||
8919453c WD |
8157 | if (GET_MODE_SIZE (mode) == 16) |
8158 | { | |
8159 | /* 128-bit operations on general registers require 2 instructions. */ | |
8160 | if (from == GENERAL_REGS && to == GENERAL_REGS) | |
8161 | return regmove_cost->GP2GP * 2; | |
8162 | else if (from == GENERAL_REGS) | |
8163 | return regmove_cost->GP2FP * 2; | |
8164 | else if (to == GENERAL_REGS) | |
8165 | return regmove_cost->FP2GP * 2; | |
8166 | ||
8167 | /* When AdvSIMD instructions are disabled it is not possible to move | |
8168 | a 128-bit value directly between Q registers. This is handled in | |
8169 | secondary reload. A general register is used as a scratch to move | |
8170 | the upper DI value and the lower DI value is moved directly, | |
8171 | hence the cost is the sum of three moves. */ | |
8172 | if (! TARGET_SIMD) | |
8173 | return regmove_cost->GP2FP + regmove_cost->FP2GP + regmove_cost->FP2FP; | |
8174 | ||
8175 | return regmove_cost->FP2FP; | |
8176 | } | |
8177 | ||
43e9d192 IB |
8178 | if (from == GENERAL_REGS && to == GENERAL_REGS) |
8179 | return regmove_cost->GP2GP; | |
8180 | else if (from == GENERAL_REGS) | |
8181 | return regmove_cost->GP2FP; | |
8182 | else if (to == GENERAL_REGS) | |
8183 | return regmove_cost->FP2GP; | |
8184 | ||
43e9d192 IB |
8185 | return regmove_cost->FP2FP; |
8186 | } | |
8187 | ||
8188 | static int | |
ef4bddc2 | 8189 | aarch64_memory_move_cost (machine_mode mode ATTRIBUTE_UNUSED, |
43e9d192 IB |
8190 | reg_class_t rclass ATTRIBUTE_UNUSED, |
8191 | bool in ATTRIBUTE_UNUSED) | |
8192 | { | |
b175b679 | 8193 | return aarch64_tune_params.memmov_cost; |
43e9d192 IB |
8194 | } |
8195 | ||
0c30e0f3 EM |
8196 | /* Return true if it is safe and beneficial to use the approximate rsqrt optabs |
8197 | to optimize 1.0/sqrt. */ | |
ee62a5a6 RS |
8198 | |
8199 | static bool | |
9acc9cbe | 8200 | use_rsqrt_p (machine_mode mode) |
ee62a5a6 RS |
8201 | { |
8202 | return (!flag_trapping_math | |
8203 | && flag_unsafe_math_optimizations | |
9acc9cbe EM |
8204 | && ((aarch64_tune_params.approx_modes->recip_sqrt |
8205 | & AARCH64_APPROX_MODE (mode)) | |
1a33079e | 8206 | || flag_mrecip_low_precision_sqrt)); |
ee62a5a6 RS |
8207 | } |
8208 | ||
0c30e0f3 EM |
8209 | /* Function to decide when to use the approximate reciprocal square root |
8210 | builtin. */ | |
a6fc00da BH |
8211 | |
8212 | static tree | |
ee62a5a6 | 8213 | aarch64_builtin_reciprocal (tree fndecl) |
a6fc00da | 8214 | { |
9acc9cbe EM |
8215 | machine_mode mode = TYPE_MODE (TREE_TYPE (fndecl)); |
8216 | ||
8217 | if (!use_rsqrt_p (mode)) | |
a6fc00da | 8218 | return NULL_TREE; |
ee62a5a6 | 8219 | return aarch64_builtin_rsqrt (DECL_FUNCTION_CODE (fndecl)); |
a6fc00da BH |
8220 | } |
8221 | ||
8222 | typedef rtx (*rsqrte_type) (rtx, rtx); | |
8223 | ||
98daafa0 EM |
8224 | /* Select reciprocal square root initial estimate insn depending on machine |
8225 | mode. */ | |
a6fc00da | 8226 | |
98daafa0 | 8227 | static rsqrte_type |
a6fc00da BH |
8228 | get_rsqrte_type (machine_mode mode) |
8229 | { | |
8230 | switch (mode) | |
8231 | { | |
4e10a5a7 RS |
8232 | case E_DFmode: return gen_aarch64_rsqrtedf; |
8233 | case E_SFmode: return gen_aarch64_rsqrtesf; | |
8234 | case E_V2DFmode: return gen_aarch64_rsqrtev2df; | |
8235 | case E_V2SFmode: return gen_aarch64_rsqrtev2sf; | |
8236 | case E_V4SFmode: return gen_aarch64_rsqrtev4sf; | |
a6fc00da BH |
8237 | default: gcc_unreachable (); |
8238 | } | |
8239 | } | |
8240 | ||
8241 | typedef rtx (*rsqrts_type) (rtx, rtx, rtx); | |
8242 | ||
98daafa0 | 8243 | /* Select reciprocal square root series step insn depending on machine mode. */ |
a6fc00da | 8244 | |
98daafa0 | 8245 | static rsqrts_type |
a6fc00da BH |
8246 | get_rsqrts_type (machine_mode mode) |
8247 | { | |
8248 | switch (mode) | |
8249 | { | |
4e10a5a7 RS |
8250 | case E_DFmode: return gen_aarch64_rsqrtsdf; |
8251 | case E_SFmode: return gen_aarch64_rsqrtssf; | |
8252 | case E_V2DFmode: return gen_aarch64_rsqrtsv2df; | |
8253 | case E_V2SFmode: return gen_aarch64_rsqrtsv2sf; | |
8254 | case E_V4SFmode: return gen_aarch64_rsqrtsv4sf; | |
a6fc00da BH |
8255 | default: gcc_unreachable (); |
8256 | } | |
8257 | } | |
8258 | ||
98daafa0 EM |
8259 | /* Emit instruction sequence to compute either the approximate square root |
8260 | or its approximate reciprocal, depending on the flag RECP, and return | |
8261 | whether the sequence was emitted or not. */ | |
a6fc00da | 8262 | |
98daafa0 EM |
8263 | bool |
8264 | aarch64_emit_approx_sqrt (rtx dst, rtx src, bool recp) | |
a6fc00da | 8265 | { |
98daafa0 | 8266 | machine_mode mode = GET_MODE (dst); |
daef0a8c JW |
8267 | |
8268 | if (GET_MODE_INNER (mode) == HFmode) | |
2e19adc8 RE |
8269 | { |
8270 | gcc_assert (!recp); | |
8271 | return false; | |
8272 | } | |
8273 | ||
8274 | machine_mode mmsk | |
304b9962 | 8275 | = mode_for_vector (int_mode_for_mode (GET_MODE_INNER (mode)).require (), |
2e19adc8 RE |
8276 | GET_MODE_NUNITS (mode)); |
8277 | if (!recp) | |
8278 | { | |
8279 | if (!(flag_mlow_precision_sqrt | |
8280 | || (aarch64_tune_params.approx_modes->sqrt | |
8281 | & AARCH64_APPROX_MODE (mode)))) | |
8282 | return false; | |
8283 | ||
8284 | if (flag_finite_math_only | |
8285 | || flag_trapping_math | |
8286 | || !flag_unsafe_math_optimizations | |
8287 | || optimize_function_for_size_p (cfun)) | |
8288 | return false; | |
8289 | } | |
8290 | else | |
8291 | /* Caller assumes we cannot fail. */ | |
8292 | gcc_assert (use_rsqrt_p (mode)); | |
daef0a8c | 8293 | |
a6fc00da | 8294 | |
98daafa0 EM |
8295 | rtx xmsk = gen_reg_rtx (mmsk); |
8296 | if (!recp) | |
2e19adc8 RE |
8297 | /* When calculating the approximate square root, compare the |
8298 | argument with 0.0 and create a mask. */ | |
8299 | emit_insn (gen_rtx_SET (xmsk, | |
8300 | gen_rtx_NEG (mmsk, | |
8301 | gen_rtx_EQ (mmsk, src, | |
8302 | CONST0_RTX (mode))))); | |
a6fc00da | 8303 | |
98daafa0 EM |
8304 | /* Estimate the approximate reciprocal square root. */ |
8305 | rtx xdst = gen_reg_rtx (mode); | |
8306 | emit_insn ((*get_rsqrte_type (mode)) (xdst, src)); | |
a6fc00da | 8307 | |
98daafa0 EM |
8308 | /* Iterate over the series twice for SF and thrice for DF. */ |
8309 | int iterations = (GET_MODE_INNER (mode) == DFmode) ? 3 : 2; | |
a6fc00da | 8310 | |
98daafa0 EM |
8311 | /* Optionally iterate over the series once less for faster performance |
8312 | while sacrificing the accuracy. */ | |
8313 | if ((recp && flag_mrecip_low_precision_sqrt) | |
8314 | || (!recp && flag_mlow_precision_sqrt)) | |
a6fc00da BH |
8315 | iterations--; |
8316 | ||
98daafa0 EM |
8317 | /* Iterate over the series to calculate the approximate reciprocal square |
8318 | root. */ | |
8319 | rtx x1 = gen_reg_rtx (mode); | |
8320 | while (iterations--) | |
a6fc00da | 8321 | { |
a6fc00da | 8322 | rtx x2 = gen_reg_rtx (mode); |
98daafa0 EM |
8323 | emit_set_insn (x2, gen_rtx_MULT (mode, xdst, xdst)); |
8324 | ||
8325 | emit_insn ((*get_rsqrts_type (mode)) (x1, src, x2)); | |
a6fc00da | 8326 | |
98daafa0 EM |
8327 | if (iterations > 0) |
8328 | emit_set_insn (xdst, gen_rtx_MULT (mode, xdst, x1)); | |
8329 | } | |
8330 | ||
8331 | if (!recp) | |
8332 | { | |
8333 | /* Qualify the approximate reciprocal square root when the argument is | |
8334 | 0.0 by squashing the intermediary result to 0.0. */ | |
8335 | rtx xtmp = gen_reg_rtx (mmsk); | |
8336 | emit_set_insn (xtmp, gen_rtx_AND (mmsk, gen_rtx_NOT (mmsk, xmsk), | |
8337 | gen_rtx_SUBREG (mmsk, xdst, 0))); | |
8338 | emit_move_insn (xdst, gen_rtx_SUBREG (mode, xtmp, 0)); | |
a6fc00da | 8339 | |
98daafa0 EM |
8340 | /* Calculate the approximate square root. */ |
8341 | emit_set_insn (xdst, gen_rtx_MULT (mode, xdst, src)); | |
a6fc00da BH |
8342 | } |
8343 | ||
98daafa0 EM |
8344 | /* Finalize the approximation. */ |
8345 | emit_set_insn (dst, gen_rtx_MULT (mode, xdst, x1)); | |
8346 | ||
8347 | return true; | |
a6fc00da BH |
8348 | } |
8349 | ||
79a2bc2d EM |
8350 | typedef rtx (*recpe_type) (rtx, rtx); |
8351 | ||
8352 | /* Select reciprocal initial estimate insn depending on machine mode. */ | |
8353 | ||
8354 | static recpe_type | |
8355 | get_recpe_type (machine_mode mode) | |
8356 | { | |
8357 | switch (mode) | |
8358 | { | |
4e10a5a7 RS |
8359 | case E_SFmode: return (gen_aarch64_frecpesf); |
8360 | case E_V2SFmode: return (gen_aarch64_frecpev2sf); | |
8361 | case E_V4SFmode: return (gen_aarch64_frecpev4sf); | |
8362 | case E_DFmode: return (gen_aarch64_frecpedf); | |
8363 | case E_V2DFmode: return (gen_aarch64_frecpev2df); | |
8364 | default: gcc_unreachable (); | |
79a2bc2d EM |
8365 | } |
8366 | } | |
8367 | ||
8368 | typedef rtx (*recps_type) (rtx, rtx, rtx); | |
8369 | ||
8370 | /* Select reciprocal series step insn depending on machine mode. */ | |
8371 | ||
8372 | static recps_type | |
8373 | get_recps_type (machine_mode mode) | |
8374 | { | |
8375 | switch (mode) | |
8376 | { | |
4e10a5a7 RS |
8377 | case E_SFmode: return (gen_aarch64_frecpssf); |
8378 | case E_V2SFmode: return (gen_aarch64_frecpsv2sf); | |
8379 | case E_V4SFmode: return (gen_aarch64_frecpsv4sf); | |
8380 | case E_DFmode: return (gen_aarch64_frecpsdf); | |
8381 | case E_V2DFmode: return (gen_aarch64_frecpsv2df); | |
8382 | default: gcc_unreachable (); | |
79a2bc2d EM |
8383 | } |
8384 | } | |
8385 | ||
8386 | /* Emit the instruction sequence to compute the approximation for the division | |
8387 | of NUM by DEN in QUO and return whether the sequence was emitted or not. */ | |
8388 | ||
8389 | bool | |
8390 | aarch64_emit_approx_div (rtx quo, rtx num, rtx den) | |
8391 | { | |
8392 | machine_mode mode = GET_MODE (quo); | |
33d72b63 JW |
8393 | |
8394 | if (GET_MODE_INNER (mode) == HFmode) | |
8395 | return false; | |
8396 | ||
79a2bc2d EM |
8397 | bool use_approx_division_p = (flag_mlow_precision_div |
8398 | || (aarch64_tune_params.approx_modes->division | |
8399 | & AARCH64_APPROX_MODE (mode))); | |
8400 | ||
8401 | if (!flag_finite_math_only | |
8402 | || flag_trapping_math | |
8403 | || !flag_unsafe_math_optimizations | |
8404 | || optimize_function_for_size_p (cfun) | |
8405 | || !use_approx_division_p) | |
8406 | return false; | |
8407 | ||
8408 | /* Estimate the approximate reciprocal. */ | |
8409 | rtx xrcp = gen_reg_rtx (mode); | |
8410 | emit_insn ((*get_recpe_type (mode)) (xrcp, den)); | |
8411 | ||
8412 | /* Iterate over the series twice for SF and thrice for DF. */ | |
8413 | int iterations = (GET_MODE_INNER (mode) == DFmode) ? 3 : 2; | |
8414 | ||
8415 | /* Optionally iterate over the series once less for faster performance, | |
8416 | while sacrificing the accuracy. */ | |
8417 | if (flag_mlow_precision_div) | |
8418 | iterations--; | |
8419 | ||
8420 | /* Iterate over the series to calculate the approximate reciprocal. */ | |
8421 | rtx xtmp = gen_reg_rtx (mode); | |
8422 | while (iterations--) | |
8423 | { | |
8424 | emit_insn ((*get_recps_type (mode)) (xtmp, xrcp, den)); | |
8425 | ||
8426 | if (iterations > 0) | |
8427 | emit_set_insn (xrcp, gen_rtx_MULT (mode, xrcp, xtmp)); | |
8428 | } | |
8429 | ||
8430 | if (num != CONST1_RTX (mode)) | |
8431 | { | |
8432 | /* As the approximate reciprocal of DEN is already calculated, only | |
8433 | calculate the approximate division when NUM is not 1.0. */ | |
8434 | rtx xnum = force_reg (mode, num); | |
8435 | emit_set_insn (xrcp, gen_rtx_MULT (mode, xrcp, xnum)); | |
8436 | } | |
8437 | ||
8438 | /* Finalize the approximation. */ | |
8439 | emit_set_insn (quo, gen_rtx_MULT (mode, xrcp, xtmp)); | |
8440 | return true; | |
8441 | } | |
8442 | ||
d126a4ae AP |
8443 | /* Return the number of instructions that can be issued per cycle. */ |
8444 | static int | |
8445 | aarch64_sched_issue_rate (void) | |
8446 | { | |
b175b679 | 8447 | return aarch64_tune_params.issue_rate; |
d126a4ae AP |
8448 | } |
8449 | ||
d03f7e44 MK |
8450 | static int |
8451 | aarch64_sched_first_cycle_multipass_dfa_lookahead (void) | |
8452 | { | |
8453 | int issue_rate = aarch64_sched_issue_rate (); | |
8454 | ||
8455 | return issue_rate > 1 && !sched_fusion ? issue_rate : 0; | |
8456 | } | |
8457 | ||
2d6bc7fa KT |
8458 | |
8459 | /* Implement TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD as | |
8460 | autopref_multipass_dfa_lookahead_guard from haifa-sched.c. It only | |
8461 | has an effect if PARAM_SCHED_AUTOPREF_QUEUE_DEPTH > 0. */ | |
8462 | ||
8463 | static int | |
8464 | aarch64_first_cycle_multipass_dfa_lookahead_guard (rtx_insn *insn, | |
8465 | int ready_index) | |
8466 | { | |
8467 | return autopref_multipass_dfa_lookahead_guard (insn, ready_index); | |
8468 | } | |
8469 | ||
8470 | ||
8990e73a TB |
8471 | /* Vectorizer cost model target hooks. */ |
8472 | ||
8473 | /* Implement targetm.vectorize.builtin_vectorization_cost. */ | |
8474 | static int | |
8475 | aarch64_builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost, | |
8476 | tree vectype, | |
8477 | int misalign ATTRIBUTE_UNUSED) | |
8478 | { | |
8479 | unsigned elements; | |
cd8ae5ed AP |
8480 | const cpu_vector_cost *costs = aarch64_tune_params.vec_costs; |
8481 | bool fp = false; | |
8482 | ||
8483 | if (vectype != NULL) | |
8484 | fp = FLOAT_TYPE_P (vectype); | |
8990e73a TB |
8485 | |
8486 | switch (type_of_cost) | |
8487 | { | |
8488 | case scalar_stmt: | |
cd8ae5ed | 8489 | return fp ? costs->scalar_fp_stmt_cost : costs->scalar_int_stmt_cost; |
8990e73a TB |
8490 | |
8491 | case scalar_load: | |
cd8ae5ed | 8492 | return costs->scalar_load_cost; |
8990e73a TB |
8493 | |
8494 | case scalar_store: | |
cd8ae5ed | 8495 | return costs->scalar_store_cost; |
8990e73a TB |
8496 | |
8497 | case vector_stmt: | |
cd8ae5ed | 8498 | return fp ? costs->vec_fp_stmt_cost : costs->vec_int_stmt_cost; |
8990e73a TB |
8499 | |
8500 | case vector_load: | |
cd8ae5ed | 8501 | return costs->vec_align_load_cost; |
8990e73a TB |
8502 | |
8503 | case vector_store: | |
cd8ae5ed | 8504 | return costs->vec_store_cost; |
8990e73a TB |
8505 | |
8506 | case vec_to_scalar: | |
cd8ae5ed | 8507 | return costs->vec_to_scalar_cost; |
8990e73a TB |
8508 | |
8509 | case scalar_to_vec: | |
cd8ae5ed | 8510 | return costs->scalar_to_vec_cost; |
8990e73a TB |
8511 | |
8512 | case unaligned_load: | |
cd8ae5ed | 8513 | return costs->vec_unalign_load_cost; |
8990e73a TB |
8514 | |
8515 | case unaligned_store: | |
cd8ae5ed | 8516 | return costs->vec_unalign_store_cost; |
8990e73a TB |
8517 | |
8518 | case cond_branch_taken: | |
cd8ae5ed | 8519 | return costs->cond_taken_branch_cost; |
8990e73a TB |
8520 | |
8521 | case cond_branch_not_taken: | |
cd8ae5ed | 8522 | return costs->cond_not_taken_branch_cost; |
8990e73a TB |
8523 | |
8524 | case vec_perm: | |
cd8ae5ed | 8525 | return costs->vec_permute_cost; |
c428f91c | 8526 | |
8990e73a | 8527 | case vec_promote_demote: |
cd8ae5ed | 8528 | return fp ? costs->vec_fp_stmt_cost : costs->vec_int_stmt_cost; |
8990e73a TB |
8529 | |
8530 | case vec_construct: | |
8531 | elements = TYPE_VECTOR_SUBPARTS (vectype); | |
8532 | return elements / 2 + 1; | |
8533 | ||
8534 | default: | |
8535 | gcc_unreachable (); | |
8536 | } | |
8537 | } | |
8538 | ||
8539 | /* Implement targetm.vectorize.add_stmt_cost. */ | |
8540 | static unsigned | |
8541 | aarch64_add_stmt_cost (void *data, int count, enum vect_cost_for_stmt kind, | |
8542 | struct _stmt_vec_info *stmt_info, int misalign, | |
8543 | enum vect_cost_model_location where) | |
8544 | { | |
8545 | unsigned *cost = (unsigned *) data; | |
8546 | unsigned retval = 0; | |
8547 | ||
8548 | if (flag_vect_cost_model) | |
8549 | { | |
8550 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; | |
8551 | int stmt_cost = | |
8552 | aarch64_builtin_vectorization_cost (kind, vectype, misalign); | |
8553 | ||
8554 | /* Statements in an inner loop relative to the loop being | |
8555 | vectorized are weighted more heavily. The value here is | |
058e4c71 | 8556 | arbitrary and could potentially be improved with analysis. */ |
8990e73a | 8557 | if (where == vect_body && stmt_info && stmt_in_inner_loop_p (stmt_info)) |
058e4c71 | 8558 | count *= 50; /* FIXME */ |
8990e73a TB |
8559 | |
8560 | retval = (unsigned) (count * stmt_cost); | |
8561 | cost[where] += retval; | |
8562 | } | |
8563 | ||
8564 | return retval; | |
8565 | } | |
8566 | ||
0cfff2a1 | 8567 | static void initialize_aarch64_code_model (struct gcc_options *); |
43e9d192 | 8568 | |
0cfff2a1 KT |
8569 | /* Parse the TO_PARSE string and put the architecture struct that it |
8570 | selects into RES and the architectural features into ISA_FLAGS. | |
8571 | Return an aarch64_parse_opt_result describing the parse result. | |
8572 | If there is an error parsing, RES and ISA_FLAGS are left unchanged. */ | |
43e9d192 | 8573 | |
0cfff2a1 KT |
8574 | static enum aarch64_parse_opt_result |
8575 | aarch64_parse_arch (const char *to_parse, const struct processor **res, | |
8576 | unsigned long *isa_flags) | |
43e9d192 IB |
8577 | { |
8578 | char *ext; | |
8579 | const struct processor *arch; | |
0cfff2a1 | 8580 | char *str = (char *) alloca (strlen (to_parse) + 1); |
43e9d192 IB |
8581 | size_t len; |
8582 | ||
0cfff2a1 | 8583 | strcpy (str, to_parse); |
43e9d192 IB |
8584 | |
8585 | ext = strchr (str, '+'); | |
8586 | ||
8587 | if (ext != NULL) | |
8588 | len = ext - str; | |
8589 | else | |
8590 | len = strlen (str); | |
8591 | ||
8592 | if (len == 0) | |
0cfff2a1 KT |
8593 | return AARCH64_PARSE_MISSING_ARG; |
8594 | ||
43e9d192 | 8595 | |
0cfff2a1 | 8596 | /* Loop through the list of supported ARCHes to find a match. */ |
43e9d192 IB |
8597 | for (arch = all_architectures; arch->name != NULL; arch++) |
8598 | { | |
8599 | if (strlen (arch->name) == len && strncmp (arch->name, str, len) == 0) | |
8600 | { | |
0cfff2a1 | 8601 | unsigned long isa_temp = arch->flags; |
43e9d192 IB |
8602 | |
8603 | if (ext != NULL) | |
8604 | { | |
0cfff2a1 KT |
8605 | /* TO_PARSE string contains at least one extension. */ |
8606 | enum aarch64_parse_opt_result ext_res | |
8607 | = aarch64_parse_extension (ext, &isa_temp); | |
43e9d192 | 8608 | |
0cfff2a1 KT |
8609 | if (ext_res != AARCH64_PARSE_OK) |
8610 | return ext_res; | |
ffee7aa9 | 8611 | } |
0cfff2a1 KT |
8612 | /* Extension parsing was successful. Confirm the result |
8613 | arch and ISA flags. */ | |
8614 | *res = arch; | |
8615 | *isa_flags = isa_temp; | |
8616 | return AARCH64_PARSE_OK; | |
43e9d192 IB |
8617 | } |
8618 | } | |
8619 | ||
8620 | /* ARCH name not found in list. */ | |
0cfff2a1 | 8621 | return AARCH64_PARSE_INVALID_ARG; |
43e9d192 IB |
8622 | } |
8623 | ||
0cfff2a1 KT |
8624 | /* Parse the TO_PARSE string and put the result tuning in RES and the |
8625 | architecture flags in ISA_FLAGS. Return an aarch64_parse_opt_result | |
8626 | describing the parse result. If there is an error parsing, RES and | |
8627 | ISA_FLAGS are left unchanged. */ | |
43e9d192 | 8628 | |
0cfff2a1 KT |
8629 | static enum aarch64_parse_opt_result |
8630 | aarch64_parse_cpu (const char *to_parse, const struct processor **res, | |
8631 | unsigned long *isa_flags) | |
43e9d192 IB |
8632 | { |
8633 | char *ext; | |
8634 | const struct processor *cpu; | |
0cfff2a1 | 8635 | char *str = (char *) alloca (strlen (to_parse) + 1); |
43e9d192 IB |
8636 | size_t len; |
8637 | ||
0cfff2a1 | 8638 | strcpy (str, to_parse); |
43e9d192 IB |
8639 | |
8640 | ext = strchr (str, '+'); | |
8641 | ||
8642 | if (ext != NULL) | |
8643 | len = ext - str; | |
8644 | else | |
8645 | len = strlen (str); | |
8646 | ||
8647 | if (len == 0) | |
0cfff2a1 KT |
8648 | return AARCH64_PARSE_MISSING_ARG; |
8649 | ||
43e9d192 IB |
8650 | |
8651 | /* Loop through the list of supported CPUs to find a match. */ | |
8652 | for (cpu = all_cores; cpu->name != NULL; cpu++) | |
8653 | { | |
8654 | if (strlen (cpu->name) == len && strncmp (cpu->name, str, len) == 0) | |
8655 | { | |
0cfff2a1 KT |
8656 | unsigned long isa_temp = cpu->flags; |
8657 | ||
43e9d192 IB |
8658 | |
8659 | if (ext != NULL) | |
8660 | { | |
0cfff2a1 KT |
8661 | /* TO_PARSE string contains at least one extension. */ |
8662 | enum aarch64_parse_opt_result ext_res | |
8663 | = aarch64_parse_extension (ext, &isa_temp); | |
43e9d192 | 8664 | |
0cfff2a1 KT |
8665 | if (ext_res != AARCH64_PARSE_OK) |
8666 | return ext_res; | |
8667 | } | |
8668 | /* Extension parsing was successfull. Confirm the result | |
8669 | cpu and ISA flags. */ | |
8670 | *res = cpu; | |
8671 | *isa_flags = isa_temp; | |
8672 | return AARCH64_PARSE_OK; | |
43e9d192 IB |
8673 | } |
8674 | } | |
8675 | ||
8676 | /* CPU name not found in list. */ | |
0cfff2a1 | 8677 | return AARCH64_PARSE_INVALID_ARG; |
43e9d192 IB |
8678 | } |
8679 | ||
0cfff2a1 KT |
8680 | /* Parse the TO_PARSE string and put the cpu it selects into RES. |
8681 | Return an aarch64_parse_opt_result describing the parse result. | |
8682 | If the parsing fails the RES does not change. */ | |
43e9d192 | 8683 | |
0cfff2a1 KT |
8684 | static enum aarch64_parse_opt_result |
8685 | aarch64_parse_tune (const char *to_parse, const struct processor **res) | |
43e9d192 IB |
8686 | { |
8687 | const struct processor *cpu; | |
0cfff2a1 KT |
8688 | char *str = (char *) alloca (strlen (to_parse) + 1); |
8689 | ||
8690 | strcpy (str, to_parse); | |
43e9d192 IB |
8691 | |
8692 | /* Loop through the list of supported CPUs to find a match. */ | |
8693 | for (cpu = all_cores; cpu->name != NULL; cpu++) | |
8694 | { | |
8695 | if (strcmp (cpu->name, str) == 0) | |
8696 | { | |
0cfff2a1 KT |
8697 | *res = cpu; |
8698 | return AARCH64_PARSE_OK; | |
43e9d192 IB |
8699 | } |
8700 | } | |
8701 | ||
8702 | /* CPU name not found in list. */ | |
0cfff2a1 | 8703 | return AARCH64_PARSE_INVALID_ARG; |
43e9d192 IB |
8704 | } |
8705 | ||
8dec06f2 JG |
8706 | /* Parse TOKEN, which has length LENGTH to see if it is an option |
8707 | described in FLAG. If it is, return the index bit for that fusion type. | |
8708 | If not, error (printing OPTION_NAME) and return zero. */ | |
8709 | ||
8710 | static unsigned int | |
8711 | aarch64_parse_one_option_token (const char *token, | |
8712 | size_t length, | |
8713 | const struct aarch64_flag_desc *flag, | |
8714 | const char *option_name) | |
8715 | { | |
8716 | for (; flag->name != NULL; flag++) | |
8717 | { | |
8718 | if (length == strlen (flag->name) | |
8719 | && !strncmp (flag->name, token, length)) | |
8720 | return flag->flag; | |
8721 | } | |
8722 | ||
8723 | error ("unknown flag passed in -moverride=%s (%s)", option_name, token); | |
8724 | return 0; | |
8725 | } | |
8726 | ||
8727 | /* Parse OPTION which is a comma-separated list of flags to enable. | |
8728 | FLAGS gives the list of flags we understand, INITIAL_STATE gives any | |
8729 | default state we inherit from the CPU tuning structures. OPTION_NAME | |
8730 | gives the top-level option we are parsing in the -moverride string, | |
8731 | for use in error messages. */ | |
8732 | ||
8733 | static unsigned int | |
8734 | aarch64_parse_boolean_options (const char *option, | |
8735 | const struct aarch64_flag_desc *flags, | |
8736 | unsigned int initial_state, | |
8737 | const char *option_name) | |
8738 | { | |
8739 | const char separator = '.'; | |
8740 | const char* specs = option; | |
8741 | const char* ntoken = option; | |
8742 | unsigned int found_flags = initial_state; | |
8743 | ||
8744 | while ((ntoken = strchr (specs, separator))) | |
8745 | { | |
8746 | size_t token_length = ntoken - specs; | |
8747 | unsigned token_ops = aarch64_parse_one_option_token (specs, | |
8748 | token_length, | |
8749 | flags, | |
8750 | option_name); | |
8751 | /* If we find "none" (or, for simplicity's sake, an error) anywhere | |
8752 | in the token stream, reset the supported operations. So: | |
8753 | ||
8754 | adrp+add.cmp+branch.none.adrp+add | |
8755 | ||
8756 | would have the result of turning on only adrp+add fusion. */ | |
8757 | if (!token_ops) | |
8758 | found_flags = 0; | |
8759 | ||
8760 | found_flags |= token_ops; | |
8761 | specs = ++ntoken; | |
8762 | } | |
8763 | ||
8764 | /* We ended with a comma, print something. */ | |
8765 | if (!(*specs)) | |
8766 | { | |
8767 | error ("%s string ill-formed\n", option_name); | |
8768 | return 0; | |
8769 | } | |
8770 | ||
8771 | /* We still have one more token to parse. */ | |
8772 | size_t token_length = strlen (specs); | |
8773 | unsigned token_ops = aarch64_parse_one_option_token (specs, | |
8774 | token_length, | |
8775 | flags, | |
8776 | option_name); | |
8777 | if (!token_ops) | |
8778 | found_flags = 0; | |
8779 | ||
8780 | found_flags |= token_ops; | |
8781 | return found_flags; | |
8782 | } | |
8783 | ||
8784 | /* Support for overriding instruction fusion. */ | |
8785 | ||
8786 | static void | |
8787 | aarch64_parse_fuse_string (const char *fuse_string, | |
8788 | struct tune_params *tune) | |
8789 | { | |
8790 | tune->fusible_ops = aarch64_parse_boolean_options (fuse_string, | |
8791 | aarch64_fusible_pairs, | |
8792 | tune->fusible_ops, | |
8793 | "fuse="); | |
8794 | } | |
8795 | ||
8796 | /* Support for overriding other tuning flags. */ | |
8797 | ||
8798 | static void | |
8799 | aarch64_parse_tune_string (const char *tune_string, | |
8800 | struct tune_params *tune) | |
8801 | { | |
8802 | tune->extra_tuning_flags | |
8803 | = aarch64_parse_boolean_options (tune_string, | |
8804 | aarch64_tuning_flags, | |
8805 | tune->extra_tuning_flags, | |
8806 | "tune="); | |
8807 | } | |
8808 | ||
8809 | /* Parse TOKEN, which has length LENGTH to see if it is a tuning option | |
8810 | we understand. If it is, extract the option string and handoff to | |
8811 | the appropriate function. */ | |
8812 | ||
8813 | void | |
8814 | aarch64_parse_one_override_token (const char* token, | |
8815 | size_t length, | |
8816 | struct tune_params *tune) | |
8817 | { | |
8818 | const struct aarch64_tuning_override_function *fn | |
8819 | = aarch64_tuning_override_functions; | |
8820 | ||
8821 | const char *option_part = strchr (token, '='); | |
8822 | if (!option_part) | |
8823 | { | |
8824 | error ("tuning string missing in option (%s)", token); | |
8825 | return; | |
8826 | } | |
8827 | ||
8828 | /* Get the length of the option name. */ | |
8829 | length = option_part - token; | |
8830 | /* Skip the '=' to get to the option string. */ | |
8831 | option_part++; | |
8832 | ||
8833 | for (; fn->name != NULL; fn++) | |
8834 | { | |
8835 | if (!strncmp (fn->name, token, length)) | |
8836 | { | |
8837 | fn->parse_override (option_part, tune); | |
8838 | return; | |
8839 | } | |
8840 | } | |
8841 | ||
8842 | error ("unknown tuning option (%s)",token); | |
8843 | return; | |
8844 | } | |
8845 | ||
5eee3c34 JW |
8846 | /* A checking mechanism for the implementation of the tls size. */ |
8847 | ||
8848 | static void | |
8849 | initialize_aarch64_tls_size (struct gcc_options *opts) | |
8850 | { | |
8851 | if (aarch64_tls_size == 0) | |
8852 | aarch64_tls_size = 24; | |
8853 | ||
8854 | switch (opts->x_aarch64_cmodel_var) | |
8855 | { | |
8856 | case AARCH64_CMODEL_TINY: | |
8857 | /* Both the default and maximum TLS size allowed under tiny is 1M which | |
8858 | needs two instructions to address, so we clamp the size to 24. */ | |
8859 | if (aarch64_tls_size > 24) | |
8860 | aarch64_tls_size = 24; | |
8861 | break; | |
8862 | case AARCH64_CMODEL_SMALL: | |
8863 | /* The maximum TLS size allowed under small is 4G. */ | |
8864 | if (aarch64_tls_size > 32) | |
8865 | aarch64_tls_size = 32; | |
8866 | break; | |
8867 | case AARCH64_CMODEL_LARGE: | |
8868 | /* The maximum TLS size allowed under large is 16E. | |
8869 | FIXME: 16E should be 64bit, we only support 48bit offset now. */ | |
8870 | if (aarch64_tls_size > 48) | |
8871 | aarch64_tls_size = 48; | |
8872 | break; | |
8873 | default: | |
8874 | gcc_unreachable (); | |
8875 | } | |
8876 | ||
8877 | return; | |
8878 | } | |
8879 | ||
8dec06f2 JG |
8880 | /* Parse STRING looking for options in the format: |
8881 | string :: option:string | |
8882 | option :: name=substring | |
8883 | name :: {a-z} | |
8884 | substring :: defined by option. */ | |
8885 | ||
8886 | static void | |
8887 | aarch64_parse_override_string (const char* input_string, | |
8888 | struct tune_params* tune) | |
8889 | { | |
8890 | const char separator = ':'; | |
8891 | size_t string_length = strlen (input_string) + 1; | |
8892 | char *string_root = (char *) xmalloc (sizeof (*string_root) * string_length); | |
8893 | char *string = string_root; | |
8894 | strncpy (string, input_string, string_length); | |
8895 | string[string_length - 1] = '\0'; | |
8896 | ||
8897 | char* ntoken = string; | |
8898 | ||
8899 | while ((ntoken = strchr (string, separator))) | |
8900 | { | |
8901 | size_t token_length = ntoken - string; | |
8902 | /* Make this substring look like a string. */ | |
8903 | *ntoken = '\0'; | |
8904 | aarch64_parse_one_override_token (string, token_length, tune); | |
8905 | string = ++ntoken; | |
8906 | } | |
8907 | ||
8908 | /* One last option to parse. */ | |
8909 | aarch64_parse_one_override_token (string, strlen (string), tune); | |
8910 | free (string_root); | |
8911 | } | |
43e9d192 | 8912 | |
43e9d192 IB |
8913 | |
8914 | static void | |
0cfff2a1 | 8915 | aarch64_override_options_after_change_1 (struct gcc_options *opts) |
43e9d192 | 8916 | { |
a3dc8760 NC |
8917 | /* The logic here is that if we are disabling all frame pointer generation |
8918 | then we do not need to disable leaf frame pointer generation as a | |
8919 | separate operation. But if we are *only* disabling leaf frame pointer | |
8920 | generation then we set flag_omit_frame_pointer to true, but in | |
8921 | aarch64_frame_pointer_required we return false only for leaf functions. | |
8922 | ||
8923 | PR 70044: We have to be careful about being called multiple times for the | |
8924 | same function. Once we have decided to set flag_omit_frame_pointer just | |
8925 | so that we can omit leaf frame pointers, we must then not interpret a | |
8926 | second call as meaning that all frame pointer generation should be | |
8927 | omitted. We do this by setting flag_omit_frame_pointer to a special, | |
8928 | non-zero value. */ | |
8929 | if (opts->x_flag_omit_frame_pointer == 2) | |
8930 | opts->x_flag_omit_frame_pointer = 0; | |
8931 | ||
0cfff2a1 KT |
8932 | if (opts->x_flag_omit_frame_pointer) |
8933 | opts->x_flag_omit_leaf_frame_pointer = false; | |
8934 | else if (opts->x_flag_omit_leaf_frame_pointer) | |
a3dc8760 | 8935 | opts->x_flag_omit_frame_pointer = 2; |
43e9d192 | 8936 | |
1be34295 | 8937 | /* If not optimizing for size, set the default |
0cfff2a1 KT |
8938 | alignment to what the target wants. */ |
8939 | if (!opts->x_optimize_size) | |
43e9d192 | 8940 | { |
0cfff2a1 KT |
8941 | if (opts->x_align_loops <= 0) |
8942 | opts->x_align_loops = aarch64_tune_params.loop_align; | |
8943 | if (opts->x_align_jumps <= 0) | |
8944 | opts->x_align_jumps = aarch64_tune_params.jump_align; | |
8945 | if (opts->x_align_functions <= 0) | |
8946 | opts->x_align_functions = aarch64_tune_params.function_align; | |
43e9d192 | 8947 | } |
b4f50fd4 | 8948 | |
9ee6540a WD |
8949 | /* We default to no pc-relative literal loads. */ |
8950 | ||
8951 | aarch64_pcrelative_literal_loads = false; | |
8952 | ||
8953 | /* If -mpc-relative-literal-loads is set on the command line, this | |
b4f50fd4 | 8954 | implies that the user asked for PC relative literal loads. */ |
9ee6540a WD |
8955 | if (opts->x_pcrelative_literal_loads == 1) |
8956 | aarch64_pcrelative_literal_loads = true; | |
b4f50fd4 | 8957 | |
48bb1a55 CL |
8958 | /* This is PR70113. When building the Linux kernel with |
8959 | CONFIG_ARM64_ERRATUM_843419, support for relocations | |
8960 | R_AARCH64_ADR_PREL_PG_HI21 and R_AARCH64_ADR_PREL_PG_HI21_NC is | |
8961 | removed from the kernel to avoid loading objects with possibly | |
9ee6540a | 8962 | offending sequences. Without -mpc-relative-literal-loads we would |
48bb1a55 CL |
8963 | generate such relocations, preventing the kernel build from |
8964 | succeeding. */ | |
9ee6540a WD |
8965 | if (opts->x_pcrelative_literal_loads == 2 |
8966 | && TARGET_FIX_ERR_A53_843419) | |
8967 | aarch64_pcrelative_literal_loads = true; | |
8968 | ||
8969 | /* In the tiny memory model it makes no sense to disallow PC relative | |
8970 | literal pool loads. */ | |
8971 | if (aarch64_cmodel == AARCH64_CMODEL_TINY | |
8972 | || aarch64_cmodel == AARCH64_CMODEL_TINY_PIC) | |
8973 | aarch64_pcrelative_literal_loads = true; | |
98daafa0 EM |
8974 | |
8975 | /* When enabling the lower precision Newton series for the square root, also | |
8976 | enable it for the reciprocal square root, since the latter is an | |
8977 | intermediary step for the former. */ | |
8978 | if (flag_mlow_precision_sqrt) | |
8979 | flag_mrecip_low_precision_sqrt = true; | |
0cfff2a1 | 8980 | } |
43e9d192 | 8981 | |
0cfff2a1 KT |
8982 | /* 'Unpack' up the internal tuning structs and update the options |
8983 | in OPTS. The caller must have set up selected_tune and selected_arch | |
8984 | as all the other target-specific codegen decisions are | |
8985 | derived from them. */ | |
8986 | ||
e4ea20c8 | 8987 | void |
0cfff2a1 KT |
8988 | aarch64_override_options_internal (struct gcc_options *opts) |
8989 | { | |
8990 | aarch64_tune_flags = selected_tune->flags; | |
8991 | aarch64_tune = selected_tune->sched_core; | |
8992 | /* Make a copy of the tuning parameters attached to the core, which | |
8993 | we may later overwrite. */ | |
8994 | aarch64_tune_params = *(selected_tune->tune); | |
8995 | aarch64_architecture_version = selected_arch->architecture_version; | |
8996 | ||
8997 | if (opts->x_aarch64_override_tune_string) | |
8998 | aarch64_parse_override_string (opts->x_aarch64_override_tune_string, | |
8999 | &aarch64_tune_params); | |
9000 | ||
9001 | /* This target defaults to strict volatile bitfields. */ | |
9002 | if (opts->x_flag_strict_volatile_bitfields < 0 && abi_version_at_least (2)) | |
9003 | opts->x_flag_strict_volatile_bitfields = 1; | |
9004 | ||
0cfff2a1 | 9005 | initialize_aarch64_code_model (opts); |
5eee3c34 | 9006 | initialize_aarch64_tls_size (opts); |
63892fa2 | 9007 | |
2d6bc7fa KT |
9008 | int queue_depth = 0; |
9009 | switch (aarch64_tune_params.autoprefetcher_model) | |
9010 | { | |
9011 | case tune_params::AUTOPREFETCHER_OFF: | |
9012 | queue_depth = -1; | |
9013 | break; | |
9014 | case tune_params::AUTOPREFETCHER_WEAK: | |
9015 | queue_depth = 0; | |
9016 | break; | |
9017 | case tune_params::AUTOPREFETCHER_STRONG: | |
9018 | queue_depth = max_insn_queue_index + 1; | |
9019 | break; | |
9020 | default: | |
9021 | gcc_unreachable (); | |
9022 | } | |
9023 | ||
9024 | /* We don't mind passing in global_options_set here as we don't use | |
9025 | the *options_set structs anyway. */ | |
9026 | maybe_set_param_value (PARAM_SCHED_AUTOPREF_QUEUE_DEPTH, | |
9027 | queue_depth, | |
9028 | opts->x_param_values, | |
9029 | global_options_set.x_param_values); | |
9030 | ||
9d2c6e2e MK |
9031 | /* Set up parameters to be used in prefetching algorithm. Do not |
9032 | override the defaults unless we are tuning for a core we have | |
9033 | researched values for. */ | |
9034 | if (aarch64_tune_params.prefetch->num_slots > 0) | |
9035 | maybe_set_param_value (PARAM_SIMULTANEOUS_PREFETCHES, | |
9036 | aarch64_tune_params.prefetch->num_slots, | |
9037 | opts->x_param_values, | |
9038 | global_options_set.x_param_values); | |
9039 | if (aarch64_tune_params.prefetch->l1_cache_size >= 0) | |
9040 | maybe_set_param_value (PARAM_L1_CACHE_SIZE, | |
9041 | aarch64_tune_params.prefetch->l1_cache_size, | |
9042 | opts->x_param_values, | |
9043 | global_options_set.x_param_values); | |
9044 | if (aarch64_tune_params.prefetch->l1_cache_line_size >= 0) | |
50487d79 | 9045 | maybe_set_param_value (PARAM_L1_CACHE_LINE_SIZE, |
9d2c6e2e MK |
9046 | aarch64_tune_params.prefetch->l1_cache_line_size, |
9047 | opts->x_param_values, | |
9048 | global_options_set.x_param_values); | |
9049 | if (aarch64_tune_params.prefetch->l2_cache_size >= 0) | |
9050 | maybe_set_param_value (PARAM_L2_CACHE_SIZE, | |
9051 | aarch64_tune_params.prefetch->l2_cache_size, | |
50487d79 EM |
9052 | opts->x_param_values, |
9053 | global_options_set.x_param_values); | |
9054 | ||
16b2cafd MK |
9055 | /* Enable sw prefetching at specified optimization level for |
9056 | CPUS that have prefetch. Lower optimization level threshold by 1 | |
9057 | when profiling is enabled. */ | |
9058 | if (opts->x_flag_prefetch_loop_arrays < 0 | |
9059 | && !opts->x_optimize_size | |
9060 | && aarch64_tune_params.prefetch->default_opt_level >= 0 | |
9061 | && opts->x_optimize >= aarch64_tune_params.prefetch->default_opt_level) | |
9062 | opts->x_flag_prefetch_loop_arrays = 1; | |
9063 | ||
0cfff2a1 KT |
9064 | aarch64_override_options_after_change_1 (opts); |
9065 | } | |
43e9d192 | 9066 | |
01f44038 KT |
9067 | /* Print a hint with a suggestion for a core or architecture name that |
9068 | most closely resembles what the user passed in STR. ARCH is true if | |
9069 | the user is asking for an architecture name. ARCH is false if the user | |
9070 | is asking for a core name. */ | |
9071 | ||
9072 | static void | |
9073 | aarch64_print_hint_for_core_or_arch (const char *str, bool arch) | |
9074 | { | |
9075 | auto_vec<const char *> candidates; | |
9076 | const struct processor *entry = arch ? all_architectures : all_cores; | |
9077 | for (; entry->name != NULL; entry++) | |
9078 | candidates.safe_push (entry->name); | |
9079 | char *s; | |
9080 | const char *hint = candidates_list_and_hint (str, s, candidates); | |
9081 | if (hint) | |
9082 | inform (input_location, "valid arguments are: %s;" | |
9083 | " did you mean %qs?", s, hint); | |
9084 | XDELETEVEC (s); | |
9085 | } | |
9086 | ||
9087 | /* Print a hint with a suggestion for a core name that most closely resembles | |
9088 | what the user passed in STR. */ | |
9089 | ||
9090 | inline static void | |
9091 | aarch64_print_hint_for_core (const char *str) | |
9092 | { | |
9093 | aarch64_print_hint_for_core_or_arch (str, false); | |
9094 | } | |
9095 | ||
9096 | /* Print a hint with a suggestion for an architecture name that most closely | |
9097 | resembles what the user passed in STR. */ | |
9098 | ||
9099 | inline static void | |
9100 | aarch64_print_hint_for_arch (const char *str) | |
9101 | { | |
9102 | aarch64_print_hint_for_core_or_arch (str, true); | |
9103 | } | |
9104 | ||
0cfff2a1 KT |
9105 | /* Validate a command-line -mcpu option. Parse the cpu and extensions (if any) |
9106 | specified in STR and throw errors if appropriate. Put the results if | |
361fb3ee KT |
9107 | they are valid in RES and ISA_FLAGS. Return whether the option is |
9108 | valid. */ | |
43e9d192 | 9109 | |
361fb3ee | 9110 | static bool |
0cfff2a1 KT |
9111 | aarch64_validate_mcpu (const char *str, const struct processor **res, |
9112 | unsigned long *isa_flags) | |
9113 | { | |
9114 | enum aarch64_parse_opt_result parse_res | |
9115 | = aarch64_parse_cpu (str, res, isa_flags); | |
9116 | ||
9117 | if (parse_res == AARCH64_PARSE_OK) | |
361fb3ee | 9118 | return true; |
0cfff2a1 KT |
9119 | |
9120 | switch (parse_res) | |
9121 | { | |
9122 | case AARCH64_PARSE_MISSING_ARG: | |
fb241da2 | 9123 | error ("missing cpu name in %<-mcpu=%s%>", str); |
0cfff2a1 KT |
9124 | break; |
9125 | case AARCH64_PARSE_INVALID_ARG: | |
9126 | error ("unknown value %qs for -mcpu", str); | |
01f44038 | 9127 | aarch64_print_hint_for_core (str); |
0cfff2a1 KT |
9128 | break; |
9129 | case AARCH64_PARSE_INVALID_FEATURE: | |
fb241da2 | 9130 | error ("invalid feature modifier in %<-mcpu=%s%>", str); |
0cfff2a1 KT |
9131 | break; |
9132 | default: | |
9133 | gcc_unreachable (); | |
9134 | } | |
361fb3ee KT |
9135 | |
9136 | return false; | |
0cfff2a1 KT |
9137 | } |
9138 | ||
9139 | /* Validate a command-line -march option. Parse the arch and extensions | |
9140 | (if any) specified in STR and throw errors if appropriate. Put the | |
361fb3ee KT |
9141 | results, if they are valid, in RES and ISA_FLAGS. Return whether the |
9142 | option is valid. */ | |
0cfff2a1 | 9143 | |
361fb3ee | 9144 | static bool |
0cfff2a1 | 9145 | aarch64_validate_march (const char *str, const struct processor **res, |
01f44038 | 9146 | unsigned long *isa_flags) |
0cfff2a1 KT |
9147 | { |
9148 | enum aarch64_parse_opt_result parse_res | |
9149 | = aarch64_parse_arch (str, res, isa_flags); | |
9150 | ||
9151 | if (parse_res == AARCH64_PARSE_OK) | |
361fb3ee | 9152 | return true; |
0cfff2a1 KT |
9153 | |
9154 | switch (parse_res) | |
9155 | { | |
9156 | case AARCH64_PARSE_MISSING_ARG: | |
fb241da2 | 9157 | error ("missing arch name in %<-march=%s%>", str); |
0cfff2a1 KT |
9158 | break; |
9159 | case AARCH64_PARSE_INVALID_ARG: | |
9160 | error ("unknown value %qs for -march", str); | |
01f44038 | 9161 | aarch64_print_hint_for_arch (str); |
0cfff2a1 KT |
9162 | break; |
9163 | case AARCH64_PARSE_INVALID_FEATURE: | |
fb241da2 | 9164 | error ("invalid feature modifier in %<-march=%s%>", str); |
0cfff2a1 KT |
9165 | break; |
9166 | default: | |
9167 | gcc_unreachable (); | |
9168 | } | |
361fb3ee KT |
9169 | |
9170 | return false; | |
0cfff2a1 KT |
9171 | } |
9172 | ||
9173 | /* Validate a command-line -mtune option. Parse the cpu | |
9174 | specified in STR and throw errors if appropriate. Put the | |
361fb3ee KT |
9175 | result, if it is valid, in RES. Return whether the option is |
9176 | valid. */ | |
0cfff2a1 | 9177 | |
361fb3ee | 9178 | static bool |
0cfff2a1 KT |
9179 | aarch64_validate_mtune (const char *str, const struct processor **res) |
9180 | { | |
9181 | enum aarch64_parse_opt_result parse_res | |
9182 | = aarch64_parse_tune (str, res); | |
9183 | ||
9184 | if (parse_res == AARCH64_PARSE_OK) | |
361fb3ee | 9185 | return true; |
0cfff2a1 KT |
9186 | |
9187 | switch (parse_res) | |
9188 | { | |
9189 | case AARCH64_PARSE_MISSING_ARG: | |
fb241da2 | 9190 | error ("missing cpu name in %<-mtune=%s%>", str); |
0cfff2a1 KT |
9191 | break; |
9192 | case AARCH64_PARSE_INVALID_ARG: | |
9193 | error ("unknown value %qs for -mtune", str); | |
01f44038 | 9194 | aarch64_print_hint_for_core (str); |
0cfff2a1 KT |
9195 | break; |
9196 | default: | |
9197 | gcc_unreachable (); | |
9198 | } | |
361fb3ee KT |
9199 | return false; |
9200 | } | |
9201 | ||
9202 | /* Return the CPU corresponding to the enum CPU. | |
9203 | If it doesn't specify a cpu, return the default. */ | |
9204 | ||
9205 | static const struct processor * | |
9206 | aarch64_get_tune_cpu (enum aarch64_processor cpu) | |
9207 | { | |
9208 | if (cpu != aarch64_none) | |
9209 | return &all_cores[cpu]; | |
9210 | ||
9211 | /* The & 0x3f is to extract the bottom 6 bits that encode the | |
9212 | default cpu as selected by the --with-cpu GCC configure option | |
9213 | in config.gcc. | |
9214 | ???: The whole TARGET_CPU_DEFAULT and AARCH64_CPU_DEFAULT_FLAGS | |
9215 | flags mechanism should be reworked to make it more sane. */ | |
9216 | return &all_cores[TARGET_CPU_DEFAULT & 0x3f]; | |
9217 | } | |
9218 | ||
9219 | /* Return the architecture corresponding to the enum ARCH. | |
9220 | If it doesn't specify a valid architecture, return the default. */ | |
9221 | ||
9222 | static const struct processor * | |
9223 | aarch64_get_arch (enum aarch64_arch arch) | |
9224 | { | |
9225 | if (arch != aarch64_no_arch) | |
9226 | return &all_architectures[arch]; | |
9227 | ||
9228 | const struct processor *cpu = &all_cores[TARGET_CPU_DEFAULT & 0x3f]; | |
9229 | ||
9230 | return &all_architectures[cpu->arch]; | |
0cfff2a1 KT |
9231 | } |
9232 | ||
9233 | /* Implement TARGET_OPTION_OVERRIDE. This is called once in the beginning | |
9234 | and is used to parse the -m{cpu,tune,arch} strings and setup the initial | |
9235 | tuning structs. In particular it must set selected_tune and | |
9236 | aarch64_isa_flags that define the available ISA features and tuning | |
9237 | decisions. It must also set selected_arch as this will be used to | |
9238 | output the .arch asm tags for each function. */ | |
9239 | ||
9240 | static void | |
9241 | aarch64_override_options (void) | |
9242 | { | |
9243 | unsigned long cpu_isa = 0; | |
9244 | unsigned long arch_isa = 0; | |
9245 | aarch64_isa_flags = 0; | |
9246 | ||
361fb3ee KT |
9247 | bool valid_cpu = true; |
9248 | bool valid_tune = true; | |
9249 | bool valid_arch = true; | |
9250 | ||
0cfff2a1 KT |
9251 | selected_cpu = NULL; |
9252 | selected_arch = NULL; | |
9253 | selected_tune = NULL; | |
9254 | ||
9255 | /* -mcpu=CPU is shorthand for -march=ARCH_FOR_CPU, -mtune=CPU. | |
9256 | If either of -march or -mtune is given, they override their | |
9257 | respective component of -mcpu. */ | |
9258 | if (aarch64_cpu_string) | |
361fb3ee KT |
9259 | valid_cpu = aarch64_validate_mcpu (aarch64_cpu_string, &selected_cpu, |
9260 | &cpu_isa); | |
0cfff2a1 KT |
9261 | |
9262 | if (aarch64_arch_string) | |
361fb3ee KT |
9263 | valid_arch = aarch64_validate_march (aarch64_arch_string, &selected_arch, |
9264 | &arch_isa); | |
0cfff2a1 KT |
9265 | |
9266 | if (aarch64_tune_string) | |
361fb3ee | 9267 | valid_tune = aarch64_validate_mtune (aarch64_tune_string, &selected_tune); |
43e9d192 IB |
9268 | |
9269 | /* If the user did not specify a processor, choose the default | |
9270 | one for them. This will be the CPU set during configuration using | |
a3cd0246 | 9271 | --with-cpu, otherwise it is "generic". */ |
43e9d192 IB |
9272 | if (!selected_cpu) |
9273 | { | |
0cfff2a1 KT |
9274 | if (selected_arch) |
9275 | { | |
9276 | selected_cpu = &all_cores[selected_arch->ident]; | |
9277 | aarch64_isa_flags = arch_isa; | |
361fb3ee | 9278 | explicit_arch = selected_arch->arch; |
0cfff2a1 KT |
9279 | } |
9280 | else | |
9281 | { | |
361fb3ee KT |
9282 | /* Get default configure-time CPU. */ |
9283 | selected_cpu = aarch64_get_tune_cpu (aarch64_none); | |
0cfff2a1 KT |
9284 | aarch64_isa_flags = TARGET_CPU_DEFAULT >> 6; |
9285 | } | |
361fb3ee KT |
9286 | |
9287 | if (selected_tune) | |
9288 | explicit_tune_core = selected_tune->ident; | |
0cfff2a1 KT |
9289 | } |
9290 | /* If both -mcpu and -march are specified check that they are architecturally | |
9291 | compatible, warn if they're not and prefer the -march ISA flags. */ | |
9292 | else if (selected_arch) | |
9293 | { | |
9294 | if (selected_arch->arch != selected_cpu->arch) | |
9295 | { | |
9296 | warning (0, "switch -mcpu=%s conflicts with -march=%s switch", | |
9297 | all_architectures[selected_cpu->arch].name, | |
9298 | selected_arch->name); | |
9299 | } | |
9300 | aarch64_isa_flags = arch_isa; | |
361fb3ee KT |
9301 | explicit_arch = selected_arch->arch; |
9302 | explicit_tune_core = selected_tune ? selected_tune->ident | |
9303 | : selected_cpu->ident; | |
0cfff2a1 KT |
9304 | } |
9305 | else | |
9306 | { | |
9307 | /* -mcpu but no -march. */ | |
9308 | aarch64_isa_flags = cpu_isa; | |
361fb3ee KT |
9309 | explicit_tune_core = selected_tune ? selected_tune->ident |
9310 | : selected_cpu->ident; | |
9311 | gcc_assert (selected_cpu); | |
9312 | selected_arch = &all_architectures[selected_cpu->arch]; | |
9313 | explicit_arch = selected_arch->arch; | |
43e9d192 IB |
9314 | } |
9315 | ||
0cfff2a1 KT |
9316 | /* Set the arch as well as we will need it when outputing |
9317 | the .arch directive in assembly. */ | |
9318 | if (!selected_arch) | |
9319 | { | |
9320 | gcc_assert (selected_cpu); | |
9321 | selected_arch = &all_architectures[selected_cpu->arch]; | |
9322 | } | |
43e9d192 | 9323 | |
43e9d192 | 9324 | if (!selected_tune) |
3edaf26d | 9325 | selected_tune = selected_cpu; |
43e9d192 | 9326 | |
0cfff2a1 KT |
9327 | #ifndef HAVE_AS_MABI_OPTION |
9328 | /* The compiler may have been configured with 2.23.* binutils, which does | |
9329 | not have support for ILP32. */ | |
9330 | if (TARGET_ILP32) | |
9331 | error ("Assembler does not support -mabi=ilp32"); | |
9332 | #endif | |
43e9d192 | 9333 | |
db58fd89 JW |
9334 | if (aarch64_ra_sign_scope != AARCH64_FUNCTION_NONE && TARGET_ILP32) |
9335 | sorry ("Return address signing is only supported for -mabi=lp64"); | |
9336 | ||
361fb3ee KT |
9337 | /* Make sure we properly set up the explicit options. */ |
9338 | if ((aarch64_cpu_string && valid_cpu) | |
9339 | || (aarch64_tune_string && valid_tune)) | |
9340 | gcc_assert (explicit_tune_core != aarch64_none); | |
9341 | ||
9342 | if ((aarch64_cpu_string && valid_cpu) | |
9343 | || (aarch64_arch_string && valid_arch)) | |
9344 | gcc_assert (explicit_arch != aarch64_no_arch); | |
9345 | ||
0cfff2a1 KT |
9346 | aarch64_override_options_internal (&global_options); |
9347 | ||
9348 | /* Save these options as the default ones in case we push and pop them later | |
9349 | while processing functions with potential target attributes. */ | |
9350 | target_option_default_node = target_option_current_node | |
9351 | = build_target_option_node (&global_options); | |
43e9d192 IB |
9352 | } |
9353 | ||
9354 | /* Implement targetm.override_options_after_change. */ | |
9355 | ||
9356 | static void | |
9357 | aarch64_override_options_after_change (void) | |
9358 | { | |
0cfff2a1 | 9359 | aarch64_override_options_after_change_1 (&global_options); |
43e9d192 IB |
9360 | } |
9361 | ||
9362 | static struct machine_function * | |
9363 | aarch64_init_machine_status (void) | |
9364 | { | |
9365 | struct machine_function *machine; | |
766090c2 | 9366 | machine = ggc_cleared_alloc<machine_function> (); |
43e9d192 IB |
9367 | return machine; |
9368 | } | |
9369 | ||
9370 | void | |
9371 | aarch64_init_expanders (void) | |
9372 | { | |
9373 | init_machine_status = aarch64_init_machine_status; | |
9374 | } | |
9375 | ||
9376 | /* A checking mechanism for the implementation of the various code models. */ | |
9377 | static void | |
0cfff2a1 | 9378 | initialize_aarch64_code_model (struct gcc_options *opts) |
43e9d192 | 9379 | { |
0cfff2a1 | 9380 | if (opts->x_flag_pic) |
43e9d192 | 9381 | { |
0cfff2a1 | 9382 | switch (opts->x_aarch64_cmodel_var) |
43e9d192 IB |
9383 | { |
9384 | case AARCH64_CMODEL_TINY: | |
9385 | aarch64_cmodel = AARCH64_CMODEL_TINY_PIC; | |
9386 | break; | |
9387 | case AARCH64_CMODEL_SMALL: | |
34ecdb0f | 9388 | #ifdef HAVE_AS_SMALL_PIC_RELOCS |
1b1e81f8 JW |
9389 | aarch64_cmodel = (flag_pic == 2 |
9390 | ? AARCH64_CMODEL_SMALL_PIC | |
9391 | : AARCH64_CMODEL_SMALL_SPIC); | |
34ecdb0f JW |
9392 | #else |
9393 | aarch64_cmodel = AARCH64_CMODEL_SMALL_PIC; | |
9394 | #endif | |
43e9d192 IB |
9395 | break; |
9396 | case AARCH64_CMODEL_LARGE: | |
9397 | sorry ("code model %qs with -f%s", "large", | |
0cfff2a1 | 9398 | opts->x_flag_pic > 1 ? "PIC" : "pic"); |
1c652781 | 9399 | break; |
43e9d192 IB |
9400 | default: |
9401 | gcc_unreachable (); | |
9402 | } | |
9403 | } | |
9404 | else | |
0cfff2a1 | 9405 | aarch64_cmodel = opts->x_aarch64_cmodel_var; |
43e9d192 IB |
9406 | } |
9407 | ||
361fb3ee KT |
9408 | /* Implement TARGET_OPTION_SAVE. */ |
9409 | ||
9410 | static void | |
9411 | aarch64_option_save (struct cl_target_option *ptr, struct gcc_options *opts) | |
9412 | { | |
9413 | ptr->x_aarch64_override_tune_string = opts->x_aarch64_override_tune_string; | |
9414 | } | |
9415 | ||
9416 | /* Implements TARGET_OPTION_RESTORE. Restore the backend codegen decisions | |
9417 | using the information saved in PTR. */ | |
9418 | ||
9419 | static void | |
9420 | aarch64_option_restore (struct gcc_options *opts, struct cl_target_option *ptr) | |
9421 | { | |
9422 | opts->x_explicit_tune_core = ptr->x_explicit_tune_core; | |
9423 | selected_tune = aarch64_get_tune_cpu (ptr->x_explicit_tune_core); | |
9424 | opts->x_explicit_arch = ptr->x_explicit_arch; | |
9425 | selected_arch = aarch64_get_arch (ptr->x_explicit_arch); | |
9426 | opts->x_aarch64_override_tune_string = ptr->x_aarch64_override_tune_string; | |
9427 | ||
9428 | aarch64_override_options_internal (opts); | |
9429 | } | |
9430 | ||
9431 | /* Implement TARGET_OPTION_PRINT. */ | |
9432 | ||
9433 | static void | |
9434 | aarch64_option_print (FILE *file, int indent, struct cl_target_option *ptr) | |
9435 | { | |
9436 | const struct processor *cpu | |
9437 | = aarch64_get_tune_cpu (ptr->x_explicit_tune_core); | |
9438 | unsigned long isa_flags = ptr->x_aarch64_isa_flags; | |
9439 | const struct processor *arch = aarch64_get_arch (ptr->x_explicit_arch); | |
054b4005 | 9440 | std::string extension |
04a99ebe | 9441 | = aarch64_get_extension_string_for_isa_flags (isa_flags, arch->flags); |
361fb3ee KT |
9442 | |
9443 | fprintf (file, "%*sselected tune = %s\n", indent, "", cpu->name); | |
054b4005 JG |
9444 | fprintf (file, "%*sselected arch = %s%s\n", indent, "", |
9445 | arch->name, extension.c_str ()); | |
361fb3ee KT |
9446 | } |
9447 | ||
d78006d9 KT |
9448 | static GTY(()) tree aarch64_previous_fndecl; |
9449 | ||
e4ea20c8 KT |
9450 | void |
9451 | aarch64_reset_previous_fndecl (void) | |
9452 | { | |
9453 | aarch64_previous_fndecl = NULL; | |
9454 | } | |
9455 | ||
acfc1ac1 KT |
9456 | /* Restore or save the TREE_TARGET_GLOBALS from or to NEW_TREE. |
9457 | Used by aarch64_set_current_function and aarch64_pragma_target_parse to | |
9458 | make sure optab availability predicates are recomputed when necessary. */ | |
9459 | ||
9460 | void | |
9461 | aarch64_save_restore_target_globals (tree new_tree) | |
9462 | { | |
9463 | if (TREE_TARGET_GLOBALS (new_tree)) | |
9464 | restore_target_globals (TREE_TARGET_GLOBALS (new_tree)); | |
9465 | else if (new_tree == target_option_default_node) | |
9466 | restore_target_globals (&default_target_globals); | |
9467 | else | |
9468 | TREE_TARGET_GLOBALS (new_tree) = save_target_globals_default_opts (); | |
9469 | } | |
9470 | ||
d78006d9 KT |
9471 | /* Implement TARGET_SET_CURRENT_FUNCTION. Unpack the codegen decisions |
9472 | like tuning and ISA features from the DECL_FUNCTION_SPECIFIC_TARGET | |
9473 | of the function, if such exists. This function may be called multiple | |
9474 | times on a single function so use aarch64_previous_fndecl to avoid | |
9475 | setting up identical state. */ | |
9476 | ||
9477 | static void | |
9478 | aarch64_set_current_function (tree fndecl) | |
9479 | { | |
acfc1ac1 KT |
9480 | if (!fndecl || fndecl == aarch64_previous_fndecl) |
9481 | return; | |
9482 | ||
d78006d9 KT |
9483 | tree old_tree = (aarch64_previous_fndecl |
9484 | ? DECL_FUNCTION_SPECIFIC_TARGET (aarch64_previous_fndecl) | |
9485 | : NULL_TREE); | |
9486 | ||
acfc1ac1 | 9487 | tree new_tree = DECL_FUNCTION_SPECIFIC_TARGET (fndecl); |
d78006d9 | 9488 | |
acfc1ac1 KT |
9489 | /* If current function has no attributes but the previous one did, |
9490 | use the default node. */ | |
9491 | if (!new_tree && old_tree) | |
9492 | new_tree = target_option_default_node; | |
d78006d9 | 9493 | |
acfc1ac1 KT |
9494 | /* If nothing to do, return. #pragma GCC reset or #pragma GCC pop to |
9495 | the default have been handled by aarch64_save_restore_target_globals from | |
9496 | aarch64_pragma_target_parse. */ | |
9497 | if (old_tree == new_tree) | |
9498 | return; | |
d78006d9 | 9499 | |
acfc1ac1 | 9500 | aarch64_previous_fndecl = fndecl; |
6e17a23b | 9501 | |
acfc1ac1 KT |
9502 | /* First set the target options. */ |
9503 | cl_target_option_restore (&global_options, TREE_TARGET_OPTION (new_tree)); | |
6e17a23b | 9504 | |
acfc1ac1 | 9505 | aarch64_save_restore_target_globals (new_tree); |
d78006d9 | 9506 | } |
361fb3ee | 9507 | |
5a2c8331 KT |
9508 | /* Enum describing the various ways we can handle attributes. |
9509 | In many cases we can reuse the generic option handling machinery. */ | |
9510 | ||
9511 | enum aarch64_attr_opt_type | |
9512 | { | |
9513 | aarch64_attr_mask, /* Attribute should set a bit in target_flags. */ | |
9514 | aarch64_attr_bool, /* Attribute sets or unsets a boolean variable. */ | |
9515 | aarch64_attr_enum, /* Attribute sets an enum variable. */ | |
9516 | aarch64_attr_custom /* Attribute requires a custom handling function. */ | |
9517 | }; | |
9518 | ||
9519 | /* All the information needed to handle a target attribute. | |
9520 | NAME is the name of the attribute. | |
9c582551 | 9521 | ATTR_TYPE specifies the type of behavior of the attribute as described |
5a2c8331 KT |
9522 | in the definition of enum aarch64_attr_opt_type. |
9523 | ALLOW_NEG is true if the attribute supports a "no-" form. | |
9524 | HANDLER is the function that takes the attribute string and whether | |
9525 | it is a pragma or attribute and handles the option. It is needed only | |
9526 | when the ATTR_TYPE is aarch64_attr_custom. | |
9527 | OPT_NUM is the enum specifying the option that the attribute modifies. | |
9c582551 | 9528 | This is needed for attributes that mirror the behavior of a command-line |
5a2c8331 KT |
9529 | option, that is it has ATTR_TYPE aarch64_attr_mask, aarch64_attr_bool or |
9530 | aarch64_attr_enum. */ | |
9531 | ||
9532 | struct aarch64_attribute_info | |
9533 | { | |
9534 | const char *name; | |
9535 | enum aarch64_attr_opt_type attr_type; | |
9536 | bool allow_neg; | |
9537 | bool (*handler) (const char *, const char *); | |
9538 | enum opt_code opt_num; | |
9539 | }; | |
9540 | ||
9541 | /* Handle the ARCH_STR argument to the arch= target attribute. | |
9542 | PRAGMA_OR_ATTR is used in potential error messages. */ | |
9543 | ||
9544 | static bool | |
9545 | aarch64_handle_attr_arch (const char *str, const char *pragma_or_attr) | |
9546 | { | |
9547 | const struct processor *tmp_arch = NULL; | |
9548 | enum aarch64_parse_opt_result parse_res | |
9549 | = aarch64_parse_arch (str, &tmp_arch, &aarch64_isa_flags); | |
9550 | ||
9551 | if (parse_res == AARCH64_PARSE_OK) | |
9552 | { | |
9553 | gcc_assert (tmp_arch); | |
9554 | selected_arch = tmp_arch; | |
9555 | explicit_arch = selected_arch->arch; | |
9556 | return true; | |
9557 | } | |
9558 | ||
9559 | switch (parse_res) | |
9560 | { | |
9561 | case AARCH64_PARSE_MISSING_ARG: | |
9562 | error ("missing architecture name in 'arch' target %s", pragma_or_attr); | |
9563 | break; | |
9564 | case AARCH64_PARSE_INVALID_ARG: | |
9565 | error ("unknown value %qs for 'arch' target %s", str, pragma_or_attr); | |
01f44038 | 9566 | aarch64_print_hint_for_arch (str); |
5a2c8331 KT |
9567 | break; |
9568 | case AARCH64_PARSE_INVALID_FEATURE: | |
9569 | error ("invalid feature modifier %qs for 'arch' target %s", | |
9570 | str, pragma_or_attr); | |
9571 | break; | |
9572 | default: | |
9573 | gcc_unreachable (); | |
9574 | } | |
9575 | ||
9576 | return false; | |
9577 | } | |
9578 | ||
9579 | /* Handle the argument CPU_STR to the cpu= target attribute. | |
9580 | PRAGMA_OR_ATTR is used in potential error messages. */ | |
9581 | ||
9582 | static bool | |
9583 | aarch64_handle_attr_cpu (const char *str, const char *pragma_or_attr) | |
9584 | { | |
9585 | const struct processor *tmp_cpu = NULL; | |
9586 | enum aarch64_parse_opt_result parse_res | |
9587 | = aarch64_parse_cpu (str, &tmp_cpu, &aarch64_isa_flags); | |
9588 | ||
9589 | if (parse_res == AARCH64_PARSE_OK) | |
9590 | { | |
9591 | gcc_assert (tmp_cpu); | |
9592 | selected_tune = tmp_cpu; | |
9593 | explicit_tune_core = selected_tune->ident; | |
9594 | ||
9595 | selected_arch = &all_architectures[tmp_cpu->arch]; | |
9596 | explicit_arch = selected_arch->arch; | |
9597 | return true; | |
9598 | } | |
9599 | ||
9600 | switch (parse_res) | |
9601 | { | |
9602 | case AARCH64_PARSE_MISSING_ARG: | |
9603 | error ("missing cpu name in 'cpu' target %s", pragma_or_attr); | |
9604 | break; | |
9605 | case AARCH64_PARSE_INVALID_ARG: | |
9606 | error ("unknown value %qs for 'cpu' target %s", str, pragma_or_attr); | |
01f44038 | 9607 | aarch64_print_hint_for_core (str); |
5a2c8331 KT |
9608 | break; |
9609 | case AARCH64_PARSE_INVALID_FEATURE: | |
9610 | error ("invalid feature modifier %qs for 'cpu' target %s", | |
9611 | str, pragma_or_attr); | |
9612 | break; | |
9613 | default: | |
9614 | gcc_unreachable (); | |
9615 | } | |
9616 | ||
9617 | return false; | |
9618 | } | |
9619 | ||
9620 | /* Handle the argument STR to the tune= target attribute. | |
9621 | PRAGMA_OR_ATTR is used in potential error messages. */ | |
9622 | ||
9623 | static bool | |
9624 | aarch64_handle_attr_tune (const char *str, const char *pragma_or_attr) | |
9625 | { | |
9626 | const struct processor *tmp_tune = NULL; | |
9627 | enum aarch64_parse_opt_result parse_res | |
9628 | = aarch64_parse_tune (str, &tmp_tune); | |
9629 | ||
9630 | if (parse_res == AARCH64_PARSE_OK) | |
9631 | { | |
9632 | gcc_assert (tmp_tune); | |
9633 | selected_tune = tmp_tune; | |
9634 | explicit_tune_core = selected_tune->ident; | |
9635 | return true; | |
9636 | } | |
9637 | ||
9638 | switch (parse_res) | |
9639 | { | |
9640 | case AARCH64_PARSE_INVALID_ARG: | |
9641 | error ("unknown value %qs for 'tune' target %s", str, pragma_or_attr); | |
01f44038 | 9642 | aarch64_print_hint_for_core (str); |
5a2c8331 KT |
9643 | break; |
9644 | default: | |
9645 | gcc_unreachable (); | |
9646 | } | |
9647 | ||
9648 | return false; | |
9649 | } | |
9650 | ||
9651 | /* Parse an architecture extensions target attribute string specified in STR. | |
9652 | For example "+fp+nosimd". Show any errors if needed. Return TRUE | |
9653 | if successful. Update aarch64_isa_flags to reflect the ISA features | |
9654 | modified. | |
9655 | PRAGMA_OR_ATTR is used in potential error messages. */ | |
9656 | ||
9657 | static bool | |
9658 | aarch64_handle_attr_isa_flags (char *str, const char *pragma_or_attr) | |
9659 | { | |
9660 | enum aarch64_parse_opt_result parse_res; | |
9661 | unsigned long isa_flags = aarch64_isa_flags; | |
9662 | ||
e4ea20c8 KT |
9663 | /* We allow "+nothing" in the beginning to clear out all architectural |
9664 | features if the user wants to handpick specific features. */ | |
9665 | if (strncmp ("+nothing", str, 8) == 0) | |
9666 | { | |
9667 | isa_flags = 0; | |
9668 | str += 8; | |
9669 | } | |
9670 | ||
5a2c8331 KT |
9671 | parse_res = aarch64_parse_extension (str, &isa_flags); |
9672 | ||
9673 | if (parse_res == AARCH64_PARSE_OK) | |
9674 | { | |
9675 | aarch64_isa_flags = isa_flags; | |
9676 | return true; | |
9677 | } | |
9678 | ||
9679 | switch (parse_res) | |
9680 | { | |
9681 | case AARCH64_PARSE_MISSING_ARG: | |
9682 | error ("missing feature modifier in target %s %qs", | |
9683 | pragma_or_attr, str); | |
9684 | break; | |
9685 | ||
9686 | case AARCH64_PARSE_INVALID_FEATURE: | |
9687 | error ("invalid feature modifier in target %s %qs", | |
9688 | pragma_or_attr, str); | |
9689 | break; | |
9690 | ||
9691 | default: | |
9692 | gcc_unreachable (); | |
9693 | } | |
9694 | ||
9695 | return false; | |
9696 | } | |
9697 | ||
9698 | /* The target attributes that we support. On top of these we also support just | |
9699 | ISA extensions, like __attribute__ ((target ("+crc"))), but that case is | |
9700 | handled explicitly in aarch64_process_one_target_attr. */ | |
9701 | ||
9702 | static const struct aarch64_attribute_info aarch64_attributes[] = | |
9703 | { | |
9704 | { "general-regs-only", aarch64_attr_mask, false, NULL, | |
9705 | OPT_mgeneral_regs_only }, | |
9706 | { "fix-cortex-a53-835769", aarch64_attr_bool, true, NULL, | |
9707 | OPT_mfix_cortex_a53_835769 }, | |
48bb1a55 CL |
9708 | { "fix-cortex-a53-843419", aarch64_attr_bool, true, NULL, |
9709 | OPT_mfix_cortex_a53_843419 }, | |
5a2c8331 KT |
9710 | { "cmodel", aarch64_attr_enum, false, NULL, OPT_mcmodel_ }, |
9711 | { "strict-align", aarch64_attr_mask, false, NULL, OPT_mstrict_align }, | |
9712 | { "omit-leaf-frame-pointer", aarch64_attr_bool, true, NULL, | |
9713 | OPT_momit_leaf_frame_pointer }, | |
9714 | { "tls-dialect", aarch64_attr_enum, false, NULL, OPT_mtls_dialect_ }, | |
9715 | { "arch", aarch64_attr_custom, false, aarch64_handle_attr_arch, | |
9716 | OPT_march_ }, | |
9717 | { "cpu", aarch64_attr_custom, false, aarch64_handle_attr_cpu, OPT_mcpu_ }, | |
9718 | { "tune", aarch64_attr_custom, false, aarch64_handle_attr_tune, | |
9719 | OPT_mtune_ }, | |
db58fd89 JW |
9720 | { "sign-return-address", aarch64_attr_enum, false, NULL, |
9721 | OPT_msign_return_address_ }, | |
5a2c8331 KT |
9722 | { NULL, aarch64_attr_custom, false, NULL, OPT____ } |
9723 | }; | |
9724 | ||
9725 | /* Parse ARG_STR which contains the definition of one target attribute. | |
9726 | Show appropriate errors if any or return true if the attribute is valid. | |
9727 | PRAGMA_OR_ATTR holds the string to use in error messages about whether | |
9728 | we're processing a target attribute or pragma. */ | |
9729 | ||
9730 | static bool | |
9731 | aarch64_process_one_target_attr (char *arg_str, const char* pragma_or_attr) | |
9732 | { | |
9733 | bool invert = false; | |
9734 | ||
9735 | size_t len = strlen (arg_str); | |
9736 | ||
9737 | if (len == 0) | |
9738 | { | |
9739 | error ("malformed target %s", pragma_or_attr); | |
9740 | return false; | |
9741 | } | |
9742 | ||
9743 | char *str_to_check = (char *) alloca (len + 1); | |
9744 | strcpy (str_to_check, arg_str); | |
9745 | ||
9746 | /* Skip leading whitespace. */ | |
9747 | while (*str_to_check == ' ' || *str_to_check == '\t') | |
9748 | str_to_check++; | |
9749 | ||
9750 | /* We have something like __attribute__ ((target ("+fp+nosimd"))). | |
9751 | It is easier to detect and handle it explicitly here rather than going | |
9752 | through the machinery for the rest of the target attributes in this | |
9753 | function. */ | |
9754 | if (*str_to_check == '+') | |
9755 | return aarch64_handle_attr_isa_flags (str_to_check, pragma_or_attr); | |
9756 | ||
9757 | if (len > 3 && strncmp (str_to_check, "no-", 3) == 0) | |
9758 | { | |
9759 | invert = true; | |
9760 | str_to_check += 3; | |
9761 | } | |
9762 | char *arg = strchr (str_to_check, '='); | |
9763 | ||
9764 | /* If we found opt=foo then terminate STR_TO_CHECK at the '=' | |
9765 | and point ARG to "foo". */ | |
9766 | if (arg) | |
9767 | { | |
9768 | *arg = '\0'; | |
9769 | arg++; | |
9770 | } | |
9771 | const struct aarch64_attribute_info *p_attr; | |
16d12992 | 9772 | bool found = false; |
5a2c8331 KT |
9773 | for (p_attr = aarch64_attributes; p_attr->name; p_attr++) |
9774 | { | |
9775 | /* If the names don't match up, or the user has given an argument | |
9776 | to an attribute that doesn't accept one, or didn't give an argument | |
9777 | to an attribute that expects one, fail to match. */ | |
9778 | if (strcmp (str_to_check, p_attr->name) != 0) | |
9779 | continue; | |
9780 | ||
16d12992 | 9781 | found = true; |
5a2c8331 KT |
9782 | bool attr_need_arg_p = p_attr->attr_type == aarch64_attr_custom |
9783 | || p_attr->attr_type == aarch64_attr_enum; | |
9784 | ||
9785 | if (attr_need_arg_p ^ (arg != NULL)) | |
9786 | { | |
9787 | error ("target %s %qs does not accept an argument", | |
9788 | pragma_or_attr, str_to_check); | |
9789 | return false; | |
9790 | } | |
9791 | ||
9792 | /* If the name matches but the attribute does not allow "no-" versions | |
9793 | then we can't match. */ | |
9794 | if (invert && !p_attr->allow_neg) | |
9795 | { | |
9796 | error ("target %s %qs does not allow a negated form", | |
9797 | pragma_or_attr, str_to_check); | |
9798 | return false; | |
9799 | } | |
9800 | ||
9801 | switch (p_attr->attr_type) | |
9802 | { | |
9803 | /* Has a custom handler registered. | |
9804 | For example, cpu=, arch=, tune=. */ | |
9805 | case aarch64_attr_custom: | |
9806 | gcc_assert (p_attr->handler); | |
9807 | if (!p_attr->handler (arg, pragma_or_attr)) | |
9808 | return false; | |
9809 | break; | |
9810 | ||
9811 | /* Either set or unset a boolean option. */ | |
9812 | case aarch64_attr_bool: | |
9813 | { | |
9814 | struct cl_decoded_option decoded; | |
9815 | ||
9816 | generate_option (p_attr->opt_num, NULL, !invert, | |
9817 | CL_TARGET, &decoded); | |
9818 | aarch64_handle_option (&global_options, &global_options_set, | |
9819 | &decoded, input_location); | |
9820 | break; | |
9821 | } | |
9822 | /* Set or unset a bit in the target_flags. aarch64_handle_option | |
9823 | should know what mask to apply given the option number. */ | |
9824 | case aarch64_attr_mask: | |
9825 | { | |
9826 | struct cl_decoded_option decoded; | |
9827 | /* We only need to specify the option number. | |
9828 | aarch64_handle_option will know which mask to apply. */ | |
9829 | decoded.opt_index = p_attr->opt_num; | |
9830 | decoded.value = !invert; | |
9831 | aarch64_handle_option (&global_options, &global_options_set, | |
9832 | &decoded, input_location); | |
9833 | break; | |
9834 | } | |
9835 | /* Use the option setting machinery to set an option to an enum. */ | |
9836 | case aarch64_attr_enum: | |
9837 | { | |
9838 | gcc_assert (arg); | |
9839 | bool valid; | |
9840 | int value; | |
9841 | valid = opt_enum_arg_to_value (p_attr->opt_num, arg, | |
9842 | &value, CL_TARGET); | |
9843 | if (valid) | |
9844 | { | |
9845 | set_option (&global_options, NULL, p_attr->opt_num, value, | |
9846 | NULL, DK_UNSPECIFIED, input_location, | |
9847 | global_dc); | |
9848 | } | |
9849 | else | |
9850 | { | |
9851 | error ("target %s %s=%s is not valid", | |
9852 | pragma_or_attr, str_to_check, arg); | |
9853 | } | |
9854 | break; | |
9855 | } | |
9856 | default: | |
9857 | gcc_unreachable (); | |
9858 | } | |
9859 | } | |
9860 | ||
16d12992 KT |
9861 | /* If we reached here we either have found an attribute and validated |
9862 | it or didn't match any. If we matched an attribute but its arguments | |
9863 | were malformed we will have returned false already. */ | |
9864 | return found; | |
5a2c8331 KT |
9865 | } |
9866 | ||
9867 | /* Count how many times the character C appears in | |
9868 | NULL-terminated string STR. */ | |
9869 | ||
9870 | static unsigned int | |
9871 | num_occurences_in_str (char c, char *str) | |
9872 | { | |
9873 | unsigned int res = 0; | |
9874 | while (*str != '\0') | |
9875 | { | |
9876 | if (*str == c) | |
9877 | res++; | |
9878 | ||
9879 | str++; | |
9880 | } | |
9881 | ||
9882 | return res; | |
9883 | } | |
9884 | ||
9885 | /* Parse the tree in ARGS that contains the target attribute information | |
9886 | and update the global target options space. PRAGMA_OR_ATTR is a string | |
9887 | to be used in error messages, specifying whether this is processing | |
9888 | a target attribute or a target pragma. */ | |
9889 | ||
9890 | bool | |
9891 | aarch64_process_target_attr (tree args, const char* pragma_or_attr) | |
9892 | { | |
9893 | if (TREE_CODE (args) == TREE_LIST) | |
9894 | { | |
9895 | do | |
9896 | { | |
9897 | tree head = TREE_VALUE (args); | |
9898 | if (head) | |
9899 | { | |
9900 | if (!aarch64_process_target_attr (head, pragma_or_attr)) | |
9901 | return false; | |
9902 | } | |
9903 | args = TREE_CHAIN (args); | |
9904 | } while (args); | |
9905 | ||
9906 | return true; | |
9907 | } | |
3b6cb9e3 ML |
9908 | |
9909 | if (TREE_CODE (args) != STRING_CST) | |
9910 | { | |
9911 | error ("attribute %<target%> argument not a string"); | |
9912 | return false; | |
9913 | } | |
5a2c8331 KT |
9914 | |
9915 | size_t len = strlen (TREE_STRING_POINTER (args)); | |
9916 | char *str_to_check = (char *) alloca (len + 1); | |
9917 | strcpy (str_to_check, TREE_STRING_POINTER (args)); | |
9918 | ||
9919 | if (len == 0) | |
9920 | { | |
9921 | error ("malformed target %s value", pragma_or_attr); | |
9922 | return false; | |
9923 | } | |
9924 | ||
9925 | /* Used to catch empty spaces between commas i.e. | |
9926 | attribute ((target ("attr1,,attr2"))). */ | |
9927 | unsigned int num_commas = num_occurences_in_str (',', str_to_check); | |
9928 | ||
9929 | /* Handle multiple target attributes separated by ','. */ | |
9930 | char *token = strtok (str_to_check, ","); | |
9931 | ||
9932 | unsigned int num_attrs = 0; | |
9933 | while (token) | |
9934 | { | |
9935 | num_attrs++; | |
9936 | if (!aarch64_process_one_target_attr (token, pragma_or_attr)) | |
9937 | { | |
9938 | error ("target %s %qs is invalid", pragma_or_attr, token); | |
9939 | return false; | |
9940 | } | |
9941 | ||
9942 | token = strtok (NULL, ","); | |
9943 | } | |
9944 | ||
9945 | if (num_attrs != num_commas + 1) | |
9946 | { | |
9947 | error ("malformed target %s list %qs", | |
9948 | pragma_or_attr, TREE_STRING_POINTER (args)); | |
9949 | return false; | |
9950 | } | |
9951 | ||
9952 | return true; | |
9953 | } | |
9954 | ||
9955 | /* Implement TARGET_OPTION_VALID_ATTRIBUTE_P. This is used to | |
9956 | process attribute ((target ("..."))). */ | |
9957 | ||
9958 | static bool | |
9959 | aarch64_option_valid_attribute_p (tree fndecl, tree, tree args, int) | |
9960 | { | |
9961 | struct cl_target_option cur_target; | |
9962 | bool ret; | |
9963 | tree old_optimize; | |
9964 | tree new_target, new_optimize; | |
9965 | tree existing_target = DECL_FUNCTION_SPECIFIC_TARGET (fndecl); | |
91d0e8de KT |
9966 | |
9967 | /* If what we're processing is the current pragma string then the | |
9968 | target option node is already stored in target_option_current_node | |
9969 | by aarch64_pragma_target_parse in aarch64-c.c. Use that to avoid | |
9970 | having to re-parse the string. This is especially useful to keep | |
9971 | arm_neon.h compile times down since that header contains a lot | |
9972 | of intrinsics enclosed in pragmas. */ | |
9973 | if (!existing_target && args == current_target_pragma) | |
9974 | { | |
9975 | DECL_FUNCTION_SPECIFIC_TARGET (fndecl) = target_option_current_node; | |
9976 | return true; | |
9977 | } | |
5a2c8331 KT |
9978 | tree func_optimize = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl); |
9979 | ||
9980 | old_optimize = build_optimization_node (&global_options); | |
9981 | func_optimize = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl); | |
9982 | ||
9983 | /* If the function changed the optimization levels as well as setting | |
9984 | target options, start with the optimizations specified. */ | |
9985 | if (func_optimize && func_optimize != old_optimize) | |
9986 | cl_optimization_restore (&global_options, | |
9987 | TREE_OPTIMIZATION (func_optimize)); | |
9988 | ||
9989 | /* Save the current target options to restore at the end. */ | |
9990 | cl_target_option_save (&cur_target, &global_options); | |
9991 | ||
9992 | /* If fndecl already has some target attributes applied to it, unpack | |
9993 | them so that we add this attribute on top of them, rather than | |
9994 | overwriting them. */ | |
9995 | if (existing_target) | |
9996 | { | |
9997 | struct cl_target_option *existing_options | |
9998 | = TREE_TARGET_OPTION (existing_target); | |
9999 | ||
10000 | if (existing_options) | |
10001 | cl_target_option_restore (&global_options, existing_options); | |
10002 | } | |
10003 | else | |
10004 | cl_target_option_restore (&global_options, | |
10005 | TREE_TARGET_OPTION (target_option_current_node)); | |
10006 | ||
10007 | ||
10008 | ret = aarch64_process_target_attr (args, "attribute"); | |
10009 | ||
10010 | /* Set up any additional state. */ | |
10011 | if (ret) | |
10012 | { | |
10013 | aarch64_override_options_internal (&global_options); | |
e95a988a KT |
10014 | /* Initialize SIMD builtins if we haven't already. |
10015 | Set current_target_pragma to NULL for the duration so that | |
10016 | the builtin initialization code doesn't try to tag the functions | |
10017 | being built with the attributes specified by any current pragma, thus | |
10018 | going into an infinite recursion. */ | |
10019 | if (TARGET_SIMD) | |
10020 | { | |
10021 | tree saved_current_target_pragma = current_target_pragma; | |
10022 | current_target_pragma = NULL; | |
10023 | aarch64_init_simd_builtins (); | |
10024 | current_target_pragma = saved_current_target_pragma; | |
10025 | } | |
5a2c8331 KT |
10026 | new_target = build_target_option_node (&global_options); |
10027 | } | |
10028 | else | |
10029 | new_target = NULL; | |
10030 | ||
10031 | new_optimize = build_optimization_node (&global_options); | |
10032 | ||
10033 | if (fndecl && ret) | |
10034 | { | |
10035 | DECL_FUNCTION_SPECIFIC_TARGET (fndecl) = new_target; | |
10036 | ||
10037 | if (old_optimize != new_optimize) | |
10038 | DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl) = new_optimize; | |
10039 | } | |
10040 | ||
10041 | cl_target_option_restore (&global_options, &cur_target); | |
10042 | ||
10043 | if (old_optimize != new_optimize) | |
10044 | cl_optimization_restore (&global_options, | |
10045 | TREE_OPTIMIZATION (old_optimize)); | |
10046 | return ret; | |
10047 | } | |
10048 | ||
1fd8d40c KT |
10049 | /* Helper for aarch64_can_inline_p. In the case where CALLER and CALLEE are |
10050 | tri-bool options (yes, no, don't care) and the default value is | |
10051 | DEF, determine whether to reject inlining. */ | |
10052 | ||
10053 | static bool | |
10054 | aarch64_tribools_ok_for_inlining_p (int caller, int callee, | |
10055 | int dont_care, int def) | |
10056 | { | |
10057 | /* If the callee doesn't care, always allow inlining. */ | |
10058 | if (callee == dont_care) | |
10059 | return true; | |
10060 | ||
10061 | /* If the caller doesn't care, always allow inlining. */ | |
10062 | if (caller == dont_care) | |
10063 | return true; | |
10064 | ||
10065 | /* Otherwise, allow inlining if either the callee and caller values | |
10066 | agree, or if the callee is using the default value. */ | |
10067 | return (callee == caller || callee == def); | |
10068 | } | |
10069 | ||
10070 | /* Implement TARGET_CAN_INLINE_P. Decide whether it is valid | |
10071 | to inline CALLEE into CALLER based on target-specific info. | |
10072 | Make sure that the caller and callee have compatible architectural | |
10073 | features. Then go through the other possible target attributes | |
10074 | and see if they can block inlining. Try not to reject always_inline | |
10075 | callees unless they are incompatible architecturally. */ | |
10076 | ||
10077 | static bool | |
10078 | aarch64_can_inline_p (tree caller, tree callee) | |
10079 | { | |
10080 | tree caller_tree = DECL_FUNCTION_SPECIFIC_TARGET (caller); | |
10081 | tree callee_tree = DECL_FUNCTION_SPECIFIC_TARGET (callee); | |
10082 | ||
10083 | /* If callee has no option attributes, then it is ok to inline. */ | |
10084 | if (!callee_tree) | |
10085 | return true; | |
10086 | ||
10087 | struct cl_target_option *caller_opts | |
10088 | = TREE_TARGET_OPTION (caller_tree ? caller_tree | |
10089 | : target_option_default_node); | |
10090 | ||
10091 | struct cl_target_option *callee_opts = TREE_TARGET_OPTION (callee_tree); | |
10092 | ||
10093 | ||
10094 | /* Callee's ISA flags should be a subset of the caller's. */ | |
10095 | if ((caller_opts->x_aarch64_isa_flags & callee_opts->x_aarch64_isa_flags) | |
10096 | != callee_opts->x_aarch64_isa_flags) | |
10097 | return false; | |
10098 | ||
10099 | /* Allow non-strict aligned functions inlining into strict | |
10100 | aligned ones. */ | |
10101 | if ((TARGET_STRICT_ALIGN_P (caller_opts->x_target_flags) | |
10102 | != TARGET_STRICT_ALIGN_P (callee_opts->x_target_flags)) | |
10103 | && !(!TARGET_STRICT_ALIGN_P (callee_opts->x_target_flags) | |
10104 | && TARGET_STRICT_ALIGN_P (caller_opts->x_target_flags))) | |
10105 | return false; | |
10106 | ||
10107 | bool always_inline = lookup_attribute ("always_inline", | |
10108 | DECL_ATTRIBUTES (callee)); | |
10109 | ||
10110 | /* If the architectural features match up and the callee is always_inline | |
10111 | then the other attributes don't matter. */ | |
10112 | if (always_inline) | |
10113 | return true; | |
10114 | ||
10115 | if (caller_opts->x_aarch64_cmodel_var | |
10116 | != callee_opts->x_aarch64_cmodel_var) | |
10117 | return false; | |
10118 | ||
10119 | if (caller_opts->x_aarch64_tls_dialect | |
10120 | != callee_opts->x_aarch64_tls_dialect) | |
10121 | return false; | |
10122 | ||
10123 | /* Honour explicit requests to workaround errata. */ | |
10124 | if (!aarch64_tribools_ok_for_inlining_p ( | |
10125 | caller_opts->x_aarch64_fix_a53_err835769, | |
10126 | callee_opts->x_aarch64_fix_a53_err835769, | |
10127 | 2, TARGET_FIX_ERR_A53_835769_DEFAULT)) | |
10128 | return false; | |
10129 | ||
48bb1a55 CL |
10130 | if (!aarch64_tribools_ok_for_inlining_p ( |
10131 | caller_opts->x_aarch64_fix_a53_err843419, | |
10132 | callee_opts->x_aarch64_fix_a53_err843419, | |
10133 | 2, TARGET_FIX_ERR_A53_843419)) | |
10134 | return false; | |
10135 | ||
1fd8d40c KT |
10136 | /* If the user explicitly specified -momit-leaf-frame-pointer for the |
10137 | caller and calle and they don't match up, reject inlining. */ | |
10138 | if (!aarch64_tribools_ok_for_inlining_p ( | |
10139 | caller_opts->x_flag_omit_leaf_frame_pointer, | |
10140 | callee_opts->x_flag_omit_leaf_frame_pointer, | |
10141 | 2, 1)) | |
10142 | return false; | |
10143 | ||
10144 | /* If the callee has specific tuning overrides, respect them. */ | |
10145 | if (callee_opts->x_aarch64_override_tune_string != NULL | |
10146 | && caller_opts->x_aarch64_override_tune_string == NULL) | |
10147 | return false; | |
10148 | ||
10149 | /* If the user specified tuning override strings for the | |
10150 | caller and callee and they don't match up, reject inlining. | |
10151 | We just do a string compare here, we don't analyze the meaning | |
10152 | of the string, as it would be too costly for little gain. */ | |
10153 | if (callee_opts->x_aarch64_override_tune_string | |
10154 | && caller_opts->x_aarch64_override_tune_string | |
10155 | && (strcmp (callee_opts->x_aarch64_override_tune_string, | |
10156 | caller_opts->x_aarch64_override_tune_string) != 0)) | |
10157 | return false; | |
10158 | ||
10159 | return true; | |
10160 | } | |
10161 | ||
43e9d192 IB |
10162 | /* Return true if SYMBOL_REF X binds locally. */ |
10163 | ||
10164 | static bool | |
10165 | aarch64_symbol_binds_local_p (const_rtx x) | |
10166 | { | |
10167 | return (SYMBOL_REF_DECL (x) | |
10168 | ? targetm.binds_local_p (SYMBOL_REF_DECL (x)) | |
10169 | : SYMBOL_REF_LOCAL_P (x)); | |
10170 | } | |
10171 | ||
10172 | /* Return true if SYMBOL_REF X is thread local */ | |
10173 | static bool | |
10174 | aarch64_tls_symbol_p (rtx x) | |
10175 | { | |
10176 | if (! TARGET_HAVE_TLS) | |
10177 | return false; | |
10178 | ||
10179 | if (GET_CODE (x) != SYMBOL_REF) | |
10180 | return false; | |
10181 | ||
10182 | return SYMBOL_REF_TLS_MODEL (x) != 0; | |
10183 | } | |
10184 | ||
10185 | /* Classify a TLS symbol into one of the TLS kinds. */ | |
10186 | enum aarch64_symbol_type | |
10187 | aarch64_classify_tls_symbol (rtx x) | |
10188 | { | |
10189 | enum tls_model tls_kind = tls_symbolic_operand_type (x); | |
10190 | ||
10191 | switch (tls_kind) | |
10192 | { | |
10193 | case TLS_MODEL_GLOBAL_DYNAMIC: | |
10194 | case TLS_MODEL_LOCAL_DYNAMIC: | |
10195 | return TARGET_TLS_DESC ? SYMBOL_SMALL_TLSDESC : SYMBOL_SMALL_TLSGD; | |
10196 | ||
10197 | case TLS_MODEL_INITIAL_EXEC: | |
5ae7caad JW |
10198 | switch (aarch64_cmodel) |
10199 | { | |
10200 | case AARCH64_CMODEL_TINY: | |
10201 | case AARCH64_CMODEL_TINY_PIC: | |
10202 | return SYMBOL_TINY_TLSIE; | |
10203 | default: | |
79496620 | 10204 | return SYMBOL_SMALL_TLSIE; |
5ae7caad | 10205 | } |
43e9d192 IB |
10206 | |
10207 | case TLS_MODEL_LOCAL_EXEC: | |
cbf5629e JW |
10208 | if (aarch64_tls_size == 12) |
10209 | return SYMBOL_TLSLE12; | |
10210 | else if (aarch64_tls_size == 24) | |
10211 | return SYMBOL_TLSLE24; | |
10212 | else if (aarch64_tls_size == 32) | |
10213 | return SYMBOL_TLSLE32; | |
10214 | else if (aarch64_tls_size == 48) | |
10215 | return SYMBOL_TLSLE48; | |
10216 | else | |
10217 | gcc_unreachable (); | |
43e9d192 IB |
10218 | |
10219 | case TLS_MODEL_EMULATED: | |
10220 | case TLS_MODEL_NONE: | |
10221 | return SYMBOL_FORCE_TO_MEM; | |
10222 | ||
10223 | default: | |
10224 | gcc_unreachable (); | |
10225 | } | |
10226 | } | |
10227 | ||
10228 | /* Return the method that should be used to access SYMBOL_REF or | |
a6e0bfa7 | 10229 | LABEL_REF X. */ |
17f4d4bf | 10230 | |
43e9d192 | 10231 | enum aarch64_symbol_type |
a6e0bfa7 | 10232 | aarch64_classify_symbol (rtx x, rtx offset) |
43e9d192 IB |
10233 | { |
10234 | if (GET_CODE (x) == LABEL_REF) | |
10235 | { | |
10236 | switch (aarch64_cmodel) | |
10237 | { | |
10238 | case AARCH64_CMODEL_LARGE: | |
10239 | return SYMBOL_FORCE_TO_MEM; | |
10240 | ||
10241 | case AARCH64_CMODEL_TINY_PIC: | |
10242 | case AARCH64_CMODEL_TINY: | |
a5350ddc CSS |
10243 | return SYMBOL_TINY_ABSOLUTE; |
10244 | ||
1b1e81f8 | 10245 | case AARCH64_CMODEL_SMALL_SPIC: |
43e9d192 IB |
10246 | case AARCH64_CMODEL_SMALL_PIC: |
10247 | case AARCH64_CMODEL_SMALL: | |
10248 | return SYMBOL_SMALL_ABSOLUTE; | |
10249 | ||
10250 | default: | |
10251 | gcc_unreachable (); | |
10252 | } | |
10253 | } | |
10254 | ||
17f4d4bf | 10255 | if (GET_CODE (x) == SYMBOL_REF) |
43e9d192 | 10256 | { |
43e9d192 IB |
10257 | if (aarch64_tls_symbol_p (x)) |
10258 | return aarch64_classify_tls_symbol (x); | |
10259 | ||
17f4d4bf CSS |
10260 | switch (aarch64_cmodel) |
10261 | { | |
10262 | case AARCH64_CMODEL_TINY: | |
15f6e0da | 10263 | /* When we retrieve symbol + offset address, we have to make sure |
f8b756b7 TB |
10264 | the offset does not cause overflow of the final address. But |
10265 | we have no way of knowing the address of symbol at compile time | |
10266 | so we can't accurately say if the distance between the PC and | |
10267 | symbol + offset is outside the addressible range of +/-1M in the | |
10268 | TINY code model. So we rely on images not being greater than | |
10269 | 1M and cap the offset at 1M and anything beyond 1M will have to | |
15f6e0da RR |
10270 | be loaded using an alternative mechanism. Furthermore if the |
10271 | symbol is a weak reference to something that isn't known to | |
10272 | resolve to a symbol in this module, then force to memory. */ | |
10273 | if ((SYMBOL_REF_WEAK (x) | |
10274 | && !aarch64_symbol_binds_local_p (x)) | |
f8b756b7 | 10275 | || INTVAL (offset) < -1048575 || INTVAL (offset) > 1048575) |
a5350ddc CSS |
10276 | return SYMBOL_FORCE_TO_MEM; |
10277 | return SYMBOL_TINY_ABSOLUTE; | |
10278 | ||
17f4d4bf | 10279 | case AARCH64_CMODEL_SMALL: |
f8b756b7 TB |
10280 | /* Same reasoning as the tiny code model, but the offset cap here is |
10281 | 4G. */ | |
15f6e0da RR |
10282 | if ((SYMBOL_REF_WEAK (x) |
10283 | && !aarch64_symbol_binds_local_p (x)) | |
3ff5d1f0 TB |
10284 | || !IN_RANGE (INTVAL (offset), HOST_WIDE_INT_C (-4294967263), |
10285 | HOST_WIDE_INT_C (4294967264))) | |
17f4d4bf CSS |
10286 | return SYMBOL_FORCE_TO_MEM; |
10287 | return SYMBOL_SMALL_ABSOLUTE; | |
43e9d192 | 10288 | |
17f4d4bf | 10289 | case AARCH64_CMODEL_TINY_PIC: |
38e6c9a6 | 10290 | if (!aarch64_symbol_binds_local_p (x)) |
87dd8ab0 | 10291 | return SYMBOL_TINY_GOT; |
38e6c9a6 MS |
10292 | return SYMBOL_TINY_ABSOLUTE; |
10293 | ||
1b1e81f8 | 10294 | case AARCH64_CMODEL_SMALL_SPIC: |
17f4d4bf CSS |
10295 | case AARCH64_CMODEL_SMALL_PIC: |
10296 | if (!aarch64_symbol_binds_local_p (x)) | |
1b1e81f8 JW |
10297 | return (aarch64_cmodel == AARCH64_CMODEL_SMALL_SPIC |
10298 | ? SYMBOL_SMALL_GOT_28K : SYMBOL_SMALL_GOT_4G); | |
17f4d4bf | 10299 | return SYMBOL_SMALL_ABSOLUTE; |
43e9d192 | 10300 | |
9ee6540a WD |
10301 | case AARCH64_CMODEL_LARGE: |
10302 | /* This is alright even in PIC code as the constant | |
10303 | pool reference is always PC relative and within | |
10304 | the same translation unit. */ | |
d47d34bb | 10305 | if (!aarch64_pcrelative_literal_loads && CONSTANT_POOL_ADDRESS_P (x)) |
9ee6540a WD |
10306 | return SYMBOL_SMALL_ABSOLUTE; |
10307 | else | |
10308 | return SYMBOL_FORCE_TO_MEM; | |
10309 | ||
17f4d4bf CSS |
10310 | default: |
10311 | gcc_unreachable (); | |
10312 | } | |
43e9d192 | 10313 | } |
17f4d4bf | 10314 | |
43e9d192 IB |
10315 | /* By default push everything into the constant pool. */ |
10316 | return SYMBOL_FORCE_TO_MEM; | |
10317 | } | |
10318 | ||
43e9d192 IB |
10319 | bool |
10320 | aarch64_constant_address_p (rtx x) | |
10321 | { | |
10322 | return (CONSTANT_P (x) && memory_address_p (DImode, x)); | |
10323 | } | |
10324 | ||
10325 | bool | |
10326 | aarch64_legitimate_pic_operand_p (rtx x) | |
10327 | { | |
10328 | if (GET_CODE (x) == SYMBOL_REF | |
10329 | || (GET_CODE (x) == CONST | |
10330 | && GET_CODE (XEXP (x, 0)) == PLUS | |
10331 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF)) | |
10332 | return false; | |
10333 | ||
10334 | return true; | |
10335 | } | |
10336 | ||
3520f7cc JG |
10337 | /* Return true if X holds either a quarter-precision or |
10338 | floating-point +0.0 constant. */ | |
10339 | static bool | |
a2170965 | 10340 | aarch64_valid_floating_const (rtx x) |
3520f7cc JG |
10341 | { |
10342 | if (!CONST_DOUBLE_P (x)) | |
10343 | return false; | |
10344 | ||
a2170965 TC |
10345 | /* This call determines which constants can be used in mov<mode> |
10346 | as integer moves instead of constant loads. */ | |
10347 | if (aarch64_float_const_rtx_p (x)) | |
6a0f8c01 JW |
10348 | return true; |
10349 | ||
3520f7cc JG |
10350 | return aarch64_float_const_representable_p (x); |
10351 | } | |
10352 | ||
43e9d192 | 10353 | static bool |
ef4bddc2 | 10354 | aarch64_legitimate_constant_p (machine_mode mode, rtx x) |
43e9d192 IB |
10355 | { |
10356 | /* Do not allow vector struct mode constants. We could support | |
10357 | 0 and -1 easily, but they need support in aarch64-simd.md. */ | |
10358 | if (TARGET_SIMD && aarch64_vect_struct_mode_p (mode)) | |
10359 | return false; | |
10360 | ||
a2170965 TC |
10361 | /* For these cases we never want to use a literal load. |
10362 | As such we have to prevent the compiler from forcing these | |
10363 | to memory. */ | |
43e9d192 | 10364 | if ((GET_CODE (x) == CONST_VECTOR |
48063b9d | 10365 | && aarch64_simd_valid_immediate (x, mode, false, NULL)) |
a2170965 TC |
10366 | || CONST_INT_P (x) |
10367 | || aarch64_valid_floating_const (x) | |
10368 | || aarch64_can_const_movi_rtx_p (x, mode) | |
10369 | || aarch64_float_const_rtx_p (x)) | |
3520f7cc | 10370 | return !targetm.cannot_force_const_mem (mode, x); |
43e9d192 IB |
10371 | |
10372 | if (GET_CODE (x) == HIGH | |
10373 | && aarch64_valid_symref (XEXP (x, 0), GET_MODE (XEXP (x, 0)))) | |
10374 | return true; | |
10375 | ||
f28e54bd WD |
10376 | /* Treat symbols as constants. Avoid TLS symbols as they are complex, |
10377 | so spilling them is better than rematerialization. */ | |
10378 | if (SYMBOL_REF_P (x) && !SYMBOL_REF_TLS_MODEL (x)) | |
10379 | return true; | |
10380 | ||
43e9d192 IB |
10381 | return aarch64_constant_address_p (x); |
10382 | } | |
10383 | ||
a5bc806c | 10384 | rtx |
43e9d192 IB |
10385 | aarch64_load_tp (rtx target) |
10386 | { | |
10387 | if (!target | |
10388 | || GET_MODE (target) != Pmode | |
10389 | || !register_operand (target, Pmode)) | |
10390 | target = gen_reg_rtx (Pmode); | |
10391 | ||
10392 | /* Can return in any reg. */ | |
10393 | emit_insn (gen_aarch64_load_tp_hard (target)); | |
10394 | return target; | |
10395 | } | |
10396 | ||
43e9d192 IB |
10397 | /* On AAPCS systems, this is the "struct __va_list". */ |
10398 | static GTY(()) tree va_list_type; | |
10399 | ||
10400 | /* Implement TARGET_BUILD_BUILTIN_VA_LIST. | |
10401 | Return the type to use as __builtin_va_list. | |
10402 | ||
10403 | AAPCS64 \S 7.1.4 requires that va_list be a typedef for a type defined as: | |
10404 | ||
10405 | struct __va_list | |
10406 | { | |
10407 | void *__stack; | |
10408 | void *__gr_top; | |
10409 | void *__vr_top; | |
10410 | int __gr_offs; | |
10411 | int __vr_offs; | |
10412 | }; */ | |
10413 | ||
10414 | static tree | |
10415 | aarch64_build_builtin_va_list (void) | |
10416 | { | |
10417 | tree va_list_name; | |
10418 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
10419 | ||
10420 | /* Create the type. */ | |
10421 | va_list_type = lang_hooks.types.make_type (RECORD_TYPE); | |
10422 | /* Give it the required name. */ | |
10423 | va_list_name = build_decl (BUILTINS_LOCATION, | |
10424 | TYPE_DECL, | |
10425 | get_identifier ("__va_list"), | |
10426 | va_list_type); | |
10427 | DECL_ARTIFICIAL (va_list_name) = 1; | |
10428 | TYPE_NAME (va_list_type) = va_list_name; | |
665c56c6 | 10429 | TYPE_STUB_DECL (va_list_type) = va_list_name; |
43e9d192 IB |
10430 | |
10431 | /* Create the fields. */ | |
10432 | f_stack = build_decl (BUILTINS_LOCATION, | |
10433 | FIELD_DECL, get_identifier ("__stack"), | |
10434 | ptr_type_node); | |
10435 | f_grtop = build_decl (BUILTINS_LOCATION, | |
10436 | FIELD_DECL, get_identifier ("__gr_top"), | |
10437 | ptr_type_node); | |
10438 | f_vrtop = build_decl (BUILTINS_LOCATION, | |
10439 | FIELD_DECL, get_identifier ("__vr_top"), | |
10440 | ptr_type_node); | |
10441 | f_groff = build_decl (BUILTINS_LOCATION, | |
10442 | FIELD_DECL, get_identifier ("__gr_offs"), | |
10443 | integer_type_node); | |
10444 | f_vroff = build_decl (BUILTINS_LOCATION, | |
10445 | FIELD_DECL, get_identifier ("__vr_offs"), | |
10446 | integer_type_node); | |
10447 | ||
88e3bdd1 | 10448 | /* Tell tree-stdarg pass about our internal offset fields. |
3fd6b9cc JW |
10449 | NOTE: va_list_gpr/fpr_counter_field are only used for tree comparision |
10450 | purpose to identify whether the code is updating va_list internal | |
10451 | offset fields through irregular way. */ | |
10452 | va_list_gpr_counter_field = f_groff; | |
10453 | va_list_fpr_counter_field = f_vroff; | |
10454 | ||
43e9d192 IB |
10455 | DECL_ARTIFICIAL (f_stack) = 1; |
10456 | DECL_ARTIFICIAL (f_grtop) = 1; | |
10457 | DECL_ARTIFICIAL (f_vrtop) = 1; | |
10458 | DECL_ARTIFICIAL (f_groff) = 1; | |
10459 | DECL_ARTIFICIAL (f_vroff) = 1; | |
10460 | ||
10461 | DECL_FIELD_CONTEXT (f_stack) = va_list_type; | |
10462 | DECL_FIELD_CONTEXT (f_grtop) = va_list_type; | |
10463 | DECL_FIELD_CONTEXT (f_vrtop) = va_list_type; | |
10464 | DECL_FIELD_CONTEXT (f_groff) = va_list_type; | |
10465 | DECL_FIELD_CONTEXT (f_vroff) = va_list_type; | |
10466 | ||
10467 | TYPE_FIELDS (va_list_type) = f_stack; | |
10468 | DECL_CHAIN (f_stack) = f_grtop; | |
10469 | DECL_CHAIN (f_grtop) = f_vrtop; | |
10470 | DECL_CHAIN (f_vrtop) = f_groff; | |
10471 | DECL_CHAIN (f_groff) = f_vroff; | |
10472 | ||
10473 | /* Compute its layout. */ | |
10474 | layout_type (va_list_type); | |
10475 | ||
10476 | return va_list_type; | |
10477 | } | |
10478 | ||
10479 | /* Implement TARGET_EXPAND_BUILTIN_VA_START. */ | |
10480 | static void | |
10481 | aarch64_expand_builtin_va_start (tree valist, rtx nextarg ATTRIBUTE_UNUSED) | |
10482 | { | |
10483 | const CUMULATIVE_ARGS *cum; | |
10484 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
10485 | tree stack, grtop, vrtop, groff, vroff; | |
10486 | tree t; | |
88e3bdd1 JW |
10487 | int gr_save_area_size = cfun->va_list_gpr_size; |
10488 | int vr_save_area_size = cfun->va_list_fpr_size; | |
43e9d192 IB |
10489 | int vr_offset; |
10490 | ||
10491 | cum = &crtl->args.info; | |
88e3bdd1 JW |
10492 | if (cfun->va_list_gpr_size) |
10493 | gr_save_area_size = MIN ((NUM_ARG_REGS - cum->aapcs_ncrn) * UNITS_PER_WORD, | |
10494 | cfun->va_list_gpr_size); | |
10495 | if (cfun->va_list_fpr_size) | |
10496 | vr_save_area_size = MIN ((NUM_FP_ARG_REGS - cum->aapcs_nvrn) | |
10497 | * UNITS_PER_VREG, cfun->va_list_fpr_size); | |
43e9d192 | 10498 | |
d5726973 | 10499 | if (!TARGET_FLOAT) |
43e9d192 | 10500 | { |
261fb553 | 10501 | gcc_assert (cum->aapcs_nvrn == 0); |
43e9d192 IB |
10502 | vr_save_area_size = 0; |
10503 | } | |
10504 | ||
10505 | f_stack = TYPE_FIELDS (va_list_type_node); | |
10506 | f_grtop = DECL_CHAIN (f_stack); | |
10507 | f_vrtop = DECL_CHAIN (f_grtop); | |
10508 | f_groff = DECL_CHAIN (f_vrtop); | |
10509 | f_vroff = DECL_CHAIN (f_groff); | |
10510 | ||
10511 | stack = build3 (COMPONENT_REF, TREE_TYPE (f_stack), valist, f_stack, | |
10512 | NULL_TREE); | |
10513 | grtop = build3 (COMPONENT_REF, TREE_TYPE (f_grtop), valist, f_grtop, | |
10514 | NULL_TREE); | |
10515 | vrtop = build3 (COMPONENT_REF, TREE_TYPE (f_vrtop), valist, f_vrtop, | |
10516 | NULL_TREE); | |
10517 | groff = build3 (COMPONENT_REF, TREE_TYPE (f_groff), valist, f_groff, | |
10518 | NULL_TREE); | |
10519 | vroff = build3 (COMPONENT_REF, TREE_TYPE (f_vroff), valist, f_vroff, | |
10520 | NULL_TREE); | |
10521 | ||
10522 | /* Emit code to initialize STACK, which points to the next varargs stack | |
10523 | argument. CUM->AAPCS_STACK_SIZE gives the number of stack words used | |
10524 | by named arguments. STACK is 8-byte aligned. */ | |
10525 | t = make_tree (TREE_TYPE (stack), virtual_incoming_args_rtx); | |
10526 | if (cum->aapcs_stack_size > 0) | |
10527 | t = fold_build_pointer_plus_hwi (t, cum->aapcs_stack_size * UNITS_PER_WORD); | |
10528 | t = build2 (MODIFY_EXPR, TREE_TYPE (stack), stack, t); | |
10529 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
10530 | ||
10531 | /* Emit code to initialize GRTOP, the top of the GR save area. | |
10532 | virtual_incoming_args_rtx should have been 16 byte aligned. */ | |
10533 | t = make_tree (TREE_TYPE (grtop), virtual_incoming_args_rtx); | |
10534 | t = build2 (MODIFY_EXPR, TREE_TYPE (grtop), grtop, t); | |
10535 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
10536 | ||
10537 | /* Emit code to initialize VRTOP, the top of the VR save area. | |
10538 | This address is gr_save_area_bytes below GRTOP, rounded | |
10539 | down to the next 16-byte boundary. */ | |
10540 | t = make_tree (TREE_TYPE (vrtop), virtual_incoming_args_rtx); | |
4f59f9f2 UB |
10541 | vr_offset = ROUND_UP (gr_save_area_size, |
10542 | STACK_BOUNDARY / BITS_PER_UNIT); | |
43e9d192 IB |
10543 | |
10544 | if (vr_offset) | |
10545 | t = fold_build_pointer_plus_hwi (t, -vr_offset); | |
10546 | t = build2 (MODIFY_EXPR, TREE_TYPE (vrtop), vrtop, t); | |
10547 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
10548 | ||
10549 | /* Emit code to initialize GROFF, the offset from GRTOP of the | |
10550 | next GPR argument. */ | |
10551 | t = build2 (MODIFY_EXPR, TREE_TYPE (groff), groff, | |
10552 | build_int_cst (TREE_TYPE (groff), -gr_save_area_size)); | |
10553 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
10554 | ||
10555 | /* Likewise emit code to initialize VROFF, the offset from FTOP | |
10556 | of the next VR argument. */ | |
10557 | t = build2 (MODIFY_EXPR, TREE_TYPE (vroff), vroff, | |
10558 | build_int_cst (TREE_TYPE (vroff), -vr_save_area_size)); | |
10559 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
10560 | } | |
10561 | ||
10562 | /* Implement TARGET_GIMPLIFY_VA_ARG_EXPR. */ | |
10563 | ||
10564 | static tree | |
10565 | aarch64_gimplify_va_arg_expr (tree valist, tree type, gimple_seq *pre_p, | |
10566 | gimple_seq *post_p ATTRIBUTE_UNUSED) | |
10567 | { | |
10568 | tree addr; | |
10569 | bool indirect_p; | |
10570 | bool is_ha; /* is HFA or HVA. */ | |
10571 | bool dw_align; /* double-word align. */ | |
ef4bddc2 | 10572 | machine_mode ag_mode = VOIDmode; |
43e9d192 | 10573 | int nregs; |
ef4bddc2 | 10574 | machine_mode mode; |
43e9d192 IB |
10575 | |
10576 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
10577 | tree stack, f_top, f_off, off, arg, roundup, on_stack; | |
10578 | HOST_WIDE_INT size, rsize, adjust, align; | |
10579 | tree t, u, cond1, cond2; | |
10580 | ||
10581 | indirect_p = pass_by_reference (NULL, TYPE_MODE (type), type, false); | |
10582 | if (indirect_p) | |
10583 | type = build_pointer_type (type); | |
10584 | ||
10585 | mode = TYPE_MODE (type); | |
10586 | ||
10587 | f_stack = TYPE_FIELDS (va_list_type_node); | |
10588 | f_grtop = DECL_CHAIN (f_stack); | |
10589 | f_vrtop = DECL_CHAIN (f_grtop); | |
10590 | f_groff = DECL_CHAIN (f_vrtop); | |
10591 | f_vroff = DECL_CHAIN (f_groff); | |
10592 | ||
10593 | stack = build3 (COMPONENT_REF, TREE_TYPE (f_stack), unshare_expr (valist), | |
10594 | f_stack, NULL_TREE); | |
10595 | size = int_size_in_bytes (type); | |
985b8393 | 10596 | align = aarch64_function_arg_alignment (mode, type) / BITS_PER_UNIT; |
43e9d192 IB |
10597 | |
10598 | dw_align = false; | |
10599 | adjust = 0; | |
10600 | if (aarch64_vfp_is_call_or_return_candidate (mode, | |
10601 | type, | |
10602 | &ag_mode, | |
10603 | &nregs, | |
10604 | &is_ha)) | |
10605 | { | |
10606 | /* TYPE passed in fp/simd registers. */ | |
d5726973 | 10607 | if (!TARGET_FLOAT) |
261fb553 | 10608 | aarch64_err_no_fpadvsimd (mode, "varargs"); |
43e9d192 IB |
10609 | |
10610 | f_top = build3 (COMPONENT_REF, TREE_TYPE (f_vrtop), | |
10611 | unshare_expr (valist), f_vrtop, NULL_TREE); | |
10612 | f_off = build3 (COMPONENT_REF, TREE_TYPE (f_vroff), | |
10613 | unshare_expr (valist), f_vroff, NULL_TREE); | |
10614 | ||
10615 | rsize = nregs * UNITS_PER_VREG; | |
10616 | ||
10617 | if (is_ha) | |
10618 | { | |
10619 | if (BYTES_BIG_ENDIAN && GET_MODE_SIZE (ag_mode) < UNITS_PER_VREG) | |
10620 | adjust = UNITS_PER_VREG - GET_MODE_SIZE (ag_mode); | |
10621 | } | |
10622 | else if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
10623 | && size < UNITS_PER_VREG) | |
10624 | { | |
10625 | adjust = UNITS_PER_VREG - size; | |
10626 | } | |
10627 | } | |
10628 | else | |
10629 | { | |
10630 | /* TYPE passed in general registers. */ | |
10631 | f_top = build3 (COMPONENT_REF, TREE_TYPE (f_grtop), | |
10632 | unshare_expr (valist), f_grtop, NULL_TREE); | |
10633 | f_off = build3 (COMPONENT_REF, TREE_TYPE (f_groff), | |
10634 | unshare_expr (valist), f_groff, NULL_TREE); | |
4f59f9f2 | 10635 | rsize = ROUND_UP (size, UNITS_PER_WORD); |
43e9d192 IB |
10636 | nregs = rsize / UNITS_PER_WORD; |
10637 | ||
10638 | if (align > 8) | |
10639 | dw_align = true; | |
10640 | ||
10641 | if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
10642 | && size < UNITS_PER_WORD) | |
10643 | { | |
10644 | adjust = UNITS_PER_WORD - size; | |
10645 | } | |
10646 | } | |
10647 | ||
10648 | /* Get a local temporary for the field value. */ | |
10649 | off = get_initialized_tmp_var (f_off, pre_p, NULL); | |
10650 | ||
10651 | /* Emit code to branch if off >= 0. */ | |
10652 | t = build2 (GE_EXPR, boolean_type_node, off, | |
10653 | build_int_cst (TREE_TYPE (off), 0)); | |
10654 | cond1 = build3 (COND_EXPR, ptr_type_node, t, NULL_TREE, NULL_TREE); | |
10655 | ||
10656 | if (dw_align) | |
10657 | { | |
10658 | /* Emit: offs = (offs + 15) & -16. */ | |
10659 | t = build2 (PLUS_EXPR, TREE_TYPE (off), off, | |
10660 | build_int_cst (TREE_TYPE (off), 15)); | |
10661 | t = build2 (BIT_AND_EXPR, TREE_TYPE (off), t, | |
10662 | build_int_cst (TREE_TYPE (off), -16)); | |
10663 | roundup = build2 (MODIFY_EXPR, TREE_TYPE (off), off, t); | |
10664 | } | |
10665 | else | |
10666 | roundup = NULL; | |
10667 | ||
10668 | /* Update ap.__[g|v]r_offs */ | |
10669 | t = build2 (PLUS_EXPR, TREE_TYPE (off), off, | |
10670 | build_int_cst (TREE_TYPE (off), rsize)); | |
10671 | t = build2 (MODIFY_EXPR, TREE_TYPE (f_off), unshare_expr (f_off), t); | |
10672 | ||
10673 | /* String up. */ | |
10674 | if (roundup) | |
10675 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), roundup, t); | |
10676 | ||
10677 | /* [cond2] if (ap.__[g|v]r_offs > 0) */ | |
10678 | u = build2 (GT_EXPR, boolean_type_node, unshare_expr (f_off), | |
10679 | build_int_cst (TREE_TYPE (f_off), 0)); | |
10680 | cond2 = build3 (COND_EXPR, ptr_type_node, u, NULL_TREE, NULL_TREE); | |
10681 | ||
10682 | /* String up: make sure the assignment happens before the use. */ | |
10683 | t = build2 (COMPOUND_EXPR, TREE_TYPE (cond2), t, cond2); | |
10684 | COND_EXPR_ELSE (cond1) = t; | |
10685 | ||
10686 | /* Prepare the trees handling the argument that is passed on the stack; | |
10687 | the top level node will store in ON_STACK. */ | |
10688 | arg = get_initialized_tmp_var (stack, pre_p, NULL); | |
10689 | if (align > 8) | |
10690 | { | |
10691 | /* if (alignof(type) > 8) (arg = arg + 15) & -16; */ | |
10692 | t = fold_convert (intDI_type_node, arg); | |
10693 | t = build2 (PLUS_EXPR, TREE_TYPE (t), t, | |
10694 | build_int_cst (TREE_TYPE (t), 15)); | |
10695 | t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t, | |
10696 | build_int_cst (TREE_TYPE (t), -16)); | |
10697 | t = fold_convert (TREE_TYPE (arg), t); | |
10698 | roundup = build2 (MODIFY_EXPR, TREE_TYPE (arg), arg, t); | |
10699 | } | |
10700 | else | |
10701 | roundup = NULL; | |
10702 | /* Advance ap.__stack */ | |
10703 | t = fold_convert (intDI_type_node, arg); | |
10704 | t = build2 (PLUS_EXPR, TREE_TYPE (t), t, | |
10705 | build_int_cst (TREE_TYPE (t), size + 7)); | |
10706 | t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t, | |
10707 | build_int_cst (TREE_TYPE (t), -8)); | |
10708 | t = fold_convert (TREE_TYPE (arg), t); | |
10709 | t = build2 (MODIFY_EXPR, TREE_TYPE (stack), unshare_expr (stack), t); | |
10710 | /* String up roundup and advance. */ | |
10711 | if (roundup) | |
10712 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), roundup, t); | |
10713 | /* String up with arg */ | |
10714 | on_stack = build2 (COMPOUND_EXPR, TREE_TYPE (arg), t, arg); | |
10715 | /* Big-endianness related address adjustment. */ | |
10716 | if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
10717 | && size < UNITS_PER_WORD) | |
10718 | { | |
10719 | t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (arg), arg, | |
10720 | size_int (UNITS_PER_WORD - size)); | |
10721 | on_stack = build2 (COMPOUND_EXPR, TREE_TYPE (arg), on_stack, t); | |
10722 | } | |
10723 | ||
10724 | COND_EXPR_THEN (cond1) = unshare_expr (on_stack); | |
10725 | COND_EXPR_THEN (cond2) = unshare_expr (on_stack); | |
10726 | ||
10727 | /* Adjustment to OFFSET in the case of BIG_ENDIAN. */ | |
10728 | t = off; | |
10729 | if (adjust) | |
10730 | t = build2 (PREINCREMENT_EXPR, TREE_TYPE (off), off, | |
10731 | build_int_cst (TREE_TYPE (off), adjust)); | |
10732 | ||
10733 | t = fold_convert (sizetype, t); | |
10734 | t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (f_top), f_top, t); | |
10735 | ||
10736 | if (is_ha) | |
10737 | { | |
10738 | /* type ha; // treat as "struct {ftype field[n];}" | |
10739 | ... [computing offs] | |
10740 | for (i = 0; i <nregs; ++i, offs += 16) | |
10741 | ha.field[i] = *((ftype *)(ap.__vr_top + offs)); | |
10742 | return ha; */ | |
10743 | int i; | |
10744 | tree tmp_ha, field_t, field_ptr_t; | |
10745 | ||
10746 | /* Declare a local variable. */ | |
10747 | tmp_ha = create_tmp_var_raw (type, "ha"); | |
10748 | gimple_add_tmp_var (tmp_ha); | |
10749 | ||
10750 | /* Establish the base type. */ | |
10751 | switch (ag_mode) | |
10752 | { | |
4e10a5a7 | 10753 | case E_SFmode: |
43e9d192 IB |
10754 | field_t = float_type_node; |
10755 | field_ptr_t = float_ptr_type_node; | |
10756 | break; | |
4e10a5a7 | 10757 | case E_DFmode: |
43e9d192 IB |
10758 | field_t = double_type_node; |
10759 | field_ptr_t = double_ptr_type_node; | |
10760 | break; | |
4e10a5a7 | 10761 | case E_TFmode: |
43e9d192 IB |
10762 | field_t = long_double_type_node; |
10763 | field_ptr_t = long_double_ptr_type_node; | |
10764 | break; | |
4e10a5a7 | 10765 | case E_HFmode: |
1b62ed4f JG |
10766 | field_t = aarch64_fp16_type_node; |
10767 | field_ptr_t = aarch64_fp16_ptr_type_node; | |
43e9d192 | 10768 | break; |
4e10a5a7 RS |
10769 | case E_V2SImode: |
10770 | case E_V4SImode: | |
43e9d192 IB |
10771 | { |
10772 | tree innertype = make_signed_type (GET_MODE_PRECISION (SImode)); | |
10773 | field_t = build_vector_type_for_mode (innertype, ag_mode); | |
10774 | field_ptr_t = build_pointer_type (field_t); | |
10775 | } | |
10776 | break; | |
10777 | default: | |
10778 | gcc_assert (0); | |
10779 | } | |
10780 | ||
10781 | /* *(field_ptr_t)&ha = *((field_ptr_t)vr_saved_area */ | |
10782 | tmp_ha = build1 (ADDR_EXPR, field_ptr_t, tmp_ha); | |
10783 | addr = t; | |
10784 | t = fold_convert (field_ptr_t, addr); | |
10785 | t = build2 (MODIFY_EXPR, field_t, | |
10786 | build1 (INDIRECT_REF, field_t, tmp_ha), | |
10787 | build1 (INDIRECT_REF, field_t, t)); | |
10788 | ||
10789 | /* ha.field[i] = *((field_ptr_t)vr_saved_area + i) */ | |
10790 | for (i = 1; i < nregs; ++i) | |
10791 | { | |
10792 | addr = fold_build_pointer_plus_hwi (addr, UNITS_PER_VREG); | |
10793 | u = fold_convert (field_ptr_t, addr); | |
10794 | u = build2 (MODIFY_EXPR, field_t, | |
10795 | build2 (MEM_REF, field_t, tmp_ha, | |
10796 | build_int_cst (field_ptr_t, | |
10797 | (i * | |
10798 | int_size_in_bytes (field_t)))), | |
10799 | build1 (INDIRECT_REF, field_t, u)); | |
10800 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), t, u); | |
10801 | } | |
10802 | ||
10803 | u = fold_convert (TREE_TYPE (f_top), tmp_ha); | |
10804 | t = build2 (COMPOUND_EXPR, TREE_TYPE (f_top), t, u); | |
10805 | } | |
10806 | ||
10807 | COND_EXPR_ELSE (cond2) = t; | |
10808 | addr = fold_convert (build_pointer_type (type), cond1); | |
10809 | addr = build_va_arg_indirect_ref (addr); | |
10810 | ||
10811 | if (indirect_p) | |
10812 | addr = build_va_arg_indirect_ref (addr); | |
10813 | ||
10814 | return addr; | |
10815 | } | |
10816 | ||
10817 | /* Implement TARGET_SETUP_INCOMING_VARARGS. */ | |
10818 | ||
10819 | static void | |
ef4bddc2 | 10820 | aarch64_setup_incoming_varargs (cumulative_args_t cum_v, machine_mode mode, |
43e9d192 IB |
10821 | tree type, int *pretend_size ATTRIBUTE_UNUSED, |
10822 | int no_rtl) | |
10823 | { | |
10824 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
10825 | CUMULATIVE_ARGS local_cum; | |
88e3bdd1 JW |
10826 | int gr_saved = cfun->va_list_gpr_size; |
10827 | int vr_saved = cfun->va_list_fpr_size; | |
43e9d192 IB |
10828 | |
10829 | /* The caller has advanced CUM up to, but not beyond, the last named | |
10830 | argument. Advance a local copy of CUM past the last "real" named | |
10831 | argument, to find out how many registers are left over. */ | |
10832 | local_cum = *cum; | |
10833 | aarch64_function_arg_advance (pack_cumulative_args(&local_cum), mode, type, true); | |
10834 | ||
88e3bdd1 JW |
10835 | /* Found out how many registers we need to save. |
10836 | Honor tree-stdvar analysis results. */ | |
10837 | if (cfun->va_list_gpr_size) | |
10838 | gr_saved = MIN (NUM_ARG_REGS - local_cum.aapcs_ncrn, | |
10839 | cfun->va_list_gpr_size / UNITS_PER_WORD); | |
10840 | if (cfun->va_list_fpr_size) | |
10841 | vr_saved = MIN (NUM_FP_ARG_REGS - local_cum.aapcs_nvrn, | |
10842 | cfun->va_list_fpr_size / UNITS_PER_VREG); | |
43e9d192 | 10843 | |
d5726973 | 10844 | if (!TARGET_FLOAT) |
43e9d192 | 10845 | { |
261fb553 | 10846 | gcc_assert (local_cum.aapcs_nvrn == 0); |
43e9d192 IB |
10847 | vr_saved = 0; |
10848 | } | |
10849 | ||
10850 | if (!no_rtl) | |
10851 | { | |
10852 | if (gr_saved > 0) | |
10853 | { | |
10854 | rtx ptr, mem; | |
10855 | ||
10856 | /* virtual_incoming_args_rtx should have been 16-byte aligned. */ | |
10857 | ptr = plus_constant (Pmode, virtual_incoming_args_rtx, | |
10858 | - gr_saved * UNITS_PER_WORD); | |
10859 | mem = gen_frame_mem (BLKmode, ptr); | |
10860 | set_mem_alias_set (mem, get_varargs_alias_set ()); | |
10861 | ||
10862 | move_block_from_reg (local_cum.aapcs_ncrn + R0_REGNUM, | |
10863 | mem, gr_saved); | |
10864 | } | |
10865 | if (vr_saved > 0) | |
10866 | { | |
10867 | /* We can't use move_block_from_reg, because it will use | |
10868 | the wrong mode, storing D regs only. */ | |
ef4bddc2 | 10869 | machine_mode mode = TImode; |
88e3bdd1 | 10870 | int off, i, vr_start; |
43e9d192 IB |
10871 | |
10872 | /* Set OFF to the offset from virtual_incoming_args_rtx of | |
10873 | the first vector register. The VR save area lies below | |
10874 | the GR one, and is aligned to 16 bytes. */ | |
4f59f9f2 UB |
10875 | off = -ROUND_UP (gr_saved * UNITS_PER_WORD, |
10876 | STACK_BOUNDARY / BITS_PER_UNIT); | |
43e9d192 IB |
10877 | off -= vr_saved * UNITS_PER_VREG; |
10878 | ||
88e3bdd1 JW |
10879 | vr_start = V0_REGNUM + local_cum.aapcs_nvrn; |
10880 | for (i = 0; i < vr_saved; ++i) | |
43e9d192 IB |
10881 | { |
10882 | rtx ptr, mem; | |
10883 | ||
10884 | ptr = plus_constant (Pmode, virtual_incoming_args_rtx, off); | |
10885 | mem = gen_frame_mem (mode, ptr); | |
10886 | set_mem_alias_set (mem, get_varargs_alias_set ()); | |
88e3bdd1 | 10887 | aarch64_emit_move (mem, gen_rtx_REG (mode, vr_start + i)); |
43e9d192 IB |
10888 | off += UNITS_PER_VREG; |
10889 | } | |
10890 | } | |
10891 | } | |
10892 | ||
10893 | /* We don't save the size into *PRETEND_SIZE because we want to avoid | |
10894 | any complication of having crtl->args.pretend_args_size changed. */ | |
8799637a | 10895 | cfun->machine->frame.saved_varargs_size |
4f59f9f2 UB |
10896 | = (ROUND_UP (gr_saved * UNITS_PER_WORD, |
10897 | STACK_BOUNDARY / BITS_PER_UNIT) | |
43e9d192 IB |
10898 | + vr_saved * UNITS_PER_VREG); |
10899 | } | |
10900 | ||
10901 | static void | |
10902 | aarch64_conditional_register_usage (void) | |
10903 | { | |
10904 | int i; | |
10905 | if (!TARGET_FLOAT) | |
10906 | { | |
10907 | for (i = V0_REGNUM; i <= V31_REGNUM; i++) | |
10908 | { | |
10909 | fixed_regs[i] = 1; | |
10910 | call_used_regs[i] = 1; | |
10911 | } | |
10912 | } | |
10913 | } | |
10914 | ||
10915 | /* Walk down the type tree of TYPE counting consecutive base elements. | |
10916 | If *MODEP is VOIDmode, then set it to the first valid floating point | |
10917 | type. If a non-floating point type is found, or if a floating point | |
10918 | type that doesn't match a non-VOIDmode *MODEP is found, then return -1, | |
10919 | otherwise return the count in the sub-tree. */ | |
10920 | static int | |
ef4bddc2 | 10921 | aapcs_vfp_sub_candidate (const_tree type, machine_mode *modep) |
43e9d192 | 10922 | { |
ef4bddc2 | 10923 | machine_mode mode; |
43e9d192 IB |
10924 | HOST_WIDE_INT size; |
10925 | ||
10926 | switch (TREE_CODE (type)) | |
10927 | { | |
10928 | case REAL_TYPE: | |
10929 | mode = TYPE_MODE (type); | |
1b62ed4f JG |
10930 | if (mode != DFmode && mode != SFmode |
10931 | && mode != TFmode && mode != HFmode) | |
43e9d192 IB |
10932 | return -1; |
10933 | ||
10934 | if (*modep == VOIDmode) | |
10935 | *modep = mode; | |
10936 | ||
10937 | if (*modep == mode) | |
10938 | return 1; | |
10939 | ||
10940 | break; | |
10941 | ||
10942 | case COMPLEX_TYPE: | |
10943 | mode = TYPE_MODE (TREE_TYPE (type)); | |
1b62ed4f JG |
10944 | if (mode != DFmode && mode != SFmode |
10945 | && mode != TFmode && mode != HFmode) | |
43e9d192 IB |
10946 | return -1; |
10947 | ||
10948 | if (*modep == VOIDmode) | |
10949 | *modep = mode; | |
10950 | ||
10951 | if (*modep == mode) | |
10952 | return 2; | |
10953 | ||
10954 | break; | |
10955 | ||
10956 | case VECTOR_TYPE: | |
10957 | /* Use V2SImode and V4SImode as representatives of all 64-bit | |
10958 | and 128-bit vector types. */ | |
10959 | size = int_size_in_bytes (type); | |
10960 | switch (size) | |
10961 | { | |
10962 | case 8: | |
10963 | mode = V2SImode; | |
10964 | break; | |
10965 | case 16: | |
10966 | mode = V4SImode; | |
10967 | break; | |
10968 | default: | |
10969 | return -1; | |
10970 | } | |
10971 | ||
10972 | if (*modep == VOIDmode) | |
10973 | *modep = mode; | |
10974 | ||
10975 | /* Vector modes are considered to be opaque: two vectors are | |
10976 | equivalent for the purposes of being homogeneous aggregates | |
10977 | if they are the same size. */ | |
10978 | if (*modep == mode) | |
10979 | return 1; | |
10980 | ||
10981 | break; | |
10982 | ||
10983 | case ARRAY_TYPE: | |
10984 | { | |
10985 | int count; | |
10986 | tree index = TYPE_DOMAIN (type); | |
10987 | ||
807e902e KZ |
10988 | /* Can't handle incomplete types nor sizes that are not |
10989 | fixed. */ | |
10990 | if (!COMPLETE_TYPE_P (type) | |
10991 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
43e9d192 IB |
10992 | return -1; |
10993 | ||
10994 | count = aapcs_vfp_sub_candidate (TREE_TYPE (type), modep); | |
10995 | if (count == -1 | |
10996 | || !index | |
10997 | || !TYPE_MAX_VALUE (index) | |
cc269bb6 | 10998 | || !tree_fits_uhwi_p (TYPE_MAX_VALUE (index)) |
43e9d192 | 10999 | || !TYPE_MIN_VALUE (index) |
cc269bb6 | 11000 | || !tree_fits_uhwi_p (TYPE_MIN_VALUE (index)) |
43e9d192 IB |
11001 | || count < 0) |
11002 | return -1; | |
11003 | ||
ae7e9ddd RS |
11004 | count *= (1 + tree_to_uhwi (TYPE_MAX_VALUE (index)) |
11005 | - tree_to_uhwi (TYPE_MIN_VALUE (index))); | |
43e9d192 IB |
11006 | |
11007 | /* There must be no padding. */ | |
807e902e | 11008 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) |
43e9d192 IB |
11009 | return -1; |
11010 | ||
11011 | return count; | |
11012 | } | |
11013 | ||
11014 | case RECORD_TYPE: | |
11015 | { | |
11016 | int count = 0; | |
11017 | int sub_count; | |
11018 | tree field; | |
11019 | ||
807e902e KZ |
11020 | /* Can't handle incomplete types nor sizes that are not |
11021 | fixed. */ | |
11022 | if (!COMPLETE_TYPE_P (type) | |
11023 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
43e9d192 IB |
11024 | return -1; |
11025 | ||
11026 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
11027 | { | |
11028 | if (TREE_CODE (field) != FIELD_DECL) | |
11029 | continue; | |
11030 | ||
11031 | sub_count = aapcs_vfp_sub_candidate (TREE_TYPE (field), modep); | |
11032 | if (sub_count < 0) | |
11033 | return -1; | |
11034 | count += sub_count; | |
11035 | } | |
11036 | ||
11037 | /* There must be no padding. */ | |
807e902e | 11038 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) |
43e9d192 IB |
11039 | return -1; |
11040 | ||
11041 | return count; | |
11042 | } | |
11043 | ||
11044 | case UNION_TYPE: | |
11045 | case QUAL_UNION_TYPE: | |
11046 | { | |
11047 | /* These aren't very interesting except in a degenerate case. */ | |
11048 | int count = 0; | |
11049 | int sub_count; | |
11050 | tree field; | |
11051 | ||
807e902e KZ |
11052 | /* Can't handle incomplete types nor sizes that are not |
11053 | fixed. */ | |
11054 | if (!COMPLETE_TYPE_P (type) | |
11055 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
43e9d192 IB |
11056 | return -1; |
11057 | ||
11058 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
11059 | { | |
11060 | if (TREE_CODE (field) != FIELD_DECL) | |
11061 | continue; | |
11062 | ||
11063 | sub_count = aapcs_vfp_sub_candidate (TREE_TYPE (field), modep); | |
11064 | if (sub_count < 0) | |
11065 | return -1; | |
11066 | count = count > sub_count ? count : sub_count; | |
11067 | } | |
11068 | ||
11069 | /* There must be no padding. */ | |
807e902e | 11070 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) |
43e9d192 IB |
11071 | return -1; |
11072 | ||
11073 | return count; | |
11074 | } | |
11075 | ||
11076 | default: | |
11077 | break; | |
11078 | } | |
11079 | ||
11080 | return -1; | |
11081 | } | |
11082 | ||
b6ec6215 KT |
11083 | /* Return TRUE if the type, as described by TYPE and MODE, is a short vector |
11084 | type as described in AAPCS64 \S 4.1.2. | |
11085 | ||
11086 | See the comment above aarch64_composite_type_p for the notes on MODE. */ | |
11087 | ||
11088 | static bool | |
11089 | aarch64_short_vector_p (const_tree type, | |
11090 | machine_mode mode) | |
11091 | { | |
11092 | HOST_WIDE_INT size = -1; | |
11093 | ||
11094 | if (type && TREE_CODE (type) == VECTOR_TYPE) | |
11095 | size = int_size_in_bytes (type); | |
11096 | else if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT | |
11097 | || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) | |
11098 | size = GET_MODE_SIZE (mode); | |
11099 | ||
11100 | return (size == 8 || size == 16); | |
11101 | } | |
11102 | ||
43e9d192 IB |
11103 | /* Return TRUE if the type, as described by TYPE and MODE, is a composite |
11104 | type as described in AAPCS64 \S 4.3. This includes aggregate, union and | |
11105 | array types. The C99 floating-point complex types are also considered | |
11106 | as composite types, according to AAPCS64 \S 7.1.1. The complex integer | |
11107 | types, which are GCC extensions and out of the scope of AAPCS64, are | |
11108 | treated as composite types here as well. | |
11109 | ||
11110 | Note that MODE itself is not sufficient in determining whether a type | |
11111 | is such a composite type or not. This is because | |
11112 | stor-layout.c:compute_record_mode may have already changed the MODE | |
11113 | (BLKmode) of a RECORD_TYPE TYPE to some other mode. For example, a | |
11114 | structure with only one field may have its MODE set to the mode of the | |
11115 | field. Also an integer mode whose size matches the size of the | |
11116 | RECORD_TYPE type may be used to substitute the original mode | |
11117 | (i.e. BLKmode) in certain circumstances. In other words, MODE cannot be | |
11118 | solely relied on. */ | |
11119 | ||
11120 | static bool | |
11121 | aarch64_composite_type_p (const_tree type, | |
ef4bddc2 | 11122 | machine_mode mode) |
43e9d192 | 11123 | { |
b6ec6215 KT |
11124 | if (aarch64_short_vector_p (type, mode)) |
11125 | return false; | |
11126 | ||
43e9d192 IB |
11127 | if (type && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == COMPLEX_TYPE)) |
11128 | return true; | |
11129 | ||
11130 | if (mode == BLKmode | |
11131 | || GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT | |
11132 | || GET_MODE_CLASS (mode) == MODE_COMPLEX_INT) | |
11133 | return true; | |
11134 | ||
11135 | return false; | |
11136 | } | |
11137 | ||
43e9d192 IB |
11138 | /* Return TRUE if an argument, whose type is described by TYPE and MODE, |
11139 | shall be passed or returned in simd/fp register(s) (providing these | |
11140 | parameter passing registers are available). | |
11141 | ||
11142 | Upon successful return, *COUNT returns the number of needed registers, | |
11143 | *BASE_MODE returns the mode of the individual register and when IS_HAF | |
11144 | is not NULL, *IS_HA indicates whether or not the argument is a homogeneous | |
11145 | floating-point aggregate or a homogeneous short-vector aggregate. */ | |
11146 | ||
11147 | static bool | |
ef4bddc2 | 11148 | aarch64_vfp_is_call_or_return_candidate (machine_mode mode, |
43e9d192 | 11149 | const_tree type, |
ef4bddc2 | 11150 | machine_mode *base_mode, |
43e9d192 IB |
11151 | int *count, |
11152 | bool *is_ha) | |
11153 | { | |
ef4bddc2 | 11154 | machine_mode new_mode = VOIDmode; |
43e9d192 IB |
11155 | bool composite_p = aarch64_composite_type_p (type, mode); |
11156 | ||
11157 | if (is_ha != NULL) *is_ha = false; | |
11158 | ||
11159 | if ((!composite_p && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
11160 | || aarch64_short_vector_p (type, mode)) | |
11161 | { | |
11162 | *count = 1; | |
11163 | new_mode = mode; | |
11164 | } | |
11165 | else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT) | |
11166 | { | |
11167 | if (is_ha != NULL) *is_ha = true; | |
11168 | *count = 2; | |
11169 | new_mode = GET_MODE_INNER (mode); | |
11170 | } | |
11171 | else if (type && composite_p) | |
11172 | { | |
11173 | int ag_count = aapcs_vfp_sub_candidate (type, &new_mode); | |
11174 | ||
11175 | if (ag_count > 0 && ag_count <= HA_MAX_NUM_FLDS) | |
11176 | { | |
11177 | if (is_ha != NULL) *is_ha = true; | |
11178 | *count = ag_count; | |
11179 | } | |
11180 | else | |
11181 | return false; | |
11182 | } | |
11183 | else | |
11184 | return false; | |
11185 | ||
11186 | *base_mode = new_mode; | |
11187 | return true; | |
11188 | } | |
11189 | ||
11190 | /* Implement TARGET_STRUCT_VALUE_RTX. */ | |
11191 | ||
11192 | static rtx | |
11193 | aarch64_struct_value_rtx (tree fndecl ATTRIBUTE_UNUSED, | |
11194 | int incoming ATTRIBUTE_UNUSED) | |
11195 | { | |
11196 | return gen_rtx_REG (Pmode, AARCH64_STRUCT_VALUE_REGNUM); | |
11197 | } | |
11198 | ||
11199 | /* Implements target hook vector_mode_supported_p. */ | |
11200 | static bool | |
ef4bddc2 | 11201 | aarch64_vector_mode_supported_p (machine_mode mode) |
43e9d192 IB |
11202 | { |
11203 | if (TARGET_SIMD | |
11204 | && (mode == V4SImode || mode == V8HImode | |
11205 | || mode == V16QImode || mode == V2DImode | |
11206 | || mode == V2SImode || mode == V4HImode | |
11207 | || mode == V8QImode || mode == V2SFmode | |
ad7d90cc | 11208 | || mode == V4SFmode || mode == V2DFmode |
71a11456 | 11209 | || mode == V4HFmode || mode == V8HFmode |
ad7d90cc | 11210 | || mode == V1DFmode)) |
43e9d192 IB |
11211 | return true; |
11212 | ||
11213 | return false; | |
11214 | } | |
11215 | ||
b7342d25 IB |
11216 | /* Return appropriate SIMD container |
11217 | for MODE within a vector of WIDTH bits. */ | |
ef4bddc2 RS |
11218 | static machine_mode |
11219 | aarch64_simd_container_mode (machine_mode mode, unsigned width) | |
43e9d192 | 11220 | { |
b7342d25 | 11221 | gcc_assert (width == 64 || width == 128); |
43e9d192 | 11222 | if (TARGET_SIMD) |
b7342d25 IB |
11223 | { |
11224 | if (width == 128) | |
11225 | switch (mode) | |
11226 | { | |
4e10a5a7 | 11227 | case E_DFmode: |
b7342d25 | 11228 | return V2DFmode; |
4e10a5a7 | 11229 | case E_SFmode: |
b7342d25 | 11230 | return V4SFmode; |
4e10a5a7 | 11231 | case E_HFmode: |
b719f884 | 11232 | return V8HFmode; |
4e10a5a7 | 11233 | case E_SImode: |
b7342d25 | 11234 | return V4SImode; |
4e10a5a7 | 11235 | case E_HImode: |
b7342d25 | 11236 | return V8HImode; |
4e10a5a7 | 11237 | case E_QImode: |
b7342d25 | 11238 | return V16QImode; |
4e10a5a7 | 11239 | case E_DImode: |
b7342d25 IB |
11240 | return V2DImode; |
11241 | default: | |
11242 | break; | |
11243 | } | |
11244 | else | |
11245 | switch (mode) | |
11246 | { | |
4e10a5a7 | 11247 | case E_SFmode: |
b7342d25 | 11248 | return V2SFmode; |
4e10a5a7 | 11249 | case E_HFmode: |
b719f884 | 11250 | return V4HFmode; |
4e10a5a7 | 11251 | case E_SImode: |
b7342d25 | 11252 | return V2SImode; |
4e10a5a7 | 11253 | case E_HImode: |
b7342d25 | 11254 | return V4HImode; |
4e10a5a7 | 11255 | case E_QImode: |
b7342d25 IB |
11256 | return V8QImode; |
11257 | default: | |
11258 | break; | |
11259 | } | |
11260 | } | |
43e9d192 IB |
11261 | return word_mode; |
11262 | } | |
11263 | ||
b7342d25 | 11264 | /* Return 128-bit container as the preferred SIMD mode for MODE. */ |
ef4bddc2 RS |
11265 | static machine_mode |
11266 | aarch64_preferred_simd_mode (machine_mode mode) | |
b7342d25 IB |
11267 | { |
11268 | return aarch64_simd_container_mode (mode, 128); | |
11269 | } | |
11270 | ||
3b357264 JG |
11271 | /* Return the bitmask of possible vector sizes for the vectorizer |
11272 | to iterate over. */ | |
11273 | static unsigned int | |
11274 | aarch64_autovectorize_vector_sizes (void) | |
11275 | { | |
11276 | return (16 | 8); | |
11277 | } | |
11278 | ||
ac2b960f YZ |
11279 | /* Implement TARGET_MANGLE_TYPE. */ |
11280 | ||
6f549691 | 11281 | static const char * |
ac2b960f YZ |
11282 | aarch64_mangle_type (const_tree type) |
11283 | { | |
11284 | /* The AArch64 ABI documents say that "__va_list" has to be | |
11285 | managled as if it is in the "std" namespace. */ | |
11286 | if (lang_hooks.types_compatible_p (CONST_CAST_TREE (type), va_list_type)) | |
11287 | return "St9__va_list"; | |
11288 | ||
c2ec330c AL |
11289 | /* Half-precision float. */ |
11290 | if (TREE_CODE (type) == REAL_TYPE && TYPE_PRECISION (type) == 16) | |
11291 | return "Dh"; | |
11292 | ||
f9d53c27 TB |
11293 | /* Mangle AArch64-specific internal types. TYPE_NAME is non-NULL_TREE for |
11294 | builtin types. */ | |
11295 | if (TYPE_NAME (type) != NULL) | |
11296 | return aarch64_mangle_builtin_type (type); | |
c6fc9e43 | 11297 | |
ac2b960f YZ |
11298 | /* Use the default mangling. */ |
11299 | return NULL; | |
11300 | } | |
11301 | ||
75cf1494 KT |
11302 | /* Find the first rtx_insn before insn that will generate an assembly |
11303 | instruction. */ | |
11304 | ||
11305 | static rtx_insn * | |
11306 | aarch64_prev_real_insn (rtx_insn *insn) | |
11307 | { | |
11308 | if (!insn) | |
11309 | return NULL; | |
11310 | ||
11311 | do | |
11312 | { | |
11313 | insn = prev_real_insn (insn); | |
11314 | } | |
11315 | while (insn && recog_memoized (insn) < 0); | |
11316 | ||
11317 | return insn; | |
11318 | } | |
11319 | ||
11320 | static bool | |
11321 | is_madd_op (enum attr_type t1) | |
11322 | { | |
11323 | unsigned int i; | |
11324 | /* A number of these may be AArch32 only. */ | |
11325 | enum attr_type mlatypes[] = { | |
11326 | TYPE_MLA, TYPE_MLAS, TYPE_SMLAD, TYPE_SMLADX, TYPE_SMLAL, TYPE_SMLALD, | |
11327 | TYPE_SMLALS, TYPE_SMLALXY, TYPE_SMLAWX, TYPE_SMLAWY, TYPE_SMLAXY, | |
11328 | TYPE_SMMLA, TYPE_UMLAL, TYPE_UMLALS,TYPE_SMLSD, TYPE_SMLSDX, TYPE_SMLSLD | |
11329 | }; | |
11330 | ||
11331 | for (i = 0; i < sizeof (mlatypes) / sizeof (enum attr_type); i++) | |
11332 | { | |
11333 | if (t1 == mlatypes[i]) | |
11334 | return true; | |
11335 | } | |
11336 | ||
11337 | return false; | |
11338 | } | |
11339 | ||
11340 | /* Check if there is a register dependency between a load and the insn | |
11341 | for which we hold recog_data. */ | |
11342 | ||
11343 | static bool | |
11344 | dep_between_memop_and_curr (rtx memop) | |
11345 | { | |
11346 | rtx load_reg; | |
11347 | int opno; | |
11348 | ||
8baff86e | 11349 | gcc_assert (GET_CODE (memop) == SET); |
75cf1494 KT |
11350 | |
11351 | if (!REG_P (SET_DEST (memop))) | |
11352 | return false; | |
11353 | ||
11354 | load_reg = SET_DEST (memop); | |
8baff86e | 11355 | for (opno = 1; opno < recog_data.n_operands; opno++) |
75cf1494 KT |
11356 | { |
11357 | rtx operand = recog_data.operand[opno]; | |
11358 | if (REG_P (operand) | |
11359 | && reg_overlap_mentioned_p (load_reg, operand)) | |
11360 | return true; | |
11361 | ||
11362 | } | |
11363 | return false; | |
11364 | } | |
11365 | ||
8baff86e KT |
11366 | |
11367 | /* When working around the Cortex-A53 erratum 835769, | |
11368 | given rtx_insn INSN, return true if it is a 64-bit multiply-accumulate | |
11369 | instruction and has a preceding memory instruction such that a NOP | |
11370 | should be inserted between them. */ | |
11371 | ||
75cf1494 KT |
11372 | bool |
11373 | aarch64_madd_needs_nop (rtx_insn* insn) | |
11374 | { | |
11375 | enum attr_type attr_type; | |
11376 | rtx_insn *prev; | |
11377 | rtx body; | |
11378 | ||
b32c1043 | 11379 | if (!TARGET_FIX_ERR_A53_835769) |
75cf1494 KT |
11380 | return false; |
11381 | ||
e322d6e3 | 11382 | if (!INSN_P (insn) || recog_memoized (insn) < 0) |
75cf1494 KT |
11383 | return false; |
11384 | ||
11385 | attr_type = get_attr_type (insn); | |
11386 | if (!is_madd_op (attr_type)) | |
11387 | return false; | |
11388 | ||
11389 | prev = aarch64_prev_real_insn (insn); | |
3fea1a75 KT |
11390 | /* aarch64_prev_real_insn can call recog_memoized on insns other than INSN. |
11391 | Restore recog state to INSN to avoid state corruption. */ | |
11392 | extract_constrain_insn_cached (insn); | |
11393 | ||
550e2205 | 11394 | if (!prev || !contains_mem_rtx_p (PATTERN (prev))) |
75cf1494 KT |
11395 | return false; |
11396 | ||
11397 | body = single_set (prev); | |
11398 | ||
11399 | /* If the previous insn is a memory op and there is no dependency between | |
8baff86e KT |
11400 | it and the DImode madd, emit a NOP between them. If body is NULL then we |
11401 | have a complex memory operation, probably a load/store pair. | |
11402 | Be conservative for now and emit a NOP. */ | |
11403 | if (GET_MODE (recog_data.operand[0]) == DImode | |
11404 | && (!body || !dep_between_memop_and_curr (body))) | |
75cf1494 KT |
11405 | return true; |
11406 | ||
11407 | return false; | |
11408 | ||
11409 | } | |
11410 | ||
8baff86e KT |
11411 | |
11412 | /* Implement FINAL_PRESCAN_INSN. */ | |
11413 | ||
75cf1494 KT |
11414 | void |
11415 | aarch64_final_prescan_insn (rtx_insn *insn) | |
11416 | { | |
11417 | if (aarch64_madd_needs_nop (insn)) | |
11418 | fprintf (asm_out_file, "\tnop // between mem op and mult-accumulate\n"); | |
11419 | } | |
11420 | ||
11421 | ||
43e9d192 | 11422 | /* Return the equivalent letter for size. */ |
81c2dfb9 | 11423 | static char |
43e9d192 IB |
11424 | sizetochar (int size) |
11425 | { | |
11426 | switch (size) | |
11427 | { | |
11428 | case 64: return 'd'; | |
11429 | case 32: return 's'; | |
11430 | case 16: return 'h'; | |
11431 | case 8 : return 'b'; | |
11432 | default: gcc_unreachable (); | |
11433 | } | |
11434 | } | |
11435 | ||
3520f7cc JG |
11436 | /* Return true iff x is a uniform vector of floating-point |
11437 | constants, and the constant can be represented in | |
11438 | quarter-precision form. Note, as aarch64_float_const_representable | |
11439 | rejects both +0.0 and -0.0, we will also reject +0.0 and -0.0. */ | |
11440 | static bool | |
11441 | aarch64_vect_float_const_representable_p (rtx x) | |
11442 | { | |
92695fbb RS |
11443 | rtx elt; |
11444 | return (GET_MODE_CLASS (GET_MODE (x)) == MODE_VECTOR_FLOAT | |
11445 | && const_vec_duplicate_p (x, &elt) | |
11446 | && aarch64_float_const_representable_p (elt)); | |
3520f7cc JG |
11447 | } |
11448 | ||
d8edd899 | 11449 | /* Return true for valid and false for invalid. */ |
3ea63f60 | 11450 | bool |
ef4bddc2 | 11451 | aarch64_simd_valid_immediate (rtx op, machine_mode mode, bool inverse, |
48063b9d | 11452 | struct simd_immediate_info *info) |
43e9d192 IB |
11453 | { |
11454 | #define CHECK(STRIDE, ELSIZE, CLASS, TEST, SHIFT, NEG) \ | |
11455 | matches = 1; \ | |
11456 | for (i = 0; i < idx; i += (STRIDE)) \ | |
11457 | if (!(TEST)) \ | |
11458 | matches = 0; \ | |
11459 | if (matches) \ | |
11460 | { \ | |
11461 | immtype = (CLASS); \ | |
11462 | elsize = (ELSIZE); \ | |
43e9d192 IB |
11463 | eshift = (SHIFT); \ |
11464 | emvn = (NEG); \ | |
11465 | break; \ | |
11466 | } | |
11467 | ||
11468 | unsigned int i, elsize = 0, idx = 0, n_elts = CONST_VECTOR_NUNITS (op); | |
cb5ca315 | 11469 | unsigned int innersize = GET_MODE_UNIT_SIZE (mode); |
43e9d192 | 11470 | unsigned char bytes[16]; |
43e9d192 IB |
11471 | int immtype = -1, matches; |
11472 | unsigned int invmask = inverse ? 0xff : 0; | |
11473 | int eshift, emvn; | |
11474 | ||
43e9d192 | 11475 | if (GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) |
3520f7cc | 11476 | { |
81c2dfb9 IB |
11477 | if (! (aarch64_simd_imm_zero_p (op, mode) |
11478 | || aarch64_vect_float_const_representable_p (op))) | |
d8edd899 | 11479 | return false; |
3520f7cc | 11480 | |
48063b9d IB |
11481 | if (info) |
11482 | { | |
e386a52f RS |
11483 | rtx elt = CONST_VECTOR_ELT (op, 0); |
11484 | scalar_float_mode elt_mode | |
11485 | = as_a <scalar_float_mode> (GET_MODE (elt)); | |
11486 | ||
11487 | info->value = elt; | |
11488 | info->element_width = GET_MODE_BITSIZE (elt_mode); | |
48063b9d IB |
11489 | info->mvn = false; |
11490 | info->shift = 0; | |
11491 | } | |
3520f7cc | 11492 | |
d8edd899 | 11493 | return true; |
3520f7cc | 11494 | } |
43e9d192 IB |
11495 | |
11496 | /* Splat vector constant out into a byte vector. */ | |
11497 | for (i = 0; i < n_elts; i++) | |
11498 | { | |
4b1e108c AL |
11499 | /* The vector is provided in gcc endian-neutral fashion. For aarch64_be, |
11500 | it must be laid out in the vector register in reverse order. */ | |
11501 | rtx el = CONST_VECTOR_ELT (op, BYTES_BIG_ENDIAN ? (n_elts - 1 - i) : i); | |
43e9d192 | 11502 | unsigned HOST_WIDE_INT elpart; |
43e9d192 | 11503 | |
ee78df47 KT |
11504 | gcc_assert (CONST_INT_P (el)); |
11505 | elpart = INTVAL (el); | |
11506 | ||
11507 | for (unsigned int byte = 0; byte < innersize; byte++) | |
11508 | { | |
11509 | bytes[idx++] = (elpart & 0xff) ^ invmask; | |
11510 | elpart >>= BITS_PER_UNIT; | |
11511 | } | |
43e9d192 | 11512 | |
43e9d192 IB |
11513 | } |
11514 | ||
11515 | /* Sanity check. */ | |
11516 | gcc_assert (idx == GET_MODE_SIZE (mode)); | |
11517 | ||
11518 | do | |
11519 | { | |
11520 | CHECK (4, 32, 0, bytes[i] == bytes[0] && bytes[i + 1] == 0 | |
11521 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 0, 0); | |
11522 | ||
11523 | CHECK (4, 32, 1, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
11524 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 8, 0); | |
11525 | ||
11526 | CHECK (4, 32, 2, bytes[i] == 0 && bytes[i + 1] == 0 | |
11527 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0, 16, 0); | |
11528 | ||
11529 | CHECK (4, 32, 3, bytes[i] == 0 && bytes[i + 1] == 0 | |
11530 | && bytes[i + 2] == 0 && bytes[i + 3] == bytes[3], 24, 0); | |
11531 | ||
11532 | CHECK (2, 16, 4, bytes[i] == bytes[0] && bytes[i + 1] == 0, 0, 0); | |
11533 | ||
11534 | CHECK (2, 16, 5, bytes[i] == 0 && bytes[i + 1] == bytes[1], 8, 0); | |
11535 | ||
11536 | CHECK (4, 32, 6, bytes[i] == bytes[0] && bytes[i + 1] == 0xff | |
11537 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 0, 1); | |
11538 | ||
11539 | CHECK (4, 32, 7, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
11540 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 8, 1); | |
11541 | ||
11542 | CHECK (4, 32, 8, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
11543 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff, 16, 1); | |
11544 | ||
11545 | CHECK (4, 32, 9, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
11546 | && bytes[i + 2] == 0xff && bytes[i + 3] == bytes[3], 24, 1); | |
11547 | ||
11548 | CHECK (2, 16, 10, bytes[i] == bytes[0] && bytes[i + 1] == 0xff, 0, 1); | |
11549 | ||
11550 | CHECK (2, 16, 11, bytes[i] == 0xff && bytes[i + 1] == bytes[1], 8, 1); | |
11551 | ||
11552 | CHECK (4, 32, 12, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
e4f0f84d | 11553 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 8, 0); |
43e9d192 IB |
11554 | |
11555 | CHECK (4, 32, 13, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
e4f0f84d | 11556 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 8, 1); |
43e9d192 IB |
11557 | |
11558 | CHECK (4, 32, 14, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
e4f0f84d | 11559 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0, 16, 0); |
43e9d192 IB |
11560 | |
11561 | CHECK (4, 32, 15, bytes[i] == 0 && bytes[i + 1] == 0 | |
e4f0f84d | 11562 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff, 16, 1); |
43e9d192 IB |
11563 | |
11564 | CHECK (1, 8, 16, bytes[i] == bytes[0], 0, 0); | |
11565 | ||
11566 | CHECK (1, 64, 17, (bytes[i] == 0 || bytes[i] == 0xff) | |
11567 | && bytes[i] == bytes[(i + 8) % idx], 0, 0); | |
11568 | } | |
11569 | while (0); | |
11570 | ||
e4f0f84d | 11571 | if (immtype == -1) |
d8edd899 | 11572 | return false; |
43e9d192 | 11573 | |
48063b9d | 11574 | if (info) |
43e9d192 | 11575 | { |
48063b9d | 11576 | info->element_width = elsize; |
48063b9d IB |
11577 | info->mvn = emvn != 0; |
11578 | info->shift = eshift; | |
11579 | ||
43e9d192 IB |
11580 | unsigned HOST_WIDE_INT imm = 0; |
11581 | ||
e4f0f84d TB |
11582 | if (immtype >= 12 && immtype <= 15) |
11583 | info->msl = true; | |
11584 | ||
43e9d192 IB |
11585 | /* Un-invert bytes of recognized vector, if necessary. */ |
11586 | if (invmask != 0) | |
11587 | for (i = 0; i < idx; i++) | |
11588 | bytes[i] ^= invmask; | |
11589 | ||
11590 | if (immtype == 17) | |
11591 | { | |
11592 | /* FIXME: Broken on 32-bit H_W_I hosts. */ | |
11593 | gcc_assert (sizeof (HOST_WIDE_INT) == 8); | |
11594 | ||
11595 | for (i = 0; i < 8; i++) | |
11596 | imm |= (unsigned HOST_WIDE_INT) (bytes[i] ? 0xff : 0) | |
11597 | << (i * BITS_PER_UNIT); | |
11598 | ||
43e9d192 | 11599 | |
48063b9d IB |
11600 | info->value = GEN_INT (imm); |
11601 | } | |
11602 | else | |
11603 | { | |
11604 | for (i = 0; i < elsize / BITS_PER_UNIT; i++) | |
11605 | imm |= (unsigned HOST_WIDE_INT) bytes[i] << (i * BITS_PER_UNIT); | |
43e9d192 IB |
11606 | |
11607 | /* Construct 'abcdefgh' because the assembler cannot handle | |
48063b9d IB |
11608 | generic constants. */ |
11609 | if (info->mvn) | |
43e9d192 | 11610 | imm = ~imm; |
48063b9d IB |
11611 | imm = (imm >> info->shift) & 0xff; |
11612 | info->value = GEN_INT (imm); | |
11613 | } | |
43e9d192 IB |
11614 | } |
11615 | ||
48063b9d | 11616 | return true; |
43e9d192 IB |
11617 | #undef CHECK |
11618 | } | |
11619 | ||
43e9d192 IB |
11620 | /* Check of immediate shift constants are within range. */ |
11621 | bool | |
ef4bddc2 | 11622 | aarch64_simd_shift_imm_p (rtx x, machine_mode mode, bool left) |
43e9d192 IB |
11623 | { |
11624 | int bit_width = GET_MODE_UNIT_SIZE (mode) * BITS_PER_UNIT; | |
11625 | if (left) | |
ddeabd3e | 11626 | return aarch64_const_vec_all_same_in_range_p (x, 0, bit_width - 1); |
43e9d192 | 11627 | else |
ddeabd3e | 11628 | return aarch64_const_vec_all_same_in_range_p (x, 1, bit_width); |
43e9d192 IB |
11629 | } |
11630 | ||
3520f7cc JG |
11631 | /* Return true if X is a uniform vector where all elements |
11632 | are either the floating-point constant 0.0 or the | |
11633 | integer constant 0. */ | |
43e9d192 | 11634 | bool |
ef4bddc2 | 11635 | aarch64_simd_imm_zero_p (rtx x, machine_mode mode) |
43e9d192 | 11636 | { |
3520f7cc | 11637 | return x == CONST0_RTX (mode); |
43e9d192 IB |
11638 | } |
11639 | ||
7325d85a KT |
11640 | |
11641 | /* Return the bitmask CONST_INT to select the bits required by a zero extract | |
11642 | operation of width WIDTH at bit position POS. */ | |
11643 | ||
11644 | rtx | |
11645 | aarch64_mask_from_zextract_ops (rtx width, rtx pos) | |
11646 | { | |
11647 | gcc_assert (CONST_INT_P (width)); | |
11648 | gcc_assert (CONST_INT_P (pos)); | |
11649 | ||
11650 | unsigned HOST_WIDE_INT mask | |
11651 | = ((unsigned HOST_WIDE_INT) 1 << UINTVAL (width)) - 1; | |
11652 | return GEN_INT (mask << UINTVAL (pos)); | |
11653 | } | |
11654 | ||
83f8c414 | 11655 | bool |
a6e0bfa7 | 11656 | aarch64_mov_operand_p (rtx x, machine_mode mode) |
83f8c414 | 11657 | { |
83f8c414 CSS |
11658 | if (GET_CODE (x) == HIGH |
11659 | && aarch64_valid_symref (XEXP (x, 0), GET_MODE (XEXP (x, 0)))) | |
11660 | return true; | |
11661 | ||
82614948 | 11662 | if (CONST_INT_P (x)) |
83f8c414 CSS |
11663 | return true; |
11664 | ||
11665 | if (GET_CODE (x) == SYMBOL_REF && mode == DImode && CONSTANT_ADDRESS_P (x)) | |
11666 | return true; | |
11667 | ||
a6e0bfa7 | 11668 | return aarch64_classify_symbolic_expression (x) |
a5350ddc | 11669 | == SYMBOL_TINY_ABSOLUTE; |
83f8c414 CSS |
11670 | } |
11671 | ||
43e9d192 IB |
11672 | /* Return a const_int vector of VAL. */ |
11673 | rtx | |
ab014eb3 | 11674 | aarch64_simd_gen_const_vector_dup (machine_mode mode, HOST_WIDE_INT val) |
43e9d192 IB |
11675 | { |
11676 | int nunits = GET_MODE_NUNITS (mode); | |
11677 | rtvec v = rtvec_alloc (nunits); | |
11678 | int i; | |
11679 | ||
ab014eb3 TC |
11680 | rtx cache = GEN_INT (val); |
11681 | ||
43e9d192 | 11682 | for (i=0; i < nunits; i++) |
ab014eb3 | 11683 | RTVEC_ELT (v, i) = cache; |
43e9d192 IB |
11684 | |
11685 | return gen_rtx_CONST_VECTOR (mode, v); | |
11686 | } | |
11687 | ||
051d0e2f SN |
11688 | /* Check OP is a legal scalar immediate for the MOVI instruction. */ |
11689 | ||
11690 | bool | |
ef4bddc2 | 11691 | aarch64_simd_scalar_immediate_valid_for_move (rtx op, machine_mode mode) |
051d0e2f | 11692 | { |
ef4bddc2 | 11693 | machine_mode vmode; |
051d0e2f SN |
11694 | |
11695 | gcc_assert (!VECTOR_MODE_P (mode)); | |
11696 | vmode = aarch64_preferred_simd_mode (mode); | |
11697 | rtx op_v = aarch64_simd_gen_const_vector_dup (vmode, INTVAL (op)); | |
48063b9d | 11698 | return aarch64_simd_valid_immediate (op_v, vmode, false, NULL); |
051d0e2f SN |
11699 | } |
11700 | ||
988fa693 JG |
11701 | /* Construct and return a PARALLEL RTX vector with elements numbering the |
11702 | lanes of either the high (HIGH == TRUE) or low (HIGH == FALSE) half of | |
11703 | the vector - from the perspective of the architecture. This does not | |
11704 | line up with GCC's perspective on lane numbers, so we end up with | |
11705 | different masks depending on our target endian-ness. The diagram | |
11706 | below may help. We must draw the distinction when building masks | |
11707 | which select one half of the vector. An instruction selecting | |
11708 | architectural low-lanes for a big-endian target, must be described using | |
11709 | a mask selecting GCC high-lanes. | |
11710 | ||
11711 | Big-Endian Little-Endian | |
11712 | ||
11713 | GCC 0 1 2 3 3 2 1 0 | |
11714 | | x | x | x | x | | x | x | x | x | | |
11715 | Architecture 3 2 1 0 3 2 1 0 | |
11716 | ||
11717 | Low Mask: { 2, 3 } { 0, 1 } | |
11718 | High Mask: { 0, 1 } { 2, 3 } | |
11719 | */ | |
11720 | ||
43e9d192 | 11721 | rtx |
ef4bddc2 | 11722 | aarch64_simd_vect_par_cnst_half (machine_mode mode, bool high) |
43e9d192 IB |
11723 | { |
11724 | int nunits = GET_MODE_NUNITS (mode); | |
11725 | rtvec v = rtvec_alloc (nunits / 2); | |
988fa693 JG |
11726 | int high_base = nunits / 2; |
11727 | int low_base = 0; | |
11728 | int base; | |
43e9d192 IB |
11729 | rtx t1; |
11730 | int i; | |
11731 | ||
988fa693 JG |
11732 | if (BYTES_BIG_ENDIAN) |
11733 | base = high ? low_base : high_base; | |
11734 | else | |
11735 | base = high ? high_base : low_base; | |
11736 | ||
11737 | for (i = 0; i < nunits / 2; i++) | |
43e9d192 IB |
11738 | RTVEC_ELT (v, i) = GEN_INT (base + i); |
11739 | ||
11740 | t1 = gen_rtx_PARALLEL (mode, v); | |
11741 | return t1; | |
11742 | } | |
11743 | ||
988fa693 JG |
11744 | /* Check OP for validity as a PARALLEL RTX vector with elements |
11745 | numbering the lanes of either the high (HIGH == TRUE) or low lanes, | |
11746 | from the perspective of the architecture. See the diagram above | |
11747 | aarch64_simd_vect_par_cnst_half for more details. */ | |
11748 | ||
11749 | bool | |
ef4bddc2 | 11750 | aarch64_simd_check_vect_par_cnst_half (rtx op, machine_mode mode, |
988fa693 JG |
11751 | bool high) |
11752 | { | |
11753 | rtx ideal = aarch64_simd_vect_par_cnst_half (mode, high); | |
11754 | HOST_WIDE_INT count_op = XVECLEN (op, 0); | |
11755 | HOST_WIDE_INT count_ideal = XVECLEN (ideal, 0); | |
11756 | int i = 0; | |
11757 | ||
11758 | if (!VECTOR_MODE_P (mode)) | |
11759 | return false; | |
11760 | ||
11761 | if (count_op != count_ideal) | |
11762 | return false; | |
11763 | ||
11764 | for (i = 0; i < count_ideal; i++) | |
11765 | { | |
11766 | rtx elt_op = XVECEXP (op, 0, i); | |
11767 | rtx elt_ideal = XVECEXP (ideal, 0, i); | |
11768 | ||
4aa81c2e | 11769 | if (!CONST_INT_P (elt_op) |
988fa693 JG |
11770 | || INTVAL (elt_ideal) != INTVAL (elt_op)) |
11771 | return false; | |
11772 | } | |
11773 | return true; | |
11774 | } | |
11775 | ||
43e9d192 IB |
11776 | /* Bounds-check lanes. Ensure OPERAND lies between LOW (inclusive) and |
11777 | HIGH (exclusive). */ | |
11778 | void | |
46ed6024 CB |
11779 | aarch64_simd_lane_bounds (rtx operand, HOST_WIDE_INT low, HOST_WIDE_INT high, |
11780 | const_tree exp) | |
43e9d192 IB |
11781 | { |
11782 | HOST_WIDE_INT lane; | |
4aa81c2e | 11783 | gcc_assert (CONST_INT_P (operand)); |
43e9d192 IB |
11784 | lane = INTVAL (operand); |
11785 | ||
11786 | if (lane < low || lane >= high) | |
46ed6024 CB |
11787 | { |
11788 | if (exp) | |
cf0c27ef | 11789 | error ("%Klane %wd out of range %wd - %wd", exp, lane, low, high - 1); |
46ed6024 | 11790 | else |
cf0c27ef | 11791 | error ("lane %wd out of range %wd - %wd", lane, low, high - 1); |
46ed6024 | 11792 | } |
43e9d192 IB |
11793 | } |
11794 | ||
43e9d192 IB |
11795 | /* Return TRUE if OP is a valid vector addressing mode. */ |
11796 | bool | |
11797 | aarch64_simd_mem_operand_p (rtx op) | |
11798 | { | |
11799 | return MEM_P (op) && (GET_CODE (XEXP (op, 0)) == POST_INC | |
4aa81c2e | 11800 | || REG_P (XEXP (op, 0))); |
43e9d192 IB |
11801 | } |
11802 | ||
2d8c6dc1 AH |
11803 | /* Emit a register copy from operand to operand, taking care not to |
11804 | early-clobber source registers in the process. | |
43e9d192 | 11805 | |
2d8c6dc1 AH |
11806 | COUNT is the number of components into which the copy needs to be |
11807 | decomposed. */ | |
43e9d192 | 11808 | void |
b8506a8a | 11809 | aarch64_simd_emit_reg_reg_move (rtx *operands, machine_mode mode, |
2d8c6dc1 | 11810 | unsigned int count) |
43e9d192 IB |
11811 | { |
11812 | unsigned int i; | |
2d8c6dc1 AH |
11813 | int rdest = REGNO (operands[0]); |
11814 | int rsrc = REGNO (operands[1]); | |
43e9d192 IB |
11815 | |
11816 | if (!reg_overlap_mentioned_p (operands[0], operands[1]) | |
2d8c6dc1 AH |
11817 | || rdest < rsrc) |
11818 | for (i = 0; i < count; i++) | |
11819 | emit_move_insn (gen_rtx_REG (mode, rdest + i), | |
11820 | gen_rtx_REG (mode, rsrc + i)); | |
43e9d192 | 11821 | else |
2d8c6dc1 AH |
11822 | for (i = 0; i < count; i++) |
11823 | emit_move_insn (gen_rtx_REG (mode, rdest + count - i - 1), | |
11824 | gen_rtx_REG (mode, rsrc + count - i - 1)); | |
43e9d192 IB |
11825 | } |
11826 | ||
668046d1 | 11827 | /* Compute and return the length of aarch64_simd_reglist<mode>, where <mode> is |
6ec0e5b9 | 11828 | one of VSTRUCT modes: OI, CI, or XI. */ |
668046d1 | 11829 | int |
b8506a8a | 11830 | aarch64_simd_attr_length_rglist (machine_mode mode) |
668046d1 DS |
11831 | { |
11832 | return (GET_MODE_SIZE (mode) / UNITS_PER_VREG) * 4; | |
11833 | } | |
11834 | ||
db0253a4 TB |
11835 | /* Implement target hook TARGET_VECTOR_ALIGNMENT. The AAPCS64 sets the maximum |
11836 | alignment of a vector to 128 bits. */ | |
11837 | static HOST_WIDE_INT | |
11838 | aarch64_simd_vector_alignment (const_tree type) | |
11839 | { | |
9439e9a1 | 11840 | HOST_WIDE_INT align = tree_to_shwi (TYPE_SIZE (type)); |
db0253a4 TB |
11841 | return MIN (align, 128); |
11842 | } | |
11843 | ||
11844 | /* Implement target hook TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE. */ | |
11845 | static bool | |
11846 | aarch64_simd_vector_alignment_reachable (const_tree type, bool is_packed) | |
11847 | { | |
11848 | if (is_packed) | |
11849 | return false; | |
11850 | ||
11851 | /* We guarantee alignment for vectors up to 128-bits. */ | |
11852 | if (tree_int_cst_compare (TYPE_SIZE (type), | |
11853 | bitsize_int (BIGGEST_ALIGNMENT)) > 0) | |
11854 | return false; | |
11855 | ||
11856 | /* Vectors whose size is <= BIGGEST_ALIGNMENT are naturally aligned. */ | |
11857 | return true; | |
11858 | } | |
11859 | ||
7df76747 N |
11860 | /* Return true if the vector misalignment factor is supported by the |
11861 | target. */ | |
11862 | static bool | |
11863 | aarch64_builtin_support_vector_misalignment (machine_mode mode, | |
11864 | const_tree type, int misalignment, | |
11865 | bool is_packed) | |
11866 | { | |
11867 | if (TARGET_SIMD && STRICT_ALIGNMENT) | |
11868 | { | |
11869 | /* Return if movmisalign pattern is not supported for this mode. */ | |
11870 | if (optab_handler (movmisalign_optab, mode) == CODE_FOR_nothing) | |
11871 | return false; | |
11872 | ||
11873 | if (misalignment == -1) | |
11874 | { | |
11875 | /* Misalignment factor is unknown at compile time but we know | |
11876 | it's word aligned. */ | |
11877 | if (aarch64_simd_vector_alignment_reachable (type, is_packed)) | |
11878 | { | |
11879 | int element_size = TREE_INT_CST_LOW (TYPE_SIZE (type)); | |
11880 | ||
11881 | if (element_size != 64) | |
11882 | return true; | |
11883 | } | |
11884 | return false; | |
11885 | } | |
11886 | } | |
11887 | return default_builtin_support_vector_misalignment (mode, type, misalignment, | |
11888 | is_packed); | |
11889 | } | |
11890 | ||
4369c11e TB |
11891 | /* If VALS is a vector constant that can be loaded into a register |
11892 | using DUP, generate instructions to do so and return an RTX to | |
11893 | assign to the register. Otherwise return NULL_RTX. */ | |
11894 | static rtx | |
11895 | aarch64_simd_dup_constant (rtx vals) | |
11896 | { | |
ef4bddc2 RS |
11897 | machine_mode mode = GET_MODE (vals); |
11898 | machine_mode inner_mode = GET_MODE_INNER (mode); | |
4369c11e | 11899 | rtx x; |
4369c11e | 11900 | |
92695fbb | 11901 | if (!const_vec_duplicate_p (vals, &x)) |
4369c11e TB |
11902 | return NULL_RTX; |
11903 | ||
11904 | /* We can load this constant by using DUP and a constant in a | |
11905 | single ARM register. This will be cheaper than a vector | |
11906 | load. */ | |
92695fbb | 11907 | x = copy_to_mode_reg (inner_mode, x); |
4369c11e TB |
11908 | return gen_rtx_VEC_DUPLICATE (mode, x); |
11909 | } | |
11910 | ||
11911 | ||
11912 | /* Generate code to load VALS, which is a PARALLEL containing only | |
11913 | constants (for vec_init) or CONST_VECTOR, efficiently into a | |
11914 | register. Returns an RTX to copy into the register, or NULL_RTX | |
11915 | for a PARALLEL that can not be converted into a CONST_VECTOR. */ | |
1df3f464 | 11916 | static rtx |
4369c11e TB |
11917 | aarch64_simd_make_constant (rtx vals) |
11918 | { | |
ef4bddc2 | 11919 | machine_mode mode = GET_MODE (vals); |
4369c11e TB |
11920 | rtx const_dup; |
11921 | rtx const_vec = NULL_RTX; | |
11922 | int n_elts = GET_MODE_NUNITS (mode); | |
11923 | int n_const = 0; | |
11924 | int i; | |
11925 | ||
11926 | if (GET_CODE (vals) == CONST_VECTOR) | |
11927 | const_vec = vals; | |
11928 | else if (GET_CODE (vals) == PARALLEL) | |
11929 | { | |
11930 | /* A CONST_VECTOR must contain only CONST_INTs and | |
11931 | CONST_DOUBLEs, but CONSTANT_P allows more (e.g. SYMBOL_REF). | |
11932 | Only store valid constants in a CONST_VECTOR. */ | |
11933 | for (i = 0; i < n_elts; ++i) | |
11934 | { | |
11935 | rtx x = XVECEXP (vals, 0, i); | |
11936 | if (CONST_INT_P (x) || CONST_DOUBLE_P (x)) | |
11937 | n_const++; | |
11938 | } | |
11939 | if (n_const == n_elts) | |
11940 | const_vec = gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0)); | |
11941 | } | |
11942 | else | |
11943 | gcc_unreachable (); | |
11944 | ||
11945 | if (const_vec != NULL_RTX | |
48063b9d | 11946 | && aarch64_simd_valid_immediate (const_vec, mode, false, NULL)) |
4369c11e TB |
11947 | /* Load using MOVI/MVNI. */ |
11948 | return const_vec; | |
11949 | else if ((const_dup = aarch64_simd_dup_constant (vals)) != NULL_RTX) | |
11950 | /* Loaded using DUP. */ | |
11951 | return const_dup; | |
11952 | else if (const_vec != NULL_RTX) | |
11953 | /* Load from constant pool. We can not take advantage of single-cycle | |
11954 | LD1 because we need a PC-relative addressing mode. */ | |
11955 | return const_vec; | |
11956 | else | |
11957 | /* A PARALLEL containing something not valid inside CONST_VECTOR. | |
11958 | We can not construct an initializer. */ | |
11959 | return NULL_RTX; | |
11960 | } | |
11961 | ||
35a093b6 JG |
11962 | /* Expand a vector initialisation sequence, such that TARGET is |
11963 | initialised to contain VALS. */ | |
11964 | ||
4369c11e TB |
11965 | void |
11966 | aarch64_expand_vector_init (rtx target, rtx vals) | |
11967 | { | |
ef4bddc2 RS |
11968 | machine_mode mode = GET_MODE (target); |
11969 | machine_mode inner_mode = GET_MODE_INNER (mode); | |
35a093b6 | 11970 | /* The number of vector elements. */ |
4369c11e | 11971 | int n_elts = GET_MODE_NUNITS (mode); |
35a093b6 | 11972 | /* The number of vector elements which are not constant. */ |
8b66a2d4 AL |
11973 | int n_var = 0; |
11974 | rtx any_const = NULL_RTX; | |
35a093b6 JG |
11975 | /* The first element of vals. */ |
11976 | rtx v0 = XVECEXP (vals, 0, 0); | |
4369c11e | 11977 | bool all_same = true; |
4369c11e | 11978 | |
35a093b6 | 11979 | /* Count the number of variable elements to initialise. */ |
8b66a2d4 | 11980 | for (int i = 0; i < n_elts; ++i) |
4369c11e | 11981 | { |
8b66a2d4 | 11982 | rtx x = XVECEXP (vals, 0, i); |
35a093b6 | 11983 | if (!(CONST_INT_P (x) || CONST_DOUBLE_P (x))) |
8b66a2d4 AL |
11984 | ++n_var; |
11985 | else | |
11986 | any_const = x; | |
4369c11e | 11987 | |
35a093b6 | 11988 | all_same &= rtx_equal_p (x, v0); |
4369c11e TB |
11989 | } |
11990 | ||
35a093b6 JG |
11991 | /* No variable elements, hand off to aarch64_simd_make_constant which knows |
11992 | how best to handle this. */ | |
4369c11e TB |
11993 | if (n_var == 0) |
11994 | { | |
11995 | rtx constant = aarch64_simd_make_constant (vals); | |
11996 | if (constant != NULL_RTX) | |
11997 | { | |
11998 | emit_move_insn (target, constant); | |
11999 | return; | |
12000 | } | |
12001 | } | |
12002 | ||
12003 | /* Splat a single non-constant element if we can. */ | |
12004 | if (all_same) | |
12005 | { | |
35a093b6 | 12006 | rtx x = copy_to_mode_reg (inner_mode, v0); |
4369c11e TB |
12007 | aarch64_emit_move (target, gen_rtx_VEC_DUPLICATE (mode, x)); |
12008 | return; | |
12009 | } | |
12010 | ||
85c1b6d7 AP |
12011 | enum insn_code icode = optab_handler (vec_set_optab, mode); |
12012 | gcc_assert (icode != CODE_FOR_nothing); | |
12013 | ||
12014 | /* If there are only variable elements, try to optimize | |
12015 | the insertion using dup for the most common element | |
12016 | followed by insertions. */ | |
12017 | ||
12018 | /* The algorithm will fill matches[*][0] with the earliest matching element, | |
12019 | and matches[X][1] with the count of duplicate elements (if X is the | |
12020 | earliest element which has duplicates). */ | |
12021 | ||
12022 | if (n_var == n_elts && n_elts <= 16) | |
12023 | { | |
12024 | int matches[16][2] = {0}; | |
12025 | for (int i = 0; i < n_elts; i++) | |
12026 | { | |
12027 | for (int j = 0; j <= i; j++) | |
12028 | { | |
12029 | if (rtx_equal_p (XVECEXP (vals, 0, i), XVECEXP (vals, 0, j))) | |
12030 | { | |
12031 | matches[i][0] = j; | |
12032 | matches[j][1]++; | |
12033 | break; | |
12034 | } | |
12035 | } | |
12036 | } | |
12037 | int maxelement = 0; | |
12038 | int maxv = 0; | |
12039 | for (int i = 0; i < n_elts; i++) | |
12040 | if (matches[i][1] > maxv) | |
12041 | { | |
12042 | maxelement = i; | |
12043 | maxv = matches[i][1]; | |
12044 | } | |
12045 | ||
12046 | /* Create a duplicate of the most common element. */ | |
12047 | rtx x = copy_to_mode_reg (inner_mode, XVECEXP (vals, 0, maxelement)); | |
12048 | aarch64_emit_move (target, gen_rtx_VEC_DUPLICATE (mode, x)); | |
12049 | ||
12050 | /* Insert the rest. */ | |
12051 | for (int i = 0; i < n_elts; i++) | |
12052 | { | |
12053 | rtx x = XVECEXP (vals, 0, i); | |
12054 | if (matches[i][0] == maxelement) | |
12055 | continue; | |
12056 | x = copy_to_mode_reg (inner_mode, x); | |
12057 | emit_insn (GEN_FCN (icode) (target, x, GEN_INT (i))); | |
12058 | } | |
12059 | return; | |
12060 | } | |
12061 | ||
35a093b6 JG |
12062 | /* Initialise a vector which is part-variable. We want to first try |
12063 | to build those lanes which are constant in the most efficient way we | |
12064 | can. */ | |
12065 | if (n_var != n_elts) | |
4369c11e TB |
12066 | { |
12067 | rtx copy = copy_rtx (vals); | |
4369c11e | 12068 | |
8b66a2d4 AL |
12069 | /* Load constant part of vector. We really don't care what goes into the |
12070 | parts we will overwrite, but we're more likely to be able to load the | |
12071 | constant efficiently if it has fewer, larger, repeating parts | |
12072 | (see aarch64_simd_valid_immediate). */ | |
12073 | for (int i = 0; i < n_elts; i++) | |
12074 | { | |
12075 | rtx x = XVECEXP (vals, 0, i); | |
12076 | if (CONST_INT_P (x) || CONST_DOUBLE_P (x)) | |
12077 | continue; | |
12078 | rtx subst = any_const; | |
12079 | for (int bit = n_elts / 2; bit > 0; bit /= 2) | |
12080 | { | |
12081 | /* Look in the copied vector, as more elements are const. */ | |
12082 | rtx test = XVECEXP (copy, 0, i ^ bit); | |
12083 | if (CONST_INT_P (test) || CONST_DOUBLE_P (test)) | |
12084 | { | |
12085 | subst = test; | |
12086 | break; | |
12087 | } | |
12088 | } | |
12089 | XVECEXP (copy, 0, i) = subst; | |
12090 | } | |
4369c11e | 12091 | aarch64_expand_vector_init (target, copy); |
35a093b6 | 12092 | } |
4369c11e | 12093 | |
35a093b6 | 12094 | /* Insert the variable lanes directly. */ |
8b66a2d4 | 12095 | for (int i = 0; i < n_elts; i++) |
35a093b6 JG |
12096 | { |
12097 | rtx x = XVECEXP (vals, 0, i); | |
12098 | if (CONST_INT_P (x) || CONST_DOUBLE_P (x)) | |
12099 | continue; | |
12100 | x = copy_to_mode_reg (inner_mode, x); | |
12101 | emit_insn (GEN_FCN (icode) (target, x, GEN_INT (i))); | |
12102 | } | |
4369c11e TB |
12103 | } |
12104 | ||
43e9d192 | 12105 | static unsigned HOST_WIDE_INT |
ef4bddc2 | 12106 | aarch64_shift_truncation_mask (machine_mode mode) |
43e9d192 IB |
12107 | { |
12108 | return | |
ac59ad4e KT |
12109 | (!SHIFT_COUNT_TRUNCATED |
12110 | || aarch64_vector_mode_supported_p (mode) | |
43e9d192 IB |
12111 | || aarch64_vect_struct_mode_p (mode)) ? 0 : (GET_MODE_BITSIZE (mode) - 1); |
12112 | } | |
12113 | ||
43e9d192 IB |
12114 | /* Select a format to encode pointers in exception handling data. */ |
12115 | int | |
12116 | aarch64_asm_preferred_eh_data_format (int code ATTRIBUTE_UNUSED, int global) | |
12117 | { | |
12118 | int type; | |
12119 | switch (aarch64_cmodel) | |
12120 | { | |
12121 | case AARCH64_CMODEL_TINY: | |
12122 | case AARCH64_CMODEL_TINY_PIC: | |
12123 | case AARCH64_CMODEL_SMALL: | |
12124 | case AARCH64_CMODEL_SMALL_PIC: | |
1b1e81f8 | 12125 | case AARCH64_CMODEL_SMALL_SPIC: |
43e9d192 IB |
12126 | /* text+got+data < 4Gb. 4-byte signed relocs are sufficient |
12127 | for everything. */ | |
12128 | type = DW_EH_PE_sdata4; | |
12129 | break; | |
12130 | default: | |
12131 | /* No assumptions here. 8-byte relocs required. */ | |
12132 | type = DW_EH_PE_sdata8; | |
12133 | break; | |
12134 | } | |
12135 | return (global ? DW_EH_PE_indirect : 0) | DW_EH_PE_pcrel | type; | |
12136 | } | |
12137 | ||
e1c1ecb0 KT |
12138 | /* The last .arch and .tune assembly strings that we printed. */ |
12139 | static std::string aarch64_last_printed_arch_string; | |
12140 | static std::string aarch64_last_printed_tune_string; | |
12141 | ||
361fb3ee KT |
12142 | /* Implement ASM_DECLARE_FUNCTION_NAME. Output the ISA features used |
12143 | by the function fndecl. */ | |
12144 | ||
12145 | void | |
12146 | aarch64_declare_function_name (FILE *stream, const char* name, | |
12147 | tree fndecl) | |
12148 | { | |
12149 | tree target_parts = DECL_FUNCTION_SPECIFIC_TARGET (fndecl); | |
12150 | ||
12151 | struct cl_target_option *targ_options; | |
12152 | if (target_parts) | |
12153 | targ_options = TREE_TARGET_OPTION (target_parts); | |
12154 | else | |
12155 | targ_options = TREE_TARGET_OPTION (target_option_current_node); | |
12156 | gcc_assert (targ_options); | |
12157 | ||
12158 | const struct processor *this_arch | |
12159 | = aarch64_get_arch (targ_options->x_explicit_arch); | |
12160 | ||
054b4005 JG |
12161 | unsigned long isa_flags = targ_options->x_aarch64_isa_flags; |
12162 | std::string extension | |
04a99ebe JG |
12163 | = aarch64_get_extension_string_for_isa_flags (isa_flags, |
12164 | this_arch->flags); | |
e1c1ecb0 KT |
12165 | /* Only update the assembler .arch string if it is distinct from the last |
12166 | such string we printed. */ | |
12167 | std::string to_print = this_arch->name + extension; | |
12168 | if (to_print != aarch64_last_printed_arch_string) | |
12169 | { | |
12170 | asm_fprintf (asm_out_file, "\t.arch %s\n", to_print.c_str ()); | |
12171 | aarch64_last_printed_arch_string = to_print; | |
12172 | } | |
361fb3ee KT |
12173 | |
12174 | /* Print the cpu name we're tuning for in the comments, might be | |
e1c1ecb0 KT |
12175 | useful to readers of the generated asm. Do it only when it changes |
12176 | from function to function and verbose assembly is requested. */ | |
361fb3ee KT |
12177 | const struct processor *this_tune |
12178 | = aarch64_get_tune_cpu (targ_options->x_explicit_tune_core); | |
12179 | ||
e1c1ecb0 KT |
12180 | if (flag_debug_asm && aarch64_last_printed_tune_string != this_tune->name) |
12181 | { | |
12182 | asm_fprintf (asm_out_file, "\t" ASM_COMMENT_START ".tune %s\n", | |
12183 | this_tune->name); | |
12184 | aarch64_last_printed_tune_string = this_tune->name; | |
12185 | } | |
361fb3ee KT |
12186 | |
12187 | /* Don't forget the type directive for ELF. */ | |
12188 | ASM_OUTPUT_TYPE_DIRECTIVE (stream, name, "function"); | |
12189 | ASM_OUTPUT_LABEL (stream, name); | |
12190 | } | |
12191 | ||
e1c1ecb0 KT |
12192 | /* Implements TARGET_ASM_FILE_START. Output the assembly header. */ |
12193 | ||
12194 | static void | |
12195 | aarch64_start_file (void) | |
12196 | { | |
12197 | struct cl_target_option *default_options | |
12198 | = TREE_TARGET_OPTION (target_option_default_node); | |
12199 | ||
12200 | const struct processor *default_arch | |
12201 | = aarch64_get_arch (default_options->x_explicit_arch); | |
12202 | unsigned long default_isa_flags = default_options->x_aarch64_isa_flags; | |
12203 | std::string extension | |
04a99ebe JG |
12204 | = aarch64_get_extension_string_for_isa_flags (default_isa_flags, |
12205 | default_arch->flags); | |
e1c1ecb0 KT |
12206 | |
12207 | aarch64_last_printed_arch_string = default_arch->name + extension; | |
12208 | aarch64_last_printed_tune_string = ""; | |
12209 | asm_fprintf (asm_out_file, "\t.arch %s\n", | |
12210 | aarch64_last_printed_arch_string.c_str ()); | |
12211 | ||
12212 | default_file_start (); | |
12213 | } | |
12214 | ||
0462169c SN |
12215 | /* Emit load exclusive. */ |
12216 | ||
12217 | static void | |
ef4bddc2 | 12218 | aarch64_emit_load_exclusive (machine_mode mode, rtx rval, |
0462169c SN |
12219 | rtx mem, rtx model_rtx) |
12220 | { | |
12221 | rtx (*gen) (rtx, rtx, rtx); | |
12222 | ||
12223 | switch (mode) | |
12224 | { | |
4e10a5a7 RS |
12225 | case E_QImode: gen = gen_aarch64_load_exclusiveqi; break; |
12226 | case E_HImode: gen = gen_aarch64_load_exclusivehi; break; | |
12227 | case E_SImode: gen = gen_aarch64_load_exclusivesi; break; | |
12228 | case E_DImode: gen = gen_aarch64_load_exclusivedi; break; | |
0462169c SN |
12229 | default: |
12230 | gcc_unreachable (); | |
12231 | } | |
12232 | ||
12233 | emit_insn (gen (rval, mem, model_rtx)); | |
12234 | } | |
12235 | ||
12236 | /* Emit store exclusive. */ | |
12237 | ||
12238 | static void | |
ef4bddc2 | 12239 | aarch64_emit_store_exclusive (machine_mode mode, rtx bval, |
0462169c SN |
12240 | rtx rval, rtx mem, rtx model_rtx) |
12241 | { | |
12242 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
12243 | ||
12244 | switch (mode) | |
12245 | { | |
4e10a5a7 RS |
12246 | case E_QImode: gen = gen_aarch64_store_exclusiveqi; break; |
12247 | case E_HImode: gen = gen_aarch64_store_exclusivehi; break; | |
12248 | case E_SImode: gen = gen_aarch64_store_exclusivesi; break; | |
12249 | case E_DImode: gen = gen_aarch64_store_exclusivedi; break; | |
0462169c SN |
12250 | default: |
12251 | gcc_unreachable (); | |
12252 | } | |
12253 | ||
12254 | emit_insn (gen (bval, rval, mem, model_rtx)); | |
12255 | } | |
12256 | ||
12257 | /* Mark the previous jump instruction as unlikely. */ | |
12258 | ||
12259 | static void | |
12260 | aarch64_emit_unlikely_jump (rtx insn) | |
12261 | { | |
f370536c | 12262 | rtx_insn *jump = emit_jump_insn (insn); |
5fa396ad | 12263 | add_reg_br_prob_note (jump, profile_probability::very_unlikely ()); |
0462169c SN |
12264 | } |
12265 | ||
12266 | /* Expand a compare and swap pattern. */ | |
12267 | ||
12268 | void | |
12269 | aarch64_expand_compare_and_swap (rtx operands[]) | |
12270 | { | |
12271 | rtx bval, rval, mem, oldval, newval, is_weak, mod_s, mod_f, x; | |
ef4bddc2 | 12272 | machine_mode mode, cmp_mode; |
b0770c0f MW |
12273 | typedef rtx (*gen_cas_fn) (rtx, rtx, rtx, rtx, rtx, rtx, rtx); |
12274 | int idx; | |
12275 | gen_cas_fn gen; | |
12276 | const gen_cas_fn split_cas[] = | |
12277 | { | |
12278 | gen_aarch64_compare_and_swapqi, | |
12279 | gen_aarch64_compare_and_swaphi, | |
12280 | gen_aarch64_compare_and_swapsi, | |
12281 | gen_aarch64_compare_and_swapdi | |
12282 | }; | |
12283 | const gen_cas_fn atomic_cas[] = | |
12284 | { | |
12285 | gen_aarch64_compare_and_swapqi_lse, | |
12286 | gen_aarch64_compare_and_swaphi_lse, | |
12287 | gen_aarch64_compare_and_swapsi_lse, | |
12288 | gen_aarch64_compare_and_swapdi_lse | |
12289 | }; | |
0462169c SN |
12290 | |
12291 | bval = operands[0]; | |
12292 | rval = operands[1]; | |
12293 | mem = operands[2]; | |
12294 | oldval = operands[3]; | |
12295 | newval = operands[4]; | |
12296 | is_weak = operands[5]; | |
12297 | mod_s = operands[6]; | |
12298 | mod_f = operands[7]; | |
12299 | mode = GET_MODE (mem); | |
12300 | cmp_mode = mode; | |
12301 | ||
12302 | /* Normally the succ memory model must be stronger than fail, but in the | |
12303 | unlikely event of fail being ACQUIRE and succ being RELEASE we need to | |
12304 | promote succ to ACQ_REL so that we don't lose the acquire semantics. */ | |
12305 | ||
46b35980 AM |
12306 | if (is_mm_acquire (memmodel_from_int (INTVAL (mod_f))) |
12307 | && is_mm_release (memmodel_from_int (INTVAL (mod_s)))) | |
0462169c SN |
12308 | mod_s = GEN_INT (MEMMODEL_ACQ_REL); |
12309 | ||
12310 | switch (mode) | |
12311 | { | |
4e10a5a7 RS |
12312 | case E_QImode: |
12313 | case E_HImode: | |
0462169c SN |
12314 | /* For short modes, we're going to perform the comparison in SImode, |
12315 | so do the zero-extension now. */ | |
12316 | cmp_mode = SImode; | |
12317 | rval = gen_reg_rtx (SImode); | |
12318 | oldval = convert_modes (SImode, mode, oldval, true); | |
12319 | /* Fall through. */ | |
12320 | ||
4e10a5a7 RS |
12321 | case E_SImode: |
12322 | case E_DImode: | |
0462169c SN |
12323 | /* Force the value into a register if needed. */ |
12324 | if (!aarch64_plus_operand (oldval, mode)) | |
12325 | oldval = force_reg (cmp_mode, oldval); | |
12326 | break; | |
12327 | ||
12328 | default: | |
12329 | gcc_unreachable (); | |
12330 | } | |
12331 | ||
12332 | switch (mode) | |
12333 | { | |
4e10a5a7 RS |
12334 | case E_QImode: idx = 0; break; |
12335 | case E_HImode: idx = 1; break; | |
12336 | case E_SImode: idx = 2; break; | |
12337 | case E_DImode: idx = 3; break; | |
0462169c SN |
12338 | default: |
12339 | gcc_unreachable (); | |
12340 | } | |
b0770c0f MW |
12341 | if (TARGET_LSE) |
12342 | gen = atomic_cas[idx]; | |
12343 | else | |
12344 | gen = split_cas[idx]; | |
0462169c SN |
12345 | |
12346 | emit_insn (gen (rval, mem, oldval, newval, is_weak, mod_s, mod_f)); | |
12347 | ||
12348 | if (mode == QImode || mode == HImode) | |
12349 | emit_move_insn (operands[1], gen_lowpart (mode, rval)); | |
12350 | ||
12351 | x = gen_rtx_REG (CCmode, CC_REGNUM); | |
12352 | x = gen_rtx_EQ (SImode, x, const0_rtx); | |
f7df4a84 | 12353 | emit_insn (gen_rtx_SET (bval, x)); |
0462169c SN |
12354 | } |
12355 | ||
641c2f8b MW |
12356 | /* Test whether the target supports using a atomic load-operate instruction. |
12357 | CODE is the operation and AFTER is TRUE if the data in memory after the | |
12358 | operation should be returned and FALSE if the data before the operation | |
12359 | should be returned. Returns FALSE if the operation isn't supported by the | |
12360 | architecture. */ | |
12361 | ||
12362 | bool | |
12363 | aarch64_atomic_ldop_supported_p (enum rtx_code code) | |
12364 | { | |
12365 | if (!TARGET_LSE) | |
12366 | return false; | |
12367 | ||
12368 | switch (code) | |
12369 | { | |
12370 | case SET: | |
12371 | case AND: | |
12372 | case IOR: | |
12373 | case XOR: | |
12374 | case MINUS: | |
12375 | case PLUS: | |
12376 | return true; | |
12377 | default: | |
12378 | return false; | |
12379 | } | |
12380 | } | |
12381 | ||
f70fb3b6 MW |
12382 | /* Emit a barrier, that is appropriate for memory model MODEL, at the end of a |
12383 | sequence implementing an atomic operation. */ | |
12384 | ||
12385 | static void | |
12386 | aarch64_emit_post_barrier (enum memmodel model) | |
12387 | { | |
12388 | const enum memmodel base_model = memmodel_base (model); | |
12389 | ||
12390 | if (is_mm_sync (model) | |
12391 | && (base_model == MEMMODEL_ACQUIRE | |
12392 | || base_model == MEMMODEL_ACQ_REL | |
12393 | || base_model == MEMMODEL_SEQ_CST)) | |
12394 | { | |
12395 | emit_insn (gen_mem_thread_fence (GEN_INT (MEMMODEL_SEQ_CST))); | |
12396 | } | |
12397 | } | |
12398 | ||
b0770c0f MW |
12399 | /* Emit an atomic compare-and-swap operation. RVAL is the destination register |
12400 | for the data in memory. EXPECTED is the value expected to be in memory. | |
12401 | DESIRED is the value to store to memory. MEM is the memory location. MODEL | |
12402 | is the memory ordering to use. */ | |
12403 | ||
12404 | void | |
12405 | aarch64_gen_atomic_cas (rtx rval, rtx mem, | |
12406 | rtx expected, rtx desired, | |
12407 | rtx model) | |
12408 | { | |
12409 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
12410 | machine_mode mode; | |
12411 | ||
12412 | mode = GET_MODE (mem); | |
12413 | ||
12414 | switch (mode) | |
12415 | { | |
4e10a5a7 RS |
12416 | case E_QImode: gen = gen_aarch64_atomic_casqi; break; |
12417 | case E_HImode: gen = gen_aarch64_atomic_cashi; break; | |
12418 | case E_SImode: gen = gen_aarch64_atomic_cassi; break; | |
12419 | case E_DImode: gen = gen_aarch64_atomic_casdi; break; | |
b0770c0f MW |
12420 | default: |
12421 | gcc_unreachable (); | |
12422 | } | |
12423 | ||
12424 | /* Move the expected value into the CAS destination register. */ | |
12425 | emit_insn (gen_rtx_SET (rval, expected)); | |
12426 | ||
12427 | /* Emit the CAS. */ | |
12428 | emit_insn (gen (rval, mem, desired, model)); | |
12429 | ||
12430 | /* Compare the expected value with the value loaded by the CAS, to establish | |
12431 | whether the swap was made. */ | |
12432 | aarch64_gen_compare_reg (EQ, rval, expected); | |
12433 | } | |
12434 | ||
0462169c SN |
12435 | /* Split a compare and swap pattern. */ |
12436 | ||
12437 | void | |
12438 | aarch64_split_compare_and_swap (rtx operands[]) | |
12439 | { | |
12440 | rtx rval, mem, oldval, newval, scratch; | |
ef4bddc2 | 12441 | machine_mode mode; |
0462169c | 12442 | bool is_weak; |
5d8a22a5 DM |
12443 | rtx_code_label *label1, *label2; |
12444 | rtx x, cond; | |
ab876106 MW |
12445 | enum memmodel model; |
12446 | rtx model_rtx; | |
0462169c SN |
12447 | |
12448 | rval = operands[0]; | |
12449 | mem = operands[1]; | |
12450 | oldval = operands[2]; | |
12451 | newval = operands[3]; | |
12452 | is_weak = (operands[4] != const0_rtx); | |
ab876106 | 12453 | model_rtx = operands[5]; |
0462169c SN |
12454 | scratch = operands[7]; |
12455 | mode = GET_MODE (mem); | |
ab876106 | 12456 | model = memmodel_from_int (INTVAL (model_rtx)); |
0462169c | 12457 | |
17f47f86 KT |
12458 | /* When OLDVAL is zero and we want the strong version we can emit a tighter |
12459 | loop: | |
12460 | .label1: | |
12461 | LD[A]XR rval, [mem] | |
12462 | CBNZ rval, .label2 | |
12463 | ST[L]XR scratch, newval, [mem] | |
12464 | CBNZ scratch, .label1 | |
12465 | .label2: | |
12466 | CMP rval, 0. */ | |
12467 | bool strong_zero_p = !is_weak && oldval == const0_rtx; | |
12468 | ||
5d8a22a5 | 12469 | label1 = NULL; |
0462169c SN |
12470 | if (!is_weak) |
12471 | { | |
12472 | label1 = gen_label_rtx (); | |
12473 | emit_label (label1); | |
12474 | } | |
12475 | label2 = gen_label_rtx (); | |
12476 | ||
ab876106 MW |
12477 | /* The initial load can be relaxed for a __sync operation since a final |
12478 | barrier will be emitted to stop code hoisting. */ | |
12479 | if (is_mm_sync (model)) | |
12480 | aarch64_emit_load_exclusive (mode, rval, mem, | |
12481 | GEN_INT (MEMMODEL_RELAXED)); | |
12482 | else | |
12483 | aarch64_emit_load_exclusive (mode, rval, mem, model_rtx); | |
0462169c | 12484 | |
17f47f86 KT |
12485 | if (strong_zero_p) |
12486 | { | |
12487 | x = gen_rtx_NE (VOIDmode, rval, const0_rtx); | |
12488 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
12489 | gen_rtx_LABEL_REF (Pmode, label2), pc_rtx); | |
12490 | aarch64_emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); | |
12491 | } | |
12492 | else | |
12493 | { | |
12494 | cond = aarch64_gen_compare_reg (NE, rval, oldval); | |
12495 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
12496 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
12497 | gen_rtx_LABEL_REF (Pmode, label2), pc_rtx); | |
12498 | aarch64_emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); | |
12499 | } | |
0462169c | 12500 | |
ab876106 | 12501 | aarch64_emit_store_exclusive (mode, scratch, mem, newval, model_rtx); |
0462169c SN |
12502 | |
12503 | if (!is_weak) | |
12504 | { | |
12505 | x = gen_rtx_NE (VOIDmode, scratch, const0_rtx); | |
12506 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
12507 | gen_rtx_LABEL_REF (Pmode, label1), pc_rtx); | |
f7df4a84 | 12508 | aarch64_emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); |
0462169c SN |
12509 | } |
12510 | else | |
12511 | { | |
12512 | cond = gen_rtx_REG (CCmode, CC_REGNUM); | |
12513 | x = gen_rtx_COMPARE (CCmode, scratch, const0_rtx); | |
f7df4a84 | 12514 | emit_insn (gen_rtx_SET (cond, x)); |
0462169c SN |
12515 | } |
12516 | ||
12517 | emit_label (label2); | |
17f47f86 KT |
12518 | /* If we used a CBNZ in the exchange loop emit an explicit compare with RVAL |
12519 | to set the condition flags. If this is not used it will be removed by | |
12520 | later passes. */ | |
12521 | if (strong_zero_p) | |
12522 | { | |
12523 | cond = gen_rtx_REG (CCmode, CC_REGNUM); | |
12524 | x = gen_rtx_COMPARE (CCmode, rval, const0_rtx); | |
12525 | emit_insn (gen_rtx_SET (cond, x)); | |
12526 | } | |
ab876106 MW |
12527 | /* Emit any final barrier needed for a __sync operation. */ |
12528 | if (is_mm_sync (model)) | |
12529 | aarch64_emit_post_barrier (model); | |
0462169c SN |
12530 | } |
12531 | ||
68729b06 MW |
12532 | /* Emit a BIC instruction. */ |
12533 | ||
12534 | static void | |
12535 | aarch64_emit_bic (machine_mode mode, rtx dst, rtx s1, rtx s2, int shift) | |
12536 | { | |
12537 | rtx shift_rtx = GEN_INT (shift); | |
12538 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
12539 | ||
12540 | switch (mode) | |
12541 | { | |
4e10a5a7 RS |
12542 | case E_SImode: gen = gen_and_one_cmpl_lshrsi3; break; |
12543 | case E_DImode: gen = gen_and_one_cmpl_lshrdi3; break; | |
68729b06 MW |
12544 | default: |
12545 | gcc_unreachable (); | |
12546 | } | |
12547 | ||
12548 | emit_insn (gen (dst, s2, shift_rtx, s1)); | |
12549 | } | |
12550 | ||
9cd7b720 MW |
12551 | /* Emit an atomic swap. */ |
12552 | ||
12553 | static void | |
12554 | aarch64_emit_atomic_swap (machine_mode mode, rtx dst, rtx value, | |
12555 | rtx mem, rtx model) | |
12556 | { | |
12557 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
12558 | ||
12559 | switch (mode) | |
12560 | { | |
4e10a5a7 RS |
12561 | case E_QImode: gen = gen_aarch64_atomic_swpqi; break; |
12562 | case E_HImode: gen = gen_aarch64_atomic_swphi; break; | |
12563 | case E_SImode: gen = gen_aarch64_atomic_swpsi; break; | |
12564 | case E_DImode: gen = gen_aarch64_atomic_swpdi; break; | |
9cd7b720 MW |
12565 | default: |
12566 | gcc_unreachable (); | |
12567 | } | |
12568 | ||
12569 | emit_insn (gen (dst, mem, value, model)); | |
12570 | } | |
12571 | ||
641c2f8b MW |
12572 | /* Operations supported by aarch64_emit_atomic_load_op. */ |
12573 | ||
12574 | enum aarch64_atomic_load_op_code | |
12575 | { | |
12576 | AARCH64_LDOP_PLUS, /* A + B */ | |
12577 | AARCH64_LDOP_XOR, /* A ^ B */ | |
12578 | AARCH64_LDOP_OR, /* A | B */ | |
12579 | AARCH64_LDOP_BIC /* A & ~B */ | |
12580 | }; | |
12581 | ||
12582 | /* Emit an atomic load-operate. */ | |
12583 | ||
12584 | static void | |
12585 | aarch64_emit_atomic_load_op (enum aarch64_atomic_load_op_code code, | |
12586 | machine_mode mode, rtx dst, rtx src, | |
12587 | rtx mem, rtx model) | |
12588 | { | |
12589 | typedef rtx (*aarch64_atomic_load_op_fn) (rtx, rtx, rtx, rtx); | |
12590 | const aarch64_atomic_load_op_fn plus[] = | |
12591 | { | |
12592 | gen_aarch64_atomic_loadaddqi, | |
12593 | gen_aarch64_atomic_loadaddhi, | |
12594 | gen_aarch64_atomic_loadaddsi, | |
12595 | gen_aarch64_atomic_loadadddi | |
12596 | }; | |
12597 | const aarch64_atomic_load_op_fn eor[] = | |
12598 | { | |
12599 | gen_aarch64_atomic_loadeorqi, | |
12600 | gen_aarch64_atomic_loadeorhi, | |
12601 | gen_aarch64_atomic_loadeorsi, | |
12602 | gen_aarch64_atomic_loadeordi | |
12603 | }; | |
12604 | const aarch64_atomic_load_op_fn ior[] = | |
12605 | { | |
12606 | gen_aarch64_atomic_loadsetqi, | |
12607 | gen_aarch64_atomic_loadsethi, | |
12608 | gen_aarch64_atomic_loadsetsi, | |
12609 | gen_aarch64_atomic_loadsetdi | |
12610 | }; | |
12611 | const aarch64_atomic_load_op_fn bic[] = | |
12612 | { | |
12613 | gen_aarch64_atomic_loadclrqi, | |
12614 | gen_aarch64_atomic_loadclrhi, | |
12615 | gen_aarch64_atomic_loadclrsi, | |
12616 | gen_aarch64_atomic_loadclrdi | |
12617 | }; | |
12618 | aarch64_atomic_load_op_fn gen; | |
12619 | int idx = 0; | |
12620 | ||
12621 | switch (mode) | |
12622 | { | |
4e10a5a7 RS |
12623 | case E_QImode: idx = 0; break; |
12624 | case E_HImode: idx = 1; break; | |
12625 | case E_SImode: idx = 2; break; | |
12626 | case E_DImode: idx = 3; break; | |
641c2f8b MW |
12627 | default: |
12628 | gcc_unreachable (); | |
12629 | } | |
12630 | ||
12631 | switch (code) | |
12632 | { | |
12633 | case AARCH64_LDOP_PLUS: gen = plus[idx]; break; | |
12634 | case AARCH64_LDOP_XOR: gen = eor[idx]; break; | |
12635 | case AARCH64_LDOP_OR: gen = ior[idx]; break; | |
12636 | case AARCH64_LDOP_BIC: gen = bic[idx]; break; | |
12637 | default: | |
12638 | gcc_unreachable (); | |
12639 | } | |
12640 | ||
12641 | emit_insn (gen (dst, mem, src, model)); | |
12642 | } | |
12643 | ||
12644 | /* Emit an atomic load+operate. CODE is the operation. OUT_DATA is the | |
68729b06 MW |
12645 | location to store the data read from memory. OUT_RESULT is the location to |
12646 | store the result of the operation. MEM is the memory location to read and | |
12647 | modify. MODEL_RTX is the memory ordering to use. VALUE is the second | |
12648 | operand for the operation. Either OUT_DATA or OUT_RESULT, but not both, can | |
12649 | be NULL. */ | |
9cd7b720 MW |
12650 | |
12651 | void | |
68729b06 | 12652 | aarch64_gen_atomic_ldop (enum rtx_code code, rtx out_data, rtx out_result, |
9cd7b720 MW |
12653 | rtx mem, rtx value, rtx model_rtx) |
12654 | { | |
12655 | machine_mode mode = GET_MODE (mem); | |
641c2f8b MW |
12656 | machine_mode wmode = (mode == DImode ? DImode : SImode); |
12657 | const bool short_mode = (mode < SImode); | |
12658 | aarch64_atomic_load_op_code ldop_code; | |
12659 | rtx src; | |
12660 | rtx x; | |
12661 | ||
12662 | if (out_data) | |
12663 | out_data = gen_lowpart (mode, out_data); | |
9cd7b720 | 12664 | |
68729b06 MW |
12665 | if (out_result) |
12666 | out_result = gen_lowpart (mode, out_result); | |
12667 | ||
641c2f8b MW |
12668 | /* Make sure the value is in a register, putting it into a destination |
12669 | register if it needs to be manipulated. */ | |
12670 | if (!register_operand (value, mode) | |
12671 | || code == AND || code == MINUS) | |
12672 | { | |
68729b06 | 12673 | src = out_result ? out_result : out_data; |
641c2f8b MW |
12674 | emit_move_insn (src, gen_lowpart (mode, value)); |
12675 | } | |
12676 | else | |
12677 | src = value; | |
12678 | gcc_assert (register_operand (src, mode)); | |
9cd7b720 | 12679 | |
641c2f8b MW |
12680 | /* Preprocess the data for the operation as necessary. If the operation is |
12681 | a SET then emit a swap instruction and finish. */ | |
9cd7b720 MW |
12682 | switch (code) |
12683 | { | |
12684 | case SET: | |
641c2f8b | 12685 | aarch64_emit_atomic_swap (mode, out_data, src, mem, model_rtx); |
9cd7b720 MW |
12686 | return; |
12687 | ||
641c2f8b MW |
12688 | case MINUS: |
12689 | /* Negate the value and treat it as a PLUS. */ | |
12690 | { | |
12691 | rtx neg_src; | |
12692 | ||
12693 | /* Resize the value if necessary. */ | |
12694 | if (short_mode) | |
12695 | src = gen_lowpart (wmode, src); | |
12696 | ||
12697 | neg_src = gen_rtx_NEG (wmode, src); | |
12698 | emit_insn (gen_rtx_SET (src, neg_src)); | |
12699 | ||
12700 | if (short_mode) | |
12701 | src = gen_lowpart (mode, src); | |
12702 | } | |
12703 | /* Fall-through. */ | |
12704 | case PLUS: | |
12705 | ldop_code = AARCH64_LDOP_PLUS; | |
12706 | break; | |
12707 | ||
12708 | case IOR: | |
12709 | ldop_code = AARCH64_LDOP_OR; | |
12710 | break; | |
12711 | ||
12712 | case XOR: | |
12713 | ldop_code = AARCH64_LDOP_XOR; | |
12714 | break; | |
12715 | ||
12716 | case AND: | |
12717 | { | |
12718 | rtx not_src; | |
12719 | ||
12720 | /* Resize the value if necessary. */ | |
12721 | if (short_mode) | |
12722 | src = gen_lowpart (wmode, src); | |
12723 | ||
12724 | not_src = gen_rtx_NOT (wmode, src); | |
12725 | emit_insn (gen_rtx_SET (src, not_src)); | |
12726 | ||
12727 | if (short_mode) | |
12728 | src = gen_lowpart (mode, src); | |
12729 | } | |
12730 | ldop_code = AARCH64_LDOP_BIC; | |
12731 | break; | |
12732 | ||
9cd7b720 MW |
12733 | default: |
12734 | /* The operation can't be done with atomic instructions. */ | |
12735 | gcc_unreachable (); | |
12736 | } | |
641c2f8b MW |
12737 | |
12738 | aarch64_emit_atomic_load_op (ldop_code, mode, out_data, src, mem, model_rtx); | |
68729b06 MW |
12739 | |
12740 | /* If necessary, calculate the data in memory after the update by redoing the | |
12741 | operation from values in registers. */ | |
12742 | if (!out_result) | |
12743 | return; | |
12744 | ||
12745 | if (short_mode) | |
12746 | { | |
12747 | src = gen_lowpart (wmode, src); | |
12748 | out_data = gen_lowpart (wmode, out_data); | |
12749 | out_result = gen_lowpart (wmode, out_result); | |
12750 | } | |
12751 | ||
12752 | x = NULL_RTX; | |
12753 | ||
12754 | switch (code) | |
12755 | { | |
12756 | case MINUS: | |
12757 | case PLUS: | |
12758 | x = gen_rtx_PLUS (wmode, out_data, src); | |
12759 | break; | |
12760 | case IOR: | |
12761 | x = gen_rtx_IOR (wmode, out_data, src); | |
12762 | break; | |
12763 | case XOR: | |
12764 | x = gen_rtx_XOR (wmode, out_data, src); | |
12765 | break; | |
12766 | case AND: | |
12767 | aarch64_emit_bic (wmode, out_result, out_data, src, 0); | |
12768 | return; | |
12769 | default: | |
12770 | gcc_unreachable (); | |
12771 | } | |
12772 | ||
12773 | emit_set_insn (out_result, x); | |
12774 | ||
12775 | return; | |
9cd7b720 MW |
12776 | } |
12777 | ||
0462169c SN |
12778 | /* Split an atomic operation. */ |
12779 | ||
12780 | void | |
12781 | aarch64_split_atomic_op (enum rtx_code code, rtx old_out, rtx new_out, rtx mem, | |
9cd7b720 | 12782 | rtx value, rtx model_rtx, rtx cond) |
0462169c | 12783 | { |
ef4bddc2 RS |
12784 | machine_mode mode = GET_MODE (mem); |
12785 | machine_mode wmode = (mode == DImode ? DImode : SImode); | |
f70fb3b6 MW |
12786 | const enum memmodel model = memmodel_from_int (INTVAL (model_rtx)); |
12787 | const bool is_sync = is_mm_sync (model); | |
5d8a22a5 DM |
12788 | rtx_code_label *label; |
12789 | rtx x; | |
0462169c | 12790 | |
9cd7b720 | 12791 | /* Split the atomic operation into a sequence. */ |
0462169c SN |
12792 | label = gen_label_rtx (); |
12793 | emit_label (label); | |
12794 | ||
12795 | if (new_out) | |
12796 | new_out = gen_lowpart (wmode, new_out); | |
12797 | if (old_out) | |
12798 | old_out = gen_lowpart (wmode, old_out); | |
12799 | else | |
12800 | old_out = new_out; | |
12801 | value = simplify_gen_subreg (wmode, value, mode, 0); | |
12802 | ||
f70fb3b6 MW |
12803 | /* The initial load can be relaxed for a __sync operation since a final |
12804 | barrier will be emitted to stop code hoisting. */ | |
12805 | if (is_sync) | |
12806 | aarch64_emit_load_exclusive (mode, old_out, mem, | |
12807 | GEN_INT (MEMMODEL_RELAXED)); | |
12808 | else | |
12809 | aarch64_emit_load_exclusive (mode, old_out, mem, model_rtx); | |
0462169c SN |
12810 | |
12811 | switch (code) | |
12812 | { | |
12813 | case SET: | |
12814 | new_out = value; | |
12815 | break; | |
12816 | ||
12817 | case NOT: | |
12818 | x = gen_rtx_AND (wmode, old_out, value); | |
f7df4a84 | 12819 | emit_insn (gen_rtx_SET (new_out, x)); |
0462169c | 12820 | x = gen_rtx_NOT (wmode, new_out); |
f7df4a84 | 12821 | emit_insn (gen_rtx_SET (new_out, x)); |
0462169c SN |
12822 | break; |
12823 | ||
12824 | case MINUS: | |
12825 | if (CONST_INT_P (value)) | |
12826 | { | |
12827 | value = GEN_INT (-INTVAL (value)); | |
12828 | code = PLUS; | |
12829 | } | |
12830 | /* Fall through. */ | |
12831 | ||
12832 | default: | |
12833 | x = gen_rtx_fmt_ee (code, wmode, old_out, value); | |
f7df4a84 | 12834 | emit_insn (gen_rtx_SET (new_out, x)); |
0462169c SN |
12835 | break; |
12836 | } | |
12837 | ||
12838 | aarch64_emit_store_exclusive (mode, cond, mem, | |
12839 | gen_lowpart (mode, new_out), model_rtx); | |
12840 | ||
12841 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
12842 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
12843 | gen_rtx_LABEL_REF (Pmode, label), pc_rtx); | |
f7df4a84 | 12844 | aarch64_emit_unlikely_jump (gen_rtx_SET (pc_rtx, x)); |
f70fb3b6 MW |
12845 | |
12846 | /* Emit any final barrier needed for a __sync operation. */ | |
12847 | if (is_sync) | |
12848 | aarch64_emit_post_barrier (model); | |
0462169c SN |
12849 | } |
12850 | ||
c2ec330c AL |
12851 | static void |
12852 | aarch64_init_libfuncs (void) | |
12853 | { | |
12854 | /* Half-precision float operations. The compiler handles all operations | |
12855 | with NULL libfuncs by converting to SFmode. */ | |
12856 | ||
12857 | /* Conversions. */ | |
12858 | set_conv_libfunc (trunc_optab, HFmode, SFmode, "__gnu_f2h_ieee"); | |
12859 | set_conv_libfunc (sext_optab, SFmode, HFmode, "__gnu_h2f_ieee"); | |
12860 | ||
12861 | /* Arithmetic. */ | |
12862 | set_optab_libfunc (add_optab, HFmode, NULL); | |
12863 | set_optab_libfunc (sdiv_optab, HFmode, NULL); | |
12864 | set_optab_libfunc (smul_optab, HFmode, NULL); | |
12865 | set_optab_libfunc (neg_optab, HFmode, NULL); | |
12866 | set_optab_libfunc (sub_optab, HFmode, NULL); | |
12867 | ||
12868 | /* Comparisons. */ | |
12869 | set_optab_libfunc (eq_optab, HFmode, NULL); | |
12870 | set_optab_libfunc (ne_optab, HFmode, NULL); | |
12871 | set_optab_libfunc (lt_optab, HFmode, NULL); | |
12872 | set_optab_libfunc (le_optab, HFmode, NULL); | |
12873 | set_optab_libfunc (ge_optab, HFmode, NULL); | |
12874 | set_optab_libfunc (gt_optab, HFmode, NULL); | |
12875 | set_optab_libfunc (unord_optab, HFmode, NULL); | |
12876 | } | |
12877 | ||
43e9d192 | 12878 | /* Target hook for c_mode_for_suffix. */ |
ef4bddc2 | 12879 | static machine_mode |
43e9d192 IB |
12880 | aarch64_c_mode_for_suffix (char suffix) |
12881 | { | |
12882 | if (suffix == 'q') | |
12883 | return TFmode; | |
12884 | ||
12885 | return VOIDmode; | |
12886 | } | |
12887 | ||
3520f7cc JG |
12888 | /* We can only represent floating point constants which will fit in |
12889 | "quarter-precision" values. These values are characterised by | |
12890 | a sign bit, a 4-bit mantissa and a 3-bit exponent. And are given | |
12891 | by: | |
12892 | ||
12893 | (-1)^s * (n/16) * 2^r | |
12894 | ||
12895 | Where: | |
12896 | 's' is the sign bit. | |
12897 | 'n' is an integer in the range 16 <= n <= 31. | |
12898 | 'r' is an integer in the range -3 <= r <= 4. */ | |
12899 | ||
12900 | /* Return true iff X can be represented by a quarter-precision | |
12901 | floating point immediate operand X. Note, we cannot represent 0.0. */ | |
12902 | bool | |
12903 | aarch64_float_const_representable_p (rtx x) | |
12904 | { | |
12905 | /* This represents our current view of how many bits | |
12906 | make up the mantissa. */ | |
12907 | int point_pos = 2 * HOST_BITS_PER_WIDE_INT - 1; | |
ba96cdfb | 12908 | int exponent; |
3520f7cc | 12909 | unsigned HOST_WIDE_INT mantissa, mask; |
3520f7cc | 12910 | REAL_VALUE_TYPE r, m; |
807e902e | 12911 | bool fail; |
3520f7cc JG |
12912 | |
12913 | if (!CONST_DOUBLE_P (x)) | |
12914 | return false; | |
12915 | ||
c2ec330c AL |
12916 | /* We don't support HFmode constants yet. */ |
12917 | if (GET_MODE (x) == VOIDmode || GET_MODE (x) == HFmode) | |
94bfa2da TV |
12918 | return false; |
12919 | ||
34a72c33 | 12920 | r = *CONST_DOUBLE_REAL_VALUE (x); |
3520f7cc JG |
12921 | |
12922 | /* We cannot represent infinities, NaNs or +/-zero. We won't | |
12923 | know if we have +zero until we analyse the mantissa, but we | |
12924 | can reject the other invalid values. */ | |
12925 | if (REAL_VALUE_ISINF (r) || REAL_VALUE_ISNAN (r) | |
12926 | || REAL_VALUE_MINUS_ZERO (r)) | |
12927 | return false; | |
12928 | ||
ba96cdfb | 12929 | /* Extract exponent. */ |
3520f7cc JG |
12930 | r = real_value_abs (&r); |
12931 | exponent = REAL_EXP (&r); | |
12932 | ||
12933 | /* For the mantissa, we expand into two HOST_WIDE_INTS, apart from the | |
12934 | highest (sign) bit, with a fixed binary point at bit point_pos. | |
12935 | m1 holds the low part of the mantissa, m2 the high part. | |
12936 | WARNING: If we ever have a representation using more than 2 * H_W_I - 1 | |
12937 | bits for the mantissa, this can fail (low bits will be lost). */ | |
12938 | real_ldexp (&m, &r, point_pos - exponent); | |
807e902e | 12939 | wide_int w = real_to_integer (&m, &fail, HOST_BITS_PER_WIDE_INT * 2); |
3520f7cc JG |
12940 | |
12941 | /* If the low part of the mantissa has bits set we cannot represent | |
12942 | the value. */ | |
d9074b29 | 12943 | if (w.ulow () != 0) |
3520f7cc JG |
12944 | return false; |
12945 | /* We have rejected the lower HOST_WIDE_INT, so update our | |
12946 | understanding of how many bits lie in the mantissa and | |
12947 | look only at the high HOST_WIDE_INT. */ | |
807e902e | 12948 | mantissa = w.elt (1); |
3520f7cc JG |
12949 | point_pos -= HOST_BITS_PER_WIDE_INT; |
12950 | ||
12951 | /* We can only represent values with a mantissa of the form 1.xxxx. */ | |
12952 | mask = ((unsigned HOST_WIDE_INT)1 << (point_pos - 5)) - 1; | |
12953 | if ((mantissa & mask) != 0) | |
12954 | return false; | |
12955 | ||
12956 | /* Having filtered unrepresentable values, we may now remove all | |
12957 | but the highest 5 bits. */ | |
12958 | mantissa >>= point_pos - 5; | |
12959 | ||
12960 | /* We cannot represent the value 0.0, so reject it. This is handled | |
12961 | elsewhere. */ | |
12962 | if (mantissa == 0) | |
12963 | return false; | |
12964 | ||
12965 | /* Then, as bit 4 is always set, we can mask it off, leaving | |
12966 | the mantissa in the range [0, 15]. */ | |
12967 | mantissa &= ~(1 << 4); | |
12968 | gcc_assert (mantissa <= 15); | |
12969 | ||
12970 | /* GCC internally does not use IEEE754-like encoding (where normalized | |
12971 | significands are in the range [1, 2). GCC uses [0.5, 1) (see real.c). | |
12972 | Our mantissa values are shifted 4 places to the left relative to | |
12973 | normalized IEEE754 so we must modify the exponent returned by REAL_EXP | |
12974 | by 5 places to correct for GCC's representation. */ | |
12975 | exponent = 5 - exponent; | |
12976 | ||
12977 | return (exponent >= 0 && exponent <= 7); | |
12978 | } | |
12979 | ||
12980 | char* | |
81c2dfb9 | 12981 | aarch64_output_simd_mov_immediate (rtx const_vector, |
ef4bddc2 | 12982 | machine_mode mode, |
3520f7cc JG |
12983 | unsigned width) |
12984 | { | |
3ea63f60 | 12985 | bool is_valid; |
3520f7cc | 12986 | static char templ[40]; |
3520f7cc | 12987 | const char *mnemonic; |
e4f0f84d | 12988 | const char *shift_op; |
3520f7cc | 12989 | unsigned int lane_count = 0; |
81c2dfb9 | 12990 | char element_char; |
3520f7cc | 12991 | |
e4f0f84d | 12992 | struct simd_immediate_info info = { NULL_RTX, 0, 0, false, false }; |
48063b9d IB |
12993 | |
12994 | /* This will return true to show const_vector is legal for use as either | |
12995 | a AdvSIMD MOVI instruction (or, implicitly, MVNI) immediate. It will | |
12996 | also update INFO to show how the immediate should be generated. */ | |
81c2dfb9 | 12997 | is_valid = aarch64_simd_valid_immediate (const_vector, mode, false, &info); |
3520f7cc JG |
12998 | gcc_assert (is_valid); |
12999 | ||
81c2dfb9 | 13000 | element_char = sizetochar (info.element_width); |
48063b9d IB |
13001 | lane_count = width / info.element_width; |
13002 | ||
3520f7cc | 13003 | mode = GET_MODE_INNER (mode); |
0d8e1702 | 13004 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
3520f7cc | 13005 | { |
48063b9d | 13006 | gcc_assert (info.shift == 0 && ! info.mvn); |
0d8e1702 KT |
13007 | /* For FP zero change it to a CONST_INT 0 and use the integer SIMD |
13008 | move immediate path. */ | |
48063b9d IB |
13009 | if (aarch64_float_const_zero_rtx_p (info.value)) |
13010 | info.value = GEN_INT (0); | |
13011 | else | |
13012 | { | |
83faf7d0 | 13013 | const unsigned int buf_size = 20; |
48063b9d | 13014 | char float_buf[buf_size] = {'\0'}; |
34a72c33 RS |
13015 | real_to_decimal_for_mode (float_buf, |
13016 | CONST_DOUBLE_REAL_VALUE (info.value), | |
13017 | buf_size, buf_size, 1, mode); | |
48063b9d IB |
13018 | |
13019 | if (lane_count == 1) | |
13020 | snprintf (templ, sizeof (templ), "fmov\t%%d0, %s", float_buf); | |
13021 | else | |
13022 | snprintf (templ, sizeof (templ), "fmov\t%%0.%d%c, %s", | |
81c2dfb9 | 13023 | lane_count, element_char, float_buf); |
48063b9d IB |
13024 | return templ; |
13025 | } | |
3520f7cc | 13026 | } |
3520f7cc | 13027 | |
48063b9d | 13028 | mnemonic = info.mvn ? "mvni" : "movi"; |
e4f0f84d | 13029 | shift_op = info.msl ? "msl" : "lsl"; |
3520f7cc | 13030 | |
0d8e1702 | 13031 | gcc_assert (CONST_INT_P (info.value)); |
3520f7cc | 13032 | if (lane_count == 1) |
48063b9d IB |
13033 | snprintf (templ, sizeof (templ), "%s\t%%d0, " HOST_WIDE_INT_PRINT_HEX, |
13034 | mnemonic, UINTVAL (info.value)); | |
13035 | else if (info.shift) | |
13036 | snprintf (templ, sizeof (templ), "%s\t%%0.%d%c, " HOST_WIDE_INT_PRINT_HEX | |
e4f0f84d TB |
13037 | ", %s %d", mnemonic, lane_count, element_char, |
13038 | UINTVAL (info.value), shift_op, info.shift); | |
3520f7cc | 13039 | else |
48063b9d | 13040 | snprintf (templ, sizeof (templ), "%s\t%%0.%d%c, " HOST_WIDE_INT_PRINT_HEX, |
81c2dfb9 | 13041 | mnemonic, lane_count, element_char, UINTVAL (info.value)); |
3520f7cc JG |
13042 | return templ; |
13043 | } | |
13044 | ||
b7342d25 | 13045 | char* |
a2170965 | 13046 | aarch64_output_scalar_simd_mov_immediate (rtx immediate, machine_mode mode) |
b7342d25 | 13047 | { |
a2170965 TC |
13048 | |
13049 | /* If a floating point number was passed and we desire to use it in an | |
13050 | integer mode do the conversion to integer. */ | |
13051 | if (CONST_DOUBLE_P (immediate) && GET_MODE_CLASS (mode) == MODE_INT) | |
13052 | { | |
13053 | unsigned HOST_WIDE_INT ival; | |
13054 | if (!aarch64_reinterpret_float_as_int (immediate, &ival)) | |
13055 | gcc_unreachable (); | |
13056 | immediate = gen_int_mode (ival, mode); | |
13057 | } | |
13058 | ||
ef4bddc2 | 13059 | machine_mode vmode; |
a2170965 TC |
13060 | /* use a 64 bit mode for everything except for DI/DF mode, where we use |
13061 | a 128 bit vector mode. */ | |
13062 | int width = GET_MODE_BITSIZE (mode) == 64 ? 128 : 64; | |
b7342d25 IB |
13063 | |
13064 | gcc_assert (!VECTOR_MODE_P (mode)); | |
a2170965 | 13065 | vmode = aarch64_simd_container_mode (mode, width); |
b7342d25 | 13066 | rtx v_op = aarch64_simd_gen_const_vector_dup (vmode, INTVAL (immediate)); |
a2170965 | 13067 | return aarch64_output_simd_mov_immediate (v_op, vmode, width); |
b7342d25 IB |
13068 | } |
13069 | ||
88b08073 JG |
13070 | /* Split operands into moves from op[1] + op[2] into op[0]. */ |
13071 | ||
13072 | void | |
13073 | aarch64_split_combinev16qi (rtx operands[3]) | |
13074 | { | |
13075 | unsigned int dest = REGNO (operands[0]); | |
13076 | unsigned int src1 = REGNO (operands[1]); | |
13077 | unsigned int src2 = REGNO (operands[2]); | |
ef4bddc2 | 13078 | machine_mode halfmode = GET_MODE (operands[1]); |
88b08073 JG |
13079 | unsigned int halfregs = HARD_REGNO_NREGS (src1, halfmode); |
13080 | rtx destlo, desthi; | |
13081 | ||
13082 | gcc_assert (halfmode == V16QImode); | |
13083 | ||
13084 | if (src1 == dest && src2 == dest + halfregs) | |
13085 | { | |
13086 | /* No-op move. Can't split to nothing; emit something. */ | |
13087 | emit_note (NOTE_INSN_DELETED); | |
13088 | return; | |
13089 | } | |
13090 | ||
13091 | /* Preserve register attributes for variable tracking. */ | |
13092 | destlo = gen_rtx_REG_offset (operands[0], halfmode, dest, 0); | |
13093 | desthi = gen_rtx_REG_offset (operands[0], halfmode, dest + halfregs, | |
13094 | GET_MODE_SIZE (halfmode)); | |
13095 | ||
13096 | /* Special case of reversed high/low parts. */ | |
13097 | if (reg_overlap_mentioned_p (operands[2], destlo) | |
13098 | && reg_overlap_mentioned_p (operands[1], desthi)) | |
13099 | { | |
13100 | emit_insn (gen_xorv16qi3 (operands[1], operands[1], operands[2])); | |
13101 | emit_insn (gen_xorv16qi3 (operands[2], operands[1], operands[2])); | |
13102 | emit_insn (gen_xorv16qi3 (operands[1], operands[1], operands[2])); | |
13103 | } | |
13104 | else if (!reg_overlap_mentioned_p (operands[2], destlo)) | |
13105 | { | |
13106 | /* Try to avoid unnecessary moves if part of the result | |
13107 | is in the right place already. */ | |
13108 | if (src1 != dest) | |
13109 | emit_move_insn (destlo, operands[1]); | |
13110 | if (src2 != dest + halfregs) | |
13111 | emit_move_insn (desthi, operands[2]); | |
13112 | } | |
13113 | else | |
13114 | { | |
13115 | if (src2 != dest + halfregs) | |
13116 | emit_move_insn (desthi, operands[2]); | |
13117 | if (src1 != dest) | |
13118 | emit_move_insn (destlo, operands[1]); | |
13119 | } | |
13120 | } | |
13121 | ||
13122 | /* vec_perm support. */ | |
13123 | ||
13124 | #define MAX_VECT_LEN 16 | |
13125 | ||
13126 | struct expand_vec_perm_d | |
13127 | { | |
13128 | rtx target, op0, op1; | |
13129 | unsigned char perm[MAX_VECT_LEN]; | |
ef4bddc2 | 13130 | machine_mode vmode; |
88b08073 JG |
13131 | unsigned char nelt; |
13132 | bool one_vector_p; | |
13133 | bool testing_p; | |
13134 | }; | |
13135 | ||
13136 | /* Generate a variable permutation. */ | |
13137 | ||
13138 | static void | |
13139 | aarch64_expand_vec_perm_1 (rtx target, rtx op0, rtx op1, rtx sel) | |
13140 | { | |
ef4bddc2 | 13141 | machine_mode vmode = GET_MODE (target); |
88b08073 JG |
13142 | bool one_vector_p = rtx_equal_p (op0, op1); |
13143 | ||
13144 | gcc_checking_assert (vmode == V8QImode || vmode == V16QImode); | |
13145 | gcc_checking_assert (GET_MODE (op0) == vmode); | |
13146 | gcc_checking_assert (GET_MODE (op1) == vmode); | |
13147 | gcc_checking_assert (GET_MODE (sel) == vmode); | |
13148 | gcc_checking_assert (TARGET_SIMD); | |
13149 | ||
13150 | if (one_vector_p) | |
13151 | { | |
13152 | if (vmode == V8QImode) | |
13153 | { | |
13154 | /* Expand the argument to a V16QI mode by duplicating it. */ | |
13155 | rtx pair = gen_reg_rtx (V16QImode); | |
13156 | emit_insn (gen_aarch64_combinev8qi (pair, op0, op0)); | |
13157 | emit_insn (gen_aarch64_tbl1v8qi (target, pair, sel)); | |
13158 | } | |
13159 | else | |
13160 | { | |
13161 | emit_insn (gen_aarch64_tbl1v16qi (target, op0, sel)); | |
13162 | } | |
13163 | } | |
13164 | else | |
13165 | { | |
13166 | rtx pair; | |
13167 | ||
13168 | if (vmode == V8QImode) | |
13169 | { | |
13170 | pair = gen_reg_rtx (V16QImode); | |
13171 | emit_insn (gen_aarch64_combinev8qi (pair, op0, op1)); | |
13172 | emit_insn (gen_aarch64_tbl1v8qi (target, pair, sel)); | |
13173 | } | |
13174 | else | |
13175 | { | |
13176 | pair = gen_reg_rtx (OImode); | |
13177 | emit_insn (gen_aarch64_combinev16qi (pair, op0, op1)); | |
13178 | emit_insn (gen_aarch64_tbl2v16qi (target, pair, sel)); | |
13179 | } | |
13180 | } | |
13181 | } | |
13182 | ||
13183 | void | |
13184 | aarch64_expand_vec_perm (rtx target, rtx op0, rtx op1, rtx sel) | |
13185 | { | |
ef4bddc2 | 13186 | machine_mode vmode = GET_MODE (target); |
c9d1a16a | 13187 | unsigned int nelt = GET_MODE_NUNITS (vmode); |
88b08073 | 13188 | bool one_vector_p = rtx_equal_p (op0, op1); |
f7c4e5b8 | 13189 | rtx mask; |
88b08073 JG |
13190 | |
13191 | /* The TBL instruction does not use a modulo index, so we must take care | |
13192 | of that ourselves. */ | |
f7c4e5b8 AL |
13193 | mask = aarch64_simd_gen_const_vector_dup (vmode, |
13194 | one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
88b08073 JG |
13195 | sel = expand_simple_binop (vmode, AND, sel, mask, NULL, 0, OPTAB_LIB_WIDEN); |
13196 | ||
f7c4e5b8 AL |
13197 | /* For big-endian, we also need to reverse the index within the vector |
13198 | (but not which vector). */ | |
13199 | if (BYTES_BIG_ENDIAN) | |
13200 | { | |
13201 | /* If one_vector_p, mask is a vector of (nelt - 1)'s already. */ | |
13202 | if (!one_vector_p) | |
13203 | mask = aarch64_simd_gen_const_vector_dup (vmode, nelt - 1); | |
13204 | sel = expand_simple_binop (vmode, XOR, sel, mask, | |
13205 | NULL, 0, OPTAB_LIB_WIDEN); | |
13206 | } | |
88b08073 JG |
13207 | aarch64_expand_vec_perm_1 (target, op0, op1, sel); |
13208 | } | |
13209 | ||
cc4d934f JG |
13210 | /* Recognize patterns suitable for the TRN instructions. */ |
13211 | static bool | |
13212 | aarch64_evpc_trn (struct expand_vec_perm_d *d) | |
13213 | { | |
13214 | unsigned int i, odd, mask, nelt = d->nelt; | |
13215 | rtx out, in0, in1, x; | |
13216 | rtx (*gen) (rtx, rtx, rtx); | |
ef4bddc2 | 13217 | machine_mode vmode = d->vmode; |
cc4d934f JG |
13218 | |
13219 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
13220 | return false; | |
13221 | ||
13222 | /* Note that these are little-endian tests. | |
13223 | We correct for big-endian later. */ | |
13224 | if (d->perm[0] == 0) | |
13225 | odd = 0; | |
13226 | else if (d->perm[0] == 1) | |
13227 | odd = 1; | |
13228 | else | |
13229 | return false; | |
13230 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
13231 | ||
13232 | for (i = 0; i < nelt; i += 2) | |
13233 | { | |
13234 | if (d->perm[i] != i + odd) | |
13235 | return false; | |
13236 | if (d->perm[i + 1] != ((i + nelt + odd) & mask)) | |
13237 | return false; | |
13238 | } | |
13239 | ||
13240 | /* Success! */ | |
13241 | if (d->testing_p) | |
13242 | return true; | |
13243 | ||
13244 | in0 = d->op0; | |
13245 | in1 = d->op1; | |
13246 | if (BYTES_BIG_ENDIAN) | |
13247 | { | |
13248 | x = in0, in0 = in1, in1 = x; | |
13249 | odd = !odd; | |
13250 | } | |
13251 | out = d->target; | |
13252 | ||
13253 | if (odd) | |
13254 | { | |
13255 | switch (vmode) | |
13256 | { | |
4e10a5a7 RS |
13257 | case E_V16QImode: gen = gen_aarch64_trn2v16qi; break; |
13258 | case E_V8QImode: gen = gen_aarch64_trn2v8qi; break; | |
13259 | case E_V8HImode: gen = gen_aarch64_trn2v8hi; break; | |
13260 | case E_V4HImode: gen = gen_aarch64_trn2v4hi; break; | |
13261 | case E_V4SImode: gen = gen_aarch64_trn2v4si; break; | |
13262 | case E_V2SImode: gen = gen_aarch64_trn2v2si; break; | |
13263 | case E_V2DImode: gen = gen_aarch64_trn2v2di; break; | |
13264 | case E_V4HFmode: gen = gen_aarch64_trn2v4hf; break; | |
13265 | case E_V8HFmode: gen = gen_aarch64_trn2v8hf; break; | |
13266 | case E_V4SFmode: gen = gen_aarch64_trn2v4sf; break; | |
13267 | case E_V2SFmode: gen = gen_aarch64_trn2v2sf; break; | |
13268 | case E_V2DFmode: gen = gen_aarch64_trn2v2df; break; | |
cc4d934f JG |
13269 | default: |
13270 | return false; | |
13271 | } | |
13272 | } | |
13273 | else | |
13274 | { | |
13275 | switch (vmode) | |
13276 | { | |
4e10a5a7 RS |
13277 | case E_V16QImode: gen = gen_aarch64_trn1v16qi; break; |
13278 | case E_V8QImode: gen = gen_aarch64_trn1v8qi; break; | |
13279 | case E_V8HImode: gen = gen_aarch64_trn1v8hi; break; | |
13280 | case E_V4HImode: gen = gen_aarch64_trn1v4hi; break; | |
13281 | case E_V4SImode: gen = gen_aarch64_trn1v4si; break; | |
13282 | case E_V2SImode: gen = gen_aarch64_trn1v2si; break; | |
13283 | case E_V2DImode: gen = gen_aarch64_trn1v2di; break; | |
13284 | case E_V4HFmode: gen = gen_aarch64_trn1v4hf; break; | |
13285 | case E_V8HFmode: gen = gen_aarch64_trn1v8hf; break; | |
13286 | case E_V4SFmode: gen = gen_aarch64_trn1v4sf; break; | |
13287 | case E_V2SFmode: gen = gen_aarch64_trn1v2sf; break; | |
13288 | case E_V2DFmode: gen = gen_aarch64_trn1v2df; break; | |
cc4d934f JG |
13289 | default: |
13290 | return false; | |
13291 | } | |
13292 | } | |
13293 | ||
13294 | emit_insn (gen (out, in0, in1)); | |
13295 | return true; | |
13296 | } | |
13297 | ||
13298 | /* Recognize patterns suitable for the UZP instructions. */ | |
13299 | static bool | |
13300 | aarch64_evpc_uzp (struct expand_vec_perm_d *d) | |
13301 | { | |
13302 | unsigned int i, odd, mask, nelt = d->nelt; | |
13303 | rtx out, in0, in1, x; | |
13304 | rtx (*gen) (rtx, rtx, rtx); | |
ef4bddc2 | 13305 | machine_mode vmode = d->vmode; |
cc4d934f JG |
13306 | |
13307 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
13308 | return false; | |
13309 | ||
13310 | /* Note that these are little-endian tests. | |
13311 | We correct for big-endian later. */ | |
13312 | if (d->perm[0] == 0) | |
13313 | odd = 0; | |
13314 | else if (d->perm[0] == 1) | |
13315 | odd = 1; | |
13316 | else | |
13317 | return false; | |
13318 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
13319 | ||
13320 | for (i = 0; i < nelt; i++) | |
13321 | { | |
13322 | unsigned elt = (i * 2 + odd) & mask; | |
13323 | if (d->perm[i] != elt) | |
13324 | return false; | |
13325 | } | |
13326 | ||
13327 | /* Success! */ | |
13328 | if (d->testing_p) | |
13329 | return true; | |
13330 | ||
13331 | in0 = d->op0; | |
13332 | in1 = d->op1; | |
13333 | if (BYTES_BIG_ENDIAN) | |
13334 | { | |
13335 | x = in0, in0 = in1, in1 = x; | |
13336 | odd = !odd; | |
13337 | } | |
13338 | out = d->target; | |
13339 | ||
13340 | if (odd) | |
13341 | { | |
13342 | switch (vmode) | |
13343 | { | |
4e10a5a7 RS |
13344 | case E_V16QImode: gen = gen_aarch64_uzp2v16qi; break; |
13345 | case E_V8QImode: gen = gen_aarch64_uzp2v8qi; break; | |
13346 | case E_V8HImode: gen = gen_aarch64_uzp2v8hi; break; | |
13347 | case E_V4HImode: gen = gen_aarch64_uzp2v4hi; break; | |
13348 | case E_V4SImode: gen = gen_aarch64_uzp2v4si; break; | |
13349 | case E_V2SImode: gen = gen_aarch64_uzp2v2si; break; | |
13350 | case E_V2DImode: gen = gen_aarch64_uzp2v2di; break; | |
13351 | case E_V4HFmode: gen = gen_aarch64_uzp2v4hf; break; | |
13352 | case E_V8HFmode: gen = gen_aarch64_uzp2v8hf; break; | |
13353 | case E_V4SFmode: gen = gen_aarch64_uzp2v4sf; break; | |
13354 | case E_V2SFmode: gen = gen_aarch64_uzp2v2sf; break; | |
13355 | case E_V2DFmode: gen = gen_aarch64_uzp2v2df; break; | |
cc4d934f JG |
13356 | default: |
13357 | return false; | |
13358 | } | |
13359 | } | |
13360 | else | |
13361 | { | |
13362 | switch (vmode) | |
13363 | { | |
4e10a5a7 RS |
13364 | case E_V16QImode: gen = gen_aarch64_uzp1v16qi; break; |
13365 | case E_V8QImode: gen = gen_aarch64_uzp1v8qi; break; | |
13366 | case E_V8HImode: gen = gen_aarch64_uzp1v8hi; break; | |
13367 | case E_V4HImode: gen = gen_aarch64_uzp1v4hi; break; | |
13368 | case E_V4SImode: gen = gen_aarch64_uzp1v4si; break; | |
13369 | case E_V2SImode: gen = gen_aarch64_uzp1v2si; break; | |
13370 | case E_V2DImode: gen = gen_aarch64_uzp1v2di; break; | |
13371 | case E_V4HFmode: gen = gen_aarch64_uzp1v4hf; break; | |
13372 | case E_V8HFmode: gen = gen_aarch64_uzp1v8hf; break; | |
13373 | case E_V4SFmode: gen = gen_aarch64_uzp1v4sf; break; | |
13374 | case E_V2SFmode: gen = gen_aarch64_uzp1v2sf; break; | |
13375 | case E_V2DFmode: gen = gen_aarch64_uzp1v2df; break; | |
cc4d934f JG |
13376 | default: |
13377 | return false; | |
13378 | } | |
13379 | } | |
13380 | ||
13381 | emit_insn (gen (out, in0, in1)); | |
13382 | return true; | |
13383 | } | |
13384 | ||
13385 | /* Recognize patterns suitable for the ZIP instructions. */ | |
13386 | static bool | |
13387 | aarch64_evpc_zip (struct expand_vec_perm_d *d) | |
13388 | { | |
13389 | unsigned int i, high, mask, nelt = d->nelt; | |
13390 | rtx out, in0, in1, x; | |
13391 | rtx (*gen) (rtx, rtx, rtx); | |
ef4bddc2 | 13392 | machine_mode vmode = d->vmode; |
cc4d934f JG |
13393 | |
13394 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
13395 | return false; | |
13396 | ||
13397 | /* Note that these are little-endian tests. | |
13398 | We correct for big-endian later. */ | |
13399 | high = nelt / 2; | |
13400 | if (d->perm[0] == high) | |
13401 | /* Do Nothing. */ | |
13402 | ; | |
13403 | else if (d->perm[0] == 0) | |
13404 | high = 0; | |
13405 | else | |
13406 | return false; | |
13407 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
13408 | ||
13409 | for (i = 0; i < nelt / 2; i++) | |
13410 | { | |
13411 | unsigned elt = (i + high) & mask; | |
13412 | if (d->perm[i * 2] != elt) | |
13413 | return false; | |
13414 | elt = (elt + nelt) & mask; | |
13415 | if (d->perm[i * 2 + 1] != elt) | |
13416 | return false; | |
13417 | } | |
13418 | ||
13419 | /* Success! */ | |
13420 | if (d->testing_p) | |
13421 | return true; | |
13422 | ||
13423 | in0 = d->op0; | |
13424 | in1 = d->op1; | |
13425 | if (BYTES_BIG_ENDIAN) | |
13426 | { | |
13427 | x = in0, in0 = in1, in1 = x; | |
13428 | high = !high; | |
13429 | } | |
13430 | out = d->target; | |
13431 | ||
13432 | if (high) | |
13433 | { | |
13434 | switch (vmode) | |
13435 | { | |
4e10a5a7 RS |
13436 | case E_V16QImode: gen = gen_aarch64_zip2v16qi; break; |
13437 | case E_V8QImode: gen = gen_aarch64_zip2v8qi; break; | |
13438 | case E_V8HImode: gen = gen_aarch64_zip2v8hi; break; | |
13439 | case E_V4HImode: gen = gen_aarch64_zip2v4hi; break; | |
13440 | case E_V4SImode: gen = gen_aarch64_zip2v4si; break; | |
13441 | case E_V2SImode: gen = gen_aarch64_zip2v2si; break; | |
13442 | case E_V2DImode: gen = gen_aarch64_zip2v2di; break; | |
13443 | case E_V4HFmode: gen = gen_aarch64_zip2v4hf; break; | |
13444 | case E_V8HFmode: gen = gen_aarch64_zip2v8hf; break; | |
13445 | case E_V4SFmode: gen = gen_aarch64_zip2v4sf; break; | |
13446 | case E_V2SFmode: gen = gen_aarch64_zip2v2sf; break; | |
13447 | case E_V2DFmode: gen = gen_aarch64_zip2v2df; break; | |
cc4d934f JG |
13448 | default: |
13449 | return false; | |
13450 | } | |
13451 | } | |
13452 | else | |
13453 | { | |
13454 | switch (vmode) | |
13455 | { | |
4e10a5a7 RS |
13456 | case E_V16QImode: gen = gen_aarch64_zip1v16qi; break; |
13457 | case E_V8QImode: gen = gen_aarch64_zip1v8qi; break; | |
13458 | case E_V8HImode: gen = gen_aarch64_zip1v8hi; break; | |
13459 | case E_V4HImode: gen = gen_aarch64_zip1v4hi; break; | |
13460 | case E_V4SImode: gen = gen_aarch64_zip1v4si; break; | |
13461 | case E_V2SImode: gen = gen_aarch64_zip1v2si; break; | |
13462 | case E_V2DImode: gen = gen_aarch64_zip1v2di; break; | |
13463 | case E_V4HFmode: gen = gen_aarch64_zip1v4hf; break; | |
13464 | case E_V8HFmode: gen = gen_aarch64_zip1v8hf; break; | |
13465 | case E_V4SFmode: gen = gen_aarch64_zip1v4sf; break; | |
13466 | case E_V2SFmode: gen = gen_aarch64_zip1v2sf; break; | |
13467 | case E_V2DFmode: gen = gen_aarch64_zip1v2df; break; | |
cc4d934f JG |
13468 | default: |
13469 | return false; | |
13470 | } | |
13471 | } | |
13472 | ||
13473 | emit_insn (gen (out, in0, in1)); | |
13474 | return true; | |
13475 | } | |
13476 | ||
ae0533da AL |
13477 | /* Recognize patterns for the EXT insn. */ |
13478 | ||
13479 | static bool | |
13480 | aarch64_evpc_ext (struct expand_vec_perm_d *d) | |
13481 | { | |
13482 | unsigned int i, nelt = d->nelt; | |
13483 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
13484 | rtx offset; | |
13485 | ||
13486 | unsigned int location = d->perm[0]; /* Always < nelt. */ | |
13487 | ||
13488 | /* Check if the extracted indices are increasing by one. */ | |
13489 | for (i = 1; i < nelt; i++) | |
13490 | { | |
13491 | unsigned int required = location + i; | |
13492 | if (d->one_vector_p) | |
13493 | { | |
13494 | /* We'll pass the same vector in twice, so allow indices to wrap. */ | |
13495 | required &= (nelt - 1); | |
13496 | } | |
13497 | if (d->perm[i] != required) | |
13498 | return false; | |
13499 | } | |
13500 | ||
ae0533da AL |
13501 | switch (d->vmode) |
13502 | { | |
4e10a5a7 RS |
13503 | case E_V16QImode: gen = gen_aarch64_extv16qi; break; |
13504 | case E_V8QImode: gen = gen_aarch64_extv8qi; break; | |
13505 | case E_V4HImode: gen = gen_aarch64_extv4hi; break; | |
13506 | case E_V8HImode: gen = gen_aarch64_extv8hi; break; | |
13507 | case E_V2SImode: gen = gen_aarch64_extv2si; break; | |
13508 | case E_V4SImode: gen = gen_aarch64_extv4si; break; | |
13509 | case E_V4HFmode: gen = gen_aarch64_extv4hf; break; | |
13510 | case E_V8HFmode: gen = gen_aarch64_extv8hf; break; | |
13511 | case E_V2SFmode: gen = gen_aarch64_extv2sf; break; | |
13512 | case E_V4SFmode: gen = gen_aarch64_extv4sf; break; | |
13513 | case E_V2DImode: gen = gen_aarch64_extv2di; break; | |
13514 | case E_V2DFmode: gen = gen_aarch64_extv2df; break; | |
ae0533da AL |
13515 | default: |
13516 | return false; | |
13517 | } | |
13518 | ||
13519 | /* Success! */ | |
13520 | if (d->testing_p) | |
13521 | return true; | |
13522 | ||
b31e65bb AL |
13523 | /* The case where (location == 0) is a no-op for both big- and little-endian, |
13524 | and is removed by the mid-end at optimization levels -O1 and higher. */ | |
13525 | ||
13526 | if (BYTES_BIG_ENDIAN && (location != 0)) | |
ae0533da AL |
13527 | { |
13528 | /* After setup, we want the high elements of the first vector (stored | |
13529 | at the LSB end of the register), and the low elements of the second | |
13530 | vector (stored at the MSB end of the register). So swap. */ | |
cb5c6c29 | 13531 | std::swap (d->op0, d->op1); |
ae0533da AL |
13532 | /* location != 0 (above), so safe to assume (nelt - location) < nelt. */ |
13533 | location = nelt - location; | |
13534 | } | |
13535 | ||
13536 | offset = GEN_INT (location); | |
13537 | emit_insn (gen (d->target, d->op0, d->op1, offset)); | |
13538 | return true; | |
13539 | } | |
13540 | ||
923fcec3 AL |
13541 | /* Recognize patterns for the REV insns. */ |
13542 | ||
13543 | static bool | |
13544 | aarch64_evpc_rev (struct expand_vec_perm_d *d) | |
13545 | { | |
13546 | unsigned int i, j, diff, nelt = d->nelt; | |
13547 | rtx (*gen) (rtx, rtx); | |
13548 | ||
13549 | if (!d->one_vector_p) | |
13550 | return false; | |
13551 | ||
13552 | diff = d->perm[0]; | |
13553 | switch (diff) | |
13554 | { | |
13555 | case 7: | |
13556 | switch (d->vmode) | |
13557 | { | |
4e10a5a7 RS |
13558 | case E_V16QImode: gen = gen_aarch64_rev64v16qi; break; |
13559 | case E_V8QImode: gen = gen_aarch64_rev64v8qi; break; | |
923fcec3 AL |
13560 | default: |
13561 | return false; | |
13562 | } | |
13563 | break; | |
13564 | case 3: | |
13565 | switch (d->vmode) | |
13566 | { | |
4e10a5a7 RS |
13567 | case E_V16QImode: gen = gen_aarch64_rev32v16qi; break; |
13568 | case E_V8QImode: gen = gen_aarch64_rev32v8qi; break; | |
13569 | case E_V8HImode: gen = gen_aarch64_rev64v8hi; break; | |
13570 | case E_V4HImode: gen = gen_aarch64_rev64v4hi; break; | |
923fcec3 AL |
13571 | default: |
13572 | return false; | |
13573 | } | |
13574 | break; | |
13575 | case 1: | |
13576 | switch (d->vmode) | |
13577 | { | |
4e10a5a7 RS |
13578 | case E_V16QImode: gen = gen_aarch64_rev16v16qi; break; |
13579 | case E_V8QImode: gen = gen_aarch64_rev16v8qi; break; | |
13580 | case E_V8HImode: gen = gen_aarch64_rev32v8hi; break; | |
13581 | case E_V4HImode: gen = gen_aarch64_rev32v4hi; break; | |
13582 | case E_V4SImode: gen = gen_aarch64_rev64v4si; break; | |
13583 | case E_V2SImode: gen = gen_aarch64_rev64v2si; break; | |
13584 | case E_V4SFmode: gen = gen_aarch64_rev64v4sf; break; | |
13585 | case E_V2SFmode: gen = gen_aarch64_rev64v2sf; break; | |
13586 | case E_V8HFmode: gen = gen_aarch64_rev64v8hf; break; | |
13587 | case E_V4HFmode: gen = gen_aarch64_rev64v4hf; break; | |
923fcec3 AL |
13588 | default: |
13589 | return false; | |
13590 | } | |
13591 | break; | |
13592 | default: | |
13593 | return false; | |
13594 | } | |
13595 | ||
13596 | for (i = 0; i < nelt ; i += diff + 1) | |
13597 | for (j = 0; j <= diff; j += 1) | |
13598 | { | |
13599 | /* This is guaranteed to be true as the value of diff | |
13600 | is 7, 3, 1 and we should have enough elements in the | |
13601 | queue to generate this. Getting a vector mask with a | |
13602 | value of diff other than these values implies that | |
13603 | something is wrong by the time we get here. */ | |
13604 | gcc_assert (i + j < nelt); | |
13605 | if (d->perm[i + j] != i + diff - j) | |
13606 | return false; | |
13607 | } | |
13608 | ||
13609 | /* Success! */ | |
13610 | if (d->testing_p) | |
13611 | return true; | |
13612 | ||
13613 | emit_insn (gen (d->target, d->op0)); | |
13614 | return true; | |
13615 | } | |
13616 | ||
91bd4114 JG |
13617 | static bool |
13618 | aarch64_evpc_dup (struct expand_vec_perm_d *d) | |
13619 | { | |
13620 | rtx (*gen) (rtx, rtx, rtx); | |
13621 | rtx out = d->target; | |
13622 | rtx in0; | |
ef4bddc2 | 13623 | machine_mode vmode = d->vmode; |
91bd4114 JG |
13624 | unsigned int i, elt, nelt = d->nelt; |
13625 | rtx lane; | |
13626 | ||
91bd4114 JG |
13627 | elt = d->perm[0]; |
13628 | for (i = 1; i < nelt; i++) | |
13629 | { | |
13630 | if (elt != d->perm[i]) | |
13631 | return false; | |
13632 | } | |
13633 | ||
13634 | /* The generic preparation in aarch64_expand_vec_perm_const_1 | |
13635 | swaps the operand order and the permute indices if it finds | |
13636 | d->perm[0] to be in the second operand. Thus, we can always | |
13637 | use d->op0 and need not do any extra arithmetic to get the | |
13638 | correct lane number. */ | |
13639 | in0 = d->op0; | |
f901401e | 13640 | lane = GEN_INT (elt); /* The pattern corrects for big-endian. */ |
91bd4114 JG |
13641 | |
13642 | switch (vmode) | |
13643 | { | |
4e10a5a7 RS |
13644 | case E_V16QImode: gen = gen_aarch64_dup_lanev16qi; break; |
13645 | case E_V8QImode: gen = gen_aarch64_dup_lanev8qi; break; | |
13646 | case E_V8HImode: gen = gen_aarch64_dup_lanev8hi; break; | |
13647 | case E_V4HImode: gen = gen_aarch64_dup_lanev4hi; break; | |
13648 | case E_V4SImode: gen = gen_aarch64_dup_lanev4si; break; | |
13649 | case E_V2SImode: gen = gen_aarch64_dup_lanev2si; break; | |
13650 | case E_V2DImode: gen = gen_aarch64_dup_lanev2di; break; | |
13651 | case E_V8HFmode: gen = gen_aarch64_dup_lanev8hf; break; | |
13652 | case E_V4HFmode: gen = gen_aarch64_dup_lanev4hf; break; | |
13653 | case E_V4SFmode: gen = gen_aarch64_dup_lanev4sf; break; | |
13654 | case E_V2SFmode: gen = gen_aarch64_dup_lanev2sf; break; | |
13655 | case E_V2DFmode: gen = gen_aarch64_dup_lanev2df; break; | |
91bd4114 JG |
13656 | default: |
13657 | return false; | |
13658 | } | |
13659 | ||
13660 | emit_insn (gen (out, in0, lane)); | |
13661 | return true; | |
13662 | } | |
13663 | ||
88b08073 JG |
13664 | static bool |
13665 | aarch64_evpc_tbl (struct expand_vec_perm_d *d) | |
13666 | { | |
13667 | rtx rperm[MAX_VECT_LEN], sel; | |
ef4bddc2 | 13668 | machine_mode vmode = d->vmode; |
88b08073 JG |
13669 | unsigned int i, nelt = d->nelt; |
13670 | ||
88b08073 JG |
13671 | if (d->testing_p) |
13672 | return true; | |
13673 | ||
13674 | /* Generic code will try constant permutation twice. Once with the | |
13675 | original mode and again with the elements lowered to QImode. | |
13676 | So wait and don't do the selector expansion ourselves. */ | |
13677 | if (vmode != V8QImode && vmode != V16QImode) | |
13678 | return false; | |
13679 | ||
13680 | for (i = 0; i < nelt; ++i) | |
bbcc9c00 TB |
13681 | { |
13682 | int nunits = GET_MODE_NUNITS (vmode); | |
13683 | ||
13684 | /* If big-endian and two vectors we end up with a weird mixed-endian | |
13685 | mode on NEON. Reverse the index within each word but not the word | |
13686 | itself. */ | |
13687 | rperm[i] = GEN_INT (BYTES_BIG_ENDIAN ? d->perm[i] ^ (nunits - 1) | |
13688 | : d->perm[i]); | |
13689 | } | |
88b08073 JG |
13690 | sel = gen_rtx_CONST_VECTOR (vmode, gen_rtvec_v (nelt, rperm)); |
13691 | sel = force_reg (vmode, sel); | |
13692 | ||
13693 | aarch64_expand_vec_perm_1 (d->target, d->op0, d->op1, sel); | |
13694 | return true; | |
13695 | } | |
13696 | ||
13697 | static bool | |
13698 | aarch64_expand_vec_perm_const_1 (struct expand_vec_perm_d *d) | |
13699 | { | |
13700 | /* The pattern matching functions above are written to look for a small | |
13701 | number to begin the sequence (0, 1, N/2). If we begin with an index | |
13702 | from the second operand, we can swap the operands. */ | |
13703 | if (d->perm[0] >= d->nelt) | |
13704 | { | |
13705 | unsigned i, nelt = d->nelt; | |
88b08073 | 13706 | |
0696116a | 13707 | gcc_assert (nelt == (nelt & -nelt)); |
88b08073 | 13708 | for (i = 0; i < nelt; ++i) |
0696116a | 13709 | d->perm[i] ^= nelt; /* Keep the same index, but in the other vector. */ |
88b08073 | 13710 | |
cb5c6c29 | 13711 | std::swap (d->op0, d->op1); |
88b08073 JG |
13712 | } |
13713 | ||
13714 | if (TARGET_SIMD) | |
cc4d934f | 13715 | { |
923fcec3 AL |
13716 | if (aarch64_evpc_rev (d)) |
13717 | return true; | |
13718 | else if (aarch64_evpc_ext (d)) | |
ae0533da | 13719 | return true; |
f901401e AL |
13720 | else if (aarch64_evpc_dup (d)) |
13721 | return true; | |
ae0533da | 13722 | else if (aarch64_evpc_zip (d)) |
cc4d934f JG |
13723 | return true; |
13724 | else if (aarch64_evpc_uzp (d)) | |
13725 | return true; | |
13726 | else if (aarch64_evpc_trn (d)) | |
13727 | return true; | |
13728 | return aarch64_evpc_tbl (d); | |
13729 | } | |
88b08073 JG |
13730 | return false; |
13731 | } | |
13732 | ||
13733 | /* Expand a vec_perm_const pattern. */ | |
13734 | ||
13735 | bool | |
13736 | aarch64_expand_vec_perm_const (rtx target, rtx op0, rtx op1, rtx sel) | |
13737 | { | |
13738 | struct expand_vec_perm_d d; | |
13739 | int i, nelt, which; | |
13740 | ||
13741 | d.target = target; | |
13742 | d.op0 = op0; | |
13743 | d.op1 = op1; | |
13744 | ||
13745 | d.vmode = GET_MODE (target); | |
13746 | gcc_assert (VECTOR_MODE_P (d.vmode)); | |
13747 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
13748 | d.testing_p = false; | |
13749 | ||
13750 | for (i = which = 0; i < nelt; ++i) | |
13751 | { | |
13752 | rtx e = XVECEXP (sel, 0, i); | |
13753 | int ei = INTVAL (e) & (2 * nelt - 1); | |
13754 | which |= (ei < nelt ? 1 : 2); | |
13755 | d.perm[i] = ei; | |
13756 | } | |
13757 | ||
13758 | switch (which) | |
13759 | { | |
13760 | default: | |
13761 | gcc_unreachable (); | |
13762 | ||
13763 | case 3: | |
13764 | d.one_vector_p = false; | |
13765 | if (!rtx_equal_p (op0, op1)) | |
13766 | break; | |
13767 | ||
13768 | /* The elements of PERM do not suggest that only the first operand | |
13769 | is used, but both operands are identical. Allow easier matching | |
13770 | of the permutation by folding the permutation into the single | |
13771 | input vector. */ | |
13772 | /* Fall Through. */ | |
13773 | case 2: | |
13774 | for (i = 0; i < nelt; ++i) | |
13775 | d.perm[i] &= nelt - 1; | |
13776 | d.op0 = op1; | |
13777 | d.one_vector_p = true; | |
13778 | break; | |
13779 | ||
13780 | case 1: | |
13781 | d.op1 = op0; | |
13782 | d.one_vector_p = true; | |
13783 | break; | |
13784 | } | |
13785 | ||
13786 | return aarch64_expand_vec_perm_const_1 (&d); | |
13787 | } | |
13788 | ||
13789 | static bool | |
ef4bddc2 | 13790 | aarch64_vectorize_vec_perm_const_ok (machine_mode vmode, |
88b08073 JG |
13791 | const unsigned char *sel) |
13792 | { | |
13793 | struct expand_vec_perm_d d; | |
13794 | unsigned int i, nelt, which; | |
13795 | bool ret; | |
13796 | ||
13797 | d.vmode = vmode; | |
13798 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
13799 | d.testing_p = true; | |
13800 | memcpy (d.perm, sel, nelt); | |
13801 | ||
13802 | /* Calculate whether all elements are in one vector. */ | |
13803 | for (i = which = 0; i < nelt; ++i) | |
13804 | { | |
13805 | unsigned char e = d.perm[i]; | |
13806 | gcc_assert (e < 2 * nelt); | |
13807 | which |= (e < nelt ? 1 : 2); | |
13808 | } | |
13809 | ||
13810 | /* If all elements are from the second vector, reindex as if from the | |
13811 | first vector. */ | |
13812 | if (which == 2) | |
13813 | for (i = 0; i < nelt; ++i) | |
13814 | d.perm[i] -= nelt; | |
13815 | ||
13816 | /* Check whether the mask can be applied to a single vector. */ | |
13817 | d.one_vector_p = (which != 3); | |
13818 | ||
13819 | d.target = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 1); | |
13820 | d.op1 = d.op0 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 2); | |
13821 | if (!d.one_vector_p) | |
13822 | d.op1 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 3); | |
13823 | ||
13824 | start_sequence (); | |
13825 | ret = aarch64_expand_vec_perm_const_1 (&d); | |
13826 | end_sequence (); | |
13827 | ||
13828 | return ret; | |
13829 | } | |
13830 | ||
668046d1 | 13831 | rtx |
b8506a8a | 13832 | aarch64_reverse_mask (machine_mode mode) |
668046d1 DS |
13833 | { |
13834 | /* We have to reverse each vector because we dont have | |
13835 | a permuted load that can reverse-load according to ABI rules. */ | |
13836 | rtx mask; | |
13837 | rtvec v = rtvec_alloc (16); | |
13838 | int i, j; | |
13839 | int nunits = GET_MODE_NUNITS (mode); | |
13840 | int usize = GET_MODE_UNIT_SIZE (mode); | |
13841 | ||
13842 | gcc_assert (BYTES_BIG_ENDIAN); | |
13843 | gcc_assert (AARCH64_VALID_SIMD_QREG_MODE (mode)); | |
13844 | ||
13845 | for (i = 0; i < nunits; i++) | |
13846 | for (j = 0; j < usize; j++) | |
13847 | RTVEC_ELT (v, i * usize + j) = GEN_INT ((i + 1) * usize - 1 - j); | |
13848 | mask = gen_rtx_CONST_VECTOR (V16QImode, v); | |
13849 | return force_reg (V16QImode, mask); | |
13850 | } | |
13851 | ||
61f17a5c WD |
13852 | /* Implement MODES_TIEABLE_P. In principle we should always return true. |
13853 | However due to issues with register allocation it is preferable to avoid | |
13854 | tieing integer scalar and FP scalar modes. Executing integer operations | |
13855 | in general registers is better than treating them as scalar vector | |
13856 | operations. This reduces latency and avoids redundant int<->FP moves. | |
13857 | So tie modes if they are either the same class, or vector modes with | |
13858 | other vector modes, vector structs or any scalar mode. | |
13859 | */ | |
97e1ad78 JG |
13860 | |
13861 | bool | |
ef4bddc2 | 13862 | aarch64_modes_tieable_p (machine_mode mode1, machine_mode mode2) |
97e1ad78 JG |
13863 | { |
13864 | if (GET_MODE_CLASS (mode1) == GET_MODE_CLASS (mode2)) | |
13865 | return true; | |
13866 | ||
13867 | /* We specifically want to allow elements of "structure" modes to | |
13868 | be tieable to the structure. This more general condition allows | |
13869 | other rarer situations too. */ | |
61f17a5c WD |
13870 | if (aarch64_vector_mode_p (mode1) && aarch64_vector_mode_p (mode2)) |
13871 | return true; | |
13872 | ||
13873 | /* Also allow any scalar modes with vectors. */ | |
13874 | if (aarch64_vector_mode_supported_p (mode1) | |
13875 | || aarch64_vector_mode_supported_p (mode2)) | |
97e1ad78 JG |
13876 | return true; |
13877 | ||
13878 | return false; | |
13879 | } | |
13880 | ||
e2c75eea JG |
13881 | /* Return a new RTX holding the result of moving POINTER forward by |
13882 | AMOUNT bytes. */ | |
13883 | ||
13884 | static rtx | |
13885 | aarch64_move_pointer (rtx pointer, int amount) | |
13886 | { | |
13887 | rtx next = plus_constant (Pmode, XEXP (pointer, 0), amount); | |
13888 | ||
13889 | return adjust_automodify_address (pointer, GET_MODE (pointer), | |
13890 | next, amount); | |
13891 | } | |
13892 | ||
13893 | /* Return a new RTX holding the result of moving POINTER forward by the | |
13894 | size of the mode it points to. */ | |
13895 | ||
13896 | static rtx | |
13897 | aarch64_progress_pointer (rtx pointer) | |
13898 | { | |
13899 | HOST_WIDE_INT amount = GET_MODE_SIZE (GET_MODE (pointer)); | |
13900 | ||
13901 | return aarch64_move_pointer (pointer, amount); | |
13902 | } | |
13903 | ||
13904 | /* Copy one MODE sized block from SRC to DST, then progress SRC and DST by | |
13905 | MODE bytes. */ | |
13906 | ||
13907 | static void | |
13908 | aarch64_copy_one_block_and_progress_pointers (rtx *src, rtx *dst, | |
ef4bddc2 | 13909 | machine_mode mode) |
e2c75eea JG |
13910 | { |
13911 | rtx reg = gen_reg_rtx (mode); | |
13912 | ||
13913 | /* "Cast" the pointers to the correct mode. */ | |
13914 | *src = adjust_address (*src, mode, 0); | |
13915 | *dst = adjust_address (*dst, mode, 0); | |
13916 | /* Emit the memcpy. */ | |
13917 | emit_move_insn (reg, *src); | |
13918 | emit_move_insn (*dst, reg); | |
13919 | /* Move the pointers forward. */ | |
13920 | *src = aarch64_progress_pointer (*src); | |
13921 | *dst = aarch64_progress_pointer (*dst); | |
13922 | } | |
13923 | ||
13924 | /* Expand movmem, as if from a __builtin_memcpy. Return true if | |
13925 | we succeed, otherwise return false. */ | |
13926 | ||
13927 | bool | |
13928 | aarch64_expand_movmem (rtx *operands) | |
13929 | { | |
13930 | unsigned int n; | |
13931 | rtx dst = operands[0]; | |
13932 | rtx src = operands[1]; | |
13933 | rtx base; | |
13934 | bool speed_p = !optimize_function_for_size_p (cfun); | |
13935 | ||
13936 | /* When optimizing for size, give a better estimate of the length of a | |
13937 | memcpy call, but use the default otherwise. */ | |
13938 | unsigned int max_instructions = (speed_p ? 15 : AARCH64_CALL_RATIO) / 2; | |
13939 | ||
13940 | /* We can't do anything smart if the amount to copy is not constant. */ | |
13941 | if (!CONST_INT_P (operands[2])) | |
13942 | return false; | |
13943 | ||
13944 | n = UINTVAL (operands[2]); | |
13945 | ||
13946 | /* Try to keep the number of instructions low. For cases below 16 bytes we | |
13947 | need to make at most two moves. For cases above 16 bytes it will be one | |
13948 | move for each 16 byte chunk, then at most two additional moves. */ | |
13949 | if (((n / 16) + (n % 16 ? 2 : 0)) > max_instructions) | |
13950 | return false; | |
13951 | ||
13952 | base = copy_to_mode_reg (Pmode, XEXP (dst, 0)); | |
13953 | dst = adjust_automodify_address (dst, VOIDmode, base, 0); | |
13954 | ||
13955 | base = copy_to_mode_reg (Pmode, XEXP (src, 0)); | |
13956 | src = adjust_automodify_address (src, VOIDmode, base, 0); | |
13957 | ||
13958 | /* Simple cases. Copy 0-3 bytes, as (if applicable) a 2-byte, then a | |
13959 | 1-byte chunk. */ | |
13960 | if (n < 4) | |
13961 | { | |
13962 | if (n >= 2) | |
13963 | { | |
13964 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, HImode); | |
13965 | n -= 2; | |
13966 | } | |
13967 | ||
13968 | if (n == 1) | |
13969 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, QImode); | |
13970 | ||
13971 | return true; | |
13972 | } | |
13973 | ||
13974 | /* Copy 4-8 bytes. First a 4-byte chunk, then (if applicable) a second | |
13975 | 4-byte chunk, partially overlapping with the previously copied chunk. */ | |
13976 | if (n < 8) | |
13977 | { | |
13978 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
13979 | n -= 4; | |
13980 | if (n > 0) | |
13981 | { | |
13982 | int move = n - 4; | |
13983 | ||
13984 | src = aarch64_move_pointer (src, move); | |
13985 | dst = aarch64_move_pointer (dst, move); | |
13986 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
13987 | } | |
13988 | return true; | |
13989 | } | |
13990 | ||
13991 | /* Copy more than 8 bytes. Copy chunks of 16 bytes until we run out of | |
13992 | them, then (if applicable) an 8-byte chunk. */ | |
13993 | while (n >= 8) | |
13994 | { | |
13995 | if (n / 16) | |
13996 | { | |
13997 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, TImode); | |
13998 | n -= 16; | |
13999 | } | |
14000 | else | |
14001 | { | |
14002 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, DImode); | |
14003 | n -= 8; | |
14004 | } | |
14005 | } | |
14006 | ||
14007 | /* Finish the final bytes of the copy. We can always do this in one | |
14008 | instruction. We either copy the exact amount we need, or partially | |
14009 | overlap with the previous chunk we copied and copy 8-bytes. */ | |
14010 | if (n == 0) | |
14011 | return true; | |
14012 | else if (n == 1) | |
14013 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, QImode); | |
14014 | else if (n == 2) | |
14015 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, HImode); | |
14016 | else if (n == 4) | |
14017 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
14018 | else | |
14019 | { | |
14020 | if (n == 3) | |
14021 | { | |
14022 | src = aarch64_move_pointer (src, -1); | |
14023 | dst = aarch64_move_pointer (dst, -1); | |
14024 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
14025 | } | |
14026 | else | |
14027 | { | |
14028 | int move = n - 8; | |
14029 | ||
14030 | src = aarch64_move_pointer (src, move); | |
14031 | dst = aarch64_move_pointer (dst, move); | |
14032 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, DImode); | |
14033 | } | |
14034 | } | |
14035 | ||
14036 | return true; | |
14037 | } | |
14038 | ||
141a3ccf KT |
14039 | /* Split a DImode store of a CONST_INT SRC to MEM DST as two |
14040 | SImode stores. Handle the case when the constant has identical | |
14041 | bottom and top halves. This is beneficial when the two stores can be | |
14042 | merged into an STP and we avoid synthesising potentially expensive | |
14043 | immediates twice. Return true if such a split is possible. */ | |
14044 | ||
14045 | bool | |
14046 | aarch64_split_dimode_const_store (rtx dst, rtx src) | |
14047 | { | |
14048 | rtx lo = gen_lowpart (SImode, src); | |
14049 | rtx hi = gen_highpart_mode (SImode, DImode, src); | |
14050 | ||
14051 | bool size_p = optimize_function_for_size_p (cfun); | |
14052 | ||
14053 | if (!rtx_equal_p (lo, hi)) | |
14054 | return false; | |
14055 | ||
14056 | unsigned int orig_cost | |
14057 | = aarch64_internal_mov_immediate (NULL_RTX, src, false, DImode); | |
14058 | unsigned int lo_cost | |
14059 | = aarch64_internal_mov_immediate (NULL_RTX, lo, false, SImode); | |
14060 | ||
14061 | /* We want to transform: | |
14062 | MOV x1, 49370 | |
14063 | MOVK x1, 0x140, lsl 16 | |
14064 | MOVK x1, 0xc0da, lsl 32 | |
14065 | MOVK x1, 0x140, lsl 48 | |
14066 | STR x1, [x0] | |
14067 | into: | |
14068 | MOV w1, 49370 | |
14069 | MOVK w1, 0x140, lsl 16 | |
14070 | STP w1, w1, [x0] | |
14071 | So we want to perform this only when we save two instructions | |
14072 | or more. When optimizing for size, however, accept any code size | |
14073 | savings we can. */ | |
14074 | if (size_p && orig_cost <= lo_cost) | |
14075 | return false; | |
14076 | ||
14077 | if (!size_p | |
14078 | && (orig_cost <= lo_cost + 1)) | |
14079 | return false; | |
14080 | ||
14081 | rtx mem_lo = adjust_address (dst, SImode, 0); | |
14082 | if (!aarch64_mem_pair_operand (mem_lo, SImode)) | |
14083 | return false; | |
14084 | ||
14085 | rtx tmp_reg = gen_reg_rtx (SImode); | |
14086 | aarch64_expand_mov_immediate (tmp_reg, lo); | |
14087 | rtx mem_hi = aarch64_move_pointer (mem_lo, GET_MODE_SIZE (SImode)); | |
14088 | /* Don't emit an explicit store pair as this may not be always profitable. | |
14089 | Let the sched-fusion logic decide whether to merge them. */ | |
14090 | emit_move_insn (mem_lo, tmp_reg); | |
14091 | emit_move_insn (mem_hi, tmp_reg); | |
14092 | ||
14093 | return true; | |
14094 | } | |
14095 | ||
a3125fc2 CL |
14096 | /* Implement the TARGET_ASAN_SHADOW_OFFSET hook. */ |
14097 | ||
14098 | static unsigned HOST_WIDE_INT | |
14099 | aarch64_asan_shadow_offset (void) | |
14100 | { | |
14101 | return (HOST_WIDE_INT_1 << 36); | |
14102 | } | |
14103 | ||
d3006da6 | 14104 | static bool |
445d7826 | 14105 | aarch64_use_by_pieces_infrastructure_p (unsigned HOST_WIDE_INT size, |
d3006da6 JG |
14106 | unsigned int align, |
14107 | enum by_pieces_operation op, | |
14108 | bool speed_p) | |
14109 | { | |
14110 | /* STORE_BY_PIECES can be used when copying a constant string, but | |
14111 | in that case each 64-bit chunk takes 5 insns instead of 2 (LDR/STR). | |
14112 | For now we always fail this and let the move_by_pieces code copy | |
14113 | the string from read-only memory. */ | |
14114 | if (op == STORE_BY_PIECES) | |
14115 | return false; | |
14116 | ||
14117 | return default_use_by_pieces_infrastructure_p (size, align, op, speed_p); | |
14118 | } | |
14119 | ||
5f3bc026 | 14120 | static rtx |
cb4347e8 | 14121 | aarch64_gen_ccmp_first (rtx_insn **prep_seq, rtx_insn **gen_seq, |
5f3bc026 ZC |
14122 | int code, tree treeop0, tree treeop1) |
14123 | { | |
c8012fbc WD |
14124 | machine_mode op_mode, cmp_mode, cc_mode = CCmode; |
14125 | rtx op0, op1; | |
5f3bc026 | 14126 | int unsignedp = TYPE_UNSIGNED (TREE_TYPE (treeop0)); |
c8012fbc | 14127 | insn_code icode; |
5f3bc026 ZC |
14128 | struct expand_operand ops[4]; |
14129 | ||
5f3bc026 ZC |
14130 | start_sequence (); |
14131 | expand_operands (treeop0, treeop1, NULL_RTX, &op0, &op1, EXPAND_NORMAL); | |
14132 | ||
14133 | op_mode = GET_MODE (op0); | |
14134 | if (op_mode == VOIDmode) | |
14135 | op_mode = GET_MODE (op1); | |
14136 | ||
14137 | switch (op_mode) | |
14138 | { | |
4e10a5a7 RS |
14139 | case E_QImode: |
14140 | case E_HImode: | |
14141 | case E_SImode: | |
5f3bc026 ZC |
14142 | cmp_mode = SImode; |
14143 | icode = CODE_FOR_cmpsi; | |
14144 | break; | |
14145 | ||
4e10a5a7 | 14146 | case E_DImode: |
5f3bc026 ZC |
14147 | cmp_mode = DImode; |
14148 | icode = CODE_FOR_cmpdi; | |
14149 | break; | |
14150 | ||
4e10a5a7 | 14151 | case E_SFmode: |
786e3c06 WD |
14152 | cmp_mode = SFmode; |
14153 | cc_mode = aarch64_select_cc_mode ((rtx_code) code, op0, op1); | |
14154 | icode = cc_mode == CCFPEmode ? CODE_FOR_fcmpesf : CODE_FOR_fcmpsf; | |
14155 | break; | |
14156 | ||
4e10a5a7 | 14157 | case E_DFmode: |
786e3c06 WD |
14158 | cmp_mode = DFmode; |
14159 | cc_mode = aarch64_select_cc_mode ((rtx_code) code, op0, op1); | |
14160 | icode = cc_mode == CCFPEmode ? CODE_FOR_fcmpedf : CODE_FOR_fcmpdf; | |
14161 | break; | |
14162 | ||
5f3bc026 ZC |
14163 | default: |
14164 | end_sequence (); | |
14165 | return NULL_RTX; | |
14166 | } | |
14167 | ||
c8012fbc WD |
14168 | op0 = prepare_operand (icode, op0, 0, op_mode, cmp_mode, unsignedp); |
14169 | op1 = prepare_operand (icode, op1, 1, op_mode, cmp_mode, unsignedp); | |
5f3bc026 ZC |
14170 | if (!op0 || !op1) |
14171 | { | |
14172 | end_sequence (); | |
14173 | return NULL_RTX; | |
14174 | } | |
14175 | *prep_seq = get_insns (); | |
14176 | end_sequence (); | |
14177 | ||
c8012fbc WD |
14178 | create_fixed_operand (&ops[0], op0); |
14179 | create_fixed_operand (&ops[1], op1); | |
5f3bc026 ZC |
14180 | |
14181 | start_sequence (); | |
c8012fbc | 14182 | if (!maybe_expand_insn (icode, 2, ops)) |
5f3bc026 ZC |
14183 | { |
14184 | end_sequence (); | |
14185 | return NULL_RTX; | |
14186 | } | |
14187 | *gen_seq = get_insns (); | |
14188 | end_sequence (); | |
14189 | ||
c8012fbc WD |
14190 | return gen_rtx_fmt_ee ((rtx_code) code, cc_mode, |
14191 | gen_rtx_REG (cc_mode, CC_REGNUM), const0_rtx); | |
5f3bc026 ZC |
14192 | } |
14193 | ||
14194 | static rtx | |
cb4347e8 TS |
14195 | aarch64_gen_ccmp_next (rtx_insn **prep_seq, rtx_insn **gen_seq, rtx prev, |
14196 | int cmp_code, tree treeop0, tree treeop1, int bit_code) | |
5f3bc026 | 14197 | { |
c8012fbc WD |
14198 | rtx op0, op1, target; |
14199 | machine_mode op_mode, cmp_mode, cc_mode = CCmode; | |
5f3bc026 | 14200 | int unsignedp = TYPE_UNSIGNED (TREE_TYPE (treeop0)); |
c8012fbc | 14201 | insn_code icode; |
5f3bc026 | 14202 | struct expand_operand ops[6]; |
c8012fbc | 14203 | int aarch64_cond; |
5f3bc026 | 14204 | |
cb4347e8 | 14205 | push_to_sequence (*prep_seq); |
5f3bc026 ZC |
14206 | expand_operands (treeop0, treeop1, NULL_RTX, &op0, &op1, EXPAND_NORMAL); |
14207 | ||
14208 | op_mode = GET_MODE (op0); | |
14209 | if (op_mode == VOIDmode) | |
14210 | op_mode = GET_MODE (op1); | |
14211 | ||
14212 | switch (op_mode) | |
14213 | { | |
4e10a5a7 RS |
14214 | case E_QImode: |
14215 | case E_HImode: | |
14216 | case E_SImode: | |
5f3bc026 | 14217 | cmp_mode = SImode; |
c8012fbc | 14218 | icode = CODE_FOR_ccmpsi; |
5f3bc026 ZC |
14219 | break; |
14220 | ||
4e10a5a7 | 14221 | case E_DImode: |
5f3bc026 | 14222 | cmp_mode = DImode; |
c8012fbc | 14223 | icode = CODE_FOR_ccmpdi; |
5f3bc026 ZC |
14224 | break; |
14225 | ||
4e10a5a7 | 14226 | case E_SFmode: |
786e3c06 WD |
14227 | cmp_mode = SFmode; |
14228 | cc_mode = aarch64_select_cc_mode ((rtx_code) cmp_code, op0, op1); | |
14229 | icode = cc_mode == CCFPEmode ? CODE_FOR_fccmpesf : CODE_FOR_fccmpsf; | |
14230 | break; | |
14231 | ||
4e10a5a7 | 14232 | case E_DFmode: |
786e3c06 WD |
14233 | cmp_mode = DFmode; |
14234 | cc_mode = aarch64_select_cc_mode ((rtx_code) cmp_code, op0, op1); | |
14235 | icode = cc_mode == CCFPEmode ? CODE_FOR_fccmpedf : CODE_FOR_fccmpdf; | |
14236 | break; | |
14237 | ||
5f3bc026 ZC |
14238 | default: |
14239 | end_sequence (); | |
14240 | return NULL_RTX; | |
14241 | } | |
14242 | ||
14243 | op0 = prepare_operand (icode, op0, 2, op_mode, cmp_mode, unsignedp); | |
14244 | op1 = prepare_operand (icode, op1, 3, op_mode, cmp_mode, unsignedp); | |
14245 | if (!op0 || !op1) | |
14246 | { | |
14247 | end_sequence (); | |
14248 | return NULL_RTX; | |
14249 | } | |
14250 | *prep_seq = get_insns (); | |
14251 | end_sequence (); | |
14252 | ||
14253 | target = gen_rtx_REG (cc_mode, CC_REGNUM); | |
c8012fbc | 14254 | aarch64_cond = aarch64_get_condition_code_1 (cc_mode, (rtx_code) cmp_code); |
5f3bc026 | 14255 | |
c8012fbc WD |
14256 | if (bit_code != AND) |
14257 | { | |
14258 | prev = gen_rtx_fmt_ee (REVERSE_CONDITION (GET_CODE (prev), | |
14259 | GET_MODE (XEXP (prev, 0))), | |
14260 | VOIDmode, XEXP (prev, 0), const0_rtx); | |
14261 | aarch64_cond = AARCH64_INVERSE_CONDITION_CODE (aarch64_cond); | |
14262 | } | |
14263 | ||
14264 | create_fixed_operand (&ops[0], XEXP (prev, 0)); | |
5f3bc026 ZC |
14265 | create_fixed_operand (&ops[1], target); |
14266 | create_fixed_operand (&ops[2], op0); | |
14267 | create_fixed_operand (&ops[3], op1); | |
c8012fbc WD |
14268 | create_fixed_operand (&ops[4], prev); |
14269 | create_fixed_operand (&ops[5], GEN_INT (aarch64_cond)); | |
5f3bc026 | 14270 | |
cb4347e8 | 14271 | push_to_sequence (*gen_seq); |
5f3bc026 ZC |
14272 | if (!maybe_expand_insn (icode, 6, ops)) |
14273 | { | |
14274 | end_sequence (); | |
14275 | return NULL_RTX; | |
14276 | } | |
14277 | ||
14278 | *gen_seq = get_insns (); | |
14279 | end_sequence (); | |
14280 | ||
c8012fbc | 14281 | return gen_rtx_fmt_ee ((rtx_code) cmp_code, VOIDmode, target, const0_rtx); |
5f3bc026 ZC |
14282 | } |
14283 | ||
14284 | #undef TARGET_GEN_CCMP_FIRST | |
14285 | #define TARGET_GEN_CCMP_FIRST aarch64_gen_ccmp_first | |
14286 | ||
14287 | #undef TARGET_GEN_CCMP_NEXT | |
14288 | #define TARGET_GEN_CCMP_NEXT aarch64_gen_ccmp_next | |
14289 | ||
6a569cdd KT |
14290 | /* Implement TARGET_SCHED_MACRO_FUSION_P. Return true if target supports |
14291 | instruction fusion of some sort. */ | |
14292 | ||
14293 | static bool | |
14294 | aarch64_macro_fusion_p (void) | |
14295 | { | |
b175b679 | 14296 | return aarch64_tune_params.fusible_ops != AARCH64_FUSE_NOTHING; |
6a569cdd KT |
14297 | } |
14298 | ||
14299 | ||
14300 | /* Implement TARGET_SCHED_MACRO_FUSION_PAIR_P. Return true if PREV and CURR | |
14301 | should be kept together during scheduling. */ | |
14302 | ||
14303 | static bool | |
14304 | aarch_macro_fusion_pair_p (rtx_insn *prev, rtx_insn *curr) | |
14305 | { | |
14306 | rtx set_dest; | |
14307 | rtx prev_set = single_set (prev); | |
14308 | rtx curr_set = single_set (curr); | |
14309 | /* prev and curr are simple SET insns i.e. no flag setting or branching. */ | |
14310 | bool simple_sets_p = prev_set && curr_set && !any_condjump_p (curr); | |
14311 | ||
14312 | if (!aarch64_macro_fusion_p ()) | |
14313 | return false; | |
14314 | ||
d7b03373 | 14315 | if (simple_sets_p && aarch64_fusion_enabled_p (AARCH64_FUSE_MOV_MOVK)) |
6a569cdd KT |
14316 | { |
14317 | /* We are trying to match: | |
14318 | prev (mov) == (set (reg r0) (const_int imm16)) | |
14319 | curr (movk) == (set (zero_extract (reg r0) | |
14320 | (const_int 16) | |
14321 | (const_int 16)) | |
14322 | (const_int imm16_1)) */ | |
14323 | ||
14324 | set_dest = SET_DEST (curr_set); | |
14325 | ||
14326 | if (GET_CODE (set_dest) == ZERO_EXTRACT | |
14327 | && CONST_INT_P (SET_SRC (curr_set)) | |
14328 | && CONST_INT_P (SET_SRC (prev_set)) | |
14329 | && CONST_INT_P (XEXP (set_dest, 2)) | |
14330 | && INTVAL (XEXP (set_dest, 2)) == 16 | |
14331 | && REG_P (XEXP (set_dest, 0)) | |
14332 | && REG_P (SET_DEST (prev_set)) | |
14333 | && REGNO (XEXP (set_dest, 0)) == REGNO (SET_DEST (prev_set))) | |
14334 | { | |
14335 | return true; | |
14336 | } | |
14337 | } | |
14338 | ||
d7b03373 | 14339 | if (simple_sets_p && aarch64_fusion_enabled_p (AARCH64_FUSE_ADRP_ADD)) |
9bbe08fe KT |
14340 | { |
14341 | ||
14342 | /* We're trying to match: | |
14343 | prev (adrp) == (set (reg r1) | |
14344 | (high (symbol_ref ("SYM")))) | |
14345 | curr (add) == (set (reg r0) | |
14346 | (lo_sum (reg r1) | |
14347 | (symbol_ref ("SYM")))) | |
14348 | Note that r0 need not necessarily be the same as r1, especially | |
14349 | during pre-regalloc scheduling. */ | |
14350 | ||
14351 | if (satisfies_constraint_Ush (SET_SRC (prev_set)) | |
14352 | && REG_P (SET_DEST (prev_set)) && REG_P (SET_DEST (curr_set))) | |
14353 | { | |
14354 | if (GET_CODE (SET_SRC (curr_set)) == LO_SUM | |
14355 | && REG_P (XEXP (SET_SRC (curr_set), 0)) | |
14356 | && REGNO (XEXP (SET_SRC (curr_set), 0)) | |
14357 | == REGNO (SET_DEST (prev_set)) | |
14358 | && rtx_equal_p (XEXP (SET_SRC (prev_set), 0), | |
14359 | XEXP (SET_SRC (curr_set), 1))) | |
14360 | return true; | |
14361 | } | |
14362 | } | |
14363 | ||
d7b03373 | 14364 | if (simple_sets_p && aarch64_fusion_enabled_p (AARCH64_FUSE_MOVK_MOVK)) |
cd0cb232 KT |
14365 | { |
14366 | ||
14367 | /* We're trying to match: | |
14368 | prev (movk) == (set (zero_extract (reg r0) | |
14369 | (const_int 16) | |
14370 | (const_int 32)) | |
14371 | (const_int imm16_1)) | |
14372 | curr (movk) == (set (zero_extract (reg r0) | |
14373 | (const_int 16) | |
14374 | (const_int 48)) | |
14375 | (const_int imm16_2)) */ | |
14376 | ||
14377 | if (GET_CODE (SET_DEST (prev_set)) == ZERO_EXTRACT | |
14378 | && GET_CODE (SET_DEST (curr_set)) == ZERO_EXTRACT | |
14379 | && REG_P (XEXP (SET_DEST (prev_set), 0)) | |
14380 | && REG_P (XEXP (SET_DEST (curr_set), 0)) | |
14381 | && REGNO (XEXP (SET_DEST (prev_set), 0)) | |
14382 | == REGNO (XEXP (SET_DEST (curr_set), 0)) | |
14383 | && CONST_INT_P (XEXP (SET_DEST (prev_set), 2)) | |
14384 | && CONST_INT_P (XEXP (SET_DEST (curr_set), 2)) | |
14385 | && INTVAL (XEXP (SET_DEST (prev_set), 2)) == 32 | |
14386 | && INTVAL (XEXP (SET_DEST (curr_set), 2)) == 48 | |
14387 | && CONST_INT_P (SET_SRC (prev_set)) | |
14388 | && CONST_INT_P (SET_SRC (curr_set))) | |
14389 | return true; | |
14390 | ||
14391 | } | |
d7b03373 | 14392 | if (simple_sets_p && aarch64_fusion_enabled_p (AARCH64_FUSE_ADRP_LDR)) |
d8354ad7 KT |
14393 | { |
14394 | /* We're trying to match: | |
14395 | prev (adrp) == (set (reg r0) | |
14396 | (high (symbol_ref ("SYM")))) | |
14397 | curr (ldr) == (set (reg r1) | |
14398 | (mem (lo_sum (reg r0) | |
14399 | (symbol_ref ("SYM"))))) | |
14400 | or | |
14401 | curr (ldr) == (set (reg r1) | |
14402 | (zero_extend (mem | |
14403 | (lo_sum (reg r0) | |
14404 | (symbol_ref ("SYM")))))) */ | |
14405 | if (satisfies_constraint_Ush (SET_SRC (prev_set)) | |
14406 | && REG_P (SET_DEST (prev_set)) && REG_P (SET_DEST (curr_set))) | |
14407 | { | |
14408 | rtx curr_src = SET_SRC (curr_set); | |
14409 | ||
14410 | if (GET_CODE (curr_src) == ZERO_EXTEND) | |
14411 | curr_src = XEXP (curr_src, 0); | |
14412 | ||
14413 | if (MEM_P (curr_src) && GET_CODE (XEXP (curr_src, 0)) == LO_SUM | |
14414 | && REG_P (XEXP (XEXP (curr_src, 0), 0)) | |
14415 | && REGNO (XEXP (XEXP (curr_src, 0), 0)) | |
14416 | == REGNO (SET_DEST (prev_set)) | |
14417 | && rtx_equal_p (XEXP (XEXP (curr_src, 0), 1), | |
14418 | XEXP (SET_SRC (prev_set), 0))) | |
14419 | return true; | |
14420 | } | |
14421 | } | |
cd0cb232 | 14422 | |
d7b03373 | 14423 | if (aarch64_fusion_enabled_p (AARCH64_FUSE_AES_AESMC) |
00a8574a WD |
14424 | && aarch_crypto_can_dual_issue (prev, curr)) |
14425 | return true; | |
14426 | ||
d7b03373 | 14427 | if (aarch64_fusion_enabled_p (AARCH64_FUSE_CMP_BRANCH) |
3759108f AP |
14428 | && any_condjump_p (curr)) |
14429 | { | |
14430 | enum attr_type prev_type = get_attr_type (prev); | |
14431 | ||
509f819a N |
14432 | unsigned int condreg1, condreg2; |
14433 | rtx cc_reg_1; | |
14434 | aarch64_fixed_condition_code_regs (&condreg1, &condreg2); | |
14435 | cc_reg_1 = gen_rtx_REG (CCmode, condreg1); | |
14436 | ||
14437 | if (reg_referenced_p (cc_reg_1, PATTERN (curr)) | |
14438 | && prev | |
14439 | && modified_in_p (cc_reg_1, prev)) | |
14440 | { | |
14441 | /* FIXME: this misses some which is considered simple arthematic | |
14442 | instructions for ThunderX. Simple shifts are missed here. */ | |
14443 | if (prev_type == TYPE_ALUS_SREG | |
14444 | || prev_type == TYPE_ALUS_IMM | |
14445 | || prev_type == TYPE_LOGICS_REG | |
14446 | || prev_type == TYPE_LOGICS_IMM) | |
14447 | return true; | |
14448 | } | |
3759108f AP |
14449 | } |
14450 | ||
bee7e0fc AP |
14451 | if (prev_set |
14452 | && curr_set | |
14453 | && aarch64_fusion_enabled_p (AARCH64_FUSE_ALU_BRANCH) | |
00c7c57f JB |
14454 | && any_condjump_p (curr)) |
14455 | { | |
14456 | /* We're trying to match: | |
14457 | prev (alu_insn) == (set (r0) plus ((r0) (r1/imm))) | |
14458 | curr (cbz) == (set (pc) (if_then_else (eq/ne) (r0) | |
14459 | (const_int 0)) | |
14460 | (label_ref ("SYM")) | |
14461 | (pc)) */ | |
14462 | if (SET_DEST (curr_set) == (pc_rtx) | |
14463 | && GET_CODE (SET_SRC (curr_set)) == IF_THEN_ELSE | |
14464 | && REG_P (XEXP (XEXP (SET_SRC (curr_set), 0), 0)) | |
14465 | && REG_P (SET_DEST (prev_set)) | |
14466 | && REGNO (SET_DEST (prev_set)) | |
14467 | == REGNO (XEXP (XEXP (SET_SRC (curr_set), 0), 0))) | |
14468 | { | |
14469 | /* Fuse ALU operations followed by conditional branch instruction. */ | |
14470 | switch (get_attr_type (prev)) | |
14471 | { | |
14472 | case TYPE_ALU_IMM: | |
14473 | case TYPE_ALU_SREG: | |
14474 | case TYPE_ADC_REG: | |
14475 | case TYPE_ADC_IMM: | |
14476 | case TYPE_ADCS_REG: | |
14477 | case TYPE_ADCS_IMM: | |
14478 | case TYPE_LOGIC_REG: | |
14479 | case TYPE_LOGIC_IMM: | |
14480 | case TYPE_CSEL: | |
14481 | case TYPE_ADR: | |
14482 | case TYPE_MOV_IMM: | |
14483 | case TYPE_SHIFT_REG: | |
14484 | case TYPE_SHIFT_IMM: | |
14485 | case TYPE_BFM: | |
14486 | case TYPE_RBIT: | |
14487 | case TYPE_REV: | |
14488 | case TYPE_EXTEND: | |
14489 | return true; | |
14490 | ||
14491 | default:; | |
14492 | } | |
14493 | } | |
14494 | } | |
14495 | ||
6a569cdd KT |
14496 | return false; |
14497 | } | |
14498 | ||
f2879a90 KT |
14499 | /* Return true iff the instruction fusion described by OP is enabled. */ |
14500 | ||
14501 | bool | |
14502 | aarch64_fusion_enabled_p (enum aarch64_fusion_pairs op) | |
14503 | { | |
14504 | return (aarch64_tune_params.fusible_ops & op) != 0; | |
14505 | } | |
14506 | ||
350013bc BC |
14507 | /* If MEM is in the form of [base+offset], extract the two parts |
14508 | of address and set to BASE and OFFSET, otherwise return false | |
14509 | after clearing BASE and OFFSET. */ | |
14510 | ||
14511 | bool | |
14512 | extract_base_offset_in_addr (rtx mem, rtx *base, rtx *offset) | |
14513 | { | |
14514 | rtx addr; | |
14515 | ||
14516 | gcc_assert (MEM_P (mem)); | |
14517 | ||
14518 | addr = XEXP (mem, 0); | |
14519 | ||
14520 | if (REG_P (addr)) | |
14521 | { | |
14522 | *base = addr; | |
14523 | *offset = const0_rtx; | |
14524 | return true; | |
14525 | } | |
14526 | ||
14527 | if (GET_CODE (addr) == PLUS | |
14528 | && REG_P (XEXP (addr, 0)) && CONST_INT_P (XEXP (addr, 1))) | |
14529 | { | |
14530 | *base = XEXP (addr, 0); | |
14531 | *offset = XEXP (addr, 1); | |
14532 | return true; | |
14533 | } | |
14534 | ||
14535 | *base = NULL_RTX; | |
14536 | *offset = NULL_RTX; | |
14537 | ||
14538 | return false; | |
14539 | } | |
14540 | ||
14541 | /* Types for scheduling fusion. */ | |
14542 | enum sched_fusion_type | |
14543 | { | |
14544 | SCHED_FUSION_NONE = 0, | |
14545 | SCHED_FUSION_LD_SIGN_EXTEND, | |
14546 | SCHED_FUSION_LD_ZERO_EXTEND, | |
14547 | SCHED_FUSION_LD, | |
14548 | SCHED_FUSION_ST, | |
14549 | SCHED_FUSION_NUM | |
14550 | }; | |
14551 | ||
14552 | /* If INSN is a load or store of address in the form of [base+offset], | |
14553 | extract the two parts and set to BASE and OFFSET. Return scheduling | |
14554 | fusion type this INSN is. */ | |
14555 | ||
14556 | static enum sched_fusion_type | |
14557 | fusion_load_store (rtx_insn *insn, rtx *base, rtx *offset) | |
14558 | { | |
14559 | rtx x, dest, src; | |
14560 | enum sched_fusion_type fusion = SCHED_FUSION_LD; | |
14561 | ||
14562 | gcc_assert (INSN_P (insn)); | |
14563 | x = PATTERN (insn); | |
14564 | if (GET_CODE (x) != SET) | |
14565 | return SCHED_FUSION_NONE; | |
14566 | ||
14567 | src = SET_SRC (x); | |
14568 | dest = SET_DEST (x); | |
14569 | ||
abc52318 KT |
14570 | machine_mode dest_mode = GET_MODE (dest); |
14571 | ||
14572 | if (!aarch64_mode_valid_for_sched_fusion_p (dest_mode)) | |
350013bc BC |
14573 | return SCHED_FUSION_NONE; |
14574 | ||
14575 | if (GET_CODE (src) == SIGN_EXTEND) | |
14576 | { | |
14577 | fusion = SCHED_FUSION_LD_SIGN_EXTEND; | |
14578 | src = XEXP (src, 0); | |
14579 | if (GET_CODE (src) != MEM || GET_MODE (src) != SImode) | |
14580 | return SCHED_FUSION_NONE; | |
14581 | } | |
14582 | else if (GET_CODE (src) == ZERO_EXTEND) | |
14583 | { | |
14584 | fusion = SCHED_FUSION_LD_ZERO_EXTEND; | |
14585 | src = XEXP (src, 0); | |
14586 | if (GET_CODE (src) != MEM || GET_MODE (src) != SImode) | |
14587 | return SCHED_FUSION_NONE; | |
14588 | } | |
14589 | ||
14590 | if (GET_CODE (src) == MEM && REG_P (dest)) | |
14591 | extract_base_offset_in_addr (src, base, offset); | |
14592 | else if (GET_CODE (dest) == MEM && (REG_P (src) || src == const0_rtx)) | |
14593 | { | |
14594 | fusion = SCHED_FUSION_ST; | |
14595 | extract_base_offset_in_addr (dest, base, offset); | |
14596 | } | |
14597 | else | |
14598 | return SCHED_FUSION_NONE; | |
14599 | ||
14600 | if (*base == NULL_RTX || *offset == NULL_RTX) | |
14601 | fusion = SCHED_FUSION_NONE; | |
14602 | ||
14603 | return fusion; | |
14604 | } | |
14605 | ||
14606 | /* Implement the TARGET_SCHED_FUSION_PRIORITY hook. | |
14607 | ||
14608 | Currently we only support to fuse ldr or str instructions, so FUSION_PRI | |
14609 | and PRI are only calculated for these instructions. For other instruction, | |
14610 | FUSION_PRI and PRI are simply set to MAX_PRI - 1. In the future, other | |
14611 | type instruction fusion can be added by returning different priorities. | |
14612 | ||
14613 | It's important that irrelevant instructions get the largest FUSION_PRI. */ | |
14614 | ||
14615 | static void | |
14616 | aarch64_sched_fusion_priority (rtx_insn *insn, int max_pri, | |
14617 | int *fusion_pri, int *pri) | |
14618 | { | |
14619 | int tmp, off_val; | |
14620 | rtx base, offset; | |
14621 | enum sched_fusion_type fusion; | |
14622 | ||
14623 | gcc_assert (INSN_P (insn)); | |
14624 | ||
14625 | tmp = max_pri - 1; | |
14626 | fusion = fusion_load_store (insn, &base, &offset); | |
14627 | if (fusion == SCHED_FUSION_NONE) | |
14628 | { | |
14629 | *pri = tmp; | |
14630 | *fusion_pri = tmp; | |
14631 | return; | |
14632 | } | |
14633 | ||
14634 | /* Set FUSION_PRI according to fusion type and base register. */ | |
14635 | *fusion_pri = tmp - fusion * FIRST_PSEUDO_REGISTER - REGNO (base); | |
14636 | ||
14637 | /* Calculate PRI. */ | |
14638 | tmp /= 2; | |
14639 | ||
14640 | /* INSN with smaller offset goes first. */ | |
14641 | off_val = (int)(INTVAL (offset)); | |
14642 | if (off_val >= 0) | |
14643 | tmp -= (off_val & 0xfffff); | |
14644 | else | |
14645 | tmp += ((- off_val) & 0xfffff); | |
14646 | ||
14647 | *pri = tmp; | |
14648 | return; | |
14649 | } | |
14650 | ||
9bca63d4 WD |
14651 | /* Implement the TARGET_SCHED_ADJUST_PRIORITY hook. |
14652 | Adjust priority of sha1h instructions so they are scheduled before | |
14653 | other SHA1 instructions. */ | |
14654 | ||
14655 | static int | |
14656 | aarch64_sched_adjust_priority (rtx_insn *insn, int priority) | |
14657 | { | |
14658 | rtx x = PATTERN (insn); | |
14659 | ||
14660 | if (GET_CODE (x) == SET) | |
14661 | { | |
14662 | x = SET_SRC (x); | |
14663 | ||
14664 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_SHA1H) | |
14665 | return priority + 10; | |
14666 | } | |
14667 | ||
14668 | return priority; | |
14669 | } | |
14670 | ||
350013bc BC |
14671 | /* Given OPERANDS of consecutive load/store, check if we can merge |
14672 | them into ldp/stp. LOAD is true if they are load instructions. | |
14673 | MODE is the mode of memory operands. */ | |
14674 | ||
14675 | bool | |
14676 | aarch64_operands_ok_for_ldpstp (rtx *operands, bool load, | |
b8506a8a | 14677 | machine_mode mode) |
350013bc BC |
14678 | { |
14679 | HOST_WIDE_INT offval_1, offval_2, msize; | |
14680 | enum reg_class rclass_1, rclass_2; | |
14681 | rtx mem_1, mem_2, reg_1, reg_2, base_1, base_2, offset_1, offset_2; | |
14682 | ||
14683 | if (load) | |
14684 | { | |
14685 | mem_1 = operands[1]; | |
14686 | mem_2 = operands[3]; | |
14687 | reg_1 = operands[0]; | |
14688 | reg_2 = operands[2]; | |
14689 | gcc_assert (REG_P (reg_1) && REG_P (reg_2)); | |
14690 | if (REGNO (reg_1) == REGNO (reg_2)) | |
14691 | return false; | |
14692 | } | |
14693 | else | |
14694 | { | |
14695 | mem_1 = operands[0]; | |
14696 | mem_2 = operands[2]; | |
14697 | reg_1 = operands[1]; | |
14698 | reg_2 = operands[3]; | |
14699 | } | |
14700 | ||
bf84ac44 AP |
14701 | /* The mems cannot be volatile. */ |
14702 | if (MEM_VOLATILE_P (mem_1) || MEM_VOLATILE_P (mem_2)) | |
14703 | return false; | |
14704 | ||
54700e2e AP |
14705 | /* If we have SImode and slow unaligned ldp, |
14706 | check the alignment to be at least 8 byte. */ | |
14707 | if (mode == SImode | |
14708 | && (aarch64_tune_params.extra_tuning_flags | |
14709 | & AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW) | |
14710 | && !optimize_size | |
14711 | && MEM_ALIGN (mem_1) < 8 * BITS_PER_UNIT) | |
14712 | return false; | |
14713 | ||
350013bc BC |
14714 | /* Check if the addresses are in the form of [base+offset]. */ |
14715 | extract_base_offset_in_addr (mem_1, &base_1, &offset_1); | |
14716 | if (base_1 == NULL_RTX || offset_1 == NULL_RTX) | |
14717 | return false; | |
14718 | extract_base_offset_in_addr (mem_2, &base_2, &offset_2); | |
14719 | if (base_2 == NULL_RTX || offset_2 == NULL_RTX) | |
14720 | return false; | |
14721 | ||
14722 | /* Check if the bases are same. */ | |
14723 | if (!rtx_equal_p (base_1, base_2)) | |
14724 | return false; | |
14725 | ||
14726 | offval_1 = INTVAL (offset_1); | |
14727 | offval_2 = INTVAL (offset_2); | |
14728 | msize = GET_MODE_SIZE (mode); | |
14729 | /* Check if the offsets are consecutive. */ | |
14730 | if (offval_1 != (offval_2 + msize) && offval_2 != (offval_1 + msize)) | |
14731 | return false; | |
14732 | ||
14733 | /* Check if the addresses are clobbered by load. */ | |
14734 | if (load) | |
14735 | { | |
14736 | if (reg_mentioned_p (reg_1, mem_1)) | |
14737 | return false; | |
14738 | ||
14739 | /* In increasing order, the last load can clobber the address. */ | |
14740 | if (offval_1 > offval_2 && reg_mentioned_p (reg_2, mem_2)) | |
14741 | return false; | |
14742 | } | |
14743 | ||
14744 | if (REG_P (reg_1) && FP_REGNUM_P (REGNO (reg_1))) | |
14745 | rclass_1 = FP_REGS; | |
14746 | else | |
14747 | rclass_1 = GENERAL_REGS; | |
14748 | ||
14749 | if (REG_P (reg_2) && FP_REGNUM_P (REGNO (reg_2))) | |
14750 | rclass_2 = FP_REGS; | |
14751 | else | |
14752 | rclass_2 = GENERAL_REGS; | |
14753 | ||
14754 | /* Check if the registers are of same class. */ | |
14755 | if (rclass_1 != rclass_2) | |
14756 | return false; | |
14757 | ||
14758 | return true; | |
14759 | } | |
14760 | ||
14761 | /* Given OPERANDS of consecutive load/store, check if we can merge | |
14762 | them into ldp/stp by adjusting the offset. LOAD is true if they | |
14763 | are load instructions. MODE is the mode of memory operands. | |
14764 | ||
14765 | Given below consecutive stores: | |
14766 | ||
14767 | str w1, [xb, 0x100] | |
14768 | str w1, [xb, 0x104] | |
14769 | str w1, [xb, 0x108] | |
14770 | str w1, [xb, 0x10c] | |
14771 | ||
14772 | Though the offsets are out of the range supported by stp, we can | |
14773 | still pair them after adjusting the offset, like: | |
14774 | ||
14775 | add scratch, xb, 0x100 | |
14776 | stp w1, w1, [scratch] | |
14777 | stp w1, w1, [scratch, 0x8] | |
14778 | ||
14779 | The peephole patterns detecting this opportunity should guarantee | |
14780 | the scratch register is avaliable. */ | |
14781 | ||
14782 | bool | |
14783 | aarch64_operands_adjust_ok_for_ldpstp (rtx *operands, bool load, | |
b8506a8a | 14784 | machine_mode mode) |
350013bc BC |
14785 | { |
14786 | enum reg_class rclass_1, rclass_2, rclass_3, rclass_4; | |
14787 | HOST_WIDE_INT offval_1, offval_2, offval_3, offval_4, msize; | |
14788 | rtx mem_1, mem_2, mem_3, mem_4, reg_1, reg_2, reg_3, reg_4; | |
14789 | rtx base_1, base_2, base_3, base_4, offset_1, offset_2, offset_3, offset_4; | |
14790 | ||
14791 | if (load) | |
14792 | { | |
14793 | reg_1 = operands[0]; | |
14794 | mem_1 = operands[1]; | |
14795 | reg_2 = operands[2]; | |
14796 | mem_2 = operands[3]; | |
14797 | reg_3 = operands[4]; | |
14798 | mem_3 = operands[5]; | |
14799 | reg_4 = operands[6]; | |
14800 | mem_4 = operands[7]; | |
14801 | gcc_assert (REG_P (reg_1) && REG_P (reg_2) | |
14802 | && REG_P (reg_3) && REG_P (reg_4)); | |
14803 | if (REGNO (reg_1) == REGNO (reg_2) || REGNO (reg_3) == REGNO (reg_4)) | |
14804 | return false; | |
14805 | } | |
14806 | else | |
14807 | { | |
14808 | mem_1 = operands[0]; | |
14809 | reg_1 = operands[1]; | |
14810 | mem_2 = operands[2]; | |
14811 | reg_2 = operands[3]; | |
14812 | mem_3 = operands[4]; | |
14813 | reg_3 = operands[5]; | |
14814 | mem_4 = operands[6]; | |
14815 | reg_4 = operands[7]; | |
14816 | } | |
14817 | /* Skip if memory operand is by itslef valid for ldp/stp. */ | |
14818 | if (!MEM_P (mem_1) || aarch64_mem_pair_operand (mem_1, mode)) | |
14819 | return false; | |
14820 | ||
bf84ac44 AP |
14821 | /* The mems cannot be volatile. */ |
14822 | if (MEM_VOLATILE_P (mem_1) || MEM_VOLATILE_P (mem_2) | |
14823 | || MEM_VOLATILE_P (mem_3) ||MEM_VOLATILE_P (mem_4)) | |
14824 | return false; | |
14825 | ||
350013bc BC |
14826 | /* Check if the addresses are in the form of [base+offset]. */ |
14827 | extract_base_offset_in_addr (mem_1, &base_1, &offset_1); | |
14828 | if (base_1 == NULL_RTX || offset_1 == NULL_RTX) | |
14829 | return false; | |
14830 | extract_base_offset_in_addr (mem_2, &base_2, &offset_2); | |
14831 | if (base_2 == NULL_RTX || offset_2 == NULL_RTX) | |
14832 | return false; | |
14833 | extract_base_offset_in_addr (mem_3, &base_3, &offset_3); | |
14834 | if (base_3 == NULL_RTX || offset_3 == NULL_RTX) | |
14835 | return false; | |
14836 | extract_base_offset_in_addr (mem_4, &base_4, &offset_4); | |
14837 | if (base_4 == NULL_RTX || offset_4 == NULL_RTX) | |
14838 | return false; | |
14839 | ||
14840 | /* Check if the bases are same. */ | |
14841 | if (!rtx_equal_p (base_1, base_2) | |
14842 | || !rtx_equal_p (base_2, base_3) | |
14843 | || !rtx_equal_p (base_3, base_4)) | |
14844 | return false; | |
14845 | ||
14846 | offval_1 = INTVAL (offset_1); | |
14847 | offval_2 = INTVAL (offset_2); | |
14848 | offval_3 = INTVAL (offset_3); | |
14849 | offval_4 = INTVAL (offset_4); | |
14850 | msize = GET_MODE_SIZE (mode); | |
14851 | /* Check if the offsets are consecutive. */ | |
14852 | if ((offval_1 != (offval_2 + msize) | |
14853 | || offval_1 != (offval_3 + msize * 2) | |
14854 | || offval_1 != (offval_4 + msize * 3)) | |
14855 | && (offval_4 != (offval_3 + msize) | |
14856 | || offval_4 != (offval_2 + msize * 2) | |
14857 | || offval_4 != (offval_1 + msize * 3))) | |
14858 | return false; | |
14859 | ||
14860 | /* Check if the addresses are clobbered by load. */ | |
14861 | if (load) | |
14862 | { | |
14863 | if (reg_mentioned_p (reg_1, mem_1) | |
14864 | || reg_mentioned_p (reg_2, mem_2) | |
14865 | || reg_mentioned_p (reg_3, mem_3)) | |
14866 | return false; | |
14867 | ||
14868 | /* In increasing order, the last load can clobber the address. */ | |
14869 | if (offval_1 > offval_2 && reg_mentioned_p (reg_4, mem_4)) | |
14870 | return false; | |
14871 | } | |
14872 | ||
54700e2e AP |
14873 | /* If we have SImode and slow unaligned ldp, |
14874 | check the alignment to be at least 8 byte. */ | |
14875 | if (mode == SImode | |
14876 | && (aarch64_tune_params.extra_tuning_flags | |
14877 | & AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW) | |
14878 | && !optimize_size | |
14879 | && MEM_ALIGN (mem_1) < 8 * BITS_PER_UNIT) | |
14880 | return false; | |
14881 | ||
350013bc BC |
14882 | if (REG_P (reg_1) && FP_REGNUM_P (REGNO (reg_1))) |
14883 | rclass_1 = FP_REGS; | |
14884 | else | |
14885 | rclass_1 = GENERAL_REGS; | |
14886 | ||
14887 | if (REG_P (reg_2) && FP_REGNUM_P (REGNO (reg_2))) | |
14888 | rclass_2 = FP_REGS; | |
14889 | else | |
14890 | rclass_2 = GENERAL_REGS; | |
14891 | ||
14892 | if (REG_P (reg_3) && FP_REGNUM_P (REGNO (reg_3))) | |
14893 | rclass_3 = FP_REGS; | |
14894 | else | |
14895 | rclass_3 = GENERAL_REGS; | |
14896 | ||
14897 | if (REG_P (reg_4) && FP_REGNUM_P (REGNO (reg_4))) | |
14898 | rclass_4 = FP_REGS; | |
14899 | else | |
14900 | rclass_4 = GENERAL_REGS; | |
14901 | ||
14902 | /* Check if the registers are of same class. */ | |
14903 | if (rclass_1 != rclass_2 || rclass_2 != rclass_3 || rclass_3 != rclass_4) | |
14904 | return false; | |
14905 | ||
14906 | return true; | |
14907 | } | |
14908 | ||
14909 | /* Given OPERANDS of consecutive load/store, this function pairs them | |
14910 | into ldp/stp after adjusting the offset. It depends on the fact | |
14911 | that addresses of load/store instructions are in increasing order. | |
14912 | MODE is the mode of memory operands. CODE is the rtl operator | |
14913 | which should be applied to all memory operands, it's SIGN_EXTEND, | |
14914 | ZERO_EXTEND or UNKNOWN. */ | |
14915 | ||
14916 | bool | |
14917 | aarch64_gen_adjusted_ldpstp (rtx *operands, bool load, | |
b8506a8a | 14918 | machine_mode mode, RTX_CODE code) |
350013bc BC |
14919 | { |
14920 | rtx base, offset, t1, t2; | |
14921 | rtx mem_1, mem_2, mem_3, mem_4; | |
14922 | HOST_WIDE_INT off_val, abs_off, adj_off, new_off, stp_off_limit, msize; | |
14923 | ||
14924 | if (load) | |
14925 | { | |
14926 | mem_1 = operands[1]; | |
14927 | mem_2 = operands[3]; | |
14928 | mem_3 = operands[5]; | |
14929 | mem_4 = operands[7]; | |
14930 | } | |
14931 | else | |
14932 | { | |
14933 | mem_1 = operands[0]; | |
14934 | mem_2 = operands[2]; | |
14935 | mem_3 = operands[4]; | |
14936 | mem_4 = operands[6]; | |
14937 | gcc_assert (code == UNKNOWN); | |
14938 | } | |
14939 | ||
14940 | extract_base_offset_in_addr (mem_1, &base, &offset); | |
14941 | gcc_assert (base != NULL_RTX && offset != NULL_RTX); | |
14942 | ||
14943 | /* Adjust offset thus it can fit in ldp/stp instruction. */ | |
14944 | msize = GET_MODE_SIZE (mode); | |
14945 | stp_off_limit = msize * 0x40; | |
14946 | off_val = INTVAL (offset); | |
14947 | abs_off = (off_val < 0) ? -off_val : off_val; | |
14948 | new_off = abs_off % stp_off_limit; | |
14949 | adj_off = abs_off - new_off; | |
14950 | ||
14951 | /* Further adjust to make sure all offsets are OK. */ | |
14952 | if ((new_off + msize * 2) >= stp_off_limit) | |
14953 | { | |
14954 | adj_off += stp_off_limit; | |
14955 | new_off -= stp_off_limit; | |
14956 | } | |
14957 | ||
14958 | /* Make sure the adjustment can be done with ADD/SUB instructions. */ | |
14959 | if (adj_off >= 0x1000) | |
14960 | return false; | |
14961 | ||
14962 | if (off_val < 0) | |
14963 | { | |
14964 | adj_off = -adj_off; | |
14965 | new_off = -new_off; | |
14966 | } | |
14967 | ||
14968 | /* Create new memory references. */ | |
14969 | mem_1 = change_address (mem_1, VOIDmode, | |
14970 | plus_constant (DImode, operands[8], new_off)); | |
14971 | ||
14972 | /* Check if the adjusted address is OK for ldp/stp. */ | |
14973 | if (!aarch64_mem_pair_operand (mem_1, mode)) | |
14974 | return false; | |
14975 | ||
14976 | msize = GET_MODE_SIZE (mode); | |
14977 | mem_2 = change_address (mem_2, VOIDmode, | |
14978 | plus_constant (DImode, | |
14979 | operands[8], | |
14980 | new_off + msize)); | |
14981 | mem_3 = change_address (mem_3, VOIDmode, | |
14982 | plus_constant (DImode, | |
14983 | operands[8], | |
14984 | new_off + msize * 2)); | |
14985 | mem_4 = change_address (mem_4, VOIDmode, | |
14986 | plus_constant (DImode, | |
14987 | operands[8], | |
14988 | new_off + msize * 3)); | |
14989 | ||
14990 | if (code == ZERO_EXTEND) | |
14991 | { | |
14992 | mem_1 = gen_rtx_ZERO_EXTEND (DImode, mem_1); | |
14993 | mem_2 = gen_rtx_ZERO_EXTEND (DImode, mem_2); | |
14994 | mem_3 = gen_rtx_ZERO_EXTEND (DImode, mem_3); | |
14995 | mem_4 = gen_rtx_ZERO_EXTEND (DImode, mem_4); | |
14996 | } | |
14997 | else if (code == SIGN_EXTEND) | |
14998 | { | |
14999 | mem_1 = gen_rtx_SIGN_EXTEND (DImode, mem_1); | |
15000 | mem_2 = gen_rtx_SIGN_EXTEND (DImode, mem_2); | |
15001 | mem_3 = gen_rtx_SIGN_EXTEND (DImode, mem_3); | |
15002 | mem_4 = gen_rtx_SIGN_EXTEND (DImode, mem_4); | |
15003 | } | |
15004 | ||
15005 | if (load) | |
15006 | { | |
15007 | operands[1] = mem_1; | |
15008 | operands[3] = mem_2; | |
15009 | operands[5] = mem_3; | |
15010 | operands[7] = mem_4; | |
15011 | } | |
15012 | else | |
15013 | { | |
15014 | operands[0] = mem_1; | |
15015 | operands[2] = mem_2; | |
15016 | operands[4] = mem_3; | |
15017 | operands[6] = mem_4; | |
15018 | } | |
15019 | ||
15020 | /* Emit adjusting instruction. */ | |
f7df4a84 | 15021 | emit_insn (gen_rtx_SET (operands[8], plus_constant (DImode, base, adj_off))); |
350013bc | 15022 | /* Emit ldp/stp instructions. */ |
f7df4a84 RS |
15023 | t1 = gen_rtx_SET (operands[0], operands[1]); |
15024 | t2 = gen_rtx_SET (operands[2], operands[3]); | |
350013bc | 15025 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, t1, t2))); |
f7df4a84 RS |
15026 | t1 = gen_rtx_SET (operands[4], operands[5]); |
15027 | t2 = gen_rtx_SET (operands[6], operands[7]); | |
350013bc BC |
15028 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, t1, t2))); |
15029 | return true; | |
15030 | } | |
15031 | ||
1b1e81f8 JW |
15032 | /* Return 1 if pseudo register should be created and used to hold |
15033 | GOT address for PIC code. */ | |
15034 | ||
15035 | bool | |
15036 | aarch64_use_pseudo_pic_reg (void) | |
15037 | { | |
15038 | return aarch64_cmodel == AARCH64_CMODEL_SMALL_SPIC; | |
15039 | } | |
15040 | ||
7b841a12 JW |
15041 | /* Implement TARGET_UNSPEC_MAY_TRAP_P. */ |
15042 | ||
15043 | static int | |
15044 | aarch64_unspec_may_trap_p (const_rtx x, unsigned flags) | |
15045 | { | |
15046 | switch (XINT (x, 1)) | |
15047 | { | |
15048 | case UNSPEC_GOTSMALLPIC: | |
15049 | case UNSPEC_GOTSMALLPIC28K: | |
15050 | case UNSPEC_GOTTINYPIC: | |
15051 | return 0; | |
15052 | default: | |
15053 | break; | |
15054 | } | |
15055 | ||
15056 | return default_unspec_may_trap_p (x, flags); | |
15057 | } | |
15058 | ||
39252973 KT |
15059 | |
15060 | /* If X is a positive CONST_DOUBLE with a value that is a power of 2 | |
15061 | return the log2 of that value. Otherwise return -1. */ | |
15062 | ||
15063 | int | |
15064 | aarch64_fpconst_pow_of_2 (rtx x) | |
15065 | { | |
15066 | const REAL_VALUE_TYPE *r; | |
15067 | ||
15068 | if (!CONST_DOUBLE_P (x)) | |
15069 | return -1; | |
15070 | ||
15071 | r = CONST_DOUBLE_REAL_VALUE (x); | |
15072 | ||
15073 | if (REAL_VALUE_NEGATIVE (*r) | |
15074 | || REAL_VALUE_ISNAN (*r) | |
15075 | || REAL_VALUE_ISINF (*r) | |
15076 | || !real_isinteger (r, DFmode)) | |
15077 | return -1; | |
15078 | ||
15079 | return exact_log2 (real_to_integer (r)); | |
15080 | } | |
15081 | ||
15082 | /* If X is a vector of equal CONST_DOUBLE values and that value is | |
15083 | Y, return the aarch64_fpconst_pow_of_2 of Y. Otherwise return -1. */ | |
15084 | ||
15085 | int | |
15086 | aarch64_vec_fpconst_pow_of_2 (rtx x) | |
15087 | { | |
15088 | if (GET_CODE (x) != CONST_VECTOR) | |
15089 | return -1; | |
15090 | ||
15091 | if (GET_MODE_CLASS (GET_MODE (x)) != MODE_VECTOR_FLOAT) | |
15092 | return -1; | |
15093 | ||
15094 | int firstval = aarch64_fpconst_pow_of_2 (CONST_VECTOR_ELT (x, 0)); | |
15095 | if (firstval <= 0) | |
15096 | return -1; | |
15097 | ||
15098 | for (int i = 1; i < CONST_VECTOR_NUNITS (x); i++) | |
15099 | if (aarch64_fpconst_pow_of_2 (CONST_VECTOR_ELT (x, i)) != firstval) | |
15100 | return -1; | |
15101 | ||
15102 | return firstval; | |
15103 | } | |
15104 | ||
11e554b3 JG |
15105 | /* Implement TARGET_PROMOTED_TYPE to promote 16-bit floating point types |
15106 | to float. | |
15107 | ||
15108 | __fp16 always promotes through this hook. | |
15109 | _Float16 may promote if TARGET_FLT_EVAL_METHOD is 16, but we do that | |
15110 | through the generic excess precision logic rather than here. */ | |
15111 | ||
c2ec330c AL |
15112 | static tree |
15113 | aarch64_promoted_type (const_tree t) | |
15114 | { | |
11e554b3 JG |
15115 | if (SCALAR_FLOAT_TYPE_P (t) |
15116 | && TYPE_MAIN_VARIANT (t) == aarch64_fp16_type_node) | |
c2ec330c | 15117 | return float_type_node; |
11e554b3 | 15118 | |
c2ec330c AL |
15119 | return NULL_TREE; |
15120 | } | |
ee62a5a6 RS |
15121 | |
15122 | /* Implement the TARGET_OPTAB_SUPPORTED_P hook. */ | |
15123 | ||
15124 | static bool | |
9acc9cbe | 15125 | aarch64_optab_supported_p (int op, machine_mode mode1, machine_mode, |
ee62a5a6 RS |
15126 | optimization_type opt_type) |
15127 | { | |
15128 | switch (op) | |
15129 | { | |
15130 | case rsqrt_optab: | |
9acc9cbe | 15131 | return opt_type == OPTIMIZE_FOR_SPEED && use_rsqrt_p (mode1); |
ee62a5a6 RS |
15132 | |
15133 | default: | |
15134 | return true; | |
15135 | } | |
15136 | } | |
15137 | ||
11e554b3 JG |
15138 | /* Implement TARGET_LIBGCC_FLOATING_POINT_MODE_SUPPORTED_P - return TRUE |
15139 | if MODE is HFmode, and punt to the generic implementation otherwise. */ | |
15140 | ||
15141 | static bool | |
7c5bd57a | 15142 | aarch64_libgcc_floating_mode_supported_p (scalar_float_mode mode) |
11e554b3 JG |
15143 | { |
15144 | return (mode == HFmode | |
15145 | ? true | |
15146 | : default_libgcc_floating_mode_supported_p (mode)); | |
15147 | } | |
15148 | ||
2e5f8203 JG |
15149 | /* Implement TARGET_SCALAR_MODE_SUPPORTED_P - return TRUE |
15150 | if MODE is HFmode, and punt to the generic implementation otherwise. */ | |
15151 | ||
15152 | static bool | |
15153 | aarch64_scalar_mode_supported_p (machine_mode mode) | |
15154 | { | |
15155 | return (mode == HFmode | |
15156 | ? true | |
15157 | : default_scalar_mode_supported_p (mode)); | |
15158 | } | |
15159 | ||
11e554b3 JG |
15160 | /* Set the value of FLT_EVAL_METHOD. |
15161 | ISO/IEC TS 18661-3 defines two values that we'd like to make use of: | |
15162 | ||
15163 | 0: evaluate all operations and constants, whose semantic type has at | |
15164 | most the range and precision of type float, to the range and | |
15165 | precision of float; evaluate all other operations and constants to | |
15166 | the range and precision of the semantic type; | |
15167 | ||
15168 | N, where _FloatN is a supported interchange floating type | |
15169 | evaluate all operations and constants, whose semantic type has at | |
15170 | most the range and precision of _FloatN type, to the range and | |
15171 | precision of the _FloatN type; evaluate all other operations and | |
15172 | constants to the range and precision of the semantic type; | |
15173 | ||
15174 | If we have the ARMv8.2-A extensions then we support _Float16 in native | |
15175 | precision, so we should set this to 16. Otherwise, we support the type, | |
15176 | but want to evaluate expressions in float precision, so set this to | |
15177 | 0. */ | |
15178 | ||
15179 | static enum flt_eval_method | |
15180 | aarch64_excess_precision (enum excess_precision_type type) | |
15181 | { | |
15182 | switch (type) | |
15183 | { | |
15184 | case EXCESS_PRECISION_TYPE_FAST: | |
15185 | case EXCESS_PRECISION_TYPE_STANDARD: | |
15186 | /* We can calculate either in 16-bit range and precision or | |
15187 | 32-bit range and precision. Make that decision based on whether | |
15188 | we have native support for the ARMv8.2-A 16-bit floating-point | |
15189 | instructions or not. */ | |
15190 | return (TARGET_FP_F16INST | |
15191 | ? FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16 | |
15192 | : FLT_EVAL_METHOD_PROMOTE_TO_FLOAT); | |
15193 | case EXCESS_PRECISION_TYPE_IMPLICIT: | |
15194 | return FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16; | |
15195 | default: | |
15196 | gcc_unreachable (); | |
15197 | } | |
15198 | return FLT_EVAL_METHOD_UNPREDICTABLE; | |
15199 | } | |
15200 | ||
b48d6421 KT |
15201 | /* Implement TARGET_SCHED_CAN_SPECULATE_INSN. Return true if INSN can be |
15202 | scheduled for speculative execution. Reject the long-running division | |
15203 | and square-root instructions. */ | |
15204 | ||
15205 | static bool | |
15206 | aarch64_sched_can_speculate_insn (rtx_insn *insn) | |
15207 | { | |
15208 | switch (get_attr_type (insn)) | |
15209 | { | |
15210 | case TYPE_SDIV: | |
15211 | case TYPE_UDIV: | |
15212 | case TYPE_FDIVS: | |
15213 | case TYPE_FDIVD: | |
15214 | case TYPE_FSQRTS: | |
15215 | case TYPE_FSQRTD: | |
15216 | case TYPE_NEON_FP_SQRT_S: | |
15217 | case TYPE_NEON_FP_SQRT_D: | |
15218 | case TYPE_NEON_FP_SQRT_S_Q: | |
15219 | case TYPE_NEON_FP_SQRT_D_Q: | |
15220 | case TYPE_NEON_FP_DIV_S: | |
15221 | case TYPE_NEON_FP_DIV_D: | |
15222 | case TYPE_NEON_FP_DIV_S_Q: | |
15223 | case TYPE_NEON_FP_DIV_D_Q: | |
15224 | return false; | |
15225 | default: | |
15226 | return true; | |
15227 | } | |
15228 | } | |
15229 | ||
51b86113 DM |
15230 | /* Target-specific selftests. */ |
15231 | ||
15232 | #if CHECKING_P | |
15233 | ||
15234 | namespace selftest { | |
15235 | ||
15236 | /* Selftest for the RTL loader. | |
15237 | Verify that the RTL loader copes with a dump from | |
15238 | print_rtx_function. This is essentially just a test that class | |
15239 | function_reader can handle a real dump, but it also verifies | |
15240 | that lookup_reg_by_dump_name correctly handles hard regs. | |
15241 | The presence of hard reg names in the dump means that the test is | |
15242 | target-specific, hence it is in this file. */ | |
15243 | ||
15244 | static void | |
15245 | aarch64_test_loading_full_dump () | |
15246 | { | |
15247 | rtl_dump_test t (SELFTEST_LOCATION, locate_file ("aarch64/times-two.rtl")); | |
15248 | ||
15249 | ASSERT_STREQ ("times_two", IDENTIFIER_POINTER (DECL_NAME (cfun->decl))); | |
15250 | ||
15251 | rtx_insn *insn_1 = get_insn_by_uid (1); | |
15252 | ASSERT_EQ (NOTE, GET_CODE (insn_1)); | |
15253 | ||
15254 | rtx_insn *insn_15 = get_insn_by_uid (15); | |
15255 | ASSERT_EQ (INSN, GET_CODE (insn_15)); | |
15256 | ASSERT_EQ (USE, GET_CODE (PATTERN (insn_15))); | |
15257 | ||
15258 | /* Verify crtl->return_rtx. */ | |
15259 | ASSERT_EQ (REG, GET_CODE (crtl->return_rtx)); | |
15260 | ASSERT_EQ (0, REGNO (crtl->return_rtx)); | |
15261 | ASSERT_EQ (SImode, GET_MODE (crtl->return_rtx)); | |
15262 | } | |
15263 | ||
15264 | /* Run all target-specific selftests. */ | |
15265 | ||
15266 | static void | |
15267 | aarch64_run_selftests (void) | |
15268 | { | |
15269 | aarch64_test_loading_full_dump (); | |
15270 | } | |
15271 | ||
15272 | } // namespace selftest | |
15273 | ||
15274 | #endif /* #if CHECKING_P */ | |
15275 | ||
43e9d192 IB |
15276 | #undef TARGET_ADDRESS_COST |
15277 | #define TARGET_ADDRESS_COST aarch64_address_cost | |
15278 | ||
15279 | /* This hook will determines whether unnamed bitfields affect the alignment | |
15280 | of the containing structure. The hook returns true if the structure | |
15281 | should inherit the alignment requirements of an unnamed bitfield's | |
15282 | type. */ | |
15283 | #undef TARGET_ALIGN_ANON_BITFIELD | |
15284 | #define TARGET_ALIGN_ANON_BITFIELD hook_bool_void_true | |
15285 | ||
15286 | #undef TARGET_ASM_ALIGNED_DI_OP | |
15287 | #define TARGET_ASM_ALIGNED_DI_OP "\t.xword\t" | |
15288 | ||
15289 | #undef TARGET_ASM_ALIGNED_HI_OP | |
15290 | #define TARGET_ASM_ALIGNED_HI_OP "\t.hword\t" | |
15291 | ||
15292 | #undef TARGET_ASM_ALIGNED_SI_OP | |
15293 | #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t" | |
15294 | ||
15295 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
15296 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK \ | |
15297 | hook_bool_const_tree_hwi_hwi_const_tree_true | |
15298 | ||
e1c1ecb0 KT |
15299 | #undef TARGET_ASM_FILE_START |
15300 | #define TARGET_ASM_FILE_START aarch64_start_file | |
15301 | ||
43e9d192 IB |
15302 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
15303 | #define TARGET_ASM_OUTPUT_MI_THUNK aarch64_output_mi_thunk | |
15304 | ||
15305 | #undef TARGET_ASM_SELECT_RTX_SECTION | |
15306 | #define TARGET_ASM_SELECT_RTX_SECTION aarch64_select_rtx_section | |
15307 | ||
15308 | #undef TARGET_ASM_TRAMPOLINE_TEMPLATE | |
15309 | #define TARGET_ASM_TRAMPOLINE_TEMPLATE aarch64_asm_trampoline_template | |
15310 | ||
15311 | #undef TARGET_BUILD_BUILTIN_VA_LIST | |
15312 | #define TARGET_BUILD_BUILTIN_VA_LIST aarch64_build_builtin_va_list | |
15313 | ||
15314 | #undef TARGET_CALLEE_COPIES | |
15315 | #define TARGET_CALLEE_COPIES hook_bool_CUMULATIVE_ARGS_mode_tree_bool_false | |
15316 | ||
15317 | #undef TARGET_CAN_ELIMINATE | |
15318 | #define TARGET_CAN_ELIMINATE aarch64_can_eliminate | |
15319 | ||
1fd8d40c KT |
15320 | #undef TARGET_CAN_INLINE_P |
15321 | #define TARGET_CAN_INLINE_P aarch64_can_inline_p | |
15322 | ||
43e9d192 IB |
15323 | #undef TARGET_CANNOT_FORCE_CONST_MEM |
15324 | #define TARGET_CANNOT_FORCE_CONST_MEM aarch64_cannot_force_const_mem | |
15325 | ||
50487d79 EM |
15326 | #undef TARGET_CASE_VALUES_THRESHOLD |
15327 | #define TARGET_CASE_VALUES_THRESHOLD aarch64_case_values_threshold | |
15328 | ||
43e9d192 IB |
15329 | #undef TARGET_CONDITIONAL_REGISTER_USAGE |
15330 | #define TARGET_CONDITIONAL_REGISTER_USAGE aarch64_conditional_register_usage | |
15331 | ||
15332 | /* Only the least significant bit is used for initialization guard | |
15333 | variables. */ | |
15334 | #undef TARGET_CXX_GUARD_MASK_BIT | |
15335 | #define TARGET_CXX_GUARD_MASK_BIT hook_bool_void_true | |
15336 | ||
15337 | #undef TARGET_C_MODE_FOR_SUFFIX | |
15338 | #define TARGET_C_MODE_FOR_SUFFIX aarch64_c_mode_for_suffix | |
15339 | ||
15340 | #ifdef TARGET_BIG_ENDIAN_DEFAULT | |
15341 | #undef TARGET_DEFAULT_TARGET_FLAGS | |
15342 | #define TARGET_DEFAULT_TARGET_FLAGS (MASK_BIG_END) | |
15343 | #endif | |
15344 | ||
15345 | #undef TARGET_CLASS_MAX_NREGS | |
15346 | #define TARGET_CLASS_MAX_NREGS aarch64_class_max_nregs | |
15347 | ||
119103ca JG |
15348 | #undef TARGET_BUILTIN_DECL |
15349 | #define TARGET_BUILTIN_DECL aarch64_builtin_decl | |
15350 | ||
a6fc00da BH |
15351 | #undef TARGET_BUILTIN_RECIPROCAL |
15352 | #define TARGET_BUILTIN_RECIPROCAL aarch64_builtin_reciprocal | |
15353 | ||
11e554b3 JG |
15354 | #undef TARGET_C_EXCESS_PRECISION |
15355 | #define TARGET_C_EXCESS_PRECISION aarch64_excess_precision | |
15356 | ||
43e9d192 IB |
15357 | #undef TARGET_EXPAND_BUILTIN |
15358 | #define TARGET_EXPAND_BUILTIN aarch64_expand_builtin | |
15359 | ||
15360 | #undef TARGET_EXPAND_BUILTIN_VA_START | |
15361 | #define TARGET_EXPAND_BUILTIN_VA_START aarch64_expand_builtin_va_start | |
15362 | ||
9697e620 JG |
15363 | #undef TARGET_FOLD_BUILTIN |
15364 | #define TARGET_FOLD_BUILTIN aarch64_fold_builtin | |
15365 | ||
43e9d192 IB |
15366 | #undef TARGET_FUNCTION_ARG |
15367 | #define TARGET_FUNCTION_ARG aarch64_function_arg | |
15368 | ||
15369 | #undef TARGET_FUNCTION_ARG_ADVANCE | |
15370 | #define TARGET_FUNCTION_ARG_ADVANCE aarch64_function_arg_advance | |
15371 | ||
15372 | #undef TARGET_FUNCTION_ARG_BOUNDARY | |
15373 | #define TARGET_FUNCTION_ARG_BOUNDARY aarch64_function_arg_boundary | |
15374 | ||
15375 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL | |
15376 | #define TARGET_FUNCTION_OK_FOR_SIBCALL aarch64_function_ok_for_sibcall | |
15377 | ||
15378 | #undef TARGET_FUNCTION_VALUE | |
15379 | #define TARGET_FUNCTION_VALUE aarch64_function_value | |
15380 | ||
15381 | #undef TARGET_FUNCTION_VALUE_REGNO_P | |
15382 | #define TARGET_FUNCTION_VALUE_REGNO_P aarch64_function_value_regno_p | |
15383 | ||
15384 | #undef TARGET_FRAME_POINTER_REQUIRED | |
15385 | #define TARGET_FRAME_POINTER_REQUIRED aarch64_frame_pointer_required | |
15386 | ||
fc72cba7 AL |
15387 | #undef TARGET_GIMPLE_FOLD_BUILTIN |
15388 | #define TARGET_GIMPLE_FOLD_BUILTIN aarch64_gimple_fold_builtin | |
0ac198d3 | 15389 | |
43e9d192 IB |
15390 | #undef TARGET_GIMPLIFY_VA_ARG_EXPR |
15391 | #define TARGET_GIMPLIFY_VA_ARG_EXPR aarch64_gimplify_va_arg_expr | |
15392 | ||
15393 | #undef TARGET_INIT_BUILTINS | |
15394 | #define TARGET_INIT_BUILTINS aarch64_init_builtins | |
15395 | ||
c64f7d37 WD |
15396 | #undef TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS |
15397 | #define TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS \ | |
15398 | aarch64_ira_change_pseudo_allocno_class | |
15399 | ||
43e9d192 IB |
15400 | #undef TARGET_LEGITIMATE_ADDRESS_P |
15401 | #define TARGET_LEGITIMATE_ADDRESS_P aarch64_legitimate_address_hook_p | |
15402 | ||
15403 | #undef TARGET_LEGITIMATE_CONSTANT_P | |
15404 | #define TARGET_LEGITIMATE_CONSTANT_P aarch64_legitimate_constant_p | |
15405 | ||
491ec060 WD |
15406 | #undef TARGET_LEGITIMIZE_ADDRESS_DISPLACEMENT |
15407 | #define TARGET_LEGITIMIZE_ADDRESS_DISPLACEMENT \ | |
15408 | aarch64_legitimize_address_displacement | |
15409 | ||
43e9d192 IB |
15410 | #undef TARGET_LIBGCC_CMP_RETURN_MODE |
15411 | #define TARGET_LIBGCC_CMP_RETURN_MODE aarch64_libgcc_cmp_return_mode | |
15412 | ||
11e554b3 JG |
15413 | #undef TARGET_LIBGCC_FLOATING_MODE_SUPPORTED_P |
15414 | #define TARGET_LIBGCC_FLOATING_MODE_SUPPORTED_P \ | |
15415 | aarch64_libgcc_floating_mode_supported_p | |
15416 | ||
ac2b960f YZ |
15417 | #undef TARGET_MANGLE_TYPE |
15418 | #define TARGET_MANGLE_TYPE aarch64_mangle_type | |
15419 | ||
43e9d192 IB |
15420 | #undef TARGET_MEMORY_MOVE_COST |
15421 | #define TARGET_MEMORY_MOVE_COST aarch64_memory_move_cost | |
15422 | ||
26e0ff94 WD |
15423 | #undef TARGET_MIN_DIVISIONS_FOR_RECIP_MUL |
15424 | #define TARGET_MIN_DIVISIONS_FOR_RECIP_MUL aarch64_min_divisions_for_recip_mul | |
15425 | ||
43e9d192 IB |
15426 | #undef TARGET_MUST_PASS_IN_STACK |
15427 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
15428 | ||
15429 | /* This target hook should return true if accesses to volatile bitfields | |
15430 | should use the narrowest mode possible. It should return false if these | |
15431 | accesses should use the bitfield container type. */ | |
15432 | #undef TARGET_NARROW_VOLATILE_BITFIELD | |
15433 | #define TARGET_NARROW_VOLATILE_BITFIELD hook_bool_void_false | |
15434 | ||
15435 | #undef TARGET_OPTION_OVERRIDE | |
15436 | #define TARGET_OPTION_OVERRIDE aarch64_override_options | |
15437 | ||
15438 | #undef TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE | |
15439 | #define TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE \ | |
15440 | aarch64_override_options_after_change | |
15441 | ||
361fb3ee KT |
15442 | #undef TARGET_OPTION_SAVE |
15443 | #define TARGET_OPTION_SAVE aarch64_option_save | |
15444 | ||
15445 | #undef TARGET_OPTION_RESTORE | |
15446 | #define TARGET_OPTION_RESTORE aarch64_option_restore | |
15447 | ||
15448 | #undef TARGET_OPTION_PRINT | |
15449 | #define TARGET_OPTION_PRINT aarch64_option_print | |
15450 | ||
5a2c8331 KT |
15451 | #undef TARGET_OPTION_VALID_ATTRIBUTE_P |
15452 | #define TARGET_OPTION_VALID_ATTRIBUTE_P aarch64_option_valid_attribute_p | |
15453 | ||
d78006d9 KT |
15454 | #undef TARGET_SET_CURRENT_FUNCTION |
15455 | #define TARGET_SET_CURRENT_FUNCTION aarch64_set_current_function | |
15456 | ||
43e9d192 IB |
15457 | #undef TARGET_PASS_BY_REFERENCE |
15458 | #define TARGET_PASS_BY_REFERENCE aarch64_pass_by_reference | |
15459 | ||
15460 | #undef TARGET_PREFERRED_RELOAD_CLASS | |
15461 | #define TARGET_PREFERRED_RELOAD_CLASS aarch64_preferred_reload_class | |
15462 | ||
cee66c68 WD |
15463 | #undef TARGET_SCHED_REASSOCIATION_WIDTH |
15464 | #define TARGET_SCHED_REASSOCIATION_WIDTH aarch64_reassociation_width | |
15465 | ||
c2ec330c AL |
15466 | #undef TARGET_PROMOTED_TYPE |
15467 | #define TARGET_PROMOTED_TYPE aarch64_promoted_type | |
15468 | ||
43e9d192 IB |
15469 | #undef TARGET_SECONDARY_RELOAD |
15470 | #define TARGET_SECONDARY_RELOAD aarch64_secondary_reload | |
15471 | ||
15472 | #undef TARGET_SHIFT_TRUNCATION_MASK | |
15473 | #define TARGET_SHIFT_TRUNCATION_MASK aarch64_shift_truncation_mask | |
15474 | ||
15475 | #undef TARGET_SETUP_INCOMING_VARARGS | |
15476 | #define TARGET_SETUP_INCOMING_VARARGS aarch64_setup_incoming_varargs | |
15477 | ||
15478 | #undef TARGET_STRUCT_VALUE_RTX | |
15479 | #define TARGET_STRUCT_VALUE_RTX aarch64_struct_value_rtx | |
15480 | ||
15481 | #undef TARGET_REGISTER_MOVE_COST | |
15482 | #define TARGET_REGISTER_MOVE_COST aarch64_register_move_cost | |
15483 | ||
15484 | #undef TARGET_RETURN_IN_MEMORY | |
15485 | #define TARGET_RETURN_IN_MEMORY aarch64_return_in_memory | |
15486 | ||
15487 | #undef TARGET_RETURN_IN_MSB | |
15488 | #define TARGET_RETURN_IN_MSB aarch64_return_in_msb | |
15489 | ||
15490 | #undef TARGET_RTX_COSTS | |
7cc2145f | 15491 | #define TARGET_RTX_COSTS aarch64_rtx_costs_wrapper |
43e9d192 | 15492 | |
2e5f8203 JG |
15493 | #undef TARGET_SCALAR_MODE_SUPPORTED_P |
15494 | #define TARGET_SCALAR_MODE_SUPPORTED_P aarch64_scalar_mode_supported_p | |
15495 | ||
d126a4ae AP |
15496 | #undef TARGET_SCHED_ISSUE_RATE |
15497 | #define TARGET_SCHED_ISSUE_RATE aarch64_sched_issue_rate | |
15498 | ||
d03f7e44 MK |
15499 | #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD |
15500 | #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD \ | |
15501 | aarch64_sched_first_cycle_multipass_dfa_lookahead | |
15502 | ||
2d6bc7fa KT |
15503 | #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD |
15504 | #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD \ | |
15505 | aarch64_first_cycle_multipass_dfa_lookahead_guard | |
15506 | ||
827ab47a KT |
15507 | #undef TARGET_SHRINK_WRAP_GET_SEPARATE_COMPONENTS |
15508 | #define TARGET_SHRINK_WRAP_GET_SEPARATE_COMPONENTS \ | |
15509 | aarch64_get_separate_components | |
15510 | ||
15511 | #undef TARGET_SHRINK_WRAP_COMPONENTS_FOR_BB | |
15512 | #define TARGET_SHRINK_WRAP_COMPONENTS_FOR_BB \ | |
15513 | aarch64_components_for_bb | |
15514 | ||
15515 | #undef TARGET_SHRINK_WRAP_DISQUALIFY_COMPONENTS | |
15516 | #define TARGET_SHRINK_WRAP_DISQUALIFY_COMPONENTS \ | |
15517 | aarch64_disqualify_components | |
15518 | ||
15519 | #undef TARGET_SHRINK_WRAP_EMIT_PROLOGUE_COMPONENTS | |
15520 | #define TARGET_SHRINK_WRAP_EMIT_PROLOGUE_COMPONENTS \ | |
15521 | aarch64_emit_prologue_components | |
15522 | ||
15523 | #undef TARGET_SHRINK_WRAP_EMIT_EPILOGUE_COMPONENTS | |
15524 | #define TARGET_SHRINK_WRAP_EMIT_EPILOGUE_COMPONENTS \ | |
15525 | aarch64_emit_epilogue_components | |
15526 | ||
15527 | #undef TARGET_SHRINK_WRAP_SET_HANDLED_COMPONENTS | |
15528 | #define TARGET_SHRINK_WRAP_SET_HANDLED_COMPONENTS \ | |
15529 | aarch64_set_handled_components | |
15530 | ||
43e9d192 IB |
15531 | #undef TARGET_TRAMPOLINE_INIT |
15532 | #define TARGET_TRAMPOLINE_INIT aarch64_trampoline_init | |
15533 | ||
15534 | #undef TARGET_USE_BLOCKS_FOR_CONSTANT_P | |
15535 | #define TARGET_USE_BLOCKS_FOR_CONSTANT_P aarch64_use_blocks_for_constant_p | |
15536 | ||
15537 | #undef TARGET_VECTOR_MODE_SUPPORTED_P | |
15538 | #define TARGET_VECTOR_MODE_SUPPORTED_P aarch64_vector_mode_supported_p | |
15539 | ||
7df76747 N |
15540 | #undef TARGET_VECTORIZE_SUPPORT_VECTOR_MISALIGNMENT |
15541 | #define TARGET_VECTORIZE_SUPPORT_VECTOR_MISALIGNMENT \ | |
15542 | aarch64_builtin_support_vector_misalignment | |
15543 | ||
43e9d192 IB |
15544 | #undef TARGET_ARRAY_MODE_SUPPORTED_P |
15545 | #define TARGET_ARRAY_MODE_SUPPORTED_P aarch64_array_mode_supported_p | |
15546 | ||
8990e73a TB |
15547 | #undef TARGET_VECTORIZE_ADD_STMT_COST |
15548 | #define TARGET_VECTORIZE_ADD_STMT_COST aarch64_add_stmt_cost | |
15549 | ||
15550 | #undef TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST | |
15551 | #define TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST \ | |
15552 | aarch64_builtin_vectorization_cost | |
15553 | ||
43e9d192 IB |
15554 | #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE |
15555 | #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE aarch64_preferred_simd_mode | |
15556 | ||
42fc9a7f JG |
15557 | #undef TARGET_VECTORIZE_BUILTINS |
15558 | #define TARGET_VECTORIZE_BUILTINS | |
15559 | ||
15560 | #undef TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION | |
15561 | #define TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION \ | |
15562 | aarch64_builtin_vectorized_function | |
15563 | ||
3b357264 JG |
15564 | #undef TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES |
15565 | #define TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES \ | |
15566 | aarch64_autovectorize_vector_sizes | |
15567 | ||
aa87aced KV |
15568 | #undef TARGET_ATOMIC_ASSIGN_EXPAND_FENV |
15569 | #define TARGET_ATOMIC_ASSIGN_EXPAND_FENV \ | |
15570 | aarch64_atomic_assign_expand_fenv | |
15571 | ||
43e9d192 IB |
15572 | /* Section anchor support. */ |
15573 | ||
15574 | #undef TARGET_MIN_ANCHOR_OFFSET | |
15575 | #define TARGET_MIN_ANCHOR_OFFSET -256 | |
15576 | ||
15577 | /* Limit the maximum anchor offset to 4k-1, since that's the limit for a | |
15578 | byte offset; we can do much more for larger data types, but have no way | |
15579 | to determine the size of the access. We assume accesses are aligned. */ | |
15580 | #undef TARGET_MAX_ANCHOR_OFFSET | |
15581 | #define TARGET_MAX_ANCHOR_OFFSET 4095 | |
15582 | ||
db0253a4 TB |
15583 | #undef TARGET_VECTOR_ALIGNMENT |
15584 | #define TARGET_VECTOR_ALIGNMENT aarch64_simd_vector_alignment | |
15585 | ||
15586 | #undef TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE | |
15587 | #define TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE \ | |
15588 | aarch64_simd_vector_alignment_reachable | |
15589 | ||
88b08073 JG |
15590 | /* vec_perm support. */ |
15591 | ||
15592 | #undef TARGET_VECTORIZE_VEC_PERM_CONST_OK | |
15593 | #define TARGET_VECTORIZE_VEC_PERM_CONST_OK \ | |
15594 | aarch64_vectorize_vec_perm_const_ok | |
15595 | ||
c2ec330c AL |
15596 | #undef TARGET_INIT_LIBFUNCS |
15597 | #define TARGET_INIT_LIBFUNCS aarch64_init_libfuncs | |
70f09188 | 15598 | |
706b2314 | 15599 | #undef TARGET_FIXED_CONDITION_CODE_REGS |
70f09188 AP |
15600 | #define TARGET_FIXED_CONDITION_CODE_REGS aarch64_fixed_condition_code_regs |
15601 | ||
5cb74e90 RR |
15602 | #undef TARGET_FLAGS_REGNUM |
15603 | #define TARGET_FLAGS_REGNUM CC_REGNUM | |
15604 | ||
78607708 TV |
15605 | #undef TARGET_CALL_FUSAGE_CONTAINS_NON_CALLEE_CLOBBERS |
15606 | #define TARGET_CALL_FUSAGE_CONTAINS_NON_CALLEE_CLOBBERS true | |
15607 | ||
a3125fc2 CL |
15608 | #undef TARGET_ASAN_SHADOW_OFFSET |
15609 | #define TARGET_ASAN_SHADOW_OFFSET aarch64_asan_shadow_offset | |
15610 | ||
0c4ec427 RE |
15611 | #undef TARGET_LEGITIMIZE_ADDRESS |
15612 | #define TARGET_LEGITIMIZE_ADDRESS aarch64_legitimize_address | |
15613 | ||
d3006da6 JG |
15614 | #undef TARGET_USE_BY_PIECES_INFRASTRUCTURE_P |
15615 | #define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \ | |
15616 | aarch64_use_by_pieces_infrastructure_p | |
15617 | ||
b48d6421 KT |
15618 | #undef TARGET_SCHED_CAN_SPECULATE_INSN |
15619 | #define TARGET_SCHED_CAN_SPECULATE_INSN aarch64_sched_can_speculate_insn | |
15620 | ||
594bdd53 FY |
15621 | #undef TARGET_CAN_USE_DOLOOP_P |
15622 | #define TARGET_CAN_USE_DOLOOP_P can_use_doloop_if_innermost | |
15623 | ||
9bca63d4 WD |
15624 | #undef TARGET_SCHED_ADJUST_PRIORITY |
15625 | #define TARGET_SCHED_ADJUST_PRIORITY aarch64_sched_adjust_priority | |
15626 | ||
6a569cdd KT |
15627 | #undef TARGET_SCHED_MACRO_FUSION_P |
15628 | #define TARGET_SCHED_MACRO_FUSION_P aarch64_macro_fusion_p | |
15629 | ||
15630 | #undef TARGET_SCHED_MACRO_FUSION_PAIR_P | |
15631 | #define TARGET_SCHED_MACRO_FUSION_PAIR_P aarch_macro_fusion_pair_p | |
15632 | ||
350013bc BC |
15633 | #undef TARGET_SCHED_FUSION_PRIORITY |
15634 | #define TARGET_SCHED_FUSION_PRIORITY aarch64_sched_fusion_priority | |
15635 | ||
7b841a12 JW |
15636 | #undef TARGET_UNSPEC_MAY_TRAP_P |
15637 | #define TARGET_UNSPEC_MAY_TRAP_P aarch64_unspec_may_trap_p | |
15638 | ||
1b1e81f8 JW |
15639 | #undef TARGET_USE_PSEUDO_PIC_REG |
15640 | #define TARGET_USE_PSEUDO_PIC_REG aarch64_use_pseudo_pic_reg | |
15641 | ||
cc8ca59e JB |
15642 | #undef TARGET_PRINT_OPERAND |
15643 | #define TARGET_PRINT_OPERAND aarch64_print_operand | |
15644 | ||
15645 | #undef TARGET_PRINT_OPERAND_ADDRESS | |
15646 | #define TARGET_PRINT_OPERAND_ADDRESS aarch64_print_operand_address | |
15647 | ||
ee62a5a6 RS |
15648 | #undef TARGET_OPTAB_SUPPORTED_P |
15649 | #define TARGET_OPTAB_SUPPORTED_P aarch64_optab_supported_p | |
15650 | ||
43203dea RR |
15651 | #undef TARGET_OMIT_STRUCT_RETURN_REG |
15652 | #define TARGET_OMIT_STRUCT_RETURN_REG true | |
15653 | ||
f46fe37e EB |
15654 | /* The architecture reserves bits 0 and 1 so use bit 2 for descriptors. */ |
15655 | #undef TARGET_CUSTOM_FUNCTION_DESCRIPTORS | |
15656 | #define TARGET_CUSTOM_FUNCTION_DESCRIPTORS 4 | |
15657 | ||
51b86113 DM |
15658 | #if CHECKING_P |
15659 | #undef TARGET_RUN_TARGET_SELFTESTS | |
15660 | #define TARGET_RUN_TARGET_SELFTESTS selftest::aarch64_run_selftests | |
15661 | #endif /* #if CHECKING_P */ | |
15662 | ||
43e9d192 IB |
15663 | struct gcc_target targetm = TARGET_INITIALIZER; |
15664 | ||
15665 | #include "gt-aarch64.h" |