]>
Commit | Line | Data |
---|---|---|
bdb7bf8a | 1 | /* Machine description for AArch64 architecture. |
5624e564 | 2 | Copyright (C) 2009-2015 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" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
24 | #include "tm.h" | |
25 | #include "insn-codes.h" | |
26 | #include "rtl.h" | |
27 | #include "insn-attr.h" | |
40e23961 MC |
28 | #include "hash-set.h" |
29 | #include "machmode.h" | |
30 | #include "vec.h" | |
31 | #include "double-int.h" | |
32 | #include "input.h" | |
33 | #include "alias.h" | |
34 | #include "symtab.h" | |
35 | #include "wide-int.h" | |
36 | #include "inchash.h" | |
43e9d192 | 37 | #include "tree.h" |
40e23961 | 38 | #include "fold-const.h" |
d8a2d370 DN |
39 | #include "stringpool.h" |
40 | #include "stor-layout.h" | |
41 | #include "calls.h" | |
42 | #include "varasm.h" | |
43e9d192 | 43 | #include "regs.h" |
60393bbc AM |
44 | #include "dominance.h" |
45 | #include "cfg.h" | |
46 | #include "cfgrtl.h" | |
47 | #include "cfganal.h" | |
48 | #include "lcm.h" | |
49 | #include "cfgbuild.h" | |
50 | #include "cfgcleanup.h" | |
51 | #include "predict.h" | |
52 | #include "basic-block.h" | |
43e9d192 IB |
53 | #include "df.h" |
54 | #include "hard-reg-set.h" | |
55 | #include "output.h" | |
36566b39 PK |
56 | #include "hashtab.h" |
57 | #include "function.h" | |
58 | #include "flags.h" | |
59 | #include "statistics.h" | |
60 | #include "real.h" | |
61 | #include "fixed-value.h" | |
62 | #include "insn-config.h" | |
63 | #include "expmed.h" | |
64 | #include "dojump.h" | |
65 | #include "explow.h" | |
66 | #include "emit-rtl.h" | |
67 | #include "stmt.h" | |
43e9d192 IB |
68 | #include "expr.h" |
69 | #include "reload.h" | |
70 | #include "toplev.h" | |
71 | #include "target.h" | |
72 | #include "target-def.h" | |
73 | #include "targhooks.h" | |
74 | #include "ggc.h" | |
43e9d192 IB |
75 | #include "tm_p.h" |
76 | #include "recog.h" | |
77 | #include "langhooks.h" | |
78 | #include "diagnostic-core.h" | |
2fb9a547 | 79 | #include "hash-table.h" |
2fb9a547 AM |
80 | #include "tree-ssa-alias.h" |
81 | #include "internal-fn.h" | |
82 | #include "gimple-fold.h" | |
83 | #include "tree-eh.h" | |
84 | #include "gimple-expr.h" | |
85 | #include "is-a.h" | |
18f429e2 | 86 | #include "gimple.h" |
45b0be94 | 87 | #include "gimplify.h" |
43e9d192 IB |
88 | #include "optabs.h" |
89 | #include "dwarf2.h" | |
8990e73a TB |
90 | #include "cfgloop.h" |
91 | #include "tree-vectorizer.h" | |
d1bcc29f | 92 | #include "aarch64-cost-tables.h" |
0ee859b5 | 93 | #include "dumpfile.h" |
9b2b7279 | 94 | #include "builtins.h" |
8baff86e | 95 | #include "rtl-iter.h" |
9bbe08fe | 96 | #include "tm-constrs.h" |
d03f7e44 | 97 | #include "sched-int.h" |
43e9d192 | 98 | |
28514dda YZ |
99 | /* Defined for convenience. */ |
100 | #define POINTER_BYTES (POINTER_SIZE / BITS_PER_UNIT) | |
101 | ||
43e9d192 IB |
102 | /* Classifies an address. |
103 | ||
104 | ADDRESS_REG_IMM | |
105 | A simple base register plus immediate offset. | |
106 | ||
107 | ADDRESS_REG_WB | |
108 | A base register indexed by immediate offset with writeback. | |
109 | ||
110 | ADDRESS_REG_REG | |
111 | A base register indexed by (optionally scaled) register. | |
112 | ||
113 | ADDRESS_REG_UXTW | |
114 | A base register indexed by (optionally scaled) zero-extended register. | |
115 | ||
116 | ADDRESS_REG_SXTW | |
117 | A base register indexed by (optionally scaled) sign-extended register. | |
118 | ||
119 | ADDRESS_LO_SUM | |
120 | A LO_SUM rtx with a base register and "LO12" symbol relocation. | |
121 | ||
122 | ADDRESS_SYMBOLIC: | |
123 | A constant symbolic address, in pc-relative literal pool. */ | |
124 | ||
125 | enum aarch64_address_type { | |
126 | ADDRESS_REG_IMM, | |
127 | ADDRESS_REG_WB, | |
128 | ADDRESS_REG_REG, | |
129 | ADDRESS_REG_UXTW, | |
130 | ADDRESS_REG_SXTW, | |
131 | ADDRESS_LO_SUM, | |
132 | ADDRESS_SYMBOLIC | |
133 | }; | |
134 | ||
135 | struct aarch64_address_info { | |
136 | enum aarch64_address_type type; | |
137 | rtx base; | |
138 | rtx offset; | |
139 | int shift; | |
140 | enum aarch64_symbol_type symbol_type; | |
141 | }; | |
142 | ||
48063b9d IB |
143 | struct simd_immediate_info |
144 | { | |
145 | rtx value; | |
146 | int shift; | |
147 | int element_width; | |
48063b9d | 148 | bool mvn; |
e4f0f84d | 149 | bool msl; |
48063b9d IB |
150 | }; |
151 | ||
43e9d192 IB |
152 | /* The current code model. */ |
153 | enum aarch64_code_model aarch64_cmodel; | |
154 | ||
155 | #ifdef HAVE_AS_TLS | |
156 | #undef TARGET_HAVE_TLS | |
157 | #define TARGET_HAVE_TLS 1 | |
158 | #endif | |
159 | ||
ef4bddc2 RS |
160 | static bool aarch64_composite_type_p (const_tree, machine_mode); |
161 | static bool aarch64_vfp_is_call_or_return_candidate (machine_mode, | |
43e9d192 | 162 | const_tree, |
ef4bddc2 | 163 | machine_mode *, int *, |
43e9d192 IB |
164 | bool *); |
165 | static void aarch64_elf_asm_constructor (rtx, int) ATTRIBUTE_UNUSED; | |
166 | static void aarch64_elf_asm_destructor (rtx, int) ATTRIBUTE_UNUSED; | |
43e9d192 | 167 | static void aarch64_override_options_after_change (void); |
ef4bddc2 | 168 | static bool aarch64_vector_mode_supported_p (machine_mode); |
43e9d192 | 169 | static unsigned bit_count (unsigned HOST_WIDE_INT); |
ef4bddc2 | 170 | static bool aarch64_vectorize_vec_perm_const_ok (machine_mode vmode, |
88b08073 | 171 | const unsigned char *sel); |
ef4bddc2 | 172 | static int aarch64_address_cost (rtx, machine_mode, addr_space_t, bool); |
88b08073 | 173 | |
0c6caaf8 RL |
174 | /* Major revision number of the ARM Architecture implemented by the target. */ |
175 | unsigned aarch64_architecture_version; | |
176 | ||
43e9d192 | 177 | /* The processor for which instructions should be scheduled. */ |
02fdbd5b | 178 | enum aarch64_processor aarch64_tune = cortexa53; |
43e9d192 IB |
179 | |
180 | /* The current tuning set. */ | |
181 | const struct tune_params *aarch64_tune_params; | |
182 | ||
183 | /* Mask to specify which instructions we are allowed to generate. */ | |
184 | unsigned long aarch64_isa_flags = 0; | |
185 | ||
186 | /* Mask to specify which instruction scheduling options should be used. */ | |
187 | unsigned long aarch64_tune_flags = 0; | |
188 | ||
189 | /* Tuning parameters. */ | |
190 | ||
43e9d192 IB |
191 | static const struct cpu_addrcost_table generic_addrcost_table = |
192 | { | |
67747367 | 193 | { |
bd95e655 JG |
194 | 0, /* hi */ |
195 | 0, /* si */ | |
196 | 0, /* di */ | |
197 | 0, /* ti */ | |
67747367 | 198 | }, |
bd95e655 JG |
199 | 0, /* pre_modify */ |
200 | 0, /* post_modify */ | |
201 | 0, /* register_offset */ | |
202 | 0, /* register_extend */ | |
203 | 0 /* imm_offset */ | |
43e9d192 IB |
204 | }; |
205 | ||
60bff090 JG |
206 | static const struct cpu_addrcost_table cortexa57_addrcost_table = |
207 | { | |
60bff090 | 208 | { |
bd95e655 JG |
209 | 1, /* hi */ |
210 | 0, /* si */ | |
211 | 0, /* di */ | |
212 | 1, /* ti */ | |
60bff090 | 213 | }, |
bd95e655 JG |
214 | 0, /* pre_modify */ |
215 | 0, /* post_modify */ | |
216 | 0, /* register_offset */ | |
217 | 0, /* register_extend */ | |
218 | 0, /* imm_offset */ | |
60bff090 JG |
219 | }; |
220 | ||
381e27aa PT |
221 | static const struct cpu_addrcost_table xgene1_addrcost_table = |
222 | { | |
381e27aa | 223 | { |
bd95e655 JG |
224 | 1, /* hi */ |
225 | 0, /* si */ | |
226 | 0, /* di */ | |
227 | 1, /* ti */ | |
381e27aa | 228 | }, |
bd95e655 JG |
229 | 1, /* pre_modify */ |
230 | 0, /* post_modify */ | |
231 | 0, /* register_offset */ | |
232 | 1, /* register_extend */ | |
233 | 0, /* imm_offset */ | |
381e27aa PT |
234 | }; |
235 | ||
43e9d192 IB |
236 | static const struct cpu_regmove_cost generic_regmove_cost = |
237 | { | |
bd95e655 | 238 | 1, /* GP2GP */ |
3969c510 WD |
239 | /* Avoid the use of slow int<->fp moves for spilling by setting |
240 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
241 | 5, /* GP2FP */ |
242 | 5, /* FP2GP */ | |
243 | 2 /* FP2FP */ | |
43e9d192 IB |
244 | }; |
245 | ||
e4a9c55a WD |
246 | static const struct cpu_regmove_cost cortexa57_regmove_cost = |
247 | { | |
bd95e655 | 248 | 1, /* GP2GP */ |
e4a9c55a WD |
249 | /* Avoid the use of slow int<->fp moves for spilling by setting |
250 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
251 | 5, /* GP2FP */ |
252 | 5, /* FP2GP */ | |
253 | 2 /* FP2FP */ | |
e4a9c55a WD |
254 | }; |
255 | ||
256 | static const struct cpu_regmove_cost cortexa53_regmove_cost = | |
257 | { | |
bd95e655 | 258 | 1, /* GP2GP */ |
e4a9c55a WD |
259 | /* Avoid the use of slow int<->fp moves for spilling by setting |
260 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
261 | 5, /* GP2FP */ |
262 | 5, /* FP2GP */ | |
263 | 2 /* FP2FP */ | |
e4a9c55a WD |
264 | }; |
265 | ||
d1bcc29f AP |
266 | static const struct cpu_regmove_cost thunderx_regmove_cost = |
267 | { | |
bd95e655 JG |
268 | 2, /* GP2GP */ |
269 | 2, /* GP2FP */ | |
270 | 6, /* FP2GP */ | |
271 | 4 /* FP2FP */ | |
d1bcc29f AP |
272 | }; |
273 | ||
381e27aa PT |
274 | static const struct cpu_regmove_cost xgene1_regmove_cost = |
275 | { | |
bd95e655 | 276 | 1, /* GP2GP */ |
381e27aa PT |
277 | /* Avoid the use of slow int<->fp moves for spilling by setting |
278 | their cost higher than memmov_cost. */ | |
bd95e655 JG |
279 | 8, /* GP2FP */ |
280 | 8, /* FP2GP */ | |
281 | 2 /* FP2FP */ | |
381e27aa PT |
282 | }; |
283 | ||
8990e73a | 284 | /* Generic costs for vector insn classes. */ |
8990e73a TB |
285 | static const struct cpu_vector_cost generic_vector_cost = |
286 | { | |
bd95e655 JG |
287 | 1, /* scalar_stmt_cost */ |
288 | 1, /* scalar_load_cost */ | |
289 | 1, /* scalar_store_cost */ | |
290 | 1, /* vec_stmt_cost */ | |
291 | 1, /* vec_to_scalar_cost */ | |
292 | 1, /* scalar_to_vec_cost */ | |
293 | 1, /* vec_align_load_cost */ | |
294 | 1, /* vec_unalign_load_cost */ | |
295 | 1, /* vec_unalign_store_cost */ | |
296 | 1, /* vec_store_cost */ | |
297 | 3, /* cond_taken_branch_cost */ | |
298 | 1 /* cond_not_taken_branch_cost */ | |
8990e73a TB |
299 | }; |
300 | ||
60bff090 | 301 | /* Generic costs for vector insn classes. */ |
60bff090 JG |
302 | static const struct cpu_vector_cost cortexa57_vector_cost = |
303 | { | |
bd95e655 JG |
304 | 1, /* scalar_stmt_cost */ |
305 | 4, /* scalar_load_cost */ | |
306 | 1, /* scalar_store_cost */ | |
307 | 3, /* vec_stmt_cost */ | |
308 | 8, /* vec_to_scalar_cost */ | |
309 | 8, /* scalar_to_vec_cost */ | |
310 | 5, /* vec_align_load_cost */ | |
311 | 5, /* vec_unalign_load_cost */ | |
312 | 1, /* vec_unalign_store_cost */ | |
313 | 1, /* vec_store_cost */ | |
314 | 1, /* cond_taken_branch_cost */ | |
315 | 1 /* cond_not_taken_branch_cost */ | |
60bff090 JG |
316 | }; |
317 | ||
381e27aa | 318 | /* Generic costs for vector insn classes. */ |
381e27aa PT |
319 | static const struct cpu_vector_cost xgene1_vector_cost = |
320 | { | |
bd95e655 JG |
321 | 1, /* scalar_stmt_cost */ |
322 | 5, /* scalar_load_cost */ | |
323 | 1, /* scalar_store_cost */ | |
324 | 2, /* vec_stmt_cost */ | |
325 | 4, /* vec_to_scalar_cost */ | |
326 | 4, /* scalar_to_vec_cost */ | |
327 | 10, /* vec_align_load_cost */ | |
328 | 10, /* vec_unalign_load_cost */ | |
329 | 2, /* vec_unalign_store_cost */ | |
330 | 2, /* vec_store_cost */ | |
331 | 2, /* cond_taken_branch_cost */ | |
332 | 1 /* cond_not_taken_branch_cost */ | |
381e27aa PT |
333 | }; |
334 | ||
6a569cdd KT |
335 | #define AARCH64_FUSE_NOTHING (0) |
336 | #define AARCH64_FUSE_MOV_MOVK (1 << 0) | |
9bbe08fe | 337 | #define AARCH64_FUSE_ADRP_ADD (1 << 1) |
cd0cb232 | 338 | #define AARCH64_FUSE_MOVK_MOVK (1 << 2) |
d8354ad7 | 339 | #define AARCH64_FUSE_ADRP_LDR (1 << 3) |
3759108f | 340 | #define AARCH64_FUSE_CMP_BRANCH (1 << 4) |
6a569cdd | 341 | |
43e9d192 IB |
342 | static const struct tune_params generic_tunings = |
343 | { | |
4e2cd668 | 344 | &cortexa57_extra_costs, |
43e9d192 IB |
345 | &generic_addrcost_table, |
346 | &generic_regmove_cost, | |
8990e73a | 347 | &generic_vector_cost, |
bd95e655 JG |
348 | 4, /* memmov_cost */ |
349 | 2, /* issue_rate */ | |
350 | AARCH64_FUSE_NOTHING, /* fuseable_ops */ | |
0b82a5a2 WD |
351 | 8, /* function_align. */ |
352 | 8, /* jump_align. */ | |
353 | 4, /* loop_align. */ | |
cee66c68 WD |
354 | 2, /* int_reassoc_width. */ |
355 | 4, /* fp_reassoc_width. */ | |
356 | 1 /* vec_reassoc_width. */ | |
43e9d192 IB |
357 | }; |
358 | ||
984239ad KT |
359 | static const struct tune_params cortexa53_tunings = |
360 | { | |
361 | &cortexa53_extra_costs, | |
362 | &generic_addrcost_table, | |
e4a9c55a | 363 | &cortexa53_regmove_cost, |
984239ad | 364 | &generic_vector_cost, |
bd95e655 JG |
365 | 4, /* memmov_cost */ |
366 | 2, /* issue_rate */ | |
367 | (AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD | |
368 | | AARCH64_FUSE_MOVK_MOVK | AARCH64_FUSE_ADRP_LDR), /* fuseable_ops */ | |
0b82a5a2 WD |
369 | 8, /* function_align. */ |
370 | 8, /* jump_align. */ | |
371 | 4, /* loop_align. */ | |
cee66c68 WD |
372 | 2, /* int_reassoc_width. */ |
373 | 4, /* fp_reassoc_width. */ | |
374 | 1 /* vec_reassoc_width. */ | |
984239ad KT |
375 | }; |
376 | ||
4fd92af6 KT |
377 | static const struct tune_params cortexa57_tunings = |
378 | { | |
379 | &cortexa57_extra_costs, | |
60bff090 | 380 | &cortexa57_addrcost_table, |
e4a9c55a | 381 | &cortexa57_regmove_cost, |
60bff090 | 382 | &cortexa57_vector_cost, |
bd95e655 JG |
383 | 4, /* memmov_cost */ |
384 | 3, /* issue_rate */ | |
385 | (AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD | |
386 | | AARCH64_FUSE_MOVK_MOVK), /* fuseable_ops */ | |
0b82a5a2 WD |
387 | 16, /* function_align. */ |
388 | 8, /* jump_align. */ | |
389 | 4, /* loop_align. */ | |
cee66c68 WD |
390 | 2, /* int_reassoc_width. */ |
391 | 4, /* fp_reassoc_width. */ | |
392 | 1 /* vec_reassoc_width. */ | |
4fd92af6 KT |
393 | }; |
394 | ||
d1bcc29f AP |
395 | static const struct tune_params thunderx_tunings = |
396 | { | |
397 | &thunderx_extra_costs, | |
398 | &generic_addrcost_table, | |
399 | &thunderx_regmove_cost, | |
400 | &generic_vector_cost, | |
bd95e655 JG |
401 | 6, /* memmov_cost */ |
402 | 2, /* issue_rate */ | |
403 | AARCH64_FUSE_CMP_BRANCH, /* fuseable_ops */ | |
0b82a5a2 WD |
404 | 8, /* function_align. */ |
405 | 8, /* jump_align. */ | |
406 | 8, /* loop_align. */ | |
cee66c68 WD |
407 | 2, /* int_reassoc_width. */ |
408 | 4, /* fp_reassoc_width. */ | |
409 | 1 /* vec_reassoc_width. */ | |
d1bcc29f AP |
410 | }; |
411 | ||
381e27aa PT |
412 | static const struct tune_params xgene1_tunings = |
413 | { | |
414 | &xgene1_extra_costs, | |
415 | &xgene1_addrcost_table, | |
416 | &xgene1_regmove_cost, | |
417 | &xgene1_vector_cost, | |
bd95e655 JG |
418 | 6, /* memmov_cost */ |
419 | 4, /* issue_rate */ | |
420 | AARCH64_FUSE_NOTHING, /* fuseable_ops */ | |
381e27aa PT |
421 | 16, /* function_align. */ |
422 | 8, /* jump_align. */ | |
423 | 16, /* loop_align. */ | |
424 | 2, /* int_reassoc_width. */ | |
425 | 4, /* fp_reassoc_width. */ | |
426 | 1 /* vec_reassoc_width. */ | |
427 | }; | |
428 | ||
43e9d192 IB |
429 | /* A processor implementing AArch64. */ |
430 | struct processor | |
431 | { | |
432 | const char *const name; | |
433 | enum aarch64_processor core; | |
434 | const char *arch; | |
0c6caaf8 | 435 | unsigned architecture_version; |
43e9d192 IB |
436 | const unsigned long flags; |
437 | const struct tune_params *const tune; | |
438 | }; | |
439 | ||
440 | /* Processor cores implementing AArch64. */ | |
441 | static const struct processor all_cores[] = | |
442 | { | |
d86cb6d5 | 443 | #define AARCH64_CORE(NAME, IDENT, SCHED, ARCH, FLAGS, COSTS) \ |
faa54226 | 444 | {NAME, SCHED, #ARCH, ARCH, FLAGS, &COSTS##_tunings}, |
43e9d192 IB |
445 | #include "aarch64-cores.def" |
446 | #undef AARCH64_CORE | |
faa54226 | 447 | {"generic", cortexa53, "8", 8, AARCH64_FL_FOR_ARCH8, &generic_tunings}, |
0c6caaf8 | 448 | {NULL, aarch64_none, NULL, 0, 0, NULL} |
43e9d192 IB |
449 | }; |
450 | ||
451 | /* Architectures implementing AArch64. */ | |
452 | static const struct processor all_architectures[] = | |
453 | { | |
454 | #define AARCH64_ARCH(NAME, CORE, ARCH, FLAGS) \ | |
0c6caaf8 | 455 | {NAME, CORE, #ARCH, ARCH, FLAGS, NULL}, |
43e9d192 IB |
456 | #include "aarch64-arches.def" |
457 | #undef AARCH64_ARCH | |
0c6caaf8 | 458 | {NULL, aarch64_none, NULL, 0, 0, NULL} |
43e9d192 IB |
459 | }; |
460 | ||
461 | /* Target specification. These are populated as commandline arguments | |
462 | are processed, or NULL if not specified. */ | |
463 | static const struct processor *selected_arch; | |
464 | static const struct processor *selected_cpu; | |
465 | static const struct processor *selected_tune; | |
466 | ||
467 | #define AARCH64_CPU_DEFAULT_FLAGS ((selected_cpu) ? selected_cpu->flags : 0) | |
468 | ||
469 | /* An ISA extension in the co-processor and main instruction set space. */ | |
470 | struct aarch64_option_extension | |
471 | { | |
472 | const char *const name; | |
473 | const unsigned long flags_on; | |
474 | const unsigned long flags_off; | |
475 | }; | |
476 | ||
477 | /* ISA extensions in AArch64. */ | |
478 | static const struct aarch64_option_extension all_extensions[] = | |
479 | { | |
480 | #define AARCH64_OPT_EXTENSION(NAME, FLAGS_ON, FLAGS_OFF) \ | |
481 | {NAME, FLAGS_ON, FLAGS_OFF}, | |
482 | #include "aarch64-option-extensions.def" | |
483 | #undef AARCH64_OPT_EXTENSION | |
484 | {NULL, 0, 0} | |
485 | }; | |
486 | ||
487 | /* Used to track the size of an address when generating a pre/post | |
488 | increment address. */ | |
ef4bddc2 | 489 | static machine_mode aarch64_memory_reference_mode; |
43e9d192 | 490 | |
43e9d192 IB |
491 | /* A table of valid AArch64 "bitmask immediate" values for |
492 | logical instructions. */ | |
493 | ||
494 | #define AARCH64_NUM_BITMASKS 5334 | |
495 | static unsigned HOST_WIDE_INT aarch64_bitmasks[AARCH64_NUM_BITMASKS]; | |
496 | ||
43e9d192 IB |
497 | typedef enum aarch64_cond_code |
498 | { | |
499 | AARCH64_EQ = 0, AARCH64_NE, AARCH64_CS, AARCH64_CC, AARCH64_MI, AARCH64_PL, | |
500 | AARCH64_VS, AARCH64_VC, AARCH64_HI, AARCH64_LS, AARCH64_GE, AARCH64_LT, | |
501 | AARCH64_GT, AARCH64_LE, AARCH64_AL, AARCH64_NV | |
502 | } | |
503 | aarch64_cc; | |
504 | ||
505 | #define AARCH64_INVERSE_CONDITION_CODE(X) ((aarch64_cc) (((int) X) ^ 1)) | |
506 | ||
507 | /* The condition codes of the processor, and the inverse function. */ | |
508 | static const char * const aarch64_condition_codes[] = | |
509 | { | |
510 | "eq", "ne", "cs", "cc", "mi", "pl", "vs", "vc", | |
511 | "hi", "ls", "ge", "lt", "gt", "le", "al", "nv" | |
512 | }; | |
513 | ||
26e0ff94 WD |
514 | static unsigned int |
515 | aarch64_min_divisions_for_recip_mul (enum machine_mode mode ATTRIBUTE_UNUSED) | |
516 | { | |
517 | return 2; | |
518 | } | |
519 | ||
cee66c68 WD |
520 | static int |
521 | aarch64_reassociation_width (unsigned opc ATTRIBUTE_UNUSED, | |
522 | enum machine_mode mode) | |
523 | { | |
524 | if (VECTOR_MODE_P (mode)) | |
525 | return aarch64_tune_params->vec_reassoc_width; | |
526 | if (INTEGRAL_MODE_P (mode)) | |
527 | return aarch64_tune_params->int_reassoc_width; | |
528 | if (FLOAT_MODE_P (mode)) | |
529 | return aarch64_tune_params->fp_reassoc_width; | |
530 | return 1; | |
531 | } | |
532 | ||
43e9d192 IB |
533 | /* Provide a mapping from gcc register numbers to dwarf register numbers. */ |
534 | unsigned | |
535 | aarch64_dbx_register_number (unsigned regno) | |
536 | { | |
537 | if (GP_REGNUM_P (regno)) | |
538 | return AARCH64_DWARF_R0 + regno - R0_REGNUM; | |
539 | else if (regno == SP_REGNUM) | |
540 | return AARCH64_DWARF_SP; | |
541 | else if (FP_REGNUM_P (regno)) | |
542 | return AARCH64_DWARF_V0 + regno - V0_REGNUM; | |
543 | ||
544 | /* Return values >= DWARF_FRAME_REGISTERS indicate that there is no | |
545 | equivalent DWARF register. */ | |
546 | return DWARF_FRAME_REGISTERS; | |
547 | } | |
548 | ||
549 | /* Return TRUE if MODE is any of the large INT modes. */ | |
550 | static bool | |
ef4bddc2 | 551 | aarch64_vect_struct_mode_p (machine_mode mode) |
43e9d192 IB |
552 | { |
553 | return mode == OImode || mode == CImode || mode == XImode; | |
554 | } | |
555 | ||
556 | /* Return TRUE if MODE is any of the vector modes. */ | |
557 | static bool | |
ef4bddc2 | 558 | aarch64_vector_mode_p (machine_mode mode) |
43e9d192 IB |
559 | { |
560 | return aarch64_vector_mode_supported_p (mode) | |
561 | || aarch64_vect_struct_mode_p (mode); | |
562 | } | |
563 | ||
564 | /* Implement target hook TARGET_ARRAY_MODE_SUPPORTED_P. */ | |
565 | static bool | |
ef4bddc2 | 566 | aarch64_array_mode_supported_p (machine_mode mode, |
43e9d192 IB |
567 | unsigned HOST_WIDE_INT nelems) |
568 | { | |
569 | if (TARGET_SIMD | |
570 | && AARCH64_VALID_SIMD_QREG_MODE (mode) | |
571 | && (nelems >= 2 && nelems <= 4)) | |
572 | return true; | |
573 | ||
574 | return false; | |
575 | } | |
576 | ||
577 | /* Implement HARD_REGNO_NREGS. */ | |
578 | ||
579 | int | |
ef4bddc2 | 580 | aarch64_hard_regno_nregs (unsigned regno, machine_mode mode) |
43e9d192 IB |
581 | { |
582 | switch (aarch64_regno_regclass (regno)) | |
583 | { | |
584 | case FP_REGS: | |
585 | case FP_LO_REGS: | |
586 | return (GET_MODE_SIZE (mode) + UNITS_PER_VREG - 1) / UNITS_PER_VREG; | |
587 | default: | |
588 | return (GET_MODE_SIZE (mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
589 | } | |
590 | gcc_unreachable (); | |
591 | } | |
592 | ||
593 | /* Implement HARD_REGNO_MODE_OK. */ | |
594 | ||
595 | int | |
ef4bddc2 | 596 | aarch64_hard_regno_mode_ok (unsigned regno, machine_mode mode) |
43e9d192 IB |
597 | { |
598 | if (GET_MODE_CLASS (mode) == MODE_CC) | |
599 | return regno == CC_REGNUM; | |
600 | ||
9259db42 YZ |
601 | if (regno == SP_REGNUM) |
602 | /* The purpose of comparing with ptr_mode is to support the | |
603 | global register variable associated with the stack pointer | |
604 | register via the syntax of asm ("wsp") in ILP32. */ | |
605 | return mode == Pmode || mode == ptr_mode; | |
606 | ||
607 | if (regno == FRAME_POINTER_REGNUM || regno == ARG_POINTER_REGNUM) | |
43e9d192 IB |
608 | return mode == Pmode; |
609 | ||
610 | if (GP_REGNUM_P (regno) && ! aarch64_vect_struct_mode_p (mode)) | |
611 | return 1; | |
612 | ||
613 | if (FP_REGNUM_P (regno)) | |
614 | { | |
615 | if (aarch64_vect_struct_mode_p (mode)) | |
616 | return | |
617 | (regno + aarch64_hard_regno_nregs (regno, mode) - 1) <= V31_REGNUM; | |
618 | else | |
619 | return 1; | |
620 | } | |
621 | ||
622 | return 0; | |
623 | } | |
624 | ||
73d9ac6a | 625 | /* Implement HARD_REGNO_CALLER_SAVE_MODE. */ |
ef4bddc2 | 626 | machine_mode |
73d9ac6a | 627 | aarch64_hard_regno_caller_save_mode (unsigned regno, unsigned nregs, |
ef4bddc2 | 628 | machine_mode mode) |
73d9ac6a IB |
629 | { |
630 | /* Handle modes that fit within single registers. */ | |
631 | if (nregs == 1 && GET_MODE_SIZE (mode) <= 16) | |
632 | { | |
633 | if (GET_MODE_SIZE (mode) >= 4) | |
634 | return mode; | |
635 | else | |
636 | return SImode; | |
637 | } | |
638 | /* Fall back to generic for multi-reg and very large modes. */ | |
639 | else | |
640 | return choose_hard_reg_mode (regno, nregs, false); | |
641 | } | |
642 | ||
43e9d192 IB |
643 | /* Return true if calls to DECL should be treated as |
644 | long-calls (ie called via a register). */ | |
645 | static bool | |
646 | aarch64_decl_is_long_call_p (const_tree decl ATTRIBUTE_UNUSED) | |
647 | { | |
648 | return false; | |
649 | } | |
650 | ||
651 | /* Return true if calls to symbol-ref SYM should be treated as | |
652 | long-calls (ie called via a register). */ | |
653 | bool | |
654 | aarch64_is_long_call_p (rtx sym) | |
655 | { | |
656 | return aarch64_decl_is_long_call_p (SYMBOL_REF_DECL (sym)); | |
657 | } | |
658 | ||
659 | /* Return true if the offsets to a zero/sign-extract operation | |
660 | represent an expression that matches an extend operation. The | |
661 | operands represent the paramters from | |
662 | ||
4745e701 | 663 | (extract:MODE (mult (reg) (MULT_IMM)) (EXTRACT_IMM) (const_int 0)). */ |
43e9d192 | 664 | bool |
ef4bddc2 | 665 | aarch64_is_extend_from_extract (machine_mode mode, rtx mult_imm, |
43e9d192 IB |
666 | rtx extract_imm) |
667 | { | |
668 | HOST_WIDE_INT mult_val, extract_val; | |
669 | ||
670 | if (! CONST_INT_P (mult_imm) || ! CONST_INT_P (extract_imm)) | |
671 | return false; | |
672 | ||
673 | mult_val = INTVAL (mult_imm); | |
674 | extract_val = INTVAL (extract_imm); | |
675 | ||
676 | if (extract_val > 8 | |
677 | && extract_val < GET_MODE_BITSIZE (mode) | |
678 | && exact_log2 (extract_val & ~7) > 0 | |
679 | && (extract_val & 7) <= 4 | |
680 | && mult_val == (1 << (extract_val & 7))) | |
681 | return true; | |
682 | ||
683 | return false; | |
684 | } | |
685 | ||
686 | /* Emit an insn that's a simple single-set. Both the operands must be | |
687 | known to be valid. */ | |
688 | inline static rtx | |
689 | emit_set_insn (rtx x, rtx y) | |
690 | { | |
691 | return emit_insn (gen_rtx_SET (VOIDmode, x, y)); | |
692 | } | |
693 | ||
694 | /* X and Y are two things to compare using CODE. Emit the compare insn and | |
695 | return the rtx for register 0 in the proper mode. */ | |
696 | rtx | |
697 | aarch64_gen_compare_reg (RTX_CODE code, rtx x, rtx y) | |
698 | { | |
ef4bddc2 | 699 | machine_mode mode = SELECT_CC_MODE (code, x, y); |
43e9d192 IB |
700 | rtx cc_reg = gen_rtx_REG (mode, CC_REGNUM); |
701 | ||
702 | emit_set_insn (cc_reg, gen_rtx_COMPARE (mode, x, y)); | |
703 | return cc_reg; | |
704 | } | |
705 | ||
706 | /* Build the SYMBOL_REF for __tls_get_addr. */ | |
707 | ||
708 | static GTY(()) rtx tls_get_addr_libfunc; | |
709 | ||
710 | rtx | |
711 | aarch64_tls_get_addr (void) | |
712 | { | |
713 | if (!tls_get_addr_libfunc) | |
714 | tls_get_addr_libfunc = init_one_libfunc ("__tls_get_addr"); | |
715 | return tls_get_addr_libfunc; | |
716 | } | |
717 | ||
718 | /* Return the TLS model to use for ADDR. */ | |
719 | ||
720 | static enum tls_model | |
721 | tls_symbolic_operand_type (rtx addr) | |
722 | { | |
723 | enum tls_model tls_kind = TLS_MODEL_NONE; | |
724 | rtx sym, addend; | |
725 | ||
726 | if (GET_CODE (addr) == CONST) | |
727 | { | |
728 | split_const (addr, &sym, &addend); | |
729 | if (GET_CODE (sym) == SYMBOL_REF) | |
730 | tls_kind = SYMBOL_REF_TLS_MODEL (sym); | |
731 | } | |
732 | else if (GET_CODE (addr) == SYMBOL_REF) | |
733 | tls_kind = SYMBOL_REF_TLS_MODEL (addr); | |
734 | ||
735 | return tls_kind; | |
736 | } | |
737 | ||
738 | /* We'll allow lo_sum's in addresses in our legitimate addresses | |
739 | so that combine would take care of combining addresses where | |
740 | necessary, but for generation purposes, we'll generate the address | |
741 | as : | |
742 | RTL Absolute | |
743 | tmp = hi (symbol_ref); adrp x1, foo | |
744 | dest = lo_sum (tmp, symbol_ref); add dest, x1, :lo_12:foo | |
745 | nop | |
746 | ||
747 | PIC TLS | |
748 | adrp x1, :got:foo adrp tmp, :tlsgd:foo | |
749 | ldr x1, [:got_lo12:foo] add dest, tmp, :tlsgd_lo12:foo | |
750 | bl __tls_get_addr | |
751 | nop | |
752 | ||
753 | Load TLS symbol, depending on TLS mechanism and TLS access model. | |
754 | ||
755 | Global Dynamic - Traditional TLS: | |
756 | adrp tmp, :tlsgd:imm | |
757 | add dest, tmp, #:tlsgd_lo12:imm | |
758 | bl __tls_get_addr | |
759 | ||
760 | Global Dynamic - TLS Descriptors: | |
761 | adrp dest, :tlsdesc:imm | |
762 | ldr tmp, [dest, #:tlsdesc_lo12:imm] | |
763 | add dest, dest, #:tlsdesc_lo12:imm | |
764 | blr tmp | |
765 | mrs tp, tpidr_el0 | |
766 | add dest, dest, tp | |
767 | ||
768 | Initial Exec: | |
769 | mrs tp, tpidr_el0 | |
770 | adrp tmp, :gottprel:imm | |
771 | ldr dest, [tmp, #:gottprel_lo12:imm] | |
772 | add dest, dest, tp | |
773 | ||
774 | Local Exec: | |
775 | mrs tp, tpidr_el0 | |
0699caae RL |
776 | add t0, tp, #:tprel_hi12:imm, lsl #12 |
777 | add t0, t0, #:tprel_lo12_nc:imm | |
43e9d192 IB |
778 | */ |
779 | ||
780 | static void | |
781 | aarch64_load_symref_appropriately (rtx dest, rtx imm, | |
782 | enum aarch64_symbol_type type) | |
783 | { | |
784 | switch (type) | |
785 | { | |
786 | case SYMBOL_SMALL_ABSOLUTE: | |
787 | { | |
28514dda | 788 | /* In ILP32, the mode of dest can be either SImode or DImode. */ |
43e9d192 | 789 | rtx tmp_reg = dest; |
ef4bddc2 | 790 | machine_mode mode = GET_MODE (dest); |
28514dda YZ |
791 | |
792 | gcc_assert (mode == Pmode || mode == ptr_mode); | |
793 | ||
43e9d192 | 794 | if (can_create_pseudo_p ()) |
28514dda | 795 | tmp_reg = gen_reg_rtx (mode); |
43e9d192 | 796 | |
28514dda | 797 | emit_move_insn (tmp_reg, gen_rtx_HIGH (mode, imm)); |
43e9d192 IB |
798 | emit_insn (gen_add_losym (dest, tmp_reg, imm)); |
799 | return; | |
800 | } | |
801 | ||
a5350ddc CSS |
802 | case SYMBOL_TINY_ABSOLUTE: |
803 | emit_insn (gen_rtx_SET (Pmode, dest, imm)); | |
804 | return; | |
805 | ||
43e9d192 IB |
806 | case SYMBOL_SMALL_GOT: |
807 | { | |
28514dda YZ |
808 | /* In ILP32, the mode of dest can be either SImode or DImode, |
809 | while the got entry is always of SImode size. The mode of | |
810 | dest depends on how dest is used: if dest is assigned to a | |
811 | pointer (e.g. in the memory), it has SImode; it may have | |
812 | DImode if dest is dereferenced to access the memeory. | |
813 | This is why we have to handle three different ldr_got_small | |
814 | patterns here (two patterns for ILP32). */ | |
43e9d192 | 815 | rtx tmp_reg = dest; |
ef4bddc2 | 816 | machine_mode mode = GET_MODE (dest); |
28514dda | 817 | |
43e9d192 | 818 | if (can_create_pseudo_p ()) |
28514dda YZ |
819 | tmp_reg = gen_reg_rtx (mode); |
820 | ||
821 | emit_move_insn (tmp_reg, gen_rtx_HIGH (mode, imm)); | |
822 | if (mode == ptr_mode) | |
823 | { | |
824 | if (mode == DImode) | |
825 | emit_insn (gen_ldr_got_small_di (dest, tmp_reg, imm)); | |
826 | else | |
827 | emit_insn (gen_ldr_got_small_si (dest, tmp_reg, imm)); | |
828 | } | |
829 | else | |
830 | { | |
831 | gcc_assert (mode == Pmode); | |
832 | emit_insn (gen_ldr_got_small_sidi (dest, tmp_reg, imm)); | |
833 | } | |
834 | ||
43e9d192 IB |
835 | return; |
836 | } | |
837 | ||
838 | case SYMBOL_SMALL_TLSGD: | |
839 | { | |
5d8a22a5 | 840 | rtx_insn *insns; |
43e9d192 IB |
841 | rtx result = gen_rtx_REG (Pmode, R0_REGNUM); |
842 | ||
843 | start_sequence (); | |
78607708 | 844 | aarch64_emit_call_insn (gen_tlsgd_small (result, imm)); |
43e9d192 IB |
845 | insns = get_insns (); |
846 | end_sequence (); | |
847 | ||
848 | RTL_CONST_CALL_P (insns) = 1; | |
849 | emit_libcall_block (insns, dest, result, imm); | |
850 | return; | |
851 | } | |
852 | ||
853 | case SYMBOL_SMALL_TLSDESC: | |
854 | { | |
ef4bddc2 | 855 | machine_mode mode = GET_MODE (dest); |
621ad2de | 856 | rtx x0 = gen_rtx_REG (mode, R0_REGNUM); |
43e9d192 IB |
857 | rtx tp; |
858 | ||
621ad2de AP |
859 | gcc_assert (mode == Pmode || mode == ptr_mode); |
860 | ||
861 | /* In ILP32, the got entry is always of SImode size. Unlike | |
862 | small GOT, the dest is fixed at reg 0. */ | |
863 | if (TARGET_ILP32) | |
864 | emit_insn (gen_tlsdesc_small_si (imm)); | |
865 | else | |
866 | emit_insn (gen_tlsdesc_small_di (imm)); | |
43e9d192 | 867 | tp = aarch64_load_tp (NULL); |
621ad2de AP |
868 | |
869 | if (mode != Pmode) | |
870 | tp = gen_lowpart (mode, tp); | |
871 | ||
872 | emit_insn (gen_rtx_SET (mode, dest, gen_rtx_PLUS (mode, tp, x0))); | |
43e9d192 IB |
873 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); |
874 | return; | |
875 | } | |
876 | ||
877 | case SYMBOL_SMALL_GOTTPREL: | |
878 | { | |
621ad2de AP |
879 | /* In ILP32, the mode of dest can be either SImode or DImode, |
880 | while the got entry is always of SImode size. The mode of | |
881 | dest depends on how dest is used: if dest is assigned to a | |
882 | pointer (e.g. in the memory), it has SImode; it may have | |
883 | DImode if dest is dereferenced to access the memeory. | |
884 | This is why we have to handle three different tlsie_small | |
885 | patterns here (two patterns for ILP32). */ | |
ef4bddc2 | 886 | machine_mode mode = GET_MODE (dest); |
621ad2de | 887 | rtx tmp_reg = gen_reg_rtx (mode); |
43e9d192 | 888 | rtx tp = aarch64_load_tp (NULL); |
621ad2de AP |
889 | |
890 | if (mode == ptr_mode) | |
891 | { | |
892 | if (mode == DImode) | |
893 | emit_insn (gen_tlsie_small_di (tmp_reg, imm)); | |
894 | else | |
895 | { | |
896 | emit_insn (gen_tlsie_small_si (tmp_reg, imm)); | |
897 | tp = gen_lowpart (mode, tp); | |
898 | } | |
899 | } | |
900 | else | |
901 | { | |
902 | gcc_assert (mode == Pmode); | |
903 | emit_insn (gen_tlsie_small_sidi (tmp_reg, imm)); | |
904 | } | |
905 | ||
906 | emit_insn (gen_rtx_SET (mode, dest, gen_rtx_PLUS (mode, tp, tmp_reg))); | |
43e9d192 IB |
907 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); |
908 | return; | |
909 | } | |
910 | ||
911 | case SYMBOL_SMALL_TPREL: | |
912 | { | |
913 | rtx tp = aarch64_load_tp (NULL); | |
914 | emit_insn (gen_tlsle_small (dest, tp, imm)); | |
915 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); | |
916 | return; | |
917 | } | |
918 | ||
87dd8ab0 MS |
919 | case SYMBOL_TINY_GOT: |
920 | emit_insn (gen_ldr_got_tiny (dest, imm)); | |
921 | return; | |
922 | ||
43e9d192 IB |
923 | default: |
924 | gcc_unreachable (); | |
925 | } | |
926 | } | |
927 | ||
928 | /* Emit a move from SRC to DEST. Assume that the move expanders can | |
929 | handle all moves if !can_create_pseudo_p (). The distinction is | |
930 | important because, unlike emit_move_insn, the move expanders know | |
931 | how to force Pmode objects into the constant pool even when the | |
932 | constant pool address is not itself legitimate. */ | |
933 | static rtx | |
934 | aarch64_emit_move (rtx dest, rtx src) | |
935 | { | |
936 | return (can_create_pseudo_p () | |
937 | ? emit_move_insn (dest, src) | |
938 | : emit_move_insn_1 (dest, src)); | |
939 | } | |
940 | ||
030d03b8 RE |
941 | /* Split a 128-bit move operation into two 64-bit move operations, |
942 | taking care to handle partial overlap of register to register | |
943 | copies. Special cases are needed when moving between GP regs and | |
944 | FP regs. SRC can be a register, constant or memory; DST a register | |
945 | or memory. If either operand is memory it must not have any side | |
946 | effects. */ | |
43e9d192 IB |
947 | void |
948 | aarch64_split_128bit_move (rtx dst, rtx src) | |
949 | { | |
030d03b8 RE |
950 | rtx dst_lo, dst_hi; |
951 | rtx src_lo, src_hi; | |
43e9d192 | 952 | |
ef4bddc2 | 953 | machine_mode mode = GET_MODE (dst); |
12dc6974 | 954 | |
030d03b8 RE |
955 | gcc_assert (mode == TImode || mode == TFmode); |
956 | gcc_assert (!(side_effects_p (src) || side_effects_p (dst))); | |
957 | gcc_assert (mode == GET_MODE (src) || GET_MODE (src) == VOIDmode); | |
43e9d192 IB |
958 | |
959 | if (REG_P (dst) && REG_P (src)) | |
960 | { | |
030d03b8 RE |
961 | int src_regno = REGNO (src); |
962 | int dst_regno = REGNO (dst); | |
43e9d192 | 963 | |
030d03b8 | 964 | /* Handle FP <-> GP regs. */ |
43e9d192 IB |
965 | if (FP_REGNUM_P (dst_regno) && GP_REGNUM_P (src_regno)) |
966 | { | |
030d03b8 RE |
967 | src_lo = gen_lowpart (word_mode, src); |
968 | src_hi = gen_highpart (word_mode, src); | |
969 | ||
970 | if (mode == TImode) | |
971 | { | |
972 | emit_insn (gen_aarch64_movtilow_di (dst, src_lo)); | |
973 | emit_insn (gen_aarch64_movtihigh_di (dst, src_hi)); | |
974 | } | |
975 | else | |
976 | { | |
977 | emit_insn (gen_aarch64_movtflow_di (dst, src_lo)); | |
978 | emit_insn (gen_aarch64_movtfhigh_di (dst, src_hi)); | |
979 | } | |
980 | return; | |
43e9d192 IB |
981 | } |
982 | else if (GP_REGNUM_P (dst_regno) && FP_REGNUM_P (src_regno)) | |
983 | { | |
030d03b8 RE |
984 | dst_lo = gen_lowpart (word_mode, dst); |
985 | dst_hi = gen_highpart (word_mode, dst); | |
986 | ||
987 | if (mode == TImode) | |
988 | { | |
989 | emit_insn (gen_aarch64_movdi_tilow (dst_lo, src)); | |
990 | emit_insn (gen_aarch64_movdi_tihigh (dst_hi, src)); | |
991 | } | |
992 | else | |
993 | { | |
994 | emit_insn (gen_aarch64_movdi_tflow (dst_lo, src)); | |
995 | emit_insn (gen_aarch64_movdi_tfhigh (dst_hi, src)); | |
996 | } | |
997 | return; | |
43e9d192 | 998 | } |
43e9d192 IB |
999 | } |
1000 | ||
030d03b8 RE |
1001 | dst_lo = gen_lowpart (word_mode, dst); |
1002 | dst_hi = gen_highpart (word_mode, dst); | |
1003 | src_lo = gen_lowpart (word_mode, src); | |
1004 | src_hi = gen_highpart_mode (word_mode, mode, src); | |
1005 | ||
1006 | /* At most one pairing may overlap. */ | |
1007 | if (reg_overlap_mentioned_p (dst_lo, src_hi)) | |
1008 | { | |
1009 | aarch64_emit_move (dst_hi, src_hi); | |
1010 | aarch64_emit_move (dst_lo, src_lo); | |
1011 | } | |
1012 | else | |
1013 | { | |
1014 | aarch64_emit_move (dst_lo, src_lo); | |
1015 | aarch64_emit_move (dst_hi, src_hi); | |
1016 | } | |
43e9d192 IB |
1017 | } |
1018 | ||
1019 | bool | |
1020 | aarch64_split_128bit_move_p (rtx dst, rtx src) | |
1021 | { | |
1022 | return (! REG_P (src) | |
1023 | || ! (FP_REGNUM_P (REGNO (dst)) && FP_REGNUM_P (REGNO (src)))); | |
1024 | } | |
1025 | ||
8b033a8a SN |
1026 | /* Split a complex SIMD combine. */ |
1027 | ||
1028 | void | |
1029 | aarch64_split_simd_combine (rtx dst, rtx src1, rtx src2) | |
1030 | { | |
ef4bddc2 RS |
1031 | machine_mode src_mode = GET_MODE (src1); |
1032 | machine_mode dst_mode = GET_MODE (dst); | |
8b033a8a SN |
1033 | |
1034 | gcc_assert (VECTOR_MODE_P (dst_mode)); | |
1035 | ||
1036 | if (REG_P (dst) && REG_P (src1) && REG_P (src2)) | |
1037 | { | |
1038 | rtx (*gen) (rtx, rtx, rtx); | |
1039 | ||
1040 | switch (src_mode) | |
1041 | { | |
1042 | case V8QImode: | |
1043 | gen = gen_aarch64_simd_combinev8qi; | |
1044 | break; | |
1045 | case V4HImode: | |
1046 | gen = gen_aarch64_simd_combinev4hi; | |
1047 | break; | |
1048 | case V2SImode: | |
1049 | gen = gen_aarch64_simd_combinev2si; | |
1050 | break; | |
1051 | case V2SFmode: | |
1052 | gen = gen_aarch64_simd_combinev2sf; | |
1053 | break; | |
1054 | case DImode: | |
1055 | gen = gen_aarch64_simd_combinedi; | |
1056 | break; | |
1057 | case DFmode: | |
1058 | gen = gen_aarch64_simd_combinedf; | |
1059 | break; | |
1060 | default: | |
1061 | gcc_unreachable (); | |
1062 | } | |
1063 | ||
1064 | emit_insn (gen (dst, src1, src2)); | |
1065 | return; | |
1066 | } | |
1067 | } | |
1068 | ||
fd4842cd SN |
1069 | /* Split a complex SIMD move. */ |
1070 | ||
1071 | void | |
1072 | aarch64_split_simd_move (rtx dst, rtx src) | |
1073 | { | |
ef4bddc2 RS |
1074 | machine_mode src_mode = GET_MODE (src); |
1075 | machine_mode dst_mode = GET_MODE (dst); | |
fd4842cd SN |
1076 | |
1077 | gcc_assert (VECTOR_MODE_P (dst_mode)); | |
1078 | ||
1079 | if (REG_P (dst) && REG_P (src)) | |
1080 | { | |
c59b7e28 SN |
1081 | rtx (*gen) (rtx, rtx); |
1082 | ||
fd4842cd SN |
1083 | gcc_assert (VECTOR_MODE_P (src_mode)); |
1084 | ||
1085 | switch (src_mode) | |
1086 | { | |
1087 | case V16QImode: | |
c59b7e28 | 1088 | gen = gen_aarch64_split_simd_movv16qi; |
fd4842cd SN |
1089 | break; |
1090 | case V8HImode: | |
c59b7e28 | 1091 | gen = gen_aarch64_split_simd_movv8hi; |
fd4842cd SN |
1092 | break; |
1093 | case V4SImode: | |
c59b7e28 | 1094 | gen = gen_aarch64_split_simd_movv4si; |
fd4842cd SN |
1095 | break; |
1096 | case V2DImode: | |
c59b7e28 | 1097 | gen = gen_aarch64_split_simd_movv2di; |
fd4842cd SN |
1098 | break; |
1099 | case V4SFmode: | |
c59b7e28 | 1100 | gen = gen_aarch64_split_simd_movv4sf; |
fd4842cd SN |
1101 | break; |
1102 | case V2DFmode: | |
c59b7e28 | 1103 | gen = gen_aarch64_split_simd_movv2df; |
fd4842cd SN |
1104 | break; |
1105 | default: | |
1106 | gcc_unreachable (); | |
1107 | } | |
c59b7e28 SN |
1108 | |
1109 | emit_insn (gen (dst, src)); | |
fd4842cd SN |
1110 | return; |
1111 | } | |
1112 | } | |
1113 | ||
43e9d192 | 1114 | static rtx |
ef4bddc2 | 1115 | aarch64_force_temporary (machine_mode mode, rtx x, rtx value) |
43e9d192 IB |
1116 | { |
1117 | if (can_create_pseudo_p ()) | |
e18b4a81 | 1118 | return force_reg (mode, value); |
43e9d192 IB |
1119 | else |
1120 | { | |
1121 | x = aarch64_emit_move (x, value); | |
1122 | return x; | |
1123 | } | |
1124 | } | |
1125 | ||
1126 | ||
1127 | static rtx | |
ef4bddc2 | 1128 | aarch64_add_offset (machine_mode mode, rtx temp, rtx reg, HOST_WIDE_INT offset) |
43e9d192 | 1129 | { |
9c023bf0 | 1130 | if (!aarch64_plus_immediate (GEN_INT (offset), mode)) |
43e9d192 IB |
1131 | { |
1132 | rtx high; | |
1133 | /* Load the full offset into a register. This | |
1134 | might be improvable in the future. */ | |
1135 | high = GEN_INT (offset); | |
1136 | offset = 0; | |
e18b4a81 YZ |
1137 | high = aarch64_force_temporary (mode, temp, high); |
1138 | reg = aarch64_force_temporary (mode, temp, | |
1139 | gen_rtx_PLUS (mode, high, reg)); | |
43e9d192 IB |
1140 | } |
1141 | return plus_constant (mode, reg, offset); | |
1142 | } | |
1143 | ||
82614948 RR |
1144 | static int |
1145 | aarch64_internal_mov_immediate (rtx dest, rtx imm, bool generate, | |
1146 | machine_mode mode) | |
43e9d192 | 1147 | { |
43e9d192 IB |
1148 | unsigned HOST_WIDE_INT mask; |
1149 | int i; | |
1150 | bool first; | |
1151 | unsigned HOST_WIDE_INT val; | |
1152 | bool subtargets; | |
1153 | rtx subtarget; | |
c747993a | 1154 | int one_match, zero_match, first_not_ffff_match; |
82614948 | 1155 | int num_insns = 0; |
43e9d192 IB |
1156 | |
1157 | if (CONST_INT_P (imm) && aarch64_move_imm (INTVAL (imm), mode)) | |
1158 | { | |
82614948 | 1159 | if (generate) |
43e9d192 | 1160 | emit_insn (gen_rtx_SET (VOIDmode, dest, imm)); |
82614948 RR |
1161 | num_insns++; |
1162 | return num_insns; | |
43e9d192 IB |
1163 | } |
1164 | ||
1165 | if (mode == SImode) | |
1166 | { | |
1167 | /* We know we can't do this in 1 insn, and we must be able to do it | |
1168 | in two; so don't mess around looking for sequences that don't buy | |
1169 | us anything. */ | |
82614948 RR |
1170 | if (generate) |
1171 | { | |
1172 | emit_insn (gen_rtx_SET (VOIDmode, dest, | |
1173 | GEN_INT (INTVAL (imm) & 0xffff))); | |
1174 | emit_insn (gen_insv_immsi (dest, GEN_INT (16), | |
1175 | GEN_INT ((INTVAL (imm) >> 16) & 0xffff))); | |
1176 | } | |
1177 | num_insns += 2; | |
1178 | return num_insns; | |
43e9d192 IB |
1179 | } |
1180 | ||
1181 | /* Remaining cases are all for DImode. */ | |
1182 | ||
1183 | val = INTVAL (imm); | |
1184 | subtargets = optimize && can_create_pseudo_p (); | |
1185 | ||
1186 | one_match = 0; | |
1187 | zero_match = 0; | |
1188 | mask = 0xffff; | |
c747993a | 1189 | first_not_ffff_match = -1; |
43e9d192 IB |
1190 | |
1191 | for (i = 0; i < 64; i += 16, mask <<= 16) | |
1192 | { | |
c747993a | 1193 | if ((val & mask) == mask) |
43e9d192 | 1194 | one_match++; |
c747993a IB |
1195 | else |
1196 | { | |
1197 | if (first_not_ffff_match < 0) | |
1198 | first_not_ffff_match = i; | |
1199 | if ((val & mask) == 0) | |
1200 | zero_match++; | |
1201 | } | |
43e9d192 IB |
1202 | } |
1203 | ||
1204 | if (one_match == 2) | |
1205 | { | |
c747993a IB |
1206 | /* Set one of the quarters and then insert back into result. */ |
1207 | mask = 0xffffll << first_not_ffff_match; | |
82614948 RR |
1208 | if (generate) |
1209 | { | |
1210 | emit_insn (gen_rtx_SET (VOIDmode, dest, GEN_INT (val | mask))); | |
1211 | emit_insn (gen_insv_immdi (dest, GEN_INT (first_not_ffff_match), | |
1212 | GEN_INT ((val >> first_not_ffff_match) | |
1213 | & 0xffff))); | |
1214 | } | |
1215 | num_insns += 2; | |
1216 | return num_insns; | |
c747993a IB |
1217 | } |
1218 | ||
43e9d192 IB |
1219 | if (zero_match == 2) |
1220 | goto simple_sequence; | |
1221 | ||
1222 | mask = 0x0ffff0000UL; | |
1223 | for (i = 16; i < 64; i += 16, mask <<= 16) | |
1224 | { | |
1225 | HOST_WIDE_INT comp = mask & ~(mask - 1); | |
1226 | ||
1227 | if (aarch64_uimm12_shift (val - (val & mask))) | |
1228 | { | |
82614948 RR |
1229 | if (generate) |
1230 | { | |
1231 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1232 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1233 | GEN_INT (val & mask))); | |
1234 | emit_insn (gen_adddi3 (dest, subtarget, | |
1235 | GEN_INT (val - (val & mask)))); | |
1236 | } | |
1237 | num_insns += 2; | |
1238 | return num_insns; | |
43e9d192 IB |
1239 | } |
1240 | else if (aarch64_uimm12_shift (-(val - ((val + comp) & mask)))) | |
1241 | { | |
82614948 RR |
1242 | if (generate) |
1243 | { | |
1244 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1245 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1246 | GEN_INT ((val + comp) & mask))); | |
1247 | emit_insn (gen_adddi3 (dest, subtarget, | |
1248 | GEN_INT (val - ((val + comp) & mask)))); | |
1249 | } | |
1250 | num_insns += 2; | |
1251 | return num_insns; | |
43e9d192 IB |
1252 | } |
1253 | else if (aarch64_uimm12_shift (val - ((val - comp) | ~mask))) | |
1254 | { | |
82614948 RR |
1255 | if (generate) |
1256 | { | |
1257 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1258 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1259 | GEN_INT ((val - comp) | ~mask))); | |
1260 | emit_insn (gen_adddi3 (dest, subtarget, | |
1261 | GEN_INT (val - ((val - comp) | ~mask)))); | |
1262 | } | |
1263 | num_insns += 2; | |
1264 | return num_insns; | |
43e9d192 IB |
1265 | } |
1266 | else if (aarch64_uimm12_shift (-(val - (val | ~mask)))) | |
1267 | { | |
82614948 RR |
1268 | if (generate) |
1269 | { | |
1270 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1271 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1272 | GEN_INT (val | ~mask))); | |
1273 | emit_insn (gen_adddi3 (dest, subtarget, | |
1274 | GEN_INT (val - (val | ~mask)))); | |
1275 | } | |
1276 | num_insns += 2; | |
1277 | return num_insns; | |
43e9d192 IB |
1278 | } |
1279 | } | |
1280 | ||
1281 | /* See if we can do it by arithmetically combining two | |
1282 | immediates. */ | |
1283 | for (i = 0; i < AARCH64_NUM_BITMASKS; i++) | |
1284 | { | |
1285 | int j; | |
1286 | mask = 0xffff; | |
1287 | ||
1288 | if (aarch64_uimm12_shift (val - aarch64_bitmasks[i]) | |
1289 | || aarch64_uimm12_shift (-val + aarch64_bitmasks[i])) | |
1290 | { | |
82614948 RR |
1291 | if (generate) |
1292 | { | |
1293 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1294 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1295 | GEN_INT (aarch64_bitmasks[i]))); | |
1296 | emit_insn (gen_adddi3 (dest, subtarget, | |
1297 | GEN_INT (val - aarch64_bitmasks[i]))); | |
1298 | } | |
1299 | num_insns += 2; | |
1300 | return num_insns; | |
43e9d192 IB |
1301 | } |
1302 | ||
1303 | for (j = 0; j < 64; j += 16, mask <<= 16) | |
1304 | { | |
1305 | if ((aarch64_bitmasks[i] & ~mask) == (val & ~mask)) | |
1306 | { | |
82614948 RR |
1307 | if (generate) |
1308 | { | |
1309 | emit_insn (gen_rtx_SET (VOIDmode, dest, | |
1310 | GEN_INT (aarch64_bitmasks[i]))); | |
1311 | emit_insn (gen_insv_immdi (dest, GEN_INT (j), | |
1312 | GEN_INT ((val >> j) & 0xffff))); | |
1313 | } | |
1314 | num_insns += 2; | |
1315 | return num_insns; | |
43e9d192 IB |
1316 | } |
1317 | } | |
1318 | } | |
1319 | ||
1320 | /* See if we can do it by logically combining two immediates. */ | |
1321 | for (i = 0; i < AARCH64_NUM_BITMASKS; i++) | |
1322 | { | |
1323 | if ((aarch64_bitmasks[i] & val) == aarch64_bitmasks[i]) | |
1324 | { | |
1325 | int j; | |
1326 | ||
1327 | for (j = i + 1; j < AARCH64_NUM_BITMASKS; j++) | |
1328 | if (val == (aarch64_bitmasks[i] | aarch64_bitmasks[j])) | |
1329 | { | |
82614948 RR |
1330 | if (generate) |
1331 | { | |
1332 | subtarget = subtargets ? gen_reg_rtx (mode) : dest; | |
1333 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1334 | GEN_INT (aarch64_bitmasks[i]))); | |
1335 | emit_insn (gen_iordi3 (dest, subtarget, | |
1336 | GEN_INT (aarch64_bitmasks[j]))); | |
1337 | } | |
1338 | num_insns += 2; | |
1339 | return num_insns; | |
43e9d192 IB |
1340 | } |
1341 | } | |
1342 | else if ((val & aarch64_bitmasks[i]) == val) | |
1343 | { | |
1344 | int j; | |
1345 | ||
1346 | for (j = i + 1; j < AARCH64_NUM_BITMASKS; j++) | |
1347 | if (val == (aarch64_bitmasks[j] & aarch64_bitmasks[i])) | |
1348 | { | |
82614948 RR |
1349 | if (generate) |
1350 | { | |
1351 | subtarget = subtargets ? gen_reg_rtx (mode) : dest; | |
1352 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1353 | GEN_INT (aarch64_bitmasks[j]))); | |
1354 | emit_insn (gen_anddi3 (dest, subtarget, | |
1355 | GEN_INT (aarch64_bitmasks[i]))); | |
1356 | } | |
1357 | num_insns += 2; | |
1358 | return num_insns; | |
43e9d192 IB |
1359 | } |
1360 | } | |
1361 | } | |
1362 | ||
2c274197 KT |
1363 | if (one_match > zero_match) |
1364 | { | |
1365 | /* Set either first three quarters or all but the third. */ | |
1366 | mask = 0xffffll << (16 - first_not_ffff_match); | |
82614948 RR |
1367 | if (generate) |
1368 | emit_insn (gen_rtx_SET (VOIDmode, dest, | |
1369 | GEN_INT (val | mask | 0xffffffff00000000ull))); | |
1370 | num_insns ++; | |
2c274197 KT |
1371 | |
1372 | /* Now insert other two quarters. */ | |
1373 | for (i = first_not_ffff_match + 16, mask <<= (first_not_ffff_match << 1); | |
1374 | i < 64; i += 16, mask <<= 16) | |
1375 | { | |
1376 | if ((val & mask) != mask) | |
82614948 RR |
1377 | { |
1378 | if (generate) | |
1379 | emit_insn (gen_insv_immdi (dest, GEN_INT (i), | |
1380 | GEN_INT ((val >> i) & 0xffff))); | |
1381 | num_insns ++; | |
1382 | } | |
2c274197 | 1383 | } |
82614948 | 1384 | return num_insns; |
2c274197 KT |
1385 | } |
1386 | ||
43e9d192 IB |
1387 | simple_sequence: |
1388 | first = true; | |
1389 | mask = 0xffff; | |
1390 | for (i = 0; i < 64; i += 16, mask <<= 16) | |
1391 | { | |
1392 | if ((val & mask) != 0) | |
1393 | { | |
1394 | if (first) | |
1395 | { | |
82614948 RR |
1396 | if (generate) |
1397 | emit_insn (gen_rtx_SET (VOIDmode, dest, | |
1398 | GEN_INT (val & mask))); | |
1399 | num_insns ++; | |
43e9d192 IB |
1400 | first = false; |
1401 | } | |
1402 | else | |
82614948 RR |
1403 | { |
1404 | if (generate) | |
1405 | emit_insn (gen_insv_immdi (dest, GEN_INT (i), | |
1406 | GEN_INT ((val >> i) & 0xffff))); | |
1407 | num_insns ++; | |
1408 | } | |
1409 | } | |
1410 | } | |
1411 | ||
1412 | return num_insns; | |
1413 | } | |
1414 | ||
1415 | ||
1416 | void | |
1417 | aarch64_expand_mov_immediate (rtx dest, rtx imm) | |
1418 | { | |
1419 | machine_mode mode = GET_MODE (dest); | |
1420 | ||
1421 | gcc_assert (mode == SImode || mode == DImode); | |
1422 | ||
1423 | /* Check on what type of symbol it is. */ | |
1424 | if (GET_CODE (imm) == SYMBOL_REF | |
1425 | || GET_CODE (imm) == LABEL_REF | |
1426 | || GET_CODE (imm) == CONST) | |
1427 | { | |
1428 | rtx mem, base, offset; | |
1429 | enum aarch64_symbol_type sty; | |
1430 | ||
1431 | /* If we have (const (plus symbol offset)), separate out the offset | |
1432 | before we start classifying the symbol. */ | |
1433 | split_const (imm, &base, &offset); | |
1434 | ||
f8b756b7 | 1435 | sty = aarch64_classify_symbol (base, offset, SYMBOL_CONTEXT_ADR); |
82614948 RR |
1436 | switch (sty) |
1437 | { | |
1438 | case SYMBOL_FORCE_TO_MEM: | |
1439 | if (offset != const0_rtx | |
1440 | && targetm.cannot_force_const_mem (mode, imm)) | |
1441 | { | |
1442 | gcc_assert (can_create_pseudo_p ()); | |
1443 | base = aarch64_force_temporary (mode, dest, base); | |
1444 | base = aarch64_add_offset (mode, NULL, base, INTVAL (offset)); | |
1445 | aarch64_emit_move (dest, base); | |
1446 | return; | |
1447 | } | |
1448 | mem = force_const_mem (ptr_mode, imm); | |
1449 | gcc_assert (mem); | |
1450 | if (mode != ptr_mode) | |
1451 | mem = gen_rtx_ZERO_EXTEND (mode, mem); | |
1452 | emit_insn (gen_rtx_SET (VOIDmode, dest, mem)); | |
1453 | return; | |
1454 | ||
1455 | case SYMBOL_SMALL_TLSGD: | |
1456 | case SYMBOL_SMALL_TLSDESC: | |
1457 | case SYMBOL_SMALL_GOTTPREL: | |
1458 | case SYMBOL_SMALL_GOT: | |
1459 | case SYMBOL_TINY_GOT: | |
1460 | if (offset != const0_rtx) | |
1461 | { | |
1462 | gcc_assert(can_create_pseudo_p ()); | |
1463 | base = aarch64_force_temporary (mode, dest, base); | |
1464 | base = aarch64_add_offset (mode, NULL, base, INTVAL (offset)); | |
1465 | aarch64_emit_move (dest, base); | |
1466 | return; | |
1467 | } | |
1468 | /* FALLTHRU */ | |
1469 | ||
1470 | case SYMBOL_SMALL_TPREL: | |
1471 | case SYMBOL_SMALL_ABSOLUTE: | |
1472 | case SYMBOL_TINY_ABSOLUTE: | |
1473 | aarch64_load_symref_appropriately (dest, imm, sty); | |
1474 | return; | |
1475 | ||
1476 | default: | |
1477 | gcc_unreachable (); | |
1478 | } | |
1479 | } | |
1480 | ||
1481 | if (!CONST_INT_P (imm)) | |
1482 | { | |
1483 | if (GET_CODE (imm) == HIGH) | |
1484 | emit_insn (gen_rtx_SET (VOIDmode, dest, imm)); | |
1485 | else | |
1486 | { | |
1487 | rtx mem = force_const_mem (mode, imm); | |
1488 | gcc_assert (mem); | |
1489 | emit_insn (gen_rtx_SET (VOIDmode, dest, mem)); | |
43e9d192 | 1490 | } |
82614948 RR |
1491 | |
1492 | return; | |
43e9d192 | 1493 | } |
82614948 RR |
1494 | |
1495 | aarch64_internal_mov_immediate (dest, imm, true, GET_MODE (dest)); | |
43e9d192 IB |
1496 | } |
1497 | ||
1498 | static bool | |
fee9ba42 JW |
1499 | aarch64_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED, |
1500 | tree exp ATTRIBUTE_UNUSED) | |
43e9d192 | 1501 | { |
fee9ba42 | 1502 | /* Currently, always true. */ |
43e9d192 IB |
1503 | return true; |
1504 | } | |
1505 | ||
1506 | /* Implement TARGET_PASS_BY_REFERENCE. */ | |
1507 | ||
1508 | static bool | |
1509 | aarch64_pass_by_reference (cumulative_args_t pcum ATTRIBUTE_UNUSED, | |
ef4bddc2 | 1510 | machine_mode mode, |
43e9d192 IB |
1511 | const_tree type, |
1512 | bool named ATTRIBUTE_UNUSED) | |
1513 | { | |
1514 | HOST_WIDE_INT size; | |
ef4bddc2 | 1515 | machine_mode dummymode; |
43e9d192 IB |
1516 | int nregs; |
1517 | ||
1518 | /* GET_MODE_SIZE (BLKmode) is useless since it is 0. */ | |
1519 | size = (mode == BLKmode && type) | |
1520 | ? int_size_in_bytes (type) : (int) GET_MODE_SIZE (mode); | |
1521 | ||
aadc1c43 MHD |
1522 | /* Aggregates are passed by reference based on their size. */ |
1523 | if (type && AGGREGATE_TYPE_P (type)) | |
43e9d192 | 1524 | { |
aadc1c43 | 1525 | size = int_size_in_bytes (type); |
43e9d192 IB |
1526 | } |
1527 | ||
1528 | /* Variable sized arguments are always returned by reference. */ | |
1529 | if (size < 0) | |
1530 | return true; | |
1531 | ||
1532 | /* Can this be a candidate to be passed in fp/simd register(s)? */ | |
1533 | if (aarch64_vfp_is_call_or_return_candidate (mode, type, | |
1534 | &dummymode, &nregs, | |
1535 | NULL)) | |
1536 | return false; | |
1537 | ||
1538 | /* Arguments which are variable sized or larger than 2 registers are | |
1539 | passed by reference unless they are a homogenous floating point | |
1540 | aggregate. */ | |
1541 | return size > 2 * UNITS_PER_WORD; | |
1542 | } | |
1543 | ||
1544 | /* Return TRUE if VALTYPE is padded to its least significant bits. */ | |
1545 | static bool | |
1546 | aarch64_return_in_msb (const_tree valtype) | |
1547 | { | |
ef4bddc2 | 1548 | machine_mode dummy_mode; |
43e9d192 IB |
1549 | int dummy_int; |
1550 | ||
1551 | /* Never happens in little-endian mode. */ | |
1552 | if (!BYTES_BIG_ENDIAN) | |
1553 | return false; | |
1554 | ||
1555 | /* Only composite types smaller than or equal to 16 bytes can | |
1556 | be potentially returned in registers. */ | |
1557 | if (!aarch64_composite_type_p (valtype, TYPE_MODE (valtype)) | |
1558 | || int_size_in_bytes (valtype) <= 0 | |
1559 | || int_size_in_bytes (valtype) > 16) | |
1560 | return false; | |
1561 | ||
1562 | /* But not a composite that is an HFA (Homogeneous Floating-point Aggregate) | |
1563 | or an HVA (Homogeneous Short-Vector Aggregate); such a special composite | |
1564 | is always passed/returned in the least significant bits of fp/simd | |
1565 | register(s). */ | |
1566 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (valtype), valtype, | |
1567 | &dummy_mode, &dummy_int, NULL)) | |
1568 | return false; | |
1569 | ||
1570 | return true; | |
1571 | } | |
1572 | ||
1573 | /* Implement TARGET_FUNCTION_VALUE. | |
1574 | Define how to find the value returned by a function. */ | |
1575 | ||
1576 | static rtx | |
1577 | aarch64_function_value (const_tree type, const_tree func, | |
1578 | bool outgoing ATTRIBUTE_UNUSED) | |
1579 | { | |
ef4bddc2 | 1580 | machine_mode mode; |
43e9d192 IB |
1581 | int unsignedp; |
1582 | int count; | |
ef4bddc2 | 1583 | machine_mode ag_mode; |
43e9d192 IB |
1584 | |
1585 | mode = TYPE_MODE (type); | |
1586 | if (INTEGRAL_TYPE_P (type)) | |
1587 | mode = promote_function_mode (type, mode, &unsignedp, func, 1); | |
1588 | ||
1589 | if (aarch64_return_in_msb (type)) | |
1590 | { | |
1591 | HOST_WIDE_INT size = int_size_in_bytes (type); | |
1592 | ||
1593 | if (size % UNITS_PER_WORD != 0) | |
1594 | { | |
1595 | size += UNITS_PER_WORD - size % UNITS_PER_WORD; | |
1596 | mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0); | |
1597 | } | |
1598 | } | |
1599 | ||
1600 | if (aarch64_vfp_is_call_or_return_candidate (mode, type, | |
1601 | &ag_mode, &count, NULL)) | |
1602 | { | |
1603 | if (!aarch64_composite_type_p (type, mode)) | |
1604 | { | |
1605 | gcc_assert (count == 1 && mode == ag_mode); | |
1606 | return gen_rtx_REG (mode, V0_REGNUM); | |
1607 | } | |
1608 | else | |
1609 | { | |
1610 | int i; | |
1611 | rtx par; | |
1612 | ||
1613 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (count)); | |
1614 | for (i = 0; i < count; i++) | |
1615 | { | |
1616 | rtx tmp = gen_rtx_REG (ag_mode, V0_REGNUM + i); | |
1617 | tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp, | |
1618 | GEN_INT (i * GET_MODE_SIZE (ag_mode))); | |
1619 | XVECEXP (par, 0, i) = tmp; | |
1620 | } | |
1621 | return par; | |
1622 | } | |
1623 | } | |
1624 | else | |
1625 | return gen_rtx_REG (mode, R0_REGNUM); | |
1626 | } | |
1627 | ||
1628 | /* Implements TARGET_FUNCTION_VALUE_REGNO_P. | |
1629 | Return true if REGNO is the number of a hard register in which the values | |
1630 | of called function may come back. */ | |
1631 | ||
1632 | static bool | |
1633 | aarch64_function_value_regno_p (const unsigned int regno) | |
1634 | { | |
1635 | /* Maximum of 16 bytes can be returned in the general registers. Examples | |
1636 | of 16-byte return values are: 128-bit integers and 16-byte small | |
1637 | structures (excluding homogeneous floating-point aggregates). */ | |
1638 | if (regno == R0_REGNUM || regno == R1_REGNUM) | |
1639 | return true; | |
1640 | ||
1641 | /* Up to four fp/simd registers can return a function value, e.g. a | |
1642 | homogeneous floating-point aggregate having four members. */ | |
1643 | if (regno >= V0_REGNUM && regno < V0_REGNUM + HA_MAX_NUM_FLDS) | |
1644 | return !TARGET_GENERAL_REGS_ONLY; | |
1645 | ||
1646 | return false; | |
1647 | } | |
1648 | ||
1649 | /* Implement TARGET_RETURN_IN_MEMORY. | |
1650 | ||
1651 | If the type T of the result of a function is such that | |
1652 | void func (T arg) | |
1653 | would require that arg be passed as a value in a register (or set of | |
1654 | registers) according to the parameter passing rules, then the result | |
1655 | is returned in the same registers as would be used for such an | |
1656 | argument. */ | |
1657 | ||
1658 | static bool | |
1659 | aarch64_return_in_memory (const_tree type, const_tree fndecl ATTRIBUTE_UNUSED) | |
1660 | { | |
1661 | HOST_WIDE_INT size; | |
ef4bddc2 | 1662 | machine_mode ag_mode; |
43e9d192 IB |
1663 | int count; |
1664 | ||
1665 | if (!AGGREGATE_TYPE_P (type) | |
1666 | && TREE_CODE (type) != COMPLEX_TYPE | |
1667 | && TREE_CODE (type) != VECTOR_TYPE) | |
1668 | /* Simple scalar types always returned in registers. */ | |
1669 | return false; | |
1670 | ||
1671 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (type), | |
1672 | type, | |
1673 | &ag_mode, | |
1674 | &count, | |
1675 | NULL)) | |
1676 | return false; | |
1677 | ||
1678 | /* Types larger than 2 registers returned in memory. */ | |
1679 | size = int_size_in_bytes (type); | |
1680 | return (size < 0 || size > 2 * UNITS_PER_WORD); | |
1681 | } | |
1682 | ||
1683 | static bool | |
ef4bddc2 | 1684 | aarch64_vfp_is_call_candidate (cumulative_args_t pcum_v, machine_mode mode, |
43e9d192 IB |
1685 | const_tree type, int *nregs) |
1686 | { | |
1687 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
1688 | return aarch64_vfp_is_call_or_return_candidate (mode, | |
1689 | type, | |
1690 | &pcum->aapcs_vfp_rmode, | |
1691 | nregs, | |
1692 | NULL); | |
1693 | } | |
1694 | ||
1695 | /* Given MODE and TYPE of a function argument, return the alignment in | |
1696 | bits. The idea is to suppress any stronger alignment requested by | |
1697 | the user and opt for the natural alignment (specified in AAPCS64 \S 4.1). | |
1698 | This is a helper function for local use only. */ | |
1699 | ||
1700 | static unsigned int | |
ef4bddc2 | 1701 | aarch64_function_arg_alignment (machine_mode mode, const_tree type) |
43e9d192 IB |
1702 | { |
1703 | unsigned int alignment; | |
1704 | ||
1705 | if (type) | |
1706 | { | |
1707 | if (!integer_zerop (TYPE_SIZE (type))) | |
1708 | { | |
1709 | if (TYPE_MODE (type) == mode) | |
1710 | alignment = TYPE_ALIGN (type); | |
1711 | else | |
1712 | alignment = GET_MODE_ALIGNMENT (mode); | |
1713 | } | |
1714 | else | |
1715 | alignment = 0; | |
1716 | } | |
1717 | else | |
1718 | alignment = GET_MODE_ALIGNMENT (mode); | |
1719 | ||
1720 | return alignment; | |
1721 | } | |
1722 | ||
1723 | /* Layout a function argument according to the AAPCS64 rules. The rule | |
1724 | numbers refer to the rule numbers in the AAPCS64. */ | |
1725 | ||
1726 | static void | |
ef4bddc2 | 1727 | aarch64_layout_arg (cumulative_args_t pcum_v, machine_mode mode, |
43e9d192 IB |
1728 | const_tree type, |
1729 | bool named ATTRIBUTE_UNUSED) | |
1730 | { | |
1731 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
1732 | int ncrn, nvrn, nregs; | |
1733 | bool allocate_ncrn, allocate_nvrn; | |
3abf17cf | 1734 | HOST_WIDE_INT size; |
43e9d192 IB |
1735 | |
1736 | /* We need to do this once per argument. */ | |
1737 | if (pcum->aapcs_arg_processed) | |
1738 | return; | |
1739 | ||
1740 | pcum->aapcs_arg_processed = true; | |
1741 | ||
3abf17cf YZ |
1742 | /* Size in bytes, rounded to the nearest multiple of 8 bytes. */ |
1743 | size | |
1744 | = AARCH64_ROUND_UP (type ? int_size_in_bytes (type) : GET_MODE_SIZE (mode), | |
1745 | UNITS_PER_WORD); | |
1746 | ||
43e9d192 IB |
1747 | allocate_ncrn = (type) ? !(FLOAT_TYPE_P (type)) : !FLOAT_MODE_P (mode); |
1748 | allocate_nvrn = aarch64_vfp_is_call_candidate (pcum_v, | |
1749 | mode, | |
1750 | type, | |
1751 | &nregs); | |
1752 | ||
1753 | /* allocate_ncrn may be false-positive, but allocate_nvrn is quite reliable. | |
1754 | The following code thus handles passing by SIMD/FP registers first. */ | |
1755 | ||
1756 | nvrn = pcum->aapcs_nvrn; | |
1757 | ||
1758 | /* C1 - C5 for floating point, homogenous floating point aggregates (HFA) | |
1759 | and homogenous short-vector aggregates (HVA). */ | |
1760 | if (allocate_nvrn) | |
1761 | { | |
1762 | if (nvrn + nregs <= NUM_FP_ARG_REGS) | |
1763 | { | |
1764 | pcum->aapcs_nextnvrn = nvrn + nregs; | |
1765 | if (!aarch64_composite_type_p (type, mode)) | |
1766 | { | |
1767 | gcc_assert (nregs == 1); | |
1768 | pcum->aapcs_reg = gen_rtx_REG (mode, V0_REGNUM + nvrn); | |
1769 | } | |
1770 | else | |
1771 | { | |
1772 | rtx par; | |
1773 | int i; | |
1774 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (nregs)); | |
1775 | for (i = 0; i < nregs; i++) | |
1776 | { | |
1777 | rtx tmp = gen_rtx_REG (pcum->aapcs_vfp_rmode, | |
1778 | V0_REGNUM + nvrn + i); | |
1779 | tmp = gen_rtx_EXPR_LIST | |
1780 | (VOIDmode, tmp, | |
1781 | GEN_INT (i * GET_MODE_SIZE (pcum->aapcs_vfp_rmode))); | |
1782 | XVECEXP (par, 0, i) = tmp; | |
1783 | } | |
1784 | pcum->aapcs_reg = par; | |
1785 | } | |
1786 | return; | |
1787 | } | |
1788 | else | |
1789 | { | |
1790 | /* C.3 NSRN is set to 8. */ | |
1791 | pcum->aapcs_nextnvrn = NUM_FP_ARG_REGS; | |
1792 | goto on_stack; | |
1793 | } | |
1794 | } | |
1795 | ||
1796 | ncrn = pcum->aapcs_ncrn; | |
3abf17cf | 1797 | nregs = size / UNITS_PER_WORD; |
43e9d192 IB |
1798 | |
1799 | /* C6 - C9. though the sign and zero extension semantics are | |
1800 | handled elsewhere. This is the case where the argument fits | |
1801 | entirely general registers. */ | |
1802 | if (allocate_ncrn && (ncrn + nregs <= NUM_ARG_REGS)) | |
1803 | { | |
1804 | unsigned int alignment = aarch64_function_arg_alignment (mode, type); | |
1805 | ||
1806 | gcc_assert (nregs == 0 || nregs == 1 || nregs == 2); | |
1807 | ||
1808 | /* C.8 if the argument has an alignment of 16 then the NGRN is | |
1809 | rounded up to the next even number. */ | |
1810 | if (nregs == 2 && alignment == 16 * BITS_PER_UNIT && ncrn % 2) | |
1811 | { | |
1812 | ++ncrn; | |
1813 | gcc_assert (ncrn + nregs <= NUM_ARG_REGS); | |
1814 | } | |
1815 | /* NREGS can be 0 when e.g. an empty structure is to be passed. | |
1816 | A reg is still generated for it, but the caller should be smart | |
1817 | enough not to use it. */ | |
1818 | if (nregs == 0 || nregs == 1 || GET_MODE_CLASS (mode) == MODE_INT) | |
1819 | { | |
1820 | pcum->aapcs_reg = gen_rtx_REG (mode, R0_REGNUM + ncrn); | |
1821 | } | |
1822 | else | |
1823 | { | |
1824 | rtx par; | |
1825 | int i; | |
1826 | ||
1827 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (nregs)); | |
1828 | for (i = 0; i < nregs; i++) | |
1829 | { | |
1830 | rtx tmp = gen_rtx_REG (word_mode, R0_REGNUM + ncrn + i); | |
1831 | tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp, | |
1832 | GEN_INT (i * UNITS_PER_WORD)); | |
1833 | XVECEXP (par, 0, i) = tmp; | |
1834 | } | |
1835 | pcum->aapcs_reg = par; | |
1836 | } | |
1837 | ||
1838 | pcum->aapcs_nextncrn = ncrn + nregs; | |
1839 | return; | |
1840 | } | |
1841 | ||
1842 | /* C.11 */ | |
1843 | pcum->aapcs_nextncrn = NUM_ARG_REGS; | |
1844 | ||
1845 | /* The argument is passed on stack; record the needed number of words for | |
3abf17cf | 1846 | this argument and align the total size if necessary. */ |
43e9d192 | 1847 | on_stack: |
3abf17cf | 1848 | pcum->aapcs_stack_words = size / UNITS_PER_WORD; |
43e9d192 IB |
1849 | if (aarch64_function_arg_alignment (mode, type) == 16 * BITS_PER_UNIT) |
1850 | pcum->aapcs_stack_size = AARCH64_ROUND_UP (pcum->aapcs_stack_size, | |
3abf17cf | 1851 | 16 / UNITS_PER_WORD); |
43e9d192 IB |
1852 | return; |
1853 | } | |
1854 | ||
1855 | /* Implement TARGET_FUNCTION_ARG. */ | |
1856 | ||
1857 | static rtx | |
ef4bddc2 | 1858 | aarch64_function_arg (cumulative_args_t pcum_v, machine_mode mode, |
43e9d192 IB |
1859 | const_tree type, bool named) |
1860 | { | |
1861 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
1862 | gcc_assert (pcum->pcs_variant == ARM_PCS_AAPCS64); | |
1863 | ||
1864 | if (mode == VOIDmode) | |
1865 | return NULL_RTX; | |
1866 | ||
1867 | aarch64_layout_arg (pcum_v, mode, type, named); | |
1868 | return pcum->aapcs_reg; | |
1869 | } | |
1870 | ||
1871 | void | |
1872 | aarch64_init_cumulative_args (CUMULATIVE_ARGS *pcum, | |
1873 | const_tree fntype ATTRIBUTE_UNUSED, | |
1874 | rtx libname ATTRIBUTE_UNUSED, | |
1875 | const_tree fndecl ATTRIBUTE_UNUSED, | |
1876 | unsigned n_named ATTRIBUTE_UNUSED) | |
1877 | { | |
1878 | pcum->aapcs_ncrn = 0; | |
1879 | pcum->aapcs_nvrn = 0; | |
1880 | pcum->aapcs_nextncrn = 0; | |
1881 | pcum->aapcs_nextnvrn = 0; | |
1882 | pcum->pcs_variant = ARM_PCS_AAPCS64; | |
1883 | pcum->aapcs_reg = NULL_RTX; | |
1884 | pcum->aapcs_arg_processed = false; | |
1885 | pcum->aapcs_stack_words = 0; | |
1886 | pcum->aapcs_stack_size = 0; | |
1887 | ||
1888 | return; | |
1889 | } | |
1890 | ||
1891 | static void | |
1892 | aarch64_function_arg_advance (cumulative_args_t pcum_v, | |
ef4bddc2 | 1893 | machine_mode mode, |
43e9d192 IB |
1894 | const_tree type, |
1895 | bool named) | |
1896 | { | |
1897 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
1898 | if (pcum->pcs_variant == ARM_PCS_AAPCS64) | |
1899 | { | |
1900 | aarch64_layout_arg (pcum_v, mode, type, named); | |
1901 | gcc_assert ((pcum->aapcs_reg != NULL_RTX) | |
1902 | != (pcum->aapcs_stack_words != 0)); | |
1903 | pcum->aapcs_arg_processed = false; | |
1904 | pcum->aapcs_ncrn = pcum->aapcs_nextncrn; | |
1905 | pcum->aapcs_nvrn = pcum->aapcs_nextnvrn; | |
1906 | pcum->aapcs_stack_size += pcum->aapcs_stack_words; | |
1907 | pcum->aapcs_stack_words = 0; | |
1908 | pcum->aapcs_reg = NULL_RTX; | |
1909 | } | |
1910 | } | |
1911 | ||
1912 | bool | |
1913 | aarch64_function_arg_regno_p (unsigned regno) | |
1914 | { | |
1915 | return ((GP_REGNUM_P (regno) && regno < R0_REGNUM + NUM_ARG_REGS) | |
1916 | || (FP_REGNUM_P (regno) && regno < V0_REGNUM + NUM_FP_ARG_REGS)); | |
1917 | } | |
1918 | ||
1919 | /* Implement FUNCTION_ARG_BOUNDARY. Every parameter gets at least | |
1920 | PARM_BOUNDARY bits of alignment, but will be given anything up | |
1921 | to STACK_BOUNDARY bits if the type requires it. This makes sure | |
1922 | that both before and after the layout of each argument, the Next | |
1923 | Stacked Argument Address (NSAA) will have a minimum alignment of | |
1924 | 8 bytes. */ | |
1925 | ||
1926 | static unsigned int | |
ef4bddc2 | 1927 | aarch64_function_arg_boundary (machine_mode mode, const_tree type) |
43e9d192 IB |
1928 | { |
1929 | unsigned int alignment = aarch64_function_arg_alignment (mode, type); | |
1930 | ||
1931 | if (alignment < PARM_BOUNDARY) | |
1932 | alignment = PARM_BOUNDARY; | |
1933 | if (alignment > STACK_BOUNDARY) | |
1934 | alignment = STACK_BOUNDARY; | |
1935 | return alignment; | |
1936 | } | |
1937 | ||
1938 | /* For use by FUNCTION_ARG_PADDING (MODE, TYPE). | |
1939 | ||
1940 | Return true if an argument passed on the stack should be padded upwards, | |
1941 | i.e. if the least-significant byte of the stack slot has useful data. | |
1942 | ||
1943 | Small aggregate types are placed in the lowest memory address. | |
1944 | ||
1945 | The related parameter passing rules are B.4, C.3, C.5 and C.14. */ | |
1946 | ||
1947 | bool | |
ef4bddc2 | 1948 | aarch64_pad_arg_upward (machine_mode mode, const_tree type) |
43e9d192 IB |
1949 | { |
1950 | /* On little-endian targets, the least significant byte of every stack | |
1951 | argument is passed at the lowest byte address of the stack slot. */ | |
1952 | if (!BYTES_BIG_ENDIAN) | |
1953 | return true; | |
1954 | ||
00edcfbe | 1955 | /* Otherwise, integral, floating-point and pointer types are padded downward: |
43e9d192 IB |
1956 | the least significant byte of a stack argument is passed at the highest |
1957 | byte address of the stack slot. */ | |
1958 | if (type | |
00edcfbe YZ |
1959 | ? (INTEGRAL_TYPE_P (type) || SCALAR_FLOAT_TYPE_P (type) |
1960 | || POINTER_TYPE_P (type)) | |
43e9d192 IB |
1961 | : (SCALAR_INT_MODE_P (mode) || SCALAR_FLOAT_MODE_P (mode))) |
1962 | return false; | |
1963 | ||
1964 | /* Everything else padded upward, i.e. data in first byte of stack slot. */ | |
1965 | return true; | |
1966 | } | |
1967 | ||
1968 | /* Similarly, for use by BLOCK_REG_PADDING (MODE, TYPE, FIRST). | |
1969 | ||
1970 | It specifies padding for the last (may also be the only) | |
1971 | element of a block move between registers and memory. If | |
1972 | assuming the block is in the memory, padding upward means that | |
1973 | the last element is padded after its highest significant byte, | |
1974 | while in downward padding, the last element is padded at the | |
1975 | its least significant byte side. | |
1976 | ||
1977 | Small aggregates and small complex types are always padded | |
1978 | upwards. | |
1979 | ||
1980 | We don't need to worry about homogeneous floating-point or | |
1981 | short-vector aggregates; their move is not affected by the | |
1982 | padding direction determined here. Regardless of endianness, | |
1983 | each element of such an aggregate is put in the least | |
1984 | significant bits of a fp/simd register. | |
1985 | ||
1986 | Return !BYTES_BIG_ENDIAN if the least significant byte of the | |
1987 | register has useful data, and return the opposite if the most | |
1988 | significant byte does. */ | |
1989 | ||
1990 | bool | |
ef4bddc2 | 1991 | aarch64_pad_reg_upward (machine_mode mode, const_tree type, |
43e9d192 IB |
1992 | bool first ATTRIBUTE_UNUSED) |
1993 | { | |
1994 | ||
1995 | /* Small composite types are always padded upward. */ | |
1996 | if (BYTES_BIG_ENDIAN && aarch64_composite_type_p (type, mode)) | |
1997 | { | |
1998 | HOST_WIDE_INT size = (type ? int_size_in_bytes (type) | |
1999 | : GET_MODE_SIZE (mode)); | |
2000 | if (size < 2 * UNITS_PER_WORD) | |
2001 | return true; | |
2002 | } | |
2003 | ||
2004 | /* Otherwise, use the default padding. */ | |
2005 | return !BYTES_BIG_ENDIAN; | |
2006 | } | |
2007 | ||
ef4bddc2 | 2008 | static machine_mode |
43e9d192 IB |
2009 | aarch64_libgcc_cmp_return_mode (void) |
2010 | { | |
2011 | return SImode; | |
2012 | } | |
2013 | ||
2014 | static bool | |
2015 | aarch64_frame_pointer_required (void) | |
2016 | { | |
0b7f8166 MS |
2017 | /* In aarch64_override_options_after_change |
2018 | flag_omit_leaf_frame_pointer turns off the frame pointer by | |
2019 | default. Turn it back on now if we've not got a leaf | |
2020 | function. */ | |
2021 | if (flag_omit_leaf_frame_pointer | |
2022 | && (!crtl->is_leaf || df_regs_ever_live_p (LR_REGNUM))) | |
2023 | return true; | |
43e9d192 | 2024 | |
0b7f8166 | 2025 | return false; |
43e9d192 IB |
2026 | } |
2027 | ||
2028 | /* Mark the registers that need to be saved by the callee and calculate | |
2029 | the size of the callee-saved registers area and frame record (both FP | |
2030 | and LR may be omitted). */ | |
2031 | static void | |
2032 | aarch64_layout_frame (void) | |
2033 | { | |
2034 | HOST_WIDE_INT offset = 0; | |
2035 | int regno; | |
2036 | ||
2037 | if (reload_completed && cfun->machine->frame.laid_out) | |
2038 | return; | |
2039 | ||
97826595 MS |
2040 | #define SLOT_NOT_REQUIRED (-2) |
2041 | #define SLOT_REQUIRED (-1) | |
2042 | ||
363ffa50 JW |
2043 | cfun->machine->frame.wb_candidate1 = FIRST_PSEUDO_REGISTER; |
2044 | cfun->machine->frame.wb_candidate2 = FIRST_PSEUDO_REGISTER; | |
2045 | ||
43e9d192 IB |
2046 | /* First mark all the registers that really need to be saved... */ |
2047 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) | |
97826595 | 2048 | cfun->machine->frame.reg_offset[regno] = SLOT_NOT_REQUIRED; |
43e9d192 IB |
2049 | |
2050 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
97826595 | 2051 | cfun->machine->frame.reg_offset[regno] = SLOT_NOT_REQUIRED; |
43e9d192 IB |
2052 | |
2053 | /* ... that includes the eh data registers (if needed)... */ | |
2054 | if (crtl->calls_eh_return) | |
2055 | for (regno = 0; EH_RETURN_DATA_REGNO (regno) != INVALID_REGNUM; regno++) | |
97826595 MS |
2056 | cfun->machine->frame.reg_offset[EH_RETURN_DATA_REGNO (regno)] |
2057 | = SLOT_REQUIRED; | |
43e9d192 IB |
2058 | |
2059 | /* ... and any callee saved register that dataflow says is live. */ | |
2060 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) | |
2061 | if (df_regs_ever_live_p (regno) | |
1c923b60 JW |
2062 | && (regno == R30_REGNUM |
2063 | || !call_used_regs[regno])) | |
97826595 | 2064 | cfun->machine->frame.reg_offset[regno] = SLOT_REQUIRED; |
43e9d192 IB |
2065 | |
2066 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
2067 | if (df_regs_ever_live_p (regno) | |
2068 | && !call_used_regs[regno]) | |
97826595 | 2069 | cfun->machine->frame.reg_offset[regno] = SLOT_REQUIRED; |
43e9d192 IB |
2070 | |
2071 | if (frame_pointer_needed) | |
2072 | { | |
2e1cdae5 | 2073 | /* FP and LR are placed in the linkage record. */ |
43e9d192 | 2074 | cfun->machine->frame.reg_offset[R29_REGNUM] = 0; |
363ffa50 | 2075 | cfun->machine->frame.wb_candidate1 = R29_REGNUM; |
2e1cdae5 | 2076 | cfun->machine->frame.reg_offset[R30_REGNUM] = UNITS_PER_WORD; |
363ffa50 | 2077 | cfun->machine->frame.wb_candidate2 = R30_REGNUM; |
43e9d192 | 2078 | cfun->machine->frame.hardfp_offset = 2 * UNITS_PER_WORD; |
2e1cdae5 | 2079 | offset += 2 * UNITS_PER_WORD; |
43e9d192 IB |
2080 | } |
2081 | ||
2082 | /* Now assign stack slots for them. */ | |
2e1cdae5 | 2083 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) |
97826595 | 2084 | if (cfun->machine->frame.reg_offset[regno] == SLOT_REQUIRED) |
43e9d192 IB |
2085 | { |
2086 | cfun->machine->frame.reg_offset[regno] = offset; | |
363ffa50 JW |
2087 | if (cfun->machine->frame.wb_candidate1 == FIRST_PSEUDO_REGISTER) |
2088 | cfun->machine->frame.wb_candidate1 = regno; | |
2089 | else if (cfun->machine->frame.wb_candidate2 == FIRST_PSEUDO_REGISTER) | |
2090 | cfun->machine->frame.wb_candidate2 = regno; | |
43e9d192 IB |
2091 | offset += UNITS_PER_WORD; |
2092 | } | |
2093 | ||
2094 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
97826595 | 2095 | if (cfun->machine->frame.reg_offset[regno] == SLOT_REQUIRED) |
43e9d192 IB |
2096 | { |
2097 | cfun->machine->frame.reg_offset[regno] = offset; | |
363ffa50 JW |
2098 | if (cfun->machine->frame.wb_candidate1 == FIRST_PSEUDO_REGISTER) |
2099 | cfun->machine->frame.wb_candidate1 = regno; | |
2100 | else if (cfun->machine->frame.wb_candidate2 == FIRST_PSEUDO_REGISTER | |
2101 | && cfun->machine->frame.wb_candidate1 >= V0_REGNUM) | |
2102 | cfun->machine->frame.wb_candidate2 = regno; | |
43e9d192 IB |
2103 | offset += UNITS_PER_WORD; |
2104 | } | |
2105 | ||
43e9d192 IB |
2106 | cfun->machine->frame.padding0 = |
2107 | (AARCH64_ROUND_UP (offset, STACK_BOUNDARY / BITS_PER_UNIT) - offset); | |
2108 | offset = AARCH64_ROUND_UP (offset, STACK_BOUNDARY / BITS_PER_UNIT); | |
2109 | ||
2110 | cfun->machine->frame.saved_regs_size = offset; | |
1c960e02 MS |
2111 | |
2112 | cfun->machine->frame.hard_fp_offset | |
2113 | = AARCH64_ROUND_UP (cfun->machine->frame.saved_varargs_size | |
2114 | + get_frame_size () | |
2115 | + cfun->machine->frame.saved_regs_size, | |
2116 | STACK_BOUNDARY / BITS_PER_UNIT); | |
2117 | ||
2118 | cfun->machine->frame.frame_size | |
2119 | = AARCH64_ROUND_UP (cfun->machine->frame.hard_fp_offset | |
2120 | + crtl->outgoing_args_size, | |
2121 | STACK_BOUNDARY / BITS_PER_UNIT); | |
2122 | ||
43e9d192 IB |
2123 | cfun->machine->frame.laid_out = true; |
2124 | } | |
2125 | ||
43e9d192 IB |
2126 | static bool |
2127 | aarch64_register_saved_on_entry (int regno) | |
2128 | { | |
97826595 | 2129 | return cfun->machine->frame.reg_offset[regno] >= 0; |
43e9d192 IB |
2130 | } |
2131 | ||
64dedd72 JW |
2132 | static unsigned |
2133 | aarch64_next_callee_save (unsigned regno, unsigned limit) | |
2134 | { | |
2135 | while (regno <= limit && !aarch64_register_saved_on_entry (regno)) | |
2136 | regno ++; | |
2137 | return regno; | |
2138 | } | |
43e9d192 | 2139 | |
c5e1f66e | 2140 | static void |
ef4bddc2 | 2141 | aarch64_pushwb_single_reg (machine_mode mode, unsigned regno, |
c5e1f66e JW |
2142 | HOST_WIDE_INT adjustment) |
2143 | { | |
2144 | rtx base_rtx = stack_pointer_rtx; | |
2145 | rtx insn, reg, mem; | |
2146 | ||
2147 | reg = gen_rtx_REG (mode, regno); | |
2148 | mem = gen_rtx_PRE_MODIFY (Pmode, base_rtx, | |
2149 | plus_constant (Pmode, base_rtx, -adjustment)); | |
2150 | mem = gen_rtx_MEM (mode, mem); | |
2151 | ||
2152 | insn = emit_move_insn (mem, reg); | |
2153 | RTX_FRAME_RELATED_P (insn) = 1; | |
2154 | } | |
2155 | ||
80c11907 | 2156 | static rtx |
ef4bddc2 | 2157 | aarch64_gen_storewb_pair (machine_mode mode, rtx base, rtx reg, rtx reg2, |
80c11907 JW |
2158 | HOST_WIDE_INT adjustment) |
2159 | { | |
2160 | switch (mode) | |
2161 | { | |
2162 | case DImode: | |
2163 | return gen_storewb_pairdi_di (base, base, reg, reg2, | |
2164 | GEN_INT (-adjustment), | |
2165 | GEN_INT (UNITS_PER_WORD - adjustment)); | |
2166 | case DFmode: | |
2167 | return gen_storewb_pairdf_di (base, base, reg, reg2, | |
2168 | GEN_INT (-adjustment), | |
2169 | GEN_INT (UNITS_PER_WORD - adjustment)); | |
2170 | default: | |
2171 | gcc_unreachable (); | |
2172 | } | |
2173 | } | |
2174 | ||
2175 | static void | |
ef4bddc2 | 2176 | aarch64_pushwb_pair_reg (machine_mode mode, unsigned regno1, |
80c11907 JW |
2177 | unsigned regno2, HOST_WIDE_INT adjustment) |
2178 | { | |
5d8a22a5 | 2179 | rtx_insn *insn; |
80c11907 JW |
2180 | rtx reg1 = gen_rtx_REG (mode, regno1); |
2181 | rtx reg2 = gen_rtx_REG (mode, regno2); | |
2182 | ||
2183 | insn = emit_insn (aarch64_gen_storewb_pair (mode, stack_pointer_rtx, reg1, | |
2184 | reg2, adjustment)); | |
2185 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 2)) = 1; | |
80c11907 JW |
2186 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; |
2187 | RTX_FRAME_RELATED_P (insn) = 1; | |
2188 | } | |
2189 | ||
159313d9 | 2190 | static rtx |
ef4bddc2 | 2191 | aarch64_gen_loadwb_pair (machine_mode mode, rtx base, rtx reg, rtx reg2, |
159313d9 JW |
2192 | HOST_WIDE_INT adjustment) |
2193 | { | |
2194 | switch (mode) | |
2195 | { | |
2196 | case DImode: | |
2197 | return gen_loadwb_pairdi_di (base, base, reg, reg2, GEN_INT (adjustment), | |
3e322b3f | 2198 | GEN_INT (UNITS_PER_WORD)); |
159313d9 JW |
2199 | case DFmode: |
2200 | return gen_loadwb_pairdf_di (base, base, reg, reg2, GEN_INT (adjustment), | |
3e322b3f | 2201 | GEN_INT (UNITS_PER_WORD)); |
159313d9 JW |
2202 | default: |
2203 | gcc_unreachable (); | |
2204 | } | |
2205 | } | |
2206 | ||
72df5c1f | 2207 | static rtx |
ef4bddc2 | 2208 | aarch64_gen_store_pair (machine_mode mode, rtx mem1, rtx reg1, rtx mem2, |
72df5c1f JW |
2209 | rtx reg2) |
2210 | { | |
2211 | switch (mode) | |
2212 | { | |
2213 | case DImode: | |
2214 | return gen_store_pairdi (mem1, reg1, mem2, reg2); | |
2215 | ||
2216 | case DFmode: | |
2217 | return gen_store_pairdf (mem1, reg1, mem2, reg2); | |
2218 | ||
2219 | default: | |
2220 | gcc_unreachable (); | |
2221 | } | |
2222 | } | |
2223 | ||
2224 | static rtx | |
ef4bddc2 | 2225 | aarch64_gen_load_pair (machine_mode mode, rtx reg1, rtx mem1, rtx reg2, |
72df5c1f JW |
2226 | rtx mem2) |
2227 | { | |
2228 | switch (mode) | |
2229 | { | |
2230 | case DImode: | |
2231 | return gen_load_pairdi (reg1, mem1, reg2, mem2); | |
2232 | ||
2233 | case DFmode: | |
2234 | return gen_load_pairdf (reg1, mem1, reg2, mem2); | |
2235 | ||
2236 | default: | |
2237 | gcc_unreachable (); | |
2238 | } | |
2239 | } | |
2240 | ||
43e9d192 | 2241 | |
43e9d192 | 2242 | static void |
ef4bddc2 | 2243 | aarch64_save_callee_saves (machine_mode mode, HOST_WIDE_INT start_offset, |
ae13fce3 | 2244 | unsigned start, unsigned limit, bool skip_wb) |
43e9d192 | 2245 | { |
5d8a22a5 | 2246 | rtx_insn *insn; |
ef4bddc2 | 2247 | rtx (*gen_mem_ref) (machine_mode, rtx) = (frame_pointer_needed |
a007a21c | 2248 | ? gen_frame_mem : gen_rtx_MEM); |
43e9d192 IB |
2249 | unsigned regno; |
2250 | unsigned regno2; | |
2251 | ||
0ec74a1e | 2252 | for (regno = aarch64_next_callee_save (start, limit); |
64dedd72 JW |
2253 | regno <= limit; |
2254 | regno = aarch64_next_callee_save (regno + 1, limit)) | |
43e9d192 | 2255 | { |
ae13fce3 JW |
2256 | rtx reg, mem; |
2257 | HOST_WIDE_INT offset; | |
64dedd72 | 2258 | |
ae13fce3 JW |
2259 | if (skip_wb |
2260 | && (regno == cfun->machine->frame.wb_candidate1 | |
2261 | || regno == cfun->machine->frame.wb_candidate2)) | |
2262 | continue; | |
2263 | ||
2264 | reg = gen_rtx_REG (mode, regno); | |
2265 | offset = start_offset + cfun->machine->frame.reg_offset[regno]; | |
0ec74a1e JW |
2266 | mem = gen_mem_ref (mode, plus_constant (Pmode, stack_pointer_rtx, |
2267 | offset)); | |
64dedd72 JW |
2268 | |
2269 | regno2 = aarch64_next_callee_save (regno + 1, limit); | |
2270 | ||
2271 | if (regno2 <= limit | |
2272 | && ((cfun->machine->frame.reg_offset[regno] + UNITS_PER_WORD) | |
2273 | == cfun->machine->frame.reg_offset[regno2])) | |
2274 | ||
43e9d192 | 2275 | { |
0ec74a1e | 2276 | rtx reg2 = gen_rtx_REG (mode, regno2); |
64dedd72 JW |
2277 | rtx mem2; |
2278 | ||
2279 | offset = start_offset + cfun->machine->frame.reg_offset[regno2]; | |
8ed2fc62 JW |
2280 | mem2 = gen_mem_ref (mode, plus_constant (Pmode, stack_pointer_rtx, |
2281 | offset)); | |
2282 | insn = emit_insn (aarch64_gen_store_pair (mode, mem, reg, mem2, | |
2283 | reg2)); | |
0b4a9743 | 2284 | |
64dedd72 JW |
2285 | /* The first part of a frame-related parallel insn is |
2286 | always assumed to be relevant to the frame | |
2287 | calculations; subsequent parts, are only | |
2288 | frame-related if explicitly marked. */ | |
2289 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; | |
2290 | regno = regno2; | |
2291 | } | |
2292 | else | |
8ed2fc62 JW |
2293 | insn = emit_move_insn (mem, reg); |
2294 | ||
2295 | RTX_FRAME_RELATED_P (insn) = 1; | |
2296 | } | |
2297 | } | |
2298 | ||
2299 | static void | |
ef4bddc2 | 2300 | aarch64_restore_callee_saves (machine_mode mode, |
8ed2fc62 | 2301 | HOST_WIDE_INT start_offset, unsigned start, |
dd991abb | 2302 | unsigned limit, bool skip_wb, rtx *cfi_ops) |
8ed2fc62 | 2303 | { |
8ed2fc62 | 2304 | rtx base_rtx = stack_pointer_rtx; |
ef4bddc2 | 2305 | rtx (*gen_mem_ref) (machine_mode, rtx) = (frame_pointer_needed |
8ed2fc62 JW |
2306 | ? gen_frame_mem : gen_rtx_MEM); |
2307 | unsigned regno; | |
2308 | unsigned regno2; | |
2309 | HOST_WIDE_INT offset; | |
2310 | ||
2311 | for (regno = aarch64_next_callee_save (start, limit); | |
2312 | regno <= limit; | |
2313 | regno = aarch64_next_callee_save (regno + 1, limit)) | |
2314 | { | |
ae13fce3 | 2315 | rtx reg, mem; |
8ed2fc62 | 2316 | |
ae13fce3 JW |
2317 | if (skip_wb |
2318 | && (regno == cfun->machine->frame.wb_candidate1 | |
2319 | || regno == cfun->machine->frame.wb_candidate2)) | |
2320 | continue; | |
2321 | ||
2322 | reg = gen_rtx_REG (mode, regno); | |
8ed2fc62 JW |
2323 | offset = start_offset + cfun->machine->frame.reg_offset[regno]; |
2324 | mem = gen_mem_ref (mode, plus_constant (Pmode, base_rtx, offset)); | |
2325 | ||
2326 | regno2 = aarch64_next_callee_save (regno + 1, limit); | |
2327 | ||
2328 | if (regno2 <= limit | |
2329 | && ((cfun->machine->frame.reg_offset[regno] + UNITS_PER_WORD) | |
2330 | == cfun->machine->frame.reg_offset[regno2])) | |
64dedd72 | 2331 | { |
8ed2fc62 JW |
2332 | rtx reg2 = gen_rtx_REG (mode, regno2); |
2333 | rtx mem2; | |
2334 | ||
2335 | offset = start_offset + cfun->machine->frame.reg_offset[regno2]; | |
2336 | mem2 = gen_mem_ref (mode, plus_constant (Pmode, base_rtx, offset)); | |
dd991abb | 2337 | emit_insn (aarch64_gen_load_pair (mode, reg, mem, reg2, mem2)); |
8ed2fc62 | 2338 | |
dd991abb | 2339 | *cfi_ops = alloc_reg_note (REG_CFA_RESTORE, reg2, *cfi_ops); |
8ed2fc62 | 2340 | regno = regno2; |
43e9d192 | 2341 | } |
8ed2fc62 | 2342 | else |
dd991abb RH |
2343 | emit_move_insn (reg, mem); |
2344 | *cfi_ops = alloc_reg_note (REG_CFA_RESTORE, reg, *cfi_ops); | |
43e9d192 | 2345 | } |
43e9d192 IB |
2346 | } |
2347 | ||
2348 | /* AArch64 stack frames generated by this compiler look like: | |
2349 | ||
2350 | +-------------------------------+ | |
2351 | | | | |
2352 | | incoming stack arguments | | |
2353 | | | | |
34834420 MS |
2354 | +-------------------------------+ |
2355 | | | <-- incoming stack pointer (aligned) | |
43e9d192 IB |
2356 | | callee-allocated save area | |
2357 | | for register varargs | | |
2358 | | | | |
34834420 MS |
2359 | +-------------------------------+ |
2360 | | local variables | <-- frame_pointer_rtx | |
43e9d192 IB |
2361 | | | |
2362 | +-------------------------------+ | |
454fdba9 RL |
2363 | | padding0 | \ |
2364 | +-------------------------------+ | | |
454fdba9 | 2365 | | callee-saved registers | | frame.saved_regs_size |
454fdba9 RL |
2366 | +-------------------------------+ | |
2367 | | LR' | | | |
2368 | +-------------------------------+ | | |
34834420 MS |
2369 | | FP' | / <- hard_frame_pointer_rtx (aligned) |
2370 | +-------------------------------+ | |
43e9d192 IB |
2371 | | dynamic allocation | |
2372 | +-------------------------------+ | |
34834420 MS |
2373 | | padding | |
2374 | +-------------------------------+ | |
2375 | | outgoing stack arguments | <-- arg_pointer | |
2376 | | | | |
2377 | +-------------------------------+ | |
2378 | | | <-- stack_pointer_rtx (aligned) | |
43e9d192 | 2379 | |
34834420 MS |
2380 | Dynamic stack allocations via alloca() decrease stack_pointer_rtx |
2381 | but leave frame_pointer_rtx and hard_frame_pointer_rtx | |
2382 | unchanged. */ | |
43e9d192 IB |
2383 | |
2384 | /* Generate the prologue instructions for entry into a function. | |
2385 | Establish the stack frame by decreasing the stack pointer with a | |
2386 | properly calculated size and, if necessary, create a frame record | |
2387 | filled with the values of LR and previous frame pointer. The | |
6991c977 | 2388 | current FP is also set up if it is in use. */ |
43e9d192 IB |
2389 | |
2390 | void | |
2391 | aarch64_expand_prologue (void) | |
2392 | { | |
2393 | /* sub sp, sp, #<frame_size> | |
2394 | stp {fp, lr}, [sp, #<frame_size> - 16] | |
2395 | add fp, sp, #<frame_size> - hardfp_offset | |
2396 | stp {cs_reg}, [fp, #-16] etc. | |
2397 | ||
2398 | sub sp, sp, <final_adjustment_if_any> | |
2399 | */ | |
43e9d192 | 2400 | HOST_WIDE_INT frame_size, offset; |
1c960e02 | 2401 | HOST_WIDE_INT fp_offset; /* Offset from hard FP to SP. */ |
dd991abb | 2402 | HOST_WIDE_INT hard_fp_offset; |
5d8a22a5 | 2403 | rtx_insn *insn; |
43e9d192 IB |
2404 | |
2405 | aarch64_layout_frame (); | |
43e9d192 | 2406 | |
dd991abb RH |
2407 | offset = frame_size = cfun->machine->frame.frame_size; |
2408 | hard_fp_offset = cfun->machine->frame.hard_fp_offset; | |
2409 | fp_offset = frame_size - hard_fp_offset; | |
43e9d192 | 2410 | |
dd991abb RH |
2411 | if (flag_stack_usage_info) |
2412 | current_function_static_stack_size = frame_size; | |
43e9d192 | 2413 | |
44c0e7b9 | 2414 | /* Store pairs and load pairs have a range only -512 to 504. */ |
43e9d192 IB |
2415 | if (offset >= 512) |
2416 | { | |
2417 | /* When the frame has a large size, an initial decrease is done on | |
2418 | the stack pointer to jump over the callee-allocated save area for | |
2419 | register varargs, the local variable area and/or the callee-saved | |
2420 | register area. This will allow the pre-index write-back | |
2421 | store pair instructions to be used for setting up the stack frame | |
2422 | efficiently. */ | |
dd991abb | 2423 | offset = hard_fp_offset; |
43e9d192 IB |
2424 | if (offset >= 512) |
2425 | offset = cfun->machine->frame.saved_regs_size; | |
2426 | ||
2427 | frame_size -= (offset + crtl->outgoing_args_size); | |
2428 | fp_offset = 0; | |
2429 | ||
2430 | if (frame_size >= 0x1000000) | |
2431 | { | |
2432 | rtx op0 = gen_rtx_REG (Pmode, IP0_REGNUM); | |
2433 | emit_move_insn (op0, GEN_INT (-frame_size)); | |
dd991abb RH |
2434 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, op0)); |
2435 | ||
2436 | add_reg_note (insn, REG_CFA_ADJUST_CFA, | |
2437 | gen_rtx_SET (VOIDmode, stack_pointer_rtx, | |
2438 | plus_constant (Pmode, stack_pointer_rtx, | |
2439 | -frame_size))); | |
2440 | RTX_FRAME_RELATED_P (insn) = 1; | |
43e9d192 IB |
2441 | } |
2442 | else if (frame_size > 0) | |
2443 | { | |
dd991abb RH |
2444 | int hi_ofs = frame_size & 0xfff000; |
2445 | int lo_ofs = frame_size & 0x000fff; | |
2446 | ||
2447 | if (hi_ofs) | |
43e9d192 IB |
2448 | { |
2449 | insn = emit_insn (gen_add2_insn | |
dd991abb | 2450 | (stack_pointer_rtx, GEN_INT (-hi_ofs))); |
43e9d192 IB |
2451 | RTX_FRAME_RELATED_P (insn) = 1; |
2452 | } | |
dd991abb | 2453 | if (lo_ofs) |
43e9d192 IB |
2454 | { |
2455 | insn = emit_insn (gen_add2_insn | |
dd991abb | 2456 | (stack_pointer_rtx, GEN_INT (-lo_ofs))); |
43e9d192 IB |
2457 | RTX_FRAME_RELATED_P (insn) = 1; |
2458 | } | |
2459 | } | |
2460 | } | |
2461 | else | |
2462 | frame_size = -1; | |
2463 | ||
2464 | if (offset > 0) | |
2465 | { | |
ae13fce3 JW |
2466 | bool skip_wb = false; |
2467 | ||
43e9d192 IB |
2468 | if (frame_pointer_needed) |
2469 | { | |
c5e1f66e JW |
2470 | skip_wb = true; |
2471 | ||
43e9d192 IB |
2472 | if (fp_offset) |
2473 | { | |
2474 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, | |
2475 | GEN_INT (-offset))); | |
2476 | RTX_FRAME_RELATED_P (insn) = 1; | |
80c11907 JW |
2477 | |
2478 | aarch64_save_callee_saves (DImode, fp_offset, R29_REGNUM, | |
c5e1f66e | 2479 | R30_REGNUM, false); |
43e9d192 IB |
2480 | } |
2481 | else | |
80c11907 | 2482 | aarch64_pushwb_pair_reg (DImode, R29_REGNUM, R30_REGNUM, offset); |
43e9d192 IB |
2483 | |
2484 | /* Set up frame pointer to point to the location of the | |
2485 | previous frame pointer on the stack. */ | |
2486 | insn = emit_insn (gen_add3_insn (hard_frame_pointer_rtx, | |
2487 | stack_pointer_rtx, | |
2488 | GEN_INT (fp_offset))); | |
43e9d192 | 2489 | RTX_FRAME_RELATED_P (insn) = 1; |
dd991abb | 2490 | emit_insn (gen_stack_tie (stack_pointer_rtx, hard_frame_pointer_rtx)); |
43e9d192 IB |
2491 | } |
2492 | else | |
2493 | { | |
c5e1f66e JW |
2494 | unsigned reg1 = cfun->machine->frame.wb_candidate1; |
2495 | unsigned reg2 = cfun->machine->frame.wb_candidate2; | |
80c11907 | 2496 | |
c5e1f66e JW |
2497 | if (fp_offset |
2498 | || reg1 == FIRST_PSEUDO_REGISTER | |
2499 | || (reg2 == FIRST_PSEUDO_REGISTER | |
2500 | && offset >= 256)) | |
2501 | { | |
2502 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, | |
2503 | GEN_INT (-offset))); | |
2504 | RTX_FRAME_RELATED_P (insn) = 1; | |
2505 | } | |
2506 | else | |
2507 | { | |
ef4bddc2 | 2508 | machine_mode mode1 = (reg1 <= R30_REGNUM) ? DImode : DFmode; |
c5e1f66e JW |
2509 | |
2510 | skip_wb = true; | |
2511 | ||
2512 | if (reg2 == FIRST_PSEUDO_REGISTER) | |
2513 | aarch64_pushwb_single_reg (mode1, reg1, offset); | |
2514 | else | |
2515 | aarch64_pushwb_pair_reg (mode1, reg1, reg2, offset); | |
2516 | } | |
43e9d192 IB |
2517 | } |
2518 | ||
c5e1f66e JW |
2519 | aarch64_save_callee_saves (DImode, fp_offset, R0_REGNUM, R30_REGNUM, |
2520 | skip_wb); | |
ae13fce3 JW |
2521 | aarch64_save_callee_saves (DFmode, fp_offset, V0_REGNUM, V31_REGNUM, |
2522 | skip_wb); | |
43e9d192 IB |
2523 | } |
2524 | ||
2525 | /* when offset >= 512, | |
2526 | sub sp, sp, #<outgoing_args_size> */ | |
2527 | if (frame_size > -1) | |
2528 | { | |
2529 | if (crtl->outgoing_args_size > 0) | |
2530 | { | |
2531 | insn = emit_insn (gen_add2_insn | |
2532 | (stack_pointer_rtx, | |
2533 | GEN_INT (- crtl->outgoing_args_size))); | |
2534 | RTX_FRAME_RELATED_P (insn) = 1; | |
2535 | } | |
2536 | } | |
2537 | } | |
2538 | ||
4f942779 RL |
2539 | /* Return TRUE if we can use a simple_return insn. |
2540 | ||
2541 | This function checks whether the callee saved stack is empty, which | |
2542 | means no restore actions are need. The pro_and_epilogue will use | |
2543 | this to check whether shrink-wrapping opt is feasible. */ | |
2544 | ||
2545 | bool | |
2546 | aarch64_use_return_insn_p (void) | |
2547 | { | |
2548 | if (!reload_completed) | |
2549 | return false; | |
2550 | ||
2551 | if (crtl->profile) | |
2552 | return false; | |
2553 | ||
2554 | aarch64_layout_frame (); | |
2555 | ||
2556 | return cfun->machine->frame.frame_size == 0; | |
2557 | } | |
2558 | ||
43e9d192 IB |
2559 | /* Generate the epilogue instructions for returning from a function. */ |
2560 | void | |
2561 | aarch64_expand_epilogue (bool for_sibcall) | |
2562 | { | |
1c960e02 | 2563 | HOST_WIDE_INT frame_size, offset; |
43e9d192 | 2564 | HOST_WIDE_INT fp_offset; |
dd991abb | 2565 | HOST_WIDE_INT hard_fp_offset; |
5d8a22a5 | 2566 | rtx_insn *insn; |
7e8c2bd5 JW |
2567 | /* We need to add memory barrier to prevent read from deallocated stack. */ |
2568 | bool need_barrier_p = (get_frame_size () != 0 | |
2569 | || cfun->machine->frame.saved_varargs_size); | |
43e9d192 IB |
2570 | |
2571 | aarch64_layout_frame (); | |
43e9d192 | 2572 | |
1c960e02 | 2573 | offset = frame_size = cfun->machine->frame.frame_size; |
dd991abb RH |
2574 | hard_fp_offset = cfun->machine->frame.hard_fp_offset; |
2575 | fp_offset = frame_size - hard_fp_offset; | |
44c0e7b9 YZ |
2576 | |
2577 | /* Store pairs and load pairs have a range only -512 to 504. */ | |
43e9d192 IB |
2578 | if (offset >= 512) |
2579 | { | |
dd991abb | 2580 | offset = hard_fp_offset; |
43e9d192 IB |
2581 | if (offset >= 512) |
2582 | offset = cfun->machine->frame.saved_regs_size; | |
2583 | ||
2584 | frame_size -= (offset + crtl->outgoing_args_size); | |
2585 | fp_offset = 0; | |
2586 | if (!frame_pointer_needed && crtl->outgoing_args_size > 0) | |
2587 | { | |
2588 | insn = emit_insn (gen_add2_insn | |
2589 | (stack_pointer_rtx, | |
2590 | GEN_INT (crtl->outgoing_args_size))); | |
2591 | RTX_FRAME_RELATED_P (insn) = 1; | |
2592 | } | |
2593 | } | |
2594 | else | |
2595 | frame_size = -1; | |
2596 | ||
2597 | /* If there were outgoing arguments or we've done dynamic stack | |
2598 | allocation, then restore the stack pointer from the frame | |
2599 | pointer. This is at most one insn and more efficient than using | |
2600 | GCC's internal mechanism. */ | |
2601 | if (frame_pointer_needed | |
2602 | && (crtl->outgoing_args_size || cfun->calls_alloca)) | |
2603 | { | |
7e8c2bd5 JW |
2604 | if (cfun->calls_alloca) |
2605 | emit_insn (gen_stack_tie (stack_pointer_rtx, stack_pointer_rtx)); | |
2606 | ||
43e9d192 IB |
2607 | insn = emit_insn (gen_add3_insn (stack_pointer_rtx, |
2608 | hard_frame_pointer_rtx, | |
8f454e9f JW |
2609 | GEN_INT (0))); |
2610 | offset = offset - fp_offset; | |
43e9d192 IB |
2611 | } |
2612 | ||
43e9d192 IB |
2613 | if (offset > 0) |
2614 | { | |
4b92caa1 JW |
2615 | unsigned reg1 = cfun->machine->frame.wb_candidate1; |
2616 | unsigned reg2 = cfun->machine->frame.wb_candidate2; | |
2617 | bool skip_wb = true; | |
dd991abb | 2618 | rtx cfi_ops = NULL; |
4b92caa1 | 2619 | |
43e9d192 | 2620 | if (frame_pointer_needed) |
4b92caa1 JW |
2621 | fp_offset = 0; |
2622 | else if (fp_offset | |
2623 | || reg1 == FIRST_PSEUDO_REGISTER | |
2624 | || (reg2 == FIRST_PSEUDO_REGISTER | |
2625 | && offset >= 256)) | |
2626 | skip_wb = false; | |
2627 | ||
2628 | aarch64_restore_callee_saves (DImode, fp_offset, R0_REGNUM, R30_REGNUM, | |
dd991abb | 2629 | skip_wb, &cfi_ops); |
4b92caa1 | 2630 | aarch64_restore_callee_saves (DFmode, fp_offset, V0_REGNUM, V31_REGNUM, |
dd991abb | 2631 | skip_wb, &cfi_ops); |
4b92caa1 | 2632 | |
7e8c2bd5 JW |
2633 | if (need_barrier_p) |
2634 | emit_insn (gen_stack_tie (stack_pointer_rtx, stack_pointer_rtx)); | |
2635 | ||
4b92caa1 | 2636 | if (skip_wb) |
43e9d192 | 2637 | { |
ef4bddc2 | 2638 | machine_mode mode1 = (reg1 <= R30_REGNUM) ? DImode : DFmode; |
dd991abb | 2639 | rtx rreg1 = gen_rtx_REG (mode1, reg1); |
4b92caa1 | 2640 | |
dd991abb | 2641 | cfi_ops = alloc_reg_note (REG_CFA_RESTORE, rreg1, cfi_ops); |
4b92caa1 | 2642 | if (reg2 == FIRST_PSEUDO_REGISTER) |
dd991abb RH |
2643 | { |
2644 | rtx mem = plus_constant (Pmode, stack_pointer_rtx, offset); | |
2645 | mem = gen_rtx_POST_MODIFY (Pmode, stack_pointer_rtx, mem); | |
2646 | mem = gen_rtx_MEM (mode1, mem); | |
2647 | insn = emit_move_insn (rreg1, mem); | |
2648 | } | |
4b92caa1 JW |
2649 | else |
2650 | { | |
dd991abb | 2651 | rtx rreg2 = gen_rtx_REG (mode1, reg2); |
4b92caa1 | 2652 | |
dd991abb RH |
2653 | cfi_ops = alloc_reg_note (REG_CFA_RESTORE, rreg2, cfi_ops); |
2654 | insn = emit_insn (aarch64_gen_loadwb_pair | |
2655 | (mode1, stack_pointer_rtx, rreg1, | |
2656 | rreg2, offset)); | |
4b92caa1 | 2657 | } |
43e9d192 | 2658 | } |
43e9d192 IB |
2659 | else |
2660 | { | |
2661 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, | |
2662 | GEN_INT (offset))); | |
43e9d192 | 2663 | } |
43e9d192 | 2664 | |
dd991abb RH |
2665 | /* Reset the CFA to be SP + FRAME_SIZE. */ |
2666 | rtx new_cfa = stack_pointer_rtx; | |
2667 | if (frame_size > 0) | |
2668 | new_cfa = plus_constant (Pmode, new_cfa, frame_size); | |
2669 | cfi_ops = alloc_reg_note (REG_CFA_DEF_CFA, new_cfa, cfi_ops); | |
2670 | REG_NOTES (insn) = cfi_ops; | |
43e9d192 | 2671 | RTX_FRAME_RELATED_P (insn) = 1; |
43e9d192 IB |
2672 | } |
2673 | ||
dd991abb | 2674 | if (frame_size > 0) |
43e9d192 | 2675 | { |
7e8c2bd5 JW |
2676 | if (need_barrier_p) |
2677 | emit_insn (gen_stack_tie (stack_pointer_rtx, stack_pointer_rtx)); | |
2678 | ||
43e9d192 IB |
2679 | if (frame_size >= 0x1000000) |
2680 | { | |
2681 | rtx op0 = gen_rtx_REG (Pmode, IP0_REGNUM); | |
2682 | emit_move_insn (op0, GEN_INT (frame_size)); | |
dd991abb | 2683 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, op0)); |
43e9d192 | 2684 | } |
dd991abb | 2685 | else |
43e9d192 | 2686 | { |
dd991abb RH |
2687 | int hi_ofs = frame_size & 0xfff000; |
2688 | int lo_ofs = frame_size & 0x000fff; | |
2689 | ||
2690 | if (hi_ofs && lo_ofs) | |
43e9d192 IB |
2691 | { |
2692 | insn = emit_insn (gen_add2_insn | |
dd991abb | 2693 | (stack_pointer_rtx, GEN_INT (hi_ofs))); |
43e9d192 | 2694 | RTX_FRAME_RELATED_P (insn) = 1; |
dd991abb | 2695 | frame_size = lo_ofs; |
43e9d192 | 2696 | } |
dd991abb RH |
2697 | insn = emit_insn (gen_add2_insn |
2698 | (stack_pointer_rtx, GEN_INT (frame_size))); | |
43e9d192 IB |
2699 | } |
2700 | ||
dd991abb RH |
2701 | /* Reset the CFA to be SP + 0. */ |
2702 | add_reg_note (insn, REG_CFA_DEF_CFA, stack_pointer_rtx); | |
2703 | RTX_FRAME_RELATED_P (insn) = 1; | |
2704 | } | |
2705 | ||
2706 | /* Stack adjustment for exception handler. */ | |
2707 | if (crtl->calls_eh_return) | |
2708 | { | |
2709 | /* We need to unwind the stack by the offset computed by | |
2710 | EH_RETURN_STACKADJ_RTX. We have already reset the CFA | |
2711 | to be SP; letting the CFA move during this adjustment | |
2712 | is just as correct as retaining the CFA from the body | |
2713 | of the function. Therefore, do nothing special. */ | |
2714 | emit_insn (gen_add2_insn (stack_pointer_rtx, EH_RETURN_STACKADJ_RTX)); | |
43e9d192 IB |
2715 | } |
2716 | ||
2717 | emit_use (gen_rtx_REG (DImode, LR_REGNUM)); | |
2718 | if (!for_sibcall) | |
2719 | emit_jump_insn (ret_rtx); | |
2720 | } | |
2721 | ||
2722 | /* Return the place to copy the exception unwinding return address to. | |
2723 | This will probably be a stack slot, but could (in theory be the | |
2724 | return register). */ | |
2725 | rtx | |
2726 | aarch64_final_eh_return_addr (void) | |
2727 | { | |
1c960e02 MS |
2728 | HOST_WIDE_INT fp_offset; |
2729 | ||
43e9d192 | 2730 | aarch64_layout_frame (); |
1c960e02 MS |
2731 | |
2732 | fp_offset = cfun->machine->frame.frame_size | |
2733 | - cfun->machine->frame.hard_fp_offset; | |
43e9d192 IB |
2734 | |
2735 | if (cfun->machine->frame.reg_offset[LR_REGNUM] < 0) | |
2736 | return gen_rtx_REG (DImode, LR_REGNUM); | |
2737 | ||
2738 | /* DSE and CSELIB do not detect an alias between sp+k1 and fp+k2. This can | |
2739 | result in a store to save LR introduced by builtin_eh_return () being | |
2740 | incorrectly deleted because the alias is not detected. | |
2741 | So in the calculation of the address to copy the exception unwinding | |
2742 | return address to, we note 2 cases. | |
2743 | If FP is needed and the fp_offset is 0, it means that SP = FP and hence | |
2744 | we return a SP-relative location since all the addresses are SP-relative | |
2745 | in this case. This prevents the store from being optimized away. | |
2746 | If the fp_offset is not 0, then the addresses will be FP-relative and | |
2747 | therefore we return a FP-relative location. */ | |
2748 | ||
2749 | if (frame_pointer_needed) | |
2750 | { | |
2751 | if (fp_offset) | |
2752 | return gen_frame_mem (DImode, | |
2753 | plus_constant (Pmode, hard_frame_pointer_rtx, UNITS_PER_WORD)); | |
2754 | else | |
2755 | return gen_frame_mem (DImode, | |
2756 | plus_constant (Pmode, stack_pointer_rtx, UNITS_PER_WORD)); | |
2757 | } | |
2758 | ||
2759 | /* If FP is not needed, we calculate the location of LR, which would be | |
2760 | at the top of the saved registers block. */ | |
2761 | ||
2762 | return gen_frame_mem (DImode, | |
2763 | plus_constant (Pmode, | |
2764 | stack_pointer_rtx, | |
2765 | fp_offset | |
2766 | + cfun->machine->frame.saved_regs_size | |
2767 | - 2 * UNITS_PER_WORD)); | |
2768 | } | |
2769 | ||
9dfc162c JG |
2770 | /* Possibly output code to build up a constant in a register. For |
2771 | the benefit of the costs infrastructure, returns the number of | |
2772 | instructions which would be emitted. GENERATE inhibits or | |
2773 | enables code generation. */ | |
2774 | ||
2775 | static int | |
2776 | aarch64_build_constant (int regnum, HOST_WIDE_INT val, bool generate) | |
43e9d192 | 2777 | { |
9dfc162c JG |
2778 | int insns = 0; |
2779 | ||
43e9d192 | 2780 | if (aarch64_bitmask_imm (val, DImode)) |
9dfc162c JG |
2781 | { |
2782 | if (generate) | |
2783 | emit_move_insn (gen_rtx_REG (Pmode, regnum), GEN_INT (val)); | |
2784 | insns = 1; | |
2785 | } | |
43e9d192 IB |
2786 | else |
2787 | { | |
2788 | int i; | |
2789 | int ncount = 0; | |
2790 | int zcount = 0; | |
2791 | HOST_WIDE_INT valp = val >> 16; | |
2792 | HOST_WIDE_INT valm; | |
2793 | HOST_WIDE_INT tval; | |
2794 | ||
2795 | for (i = 16; i < 64; i += 16) | |
2796 | { | |
2797 | valm = (valp & 0xffff); | |
2798 | ||
2799 | if (valm != 0) | |
2800 | ++ zcount; | |
2801 | ||
2802 | if (valm != 0xffff) | |
2803 | ++ ncount; | |
2804 | ||
2805 | valp >>= 16; | |
2806 | } | |
2807 | ||
2808 | /* zcount contains the number of additional MOVK instructions | |
2809 | required if the constant is built up with an initial MOVZ instruction, | |
2810 | while ncount is the number of MOVK instructions required if starting | |
2811 | with a MOVN instruction. Choose the sequence that yields the fewest | |
2812 | number of instructions, preferring MOVZ instructions when they are both | |
2813 | the same. */ | |
2814 | if (ncount < zcount) | |
2815 | { | |
9dfc162c JG |
2816 | if (generate) |
2817 | emit_move_insn (gen_rtx_REG (Pmode, regnum), | |
2818 | GEN_INT (val | ~(HOST_WIDE_INT) 0xffff)); | |
43e9d192 | 2819 | tval = 0xffff; |
9dfc162c | 2820 | insns++; |
43e9d192 IB |
2821 | } |
2822 | else | |
2823 | { | |
9dfc162c JG |
2824 | if (generate) |
2825 | emit_move_insn (gen_rtx_REG (Pmode, regnum), | |
2826 | GEN_INT (val & 0xffff)); | |
43e9d192 | 2827 | tval = 0; |
9dfc162c | 2828 | insns++; |
43e9d192 IB |
2829 | } |
2830 | ||
2831 | val >>= 16; | |
2832 | ||
2833 | for (i = 16; i < 64; i += 16) | |
2834 | { | |
2835 | if ((val & 0xffff) != tval) | |
9dfc162c JG |
2836 | { |
2837 | if (generate) | |
2838 | emit_insn (gen_insv_immdi (gen_rtx_REG (Pmode, regnum), | |
2839 | GEN_INT (i), | |
2840 | GEN_INT (val & 0xffff))); | |
2841 | insns++; | |
2842 | } | |
43e9d192 IB |
2843 | val >>= 16; |
2844 | } | |
2845 | } | |
9dfc162c | 2846 | return insns; |
43e9d192 IB |
2847 | } |
2848 | ||
2849 | static void | |
d9600ae5 | 2850 | aarch64_add_constant (int regnum, int scratchreg, HOST_WIDE_INT delta) |
43e9d192 IB |
2851 | { |
2852 | HOST_WIDE_INT mdelta = delta; | |
d9600ae5 SN |
2853 | rtx this_rtx = gen_rtx_REG (Pmode, regnum); |
2854 | rtx scratch_rtx = gen_rtx_REG (Pmode, scratchreg); | |
43e9d192 IB |
2855 | |
2856 | if (mdelta < 0) | |
2857 | mdelta = -mdelta; | |
2858 | ||
2859 | if (mdelta >= 4096 * 4096) | |
2860 | { | |
9dfc162c | 2861 | (void) aarch64_build_constant (scratchreg, delta, true); |
d9600ae5 | 2862 | emit_insn (gen_add3_insn (this_rtx, this_rtx, scratch_rtx)); |
43e9d192 IB |
2863 | } |
2864 | else if (mdelta > 0) | |
2865 | { | |
43e9d192 | 2866 | if (mdelta >= 4096) |
d9600ae5 SN |
2867 | { |
2868 | emit_insn (gen_rtx_SET (Pmode, scratch_rtx, GEN_INT (mdelta / 4096))); | |
2869 | rtx shift = gen_rtx_ASHIFT (Pmode, scratch_rtx, GEN_INT (12)); | |
2870 | if (delta < 0) | |
2871 | emit_insn (gen_rtx_SET (Pmode, this_rtx, | |
2872 | gen_rtx_MINUS (Pmode, this_rtx, shift))); | |
2873 | else | |
2874 | emit_insn (gen_rtx_SET (Pmode, this_rtx, | |
2875 | gen_rtx_PLUS (Pmode, this_rtx, shift))); | |
2876 | } | |
43e9d192 | 2877 | if (mdelta % 4096 != 0) |
d9600ae5 SN |
2878 | { |
2879 | scratch_rtx = GEN_INT ((delta < 0 ? -1 : 1) * (mdelta % 4096)); | |
2880 | emit_insn (gen_rtx_SET (Pmode, this_rtx, | |
2881 | gen_rtx_PLUS (Pmode, this_rtx, scratch_rtx))); | |
2882 | } | |
43e9d192 IB |
2883 | } |
2884 | } | |
2885 | ||
2886 | /* Output code to add DELTA to the first argument, and then jump | |
2887 | to FUNCTION. Used for C++ multiple inheritance. */ | |
2888 | static void | |
2889 | aarch64_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED, | |
2890 | HOST_WIDE_INT delta, | |
2891 | HOST_WIDE_INT vcall_offset, | |
2892 | tree function) | |
2893 | { | |
2894 | /* The this pointer is always in x0. Note that this differs from | |
2895 | Arm where the this pointer maybe bumped to r1 if r0 is required | |
2896 | to return a pointer to an aggregate. On AArch64 a result value | |
2897 | pointer will be in x8. */ | |
2898 | int this_regno = R0_REGNUM; | |
5d8a22a5 DM |
2899 | rtx this_rtx, temp0, temp1, addr, funexp; |
2900 | rtx_insn *insn; | |
43e9d192 | 2901 | |
75f1d6fc SN |
2902 | reload_completed = 1; |
2903 | emit_note (NOTE_INSN_PROLOGUE_END); | |
43e9d192 IB |
2904 | |
2905 | if (vcall_offset == 0) | |
d9600ae5 | 2906 | aarch64_add_constant (this_regno, IP1_REGNUM, delta); |
43e9d192 IB |
2907 | else |
2908 | { | |
28514dda | 2909 | gcc_assert ((vcall_offset & (POINTER_BYTES - 1)) == 0); |
43e9d192 | 2910 | |
75f1d6fc SN |
2911 | this_rtx = gen_rtx_REG (Pmode, this_regno); |
2912 | temp0 = gen_rtx_REG (Pmode, IP0_REGNUM); | |
2913 | temp1 = gen_rtx_REG (Pmode, IP1_REGNUM); | |
43e9d192 | 2914 | |
75f1d6fc SN |
2915 | addr = this_rtx; |
2916 | if (delta != 0) | |
2917 | { | |
2918 | if (delta >= -256 && delta < 256) | |
2919 | addr = gen_rtx_PRE_MODIFY (Pmode, this_rtx, | |
2920 | plus_constant (Pmode, this_rtx, delta)); | |
2921 | else | |
d9600ae5 | 2922 | aarch64_add_constant (this_regno, IP1_REGNUM, delta); |
43e9d192 IB |
2923 | } |
2924 | ||
28514dda YZ |
2925 | if (Pmode == ptr_mode) |
2926 | aarch64_emit_move (temp0, gen_rtx_MEM (ptr_mode, addr)); | |
2927 | else | |
2928 | aarch64_emit_move (temp0, | |
2929 | gen_rtx_ZERO_EXTEND (Pmode, | |
2930 | gen_rtx_MEM (ptr_mode, addr))); | |
75f1d6fc | 2931 | |
28514dda | 2932 | if (vcall_offset >= -256 && vcall_offset < 4096 * POINTER_BYTES) |
75f1d6fc | 2933 | addr = plus_constant (Pmode, temp0, vcall_offset); |
43e9d192 IB |
2934 | else |
2935 | { | |
9dfc162c | 2936 | (void) aarch64_build_constant (IP1_REGNUM, vcall_offset, true); |
75f1d6fc | 2937 | addr = gen_rtx_PLUS (Pmode, temp0, temp1); |
43e9d192 IB |
2938 | } |
2939 | ||
28514dda YZ |
2940 | if (Pmode == ptr_mode) |
2941 | aarch64_emit_move (temp1, gen_rtx_MEM (ptr_mode,addr)); | |
2942 | else | |
2943 | aarch64_emit_move (temp1, | |
2944 | gen_rtx_SIGN_EXTEND (Pmode, | |
2945 | gen_rtx_MEM (ptr_mode, addr))); | |
2946 | ||
75f1d6fc | 2947 | emit_insn (gen_add2_insn (this_rtx, temp1)); |
43e9d192 IB |
2948 | } |
2949 | ||
75f1d6fc SN |
2950 | /* Generate a tail call to the target function. */ |
2951 | if (!TREE_USED (function)) | |
2952 | { | |
2953 | assemble_external (function); | |
2954 | TREE_USED (function) = 1; | |
2955 | } | |
2956 | funexp = XEXP (DECL_RTL (function), 0); | |
2957 | funexp = gen_rtx_MEM (FUNCTION_MODE, funexp); | |
2958 | insn = emit_call_insn (gen_sibcall (funexp, const0_rtx, NULL_RTX)); | |
2959 | SIBLING_CALL_P (insn) = 1; | |
2960 | ||
2961 | insn = get_insns (); | |
2962 | shorten_branches (insn); | |
2963 | final_start_function (insn, file, 1); | |
2964 | final (insn, file, 1); | |
43e9d192 | 2965 | final_end_function (); |
75f1d6fc SN |
2966 | |
2967 | /* Stop pretending to be a post-reload pass. */ | |
2968 | reload_completed = 0; | |
43e9d192 IB |
2969 | } |
2970 | ||
43e9d192 IB |
2971 | static bool |
2972 | aarch64_tls_referenced_p (rtx x) | |
2973 | { | |
2974 | if (!TARGET_HAVE_TLS) | |
2975 | return false; | |
e7de8563 RS |
2976 | subrtx_iterator::array_type array; |
2977 | FOR_EACH_SUBRTX (iter, array, x, ALL) | |
2978 | { | |
2979 | const_rtx x = *iter; | |
2980 | if (GET_CODE (x) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (x) != 0) | |
2981 | return true; | |
2982 | /* Don't recurse into UNSPEC_TLS looking for TLS symbols; these are | |
2983 | TLS offsets, not real symbol references. */ | |
2984 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
2985 | iter.skip_subrtxes (); | |
2986 | } | |
2987 | return false; | |
43e9d192 IB |
2988 | } |
2989 | ||
2990 | ||
2991 | static int | |
2992 | aarch64_bitmasks_cmp (const void *i1, const void *i2) | |
2993 | { | |
2994 | const unsigned HOST_WIDE_INT *imm1 = (const unsigned HOST_WIDE_INT *) i1; | |
2995 | const unsigned HOST_WIDE_INT *imm2 = (const unsigned HOST_WIDE_INT *) i2; | |
2996 | ||
2997 | if (*imm1 < *imm2) | |
2998 | return -1; | |
2999 | if (*imm1 > *imm2) | |
3000 | return +1; | |
3001 | return 0; | |
3002 | } | |
3003 | ||
3004 | ||
3005 | static void | |
3006 | aarch64_build_bitmask_table (void) | |
3007 | { | |
3008 | unsigned HOST_WIDE_INT mask, imm; | |
3009 | unsigned int log_e, e, s, r; | |
3010 | unsigned int nimms = 0; | |
3011 | ||
3012 | for (log_e = 1; log_e <= 6; log_e++) | |
3013 | { | |
3014 | e = 1 << log_e; | |
3015 | if (e == 64) | |
3016 | mask = ~(HOST_WIDE_INT) 0; | |
3017 | else | |
3018 | mask = ((HOST_WIDE_INT) 1 << e) - 1; | |
3019 | for (s = 1; s < e; s++) | |
3020 | { | |
3021 | for (r = 0; r < e; r++) | |
3022 | { | |
3023 | /* set s consecutive bits to 1 (s < 64) */ | |
3024 | imm = ((unsigned HOST_WIDE_INT)1 << s) - 1; | |
3025 | /* rotate right by r */ | |
3026 | if (r != 0) | |
3027 | imm = ((imm >> r) | (imm << (e - r))) & mask; | |
3028 | /* replicate the constant depending on SIMD size */ | |
3029 | switch (log_e) { | |
3030 | case 1: imm |= (imm << 2); | |
3031 | case 2: imm |= (imm << 4); | |
3032 | case 3: imm |= (imm << 8); | |
3033 | case 4: imm |= (imm << 16); | |
3034 | case 5: imm |= (imm << 32); | |
3035 | case 6: | |
3036 | break; | |
3037 | default: | |
3038 | gcc_unreachable (); | |
3039 | } | |
3040 | gcc_assert (nimms < AARCH64_NUM_BITMASKS); | |
3041 | aarch64_bitmasks[nimms++] = imm; | |
3042 | } | |
3043 | } | |
3044 | } | |
3045 | ||
3046 | gcc_assert (nimms == AARCH64_NUM_BITMASKS); | |
3047 | qsort (aarch64_bitmasks, nimms, sizeof (aarch64_bitmasks[0]), | |
3048 | aarch64_bitmasks_cmp); | |
3049 | } | |
3050 | ||
3051 | ||
3052 | /* Return true if val can be encoded as a 12-bit unsigned immediate with | |
3053 | a left shift of 0 or 12 bits. */ | |
3054 | bool | |
3055 | aarch64_uimm12_shift (HOST_WIDE_INT val) | |
3056 | { | |
3057 | return ((val & (((HOST_WIDE_INT) 0xfff) << 0)) == val | |
3058 | || (val & (((HOST_WIDE_INT) 0xfff) << 12)) == val | |
3059 | ); | |
3060 | } | |
3061 | ||
3062 | ||
3063 | /* Return true if val is an immediate that can be loaded into a | |
3064 | register by a MOVZ instruction. */ | |
3065 | static bool | |
ef4bddc2 | 3066 | aarch64_movw_imm (HOST_WIDE_INT val, machine_mode mode) |
43e9d192 IB |
3067 | { |
3068 | if (GET_MODE_SIZE (mode) > 4) | |
3069 | { | |
3070 | if ((val & (((HOST_WIDE_INT) 0xffff) << 32)) == val | |
3071 | || (val & (((HOST_WIDE_INT) 0xffff) << 48)) == val) | |
3072 | return 1; | |
3073 | } | |
3074 | else | |
3075 | { | |
3076 | /* Ignore sign extension. */ | |
3077 | val &= (HOST_WIDE_INT) 0xffffffff; | |
3078 | } | |
3079 | return ((val & (((HOST_WIDE_INT) 0xffff) << 0)) == val | |
3080 | || (val & (((HOST_WIDE_INT) 0xffff) << 16)) == val); | |
3081 | } | |
3082 | ||
3083 | ||
3084 | /* Return true if val is a valid bitmask immediate. */ | |
3085 | bool | |
ef4bddc2 | 3086 | aarch64_bitmask_imm (HOST_WIDE_INT val, machine_mode mode) |
43e9d192 IB |
3087 | { |
3088 | if (GET_MODE_SIZE (mode) < 8) | |
3089 | { | |
3090 | /* Replicate bit pattern. */ | |
3091 | val &= (HOST_WIDE_INT) 0xffffffff; | |
3092 | val |= val << 32; | |
3093 | } | |
3094 | return bsearch (&val, aarch64_bitmasks, AARCH64_NUM_BITMASKS, | |
3095 | sizeof (aarch64_bitmasks[0]), aarch64_bitmasks_cmp) != NULL; | |
3096 | } | |
3097 | ||
3098 | ||
3099 | /* Return true if val is an immediate that can be loaded into a | |
3100 | register in a single instruction. */ | |
3101 | bool | |
ef4bddc2 | 3102 | aarch64_move_imm (HOST_WIDE_INT val, machine_mode mode) |
43e9d192 IB |
3103 | { |
3104 | if (aarch64_movw_imm (val, mode) || aarch64_movw_imm (~val, mode)) | |
3105 | return 1; | |
3106 | return aarch64_bitmask_imm (val, mode); | |
3107 | } | |
3108 | ||
3109 | static bool | |
ef4bddc2 | 3110 | aarch64_cannot_force_const_mem (machine_mode mode ATTRIBUTE_UNUSED, rtx x) |
43e9d192 IB |
3111 | { |
3112 | rtx base, offset; | |
7eda14e1 | 3113 | |
43e9d192 IB |
3114 | if (GET_CODE (x) == HIGH) |
3115 | return true; | |
3116 | ||
3117 | split_const (x, &base, &offset); | |
3118 | if (GET_CODE (base) == SYMBOL_REF || GET_CODE (base) == LABEL_REF) | |
28514dda | 3119 | { |
f8b756b7 | 3120 | if (aarch64_classify_symbol (base, offset, SYMBOL_CONTEXT_ADR) |
28514dda YZ |
3121 | != SYMBOL_FORCE_TO_MEM) |
3122 | return true; | |
3123 | else | |
3124 | /* Avoid generating a 64-bit relocation in ILP32; leave | |
3125 | to aarch64_expand_mov_immediate to handle it properly. */ | |
3126 | return mode != ptr_mode; | |
3127 | } | |
43e9d192 IB |
3128 | |
3129 | return aarch64_tls_referenced_p (x); | |
3130 | } | |
3131 | ||
3132 | /* Return true if register REGNO is a valid index register. | |
3133 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
3134 | ||
3135 | bool | |
3136 | aarch64_regno_ok_for_index_p (int regno, bool strict_p) | |
3137 | { | |
3138 | if (!HARD_REGISTER_NUM_P (regno)) | |
3139 | { | |
3140 | if (!strict_p) | |
3141 | return true; | |
3142 | ||
3143 | if (!reg_renumber) | |
3144 | return false; | |
3145 | ||
3146 | regno = reg_renumber[regno]; | |
3147 | } | |
3148 | return GP_REGNUM_P (regno); | |
3149 | } | |
3150 | ||
3151 | /* Return true if register REGNO is a valid base register for mode MODE. | |
3152 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
3153 | ||
3154 | bool | |
3155 | aarch64_regno_ok_for_base_p (int regno, bool strict_p) | |
3156 | { | |
3157 | if (!HARD_REGISTER_NUM_P (regno)) | |
3158 | { | |
3159 | if (!strict_p) | |
3160 | return true; | |
3161 | ||
3162 | if (!reg_renumber) | |
3163 | return false; | |
3164 | ||
3165 | regno = reg_renumber[regno]; | |
3166 | } | |
3167 | ||
3168 | /* The fake registers will be eliminated to either the stack or | |
3169 | hard frame pointer, both of which are usually valid base registers. | |
3170 | Reload deals with the cases where the eliminated form isn't valid. */ | |
3171 | return (GP_REGNUM_P (regno) | |
3172 | || regno == SP_REGNUM | |
3173 | || regno == FRAME_POINTER_REGNUM | |
3174 | || regno == ARG_POINTER_REGNUM); | |
3175 | } | |
3176 | ||
3177 | /* Return true if X is a valid base register for mode MODE. | |
3178 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
3179 | ||
3180 | static bool | |
3181 | aarch64_base_register_rtx_p (rtx x, bool strict_p) | |
3182 | { | |
3183 | if (!strict_p && GET_CODE (x) == SUBREG) | |
3184 | x = SUBREG_REG (x); | |
3185 | ||
3186 | return (REG_P (x) && aarch64_regno_ok_for_base_p (REGNO (x), strict_p)); | |
3187 | } | |
3188 | ||
3189 | /* Return true if address offset is a valid index. If it is, fill in INFO | |
3190 | appropriately. STRICT_P is true if REG_OK_STRICT is in effect. */ | |
3191 | ||
3192 | static bool | |
3193 | aarch64_classify_index (struct aarch64_address_info *info, rtx x, | |
ef4bddc2 | 3194 | machine_mode mode, bool strict_p) |
43e9d192 IB |
3195 | { |
3196 | enum aarch64_address_type type; | |
3197 | rtx index; | |
3198 | int shift; | |
3199 | ||
3200 | /* (reg:P) */ | |
3201 | if ((REG_P (x) || GET_CODE (x) == SUBREG) | |
3202 | && GET_MODE (x) == Pmode) | |
3203 | { | |
3204 | type = ADDRESS_REG_REG; | |
3205 | index = x; | |
3206 | shift = 0; | |
3207 | } | |
3208 | /* (sign_extend:DI (reg:SI)) */ | |
3209 | else if ((GET_CODE (x) == SIGN_EXTEND | |
3210 | || GET_CODE (x) == ZERO_EXTEND) | |
3211 | && GET_MODE (x) == DImode | |
3212 | && GET_MODE (XEXP (x, 0)) == SImode) | |
3213 | { | |
3214 | type = (GET_CODE (x) == SIGN_EXTEND) | |
3215 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3216 | index = XEXP (x, 0); | |
3217 | shift = 0; | |
3218 | } | |
3219 | /* (mult:DI (sign_extend:DI (reg:SI)) (const_int scale)) */ | |
3220 | else if (GET_CODE (x) == MULT | |
3221 | && (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
3222 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
3223 | && GET_MODE (XEXP (x, 0)) == DImode | |
3224 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == SImode | |
3225 | && CONST_INT_P (XEXP (x, 1))) | |
3226 | { | |
3227 | type = (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
3228 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3229 | index = XEXP (XEXP (x, 0), 0); | |
3230 | shift = exact_log2 (INTVAL (XEXP (x, 1))); | |
3231 | } | |
3232 | /* (ashift:DI (sign_extend:DI (reg:SI)) (const_int shift)) */ | |
3233 | else if (GET_CODE (x) == ASHIFT | |
3234 | && (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
3235 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
3236 | && GET_MODE (XEXP (x, 0)) == DImode | |
3237 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == SImode | |
3238 | && CONST_INT_P (XEXP (x, 1))) | |
3239 | { | |
3240 | type = (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
3241 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3242 | index = XEXP (XEXP (x, 0), 0); | |
3243 | shift = INTVAL (XEXP (x, 1)); | |
3244 | } | |
3245 | /* (sign_extract:DI (mult:DI (reg:DI) (const_int scale)) 32+shift 0) */ | |
3246 | else if ((GET_CODE (x) == SIGN_EXTRACT | |
3247 | || GET_CODE (x) == ZERO_EXTRACT) | |
3248 | && GET_MODE (x) == DImode | |
3249 | && GET_CODE (XEXP (x, 0)) == MULT | |
3250 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3251 | && CONST_INT_P (XEXP (XEXP (x, 0), 1))) | |
3252 | { | |
3253 | type = (GET_CODE (x) == SIGN_EXTRACT) | |
3254 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3255 | index = XEXP (XEXP (x, 0), 0); | |
3256 | shift = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1))); | |
3257 | if (INTVAL (XEXP (x, 1)) != 32 + shift | |
3258 | || INTVAL (XEXP (x, 2)) != 0) | |
3259 | shift = -1; | |
3260 | } | |
3261 | /* (and:DI (mult:DI (reg:DI) (const_int scale)) | |
3262 | (const_int 0xffffffff<<shift)) */ | |
3263 | else if (GET_CODE (x) == AND | |
3264 | && GET_MODE (x) == DImode | |
3265 | && GET_CODE (XEXP (x, 0)) == MULT | |
3266 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3267 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
3268 | && CONST_INT_P (XEXP (x, 1))) | |
3269 | { | |
3270 | type = ADDRESS_REG_UXTW; | |
3271 | index = XEXP (XEXP (x, 0), 0); | |
3272 | shift = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1))); | |
3273 | if (INTVAL (XEXP (x, 1)) != (HOST_WIDE_INT)0xffffffff << shift) | |
3274 | shift = -1; | |
3275 | } | |
3276 | /* (sign_extract:DI (ashift:DI (reg:DI) (const_int shift)) 32+shift 0) */ | |
3277 | else if ((GET_CODE (x) == SIGN_EXTRACT | |
3278 | || GET_CODE (x) == ZERO_EXTRACT) | |
3279 | && GET_MODE (x) == DImode | |
3280 | && GET_CODE (XEXP (x, 0)) == ASHIFT | |
3281 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3282 | && CONST_INT_P (XEXP (XEXP (x, 0), 1))) | |
3283 | { | |
3284 | type = (GET_CODE (x) == SIGN_EXTRACT) | |
3285 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
3286 | index = XEXP (XEXP (x, 0), 0); | |
3287 | shift = INTVAL (XEXP (XEXP (x, 0), 1)); | |
3288 | if (INTVAL (XEXP (x, 1)) != 32 + shift | |
3289 | || INTVAL (XEXP (x, 2)) != 0) | |
3290 | shift = -1; | |
3291 | } | |
3292 | /* (and:DI (ashift:DI (reg:DI) (const_int shift)) | |
3293 | (const_int 0xffffffff<<shift)) */ | |
3294 | else if (GET_CODE (x) == AND | |
3295 | && GET_MODE (x) == DImode | |
3296 | && GET_CODE (XEXP (x, 0)) == ASHIFT | |
3297 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
3298 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
3299 | && CONST_INT_P (XEXP (x, 1))) | |
3300 | { | |
3301 | type = ADDRESS_REG_UXTW; | |
3302 | index = XEXP (XEXP (x, 0), 0); | |
3303 | shift = INTVAL (XEXP (XEXP (x, 0), 1)); | |
3304 | if (INTVAL (XEXP (x, 1)) != (HOST_WIDE_INT)0xffffffff << shift) | |
3305 | shift = -1; | |
3306 | } | |
3307 | /* (mult:P (reg:P) (const_int scale)) */ | |
3308 | else if (GET_CODE (x) == MULT | |
3309 | && GET_MODE (x) == Pmode | |
3310 | && GET_MODE (XEXP (x, 0)) == Pmode | |
3311 | && CONST_INT_P (XEXP (x, 1))) | |
3312 | { | |
3313 | type = ADDRESS_REG_REG; | |
3314 | index = XEXP (x, 0); | |
3315 | shift = exact_log2 (INTVAL (XEXP (x, 1))); | |
3316 | } | |
3317 | /* (ashift:P (reg:P) (const_int shift)) */ | |
3318 | else if (GET_CODE (x) == ASHIFT | |
3319 | && GET_MODE (x) == Pmode | |
3320 | && GET_MODE (XEXP (x, 0)) == Pmode | |
3321 | && CONST_INT_P (XEXP (x, 1))) | |
3322 | { | |
3323 | type = ADDRESS_REG_REG; | |
3324 | index = XEXP (x, 0); | |
3325 | shift = INTVAL (XEXP (x, 1)); | |
3326 | } | |
3327 | else | |
3328 | return false; | |
3329 | ||
3330 | if (GET_CODE (index) == SUBREG) | |
3331 | index = SUBREG_REG (index); | |
3332 | ||
3333 | if ((shift == 0 || | |
3334 | (shift > 0 && shift <= 3 | |
3335 | && (1 << shift) == GET_MODE_SIZE (mode))) | |
3336 | && REG_P (index) | |
3337 | && aarch64_regno_ok_for_index_p (REGNO (index), strict_p)) | |
3338 | { | |
3339 | info->type = type; | |
3340 | info->offset = index; | |
3341 | info->shift = shift; | |
3342 | return true; | |
3343 | } | |
3344 | ||
3345 | return false; | |
3346 | } | |
3347 | ||
44707478 | 3348 | bool |
ef4bddc2 | 3349 | aarch64_offset_7bit_signed_scaled_p (machine_mode mode, HOST_WIDE_INT offset) |
43e9d192 IB |
3350 | { |
3351 | return (offset >= -64 * GET_MODE_SIZE (mode) | |
3352 | && offset < 64 * GET_MODE_SIZE (mode) | |
3353 | && offset % GET_MODE_SIZE (mode) == 0); | |
3354 | } | |
3355 | ||
3356 | static inline bool | |
ef4bddc2 | 3357 | offset_9bit_signed_unscaled_p (machine_mode mode ATTRIBUTE_UNUSED, |
43e9d192 IB |
3358 | HOST_WIDE_INT offset) |
3359 | { | |
3360 | return offset >= -256 && offset < 256; | |
3361 | } | |
3362 | ||
3363 | static inline bool | |
ef4bddc2 | 3364 | offset_12bit_unsigned_scaled_p (machine_mode mode, HOST_WIDE_INT offset) |
43e9d192 IB |
3365 | { |
3366 | return (offset >= 0 | |
3367 | && offset < 4096 * GET_MODE_SIZE (mode) | |
3368 | && offset % GET_MODE_SIZE (mode) == 0); | |
3369 | } | |
3370 | ||
3371 | /* Return true if X is a valid address for machine mode MODE. If it is, | |
3372 | fill in INFO appropriately. STRICT_P is true if REG_OK_STRICT is in | |
3373 | effect. OUTER_CODE is PARALLEL for a load/store pair. */ | |
3374 | ||
3375 | static bool | |
3376 | aarch64_classify_address (struct aarch64_address_info *info, | |
ef4bddc2 | 3377 | rtx x, machine_mode mode, |
43e9d192 IB |
3378 | RTX_CODE outer_code, bool strict_p) |
3379 | { | |
3380 | enum rtx_code code = GET_CODE (x); | |
3381 | rtx op0, op1; | |
2d8c6dc1 AH |
3382 | |
3383 | /* On BE, we use load/store pair for all large int mode load/stores. */ | |
3384 | bool load_store_pair_p = (outer_code == PARALLEL | |
3385 | || (BYTES_BIG_ENDIAN | |
3386 | && aarch64_vect_struct_mode_p (mode))); | |
3387 | ||
43e9d192 | 3388 | bool allow_reg_index_p = |
2d8c6dc1 AH |
3389 | !load_store_pair_p |
3390 | && (GET_MODE_SIZE (mode) != 16 || aarch64_vector_mode_supported_p (mode)) | |
3391 | && !aarch64_vect_struct_mode_p (mode); | |
3392 | ||
3393 | /* On LE, for AdvSIMD, don't support anything other than POST_INC or | |
3394 | REG addressing. */ | |
3395 | if (aarch64_vect_struct_mode_p (mode) && !BYTES_BIG_ENDIAN | |
43e9d192 IB |
3396 | && (code != POST_INC && code != REG)) |
3397 | return false; | |
3398 | ||
3399 | switch (code) | |
3400 | { | |
3401 | case REG: | |
3402 | case SUBREG: | |
3403 | info->type = ADDRESS_REG_IMM; | |
3404 | info->base = x; | |
3405 | info->offset = const0_rtx; | |
3406 | return aarch64_base_register_rtx_p (x, strict_p); | |
3407 | ||
3408 | case PLUS: | |
3409 | op0 = XEXP (x, 0); | |
3410 | op1 = XEXP (x, 1); | |
15c0c5c9 JW |
3411 | |
3412 | if (! strict_p | |
4aa81c2e | 3413 | && REG_P (op0) |
15c0c5c9 JW |
3414 | && (op0 == virtual_stack_vars_rtx |
3415 | || op0 == frame_pointer_rtx | |
3416 | || op0 == arg_pointer_rtx) | |
4aa81c2e | 3417 | && CONST_INT_P (op1)) |
15c0c5c9 JW |
3418 | { |
3419 | info->type = ADDRESS_REG_IMM; | |
3420 | info->base = op0; | |
3421 | info->offset = op1; | |
3422 | ||
3423 | return true; | |
3424 | } | |
3425 | ||
43e9d192 IB |
3426 | if (GET_MODE_SIZE (mode) != 0 |
3427 | && CONST_INT_P (op1) | |
3428 | && aarch64_base_register_rtx_p (op0, strict_p)) | |
3429 | { | |
3430 | HOST_WIDE_INT offset = INTVAL (op1); | |
3431 | ||
3432 | info->type = ADDRESS_REG_IMM; | |
3433 | info->base = op0; | |
3434 | info->offset = op1; | |
3435 | ||
3436 | /* TImode and TFmode values are allowed in both pairs of X | |
3437 | registers and individual Q registers. The available | |
3438 | address modes are: | |
3439 | X,X: 7-bit signed scaled offset | |
3440 | Q: 9-bit signed offset | |
3441 | We conservatively require an offset representable in either mode. | |
3442 | */ | |
3443 | if (mode == TImode || mode == TFmode) | |
44707478 | 3444 | return (aarch64_offset_7bit_signed_scaled_p (mode, offset) |
43e9d192 IB |
3445 | && offset_9bit_signed_unscaled_p (mode, offset)); |
3446 | ||
2d8c6dc1 AH |
3447 | /* A 7bit offset check because OImode will emit a ldp/stp |
3448 | instruction (only big endian will get here). | |
3449 | For ldp/stp instructions, the offset is scaled for the size of a | |
3450 | single element of the pair. */ | |
3451 | if (mode == OImode) | |
3452 | return aarch64_offset_7bit_signed_scaled_p (TImode, offset); | |
3453 | ||
3454 | /* Three 9/12 bit offsets checks because CImode will emit three | |
3455 | ldr/str instructions (only big endian will get here). */ | |
3456 | if (mode == CImode) | |
3457 | return (aarch64_offset_7bit_signed_scaled_p (TImode, offset) | |
3458 | && (offset_9bit_signed_unscaled_p (V16QImode, offset + 32) | |
3459 | || offset_12bit_unsigned_scaled_p (V16QImode, | |
3460 | offset + 32))); | |
3461 | ||
3462 | /* Two 7bit offsets checks because XImode will emit two ldp/stp | |
3463 | instructions (only big endian will get here). */ | |
3464 | if (mode == XImode) | |
3465 | return (aarch64_offset_7bit_signed_scaled_p (TImode, offset) | |
3466 | && aarch64_offset_7bit_signed_scaled_p (TImode, | |
3467 | offset + 32)); | |
3468 | ||
3469 | if (load_store_pair_p) | |
43e9d192 | 3470 | return ((GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) |
44707478 | 3471 | && aarch64_offset_7bit_signed_scaled_p (mode, offset)); |
43e9d192 IB |
3472 | else |
3473 | return (offset_9bit_signed_unscaled_p (mode, offset) | |
3474 | || offset_12bit_unsigned_scaled_p (mode, offset)); | |
3475 | } | |
3476 | ||
3477 | if (allow_reg_index_p) | |
3478 | { | |
3479 | /* Look for base + (scaled/extended) index register. */ | |
3480 | if (aarch64_base_register_rtx_p (op0, strict_p) | |
3481 | && aarch64_classify_index (info, op1, mode, strict_p)) | |
3482 | { | |
3483 | info->base = op0; | |
3484 | return true; | |
3485 | } | |
3486 | if (aarch64_base_register_rtx_p (op1, strict_p) | |
3487 | && aarch64_classify_index (info, op0, mode, strict_p)) | |
3488 | { | |
3489 | info->base = op1; | |
3490 | return true; | |
3491 | } | |
3492 | } | |
3493 | ||
3494 | return false; | |
3495 | ||
3496 | case POST_INC: | |
3497 | case POST_DEC: | |
3498 | case PRE_INC: | |
3499 | case PRE_DEC: | |
3500 | info->type = ADDRESS_REG_WB; | |
3501 | info->base = XEXP (x, 0); | |
3502 | info->offset = NULL_RTX; | |
3503 | return aarch64_base_register_rtx_p (info->base, strict_p); | |
3504 | ||
3505 | case POST_MODIFY: | |
3506 | case PRE_MODIFY: | |
3507 | info->type = ADDRESS_REG_WB; | |
3508 | info->base = XEXP (x, 0); | |
3509 | if (GET_CODE (XEXP (x, 1)) == PLUS | |
3510 | && CONST_INT_P (XEXP (XEXP (x, 1), 1)) | |
3511 | && rtx_equal_p (XEXP (XEXP (x, 1), 0), info->base) | |
3512 | && aarch64_base_register_rtx_p (info->base, strict_p)) | |
3513 | { | |
3514 | HOST_WIDE_INT offset; | |
3515 | info->offset = XEXP (XEXP (x, 1), 1); | |
3516 | offset = INTVAL (info->offset); | |
3517 | ||
3518 | /* TImode and TFmode values are allowed in both pairs of X | |
3519 | registers and individual Q registers. The available | |
3520 | address modes are: | |
3521 | X,X: 7-bit signed scaled offset | |
3522 | Q: 9-bit signed offset | |
3523 | We conservatively require an offset representable in either mode. | |
3524 | */ | |
3525 | if (mode == TImode || mode == TFmode) | |
44707478 | 3526 | return (aarch64_offset_7bit_signed_scaled_p (mode, offset) |
43e9d192 IB |
3527 | && offset_9bit_signed_unscaled_p (mode, offset)); |
3528 | ||
2d8c6dc1 | 3529 | if (load_store_pair_p) |
43e9d192 | 3530 | return ((GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) |
44707478 | 3531 | && aarch64_offset_7bit_signed_scaled_p (mode, offset)); |
43e9d192 IB |
3532 | else |
3533 | return offset_9bit_signed_unscaled_p (mode, offset); | |
3534 | } | |
3535 | return false; | |
3536 | ||
3537 | case CONST: | |
3538 | case SYMBOL_REF: | |
3539 | case LABEL_REF: | |
79517551 SN |
3540 | /* load literal: pc-relative constant pool entry. Only supported |
3541 | for SI mode or larger. */ | |
43e9d192 | 3542 | info->type = ADDRESS_SYMBOLIC; |
2d8c6dc1 AH |
3543 | |
3544 | if (!load_store_pair_p && GET_MODE_SIZE (mode) >= 4) | |
43e9d192 IB |
3545 | { |
3546 | rtx sym, addend; | |
3547 | ||
3548 | split_const (x, &sym, &addend); | |
3549 | return (GET_CODE (sym) == LABEL_REF | |
3550 | || (GET_CODE (sym) == SYMBOL_REF | |
3551 | && CONSTANT_POOL_ADDRESS_P (sym))); | |
3552 | } | |
3553 | return false; | |
3554 | ||
3555 | case LO_SUM: | |
3556 | info->type = ADDRESS_LO_SUM; | |
3557 | info->base = XEXP (x, 0); | |
3558 | info->offset = XEXP (x, 1); | |
3559 | if (allow_reg_index_p | |
3560 | && aarch64_base_register_rtx_p (info->base, strict_p)) | |
3561 | { | |
3562 | rtx sym, offs; | |
3563 | split_const (info->offset, &sym, &offs); | |
3564 | if (GET_CODE (sym) == SYMBOL_REF | |
f8b756b7 | 3565 | && (aarch64_classify_symbol (sym, offs, SYMBOL_CONTEXT_MEM) |
43e9d192 IB |
3566 | == SYMBOL_SMALL_ABSOLUTE)) |
3567 | { | |
3568 | /* The symbol and offset must be aligned to the access size. */ | |
3569 | unsigned int align; | |
3570 | unsigned int ref_size; | |
3571 | ||
3572 | if (CONSTANT_POOL_ADDRESS_P (sym)) | |
3573 | align = GET_MODE_ALIGNMENT (get_pool_mode (sym)); | |
3574 | else if (TREE_CONSTANT_POOL_ADDRESS_P (sym)) | |
3575 | { | |
3576 | tree exp = SYMBOL_REF_DECL (sym); | |
3577 | align = TYPE_ALIGN (TREE_TYPE (exp)); | |
3578 | align = CONSTANT_ALIGNMENT (exp, align); | |
3579 | } | |
3580 | else if (SYMBOL_REF_DECL (sym)) | |
3581 | align = DECL_ALIGN (SYMBOL_REF_DECL (sym)); | |
6c031d8d KV |
3582 | else if (SYMBOL_REF_HAS_BLOCK_INFO_P (sym) |
3583 | && SYMBOL_REF_BLOCK (sym) != NULL) | |
3584 | align = SYMBOL_REF_BLOCK (sym)->alignment; | |
43e9d192 IB |
3585 | else |
3586 | align = BITS_PER_UNIT; | |
3587 | ||
3588 | ref_size = GET_MODE_SIZE (mode); | |
3589 | if (ref_size == 0) | |
3590 | ref_size = GET_MODE_SIZE (DImode); | |
3591 | ||
3592 | return ((INTVAL (offs) & (ref_size - 1)) == 0 | |
3593 | && ((align / BITS_PER_UNIT) & (ref_size - 1)) == 0); | |
3594 | } | |
3595 | } | |
3596 | return false; | |
3597 | ||
3598 | default: | |
3599 | return false; | |
3600 | } | |
3601 | } | |
3602 | ||
3603 | bool | |
3604 | aarch64_symbolic_address_p (rtx x) | |
3605 | { | |
3606 | rtx offset; | |
3607 | ||
3608 | split_const (x, &x, &offset); | |
3609 | return GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == LABEL_REF; | |
3610 | } | |
3611 | ||
3612 | /* Classify the base of symbolic expression X, given that X appears in | |
3613 | context CONTEXT. */ | |
da4f13a4 MS |
3614 | |
3615 | enum aarch64_symbol_type | |
3616 | aarch64_classify_symbolic_expression (rtx x, | |
3617 | enum aarch64_symbol_context context) | |
43e9d192 IB |
3618 | { |
3619 | rtx offset; | |
da4f13a4 | 3620 | |
43e9d192 | 3621 | split_const (x, &x, &offset); |
f8b756b7 | 3622 | return aarch64_classify_symbol (x, offset, context); |
43e9d192 IB |
3623 | } |
3624 | ||
3625 | ||
3626 | /* Return TRUE if X is a legitimate address for accessing memory in | |
3627 | mode MODE. */ | |
3628 | static bool | |
ef4bddc2 | 3629 | aarch64_legitimate_address_hook_p (machine_mode mode, rtx x, bool strict_p) |
43e9d192 IB |
3630 | { |
3631 | struct aarch64_address_info addr; | |
3632 | ||
3633 | return aarch64_classify_address (&addr, x, mode, MEM, strict_p); | |
3634 | } | |
3635 | ||
3636 | /* Return TRUE if X is a legitimate address for accessing memory in | |
3637 | mode MODE. OUTER_CODE will be PARALLEL if this is a load/store | |
3638 | pair operation. */ | |
3639 | bool | |
ef4bddc2 | 3640 | aarch64_legitimate_address_p (machine_mode mode, rtx x, |
aef66c94 | 3641 | RTX_CODE outer_code, bool strict_p) |
43e9d192 IB |
3642 | { |
3643 | struct aarch64_address_info addr; | |
3644 | ||
3645 | return aarch64_classify_address (&addr, x, mode, outer_code, strict_p); | |
3646 | } | |
3647 | ||
3648 | /* Return TRUE if rtx X is immediate constant 0.0 */ | |
3649 | bool | |
3520f7cc | 3650 | aarch64_float_const_zero_rtx_p (rtx x) |
43e9d192 IB |
3651 | { |
3652 | REAL_VALUE_TYPE r; | |
3653 | ||
3654 | if (GET_MODE (x) == VOIDmode) | |
3655 | return false; | |
3656 | ||
3657 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
3658 | if (REAL_VALUE_MINUS_ZERO (r)) | |
3659 | return !HONOR_SIGNED_ZEROS (GET_MODE (x)); | |
3660 | return REAL_VALUES_EQUAL (r, dconst0); | |
3661 | } | |
3662 | ||
70f09188 AP |
3663 | /* Return the fixed registers used for condition codes. */ |
3664 | ||
3665 | static bool | |
3666 | aarch64_fixed_condition_code_regs (unsigned int *p1, unsigned int *p2) | |
3667 | { | |
3668 | *p1 = CC_REGNUM; | |
3669 | *p2 = INVALID_REGNUM; | |
3670 | return true; | |
3671 | } | |
3672 | ||
78607708 TV |
3673 | /* Emit call insn with PAT and do aarch64-specific handling. */ |
3674 | ||
d07a3fed | 3675 | void |
78607708 TV |
3676 | aarch64_emit_call_insn (rtx pat) |
3677 | { | |
3678 | rtx insn = emit_call_insn (pat); | |
3679 | ||
3680 | rtx *fusage = &CALL_INSN_FUNCTION_USAGE (insn); | |
3681 | clobber_reg (fusage, gen_rtx_REG (word_mode, IP0_REGNUM)); | |
3682 | clobber_reg (fusage, gen_rtx_REG (word_mode, IP1_REGNUM)); | |
3683 | } | |
3684 | ||
ef4bddc2 | 3685 | machine_mode |
43e9d192 IB |
3686 | aarch64_select_cc_mode (RTX_CODE code, rtx x, rtx y) |
3687 | { | |
3688 | /* All floating point compares return CCFP if it is an equality | |
3689 | comparison, and CCFPE otherwise. */ | |
3690 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
3691 | { | |
3692 | switch (code) | |
3693 | { | |
3694 | case EQ: | |
3695 | case NE: | |
3696 | case UNORDERED: | |
3697 | case ORDERED: | |
3698 | case UNLT: | |
3699 | case UNLE: | |
3700 | case UNGT: | |
3701 | case UNGE: | |
3702 | case UNEQ: | |
3703 | case LTGT: | |
3704 | return CCFPmode; | |
3705 | ||
3706 | case LT: | |
3707 | case LE: | |
3708 | case GT: | |
3709 | case GE: | |
3710 | return CCFPEmode; | |
3711 | ||
3712 | default: | |
3713 | gcc_unreachable (); | |
3714 | } | |
3715 | } | |
3716 | ||
3717 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
3718 | && y == const0_rtx | |
3719 | && (code == EQ || code == NE || code == LT || code == GE) | |
b056c910 N |
3720 | && (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS || GET_CODE (x) == AND |
3721 | || GET_CODE (x) == NEG)) | |
43e9d192 IB |
3722 | return CC_NZmode; |
3723 | ||
1c992d1e | 3724 | /* A compare with a shifted operand. Because of canonicalization, |
43e9d192 IB |
3725 | the comparison will have to be swapped when we emit the assembly |
3726 | code. */ | |
3727 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
4aa81c2e | 3728 | && (REG_P (y) || GET_CODE (y) == SUBREG) |
43e9d192 IB |
3729 | && (GET_CODE (x) == ASHIFT || GET_CODE (x) == ASHIFTRT |
3730 | || GET_CODE (x) == LSHIFTRT | |
1c992d1e | 3731 | || GET_CODE (x) == ZERO_EXTEND || GET_CODE (x) == SIGN_EXTEND)) |
43e9d192 IB |
3732 | return CC_SWPmode; |
3733 | ||
1c992d1e RE |
3734 | /* Similarly for a negated operand, but we can only do this for |
3735 | equalities. */ | |
3736 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
4aa81c2e | 3737 | && (REG_P (y) || GET_CODE (y) == SUBREG) |
1c992d1e RE |
3738 | && (code == EQ || code == NE) |
3739 | && GET_CODE (x) == NEG) | |
3740 | return CC_Zmode; | |
3741 | ||
43e9d192 IB |
3742 | /* A compare of a mode narrower than SI mode against zero can be done |
3743 | by extending the value in the comparison. */ | |
3744 | if ((GET_MODE (x) == QImode || GET_MODE (x) == HImode) | |
3745 | && y == const0_rtx) | |
3746 | /* Only use sign-extension if we really need it. */ | |
3747 | return ((code == GT || code == GE || code == LE || code == LT) | |
3748 | ? CC_SESWPmode : CC_ZESWPmode); | |
3749 | ||
3750 | /* For everything else, return CCmode. */ | |
3751 | return CCmode; | |
3752 | } | |
3753 | ||
3dfa7055 ZC |
3754 | static int |
3755 | aarch64_get_condition_code_1 (enum machine_mode, enum rtx_code); | |
3756 | ||
cd5660ab | 3757 | int |
43e9d192 IB |
3758 | aarch64_get_condition_code (rtx x) |
3759 | { | |
ef4bddc2 | 3760 | machine_mode mode = GET_MODE (XEXP (x, 0)); |
43e9d192 IB |
3761 | enum rtx_code comp_code = GET_CODE (x); |
3762 | ||
3763 | if (GET_MODE_CLASS (mode) != MODE_CC) | |
3764 | mode = SELECT_CC_MODE (comp_code, XEXP (x, 0), XEXP (x, 1)); | |
3dfa7055 ZC |
3765 | return aarch64_get_condition_code_1 (mode, comp_code); |
3766 | } | |
43e9d192 | 3767 | |
3dfa7055 ZC |
3768 | static int |
3769 | aarch64_get_condition_code_1 (enum machine_mode mode, enum rtx_code comp_code) | |
3770 | { | |
3771 | int ne = -1, eq = -1; | |
43e9d192 IB |
3772 | switch (mode) |
3773 | { | |
3774 | case CCFPmode: | |
3775 | case CCFPEmode: | |
3776 | switch (comp_code) | |
3777 | { | |
3778 | case GE: return AARCH64_GE; | |
3779 | case GT: return AARCH64_GT; | |
3780 | case LE: return AARCH64_LS; | |
3781 | case LT: return AARCH64_MI; | |
3782 | case NE: return AARCH64_NE; | |
3783 | case EQ: return AARCH64_EQ; | |
3784 | case ORDERED: return AARCH64_VC; | |
3785 | case UNORDERED: return AARCH64_VS; | |
3786 | case UNLT: return AARCH64_LT; | |
3787 | case UNLE: return AARCH64_LE; | |
3788 | case UNGT: return AARCH64_HI; | |
3789 | case UNGE: return AARCH64_PL; | |
cd5660ab | 3790 | default: return -1; |
43e9d192 IB |
3791 | } |
3792 | break; | |
3793 | ||
3dfa7055 ZC |
3794 | case CC_DNEmode: |
3795 | ne = AARCH64_NE; | |
3796 | eq = AARCH64_EQ; | |
3797 | break; | |
3798 | ||
3799 | case CC_DEQmode: | |
3800 | ne = AARCH64_EQ; | |
3801 | eq = AARCH64_NE; | |
3802 | break; | |
3803 | ||
3804 | case CC_DGEmode: | |
3805 | ne = AARCH64_GE; | |
3806 | eq = AARCH64_LT; | |
3807 | break; | |
3808 | ||
3809 | case CC_DLTmode: | |
3810 | ne = AARCH64_LT; | |
3811 | eq = AARCH64_GE; | |
3812 | break; | |
3813 | ||
3814 | case CC_DGTmode: | |
3815 | ne = AARCH64_GT; | |
3816 | eq = AARCH64_LE; | |
3817 | break; | |
3818 | ||
3819 | case CC_DLEmode: | |
3820 | ne = AARCH64_LE; | |
3821 | eq = AARCH64_GT; | |
3822 | break; | |
3823 | ||
3824 | case CC_DGEUmode: | |
3825 | ne = AARCH64_CS; | |
3826 | eq = AARCH64_CC; | |
3827 | break; | |
3828 | ||
3829 | case CC_DLTUmode: | |
3830 | ne = AARCH64_CC; | |
3831 | eq = AARCH64_CS; | |
3832 | break; | |
3833 | ||
3834 | case CC_DGTUmode: | |
3835 | ne = AARCH64_HI; | |
3836 | eq = AARCH64_LS; | |
3837 | break; | |
3838 | ||
3839 | case CC_DLEUmode: | |
3840 | ne = AARCH64_LS; | |
3841 | eq = AARCH64_HI; | |
3842 | break; | |
3843 | ||
43e9d192 IB |
3844 | case CCmode: |
3845 | switch (comp_code) | |
3846 | { | |
3847 | case NE: return AARCH64_NE; | |
3848 | case EQ: return AARCH64_EQ; | |
3849 | case GE: return AARCH64_GE; | |
3850 | case GT: return AARCH64_GT; | |
3851 | case LE: return AARCH64_LE; | |
3852 | case LT: return AARCH64_LT; | |
3853 | case GEU: return AARCH64_CS; | |
3854 | case GTU: return AARCH64_HI; | |
3855 | case LEU: return AARCH64_LS; | |
3856 | case LTU: return AARCH64_CC; | |
cd5660ab | 3857 | default: return -1; |
43e9d192 IB |
3858 | } |
3859 | break; | |
3860 | ||
3861 | case CC_SWPmode: | |
3862 | case CC_ZESWPmode: | |
3863 | case CC_SESWPmode: | |
3864 | switch (comp_code) | |
3865 | { | |
3866 | case NE: return AARCH64_NE; | |
3867 | case EQ: return AARCH64_EQ; | |
3868 | case GE: return AARCH64_LE; | |
3869 | case GT: return AARCH64_LT; | |
3870 | case LE: return AARCH64_GE; | |
3871 | case LT: return AARCH64_GT; | |
3872 | case GEU: return AARCH64_LS; | |
3873 | case GTU: return AARCH64_CC; | |
3874 | case LEU: return AARCH64_CS; | |
3875 | case LTU: return AARCH64_HI; | |
cd5660ab | 3876 | default: return -1; |
43e9d192 IB |
3877 | } |
3878 | break; | |
3879 | ||
3880 | case CC_NZmode: | |
3881 | switch (comp_code) | |
3882 | { | |
3883 | case NE: return AARCH64_NE; | |
3884 | case EQ: return AARCH64_EQ; | |
3885 | case GE: return AARCH64_PL; | |
3886 | case LT: return AARCH64_MI; | |
cd5660ab | 3887 | default: return -1; |
43e9d192 IB |
3888 | } |
3889 | break; | |
3890 | ||
1c992d1e RE |
3891 | case CC_Zmode: |
3892 | switch (comp_code) | |
3893 | { | |
3894 | case NE: return AARCH64_NE; | |
3895 | case EQ: return AARCH64_EQ; | |
cd5660ab | 3896 | default: return -1; |
1c992d1e RE |
3897 | } |
3898 | break; | |
3899 | ||
43e9d192 | 3900 | default: |
cd5660ab | 3901 | return -1; |
43e9d192 IB |
3902 | break; |
3903 | } | |
3dfa7055 ZC |
3904 | |
3905 | if (comp_code == NE) | |
3906 | return ne; | |
3907 | ||
3908 | if (comp_code == EQ) | |
3909 | return eq; | |
3910 | ||
3911 | return -1; | |
43e9d192 IB |
3912 | } |
3913 | ||
ddeabd3e AL |
3914 | bool |
3915 | aarch64_const_vec_all_same_in_range_p (rtx x, | |
3916 | HOST_WIDE_INT minval, | |
3917 | HOST_WIDE_INT maxval) | |
3918 | { | |
3919 | HOST_WIDE_INT firstval; | |
3920 | int count, i; | |
3921 | ||
3922 | if (GET_CODE (x) != CONST_VECTOR | |
3923 | || GET_MODE_CLASS (GET_MODE (x)) != MODE_VECTOR_INT) | |
3924 | return false; | |
3925 | ||
3926 | firstval = INTVAL (CONST_VECTOR_ELT (x, 0)); | |
3927 | if (firstval < minval || firstval > maxval) | |
3928 | return false; | |
3929 | ||
3930 | count = CONST_VECTOR_NUNITS (x); | |
3931 | for (i = 1; i < count; i++) | |
3932 | if (INTVAL (CONST_VECTOR_ELT (x, i)) != firstval) | |
3933 | return false; | |
3934 | ||
3935 | return true; | |
3936 | } | |
3937 | ||
3938 | bool | |
3939 | aarch64_const_vec_all_same_int_p (rtx x, HOST_WIDE_INT val) | |
3940 | { | |
3941 | return aarch64_const_vec_all_same_in_range_p (x, val, val); | |
3942 | } | |
3943 | ||
43e9d192 IB |
3944 | static unsigned |
3945 | bit_count (unsigned HOST_WIDE_INT value) | |
3946 | { | |
3947 | unsigned count = 0; | |
3948 | ||
3949 | while (value) | |
3950 | { | |
3951 | count++; | |
3952 | value &= value - 1; | |
3953 | } | |
3954 | ||
3955 | return count; | |
3956 | } | |
3957 | ||
cf670503 ZC |
3958 | /* N Z C V. */ |
3959 | #define AARCH64_CC_V 1 | |
3960 | #define AARCH64_CC_C (1 << 1) | |
3961 | #define AARCH64_CC_Z (1 << 2) | |
3962 | #define AARCH64_CC_N (1 << 3) | |
3963 | ||
3964 | /* N Z C V flags for ccmp. The first code is for AND op and the other | |
3965 | is for IOR op. Indexed by AARCH64_COND_CODE. */ | |
3966 | static const int aarch64_nzcv_codes[][2] = | |
3967 | { | |
3968 | {AARCH64_CC_Z, 0}, /* EQ, Z == 1. */ | |
3969 | {0, AARCH64_CC_Z}, /* NE, Z == 0. */ | |
3970 | {AARCH64_CC_C, 0}, /* CS, C == 1. */ | |
3971 | {0, AARCH64_CC_C}, /* CC, C == 0. */ | |
3972 | {AARCH64_CC_N, 0}, /* MI, N == 1. */ | |
3973 | {0, AARCH64_CC_N}, /* PL, N == 0. */ | |
3974 | {AARCH64_CC_V, 0}, /* VS, V == 1. */ | |
3975 | {0, AARCH64_CC_V}, /* VC, V == 0. */ | |
3976 | {AARCH64_CC_C, 0}, /* HI, C ==1 && Z == 0. */ | |
3977 | {0, AARCH64_CC_C}, /* LS, !(C == 1 && Z == 0). */ | |
3978 | {0, AARCH64_CC_V}, /* GE, N == V. */ | |
3979 | {AARCH64_CC_V, 0}, /* LT, N != V. */ | |
3980 | {0, AARCH64_CC_Z}, /* GT, Z == 0 && N == V. */ | |
3981 | {AARCH64_CC_Z, 0}, /* LE, !(Z == 0 && N == V). */ | |
3982 | {0, 0}, /* AL, Any. */ | |
3983 | {0, 0}, /* NV, Any. */ | |
3984 | }; | |
3985 | ||
3986 | int | |
3987 | aarch64_ccmp_mode_to_code (enum machine_mode mode) | |
3988 | { | |
3989 | switch (mode) | |
3990 | { | |
3991 | case CC_DNEmode: | |
3992 | return NE; | |
3993 | ||
3994 | case CC_DEQmode: | |
3995 | return EQ; | |
3996 | ||
3997 | case CC_DLEmode: | |
3998 | return LE; | |
3999 | ||
4000 | case CC_DGTmode: | |
4001 | return GT; | |
4002 | ||
4003 | case CC_DLTmode: | |
4004 | return LT; | |
4005 | ||
4006 | case CC_DGEmode: | |
4007 | return GE; | |
4008 | ||
4009 | case CC_DLEUmode: | |
4010 | return LEU; | |
4011 | ||
4012 | case CC_DGTUmode: | |
4013 | return GTU; | |
4014 | ||
4015 | case CC_DLTUmode: | |
4016 | return LTU; | |
4017 | ||
4018 | case CC_DGEUmode: | |
4019 | return GEU; | |
4020 | ||
4021 | default: | |
4022 | gcc_unreachable (); | |
4023 | } | |
4024 | } | |
4025 | ||
4026 | ||
43e9d192 IB |
4027 | void |
4028 | aarch64_print_operand (FILE *f, rtx x, char code) | |
4029 | { | |
4030 | switch (code) | |
4031 | { | |
f541a481 KT |
4032 | /* An integer or symbol address without a preceding # sign. */ |
4033 | case 'c': | |
4034 | switch (GET_CODE (x)) | |
4035 | { | |
4036 | case CONST_INT: | |
4037 | fprintf (f, HOST_WIDE_INT_PRINT_DEC, INTVAL (x)); | |
4038 | break; | |
4039 | ||
4040 | case SYMBOL_REF: | |
4041 | output_addr_const (f, x); | |
4042 | break; | |
4043 | ||
4044 | case CONST: | |
4045 | if (GET_CODE (XEXP (x, 0)) == PLUS | |
4046 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF) | |
4047 | { | |
4048 | output_addr_const (f, x); | |
4049 | break; | |
4050 | } | |
4051 | /* Fall through. */ | |
4052 | ||
4053 | default: | |
4054 | output_operand_lossage ("Unsupported operand for code '%c'", code); | |
4055 | } | |
4056 | break; | |
4057 | ||
43e9d192 IB |
4058 | case 'e': |
4059 | /* Print the sign/zero-extend size as a character 8->b, 16->h, 32->w. */ | |
4060 | { | |
4061 | int n; | |
4062 | ||
4aa81c2e | 4063 | if (!CONST_INT_P (x) |
43e9d192 IB |
4064 | || (n = exact_log2 (INTVAL (x) & ~7)) <= 0) |
4065 | { | |
4066 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4067 | return; | |
4068 | } | |
4069 | ||
4070 | switch (n) | |
4071 | { | |
4072 | case 3: | |
4073 | fputc ('b', f); | |
4074 | break; | |
4075 | case 4: | |
4076 | fputc ('h', f); | |
4077 | break; | |
4078 | case 5: | |
4079 | fputc ('w', f); | |
4080 | break; | |
4081 | default: | |
4082 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4083 | return; | |
4084 | } | |
4085 | } | |
4086 | break; | |
4087 | ||
4088 | case 'p': | |
4089 | { | |
4090 | int n; | |
4091 | ||
4092 | /* Print N such that 2^N == X. */ | |
4aa81c2e | 4093 | if (!CONST_INT_P (x) || (n = exact_log2 (INTVAL (x))) < 0) |
43e9d192 IB |
4094 | { |
4095 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4096 | return; | |
4097 | } | |
4098 | ||
4099 | asm_fprintf (f, "%d", n); | |
4100 | } | |
4101 | break; | |
4102 | ||
4103 | case 'P': | |
4104 | /* Print the number of non-zero bits in X (a const_int). */ | |
4aa81c2e | 4105 | if (!CONST_INT_P (x)) |
43e9d192 IB |
4106 | { |
4107 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4108 | return; | |
4109 | } | |
4110 | ||
4111 | asm_fprintf (f, "%u", bit_count (INTVAL (x))); | |
4112 | break; | |
4113 | ||
4114 | case 'H': | |
4115 | /* Print the higher numbered register of a pair (TImode) of regs. */ | |
4aa81c2e | 4116 | if (!REG_P (x) || !GP_REGNUM_P (REGNO (x) + 1)) |
43e9d192 IB |
4117 | { |
4118 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4119 | return; | |
4120 | } | |
4121 | ||
01a3a324 | 4122 | asm_fprintf (f, "%s", reg_names [REGNO (x) + 1]); |
43e9d192 IB |
4123 | break; |
4124 | ||
43e9d192 | 4125 | case 'm': |
cd5660ab KT |
4126 | { |
4127 | int cond_code; | |
4128 | /* Print a condition (eq, ne, etc). */ | |
43e9d192 | 4129 | |
cd5660ab KT |
4130 | /* CONST_TRUE_RTX means always -- that's the default. */ |
4131 | if (x == const_true_rtx) | |
43e9d192 | 4132 | return; |
43e9d192 | 4133 | |
cd5660ab KT |
4134 | if (!COMPARISON_P (x)) |
4135 | { | |
4136 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4137 | return; | |
4138 | } | |
4139 | ||
4140 | cond_code = aarch64_get_condition_code (x); | |
4141 | gcc_assert (cond_code >= 0); | |
4142 | fputs (aarch64_condition_codes[cond_code], f); | |
4143 | } | |
43e9d192 IB |
4144 | break; |
4145 | ||
4146 | case 'M': | |
cd5660ab KT |
4147 | { |
4148 | int cond_code; | |
4149 | /* Print the inverse of a condition (eq <-> ne, etc). */ | |
43e9d192 | 4150 | |
cd5660ab KT |
4151 | /* CONST_TRUE_RTX means never -- that's the default. */ |
4152 | if (x == const_true_rtx) | |
4153 | { | |
4154 | fputs ("nv", f); | |
4155 | return; | |
4156 | } | |
43e9d192 | 4157 | |
cd5660ab KT |
4158 | if (!COMPARISON_P (x)) |
4159 | { | |
4160 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4161 | return; | |
4162 | } | |
4163 | cond_code = aarch64_get_condition_code (x); | |
4164 | gcc_assert (cond_code >= 0); | |
4165 | fputs (aarch64_condition_codes[AARCH64_INVERSE_CONDITION_CODE | |
4166 | (cond_code)], f); | |
4167 | } | |
43e9d192 IB |
4168 | break; |
4169 | ||
4170 | case 'b': | |
4171 | case 'h': | |
4172 | case 's': | |
4173 | case 'd': | |
4174 | case 'q': | |
4175 | /* Print a scalar FP/SIMD register name. */ | |
4176 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) | |
4177 | { | |
4178 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
4179 | return; | |
4180 | } | |
50ce6f88 | 4181 | asm_fprintf (f, "%c%d", code, REGNO (x) - V0_REGNUM); |
43e9d192 IB |
4182 | break; |
4183 | ||
4184 | case 'S': | |
4185 | case 'T': | |
4186 | case 'U': | |
4187 | case 'V': | |
4188 | /* Print the first FP/SIMD register name in a list. */ | |
4189 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) | |
4190 | { | |
4191 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
4192 | return; | |
4193 | } | |
50ce6f88 | 4194 | asm_fprintf (f, "v%d", REGNO (x) - V0_REGNUM + (code - 'S')); |
43e9d192 IB |
4195 | break; |
4196 | ||
2d8c6dc1 AH |
4197 | case 'R': |
4198 | /* Print a scalar FP/SIMD register name + 1. */ | |
4199 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) | |
4200 | { | |
4201 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
4202 | return; | |
4203 | } | |
4204 | asm_fprintf (f, "q%d", REGNO (x) - V0_REGNUM + 1); | |
4205 | break; | |
4206 | ||
a05c0ddf | 4207 | case 'X': |
50d38551 | 4208 | /* Print bottom 16 bits of integer constant in hex. */ |
4aa81c2e | 4209 | if (!CONST_INT_P (x)) |
a05c0ddf IB |
4210 | { |
4211 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4212 | return; | |
4213 | } | |
50d38551 | 4214 | asm_fprintf (f, "0x%wx", UINTVAL (x) & 0xffff); |
a05c0ddf IB |
4215 | break; |
4216 | ||
43e9d192 IB |
4217 | case 'w': |
4218 | case 'x': | |
4219 | /* Print a general register name or the zero register (32-bit or | |
4220 | 64-bit). */ | |
3520f7cc JG |
4221 | if (x == const0_rtx |
4222 | || (CONST_DOUBLE_P (x) && aarch64_float_const_zero_rtx_p (x))) | |
43e9d192 | 4223 | { |
50ce6f88 | 4224 | asm_fprintf (f, "%czr", code); |
43e9d192 IB |
4225 | break; |
4226 | } | |
4227 | ||
4228 | if (REG_P (x) && GP_REGNUM_P (REGNO (x))) | |
4229 | { | |
50ce6f88 | 4230 | asm_fprintf (f, "%c%d", code, REGNO (x) - R0_REGNUM); |
43e9d192 IB |
4231 | break; |
4232 | } | |
4233 | ||
4234 | if (REG_P (x) && REGNO (x) == SP_REGNUM) | |
4235 | { | |
50ce6f88 | 4236 | asm_fprintf (f, "%ssp", code == 'w' ? "w" : ""); |
43e9d192 IB |
4237 | break; |
4238 | } | |
4239 | ||
4240 | /* Fall through */ | |
4241 | ||
4242 | case 0: | |
4243 | /* Print a normal operand, if it's a general register, then we | |
4244 | assume DImode. */ | |
4245 | if (x == NULL) | |
4246 | { | |
4247 | output_operand_lossage ("missing operand"); | |
4248 | return; | |
4249 | } | |
4250 | ||
4251 | switch (GET_CODE (x)) | |
4252 | { | |
4253 | case REG: | |
01a3a324 | 4254 | asm_fprintf (f, "%s", reg_names [REGNO (x)]); |
43e9d192 IB |
4255 | break; |
4256 | ||
4257 | case MEM: | |
4258 | aarch64_memory_reference_mode = GET_MODE (x); | |
4259 | output_address (XEXP (x, 0)); | |
4260 | break; | |
4261 | ||
4262 | case LABEL_REF: | |
4263 | case SYMBOL_REF: | |
4264 | output_addr_const (asm_out_file, x); | |
4265 | break; | |
4266 | ||
4267 | case CONST_INT: | |
4268 | asm_fprintf (f, "%wd", INTVAL (x)); | |
4269 | break; | |
4270 | ||
4271 | case CONST_VECTOR: | |
3520f7cc JG |
4272 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_VECTOR_INT) |
4273 | { | |
ddeabd3e AL |
4274 | gcc_assert ( |
4275 | aarch64_const_vec_all_same_in_range_p (x, | |
4276 | HOST_WIDE_INT_MIN, | |
4277 | HOST_WIDE_INT_MAX)); | |
3520f7cc JG |
4278 | asm_fprintf (f, "%wd", INTVAL (CONST_VECTOR_ELT (x, 0))); |
4279 | } | |
4280 | else if (aarch64_simd_imm_zero_p (x, GET_MODE (x))) | |
4281 | { | |
4282 | fputc ('0', f); | |
4283 | } | |
4284 | else | |
4285 | gcc_unreachable (); | |
43e9d192 IB |
4286 | break; |
4287 | ||
3520f7cc JG |
4288 | case CONST_DOUBLE: |
4289 | /* CONST_DOUBLE can represent a double-width integer. | |
4290 | In this case, the mode of x is VOIDmode. */ | |
4291 | if (GET_MODE (x) == VOIDmode) | |
4292 | ; /* Do Nothing. */ | |
4293 | else if (aarch64_float_const_zero_rtx_p (x)) | |
4294 | { | |
4295 | fputc ('0', f); | |
4296 | break; | |
4297 | } | |
4298 | else if (aarch64_float_const_representable_p (x)) | |
4299 | { | |
4300 | #define buf_size 20 | |
4301 | char float_buf[buf_size] = {'\0'}; | |
4302 | REAL_VALUE_TYPE r; | |
4303 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
4304 | real_to_decimal_for_mode (float_buf, &r, | |
4305 | buf_size, buf_size, | |
4306 | 1, GET_MODE (x)); | |
4307 | asm_fprintf (asm_out_file, "%s", float_buf); | |
4308 | break; | |
4309 | #undef buf_size | |
4310 | } | |
4311 | output_operand_lossage ("invalid constant"); | |
4312 | return; | |
43e9d192 IB |
4313 | default: |
4314 | output_operand_lossage ("invalid operand"); | |
4315 | return; | |
4316 | } | |
4317 | break; | |
4318 | ||
4319 | case 'A': | |
4320 | if (GET_CODE (x) == HIGH) | |
4321 | x = XEXP (x, 0); | |
4322 | ||
4323 | switch (aarch64_classify_symbolic_expression (x, SYMBOL_CONTEXT_ADR)) | |
4324 | { | |
4325 | case SYMBOL_SMALL_GOT: | |
4326 | asm_fprintf (asm_out_file, ":got:"); | |
4327 | break; | |
4328 | ||
4329 | case SYMBOL_SMALL_TLSGD: | |
4330 | asm_fprintf (asm_out_file, ":tlsgd:"); | |
4331 | break; | |
4332 | ||
4333 | case SYMBOL_SMALL_TLSDESC: | |
4334 | asm_fprintf (asm_out_file, ":tlsdesc:"); | |
4335 | break; | |
4336 | ||
4337 | case SYMBOL_SMALL_GOTTPREL: | |
4338 | asm_fprintf (asm_out_file, ":gottprel:"); | |
4339 | break; | |
4340 | ||
4341 | case SYMBOL_SMALL_TPREL: | |
4342 | asm_fprintf (asm_out_file, ":tprel:"); | |
4343 | break; | |
4344 | ||
87dd8ab0 MS |
4345 | case SYMBOL_TINY_GOT: |
4346 | gcc_unreachable (); | |
4347 | break; | |
4348 | ||
43e9d192 IB |
4349 | default: |
4350 | break; | |
4351 | } | |
4352 | output_addr_const (asm_out_file, x); | |
4353 | break; | |
4354 | ||
4355 | case 'L': | |
4356 | switch (aarch64_classify_symbolic_expression (x, SYMBOL_CONTEXT_ADR)) | |
4357 | { | |
4358 | case SYMBOL_SMALL_GOT: | |
4359 | asm_fprintf (asm_out_file, ":lo12:"); | |
4360 | break; | |
4361 | ||
4362 | case SYMBOL_SMALL_TLSGD: | |
4363 | asm_fprintf (asm_out_file, ":tlsgd_lo12:"); | |
4364 | break; | |
4365 | ||
4366 | case SYMBOL_SMALL_TLSDESC: | |
4367 | asm_fprintf (asm_out_file, ":tlsdesc_lo12:"); | |
4368 | break; | |
4369 | ||
4370 | case SYMBOL_SMALL_GOTTPREL: | |
4371 | asm_fprintf (asm_out_file, ":gottprel_lo12:"); | |
4372 | break; | |
4373 | ||
4374 | case SYMBOL_SMALL_TPREL: | |
4375 | asm_fprintf (asm_out_file, ":tprel_lo12_nc:"); | |
4376 | break; | |
4377 | ||
87dd8ab0 MS |
4378 | case SYMBOL_TINY_GOT: |
4379 | asm_fprintf (asm_out_file, ":got:"); | |
4380 | break; | |
4381 | ||
43e9d192 IB |
4382 | default: |
4383 | break; | |
4384 | } | |
4385 | output_addr_const (asm_out_file, x); | |
4386 | break; | |
4387 | ||
4388 | case 'G': | |
4389 | ||
4390 | switch (aarch64_classify_symbolic_expression (x, SYMBOL_CONTEXT_ADR)) | |
4391 | { | |
4392 | case SYMBOL_SMALL_TPREL: | |
4393 | asm_fprintf (asm_out_file, ":tprel_hi12:"); | |
4394 | break; | |
4395 | default: | |
4396 | break; | |
4397 | } | |
4398 | output_addr_const (asm_out_file, x); | |
4399 | break; | |
4400 | ||
cf670503 ZC |
4401 | case 'K': |
4402 | { | |
4403 | int cond_code; | |
4404 | /* Print nzcv. */ | |
4405 | ||
4406 | if (!COMPARISON_P (x)) | |
4407 | { | |
4408 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4409 | return; | |
4410 | } | |
4411 | ||
4412 | cond_code = aarch64_get_condition_code_1 (CCmode, GET_CODE (x)); | |
4413 | gcc_assert (cond_code >= 0); | |
4414 | asm_fprintf (f, "%d", aarch64_nzcv_codes[cond_code][0]); | |
4415 | } | |
4416 | break; | |
4417 | ||
4418 | case 'k': | |
4419 | { | |
4420 | int cond_code; | |
4421 | /* Print nzcv. */ | |
4422 | ||
4423 | if (!COMPARISON_P (x)) | |
4424 | { | |
4425 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
4426 | return; | |
4427 | } | |
4428 | ||
4429 | cond_code = aarch64_get_condition_code_1 (CCmode, GET_CODE (x)); | |
4430 | gcc_assert (cond_code >= 0); | |
4431 | asm_fprintf (f, "%d", aarch64_nzcv_codes[cond_code][1]); | |
4432 | } | |
4433 | break; | |
4434 | ||
43e9d192 IB |
4435 | default: |
4436 | output_operand_lossage ("invalid operand prefix '%%%c'", code); | |
4437 | return; | |
4438 | } | |
4439 | } | |
4440 | ||
4441 | void | |
4442 | aarch64_print_operand_address (FILE *f, rtx x) | |
4443 | { | |
4444 | struct aarch64_address_info addr; | |
4445 | ||
4446 | if (aarch64_classify_address (&addr, x, aarch64_memory_reference_mode, | |
4447 | MEM, true)) | |
4448 | switch (addr.type) | |
4449 | { | |
4450 | case ADDRESS_REG_IMM: | |
4451 | if (addr.offset == const0_rtx) | |
01a3a324 | 4452 | asm_fprintf (f, "[%s]", reg_names [REGNO (addr.base)]); |
43e9d192 | 4453 | else |
16a3246f | 4454 | asm_fprintf (f, "[%s, %wd]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4455 | INTVAL (addr.offset)); |
4456 | return; | |
4457 | ||
4458 | case ADDRESS_REG_REG: | |
4459 | if (addr.shift == 0) | |
16a3246f | 4460 | asm_fprintf (f, "[%s, %s]", reg_names [REGNO (addr.base)], |
01a3a324 | 4461 | reg_names [REGNO (addr.offset)]); |
43e9d192 | 4462 | else |
16a3246f | 4463 | asm_fprintf (f, "[%s, %s, lsl %u]", reg_names [REGNO (addr.base)], |
01a3a324 | 4464 | reg_names [REGNO (addr.offset)], addr.shift); |
43e9d192 IB |
4465 | return; |
4466 | ||
4467 | case ADDRESS_REG_UXTW: | |
4468 | if (addr.shift == 0) | |
16a3246f | 4469 | asm_fprintf (f, "[%s, w%d, uxtw]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4470 | REGNO (addr.offset) - R0_REGNUM); |
4471 | else | |
16a3246f | 4472 | asm_fprintf (f, "[%s, w%d, uxtw %u]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4473 | REGNO (addr.offset) - R0_REGNUM, addr.shift); |
4474 | return; | |
4475 | ||
4476 | case ADDRESS_REG_SXTW: | |
4477 | if (addr.shift == 0) | |
16a3246f | 4478 | asm_fprintf (f, "[%s, w%d, sxtw]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4479 | REGNO (addr.offset) - R0_REGNUM); |
4480 | else | |
16a3246f | 4481 | asm_fprintf (f, "[%s, w%d, sxtw %u]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4482 | REGNO (addr.offset) - R0_REGNUM, addr.shift); |
4483 | return; | |
4484 | ||
4485 | case ADDRESS_REG_WB: | |
4486 | switch (GET_CODE (x)) | |
4487 | { | |
4488 | case PRE_INC: | |
16a3246f | 4489 | asm_fprintf (f, "[%s, %d]!", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4490 | GET_MODE_SIZE (aarch64_memory_reference_mode)); |
4491 | return; | |
4492 | case POST_INC: | |
16a3246f | 4493 | asm_fprintf (f, "[%s], %d", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4494 | GET_MODE_SIZE (aarch64_memory_reference_mode)); |
4495 | return; | |
4496 | case PRE_DEC: | |
16a3246f | 4497 | asm_fprintf (f, "[%s, -%d]!", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4498 | GET_MODE_SIZE (aarch64_memory_reference_mode)); |
4499 | return; | |
4500 | case POST_DEC: | |
16a3246f | 4501 | asm_fprintf (f, "[%s], -%d", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4502 | GET_MODE_SIZE (aarch64_memory_reference_mode)); |
4503 | return; | |
4504 | case PRE_MODIFY: | |
16a3246f | 4505 | asm_fprintf (f, "[%s, %wd]!", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4506 | INTVAL (addr.offset)); |
4507 | return; | |
4508 | case POST_MODIFY: | |
16a3246f | 4509 | asm_fprintf (f, "[%s], %wd", reg_names [REGNO (addr.base)], |
43e9d192 IB |
4510 | INTVAL (addr.offset)); |
4511 | return; | |
4512 | default: | |
4513 | break; | |
4514 | } | |
4515 | break; | |
4516 | ||
4517 | case ADDRESS_LO_SUM: | |
16a3246f | 4518 | asm_fprintf (f, "[%s, #:lo12:", reg_names [REGNO (addr.base)]); |
43e9d192 IB |
4519 | output_addr_const (f, addr.offset); |
4520 | asm_fprintf (f, "]"); | |
4521 | return; | |
4522 | ||
4523 | case ADDRESS_SYMBOLIC: | |
4524 | break; | |
4525 | } | |
4526 | ||
4527 | output_addr_const (f, x); | |
4528 | } | |
4529 | ||
43e9d192 IB |
4530 | bool |
4531 | aarch64_label_mentioned_p (rtx x) | |
4532 | { | |
4533 | const char *fmt; | |
4534 | int i; | |
4535 | ||
4536 | if (GET_CODE (x) == LABEL_REF) | |
4537 | return true; | |
4538 | ||
4539 | /* UNSPEC_TLS entries for a symbol include a LABEL_REF for the | |
4540 | referencing instruction, but they are constant offsets, not | |
4541 | symbols. */ | |
4542 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
4543 | return false; | |
4544 | ||
4545 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
4546 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
4547 | { | |
4548 | if (fmt[i] == 'E') | |
4549 | { | |
4550 | int j; | |
4551 | ||
4552 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
4553 | if (aarch64_label_mentioned_p (XVECEXP (x, i, j))) | |
4554 | return 1; | |
4555 | } | |
4556 | else if (fmt[i] == 'e' && aarch64_label_mentioned_p (XEXP (x, i))) | |
4557 | return 1; | |
4558 | } | |
4559 | ||
4560 | return 0; | |
4561 | } | |
4562 | ||
4563 | /* Implement REGNO_REG_CLASS. */ | |
4564 | ||
4565 | enum reg_class | |
4566 | aarch64_regno_regclass (unsigned regno) | |
4567 | { | |
4568 | if (GP_REGNUM_P (regno)) | |
a4a182c6 | 4569 | return GENERAL_REGS; |
43e9d192 IB |
4570 | |
4571 | if (regno == SP_REGNUM) | |
4572 | return STACK_REG; | |
4573 | ||
4574 | if (regno == FRAME_POINTER_REGNUM | |
4575 | || regno == ARG_POINTER_REGNUM) | |
f24bb080 | 4576 | return POINTER_REGS; |
43e9d192 IB |
4577 | |
4578 | if (FP_REGNUM_P (regno)) | |
4579 | return FP_LO_REGNUM_P (regno) ? FP_LO_REGS : FP_REGS; | |
4580 | ||
4581 | return NO_REGS; | |
4582 | } | |
4583 | ||
0c4ec427 | 4584 | static rtx |
ef4bddc2 | 4585 | aarch64_legitimize_address (rtx x, rtx /* orig_x */, machine_mode mode) |
0c4ec427 RE |
4586 | { |
4587 | /* Try to split X+CONST into Y=X+(CONST & ~mask), Y+(CONST&mask), | |
4588 | where mask is selected by alignment and size of the offset. | |
4589 | We try to pick as large a range for the offset as possible to | |
4590 | maximize the chance of a CSE. However, for aligned addresses | |
4591 | we limit the range to 4k so that structures with different sized | |
4592 | elements are likely to use the same base. */ | |
4593 | ||
4594 | if (GET_CODE (x) == PLUS && CONST_INT_P (XEXP (x, 1))) | |
4595 | { | |
4596 | HOST_WIDE_INT offset = INTVAL (XEXP (x, 1)); | |
4597 | HOST_WIDE_INT base_offset; | |
4598 | ||
4599 | /* Does it look like we'll need a load/store-pair operation? */ | |
4600 | if (GET_MODE_SIZE (mode) > 16 | |
4601 | || mode == TImode) | |
4602 | base_offset = ((offset + 64 * GET_MODE_SIZE (mode)) | |
4603 | & ~((128 * GET_MODE_SIZE (mode)) - 1)); | |
4604 | /* For offsets aren't a multiple of the access size, the limit is | |
4605 | -256...255. */ | |
4606 | else if (offset & (GET_MODE_SIZE (mode) - 1)) | |
4607 | base_offset = (offset + 0x100) & ~0x1ff; | |
4608 | else | |
4609 | base_offset = offset & ~0xfff; | |
4610 | ||
4611 | if (base_offset == 0) | |
4612 | return x; | |
4613 | ||
4614 | offset -= base_offset; | |
4615 | rtx base_reg = gen_reg_rtx (Pmode); | |
4616 | rtx val = force_operand (plus_constant (Pmode, XEXP (x, 0), base_offset), | |
4617 | NULL_RTX); | |
4618 | emit_move_insn (base_reg, val); | |
4619 | x = plus_constant (Pmode, base_reg, offset); | |
4620 | } | |
4621 | ||
4622 | return x; | |
4623 | } | |
4624 | ||
43e9d192 IB |
4625 | /* Try a machine-dependent way of reloading an illegitimate address |
4626 | operand. If we find one, push the reload and return the new rtx. */ | |
4627 | ||
4628 | rtx | |
4629 | aarch64_legitimize_reload_address (rtx *x_p, | |
ef4bddc2 | 4630 | machine_mode mode, |
43e9d192 IB |
4631 | int opnum, int type, |
4632 | int ind_levels ATTRIBUTE_UNUSED) | |
4633 | { | |
4634 | rtx x = *x_p; | |
4635 | ||
348d4b0a BC |
4636 | /* Do not allow mem (plus (reg, const)) if vector struct mode. */ |
4637 | if (aarch64_vect_struct_mode_p (mode) | |
43e9d192 IB |
4638 | && GET_CODE (x) == PLUS |
4639 | && REG_P (XEXP (x, 0)) | |
4640 | && CONST_INT_P (XEXP (x, 1))) | |
4641 | { | |
4642 | rtx orig_rtx = x; | |
4643 | x = copy_rtx (x); | |
4644 | push_reload (orig_rtx, NULL_RTX, x_p, NULL, | |
4645 | BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0, | |
4646 | opnum, (enum reload_type) type); | |
4647 | return x; | |
4648 | } | |
4649 | ||
4650 | /* We must recognize output that we have already generated ourselves. */ | |
4651 | if (GET_CODE (x) == PLUS | |
4652 | && GET_CODE (XEXP (x, 0)) == PLUS | |
4653 | && REG_P (XEXP (XEXP (x, 0), 0)) | |
4654 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
4655 | && CONST_INT_P (XEXP (x, 1))) | |
4656 | { | |
4657 | push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL, | |
4658 | BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0, | |
4659 | opnum, (enum reload_type) type); | |
4660 | return x; | |
4661 | } | |
4662 | ||
4663 | /* We wish to handle large displacements off a base register by splitting | |
4664 | the addend across an add and the mem insn. This can cut the number of | |
4665 | extra insns needed from 3 to 1. It is only useful for load/store of a | |
4666 | single register with 12 bit offset field. */ | |
4667 | if (GET_CODE (x) == PLUS | |
4668 | && REG_P (XEXP (x, 0)) | |
4669 | && CONST_INT_P (XEXP (x, 1)) | |
4670 | && HARD_REGISTER_P (XEXP (x, 0)) | |
4671 | && mode != TImode | |
4672 | && mode != TFmode | |
4673 | && aarch64_regno_ok_for_base_p (REGNO (XEXP (x, 0)), true)) | |
4674 | { | |
4675 | HOST_WIDE_INT val = INTVAL (XEXP (x, 1)); | |
4676 | HOST_WIDE_INT low = val & 0xfff; | |
4677 | HOST_WIDE_INT high = val - low; | |
4678 | HOST_WIDE_INT offs; | |
4679 | rtx cst; | |
ef4bddc2 | 4680 | machine_mode xmode = GET_MODE (x); |
28514dda YZ |
4681 | |
4682 | /* In ILP32, xmode can be either DImode or SImode. */ | |
4683 | gcc_assert (xmode == DImode || xmode == SImode); | |
43e9d192 IB |
4684 | |
4685 | /* Reload non-zero BLKmode offsets. This is because we cannot ascertain | |
4686 | BLKmode alignment. */ | |
4687 | if (GET_MODE_SIZE (mode) == 0) | |
4688 | return NULL_RTX; | |
4689 | ||
4690 | offs = low % GET_MODE_SIZE (mode); | |
4691 | ||
4692 | /* Align misaligned offset by adjusting high part to compensate. */ | |
4693 | if (offs != 0) | |
4694 | { | |
4695 | if (aarch64_uimm12_shift (high + offs)) | |
4696 | { | |
4697 | /* Align down. */ | |
4698 | low = low - offs; | |
4699 | high = high + offs; | |
4700 | } | |
4701 | else | |
4702 | { | |
4703 | /* Align up. */ | |
4704 | offs = GET_MODE_SIZE (mode) - offs; | |
4705 | low = low + offs; | |
4706 | high = high + (low & 0x1000) - offs; | |
4707 | low &= 0xfff; | |
4708 | } | |
4709 | } | |
4710 | ||
4711 | /* Check for overflow. */ | |
4712 | if (high + low != val) | |
4713 | return NULL_RTX; | |
4714 | ||
4715 | cst = GEN_INT (high); | |
4716 | if (!aarch64_uimm12_shift (high)) | |
28514dda | 4717 | cst = force_const_mem (xmode, cst); |
43e9d192 IB |
4718 | |
4719 | /* Reload high part into base reg, leaving the low part | |
acf67b8e KT |
4720 | in the mem instruction. |
4721 | Note that replacing this gen_rtx_PLUS with plus_constant is | |
4722 | wrong in this case because we rely on the | |
4723 | (plus (plus reg c1) c2) structure being preserved so that | |
4724 | XEXP (*p, 0) in push_reload below uses the correct term. */ | |
4d0e5b54 JG |
4725 | x = gen_rtx_PLUS (xmode, |
4726 | gen_rtx_PLUS (xmode, XEXP (x, 0), cst), | |
4727 | GEN_INT (low)); | |
43e9d192 IB |
4728 | |
4729 | push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL, | |
28514dda | 4730 | BASE_REG_CLASS, xmode, VOIDmode, 0, 0, |
43e9d192 IB |
4731 | opnum, (enum reload_type) type); |
4732 | return x; | |
4733 | } | |
4734 | ||
4735 | return NULL_RTX; | |
4736 | } | |
4737 | ||
4738 | ||
4739 | static reg_class_t | |
4740 | aarch64_secondary_reload (bool in_p ATTRIBUTE_UNUSED, rtx x, | |
4741 | reg_class_t rclass, | |
ef4bddc2 | 4742 | machine_mode mode, |
43e9d192 IB |
4743 | secondary_reload_info *sri) |
4744 | { | |
43e9d192 IB |
4745 | /* Without the TARGET_SIMD instructions we cannot move a Q register |
4746 | to a Q register directly. We need a scratch. */ | |
4747 | if (REG_P (x) && (mode == TFmode || mode == TImode) && mode == GET_MODE (x) | |
4748 | && FP_REGNUM_P (REGNO (x)) && !TARGET_SIMD | |
4749 | && reg_class_subset_p (rclass, FP_REGS)) | |
4750 | { | |
4751 | if (mode == TFmode) | |
4752 | sri->icode = CODE_FOR_aarch64_reload_movtf; | |
4753 | else if (mode == TImode) | |
4754 | sri->icode = CODE_FOR_aarch64_reload_movti; | |
4755 | return NO_REGS; | |
4756 | } | |
4757 | ||
4758 | /* A TFmode or TImode memory access should be handled via an FP_REGS | |
4759 | because AArch64 has richer addressing modes for LDR/STR instructions | |
4760 | than LDP/STP instructions. */ | |
a4a182c6 | 4761 | if (!TARGET_GENERAL_REGS_ONLY && rclass == GENERAL_REGS |
43e9d192 IB |
4762 | && GET_MODE_SIZE (mode) == 16 && MEM_P (x)) |
4763 | return FP_REGS; | |
4764 | ||
4765 | if (rclass == FP_REGS && (mode == TImode || mode == TFmode) && CONSTANT_P(x)) | |
a4a182c6 | 4766 | return GENERAL_REGS; |
43e9d192 IB |
4767 | |
4768 | return NO_REGS; | |
4769 | } | |
4770 | ||
4771 | static bool | |
4772 | aarch64_can_eliminate (const int from, const int to) | |
4773 | { | |
4774 | /* If we need a frame pointer, we must eliminate FRAME_POINTER_REGNUM into | |
4775 | HARD_FRAME_POINTER_REGNUM and not into STACK_POINTER_REGNUM. */ | |
4776 | ||
4777 | if (frame_pointer_needed) | |
4778 | { | |
4779 | if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
4780 | return true; | |
4781 | if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) | |
4782 | return false; | |
4783 | if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM | |
4784 | && !cfun->calls_alloca) | |
4785 | return true; | |
4786 | if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
4787 | return true; | |
0b7f8166 MS |
4788 | |
4789 | return false; | |
43e9d192 | 4790 | } |
1c923b60 JW |
4791 | else |
4792 | { | |
4793 | /* If we decided that we didn't need a leaf frame pointer but then used | |
4794 | LR in the function, then we'll want a frame pointer after all, so | |
4795 | prevent this elimination to ensure a frame pointer is used. */ | |
4796 | if (to == STACK_POINTER_REGNUM | |
4797 | && flag_omit_leaf_frame_pointer | |
4798 | && df_regs_ever_live_p (LR_REGNUM)) | |
4799 | return false; | |
4800 | } | |
777e6976 | 4801 | |
43e9d192 IB |
4802 | return true; |
4803 | } | |
4804 | ||
4805 | HOST_WIDE_INT | |
4806 | aarch64_initial_elimination_offset (unsigned from, unsigned to) | |
4807 | { | |
43e9d192 | 4808 | aarch64_layout_frame (); |
78c29983 MS |
4809 | |
4810 | if (to == HARD_FRAME_POINTER_REGNUM) | |
4811 | { | |
4812 | if (from == ARG_POINTER_REGNUM) | |
1c960e02 | 4813 | return cfun->machine->frame.frame_size - crtl->outgoing_args_size; |
78c29983 MS |
4814 | |
4815 | if (from == FRAME_POINTER_REGNUM) | |
1c960e02 MS |
4816 | return (cfun->machine->frame.hard_fp_offset |
4817 | - cfun->machine->frame.saved_varargs_size); | |
78c29983 MS |
4818 | } |
4819 | ||
4820 | if (to == STACK_POINTER_REGNUM) | |
4821 | { | |
4822 | if (from == FRAME_POINTER_REGNUM) | |
1c960e02 MS |
4823 | return (cfun->machine->frame.frame_size |
4824 | - cfun->machine->frame.saved_varargs_size); | |
78c29983 MS |
4825 | } |
4826 | ||
1c960e02 | 4827 | return cfun->machine->frame.frame_size; |
43e9d192 IB |
4828 | } |
4829 | ||
43e9d192 IB |
4830 | /* Implement RETURN_ADDR_RTX. We do not support moving back to a |
4831 | previous frame. */ | |
4832 | ||
4833 | rtx | |
4834 | aarch64_return_addr (int count, rtx frame ATTRIBUTE_UNUSED) | |
4835 | { | |
4836 | if (count != 0) | |
4837 | return const0_rtx; | |
4838 | return get_hard_reg_initial_val (Pmode, LR_REGNUM); | |
4839 | } | |
4840 | ||
4841 | ||
4842 | static void | |
4843 | aarch64_asm_trampoline_template (FILE *f) | |
4844 | { | |
28514dda YZ |
4845 | if (TARGET_ILP32) |
4846 | { | |
4847 | asm_fprintf (f, "\tldr\tw%d, .+16\n", IP1_REGNUM - R0_REGNUM); | |
4848 | asm_fprintf (f, "\tldr\tw%d, .+16\n", STATIC_CHAIN_REGNUM - R0_REGNUM); | |
4849 | } | |
4850 | else | |
4851 | { | |
4852 | asm_fprintf (f, "\tldr\t%s, .+16\n", reg_names [IP1_REGNUM]); | |
4853 | asm_fprintf (f, "\tldr\t%s, .+20\n", reg_names [STATIC_CHAIN_REGNUM]); | |
4854 | } | |
01a3a324 | 4855 | asm_fprintf (f, "\tbr\t%s\n", reg_names [IP1_REGNUM]); |
43e9d192 | 4856 | assemble_aligned_integer (4, const0_rtx); |
28514dda YZ |
4857 | assemble_aligned_integer (POINTER_BYTES, const0_rtx); |
4858 | assemble_aligned_integer (POINTER_BYTES, const0_rtx); | |
43e9d192 IB |
4859 | } |
4860 | ||
4861 | static void | |
4862 | aarch64_trampoline_init (rtx m_tramp, tree fndecl, rtx chain_value) | |
4863 | { | |
4864 | rtx fnaddr, mem, a_tramp; | |
28514dda | 4865 | const int tramp_code_sz = 16; |
43e9d192 IB |
4866 | |
4867 | /* Don't need to copy the trailing D-words, we fill those in below. */ | |
4868 | emit_block_move (m_tramp, assemble_trampoline_template (), | |
28514dda YZ |
4869 | GEN_INT (tramp_code_sz), BLOCK_OP_NORMAL); |
4870 | mem = adjust_address (m_tramp, ptr_mode, tramp_code_sz); | |
43e9d192 | 4871 | fnaddr = XEXP (DECL_RTL (fndecl), 0); |
28514dda YZ |
4872 | if (GET_MODE (fnaddr) != ptr_mode) |
4873 | fnaddr = convert_memory_address (ptr_mode, fnaddr); | |
43e9d192 IB |
4874 | emit_move_insn (mem, fnaddr); |
4875 | ||
28514dda | 4876 | mem = adjust_address (m_tramp, ptr_mode, tramp_code_sz + POINTER_BYTES); |
43e9d192 IB |
4877 | emit_move_insn (mem, chain_value); |
4878 | ||
4879 | /* XXX We should really define a "clear_cache" pattern and use | |
4880 | gen_clear_cache(). */ | |
4881 | a_tramp = XEXP (m_tramp, 0); | |
4882 | emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__clear_cache"), | |
28514dda YZ |
4883 | LCT_NORMAL, VOIDmode, 2, a_tramp, ptr_mode, |
4884 | plus_constant (ptr_mode, a_tramp, TRAMPOLINE_SIZE), | |
4885 | ptr_mode); | |
43e9d192 IB |
4886 | } |
4887 | ||
4888 | static unsigned char | |
ef4bddc2 | 4889 | aarch64_class_max_nregs (reg_class_t regclass, machine_mode mode) |
43e9d192 IB |
4890 | { |
4891 | switch (regclass) | |
4892 | { | |
fee9ba42 | 4893 | case CALLER_SAVE_REGS: |
43e9d192 IB |
4894 | case POINTER_REGS: |
4895 | case GENERAL_REGS: | |
4896 | case ALL_REGS: | |
4897 | case FP_REGS: | |
4898 | case FP_LO_REGS: | |
4899 | return | |
4900 | aarch64_vector_mode_p (mode) ? (GET_MODE_SIZE (mode) + 15) / 16 : | |
aef66c94 | 4901 | (GET_MODE_SIZE (mode) + 7) / 8; |
43e9d192 IB |
4902 | case STACK_REG: |
4903 | return 1; | |
4904 | ||
4905 | case NO_REGS: | |
4906 | return 0; | |
4907 | ||
4908 | default: | |
4909 | break; | |
4910 | } | |
4911 | gcc_unreachable (); | |
4912 | } | |
4913 | ||
4914 | static reg_class_t | |
78d8b9f0 | 4915 | aarch64_preferred_reload_class (rtx x, reg_class_t regclass) |
43e9d192 | 4916 | { |
51bb310d | 4917 | if (regclass == POINTER_REGS) |
78d8b9f0 IB |
4918 | return GENERAL_REGS; |
4919 | ||
51bb310d MS |
4920 | if (regclass == STACK_REG) |
4921 | { | |
4922 | if (REG_P(x) | |
4923 | && reg_class_subset_p (REGNO_REG_CLASS (REGNO (x)), POINTER_REGS)) | |
4924 | return regclass; | |
4925 | ||
4926 | return NO_REGS; | |
4927 | } | |
4928 | ||
78d8b9f0 IB |
4929 | /* If it's an integer immediate that MOVI can't handle, then |
4930 | FP_REGS is not an option, so we return NO_REGS instead. */ | |
4931 | if (CONST_INT_P (x) && reg_class_subset_p (regclass, FP_REGS) | |
4932 | && !aarch64_simd_imm_scalar_p (x, GET_MODE (x))) | |
4933 | return NO_REGS; | |
4934 | ||
27bd251b IB |
4935 | /* Register eliminiation can result in a request for |
4936 | SP+constant->FP_REGS. We cannot support such operations which | |
4937 | use SP as source and an FP_REG as destination, so reject out | |
4938 | right now. */ | |
4939 | if (! reg_class_subset_p (regclass, GENERAL_REGS) && GET_CODE (x) == PLUS) | |
4940 | { | |
4941 | rtx lhs = XEXP (x, 0); | |
4942 | ||
4943 | /* Look through a possible SUBREG introduced by ILP32. */ | |
4944 | if (GET_CODE (lhs) == SUBREG) | |
4945 | lhs = SUBREG_REG (lhs); | |
4946 | ||
4947 | gcc_assert (REG_P (lhs)); | |
4948 | gcc_assert (reg_class_subset_p (REGNO_REG_CLASS (REGNO (lhs)), | |
4949 | POINTER_REGS)); | |
4950 | return NO_REGS; | |
4951 | } | |
4952 | ||
78d8b9f0 | 4953 | return regclass; |
43e9d192 IB |
4954 | } |
4955 | ||
4956 | void | |
4957 | aarch64_asm_output_labelref (FILE* f, const char *name) | |
4958 | { | |
4959 | asm_fprintf (f, "%U%s", name); | |
4960 | } | |
4961 | ||
4962 | static void | |
4963 | aarch64_elf_asm_constructor (rtx symbol, int priority) | |
4964 | { | |
4965 | if (priority == DEFAULT_INIT_PRIORITY) | |
4966 | default_ctor_section_asm_out_constructor (symbol, priority); | |
4967 | else | |
4968 | { | |
4969 | section *s; | |
4970 | char buf[18]; | |
4971 | snprintf (buf, sizeof (buf), ".init_array.%.5u", priority); | |
4972 | s = get_section (buf, SECTION_WRITE, NULL); | |
4973 | switch_to_section (s); | |
4974 | assemble_align (POINTER_SIZE); | |
28514dda | 4975 | assemble_aligned_integer (POINTER_BYTES, symbol); |
43e9d192 IB |
4976 | } |
4977 | } | |
4978 | ||
4979 | static void | |
4980 | aarch64_elf_asm_destructor (rtx symbol, int priority) | |
4981 | { | |
4982 | if (priority == DEFAULT_INIT_PRIORITY) | |
4983 | default_dtor_section_asm_out_destructor (symbol, priority); | |
4984 | else | |
4985 | { | |
4986 | section *s; | |
4987 | char buf[18]; | |
4988 | snprintf (buf, sizeof (buf), ".fini_array.%.5u", priority); | |
4989 | s = get_section (buf, SECTION_WRITE, NULL); | |
4990 | switch_to_section (s); | |
4991 | assemble_align (POINTER_SIZE); | |
28514dda | 4992 | assemble_aligned_integer (POINTER_BYTES, symbol); |
43e9d192 IB |
4993 | } |
4994 | } | |
4995 | ||
4996 | const char* | |
4997 | aarch64_output_casesi (rtx *operands) | |
4998 | { | |
4999 | char buf[100]; | |
5000 | char label[100]; | |
b32d5189 | 5001 | rtx diff_vec = PATTERN (NEXT_INSN (as_a <rtx_insn *> (operands[2]))); |
43e9d192 IB |
5002 | int index; |
5003 | static const char *const patterns[4][2] = | |
5004 | { | |
5005 | { | |
5006 | "ldrb\t%w3, [%0,%w1,uxtw]", | |
5007 | "add\t%3, %4, %w3, sxtb #2" | |
5008 | }, | |
5009 | { | |
5010 | "ldrh\t%w3, [%0,%w1,uxtw #1]", | |
5011 | "add\t%3, %4, %w3, sxth #2" | |
5012 | }, | |
5013 | { | |
5014 | "ldr\t%w3, [%0,%w1,uxtw #2]", | |
5015 | "add\t%3, %4, %w3, sxtw #2" | |
5016 | }, | |
5017 | /* We assume that DImode is only generated when not optimizing and | |
5018 | that we don't really need 64-bit address offsets. That would | |
5019 | imply an object file with 8GB of code in a single function! */ | |
5020 | { | |
5021 | "ldr\t%w3, [%0,%w1,uxtw #2]", | |
5022 | "add\t%3, %4, %w3, sxtw #2" | |
5023 | } | |
5024 | }; | |
5025 | ||
5026 | gcc_assert (GET_CODE (diff_vec) == ADDR_DIFF_VEC); | |
5027 | ||
5028 | index = exact_log2 (GET_MODE_SIZE (GET_MODE (diff_vec))); | |
5029 | ||
5030 | gcc_assert (index >= 0 && index <= 3); | |
5031 | ||
5032 | /* Need to implement table size reduction, by chaning the code below. */ | |
5033 | output_asm_insn (patterns[index][0], operands); | |
5034 | ASM_GENERATE_INTERNAL_LABEL (label, "Lrtx", CODE_LABEL_NUMBER (operands[2])); | |
5035 | snprintf (buf, sizeof (buf), | |
5036 | "adr\t%%4, %s", targetm.strip_name_encoding (label)); | |
5037 | output_asm_insn (buf, operands); | |
5038 | output_asm_insn (patterns[index][1], operands); | |
5039 | output_asm_insn ("br\t%3", operands); | |
5040 | assemble_label (asm_out_file, label); | |
5041 | return ""; | |
5042 | } | |
5043 | ||
5044 | ||
5045 | /* Return size in bits of an arithmetic operand which is shifted/scaled and | |
5046 | masked such that it is suitable for a UXTB, UXTH, or UXTW extend | |
5047 | operator. */ | |
5048 | ||
5049 | int | |
5050 | aarch64_uxt_size (int shift, HOST_WIDE_INT mask) | |
5051 | { | |
5052 | if (shift >= 0 && shift <= 3) | |
5053 | { | |
5054 | int size; | |
5055 | for (size = 8; size <= 32; size *= 2) | |
5056 | { | |
5057 | HOST_WIDE_INT bits = ((HOST_WIDE_INT)1U << size) - 1; | |
5058 | if (mask == bits << shift) | |
5059 | return size; | |
5060 | } | |
5061 | } | |
5062 | return 0; | |
5063 | } | |
5064 | ||
5065 | static bool | |
ef4bddc2 | 5066 | aarch64_use_blocks_for_constant_p (machine_mode mode ATTRIBUTE_UNUSED, |
43e9d192 IB |
5067 | const_rtx x ATTRIBUTE_UNUSED) |
5068 | { | |
5069 | /* We can't use blocks for constants when we're using a per-function | |
5070 | constant pool. */ | |
5071 | return false; | |
5072 | } | |
5073 | ||
5074 | static section * | |
ef4bddc2 | 5075 | aarch64_select_rtx_section (machine_mode mode ATTRIBUTE_UNUSED, |
43e9d192 IB |
5076 | rtx x ATTRIBUTE_UNUSED, |
5077 | unsigned HOST_WIDE_INT align ATTRIBUTE_UNUSED) | |
5078 | { | |
5079 | /* Force all constant pool entries into the current function section. */ | |
5080 | return function_section (current_function_decl); | |
5081 | } | |
5082 | ||
5083 | ||
5084 | /* Costs. */ | |
5085 | ||
5086 | /* Helper function for rtx cost calculation. Strip a shift expression | |
5087 | from X. Returns the inner operand if successful, or the original | |
5088 | expression on failure. */ | |
5089 | static rtx | |
5090 | aarch64_strip_shift (rtx x) | |
5091 | { | |
5092 | rtx op = x; | |
5093 | ||
57b77d46 RE |
5094 | /* We accept both ROTATERT and ROTATE: since the RHS must be a constant |
5095 | we can convert both to ROR during final output. */ | |
43e9d192 IB |
5096 | if ((GET_CODE (op) == ASHIFT |
5097 | || GET_CODE (op) == ASHIFTRT | |
57b77d46 RE |
5098 | || GET_CODE (op) == LSHIFTRT |
5099 | || GET_CODE (op) == ROTATERT | |
5100 | || GET_CODE (op) == ROTATE) | |
43e9d192 IB |
5101 | && CONST_INT_P (XEXP (op, 1))) |
5102 | return XEXP (op, 0); | |
5103 | ||
5104 | if (GET_CODE (op) == MULT | |
5105 | && CONST_INT_P (XEXP (op, 1)) | |
5106 | && ((unsigned) exact_log2 (INTVAL (XEXP (op, 1)))) < 64) | |
5107 | return XEXP (op, 0); | |
5108 | ||
5109 | return x; | |
5110 | } | |
5111 | ||
4745e701 | 5112 | /* Helper function for rtx cost calculation. Strip an extend |
43e9d192 IB |
5113 | expression from X. Returns the inner operand if successful, or the |
5114 | original expression on failure. We deal with a number of possible | |
5115 | canonicalization variations here. */ | |
5116 | static rtx | |
4745e701 | 5117 | aarch64_strip_extend (rtx x) |
43e9d192 IB |
5118 | { |
5119 | rtx op = x; | |
5120 | ||
5121 | /* Zero and sign extraction of a widened value. */ | |
5122 | if ((GET_CODE (op) == ZERO_EXTRACT || GET_CODE (op) == SIGN_EXTRACT) | |
5123 | && XEXP (op, 2) == const0_rtx | |
4745e701 | 5124 | && GET_CODE (XEXP (op, 0)) == MULT |
43e9d192 IB |
5125 | && aarch64_is_extend_from_extract (GET_MODE (op), XEXP (XEXP (op, 0), 1), |
5126 | XEXP (op, 1))) | |
5127 | return XEXP (XEXP (op, 0), 0); | |
5128 | ||
5129 | /* It can also be represented (for zero-extend) as an AND with an | |
5130 | immediate. */ | |
5131 | if (GET_CODE (op) == AND | |
5132 | && GET_CODE (XEXP (op, 0)) == MULT | |
5133 | && CONST_INT_P (XEXP (XEXP (op, 0), 1)) | |
5134 | && CONST_INT_P (XEXP (op, 1)) | |
5135 | && aarch64_uxt_size (exact_log2 (INTVAL (XEXP (XEXP (op, 0), 1))), | |
5136 | INTVAL (XEXP (op, 1))) != 0) | |
5137 | return XEXP (XEXP (op, 0), 0); | |
5138 | ||
5139 | /* Now handle extended register, as this may also have an optional | |
5140 | left shift by 1..4. */ | |
5141 | if (GET_CODE (op) == ASHIFT | |
5142 | && CONST_INT_P (XEXP (op, 1)) | |
5143 | && ((unsigned HOST_WIDE_INT) INTVAL (XEXP (op, 1))) <= 4) | |
5144 | op = XEXP (op, 0); | |
5145 | ||
5146 | if (GET_CODE (op) == ZERO_EXTEND | |
5147 | || GET_CODE (op) == SIGN_EXTEND) | |
5148 | op = XEXP (op, 0); | |
5149 | ||
5150 | if (op != x) | |
5151 | return op; | |
5152 | ||
4745e701 JG |
5153 | return x; |
5154 | } | |
5155 | ||
5156 | /* Helper function for rtx cost calculation. Calculate the cost of | |
5157 | a MULT, which may be part of a multiply-accumulate rtx. Return | |
5158 | the calculated cost of the expression, recursing manually in to | |
5159 | operands where needed. */ | |
5160 | ||
5161 | static int | |
5162 | aarch64_rtx_mult_cost (rtx x, int code, int outer, bool speed) | |
5163 | { | |
5164 | rtx op0, op1; | |
5165 | const struct cpu_cost_table *extra_cost | |
5166 | = aarch64_tune_params->insn_extra_cost; | |
5167 | int cost = 0; | |
5168 | bool maybe_fma = (outer == PLUS || outer == MINUS); | |
ef4bddc2 | 5169 | machine_mode mode = GET_MODE (x); |
4745e701 JG |
5170 | |
5171 | gcc_checking_assert (code == MULT); | |
5172 | ||
5173 | op0 = XEXP (x, 0); | |
5174 | op1 = XEXP (x, 1); | |
5175 | ||
5176 | if (VECTOR_MODE_P (mode)) | |
5177 | mode = GET_MODE_INNER (mode); | |
5178 | ||
5179 | /* Integer multiply/fma. */ | |
5180 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
5181 | { | |
5182 | /* The multiply will be canonicalized as a shift, cost it as such. */ | |
5183 | if (CONST_INT_P (op1) | |
5184 | && exact_log2 (INTVAL (op1)) > 0) | |
5185 | { | |
5186 | if (speed) | |
5187 | { | |
5188 | if (maybe_fma) | |
5189 | /* ADD (shifted register). */ | |
5190 | cost += extra_cost->alu.arith_shift; | |
5191 | else | |
5192 | /* LSL (immediate). */ | |
5193 | cost += extra_cost->alu.shift; | |
5194 | } | |
5195 | ||
5196 | cost += rtx_cost (op0, GET_CODE (op0), 0, speed); | |
5197 | ||
5198 | return cost; | |
5199 | } | |
5200 | ||
5201 | /* Integer multiplies or FMAs have zero/sign extending variants. */ | |
5202 | if ((GET_CODE (op0) == ZERO_EXTEND | |
5203 | && GET_CODE (op1) == ZERO_EXTEND) | |
5204 | || (GET_CODE (op0) == SIGN_EXTEND | |
5205 | && GET_CODE (op1) == SIGN_EXTEND)) | |
5206 | { | |
5207 | cost += rtx_cost (XEXP (op0, 0), MULT, 0, speed) | |
5208 | + rtx_cost (XEXP (op1, 0), MULT, 1, speed); | |
5209 | ||
5210 | if (speed) | |
5211 | { | |
5212 | if (maybe_fma) | |
5213 | /* MADD/SMADDL/UMADDL. */ | |
5214 | cost += extra_cost->mult[0].extend_add; | |
5215 | else | |
5216 | /* MUL/SMULL/UMULL. */ | |
5217 | cost += extra_cost->mult[0].extend; | |
5218 | } | |
5219 | ||
5220 | return cost; | |
5221 | } | |
5222 | ||
5223 | /* This is either an integer multiply or an FMA. In both cases | |
5224 | we want to recurse and cost the operands. */ | |
5225 | cost += rtx_cost (op0, MULT, 0, speed) | |
5226 | + rtx_cost (op1, MULT, 1, speed); | |
5227 | ||
5228 | if (speed) | |
5229 | { | |
5230 | if (maybe_fma) | |
5231 | /* MADD. */ | |
5232 | cost += extra_cost->mult[mode == DImode].add; | |
5233 | else | |
5234 | /* MUL. */ | |
5235 | cost += extra_cost->mult[mode == DImode].simple; | |
5236 | } | |
5237 | ||
5238 | return cost; | |
5239 | } | |
5240 | else | |
5241 | { | |
5242 | if (speed) | |
5243 | { | |
3d840f7d | 5244 | /* Floating-point FMA/FMUL can also support negations of the |
4745e701 JG |
5245 | operands. */ |
5246 | if (GET_CODE (op0) == NEG) | |
3d840f7d | 5247 | op0 = XEXP (op0, 0); |
4745e701 | 5248 | if (GET_CODE (op1) == NEG) |
3d840f7d | 5249 | op1 = XEXP (op1, 0); |
4745e701 JG |
5250 | |
5251 | if (maybe_fma) | |
5252 | /* FMADD/FNMADD/FNMSUB/FMSUB. */ | |
5253 | cost += extra_cost->fp[mode == DFmode].fma; | |
5254 | else | |
3d840f7d | 5255 | /* FMUL/FNMUL. */ |
4745e701 JG |
5256 | cost += extra_cost->fp[mode == DFmode].mult; |
5257 | } | |
5258 | ||
5259 | cost += rtx_cost (op0, MULT, 0, speed) | |
5260 | + rtx_cost (op1, MULT, 1, speed); | |
5261 | return cost; | |
5262 | } | |
43e9d192 IB |
5263 | } |
5264 | ||
67747367 JG |
5265 | static int |
5266 | aarch64_address_cost (rtx x, | |
ef4bddc2 | 5267 | machine_mode mode, |
67747367 JG |
5268 | addr_space_t as ATTRIBUTE_UNUSED, |
5269 | bool speed) | |
5270 | { | |
5271 | enum rtx_code c = GET_CODE (x); | |
5272 | const struct cpu_addrcost_table *addr_cost = aarch64_tune_params->addr_cost; | |
5273 | struct aarch64_address_info info; | |
5274 | int cost = 0; | |
5275 | info.shift = 0; | |
5276 | ||
5277 | if (!aarch64_classify_address (&info, x, mode, c, false)) | |
5278 | { | |
5279 | if (GET_CODE (x) == CONST || GET_CODE (x) == SYMBOL_REF) | |
5280 | { | |
5281 | /* This is a CONST or SYMBOL ref which will be split | |
5282 | in a different way depending on the code model in use. | |
5283 | Cost it through the generic infrastructure. */ | |
5284 | int cost_symbol_ref = rtx_cost (x, MEM, 1, speed); | |
5285 | /* Divide through by the cost of one instruction to | |
5286 | bring it to the same units as the address costs. */ | |
5287 | cost_symbol_ref /= COSTS_N_INSNS (1); | |
5288 | /* The cost is then the cost of preparing the address, | |
5289 | followed by an immediate (possibly 0) offset. */ | |
5290 | return cost_symbol_ref + addr_cost->imm_offset; | |
5291 | } | |
5292 | else | |
5293 | { | |
5294 | /* This is most likely a jump table from a case | |
5295 | statement. */ | |
5296 | return addr_cost->register_offset; | |
5297 | } | |
5298 | } | |
5299 | ||
5300 | switch (info.type) | |
5301 | { | |
5302 | case ADDRESS_LO_SUM: | |
5303 | case ADDRESS_SYMBOLIC: | |
5304 | case ADDRESS_REG_IMM: | |
5305 | cost += addr_cost->imm_offset; | |
5306 | break; | |
5307 | ||
5308 | case ADDRESS_REG_WB: | |
5309 | if (c == PRE_INC || c == PRE_DEC || c == PRE_MODIFY) | |
5310 | cost += addr_cost->pre_modify; | |
5311 | else if (c == POST_INC || c == POST_DEC || c == POST_MODIFY) | |
5312 | cost += addr_cost->post_modify; | |
5313 | else | |
5314 | gcc_unreachable (); | |
5315 | ||
5316 | break; | |
5317 | ||
5318 | case ADDRESS_REG_REG: | |
5319 | cost += addr_cost->register_offset; | |
5320 | break; | |
5321 | ||
5322 | case ADDRESS_REG_UXTW: | |
5323 | case ADDRESS_REG_SXTW: | |
5324 | cost += addr_cost->register_extend; | |
5325 | break; | |
5326 | ||
5327 | default: | |
5328 | gcc_unreachable (); | |
5329 | } | |
5330 | ||
5331 | ||
5332 | if (info.shift > 0) | |
5333 | { | |
5334 | /* For the sake of calculating the cost of the shifted register | |
5335 | component, we can treat same sized modes in the same way. */ | |
5336 | switch (GET_MODE_BITSIZE (mode)) | |
5337 | { | |
5338 | case 16: | |
5339 | cost += addr_cost->addr_scale_costs.hi; | |
5340 | break; | |
5341 | ||
5342 | case 32: | |
5343 | cost += addr_cost->addr_scale_costs.si; | |
5344 | break; | |
5345 | ||
5346 | case 64: | |
5347 | cost += addr_cost->addr_scale_costs.di; | |
5348 | break; | |
5349 | ||
5350 | /* We can't tell, or this is a 128-bit vector. */ | |
5351 | default: | |
5352 | cost += addr_cost->addr_scale_costs.ti; | |
5353 | break; | |
5354 | } | |
5355 | } | |
5356 | ||
5357 | return cost; | |
5358 | } | |
5359 | ||
7cc2145f JG |
5360 | /* Return true if the RTX X in mode MODE is a zero or sign extract |
5361 | usable in an ADD or SUB (extended register) instruction. */ | |
5362 | static bool | |
ef4bddc2 | 5363 | aarch64_rtx_arith_op_extract_p (rtx x, machine_mode mode) |
7cc2145f JG |
5364 | { |
5365 | /* Catch add with a sign extract. | |
5366 | This is add_<optab><mode>_multp2. */ | |
5367 | if (GET_CODE (x) == SIGN_EXTRACT | |
5368 | || GET_CODE (x) == ZERO_EXTRACT) | |
5369 | { | |
5370 | rtx op0 = XEXP (x, 0); | |
5371 | rtx op1 = XEXP (x, 1); | |
5372 | rtx op2 = XEXP (x, 2); | |
5373 | ||
5374 | if (GET_CODE (op0) == MULT | |
5375 | && CONST_INT_P (op1) | |
5376 | && op2 == const0_rtx | |
5377 | && CONST_INT_P (XEXP (op0, 1)) | |
5378 | && aarch64_is_extend_from_extract (mode, | |
5379 | XEXP (op0, 1), | |
5380 | op1)) | |
5381 | { | |
5382 | return true; | |
5383 | } | |
5384 | } | |
5385 | ||
5386 | return false; | |
5387 | } | |
5388 | ||
61263118 KT |
5389 | static bool |
5390 | aarch64_frint_unspec_p (unsigned int u) | |
5391 | { | |
5392 | switch (u) | |
5393 | { | |
5394 | case UNSPEC_FRINTZ: | |
5395 | case UNSPEC_FRINTP: | |
5396 | case UNSPEC_FRINTM: | |
5397 | case UNSPEC_FRINTA: | |
5398 | case UNSPEC_FRINTN: | |
5399 | case UNSPEC_FRINTX: | |
5400 | case UNSPEC_FRINTI: | |
5401 | return true; | |
5402 | ||
5403 | default: | |
5404 | return false; | |
5405 | } | |
5406 | } | |
5407 | ||
2d5ffe46 AP |
5408 | /* Calculate the cost of calculating (if_then_else (OP0) (OP1) (OP2)), |
5409 | storing it in *COST. Result is true if the total cost of the operation | |
5410 | has now been calculated. */ | |
5411 | static bool | |
5412 | aarch64_if_then_else_costs (rtx op0, rtx op1, rtx op2, int *cost, bool speed) | |
5413 | { | |
b9e3afe9 AP |
5414 | rtx inner; |
5415 | rtx comparator; | |
5416 | enum rtx_code cmpcode; | |
5417 | ||
5418 | if (COMPARISON_P (op0)) | |
5419 | { | |
5420 | inner = XEXP (op0, 0); | |
5421 | comparator = XEXP (op0, 1); | |
5422 | cmpcode = GET_CODE (op0); | |
5423 | } | |
5424 | else | |
5425 | { | |
5426 | inner = op0; | |
5427 | comparator = const0_rtx; | |
5428 | cmpcode = NE; | |
5429 | } | |
5430 | ||
2d5ffe46 AP |
5431 | if (GET_CODE (op1) == PC || GET_CODE (op2) == PC) |
5432 | { | |
5433 | /* Conditional branch. */ | |
b9e3afe9 | 5434 | if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_CC) |
2d5ffe46 AP |
5435 | return true; |
5436 | else | |
5437 | { | |
b9e3afe9 | 5438 | if (cmpcode == NE || cmpcode == EQ) |
2d5ffe46 | 5439 | { |
2d5ffe46 AP |
5440 | if (comparator == const0_rtx) |
5441 | { | |
5442 | /* TBZ/TBNZ/CBZ/CBNZ. */ | |
5443 | if (GET_CODE (inner) == ZERO_EXTRACT) | |
5444 | /* TBZ/TBNZ. */ | |
5445 | *cost += rtx_cost (XEXP (inner, 0), ZERO_EXTRACT, | |
5446 | 0, speed); | |
5447 | else | |
5448 | /* CBZ/CBNZ. */ | |
b9e3afe9 | 5449 | *cost += rtx_cost (inner, cmpcode, 0, speed); |
2d5ffe46 AP |
5450 | |
5451 | return true; | |
5452 | } | |
5453 | } | |
b9e3afe9 | 5454 | else if (cmpcode == LT || cmpcode == GE) |
2d5ffe46 | 5455 | { |
2d5ffe46 AP |
5456 | /* TBZ/TBNZ. */ |
5457 | if (comparator == const0_rtx) | |
5458 | return true; | |
5459 | } | |
5460 | } | |
5461 | } | |
b9e3afe9 | 5462 | else if (GET_MODE_CLASS (GET_MODE (inner)) == MODE_CC) |
2d5ffe46 AP |
5463 | { |
5464 | /* It's a conditional operation based on the status flags, | |
5465 | so it must be some flavor of CSEL. */ | |
5466 | ||
5467 | /* CSNEG, CSINV, and CSINC are handled for free as part of CSEL. */ | |
5468 | if (GET_CODE (op1) == NEG | |
5469 | || GET_CODE (op1) == NOT | |
5470 | || (GET_CODE (op1) == PLUS && XEXP (op1, 1) == const1_rtx)) | |
5471 | op1 = XEXP (op1, 0); | |
5472 | ||
5473 | *cost += rtx_cost (op1, IF_THEN_ELSE, 1, speed); | |
5474 | *cost += rtx_cost (op2, IF_THEN_ELSE, 2, speed); | |
5475 | return true; | |
5476 | } | |
5477 | ||
5478 | /* We don't know what this is, cost all operands. */ | |
5479 | return false; | |
5480 | } | |
5481 | ||
43e9d192 IB |
5482 | /* Calculate the cost of calculating X, storing it in *COST. Result |
5483 | is true if the total cost of the operation has now been calculated. */ | |
5484 | static bool | |
5485 | aarch64_rtx_costs (rtx x, int code, int outer ATTRIBUTE_UNUSED, | |
5486 | int param ATTRIBUTE_UNUSED, int *cost, bool speed) | |
5487 | { | |
a8eecd00 | 5488 | rtx op0, op1, op2; |
73250c4c | 5489 | const struct cpu_cost_table *extra_cost |
43e9d192 | 5490 | = aarch64_tune_params->insn_extra_cost; |
ef4bddc2 | 5491 | machine_mode mode = GET_MODE (x); |
43e9d192 | 5492 | |
7fc5ef02 JG |
5493 | /* By default, assume that everything has equivalent cost to the |
5494 | cheapest instruction. Any additional costs are applied as a delta | |
5495 | above this default. */ | |
5496 | *cost = COSTS_N_INSNS (1); | |
5497 | ||
5498 | /* TODO: The cost infrastructure currently does not handle | |
5499 | vector operations. Assume that all vector operations | |
5500 | are equally expensive. */ | |
5501 | if (VECTOR_MODE_P (mode)) | |
5502 | { | |
5503 | if (speed) | |
5504 | *cost += extra_cost->vect.alu; | |
5505 | return true; | |
5506 | } | |
5507 | ||
43e9d192 IB |
5508 | switch (code) |
5509 | { | |
5510 | case SET: | |
ba123b0d JG |
5511 | /* The cost depends entirely on the operands to SET. */ |
5512 | *cost = 0; | |
43e9d192 IB |
5513 | op0 = SET_DEST (x); |
5514 | op1 = SET_SRC (x); | |
5515 | ||
5516 | switch (GET_CODE (op0)) | |
5517 | { | |
5518 | case MEM: | |
5519 | if (speed) | |
2961177e JG |
5520 | { |
5521 | rtx address = XEXP (op0, 0); | |
5522 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
5523 | *cost += extra_cost->ldst.store; | |
5524 | else if (mode == SFmode) | |
5525 | *cost += extra_cost->ldst.storef; | |
5526 | else if (mode == DFmode) | |
5527 | *cost += extra_cost->ldst.stored; | |
5528 | ||
5529 | *cost += | |
5530 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
5531 | 0, speed)); | |
5532 | } | |
43e9d192 | 5533 | |
ba123b0d | 5534 | *cost += rtx_cost (op1, SET, 1, speed); |
43e9d192 IB |
5535 | return true; |
5536 | ||
5537 | case SUBREG: | |
5538 | if (! REG_P (SUBREG_REG (op0))) | |
5539 | *cost += rtx_cost (SUBREG_REG (op0), SET, 0, speed); | |
ba123b0d | 5540 | |
43e9d192 IB |
5541 | /* Fall through. */ |
5542 | case REG: | |
ba123b0d JG |
5543 | /* const0_rtx is in general free, but we will use an |
5544 | instruction to set a register to 0. */ | |
5545 | if (REG_P (op1) || op1 == const0_rtx) | |
5546 | { | |
5547 | /* The cost is 1 per register copied. */ | |
5548 | int n_minus_1 = (GET_MODE_SIZE (GET_MODE (op0)) - 1) | |
5549 | / UNITS_PER_WORD; | |
5550 | *cost = COSTS_N_INSNS (n_minus_1 + 1); | |
5551 | } | |
5552 | else | |
5553 | /* Cost is just the cost of the RHS of the set. */ | |
5554 | *cost += rtx_cost (op1, SET, 1, speed); | |
43e9d192 IB |
5555 | return true; |
5556 | ||
ba123b0d | 5557 | case ZERO_EXTRACT: |
43e9d192 | 5558 | case SIGN_EXTRACT: |
ba123b0d JG |
5559 | /* Bit-field insertion. Strip any redundant widening of |
5560 | the RHS to meet the width of the target. */ | |
43e9d192 IB |
5561 | if (GET_CODE (op1) == SUBREG) |
5562 | op1 = SUBREG_REG (op1); | |
5563 | if ((GET_CODE (op1) == ZERO_EXTEND | |
5564 | || GET_CODE (op1) == SIGN_EXTEND) | |
4aa81c2e | 5565 | && CONST_INT_P (XEXP (op0, 1)) |
43e9d192 IB |
5566 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (op1, 0))) |
5567 | >= INTVAL (XEXP (op0, 1)))) | |
5568 | op1 = XEXP (op1, 0); | |
ba123b0d JG |
5569 | |
5570 | if (CONST_INT_P (op1)) | |
5571 | { | |
5572 | /* MOV immediate is assumed to always be cheap. */ | |
5573 | *cost = COSTS_N_INSNS (1); | |
5574 | } | |
5575 | else | |
5576 | { | |
5577 | /* BFM. */ | |
5578 | if (speed) | |
5579 | *cost += extra_cost->alu.bfi; | |
5580 | *cost += rtx_cost (op1, (enum rtx_code) code, 1, speed); | |
5581 | } | |
5582 | ||
43e9d192 IB |
5583 | return true; |
5584 | ||
5585 | default: | |
ba123b0d JG |
5586 | /* We can't make sense of this, assume default cost. */ |
5587 | *cost = COSTS_N_INSNS (1); | |
61263118 | 5588 | return false; |
43e9d192 IB |
5589 | } |
5590 | return false; | |
5591 | ||
9dfc162c JG |
5592 | case CONST_INT: |
5593 | /* If an instruction can incorporate a constant within the | |
5594 | instruction, the instruction's expression avoids calling | |
5595 | rtx_cost() on the constant. If rtx_cost() is called on a | |
5596 | constant, then it is usually because the constant must be | |
5597 | moved into a register by one or more instructions. | |
5598 | ||
5599 | The exception is constant 0, which can be expressed | |
5600 | as XZR/WZR and is therefore free. The exception to this is | |
5601 | if we have (set (reg) (const0_rtx)) in which case we must cost | |
5602 | the move. However, we can catch that when we cost the SET, so | |
5603 | we don't need to consider that here. */ | |
5604 | if (x == const0_rtx) | |
5605 | *cost = 0; | |
5606 | else | |
5607 | { | |
5608 | /* To an approximation, building any other constant is | |
5609 | proportionally expensive to the number of instructions | |
5610 | required to build that constant. This is true whether we | |
5611 | are compiling for SPEED or otherwise. */ | |
82614948 RR |
5612 | *cost = COSTS_N_INSNS (aarch64_internal_mov_immediate |
5613 | (NULL_RTX, x, false, mode)); | |
9dfc162c JG |
5614 | } |
5615 | return true; | |
5616 | ||
5617 | case CONST_DOUBLE: | |
5618 | if (speed) | |
5619 | { | |
5620 | /* mov[df,sf]_aarch64. */ | |
5621 | if (aarch64_float_const_representable_p (x)) | |
5622 | /* FMOV (scalar immediate). */ | |
5623 | *cost += extra_cost->fp[mode == DFmode].fpconst; | |
5624 | else if (!aarch64_float_const_zero_rtx_p (x)) | |
5625 | { | |
5626 | /* This will be a load from memory. */ | |
5627 | if (mode == DFmode) | |
5628 | *cost += extra_cost->ldst.loadd; | |
5629 | else | |
5630 | *cost += extra_cost->ldst.loadf; | |
5631 | } | |
5632 | else | |
5633 | /* Otherwise this is +0.0. We get this using MOVI d0, #0 | |
5634 | or MOV v0.s[0], wzr - neither of which are modeled by the | |
5635 | cost tables. Just use the default cost. */ | |
5636 | { | |
5637 | } | |
5638 | } | |
5639 | ||
5640 | return true; | |
5641 | ||
43e9d192 IB |
5642 | case MEM: |
5643 | if (speed) | |
2961177e JG |
5644 | { |
5645 | /* For loads we want the base cost of a load, plus an | |
5646 | approximation for the additional cost of the addressing | |
5647 | mode. */ | |
5648 | rtx address = XEXP (x, 0); | |
5649 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
5650 | *cost += extra_cost->ldst.load; | |
5651 | else if (mode == SFmode) | |
5652 | *cost += extra_cost->ldst.loadf; | |
5653 | else if (mode == DFmode) | |
5654 | *cost += extra_cost->ldst.loadd; | |
5655 | ||
5656 | *cost += | |
5657 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
5658 | 0, speed)); | |
5659 | } | |
43e9d192 IB |
5660 | |
5661 | return true; | |
5662 | ||
5663 | case NEG: | |
4745e701 JG |
5664 | op0 = XEXP (x, 0); |
5665 | ||
5666 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT) | |
5667 | { | |
5668 | if (GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMPARE | |
5669 | || GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMM_COMPARE) | |
5670 | { | |
5671 | /* CSETM. */ | |
5672 | *cost += rtx_cost (XEXP (op0, 0), NEG, 0, speed); | |
5673 | return true; | |
5674 | } | |
5675 | ||
5676 | /* Cost this as SUB wzr, X. */ | |
5677 | op0 = CONST0_RTX (GET_MODE (x)); | |
5678 | op1 = XEXP (x, 0); | |
5679 | goto cost_minus; | |
5680 | } | |
5681 | ||
5682 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
5683 | { | |
5684 | /* Support (neg(fma...)) as a single instruction only if | |
5685 | sign of zeros is unimportant. This matches the decision | |
5686 | making in aarch64.md. */ | |
5687 | if (GET_CODE (op0) == FMA && !HONOR_SIGNED_ZEROS (GET_MODE (op0))) | |
5688 | { | |
5689 | /* FNMADD. */ | |
5690 | *cost = rtx_cost (op0, NEG, 0, speed); | |
5691 | return true; | |
5692 | } | |
5693 | if (speed) | |
5694 | /* FNEG. */ | |
5695 | *cost += extra_cost->fp[mode == DFmode].neg; | |
5696 | return false; | |
5697 | } | |
5698 | ||
5699 | return false; | |
43e9d192 | 5700 | |
781aeb73 KT |
5701 | case CLRSB: |
5702 | case CLZ: | |
5703 | if (speed) | |
5704 | *cost += extra_cost->alu.clz; | |
5705 | ||
5706 | return false; | |
5707 | ||
43e9d192 IB |
5708 | case COMPARE: |
5709 | op0 = XEXP (x, 0); | |
5710 | op1 = XEXP (x, 1); | |
5711 | ||
5712 | if (op1 == const0_rtx | |
5713 | && GET_CODE (op0) == AND) | |
5714 | { | |
5715 | x = op0; | |
5716 | goto cost_logic; | |
5717 | } | |
5718 | ||
a8eecd00 JG |
5719 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT) |
5720 | { | |
5721 | /* TODO: A write to the CC flags possibly costs extra, this | |
5722 | needs encoding in the cost tables. */ | |
5723 | ||
5724 | /* CC_ZESWPmode supports zero extend for free. */ | |
5725 | if (GET_MODE (x) == CC_ZESWPmode && GET_CODE (op0) == ZERO_EXTEND) | |
5726 | op0 = XEXP (op0, 0); | |
5727 | ||
5728 | /* ANDS. */ | |
5729 | if (GET_CODE (op0) == AND) | |
5730 | { | |
5731 | x = op0; | |
5732 | goto cost_logic; | |
5733 | } | |
5734 | ||
5735 | if (GET_CODE (op0) == PLUS) | |
5736 | { | |
5737 | /* ADDS (and CMN alias). */ | |
5738 | x = op0; | |
5739 | goto cost_plus; | |
5740 | } | |
5741 | ||
5742 | if (GET_CODE (op0) == MINUS) | |
5743 | { | |
5744 | /* SUBS. */ | |
5745 | x = op0; | |
5746 | goto cost_minus; | |
5747 | } | |
5748 | ||
5749 | if (GET_CODE (op1) == NEG) | |
5750 | { | |
5751 | /* CMN. */ | |
5752 | if (speed) | |
5753 | *cost += extra_cost->alu.arith; | |
5754 | ||
5755 | *cost += rtx_cost (op0, COMPARE, 0, speed); | |
5756 | *cost += rtx_cost (XEXP (op1, 0), NEG, 1, speed); | |
5757 | return true; | |
5758 | } | |
5759 | ||
5760 | /* CMP. | |
5761 | ||
5762 | Compare can freely swap the order of operands, and | |
5763 | canonicalization puts the more complex operation first. | |
5764 | But the integer MINUS logic expects the shift/extend | |
5765 | operation in op1. */ | |
5766 | if (! (REG_P (op0) | |
5767 | || (GET_CODE (op0) == SUBREG && REG_P (SUBREG_REG (op0))))) | |
5768 | { | |
5769 | op0 = XEXP (x, 1); | |
5770 | op1 = XEXP (x, 0); | |
5771 | } | |
5772 | goto cost_minus; | |
5773 | } | |
5774 | ||
5775 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT) | |
5776 | { | |
5777 | /* FCMP. */ | |
5778 | if (speed) | |
5779 | *cost += extra_cost->fp[mode == DFmode].compare; | |
5780 | ||
5781 | if (CONST_DOUBLE_P (op1) && aarch64_float_const_zero_rtx_p (op1)) | |
5782 | { | |
5783 | /* FCMP supports constant 0.0 for no extra cost. */ | |
5784 | return true; | |
5785 | } | |
5786 | return false; | |
5787 | } | |
5788 | ||
5789 | return false; | |
43e9d192 IB |
5790 | |
5791 | case MINUS: | |
4745e701 JG |
5792 | { |
5793 | op0 = XEXP (x, 0); | |
5794 | op1 = XEXP (x, 1); | |
5795 | ||
5796 | cost_minus: | |
5797 | /* Detect valid immediates. */ | |
5798 | if ((GET_MODE_CLASS (mode) == MODE_INT | |
5799 | || (GET_MODE_CLASS (mode) == MODE_CC | |
5800 | && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT)) | |
5801 | && CONST_INT_P (op1) | |
5802 | && aarch64_uimm12_shift (INTVAL (op1))) | |
5803 | { | |
5804 | *cost += rtx_cost (op0, MINUS, 0, speed); | |
43e9d192 | 5805 | |
4745e701 JG |
5806 | if (speed) |
5807 | /* SUB(S) (immediate). */ | |
5808 | *cost += extra_cost->alu.arith; | |
5809 | return true; | |
5810 | ||
5811 | } | |
5812 | ||
7cc2145f JG |
5813 | /* Look for SUB (extended register). */ |
5814 | if (aarch64_rtx_arith_op_extract_p (op1, mode)) | |
5815 | { | |
5816 | if (speed) | |
5817 | *cost += extra_cost->alu.arith_shift; | |
5818 | ||
5819 | *cost += rtx_cost (XEXP (XEXP (op1, 0), 0), | |
5820 | (enum rtx_code) GET_CODE (op1), | |
5821 | 0, speed); | |
5822 | return true; | |
5823 | } | |
5824 | ||
4745e701 JG |
5825 | rtx new_op1 = aarch64_strip_extend (op1); |
5826 | ||
5827 | /* Cost this as an FMA-alike operation. */ | |
5828 | if ((GET_CODE (new_op1) == MULT | |
5829 | || GET_CODE (new_op1) == ASHIFT) | |
5830 | && code != COMPARE) | |
5831 | { | |
5832 | *cost += aarch64_rtx_mult_cost (new_op1, MULT, | |
5833 | (enum rtx_code) code, | |
5834 | speed); | |
43e9d192 | 5835 | *cost += rtx_cost (op0, MINUS, 0, speed); |
4745e701 JG |
5836 | return true; |
5837 | } | |
43e9d192 | 5838 | |
4745e701 | 5839 | *cost += rtx_cost (new_op1, MINUS, 1, speed); |
43e9d192 | 5840 | |
4745e701 JG |
5841 | if (speed) |
5842 | { | |
5843 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
5844 | /* SUB(S). */ | |
5845 | *cost += extra_cost->alu.arith; | |
5846 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5847 | /* FSUB. */ | |
5848 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
5849 | } | |
5850 | return true; | |
5851 | } | |
43e9d192 IB |
5852 | |
5853 | case PLUS: | |
4745e701 JG |
5854 | { |
5855 | rtx new_op0; | |
43e9d192 | 5856 | |
4745e701 JG |
5857 | op0 = XEXP (x, 0); |
5858 | op1 = XEXP (x, 1); | |
43e9d192 | 5859 | |
a8eecd00 | 5860 | cost_plus: |
4745e701 JG |
5861 | if (GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMPARE |
5862 | || GET_RTX_CLASS (GET_CODE (op0)) == RTX_COMM_COMPARE) | |
5863 | { | |
5864 | /* CSINC. */ | |
5865 | *cost += rtx_cost (XEXP (op0, 0), PLUS, 0, speed); | |
5866 | *cost += rtx_cost (op1, PLUS, 1, speed); | |
5867 | return true; | |
5868 | } | |
43e9d192 | 5869 | |
4745e701 JG |
5870 | if (GET_MODE_CLASS (mode) == MODE_INT |
5871 | && CONST_INT_P (op1) | |
5872 | && aarch64_uimm12_shift (INTVAL (op1))) | |
5873 | { | |
5874 | *cost += rtx_cost (op0, PLUS, 0, speed); | |
43e9d192 | 5875 | |
4745e701 JG |
5876 | if (speed) |
5877 | /* ADD (immediate). */ | |
5878 | *cost += extra_cost->alu.arith; | |
5879 | return true; | |
5880 | } | |
5881 | ||
7cc2145f JG |
5882 | /* Look for ADD (extended register). */ |
5883 | if (aarch64_rtx_arith_op_extract_p (op0, mode)) | |
5884 | { | |
5885 | if (speed) | |
5886 | *cost += extra_cost->alu.arith_shift; | |
5887 | ||
5888 | *cost += rtx_cost (XEXP (XEXP (op0, 0), 0), | |
5889 | (enum rtx_code) GET_CODE (op0), | |
5890 | 0, speed); | |
5891 | return true; | |
5892 | } | |
5893 | ||
4745e701 JG |
5894 | /* Strip any extend, leave shifts behind as we will |
5895 | cost them through mult_cost. */ | |
5896 | new_op0 = aarch64_strip_extend (op0); | |
5897 | ||
5898 | if (GET_CODE (new_op0) == MULT | |
5899 | || GET_CODE (new_op0) == ASHIFT) | |
5900 | { | |
5901 | *cost += aarch64_rtx_mult_cost (new_op0, MULT, PLUS, | |
5902 | speed); | |
5903 | *cost += rtx_cost (op1, PLUS, 1, speed); | |
5904 | return true; | |
5905 | } | |
5906 | ||
5907 | *cost += (rtx_cost (new_op0, PLUS, 0, speed) | |
5908 | + rtx_cost (op1, PLUS, 1, speed)); | |
5909 | ||
5910 | if (speed) | |
5911 | { | |
5912 | if (GET_MODE_CLASS (mode) == MODE_INT) | |
5913 | /* ADD. */ | |
5914 | *cost += extra_cost->alu.arith; | |
5915 | else if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5916 | /* FADD. */ | |
5917 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
5918 | } | |
5919 | return true; | |
5920 | } | |
43e9d192 | 5921 | |
18b42b2a KT |
5922 | case BSWAP: |
5923 | *cost = COSTS_N_INSNS (1); | |
5924 | ||
5925 | if (speed) | |
5926 | *cost += extra_cost->alu.rev; | |
5927 | ||
5928 | return false; | |
5929 | ||
43e9d192 | 5930 | case IOR: |
f7d5cf8d KT |
5931 | if (aarch_rev16_p (x)) |
5932 | { | |
5933 | *cost = COSTS_N_INSNS (1); | |
5934 | ||
5935 | if (speed) | |
5936 | *cost += extra_cost->alu.rev; | |
5937 | ||
5938 | return true; | |
5939 | } | |
5940 | /* Fall through. */ | |
43e9d192 IB |
5941 | case XOR: |
5942 | case AND: | |
5943 | cost_logic: | |
5944 | op0 = XEXP (x, 0); | |
5945 | op1 = XEXP (x, 1); | |
5946 | ||
268c3b47 JG |
5947 | if (code == AND |
5948 | && GET_CODE (op0) == MULT | |
5949 | && CONST_INT_P (XEXP (op0, 1)) | |
5950 | && CONST_INT_P (op1) | |
5951 | && aarch64_uxt_size (exact_log2 (INTVAL (XEXP (op0, 1))), | |
5952 | INTVAL (op1)) != 0) | |
5953 | { | |
5954 | /* This is a UBFM/SBFM. */ | |
5955 | *cost += rtx_cost (XEXP (op0, 0), ZERO_EXTRACT, 0, speed); | |
5956 | if (speed) | |
5957 | *cost += extra_cost->alu.bfx; | |
5958 | return true; | |
5959 | } | |
5960 | ||
43e9d192 IB |
5961 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT) |
5962 | { | |
268c3b47 JG |
5963 | /* We possibly get the immediate for free, this is not |
5964 | modelled. */ | |
43e9d192 IB |
5965 | if (CONST_INT_P (op1) |
5966 | && aarch64_bitmask_imm (INTVAL (op1), GET_MODE (x))) | |
5967 | { | |
268c3b47 JG |
5968 | *cost += rtx_cost (op0, (enum rtx_code) code, 0, speed); |
5969 | ||
5970 | if (speed) | |
5971 | *cost += extra_cost->alu.logical; | |
5972 | ||
5973 | return true; | |
43e9d192 IB |
5974 | } |
5975 | else | |
5976 | { | |
268c3b47 JG |
5977 | rtx new_op0 = op0; |
5978 | ||
5979 | /* Handle ORN, EON, or BIC. */ | |
43e9d192 IB |
5980 | if (GET_CODE (op0) == NOT) |
5981 | op0 = XEXP (op0, 0); | |
268c3b47 JG |
5982 | |
5983 | new_op0 = aarch64_strip_shift (op0); | |
5984 | ||
5985 | /* If we had a shift on op0 then this is a logical-shift- | |
5986 | by-register/immediate operation. Otherwise, this is just | |
5987 | a logical operation. */ | |
5988 | if (speed) | |
5989 | { | |
5990 | if (new_op0 != op0) | |
5991 | { | |
5992 | /* Shift by immediate. */ | |
5993 | if (CONST_INT_P (XEXP (op0, 1))) | |
5994 | *cost += extra_cost->alu.log_shift; | |
5995 | else | |
5996 | *cost += extra_cost->alu.log_shift_reg; | |
5997 | } | |
5998 | else | |
5999 | *cost += extra_cost->alu.logical; | |
6000 | } | |
6001 | ||
6002 | /* In both cases we want to cost both operands. */ | |
6003 | *cost += rtx_cost (new_op0, (enum rtx_code) code, 0, speed) | |
6004 | + rtx_cost (op1, (enum rtx_code) code, 1, speed); | |
6005 | ||
6006 | return true; | |
43e9d192 | 6007 | } |
43e9d192 IB |
6008 | } |
6009 | return false; | |
6010 | ||
268c3b47 JG |
6011 | case NOT: |
6012 | /* MVN. */ | |
6013 | if (speed) | |
6014 | *cost += extra_cost->alu.logical; | |
6015 | ||
6016 | /* The logical instruction could have the shifted register form, | |
6017 | but the cost is the same if the shift is processed as a separate | |
6018 | instruction, so we don't bother with it here. */ | |
6019 | return false; | |
6020 | ||
43e9d192 | 6021 | case ZERO_EXTEND: |
b1685e62 JG |
6022 | |
6023 | op0 = XEXP (x, 0); | |
6024 | /* If a value is written in SI mode, then zero extended to DI | |
6025 | mode, the operation will in general be free as a write to | |
6026 | a 'w' register implicitly zeroes the upper bits of an 'x' | |
6027 | register. However, if this is | |
6028 | ||
6029 | (set (reg) (zero_extend (reg))) | |
6030 | ||
6031 | we must cost the explicit register move. */ | |
6032 | if (mode == DImode | |
6033 | && GET_MODE (op0) == SImode | |
6034 | && outer == SET) | |
6035 | { | |
6036 | int op_cost = rtx_cost (XEXP (x, 0), ZERO_EXTEND, 0, speed); | |
6037 | ||
6038 | if (!op_cost && speed) | |
6039 | /* MOV. */ | |
6040 | *cost += extra_cost->alu.extend; | |
6041 | else | |
6042 | /* Free, the cost is that of the SI mode operation. */ | |
6043 | *cost = op_cost; | |
6044 | ||
6045 | return true; | |
6046 | } | |
6047 | else if (MEM_P (XEXP (x, 0))) | |
43e9d192 | 6048 | { |
b1685e62 JG |
6049 | /* All loads can zero extend to any size for free. */ |
6050 | *cost = rtx_cost (XEXP (x, 0), ZERO_EXTEND, param, speed); | |
43e9d192 IB |
6051 | return true; |
6052 | } | |
b1685e62 JG |
6053 | |
6054 | /* UXTB/UXTH. */ | |
6055 | if (speed) | |
6056 | *cost += extra_cost->alu.extend; | |
6057 | ||
43e9d192 IB |
6058 | return false; |
6059 | ||
6060 | case SIGN_EXTEND: | |
b1685e62 | 6061 | if (MEM_P (XEXP (x, 0))) |
43e9d192 | 6062 | { |
b1685e62 JG |
6063 | /* LDRSH. */ |
6064 | if (speed) | |
6065 | { | |
6066 | rtx address = XEXP (XEXP (x, 0), 0); | |
6067 | *cost += extra_cost->ldst.load_sign_extend; | |
6068 | ||
6069 | *cost += | |
6070 | COSTS_N_INSNS (aarch64_address_cost (address, mode, | |
6071 | 0, speed)); | |
6072 | } | |
43e9d192 IB |
6073 | return true; |
6074 | } | |
b1685e62 JG |
6075 | |
6076 | if (speed) | |
6077 | *cost += extra_cost->alu.extend; | |
43e9d192 IB |
6078 | return false; |
6079 | ||
ba0cfa17 JG |
6080 | case ASHIFT: |
6081 | op0 = XEXP (x, 0); | |
6082 | op1 = XEXP (x, 1); | |
6083 | ||
6084 | if (CONST_INT_P (op1)) | |
6085 | { | |
6086 | /* LSL (immediate), UBMF, UBFIZ and friends. These are all | |
6087 | aliases. */ | |
6088 | if (speed) | |
6089 | *cost += extra_cost->alu.shift; | |
6090 | ||
6091 | /* We can incorporate zero/sign extend for free. */ | |
6092 | if (GET_CODE (op0) == ZERO_EXTEND | |
6093 | || GET_CODE (op0) == SIGN_EXTEND) | |
6094 | op0 = XEXP (op0, 0); | |
6095 | ||
6096 | *cost += rtx_cost (op0, ASHIFT, 0, speed); | |
6097 | return true; | |
6098 | } | |
6099 | else | |
6100 | { | |
6101 | /* LSLV. */ | |
6102 | if (speed) | |
6103 | *cost += extra_cost->alu.shift_reg; | |
6104 | ||
6105 | return false; /* All arguments need to be in registers. */ | |
6106 | } | |
6107 | ||
43e9d192 | 6108 | case ROTATE: |
43e9d192 IB |
6109 | case ROTATERT: |
6110 | case LSHIFTRT: | |
43e9d192 | 6111 | case ASHIFTRT: |
ba0cfa17 JG |
6112 | op0 = XEXP (x, 0); |
6113 | op1 = XEXP (x, 1); | |
43e9d192 | 6114 | |
ba0cfa17 JG |
6115 | if (CONST_INT_P (op1)) |
6116 | { | |
6117 | /* ASR (immediate) and friends. */ | |
6118 | if (speed) | |
6119 | *cost += extra_cost->alu.shift; | |
43e9d192 | 6120 | |
ba0cfa17 JG |
6121 | *cost += rtx_cost (op0, (enum rtx_code) code, 0, speed); |
6122 | return true; | |
6123 | } | |
6124 | else | |
6125 | { | |
6126 | ||
6127 | /* ASR (register) and friends. */ | |
6128 | if (speed) | |
6129 | *cost += extra_cost->alu.shift_reg; | |
6130 | ||
6131 | return false; /* All arguments need to be in registers. */ | |
6132 | } | |
43e9d192 | 6133 | |
909734be JG |
6134 | case SYMBOL_REF: |
6135 | ||
6136 | if (aarch64_cmodel == AARCH64_CMODEL_LARGE) | |
6137 | { | |
6138 | /* LDR. */ | |
6139 | if (speed) | |
6140 | *cost += extra_cost->ldst.load; | |
6141 | } | |
6142 | else if (aarch64_cmodel == AARCH64_CMODEL_SMALL | |
6143 | || aarch64_cmodel == AARCH64_CMODEL_SMALL_PIC) | |
6144 | { | |
6145 | /* ADRP, followed by ADD. */ | |
6146 | *cost += COSTS_N_INSNS (1); | |
6147 | if (speed) | |
6148 | *cost += 2 * extra_cost->alu.arith; | |
6149 | } | |
6150 | else if (aarch64_cmodel == AARCH64_CMODEL_TINY | |
6151 | || aarch64_cmodel == AARCH64_CMODEL_TINY_PIC) | |
6152 | { | |
6153 | /* ADR. */ | |
6154 | if (speed) | |
6155 | *cost += extra_cost->alu.arith; | |
6156 | } | |
6157 | ||
6158 | if (flag_pic) | |
6159 | { | |
6160 | /* One extra load instruction, after accessing the GOT. */ | |
6161 | *cost += COSTS_N_INSNS (1); | |
6162 | if (speed) | |
6163 | *cost += extra_cost->ldst.load; | |
6164 | } | |
43e9d192 IB |
6165 | return true; |
6166 | ||
909734be | 6167 | case HIGH: |
43e9d192 | 6168 | case LO_SUM: |
909734be JG |
6169 | /* ADRP/ADD (immediate). */ |
6170 | if (speed) | |
6171 | *cost += extra_cost->alu.arith; | |
43e9d192 IB |
6172 | return true; |
6173 | ||
6174 | case ZERO_EXTRACT: | |
6175 | case SIGN_EXTRACT: | |
7cc2145f JG |
6176 | /* UBFX/SBFX. */ |
6177 | if (speed) | |
6178 | *cost += extra_cost->alu.bfx; | |
6179 | ||
6180 | /* We can trust that the immediates used will be correct (there | |
6181 | are no by-register forms), so we need only cost op0. */ | |
6182 | *cost += rtx_cost (XEXP (x, 0), (enum rtx_code) code, 0, speed); | |
43e9d192 IB |
6183 | return true; |
6184 | ||
6185 | case MULT: | |
4745e701 JG |
6186 | *cost += aarch64_rtx_mult_cost (x, MULT, 0, speed); |
6187 | /* aarch64_rtx_mult_cost always handles recursion to its | |
6188 | operands. */ | |
6189 | return true; | |
43e9d192 IB |
6190 | |
6191 | case MOD: | |
6192 | case UMOD: | |
43e9d192 IB |
6193 | if (speed) |
6194 | { | |
6195 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT) | |
73250c4c KT |
6196 | *cost += (extra_cost->mult[GET_MODE (x) == DImode].add |
6197 | + extra_cost->mult[GET_MODE (x) == DImode].idiv); | |
43e9d192 | 6198 | else if (GET_MODE (x) == DFmode) |
73250c4c KT |
6199 | *cost += (extra_cost->fp[1].mult |
6200 | + extra_cost->fp[1].div); | |
43e9d192 | 6201 | else if (GET_MODE (x) == SFmode) |
73250c4c KT |
6202 | *cost += (extra_cost->fp[0].mult |
6203 | + extra_cost->fp[0].div); | |
43e9d192 IB |
6204 | } |
6205 | return false; /* All arguments need to be in registers. */ | |
6206 | ||
6207 | case DIV: | |
6208 | case UDIV: | |
4105fe38 | 6209 | case SQRT: |
43e9d192 IB |
6210 | if (speed) |
6211 | { | |
4105fe38 JG |
6212 | if (GET_MODE_CLASS (mode) == MODE_INT) |
6213 | /* There is no integer SQRT, so only DIV and UDIV can get | |
6214 | here. */ | |
6215 | *cost += extra_cost->mult[mode == DImode].idiv; | |
6216 | else | |
6217 | *cost += extra_cost->fp[mode == DFmode].div; | |
43e9d192 IB |
6218 | } |
6219 | return false; /* All arguments need to be in registers. */ | |
6220 | ||
a8eecd00 | 6221 | case IF_THEN_ELSE: |
2d5ffe46 AP |
6222 | return aarch64_if_then_else_costs (XEXP (x, 0), XEXP (x, 1), |
6223 | XEXP (x, 2), cost, speed); | |
a8eecd00 JG |
6224 | |
6225 | case EQ: | |
6226 | case NE: | |
6227 | case GT: | |
6228 | case GTU: | |
6229 | case LT: | |
6230 | case LTU: | |
6231 | case GE: | |
6232 | case GEU: | |
6233 | case LE: | |
6234 | case LEU: | |
6235 | ||
6236 | return false; /* All arguments must be in registers. */ | |
6237 | ||
b292109f JG |
6238 | case FMA: |
6239 | op0 = XEXP (x, 0); | |
6240 | op1 = XEXP (x, 1); | |
6241 | op2 = XEXP (x, 2); | |
6242 | ||
6243 | if (speed) | |
6244 | *cost += extra_cost->fp[mode == DFmode].fma; | |
6245 | ||
6246 | /* FMSUB, FNMADD, and FNMSUB are free. */ | |
6247 | if (GET_CODE (op0) == NEG) | |
6248 | op0 = XEXP (op0, 0); | |
6249 | ||
6250 | if (GET_CODE (op2) == NEG) | |
6251 | op2 = XEXP (op2, 0); | |
6252 | ||
6253 | /* aarch64_fnma4_elt_to_64v2df has the NEG as operand 1, | |
6254 | and the by-element operand as operand 0. */ | |
6255 | if (GET_CODE (op1) == NEG) | |
6256 | op1 = XEXP (op1, 0); | |
6257 | ||
6258 | /* Catch vector-by-element operations. The by-element operand can | |
6259 | either be (vec_duplicate (vec_select (x))) or just | |
6260 | (vec_select (x)), depending on whether we are multiplying by | |
6261 | a vector or a scalar. | |
6262 | ||
6263 | Canonicalization is not very good in these cases, FMA4 will put the | |
6264 | by-element operand as operand 0, FNMA4 will have it as operand 1. */ | |
6265 | if (GET_CODE (op0) == VEC_DUPLICATE) | |
6266 | op0 = XEXP (op0, 0); | |
6267 | else if (GET_CODE (op1) == VEC_DUPLICATE) | |
6268 | op1 = XEXP (op1, 0); | |
6269 | ||
6270 | if (GET_CODE (op0) == VEC_SELECT) | |
6271 | op0 = XEXP (op0, 0); | |
6272 | else if (GET_CODE (op1) == VEC_SELECT) | |
6273 | op1 = XEXP (op1, 0); | |
6274 | ||
6275 | /* If the remaining parameters are not registers, | |
6276 | get the cost to put them into registers. */ | |
6277 | *cost += rtx_cost (op0, FMA, 0, speed); | |
6278 | *cost += rtx_cost (op1, FMA, 1, speed); | |
6279 | *cost += rtx_cost (op2, FMA, 2, speed); | |
6280 | return true; | |
6281 | ||
6282 | case FLOAT_EXTEND: | |
6283 | if (speed) | |
6284 | *cost += extra_cost->fp[mode == DFmode].widen; | |
6285 | return false; | |
6286 | ||
6287 | case FLOAT_TRUNCATE: | |
6288 | if (speed) | |
6289 | *cost += extra_cost->fp[mode == DFmode].narrow; | |
6290 | return false; | |
6291 | ||
61263118 KT |
6292 | case FIX: |
6293 | case UNSIGNED_FIX: | |
6294 | x = XEXP (x, 0); | |
6295 | /* Strip the rounding part. They will all be implemented | |
6296 | by the fcvt* family of instructions anyway. */ | |
6297 | if (GET_CODE (x) == UNSPEC) | |
6298 | { | |
6299 | unsigned int uns_code = XINT (x, 1); | |
6300 | ||
6301 | if (uns_code == UNSPEC_FRINTA | |
6302 | || uns_code == UNSPEC_FRINTM | |
6303 | || uns_code == UNSPEC_FRINTN | |
6304 | || uns_code == UNSPEC_FRINTP | |
6305 | || uns_code == UNSPEC_FRINTZ) | |
6306 | x = XVECEXP (x, 0, 0); | |
6307 | } | |
6308 | ||
6309 | if (speed) | |
6310 | *cost += extra_cost->fp[GET_MODE (x) == DFmode].toint; | |
6311 | ||
6312 | *cost += rtx_cost (x, (enum rtx_code) code, 0, speed); | |
6313 | return true; | |
6314 | ||
b292109f JG |
6315 | case ABS: |
6316 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
6317 | { | |
6318 | /* FABS and FNEG are analogous. */ | |
6319 | if (speed) | |
6320 | *cost += extra_cost->fp[mode == DFmode].neg; | |
6321 | } | |
6322 | else | |
6323 | { | |
6324 | /* Integer ABS will either be split to | |
6325 | two arithmetic instructions, or will be an ABS | |
6326 | (scalar), which we don't model. */ | |
6327 | *cost = COSTS_N_INSNS (2); | |
6328 | if (speed) | |
6329 | *cost += 2 * extra_cost->alu.arith; | |
6330 | } | |
6331 | return false; | |
6332 | ||
6333 | case SMAX: | |
6334 | case SMIN: | |
6335 | if (speed) | |
6336 | { | |
6337 | /* FMAXNM/FMINNM/FMAX/FMIN. | |
6338 | TODO: This may not be accurate for all implementations, but | |
6339 | we do not model this in the cost tables. */ | |
6340 | *cost += extra_cost->fp[mode == DFmode].addsub; | |
6341 | } | |
6342 | return false; | |
6343 | ||
61263118 KT |
6344 | case UNSPEC: |
6345 | /* The floating point round to integer frint* instructions. */ | |
6346 | if (aarch64_frint_unspec_p (XINT (x, 1))) | |
6347 | { | |
6348 | if (speed) | |
6349 | *cost += extra_cost->fp[mode == DFmode].roundint; | |
6350 | ||
6351 | return false; | |
6352 | } | |
781aeb73 KT |
6353 | |
6354 | if (XINT (x, 1) == UNSPEC_RBIT) | |
6355 | { | |
6356 | if (speed) | |
6357 | *cost += extra_cost->alu.rev; | |
6358 | ||
6359 | return false; | |
6360 | } | |
61263118 KT |
6361 | break; |
6362 | ||
fb620c4a JG |
6363 | case TRUNCATE: |
6364 | ||
6365 | /* Decompose <su>muldi3_highpart. */ | |
6366 | if (/* (truncate:DI */ | |
6367 | mode == DImode | |
6368 | /* (lshiftrt:TI */ | |
6369 | && GET_MODE (XEXP (x, 0)) == TImode | |
6370 | && GET_CODE (XEXP (x, 0)) == LSHIFTRT | |
6371 | /* (mult:TI */ | |
6372 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT | |
6373 | /* (ANY_EXTEND:TI (reg:DI)) | |
6374 | (ANY_EXTEND:TI (reg:DI))) */ | |
6375 | && ((GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == ZERO_EXTEND | |
6376 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == ZERO_EXTEND) | |
6377 | || (GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == SIGN_EXTEND | |
6378 | && GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == SIGN_EXTEND)) | |
6379 | && GET_MODE (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 0)) == DImode | |
6380 | && GET_MODE (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 1), 0)) == DImode | |
6381 | /* (const_int 64) */ | |
6382 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
6383 | && UINTVAL (XEXP (XEXP (x, 0), 1)) == 64) | |
6384 | { | |
6385 | /* UMULH/SMULH. */ | |
6386 | if (speed) | |
6387 | *cost += extra_cost->mult[mode == DImode].extend; | |
6388 | *cost += rtx_cost (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 0), 0), | |
6389 | MULT, 0, speed); | |
6390 | *cost += rtx_cost (XEXP (XEXP (XEXP (XEXP (x, 0), 0), 1), 0), | |
6391 | MULT, 1, speed); | |
6392 | return true; | |
6393 | } | |
6394 | ||
6395 | /* Fall through. */ | |
43e9d192 | 6396 | default: |
61263118 | 6397 | break; |
43e9d192 | 6398 | } |
61263118 KT |
6399 | |
6400 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
6401 | fprintf (dump_file, | |
6402 | "\nFailed to cost RTX. Assuming default cost.\n"); | |
6403 | ||
6404 | return true; | |
43e9d192 IB |
6405 | } |
6406 | ||
0ee859b5 JG |
6407 | /* Wrapper around aarch64_rtx_costs, dumps the partial, or total cost |
6408 | calculated for X. This cost is stored in *COST. Returns true | |
6409 | if the total cost of X was calculated. */ | |
6410 | static bool | |
6411 | aarch64_rtx_costs_wrapper (rtx x, int code, int outer, | |
6412 | int param, int *cost, bool speed) | |
6413 | { | |
6414 | bool result = aarch64_rtx_costs (x, code, outer, param, cost, speed); | |
6415 | ||
6416 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
6417 | { | |
6418 | print_rtl_single (dump_file, x); | |
6419 | fprintf (dump_file, "\n%s cost: %d (%s)\n", | |
6420 | speed ? "Hot" : "Cold", | |
6421 | *cost, result ? "final" : "partial"); | |
6422 | } | |
6423 | ||
6424 | return result; | |
6425 | } | |
6426 | ||
43e9d192 | 6427 | static int |
ef4bddc2 | 6428 | aarch64_register_move_cost (machine_mode mode, |
8a3a7e67 | 6429 | reg_class_t from_i, reg_class_t to_i) |
43e9d192 | 6430 | { |
8a3a7e67 RH |
6431 | enum reg_class from = (enum reg_class) from_i; |
6432 | enum reg_class to = (enum reg_class) to_i; | |
43e9d192 IB |
6433 | const struct cpu_regmove_cost *regmove_cost |
6434 | = aarch64_tune_params->regmove_cost; | |
6435 | ||
3be07662 WD |
6436 | /* Caller save and pointer regs are equivalent to GENERAL_REGS. */ |
6437 | if (to == CALLER_SAVE_REGS || to == POINTER_REGS) | |
6438 | to = GENERAL_REGS; | |
6439 | ||
6440 | if (from == CALLER_SAVE_REGS || from == POINTER_REGS) | |
6441 | from = GENERAL_REGS; | |
6442 | ||
6ee70f81 AP |
6443 | /* Moving between GPR and stack cost is the same as GP2GP. */ |
6444 | if ((from == GENERAL_REGS && to == STACK_REG) | |
6445 | || (to == GENERAL_REGS && from == STACK_REG)) | |
6446 | return regmove_cost->GP2GP; | |
6447 | ||
6448 | /* To/From the stack register, we move via the gprs. */ | |
6449 | if (to == STACK_REG || from == STACK_REG) | |
6450 | return aarch64_register_move_cost (mode, from, GENERAL_REGS) | |
6451 | + aarch64_register_move_cost (mode, GENERAL_REGS, to); | |
6452 | ||
8919453c WD |
6453 | if (GET_MODE_SIZE (mode) == 16) |
6454 | { | |
6455 | /* 128-bit operations on general registers require 2 instructions. */ | |
6456 | if (from == GENERAL_REGS && to == GENERAL_REGS) | |
6457 | return regmove_cost->GP2GP * 2; | |
6458 | else if (from == GENERAL_REGS) | |
6459 | return regmove_cost->GP2FP * 2; | |
6460 | else if (to == GENERAL_REGS) | |
6461 | return regmove_cost->FP2GP * 2; | |
6462 | ||
6463 | /* When AdvSIMD instructions are disabled it is not possible to move | |
6464 | a 128-bit value directly between Q registers. This is handled in | |
6465 | secondary reload. A general register is used as a scratch to move | |
6466 | the upper DI value and the lower DI value is moved directly, | |
6467 | hence the cost is the sum of three moves. */ | |
6468 | if (! TARGET_SIMD) | |
6469 | return regmove_cost->GP2FP + regmove_cost->FP2GP + regmove_cost->FP2FP; | |
6470 | ||
6471 | return regmove_cost->FP2FP; | |
6472 | } | |
6473 | ||
43e9d192 IB |
6474 | if (from == GENERAL_REGS && to == GENERAL_REGS) |
6475 | return regmove_cost->GP2GP; | |
6476 | else if (from == GENERAL_REGS) | |
6477 | return regmove_cost->GP2FP; | |
6478 | else if (to == GENERAL_REGS) | |
6479 | return regmove_cost->FP2GP; | |
6480 | ||
43e9d192 IB |
6481 | return regmove_cost->FP2FP; |
6482 | } | |
6483 | ||
6484 | static int | |
ef4bddc2 | 6485 | aarch64_memory_move_cost (machine_mode mode ATTRIBUTE_UNUSED, |
43e9d192 IB |
6486 | reg_class_t rclass ATTRIBUTE_UNUSED, |
6487 | bool in ATTRIBUTE_UNUSED) | |
6488 | { | |
6489 | return aarch64_tune_params->memmov_cost; | |
6490 | } | |
6491 | ||
d126a4ae AP |
6492 | /* Return the number of instructions that can be issued per cycle. */ |
6493 | static int | |
6494 | aarch64_sched_issue_rate (void) | |
6495 | { | |
6496 | return aarch64_tune_params->issue_rate; | |
6497 | } | |
6498 | ||
d03f7e44 MK |
6499 | static int |
6500 | aarch64_sched_first_cycle_multipass_dfa_lookahead (void) | |
6501 | { | |
6502 | int issue_rate = aarch64_sched_issue_rate (); | |
6503 | ||
6504 | return issue_rate > 1 && !sched_fusion ? issue_rate : 0; | |
6505 | } | |
6506 | ||
8990e73a TB |
6507 | /* Vectorizer cost model target hooks. */ |
6508 | ||
6509 | /* Implement targetm.vectorize.builtin_vectorization_cost. */ | |
6510 | static int | |
6511 | aarch64_builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost, | |
6512 | tree vectype, | |
6513 | int misalign ATTRIBUTE_UNUSED) | |
6514 | { | |
6515 | unsigned elements; | |
6516 | ||
6517 | switch (type_of_cost) | |
6518 | { | |
6519 | case scalar_stmt: | |
6520 | return aarch64_tune_params->vec_costs->scalar_stmt_cost; | |
6521 | ||
6522 | case scalar_load: | |
6523 | return aarch64_tune_params->vec_costs->scalar_load_cost; | |
6524 | ||
6525 | case scalar_store: | |
6526 | return aarch64_tune_params->vec_costs->scalar_store_cost; | |
6527 | ||
6528 | case vector_stmt: | |
6529 | return aarch64_tune_params->vec_costs->vec_stmt_cost; | |
6530 | ||
6531 | case vector_load: | |
6532 | return aarch64_tune_params->vec_costs->vec_align_load_cost; | |
6533 | ||
6534 | case vector_store: | |
6535 | return aarch64_tune_params->vec_costs->vec_store_cost; | |
6536 | ||
6537 | case vec_to_scalar: | |
6538 | return aarch64_tune_params->vec_costs->vec_to_scalar_cost; | |
6539 | ||
6540 | case scalar_to_vec: | |
6541 | return aarch64_tune_params->vec_costs->scalar_to_vec_cost; | |
6542 | ||
6543 | case unaligned_load: | |
6544 | return aarch64_tune_params->vec_costs->vec_unalign_load_cost; | |
6545 | ||
6546 | case unaligned_store: | |
6547 | return aarch64_tune_params->vec_costs->vec_unalign_store_cost; | |
6548 | ||
6549 | case cond_branch_taken: | |
6550 | return aarch64_tune_params->vec_costs->cond_taken_branch_cost; | |
6551 | ||
6552 | case cond_branch_not_taken: | |
6553 | return aarch64_tune_params->vec_costs->cond_not_taken_branch_cost; | |
6554 | ||
6555 | case vec_perm: | |
6556 | case vec_promote_demote: | |
6557 | return aarch64_tune_params->vec_costs->vec_stmt_cost; | |
6558 | ||
6559 | case vec_construct: | |
6560 | elements = TYPE_VECTOR_SUBPARTS (vectype); | |
6561 | return elements / 2 + 1; | |
6562 | ||
6563 | default: | |
6564 | gcc_unreachable (); | |
6565 | } | |
6566 | } | |
6567 | ||
6568 | /* Implement targetm.vectorize.add_stmt_cost. */ | |
6569 | static unsigned | |
6570 | aarch64_add_stmt_cost (void *data, int count, enum vect_cost_for_stmt kind, | |
6571 | struct _stmt_vec_info *stmt_info, int misalign, | |
6572 | enum vect_cost_model_location where) | |
6573 | { | |
6574 | unsigned *cost = (unsigned *) data; | |
6575 | unsigned retval = 0; | |
6576 | ||
6577 | if (flag_vect_cost_model) | |
6578 | { | |
6579 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; | |
6580 | int stmt_cost = | |
6581 | aarch64_builtin_vectorization_cost (kind, vectype, misalign); | |
6582 | ||
6583 | /* Statements in an inner loop relative to the loop being | |
6584 | vectorized are weighted more heavily. The value here is | |
6585 | a function (linear for now) of the loop nest level. */ | |
6586 | if (where == vect_body && stmt_info && stmt_in_inner_loop_p (stmt_info)) | |
6587 | { | |
6588 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
6589 | struct loop *loop = LOOP_VINFO_LOOP (loop_info); | |
6590 | unsigned nest_level = loop_depth (loop); | |
6591 | ||
6592 | count *= nest_level; | |
6593 | } | |
6594 | ||
6595 | retval = (unsigned) (count * stmt_cost); | |
6596 | cost[where] += retval; | |
6597 | } | |
6598 | ||
6599 | return retval; | |
6600 | } | |
6601 | ||
43e9d192 IB |
6602 | static void initialize_aarch64_code_model (void); |
6603 | ||
6604 | /* Parse the architecture extension string. */ | |
6605 | ||
6606 | static void | |
6607 | aarch64_parse_extension (char *str) | |
6608 | { | |
6609 | /* The extension string is parsed left to right. */ | |
6610 | const struct aarch64_option_extension *opt = NULL; | |
6611 | ||
6612 | /* Flag to say whether we are adding or removing an extension. */ | |
6613 | int adding_ext = -1; | |
6614 | ||
6615 | while (str != NULL && *str != 0) | |
6616 | { | |
6617 | char *ext; | |
6618 | size_t len; | |
6619 | ||
6620 | str++; | |
6621 | ext = strchr (str, '+'); | |
6622 | ||
6623 | if (ext != NULL) | |
6624 | len = ext - str; | |
6625 | else | |
6626 | len = strlen (str); | |
6627 | ||
6628 | if (len >= 2 && strncmp (str, "no", 2) == 0) | |
6629 | { | |
6630 | adding_ext = 0; | |
6631 | len -= 2; | |
6632 | str += 2; | |
6633 | } | |
6634 | else if (len > 0) | |
6635 | adding_ext = 1; | |
6636 | ||
6637 | if (len == 0) | |
6638 | { | |
217d0904 KT |
6639 | error ("missing feature modifier after %qs", adding_ext ? "+" |
6640 | : "+no"); | |
43e9d192 IB |
6641 | return; |
6642 | } | |
6643 | ||
6644 | /* Scan over the extensions table trying to find an exact match. */ | |
6645 | for (opt = all_extensions; opt->name != NULL; opt++) | |
6646 | { | |
6647 | if (strlen (opt->name) == len && strncmp (opt->name, str, len) == 0) | |
6648 | { | |
6649 | /* Add or remove the extension. */ | |
6650 | if (adding_ext) | |
6651 | aarch64_isa_flags |= opt->flags_on; | |
6652 | else | |
6653 | aarch64_isa_flags &= ~(opt->flags_off); | |
6654 | break; | |
6655 | } | |
6656 | } | |
6657 | ||
6658 | if (opt->name == NULL) | |
6659 | { | |
6660 | /* Extension not found in list. */ | |
6661 | error ("unknown feature modifier %qs", str); | |
6662 | return; | |
6663 | } | |
6664 | ||
6665 | str = ext; | |
6666 | }; | |
6667 | ||
6668 | return; | |
6669 | } | |
6670 | ||
6671 | /* Parse the ARCH string. */ | |
6672 | ||
6673 | static void | |
6674 | aarch64_parse_arch (void) | |
6675 | { | |
6676 | char *ext; | |
6677 | const struct processor *arch; | |
6678 | char *str = (char *) alloca (strlen (aarch64_arch_string) + 1); | |
6679 | size_t len; | |
6680 | ||
6681 | strcpy (str, aarch64_arch_string); | |
6682 | ||
6683 | ext = strchr (str, '+'); | |
6684 | ||
6685 | if (ext != NULL) | |
6686 | len = ext - str; | |
6687 | else | |
6688 | len = strlen (str); | |
6689 | ||
6690 | if (len == 0) | |
6691 | { | |
6692 | error ("missing arch name in -march=%qs", str); | |
6693 | return; | |
6694 | } | |
6695 | ||
6696 | /* Loop through the list of supported ARCHs to find a match. */ | |
6697 | for (arch = all_architectures; arch->name != NULL; arch++) | |
6698 | { | |
6699 | if (strlen (arch->name) == len && strncmp (arch->name, str, len) == 0) | |
6700 | { | |
6701 | selected_arch = arch; | |
6702 | aarch64_isa_flags = selected_arch->flags; | |
ffee7aa9 JG |
6703 | |
6704 | if (!selected_cpu) | |
6705 | selected_cpu = &all_cores[selected_arch->core]; | |
43e9d192 IB |
6706 | |
6707 | if (ext != NULL) | |
6708 | { | |
6709 | /* ARCH string contains at least one extension. */ | |
6710 | aarch64_parse_extension (ext); | |
6711 | } | |
6712 | ||
ffee7aa9 JG |
6713 | if (strcmp (selected_arch->arch, selected_cpu->arch)) |
6714 | { | |
6715 | warning (0, "switch -mcpu=%s conflicts with -march=%s switch", | |
6716 | selected_cpu->name, selected_arch->name); | |
6717 | } | |
6718 | ||
43e9d192 IB |
6719 | return; |
6720 | } | |
6721 | } | |
6722 | ||
6723 | /* ARCH name not found in list. */ | |
6724 | error ("unknown value %qs for -march", str); | |
6725 | return; | |
6726 | } | |
6727 | ||
6728 | /* Parse the CPU string. */ | |
6729 | ||
6730 | static void | |
6731 | aarch64_parse_cpu (void) | |
6732 | { | |
6733 | char *ext; | |
6734 | const struct processor *cpu; | |
6735 | char *str = (char *) alloca (strlen (aarch64_cpu_string) + 1); | |
6736 | size_t len; | |
6737 | ||
6738 | strcpy (str, aarch64_cpu_string); | |
6739 | ||
6740 | ext = strchr (str, '+'); | |
6741 | ||
6742 | if (ext != NULL) | |
6743 | len = ext - str; | |
6744 | else | |
6745 | len = strlen (str); | |
6746 | ||
6747 | if (len == 0) | |
6748 | { | |
6749 | error ("missing cpu name in -mcpu=%qs", str); | |
6750 | return; | |
6751 | } | |
6752 | ||
6753 | /* Loop through the list of supported CPUs to find a match. */ | |
6754 | for (cpu = all_cores; cpu->name != NULL; cpu++) | |
6755 | { | |
6756 | if (strlen (cpu->name) == len && strncmp (cpu->name, str, len) == 0) | |
6757 | { | |
6758 | selected_cpu = cpu; | |
6759 | aarch64_isa_flags = selected_cpu->flags; | |
6760 | ||
6761 | if (ext != NULL) | |
6762 | { | |
6763 | /* CPU string contains at least one extension. */ | |
6764 | aarch64_parse_extension (ext); | |
6765 | } | |
6766 | ||
6767 | return; | |
6768 | } | |
6769 | } | |
6770 | ||
6771 | /* CPU name not found in list. */ | |
6772 | error ("unknown value %qs for -mcpu", str); | |
6773 | return; | |
6774 | } | |
6775 | ||
6776 | /* Parse the TUNE string. */ | |
6777 | ||
6778 | static void | |
6779 | aarch64_parse_tune (void) | |
6780 | { | |
6781 | const struct processor *cpu; | |
6782 | char *str = (char *) alloca (strlen (aarch64_tune_string) + 1); | |
6783 | strcpy (str, aarch64_tune_string); | |
6784 | ||
6785 | /* Loop through the list of supported CPUs to find a match. */ | |
6786 | for (cpu = all_cores; cpu->name != NULL; cpu++) | |
6787 | { | |
6788 | if (strcmp (cpu->name, str) == 0) | |
6789 | { | |
6790 | selected_tune = cpu; | |
6791 | return; | |
6792 | } | |
6793 | } | |
6794 | ||
6795 | /* CPU name not found in list. */ | |
6796 | error ("unknown value %qs for -mtune", str); | |
6797 | return; | |
6798 | } | |
6799 | ||
6800 | ||
6801 | /* Implement TARGET_OPTION_OVERRIDE. */ | |
6802 | ||
6803 | static void | |
6804 | aarch64_override_options (void) | |
6805 | { | |
ffee7aa9 JG |
6806 | /* -mcpu=CPU is shorthand for -march=ARCH_FOR_CPU, -mtune=CPU. |
6807 | If either of -march or -mtune is given, they override their | |
6808 | respective component of -mcpu. | |
43e9d192 | 6809 | |
ffee7aa9 JG |
6810 | So, first parse AARCH64_CPU_STRING, then the others, be careful |
6811 | with -march as, if -mcpu is not present on the command line, march | |
6812 | must set a sensible default CPU. */ | |
6813 | if (aarch64_cpu_string) | |
43e9d192 | 6814 | { |
ffee7aa9 | 6815 | aarch64_parse_cpu (); |
43e9d192 IB |
6816 | } |
6817 | ||
ffee7aa9 | 6818 | if (aarch64_arch_string) |
43e9d192 | 6819 | { |
ffee7aa9 | 6820 | aarch64_parse_arch (); |
43e9d192 IB |
6821 | } |
6822 | ||
6823 | if (aarch64_tune_string) | |
6824 | { | |
6825 | aarch64_parse_tune (); | |
6826 | } | |
6827 | ||
63892fa2 KV |
6828 | #ifndef HAVE_AS_MABI_OPTION |
6829 | /* The compiler may have been configured with 2.23.* binutils, which does | |
6830 | not have support for ILP32. */ | |
6831 | if (TARGET_ILP32) | |
6832 | error ("Assembler does not support -mabi=ilp32"); | |
6833 | #endif | |
6834 | ||
43e9d192 IB |
6835 | initialize_aarch64_code_model (); |
6836 | ||
6837 | aarch64_build_bitmask_table (); | |
6838 | ||
6839 | /* This target defaults to strict volatile bitfields. */ | |
6840 | if (flag_strict_volatile_bitfields < 0 && abi_version_at_least (2)) | |
6841 | flag_strict_volatile_bitfields = 1; | |
6842 | ||
6843 | /* If the user did not specify a processor, choose the default | |
6844 | one for them. This will be the CPU set during configuration using | |
a3cd0246 | 6845 | --with-cpu, otherwise it is "generic". */ |
43e9d192 IB |
6846 | if (!selected_cpu) |
6847 | { | |
6848 | selected_cpu = &all_cores[TARGET_CPU_DEFAULT & 0x3f]; | |
6849 | aarch64_isa_flags = TARGET_CPU_DEFAULT >> 6; | |
6850 | } | |
6851 | ||
6852 | gcc_assert (selected_cpu); | |
6853 | ||
43e9d192 | 6854 | if (!selected_tune) |
3edaf26d | 6855 | selected_tune = selected_cpu; |
43e9d192 IB |
6856 | |
6857 | aarch64_tune_flags = selected_tune->flags; | |
6858 | aarch64_tune = selected_tune->core; | |
6859 | aarch64_tune_params = selected_tune->tune; | |
0c6caaf8 | 6860 | aarch64_architecture_version = selected_cpu->architecture_version; |
43e9d192 | 6861 | |
5e396da6 KT |
6862 | if (aarch64_fix_a53_err835769 == 2) |
6863 | { | |
6864 | #ifdef TARGET_FIX_ERR_A53_835769_DEFAULT | |
6865 | aarch64_fix_a53_err835769 = 1; | |
6866 | #else | |
6867 | aarch64_fix_a53_err835769 = 0; | |
6868 | #endif | |
6869 | } | |
6870 | ||
b4917c98 AP |
6871 | /* If not opzimizing for size, set the default |
6872 | alignment to what the target wants */ | |
6873 | if (!optimize_size) | |
6874 | { | |
6875 | if (align_loops <= 0) | |
0b82a5a2 | 6876 | align_loops = aarch64_tune_params->loop_align; |
b4917c98 | 6877 | if (align_jumps <= 0) |
0b82a5a2 | 6878 | align_jumps = aarch64_tune_params->jump_align; |
b4917c98 | 6879 | if (align_functions <= 0) |
0b82a5a2 | 6880 | align_functions = aarch64_tune_params->function_align; |
b4917c98 AP |
6881 | } |
6882 | ||
43e9d192 IB |
6883 | aarch64_override_options_after_change (); |
6884 | } | |
6885 | ||
6886 | /* Implement targetm.override_options_after_change. */ | |
6887 | ||
6888 | static void | |
6889 | aarch64_override_options_after_change (void) | |
6890 | { | |
0b7f8166 MS |
6891 | if (flag_omit_frame_pointer) |
6892 | flag_omit_leaf_frame_pointer = false; | |
6893 | else if (flag_omit_leaf_frame_pointer) | |
6894 | flag_omit_frame_pointer = true; | |
43e9d192 IB |
6895 | } |
6896 | ||
6897 | static struct machine_function * | |
6898 | aarch64_init_machine_status (void) | |
6899 | { | |
6900 | struct machine_function *machine; | |
766090c2 | 6901 | machine = ggc_cleared_alloc<machine_function> (); |
43e9d192 IB |
6902 | return machine; |
6903 | } | |
6904 | ||
6905 | void | |
6906 | aarch64_init_expanders (void) | |
6907 | { | |
6908 | init_machine_status = aarch64_init_machine_status; | |
6909 | } | |
6910 | ||
6911 | /* A checking mechanism for the implementation of the various code models. */ | |
6912 | static void | |
6913 | initialize_aarch64_code_model (void) | |
6914 | { | |
6915 | if (flag_pic) | |
6916 | { | |
6917 | switch (aarch64_cmodel_var) | |
6918 | { | |
6919 | case AARCH64_CMODEL_TINY: | |
6920 | aarch64_cmodel = AARCH64_CMODEL_TINY_PIC; | |
6921 | break; | |
6922 | case AARCH64_CMODEL_SMALL: | |
6923 | aarch64_cmodel = AARCH64_CMODEL_SMALL_PIC; | |
6924 | break; | |
6925 | case AARCH64_CMODEL_LARGE: | |
6926 | sorry ("code model %qs with -f%s", "large", | |
6927 | flag_pic > 1 ? "PIC" : "pic"); | |
6928 | default: | |
6929 | gcc_unreachable (); | |
6930 | } | |
6931 | } | |
6932 | else | |
6933 | aarch64_cmodel = aarch64_cmodel_var; | |
6934 | } | |
6935 | ||
6936 | /* Return true if SYMBOL_REF X binds locally. */ | |
6937 | ||
6938 | static bool | |
6939 | aarch64_symbol_binds_local_p (const_rtx x) | |
6940 | { | |
6941 | return (SYMBOL_REF_DECL (x) | |
6942 | ? targetm.binds_local_p (SYMBOL_REF_DECL (x)) | |
6943 | : SYMBOL_REF_LOCAL_P (x)); | |
6944 | } | |
6945 | ||
6946 | /* Return true if SYMBOL_REF X is thread local */ | |
6947 | static bool | |
6948 | aarch64_tls_symbol_p (rtx x) | |
6949 | { | |
6950 | if (! TARGET_HAVE_TLS) | |
6951 | return false; | |
6952 | ||
6953 | if (GET_CODE (x) != SYMBOL_REF) | |
6954 | return false; | |
6955 | ||
6956 | return SYMBOL_REF_TLS_MODEL (x) != 0; | |
6957 | } | |
6958 | ||
6959 | /* Classify a TLS symbol into one of the TLS kinds. */ | |
6960 | enum aarch64_symbol_type | |
6961 | aarch64_classify_tls_symbol (rtx x) | |
6962 | { | |
6963 | enum tls_model tls_kind = tls_symbolic_operand_type (x); | |
6964 | ||
6965 | switch (tls_kind) | |
6966 | { | |
6967 | case TLS_MODEL_GLOBAL_DYNAMIC: | |
6968 | case TLS_MODEL_LOCAL_DYNAMIC: | |
6969 | return TARGET_TLS_DESC ? SYMBOL_SMALL_TLSDESC : SYMBOL_SMALL_TLSGD; | |
6970 | ||
6971 | case TLS_MODEL_INITIAL_EXEC: | |
6972 | return SYMBOL_SMALL_GOTTPREL; | |
6973 | ||
6974 | case TLS_MODEL_LOCAL_EXEC: | |
6975 | return SYMBOL_SMALL_TPREL; | |
6976 | ||
6977 | case TLS_MODEL_EMULATED: | |
6978 | case TLS_MODEL_NONE: | |
6979 | return SYMBOL_FORCE_TO_MEM; | |
6980 | ||
6981 | default: | |
6982 | gcc_unreachable (); | |
6983 | } | |
6984 | } | |
6985 | ||
6986 | /* Return the method that should be used to access SYMBOL_REF or | |
6987 | LABEL_REF X in context CONTEXT. */ | |
17f4d4bf | 6988 | |
43e9d192 | 6989 | enum aarch64_symbol_type |
f8b756b7 | 6990 | aarch64_classify_symbol (rtx x, rtx offset, |
43e9d192 IB |
6991 | enum aarch64_symbol_context context ATTRIBUTE_UNUSED) |
6992 | { | |
6993 | if (GET_CODE (x) == LABEL_REF) | |
6994 | { | |
6995 | switch (aarch64_cmodel) | |
6996 | { | |
6997 | case AARCH64_CMODEL_LARGE: | |
6998 | return SYMBOL_FORCE_TO_MEM; | |
6999 | ||
7000 | case AARCH64_CMODEL_TINY_PIC: | |
7001 | case AARCH64_CMODEL_TINY: | |
a5350ddc CSS |
7002 | return SYMBOL_TINY_ABSOLUTE; |
7003 | ||
43e9d192 IB |
7004 | case AARCH64_CMODEL_SMALL_PIC: |
7005 | case AARCH64_CMODEL_SMALL: | |
7006 | return SYMBOL_SMALL_ABSOLUTE; | |
7007 | ||
7008 | default: | |
7009 | gcc_unreachable (); | |
7010 | } | |
7011 | } | |
7012 | ||
17f4d4bf | 7013 | if (GET_CODE (x) == SYMBOL_REF) |
43e9d192 | 7014 | { |
4a985a37 MS |
7015 | if (aarch64_cmodel == AARCH64_CMODEL_LARGE) |
7016 | return SYMBOL_FORCE_TO_MEM; | |
43e9d192 IB |
7017 | |
7018 | if (aarch64_tls_symbol_p (x)) | |
7019 | return aarch64_classify_tls_symbol (x); | |
7020 | ||
17f4d4bf CSS |
7021 | switch (aarch64_cmodel) |
7022 | { | |
7023 | case AARCH64_CMODEL_TINY: | |
f8b756b7 TB |
7024 | /* When we retreive symbol + offset address, we have to make sure |
7025 | the offset does not cause overflow of the final address. But | |
7026 | we have no way of knowing the address of symbol at compile time | |
7027 | so we can't accurately say if the distance between the PC and | |
7028 | symbol + offset is outside the addressible range of +/-1M in the | |
7029 | TINY code model. So we rely on images not being greater than | |
7030 | 1M and cap the offset at 1M and anything beyond 1M will have to | |
7031 | be loaded using an alternative mechanism. */ | |
7032 | if (SYMBOL_REF_WEAK (x) | |
7033 | || INTVAL (offset) < -1048575 || INTVAL (offset) > 1048575) | |
a5350ddc CSS |
7034 | return SYMBOL_FORCE_TO_MEM; |
7035 | return SYMBOL_TINY_ABSOLUTE; | |
7036 | ||
17f4d4bf | 7037 | case AARCH64_CMODEL_SMALL: |
f8b756b7 TB |
7038 | /* Same reasoning as the tiny code model, but the offset cap here is |
7039 | 4G. */ | |
7040 | if (SYMBOL_REF_WEAK (x) | |
3ff5d1f0 TB |
7041 | || !IN_RANGE (INTVAL (offset), HOST_WIDE_INT_C (-4294967263), |
7042 | HOST_WIDE_INT_C (4294967264))) | |
17f4d4bf CSS |
7043 | return SYMBOL_FORCE_TO_MEM; |
7044 | return SYMBOL_SMALL_ABSOLUTE; | |
43e9d192 | 7045 | |
17f4d4bf | 7046 | case AARCH64_CMODEL_TINY_PIC: |
38e6c9a6 | 7047 | if (!aarch64_symbol_binds_local_p (x)) |
87dd8ab0 | 7048 | return SYMBOL_TINY_GOT; |
38e6c9a6 MS |
7049 | return SYMBOL_TINY_ABSOLUTE; |
7050 | ||
17f4d4bf CSS |
7051 | case AARCH64_CMODEL_SMALL_PIC: |
7052 | if (!aarch64_symbol_binds_local_p (x)) | |
7053 | return SYMBOL_SMALL_GOT; | |
7054 | return SYMBOL_SMALL_ABSOLUTE; | |
43e9d192 | 7055 | |
17f4d4bf CSS |
7056 | default: |
7057 | gcc_unreachable (); | |
7058 | } | |
43e9d192 | 7059 | } |
17f4d4bf | 7060 | |
43e9d192 IB |
7061 | /* By default push everything into the constant pool. */ |
7062 | return SYMBOL_FORCE_TO_MEM; | |
7063 | } | |
7064 | ||
43e9d192 IB |
7065 | bool |
7066 | aarch64_constant_address_p (rtx x) | |
7067 | { | |
7068 | return (CONSTANT_P (x) && memory_address_p (DImode, x)); | |
7069 | } | |
7070 | ||
7071 | bool | |
7072 | aarch64_legitimate_pic_operand_p (rtx x) | |
7073 | { | |
7074 | if (GET_CODE (x) == SYMBOL_REF | |
7075 | || (GET_CODE (x) == CONST | |
7076 | && GET_CODE (XEXP (x, 0)) == PLUS | |
7077 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF)) | |
7078 | return false; | |
7079 | ||
7080 | return true; | |
7081 | } | |
7082 | ||
3520f7cc JG |
7083 | /* Return true if X holds either a quarter-precision or |
7084 | floating-point +0.0 constant. */ | |
7085 | static bool | |
ef4bddc2 | 7086 | aarch64_valid_floating_const (machine_mode mode, rtx x) |
3520f7cc JG |
7087 | { |
7088 | if (!CONST_DOUBLE_P (x)) | |
7089 | return false; | |
7090 | ||
7091 | /* TODO: We could handle moving 0.0 to a TFmode register, | |
7092 | but first we would like to refactor the movtf_aarch64 | |
7093 | to be more amicable to split moves properly and | |
7094 | correctly gate on TARGET_SIMD. For now - reject all | |
7095 | constants which are not to SFmode or DFmode registers. */ | |
7096 | if (!(mode == SFmode || mode == DFmode)) | |
7097 | return false; | |
7098 | ||
7099 | if (aarch64_float_const_zero_rtx_p (x)) | |
7100 | return true; | |
7101 | return aarch64_float_const_representable_p (x); | |
7102 | } | |
7103 | ||
43e9d192 | 7104 | static bool |
ef4bddc2 | 7105 | aarch64_legitimate_constant_p (machine_mode mode, rtx x) |
43e9d192 IB |
7106 | { |
7107 | /* Do not allow vector struct mode constants. We could support | |
7108 | 0 and -1 easily, but they need support in aarch64-simd.md. */ | |
7109 | if (TARGET_SIMD && aarch64_vect_struct_mode_p (mode)) | |
7110 | return false; | |
7111 | ||
7112 | /* This could probably go away because | |
7113 | we now decompose CONST_INTs according to expand_mov_immediate. */ | |
7114 | if ((GET_CODE (x) == CONST_VECTOR | |
48063b9d | 7115 | && aarch64_simd_valid_immediate (x, mode, false, NULL)) |
3520f7cc JG |
7116 | || CONST_INT_P (x) || aarch64_valid_floating_const (mode, x)) |
7117 | return !targetm.cannot_force_const_mem (mode, x); | |
43e9d192 IB |
7118 | |
7119 | if (GET_CODE (x) == HIGH | |
7120 | && aarch64_valid_symref (XEXP (x, 0), GET_MODE (XEXP (x, 0)))) | |
7121 | return true; | |
7122 | ||
7123 | return aarch64_constant_address_p (x); | |
7124 | } | |
7125 | ||
a5bc806c | 7126 | rtx |
43e9d192 IB |
7127 | aarch64_load_tp (rtx target) |
7128 | { | |
7129 | if (!target | |
7130 | || GET_MODE (target) != Pmode | |
7131 | || !register_operand (target, Pmode)) | |
7132 | target = gen_reg_rtx (Pmode); | |
7133 | ||
7134 | /* Can return in any reg. */ | |
7135 | emit_insn (gen_aarch64_load_tp_hard (target)); | |
7136 | return target; | |
7137 | } | |
7138 | ||
43e9d192 IB |
7139 | /* On AAPCS systems, this is the "struct __va_list". */ |
7140 | static GTY(()) tree va_list_type; | |
7141 | ||
7142 | /* Implement TARGET_BUILD_BUILTIN_VA_LIST. | |
7143 | Return the type to use as __builtin_va_list. | |
7144 | ||
7145 | AAPCS64 \S 7.1.4 requires that va_list be a typedef for a type defined as: | |
7146 | ||
7147 | struct __va_list | |
7148 | { | |
7149 | void *__stack; | |
7150 | void *__gr_top; | |
7151 | void *__vr_top; | |
7152 | int __gr_offs; | |
7153 | int __vr_offs; | |
7154 | }; */ | |
7155 | ||
7156 | static tree | |
7157 | aarch64_build_builtin_va_list (void) | |
7158 | { | |
7159 | tree va_list_name; | |
7160 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
7161 | ||
7162 | /* Create the type. */ | |
7163 | va_list_type = lang_hooks.types.make_type (RECORD_TYPE); | |
7164 | /* Give it the required name. */ | |
7165 | va_list_name = build_decl (BUILTINS_LOCATION, | |
7166 | TYPE_DECL, | |
7167 | get_identifier ("__va_list"), | |
7168 | va_list_type); | |
7169 | DECL_ARTIFICIAL (va_list_name) = 1; | |
7170 | TYPE_NAME (va_list_type) = va_list_name; | |
665c56c6 | 7171 | TYPE_STUB_DECL (va_list_type) = va_list_name; |
43e9d192 IB |
7172 | |
7173 | /* Create the fields. */ | |
7174 | f_stack = build_decl (BUILTINS_LOCATION, | |
7175 | FIELD_DECL, get_identifier ("__stack"), | |
7176 | ptr_type_node); | |
7177 | f_grtop = build_decl (BUILTINS_LOCATION, | |
7178 | FIELD_DECL, get_identifier ("__gr_top"), | |
7179 | ptr_type_node); | |
7180 | f_vrtop = build_decl (BUILTINS_LOCATION, | |
7181 | FIELD_DECL, get_identifier ("__vr_top"), | |
7182 | ptr_type_node); | |
7183 | f_groff = build_decl (BUILTINS_LOCATION, | |
7184 | FIELD_DECL, get_identifier ("__gr_offs"), | |
7185 | integer_type_node); | |
7186 | f_vroff = build_decl (BUILTINS_LOCATION, | |
7187 | FIELD_DECL, get_identifier ("__vr_offs"), | |
7188 | integer_type_node); | |
7189 | ||
7190 | DECL_ARTIFICIAL (f_stack) = 1; | |
7191 | DECL_ARTIFICIAL (f_grtop) = 1; | |
7192 | DECL_ARTIFICIAL (f_vrtop) = 1; | |
7193 | DECL_ARTIFICIAL (f_groff) = 1; | |
7194 | DECL_ARTIFICIAL (f_vroff) = 1; | |
7195 | ||
7196 | DECL_FIELD_CONTEXT (f_stack) = va_list_type; | |
7197 | DECL_FIELD_CONTEXT (f_grtop) = va_list_type; | |
7198 | DECL_FIELD_CONTEXT (f_vrtop) = va_list_type; | |
7199 | DECL_FIELD_CONTEXT (f_groff) = va_list_type; | |
7200 | DECL_FIELD_CONTEXT (f_vroff) = va_list_type; | |
7201 | ||
7202 | TYPE_FIELDS (va_list_type) = f_stack; | |
7203 | DECL_CHAIN (f_stack) = f_grtop; | |
7204 | DECL_CHAIN (f_grtop) = f_vrtop; | |
7205 | DECL_CHAIN (f_vrtop) = f_groff; | |
7206 | DECL_CHAIN (f_groff) = f_vroff; | |
7207 | ||
7208 | /* Compute its layout. */ | |
7209 | layout_type (va_list_type); | |
7210 | ||
7211 | return va_list_type; | |
7212 | } | |
7213 | ||
7214 | /* Implement TARGET_EXPAND_BUILTIN_VA_START. */ | |
7215 | static void | |
7216 | aarch64_expand_builtin_va_start (tree valist, rtx nextarg ATTRIBUTE_UNUSED) | |
7217 | { | |
7218 | const CUMULATIVE_ARGS *cum; | |
7219 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
7220 | tree stack, grtop, vrtop, groff, vroff; | |
7221 | tree t; | |
7222 | int gr_save_area_size; | |
7223 | int vr_save_area_size; | |
7224 | int vr_offset; | |
7225 | ||
7226 | cum = &crtl->args.info; | |
7227 | gr_save_area_size | |
7228 | = (NUM_ARG_REGS - cum->aapcs_ncrn) * UNITS_PER_WORD; | |
7229 | vr_save_area_size | |
7230 | = (NUM_FP_ARG_REGS - cum->aapcs_nvrn) * UNITS_PER_VREG; | |
7231 | ||
7232 | if (TARGET_GENERAL_REGS_ONLY) | |
7233 | { | |
7234 | if (cum->aapcs_nvrn > 0) | |
7235 | sorry ("%qs and floating point or vector arguments", | |
7236 | "-mgeneral-regs-only"); | |
7237 | vr_save_area_size = 0; | |
7238 | } | |
7239 | ||
7240 | f_stack = TYPE_FIELDS (va_list_type_node); | |
7241 | f_grtop = DECL_CHAIN (f_stack); | |
7242 | f_vrtop = DECL_CHAIN (f_grtop); | |
7243 | f_groff = DECL_CHAIN (f_vrtop); | |
7244 | f_vroff = DECL_CHAIN (f_groff); | |
7245 | ||
7246 | stack = build3 (COMPONENT_REF, TREE_TYPE (f_stack), valist, f_stack, | |
7247 | NULL_TREE); | |
7248 | grtop = build3 (COMPONENT_REF, TREE_TYPE (f_grtop), valist, f_grtop, | |
7249 | NULL_TREE); | |
7250 | vrtop = build3 (COMPONENT_REF, TREE_TYPE (f_vrtop), valist, f_vrtop, | |
7251 | NULL_TREE); | |
7252 | groff = build3 (COMPONENT_REF, TREE_TYPE (f_groff), valist, f_groff, | |
7253 | NULL_TREE); | |
7254 | vroff = build3 (COMPONENT_REF, TREE_TYPE (f_vroff), valist, f_vroff, | |
7255 | NULL_TREE); | |
7256 | ||
7257 | /* Emit code to initialize STACK, which points to the next varargs stack | |
7258 | argument. CUM->AAPCS_STACK_SIZE gives the number of stack words used | |
7259 | by named arguments. STACK is 8-byte aligned. */ | |
7260 | t = make_tree (TREE_TYPE (stack), virtual_incoming_args_rtx); | |
7261 | if (cum->aapcs_stack_size > 0) | |
7262 | t = fold_build_pointer_plus_hwi (t, cum->aapcs_stack_size * UNITS_PER_WORD); | |
7263 | t = build2 (MODIFY_EXPR, TREE_TYPE (stack), stack, t); | |
7264 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7265 | ||
7266 | /* Emit code to initialize GRTOP, the top of the GR save area. | |
7267 | virtual_incoming_args_rtx should have been 16 byte aligned. */ | |
7268 | t = make_tree (TREE_TYPE (grtop), virtual_incoming_args_rtx); | |
7269 | t = build2 (MODIFY_EXPR, TREE_TYPE (grtop), grtop, t); | |
7270 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7271 | ||
7272 | /* Emit code to initialize VRTOP, the top of the VR save area. | |
7273 | This address is gr_save_area_bytes below GRTOP, rounded | |
7274 | down to the next 16-byte boundary. */ | |
7275 | t = make_tree (TREE_TYPE (vrtop), virtual_incoming_args_rtx); | |
7276 | vr_offset = AARCH64_ROUND_UP (gr_save_area_size, | |
7277 | STACK_BOUNDARY / BITS_PER_UNIT); | |
7278 | ||
7279 | if (vr_offset) | |
7280 | t = fold_build_pointer_plus_hwi (t, -vr_offset); | |
7281 | t = build2 (MODIFY_EXPR, TREE_TYPE (vrtop), vrtop, t); | |
7282 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7283 | ||
7284 | /* Emit code to initialize GROFF, the offset from GRTOP of the | |
7285 | next GPR argument. */ | |
7286 | t = build2 (MODIFY_EXPR, TREE_TYPE (groff), groff, | |
7287 | build_int_cst (TREE_TYPE (groff), -gr_save_area_size)); | |
7288 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7289 | ||
7290 | /* Likewise emit code to initialize VROFF, the offset from FTOP | |
7291 | of the next VR argument. */ | |
7292 | t = build2 (MODIFY_EXPR, TREE_TYPE (vroff), vroff, | |
7293 | build_int_cst (TREE_TYPE (vroff), -vr_save_area_size)); | |
7294 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
7295 | } | |
7296 | ||
7297 | /* Implement TARGET_GIMPLIFY_VA_ARG_EXPR. */ | |
7298 | ||
7299 | static tree | |
7300 | aarch64_gimplify_va_arg_expr (tree valist, tree type, gimple_seq *pre_p, | |
7301 | gimple_seq *post_p ATTRIBUTE_UNUSED) | |
7302 | { | |
7303 | tree addr; | |
7304 | bool indirect_p; | |
7305 | bool is_ha; /* is HFA or HVA. */ | |
7306 | bool dw_align; /* double-word align. */ | |
ef4bddc2 | 7307 | machine_mode ag_mode = VOIDmode; |
43e9d192 | 7308 | int nregs; |
ef4bddc2 | 7309 | machine_mode mode; |
43e9d192 IB |
7310 | |
7311 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
7312 | tree stack, f_top, f_off, off, arg, roundup, on_stack; | |
7313 | HOST_WIDE_INT size, rsize, adjust, align; | |
7314 | tree t, u, cond1, cond2; | |
7315 | ||
7316 | indirect_p = pass_by_reference (NULL, TYPE_MODE (type), type, false); | |
7317 | if (indirect_p) | |
7318 | type = build_pointer_type (type); | |
7319 | ||
7320 | mode = TYPE_MODE (type); | |
7321 | ||
7322 | f_stack = TYPE_FIELDS (va_list_type_node); | |
7323 | f_grtop = DECL_CHAIN (f_stack); | |
7324 | f_vrtop = DECL_CHAIN (f_grtop); | |
7325 | f_groff = DECL_CHAIN (f_vrtop); | |
7326 | f_vroff = DECL_CHAIN (f_groff); | |
7327 | ||
7328 | stack = build3 (COMPONENT_REF, TREE_TYPE (f_stack), unshare_expr (valist), | |
7329 | f_stack, NULL_TREE); | |
7330 | size = int_size_in_bytes (type); | |
7331 | align = aarch64_function_arg_alignment (mode, type) / BITS_PER_UNIT; | |
7332 | ||
7333 | dw_align = false; | |
7334 | adjust = 0; | |
7335 | if (aarch64_vfp_is_call_or_return_candidate (mode, | |
7336 | type, | |
7337 | &ag_mode, | |
7338 | &nregs, | |
7339 | &is_ha)) | |
7340 | { | |
7341 | /* TYPE passed in fp/simd registers. */ | |
7342 | if (TARGET_GENERAL_REGS_ONLY) | |
7343 | sorry ("%qs and floating point or vector arguments", | |
7344 | "-mgeneral-regs-only"); | |
7345 | ||
7346 | f_top = build3 (COMPONENT_REF, TREE_TYPE (f_vrtop), | |
7347 | unshare_expr (valist), f_vrtop, NULL_TREE); | |
7348 | f_off = build3 (COMPONENT_REF, TREE_TYPE (f_vroff), | |
7349 | unshare_expr (valist), f_vroff, NULL_TREE); | |
7350 | ||
7351 | rsize = nregs * UNITS_PER_VREG; | |
7352 | ||
7353 | if (is_ha) | |
7354 | { | |
7355 | if (BYTES_BIG_ENDIAN && GET_MODE_SIZE (ag_mode) < UNITS_PER_VREG) | |
7356 | adjust = UNITS_PER_VREG - GET_MODE_SIZE (ag_mode); | |
7357 | } | |
7358 | else if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
7359 | && size < UNITS_PER_VREG) | |
7360 | { | |
7361 | adjust = UNITS_PER_VREG - size; | |
7362 | } | |
7363 | } | |
7364 | else | |
7365 | { | |
7366 | /* TYPE passed in general registers. */ | |
7367 | f_top = build3 (COMPONENT_REF, TREE_TYPE (f_grtop), | |
7368 | unshare_expr (valist), f_grtop, NULL_TREE); | |
7369 | f_off = build3 (COMPONENT_REF, TREE_TYPE (f_groff), | |
7370 | unshare_expr (valist), f_groff, NULL_TREE); | |
7371 | rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD; | |
7372 | nregs = rsize / UNITS_PER_WORD; | |
7373 | ||
7374 | if (align > 8) | |
7375 | dw_align = true; | |
7376 | ||
7377 | if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
7378 | && size < UNITS_PER_WORD) | |
7379 | { | |
7380 | adjust = UNITS_PER_WORD - size; | |
7381 | } | |
7382 | } | |
7383 | ||
7384 | /* Get a local temporary for the field value. */ | |
7385 | off = get_initialized_tmp_var (f_off, pre_p, NULL); | |
7386 | ||
7387 | /* Emit code to branch if off >= 0. */ | |
7388 | t = build2 (GE_EXPR, boolean_type_node, off, | |
7389 | build_int_cst (TREE_TYPE (off), 0)); | |
7390 | cond1 = build3 (COND_EXPR, ptr_type_node, t, NULL_TREE, NULL_TREE); | |
7391 | ||
7392 | if (dw_align) | |
7393 | { | |
7394 | /* Emit: offs = (offs + 15) & -16. */ | |
7395 | t = build2 (PLUS_EXPR, TREE_TYPE (off), off, | |
7396 | build_int_cst (TREE_TYPE (off), 15)); | |
7397 | t = build2 (BIT_AND_EXPR, TREE_TYPE (off), t, | |
7398 | build_int_cst (TREE_TYPE (off), -16)); | |
7399 | roundup = build2 (MODIFY_EXPR, TREE_TYPE (off), off, t); | |
7400 | } | |
7401 | else | |
7402 | roundup = NULL; | |
7403 | ||
7404 | /* Update ap.__[g|v]r_offs */ | |
7405 | t = build2 (PLUS_EXPR, TREE_TYPE (off), off, | |
7406 | build_int_cst (TREE_TYPE (off), rsize)); | |
7407 | t = build2 (MODIFY_EXPR, TREE_TYPE (f_off), unshare_expr (f_off), t); | |
7408 | ||
7409 | /* String up. */ | |
7410 | if (roundup) | |
7411 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), roundup, t); | |
7412 | ||
7413 | /* [cond2] if (ap.__[g|v]r_offs > 0) */ | |
7414 | u = build2 (GT_EXPR, boolean_type_node, unshare_expr (f_off), | |
7415 | build_int_cst (TREE_TYPE (f_off), 0)); | |
7416 | cond2 = build3 (COND_EXPR, ptr_type_node, u, NULL_TREE, NULL_TREE); | |
7417 | ||
7418 | /* String up: make sure the assignment happens before the use. */ | |
7419 | t = build2 (COMPOUND_EXPR, TREE_TYPE (cond2), t, cond2); | |
7420 | COND_EXPR_ELSE (cond1) = t; | |
7421 | ||
7422 | /* Prepare the trees handling the argument that is passed on the stack; | |
7423 | the top level node will store in ON_STACK. */ | |
7424 | arg = get_initialized_tmp_var (stack, pre_p, NULL); | |
7425 | if (align > 8) | |
7426 | { | |
7427 | /* if (alignof(type) > 8) (arg = arg + 15) & -16; */ | |
7428 | t = fold_convert (intDI_type_node, arg); | |
7429 | t = build2 (PLUS_EXPR, TREE_TYPE (t), t, | |
7430 | build_int_cst (TREE_TYPE (t), 15)); | |
7431 | t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t, | |
7432 | build_int_cst (TREE_TYPE (t), -16)); | |
7433 | t = fold_convert (TREE_TYPE (arg), t); | |
7434 | roundup = build2 (MODIFY_EXPR, TREE_TYPE (arg), arg, t); | |
7435 | } | |
7436 | else | |
7437 | roundup = NULL; | |
7438 | /* Advance ap.__stack */ | |
7439 | t = fold_convert (intDI_type_node, arg); | |
7440 | t = build2 (PLUS_EXPR, TREE_TYPE (t), t, | |
7441 | build_int_cst (TREE_TYPE (t), size + 7)); | |
7442 | t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t, | |
7443 | build_int_cst (TREE_TYPE (t), -8)); | |
7444 | t = fold_convert (TREE_TYPE (arg), t); | |
7445 | t = build2 (MODIFY_EXPR, TREE_TYPE (stack), unshare_expr (stack), t); | |
7446 | /* String up roundup and advance. */ | |
7447 | if (roundup) | |
7448 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), roundup, t); | |
7449 | /* String up with arg */ | |
7450 | on_stack = build2 (COMPOUND_EXPR, TREE_TYPE (arg), t, arg); | |
7451 | /* Big-endianness related address adjustment. */ | |
7452 | if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
7453 | && size < UNITS_PER_WORD) | |
7454 | { | |
7455 | t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (arg), arg, | |
7456 | size_int (UNITS_PER_WORD - size)); | |
7457 | on_stack = build2 (COMPOUND_EXPR, TREE_TYPE (arg), on_stack, t); | |
7458 | } | |
7459 | ||
7460 | COND_EXPR_THEN (cond1) = unshare_expr (on_stack); | |
7461 | COND_EXPR_THEN (cond2) = unshare_expr (on_stack); | |
7462 | ||
7463 | /* Adjustment to OFFSET in the case of BIG_ENDIAN. */ | |
7464 | t = off; | |
7465 | if (adjust) | |
7466 | t = build2 (PREINCREMENT_EXPR, TREE_TYPE (off), off, | |
7467 | build_int_cst (TREE_TYPE (off), adjust)); | |
7468 | ||
7469 | t = fold_convert (sizetype, t); | |
7470 | t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (f_top), f_top, t); | |
7471 | ||
7472 | if (is_ha) | |
7473 | { | |
7474 | /* type ha; // treat as "struct {ftype field[n];}" | |
7475 | ... [computing offs] | |
7476 | for (i = 0; i <nregs; ++i, offs += 16) | |
7477 | ha.field[i] = *((ftype *)(ap.__vr_top + offs)); | |
7478 | return ha; */ | |
7479 | int i; | |
7480 | tree tmp_ha, field_t, field_ptr_t; | |
7481 | ||
7482 | /* Declare a local variable. */ | |
7483 | tmp_ha = create_tmp_var_raw (type, "ha"); | |
7484 | gimple_add_tmp_var (tmp_ha); | |
7485 | ||
7486 | /* Establish the base type. */ | |
7487 | switch (ag_mode) | |
7488 | { | |
7489 | case SFmode: | |
7490 | field_t = float_type_node; | |
7491 | field_ptr_t = float_ptr_type_node; | |
7492 | break; | |
7493 | case DFmode: | |
7494 | field_t = double_type_node; | |
7495 | field_ptr_t = double_ptr_type_node; | |
7496 | break; | |
7497 | case TFmode: | |
7498 | field_t = long_double_type_node; | |
7499 | field_ptr_t = long_double_ptr_type_node; | |
7500 | break; | |
7501 | /* The half precision and quad precision are not fully supported yet. Enable | |
7502 | the following code after the support is complete. Need to find the correct | |
7503 | type node for __fp16 *. */ | |
7504 | #if 0 | |
7505 | case HFmode: | |
7506 | field_t = float_type_node; | |
7507 | field_ptr_t = float_ptr_type_node; | |
7508 | break; | |
7509 | #endif | |
7510 | case V2SImode: | |
7511 | case V4SImode: | |
7512 | { | |
7513 | tree innertype = make_signed_type (GET_MODE_PRECISION (SImode)); | |
7514 | field_t = build_vector_type_for_mode (innertype, ag_mode); | |
7515 | field_ptr_t = build_pointer_type (field_t); | |
7516 | } | |
7517 | break; | |
7518 | default: | |
7519 | gcc_assert (0); | |
7520 | } | |
7521 | ||
7522 | /* *(field_ptr_t)&ha = *((field_ptr_t)vr_saved_area */ | |
7523 | tmp_ha = build1 (ADDR_EXPR, field_ptr_t, tmp_ha); | |
7524 | addr = t; | |
7525 | t = fold_convert (field_ptr_t, addr); | |
7526 | t = build2 (MODIFY_EXPR, field_t, | |
7527 | build1 (INDIRECT_REF, field_t, tmp_ha), | |
7528 | build1 (INDIRECT_REF, field_t, t)); | |
7529 | ||
7530 | /* ha.field[i] = *((field_ptr_t)vr_saved_area + i) */ | |
7531 | for (i = 1; i < nregs; ++i) | |
7532 | { | |
7533 | addr = fold_build_pointer_plus_hwi (addr, UNITS_PER_VREG); | |
7534 | u = fold_convert (field_ptr_t, addr); | |
7535 | u = build2 (MODIFY_EXPR, field_t, | |
7536 | build2 (MEM_REF, field_t, tmp_ha, | |
7537 | build_int_cst (field_ptr_t, | |
7538 | (i * | |
7539 | int_size_in_bytes (field_t)))), | |
7540 | build1 (INDIRECT_REF, field_t, u)); | |
7541 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), t, u); | |
7542 | } | |
7543 | ||
7544 | u = fold_convert (TREE_TYPE (f_top), tmp_ha); | |
7545 | t = build2 (COMPOUND_EXPR, TREE_TYPE (f_top), t, u); | |
7546 | } | |
7547 | ||
7548 | COND_EXPR_ELSE (cond2) = t; | |
7549 | addr = fold_convert (build_pointer_type (type), cond1); | |
7550 | addr = build_va_arg_indirect_ref (addr); | |
7551 | ||
7552 | if (indirect_p) | |
7553 | addr = build_va_arg_indirect_ref (addr); | |
7554 | ||
7555 | return addr; | |
7556 | } | |
7557 | ||
7558 | /* Implement TARGET_SETUP_INCOMING_VARARGS. */ | |
7559 | ||
7560 | static void | |
ef4bddc2 | 7561 | aarch64_setup_incoming_varargs (cumulative_args_t cum_v, machine_mode mode, |
43e9d192 IB |
7562 | tree type, int *pretend_size ATTRIBUTE_UNUSED, |
7563 | int no_rtl) | |
7564 | { | |
7565 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
7566 | CUMULATIVE_ARGS local_cum; | |
7567 | int gr_saved, vr_saved; | |
7568 | ||
7569 | /* The caller has advanced CUM up to, but not beyond, the last named | |
7570 | argument. Advance a local copy of CUM past the last "real" named | |
7571 | argument, to find out how many registers are left over. */ | |
7572 | local_cum = *cum; | |
7573 | aarch64_function_arg_advance (pack_cumulative_args(&local_cum), mode, type, true); | |
7574 | ||
7575 | /* Found out how many registers we need to save. */ | |
7576 | gr_saved = NUM_ARG_REGS - local_cum.aapcs_ncrn; | |
7577 | vr_saved = NUM_FP_ARG_REGS - local_cum.aapcs_nvrn; | |
7578 | ||
7579 | if (TARGET_GENERAL_REGS_ONLY) | |
7580 | { | |
7581 | if (local_cum.aapcs_nvrn > 0) | |
7582 | sorry ("%qs and floating point or vector arguments", | |
7583 | "-mgeneral-regs-only"); | |
7584 | vr_saved = 0; | |
7585 | } | |
7586 | ||
7587 | if (!no_rtl) | |
7588 | { | |
7589 | if (gr_saved > 0) | |
7590 | { | |
7591 | rtx ptr, mem; | |
7592 | ||
7593 | /* virtual_incoming_args_rtx should have been 16-byte aligned. */ | |
7594 | ptr = plus_constant (Pmode, virtual_incoming_args_rtx, | |
7595 | - gr_saved * UNITS_PER_WORD); | |
7596 | mem = gen_frame_mem (BLKmode, ptr); | |
7597 | set_mem_alias_set (mem, get_varargs_alias_set ()); | |
7598 | ||
7599 | move_block_from_reg (local_cum.aapcs_ncrn + R0_REGNUM, | |
7600 | mem, gr_saved); | |
7601 | } | |
7602 | if (vr_saved > 0) | |
7603 | { | |
7604 | /* We can't use move_block_from_reg, because it will use | |
7605 | the wrong mode, storing D regs only. */ | |
ef4bddc2 | 7606 | machine_mode mode = TImode; |
43e9d192 IB |
7607 | int off, i; |
7608 | ||
7609 | /* Set OFF to the offset from virtual_incoming_args_rtx of | |
7610 | the first vector register. The VR save area lies below | |
7611 | the GR one, and is aligned to 16 bytes. */ | |
7612 | off = -AARCH64_ROUND_UP (gr_saved * UNITS_PER_WORD, | |
7613 | STACK_BOUNDARY / BITS_PER_UNIT); | |
7614 | off -= vr_saved * UNITS_PER_VREG; | |
7615 | ||
7616 | for (i = local_cum.aapcs_nvrn; i < NUM_FP_ARG_REGS; ++i) | |
7617 | { | |
7618 | rtx ptr, mem; | |
7619 | ||
7620 | ptr = plus_constant (Pmode, virtual_incoming_args_rtx, off); | |
7621 | mem = gen_frame_mem (mode, ptr); | |
7622 | set_mem_alias_set (mem, get_varargs_alias_set ()); | |
7623 | aarch64_emit_move (mem, gen_rtx_REG (mode, V0_REGNUM + i)); | |
7624 | off += UNITS_PER_VREG; | |
7625 | } | |
7626 | } | |
7627 | } | |
7628 | ||
7629 | /* We don't save the size into *PRETEND_SIZE because we want to avoid | |
7630 | any complication of having crtl->args.pretend_args_size changed. */ | |
8799637a | 7631 | cfun->machine->frame.saved_varargs_size |
43e9d192 IB |
7632 | = (AARCH64_ROUND_UP (gr_saved * UNITS_PER_WORD, |
7633 | STACK_BOUNDARY / BITS_PER_UNIT) | |
7634 | + vr_saved * UNITS_PER_VREG); | |
7635 | } | |
7636 | ||
7637 | static void | |
7638 | aarch64_conditional_register_usage (void) | |
7639 | { | |
7640 | int i; | |
7641 | if (!TARGET_FLOAT) | |
7642 | { | |
7643 | for (i = V0_REGNUM; i <= V31_REGNUM; i++) | |
7644 | { | |
7645 | fixed_regs[i] = 1; | |
7646 | call_used_regs[i] = 1; | |
7647 | } | |
7648 | } | |
7649 | } | |
7650 | ||
7651 | /* Walk down the type tree of TYPE counting consecutive base elements. | |
7652 | If *MODEP is VOIDmode, then set it to the first valid floating point | |
7653 | type. If a non-floating point type is found, or if a floating point | |
7654 | type that doesn't match a non-VOIDmode *MODEP is found, then return -1, | |
7655 | otherwise return the count in the sub-tree. */ | |
7656 | static int | |
ef4bddc2 | 7657 | aapcs_vfp_sub_candidate (const_tree type, machine_mode *modep) |
43e9d192 | 7658 | { |
ef4bddc2 | 7659 | machine_mode mode; |
43e9d192 IB |
7660 | HOST_WIDE_INT size; |
7661 | ||
7662 | switch (TREE_CODE (type)) | |
7663 | { | |
7664 | case REAL_TYPE: | |
7665 | mode = TYPE_MODE (type); | |
7666 | if (mode != DFmode && mode != SFmode && mode != TFmode) | |
7667 | return -1; | |
7668 | ||
7669 | if (*modep == VOIDmode) | |
7670 | *modep = mode; | |
7671 | ||
7672 | if (*modep == mode) | |
7673 | return 1; | |
7674 | ||
7675 | break; | |
7676 | ||
7677 | case COMPLEX_TYPE: | |
7678 | mode = TYPE_MODE (TREE_TYPE (type)); | |
7679 | if (mode != DFmode && mode != SFmode && mode != TFmode) | |
7680 | return -1; | |
7681 | ||
7682 | if (*modep == VOIDmode) | |
7683 | *modep = mode; | |
7684 | ||
7685 | if (*modep == mode) | |
7686 | return 2; | |
7687 | ||
7688 | break; | |
7689 | ||
7690 | case VECTOR_TYPE: | |
7691 | /* Use V2SImode and V4SImode as representatives of all 64-bit | |
7692 | and 128-bit vector types. */ | |
7693 | size = int_size_in_bytes (type); | |
7694 | switch (size) | |
7695 | { | |
7696 | case 8: | |
7697 | mode = V2SImode; | |
7698 | break; | |
7699 | case 16: | |
7700 | mode = V4SImode; | |
7701 | break; | |
7702 | default: | |
7703 | return -1; | |
7704 | } | |
7705 | ||
7706 | if (*modep == VOIDmode) | |
7707 | *modep = mode; | |
7708 | ||
7709 | /* Vector modes are considered to be opaque: two vectors are | |
7710 | equivalent for the purposes of being homogeneous aggregates | |
7711 | if they are the same size. */ | |
7712 | if (*modep == mode) | |
7713 | return 1; | |
7714 | ||
7715 | break; | |
7716 | ||
7717 | case ARRAY_TYPE: | |
7718 | { | |
7719 | int count; | |
7720 | tree index = TYPE_DOMAIN (type); | |
7721 | ||
807e902e KZ |
7722 | /* Can't handle incomplete types nor sizes that are not |
7723 | fixed. */ | |
7724 | if (!COMPLETE_TYPE_P (type) | |
7725 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
43e9d192 IB |
7726 | return -1; |
7727 | ||
7728 | count = aapcs_vfp_sub_candidate (TREE_TYPE (type), modep); | |
7729 | if (count == -1 | |
7730 | || !index | |
7731 | || !TYPE_MAX_VALUE (index) | |
cc269bb6 | 7732 | || !tree_fits_uhwi_p (TYPE_MAX_VALUE (index)) |
43e9d192 | 7733 | || !TYPE_MIN_VALUE (index) |
cc269bb6 | 7734 | || !tree_fits_uhwi_p (TYPE_MIN_VALUE (index)) |
43e9d192 IB |
7735 | || count < 0) |
7736 | return -1; | |
7737 | ||
ae7e9ddd RS |
7738 | count *= (1 + tree_to_uhwi (TYPE_MAX_VALUE (index)) |
7739 | - tree_to_uhwi (TYPE_MIN_VALUE (index))); | |
43e9d192 IB |
7740 | |
7741 | /* There must be no padding. */ | |
807e902e | 7742 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) |
43e9d192 IB |
7743 | return -1; |
7744 | ||
7745 | return count; | |
7746 | } | |
7747 | ||
7748 | case RECORD_TYPE: | |
7749 | { | |
7750 | int count = 0; | |
7751 | int sub_count; | |
7752 | tree field; | |
7753 | ||
807e902e KZ |
7754 | /* Can't handle incomplete types nor sizes that are not |
7755 | fixed. */ | |
7756 | if (!COMPLETE_TYPE_P (type) | |
7757 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
43e9d192 IB |
7758 | return -1; |
7759 | ||
7760 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
7761 | { | |
7762 | if (TREE_CODE (field) != FIELD_DECL) | |
7763 | continue; | |
7764 | ||
7765 | sub_count = aapcs_vfp_sub_candidate (TREE_TYPE (field), modep); | |
7766 | if (sub_count < 0) | |
7767 | return -1; | |
7768 | count += sub_count; | |
7769 | } | |
7770 | ||
7771 | /* There must be no padding. */ | |
807e902e | 7772 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) |
43e9d192 IB |
7773 | return -1; |
7774 | ||
7775 | return count; | |
7776 | } | |
7777 | ||
7778 | case UNION_TYPE: | |
7779 | case QUAL_UNION_TYPE: | |
7780 | { | |
7781 | /* These aren't very interesting except in a degenerate case. */ | |
7782 | int count = 0; | |
7783 | int sub_count; | |
7784 | tree field; | |
7785 | ||
807e902e KZ |
7786 | /* Can't handle incomplete types nor sizes that are not |
7787 | fixed. */ | |
7788 | if (!COMPLETE_TYPE_P (type) | |
7789 | || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST) | |
43e9d192 IB |
7790 | return -1; |
7791 | ||
7792 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
7793 | { | |
7794 | if (TREE_CODE (field) != FIELD_DECL) | |
7795 | continue; | |
7796 | ||
7797 | sub_count = aapcs_vfp_sub_candidate (TREE_TYPE (field), modep); | |
7798 | if (sub_count < 0) | |
7799 | return -1; | |
7800 | count = count > sub_count ? count : sub_count; | |
7801 | } | |
7802 | ||
7803 | /* There must be no padding. */ | |
807e902e | 7804 | if (wi::ne_p (TYPE_SIZE (type), count * GET_MODE_BITSIZE (*modep))) |
43e9d192 IB |
7805 | return -1; |
7806 | ||
7807 | return count; | |
7808 | } | |
7809 | ||
7810 | default: | |
7811 | break; | |
7812 | } | |
7813 | ||
7814 | return -1; | |
7815 | } | |
7816 | ||
7817 | /* Return TRUE if the type, as described by TYPE and MODE, is a composite | |
7818 | type as described in AAPCS64 \S 4.3. This includes aggregate, union and | |
7819 | array types. The C99 floating-point complex types are also considered | |
7820 | as composite types, according to AAPCS64 \S 7.1.1. The complex integer | |
7821 | types, which are GCC extensions and out of the scope of AAPCS64, are | |
7822 | treated as composite types here as well. | |
7823 | ||
7824 | Note that MODE itself is not sufficient in determining whether a type | |
7825 | is such a composite type or not. This is because | |
7826 | stor-layout.c:compute_record_mode may have already changed the MODE | |
7827 | (BLKmode) of a RECORD_TYPE TYPE to some other mode. For example, a | |
7828 | structure with only one field may have its MODE set to the mode of the | |
7829 | field. Also an integer mode whose size matches the size of the | |
7830 | RECORD_TYPE type may be used to substitute the original mode | |
7831 | (i.e. BLKmode) in certain circumstances. In other words, MODE cannot be | |
7832 | solely relied on. */ | |
7833 | ||
7834 | static bool | |
7835 | aarch64_composite_type_p (const_tree type, | |
ef4bddc2 | 7836 | machine_mode mode) |
43e9d192 IB |
7837 | { |
7838 | if (type && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == COMPLEX_TYPE)) | |
7839 | return true; | |
7840 | ||
7841 | if (mode == BLKmode | |
7842 | || GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT | |
7843 | || GET_MODE_CLASS (mode) == MODE_COMPLEX_INT) | |
7844 | return true; | |
7845 | ||
7846 | return false; | |
7847 | } | |
7848 | ||
7849 | /* Return TRUE if the type, as described by TYPE and MODE, is a short vector | |
7850 | type as described in AAPCS64 \S 4.1.2. | |
7851 | ||
7852 | See the comment above aarch64_composite_type_p for the notes on MODE. */ | |
7853 | ||
7854 | static bool | |
7855 | aarch64_short_vector_p (const_tree type, | |
ef4bddc2 | 7856 | machine_mode mode) |
43e9d192 IB |
7857 | { |
7858 | HOST_WIDE_INT size = -1; | |
7859 | ||
7860 | if (type && TREE_CODE (type) == VECTOR_TYPE) | |
7861 | size = int_size_in_bytes (type); | |
7862 | else if (!aarch64_composite_type_p (type, mode) | |
7863 | && (GET_MODE_CLASS (mode) == MODE_VECTOR_INT | |
7864 | || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT)) | |
7865 | size = GET_MODE_SIZE (mode); | |
7866 | ||
7867 | return (size == 8 || size == 16) ? true : false; | |
7868 | } | |
7869 | ||
7870 | /* Return TRUE if an argument, whose type is described by TYPE and MODE, | |
7871 | shall be passed or returned in simd/fp register(s) (providing these | |
7872 | parameter passing registers are available). | |
7873 | ||
7874 | Upon successful return, *COUNT returns the number of needed registers, | |
7875 | *BASE_MODE returns the mode of the individual register and when IS_HAF | |
7876 | is not NULL, *IS_HA indicates whether or not the argument is a homogeneous | |
7877 | floating-point aggregate or a homogeneous short-vector aggregate. */ | |
7878 | ||
7879 | static bool | |
ef4bddc2 | 7880 | aarch64_vfp_is_call_or_return_candidate (machine_mode mode, |
43e9d192 | 7881 | const_tree type, |
ef4bddc2 | 7882 | machine_mode *base_mode, |
43e9d192 IB |
7883 | int *count, |
7884 | bool *is_ha) | |
7885 | { | |
ef4bddc2 | 7886 | machine_mode new_mode = VOIDmode; |
43e9d192 IB |
7887 | bool composite_p = aarch64_composite_type_p (type, mode); |
7888 | ||
7889 | if (is_ha != NULL) *is_ha = false; | |
7890 | ||
7891 | if ((!composite_p && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
7892 | || aarch64_short_vector_p (type, mode)) | |
7893 | { | |
7894 | *count = 1; | |
7895 | new_mode = mode; | |
7896 | } | |
7897 | else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT) | |
7898 | { | |
7899 | if (is_ha != NULL) *is_ha = true; | |
7900 | *count = 2; | |
7901 | new_mode = GET_MODE_INNER (mode); | |
7902 | } | |
7903 | else if (type && composite_p) | |
7904 | { | |
7905 | int ag_count = aapcs_vfp_sub_candidate (type, &new_mode); | |
7906 | ||
7907 | if (ag_count > 0 && ag_count <= HA_MAX_NUM_FLDS) | |
7908 | { | |
7909 | if (is_ha != NULL) *is_ha = true; | |
7910 | *count = ag_count; | |
7911 | } | |
7912 | else | |
7913 | return false; | |
7914 | } | |
7915 | else | |
7916 | return false; | |
7917 | ||
7918 | *base_mode = new_mode; | |
7919 | return true; | |
7920 | } | |
7921 | ||
7922 | /* Implement TARGET_STRUCT_VALUE_RTX. */ | |
7923 | ||
7924 | static rtx | |
7925 | aarch64_struct_value_rtx (tree fndecl ATTRIBUTE_UNUSED, | |
7926 | int incoming ATTRIBUTE_UNUSED) | |
7927 | { | |
7928 | return gen_rtx_REG (Pmode, AARCH64_STRUCT_VALUE_REGNUM); | |
7929 | } | |
7930 | ||
7931 | /* Implements target hook vector_mode_supported_p. */ | |
7932 | static bool | |
ef4bddc2 | 7933 | aarch64_vector_mode_supported_p (machine_mode mode) |
43e9d192 IB |
7934 | { |
7935 | if (TARGET_SIMD | |
7936 | && (mode == V4SImode || mode == V8HImode | |
7937 | || mode == V16QImode || mode == V2DImode | |
7938 | || mode == V2SImode || mode == V4HImode | |
7939 | || mode == V8QImode || mode == V2SFmode | |
ad7d90cc AL |
7940 | || mode == V4SFmode || mode == V2DFmode |
7941 | || mode == V1DFmode)) | |
43e9d192 IB |
7942 | return true; |
7943 | ||
7944 | return false; | |
7945 | } | |
7946 | ||
b7342d25 IB |
7947 | /* Return appropriate SIMD container |
7948 | for MODE within a vector of WIDTH bits. */ | |
ef4bddc2 RS |
7949 | static machine_mode |
7950 | aarch64_simd_container_mode (machine_mode mode, unsigned width) | |
43e9d192 | 7951 | { |
b7342d25 | 7952 | gcc_assert (width == 64 || width == 128); |
43e9d192 | 7953 | if (TARGET_SIMD) |
b7342d25 IB |
7954 | { |
7955 | if (width == 128) | |
7956 | switch (mode) | |
7957 | { | |
7958 | case DFmode: | |
7959 | return V2DFmode; | |
7960 | case SFmode: | |
7961 | return V4SFmode; | |
7962 | case SImode: | |
7963 | return V4SImode; | |
7964 | case HImode: | |
7965 | return V8HImode; | |
7966 | case QImode: | |
7967 | return V16QImode; | |
7968 | case DImode: | |
7969 | return V2DImode; | |
7970 | default: | |
7971 | break; | |
7972 | } | |
7973 | else | |
7974 | switch (mode) | |
7975 | { | |
7976 | case SFmode: | |
7977 | return V2SFmode; | |
7978 | case SImode: | |
7979 | return V2SImode; | |
7980 | case HImode: | |
7981 | return V4HImode; | |
7982 | case QImode: | |
7983 | return V8QImode; | |
7984 | default: | |
7985 | break; | |
7986 | } | |
7987 | } | |
43e9d192 IB |
7988 | return word_mode; |
7989 | } | |
7990 | ||
b7342d25 | 7991 | /* Return 128-bit container as the preferred SIMD mode for MODE. */ |
ef4bddc2 RS |
7992 | static machine_mode |
7993 | aarch64_preferred_simd_mode (machine_mode mode) | |
b7342d25 IB |
7994 | { |
7995 | return aarch64_simd_container_mode (mode, 128); | |
7996 | } | |
7997 | ||
3b357264 JG |
7998 | /* Return the bitmask of possible vector sizes for the vectorizer |
7999 | to iterate over. */ | |
8000 | static unsigned int | |
8001 | aarch64_autovectorize_vector_sizes (void) | |
8002 | { | |
8003 | return (16 | 8); | |
8004 | } | |
8005 | ||
ac2b960f YZ |
8006 | /* Implement TARGET_MANGLE_TYPE. */ |
8007 | ||
6f549691 | 8008 | static const char * |
ac2b960f YZ |
8009 | aarch64_mangle_type (const_tree type) |
8010 | { | |
8011 | /* The AArch64 ABI documents say that "__va_list" has to be | |
8012 | managled as if it is in the "std" namespace. */ | |
8013 | if (lang_hooks.types_compatible_p (CONST_CAST_TREE (type), va_list_type)) | |
8014 | return "St9__va_list"; | |
8015 | ||
f9d53c27 TB |
8016 | /* Mangle AArch64-specific internal types. TYPE_NAME is non-NULL_TREE for |
8017 | builtin types. */ | |
8018 | if (TYPE_NAME (type) != NULL) | |
8019 | return aarch64_mangle_builtin_type (type); | |
c6fc9e43 | 8020 | |
ac2b960f YZ |
8021 | /* Use the default mangling. */ |
8022 | return NULL; | |
8023 | } | |
8024 | ||
8baff86e KT |
8025 | |
8026 | /* Return true if the rtx_insn contains a MEM RTX somewhere | |
8027 | in it. */ | |
75cf1494 KT |
8028 | |
8029 | static bool | |
8baff86e | 8030 | has_memory_op (rtx_insn *mem_insn) |
75cf1494 | 8031 | { |
8baff86e KT |
8032 | subrtx_iterator::array_type array; |
8033 | FOR_EACH_SUBRTX (iter, array, PATTERN (mem_insn), ALL) | |
8034 | if (MEM_P (*iter)) | |
8035 | return true; | |
8036 | ||
8037 | return false; | |
75cf1494 KT |
8038 | } |
8039 | ||
8040 | /* Find the first rtx_insn before insn that will generate an assembly | |
8041 | instruction. */ | |
8042 | ||
8043 | static rtx_insn * | |
8044 | aarch64_prev_real_insn (rtx_insn *insn) | |
8045 | { | |
8046 | if (!insn) | |
8047 | return NULL; | |
8048 | ||
8049 | do | |
8050 | { | |
8051 | insn = prev_real_insn (insn); | |
8052 | } | |
8053 | while (insn && recog_memoized (insn) < 0); | |
8054 | ||
8055 | return insn; | |
8056 | } | |
8057 | ||
8058 | static bool | |
8059 | is_madd_op (enum attr_type t1) | |
8060 | { | |
8061 | unsigned int i; | |
8062 | /* A number of these may be AArch32 only. */ | |
8063 | enum attr_type mlatypes[] = { | |
8064 | TYPE_MLA, TYPE_MLAS, TYPE_SMLAD, TYPE_SMLADX, TYPE_SMLAL, TYPE_SMLALD, | |
8065 | TYPE_SMLALS, TYPE_SMLALXY, TYPE_SMLAWX, TYPE_SMLAWY, TYPE_SMLAXY, | |
8066 | TYPE_SMMLA, TYPE_UMLAL, TYPE_UMLALS,TYPE_SMLSD, TYPE_SMLSDX, TYPE_SMLSLD | |
8067 | }; | |
8068 | ||
8069 | for (i = 0; i < sizeof (mlatypes) / sizeof (enum attr_type); i++) | |
8070 | { | |
8071 | if (t1 == mlatypes[i]) | |
8072 | return true; | |
8073 | } | |
8074 | ||
8075 | return false; | |
8076 | } | |
8077 | ||
8078 | /* Check if there is a register dependency between a load and the insn | |
8079 | for which we hold recog_data. */ | |
8080 | ||
8081 | static bool | |
8082 | dep_between_memop_and_curr (rtx memop) | |
8083 | { | |
8084 | rtx load_reg; | |
8085 | int opno; | |
8086 | ||
8baff86e | 8087 | gcc_assert (GET_CODE (memop) == SET); |
75cf1494 KT |
8088 | |
8089 | if (!REG_P (SET_DEST (memop))) | |
8090 | return false; | |
8091 | ||
8092 | load_reg = SET_DEST (memop); | |
8baff86e | 8093 | for (opno = 1; opno < recog_data.n_operands; opno++) |
75cf1494 KT |
8094 | { |
8095 | rtx operand = recog_data.operand[opno]; | |
8096 | if (REG_P (operand) | |
8097 | && reg_overlap_mentioned_p (load_reg, operand)) | |
8098 | return true; | |
8099 | ||
8100 | } | |
8101 | return false; | |
8102 | } | |
8103 | ||
8baff86e KT |
8104 | |
8105 | /* When working around the Cortex-A53 erratum 835769, | |
8106 | given rtx_insn INSN, return true if it is a 64-bit multiply-accumulate | |
8107 | instruction and has a preceding memory instruction such that a NOP | |
8108 | should be inserted between them. */ | |
8109 | ||
75cf1494 KT |
8110 | bool |
8111 | aarch64_madd_needs_nop (rtx_insn* insn) | |
8112 | { | |
8113 | enum attr_type attr_type; | |
8114 | rtx_insn *prev; | |
8115 | rtx body; | |
8116 | ||
8117 | if (!aarch64_fix_a53_err835769) | |
8118 | return false; | |
8119 | ||
8120 | if (recog_memoized (insn) < 0) | |
8121 | return false; | |
8122 | ||
8123 | attr_type = get_attr_type (insn); | |
8124 | if (!is_madd_op (attr_type)) | |
8125 | return false; | |
8126 | ||
8127 | prev = aarch64_prev_real_insn (insn); | |
3fea1a75 KT |
8128 | /* aarch64_prev_real_insn can call recog_memoized on insns other than INSN. |
8129 | Restore recog state to INSN to avoid state corruption. */ | |
8130 | extract_constrain_insn_cached (insn); | |
8131 | ||
8baff86e | 8132 | if (!prev || !has_memory_op (prev)) |
75cf1494 KT |
8133 | return false; |
8134 | ||
8135 | body = single_set (prev); | |
8136 | ||
8137 | /* If the previous insn is a memory op and there is no dependency between | |
8baff86e KT |
8138 | it and the DImode madd, emit a NOP between them. If body is NULL then we |
8139 | have a complex memory operation, probably a load/store pair. | |
8140 | Be conservative for now and emit a NOP. */ | |
8141 | if (GET_MODE (recog_data.operand[0]) == DImode | |
8142 | && (!body || !dep_between_memop_and_curr (body))) | |
75cf1494 KT |
8143 | return true; |
8144 | ||
8145 | return false; | |
8146 | ||
8147 | } | |
8148 | ||
8baff86e KT |
8149 | |
8150 | /* Implement FINAL_PRESCAN_INSN. */ | |
8151 | ||
75cf1494 KT |
8152 | void |
8153 | aarch64_final_prescan_insn (rtx_insn *insn) | |
8154 | { | |
8155 | if (aarch64_madd_needs_nop (insn)) | |
8156 | fprintf (asm_out_file, "\tnop // between mem op and mult-accumulate\n"); | |
8157 | } | |
8158 | ||
8159 | ||
43e9d192 | 8160 | /* Return the equivalent letter for size. */ |
81c2dfb9 | 8161 | static char |
43e9d192 IB |
8162 | sizetochar (int size) |
8163 | { | |
8164 | switch (size) | |
8165 | { | |
8166 | case 64: return 'd'; | |
8167 | case 32: return 's'; | |
8168 | case 16: return 'h'; | |
8169 | case 8 : return 'b'; | |
8170 | default: gcc_unreachable (); | |
8171 | } | |
8172 | } | |
8173 | ||
3520f7cc JG |
8174 | /* Return true iff x is a uniform vector of floating-point |
8175 | constants, and the constant can be represented in | |
8176 | quarter-precision form. Note, as aarch64_float_const_representable | |
8177 | rejects both +0.0 and -0.0, we will also reject +0.0 and -0.0. */ | |
8178 | static bool | |
8179 | aarch64_vect_float_const_representable_p (rtx x) | |
8180 | { | |
8181 | int i = 0; | |
8182 | REAL_VALUE_TYPE r0, ri; | |
8183 | rtx x0, xi; | |
8184 | ||
8185 | if (GET_MODE_CLASS (GET_MODE (x)) != MODE_VECTOR_FLOAT) | |
8186 | return false; | |
8187 | ||
8188 | x0 = CONST_VECTOR_ELT (x, 0); | |
8189 | if (!CONST_DOUBLE_P (x0)) | |
8190 | return false; | |
8191 | ||
8192 | REAL_VALUE_FROM_CONST_DOUBLE (r0, x0); | |
8193 | ||
8194 | for (i = 1; i < CONST_VECTOR_NUNITS (x); i++) | |
8195 | { | |
8196 | xi = CONST_VECTOR_ELT (x, i); | |
8197 | if (!CONST_DOUBLE_P (xi)) | |
8198 | return false; | |
8199 | ||
8200 | REAL_VALUE_FROM_CONST_DOUBLE (ri, xi); | |
8201 | if (!REAL_VALUES_EQUAL (r0, ri)) | |
8202 | return false; | |
8203 | } | |
8204 | ||
8205 | return aarch64_float_const_representable_p (x0); | |
8206 | } | |
8207 | ||
d8edd899 | 8208 | /* Return true for valid and false for invalid. */ |
3ea63f60 | 8209 | bool |
ef4bddc2 | 8210 | aarch64_simd_valid_immediate (rtx op, machine_mode mode, bool inverse, |
48063b9d | 8211 | struct simd_immediate_info *info) |
43e9d192 IB |
8212 | { |
8213 | #define CHECK(STRIDE, ELSIZE, CLASS, TEST, SHIFT, NEG) \ | |
8214 | matches = 1; \ | |
8215 | for (i = 0; i < idx; i += (STRIDE)) \ | |
8216 | if (!(TEST)) \ | |
8217 | matches = 0; \ | |
8218 | if (matches) \ | |
8219 | { \ | |
8220 | immtype = (CLASS); \ | |
8221 | elsize = (ELSIZE); \ | |
43e9d192 IB |
8222 | eshift = (SHIFT); \ |
8223 | emvn = (NEG); \ | |
8224 | break; \ | |
8225 | } | |
8226 | ||
8227 | unsigned int i, elsize = 0, idx = 0, n_elts = CONST_VECTOR_NUNITS (op); | |
8228 | unsigned int innersize = GET_MODE_SIZE (GET_MODE_INNER (mode)); | |
8229 | unsigned char bytes[16]; | |
43e9d192 IB |
8230 | int immtype = -1, matches; |
8231 | unsigned int invmask = inverse ? 0xff : 0; | |
8232 | int eshift, emvn; | |
8233 | ||
43e9d192 | 8234 | if (GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) |
3520f7cc | 8235 | { |
81c2dfb9 IB |
8236 | if (! (aarch64_simd_imm_zero_p (op, mode) |
8237 | || aarch64_vect_float_const_representable_p (op))) | |
d8edd899 | 8238 | return false; |
3520f7cc | 8239 | |
48063b9d IB |
8240 | if (info) |
8241 | { | |
8242 | info->value = CONST_VECTOR_ELT (op, 0); | |
81c2dfb9 | 8243 | info->element_width = GET_MODE_BITSIZE (GET_MODE (info->value)); |
48063b9d IB |
8244 | info->mvn = false; |
8245 | info->shift = 0; | |
8246 | } | |
3520f7cc | 8247 | |
d8edd899 | 8248 | return true; |
3520f7cc | 8249 | } |
43e9d192 IB |
8250 | |
8251 | /* Splat vector constant out into a byte vector. */ | |
8252 | for (i = 0; i < n_elts; i++) | |
8253 | { | |
4b1e108c AL |
8254 | /* The vector is provided in gcc endian-neutral fashion. For aarch64_be, |
8255 | it must be laid out in the vector register in reverse order. */ | |
8256 | rtx el = CONST_VECTOR_ELT (op, BYTES_BIG_ENDIAN ? (n_elts - 1 - i) : i); | |
43e9d192 IB |
8257 | unsigned HOST_WIDE_INT elpart; |
8258 | unsigned int part, parts; | |
8259 | ||
4aa81c2e | 8260 | if (CONST_INT_P (el)) |
43e9d192 IB |
8261 | { |
8262 | elpart = INTVAL (el); | |
8263 | parts = 1; | |
8264 | } | |
8265 | else if (GET_CODE (el) == CONST_DOUBLE) | |
8266 | { | |
8267 | elpart = CONST_DOUBLE_LOW (el); | |
8268 | parts = 2; | |
8269 | } | |
8270 | else | |
8271 | gcc_unreachable (); | |
8272 | ||
8273 | for (part = 0; part < parts; part++) | |
8274 | { | |
8275 | unsigned int byte; | |
8276 | for (byte = 0; byte < innersize; byte++) | |
8277 | { | |
8278 | bytes[idx++] = (elpart & 0xff) ^ invmask; | |
8279 | elpart >>= BITS_PER_UNIT; | |
8280 | } | |
8281 | if (GET_CODE (el) == CONST_DOUBLE) | |
8282 | elpart = CONST_DOUBLE_HIGH (el); | |
8283 | } | |
8284 | } | |
8285 | ||
8286 | /* Sanity check. */ | |
8287 | gcc_assert (idx == GET_MODE_SIZE (mode)); | |
8288 | ||
8289 | do | |
8290 | { | |
8291 | CHECK (4, 32, 0, bytes[i] == bytes[0] && bytes[i + 1] == 0 | |
8292 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 0, 0); | |
8293 | ||
8294 | CHECK (4, 32, 1, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
8295 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 8, 0); | |
8296 | ||
8297 | CHECK (4, 32, 2, bytes[i] == 0 && bytes[i + 1] == 0 | |
8298 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0, 16, 0); | |
8299 | ||
8300 | CHECK (4, 32, 3, bytes[i] == 0 && bytes[i + 1] == 0 | |
8301 | && bytes[i + 2] == 0 && bytes[i + 3] == bytes[3], 24, 0); | |
8302 | ||
8303 | CHECK (2, 16, 4, bytes[i] == bytes[0] && bytes[i + 1] == 0, 0, 0); | |
8304 | ||
8305 | CHECK (2, 16, 5, bytes[i] == 0 && bytes[i + 1] == bytes[1], 8, 0); | |
8306 | ||
8307 | CHECK (4, 32, 6, bytes[i] == bytes[0] && bytes[i + 1] == 0xff | |
8308 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 0, 1); | |
8309 | ||
8310 | CHECK (4, 32, 7, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
8311 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 8, 1); | |
8312 | ||
8313 | CHECK (4, 32, 8, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
8314 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff, 16, 1); | |
8315 | ||
8316 | CHECK (4, 32, 9, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
8317 | && bytes[i + 2] == 0xff && bytes[i + 3] == bytes[3], 24, 1); | |
8318 | ||
8319 | CHECK (2, 16, 10, bytes[i] == bytes[0] && bytes[i + 1] == 0xff, 0, 1); | |
8320 | ||
8321 | CHECK (2, 16, 11, bytes[i] == 0xff && bytes[i + 1] == bytes[1], 8, 1); | |
8322 | ||
8323 | CHECK (4, 32, 12, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
e4f0f84d | 8324 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 8, 0); |
43e9d192 IB |
8325 | |
8326 | CHECK (4, 32, 13, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
e4f0f84d | 8327 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 8, 1); |
43e9d192 IB |
8328 | |
8329 | CHECK (4, 32, 14, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
e4f0f84d | 8330 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0, 16, 0); |
43e9d192 IB |
8331 | |
8332 | CHECK (4, 32, 15, bytes[i] == 0 && bytes[i + 1] == 0 | |
e4f0f84d | 8333 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff, 16, 1); |
43e9d192 IB |
8334 | |
8335 | CHECK (1, 8, 16, bytes[i] == bytes[0], 0, 0); | |
8336 | ||
8337 | CHECK (1, 64, 17, (bytes[i] == 0 || bytes[i] == 0xff) | |
8338 | && bytes[i] == bytes[(i + 8) % idx], 0, 0); | |
8339 | } | |
8340 | while (0); | |
8341 | ||
e4f0f84d | 8342 | if (immtype == -1) |
d8edd899 | 8343 | return false; |
43e9d192 | 8344 | |
48063b9d | 8345 | if (info) |
43e9d192 | 8346 | { |
48063b9d | 8347 | info->element_width = elsize; |
48063b9d IB |
8348 | info->mvn = emvn != 0; |
8349 | info->shift = eshift; | |
8350 | ||
43e9d192 IB |
8351 | unsigned HOST_WIDE_INT imm = 0; |
8352 | ||
e4f0f84d TB |
8353 | if (immtype >= 12 && immtype <= 15) |
8354 | info->msl = true; | |
8355 | ||
43e9d192 IB |
8356 | /* Un-invert bytes of recognized vector, if necessary. */ |
8357 | if (invmask != 0) | |
8358 | for (i = 0; i < idx; i++) | |
8359 | bytes[i] ^= invmask; | |
8360 | ||
8361 | if (immtype == 17) | |
8362 | { | |
8363 | /* FIXME: Broken on 32-bit H_W_I hosts. */ | |
8364 | gcc_assert (sizeof (HOST_WIDE_INT) == 8); | |
8365 | ||
8366 | for (i = 0; i < 8; i++) | |
8367 | imm |= (unsigned HOST_WIDE_INT) (bytes[i] ? 0xff : 0) | |
8368 | << (i * BITS_PER_UNIT); | |
8369 | ||
43e9d192 | 8370 | |
48063b9d IB |
8371 | info->value = GEN_INT (imm); |
8372 | } | |
8373 | else | |
8374 | { | |
8375 | for (i = 0; i < elsize / BITS_PER_UNIT; i++) | |
8376 | imm |= (unsigned HOST_WIDE_INT) bytes[i] << (i * BITS_PER_UNIT); | |
43e9d192 IB |
8377 | |
8378 | /* Construct 'abcdefgh' because the assembler cannot handle | |
48063b9d IB |
8379 | generic constants. */ |
8380 | if (info->mvn) | |
43e9d192 | 8381 | imm = ~imm; |
48063b9d IB |
8382 | imm = (imm >> info->shift) & 0xff; |
8383 | info->value = GEN_INT (imm); | |
8384 | } | |
43e9d192 IB |
8385 | } |
8386 | ||
48063b9d | 8387 | return true; |
43e9d192 IB |
8388 | #undef CHECK |
8389 | } | |
8390 | ||
43e9d192 IB |
8391 | /* Check of immediate shift constants are within range. */ |
8392 | bool | |
ef4bddc2 | 8393 | aarch64_simd_shift_imm_p (rtx x, machine_mode mode, bool left) |
43e9d192 IB |
8394 | { |
8395 | int bit_width = GET_MODE_UNIT_SIZE (mode) * BITS_PER_UNIT; | |
8396 | if (left) | |
ddeabd3e | 8397 | return aarch64_const_vec_all_same_in_range_p (x, 0, bit_width - 1); |
43e9d192 | 8398 | else |
ddeabd3e | 8399 | return aarch64_const_vec_all_same_in_range_p (x, 1, bit_width); |
43e9d192 IB |
8400 | } |
8401 | ||
3520f7cc JG |
8402 | /* Return true if X is a uniform vector where all elements |
8403 | are either the floating-point constant 0.0 or the | |
8404 | integer constant 0. */ | |
43e9d192 | 8405 | bool |
ef4bddc2 | 8406 | aarch64_simd_imm_zero_p (rtx x, machine_mode mode) |
43e9d192 | 8407 | { |
3520f7cc | 8408 | return x == CONST0_RTX (mode); |
43e9d192 IB |
8409 | } |
8410 | ||
8411 | bool | |
ef4bddc2 | 8412 | aarch64_simd_imm_scalar_p (rtx x, machine_mode mode ATTRIBUTE_UNUSED) |
43e9d192 IB |
8413 | { |
8414 | HOST_WIDE_INT imm = INTVAL (x); | |
8415 | int i; | |
8416 | ||
8417 | for (i = 0; i < 8; i++) | |
8418 | { | |
8419 | unsigned int byte = imm & 0xff; | |
8420 | if (byte != 0xff && byte != 0) | |
8421 | return false; | |
8422 | imm >>= 8; | |
8423 | } | |
8424 | ||
8425 | return true; | |
8426 | } | |
8427 | ||
83f8c414 CSS |
8428 | bool |
8429 | aarch64_mov_operand_p (rtx x, | |
a5350ddc | 8430 | enum aarch64_symbol_context context, |
ef4bddc2 | 8431 | machine_mode mode) |
83f8c414 | 8432 | { |
83f8c414 CSS |
8433 | if (GET_CODE (x) == HIGH |
8434 | && aarch64_valid_symref (XEXP (x, 0), GET_MODE (XEXP (x, 0)))) | |
8435 | return true; | |
8436 | ||
82614948 | 8437 | if (CONST_INT_P (x)) |
83f8c414 CSS |
8438 | return true; |
8439 | ||
8440 | if (GET_CODE (x) == SYMBOL_REF && mode == DImode && CONSTANT_ADDRESS_P (x)) | |
8441 | return true; | |
8442 | ||
a5350ddc CSS |
8443 | return aarch64_classify_symbolic_expression (x, context) |
8444 | == SYMBOL_TINY_ABSOLUTE; | |
83f8c414 CSS |
8445 | } |
8446 | ||
43e9d192 IB |
8447 | /* Return a const_int vector of VAL. */ |
8448 | rtx | |
ef4bddc2 | 8449 | aarch64_simd_gen_const_vector_dup (machine_mode mode, int val) |
43e9d192 IB |
8450 | { |
8451 | int nunits = GET_MODE_NUNITS (mode); | |
8452 | rtvec v = rtvec_alloc (nunits); | |
8453 | int i; | |
8454 | ||
8455 | for (i=0; i < nunits; i++) | |
8456 | RTVEC_ELT (v, i) = GEN_INT (val); | |
8457 | ||
8458 | return gen_rtx_CONST_VECTOR (mode, v); | |
8459 | } | |
8460 | ||
051d0e2f SN |
8461 | /* Check OP is a legal scalar immediate for the MOVI instruction. */ |
8462 | ||
8463 | bool | |
ef4bddc2 | 8464 | aarch64_simd_scalar_immediate_valid_for_move (rtx op, machine_mode mode) |
051d0e2f | 8465 | { |
ef4bddc2 | 8466 | machine_mode vmode; |
051d0e2f SN |
8467 | |
8468 | gcc_assert (!VECTOR_MODE_P (mode)); | |
8469 | vmode = aarch64_preferred_simd_mode (mode); | |
8470 | rtx op_v = aarch64_simd_gen_const_vector_dup (vmode, INTVAL (op)); | |
48063b9d | 8471 | return aarch64_simd_valid_immediate (op_v, vmode, false, NULL); |
051d0e2f SN |
8472 | } |
8473 | ||
988fa693 JG |
8474 | /* Construct and return a PARALLEL RTX vector with elements numbering the |
8475 | lanes of either the high (HIGH == TRUE) or low (HIGH == FALSE) half of | |
8476 | the vector - from the perspective of the architecture. This does not | |
8477 | line up with GCC's perspective on lane numbers, so we end up with | |
8478 | different masks depending on our target endian-ness. The diagram | |
8479 | below may help. We must draw the distinction when building masks | |
8480 | which select one half of the vector. An instruction selecting | |
8481 | architectural low-lanes for a big-endian target, must be described using | |
8482 | a mask selecting GCC high-lanes. | |
8483 | ||
8484 | Big-Endian Little-Endian | |
8485 | ||
8486 | GCC 0 1 2 3 3 2 1 0 | |
8487 | | x | x | x | x | | x | x | x | x | | |
8488 | Architecture 3 2 1 0 3 2 1 0 | |
8489 | ||
8490 | Low Mask: { 2, 3 } { 0, 1 } | |
8491 | High Mask: { 0, 1 } { 2, 3 } | |
8492 | */ | |
8493 | ||
43e9d192 | 8494 | rtx |
ef4bddc2 | 8495 | aarch64_simd_vect_par_cnst_half (machine_mode mode, bool high) |
43e9d192 IB |
8496 | { |
8497 | int nunits = GET_MODE_NUNITS (mode); | |
8498 | rtvec v = rtvec_alloc (nunits / 2); | |
988fa693 JG |
8499 | int high_base = nunits / 2; |
8500 | int low_base = 0; | |
8501 | int base; | |
43e9d192 IB |
8502 | rtx t1; |
8503 | int i; | |
8504 | ||
988fa693 JG |
8505 | if (BYTES_BIG_ENDIAN) |
8506 | base = high ? low_base : high_base; | |
8507 | else | |
8508 | base = high ? high_base : low_base; | |
8509 | ||
8510 | for (i = 0; i < nunits / 2; i++) | |
43e9d192 IB |
8511 | RTVEC_ELT (v, i) = GEN_INT (base + i); |
8512 | ||
8513 | t1 = gen_rtx_PARALLEL (mode, v); | |
8514 | return t1; | |
8515 | } | |
8516 | ||
988fa693 JG |
8517 | /* Check OP for validity as a PARALLEL RTX vector with elements |
8518 | numbering the lanes of either the high (HIGH == TRUE) or low lanes, | |
8519 | from the perspective of the architecture. See the diagram above | |
8520 | aarch64_simd_vect_par_cnst_half for more details. */ | |
8521 | ||
8522 | bool | |
ef4bddc2 | 8523 | aarch64_simd_check_vect_par_cnst_half (rtx op, machine_mode mode, |
988fa693 JG |
8524 | bool high) |
8525 | { | |
8526 | rtx ideal = aarch64_simd_vect_par_cnst_half (mode, high); | |
8527 | HOST_WIDE_INT count_op = XVECLEN (op, 0); | |
8528 | HOST_WIDE_INT count_ideal = XVECLEN (ideal, 0); | |
8529 | int i = 0; | |
8530 | ||
8531 | if (!VECTOR_MODE_P (mode)) | |
8532 | return false; | |
8533 | ||
8534 | if (count_op != count_ideal) | |
8535 | return false; | |
8536 | ||
8537 | for (i = 0; i < count_ideal; i++) | |
8538 | { | |
8539 | rtx elt_op = XVECEXP (op, 0, i); | |
8540 | rtx elt_ideal = XVECEXP (ideal, 0, i); | |
8541 | ||
4aa81c2e | 8542 | if (!CONST_INT_P (elt_op) |
988fa693 JG |
8543 | || INTVAL (elt_ideal) != INTVAL (elt_op)) |
8544 | return false; | |
8545 | } | |
8546 | return true; | |
8547 | } | |
8548 | ||
43e9d192 IB |
8549 | /* Bounds-check lanes. Ensure OPERAND lies between LOW (inclusive) and |
8550 | HIGH (exclusive). */ | |
8551 | void | |
46ed6024 CB |
8552 | aarch64_simd_lane_bounds (rtx operand, HOST_WIDE_INT low, HOST_WIDE_INT high, |
8553 | const_tree exp) | |
43e9d192 IB |
8554 | { |
8555 | HOST_WIDE_INT lane; | |
4aa81c2e | 8556 | gcc_assert (CONST_INT_P (operand)); |
43e9d192 IB |
8557 | lane = INTVAL (operand); |
8558 | ||
8559 | if (lane < low || lane >= high) | |
46ed6024 CB |
8560 | { |
8561 | if (exp) | |
8562 | error ("%Klane %ld out of range %ld - %ld", exp, lane, low, high - 1); | |
8563 | else | |
8564 | error ("lane %ld out of range %ld - %ld", lane, low, high - 1); | |
8565 | } | |
43e9d192 IB |
8566 | } |
8567 | ||
43e9d192 IB |
8568 | /* Emit code to place a AdvSIMD pair result in memory locations (with equal |
8569 | registers). */ | |
8570 | void | |
ef4bddc2 | 8571 | aarch64_simd_emit_pair_result_insn (machine_mode mode, |
43e9d192 IB |
8572 | rtx (*intfn) (rtx, rtx, rtx), rtx destaddr, |
8573 | rtx op1) | |
8574 | { | |
8575 | rtx mem = gen_rtx_MEM (mode, destaddr); | |
8576 | rtx tmp1 = gen_reg_rtx (mode); | |
8577 | rtx tmp2 = gen_reg_rtx (mode); | |
8578 | ||
8579 | emit_insn (intfn (tmp1, op1, tmp2)); | |
8580 | ||
8581 | emit_move_insn (mem, tmp1); | |
8582 | mem = adjust_address (mem, mode, GET_MODE_SIZE (mode)); | |
8583 | emit_move_insn (mem, tmp2); | |
8584 | } | |
8585 | ||
8586 | /* Return TRUE if OP is a valid vector addressing mode. */ | |
8587 | bool | |
8588 | aarch64_simd_mem_operand_p (rtx op) | |
8589 | { | |
8590 | return MEM_P (op) && (GET_CODE (XEXP (op, 0)) == POST_INC | |
4aa81c2e | 8591 | || REG_P (XEXP (op, 0))); |
43e9d192 IB |
8592 | } |
8593 | ||
2d8c6dc1 AH |
8594 | /* Emit a register copy from operand to operand, taking care not to |
8595 | early-clobber source registers in the process. | |
43e9d192 | 8596 | |
2d8c6dc1 AH |
8597 | COUNT is the number of components into which the copy needs to be |
8598 | decomposed. */ | |
43e9d192 | 8599 | void |
2d8c6dc1 AH |
8600 | aarch64_simd_emit_reg_reg_move (rtx *operands, enum machine_mode mode, |
8601 | unsigned int count) | |
43e9d192 IB |
8602 | { |
8603 | unsigned int i; | |
2d8c6dc1 AH |
8604 | int rdest = REGNO (operands[0]); |
8605 | int rsrc = REGNO (operands[1]); | |
43e9d192 IB |
8606 | |
8607 | if (!reg_overlap_mentioned_p (operands[0], operands[1]) | |
2d8c6dc1 AH |
8608 | || rdest < rsrc) |
8609 | for (i = 0; i < count; i++) | |
8610 | emit_move_insn (gen_rtx_REG (mode, rdest + i), | |
8611 | gen_rtx_REG (mode, rsrc + i)); | |
43e9d192 | 8612 | else |
2d8c6dc1 AH |
8613 | for (i = 0; i < count; i++) |
8614 | emit_move_insn (gen_rtx_REG (mode, rdest + count - i - 1), | |
8615 | gen_rtx_REG (mode, rsrc + count - i - 1)); | |
43e9d192 IB |
8616 | } |
8617 | ||
8618 | /* Compute and return the length of aarch64_simd_mov<mode>, where <mode> is | |
8619 | one of VSTRUCT modes: OI, CI or XI. */ | |
8620 | int | |
647d790d | 8621 | aarch64_simd_attr_length_move (rtx_insn *insn) |
43e9d192 | 8622 | { |
ef4bddc2 | 8623 | machine_mode mode; |
43e9d192 IB |
8624 | |
8625 | extract_insn_cached (insn); | |
8626 | ||
8627 | if (REG_P (recog_data.operand[0]) && REG_P (recog_data.operand[1])) | |
8628 | { | |
8629 | mode = GET_MODE (recog_data.operand[0]); | |
8630 | switch (mode) | |
8631 | { | |
8632 | case OImode: | |
8633 | return 8; | |
8634 | case CImode: | |
8635 | return 12; | |
8636 | case XImode: | |
8637 | return 16; | |
8638 | default: | |
8639 | gcc_unreachable (); | |
8640 | } | |
8641 | } | |
8642 | return 4; | |
8643 | } | |
8644 | ||
668046d1 DS |
8645 | /* Compute and return the length of aarch64_simd_reglist<mode>, where <mode> is |
8646 | one of VSTRUCT modes: OI, CI, EI, or XI. */ | |
8647 | int | |
8648 | aarch64_simd_attr_length_rglist (enum machine_mode mode) | |
8649 | { | |
8650 | return (GET_MODE_SIZE (mode) / UNITS_PER_VREG) * 4; | |
8651 | } | |
8652 | ||
db0253a4 TB |
8653 | /* Implement target hook TARGET_VECTOR_ALIGNMENT. The AAPCS64 sets the maximum |
8654 | alignment of a vector to 128 bits. */ | |
8655 | static HOST_WIDE_INT | |
8656 | aarch64_simd_vector_alignment (const_tree type) | |
8657 | { | |
9439e9a1 | 8658 | HOST_WIDE_INT align = tree_to_shwi (TYPE_SIZE (type)); |
db0253a4 TB |
8659 | return MIN (align, 128); |
8660 | } | |
8661 | ||
8662 | /* Implement target hook TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE. */ | |
8663 | static bool | |
8664 | aarch64_simd_vector_alignment_reachable (const_tree type, bool is_packed) | |
8665 | { | |
8666 | if (is_packed) | |
8667 | return false; | |
8668 | ||
8669 | /* We guarantee alignment for vectors up to 128-bits. */ | |
8670 | if (tree_int_cst_compare (TYPE_SIZE (type), | |
8671 | bitsize_int (BIGGEST_ALIGNMENT)) > 0) | |
8672 | return false; | |
8673 | ||
8674 | /* Vectors whose size is <= BIGGEST_ALIGNMENT are naturally aligned. */ | |
8675 | return true; | |
8676 | } | |
8677 | ||
4369c11e TB |
8678 | /* If VALS is a vector constant that can be loaded into a register |
8679 | using DUP, generate instructions to do so and return an RTX to | |
8680 | assign to the register. Otherwise return NULL_RTX. */ | |
8681 | static rtx | |
8682 | aarch64_simd_dup_constant (rtx vals) | |
8683 | { | |
ef4bddc2 RS |
8684 | machine_mode mode = GET_MODE (vals); |
8685 | machine_mode inner_mode = GET_MODE_INNER (mode); | |
4369c11e TB |
8686 | int n_elts = GET_MODE_NUNITS (mode); |
8687 | bool all_same = true; | |
8688 | rtx x; | |
8689 | int i; | |
8690 | ||
8691 | if (GET_CODE (vals) != CONST_VECTOR) | |
8692 | return NULL_RTX; | |
8693 | ||
8694 | for (i = 1; i < n_elts; ++i) | |
8695 | { | |
8696 | x = CONST_VECTOR_ELT (vals, i); | |
8697 | if (!rtx_equal_p (x, CONST_VECTOR_ELT (vals, 0))) | |
8698 | all_same = false; | |
8699 | } | |
8700 | ||
8701 | if (!all_same) | |
8702 | return NULL_RTX; | |
8703 | ||
8704 | /* We can load this constant by using DUP and a constant in a | |
8705 | single ARM register. This will be cheaper than a vector | |
8706 | load. */ | |
8707 | x = copy_to_mode_reg (inner_mode, CONST_VECTOR_ELT (vals, 0)); | |
8708 | return gen_rtx_VEC_DUPLICATE (mode, x); | |
8709 | } | |
8710 | ||
8711 | ||
8712 | /* Generate code to load VALS, which is a PARALLEL containing only | |
8713 | constants (for vec_init) or CONST_VECTOR, efficiently into a | |
8714 | register. Returns an RTX to copy into the register, or NULL_RTX | |
8715 | for a PARALLEL that can not be converted into a CONST_VECTOR. */ | |
1df3f464 | 8716 | static rtx |
4369c11e TB |
8717 | aarch64_simd_make_constant (rtx vals) |
8718 | { | |
ef4bddc2 | 8719 | machine_mode mode = GET_MODE (vals); |
4369c11e TB |
8720 | rtx const_dup; |
8721 | rtx const_vec = NULL_RTX; | |
8722 | int n_elts = GET_MODE_NUNITS (mode); | |
8723 | int n_const = 0; | |
8724 | int i; | |
8725 | ||
8726 | if (GET_CODE (vals) == CONST_VECTOR) | |
8727 | const_vec = vals; | |
8728 | else if (GET_CODE (vals) == PARALLEL) | |
8729 | { | |
8730 | /* A CONST_VECTOR must contain only CONST_INTs and | |
8731 | CONST_DOUBLEs, but CONSTANT_P allows more (e.g. SYMBOL_REF). | |
8732 | Only store valid constants in a CONST_VECTOR. */ | |
8733 | for (i = 0; i < n_elts; ++i) | |
8734 | { | |
8735 | rtx x = XVECEXP (vals, 0, i); | |
8736 | if (CONST_INT_P (x) || CONST_DOUBLE_P (x)) | |
8737 | n_const++; | |
8738 | } | |
8739 | if (n_const == n_elts) | |
8740 | const_vec = gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0)); | |
8741 | } | |
8742 | else | |
8743 | gcc_unreachable (); | |
8744 | ||
8745 | if (const_vec != NULL_RTX | |
48063b9d | 8746 | && aarch64_simd_valid_immediate (const_vec, mode, false, NULL)) |
4369c11e TB |
8747 | /* Load using MOVI/MVNI. */ |
8748 | return const_vec; | |
8749 | else if ((const_dup = aarch64_simd_dup_constant (vals)) != NULL_RTX) | |
8750 | /* Loaded using DUP. */ | |
8751 | return const_dup; | |
8752 | else if (const_vec != NULL_RTX) | |
8753 | /* Load from constant pool. We can not take advantage of single-cycle | |
8754 | LD1 because we need a PC-relative addressing mode. */ | |
8755 | return const_vec; | |
8756 | else | |
8757 | /* A PARALLEL containing something not valid inside CONST_VECTOR. | |
8758 | We can not construct an initializer. */ | |
8759 | return NULL_RTX; | |
8760 | } | |
8761 | ||
8762 | void | |
8763 | aarch64_expand_vector_init (rtx target, rtx vals) | |
8764 | { | |
ef4bddc2 RS |
8765 | machine_mode mode = GET_MODE (target); |
8766 | machine_mode inner_mode = GET_MODE_INNER (mode); | |
4369c11e TB |
8767 | int n_elts = GET_MODE_NUNITS (mode); |
8768 | int n_var = 0, one_var = -1; | |
8769 | bool all_same = true; | |
8770 | rtx x, mem; | |
8771 | int i; | |
8772 | ||
8773 | x = XVECEXP (vals, 0, 0); | |
8774 | if (!CONST_INT_P (x) && !CONST_DOUBLE_P (x)) | |
8775 | n_var = 1, one_var = 0; | |
8776 | ||
8777 | for (i = 1; i < n_elts; ++i) | |
8778 | { | |
8779 | x = XVECEXP (vals, 0, i); | |
8780 | if (!CONST_INT_P (x) && !CONST_DOUBLE_P (x)) | |
8781 | ++n_var, one_var = i; | |
8782 | ||
8783 | if (!rtx_equal_p (x, XVECEXP (vals, 0, 0))) | |
8784 | all_same = false; | |
8785 | } | |
8786 | ||
8787 | if (n_var == 0) | |
8788 | { | |
8789 | rtx constant = aarch64_simd_make_constant (vals); | |
8790 | if (constant != NULL_RTX) | |
8791 | { | |
8792 | emit_move_insn (target, constant); | |
8793 | return; | |
8794 | } | |
8795 | } | |
8796 | ||
8797 | /* Splat a single non-constant element if we can. */ | |
8798 | if (all_same) | |
8799 | { | |
8800 | x = copy_to_mode_reg (inner_mode, XVECEXP (vals, 0, 0)); | |
8801 | aarch64_emit_move (target, gen_rtx_VEC_DUPLICATE (mode, x)); | |
8802 | return; | |
8803 | } | |
8804 | ||
8805 | /* One field is non-constant. Load constant then overwrite varying | |
8806 | field. This is more efficient than using the stack. */ | |
8807 | if (n_var == 1) | |
8808 | { | |
8809 | rtx copy = copy_rtx (vals); | |
8810 | rtx index = GEN_INT (one_var); | |
8811 | enum insn_code icode; | |
8812 | ||
8813 | /* Load constant part of vector, substitute neighboring value for | |
8814 | varying element. */ | |
8815 | XVECEXP (copy, 0, one_var) = XVECEXP (vals, 0, one_var ^ 1); | |
8816 | aarch64_expand_vector_init (target, copy); | |
8817 | ||
8818 | /* Insert variable. */ | |
8819 | x = copy_to_mode_reg (inner_mode, XVECEXP (vals, 0, one_var)); | |
8820 | icode = optab_handler (vec_set_optab, mode); | |
8821 | gcc_assert (icode != CODE_FOR_nothing); | |
8822 | emit_insn (GEN_FCN (icode) (target, x, index)); | |
8823 | return; | |
8824 | } | |
8825 | ||
8826 | /* Construct the vector in memory one field at a time | |
8827 | and load the whole vector. */ | |
8828 | mem = assign_stack_temp (mode, GET_MODE_SIZE (mode)); | |
8829 | for (i = 0; i < n_elts; i++) | |
8830 | emit_move_insn (adjust_address_nv (mem, inner_mode, | |
8831 | i * GET_MODE_SIZE (inner_mode)), | |
8832 | XVECEXP (vals, 0, i)); | |
8833 | emit_move_insn (target, mem); | |
8834 | ||
8835 | } | |
8836 | ||
43e9d192 | 8837 | static unsigned HOST_WIDE_INT |
ef4bddc2 | 8838 | aarch64_shift_truncation_mask (machine_mode mode) |
43e9d192 IB |
8839 | { |
8840 | return | |
8841 | (aarch64_vector_mode_supported_p (mode) | |
8842 | || aarch64_vect_struct_mode_p (mode)) ? 0 : (GET_MODE_BITSIZE (mode) - 1); | |
8843 | } | |
8844 | ||
8845 | #ifndef TLS_SECTION_ASM_FLAG | |
8846 | #define TLS_SECTION_ASM_FLAG 'T' | |
8847 | #endif | |
8848 | ||
8849 | void | |
8850 | aarch64_elf_asm_named_section (const char *name, unsigned int flags, | |
8851 | tree decl ATTRIBUTE_UNUSED) | |
8852 | { | |
8853 | char flagchars[10], *f = flagchars; | |
8854 | ||
8855 | /* If we have already declared this section, we can use an | |
8856 | abbreviated form to switch back to it -- unless this section is | |
8857 | part of a COMDAT groups, in which case GAS requires the full | |
8858 | declaration every time. */ | |
8859 | if (!(HAVE_COMDAT_GROUP && (flags & SECTION_LINKONCE)) | |
8860 | && (flags & SECTION_DECLARED)) | |
8861 | { | |
8862 | fprintf (asm_out_file, "\t.section\t%s\n", name); | |
8863 | return; | |
8864 | } | |
8865 | ||
8866 | if (!(flags & SECTION_DEBUG)) | |
8867 | *f++ = 'a'; | |
8868 | if (flags & SECTION_WRITE) | |
8869 | *f++ = 'w'; | |
8870 | if (flags & SECTION_CODE) | |
8871 | *f++ = 'x'; | |
8872 | if (flags & SECTION_SMALL) | |
8873 | *f++ = 's'; | |
8874 | if (flags & SECTION_MERGE) | |
8875 | *f++ = 'M'; | |
8876 | if (flags & SECTION_STRINGS) | |
8877 | *f++ = 'S'; | |
8878 | if (flags & SECTION_TLS) | |
8879 | *f++ = TLS_SECTION_ASM_FLAG; | |
8880 | if (HAVE_COMDAT_GROUP && (flags & SECTION_LINKONCE)) | |
8881 | *f++ = 'G'; | |
8882 | *f = '\0'; | |
8883 | ||
8884 | fprintf (asm_out_file, "\t.section\t%s,\"%s\"", name, flagchars); | |
8885 | ||
8886 | if (!(flags & SECTION_NOTYPE)) | |
8887 | { | |
8888 | const char *type; | |
8889 | const char *format; | |
8890 | ||
8891 | if (flags & SECTION_BSS) | |
8892 | type = "nobits"; | |
8893 | else | |
8894 | type = "progbits"; | |
8895 | ||
8896 | #ifdef TYPE_OPERAND_FMT | |
8897 | format = "," TYPE_OPERAND_FMT; | |
8898 | #else | |
8899 | format = ",@%s"; | |
8900 | #endif | |
8901 | ||
8902 | fprintf (asm_out_file, format, type); | |
8903 | ||
8904 | if (flags & SECTION_ENTSIZE) | |
8905 | fprintf (asm_out_file, ",%d", flags & SECTION_ENTSIZE); | |
8906 | if (HAVE_COMDAT_GROUP && (flags & SECTION_LINKONCE)) | |
8907 | { | |
8908 | if (TREE_CODE (decl) == IDENTIFIER_NODE) | |
8909 | fprintf (asm_out_file, ",%s,comdat", IDENTIFIER_POINTER (decl)); | |
8910 | else | |
8911 | fprintf (asm_out_file, ",%s,comdat", | |
8912 | IDENTIFIER_POINTER (DECL_COMDAT_GROUP (decl))); | |
8913 | } | |
8914 | } | |
8915 | ||
8916 | putc ('\n', asm_out_file); | |
8917 | } | |
8918 | ||
8919 | /* Select a format to encode pointers in exception handling data. */ | |
8920 | int | |
8921 | aarch64_asm_preferred_eh_data_format (int code ATTRIBUTE_UNUSED, int global) | |
8922 | { | |
8923 | int type; | |
8924 | switch (aarch64_cmodel) | |
8925 | { | |
8926 | case AARCH64_CMODEL_TINY: | |
8927 | case AARCH64_CMODEL_TINY_PIC: | |
8928 | case AARCH64_CMODEL_SMALL: | |
8929 | case AARCH64_CMODEL_SMALL_PIC: | |
8930 | /* text+got+data < 4Gb. 4-byte signed relocs are sufficient | |
8931 | for everything. */ | |
8932 | type = DW_EH_PE_sdata4; | |
8933 | break; | |
8934 | default: | |
8935 | /* No assumptions here. 8-byte relocs required. */ | |
8936 | type = DW_EH_PE_sdata8; | |
8937 | break; | |
8938 | } | |
8939 | return (global ? DW_EH_PE_indirect : 0) | DW_EH_PE_pcrel | type; | |
8940 | } | |
8941 | ||
0462169c SN |
8942 | /* Emit load exclusive. */ |
8943 | ||
8944 | static void | |
ef4bddc2 | 8945 | aarch64_emit_load_exclusive (machine_mode mode, rtx rval, |
0462169c SN |
8946 | rtx mem, rtx model_rtx) |
8947 | { | |
8948 | rtx (*gen) (rtx, rtx, rtx); | |
8949 | ||
8950 | switch (mode) | |
8951 | { | |
8952 | case QImode: gen = gen_aarch64_load_exclusiveqi; break; | |
8953 | case HImode: gen = gen_aarch64_load_exclusivehi; break; | |
8954 | case SImode: gen = gen_aarch64_load_exclusivesi; break; | |
8955 | case DImode: gen = gen_aarch64_load_exclusivedi; break; | |
8956 | default: | |
8957 | gcc_unreachable (); | |
8958 | } | |
8959 | ||
8960 | emit_insn (gen (rval, mem, model_rtx)); | |
8961 | } | |
8962 | ||
8963 | /* Emit store exclusive. */ | |
8964 | ||
8965 | static void | |
ef4bddc2 | 8966 | aarch64_emit_store_exclusive (machine_mode mode, rtx bval, |
0462169c SN |
8967 | rtx rval, rtx mem, rtx model_rtx) |
8968 | { | |
8969 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
8970 | ||
8971 | switch (mode) | |
8972 | { | |
8973 | case QImode: gen = gen_aarch64_store_exclusiveqi; break; | |
8974 | case HImode: gen = gen_aarch64_store_exclusivehi; break; | |
8975 | case SImode: gen = gen_aarch64_store_exclusivesi; break; | |
8976 | case DImode: gen = gen_aarch64_store_exclusivedi; break; | |
8977 | default: | |
8978 | gcc_unreachable (); | |
8979 | } | |
8980 | ||
8981 | emit_insn (gen (bval, rval, mem, model_rtx)); | |
8982 | } | |
8983 | ||
8984 | /* Mark the previous jump instruction as unlikely. */ | |
8985 | ||
8986 | static void | |
8987 | aarch64_emit_unlikely_jump (rtx insn) | |
8988 | { | |
e5af9ddd | 8989 | int very_unlikely = REG_BR_PROB_BASE / 100 - 1; |
0462169c SN |
8990 | |
8991 | insn = emit_jump_insn (insn); | |
e5af9ddd | 8992 | add_int_reg_note (insn, REG_BR_PROB, very_unlikely); |
0462169c SN |
8993 | } |
8994 | ||
8995 | /* Expand a compare and swap pattern. */ | |
8996 | ||
8997 | void | |
8998 | aarch64_expand_compare_and_swap (rtx operands[]) | |
8999 | { | |
9000 | rtx bval, rval, mem, oldval, newval, is_weak, mod_s, mod_f, x; | |
ef4bddc2 | 9001 | machine_mode mode, cmp_mode; |
0462169c SN |
9002 | rtx (*gen) (rtx, rtx, rtx, rtx, rtx, rtx, rtx); |
9003 | ||
9004 | bval = operands[0]; | |
9005 | rval = operands[1]; | |
9006 | mem = operands[2]; | |
9007 | oldval = operands[3]; | |
9008 | newval = operands[4]; | |
9009 | is_weak = operands[5]; | |
9010 | mod_s = operands[6]; | |
9011 | mod_f = operands[7]; | |
9012 | mode = GET_MODE (mem); | |
9013 | cmp_mode = mode; | |
9014 | ||
9015 | /* Normally the succ memory model must be stronger than fail, but in the | |
9016 | unlikely event of fail being ACQUIRE and succ being RELEASE we need to | |
9017 | promote succ to ACQ_REL so that we don't lose the acquire semantics. */ | |
9018 | ||
9019 | if (INTVAL (mod_f) == MEMMODEL_ACQUIRE | |
9020 | && INTVAL (mod_s) == MEMMODEL_RELEASE) | |
9021 | mod_s = GEN_INT (MEMMODEL_ACQ_REL); | |
9022 | ||
9023 | switch (mode) | |
9024 | { | |
9025 | case QImode: | |
9026 | case HImode: | |
9027 | /* For short modes, we're going to perform the comparison in SImode, | |
9028 | so do the zero-extension now. */ | |
9029 | cmp_mode = SImode; | |
9030 | rval = gen_reg_rtx (SImode); | |
9031 | oldval = convert_modes (SImode, mode, oldval, true); | |
9032 | /* Fall through. */ | |
9033 | ||
9034 | case SImode: | |
9035 | case DImode: | |
9036 | /* Force the value into a register if needed. */ | |
9037 | if (!aarch64_plus_operand (oldval, mode)) | |
9038 | oldval = force_reg (cmp_mode, oldval); | |
9039 | break; | |
9040 | ||
9041 | default: | |
9042 | gcc_unreachable (); | |
9043 | } | |
9044 | ||
9045 | switch (mode) | |
9046 | { | |
9047 | case QImode: gen = gen_atomic_compare_and_swapqi_1; break; | |
9048 | case HImode: gen = gen_atomic_compare_and_swaphi_1; break; | |
9049 | case SImode: gen = gen_atomic_compare_and_swapsi_1; break; | |
9050 | case DImode: gen = gen_atomic_compare_and_swapdi_1; break; | |
9051 | default: | |
9052 | gcc_unreachable (); | |
9053 | } | |
9054 | ||
9055 | emit_insn (gen (rval, mem, oldval, newval, is_weak, mod_s, mod_f)); | |
9056 | ||
9057 | if (mode == QImode || mode == HImode) | |
9058 | emit_move_insn (operands[1], gen_lowpart (mode, rval)); | |
9059 | ||
9060 | x = gen_rtx_REG (CCmode, CC_REGNUM); | |
9061 | x = gen_rtx_EQ (SImode, x, const0_rtx); | |
9062 | emit_insn (gen_rtx_SET (VOIDmode, bval, x)); | |
9063 | } | |
9064 | ||
9065 | /* Split a compare and swap pattern. */ | |
9066 | ||
9067 | void | |
9068 | aarch64_split_compare_and_swap (rtx operands[]) | |
9069 | { | |
9070 | rtx rval, mem, oldval, newval, scratch; | |
ef4bddc2 | 9071 | machine_mode mode; |
0462169c | 9072 | bool is_weak; |
5d8a22a5 DM |
9073 | rtx_code_label *label1, *label2; |
9074 | rtx x, cond; | |
0462169c SN |
9075 | |
9076 | rval = operands[0]; | |
9077 | mem = operands[1]; | |
9078 | oldval = operands[2]; | |
9079 | newval = operands[3]; | |
9080 | is_weak = (operands[4] != const0_rtx); | |
0462169c SN |
9081 | scratch = operands[7]; |
9082 | mode = GET_MODE (mem); | |
9083 | ||
5d8a22a5 | 9084 | label1 = NULL; |
0462169c SN |
9085 | if (!is_weak) |
9086 | { | |
9087 | label1 = gen_label_rtx (); | |
9088 | emit_label (label1); | |
9089 | } | |
9090 | label2 = gen_label_rtx (); | |
9091 | ||
9092 | aarch64_emit_load_exclusive (mode, rval, mem, operands[5]); | |
9093 | ||
9094 | cond = aarch64_gen_compare_reg (NE, rval, oldval); | |
9095 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
9096 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
9097 | gen_rtx_LABEL_REF (Pmode, label2), pc_rtx); | |
9098 | aarch64_emit_unlikely_jump (gen_rtx_SET (VOIDmode, pc_rtx, x)); | |
9099 | ||
9100 | aarch64_emit_store_exclusive (mode, scratch, mem, newval, operands[5]); | |
9101 | ||
9102 | if (!is_weak) | |
9103 | { | |
9104 | x = gen_rtx_NE (VOIDmode, scratch, const0_rtx); | |
9105 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
9106 | gen_rtx_LABEL_REF (Pmode, label1), pc_rtx); | |
9107 | aarch64_emit_unlikely_jump (gen_rtx_SET (VOIDmode, pc_rtx, x)); | |
9108 | } | |
9109 | else | |
9110 | { | |
9111 | cond = gen_rtx_REG (CCmode, CC_REGNUM); | |
9112 | x = gen_rtx_COMPARE (CCmode, scratch, const0_rtx); | |
9113 | emit_insn (gen_rtx_SET (VOIDmode, cond, x)); | |
9114 | } | |
9115 | ||
9116 | emit_label (label2); | |
9117 | } | |
9118 | ||
9119 | /* Split an atomic operation. */ | |
9120 | ||
9121 | void | |
9122 | aarch64_split_atomic_op (enum rtx_code code, rtx old_out, rtx new_out, rtx mem, | |
9123 | rtx value, rtx model_rtx, rtx cond) | |
9124 | { | |
ef4bddc2 RS |
9125 | machine_mode mode = GET_MODE (mem); |
9126 | machine_mode wmode = (mode == DImode ? DImode : SImode); | |
5d8a22a5 DM |
9127 | rtx_code_label *label; |
9128 | rtx x; | |
0462169c SN |
9129 | |
9130 | label = gen_label_rtx (); | |
9131 | emit_label (label); | |
9132 | ||
9133 | if (new_out) | |
9134 | new_out = gen_lowpart (wmode, new_out); | |
9135 | if (old_out) | |
9136 | old_out = gen_lowpart (wmode, old_out); | |
9137 | else | |
9138 | old_out = new_out; | |
9139 | value = simplify_gen_subreg (wmode, value, mode, 0); | |
9140 | ||
9141 | aarch64_emit_load_exclusive (mode, old_out, mem, model_rtx); | |
9142 | ||
9143 | switch (code) | |
9144 | { | |
9145 | case SET: | |
9146 | new_out = value; | |
9147 | break; | |
9148 | ||
9149 | case NOT: | |
9150 | x = gen_rtx_AND (wmode, old_out, value); | |
9151 | emit_insn (gen_rtx_SET (VOIDmode, new_out, x)); | |
9152 | x = gen_rtx_NOT (wmode, new_out); | |
9153 | emit_insn (gen_rtx_SET (VOIDmode, new_out, x)); | |
9154 | break; | |
9155 | ||
9156 | case MINUS: | |
9157 | if (CONST_INT_P (value)) | |
9158 | { | |
9159 | value = GEN_INT (-INTVAL (value)); | |
9160 | code = PLUS; | |
9161 | } | |
9162 | /* Fall through. */ | |
9163 | ||
9164 | default: | |
9165 | x = gen_rtx_fmt_ee (code, wmode, old_out, value); | |
9166 | emit_insn (gen_rtx_SET (VOIDmode, new_out, x)); | |
9167 | break; | |
9168 | } | |
9169 | ||
9170 | aarch64_emit_store_exclusive (mode, cond, mem, | |
9171 | gen_lowpart (mode, new_out), model_rtx); | |
9172 | ||
9173 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
9174 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
9175 | gen_rtx_LABEL_REF (Pmode, label), pc_rtx); | |
9176 | aarch64_emit_unlikely_jump (gen_rtx_SET (VOIDmode, pc_rtx, x)); | |
9177 | } | |
9178 | ||
95ca411e YZ |
9179 | static void |
9180 | aarch64_print_extension (void) | |
9181 | { | |
9182 | const struct aarch64_option_extension *opt = NULL; | |
9183 | ||
9184 | for (opt = all_extensions; opt->name != NULL; opt++) | |
9185 | if ((aarch64_isa_flags & opt->flags_on) == opt->flags_on) | |
9186 | asm_fprintf (asm_out_file, "+%s", opt->name); | |
9187 | ||
9188 | asm_fprintf (asm_out_file, "\n"); | |
9189 | } | |
9190 | ||
43e9d192 IB |
9191 | static void |
9192 | aarch64_start_file (void) | |
9193 | { | |
9194 | if (selected_arch) | |
95ca411e YZ |
9195 | { |
9196 | asm_fprintf (asm_out_file, "\t.arch %s", selected_arch->name); | |
9197 | aarch64_print_extension (); | |
9198 | } | |
43e9d192 | 9199 | else if (selected_cpu) |
95ca411e | 9200 | { |
682287fb JG |
9201 | const char *truncated_name |
9202 | = aarch64_rewrite_selected_cpu (selected_cpu->name); | |
9203 | asm_fprintf (asm_out_file, "\t.cpu %s", truncated_name); | |
95ca411e YZ |
9204 | aarch64_print_extension (); |
9205 | } | |
43e9d192 IB |
9206 | default_file_start(); |
9207 | } | |
9208 | ||
9209 | /* Target hook for c_mode_for_suffix. */ | |
ef4bddc2 | 9210 | static machine_mode |
43e9d192 IB |
9211 | aarch64_c_mode_for_suffix (char suffix) |
9212 | { | |
9213 | if (suffix == 'q') | |
9214 | return TFmode; | |
9215 | ||
9216 | return VOIDmode; | |
9217 | } | |
9218 | ||
3520f7cc JG |
9219 | /* We can only represent floating point constants which will fit in |
9220 | "quarter-precision" values. These values are characterised by | |
9221 | a sign bit, a 4-bit mantissa and a 3-bit exponent. And are given | |
9222 | by: | |
9223 | ||
9224 | (-1)^s * (n/16) * 2^r | |
9225 | ||
9226 | Where: | |
9227 | 's' is the sign bit. | |
9228 | 'n' is an integer in the range 16 <= n <= 31. | |
9229 | 'r' is an integer in the range -3 <= r <= 4. */ | |
9230 | ||
9231 | /* Return true iff X can be represented by a quarter-precision | |
9232 | floating point immediate operand X. Note, we cannot represent 0.0. */ | |
9233 | bool | |
9234 | aarch64_float_const_representable_p (rtx x) | |
9235 | { | |
9236 | /* This represents our current view of how many bits | |
9237 | make up the mantissa. */ | |
9238 | int point_pos = 2 * HOST_BITS_PER_WIDE_INT - 1; | |
ba96cdfb | 9239 | int exponent; |
3520f7cc | 9240 | unsigned HOST_WIDE_INT mantissa, mask; |
3520f7cc | 9241 | REAL_VALUE_TYPE r, m; |
807e902e | 9242 | bool fail; |
3520f7cc JG |
9243 | |
9244 | if (!CONST_DOUBLE_P (x)) | |
9245 | return false; | |
9246 | ||
94bfa2da TV |
9247 | if (GET_MODE (x) == VOIDmode) |
9248 | return false; | |
9249 | ||
3520f7cc JG |
9250 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); |
9251 | ||
9252 | /* We cannot represent infinities, NaNs or +/-zero. We won't | |
9253 | know if we have +zero until we analyse the mantissa, but we | |
9254 | can reject the other invalid values. */ | |
9255 | if (REAL_VALUE_ISINF (r) || REAL_VALUE_ISNAN (r) | |
9256 | || REAL_VALUE_MINUS_ZERO (r)) | |
9257 | return false; | |
9258 | ||
ba96cdfb | 9259 | /* Extract exponent. */ |
3520f7cc JG |
9260 | r = real_value_abs (&r); |
9261 | exponent = REAL_EXP (&r); | |
9262 | ||
9263 | /* For the mantissa, we expand into two HOST_WIDE_INTS, apart from the | |
9264 | highest (sign) bit, with a fixed binary point at bit point_pos. | |
9265 | m1 holds the low part of the mantissa, m2 the high part. | |
9266 | WARNING: If we ever have a representation using more than 2 * H_W_I - 1 | |
9267 | bits for the mantissa, this can fail (low bits will be lost). */ | |
9268 | real_ldexp (&m, &r, point_pos - exponent); | |
807e902e | 9269 | wide_int w = real_to_integer (&m, &fail, HOST_BITS_PER_WIDE_INT * 2); |
3520f7cc JG |
9270 | |
9271 | /* If the low part of the mantissa has bits set we cannot represent | |
9272 | the value. */ | |
807e902e | 9273 | if (w.elt (0) != 0) |
3520f7cc JG |
9274 | return false; |
9275 | /* We have rejected the lower HOST_WIDE_INT, so update our | |
9276 | understanding of how many bits lie in the mantissa and | |
9277 | look only at the high HOST_WIDE_INT. */ | |
807e902e | 9278 | mantissa = w.elt (1); |
3520f7cc JG |
9279 | point_pos -= HOST_BITS_PER_WIDE_INT; |
9280 | ||
9281 | /* We can only represent values with a mantissa of the form 1.xxxx. */ | |
9282 | mask = ((unsigned HOST_WIDE_INT)1 << (point_pos - 5)) - 1; | |
9283 | if ((mantissa & mask) != 0) | |
9284 | return false; | |
9285 | ||
9286 | /* Having filtered unrepresentable values, we may now remove all | |
9287 | but the highest 5 bits. */ | |
9288 | mantissa >>= point_pos - 5; | |
9289 | ||
9290 | /* We cannot represent the value 0.0, so reject it. This is handled | |
9291 | elsewhere. */ | |
9292 | if (mantissa == 0) | |
9293 | return false; | |
9294 | ||
9295 | /* Then, as bit 4 is always set, we can mask it off, leaving | |
9296 | the mantissa in the range [0, 15]. */ | |
9297 | mantissa &= ~(1 << 4); | |
9298 | gcc_assert (mantissa <= 15); | |
9299 | ||
9300 | /* GCC internally does not use IEEE754-like encoding (where normalized | |
9301 | significands are in the range [1, 2). GCC uses [0.5, 1) (see real.c). | |
9302 | Our mantissa values are shifted 4 places to the left relative to | |
9303 | normalized IEEE754 so we must modify the exponent returned by REAL_EXP | |
9304 | by 5 places to correct for GCC's representation. */ | |
9305 | exponent = 5 - exponent; | |
9306 | ||
9307 | return (exponent >= 0 && exponent <= 7); | |
9308 | } | |
9309 | ||
9310 | char* | |
81c2dfb9 | 9311 | aarch64_output_simd_mov_immediate (rtx const_vector, |
ef4bddc2 | 9312 | machine_mode mode, |
3520f7cc JG |
9313 | unsigned width) |
9314 | { | |
3ea63f60 | 9315 | bool is_valid; |
3520f7cc | 9316 | static char templ[40]; |
3520f7cc | 9317 | const char *mnemonic; |
e4f0f84d | 9318 | const char *shift_op; |
3520f7cc | 9319 | unsigned int lane_count = 0; |
81c2dfb9 | 9320 | char element_char; |
3520f7cc | 9321 | |
e4f0f84d | 9322 | struct simd_immediate_info info = { NULL_RTX, 0, 0, false, false }; |
48063b9d IB |
9323 | |
9324 | /* This will return true to show const_vector is legal for use as either | |
9325 | a AdvSIMD MOVI instruction (or, implicitly, MVNI) immediate. It will | |
9326 | also update INFO to show how the immediate should be generated. */ | |
81c2dfb9 | 9327 | is_valid = aarch64_simd_valid_immediate (const_vector, mode, false, &info); |
3520f7cc JG |
9328 | gcc_assert (is_valid); |
9329 | ||
81c2dfb9 | 9330 | element_char = sizetochar (info.element_width); |
48063b9d IB |
9331 | lane_count = width / info.element_width; |
9332 | ||
3520f7cc JG |
9333 | mode = GET_MODE_INNER (mode); |
9334 | if (mode == SFmode || mode == DFmode) | |
9335 | { | |
48063b9d IB |
9336 | gcc_assert (info.shift == 0 && ! info.mvn); |
9337 | if (aarch64_float_const_zero_rtx_p (info.value)) | |
9338 | info.value = GEN_INT (0); | |
9339 | else | |
9340 | { | |
9341 | #define buf_size 20 | |
9342 | REAL_VALUE_TYPE r; | |
9343 | REAL_VALUE_FROM_CONST_DOUBLE (r, info.value); | |
9344 | char float_buf[buf_size] = {'\0'}; | |
9345 | real_to_decimal_for_mode (float_buf, &r, buf_size, buf_size, 1, mode); | |
9346 | #undef buf_size | |
9347 | ||
9348 | if (lane_count == 1) | |
9349 | snprintf (templ, sizeof (templ), "fmov\t%%d0, %s", float_buf); | |
9350 | else | |
9351 | snprintf (templ, sizeof (templ), "fmov\t%%0.%d%c, %s", | |
81c2dfb9 | 9352 | lane_count, element_char, float_buf); |
48063b9d IB |
9353 | return templ; |
9354 | } | |
3520f7cc | 9355 | } |
3520f7cc | 9356 | |
48063b9d | 9357 | mnemonic = info.mvn ? "mvni" : "movi"; |
e4f0f84d | 9358 | shift_op = info.msl ? "msl" : "lsl"; |
3520f7cc JG |
9359 | |
9360 | if (lane_count == 1) | |
48063b9d IB |
9361 | snprintf (templ, sizeof (templ), "%s\t%%d0, " HOST_WIDE_INT_PRINT_HEX, |
9362 | mnemonic, UINTVAL (info.value)); | |
9363 | else if (info.shift) | |
9364 | snprintf (templ, sizeof (templ), "%s\t%%0.%d%c, " HOST_WIDE_INT_PRINT_HEX | |
e4f0f84d TB |
9365 | ", %s %d", mnemonic, lane_count, element_char, |
9366 | UINTVAL (info.value), shift_op, info.shift); | |
3520f7cc | 9367 | else |
48063b9d | 9368 | snprintf (templ, sizeof (templ), "%s\t%%0.%d%c, " HOST_WIDE_INT_PRINT_HEX, |
81c2dfb9 | 9369 | mnemonic, lane_count, element_char, UINTVAL (info.value)); |
3520f7cc JG |
9370 | return templ; |
9371 | } | |
9372 | ||
b7342d25 IB |
9373 | char* |
9374 | aarch64_output_scalar_simd_mov_immediate (rtx immediate, | |
ef4bddc2 | 9375 | machine_mode mode) |
b7342d25 | 9376 | { |
ef4bddc2 | 9377 | machine_mode vmode; |
b7342d25 IB |
9378 | |
9379 | gcc_assert (!VECTOR_MODE_P (mode)); | |
9380 | vmode = aarch64_simd_container_mode (mode, 64); | |
9381 | rtx v_op = aarch64_simd_gen_const_vector_dup (vmode, INTVAL (immediate)); | |
9382 | return aarch64_output_simd_mov_immediate (v_op, vmode, 64); | |
9383 | } | |
9384 | ||
88b08073 JG |
9385 | /* Split operands into moves from op[1] + op[2] into op[0]. */ |
9386 | ||
9387 | void | |
9388 | aarch64_split_combinev16qi (rtx operands[3]) | |
9389 | { | |
9390 | unsigned int dest = REGNO (operands[0]); | |
9391 | unsigned int src1 = REGNO (operands[1]); | |
9392 | unsigned int src2 = REGNO (operands[2]); | |
ef4bddc2 | 9393 | machine_mode halfmode = GET_MODE (operands[1]); |
88b08073 JG |
9394 | unsigned int halfregs = HARD_REGNO_NREGS (src1, halfmode); |
9395 | rtx destlo, desthi; | |
9396 | ||
9397 | gcc_assert (halfmode == V16QImode); | |
9398 | ||
9399 | if (src1 == dest && src2 == dest + halfregs) | |
9400 | { | |
9401 | /* No-op move. Can't split to nothing; emit something. */ | |
9402 | emit_note (NOTE_INSN_DELETED); | |
9403 | return; | |
9404 | } | |
9405 | ||
9406 | /* Preserve register attributes for variable tracking. */ | |
9407 | destlo = gen_rtx_REG_offset (operands[0], halfmode, dest, 0); | |
9408 | desthi = gen_rtx_REG_offset (operands[0], halfmode, dest + halfregs, | |
9409 | GET_MODE_SIZE (halfmode)); | |
9410 | ||
9411 | /* Special case of reversed high/low parts. */ | |
9412 | if (reg_overlap_mentioned_p (operands[2], destlo) | |
9413 | && reg_overlap_mentioned_p (operands[1], desthi)) | |
9414 | { | |
9415 | emit_insn (gen_xorv16qi3 (operands[1], operands[1], operands[2])); | |
9416 | emit_insn (gen_xorv16qi3 (operands[2], operands[1], operands[2])); | |
9417 | emit_insn (gen_xorv16qi3 (operands[1], operands[1], operands[2])); | |
9418 | } | |
9419 | else if (!reg_overlap_mentioned_p (operands[2], destlo)) | |
9420 | { | |
9421 | /* Try to avoid unnecessary moves if part of the result | |
9422 | is in the right place already. */ | |
9423 | if (src1 != dest) | |
9424 | emit_move_insn (destlo, operands[1]); | |
9425 | if (src2 != dest + halfregs) | |
9426 | emit_move_insn (desthi, operands[2]); | |
9427 | } | |
9428 | else | |
9429 | { | |
9430 | if (src2 != dest + halfregs) | |
9431 | emit_move_insn (desthi, operands[2]); | |
9432 | if (src1 != dest) | |
9433 | emit_move_insn (destlo, operands[1]); | |
9434 | } | |
9435 | } | |
9436 | ||
9437 | /* vec_perm support. */ | |
9438 | ||
9439 | #define MAX_VECT_LEN 16 | |
9440 | ||
9441 | struct expand_vec_perm_d | |
9442 | { | |
9443 | rtx target, op0, op1; | |
9444 | unsigned char perm[MAX_VECT_LEN]; | |
ef4bddc2 | 9445 | machine_mode vmode; |
88b08073 JG |
9446 | unsigned char nelt; |
9447 | bool one_vector_p; | |
9448 | bool testing_p; | |
9449 | }; | |
9450 | ||
9451 | /* Generate a variable permutation. */ | |
9452 | ||
9453 | static void | |
9454 | aarch64_expand_vec_perm_1 (rtx target, rtx op0, rtx op1, rtx sel) | |
9455 | { | |
ef4bddc2 | 9456 | machine_mode vmode = GET_MODE (target); |
88b08073 JG |
9457 | bool one_vector_p = rtx_equal_p (op0, op1); |
9458 | ||
9459 | gcc_checking_assert (vmode == V8QImode || vmode == V16QImode); | |
9460 | gcc_checking_assert (GET_MODE (op0) == vmode); | |
9461 | gcc_checking_assert (GET_MODE (op1) == vmode); | |
9462 | gcc_checking_assert (GET_MODE (sel) == vmode); | |
9463 | gcc_checking_assert (TARGET_SIMD); | |
9464 | ||
9465 | if (one_vector_p) | |
9466 | { | |
9467 | if (vmode == V8QImode) | |
9468 | { | |
9469 | /* Expand the argument to a V16QI mode by duplicating it. */ | |
9470 | rtx pair = gen_reg_rtx (V16QImode); | |
9471 | emit_insn (gen_aarch64_combinev8qi (pair, op0, op0)); | |
9472 | emit_insn (gen_aarch64_tbl1v8qi (target, pair, sel)); | |
9473 | } | |
9474 | else | |
9475 | { | |
9476 | emit_insn (gen_aarch64_tbl1v16qi (target, op0, sel)); | |
9477 | } | |
9478 | } | |
9479 | else | |
9480 | { | |
9481 | rtx pair; | |
9482 | ||
9483 | if (vmode == V8QImode) | |
9484 | { | |
9485 | pair = gen_reg_rtx (V16QImode); | |
9486 | emit_insn (gen_aarch64_combinev8qi (pair, op0, op1)); | |
9487 | emit_insn (gen_aarch64_tbl1v8qi (target, pair, sel)); | |
9488 | } | |
9489 | else | |
9490 | { | |
9491 | pair = gen_reg_rtx (OImode); | |
9492 | emit_insn (gen_aarch64_combinev16qi (pair, op0, op1)); | |
9493 | emit_insn (gen_aarch64_tbl2v16qi (target, pair, sel)); | |
9494 | } | |
9495 | } | |
9496 | } | |
9497 | ||
9498 | void | |
9499 | aarch64_expand_vec_perm (rtx target, rtx op0, rtx op1, rtx sel) | |
9500 | { | |
ef4bddc2 | 9501 | machine_mode vmode = GET_MODE (target); |
c9d1a16a | 9502 | unsigned int nelt = GET_MODE_NUNITS (vmode); |
88b08073 | 9503 | bool one_vector_p = rtx_equal_p (op0, op1); |
f7c4e5b8 | 9504 | rtx mask; |
88b08073 JG |
9505 | |
9506 | /* The TBL instruction does not use a modulo index, so we must take care | |
9507 | of that ourselves. */ | |
f7c4e5b8 AL |
9508 | mask = aarch64_simd_gen_const_vector_dup (vmode, |
9509 | one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
88b08073 JG |
9510 | sel = expand_simple_binop (vmode, AND, sel, mask, NULL, 0, OPTAB_LIB_WIDEN); |
9511 | ||
f7c4e5b8 AL |
9512 | /* For big-endian, we also need to reverse the index within the vector |
9513 | (but not which vector). */ | |
9514 | if (BYTES_BIG_ENDIAN) | |
9515 | { | |
9516 | /* If one_vector_p, mask is a vector of (nelt - 1)'s already. */ | |
9517 | if (!one_vector_p) | |
9518 | mask = aarch64_simd_gen_const_vector_dup (vmode, nelt - 1); | |
9519 | sel = expand_simple_binop (vmode, XOR, sel, mask, | |
9520 | NULL, 0, OPTAB_LIB_WIDEN); | |
9521 | } | |
88b08073 JG |
9522 | aarch64_expand_vec_perm_1 (target, op0, op1, sel); |
9523 | } | |
9524 | ||
cc4d934f JG |
9525 | /* Recognize patterns suitable for the TRN instructions. */ |
9526 | static bool | |
9527 | aarch64_evpc_trn (struct expand_vec_perm_d *d) | |
9528 | { | |
9529 | unsigned int i, odd, mask, nelt = d->nelt; | |
9530 | rtx out, in0, in1, x; | |
9531 | rtx (*gen) (rtx, rtx, rtx); | |
ef4bddc2 | 9532 | machine_mode vmode = d->vmode; |
cc4d934f JG |
9533 | |
9534 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
9535 | return false; | |
9536 | ||
9537 | /* Note that these are little-endian tests. | |
9538 | We correct for big-endian later. */ | |
9539 | if (d->perm[0] == 0) | |
9540 | odd = 0; | |
9541 | else if (d->perm[0] == 1) | |
9542 | odd = 1; | |
9543 | else | |
9544 | return false; | |
9545 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
9546 | ||
9547 | for (i = 0; i < nelt; i += 2) | |
9548 | { | |
9549 | if (d->perm[i] != i + odd) | |
9550 | return false; | |
9551 | if (d->perm[i + 1] != ((i + nelt + odd) & mask)) | |
9552 | return false; | |
9553 | } | |
9554 | ||
9555 | /* Success! */ | |
9556 | if (d->testing_p) | |
9557 | return true; | |
9558 | ||
9559 | in0 = d->op0; | |
9560 | in1 = d->op1; | |
9561 | if (BYTES_BIG_ENDIAN) | |
9562 | { | |
9563 | x = in0, in0 = in1, in1 = x; | |
9564 | odd = !odd; | |
9565 | } | |
9566 | out = d->target; | |
9567 | ||
9568 | if (odd) | |
9569 | { | |
9570 | switch (vmode) | |
9571 | { | |
9572 | case V16QImode: gen = gen_aarch64_trn2v16qi; break; | |
9573 | case V8QImode: gen = gen_aarch64_trn2v8qi; break; | |
9574 | case V8HImode: gen = gen_aarch64_trn2v8hi; break; | |
9575 | case V4HImode: gen = gen_aarch64_trn2v4hi; break; | |
9576 | case V4SImode: gen = gen_aarch64_trn2v4si; break; | |
9577 | case V2SImode: gen = gen_aarch64_trn2v2si; break; | |
9578 | case V2DImode: gen = gen_aarch64_trn2v2di; break; | |
9579 | case V4SFmode: gen = gen_aarch64_trn2v4sf; break; | |
9580 | case V2SFmode: gen = gen_aarch64_trn2v2sf; break; | |
9581 | case V2DFmode: gen = gen_aarch64_trn2v2df; break; | |
9582 | default: | |
9583 | return false; | |
9584 | } | |
9585 | } | |
9586 | else | |
9587 | { | |
9588 | switch (vmode) | |
9589 | { | |
9590 | case V16QImode: gen = gen_aarch64_trn1v16qi; break; | |
9591 | case V8QImode: gen = gen_aarch64_trn1v8qi; break; | |
9592 | case V8HImode: gen = gen_aarch64_trn1v8hi; break; | |
9593 | case V4HImode: gen = gen_aarch64_trn1v4hi; break; | |
9594 | case V4SImode: gen = gen_aarch64_trn1v4si; break; | |
9595 | case V2SImode: gen = gen_aarch64_trn1v2si; break; | |
9596 | case V2DImode: gen = gen_aarch64_trn1v2di; break; | |
9597 | case V4SFmode: gen = gen_aarch64_trn1v4sf; break; | |
9598 | case V2SFmode: gen = gen_aarch64_trn1v2sf; break; | |
9599 | case V2DFmode: gen = gen_aarch64_trn1v2df; break; | |
9600 | default: | |
9601 | return false; | |
9602 | } | |
9603 | } | |
9604 | ||
9605 | emit_insn (gen (out, in0, in1)); | |
9606 | return true; | |
9607 | } | |
9608 | ||
9609 | /* Recognize patterns suitable for the UZP instructions. */ | |
9610 | static bool | |
9611 | aarch64_evpc_uzp (struct expand_vec_perm_d *d) | |
9612 | { | |
9613 | unsigned int i, odd, mask, nelt = d->nelt; | |
9614 | rtx out, in0, in1, x; | |
9615 | rtx (*gen) (rtx, rtx, rtx); | |
ef4bddc2 | 9616 | machine_mode vmode = d->vmode; |
cc4d934f JG |
9617 | |
9618 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
9619 | return false; | |
9620 | ||
9621 | /* Note that these are little-endian tests. | |
9622 | We correct for big-endian later. */ | |
9623 | if (d->perm[0] == 0) | |
9624 | odd = 0; | |
9625 | else if (d->perm[0] == 1) | |
9626 | odd = 1; | |
9627 | else | |
9628 | return false; | |
9629 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
9630 | ||
9631 | for (i = 0; i < nelt; i++) | |
9632 | { | |
9633 | unsigned elt = (i * 2 + odd) & mask; | |
9634 | if (d->perm[i] != elt) | |
9635 | return false; | |
9636 | } | |
9637 | ||
9638 | /* Success! */ | |
9639 | if (d->testing_p) | |
9640 | return true; | |
9641 | ||
9642 | in0 = d->op0; | |
9643 | in1 = d->op1; | |
9644 | if (BYTES_BIG_ENDIAN) | |
9645 | { | |
9646 | x = in0, in0 = in1, in1 = x; | |
9647 | odd = !odd; | |
9648 | } | |
9649 | out = d->target; | |
9650 | ||
9651 | if (odd) | |
9652 | { | |
9653 | switch (vmode) | |
9654 | { | |
9655 | case V16QImode: gen = gen_aarch64_uzp2v16qi; break; | |
9656 | case V8QImode: gen = gen_aarch64_uzp2v8qi; break; | |
9657 | case V8HImode: gen = gen_aarch64_uzp2v8hi; break; | |
9658 | case V4HImode: gen = gen_aarch64_uzp2v4hi; break; | |
9659 | case V4SImode: gen = gen_aarch64_uzp2v4si; break; | |
9660 | case V2SImode: gen = gen_aarch64_uzp2v2si; break; | |
9661 | case V2DImode: gen = gen_aarch64_uzp2v2di; break; | |
9662 | case V4SFmode: gen = gen_aarch64_uzp2v4sf; break; | |
9663 | case V2SFmode: gen = gen_aarch64_uzp2v2sf; break; | |
9664 | case V2DFmode: gen = gen_aarch64_uzp2v2df; break; | |
9665 | default: | |
9666 | return false; | |
9667 | } | |
9668 | } | |
9669 | else | |
9670 | { | |
9671 | switch (vmode) | |
9672 | { | |
9673 | case V16QImode: gen = gen_aarch64_uzp1v16qi; break; | |
9674 | case V8QImode: gen = gen_aarch64_uzp1v8qi; break; | |
9675 | case V8HImode: gen = gen_aarch64_uzp1v8hi; break; | |
9676 | case V4HImode: gen = gen_aarch64_uzp1v4hi; break; | |
9677 | case V4SImode: gen = gen_aarch64_uzp1v4si; break; | |
9678 | case V2SImode: gen = gen_aarch64_uzp1v2si; break; | |
9679 | case V2DImode: gen = gen_aarch64_uzp1v2di; break; | |
9680 | case V4SFmode: gen = gen_aarch64_uzp1v4sf; break; | |
9681 | case V2SFmode: gen = gen_aarch64_uzp1v2sf; break; | |
9682 | case V2DFmode: gen = gen_aarch64_uzp1v2df; break; | |
9683 | default: | |
9684 | return false; | |
9685 | } | |
9686 | } | |
9687 | ||
9688 | emit_insn (gen (out, in0, in1)); | |
9689 | return true; | |
9690 | } | |
9691 | ||
9692 | /* Recognize patterns suitable for the ZIP instructions. */ | |
9693 | static bool | |
9694 | aarch64_evpc_zip (struct expand_vec_perm_d *d) | |
9695 | { | |
9696 | unsigned int i, high, mask, nelt = d->nelt; | |
9697 | rtx out, in0, in1, x; | |
9698 | rtx (*gen) (rtx, rtx, rtx); | |
ef4bddc2 | 9699 | machine_mode vmode = d->vmode; |
cc4d934f JG |
9700 | |
9701 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
9702 | return false; | |
9703 | ||
9704 | /* Note that these are little-endian tests. | |
9705 | We correct for big-endian later. */ | |
9706 | high = nelt / 2; | |
9707 | if (d->perm[0] == high) | |
9708 | /* Do Nothing. */ | |
9709 | ; | |
9710 | else if (d->perm[0] == 0) | |
9711 | high = 0; | |
9712 | else | |
9713 | return false; | |
9714 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
9715 | ||
9716 | for (i = 0; i < nelt / 2; i++) | |
9717 | { | |
9718 | unsigned elt = (i + high) & mask; | |
9719 | if (d->perm[i * 2] != elt) | |
9720 | return false; | |
9721 | elt = (elt + nelt) & mask; | |
9722 | if (d->perm[i * 2 + 1] != elt) | |
9723 | return false; | |
9724 | } | |
9725 | ||
9726 | /* Success! */ | |
9727 | if (d->testing_p) | |
9728 | return true; | |
9729 | ||
9730 | in0 = d->op0; | |
9731 | in1 = d->op1; | |
9732 | if (BYTES_BIG_ENDIAN) | |
9733 | { | |
9734 | x = in0, in0 = in1, in1 = x; | |
9735 | high = !high; | |
9736 | } | |
9737 | out = d->target; | |
9738 | ||
9739 | if (high) | |
9740 | { | |
9741 | switch (vmode) | |
9742 | { | |
9743 | case V16QImode: gen = gen_aarch64_zip2v16qi; break; | |
9744 | case V8QImode: gen = gen_aarch64_zip2v8qi; break; | |
9745 | case V8HImode: gen = gen_aarch64_zip2v8hi; break; | |
9746 | case V4HImode: gen = gen_aarch64_zip2v4hi; break; | |
9747 | case V4SImode: gen = gen_aarch64_zip2v4si; break; | |
9748 | case V2SImode: gen = gen_aarch64_zip2v2si; break; | |
9749 | case V2DImode: gen = gen_aarch64_zip2v2di; break; | |
9750 | case V4SFmode: gen = gen_aarch64_zip2v4sf; break; | |
9751 | case V2SFmode: gen = gen_aarch64_zip2v2sf; break; | |
9752 | case V2DFmode: gen = gen_aarch64_zip2v2df; break; | |
9753 | default: | |
9754 | return false; | |
9755 | } | |
9756 | } | |
9757 | else | |
9758 | { | |
9759 | switch (vmode) | |
9760 | { | |
9761 | case V16QImode: gen = gen_aarch64_zip1v16qi; break; | |
9762 | case V8QImode: gen = gen_aarch64_zip1v8qi; break; | |
9763 | case V8HImode: gen = gen_aarch64_zip1v8hi; break; | |
9764 | case V4HImode: gen = gen_aarch64_zip1v4hi; break; | |
9765 | case V4SImode: gen = gen_aarch64_zip1v4si; break; | |
9766 | case V2SImode: gen = gen_aarch64_zip1v2si; break; | |
9767 | case V2DImode: gen = gen_aarch64_zip1v2di; break; | |
9768 | case V4SFmode: gen = gen_aarch64_zip1v4sf; break; | |
9769 | case V2SFmode: gen = gen_aarch64_zip1v2sf; break; | |
9770 | case V2DFmode: gen = gen_aarch64_zip1v2df; break; | |
9771 | default: | |
9772 | return false; | |
9773 | } | |
9774 | } | |
9775 | ||
9776 | emit_insn (gen (out, in0, in1)); | |
9777 | return true; | |
9778 | } | |
9779 | ||
ae0533da AL |
9780 | /* Recognize patterns for the EXT insn. */ |
9781 | ||
9782 | static bool | |
9783 | aarch64_evpc_ext (struct expand_vec_perm_d *d) | |
9784 | { | |
9785 | unsigned int i, nelt = d->nelt; | |
9786 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
9787 | rtx offset; | |
9788 | ||
9789 | unsigned int location = d->perm[0]; /* Always < nelt. */ | |
9790 | ||
9791 | /* Check if the extracted indices are increasing by one. */ | |
9792 | for (i = 1; i < nelt; i++) | |
9793 | { | |
9794 | unsigned int required = location + i; | |
9795 | if (d->one_vector_p) | |
9796 | { | |
9797 | /* We'll pass the same vector in twice, so allow indices to wrap. */ | |
9798 | required &= (nelt - 1); | |
9799 | } | |
9800 | if (d->perm[i] != required) | |
9801 | return false; | |
9802 | } | |
9803 | ||
ae0533da AL |
9804 | switch (d->vmode) |
9805 | { | |
9806 | case V16QImode: gen = gen_aarch64_extv16qi; break; | |
9807 | case V8QImode: gen = gen_aarch64_extv8qi; break; | |
9808 | case V4HImode: gen = gen_aarch64_extv4hi; break; | |
9809 | case V8HImode: gen = gen_aarch64_extv8hi; break; | |
9810 | case V2SImode: gen = gen_aarch64_extv2si; break; | |
9811 | case V4SImode: gen = gen_aarch64_extv4si; break; | |
9812 | case V2SFmode: gen = gen_aarch64_extv2sf; break; | |
9813 | case V4SFmode: gen = gen_aarch64_extv4sf; break; | |
9814 | case V2DImode: gen = gen_aarch64_extv2di; break; | |
9815 | case V2DFmode: gen = gen_aarch64_extv2df; break; | |
9816 | default: | |
9817 | return false; | |
9818 | } | |
9819 | ||
9820 | /* Success! */ | |
9821 | if (d->testing_p) | |
9822 | return true; | |
9823 | ||
b31e65bb AL |
9824 | /* The case where (location == 0) is a no-op for both big- and little-endian, |
9825 | and is removed by the mid-end at optimization levels -O1 and higher. */ | |
9826 | ||
9827 | if (BYTES_BIG_ENDIAN && (location != 0)) | |
ae0533da AL |
9828 | { |
9829 | /* After setup, we want the high elements of the first vector (stored | |
9830 | at the LSB end of the register), and the low elements of the second | |
9831 | vector (stored at the MSB end of the register). So swap. */ | |
cb5c6c29 | 9832 | std::swap (d->op0, d->op1); |
ae0533da AL |
9833 | /* location != 0 (above), so safe to assume (nelt - location) < nelt. */ |
9834 | location = nelt - location; | |
9835 | } | |
9836 | ||
9837 | offset = GEN_INT (location); | |
9838 | emit_insn (gen (d->target, d->op0, d->op1, offset)); | |
9839 | return true; | |
9840 | } | |
9841 | ||
923fcec3 AL |
9842 | /* Recognize patterns for the REV insns. */ |
9843 | ||
9844 | static bool | |
9845 | aarch64_evpc_rev (struct expand_vec_perm_d *d) | |
9846 | { | |
9847 | unsigned int i, j, diff, nelt = d->nelt; | |
9848 | rtx (*gen) (rtx, rtx); | |
9849 | ||
9850 | if (!d->one_vector_p) | |
9851 | return false; | |
9852 | ||
9853 | diff = d->perm[0]; | |
9854 | switch (diff) | |
9855 | { | |
9856 | case 7: | |
9857 | switch (d->vmode) | |
9858 | { | |
9859 | case V16QImode: gen = gen_aarch64_rev64v16qi; break; | |
9860 | case V8QImode: gen = gen_aarch64_rev64v8qi; break; | |
9861 | default: | |
9862 | return false; | |
9863 | } | |
9864 | break; | |
9865 | case 3: | |
9866 | switch (d->vmode) | |
9867 | { | |
9868 | case V16QImode: gen = gen_aarch64_rev32v16qi; break; | |
9869 | case V8QImode: gen = gen_aarch64_rev32v8qi; break; | |
9870 | case V8HImode: gen = gen_aarch64_rev64v8hi; break; | |
9871 | case V4HImode: gen = gen_aarch64_rev64v4hi; break; | |
9872 | default: | |
9873 | return false; | |
9874 | } | |
9875 | break; | |
9876 | case 1: | |
9877 | switch (d->vmode) | |
9878 | { | |
9879 | case V16QImode: gen = gen_aarch64_rev16v16qi; break; | |
9880 | case V8QImode: gen = gen_aarch64_rev16v8qi; break; | |
9881 | case V8HImode: gen = gen_aarch64_rev32v8hi; break; | |
9882 | case V4HImode: gen = gen_aarch64_rev32v4hi; break; | |
9883 | case V4SImode: gen = gen_aarch64_rev64v4si; break; | |
9884 | case V2SImode: gen = gen_aarch64_rev64v2si; break; | |
9885 | case V4SFmode: gen = gen_aarch64_rev64v4sf; break; | |
9886 | case V2SFmode: gen = gen_aarch64_rev64v2sf; break; | |
9887 | default: | |
9888 | return false; | |
9889 | } | |
9890 | break; | |
9891 | default: | |
9892 | return false; | |
9893 | } | |
9894 | ||
9895 | for (i = 0; i < nelt ; i += diff + 1) | |
9896 | for (j = 0; j <= diff; j += 1) | |
9897 | { | |
9898 | /* This is guaranteed to be true as the value of diff | |
9899 | is 7, 3, 1 and we should have enough elements in the | |
9900 | queue to generate this. Getting a vector mask with a | |
9901 | value of diff other than these values implies that | |
9902 | something is wrong by the time we get here. */ | |
9903 | gcc_assert (i + j < nelt); | |
9904 | if (d->perm[i + j] != i + diff - j) | |
9905 | return false; | |
9906 | } | |
9907 | ||
9908 | /* Success! */ | |
9909 | if (d->testing_p) | |
9910 | return true; | |
9911 | ||
9912 | emit_insn (gen (d->target, d->op0)); | |
9913 | return true; | |
9914 | } | |
9915 | ||
91bd4114 JG |
9916 | static bool |
9917 | aarch64_evpc_dup (struct expand_vec_perm_d *d) | |
9918 | { | |
9919 | rtx (*gen) (rtx, rtx, rtx); | |
9920 | rtx out = d->target; | |
9921 | rtx in0; | |
ef4bddc2 | 9922 | machine_mode vmode = d->vmode; |
91bd4114 JG |
9923 | unsigned int i, elt, nelt = d->nelt; |
9924 | rtx lane; | |
9925 | ||
91bd4114 JG |
9926 | elt = d->perm[0]; |
9927 | for (i = 1; i < nelt; i++) | |
9928 | { | |
9929 | if (elt != d->perm[i]) | |
9930 | return false; | |
9931 | } | |
9932 | ||
9933 | /* The generic preparation in aarch64_expand_vec_perm_const_1 | |
9934 | swaps the operand order and the permute indices if it finds | |
9935 | d->perm[0] to be in the second operand. Thus, we can always | |
9936 | use d->op0 and need not do any extra arithmetic to get the | |
9937 | correct lane number. */ | |
9938 | in0 = d->op0; | |
f901401e | 9939 | lane = GEN_INT (elt); /* The pattern corrects for big-endian. */ |
91bd4114 JG |
9940 | |
9941 | switch (vmode) | |
9942 | { | |
9943 | case V16QImode: gen = gen_aarch64_dup_lanev16qi; break; | |
9944 | case V8QImode: gen = gen_aarch64_dup_lanev8qi; break; | |
9945 | case V8HImode: gen = gen_aarch64_dup_lanev8hi; break; | |
9946 | case V4HImode: gen = gen_aarch64_dup_lanev4hi; break; | |
9947 | case V4SImode: gen = gen_aarch64_dup_lanev4si; break; | |
9948 | case V2SImode: gen = gen_aarch64_dup_lanev2si; break; | |
9949 | case V2DImode: gen = gen_aarch64_dup_lanev2di; break; | |
9950 | case V4SFmode: gen = gen_aarch64_dup_lanev4sf; break; | |
9951 | case V2SFmode: gen = gen_aarch64_dup_lanev2sf; break; | |
9952 | case V2DFmode: gen = gen_aarch64_dup_lanev2df; break; | |
9953 | default: | |
9954 | return false; | |
9955 | } | |
9956 | ||
9957 | emit_insn (gen (out, in0, lane)); | |
9958 | return true; | |
9959 | } | |
9960 | ||
88b08073 JG |
9961 | static bool |
9962 | aarch64_evpc_tbl (struct expand_vec_perm_d *d) | |
9963 | { | |
9964 | rtx rperm[MAX_VECT_LEN], sel; | |
ef4bddc2 | 9965 | machine_mode vmode = d->vmode; |
88b08073 JG |
9966 | unsigned int i, nelt = d->nelt; |
9967 | ||
88b08073 JG |
9968 | if (d->testing_p) |
9969 | return true; | |
9970 | ||
9971 | /* Generic code will try constant permutation twice. Once with the | |
9972 | original mode and again with the elements lowered to QImode. | |
9973 | So wait and don't do the selector expansion ourselves. */ | |
9974 | if (vmode != V8QImode && vmode != V16QImode) | |
9975 | return false; | |
9976 | ||
9977 | for (i = 0; i < nelt; ++i) | |
bbcc9c00 TB |
9978 | { |
9979 | int nunits = GET_MODE_NUNITS (vmode); | |
9980 | ||
9981 | /* If big-endian and two vectors we end up with a weird mixed-endian | |
9982 | mode on NEON. Reverse the index within each word but not the word | |
9983 | itself. */ | |
9984 | rperm[i] = GEN_INT (BYTES_BIG_ENDIAN ? d->perm[i] ^ (nunits - 1) | |
9985 | : d->perm[i]); | |
9986 | } | |
88b08073 JG |
9987 | sel = gen_rtx_CONST_VECTOR (vmode, gen_rtvec_v (nelt, rperm)); |
9988 | sel = force_reg (vmode, sel); | |
9989 | ||
9990 | aarch64_expand_vec_perm_1 (d->target, d->op0, d->op1, sel); | |
9991 | return true; | |
9992 | } | |
9993 | ||
9994 | static bool | |
9995 | aarch64_expand_vec_perm_const_1 (struct expand_vec_perm_d *d) | |
9996 | { | |
9997 | /* The pattern matching functions above are written to look for a small | |
9998 | number to begin the sequence (0, 1, N/2). If we begin with an index | |
9999 | from the second operand, we can swap the operands. */ | |
10000 | if (d->perm[0] >= d->nelt) | |
10001 | { | |
10002 | unsigned i, nelt = d->nelt; | |
88b08073 | 10003 | |
0696116a | 10004 | gcc_assert (nelt == (nelt & -nelt)); |
88b08073 | 10005 | for (i = 0; i < nelt; ++i) |
0696116a | 10006 | d->perm[i] ^= nelt; /* Keep the same index, but in the other vector. */ |
88b08073 | 10007 | |
cb5c6c29 | 10008 | std::swap (d->op0, d->op1); |
88b08073 JG |
10009 | } |
10010 | ||
10011 | if (TARGET_SIMD) | |
cc4d934f | 10012 | { |
923fcec3 AL |
10013 | if (aarch64_evpc_rev (d)) |
10014 | return true; | |
10015 | else if (aarch64_evpc_ext (d)) | |
ae0533da | 10016 | return true; |
f901401e AL |
10017 | else if (aarch64_evpc_dup (d)) |
10018 | return true; | |
ae0533da | 10019 | else if (aarch64_evpc_zip (d)) |
cc4d934f JG |
10020 | return true; |
10021 | else if (aarch64_evpc_uzp (d)) | |
10022 | return true; | |
10023 | else if (aarch64_evpc_trn (d)) | |
10024 | return true; | |
10025 | return aarch64_evpc_tbl (d); | |
10026 | } | |
88b08073 JG |
10027 | return false; |
10028 | } | |
10029 | ||
10030 | /* Expand a vec_perm_const pattern. */ | |
10031 | ||
10032 | bool | |
10033 | aarch64_expand_vec_perm_const (rtx target, rtx op0, rtx op1, rtx sel) | |
10034 | { | |
10035 | struct expand_vec_perm_d d; | |
10036 | int i, nelt, which; | |
10037 | ||
10038 | d.target = target; | |
10039 | d.op0 = op0; | |
10040 | d.op1 = op1; | |
10041 | ||
10042 | d.vmode = GET_MODE (target); | |
10043 | gcc_assert (VECTOR_MODE_P (d.vmode)); | |
10044 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
10045 | d.testing_p = false; | |
10046 | ||
10047 | for (i = which = 0; i < nelt; ++i) | |
10048 | { | |
10049 | rtx e = XVECEXP (sel, 0, i); | |
10050 | int ei = INTVAL (e) & (2 * nelt - 1); | |
10051 | which |= (ei < nelt ? 1 : 2); | |
10052 | d.perm[i] = ei; | |
10053 | } | |
10054 | ||
10055 | switch (which) | |
10056 | { | |
10057 | default: | |
10058 | gcc_unreachable (); | |
10059 | ||
10060 | case 3: | |
10061 | d.one_vector_p = false; | |
10062 | if (!rtx_equal_p (op0, op1)) | |
10063 | break; | |
10064 | ||
10065 | /* The elements of PERM do not suggest that only the first operand | |
10066 | is used, but both operands are identical. Allow easier matching | |
10067 | of the permutation by folding the permutation into the single | |
10068 | input vector. */ | |
10069 | /* Fall Through. */ | |
10070 | case 2: | |
10071 | for (i = 0; i < nelt; ++i) | |
10072 | d.perm[i] &= nelt - 1; | |
10073 | d.op0 = op1; | |
10074 | d.one_vector_p = true; | |
10075 | break; | |
10076 | ||
10077 | case 1: | |
10078 | d.op1 = op0; | |
10079 | d.one_vector_p = true; | |
10080 | break; | |
10081 | } | |
10082 | ||
10083 | return aarch64_expand_vec_perm_const_1 (&d); | |
10084 | } | |
10085 | ||
10086 | static bool | |
ef4bddc2 | 10087 | aarch64_vectorize_vec_perm_const_ok (machine_mode vmode, |
88b08073 JG |
10088 | const unsigned char *sel) |
10089 | { | |
10090 | struct expand_vec_perm_d d; | |
10091 | unsigned int i, nelt, which; | |
10092 | bool ret; | |
10093 | ||
10094 | d.vmode = vmode; | |
10095 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
10096 | d.testing_p = true; | |
10097 | memcpy (d.perm, sel, nelt); | |
10098 | ||
10099 | /* Calculate whether all elements are in one vector. */ | |
10100 | for (i = which = 0; i < nelt; ++i) | |
10101 | { | |
10102 | unsigned char e = d.perm[i]; | |
10103 | gcc_assert (e < 2 * nelt); | |
10104 | which |= (e < nelt ? 1 : 2); | |
10105 | } | |
10106 | ||
10107 | /* If all elements are from the second vector, reindex as if from the | |
10108 | first vector. */ | |
10109 | if (which == 2) | |
10110 | for (i = 0; i < nelt; ++i) | |
10111 | d.perm[i] -= nelt; | |
10112 | ||
10113 | /* Check whether the mask can be applied to a single vector. */ | |
10114 | d.one_vector_p = (which != 3); | |
10115 | ||
10116 | d.target = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 1); | |
10117 | d.op1 = d.op0 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 2); | |
10118 | if (!d.one_vector_p) | |
10119 | d.op1 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 3); | |
10120 | ||
10121 | start_sequence (); | |
10122 | ret = aarch64_expand_vec_perm_const_1 (&d); | |
10123 | end_sequence (); | |
10124 | ||
10125 | return ret; | |
10126 | } | |
10127 | ||
668046d1 DS |
10128 | rtx |
10129 | aarch64_reverse_mask (enum machine_mode mode) | |
10130 | { | |
10131 | /* We have to reverse each vector because we dont have | |
10132 | a permuted load that can reverse-load according to ABI rules. */ | |
10133 | rtx mask; | |
10134 | rtvec v = rtvec_alloc (16); | |
10135 | int i, j; | |
10136 | int nunits = GET_MODE_NUNITS (mode); | |
10137 | int usize = GET_MODE_UNIT_SIZE (mode); | |
10138 | ||
10139 | gcc_assert (BYTES_BIG_ENDIAN); | |
10140 | gcc_assert (AARCH64_VALID_SIMD_QREG_MODE (mode)); | |
10141 | ||
10142 | for (i = 0; i < nunits; i++) | |
10143 | for (j = 0; j < usize; j++) | |
10144 | RTVEC_ELT (v, i * usize + j) = GEN_INT ((i + 1) * usize - 1 - j); | |
10145 | mask = gen_rtx_CONST_VECTOR (V16QImode, v); | |
10146 | return force_reg (V16QImode, mask); | |
10147 | } | |
10148 | ||
97e1ad78 JG |
10149 | /* Implement MODES_TIEABLE_P. */ |
10150 | ||
10151 | bool | |
ef4bddc2 | 10152 | aarch64_modes_tieable_p (machine_mode mode1, machine_mode mode2) |
97e1ad78 JG |
10153 | { |
10154 | if (GET_MODE_CLASS (mode1) == GET_MODE_CLASS (mode2)) | |
10155 | return true; | |
10156 | ||
10157 | /* We specifically want to allow elements of "structure" modes to | |
10158 | be tieable to the structure. This more general condition allows | |
10159 | other rarer situations too. */ | |
10160 | if (TARGET_SIMD | |
10161 | && aarch64_vector_mode_p (mode1) | |
10162 | && aarch64_vector_mode_p (mode2)) | |
10163 | return true; | |
10164 | ||
10165 | return false; | |
10166 | } | |
10167 | ||
e2c75eea JG |
10168 | /* Return a new RTX holding the result of moving POINTER forward by |
10169 | AMOUNT bytes. */ | |
10170 | ||
10171 | static rtx | |
10172 | aarch64_move_pointer (rtx pointer, int amount) | |
10173 | { | |
10174 | rtx next = plus_constant (Pmode, XEXP (pointer, 0), amount); | |
10175 | ||
10176 | return adjust_automodify_address (pointer, GET_MODE (pointer), | |
10177 | next, amount); | |
10178 | } | |
10179 | ||
10180 | /* Return a new RTX holding the result of moving POINTER forward by the | |
10181 | size of the mode it points to. */ | |
10182 | ||
10183 | static rtx | |
10184 | aarch64_progress_pointer (rtx pointer) | |
10185 | { | |
10186 | HOST_WIDE_INT amount = GET_MODE_SIZE (GET_MODE (pointer)); | |
10187 | ||
10188 | return aarch64_move_pointer (pointer, amount); | |
10189 | } | |
10190 | ||
10191 | /* Copy one MODE sized block from SRC to DST, then progress SRC and DST by | |
10192 | MODE bytes. */ | |
10193 | ||
10194 | static void | |
10195 | aarch64_copy_one_block_and_progress_pointers (rtx *src, rtx *dst, | |
ef4bddc2 | 10196 | machine_mode mode) |
e2c75eea JG |
10197 | { |
10198 | rtx reg = gen_reg_rtx (mode); | |
10199 | ||
10200 | /* "Cast" the pointers to the correct mode. */ | |
10201 | *src = adjust_address (*src, mode, 0); | |
10202 | *dst = adjust_address (*dst, mode, 0); | |
10203 | /* Emit the memcpy. */ | |
10204 | emit_move_insn (reg, *src); | |
10205 | emit_move_insn (*dst, reg); | |
10206 | /* Move the pointers forward. */ | |
10207 | *src = aarch64_progress_pointer (*src); | |
10208 | *dst = aarch64_progress_pointer (*dst); | |
10209 | } | |
10210 | ||
10211 | /* Expand movmem, as if from a __builtin_memcpy. Return true if | |
10212 | we succeed, otherwise return false. */ | |
10213 | ||
10214 | bool | |
10215 | aarch64_expand_movmem (rtx *operands) | |
10216 | { | |
10217 | unsigned int n; | |
10218 | rtx dst = operands[0]; | |
10219 | rtx src = operands[1]; | |
10220 | rtx base; | |
10221 | bool speed_p = !optimize_function_for_size_p (cfun); | |
10222 | ||
10223 | /* When optimizing for size, give a better estimate of the length of a | |
10224 | memcpy call, but use the default otherwise. */ | |
10225 | unsigned int max_instructions = (speed_p ? 15 : AARCH64_CALL_RATIO) / 2; | |
10226 | ||
10227 | /* We can't do anything smart if the amount to copy is not constant. */ | |
10228 | if (!CONST_INT_P (operands[2])) | |
10229 | return false; | |
10230 | ||
10231 | n = UINTVAL (operands[2]); | |
10232 | ||
10233 | /* Try to keep the number of instructions low. For cases below 16 bytes we | |
10234 | need to make at most two moves. For cases above 16 bytes it will be one | |
10235 | move for each 16 byte chunk, then at most two additional moves. */ | |
10236 | if (((n / 16) + (n % 16 ? 2 : 0)) > max_instructions) | |
10237 | return false; | |
10238 | ||
10239 | base = copy_to_mode_reg (Pmode, XEXP (dst, 0)); | |
10240 | dst = adjust_automodify_address (dst, VOIDmode, base, 0); | |
10241 | ||
10242 | base = copy_to_mode_reg (Pmode, XEXP (src, 0)); | |
10243 | src = adjust_automodify_address (src, VOIDmode, base, 0); | |
10244 | ||
10245 | /* Simple cases. Copy 0-3 bytes, as (if applicable) a 2-byte, then a | |
10246 | 1-byte chunk. */ | |
10247 | if (n < 4) | |
10248 | { | |
10249 | if (n >= 2) | |
10250 | { | |
10251 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, HImode); | |
10252 | n -= 2; | |
10253 | } | |
10254 | ||
10255 | if (n == 1) | |
10256 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, QImode); | |
10257 | ||
10258 | return true; | |
10259 | } | |
10260 | ||
10261 | /* Copy 4-8 bytes. First a 4-byte chunk, then (if applicable) a second | |
10262 | 4-byte chunk, partially overlapping with the previously copied chunk. */ | |
10263 | if (n < 8) | |
10264 | { | |
10265 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
10266 | n -= 4; | |
10267 | if (n > 0) | |
10268 | { | |
10269 | int move = n - 4; | |
10270 | ||
10271 | src = aarch64_move_pointer (src, move); | |
10272 | dst = aarch64_move_pointer (dst, move); | |
10273 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
10274 | } | |
10275 | return true; | |
10276 | } | |
10277 | ||
10278 | /* Copy more than 8 bytes. Copy chunks of 16 bytes until we run out of | |
10279 | them, then (if applicable) an 8-byte chunk. */ | |
10280 | while (n >= 8) | |
10281 | { | |
10282 | if (n / 16) | |
10283 | { | |
10284 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, TImode); | |
10285 | n -= 16; | |
10286 | } | |
10287 | else | |
10288 | { | |
10289 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, DImode); | |
10290 | n -= 8; | |
10291 | } | |
10292 | } | |
10293 | ||
10294 | /* Finish the final bytes of the copy. We can always do this in one | |
10295 | instruction. We either copy the exact amount we need, or partially | |
10296 | overlap with the previous chunk we copied and copy 8-bytes. */ | |
10297 | if (n == 0) | |
10298 | return true; | |
10299 | else if (n == 1) | |
10300 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, QImode); | |
10301 | else if (n == 2) | |
10302 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, HImode); | |
10303 | else if (n == 4) | |
10304 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
10305 | else | |
10306 | { | |
10307 | if (n == 3) | |
10308 | { | |
10309 | src = aarch64_move_pointer (src, -1); | |
10310 | dst = aarch64_move_pointer (dst, -1); | |
10311 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, SImode); | |
10312 | } | |
10313 | else | |
10314 | { | |
10315 | int move = n - 8; | |
10316 | ||
10317 | src = aarch64_move_pointer (src, move); | |
10318 | dst = aarch64_move_pointer (dst, move); | |
10319 | aarch64_copy_one_block_and_progress_pointers (&src, &dst, DImode); | |
10320 | } | |
10321 | } | |
10322 | ||
10323 | return true; | |
10324 | } | |
10325 | ||
a3125fc2 CL |
10326 | /* Implement the TARGET_ASAN_SHADOW_OFFSET hook. */ |
10327 | ||
10328 | static unsigned HOST_WIDE_INT | |
10329 | aarch64_asan_shadow_offset (void) | |
10330 | { | |
10331 | return (HOST_WIDE_INT_1 << 36); | |
10332 | } | |
10333 | ||
d3006da6 | 10334 | static bool |
445d7826 | 10335 | aarch64_use_by_pieces_infrastructure_p (unsigned HOST_WIDE_INT size, |
d3006da6 JG |
10336 | unsigned int align, |
10337 | enum by_pieces_operation op, | |
10338 | bool speed_p) | |
10339 | { | |
10340 | /* STORE_BY_PIECES can be used when copying a constant string, but | |
10341 | in that case each 64-bit chunk takes 5 insns instead of 2 (LDR/STR). | |
10342 | For now we always fail this and let the move_by_pieces code copy | |
10343 | the string from read-only memory. */ | |
10344 | if (op == STORE_BY_PIECES) | |
10345 | return false; | |
10346 | ||
10347 | return default_use_by_pieces_infrastructure_p (size, align, op, speed_p); | |
10348 | } | |
10349 | ||
5f3bc026 ZC |
10350 | static enum machine_mode |
10351 | aarch64_code_to_ccmode (enum rtx_code code) | |
10352 | { | |
10353 | switch (code) | |
10354 | { | |
10355 | case NE: | |
10356 | return CC_DNEmode; | |
10357 | ||
10358 | case EQ: | |
10359 | return CC_DEQmode; | |
10360 | ||
10361 | case LE: | |
10362 | return CC_DLEmode; | |
10363 | ||
10364 | case LT: | |
10365 | return CC_DLTmode; | |
10366 | ||
10367 | case GE: | |
10368 | return CC_DGEmode; | |
10369 | ||
10370 | case GT: | |
10371 | return CC_DGTmode; | |
10372 | ||
10373 | case LEU: | |
10374 | return CC_DLEUmode; | |
10375 | ||
10376 | case LTU: | |
10377 | return CC_DLTUmode; | |
10378 | ||
10379 | case GEU: | |
10380 | return CC_DGEUmode; | |
10381 | ||
10382 | case GTU: | |
10383 | return CC_DGTUmode; | |
10384 | ||
10385 | default: | |
10386 | return CCmode; | |
10387 | } | |
10388 | } | |
10389 | ||
10390 | static rtx | |
10391 | aarch64_gen_ccmp_first (rtx *prep_seq, rtx *gen_seq, | |
10392 | int code, tree treeop0, tree treeop1) | |
10393 | { | |
10394 | enum machine_mode op_mode, cmp_mode, cc_mode; | |
10395 | rtx op0, op1, cmp, target; | |
10396 | int unsignedp = TYPE_UNSIGNED (TREE_TYPE (treeop0)); | |
10397 | enum insn_code icode; | |
10398 | struct expand_operand ops[4]; | |
10399 | ||
10400 | cc_mode = aarch64_code_to_ccmode ((enum rtx_code) code); | |
10401 | if (cc_mode == CCmode) | |
10402 | return NULL_RTX; | |
10403 | ||
10404 | start_sequence (); | |
10405 | expand_operands (treeop0, treeop1, NULL_RTX, &op0, &op1, EXPAND_NORMAL); | |
10406 | ||
10407 | op_mode = GET_MODE (op0); | |
10408 | if (op_mode == VOIDmode) | |
10409 | op_mode = GET_MODE (op1); | |
10410 | ||
10411 | switch (op_mode) | |
10412 | { | |
10413 | case QImode: | |
10414 | case HImode: | |
10415 | case SImode: | |
10416 | cmp_mode = SImode; | |
10417 | icode = CODE_FOR_cmpsi; | |
10418 | break; | |
10419 | ||
10420 | case DImode: | |
10421 | cmp_mode = DImode; | |
10422 | icode = CODE_FOR_cmpdi; | |
10423 | break; | |
10424 | ||
10425 | default: | |
10426 | end_sequence (); | |
10427 | return NULL_RTX; | |
10428 | } | |
10429 | ||
10430 | op0 = prepare_operand (icode, op0, 2, op_mode, cmp_mode, unsignedp); | |
10431 | op1 = prepare_operand (icode, op1, 3, op_mode, cmp_mode, unsignedp); | |
10432 | if (!op0 || !op1) | |
10433 | { | |
10434 | end_sequence (); | |
10435 | return NULL_RTX; | |
10436 | } | |
10437 | *prep_seq = get_insns (); | |
10438 | end_sequence (); | |
10439 | ||
10440 | cmp = gen_rtx_fmt_ee ((enum rtx_code) code, cmp_mode, op0, op1); | |
10441 | target = gen_rtx_REG (CCmode, CC_REGNUM); | |
10442 | ||
10443 | create_output_operand (&ops[0], target, CCmode); | |
10444 | create_fixed_operand (&ops[1], cmp); | |
10445 | create_fixed_operand (&ops[2], op0); | |
10446 | create_fixed_operand (&ops[3], op1); | |
10447 | ||
10448 | start_sequence (); | |
10449 | if (!maybe_expand_insn (icode, 4, ops)) | |
10450 | { | |
10451 | end_sequence (); | |
10452 | return NULL_RTX; | |
10453 | } | |
10454 | *gen_seq = get_insns (); | |
10455 | end_sequence (); | |
10456 | ||
10457 | return gen_rtx_REG (cc_mode, CC_REGNUM); | |
10458 | } | |
10459 | ||
10460 | static rtx | |
10461 | aarch64_gen_ccmp_next (rtx *prep_seq, rtx *gen_seq, rtx prev, int cmp_code, | |
10462 | tree treeop0, tree treeop1, int bit_code) | |
10463 | { | |
10464 | rtx op0, op1, cmp0, cmp1, target; | |
10465 | enum machine_mode op_mode, cmp_mode, cc_mode; | |
10466 | int unsignedp = TYPE_UNSIGNED (TREE_TYPE (treeop0)); | |
10467 | enum insn_code icode = CODE_FOR_ccmp_andsi; | |
10468 | struct expand_operand ops[6]; | |
10469 | ||
10470 | cc_mode = aarch64_code_to_ccmode ((enum rtx_code) cmp_code); | |
10471 | if (cc_mode == CCmode) | |
10472 | return NULL_RTX; | |
10473 | ||
10474 | push_to_sequence ((rtx_insn*) *prep_seq); | |
10475 | expand_operands (treeop0, treeop1, NULL_RTX, &op0, &op1, EXPAND_NORMAL); | |
10476 | ||
10477 | op_mode = GET_MODE (op0); | |
10478 | if (op_mode == VOIDmode) | |
10479 | op_mode = GET_MODE (op1); | |
10480 | ||
10481 | switch (op_mode) | |
10482 | { | |
10483 | case QImode: | |
10484 | case HImode: | |
10485 | case SImode: | |
10486 | cmp_mode = SImode; | |
10487 | icode = (enum rtx_code) bit_code == AND ? CODE_FOR_ccmp_andsi | |
10488 | : CODE_FOR_ccmp_iorsi; | |
10489 | break; | |
10490 | ||
10491 | case DImode: | |
10492 | cmp_mode = DImode; | |
10493 | icode = (enum rtx_code) bit_code == AND ? CODE_FOR_ccmp_anddi | |
10494 | : CODE_FOR_ccmp_iordi; | |
10495 | break; | |
10496 | ||
10497 | default: | |
10498 | end_sequence (); | |
10499 | return NULL_RTX; | |
10500 | } | |
10501 | ||
10502 | op0 = prepare_operand (icode, op0, 2, op_mode, cmp_mode, unsignedp); | |
10503 | op1 = prepare_operand (icode, op1, 3, op_mode, cmp_mode, unsignedp); | |
10504 | if (!op0 || !op1) | |
10505 | { | |
10506 | end_sequence (); | |
10507 | return NULL_RTX; | |
10508 | } | |
10509 | *prep_seq = get_insns (); | |
10510 | end_sequence (); | |
10511 | ||
10512 | target = gen_rtx_REG (cc_mode, CC_REGNUM); | |
10513 | cmp1 = gen_rtx_fmt_ee ((enum rtx_code) cmp_code, cmp_mode, op0, op1); | |
10514 | cmp0 = gen_rtx_fmt_ee (NE, cmp_mode, prev, const0_rtx); | |
10515 | ||
10516 | create_fixed_operand (&ops[0], prev); | |
10517 | create_fixed_operand (&ops[1], target); | |
10518 | create_fixed_operand (&ops[2], op0); | |
10519 | create_fixed_operand (&ops[3], op1); | |
10520 | create_fixed_operand (&ops[4], cmp0); | |
10521 | create_fixed_operand (&ops[5], cmp1); | |
10522 | ||
10523 | push_to_sequence ((rtx_insn*) *gen_seq); | |
10524 | if (!maybe_expand_insn (icode, 6, ops)) | |
10525 | { | |
10526 | end_sequence (); | |
10527 | return NULL_RTX; | |
10528 | } | |
10529 | ||
10530 | *gen_seq = get_insns (); | |
10531 | end_sequence (); | |
10532 | ||
10533 | return target; | |
10534 | } | |
10535 | ||
10536 | #undef TARGET_GEN_CCMP_FIRST | |
10537 | #define TARGET_GEN_CCMP_FIRST aarch64_gen_ccmp_first | |
10538 | ||
10539 | #undef TARGET_GEN_CCMP_NEXT | |
10540 | #define TARGET_GEN_CCMP_NEXT aarch64_gen_ccmp_next | |
10541 | ||
6a569cdd KT |
10542 | /* Implement TARGET_SCHED_MACRO_FUSION_P. Return true if target supports |
10543 | instruction fusion of some sort. */ | |
10544 | ||
10545 | static bool | |
10546 | aarch64_macro_fusion_p (void) | |
10547 | { | |
10548 | return aarch64_tune_params->fuseable_ops != AARCH64_FUSE_NOTHING; | |
10549 | } | |
10550 | ||
10551 | ||
10552 | /* Implement TARGET_SCHED_MACRO_FUSION_PAIR_P. Return true if PREV and CURR | |
10553 | should be kept together during scheduling. */ | |
10554 | ||
10555 | static bool | |
10556 | aarch_macro_fusion_pair_p (rtx_insn *prev, rtx_insn *curr) | |
10557 | { | |
10558 | rtx set_dest; | |
10559 | rtx prev_set = single_set (prev); | |
10560 | rtx curr_set = single_set (curr); | |
10561 | /* prev and curr are simple SET insns i.e. no flag setting or branching. */ | |
10562 | bool simple_sets_p = prev_set && curr_set && !any_condjump_p (curr); | |
10563 | ||
10564 | if (!aarch64_macro_fusion_p ()) | |
10565 | return false; | |
10566 | ||
10567 | if (simple_sets_p | |
10568 | && (aarch64_tune_params->fuseable_ops & AARCH64_FUSE_MOV_MOVK)) | |
10569 | { | |
10570 | /* We are trying to match: | |
10571 | prev (mov) == (set (reg r0) (const_int imm16)) | |
10572 | curr (movk) == (set (zero_extract (reg r0) | |
10573 | (const_int 16) | |
10574 | (const_int 16)) | |
10575 | (const_int imm16_1)) */ | |
10576 | ||
10577 | set_dest = SET_DEST (curr_set); | |
10578 | ||
10579 | if (GET_CODE (set_dest) == ZERO_EXTRACT | |
10580 | && CONST_INT_P (SET_SRC (curr_set)) | |
10581 | && CONST_INT_P (SET_SRC (prev_set)) | |
10582 | && CONST_INT_P (XEXP (set_dest, 2)) | |
10583 | && INTVAL (XEXP (set_dest, 2)) == 16 | |
10584 | && REG_P (XEXP (set_dest, 0)) | |
10585 | && REG_P (SET_DEST (prev_set)) | |
10586 | && REGNO (XEXP (set_dest, 0)) == REGNO (SET_DEST (prev_set))) | |
10587 | { | |
10588 | return true; | |
10589 | } | |
10590 | } | |
10591 | ||
9bbe08fe KT |
10592 | if (simple_sets_p |
10593 | && (aarch64_tune_params->fuseable_ops & AARCH64_FUSE_ADRP_ADD)) | |
10594 | { | |
10595 | ||
10596 | /* We're trying to match: | |
10597 | prev (adrp) == (set (reg r1) | |
10598 | (high (symbol_ref ("SYM")))) | |
10599 | curr (add) == (set (reg r0) | |
10600 | (lo_sum (reg r1) | |
10601 | (symbol_ref ("SYM")))) | |
10602 | Note that r0 need not necessarily be the same as r1, especially | |
10603 | during pre-regalloc scheduling. */ | |
10604 | ||
10605 | if (satisfies_constraint_Ush (SET_SRC (prev_set)) | |
10606 | && REG_P (SET_DEST (prev_set)) && REG_P (SET_DEST (curr_set))) | |
10607 | { | |
10608 | if (GET_CODE (SET_SRC (curr_set)) == LO_SUM | |
10609 | && REG_P (XEXP (SET_SRC (curr_set), 0)) | |
10610 | && REGNO (XEXP (SET_SRC (curr_set), 0)) | |
10611 | == REGNO (SET_DEST (prev_set)) | |
10612 | && rtx_equal_p (XEXP (SET_SRC (prev_set), 0), | |
10613 | XEXP (SET_SRC (curr_set), 1))) | |
10614 | return true; | |
10615 | } | |
10616 | } | |
10617 | ||
cd0cb232 KT |
10618 | if (simple_sets_p |
10619 | && (aarch64_tune_params->fuseable_ops & AARCH64_FUSE_MOVK_MOVK)) | |
10620 | { | |
10621 | ||
10622 | /* We're trying to match: | |
10623 | prev (movk) == (set (zero_extract (reg r0) | |
10624 | (const_int 16) | |
10625 | (const_int 32)) | |
10626 | (const_int imm16_1)) | |
10627 | curr (movk) == (set (zero_extract (reg r0) | |
10628 | (const_int 16) | |
10629 | (const_int 48)) | |
10630 | (const_int imm16_2)) */ | |
10631 | ||
10632 | if (GET_CODE (SET_DEST (prev_set)) == ZERO_EXTRACT | |
10633 | && GET_CODE (SET_DEST (curr_set)) == ZERO_EXTRACT | |
10634 | && REG_P (XEXP (SET_DEST (prev_set), 0)) | |
10635 | && REG_P (XEXP (SET_DEST (curr_set), 0)) | |
10636 | && REGNO (XEXP (SET_DEST (prev_set), 0)) | |
10637 | == REGNO (XEXP (SET_DEST (curr_set), 0)) | |
10638 | && CONST_INT_P (XEXP (SET_DEST (prev_set), 2)) | |
10639 | && CONST_INT_P (XEXP (SET_DEST (curr_set), 2)) | |
10640 | && INTVAL (XEXP (SET_DEST (prev_set), 2)) == 32 | |
10641 | && INTVAL (XEXP (SET_DEST (curr_set), 2)) == 48 | |
10642 | && CONST_INT_P (SET_SRC (prev_set)) | |
10643 | && CONST_INT_P (SET_SRC (curr_set))) | |
10644 | return true; | |
10645 | ||
10646 | } | |
d8354ad7 KT |
10647 | if (simple_sets_p |
10648 | && (aarch64_tune_params->fuseable_ops & AARCH64_FUSE_ADRP_LDR)) | |
10649 | { | |
10650 | /* We're trying to match: | |
10651 | prev (adrp) == (set (reg r0) | |
10652 | (high (symbol_ref ("SYM")))) | |
10653 | curr (ldr) == (set (reg r1) | |
10654 | (mem (lo_sum (reg r0) | |
10655 | (symbol_ref ("SYM"))))) | |
10656 | or | |
10657 | curr (ldr) == (set (reg r1) | |
10658 | (zero_extend (mem | |
10659 | (lo_sum (reg r0) | |
10660 | (symbol_ref ("SYM")))))) */ | |
10661 | if (satisfies_constraint_Ush (SET_SRC (prev_set)) | |
10662 | && REG_P (SET_DEST (prev_set)) && REG_P (SET_DEST (curr_set))) | |
10663 | { | |
10664 | rtx curr_src = SET_SRC (curr_set); | |
10665 | ||
10666 | if (GET_CODE (curr_src) == ZERO_EXTEND) | |
10667 | curr_src = XEXP (curr_src, 0); | |
10668 | ||
10669 | if (MEM_P (curr_src) && GET_CODE (XEXP (curr_src, 0)) == LO_SUM | |
10670 | && REG_P (XEXP (XEXP (curr_src, 0), 0)) | |
10671 | && REGNO (XEXP (XEXP (curr_src, 0), 0)) | |
10672 | == REGNO (SET_DEST (prev_set)) | |
10673 | && rtx_equal_p (XEXP (XEXP (curr_src, 0), 1), | |
10674 | XEXP (SET_SRC (prev_set), 0))) | |
10675 | return true; | |
10676 | } | |
10677 | } | |
cd0cb232 | 10678 | |
3759108f AP |
10679 | if ((aarch64_tune_params->fuseable_ops & AARCH64_FUSE_CMP_BRANCH) |
10680 | && any_condjump_p (curr)) | |
10681 | { | |
10682 | enum attr_type prev_type = get_attr_type (prev); | |
10683 | ||
10684 | /* FIXME: this misses some which is considered simple arthematic | |
10685 | instructions for ThunderX. Simple shifts are missed here. */ | |
10686 | if (prev_type == TYPE_ALUS_SREG | |
10687 | || prev_type == TYPE_ALUS_IMM | |
10688 | || prev_type == TYPE_LOGICS_REG | |
10689 | || prev_type == TYPE_LOGICS_IMM) | |
10690 | return true; | |
10691 | } | |
10692 | ||
6a569cdd KT |
10693 | return false; |
10694 | } | |
10695 | ||
350013bc BC |
10696 | /* If MEM is in the form of [base+offset], extract the two parts |
10697 | of address and set to BASE and OFFSET, otherwise return false | |
10698 | after clearing BASE and OFFSET. */ | |
10699 | ||
10700 | bool | |
10701 | extract_base_offset_in_addr (rtx mem, rtx *base, rtx *offset) | |
10702 | { | |
10703 | rtx addr; | |
10704 | ||
10705 | gcc_assert (MEM_P (mem)); | |
10706 | ||
10707 | addr = XEXP (mem, 0); | |
10708 | ||
10709 | if (REG_P (addr)) | |
10710 | { | |
10711 | *base = addr; | |
10712 | *offset = const0_rtx; | |
10713 | return true; | |
10714 | } | |
10715 | ||
10716 | if (GET_CODE (addr) == PLUS | |
10717 | && REG_P (XEXP (addr, 0)) && CONST_INT_P (XEXP (addr, 1))) | |
10718 | { | |
10719 | *base = XEXP (addr, 0); | |
10720 | *offset = XEXP (addr, 1); | |
10721 | return true; | |
10722 | } | |
10723 | ||
10724 | *base = NULL_RTX; | |
10725 | *offset = NULL_RTX; | |
10726 | ||
10727 | return false; | |
10728 | } | |
10729 | ||
10730 | /* Types for scheduling fusion. */ | |
10731 | enum sched_fusion_type | |
10732 | { | |
10733 | SCHED_FUSION_NONE = 0, | |
10734 | SCHED_FUSION_LD_SIGN_EXTEND, | |
10735 | SCHED_FUSION_LD_ZERO_EXTEND, | |
10736 | SCHED_FUSION_LD, | |
10737 | SCHED_FUSION_ST, | |
10738 | SCHED_FUSION_NUM | |
10739 | }; | |
10740 | ||
10741 | /* If INSN is a load or store of address in the form of [base+offset], | |
10742 | extract the two parts and set to BASE and OFFSET. Return scheduling | |
10743 | fusion type this INSN is. */ | |
10744 | ||
10745 | static enum sched_fusion_type | |
10746 | fusion_load_store (rtx_insn *insn, rtx *base, rtx *offset) | |
10747 | { | |
10748 | rtx x, dest, src; | |
10749 | enum sched_fusion_type fusion = SCHED_FUSION_LD; | |
10750 | ||
10751 | gcc_assert (INSN_P (insn)); | |
10752 | x = PATTERN (insn); | |
10753 | if (GET_CODE (x) != SET) | |
10754 | return SCHED_FUSION_NONE; | |
10755 | ||
10756 | src = SET_SRC (x); | |
10757 | dest = SET_DEST (x); | |
10758 | ||
1f46bd52 AP |
10759 | if (GET_MODE (dest) != SImode && GET_MODE (dest) != DImode |
10760 | && GET_MODE (dest) != SFmode && GET_MODE (dest) != DFmode) | |
350013bc BC |
10761 | return SCHED_FUSION_NONE; |
10762 | ||
10763 | if (GET_CODE (src) == SIGN_EXTEND) | |
10764 | { | |
10765 | fusion = SCHED_FUSION_LD_SIGN_EXTEND; | |
10766 | src = XEXP (src, 0); | |
10767 | if (GET_CODE (src) != MEM || GET_MODE (src) != SImode) | |
10768 | return SCHED_FUSION_NONE; | |
10769 | } | |
10770 | else if (GET_CODE (src) == ZERO_EXTEND) | |
10771 | { | |
10772 | fusion = SCHED_FUSION_LD_ZERO_EXTEND; | |
10773 | src = XEXP (src, 0); | |
10774 | if (GET_CODE (src) != MEM || GET_MODE (src) != SImode) | |
10775 | return SCHED_FUSION_NONE; | |
10776 | } | |
10777 | ||
10778 | if (GET_CODE (src) == MEM && REG_P (dest)) | |
10779 | extract_base_offset_in_addr (src, base, offset); | |
10780 | else if (GET_CODE (dest) == MEM && (REG_P (src) || src == const0_rtx)) | |
10781 | { | |
10782 | fusion = SCHED_FUSION_ST; | |
10783 | extract_base_offset_in_addr (dest, base, offset); | |
10784 | } | |
10785 | else | |
10786 | return SCHED_FUSION_NONE; | |
10787 | ||
10788 | if (*base == NULL_RTX || *offset == NULL_RTX) | |
10789 | fusion = SCHED_FUSION_NONE; | |
10790 | ||
10791 | return fusion; | |
10792 | } | |
10793 | ||
10794 | /* Implement the TARGET_SCHED_FUSION_PRIORITY hook. | |
10795 | ||
10796 | Currently we only support to fuse ldr or str instructions, so FUSION_PRI | |
10797 | and PRI are only calculated for these instructions. For other instruction, | |
10798 | FUSION_PRI and PRI are simply set to MAX_PRI - 1. In the future, other | |
10799 | type instruction fusion can be added by returning different priorities. | |
10800 | ||
10801 | It's important that irrelevant instructions get the largest FUSION_PRI. */ | |
10802 | ||
10803 | static void | |
10804 | aarch64_sched_fusion_priority (rtx_insn *insn, int max_pri, | |
10805 | int *fusion_pri, int *pri) | |
10806 | { | |
10807 | int tmp, off_val; | |
10808 | rtx base, offset; | |
10809 | enum sched_fusion_type fusion; | |
10810 | ||
10811 | gcc_assert (INSN_P (insn)); | |
10812 | ||
10813 | tmp = max_pri - 1; | |
10814 | fusion = fusion_load_store (insn, &base, &offset); | |
10815 | if (fusion == SCHED_FUSION_NONE) | |
10816 | { | |
10817 | *pri = tmp; | |
10818 | *fusion_pri = tmp; | |
10819 | return; | |
10820 | } | |
10821 | ||
10822 | /* Set FUSION_PRI according to fusion type and base register. */ | |
10823 | *fusion_pri = tmp - fusion * FIRST_PSEUDO_REGISTER - REGNO (base); | |
10824 | ||
10825 | /* Calculate PRI. */ | |
10826 | tmp /= 2; | |
10827 | ||
10828 | /* INSN with smaller offset goes first. */ | |
10829 | off_val = (int)(INTVAL (offset)); | |
10830 | if (off_val >= 0) | |
10831 | tmp -= (off_val & 0xfffff); | |
10832 | else | |
10833 | tmp += ((- off_val) & 0xfffff); | |
10834 | ||
10835 | *pri = tmp; | |
10836 | return; | |
10837 | } | |
10838 | ||
10839 | /* Given OPERANDS of consecutive load/store, check if we can merge | |
10840 | them into ldp/stp. LOAD is true if they are load instructions. | |
10841 | MODE is the mode of memory operands. */ | |
10842 | ||
10843 | bool | |
10844 | aarch64_operands_ok_for_ldpstp (rtx *operands, bool load, | |
10845 | enum machine_mode mode) | |
10846 | { | |
10847 | HOST_WIDE_INT offval_1, offval_2, msize; | |
10848 | enum reg_class rclass_1, rclass_2; | |
10849 | rtx mem_1, mem_2, reg_1, reg_2, base_1, base_2, offset_1, offset_2; | |
10850 | ||
10851 | if (load) | |
10852 | { | |
10853 | mem_1 = operands[1]; | |
10854 | mem_2 = operands[3]; | |
10855 | reg_1 = operands[0]; | |
10856 | reg_2 = operands[2]; | |
10857 | gcc_assert (REG_P (reg_1) && REG_P (reg_2)); | |
10858 | if (REGNO (reg_1) == REGNO (reg_2)) | |
10859 | return false; | |
10860 | } | |
10861 | else | |
10862 | { | |
10863 | mem_1 = operands[0]; | |
10864 | mem_2 = operands[2]; | |
10865 | reg_1 = operands[1]; | |
10866 | reg_2 = operands[3]; | |
10867 | } | |
10868 | ||
bf84ac44 AP |
10869 | /* The mems cannot be volatile. */ |
10870 | if (MEM_VOLATILE_P (mem_1) || MEM_VOLATILE_P (mem_2)) | |
10871 | return false; | |
10872 | ||
350013bc BC |
10873 | /* Check if the addresses are in the form of [base+offset]. */ |
10874 | extract_base_offset_in_addr (mem_1, &base_1, &offset_1); | |
10875 | if (base_1 == NULL_RTX || offset_1 == NULL_RTX) | |
10876 | return false; | |
10877 | extract_base_offset_in_addr (mem_2, &base_2, &offset_2); | |
10878 | if (base_2 == NULL_RTX || offset_2 == NULL_RTX) | |
10879 | return false; | |
10880 | ||
10881 | /* Check if the bases are same. */ | |
10882 | if (!rtx_equal_p (base_1, base_2)) | |
10883 | return false; | |
10884 | ||
10885 | offval_1 = INTVAL (offset_1); | |
10886 | offval_2 = INTVAL (offset_2); | |
10887 | msize = GET_MODE_SIZE (mode); | |
10888 | /* Check if the offsets are consecutive. */ | |
10889 | if (offval_1 != (offval_2 + msize) && offval_2 != (offval_1 + msize)) | |
10890 | return false; | |
10891 | ||
10892 | /* Check if the addresses are clobbered by load. */ | |
10893 | if (load) | |
10894 | { | |
10895 | if (reg_mentioned_p (reg_1, mem_1)) | |
10896 | return false; | |
10897 | ||
10898 | /* In increasing order, the last load can clobber the address. */ | |
10899 | if (offval_1 > offval_2 && reg_mentioned_p (reg_2, mem_2)) | |
10900 | return false; | |
10901 | } | |
10902 | ||
10903 | if (REG_P (reg_1) && FP_REGNUM_P (REGNO (reg_1))) | |
10904 | rclass_1 = FP_REGS; | |
10905 | else | |
10906 | rclass_1 = GENERAL_REGS; | |
10907 | ||
10908 | if (REG_P (reg_2) && FP_REGNUM_P (REGNO (reg_2))) | |
10909 | rclass_2 = FP_REGS; | |
10910 | else | |
10911 | rclass_2 = GENERAL_REGS; | |
10912 | ||
10913 | /* Check if the registers are of same class. */ | |
10914 | if (rclass_1 != rclass_2) | |
10915 | return false; | |
10916 | ||
10917 | return true; | |
10918 | } | |
10919 | ||
10920 | /* Given OPERANDS of consecutive load/store, check if we can merge | |
10921 | them into ldp/stp by adjusting the offset. LOAD is true if they | |
10922 | are load instructions. MODE is the mode of memory operands. | |
10923 | ||
10924 | Given below consecutive stores: | |
10925 | ||
10926 | str w1, [xb, 0x100] | |
10927 | str w1, [xb, 0x104] | |
10928 | str w1, [xb, 0x108] | |
10929 | str w1, [xb, 0x10c] | |
10930 | ||
10931 | Though the offsets are out of the range supported by stp, we can | |
10932 | still pair them after adjusting the offset, like: | |
10933 | ||
10934 | add scratch, xb, 0x100 | |
10935 | stp w1, w1, [scratch] | |
10936 | stp w1, w1, [scratch, 0x8] | |
10937 | ||
10938 | The peephole patterns detecting this opportunity should guarantee | |
10939 | the scratch register is avaliable. */ | |
10940 | ||
10941 | bool | |
10942 | aarch64_operands_adjust_ok_for_ldpstp (rtx *operands, bool load, | |
10943 | enum machine_mode mode) | |
10944 | { | |
10945 | enum reg_class rclass_1, rclass_2, rclass_3, rclass_4; | |
10946 | HOST_WIDE_INT offval_1, offval_2, offval_3, offval_4, msize; | |
10947 | rtx mem_1, mem_2, mem_3, mem_4, reg_1, reg_2, reg_3, reg_4; | |
10948 | rtx base_1, base_2, base_3, base_4, offset_1, offset_2, offset_3, offset_4; | |
10949 | ||
10950 | if (load) | |
10951 | { | |
10952 | reg_1 = operands[0]; | |
10953 | mem_1 = operands[1]; | |
10954 | reg_2 = operands[2]; | |
10955 | mem_2 = operands[3]; | |
10956 | reg_3 = operands[4]; | |
10957 | mem_3 = operands[5]; | |
10958 | reg_4 = operands[6]; | |
10959 | mem_4 = operands[7]; | |
10960 | gcc_assert (REG_P (reg_1) && REG_P (reg_2) | |
10961 | && REG_P (reg_3) && REG_P (reg_4)); | |
10962 | if (REGNO (reg_1) == REGNO (reg_2) || REGNO (reg_3) == REGNO (reg_4)) | |
10963 | return false; | |
10964 | } | |
10965 | else | |
10966 | { | |
10967 | mem_1 = operands[0]; | |
10968 | reg_1 = operands[1]; | |
10969 | mem_2 = operands[2]; | |
10970 | reg_2 = operands[3]; | |
10971 | mem_3 = operands[4]; | |
10972 | reg_3 = operands[5]; | |
10973 | mem_4 = operands[6]; | |
10974 | reg_4 = operands[7]; | |
10975 | } | |
10976 | /* Skip if memory operand is by itslef valid for ldp/stp. */ | |
10977 | if (!MEM_P (mem_1) || aarch64_mem_pair_operand (mem_1, mode)) | |
10978 | return false; | |
10979 | ||
bf84ac44 AP |
10980 | /* The mems cannot be volatile. */ |
10981 | if (MEM_VOLATILE_P (mem_1) || MEM_VOLATILE_P (mem_2) | |
10982 | || MEM_VOLATILE_P (mem_3) ||MEM_VOLATILE_P (mem_4)) | |
10983 | return false; | |
10984 | ||
350013bc BC |
10985 | /* Check if the addresses are in the form of [base+offset]. */ |
10986 | extract_base_offset_in_addr (mem_1, &base_1, &offset_1); | |
10987 | if (base_1 == NULL_RTX || offset_1 == NULL_RTX) | |
10988 | return false; | |
10989 | extract_base_offset_in_addr (mem_2, &base_2, &offset_2); | |
10990 | if (base_2 == NULL_RTX || offset_2 == NULL_RTX) | |
10991 | return false; | |
10992 | extract_base_offset_in_addr (mem_3, &base_3, &offset_3); | |
10993 | if (base_3 == NULL_RTX || offset_3 == NULL_RTX) | |
10994 | return false; | |
10995 | extract_base_offset_in_addr (mem_4, &base_4, &offset_4); | |
10996 | if (base_4 == NULL_RTX || offset_4 == NULL_RTX) | |
10997 | return false; | |
10998 | ||
10999 | /* Check if the bases are same. */ | |
11000 | if (!rtx_equal_p (base_1, base_2) | |
11001 | || !rtx_equal_p (base_2, base_3) | |
11002 | || !rtx_equal_p (base_3, base_4)) | |
11003 | return false; | |
11004 | ||
11005 | offval_1 = INTVAL (offset_1); | |
11006 | offval_2 = INTVAL (offset_2); | |
11007 | offval_3 = INTVAL (offset_3); | |
11008 | offval_4 = INTVAL (offset_4); | |
11009 | msize = GET_MODE_SIZE (mode); | |
11010 | /* Check if the offsets are consecutive. */ | |
11011 | if ((offval_1 != (offval_2 + msize) | |
11012 | || offval_1 != (offval_3 + msize * 2) | |
11013 | || offval_1 != (offval_4 + msize * 3)) | |
11014 | && (offval_4 != (offval_3 + msize) | |
11015 | || offval_4 != (offval_2 + msize * 2) | |
11016 | || offval_4 != (offval_1 + msize * 3))) | |
11017 | return false; | |
11018 | ||
11019 | /* Check if the addresses are clobbered by load. */ | |
11020 | if (load) | |
11021 | { | |
11022 | if (reg_mentioned_p (reg_1, mem_1) | |
11023 | || reg_mentioned_p (reg_2, mem_2) | |
11024 | || reg_mentioned_p (reg_3, mem_3)) | |
11025 | return false; | |
11026 | ||
11027 | /* In increasing order, the last load can clobber the address. */ | |
11028 | if (offval_1 > offval_2 && reg_mentioned_p (reg_4, mem_4)) | |
11029 | return false; | |
11030 | } | |
11031 | ||
11032 | if (REG_P (reg_1) && FP_REGNUM_P (REGNO (reg_1))) | |
11033 | rclass_1 = FP_REGS; | |
11034 | else | |
11035 | rclass_1 = GENERAL_REGS; | |
11036 | ||
11037 | if (REG_P (reg_2) && FP_REGNUM_P (REGNO (reg_2))) | |
11038 | rclass_2 = FP_REGS; | |
11039 | else | |
11040 | rclass_2 = GENERAL_REGS; | |
11041 | ||
11042 | if (REG_P (reg_3) && FP_REGNUM_P (REGNO (reg_3))) | |
11043 | rclass_3 = FP_REGS; | |
11044 | else | |
11045 | rclass_3 = GENERAL_REGS; | |
11046 | ||
11047 | if (REG_P (reg_4) && FP_REGNUM_P (REGNO (reg_4))) | |
11048 | rclass_4 = FP_REGS; | |
11049 | else | |
11050 | rclass_4 = GENERAL_REGS; | |
11051 | ||
11052 | /* Check if the registers are of same class. */ | |
11053 | if (rclass_1 != rclass_2 || rclass_2 != rclass_3 || rclass_3 != rclass_4) | |
11054 | return false; | |
11055 | ||
11056 | return true; | |
11057 | } | |
11058 | ||
11059 | /* Given OPERANDS of consecutive load/store, this function pairs them | |
11060 | into ldp/stp after adjusting the offset. It depends on the fact | |
11061 | that addresses of load/store instructions are in increasing order. | |
11062 | MODE is the mode of memory operands. CODE is the rtl operator | |
11063 | which should be applied to all memory operands, it's SIGN_EXTEND, | |
11064 | ZERO_EXTEND or UNKNOWN. */ | |
11065 | ||
11066 | bool | |
11067 | aarch64_gen_adjusted_ldpstp (rtx *operands, bool load, | |
11068 | enum machine_mode mode, RTX_CODE code) | |
11069 | { | |
11070 | rtx base, offset, t1, t2; | |
11071 | rtx mem_1, mem_2, mem_3, mem_4; | |
11072 | HOST_WIDE_INT off_val, abs_off, adj_off, new_off, stp_off_limit, msize; | |
11073 | ||
11074 | if (load) | |
11075 | { | |
11076 | mem_1 = operands[1]; | |
11077 | mem_2 = operands[3]; | |
11078 | mem_3 = operands[5]; | |
11079 | mem_4 = operands[7]; | |
11080 | } | |
11081 | else | |
11082 | { | |
11083 | mem_1 = operands[0]; | |
11084 | mem_2 = operands[2]; | |
11085 | mem_3 = operands[4]; | |
11086 | mem_4 = operands[6]; | |
11087 | gcc_assert (code == UNKNOWN); | |
11088 | } | |
11089 | ||
11090 | extract_base_offset_in_addr (mem_1, &base, &offset); | |
11091 | gcc_assert (base != NULL_RTX && offset != NULL_RTX); | |
11092 | ||
11093 | /* Adjust offset thus it can fit in ldp/stp instruction. */ | |
11094 | msize = GET_MODE_SIZE (mode); | |
11095 | stp_off_limit = msize * 0x40; | |
11096 | off_val = INTVAL (offset); | |
11097 | abs_off = (off_val < 0) ? -off_val : off_val; | |
11098 | new_off = abs_off % stp_off_limit; | |
11099 | adj_off = abs_off - new_off; | |
11100 | ||
11101 | /* Further adjust to make sure all offsets are OK. */ | |
11102 | if ((new_off + msize * 2) >= stp_off_limit) | |
11103 | { | |
11104 | adj_off += stp_off_limit; | |
11105 | new_off -= stp_off_limit; | |
11106 | } | |
11107 | ||
11108 | /* Make sure the adjustment can be done with ADD/SUB instructions. */ | |
11109 | if (adj_off >= 0x1000) | |
11110 | return false; | |
11111 | ||
11112 | if (off_val < 0) | |
11113 | { | |
11114 | adj_off = -adj_off; | |
11115 | new_off = -new_off; | |
11116 | } | |
11117 | ||
11118 | /* Create new memory references. */ | |
11119 | mem_1 = change_address (mem_1, VOIDmode, | |
11120 | plus_constant (DImode, operands[8], new_off)); | |
11121 | ||
11122 | /* Check if the adjusted address is OK for ldp/stp. */ | |
11123 | if (!aarch64_mem_pair_operand (mem_1, mode)) | |
11124 | return false; | |
11125 | ||
11126 | msize = GET_MODE_SIZE (mode); | |
11127 | mem_2 = change_address (mem_2, VOIDmode, | |
11128 | plus_constant (DImode, | |
11129 | operands[8], | |
11130 | new_off + msize)); | |
11131 | mem_3 = change_address (mem_3, VOIDmode, | |
11132 | plus_constant (DImode, | |
11133 | operands[8], | |
11134 | new_off + msize * 2)); | |
11135 | mem_4 = change_address (mem_4, VOIDmode, | |
11136 | plus_constant (DImode, | |
11137 | operands[8], | |
11138 | new_off + msize * 3)); | |
11139 | ||
11140 | if (code == ZERO_EXTEND) | |
11141 | { | |
11142 | mem_1 = gen_rtx_ZERO_EXTEND (DImode, mem_1); | |
11143 | mem_2 = gen_rtx_ZERO_EXTEND (DImode, mem_2); | |
11144 | mem_3 = gen_rtx_ZERO_EXTEND (DImode, mem_3); | |
11145 | mem_4 = gen_rtx_ZERO_EXTEND (DImode, mem_4); | |
11146 | } | |
11147 | else if (code == SIGN_EXTEND) | |
11148 | { | |
11149 | mem_1 = gen_rtx_SIGN_EXTEND (DImode, mem_1); | |
11150 | mem_2 = gen_rtx_SIGN_EXTEND (DImode, mem_2); | |
11151 | mem_3 = gen_rtx_SIGN_EXTEND (DImode, mem_3); | |
11152 | mem_4 = gen_rtx_SIGN_EXTEND (DImode, mem_4); | |
11153 | } | |
11154 | ||
11155 | if (load) | |
11156 | { | |
11157 | operands[1] = mem_1; | |
11158 | operands[3] = mem_2; | |
11159 | operands[5] = mem_3; | |
11160 | operands[7] = mem_4; | |
11161 | } | |
11162 | else | |
11163 | { | |
11164 | operands[0] = mem_1; | |
11165 | operands[2] = mem_2; | |
11166 | operands[4] = mem_3; | |
11167 | operands[6] = mem_4; | |
11168 | } | |
11169 | ||
11170 | /* Emit adjusting instruction. */ | |
11171 | emit_insn (gen_rtx_SET (VOIDmode, operands[8], | |
11172 | plus_constant (DImode, base, adj_off))); | |
11173 | /* Emit ldp/stp instructions. */ | |
11174 | t1 = gen_rtx_SET (VOIDmode, operands[0], operands[1]); | |
11175 | t2 = gen_rtx_SET (VOIDmode, operands[2], operands[3]); | |
11176 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, t1, t2))); | |
11177 | t1 = gen_rtx_SET (VOIDmode, operands[4], operands[5]); | |
11178 | t2 = gen_rtx_SET (VOIDmode, operands[6], operands[7]); | |
11179 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, t1, t2))); | |
11180 | return true; | |
11181 | } | |
11182 | ||
43e9d192 IB |
11183 | #undef TARGET_ADDRESS_COST |
11184 | #define TARGET_ADDRESS_COST aarch64_address_cost | |
11185 | ||
11186 | /* This hook will determines whether unnamed bitfields affect the alignment | |
11187 | of the containing structure. The hook returns true if the structure | |
11188 | should inherit the alignment requirements of an unnamed bitfield's | |
11189 | type. */ | |
11190 | #undef TARGET_ALIGN_ANON_BITFIELD | |
11191 | #define TARGET_ALIGN_ANON_BITFIELD hook_bool_void_true | |
11192 | ||
11193 | #undef TARGET_ASM_ALIGNED_DI_OP | |
11194 | #define TARGET_ASM_ALIGNED_DI_OP "\t.xword\t" | |
11195 | ||
11196 | #undef TARGET_ASM_ALIGNED_HI_OP | |
11197 | #define TARGET_ASM_ALIGNED_HI_OP "\t.hword\t" | |
11198 | ||
11199 | #undef TARGET_ASM_ALIGNED_SI_OP | |
11200 | #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t" | |
11201 | ||
11202 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
11203 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK \ | |
11204 | hook_bool_const_tree_hwi_hwi_const_tree_true | |
11205 | ||
11206 | #undef TARGET_ASM_FILE_START | |
11207 | #define TARGET_ASM_FILE_START aarch64_start_file | |
11208 | ||
11209 | #undef TARGET_ASM_OUTPUT_MI_THUNK | |
11210 | #define TARGET_ASM_OUTPUT_MI_THUNK aarch64_output_mi_thunk | |
11211 | ||
11212 | #undef TARGET_ASM_SELECT_RTX_SECTION | |
11213 | #define TARGET_ASM_SELECT_RTX_SECTION aarch64_select_rtx_section | |
11214 | ||
11215 | #undef TARGET_ASM_TRAMPOLINE_TEMPLATE | |
11216 | #define TARGET_ASM_TRAMPOLINE_TEMPLATE aarch64_asm_trampoline_template | |
11217 | ||
11218 | #undef TARGET_BUILD_BUILTIN_VA_LIST | |
11219 | #define TARGET_BUILD_BUILTIN_VA_LIST aarch64_build_builtin_va_list | |
11220 | ||
11221 | #undef TARGET_CALLEE_COPIES | |
11222 | #define TARGET_CALLEE_COPIES hook_bool_CUMULATIVE_ARGS_mode_tree_bool_false | |
11223 | ||
11224 | #undef TARGET_CAN_ELIMINATE | |
11225 | #define TARGET_CAN_ELIMINATE aarch64_can_eliminate | |
11226 | ||
11227 | #undef TARGET_CANNOT_FORCE_CONST_MEM | |
11228 | #define TARGET_CANNOT_FORCE_CONST_MEM aarch64_cannot_force_const_mem | |
11229 | ||
11230 | #undef TARGET_CONDITIONAL_REGISTER_USAGE | |
11231 | #define TARGET_CONDITIONAL_REGISTER_USAGE aarch64_conditional_register_usage | |
11232 | ||
11233 | /* Only the least significant bit is used for initialization guard | |
11234 | variables. */ | |
11235 | #undef TARGET_CXX_GUARD_MASK_BIT | |
11236 | #define TARGET_CXX_GUARD_MASK_BIT hook_bool_void_true | |
11237 | ||
11238 | #undef TARGET_C_MODE_FOR_SUFFIX | |
11239 | #define TARGET_C_MODE_FOR_SUFFIX aarch64_c_mode_for_suffix | |
11240 | ||
11241 | #ifdef TARGET_BIG_ENDIAN_DEFAULT | |
11242 | #undef TARGET_DEFAULT_TARGET_FLAGS | |
11243 | #define TARGET_DEFAULT_TARGET_FLAGS (MASK_BIG_END) | |
11244 | #endif | |
11245 | ||
11246 | #undef TARGET_CLASS_MAX_NREGS | |
11247 | #define TARGET_CLASS_MAX_NREGS aarch64_class_max_nregs | |
11248 | ||
119103ca JG |
11249 | #undef TARGET_BUILTIN_DECL |
11250 | #define TARGET_BUILTIN_DECL aarch64_builtin_decl | |
11251 | ||
43e9d192 IB |
11252 | #undef TARGET_EXPAND_BUILTIN |
11253 | #define TARGET_EXPAND_BUILTIN aarch64_expand_builtin | |
11254 | ||
11255 | #undef TARGET_EXPAND_BUILTIN_VA_START | |
11256 | #define TARGET_EXPAND_BUILTIN_VA_START aarch64_expand_builtin_va_start | |
11257 | ||
9697e620 JG |
11258 | #undef TARGET_FOLD_BUILTIN |
11259 | #define TARGET_FOLD_BUILTIN aarch64_fold_builtin | |
11260 | ||
43e9d192 IB |
11261 | #undef TARGET_FUNCTION_ARG |
11262 | #define TARGET_FUNCTION_ARG aarch64_function_arg | |
11263 | ||
11264 | #undef TARGET_FUNCTION_ARG_ADVANCE | |
11265 | #define TARGET_FUNCTION_ARG_ADVANCE aarch64_function_arg_advance | |
11266 | ||
11267 | #undef TARGET_FUNCTION_ARG_BOUNDARY | |
11268 | #define TARGET_FUNCTION_ARG_BOUNDARY aarch64_function_arg_boundary | |
11269 | ||
11270 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL | |
11271 | #define TARGET_FUNCTION_OK_FOR_SIBCALL aarch64_function_ok_for_sibcall | |
11272 | ||
11273 | #undef TARGET_FUNCTION_VALUE | |
11274 | #define TARGET_FUNCTION_VALUE aarch64_function_value | |
11275 | ||
11276 | #undef TARGET_FUNCTION_VALUE_REGNO_P | |
11277 | #define TARGET_FUNCTION_VALUE_REGNO_P aarch64_function_value_regno_p | |
11278 | ||
11279 | #undef TARGET_FRAME_POINTER_REQUIRED | |
11280 | #define TARGET_FRAME_POINTER_REQUIRED aarch64_frame_pointer_required | |
11281 | ||
fc72cba7 AL |
11282 | #undef TARGET_GIMPLE_FOLD_BUILTIN |
11283 | #define TARGET_GIMPLE_FOLD_BUILTIN aarch64_gimple_fold_builtin | |
0ac198d3 | 11284 | |
43e9d192 IB |
11285 | #undef TARGET_GIMPLIFY_VA_ARG_EXPR |
11286 | #define TARGET_GIMPLIFY_VA_ARG_EXPR aarch64_gimplify_va_arg_expr | |
11287 | ||
11288 | #undef TARGET_INIT_BUILTINS | |
11289 | #define TARGET_INIT_BUILTINS aarch64_init_builtins | |
11290 | ||
11291 | #undef TARGET_LEGITIMATE_ADDRESS_P | |
11292 | #define TARGET_LEGITIMATE_ADDRESS_P aarch64_legitimate_address_hook_p | |
11293 | ||
11294 | #undef TARGET_LEGITIMATE_CONSTANT_P | |
11295 | #define TARGET_LEGITIMATE_CONSTANT_P aarch64_legitimate_constant_p | |
11296 | ||
11297 | #undef TARGET_LIBGCC_CMP_RETURN_MODE | |
11298 | #define TARGET_LIBGCC_CMP_RETURN_MODE aarch64_libgcc_cmp_return_mode | |
11299 | ||
38e8f663 | 11300 | #undef TARGET_LRA_P |
98d404be | 11301 | #define TARGET_LRA_P hook_bool_void_true |
38e8f663 | 11302 | |
ac2b960f YZ |
11303 | #undef TARGET_MANGLE_TYPE |
11304 | #define TARGET_MANGLE_TYPE aarch64_mangle_type | |
11305 | ||
43e9d192 IB |
11306 | #undef TARGET_MEMORY_MOVE_COST |
11307 | #define TARGET_MEMORY_MOVE_COST aarch64_memory_move_cost | |
11308 | ||
26e0ff94 WD |
11309 | #undef TARGET_MIN_DIVISIONS_FOR_RECIP_MUL |
11310 | #define TARGET_MIN_DIVISIONS_FOR_RECIP_MUL aarch64_min_divisions_for_recip_mul | |
11311 | ||
43e9d192 IB |
11312 | #undef TARGET_MUST_PASS_IN_STACK |
11313 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
11314 | ||
11315 | /* This target hook should return true if accesses to volatile bitfields | |
11316 | should use the narrowest mode possible. It should return false if these | |
11317 | accesses should use the bitfield container type. */ | |
11318 | #undef TARGET_NARROW_VOLATILE_BITFIELD | |
11319 | #define TARGET_NARROW_VOLATILE_BITFIELD hook_bool_void_false | |
11320 | ||
11321 | #undef TARGET_OPTION_OVERRIDE | |
11322 | #define TARGET_OPTION_OVERRIDE aarch64_override_options | |
11323 | ||
11324 | #undef TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE | |
11325 | #define TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE \ | |
11326 | aarch64_override_options_after_change | |
11327 | ||
11328 | #undef TARGET_PASS_BY_REFERENCE | |
11329 | #define TARGET_PASS_BY_REFERENCE aarch64_pass_by_reference | |
11330 | ||
11331 | #undef TARGET_PREFERRED_RELOAD_CLASS | |
11332 | #define TARGET_PREFERRED_RELOAD_CLASS aarch64_preferred_reload_class | |
11333 | ||
cee66c68 WD |
11334 | #undef TARGET_SCHED_REASSOCIATION_WIDTH |
11335 | #define TARGET_SCHED_REASSOCIATION_WIDTH aarch64_reassociation_width | |
11336 | ||
43e9d192 IB |
11337 | #undef TARGET_SECONDARY_RELOAD |
11338 | #define TARGET_SECONDARY_RELOAD aarch64_secondary_reload | |
11339 | ||
11340 | #undef TARGET_SHIFT_TRUNCATION_MASK | |
11341 | #define TARGET_SHIFT_TRUNCATION_MASK aarch64_shift_truncation_mask | |
11342 | ||
11343 | #undef TARGET_SETUP_INCOMING_VARARGS | |
11344 | #define TARGET_SETUP_INCOMING_VARARGS aarch64_setup_incoming_varargs | |
11345 | ||
11346 | #undef TARGET_STRUCT_VALUE_RTX | |
11347 | #define TARGET_STRUCT_VALUE_RTX aarch64_struct_value_rtx | |
11348 | ||
11349 | #undef TARGET_REGISTER_MOVE_COST | |
11350 | #define TARGET_REGISTER_MOVE_COST aarch64_register_move_cost | |
11351 | ||
11352 | #undef TARGET_RETURN_IN_MEMORY | |
11353 | #define TARGET_RETURN_IN_MEMORY aarch64_return_in_memory | |
11354 | ||
11355 | #undef TARGET_RETURN_IN_MSB | |
11356 | #define TARGET_RETURN_IN_MSB aarch64_return_in_msb | |
11357 | ||
11358 | #undef TARGET_RTX_COSTS | |
7cc2145f | 11359 | #define TARGET_RTX_COSTS aarch64_rtx_costs_wrapper |
43e9d192 | 11360 | |
d126a4ae AP |
11361 | #undef TARGET_SCHED_ISSUE_RATE |
11362 | #define TARGET_SCHED_ISSUE_RATE aarch64_sched_issue_rate | |
11363 | ||
d03f7e44 MK |
11364 | #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD |
11365 | #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD \ | |
11366 | aarch64_sched_first_cycle_multipass_dfa_lookahead | |
11367 | ||
43e9d192 IB |
11368 | #undef TARGET_TRAMPOLINE_INIT |
11369 | #define TARGET_TRAMPOLINE_INIT aarch64_trampoline_init | |
11370 | ||
11371 | #undef TARGET_USE_BLOCKS_FOR_CONSTANT_P | |
11372 | #define TARGET_USE_BLOCKS_FOR_CONSTANT_P aarch64_use_blocks_for_constant_p | |
11373 | ||
11374 | #undef TARGET_VECTOR_MODE_SUPPORTED_P | |
11375 | #define TARGET_VECTOR_MODE_SUPPORTED_P aarch64_vector_mode_supported_p | |
11376 | ||
11377 | #undef TARGET_ARRAY_MODE_SUPPORTED_P | |
11378 | #define TARGET_ARRAY_MODE_SUPPORTED_P aarch64_array_mode_supported_p | |
11379 | ||
8990e73a TB |
11380 | #undef TARGET_VECTORIZE_ADD_STMT_COST |
11381 | #define TARGET_VECTORIZE_ADD_STMT_COST aarch64_add_stmt_cost | |
11382 | ||
11383 | #undef TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST | |
11384 | #define TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST \ | |
11385 | aarch64_builtin_vectorization_cost | |
11386 | ||
43e9d192 IB |
11387 | #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE |
11388 | #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE aarch64_preferred_simd_mode | |
11389 | ||
42fc9a7f JG |
11390 | #undef TARGET_VECTORIZE_BUILTINS |
11391 | #define TARGET_VECTORIZE_BUILTINS | |
11392 | ||
11393 | #undef TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION | |
11394 | #define TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION \ | |
11395 | aarch64_builtin_vectorized_function | |
11396 | ||
3b357264 JG |
11397 | #undef TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES |
11398 | #define TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES \ | |
11399 | aarch64_autovectorize_vector_sizes | |
11400 | ||
aa87aced KV |
11401 | #undef TARGET_ATOMIC_ASSIGN_EXPAND_FENV |
11402 | #define TARGET_ATOMIC_ASSIGN_EXPAND_FENV \ | |
11403 | aarch64_atomic_assign_expand_fenv | |
11404 | ||
43e9d192 IB |
11405 | /* Section anchor support. */ |
11406 | ||
11407 | #undef TARGET_MIN_ANCHOR_OFFSET | |
11408 | #define TARGET_MIN_ANCHOR_OFFSET -256 | |
11409 | ||
11410 | /* Limit the maximum anchor offset to 4k-1, since that's the limit for a | |
11411 | byte offset; we can do much more for larger data types, but have no way | |
11412 | to determine the size of the access. We assume accesses are aligned. */ | |
11413 | #undef TARGET_MAX_ANCHOR_OFFSET | |
11414 | #define TARGET_MAX_ANCHOR_OFFSET 4095 | |
11415 | ||
db0253a4 TB |
11416 | #undef TARGET_VECTOR_ALIGNMENT |
11417 | #define TARGET_VECTOR_ALIGNMENT aarch64_simd_vector_alignment | |
11418 | ||
11419 | #undef TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE | |
11420 | #define TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE \ | |
11421 | aarch64_simd_vector_alignment_reachable | |
11422 | ||
88b08073 JG |
11423 | /* vec_perm support. */ |
11424 | ||
11425 | #undef TARGET_VECTORIZE_VEC_PERM_CONST_OK | |
11426 | #define TARGET_VECTORIZE_VEC_PERM_CONST_OK \ | |
11427 | aarch64_vectorize_vec_perm_const_ok | |
11428 | ||
70f09188 | 11429 | |
706b2314 | 11430 | #undef TARGET_FIXED_CONDITION_CODE_REGS |
70f09188 AP |
11431 | #define TARGET_FIXED_CONDITION_CODE_REGS aarch64_fixed_condition_code_regs |
11432 | ||
5cb74e90 RR |
11433 | #undef TARGET_FLAGS_REGNUM |
11434 | #define TARGET_FLAGS_REGNUM CC_REGNUM | |
11435 | ||
78607708 TV |
11436 | #undef TARGET_CALL_FUSAGE_CONTAINS_NON_CALLEE_CLOBBERS |
11437 | #define TARGET_CALL_FUSAGE_CONTAINS_NON_CALLEE_CLOBBERS true | |
11438 | ||
a3125fc2 CL |
11439 | #undef TARGET_ASAN_SHADOW_OFFSET |
11440 | #define TARGET_ASAN_SHADOW_OFFSET aarch64_asan_shadow_offset | |
11441 | ||
0c4ec427 RE |
11442 | #undef TARGET_LEGITIMIZE_ADDRESS |
11443 | #define TARGET_LEGITIMIZE_ADDRESS aarch64_legitimize_address | |
11444 | ||
d3006da6 JG |
11445 | #undef TARGET_USE_BY_PIECES_INFRASTRUCTURE_P |
11446 | #define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \ | |
11447 | aarch64_use_by_pieces_infrastructure_p | |
11448 | ||
594bdd53 FY |
11449 | #undef TARGET_CAN_USE_DOLOOP_P |
11450 | #define TARGET_CAN_USE_DOLOOP_P can_use_doloop_if_innermost | |
11451 | ||
6a569cdd KT |
11452 | #undef TARGET_SCHED_MACRO_FUSION_P |
11453 | #define TARGET_SCHED_MACRO_FUSION_P aarch64_macro_fusion_p | |
11454 | ||
11455 | #undef TARGET_SCHED_MACRO_FUSION_PAIR_P | |
11456 | #define TARGET_SCHED_MACRO_FUSION_PAIR_P aarch_macro_fusion_pair_p | |
11457 | ||
350013bc BC |
11458 | #undef TARGET_SCHED_FUSION_PRIORITY |
11459 | #define TARGET_SCHED_FUSION_PRIORITY aarch64_sched_fusion_priority | |
11460 | ||
43e9d192 IB |
11461 | struct gcc_target targetm = TARGET_INITIALIZER; |
11462 | ||
11463 | #include "gt-aarch64.h" |