1 /* Machine description for AArch64 architecture.
2 Copyright (C) 2009-2017 Free Software Foundation, Inc.
3 Contributed by ARM Ltd.
5 This file is part of GCC.
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)
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.
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/>. */
22 #define INCLUDE_STRING
24 #include "coretypes.h"
35 #include "stringpool.h"
41 #include "diagnostic.h"
42 #include "insn-attr.h"
44 #include "fold-const.h"
45 #include "stor-layout.h"
53 #include "langhooks.h"
58 #include "gimple-iterator.h"
59 #include "tree-vectorizer.h"
60 #include "aarch64-cost-tables.h"
64 #include "tm-constrs.h"
65 #include "sched-int.h"
66 #include "target-globals.h"
67 #include "common/common-target.h"
69 #include "selftest-rtl.h"
71 /* This file should be included last. */
72 #include "target-def.h"
74 /* Defined for convenience. */
75 #define POINTER_BYTES (POINTER_SIZE / BITS_PER_UNIT)
77 /* Classifies an address.
80 A simple base register plus immediate offset.
83 A base register indexed by immediate offset with writeback.
86 A base register indexed by (optionally scaled) register.
89 A base register indexed by (optionally scaled) zero-extended register.
92 A base register indexed by (optionally scaled) sign-extended register.
95 A LO_SUM rtx with a base register and "LO12" symbol relocation.
98 A constant symbolic address, in pc-relative literal pool. */
100 enum aarch64_address_type
{
110 struct aarch64_address_info
{
111 enum aarch64_address_type type
;
115 enum aarch64_symbol_type symbol_type
;
118 struct simd_immediate_info
127 /* The current code model. */
128 enum aarch64_code_model aarch64_cmodel
;
131 #undef TARGET_HAVE_TLS
132 #define TARGET_HAVE_TLS 1
135 static bool aarch64_composite_type_p (const_tree
, machine_mode
);
136 static bool aarch64_vfp_is_call_or_return_candidate (machine_mode
,
138 machine_mode
*, int *,
140 static void aarch64_elf_asm_constructor (rtx
, int) ATTRIBUTE_UNUSED
;
141 static void aarch64_elf_asm_destructor (rtx
, int) ATTRIBUTE_UNUSED
;
142 static void aarch64_override_options_after_change (void);
143 static bool aarch64_vector_mode_supported_p (machine_mode
);
144 static bool aarch64_vectorize_vec_perm_const_ok (machine_mode vmode
,
145 const unsigned char *sel
);
146 static int aarch64_address_cost (rtx
, machine_mode
, addr_space_t
, bool);
147 static bool aarch64_builtin_support_vector_misalignment (machine_mode mode
,
152 aarch64_simd_container_mode (scalar_mode mode
, unsigned width
);
154 /* Major revision number of the ARM Architecture implemented by the target. */
155 unsigned aarch64_architecture_version
;
157 /* The processor for which instructions should be scheduled. */
158 enum aarch64_processor aarch64_tune
= cortexa53
;
160 /* Mask to specify which instruction scheduling options should be used. */
161 unsigned long aarch64_tune_flags
= 0;
163 /* Global flag for PC relative loads. */
164 bool aarch64_pcrelative_literal_loads
;
166 /* Support for command line parsing of boolean flags in the tuning
168 struct aarch64_flag_desc
174 #define AARCH64_FUSION_PAIR(name, internal_name) \
175 { name, AARCH64_FUSE_##internal_name },
176 static const struct aarch64_flag_desc aarch64_fusible_pairs
[] =
178 { "none", AARCH64_FUSE_NOTHING
},
179 #include "aarch64-fusion-pairs.def"
180 { "all", AARCH64_FUSE_ALL
},
181 { NULL
, AARCH64_FUSE_NOTHING
}
184 #define AARCH64_EXTRA_TUNING_OPTION(name, internal_name) \
185 { name, AARCH64_EXTRA_TUNE_##internal_name },
186 static const struct aarch64_flag_desc aarch64_tuning_flags
[] =
188 { "none", AARCH64_EXTRA_TUNE_NONE
},
189 #include "aarch64-tuning-flags.def"
190 { "all", AARCH64_EXTRA_TUNE_ALL
},
191 { NULL
, AARCH64_EXTRA_TUNE_NONE
}
194 /* Tuning parameters. */
196 static const struct cpu_addrcost_table generic_addrcost_table
=
206 0, /* register_offset */
207 0, /* register_sextend */
208 0, /* register_zextend */
212 static const struct cpu_addrcost_table exynosm1_addrcost_table
=
222 1, /* register_offset */
223 1, /* register_sextend */
224 2, /* register_zextend */
228 static const struct cpu_addrcost_table xgene1_addrcost_table
=
238 0, /* register_offset */
239 1, /* register_sextend */
240 1, /* register_zextend */
244 static const struct cpu_addrcost_table thunderx2t99_addrcost_table
=
254 2, /* register_offset */
255 3, /* register_sextend */
256 3, /* register_zextend */
260 static const struct cpu_regmove_cost generic_regmove_cost
=
263 /* Avoid the use of slow int<->fp moves for spilling by setting
264 their cost higher than memmov_cost. */
270 static const struct cpu_regmove_cost cortexa57_regmove_cost
=
273 /* Avoid the use of slow int<->fp moves for spilling by setting
274 their cost higher than memmov_cost. */
280 static const struct cpu_regmove_cost cortexa53_regmove_cost
=
283 /* Avoid the use of slow int<->fp moves for spilling by setting
284 their cost higher than memmov_cost. */
290 static const struct cpu_regmove_cost exynosm1_regmove_cost
=
293 /* Avoid the use of slow int<->fp moves for spilling by setting
294 their cost higher than memmov_cost (actual, 4 and 9). */
300 static const struct cpu_regmove_cost thunderx_regmove_cost
=
308 static const struct cpu_regmove_cost xgene1_regmove_cost
=
311 /* Avoid the use of slow int<->fp moves for spilling by setting
312 their cost higher than memmov_cost. */
318 static const struct cpu_regmove_cost qdf24xx_regmove_cost
=
321 /* Avoid the use of int<->fp moves for spilling. */
327 static const struct cpu_regmove_cost thunderx2t99_regmove_cost
=
330 /* Avoid the use of int<->fp moves for spilling. */
336 /* Generic costs for vector insn classes. */
337 static const struct cpu_vector_cost generic_vector_cost
=
339 1, /* scalar_int_stmt_cost */
340 1, /* scalar_fp_stmt_cost */
341 1, /* scalar_load_cost */
342 1, /* scalar_store_cost */
343 1, /* vec_int_stmt_cost */
344 1, /* vec_fp_stmt_cost */
345 2, /* vec_permute_cost */
346 1, /* vec_to_scalar_cost */
347 1, /* scalar_to_vec_cost */
348 1, /* vec_align_load_cost */
349 1, /* vec_unalign_load_cost */
350 1, /* vec_unalign_store_cost */
351 1, /* vec_store_cost */
352 3, /* cond_taken_branch_cost */
353 1 /* cond_not_taken_branch_cost */
356 /* ThunderX costs for vector insn classes. */
357 static const struct cpu_vector_cost thunderx_vector_cost
=
359 1, /* scalar_int_stmt_cost */
360 1, /* scalar_fp_stmt_cost */
361 3, /* scalar_load_cost */
362 1, /* scalar_store_cost */
363 4, /* vec_int_stmt_cost */
364 1, /* vec_fp_stmt_cost */
365 4, /* vec_permute_cost */
366 2, /* vec_to_scalar_cost */
367 2, /* scalar_to_vec_cost */
368 3, /* vec_align_load_cost */
369 5, /* vec_unalign_load_cost */
370 5, /* vec_unalign_store_cost */
371 1, /* vec_store_cost */
372 3, /* cond_taken_branch_cost */
373 3 /* cond_not_taken_branch_cost */
376 /* Generic costs for vector insn classes. */
377 static const struct cpu_vector_cost cortexa57_vector_cost
=
379 1, /* scalar_int_stmt_cost */
380 1, /* scalar_fp_stmt_cost */
381 4, /* scalar_load_cost */
382 1, /* scalar_store_cost */
383 2, /* vec_int_stmt_cost */
384 2, /* vec_fp_stmt_cost */
385 3, /* vec_permute_cost */
386 8, /* vec_to_scalar_cost */
387 8, /* scalar_to_vec_cost */
388 4, /* vec_align_load_cost */
389 4, /* vec_unalign_load_cost */
390 1, /* vec_unalign_store_cost */
391 1, /* vec_store_cost */
392 1, /* cond_taken_branch_cost */
393 1 /* cond_not_taken_branch_cost */
396 static const struct cpu_vector_cost exynosm1_vector_cost
=
398 1, /* scalar_int_stmt_cost */
399 1, /* scalar_fp_stmt_cost */
400 5, /* scalar_load_cost */
401 1, /* scalar_store_cost */
402 3, /* vec_int_stmt_cost */
403 3, /* vec_fp_stmt_cost */
404 3, /* vec_permute_cost */
405 3, /* vec_to_scalar_cost */
406 3, /* scalar_to_vec_cost */
407 5, /* vec_align_load_cost */
408 5, /* vec_unalign_load_cost */
409 1, /* vec_unalign_store_cost */
410 1, /* vec_store_cost */
411 1, /* cond_taken_branch_cost */
412 1 /* cond_not_taken_branch_cost */
415 /* Generic costs for vector insn classes. */
416 static const struct cpu_vector_cost xgene1_vector_cost
=
418 1, /* scalar_int_stmt_cost */
419 1, /* scalar_fp_stmt_cost */
420 5, /* scalar_load_cost */
421 1, /* scalar_store_cost */
422 2, /* vec_int_stmt_cost */
423 2, /* vec_fp_stmt_cost */
424 2, /* vec_permute_cost */
425 4, /* vec_to_scalar_cost */
426 4, /* scalar_to_vec_cost */
427 10, /* vec_align_load_cost */
428 10, /* vec_unalign_load_cost */
429 2, /* vec_unalign_store_cost */
430 2, /* vec_store_cost */
431 2, /* cond_taken_branch_cost */
432 1 /* cond_not_taken_branch_cost */
435 /* Costs for vector insn classes for Vulcan. */
436 static const struct cpu_vector_cost thunderx2t99_vector_cost
=
438 1, /* scalar_int_stmt_cost */
439 6, /* scalar_fp_stmt_cost */
440 4, /* scalar_load_cost */
441 1, /* scalar_store_cost */
442 5, /* vec_int_stmt_cost */
443 6, /* vec_fp_stmt_cost */
444 3, /* vec_permute_cost */
445 6, /* vec_to_scalar_cost */
446 5, /* scalar_to_vec_cost */
447 8, /* vec_align_load_cost */
448 8, /* vec_unalign_load_cost */
449 4, /* vec_unalign_store_cost */
450 4, /* vec_store_cost */
451 2, /* cond_taken_branch_cost */
452 1 /* cond_not_taken_branch_cost */
455 /* Generic costs for branch instructions. */
456 static const struct cpu_branch_cost generic_branch_cost
=
458 1, /* Predictable. */
459 3 /* Unpredictable. */
462 /* Generic approximation modes. */
463 static const cpu_approx_modes generic_approx_modes
=
465 AARCH64_APPROX_NONE
, /* division */
466 AARCH64_APPROX_NONE
, /* sqrt */
467 AARCH64_APPROX_NONE
/* recip_sqrt */
470 /* Approximation modes for Exynos M1. */
471 static const cpu_approx_modes exynosm1_approx_modes
=
473 AARCH64_APPROX_NONE
, /* division */
474 AARCH64_APPROX_ALL
, /* sqrt */
475 AARCH64_APPROX_ALL
/* recip_sqrt */
478 /* Approximation modes for X-Gene 1. */
479 static const cpu_approx_modes xgene1_approx_modes
=
481 AARCH64_APPROX_NONE
, /* division */
482 AARCH64_APPROX_NONE
, /* sqrt */
483 AARCH64_APPROX_ALL
/* recip_sqrt */
486 /* Generic prefetch settings (which disable prefetch). */
487 static const cpu_prefetch_tune generic_prefetch_tune
=
490 -1, /* l1_cache_size */
491 -1, /* l1_cache_line_size */
492 -1, /* l2_cache_size */
493 -1 /* default_opt_level */
496 static const cpu_prefetch_tune exynosm1_prefetch_tune
=
499 -1, /* l1_cache_size */
500 64, /* l1_cache_line_size */
501 -1, /* l2_cache_size */
502 -1 /* default_opt_level */
505 static const cpu_prefetch_tune qdf24xx_prefetch_tune
=
508 32, /* l1_cache_size */
509 64, /* l1_cache_line_size */
510 1024, /* l2_cache_size */
511 3 /* default_opt_level */
514 static const cpu_prefetch_tune thunderxt88_prefetch_tune
=
517 32, /* l1_cache_size */
518 128, /* l1_cache_line_size */
519 16*1024, /* l2_cache_size */
520 3 /* default_opt_level */
523 static const cpu_prefetch_tune thunderx_prefetch_tune
=
526 32, /* l1_cache_size */
527 128, /* l1_cache_line_size */
528 -1, /* l2_cache_size */
529 -1 /* default_opt_level */
532 static const cpu_prefetch_tune thunderx2t99_prefetch_tune
=
535 32, /* l1_cache_size */
536 64, /* l1_cache_line_size */
537 256, /* l2_cache_size */
538 -1 /* default_opt_level */
541 static const struct tune_params generic_tunings
=
543 &cortexa57_extra_costs
,
544 &generic_addrcost_table
,
545 &generic_regmove_cost
,
546 &generic_vector_cost
,
547 &generic_branch_cost
,
548 &generic_approx_modes
,
551 (AARCH64_FUSE_AES_AESMC
), /* fusible_ops */
552 8, /* function_align. */
555 2, /* int_reassoc_width. */
556 4, /* fp_reassoc_width. */
557 1, /* vec_reassoc_width. */
558 2, /* min_div_recip_mul_sf. */
559 2, /* min_div_recip_mul_df. */
560 0, /* max_case_values. */
561 tune_params::AUTOPREFETCHER_WEAK
, /* autoprefetcher_model. */
562 (AARCH64_EXTRA_TUNE_NONE
), /* tune_flags. */
563 &generic_prefetch_tune
566 static const struct tune_params cortexa35_tunings
=
568 &cortexa53_extra_costs
,
569 &generic_addrcost_table
,
570 &cortexa53_regmove_cost
,
571 &generic_vector_cost
,
572 &generic_branch_cost
,
573 &generic_approx_modes
,
576 (AARCH64_FUSE_AES_AESMC
| AARCH64_FUSE_MOV_MOVK
| AARCH64_FUSE_ADRP_ADD
577 | AARCH64_FUSE_MOVK_MOVK
| AARCH64_FUSE_ADRP_LDR
), /* fusible_ops */
578 16, /* function_align. */
581 2, /* int_reassoc_width. */
582 4, /* fp_reassoc_width. */
583 1, /* vec_reassoc_width. */
584 2, /* min_div_recip_mul_sf. */
585 2, /* min_div_recip_mul_df. */
586 0, /* max_case_values. */
587 tune_params::AUTOPREFETCHER_WEAK
, /* autoprefetcher_model. */
588 (AARCH64_EXTRA_TUNE_NONE
), /* tune_flags. */
589 &generic_prefetch_tune
592 static const struct tune_params cortexa53_tunings
=
594 &cortexa53_extra_costs
,
595 &generic_addrcost_table
,
596 &cortexa53_regmove_cost
,
597 &generic_vector_cost
,
598 &generic_branch_cost
,
599 &generic_approx_modes
,
602 (AARCH64_FUSE_AES_AESMC
| AARCH64_FUSE_MOV_MOVK
| AARCH64_FUSE_ADRP_ADD
603 | AARCH64_FUSE_MOVK_MOVK
| AARCH64_FUSE_ADRP_LDR
), /* fusible_ops */
604 16, /* function_align. */
607 2, /* int_reassoc_width. */
608 4, /* fp_reassoc_width. */
609 1, /* vec_reassoc_width. */
610 2, /* min_div_recip_mul_sf. */
611 2, /* min_div_recip_mul_df. */
612 0, /* max_case_values. */
613 tune_params::AUTOPREFETCHER_WEAK
, /* autoprefetcher_model. */
614 (AARCH64_EXTRA_TUNE_NONE
), /* tune_flags. */
615 &generic_prefetch_tune
618 static const struct tune_params cortexa57_tunings
=
620 &cortexa57_extra_costs
,
621 &generic_addrcost_table
,
622 &cortexa57_regmove_cost
,
623 &cortexa57_vector_cost
,
624 &generic_branch_cost
,
625 &generic_approx_modes
,
628 (AARCH64_FUSE_AES_AESMC
| AARCH64_FUSE_MOV_MOVK
| AARCH64_FUSE_ADRP_ADD
629 | AARCH64_FUSE_MOVK_MOVK
), /* fusible_ops */
630 16, /* function_align. */
633 2, /* int_reassoc_width. */
634 4, /* fp_reassoc_width. */
635 1, /* vec_reassoc_width. */
636 2, /* min_div_recip_mul_sf. */
637 2, /* min_div_recip_mul_df. */
638 0, /* max_case_values. */
639 tune_params::AUTOPREFETCHER_WEAK
, /* autoprefetcher_model. */
640 (AARCH64_EXTRA_TUNE_RENAME_FMA_REGS
), /* tune_flags. */
641 &generic_prefetch_tune
644 static const struct tune_params cortexa72_tunings
=
646 &cortexa57_extra_costs
,
647 &generic_addrcost_table
,
648 &cortexa57_regmove_cost
,
649 &cortexa57_vector_cost
,
650 &generic_branch_cost
,
651 &generic_approx_modes
,
654 (AARCH64_FUSE_AES_AESMC
| AARCH64_FUSE_MOV_MOVK
| AARCH64_FUSE_ADRP_ADD
655 | AARCH64_FUSE_MOVK_MOVK
), /* fusible_ops */
656 16, /* function_align. */
659 2, /* int_reassoc_width. */
660 4, /* fp_reassoc_width. */
661 1, /* vec_reassoc_width. */
662 2, /* min_div_recip_mul_sf. */
663 2, /* min_div_recip_mul_df. */
664 0, /* max_case_values. */
665 tune_params::AUTOPREFETCHER_WEAK
, /* autoprefetcher_model. */
666 (AARCH64_EXTRA_TUNE_NONE
), /* tune_flags. */
667 &generic_prefetch_tune
670 static const struct tune_params cortexa73_tunings
=
672 &cortexa57_extra_costs
,
673 &generic_addrcost_table
,
674 &cortexa57_regmove_cost
,
675 &cortexa57_vector_cost
,
676 &generic_branch_cost
,
677 &generic_approx_modes
,
678 4, /* memmov_cost. */
680 (AARCH64_FUSE_AES_AESMC
| AARCH64_FUSE_MOV_MOVK
| AARCH64_FUSE_ADRP_ADD
681 | AARCH64_FUSE_MOVK_MOVK
| AARCH64_FUSE_ADRP_LDR
), /* fusible_ops */
682 16, /* function_align. */
685 2, /* int_reassoc_width. */
686 4, /* fp_reassoc_width. */
687 1, /* vec_reassoc_width. */
688 2, /* min_div_recip_mul_sf. */
689 2, /* min_div_recip_mul_df. */
690 0, /* max_case_values. */
691 tune_params::AUTOPREFETCHER_WEAK
, /* autoprefetcher_model. */
692 (AARCH64_EXTRA_TUNE_NONE
), /* tune_flags. */
693 &generic_prefetch_tune
698 static const struct tune_params exynosm1_tunings
=
700 &exynosm1_extra_costs
,
701 &exynosm1_addrcost_table
,
702 &exynosm1_regmove_cost
,
703 &exynosm1_vector_cost
,
704 &generic_branch_cost
,
705 &exynosm1_approx_modes
,
708 (AARCH64_FUSE_AES_AESMC
), /* fusible_ops */
709 4, /* function_align. */
712 2, /* int_reassoc_width. */
713 4, /* fp_reassoc_width. */
714 1, /* vec_reassoc_width. */
715 2, /* min_div_recip_mul_sf. */
716 2, /* min_div_recip_mul_df. */
717 48, /* max_case_values. */
718 tune_params::AUTOPREFETCHER_WEAK
, /* autoprefetcher_model. */
719 (AARCH64_EXTRA_TUNE_NONE
), /* tune_flags. */
720 &exynosm1_prefetch_tune
723 static const struct tune_params thunderxt88_tunings
=
725 &thunderx_extra_costs
,
726 &generic_addrcost_table
,
727 &thunderx_regmove_cost
,
728 &thunderx_vector_cost
,
729 &generic_branch_cost
,
730 &generic_approx_modes
,
733 AARCH64_FUSE_CMP_BRANCH
, /* fusible_ops */
734 8, /* function_align. */
737 2, /* int_reassoc_width. */
738 4, /* fp_reassoc_width. */
739 1, /* vec_reassoc_width. */
740 2, /* min_div_recip_mul_sf. */
741 2, /* min_div_recip_mul_df. */
742 0, /* max_case_values. */
743 tune_params::AUTOPREFETCHER_OFF
, /* autoprefetcher_model. */
744 (AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW
), /* tune_flags. */
745 &thunderxt88_prefetch_tune
748 static const struct tune_params thunderx_tunings
=
750 &thunderx_extra_costs
,
751 &generic_addrcost_table
,
752 &thunderx_regmove_cost
,
753 &thunderx_vector_cost
,
754 &generic_branch_cost
,
755 &generic_approx_modes
,
758 AARCH64_FUSE_CMP_BRANCH
, /* fusible_ops */
759 8, /* function_align. */
762 2, /* int_reassoc_width. */
763 4, /* fp_reassoc_width. */
764 1, /* vec_reassoc_width. */
765 2, /* min_div_recip_mul_sf. */
766 2, /* min_div_recip_mul_df. */
767 0, /* max_case_values. */
768 tune_params::AUTOPREFETCHER_OFF
, /* autoprefetcher_model. */
769 (AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW
770 | AARCH64_EXTRA_TUNE_CHEAP_SHIFT_EXTEND
), /* tune_flags. */
771 &thunderx_prefetch_tune
774 static const struct tune_params xgene1_tunings
=
777 &xgene1_addrcost_table
,
778 &xgene1_regmove_cost
,
780 &generic_branch_cost
,
781 &xgene1_approx_modes
,
784 AARCH64_FUSE_NOTHING
, /* fusible_ops */
785 16, /* function_align. */
787 16, /* loop_align. */
788 2, /* int_reassoc_width. */
789 4, /* fp_reassoc_width. */
790 1, /* vec_reassoc_width. */
791 2, /* min_div_recip_mul_sf. */
792 2, /* min_div_recip_mul_df. */
793 0, /* max_case_values. */
794 tune_params::AUTOPREFETCHER_OFF
, /* autoprefetcher_model. */
795 (AARCH64_EXTRA_TUNE_NONE
), /* tune_flags. */
796 &generic_prefetch_tune
799 static const struct tune_params qdf24xx_tunings
=
801 &qdf24xx_extra_costs
,
802 &generic_addrcost_table
,
803 &qdf24xx_regmove_cost
,
804 &generic_vector_cost
,
805 &generic_branch_cost
,
806 &generic_approx_modes
,
809 (AARCH64_FUSE_MOV_MOVK
| AARCH64_FUSE_ADRP_ADD
810 | AARCH64_FUSE_MOVK_MOVK
), /* fuseable_ops */
811 16, /* function_align. */
813 16, /* loop_align. */
814 2, /* int_reassoc_width. */
815 4, /* fp_reassoc_width. */
816 1, /* vec_reassoc_width. */
817 2, /* min_div_recip_mul_sf. */
818 2, /* min_div_recip_mul_df. */
819 0, /* max_case_values. */
820 tune_params::AUTOPREFETCHER_STRONG
, /* autoprefetcher_model. */
821 (AARCH64_EXTRA_TUNE_NONE
), /* tune_flags. */
822 &qdf24xx_prefetch_tune
825 static const struct tune_params thunderx2t99_tunings
=
827 &thunderx2t99_extra_costs
,
828 &thunderx2t99_addrcost_table
,
829 &thunderx2t99_regmove_cost
,
830 &thunderx2t99_vector_cost
,
831 &generic_branch_cost
,
832 &generic_approx_modes
,
833 4, /* memmov_cost. */
835 (AARCH64_FUSE_CMP_BRANCH
| AARCH64_FUSE_AES_AESMC
836 | AARCH64_FUSE_ALU_BRANCH
), /* fusible_ops */
837 16, /* function_align. */
839 16, /* loop_align. */
840 3, /* int_reassoc_width. */
841 2, /* fp_reassoc_width. */
842 2, /* vec_reassoc_width. */
843 2, /* min_div_recip_mul_sf. */
844 2, /* min_div_recip_mul_df. */
845 0, /* max_case_values. */
846 tune_params::AUTOPREFETCHER_WEAK
, /* autoprefetcher_model. */
847 (AARCH64_EXTRA_TUNE_NONE
), /* tune_flags. */
848 &thunderx2t99_prefetch_tune
851 /* Support for fine-grained override of the tuning structures. */
852 struct aarch64_tuning_override_function
855 void (*parse_override
)(const char*, struct tune_params
*);
858 static void aarch64_parse_fuse_string (const char*, struct tune_params
*);
859 static void aarch64_parse_tune_string (const char*, struct tune_params
*);
861 static const struct aarch64_tuning_override_function
862 aarch64_tuning_override_functions
[] =
864 { "fuse", aarch64_parse_fuse_string
},
865 { "tune", aarch64_parse_tune_string
},
869 /* A processor implementing AArch64. */
872 const char *const name
;
873 enum aarch64_processor ident
;
874 enum aarch64_processor sched_core
;
875 enum aarch64_arch arch
;
876 unsigned architecture_version
;
877 const unsigned long flags
;
878 const struct tune_params
*const tune
;
881 /* Architectures implementing AArch64. */
882 static const struct processor all_architectures
[] =
884 #define AARCH64_ARCH(NAME, CORE, ARCH_IDENT, ARCH_REV, FLAGS) \
885 {NAME, CORE, CORE, AARCH64_ARCH_##ARCH_IDENT, ARCH_REV, FLAGS, NULL},
886 #include "aarch64-arches.def"
887 {NULL
, aarch64_none
, aarch64_none
, aarch64_no_arch
, 0, 0, NULL
}
890 /* Processor cores implementing AArch64. */
891 static const struct processor all_cores
[] =
893 #define AARCH64_CORE(NAME, IDENT, SCHED, ARCH, FLAGS, COSTS, IMP, PART, VARIANT) \
894 {NAME, IDENT, SCHED, AARCH64_ARCH_##ARCH, \
895 all_architectures[AARCH64_ARCH_##ARCH].architecture_version, \
896 FLAGS, &COSTS##_tunings},
897 #include "aarch64-cores.def"
898 {"generic", generic
, cortexa53
, AARCH64_ARCH_8A
, 8,
899 AARCH64_FL_FOR_ARCH8
, &generic_tunings
},
900 {NULL
, aarch64_none
, aarch64_none
, aarch64_no_arch
, 0, 0, NULL
}
904 /* Target specification. These are populated by the -march, -mtune, -mcpu
905 handling code or by target attributes. */
906 static const struct processor
*selected_arch
;
907 static const struct processor
*selected_cpu
;
908 static const struct processor
*selected_tune
;
910 /* The current tuning set. */
911 struct tune_params aarch64_tune_params
= generic_tunings
;
913 #define AARCH64_CPU_DEFAULT_FLAGS ((selected_cpu) ? selected_cpu->flags : 0)
915 /* An ISA extension in the co-processor and main instruction set space. */
916 struct aarch64_option_extension
918 const char *const name
;
919 const unsigned long flags_on
;
920 const unsigned long flags_off
;
923 typedef enum aarch64_cond_code
925 AARCH64_EQ
= 0, AARCH64_NE
, AARCH64_CS
, AARCH64_CC
, AARCH64_MI
, AARCH64_PL
,
926 AARCH64_VS
, AARCH64_VC
, AARCH64_HI
, AARCH64_LS
, AARCH64_GE
, AARCH64_LT
,
927 AARCH64_GT
, AARCH64_LE
, AARCH64_AL
, AARCH64_NV
931 #define AARCH64_INVERSE_CONDITION_CODE(X) ((aarch64_cc) (((int) X) ^ 1))
933 /* The condition codes of the processor, and the inverse function. */
934 static const char * const aarch64_condition_codes
[] =
936 "eq", "ne", "cs", "cc", "mi", "pl", "vs", "vc",
937 "hi", "ls", "ge", "lt", "gt", "le", "al", "nv"
940 /* Generate code to enable conditional branches in functions over 1 MiB. */
942 aarch64_gen_far_branch (rtx
* operands
, int pos_label
, const char * dest
,
943 const char * branch_format
)
945 rtx_code_label
* tmp_label
= gen_label_rtx ();
948 ASM_GENERATE_INTERNAL_LABEL (label_buf
, dest
,
949 CODE_LABEL_NUMBER (tmp_label
));
950 const char *label_ptr
= targetm
.strip_name_encoding (label_buf
);
951 rtx dest_label
= operands
[pos_label
];
952 operands
[pos_label
] = tmp_label
;
954 snprintf (buffer
, sizeof (buffer
), "%s%s", branch_format
, label_ptr
);
955 output_asm_insn (buffer
, operands
);
957 snprintf (buffer
, sizeof (buffer
), "b\t%%l%d\n%s:", pos_label
, label_ptr
);
958 operands
[pos_label
] = dest_label
;
959 output_asm_insn (buffer
, operands
);
964 aarch64_err_no_fpadvsimd (machine_mode mode
, const char *msg
)
966 const char *mc
= FLOAT_MODE_P (mode
) ? "floating-point" : "vector";
967 if (TARGET_GENERAL_REGS_ONLY
)
968 error ("%qs is incompatible with %s %s", "-mgeneral-regs-only", mc
, msg
);
970 error ("%qs feature modifier is incompatible with %s %s", "+nofp", mc
, msg
);
973 /* Implement TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS.
974 The register allocator chooses ALL_REGS if FP_REGS and GENERAL_REGS have
975 the same cost even if ALL_REGS has a much larger cost. ALL_REGS is also
976 used if the cost of both FP_REGS and GENERAL_REGS is lower than the memory
977 cost (in this case the best class is the lowest cost one). Using ALL_REGS
978 irrespectively of its cost results in bad allocations with many redundant
979 int<->FP moves which are expensive on various cores.
980 To avoid this we don't allow ALL_REGS as the allocno class, but force a
981 decision between FP_REGS and GENERAL_REGS. We use the allocno class if it
982 isn't ALL_REGS. Similarly, use the best class if it isn't ALL_REGS.
983 Otherwise set the allocno class depending on the mode.
984 The result of this is that it is no longer inefficient to have a higher
985 memory move cost than the register move cost.
989 aarch64_ira_change_pseudo_allocno_class (int regno
, reg_class_t allocno_class
,
990 reg_class_t best_class
)
994 if (allocno_class
!= ALL_REGS
)
995 return allocno_class
;
997 if (best_class
!= ALL_REGS
)
1000 mode
= PSEUDO_REGNO_MODE (regno
);
1001 return FLOAT_MODE_P (mode
) || VECTOR_MODE_P (mode
) ? FP_REGS
: GENERAL_REGS
;
1005 aarch64_min_divisions_for_recip_mul (machine_mode mode
)
1007 if (GET_MODE_UNIT_SIZE (mode
) == 4)
1008 return aarch64_tune_params
.min_div_recip_mul_sf
;
1009 return aarch64_tune_params
.min_div_recip_mul_df
;
1013 aarch64_reassociation_width (unsigned opc ATTRIBUTE_UNUSED
,
1016 if (VECTOR_MODE_P (mode
))
1017 return aarch64_tune_params
.vec_reassoc_width
;
1018 if (INTEGRAL_MODE_P (mode
))
1019 return aarch64_tune_params
.int_reassoc_width
;
1020 if (FLOAT_MODE_P (mode
))
1021 return aarch64_tune_params
.fp_reassoc_width
;
1025 /* Provide a mapping from gcc register numbers to dwarf register numbers. */
1027 aarch64_dbx_register_number (unsigned regno
)
1029 if (GP_REGNUM_P (regno
))
1030 return AARCH64_DWARF_R0
+ regno
- R0_REGNUM
;
1031 else if (regno
== SP_REGNUM
)
1032 return AARCH64_DWARF_SP
;
1033 else if (FP_REGNUM_P (regno
))
1034 return AARCH64_DWARF_V0
+ regno
- V0_REGNUM
;
1036 /* Return values >= DWARF_FRAME_REGISTERS indicate that there is no
1037 equivalent DWARF register. */
1038 return DWARF_FRAME_REGISTERS
;
1041 /* Return TRUE if MODE is any of the large INT modes. */
1043 aarch64_vect_struct_mode_p (machine_mode mode
)
1045 return mode
== OImode
|| mode
== CImode
|| mode
== XImode
;
1048 /* Return TRUE if MODE is any of the vector modes. */
1050 aarch64_vector_mode_p (machine_mode mode
)
1052 return aarch64_vector_mode_supported_p (mode
)
1053 || aarch64_vect_struct_mode_p (mode
);
1056 /* Implement target hook TARGET_ARRAY_MODE_SUPPORTED_P. */
1058 aarch64_array_mode_supported_p (machine_mode mode
,
1059 unsigned HOST_WIDE_INT nelems
)
1062 && (AARCH64_VALID_SIMD_QREG_MODE (mode
)
1063 || AARCH64_VALID_SIMD_DREG_MODE (mode
))
1064 && (nelems
>= 2 && nelems
<= 4))
1070 /* Implement HARD_REGNO_NREGS. */
1073 aarch64_hard_regno_nregs (unsigned regno
, machine_mode mode
)
1075 switch (aarch64_regno_regclass (regno
))
1079 return (GET_MODE_SIZE (mode
) + UNITS_PER_VREG
- 1) / UNITS_PER_VREG
;
1081 return (GET_MODE_SIZE (mode
) + UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
1086 /* Implement TARGET_HARD_REGNO_MODE_OK. */
1089 aarch64_hard_regno_mode_ok (unsigned regno
, machine_mode mode
)
1091 if (GET_MODE_CLASS (mode
) == MODE_CC
)
1092 return regno
== CC_REGNUM
;
1094 if (regno
== SP_REGNUM
)
1095 /* The purpose of comparing with ptr_mode is to support the
1096 global register variable associated with the stack pointer
1097 register via the syntax of asm ("wsp") in ILP32. */
1098 return mode
== Pmode
|| mode
== ptr_mode
;
1100 if (regno
== FRAME_POINTER_REGNUM
|| regno
== ARG_POINTER_REGNUM
)
1101 return mode
== Pmode
;
1103 if (GP_REGNUM_P (regno
) && ! aarch64_vect_struct_mode_p (mode
))
1106 if (FP_REGNUM_P (regno
))
1108 if (aarch64_vect_struct_mode_p (mode
))
1109 return end_hard_regno (mode
, regno
) - 1 <= V31_REGNUM
;
1117 /* Implement TARGET_HARD_REGNO_CALL_PART_CLOBBERED. The callee only saves
1118 the lower 64 bits of a 128-bit register. Tell the compiler the callee
1119 clobbers the top 64 bits when restoring the bottom 64 bits. */
1122 aarch64_hard_regno_call_part_clobbered (unsigned int regno
, machine_mode mode
)
1124 return FP_REGNUM_P (regno
) && GET_MODE_SIZE (mode
) > 8;
1127 /* Implement HARD_REGNO_CALLER_SAVE_MODE. */
1129 aarch64_hard_regno_caller_save_mode (unsigned regno
, unsigned nregs
,
1132 /* Handle modes that fit within single registers. */
1133 if (nregs
== 1 && GET_MODE_SIZE (mode
) <= 16)
1135 if (GET_MODE_SIZE (mode
) >= 4)
1140 /* Fall back to generic for multi-reg and very large modes. */
1142 return choose_hard_reg_mode (regno
, nregs
, false);
1145 /* Return true if calls to DECL should be treated as
1146 long-calls (ie called via a register). */
1148 aarch64_decl_is_long_call_p (const_tree decl ATTRIBUTE_UNUSED
)
1153 /* Return true if calls to symbol-ref SYM should be treated as
1154 long-calls (ie called via a register). */
1156 aarch64_is_long_call_p (rtx sym
)
1158 return aarch64_decl_is_long_call_p (SYMBOL_REF_DECL (sym
));
1161 /* Return true if calls to symbol-ref SYM should not go through
1165 aarch64_is_noplt_call_p (rtx sym
)
1167 const_tree decl
= SYMBOL_REF_DECL (sym
);
1172 || lookup_attribute ("noplt", DECL_ATTRIBUTES (decl
)))
1173 && !targetm
.binds_local_p (decl
))
1179 /* Return true if the offsets to a zero/sign-extract operation
1180 represent an expression that matches an extend operation. The
1181 operands represent the paramters from
1183 (extract:MODE (mult (reg) (MULT_IMM)) (EXTRACT_IMM) (const_int 0)). */
1185 aarch64_is_extend_from_extract (scalar_int_mode mode
, rtx mult_imm
,
1188 HOST_WIDE_INT mult_val
, extract_val
;
1190 if (! CONST_INT_P (mult_imm
) || ! CONST_INT_P (extract_imm
))
1193 mult_val
= INTVAL (mult_imm
);
1194 extract_val
= INTVAL (extract_imm
);
1197 && extract_val
< GET_MODE_BITSIZE (mode
)
1198 && exact_log2 (extract_val
& ~7) > 0
1199 && (extract_val
& 7) <= 4
1200 && mult_val
== (1 << (extract_val
& 7)))
1206 /* Emit an insn that's a simple single-set. Both the operands must be
1207 known to be valid. */
1208 inline static rtx_insn
*
1209 emit_set_insn (rtx x
, rtx y
)
1211 return emit_insn (gen_rtx_SET (x
, y
));
1214 /* X and Y are two things to compare using CODE. Emit the compare insn and
1215 return the rtx for register 0 in the proper mode. */
1217 aarch64_gen_compare_reg (RTX_CODE code
, rtx x
, rtx y
)
1219 machine_mode mode
= SELECT_CC_MODE (code
, x
, y
);
1220 rtx cc_reg
= gen_rtx_REG (mode
, CC_REGNUM
);
1222 emit_set_insn (cc_reg
, gen_rtx_COMPARE (mode
, x
, y
));
1226 /* Build the SYMBOL_REF for __tls_get_addr. */
1228 static GTY(()) rtx tls_get_addr_libfunc
;
1231 aarch64_tls_get_addr (void)
1233 if (!tls_get_addr_libfunc
)
1234 tls_get_addr_libfunc
= init_one_libfunc ("__tls_get_addr");
1235 return tls_get_addr_libfunc
;
1238 /* Return the TLS model to use for ADDR. */
1240 static enum tls_model
1241 tls_symbolic_operand_type (rtx addr
)
1243 enum tls_model tls_kind
= TLS_MODEL_NONE
;
1246 if (GET_CODE (addr
) == CONST
)
1248 split_const (addr
, &sym
, &addend
);
1249 if (GET_CODE (sym
) == SYMBOL_REF
)
1250 tls_kind
= SYMBOL_REF_TLS_MODEL (sym
);
1252 else if (GET_CODE (addr
) == SYMBOL_REF
)
1253 tls_kind
= SYMBOL_REF_TLS_MODEL (addr
);
1258 /* We'll allow lo_sum's in addresses in our legitimate addresses
1259 so that combine would take care of combining addresses where
1260 necessary, but for generation purposes, we'll generate the address
1263 tmp = hi (symbol_ref); adrp x1, foo
1264 dest = lo_sum (tmp, symbol_ref); add dest, x1, :lo_12:foo
1268 adrp x1, :got:foo adrp tmp, :tlsgd:foo
1269 ldr x1, [:got_lo12:foo] add dest, tmp, :tlsgd_lo12:foo
1273 Load TLS symbol, depending on TLS mechanism and TLS access model.
1275 Global Dynamic - Traditional TLS:
1276 adrp tmp, :tlsgd:imm
1277 add dest, tmp, #:tlsgd_lo12:imm
1280 Global Dynamic - TLS Descriptors:
1281 adrp dest, :tlsdesc:imm
1282 ldr tmp, [dest, #:tlsdesc_lo12:imm]
1283 add dest, dest, #:tlsdesc_lo12:imm
1290 adrp tmp, :gottprel:imm
1291 ldr dest, [tmp, #:gottprel_lo12:imm]
1296 add t0, tp, #:tprel_hi12:imm, lsl #12
1297 add t0, t0, #:tprel_lo12_nc:imm
1301 aarch64_load_symref_appropriately (rtx dest
, rtx imm
,
1302 enum aarch64_symbol_type type
)
1306 case SYMBOL_SMALL_ABSOLUTE
:
1308 /* In ILP32, the mode of dest can be either SImode or DImode. */
1310 machine_mode mode
= GET_MODE (dest
);
1312 gcc_assert (mode
== Pmode
|| mode
== ptr_mode
);
1314 if (can_create_pseudo_p ())
1315 tmp_reg
= gen_reg_rtx (mode
);
1317 emit_move_insn (tmp_reg
, gen_rtx_HIGH (mode
, imm
));
1318 emit_insn (gen_add_losym (dest
, tmp_reg
, imm
));
1322 case SYMBOL_TINY_ABSOLUTE
:
1323 emit_insn (gen_rtx_SET (dest
, imm
));
1326 case SYMBOL_SMALL_GOT_28K
:
1328 machine_mode mode
= GET_MODE (dest
);
1329 rtx gp_rtx
= pic_offset_table_rtx
;
1333 /* NOTE: pic_offset_table_rtx can be NULL_RTX, because we can reach
1334 here before rtl expand. Tree IVOPT will generate rtl pattern to
1335 decide rtx costs, in which case pic_offset_table_rtx is not
1336 initialized. For that case no need to generate the first adrp
1337 instruction as the final cost for global variable access is
1341 /* -fpic for -mcmodel=small allow 32K GOT table size (but we are
1342 using the page base as GOT base, the first page may be wasted,
1343 in the worst scenario, there is only 28K space for GOT).
1345 The generate instruction sequence for accessing global variable
1348 ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym]
1350 Only one instruction needed. But we must initialize
1351 pic_offset_table_rtx properly. We generate initialize insn for
1352 every global access, and allow CSE to remove all redundant.
1354 The final instruction sequences will look like the following
1355 for multiply global variables access.
1357 adrp pic_offset_table_rtx, _GLOBAL_OFFSET_TABLE_
1359 ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym1]
1360 ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym2]
1361 ldr reg, [pic_offset_table_rtx, #:gotpage_lo15:sym3]
1364 rtx s
= gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
1365 crtl
->uses_pic_offset_table
= 1;
1366 emit_move_insn (gp_rtx
, gen_rtx_HIGH (Pmode
, s
));
1368 if (mode
!= GET_MODE (gp_rtx
))
1369 gp_rtx
= gen_lowpart (mode
, gp_rtx
);
1373 if (mode
== ptr_mode
)
1376 insn
= gen_ldr_got_small_28k_di (dest
, gp_rtx
, imm
);
1378 insn
= gen_ldr_got_small_28k_si (dest
, gp_rtx
, imm
);
1380 mem
= XVECEXP (SET_SRC (insn
), 0, 0);
1384 gcc_assert (mode
== Pmode
);
1386 insn
= gen_ldr_got_small_28k_sidi (dest
, gp_rtx
, imm
);
1387 mem
= XVECEXP (XEXP (SET_SRC (insn
), 0), 0, 0);
1390 /* The operand is expected to be MEM. Whenever the related insn
1391 pattern changed, above code which calculate mem should be
1393 gcc_assert (GET_CODE (mem
) == MEM
);
1394 MEM_READONLY_P (mem
) = 1;
1395 MEM_NOTRAP_P (mem
) = 1;
1400 case SYMBOL_SMALL_GOT_4G
:
1402 /* In ILP32, the mode of dest can be either SImode or DImode,
1403 while the got entry is always of SImode size. The mode of
1404 dest depends on how dest is used: if dest is assigned to a
1405 pointer (e.g. in the memory), it has SImode; it may have
1406 DImode if dest is dereferenced to access the memeory.
1407 This is why we have to handle three different ldr_got_small
1408 patterns here (two patterns for ILP32). */
1413 machine_mode mode
= GET_MODE (dest
);
1415 if (can_create_pseudo_p ())
1416 tmp_reg
= gen_reg_rtx (mode
);
1418 emit_move_insn (tmp_reg
, gen_rtx_HIGH (mode
, imm
));
1419 if (mode
== ptr_mode
)
1422 insn
= gen_ldr_got_small_di (dest
, tmp_reg
, imm
);
1424 insn
= gen_ldr_got_small_si (dest
, tmp_reg
, imm
);
1426 mem
= XVECEXP (SET_SRC (insn
), 0, 0);
1430 gcc_assert (mode
== Pmode
);
1432 insn
= gen_ldr_got_small_sidi (dest
, tmp_reg
, imm
);
1433 mem
= XVECEXP (XEXP (SET_SRC (insn
), 0), 0, 0);
1436 gcc_assert (GET_CODE (mem
) == MEM
);
1437 MEM_READONLY_P (mem
) = 1;
1438 MEM_NOTRAP_P (mem
) = 1;
1443 case SYMBOL_SMALL_TLSGD
:
1446 machine_mode mode
= GET_MODE (dest
);
1447 rtx result
= gen_rtx_REG (mode
, R0_REGNUM
);
1451 aarch64_emit_call_insn (gen_tlsgd_small_si (result
, imm
));
1453 aarch64_emit_call_insn (gen_tlsgd_small_di (result
, imm
));
1454 insns
= get_insns ();
1457 RTL_CONST_CALL_P (insns
) = 1;
1458 emit_libcall_block (insns
, dest
, result
, imm
);
1462 case SYMBOL_SMALL_TLSDESC
:
1464 machine_mode mode
= GET_MODE (dest
);
1465 rtx x0
= gen_rtx_REG (mode
, R0_REGNUM
);
1468 gcc_assert (mode
== Pmode
|| mode
== ptr_mode
);
1470 /* In ILP32, the got entry is always of SImode size. Unlike
1471 small GOT, the dest is fixed at reg 0. */
1473 emit_insn (gen_tlsdesc_small_si (imm
));
1475 emit_insn (gen_tlsdesc_small_di (imm
));
1476 tp
= aarch64_load_tp (NULL
);
1479 tp
= gen_lowpart (mode
, tp
);
1481 emit_insn (gen_rtx_SET (dest
, gen_rtx_PLUS (mode
, tp
, x0
)));
1482 set_unique_reg_note (get_last_insn (), REG_EQUIV
, imm
);
1486 case SYMBOL_SMALL_TLSIE
:
1488 /* In ILP32, the mode of dest can be either SImode or DImode,
1489 while the got entry is always of SImode size. The mode of
1490 dest depends on how dest is used: if dest is assigned to a
1491 pointer (e.g. in the memory), it has SImode; it may have
1492 DImode if dest is dereferenced to access the memeory.
1493 This is why we have to handle three different tlsie_small
1494 patterns here (two patterns for ILP32). */
1495 machine_mode mode
= GET_MODE (dest
);
1496 rtx tmp_reg
= gen_reg_rtx (mode
);
1497 rtx tp
= aarch64_load_tp (NULL
);
1499 if (mode
== ptr_mode
)
1502 emit_insn (gen_tlsie_small_di (tmp_reg
, imm
));
1505 emit_insn (gen_tlsie_small_si (tmp_reg
, imm
));
1506 tp
= gen_lowpart (mode
, tp
);
1511 gcc_assert (mode
== Pmode
);
1512 emit_insn (gen_tlsie_small_sidi (tmp_reg
, imm
));
1515 emit_insn (gen_rtx_SET (dest
, gen_rtx_PLUS (mode
, tp
, tmp_reg
)));
1516 set_unique_reg_note (get_last_insn (), REG_EQUIV
, imm
);
1520 case SYMBOL_TLSLE12
:
1521 case SYMBOL_TLSLE24
:
1522 case SYMBOL_TLSLE32
:
1523 case SYMBOL_TLSLE48
:
1525 machine_mode mode
= GET_MODE (dest
);
1526 rtx tp
= aarch64_load_tp (NULL
);
1529 tp
= gen_lowpart (mode
, tp
);
1533 case SYMBOL_TLSLE12
:
1534 emit_insn ((mode
== DImode
? gen_tlsle12_di
: gen_tlsle12_si
)
1537 case SYMBOL_TLSLE24
:
1538 emit_insn ((mode
== DImode
? gen_tlsle24_di
: gen_tlsle24_si
)
1541 case SYMBOL_TLSLE32
:
1542 emit_insn ((mode
== DImode
? gen_tlsle32_di
: gen_tlsle32_si
)
1544 emit_insn ((mode
== DImode
? gen_adddi3
: gen_addsi3
)
1547 case SYMBOL_TLSLE48
:
1548 emit_insn ((mode
== DImode
? gen_tlsle48_di
: gen_tlsle48_si
)
1550 emit_insn ((mode
== DImode
? gen_adddi3
: gen_addsi3
)
1557 set_unique_reg_note (get_last_insn (), REG_EQUIV
, imm
);
1561 case SYMBOL_TINY_GOT
:
1562 emit_insn (gen_ldr_got_tiny (dest
, imm
));
1565 case SYMBOL_TINY_TLSIE
:
1567 machine_mode mode
= GET_MODE (dest
);
1568 rtx tp
= aarch64_load_tp (NULL
);
1570 if (mode
== ptr_mode
)
1573 emit_insn (gen_tlsie_tiny_di (dest
, imm
, tp
));
1576 tp
= gen_lowpart (mode
, tp
);
1577 emit_insn (gen_tlsie_tiny_si (dest
, imm
, tp
));
1582 gcc_assert (mode
== Pmode
);
1583 emit_insn (gen_tlsie_tiny_sidi (dest
, imm
, tp
));
1586 set_unique_reg_note (get_last_insn (), REG_EQUIV
, imm
);
1595 /* Emit a move from SRC to DEST. Assume that the move expanders can
1596 handle all moves if !can_create_pseudo_p (). The distinction is
1597 important because, unlike emit_move_insn, the move expanders know
1598 how to force Pmode objects into the constant pool even when the
1599 constant pool address is not itself legitimate. */
1601 aarch64_emit_move (rtx dest
, rtx src
)
1603 return (can_create_pseudo_p ()
1604 ? emit_move_insn (dest
, src
)
1605 : emit_move_insn_1 (dest
, src
));
1608 /* Split a 128-bit move operation into two 64-bit move operations,
1609 taking care to handle partial overlap of register to register
1610 copies. Special cases are needed when moving between GP regs and
1611 FP regs. SRC can be a register, constant or memory; DST a register
1612 or memory. If either operand is memory it must not have any side
1615 aarch64_split_128bit_move (rtx dst
, rtx src
)
1620 machine_mode mode
= GET_MODE (dst
);
1622 gcc_assert (mode
== TImode
|| mode
== TFmode
);
1623 gcc_assert (!(side_effects_p (src
) || side_effects_p (dst
)));
1624 gcc_assert (mode
== GET_MODE (src
) || GET_MODE (src
) == VOIDmode
);
1626 if (REG_P (dst
) && REG_P (src
))
1628 int src_regno
= REGNO (src
);
1629 int dst_regno
= REGNO (dst
);
1631 /* Handle FP <-> GP regs. */
1632 if (FP_REGNUM_P (dst_regno
) && GP_REGNUM_P (src_regno
))
1634 src_lo
= gen_lowpart (word_mode
, src
);
1635 src_hi
= gen_highpart (word_mode
, src
);
1639 emit_insn (gen_aarch64_movtilow_di (dst
, src_lo
));
1640 emit_insn (gen_aarch64_movtihigh_di (dst
, src_hi
));
1644 emit_insn (gen_aarch64_movtflow_di (dst
, src_lo
));
1645 emit_insn (gen_aarch64_movtfhigh_di (dst
, src_hi
));
1649 else if (GP_REGNUM_P (dst_regno
) && FP_REGNUM_P (src_regno
))
1651 dst_lo
= gen_lowpart (word_mode
, dst
);
1652 dst_hi
= gen_highpart (word_mode
, dst
);
1656 emit_insn (gen_aarch64_movdi_tilow (dst_lo
, src
));
1657 emit_insn (gen_aarch64_movdi_tihigh (dst_hi
, src
));
1661 emit_insn (gen_aarch64_movdi_tflow (dst_lo
, src
));
1662 emit_insn (gen_aarch64_movdi_tfhigh (dst_hi
, src
));
1668 dst_lo
= gen_lowpart (word_mode
, dst
);
1669 dst_hi
= gen_highpart (word_mode
, dst
);
1670 src_lo
= gen_lowpart (word_mode
, src
);
1671 src_hi
= gen_highpart_mode (word_mode
, mode
, src
);
1673 /* At most one pairing may overlap. */
1674 if (reg_overlap_mentioned_p (dst_lo
, src_hi
))
1676 aarch64_emit_move (dst_hi
, src_hi
);
1677 aarch64_emit_move (dst_lo
, src_lo
);
1681 aarch64_emit_move (dst_lo
, src_lo
);
1682 aarch64_emit_move (dst_hi
, src_hi
);
1687 aarch64_split_128bit_move_p (rtx dst
, rtx src
)
1689 return (! REG_P (src
)
1690 || ! (FP_REGNUM_P (REGNO (dst
)) && FP_REGNUM_P (REGNO (src
))));
1693 /* Split a complex SIMD combine. */
1696 aarch64_split_simd_combine (rtx dst
, rtx src1
, rtx src2
)
1698 machine_mode src_mode
= GET_MODE (src1
);
1699 machine_mode dst_mode
= GET_MODE (dst
);
1701 gcc_assert (VECTOR_MODE_P (dst_mode
));
1702 gcc_assert (register_operand (dst
, dst_mode
)
1703 && register_operand (src1
, src_mode
)
1704 && register_operand (src2
, src_mode
));
1706 rtx (*gen
) (rtx
, rtx
, rtx
);
1711 gen
= gen_aarch64_simd_combinev8qi
;
1714 gen
= gen_aarch64_simd_combinev4hi
;
1717 gen
= gen_aarch64_simd_combinev2si
;
1720 gen
= gen_aarch64_simd_combinev4hf
;
1723 gen
= gen_aarch64_simd_combinev2sf
;
1726 gen
= gen_aarch64_simd_combinedi
;
1729 gen
= gen_aarch64_simd_combinedf
;
1735 emit_insn (gen (dst
, src1
, src2
));
1739 /* Split a complex SIMD move. */
1742 aarch64_split_simd_move (rtx dst
, rtx src
)
1744 machine_mode src_mode
= GET_MODE (src
);
1745 machine_mode dst_mode
= GET_MODE (dst
);
1747 gcc_assert (VECTOR_MODE_P (dst_mode
));
1749 if (REG_P (dst
) && REG_P (src
))
1751 rtx (*gen
) (rtx
, rtx
);
1753 gcc_assert (VECTOR_MODE_P (src_mode
));
1758 gen
= gen_aarch64_split_simd_movv16qi
;
1761 gen
= gen_aarch64_split_simd_movv8hi
;
1764 gen
= gen_aarch64_split_simd_movv4si
;
1767 gen
= gen_aarch64_split_simd_movv2di
;
1770 gen
= gen_aarch64_split_simd_movv8hf
;
1773 gen
= gen_aarch64_split_simd_movv4sf
;
1776 gen
= gen_aarch64_split_simd_movv2df
;
1782 emit_insn (gen (dst
, src
));
1788 aarch64_zero_extend_const_eq (machine_mode xmode
, rtx x
,
1789 machine_mode ymode
, rtx y
)
1791 rtx r
= simplify_const_unary_operation (ZERO_EXTEND
, xmode
, y
, ymode
);
1792 gcc_assert (r
!= NULL
);
1793 return rtx_equal_p (x
, r
);
1798 aarch64_force_temporary (machine_mode mode
, rtx x
, rtx value
)
1800 if (can_create_pseudo_p ())
1801 return force_reg (mode
, value
);
1804 x
= aarch64_emit_move (x
, value
);
1811 aarch64_add_offset (scalar_int_mode mode
, rtx temp
, rtx reg
,
1812 HOST_WIDE_INT offset
)
1814 if (!aarch64_plus_immediate (GEN_INT (offset
), mode
))
1817 /* Load the full offset into a register. This
1818 might be improvable in the future. */
1819 high
= GEN_INT (offset
);
1821 high
= aarch64_force_temporary (mode
, temp
, high
);
1822 reg
= aarch64_force_temporary (mode
, temp
,
1823 gen_rtx_PLUS (mode
, high
, reg
));
1825 return plus_constant (mode
, reg
, offset
);
1829 aarch64_internal_mov_immediate (rtx dest
, rtx imm
, bool generate
,
1830 scalar_int_mode mode
)
1833 unsigned HOST_WIDE_INT val
, val2
, mask
;
1834 int one_match
, zero_match
;
1839 if (aarch64_move_imm (val
, mode
))
1842 emit_insn (gen_rtx_SET (dest
, imm
));
1846 /* Check to see if the low 32 bits are either 0xffffXXXX or 0xXXXXffff
1847 (with XXXX non-zero). In that case check to see if the move can be done in
1849 val2
= val
& 0xffffffff;
1851 && aarch64_move_imm (val2
, SImode
)
1852 && (((val
>> 32) & 0xffff) == 0 || (val
>> 48) == 0))
1855 emit_insn (gen_rtx_SET (dest
, GEN_INT (val2
)));
1857 /* Check if we have to emit a second instruction by checking to see
1858 if any of the upper 32 bits of the original DI mode value is set. */
1862 i
= (val
>> 48) ? 48 : 32;
1865 emit_insn (gen_insv_immdi (dest
, GEN_INT (i
),
1866 GEN_INT ((val
>> i
) & 0xffff)));
1871 if ((val
>> 32) == 0 || mode
== SImode
)
1875 emit_insn (gen_rtx_SET (dest
, GEN_INT (val
& 0xffff)));
1877 emit_insn (gen_insv_immsi (dest
, GEN_INT (16),
1878 GEN_INT ((val
>> 16) & 0xffff)));
1880 emit_insn (gen_insv_immdi (dest
, GEN_INT (16),
1881 GEN_INT ((val
>> 16) & 0xffff)));
1886 /* Remaining cases are all for DImode. */
1889 zero_match
= ((val
& mask
) == 0) + ((val
& (mask
<< 16)) == 0) +
1890 ((val
& (mask
<< 32)) == 0) + ((val
& (mask
<< 48)) == 0);
1891 one_match
= ((~val
& mask
) == 0) + ((~val
& (mask
<< 16)) == 0) +
1892 ((~val
& (mask
<< 32)) == 0) + ((~val
& (mask
<< 48)) == 0);
1894 if (zero_match
!= 2 && one_match
!= 2)
1896 /* Try emitting a bitmask immediate with a movk replacing 16 bits.
1897 For a 64-bit bitmask try whether changing 16 bits to all ones or
1898 zeroes creates a valid bitmask. To check any repeated bitmask,
1899 try using 16 bits from the other 32-bit half of val. */
1901 for (i
= 0; i
< 64; i
+= 16, mask
<<= 16)
1904 if (val2
!= val
&& aarch64_bitmask_imm (val2
, mode
))
1907 if (val2
!= val
&& aarch64_bitmask_imm (val2
, mode
))
1909 val2
= val2
& ~mask
;
1910 val2
= val2
| (((val2
>> 32) | (val2
<< 32)) & mask
);
1911 if (val2
!= val
&& aarch64_bitmask_imm (val2
, mode
))
1918 emit_insn (gen_rtx_SET (dest
, GEN_INT (val2
)));
1919 emit_insn (gen_insv_immdi (dest
, GEN_INT (i
),
1920 GEN_INT ((val
>> i
) & 0xffff)));
1926 /* Generate 2-4 instructions, skipping 16 bits of all zeroes or ones which
1927 are emitted by the initial mov. If one_match > zero_match, skip set bits,
1928 otherwise skip zero bits. */
1932 val2
= one_match
> zero_match
? ~val
: val
;
1933 i
= (val2
& mask
) != 0 ? 0 : (val2
& (mask
<< 16)) != 0 ? 16 : 32;
1936 emit_insn (gen_rtx_SET (dest
, GEN_INT (one_match
> zero_match
1937 ? (val
| ~(mask
<< i
))
1938 : (val
& (mask
<< i
)))));
1939 for (i
+= 16; i
< 64; i
+= 16)
1941 if ((val2
& (mask
<< i
)) == 0)
1944 emit_insn (gen_insv_immdi (dest
, GEN_INT (i
),
1945 GEN_INT ((val
>> i
) & 0xffff)));
1954 aarch64_expand_mov_immediate (rtx dest
, rtx imm
)
1956 machine_mode mode
= GET_MODE (dest
);
1958 gcc_assert (mode
== SImode
|| mode
== DImode
);
1960 /* Check on what type of symbol it is. */
1961 scalar_int_mode int_mode
;
1962 if ((GET_CODE (imm
) == SYMBOL_REF
1963 || GET_CODE (imm
) == LABEL_REF
1964 || GET_CODE (imm
) == CONST
)
1965 && is_a
<scalar_int_mode
> (mode
, &int_mode
))
1967 rtx mem
, base
, offset
;
1968 enum aarch64_symbol_type sty
;
1970 /* If we have (const (plus symbol offset)), separate out the offset
1971 before we start classifying the symbol. */
1972 split_const (imm
, &base
, &offset
);
1974 sty
= aarch64_classify_symbol (base
, offset
);
1977 case SYMBOL_FORCE_TO_MEM
:
1978 if (offset
!= const0_rtx
1979 && targetm
.cannot_force_const_mem (int_mode
, imm
))
1981 gcc_assert (can_create_pseudo_p ());
1982 base
= aarch64_force_temporary (int_mode
, dest
, base
);
1983 base
= aarch64_add_offset (int_mode
, NULL
, base
,
1985 aarch64_emit_move (dest
, base
);
1989 mem
= force_const_mem (ptr_mode
, imm
);
1992 /* If we aren't generating PC relative literals, then
1993 we need to expand the literal pool access carefully.
1994 This is something that needs to be done in a number
1995 of places, so could well live as a separate function. */
1996 if (!aarch64_pcrelative_literal_loads
)
1998 gcc_assert (can_create_pseudo_p ());
1999 base
= gen_reg_rtx (ptr_mode
);
2000 aarch64_expand_mov_immediate (base
, XEXP (mem
, 0));
2001 if (ptr_mode
!= Pmode
)
2002 base
= convert_memory_address (Pmode
, base
);
2003 mem
= gen_rtx_MEM (ptr_mode
, base
);
2006 if (int_mode
!= ptr_mode
)
2007 mem
= gen_rtx_ZERO_EXTEND (int_mode
, mem
);
2009 emit_insn (gen_rtx_SET (dest
, mem
));
2013 case SYMBOL_SMALL_TLSGD
:
2014 case SYMBOL_SMALL_TLSDESC
:
2015 case SYMBOL_SMALL_TLSIE
:
2016 case SYMBOL_SMALL_GOT_28K
:
2017 case SYMBOL_SMALL_GOT_4G
:
2018 case SYMBOL_TINY_GOT
:
2019 case SYMBOL_TINY_TLSIE
:
2020 if (offset
!= const0_rtx
)
2022 gcc_assert(can_create_pseudo_p ());
2023 base
= aarch64_force_temporary (int_mode
, dest
, base
);
2024 base
= aarch64_add_offset (int_mode
, NULL
, base
,
2026 aarch64_emit_move (dest
, base
);
2031 case SYMBOL_SMALL_ABSOLUTE
:
2032 case SYMBOL_TINY_ABSOLUTE
:
2033 case SYMBOL_TLSLE12
:
2034 case SYMBOL_TLSLE24
:
2035 case SYMBOL_TLSLE32
:
2036 case SYMBOL_TLSLE48
:
2037 aarch64_load_symref_appropriately (dest
, imm
, sty
);
2045 if (!CONST_INT_P (imm
))
2047 if (GET_CODE (imm
) == HIGH
)
2048 emit_insn (gen_rtx_SET (dest
, imm
));
2051 rtx mem
= force_const_mem (mode
, imm
);
2053 emit_insn (gen_rtx_SET (dest
, mem
));
2059 aarch64_internal_mov_immediate (dest
, imm
, true,
2060 as_a
<scalar_int_mode
> (mode
));
2063 /* Add DELTA to REGNUM in mode MODE. SCRATCHREG can be used to hold a
2064 temporary value if necessary. FRAME_RELATED_P should be true if
2065 the RTX_FRAME_RELATED flag should be set and CFA adjustments added
2066 to the generated instructions. If SCRATCHREG is known to hold
2067 abs (delta), EMIT_MOVE_IMM can be set to false to avoid emitting the
2070 Since this function may be used to adjust the stack pointer, we must
2071 ensure that it cannot cause transient stack deallocation (for example
2072 by first incrementing SP and then decrementing when adjusting by a
2073 large immediate). */
2076 aarch64_add_constant_internal (scalar_int_mode mode
, int regnum
,
2077 int scratchreg
, HOST_WIDE_INT delta
,
2078 bool frame_related_p
, bool emit_move_imm
)
2080 HOST_WIDE_INT mdelta
= abs_hwi (delta
);
2081 rtx this_rtx
= gen_rtx_REG (mode
, regnum
);
2087 /* Single instruction adjustment. */
2088 if (aarch64_uimm12_shift (mdelta
))
2090 insn
= emit_insn (gen_add2_insn (this_rtx
, GEN_INT (delta
)));
2091 RTX_FRAME_RELATED_P (insn
) = frame_related_p
;
2095 /* Emit 2 additions/subtractions if the adjustment is less than 24 bits.
2096 Only do this if mdelta is not a 16-bit move as adjusting using a move
2098 if (mdelta
< 0x1000000 && !aarch64_move_imm (mdelta
, mode
))
2100 HOST_WIDE_INT low_off
= mdelta
& 0xfff;
2102 low_off
= delta
< 0 ? -low_off
: low_off
;
2103 insn
= emit_insn (gen_add2_insn (this_rtx
, GEN_INT (low_off
)));
2104 RTX_FRAME_RELATED_P (insn
) = frame_related_p
;
2105 insn
= emit_insn (gen_add2_insn (this_rtx
, GEN_INT (delta
- low_off
)));
2106 RTX_FRAME_RELATED_P (insn
) = frame_related_p
;
2110 /* Emit a move immediate if required and an addition/subtraction. */
2111 rtx scratch_rtx
= gen_rtx_REG (mode
, scratchreg
);
2113 aarch64_internal_mov_immediate (scratch_rtx
, GEN_INT (mdelta
), true, mode
);
2114 insn
= emit_insn (delta
< 0 ? gen_sub2_insn (this_rtx
, scratch_rtx
)
2115 : gen_add2_insn (this_rtx
, scratch_rtx
));
2116 if (frame_related_p
)
2118 RTX_FRAME_RELATED_P (insn
) = frame_related_p
;
2119 rtx adj
= plus_constant (mode
, this_rtx
, delta
);
2120 add_reg_note (insn
, REG_CFA_ADJUST_CFA
, gen_rtx_SET (this_rtx
, adj
));
2125 aarch64_add_constant (scalar_int_mode mode
, int regnum
, int scratchreg
,
2126 HOST_WIDE_INT delta
)
2128 aarch64_add_constant_internal (mode
, regnum
, scratchreg
, delta
, false, true);
2132 aarch64_add_sp (int scratchreg
, HOST_WIDE_INT delta
, bool emit_move_imm
)
2134 aarch64_add_constant_internal (Pmode
, SP_REGNUM
, scratchreg
, delta
,
2135 true, emit_move_imm
);
2139 aarch64_sub_sp (int scratchreg
, HOST_WIDE_INT delta
, bool frame_related_p
)
2141 aarch64_add_constant_internal (Pmode
, SP_REGNUM
, scratchreg
, -delta
,
2142 frame_related_p
, true);
2146 aarch64_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED
,
2147 tree exp ATTRIBUTE_UNUSED
)
2149 /* Currently, always true. */
2153 /* Implement TARGET_PASS_BY_REFERENCE. */
2156 aarch64_pass_by_reference (cumulative_args_t pcum ATTRIBUTE_UNUSED
,
2159 bool named ATTRIBUTE_UNUSED
)
2162 machine_mode dummymode
;
2165 /* GET_MODE_SIZE (BLKmode) is useless since it is 0. */
2166 size
= (mode
== BLKmode
&& type
)
2167 ? int_size_in_bytes (type
) : (int) GET_MODE_SIZE (mode
);
2169 /* Aggregates are passed by reference based on their size. */
2170 if (type
&& AGGREGATE_TYPE_P (type
))
2172 size
= int_size_in_bytes (type
);
2175 /* Variable sized arguments are always returned by reference. */
2179 /* Can this be a candidate to be passed in fp/simd register(s)? */
2180 if (aarch64_vfp_is_call_or_return_candidate (mode
, type
,
2185 /* Arguments which are variable sized or larger than 2 registers are
2186 passed by reference unless they are a homogenous floating point
2188 return size
> 2 * UNITS_PER_WORD
;
2191 /* Return TRUE if VALTYPE is padded to its least significant bits. */
2193 aarch64_return_in_msb (const_tree valtype
)
2195 machine_mode dummy_mode
;
2198 /* Never happens in little-endian mode. */
2199 if (!BYTES_BIG_ENDIAN
)
2202 /* Only composite types smaller than or equal to 16 bytes can
2203 be potentially returned in registers. */
2204 if (!aarch64_composite_type_p (valtype
, TYPE_MODE (valtype
))
2205 || int_size_in_bytes (valtype
) <= 0
2206 || int_size_in_bytes (valtype
) > 16)
2209 /* But not a composite that is an HFA (Homogeneous Floating-point Aggregate)
2210 or an HVA (Homogeneous Short-Vector Aggregate); such a special composite
2211 is always passed/returned in the least significant bits of fp/simd
2213 if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (valtype
), valtype
,
2214 &dummy_mode
, &dummy_int
, NULL
))
2220 /* Implement TARGET_FUNCTION_VALUE.
2221 Define how to find the value returned by a function. */
2224 aarch64_function_value (const_tree type
, const_tree func
,
2225 bool outgoing ATTRIBUTE_UNUSED
)
2230 machine_mode ag_mode
;
2232 mode
= TYPE_MODE (type
);
2233 if (INTEGRAL_TYPE_P (type
))
2234 mode
= promote_function_mode (type
, mode
, &unsignedp
, func
, 1);
2236 if (aarch64_return_in_msb (type
))
2238 HOST_WIDE_INT size
= int_size_in_bytes (type
);
2240 if (size
% UNITS_PER_WORD
!= 0)
2242 size
+= UNITS_PER_WORD
- size
% UNITS_PER_WORD
;
2243 mode
= int_mode_for_size (size
* BITS_PER_UNIT
, 0).require ();
2247 if (aarch64_vfp_is_call_or_return_candidate (mode
, type
,
2248 &ag_mode
, &count
, NULL
))
2250 if (!aarch64_composite_type_p (type
, mode
))
2252 gcc_assert (count
== 1 && mode
== ag_mode
);
2253 return gen_rtx_REG (mode
, V0_REGNUM
);
2260 par
= gen_rtx_PARALLEL (mode
, rtvec_alloc (count
));
2261 for (i
= 0; i
< count
; i
++)
2263 rtx tmp
= gen_rtx_REG (ag_mode
, V0_REGNUM
+ i
);
2264 tmp
= gen_rtx_EXPR_LIST (VOIDmode
, tmp
,
2265 GEN_INT (i
* GET_MODE_SIZE (ag_mode
)));
2266 XVECEXP (par
, 0, i
) = tmp
;
2272 return gen_rtx_REG (mode
, R0_REGNUM
);
2275 /* Implements TARGET_FUNCTION_VALUE_REGNO_P.
2276 Return true if REGNO is the number of a hard register in which the values
2277 of called function may come back. */
2280 aarch64_function_value_regno_p (const unsigned int regno
)
2282 /* Maximum of 16 bytes can be returned in the general registers. Examples
2283 of 16-byte return values are: 128-bit integers and 16-byte small
2284 structures (excluding homogeneous floating-point aggregates). */
2285 if (regno
== R0_REGNUM
|| regno
== R1_REGNUM
)
2288 /* Up to four fp/simd registers can return a function value, e.g. a
2289 homogeneous floating-point aggregate having four members. */
2290 if (regno
>= V0_REGNUM
&& regno
< V0_REGNUM
+ HA_MAX_NUM_FLDS
)
2291 return TARGET_FLOAT
;
2296 /* Implement TARGET_RETURN_IN_MEMORY.
2298 If the type T of the result of a function is such that
2300 would require that arg be passed as a value in a register (or set of
2301 registers) according to the parameter passing rules, then the result
2302 is returned in the same registers as would be used for such an
2306 aarch64_return_in_memory (const_tree type
, const_tree fndecl ATTRIBUTE_UNUSED
)
2309 machine_mode ag_mode
;
2312 if (!AGGREGATE_TYPE_P (type
)
2313 && TREE_CODE (type
) != COMPLEX_TYPE
2314 && TREE_CODE (type
) != VECTOR_TYPE
)
2315 /* Simple scalar types always returned in registers. */
2318 if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (type
),
2325 /* Types larger than 2 registers returned in memory. */
2326 size
= int_size_in_bytes (type
);
2327 return (size
< 0 || size
> 2 * UNITS_PER_WORD
);
2331 aarch64_vfp_is_call_candidate (cumulative_args_t pcum_v
, machine_mode mode
,
2332 const_tree type
, int *nregs
)
2334 CUMULATIVE_ARGS
*pcum
= get_cumulative_args (pcum_v
);
2335 return aarch64_vfp_is_call_or_return_candidate (mode
,
2337 &pcum
->aapcs_vfp_rmode
,
2342 /* Given MODE and TYPE of a function argument, return the alignment in
2343 bits. The idea is to suppress any stronger alignment requested by
2344 the user and opt for the natural alignment (specified in AAPCS64 \S 4.1).
2345 This is a helper function for local use only. */
2348 aarch64_function_arg_alignment (machine_mode mode
, const_tree type
)
2351 return GET_MODE_ALIGNMENT (mode
);
2353 if (integer_zerop (TYPE_SIZE (type
)))
2356 gcc_assert (TYPE_MODE (type
) == mode
);
2358 if (!AGGREGATE_TYPE_P (type
))
2359 return TYPE_ALIGN (TYPE_MAIN_VARIANT (type
));
2361 if (TREE_CODE (type
) == ARRAY_TYPE
)
2362 return TYPE_ALIGN (TREE_TYPE (type
));
2364 unsigned int alignment
= 0;
2365 for (tree field
= TYPE_FIELDS (type
); field
; field
= DECL_CHAIN (field
))
2366 if (TREE_CODE (field
) == FIELD_DECL
)
2367 alignment
= std::max (alignment
, DECL_ALIGN (field
));
2372 /* Layout a function argument according to the AAPCS64 rules. The rule
2373 numbers refer to the rule numbers in the AAPCS64. */
2376 aarch64_layout_arg (cumulative_args_t pcum_v
, machine_mode mode
,
2378 bool named ATTRIBUTE_UNUSED
)
2380 CUMULATIVE_ARGS
*pcum
= get_cumulative_args (pcum_v
);
2381 int ncrn
, nvrn
, nregs
;
2382 bool allocate_ncrn
, allocate_nvrn
;
2385 /* We need to do this once per argument. */
2386 if (pcum
->aapcs_arg_processed
)
2389 pcum
->aapcs_arg_processed
= true;
2391 /* Size in bytes, rounded to the nearest multiple of 8 bytes. */
2393 = ROUND_UP (type
? int_size_in_bytes (type
) : GET_MODE_SIZE (mode
),
2396 allocate_ncrn
= (type
) ? !(FLOAT_TYPE_P (type
)) : !FLOAT_MODE_P (mode
);
2397 allocate_nvrn
= aarch64_vfp_is_call_candidate (pcum_v
,
2402 /* allocate_ncrn may be false-positive, but allocate_nvrn is quite reliable.
2403 The following code thus handles passing by SIMD/FP registers first. */
2405 nvrn
= pcum
->aapcs_nvrn
;
2407 /* C1 - C5 for floating point, homogenous floating point aggregates (HFA)
2408 and homogenous short-vector aggregates (HVA). */
2412 aarch64_err_no_fpadvsimd (mode
, "argument");
2414 if (nvrn
+ nregs
<= NUM_FP_ARG_REGS
)
2416 pcum
->aapcs_nextnvrn
= nvrn
+ nregs
;
2417 if (!aarch64_composite_type_p (type
, mode
))
2419 gcc_assert (nregs
== 1);
2420 pcum
->aapcs_reg
= gen_rtx_REG (mode
, V0_REGNUM
+ nvrn
);
2426 par
= gen_rtx_PARALLEL (mode
, rtvec_alloc (nregs
));
2427 for (i
= 0; i
< nregs
; i
++)
2429 rtx tmp
= gen_rtx_REG (pcum
->aapcs_vfp_rmode
,
2430 V0_REGNUM
+ nvrn
+ i
);
2431 tmp
= gen_rtx_EXPR_LIST
2433 GEN_INT (i
* GET_MODE_SIZE (pcum
->aapcs_vfp_rmode
)));
2434 XVECEXP (par
, 0, i
) = tmp
;
2436 pcum
->aapcs_reg
= par
;
2442 /* C.3 NSRN is set to 8. */
2443 pcum
->aapcs_nextnvrn
= NUM_FP_ARG_REGS
;
2448 ncrn
= pcum
->aapcs_ncrn
;
2449 nregs
= size
/ UNITS_PER_WORD
;
2451 /* C6 - C9. though the sign and zero extension semantics are
2452 handled elsewhere. This is the case where the argument fits
2453 entirely general registers. */
2454 if (allocate_ncrn
&& (ncrn
+ nregs
<= NUM_ARG_REGS
))
2457 gcc_assert (nregs
== 0 || nregs
== 1 || nregs
== 2);
2459 /* C.8 if the argument has an alignment of 16 then the NGRN is
2460 rounded up to the next even number. */
2463 /* The == 16 * BITS_PER_UNIT instead of >= 16 * BITS_PER_UNIT
2464 comparison is there because for > 16 * BITS_PER_UNIT
2465 alignment nregs should be > 2 and therefore it should be
2466 passed by reference rather than value. */
2467 && aarch64_function_arg_alignment (mode
, type
) == 16 * BITS_PER_UNIT
)
2470 gcc_assert (ncrn
+ nregs
<= NUM_ARG_REGS
);
2473 /* NREGS can be 0 when e.g. an empty structure is to be passed.
2474 A reg is still generated for it, but the caller should be smart
2475 enough not to use it. */
2476 if (nregs
== 0 || nregs
== 1 || GET_MODE_CLASS (mode
) == MODE_INT
)
2477 pcum
->aapcs_reg
= gen_rtx_REG (mode
, R0_REGNUM
+ ncrn
);
2483 par
= gen_rtx_PARALLEL (mode
, rtvec_alloc (nregs
));
2484 for (i
= 0; i
< nregs
; i
++)
2486 rtx tmp
= gen_rtx_REG (word_mode
, R0_REGNUM
+ ncrn
+ i
);
2487 tmp
= gen_rtx_EXPR_LIST (VOIDmode
, tmp
,
2488 GEN_INT (i
* UNITS_PER_WORD
));
2489 XVECEXP (par
, 0, i
) = tmp
;
2491 pcum
->aapcs_reg
= par
;
2494 pcum
->aapcs_nextncrn
= ncrn
+ nregs
;
2499 pcum
->aapcs_nextncrn
= NUM_ARG_REGS
;
2501 /* The argument is passed on stack; record the needed number of words for
2502 this argument and align the total size if necessary. */
2504 pcum
->aapcs_stack_words
= size
/ UNITS_PER_WORD
;
2506 if (aarch64_function_arg_alignment (mode
, type
) == 16 * BITS_PER_UNIT
)
2507 pcum
->aapcs_stack_size
= ROUND_UP (pcum
->aapcs_stack_size
,
2508 16 / UNITS_PER_WORD
);
2512 /* Implement TARGET_FUNCTION_ARG. */
2515 aarch64_function_arg (cumulative_args_t pcum_v
, machine_mode mode
,
2516 const_tree type
, bool named
)
2518 CUMULATIVE_ARGS
*pcum
= get_cumulative_args (pcum_v
);
2519 gcc_assert (pcum
->pcs_variant
== ARM_PCS_AAPCS64
);
2521 if (mode
== VOIDmode
)
2524 aarch64_layout_arg (pcum_v
, mode
, type
, named
);
2525 return pcum
->aapcs_reg
;
2529 aarch64_init_cumulative_args (CUMULATIVE_ARGS
*pcum
,
2530 const_tree fntype ATTRIBUTE_UNUSED
,
2531 rtx libname ATTRIBUTE_UNUSED
,
2532 const_tree fndecl ATTRIBUTE_UNUSED
,
2533 unsigned n_named ATTRIBUTE_UNUSED
)
2535 pcum
->aapcs_ncrn
= 0;
2536 pcum
->aapcs_nvrn
= 0;
2537 pcum
->aapcs_nextncrn
= 0;
2538 pcum
->aapcs_nextnvrn
= 0;
2539 pcum
->pcs_variant
= ARM_PCS_AAPCS64
;
2540 pcum
->aapcs_reg
= NULL_RTX
;
2541 pcum
->aapcs_arg_processed
= false;
2542 pcum
->aapcs_stack_words
= 0;
2543 pcum
->aapcs_stack_size
= 0;
2546 && fndecl
&& TREE_PUBLIC (fndecl
)
2547 && fntype
&& fntype
!= error_mark_node
)
2549 const_tree type
= TREE_TYPE (fntype
);
2550 machine_mode mode ATTRIBUTE_UNUSED
; /* To pass pointer as argument. */
2551 int nregs ATTRIBUTE_UNUSED
; /* Likewise. */
2552 if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (type
), type
,
2553 &mode
, &nregs
, NULL
))
2554 aarch64_err_no_fpadvsimd (TYPE_MODE (type
), "return type");
2560 aarch64_function_arg_advance (cumulative_args_t pcum_v
,
2565 CUMULATIVE_ARGS
*pcum
= get_cumulative_args (pcum_v
);
2566 if (pcum
->pcs_variant
== ARM_PCS_AAPCS64
)
2568 aarch64_layout_arg (pcum_v
, mode
, type
, named
);
2569 gcc_assert ((pcum
->aapcs_reg
!= NULL_RTX
)
2570 != (pcum
->aapcs_stack_words
!= 0));
2571 pcum
->aapcs_arg_processed
= false;
2572 pcum
->aapcs_ncrn
= pcum
->aapcs_nextncrn
;
2573 pcum
->aapcs_nvrn
= pcum
->aapcs_nextnvrn
;
2574 pcum
->aapcs_stack_size
+= pcum
->aapcs_stack_words
;
2575 pcum
->aapcs_stack_words
= 0;
2576 pcum
->aapcs_reg
= NULL_RTX
;
2581 aarch64_function_arg_regno_p (unsigned regno
)
2583 return ((GP_REGNUM_P (regno
) && regno
< R0_REGNUM
+ NUM_ARG_REGS
)
2584 || (FP_REGNUM_P (regno
) && regno
< V0_REGNUM
+ NUM_FP_ARG_REGS
));
2587 /* Implement FUNCTION_ARG_BOUNDARY. Every parameter gets at least
2588 PARM_BOUNDARY bits of alignment, but will be given anything up
2589 to STACK_BOUNDARY bits if the type requires it. This makes sure
2590 that both before and after the layout of each argument, the Next
2591 Stacked Argument Address (NSAA) will have a minimum alignment of
2595 aarch64_function_arg_boundary (machine_mode mode
, const_tree type
)
2597 unsigned int alignment
= aarch64_function_arg_alignment (mode
, type
);
2598 return MIN (MAX (alignment
, PARM_BOUNDARY
), STACK_BOUNDARY
);
2601 /* Implement TARGET_FUNCTION_ARG_PADDING.
2603 Small aggregate types are placed in the lowest memory address.
2605 The related parameter passing rules are B.4, C.3, C.5 and C.14. */
2607 static pad_direction
2608 aarch64_function_arg_padding (machine_mode mode
, const_tree type
)
2610 /* On little-endian targets, the least significant byte of every stack
2611 argument is passed at the lowest byte address of the stack slot. */
2612 if (!BYTES_BIG_ENDIAN
)
2615 /* Otherwise, integral, floating-point and pointer types are padded downward:
2616 the least significant byte of a stack argument is passed at the highest
2617 byte address of the stack slot. */
2619 ? (INTEGRAL_TYPE_P (type
) || SCALAR_FLOAT_TYPE_P (type
)
2620 || POINTER_TYPE_P (type
))
2621 : (SCALAR_INT_MODE_P (mode
) || SCALAR_FLOAT_MODE_P (mode
)))
2622 return PAD_DOWNWARD
;
2624 /* Everything else padded upward, i.e. data in first byte of stack slot. */
2628 /* Similarly, for use by BLOCK_REG_PADDING (MODE, TYPE, FIRST).
2630 It specifies padding for the last (may also be the only)
2631 element of a block move between registers and memory. If
2632 assuming the block is in the memory, padding upward means that
2633 the last element is padded after its highest significant byte,
2634 while in downward padding, the last element is padded at the
2635 its least significant byte side.
2637 Small aggregates and small complex types are always padded
2640 We don't need to worry about homogeneous floating-point or
2641 short-vector aggregates; their move is not affected by the
2642 padding direction determined here. Regardless of endianness,
2643 each element of such an aggregate is put in the least
2644 significant bits of a fp/simd register.
2646 Return !BYTES_BIG_ENDIAN if the least significant byte of the
2647 register has useful data, and return the opposite if the most
2648 significant byte does. */
2651 aarch64_pad_reg_upward (machine_mode mode
, const_tree type
,
2652 bool first ATTRIBUTE_UNUSED
)
2655 /* Small composite types are always padded upward. */
2656 if (BYTES_BIG_ENDIAN
&& aarch64_composite_type_p (type
, mode
))
2658 HOST_WIDE_INT size
= (type
? int_size_in_bytes (type
)
2659 : GET_MODE_SIZE (mode
));
2660 if (size
< 2 * UNITS_PER_WORD
)
2664 /* Otherwise, use the default padding. */
2665 return !BYTES_BIG_ENDIAN
;
2668 static scalar_int_mode
2669 aarch64_libgcc_cmp_return_mode (void)
2674 #define PROBE_INTERVAL (1 << STACK_CHECK_PROBE_INTERVAL_EXP)
2676 /* We use the 12-bit shifted immediate arithmetic instructions so values
2677 must be multiple of (1 << 12), i.e. 4096. */
2678 #define ARITH_FACTOR 4096
2680 #if (PROBE_INTERVAL % ARITH_FACTOR) != 0
2681 #error Cannot use simple address calculation for stack probing
2684 /* The pair of scratch registers used for stack probing. */
2685 #define PROBE_STACK_FIRST_REG 9
2686 #define PROBE_STACK_SECOND_REG 10
2688 /* Emit code to probe a range of stack addresses from FIRST to FIRST+SIZE,
2689 inclusive. These are offsets from the current stack pointer. */
2692 aarch64_emit_probe_stack_range (HOST_WIDE_INT first
, HOST_WIDE_INT size
)
2694 rtx reg1
= gen_rtx_REG (Pmode
, PROBE_STACK_FIRST_REG
);
2696 /* See the same assertion on PROBE_INTERVAL above. */
2697 gcc_assert ((first
% ARITH_FACTOR
) == 0);
2699 /* See if we have a constant small number of probes to generate. If so,
2700 that's the easy case. */
2701 if (size
<= PROBE_INTERVAL
)
2703 const HOST_WIDE_INT base
= ROUND_UP (size
, ARITH_FACTOR
);
2705 emit_set_insn (reg1
,
2706 plus_constant (Pmode
,
2707 stack_pointer_rtx
, -(first
+ base
)));
2708 emit_stack_probe (plus_constant (Pmode
, reg1
, base
- size
));
2711 /* The run-time loop is made up of 8 insns in the generic case while the
2712 compile-time loop is made up of 4+2*(n-2) insns for n # of intervals. */
2713 else if (size
<= 4 * PROBE_INTERVAL
)
2715 HOST_WIDE_INT i
, rem
;
2717 emit_set_insn (reg1
,
2718 plus_constant (Pmode
,
2720 -(first
+ PROBE_INTERVAL
)));
2721 emit_stack_probe (reg1
);
2723 /* Probe at FIRST + N * PROBE_INTERVAL for values of N from 2 until
2724 it exceeds SIZE. If only two probes are needed, this will not
2725 generate any code. Then probe at FIRST + SIZE. */
2726 for (i
= 2 * PROBE_INTERVAL
; i
< size
; i
+= PROBE_INTERVAL
)
2728 emit_set_insn (reg1
,
2729 plus_constant (Pmode
, reg1
, -PROBE_INTERVAL
));
2730 emit_stack_probe (reg1
);
2733 rem
= size
- (i
- PROBE_INTERVAL
);
2736 const HOST_WIDE_INT base
= ROUND_UP (rem
, ARITH_FACTOR
);
2738 emit_set_insn (reg1
, plus_constant (Pmode
, reg1
, -base
));
2739 emit_stack_probe (plus_constant (Pmode
, reg1
, base
- rem
));
2742 emit_stack_probe (plus_constant (Pmode
, reg1
, -rem
));
2745 /* Otherwise, do the same as above, but in a loop. Note that we must be
2746 extra careful with variables wrapping around because we might be at
2747 the very top (or the very bottom) of the address space and we have
2748 to be able to handle this case properly; in particular, we use an
2749 equality test for the loop condition. */
2752 rtx reg2
= gen_rtx_REG (Pmode
, PROBE_STACK_SECOND_REG
);
2754 /* Step 1: round SIZE to the previous multiple of the interval. */
2756 HOST_WIDE_INT rounded_size
= size
& -PROBE_INTERVAL
;
2759 /* Step 2: compute initial and final value of the loop counter. */
2761 /* TEST_ADDR = SP + FIRST. */
2762 emit_set_insn (reg1
,
2763 plus_constant (Pmode
, stack_pointer_rtx
, -first
));
2765 /* LAST_ADDR = SP + FIRST + ROUNDED_SIZE. */
2766 HOST_WIDE_INT adjustment
= - (first
+ rounded_size
);
2767 if (! aarch64_uimm12_shift (adjustment
))
2769 aarch64_internal_mov_immediate (reg2
, GEN_INT (adjustment
),
2771 emit_set_insn (reg2
, gen_rtx_PLUS (Pmode
, stack_pointer_rtx
, reg2
));
2775 emit_set_insn (reg2
,
2776 plus_constant (Pmode
, stack_pointer_rtx
, adjustment
));
2783 TEST_ADDR = TEST_ADDR + PROBE_INTERVAL
2786 while (TEST_ADDR != LAST_ADDR)
2788 probes at FIRST + N * PROBE_INTERVAL for values of N from 1
2789 until it is equal to ROUNDED_SIZE. */
2791 emit_insn (gen_probe_stack_range (reg1
, reg1
, reg2
));
2794 /* Step 4: probe at FIRST + SIZE if we cannot assert at compile-time
2795 that SIZE is equal to ROUNDED_SIZE. */
2797 if (size
!= rounded_size
)
2799 HOST_WIDE_INT rem
= size
- rounded_size
;
2803 const HOST_WIDE_INT base
= ROUND_UP (rem
, ARITH_FACTOR
);
2805 emit_set_insn (reg2
, plus_constant (Pmode
, reg2
, -base
));
2806 emit_stack_probe (plus_constant (Pmode
, reg2
, base
- rem
));
2809 emit_stack_probe (plus_constant (Pmode
, reg2
, -rem
));
2813 /* Make sure nothing is scheduled before we are done. */
2814 emit_insn (gen_blockage ());
2817 /* Probe a range of stack addresses from REG1 to REG2 inclusive. These are
2818 absolute addresses. */
2821 aarch64_output_probe_stack_range (rtx reg1
, rtx reg2
)
2823 static int labelno
= 0;
2827 ASM_GENERATE_INTERNAL_LABEL (loop_lab
, "LPSRL", labelno
++);
2830 ASM_OUTPUT_INTERNAL_LABEL (asm_out_file
, loop_lab
);
2832 /* TEST_ADDR = TEST_ADDR + PROBE_INTERVAL. */
2834 xops
[1] = GEN_INT (PROBE_INTERVAL
);
2835 output_asm_insn ("sub\t%0, %0, %1", xops
);
2837 /* Probe at TEST_ADDR. */
2838 output_asm_insn ("str\txzr, [%0]", xops
);
2840 /* Test if TEST_ADDR == LAST_ADDR. */
2842 output_asm_insn ("cmp\t%0, %1", xops
);
2845 fputs ("\tb.ne\t", asm_out_file
);
2846 assemble_name_raw (asm_out_file
, loop_lab
);
2847 fputc ('\n', asm_out_file
);
2853 aarch64_frame_pointer_required (void)
2855 /* In aarch64_override_options_after_change
2856 flag_omit_leaf_frame_pointer turns off the frame pointer by
2857 default. Turn it back on now if we've not got a leaf
2859 if (flag_omit_leaf_frame_pointer
2860 && (!crtl
->is_leaf
|| df_regs_ever_live_p (LR_REGNUM
)))
2863 /* Force a frame pointer for EH returns so the return address is at FP+8. */
2864 if (crtl
->calls_eh_return
)
2870 /* Mark the registers that need to be saved by the callee and calculate
2871 the size of the callee-saved registers area and frame record (both FP
2872 and LR may be omitted). */
2874 aarch64_layout_frame (void)
2876 HOST_WIDE_INT offset
= 0;
2877 int regno
, last_fp_reg
= INVALID_REGNUM
;
2879 if (reload_completed
&& cfun
->machine
->frame
.laid_out
)
2882 #define SLOT_NOT_REQUIRED (-2)
2883 #define SLOT_REQUIRED (-1)
2885 cfun
->machine
->frame
.wb_candidate1
= INVALID_REGNUM
;
2886 cfun
->machine
->frame
.wb_candidate2
= INVALID_REGNUM
;
2888 /* First mark all the registers that really need to be saved... */
2889 for (regno
= R0_REGNUM
; regno
<= R30_REGNUM
; regno
++)
2890 cfun
->machine
->frame
.reg_offset
[regno
] = SLOT_NOT_REQUIRED
;
2892 for (regno
= V0_REGNUM
; regno
<= V31_REGNUM
; regno
++)
2893 cfun
->machine
->frame
.reg_offset
[regno
] = SLOT_NOT_REQUIRED
;
2895 /* ... that includes the eh data registers (if needed)... */
2896 if (crtl
->calls_eh_return
)
2897 for (regno
= 0; EH_RETURN_DATA_REGNO (regno
) != INVALID_REGNUM
; regno
++)
2898 cfun
->machine
->frame
.reg_offset
[EH_RETURN_DATA_REGNO (regno
)]
2901 /* ... and any callee saved register that dataflow says is live. */
2902 for (regno
= R0_REGNUM
; regno
<= R30_REGNUM
; regno
++)
2903 if (df_regs_ever_live_p (regno
)
2904 && (regno
== R30_REGNUM
2905 || !call_used_regs
[regno
]))
2906 cfun
->machine
->frame
.reg_offset
[regno
] = SLOT_REQUIRED
;
2908 for (regno
= V0_REGNUM
; regno
<= V31_REGNUM
; regno
++)
2909 if (df_regs_ever_live_p (regno
)
2910 && !call_used_regs
[regno
])
2912 cfun
->machine
->frame
.reg_offset
[regno
] = SLOT_REQUIRED
;
2913 last_fp_reg
= regno
;
2916 if (frame_pointer_needed
)
2918 /* FP and LR are placed in the linkage record. */
2919 cfun
->machine
->frame
.reg_offset
[R29_REGNUM
] = 0;
2920 cfun
->machine
->frame
.wb_candidate1
= R29_REGNUM
;
2921 cfun
->machine
->frame
.reg_offset
[R30_REGNUM
] = UNITS_PER_WORD
;
2922 cfun
->machine
->frame
.wb_candidate2
= R30_REGNUM
;
2923 offset
+= 2 * UNITS_PER_WORD
;
2926 /* Now assign stack slots for them. */
2927 for (regno
= R0_REGNUM
; regno
<= R30_REGNUM
; regno
++)
2928 if (cfun
->machine
->frame
.reg_offset
[regno
] == SLOT_REQUIRED
)
2930 cfun
->machine
->frame
.reg_offset
[regno
] = offset
;
2931 if (cfun
->machine
->frame
.wb_candidate1
== INVALID_REGNUM
)
2932 cfun
->machine
->frame
.wb_candidate1
= regno
;
2933 else if (cfun
->machine
->frame
.wb_candidate2
== INVALID_REGNUM
)
2934 cfun
->machine
->frame
.wb_candidate2
= regno
;
2935 offset
+= UNITS_PER_WORD
;
2938 HOST_WIDE_INT max_int_offset
= offset
;
2939 offset
= ROUND_UP (offset
, STACK_BOUNDARY
/ BITS_PER_UNIT
);
2940 bool has_align_gap
= offset
!= max_int_offset
;
2942 for (regno
= V0_REGNUM
; regno
<= V31_REGNUM
; regno
++)
2943 if (cfun
->machine
->frame
.reg_offset
[regno
] == SLOT_REQUIRED
)
2945 /* If there is an alignment gap between integer and fp callee-saves,
2946 allocate the last fp register to it if possible. */
2947 if (regno
== last_fp_reg
&& has_align_gap
&& (offset
& 8) == 0)
2949 cfun
->machine
->frame
.reg_offset
[regno
] = max_int_offset
;
2953 cfun
->machine
->frame
.reg_offset
[regno
] = offset
;
2954 if (cfun
->machine
->frame
.wb_candidate1
== INVALID_REGNUM
)
2955 cfun
->machine
->frame
.wb_candidate1
= regno
;
2956 else if (cfun
->machine
->frame
.wb_candidate2
== INVALID_REGNUM
2957 && cfun
->machine
->frame
.wb_candidate1
>= V0_REGNUM
)
2958 cfun
->machine
->frame
.wb_candidate2
= regno
;
2959 offset
+= UNITS_PER_WORD
;
2962 offset
= ROUND_UP (offset
, STACK_BOUNDARY
/ BITS_PER_UNIT
);
2964 cfun
->machine
->frame
.saved_regs_size
= offset
;
2966 HOST_WIDE_INT varargs_and_saved_regs_size
2967 = offset
+ cfun
->machine
->frame
.saved_varargs_size
;
2969 cfun
->machine
->frame
.hard_fp_offset
2970 = ROUND_UP (varargs_and_saved_regs_size
+ get_frame_size (),
2971 STACK_BOUNDARY
/ BITS_PER_UNIT
);
2973 cfun
->machine
->frame
.frame_size
2974 = ROUND_UP (cfun
->machine
->frame
.hard_fp_offset
2975 + crtl
->outgoing_args_size
,
2976 STACK_BOUNDARY
/ BITS_PER_UNIT
);
2978 cfun
->machine
->frame
.locals_offset
= cfun
->machine
->frame
.saved_varargs_size
;
2980 cfun
->machine
->frame
.initial_adjust
= 0;
2981 cfun
->machine
->frame
.final_adjust
= 0;
2982 cfun
->machine
->frame
.callee_adjust
= 0;
2983 cfun
->machine
->frame
.callee_offset
= 0;
2985 HOST_WIDE_INT max_push_offset
= 0;
2986 if (cfun
->machine
->frame
.wb_candidate2
!= INVALID_REGNUM
)
2987 max_push_offset
= 512;
2988 else if (cfun
->machine
->frame
.wb_candidate1
!= INVALID_REGNUM
)
2989 max_push_offset
= 256;
2991 if (cfun
->machine
->frame
.frame_size
< max_push_offset
2992 && crtl
->outgoing_args_size
== 0)
2994 /* Simple, small frame with no outgoing arguments:
2995 stp reg1, reg2, [sp, -frame_size]!
2996 stp reg3, reg4, [sp, 16] */
2997 cfun
->machine
->frame
.callee_adjust
= cfun
->machine
->frame
.frame_size
;
2999 else if ((crtl
->outgoing_args_size
3000 + cfun
->machine
->frame
.saved_regs_size
< 512)
3001 && !(cfun
->calls_alloca
3002 && cfun
->machine
->frame
.hard_fp_offset
< max_push_offset
))
3004 /* Frame with small outgoing arguments:
3005 sub sp, sp, frame_size
3006 stp reg1, reg2, [sp, outgoing_args_size]
3007 stp reg3, reg4, [sp, outgoing_args_size + 16] */
3008 cfun
->machine
->frame
.initial_adjust
= cfun
->machine
->frame
.frame_size
;
3009 cfun
->machine
->frame
.callee_offset
3010 = cfun
->machine
->frame
.frame_size
- cfun
->machine
->frame
.hard_fp_offset
;
3012 else if (cfun
->machine
->frame
.hard_fp_offset
< max_push_offset
)
3014 /* Frame with large outgoing arguments but a small local area:
3015 stp reg1, reg2, [sp, -hard_fp_offset]!
3016 stp reg3, reg4, [sp, 16]
3017 sub sp, sp, outgoing_args_size */
3018 cfun
->machine
->frame
.callee_adjust
= cfun
->machine
->frame
.hard_fp_offset
;
3019 cfun
->machine
->frame
.final_adjust
3020 = cfun
->machine
->frame
.frame_size
- cfun
->machine
->frame
.callee_adjust
;
3022 else if (!frame_pointer_needed
3023 && varargs_and_saved_regs_size
< max_push_offset
)
3025 /* Frame with large local area and outgoing arguments (this pushes the
3026 callee-saves first, followed by the locals and outgoing area):
3027 stp reg1, reg2, [sp, -varargs_and_saved_regs_size]!
3028 stp reg3, reg4, [sp, 16]
3029 sub sp, sp, frame_size - varargs_and_saved_regs_size */
3030 cfun
->machine
->frame
.callee_adjust
= varargs_and_saved_regs_size
;
3031 cfun
->machine
->frame
.final_adjust
3032 = cfun
->machine
->frame
.frame_size
- cfun
->machine
->frame
.callee_adjust
;
3033 cfun
->machine
->frame
.hard_fp_offset
= cfun
->machine
->frame
.callee_adjust
;
3034 cfun
->machine
->frame
.locals_offset
= cfun
->machine
->frame
.hard_fp_offset
;
3038 /* Frame with large local area and outgoing arguments using frame pointer:
3039 sub sp, sp, hard_fp_offset
3040 stp x29, x30, [sp, 0]
3042 stp reg3, reg4, [sp, 16]
3043 sub sp, sp, outgoing_args_size */
3044 cfun
->machine
->frame
.initial_adjust
= cfun
->machine
->frame
.hard_fp_offset
;
3045 cfun
->machine
->frame
.final_adjust
3046 = cfun
->machine
->frame
.frame_size
- cfun
->machine
->frame
.initial_adjust
;
3049 cfun
->machine
->frame
.laid_out
= true;
3052 /* Return true if the register REGNO is saved on entry to
3053 the current function. */
3056 aarch64_register_saved_on_entry (int regno
)
3058 return cfun
->machine
->frame
.reg_offset
[regno
] >= 0;
3061 /* Return the next register up from REGNO up to LIMIT for the callee
3065 aarch64_next_callee_save (unsigned regno
, unsigned limit
)
3067 while (regno
<= limit
&& !aarch64_register_saved_on_entry (regno
))
3072 /* Push the register number REGNO of mode MODE to the stack with write-back
3073 adjusting the stack by ADJUSTMENT. */
3076 aarch64_pushwb_single_reg (machine_mode mode
, unsigned regno
,
3077 HOST_WIDE_INT adjustment
)
3079 rtx base_rtx
= stack_pointer_rtx
;
3082 reg
= gen_rtx_REG (mode
, regno
);
3083 mem
= gen_rtx_PRE_MODIFY (Pmode
, base_rtx
,
3084 plus_constant (Pmode
, base_rtx
, -adjustment
));
3085 mem
= gen_frame_mem (mode
, mem
);
3087 insn
= emit_move_insn (mem
, reg
);
3088 RTX_FRAME_RELATED_P (insn
) = 1;
3091 /* Generate and return an instruction to store the pair of registers
3092 REG and REG2 of mode MODE to location BASE with write-back adjusting
3093 the stack location BASE by ADJUSTMENT. */
3096 aarch64_gen_storewb_pair (machine_mode mode
, rtx base
, rtx reg
, rtx reg2
,
3097 HOST_WIDE_INT adjustment
)
3102 return gen_storewb_pairdi_di (base
, base
, reg
, reg2
,
3103 GEN_INT (-adjustment
),
3104 GEN_INT (UNITS_PER_WORD
- adjustment
));
3106 return gen_storewb_pairdf_di (base
, base
, reg
, reg2
,
3107 GEN_INT (-adjustment
),
3108 GEN_INT (UNITS_PER_WORD
- adjustment
));
3114 /* Push registers numbered REGNO1 and REGNO2 to the stack, adjusting the
3115 stack pointer by ADJUSTMENT. */
3118 aarch64_push_regs (unsigned regno1
, unsigned regno2
, HOST_WIDE_INT adjustment
)
3121 machine_mode mode
= (regno1
<= R30_REGNUM
) ? E_DImode
: E_DFmode
;
3123 if (regno2
== INVALID_REGNUM
)
3124 return aarch64_pushwb_single_reg (mode
, regno1
, adjustment
);
3126 rtx reg1
= gen_rtx_REG (mode
, regno1
);
3127 rtx reg2
= gen_rtx_REG (mode
, regno2
);
3129 insn
= emit_insn (aarch64_gen_storewb_pair (mode
, stack_pointer_rtx
, reg1
,
3131 RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn
), 0, 2)) = 1;
3132 RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn
), 0, 1)) = 1;
3133 RTX_FRAME_RELATED_P (insn
) = 1;
3136 /* Load the pair of register REG, REG2 of mode MODE from stack location BASE,
3137 adjusting it by ADJUSTMENT afterwards. */
3140 aarch64_gen_loadwb_pair (machine_mode mode
, rtx base
, rtx reg
, rtx reg2
,
3141 HOST_WIDE_INT adjustment
)
3146 return gen_loadwb_pairdi_di (base
, base
, reg
, reg2
, GEN_INT (adjustment
),
3147 GEN_INT (UNITS_PER_WORD
));
3149 return gen_loadwb_pairdf_di (base
, base
, reg
, reg2
, GEN_INT (adjustment
),
3150 GEN_INT (UNITS_PER_WORD
));
3156 /* Pop the two registers numbered REGNO1, REGNO2 from the stack, adjusting it
3157 afterwards by ADJUSTMENT and writing the appropriate REG_CFA_RESTORE notes
3161 aarch64_pop_regs (unsigned regno1
, unsigned regno2
, HOST_WIDE_INT adjustment
,
3164 machine_mode mode
= (regno1
<= R30_REGNUM
) ? E_DImode
: E_DFmode
;
3165 rtx reg1
= gen_rtx_REG (mode
, regno1
);
3167 *cfi_ops
= alloc_reg_note (REG_CFA_RESTORE
, reg1
, *cfi_ops
);
3169 if (regno2
== INVALID_REGNUM
)
3171 rtx mem
= plus_constant (Pmode
, stack_pointer_rtx
, adjustment
);
3172 mem
= gen_rtx_POST_MODIFY (Pmode
, stack_pointer_rtx
, mem
);
3173 emit_move_insn (reg1
, gen_frame_mem (mode
, mem
));
3177 rtx reg2
= gen_rtx_REG (mode
, regno2
);
3178 *cfi_ops
= alloc_reg_note (REG_CFA_RESTORE
, reg2
, *cfi_ops
);
3179 emit_insn (aarch64_gen_loadwb_pair (mode
, stack_pointer_rtx
, reg1
,
3184 /* Generate and return a store pair instruction of mode MODE to store
3185 register REG1 to MEM1 and register REG2 to MEM2. */
3188 aarch64_gen_store_pair (machine_mode mode
, rtx mem1
, rtx reg1
, rtx mem2
,
3194 return gen_store_pairdi (mem1
, reg1
, mem2
, reg2
);
3197 return gen_store_pairdf (mem1
, reg1
, mem2
, reg2
);
3204 /* Generate and regurn a load pair isntruction of mode MODE to load register
3205 REG1 from MEM1 and register REG2 from MEM2. */
3208 aarch64_gen_load_pair (machine_mode mode
, rtx reg1
, rtx mem1
, rtx reg2
,
3214 return gen_load_pairdi (reg1
, mem1
, reg2
, mem2
);
3217 return gen_load_pairdf (reg1
, mem1
, reg2
, mem2
);
3224 /* Return TRUE if return address signing should be enabled for the current
3225 function, otherwise return FALSE. */
3228 aarch64_return_address_signing_enabled (void)
3230 /* This function should only be called after frame laid out. */
3231 gcc_assert (cfun
->machine
->frame
.laid_out
);
3233 /* If signing scope is AARCH64_FUNCTION_NON_LEAF, we only sign a leaf function
3234 if it's LR is pushed onto stack. */
3235 return (aarch64_ra_sign_scope
== AARCH64_FUNCTION_ALL
3236 || (aarch64_ra_sign_scope
== AARCH64_FUNCTION_NON_LEAF
3237 && cfun
->machine
->frame
.reg_offset
[LR_REGNUM
] >= 0));
3240 /* Emit code to save the callee-saved registers from register number START
3241 to LIMIT to the stack at the location starting at offset START_OFFSET,
3242 skipping any write-back candidates if SKIP_WB is true. */
3245 aarch64_save_callee_saves (machine_mode mode
, HOST_WIDE_INT start_offset
,
3246 unsigned start
, unsigned limit
, bool skip_wb
)
3252 for (regno
= aarch64_next_callee_save (start
, limit
);
3254 regno
= aarch64_next_callee_save (regno
+ 1, limit
))
3257 HOST_WIDE_INT offset
;
3260 && (regno
== cfun
->machine
->frame
.wb_candidate1
3261 || regno
== cfun
->machine
->frame
.wb_candidate2
))
3264 if (cfun
->machine
->reg_is_wrapped_separately
[regno
])
3267 reg
= gen_rtx_REG (mode
, regno
);
3268 offset
= start_offset
+ cfun
->machine
->frame
.reg_offset
[regno
];
3269 mem
= gen_frame_mem (mode
, plus_constant (Pmode
, stack_pointer_rtx
,
3272 regno2
= aarch64_next_callee_save (regno
+ 1, limit
);
3275 && !cfun
->machine
->reg_is_wrapped_separately
[regno2
]
3276 && ((cfun
->machine
->frame
.reg_offset
[regno
] + UNITS_PER_WORD
)
3277 == cfun
->machine
->frame
.reg_offset
[regno2
]))
3280 rtx reg2
= gen_rtx_REG (mode
, regno2
);
3283 offset
= start_offset
+ cfun
->machine
->frame
.reg_offset
[regno2
];
3284 mem2
= gen_frame_mem (mode
, plus_constant (Pmode
, stack_pointer_rtx
,
3286 insn
= emit_insn (aarch64_gen_store_pair (mode
, mem
, reg
, mem2
,
3289 /* The first part of a frame-related parallel insn is
3290 always assumed to be relevant to the frame
3291 calculations; subsequent parts, are only
3292 frame-related if explicitly marked. */
3293 RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn
), 0, 1)) = 1;
3297 insn
= emit_move_insn (mem
, reg
);
3299 RTX_FRAME_RELATED_P (insn
) = 1;
3303 /* Emit code to restore the callee registers of mode MODE from register
3304 number START up to and including LIMIT. Restore from the stack offset
3305 START_OFFSET, skipping any write-back candidates if SKIP_WB is true.
3306 Write the appropriate REG_CFA_RESTORE notes into CFI_OPS. */
3309 aarch64_restore_callee_saves (machine_mode mode
,
3310 HOST_WIDE_INT start_offset
, unsigned start
,
3311 unsigned limit
, bool skip_wb
, rtx
*cfi_ops
)
3313 rtx base_rtx
= stack_pointer_rtx
;
3316 HOST_WIDE_INT offset
;
3318 for (regno
= aarch64_next_callee_save (start
, limit
);
3320 regno
= aarch64_next_callee_save (regno
+ 1, limit
))
3322 if (cfun
->machine
->reg_is_wrapped_separately
[regno
])
3328 && (regno
== cfun
->machine
->frame
.wb_candidate1
3329 || regno
== cfun
->machine
->frame
.wb_candidate2
))
3332 reg
= gen_rtx_REG (mode
, regno
);
3333 offset
= start_offset
+ cfun
->machine
->frame
.reg_offset
[regno
];
3334 mem
= gen_frame_mem (mode
, plus_constant (Pmode
, base_rtx
, offset
));
3336 regno2
= aarch64_next_callee_save (regno
+ 1, limit
);
3339 && !cfun
->machine
->reg_is_wrapped_separately
[regno2
]
3340 && ((cfun
->machine
->frame
.reg_offset
[regno
] + UNITS_PER_WORD
)
3341 == cfun
->machine
->frame
.reg_offset
[regno2
]))
3343 rtx reg2
= gen_rtx_REG (mode
, regno2
);
3346 offset
= start_offset
+ cfun
->machine
->frame
.reg_offset
[regno2
];
3347 mem2
= gen_frame_mem (mode
, plus_constant (Pmode
, base_rtx
, offset
));
3348 emit_insn (aarch64_gen_load_pair (mode
, reg
, mem
, reg2
, mem2
));
3350 *cfi_ops
= alloc_reg_note (REG_CFA_RESTORE
, reg2
, *cfi_ops
);
3354 emit_move_insn (reg
, mem
);
3355 *cfi_ops
= alloc_reg_note (REG_CFA_RESTORE
, reg
, *cfi_ops
);
3360 offset_9bit_signed_unscaled_p (machine_mode mode ATTRIBUTE_UNUSED
,
3361 HOST_WIDE_INT offset
)
3363 return offset
>= -256 && offset
< 256;
3367 offset_12bit_unsigned_scaled_p (machine_mode mode
, HOST_WIDE_INT offset
)
3370 && offset
< 4096 * GET_MODE_SIZE (mode
)
3371 && offset
% GET_MODE_SIZE (mode
) == 0);
3375 aarch64_offset_7bit_signed_scaled_p (machine_mode mode
, HOST_WIDE_INT offset
)
3377 return (offset
>= -64 * GET_MODE_SIZE (mode
)
3378 && offset
< 64 * GET_MODE_SIZE (mode
)
3379 && offset
% GET_MODE_SIZE (mode
) == 0);
3382 /* Implement TARGET_SHRINK_WRAP_GET_SEPARATE_COMPONENTS. */
3385 aarch64_get_separate_components (void)
3387 aarch64_layout_frame ();
3389 sbitmap components
= sbitmap_alloc (LAST_SAVED_REGNUM
+ 1);
3390 bitmap_clear (components
);
3392 /* The registers we need saved to the frame. */
3393 for (unsigned regno
= 0; regno
<= LAST_SAVED_REGNUM
; regno
++)
3394 if (aarch64_register_saved_on_entry (regno
))
3396 HOST_WIDE_INT offset
= cfun
->machine
->frame
.reg_offset
[regno
];
3397 if (!frame_pointer_needed
)
3398 offset
+= cfun
->machine
->frame
.frame_size
3399 - cfun
->machine
->frame
.hard_fp_offset
;
3400 /* Check that we can access the stack slot of the register with one
3401 direct load with no adjustments needed. */
3402 if (offset_12bit_unsigned_scaled_p (DImode
, offset
))
3403 bitmap_set_bit (components
, regno
);
3406 /* Don't mess with the hard frame pointer. */
3407 if (frame_pointer_needed
)
3408 bitmap_clear_bit (components
, HARD_FRAME_POINTER_REGNUM
);
3410 unsigned reg1
= cfun
->machine
->frame
.wb_candidate1
;
3411 unsigned reg2
= cfun
->machine
->frame
.wb_candidate2
;
3412 /* If aarch64_layout_frame has chosen registers to store/restore with
3413 writeback don't interfere with them to avoid having to output explicit
3414 stack adjustment instructions. */
3415 if (reg2
!= INVALID_REGNUM
)
3416 bitmap_clear_bit (components
, reg2
);
3417 if (reg1
!= INVALID_REGNUM
)
3418 bitmap_clear_bit (components
, reg1
);
3420 bitmap_clear_bit (components
, LR_REGNUM
);
3421 bitmap_clear_bit (components
, SP_REGNUM
);
3426 /* Implement TARGET_SHRINK_WRAP_COMPONENTS_FOR_BB. */
3429 aarch64_components_for_bb (basic_block bb
)
3431 bitmap in
= DF_LIVE_IN (bb
);
3432 bitmap gen
= &DF_LIVE_BB_INFO (bb
)->gen
;
3433 bitmap kill
= &DF_LIVE_BB_INFO (bb
)->kill
;
3435 sbitmap components
= sbitmap_alloc (LAST_SAVED_REGNUM
+ 1);
3436 bitmap_clear (components
);
3438 /* GPRs are used in a bb if they are in the IN, GEN, or KILL sets. */
3439 for (unsigned regno
= 0; regno
<= LAST_SAVED_REGNUM
; regno
++)
3440 if ((!call_used_regs
[regno
])
3441 && (bitmap_bit_p (in
, regno
)
3442 || bitmap_bit_p (gen
, regno
)
3443 || bitmap_bit_p (kill
, regno
)))
3444 bitmap_set_bit (components
, regno
);
3449 /* Implement TARGET_SHRINK_WRAP_DISQUALIFY_COMPONENTS.
3450 Nothing to do for aarch64. */
3453 aarch64_disqualify_components (sbitmap
, edge
, sbitmap
, bool)
3457 /* Return the next set bit in BMP from START onwards. Return the total number
3458 of bits in BMP if no set bit is found at or after START. */
3461 aarch64_get_next_set_bit (sbitmap bmp
, unsigned int start
)
3463 unsigned int nbits
= SBITMAP_SIZE (bmp
);
3467 gcc_assert (start
< nbits
);
3468 for (unsigned int i
= start
; i
< nbits
; i
++)
3469 if (bitmap_bit_p (bmp
, i
))
3475 /* Do the work for aarch64_emit_prologue_components and
3476 aarch64_emit_epilogue_components. COMPONENTS is the bitmap of registers
3477 to save/restore, PROLOGUE_P indicates whether to emit the prologue sequence
3478 for these components or the epilogue sequence. That is, it determines
3479 whether we should emit stores or loads and what kind of CFA notes to attach
3480 to the insns. Otherwise the logic for the two sequences is very
3484 aarch64_process_components (sbitmap components
, bool prologue_p
)
3486 rtx ptr_reg
= gen_rtx_REG (Pmode
, frame_pointer_needed
3487 ? HARD_FRAME_POINTER_REGNUM
3488 : STACK_POINTER_REGNUM
);
3490 unsigned last_regno
= SBITMAP_SIZE (components
);
3491 unsigned regno
= aarch64_get_next_set_bit (components
, R0_REGNUM
);
3492 rtx_insn
*insn
= NULL
;
3494 while (regno
!= last_regno
)
3496 /* AAPCS64 section 5.1.2 requires only the bottom 64 bits to be saved
3497 so DFmode for the vector registers is enough. */
3498 machine_mode mode
= GP_REGNUM_P (regno
) ? E_DImode
: E_DFmode
;
3499 rtx reg
= gen_rtx_REG (mode
, regno
);
3500 HOST_WIDE_INT offset
= cfun
->machine
->frame
.reg_offset
[regno
];
3501 if (!frame_pointer_needed
)
3502 offset
+= cfun
->machine
->frame
.frame_size
3503 - cfun
->machine
->frame
.hard_fp_offset
;
3504 rtx addr
= plus_constant (Pmode
, ptr_reg
, offset
);
3505 rtx mem
= gen_frame_mem (mode
, addr
);
3507 rtx set
= prologue_p
? gen_rtx_SET (mem
, reg
) : gen_rtx_SET (reg
, mem
);
3508 unsigned regno2
= aarch64_get_next_set_bit (components
, regno
+ 1);
3509 /* No more registers to handle after REGNO.
3510 Emit a single save/restore and exit. */
3511 if (regno2
== last_regno
)
3513 insn
= emit_insn (set
);
3514 RTX_FRAME_RELATED_P (insn
) = 1;
3516 add_reg_note (insn
, REG_CFA_OFFSET
, copy_rtx (set
));
3518 add_reg_note (insn
, REG_CFA_RESTORE
, reg
);
3522 HOST_WIDE_INT offset2
= cfun
->machine
->frame
.reg_offset
[regno2
];
3523 /* The next register is not of the same class or its offset is not
3524 mergeable with the current one into a pair. */
3525 if (!satisfies_constraint_Ump (mem
)
3526 || GP_REGNUM_P (regno
) != GP_REGNUM_P (regno2
)
3527 || (offset2
- cfun
->machine
->frame
.reg_offset
[regno
])
3528 != GET_MODE_SIZE (mode
))
3530 insn
= emit_insn (set
);
3531 RTX_FRAME_RELATED_P (insn
) = 1;
3533 add_reg_note (insn
, REG_CFA_OFFSET
, copy_rtx (set
));
3535 add_reg_note (insn
, REG_CFA_RESTORE
, reg
);
3541 /* REGNO2 can be saved/restored in a pair with REGNO. */
3542 rtx reg2
= gen_rtx_REG (mode
, regno2
);
3543 if (!frame_pointer_needed
)
3544 offset2
+= cfun
->machine
->frame
.frame_size
3545 - cfun
->machine
->frame
.hard_fp_offset
;
3546 rtx addr2
= plus_constant (Pmode
, ptr_reg
, offset2
);
3547 rtx mem2
= gen_frame_mem (mode
, addr2
);
3548 rtx set2
= prologue_p
? gen_rtx_SET (mem2
, reg2
)
3549 : gen_rtx_SET (reg2
, mem2
);
3552 insn
= emit_insn (aarch64_gen_store_pair (mode
, mem
, reg
, mem2
, reg2
));
3554 insn
= emit_insn (aarch64_gen_load_pair (mode
, reg
, mem
, reg2
, mem2
));
3556 RTX_FRAME_RELATED_P (insn
) = 1;
3559 add_reg_note (insn
, REG_CFA_OFFSET
, set
);
3560 add_reg_note (insn
, REG_CFA_OFFSET
, set2
);
3564 add_reg_note (insn
, REG_CFA_RESTORE
, reg
);
3565 add_reg_note (insn
, REG_CFA_RESTORE
, reg2
);
3568 regno
= aarch64_get_next_set_bit (components
, regno2
+ 1);
3572 /* Implement TARGET_SHRINK_WRAP_EMIT_PROLOGUE_COMPONENTS. */
3575 aarch64_emit_prologue_components (sbitmap components
)
3577 aarch64_process_components (components
, true);
3580 /* Implement TARGET_SHRINK_WRAP_EMIT_EPILOGUE_COMPONENTS. */
3583 aarch64_emit_epilogue_components (sbitmap components
)
3585 aarch64_process_components (components
, false);
3588 /* Implement TARGET_SHRINK_WRAP_SET_HANDLED_COMPONENTS. */
3591 aarch64_set_handled_components (sbitmap components
)
3593 for (unsigned regno
= 0; regno
<= LAST_SAVED_REGNUM
; regno
++)
3594 if (bitmap_bit_p (components
, regno
))
3595 cfun
->machine
->reg_is_wrapped_separately
[regno
] = true;
3598 /* AArch64 stack frames generated by this compiler look like:
3600 +-------------------------------+
3602 | incoming stack arguments |
3604 +-------------------------------+
3605 | | <-- incoming stack pointer (aligned)
3606 | callee-allocated save area |
3607 | for register varargs |
3609 +-------------------------------+
3610 | local variables | <-- frame_pointer_rtx
3612 +-------------------------------+
3614 +-------------------------------+ |
3615 | callee-saved registers | | frame.saved_regs_size
3616 +-------------------------------+ |
3618 +-------------------------------+ |
3619 | FP' | / <- hard_frame_pointer_rtx (aligned)
3620 +-------------------------------+
3621 | dynamic allocation |
3622 +-------------------------------+
3624 +-------------------------------+
3625 | outgoing stack arguments | <-- arg_pointer
3627 +-------------------------------+
3628 | | <-- stack_pointer_rtx (aligned)
3630 Dynamic stack allocations via alloca() decrease stack_pointer_rtx
3631 but leave frame_pointer_rtx and hard_frame_pointer_rtx
3634 /* Generate the prologue instructions for entry into a function.
3635 Establish the stack frame by decreasing the stack pointer with a
3636 properly calculated size and, if necessary, create a frame record
3637 filled with the values of LR and previous frame pointer. The
3638 current FP is also set up if it is in use. */
3641 aarch64_expand_prologue (void)
3643 aarch64_layout_frame ();
3645 HOST_WIDE_INT frame_size
= cfun
->machine
->frame
.frame_size
;
3646 HOST_WIDE_INT initial_adjust
= cfun
->machine
->frame
.initial_adjust
;
3647 HOST_WIDE_INT callee_adjust
= cfun
->machine
->frame
.callee_adjust
;
3648 HOST_WIDE_INT final_adjust
= cfun
->machine
->frame
.final_adjust
;
3649 HOST_WIDE_INT callee_offset
= cfun
->machine
->frame
.callee_offset
;
3650 unsigned reg1
= cfun
->machine
->frame
.wb_candidate1
;
3651 unsigned reg2
= cfun
->machine
->frame
.wb_candidate2
;
3654 /* Sign return address for functions. */
3655 if (aarch64_return_address_signing_enabled ())
3657 insn
= emit_insn (gen_pacisp ());
3658 add_reg_note (insn
, REG_CFA_TOGGLE_RA_MANGLE
, const0_rtx
);
3659 RTX_FRAME_RELATED_P (insn
) = 1;
3662 if (flag_stack_usage_info
)
3663 current_function_static_stack_size
= frame_size
;
3665 if (flag_stack_check
== STATIC_BUILTIN_STACK_CHECK
)
3667 if (crtl
->is_leaf
&& !cfun
->calls_alloca
)
3669 if (frame_size
> PROBE_INTERVAL
&& frame_size
> STACK_CHECK_PROTECT
)
3670 aarch64_emit_probe_stack_range (STACK_CHECK_PROTECT
,
3671 frame_size
- STACK_CHECK_PROTECT
);
3673 else if (frame_size
> 0)
3674 aarch64_emit_probe_stack_range (STACK_CHECK_PROTECT
, frame_size
);
3677 aarch64_sub_sp (IP0_REGNUM
, initial_adjust
, true);
3679 if (callee_adjust
!= 0)
3680 aarch64_push_regs (reg1
, reg2
, callee_adjust
);
3682 if (frame_pointer_needed
)
3684 if (callee_adjust
== 0)
3685 aarch64_save_callee_saves (DImode
, callee_offset
, R29_REGNUM
,
3687 insn
= emit_insn (gen_add3_insn (hard_frame_pointer_rtx
,
3689 GEN_INT (callee_offset
)));
3690 RTX_FRAME_RELATED_P (insn
) = 1;
3691 emit_insn (gen_stack_tie (stack_pointer_rtx
, hard_frame_pointer_rtx
));
3694 aarch64_save_callee_saves (DImode
, callee_offset
, R0_REGNUM
, R30_REGNUM
,
3695 callee_adjust
!= 0 || frame_pointer_needed
);
3696 aarch64_save_callee_saves (DFmode
, callee_offset
, V0_REGNUM
, V31_REGNUM
,
3697 callee_adjust
!= 0 || frame_pointer_needed
);
3698 aarch64_sub_sp (IP1_REGNUM
, final_adjust
, !frame_pointer_needed
);
3701 /* Return TRUE if we can use a simple_return insn.
3703 This function checks whether the callee saved stack is empty, which
3704 means no restore actions are need. The pro_and_epilogue will use
3705 this to check whether shrink-wrapping opt is feasible. */
3708 aarch64_use_return_insn_p (void)
3710 if (!reload_completed
)
3716 aarch64_layout_frame ();
3718 return cfun
->machine
->frame
.frame_size
== 0;
3721 /* Generate the epilogue instructions for returning from a function.
3722 This is almost exactly the reverse of the prolog sequence, except
3723 that we need to insert barriers to avoid scheduling loads that read
3724 from a deallocated stack, and we optimize the unwind records by
3725 emitting them all together if possible. */
3727 aarch64_expand_epilogue (bool for_sibcall
)
3729 aarch64_layout_frame ();
3731 HOST_WIDE_INT initial_adjust
= cfun
->machine
->frame
.initial_adjust
;
3732 HOST_WIDE_INT callee_adjust
= cfun
->machine
->frame
.callee_adjust
;
3733 HOST_WIDE_INT final_adjust
= cfun
->machine
->frame
.final_adjust
;
3734 HOST_WIDE_INT callee_offset
= cfun
->machine
->frame
.callee_offset
;
3735 unsigned reg1
= cfun
->machine
->frame
.wb_candidate1
;
3736 unsigned reg2
= cfun
->machine
->frame
.wb_candidate2
;
3740 /* We need to add memory barrier to prevent read from deallocated stack. */
3741 bool need_barrier_p
= (get_frame_size ()
3742 + cfun
->machine
->frame
.saved_varargs_size
) != 0;
3744 /* Emit a barrier to prevent loads from a deallocated stack. */
3745 if (final_adjust
> crtl
->outgoing_args_size
|| cfun
->calls_alloca
3746 || crtl
->calls_eh_return
)
3748 emit_insn (gen_stack_tie (stack_pointer_rtx
, stack_pointer_rtx
));
3749 need_barrier_p
= false;
3752 /* Restore the stack pointer from the frame pointer if it may not
3753 be the same as the stack pointer. */
3754 if (frame_pointer_needed
&& (final_adjust
|| cfun
->calls_alloca
))
3756 insn
= emit_insn (gen_add3_insn (stack_pointer_rtx
,
3757 hard_frame_pointer_rtx
,
3758 GEN_INT (-callee_offset
)));
3759 /* If writeback is used when restoring callee-saves, the CFA
3760 is restored on the instruction doing the writeback. */
3761 RTX_FRAME_RELATED_P (insn
) = callee_adjust
== 0;
3764 aarch64_add_sp (IP1_REGNUM
, final_adjust
, df_regs_ever_live_p (IP1_REGNUM
));
3766 aarch64_restore_callee_saves (DImode
, callee_offset
, R0_REGNUM
, R30_REGNUM
,
3767 callee_adjust
!= 0, &cfi_ops
);
3768 aarch64_restore_callee_saves (DFmode
, callee_offset
, V0_REGNUM
, V31_REGNUM
,
3769 callee_adjust
!= 0, &cfi_ops
);
3772 emit_insn (gen_stack_tie (stack_pointer_rtx
, stack_pointer_rtx
));
3774 if (callee_adjust
!= 0)
3775 aarch64_pop_regs (reg1
, reg2
, callee_adjust
, &cfi_ops
);
3777 if (callee_adjust
!= 0 || initial_adjust
> 65536)
3779 /* Emit delayed restores and set the CFA to be SP + initial_adjust. */
3780 insn
= get_last_insn ();
3781 rtx new_cfa
= plus_constant (Pmode
, stack_pointer_rtx
, initial_adjust
);
3782 REG_NOTES (insn
) = alloc_reg_note (REG_CFA_DEF_CFA
, new_cfa
, cfi_ops
);
3783 RTX_FRAME_RELATED_P (insn
) = 1;
3787 aarch64_add_sp (IP0_REGNUM
, initial_adjust
, df_regs_ever_live_p (IP0_REGNUM
));
3791 /* Emit delayed restores and reset the CFA to be SP. */
3792 insn
= get_last_insn ();
3793 cfi_ops
= alloc_reg_note (REG_CFA_DEF_CFA
, stack_pointer_rtx
, cfi_ops
);
3794 REG_NOTES (insn
) = cfi_ops
;
3795 RTX_FRAME_RELATED_P (insn
) = 1;
3798 /* We prefer to emit the combined return/authenticate instruction RETAA,
3799 however there are three cases in which we must instead emit an explicit
3800 authentication instruction.
3802 1) Sibcalls don't return in a normal way, so if we're about to call one
3803 we must authenticate.
3805 2) The RETAA instruction is not available before ARMv8.3-A, so if we are
3806 generating code for !TARGET_ARMV8_3 we can't use it and must
3807 explicitly authenticate.
3809 3) On an eh_return path we make extra stack adjustments to update the
3810 canonical frame address to be the exception handler's CFA. We want
3811 to authenticate using the CFA of the function which calls eh_return.
3813 if (aarch64_return_address_signing_enabled ()
3814 && (for_sibcall
|| !TARGET_ARMV8_3
|| crtl
->calls_eh_return
))
3816 insn
= emit_insn (gen_autisp ());
3817 add_reg_note (insn
, REG_CFA_TOGGLE_RA_MANGLE
, const0_rtx
);
3818 RTX_FRAME_RELATED_P (insn
) = 1;
3821 /* Stack adjustment for exception handler. */
3822 if (crtl
->calls_eh_return
)
3824 /* We need to unwind the stack by the offset computed by
3825 EH_RETURN_STACKADJ_RTX. We have already reset the CFA
3826 to be SP; letting the CFA move during this adjustment
3827 is just as correct as retaining the CFA from the body
3828 of the function. Therefore, do nothing special. */
3829 emit_insn (gen_add2_insn (stack_pointer_rtx
, EH_RETURN_STACKADJ_RTX
));
3832 emit_use (gen_rtx_REG (DImode
, LR_REGNUM
));
3834 emit_jump_insn (ret_rtx
);
3837 /* Implement EH_RETURN_HANDLER_RTX. EH returns need to either return
3838 normally or return to a previous frame after unwinding.
3840 An EH return uses a single shared return sequence. The epilogue is
3841 exactly like a normal epilogue except that it has an extra input
3842 register (EH_RETURN_STACKADJ_RTX) which contains the stack adjustment
3843 that must be applied after the frame has been destroyed. An extra label
3844 is inserted before the epilogue which initializes this register to zero,
3845 and this is the entry point for a normal return.
3847 An actual EH return updates the return address, initializes the stack
3848 adjustment and jumps directly into the epilogue (bypassing the zeroing
3849 of the adjustment). Since the return address is typically saved on the
3850 stack when a function makes a call, the saved LR must be updated outside
3853 This poses problems as the store is generated well before the epilogue,
3854 so the offset of LR is not known yet. Also optimizations will remove the
3855 store as it appears dead, even after the epilogue is generated (as the
3856 base or offset for loading LR is different in many cases).
3858 To avoid these problems this implementation forces the frame pointer
3859 in eh_return functions so that the location of LR is fixed and known early.
3860 It also marks the store volatile, so no optimization is permitted to
3861 remove the store. */
3863 aarch64_eh_return_handler_rtx (void)
3865 rtx tmp
= gen_frame_mem (Pmode
,
3866 plus_constant (Pmode
, hard_frame_pointer_rtx
, UNITS_PER_WORD
));
3868 /* Mark the store volatile, so no optimization is permitted to remove it. */
3869 MEM_VOLATILE_P (tmp
) = true;
3873 /* Output code to add DELTA to the first argument, and then jump
3874 to FUNCTION. Used for C++ multiple inheritance. */
3876 aarch64_output_mi_thunk (FILE *file
, tree thunk ATTRIBUTE_UNUSED
,
3877 HOST_WIDE_INT delta
,
3878 HOST_WIDE_INT vcall_offset
,
3881 /* The this pointer is always in x0. Note that this differs from
3882 Arm where the this pointer maybe bumped to r1 if r0 is required
3883 to return a pointer to an aggregate. On AArch64 a result value
3884 pointer will be in x8. */
3885 int this_regno
= R0_REGNUM
;
3886 rtx this_rtx
, temp0
, temp1
, addr
, funexp
;
3889 reload_completed
= 1;
3890 emit_note (NOTE_INSN_PROLOGUE_END
);
3892 if (vcall_offset
== 0)
3893 aarch64_add_constant (Pmode
, this_regno
, IP1_REGNUM
, delta
);
3896 gcc_assert ((vcall_offset
& (POINTER_BYTES
- 1)) == 0);
3898 this_rtx
= gen_rtx_REG (Pmode
, this_regno
);
3899 temp0
= gen_rtx_REG (Pmode
, IP0_REGNUM
);
3900 temp1
= gen_rtx_REG (Pmode
, IP1_REGNUM
);
3905 if (delta
>= -256 && delta
< 256)
3906 addr
= gen_rtx_PRE_MODIFY (Pmode
, this_rtx
,
3907 plus_constant (Pmode
, this_rtx
, delta
));
3909 aarch64_add_constant (Pmode
, this_regno
, IP1_REGNUM
, delta
);
3912 if (Pmode
== ptr_mode
)
3913 aarch64_emit_move (temp0
, gen_rtx_MEM (ptr_mode
, addr
));
3915 aarch64_emit_move (temp0
,
3916 gen_rtx_ZERO_EXTEND (Pmode
,
3917 gen_rtx_MEM (ptr_mode
, addr
)));
3919 if (vcall_offset
>= -256 && vcall_offset
< 4096 * POINTER_BYTES
)
3920 addr
= plus_constant (Pmode
, temp0
, vcall_offset
);
3923 aarch64_internal_mov_immediate (temp1
, GEN_INT (vcall_offset
), true,
3925 addr
= gen_rtx_PLUS (Pmode
, temp0
, temp1
);
3928 if (Pmode
== ptr_mode
)
3929 aarch64_emit_move (temp1
, gen_rtx_MEM (ptr_mode
,addr
));
3931 aarch64_emit_move (temp1
,
3932 gen_rtx_SIGN_EXTEND (Pmode
,
3933 gen_rtx_MEM (ptr_mode
, addr
)));
3935 emit_insn (gen_add2_insn (this_rtx
, temp1
));
3938 /* Generate a tail call to the target function. */
3939 if (!TREE_USED (function
))
3941 assemble_external (function
);
3942 TREE_USED (function
) = 1;
3944 funexp
= XEXP (DECL_RTL (function
), 0);
3945 funexp
= gen_rtx_MEM (FUNCTION_MODE
, funexp
);
3946 insn
= emit_call_insn (gen_sibcall (funexp
, const0_rtx
, NULL_RTX
));
3947 SIBLING_CALL_P (insn
) = 1;
3949 insn
= get_insns ();
3950 shorten_branches (insn
);
3951 final_start_function (insn
, file
, 1);
3952 final (insn
, file
, 1);
3953 final_end_function ();
3955 /* Stop pretending to be a post-reload pass. */
3956 reload_completed
= 0;
3960 aarch64_tls_referenced_p (rtx x
)
3962 if (!TARGET_HAVE_TLS
)
3964 subrtx_iterator::array_type array
;
3965 FOR_EACH_SUBRTX (iter
, array
, x
, ALL
)
3967 const_rtx x
= *iter
;
3968 if (GET_CODE (x
) == SYMBOL_REF
&& SYMBOL_REF_TLS_MODEL (x
) != 0)
3970 /* Don't recurse into UNSPEC_TLS looking for TLS symbols; these are
3971 TLS offsets, not real symbol references. */
3972 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLS
)
3973 iter
.skip_subrtxes ();
3979 /* Return true if val can be encoded as a 12-bit unsigned immediate with
3980 a left shift of 0 or 12 bits. */
3982 aarch64_uimm12_shift (HOST_WIDE_INT val
)
3984 return ((val
& (((HOST_WIDE_INT
) 0xfff) << 0)) == val
3985 || (val
& (((HOST_WIDE_INT
) 0xfff) << 12)) == val
3990 /* Return true if val is an immediate that can be loaded into a
3991 register by a MOVZ instruction. */
3993 aarch64_movw_imm (HOST_WIDE_INT val
, scalar_int_mode mode
)
3995 if (GET_MODE_SIZE (mode
) > 4)
3997 if ((val
& (((HOST_WIDE_INT
) 0xffff) << 32)) == val
3998 || (val
& (((HOST_WIDE_INT
) 0xffff) << 48)) == val
)
4003 /* Ignore sign extension. */
4004 val
&= (HOST_WIDE_INT
) 0xffffffff;
4006 return ((val
& (((HOST_WIDE_INT
) 0xffff) << 0)) == val
4007 || (val
& (((HOST_WIDE_INT
) 0xffff) << 16)) == val
);
4010 /* Multipliers for repeating bitmasks of width 32, 16, 8, 4, and 2. */
4012 static const unsigned HOST_WIDE_INT bitmask_imm_mul
[] =
4014 0x0000000100000001ull
,
4015 0x0001000100010001ull
,
4016 0x0101010101010101ull
,
4017 0x1111111111111111ull
,
4018 0x5555555555555555ull
,
4022 /* Return true if val is a valid bitmask immediate. */
4025 aarch64_bitmask_imm (HOST_WIDE_INT val_in
, machine_mode mode
)
4027 unsigned HOST_WIDE_INT val
, tmp
, mask
, first_one
, next_one
;
4030 /* Check for a single sequence of one bits and return quickly if so.
4031 The special cases of all ones and all zeroes returns false. */
4032 val
= (unsigned HOST_WIDE_INT
) val_in
;
4033 tmp
= val
+ (val
& -val
);
4035 if (tmp
== (tmp
& -tmp
))
4036 return (val
+ 1) > 1;
4038 /* Replicate 32-bit immediates so we can treat them as 64-bit. */
4040 val
= (val
<< 32) | (val
& 0xffffffff);
4042 /* Invert if the immediate doesn't start with a zero bit - this means we
4043 only need to search for sequences of one bits. */
4047 /* Find the first set bit and set tmp to val with the first sequence of one
4048 bits removed. Return success if there is a single sequence of ones. */
4049 first_one
= val
& -val
;
4050 tmp
= val
& (val
+ first_one
);
4055 /* Find the next set bit and compute the difference in bit position. */
4056 next_one
= tmp
& -tmp
;
4057 bits
= clz_hwi (first_one
) - clz_hwi (next_one
);
4060 /* Check the bit position difference is a power of 2, and that the first
4061 sequence of one bits fits within 'bits' bits. */
4062 if ((mask
>> bits
) != 0 || bits
!= (bits
& -bits
))
4065 /* Check the sequence of one bits is repeated 64/bits times. */
4066 return val
== mask
* bitmask_imm_mul
[__builtin_clz (bits
) - 26];
4069 /* Create mask of ones, covering the lowest to highest bits set in VAL_IN.
4070 Assumed precondition: VAL_IN Is not zero. */
4072 unsigned HOST_WIDE_INT
4073 aarch64_and_split_imm1 (HOST_WIDE_INT val_in
)
4075 int lowest_bit_set
= ctz_hwi (val_in
);
4076 int highest_bit_set
= floor_log2 (val_in
);
4077 gcc_assert (val_in
!= 0);
4079 return ((HOST_WIDE_INT_UC (2) << highest_bit_set
) -
4080 (HOST_WIDE_INT_1U
<< lowest_bit_set
));
4083 /* Create constant where bits outside of lowest bit set to highest bit set
4086 unsigned HOST_WIDE_INT
4087 aarch64_and_split_imm2 (HOST_WIDE_INT val_in
)
4089 return val_in
| ~aarch64_and_split_imm1 (val_in
);
4092 /* Return true if VAL_IN is a valid 'and' bitmask immediate. */
4095 aarch64_and_bitmask_imm (unsigned HOST_WIDE_INT val_in
, machine_mode mode
)
4097 scalar_int_mode int_mode
;
4098 if (!is_a
<scalar_int_mode
> (mode
, &int_mode
))
4101 if (aarch64_bitmask_imm (val_in
, int_mode
))
4104 if (aarch64_move_imm (val_in
, int_mode
))
4107 unsigned HOST_WIDE_INT imm2
= aarch64_and_split_imm2 (val_in
);
4109 return aarch64_bitmask_imm (imm2
, int_mode
);
4112 /* Return true if val is an immediate that can be loaded into a
4113 register in a single instruction. */
4115 aarch64_move_imm (HOST_WIDE_INT val
, machine_mode mode
)
4117 scalar_int_mode int_mode
;
4118 if (!is_a
<scalar_int_mode
> (mode
, &int_mode
))
4121 if (aarch64_movw_imm (val
, int_mode
) || aarch64_movw_imm (~val
, int_mode
))
4123 return aarch64_bitmask_imm (val
, int_mode
);
4127 aarch64_cannot_force_const_mem (machine_mode mode ATTRIBUTE_UNUSED
, rtx x
)
4131 if (GET_CODE (x
) == HIGH
)
4134 split_const (x
, &base
, &offset
);
4135 if (GET_CODE (base
) == SYMBOL_REF
|| GET_CODE (base
) == LABEL_REF
)
4137 if (aarch64_classify_symbol (base
, offset
)
4138 != SYMBOL_FORCE_TO_MEM
)
4141 /* Avoid generating a 64-bit relocation in ILP32; leave
4142 to aarch64_expand_mov_immediate to handle it properly. */
4143 return mode
!= ptr_mode
;
4146 return aarch64_tls_referenced_p (x
);
4149 /* Implement TARGET_CASE_VALUES_THRESHOLD.
4150 The expansion for a table switch is quite expensive due to the number
4151 of instructions, the table lookup and hard to predict indirect jump.
4152 When optimizing for speed, and -O3 enabled, use the per-core tuning if
4153 set, otherwise use tables for > 16 cases as a tradeoff between size and
4154 performance. When optimizing for size, use the default setting. */
4157 aarch64_case_values_threshold (void)
4159 /* Use the specified limit for the number of cases before using jump
4160 tables at higher optimization levels. */
4162 && selected_cpu
->tune
->max_case_values
!= 0)
4163 return selected_cpu
->tune
->max_case_values
;
4165 return optimize_size
? default_case_values_threshold () : 17;
4168 /* Return true if register REGNO is a valid index register.
4169 STRICT_P is true if REG_OK_STRICT is in effect. */
4172 aarch64_regno_ok_for_index_p (int regno
, bool strict_p
)
4174 if (!HARD_REGISTER_NUM_P (regno
))
4182 regno
= reg_renumber
[regno
];
4184 return GP_REGNUM_P (regno
);
4187 /* Return true if register REGNO is a valid base register for mode MODE.
4188 STRICT_P is true if REG_OK_STRICT is in effect. */
4191 aarch64_regno_ok_for_base_p (int regno
, bool strict_p
)
4193 if (!HARD_REGISTER_NUM_P (regno
))
4201 regno
= reg_renumber
[regno
];
4204 /* The fake registers will be eliminated to either the stack or
4205 hard frame pointer, both of which are usually valid base registers.
4206 Reload deals with the cases where the eliminated form isn't valid. */
4207 return (GP_REGNUM_P (regno
)
4208 || regno
== SP_REGNUM
4209 || regno
== FRAME_POINTER_REGNUM
4210 || regno
== ARG_POINTER_REGNUM
);
4213 /* Return true if X is a valid base register for mode MODE.
4214 STRICT_P is true if REG_OK_STRICT is in effect. */
4217 aarch64_base_register_rtx_p (rtx x
, bool strict_p
)
4220 && GET_CODE (x
) == SUBREG
4221 && contains_reg_of_mode
[GENERAL_REGS
][GET_MODE (SUBREG_REG (x
))])
4224 return (REG_P (x
) && aarch64_regno_ok_for_base_p (REGNO (x
), strict_p
));
4227 /* Return true if address offset is a valid index. If it is, fill in INFO
4228 appropriately. STRICT_P is true if REG_OK_STRICT is in effect. */
4231 aarch64_classify_index (struct aarch64_address_info
*info
, rtx x
,
4232 machine_mode mode
, bool strict_p
)
4234 enum aarch64_address_type type
;
4239 if ((REG_P (x
) || GET_CODE (x
) == SUBREG
)
4240 && GET_MODE (x
) == Pmode
)
4242 type
= ADDRESS_REG_REG
;
4246 /* (sign_extend:DI (reg:SI)) */
4247 else if ((GET_CODE (x
) == SIGN_EXTEND
4248 || GET_CODE (x
) == ZERO_EXTEND
)
4249 && GET_MODE (x
) == DImode
4250 && GET_MODE (XEXP (x
, 0)) == SImode
)
4252 type
= (GET_CODE (x
) == SIGN_EXTEND
)
4253 ? ADDRESS_REG_SXTW
: ADDRESS_REG_UXTW
;
4254 index
= XEXP (x
, 0);
4257 /* (mult:DI (sign_extend:DI (reg:SI)) (const_int scale)) */
4258 else if (GET_CODE (x
) == MULT
4259 && (GET_CODE (XEXP (x
, 0)) == SIGN_EXTEND
4260 || GET_CODE (XEXP (x
, 0)) == ZERO_EXTEND
)
4261 && GET_MODE (XEXP (x
, 0)) == DImode
4262 && GET_MODE (XEXP (XEXP (x
, 0), 0)) == SImode
4263 && CONST_INT_P (XEXP (x
, 1)))
4265 type
= (GET_CODE (XEXP (x
, 0)) == SIGN_EXTEND
)
4266 ? ADDRESS_REG_SXTW
: ADDRESS_REG_UXTW
;
4267 index
= XEXP (XEXP (x
, 0), 0);
4268 shift
= exact_log2 (INTVAL (XEXP (x
, 1)));
4270 /* (ashift:DI (sign_extend:DI (reg:SI)) (const_int shift)) */
4271 else if (GET_CODE (x
) == ASHIFT
4272 && (GET_CODE (XEXP (x
, 0)) == SIGN_EXTEND
4273 || GET_CODE (XEXP (x
, 0)) == ZERO_EXTEND
)
4274 && GET_MODE (XEXP (x
, 0)) == DImode
4275 && GET_MODE (XEXP (XEXP (x
, 0), 0)) == SImode
4276 && CONST_INT_P (XEXP (x
, 1)))
4278 type
= (GET_CODE (XEXP (x
, 0)) == SIGN_EXTEND
)
4279 ? ADDRESS_REG_SXTW
: ADDRESS_REG_UXTW
;
4280 index
= XEXP (XEXP (x
, 0), 0);
4281 shift
= INTVAL (XEXP (x
, 1));
4283 /* (sign_extract:DI (mult:DI (reg:DI) (const_int scale)) 32+shift 0) */
4284 else if ((GET_CODE (x
) == SIGN_EXTRACT
4285 || GET_CODE (x
) == ZERO_EXTRACT
)
4286 && GET_MODE (x
) == DImode
4287 && GET_CODE (XEXP (x
, 0)) == MULT
4288 && GET_MODE (XEXP (XEXP (x
, 0), 0)) == DImode
4289 && CONST_INT_P (XEXP (XEXP (x
, 0), 1)))
4291 type
= (GET_CODE (x
) == SIGN_EXTRACT
)
4292 ? ADDRESS_REG_SXTW
: ADDRESS_REG_UXTW
;
4293 index
= XEXP (XEXP (x
, 0), 0);
4294 shift
= exact_log2 (INTVAL (XEXP (XEXP (x
, 0), 1)));
4295 if (INTVAL (XEXP (x
, 1)) != 32 + shift
4296 || INTVAL (XEXP (x
, 2)) != 0)
4299 /* (and:DI (mult:DI (reg:DI) (const_int scale))
4300 (const_int 0xffffffff<<shift)) */
4301 else if (GET_CODE (x
) == AND
4302 && GET_MODE (x
) == DImode
4303 && GET_CODE (XEXP (x
, 0)) == MULT
4304 && GET_MODE (XEXP (XEXP (x
, 0), 0)) == DImode
4305 && CONST_INT_P (XEXP (XEXP (x
, 0), 1))
4306 && CONST_INT_P (XEXP (x
, 1)))
4308 type
= ADDRESS_REG_UXTW
;
4309 index
= XEXP (XEXP (x
, 0), 0);
4310 shift
= exact_log2 (INTVAL (XEXP (XEXP (x
, 0), 1)));
4311 if (INTVAL (XEXP (x
, 1)) != (HOST_WIDE_INT
)0xffffffff << shift
)
4314 /* (sign_extract:DI (ashift:DI (reg:DI) (const_int shift)) 32+shift 0) */
4315 else if ((GET_CODE (x
) == SIGN_EXTRACT
4316 || GET_CODE (x
) == ZERO_EXTRACT
)
4317 && GET_MODE (x
) == DImode
4318 && GET_CODE (XEXP (x
, 0)) == ASHIFT
4319 && GET_MODE (XEXP (XEXP (x
, 0), 0)) == DImode
4320 && CONST_INT_P (XEXP (XEXP (x
, 0), 1)))
4322 type
= (GET_CODE (x
) == SIGN_EXTRACT
)
4323 ? ADDRESS_REG_SXTW
: ADDRESS_REG_UXTW
;
4324 index
= XEXP (XEXP (x
, 0), 0);
4325 shift
= INTVAL (XEXP (XEXP (x
, 0), 1));
4326 if (INTVAL (XEXP (x
, 1)) != 32 + shift
4327 || INTVAL (XEXP (x
, 2)) != 0)
4330 /* (and:DI (ashift:DI (reg:DI) (const_int shift))
4331 (const_int 0xffffffff<<shift)) */
4332 else if (GET_CODE (x
) == AND
4333 && GET_MODE (x
) == DImode
4334 && GET_CODE (XEXP (x
, 0)) == ASHIFT
4335 && GET_MODE (XEXP (XEXP (x
, 0), 0)) == DImode
4336 && CONST_INT_P (XEXP (XEXP (x
, 0), 1))
4337 && CONST_INT_P (XEXP (x
, 1)))
4339 type
= ADDRESS_REG_UXTW
;
4340 index
= XEXP (XEXP (x
, 0), 0);
4341 shift
= INTVAL (XEXP (XEXP (x
, 0), 1));
4342 if (INTVAL (XEXP (x
, 1)) != (HOST_WIDE_INT
)0xffffffff << shift
)
4345 /* (mult:P (reg:P) (const_int scale)) */
4346 else if (GET_CODE (x
) == MULT
4347 && GET_MODE (x
) == Pmode
4348 && GET_MODE (XEXP (x
, 0)) == Pmode
4349 && CONST_INT_P (XEXP (x
, 1)))
4351 type
= ADDRESS_REG_REG
;
4352 index
= XEXP (x
, 0);
4353 shift
= exact_log2 (INTVAL (XEXP (x
, 1)));
4355 /* (ashift:P (reg:P) (const_int shift)) */
4356 else if (GET_CODE (x
) == ASHIFT
4357 && GET_MODE (x
) == Pmode
4358 && GET_MODE (XEXP (x
, 0)) == Pmode
4359 && CONST_INT_P (XEXP (x
, 1)))
4361 type
= ADDRESS_REG_REG
;
4362 index
= XEXP (x
, 0);
4363 shift
= INTVAL (XEXP (x
, 1));
4369 && GET_CODE (index
) == SUBREG
4370 && contains_reg_of_mode
[GENERAL_REGS
][GET_MODE (SUBREG_REG (index
))])
4371 index
= SUBREG_REG (index
);
4374 (shift
> 0 && shift
<= 3
4375 && (1 << shift
) == GET_MODE_SIZE (mode
)))
4377 && aarch64_regno_ok_for_index_p (REGNO (index
), strict_p
))
4380 info
->offset
= index
;
4381 info
->shift
= shift
;
4388 /* Return true if MODE is one of the modes for which we
4389 support LDP/STP operations. */
4392 aarch64_mode_valid_for_sched_fusion_p (machine_mode mode
)
4394 return mode
== SImode
|| mode
== DImode
4395 || mode
== SFmode
|| mode
== DFmode
4396 || (aarch64_vector_mode_supported_p (mode
)
4397 && GET_MODE_SIZE (mode
) == 8);
4400 /* Return true if REGNO is a virtual pointer register, or an eliminable
4401 "soft" frame register. Like REGNO_PTR_FRAME_P except that we don't
4402 include stack_pointer or hard_frame_pointer. */
4404 virt_or_elim_regno_p (unsigned regno
)
4406 return ((regno
>= FIRST_VIRTUAL_REGISTER
4407 && regno
<= LAST_VIRTUAL_POINTER_REGISTER
)
4408 || regno
== FRAME_POINTER_REGNUM
4409 || regno
== ARG_POINTER_REGNUM
);
4412 /* Return true if X is a valid address for machine mode MODE. If it is,
4413 fill in INFO appropriately. STRICT_P is true if REG_OK_STRICT is in
4414 effect. OUTER_CODE is PARALLEL for a load/store pair. */
4417 aarch64_classify_address (struct aarch64_address_info
*info
,
4418 rtx x
, machine_mode mode
,
4419 RTX_CODE outer_code
, bool strict_p
)
4421 enum rtx_code code
= GET_CODE (x
);
4424 /* On BE, we use load/store pair for all large int mode load/stores.
4425 TI/TFmode may also use a load/store pair. */
4426 bool load_store_pair_p
= (outer_code
== PARALLEL
4429 || (BYTES_BIG_ENDIAN
4430 && aarch64_vect_struct_mode_p (mode
)));
4432 bool allow_reg_index_p
=
4434 && (GET_MODE_SIZE (mode
) != 16 || aarch64_vector_mode_supported_p (mode
))
4435 && !aarch64_vect_struct_mode_p (mode
);
4437 /* On LE, for AdvSIMD, don't support anything other than POST_INC or
4439 if (aarch64_vect_struct_mode_p (mode
) && !BYTES_BIG_ENDIAN
4440 && (code
!= POST_INC
&& code
!= REG
))
4447 info
->type
= ADDRESS_REG_IMM
;
4449 info
->offset
= const0_rtx
;
4450 return aarch64_base_register_rtx_p (x
, strict_p
);
4458 && virt_or_elim_regno_p (REGNO (op0
))
4459 && CONST_INT_P (op1
))
4461 info
->type
= ADDRESS_REG_IMM
;
4468 if (GET_MODE_SIZE (mode
) != 0
4469 && CONST_INT_P (op1
)
4470 && aarch64_base_register_rtx_p (op0
, strict_p
))
4472 HOST_WIDE_INT offset
= INTVAL (op1
);
4474 info
->type
= ADDRESS_REG_IMM
;
4478 /* TImode and TFmode values are allowed in both pairs of X
4479 registers and individual Q registers. The available
4481 X,X: 7-bit signed scaled offset
4482 Q: 9-bit signed offset
4483 We conservatively require an offset representable in either mode.
4484 When performing the check for pairs of X registers i.e. LDP/STP
4485 pass down DImode since that is the natural size of the LDP/STP
4486 instruction memory accesses. */
4487 if (mode
== TImode
|| mode
== TFmode
)
4488 return (aarch64_offset_7bit_signed_scaled_p (DImode
, offset
)
4489 && (offset_9bit_signed_unscaled_p (mode
, offset
)
4490 || offset_12bit_unsigned_scaled_p (mode
, offset
)));
4492 /* A 7bit offset check because OImode will emit a ldp/stp
4493 instruction (only big endian will get here).
4494 For ldp/stp instructions, the offset is scaled for the size of a
4495 single element of the pair. */
4497 return aarch64_offset_7bit_signed_scaled_p (TImode
, offset
);
4499 /* Three 9/12 bit offsets checks because CImode will emit three
4500 ldr/str instructions (only big endian will get here). */
4502 return (aarch64_offset_7bit_signed_scaled_p (TImode
, offset
)
4503 && (offset_9bit_signed_unscaled_p (V16QImode
, offset
+ 32)
4504 || offset_12bit_unsigned_scaled_p (V16QImode
,
4507 /* Two 7bit offsets checks because XImode will emit two ldp/stp
4508 instructions (only big endian will get here). */
4510 return (aarch64_offset_7bit_signed_scaled_p (TImode
, offset
)
4511 && aarch64_offset_7bit_signed_scaled_p (TImode
,
4514 if (load_store_pair_p
)
4515 return ((GET_MODE_SIZE (mode
) == 4 || GET_MODE_SIZE (mode
) == 8)
4516 && aarch64_offset_7bit_signed_scaled_p (mode
, offset
));
4518 return (offset_9bit_signed_unscaled_p (mode
, offset
)
4519 || offset_12bit_unsigned_scaled_p (mode
, offset
));
4522 if (allow_reg_index_p
)
4524 /* Look for base + (scaled/extended) index register. */
4525 if (aarch64_base_register_rtx_p (op0
, strict_p
)
4526 && aarch64_classify_index (info
, op1
, mode
, strict_p
))
4531 if (aarch64_base_register_rtx_p (op1
, strict_p
)
4532 && aarch64_classify_index (info
, op0
, mode
, strict_p
))
4545 info
->type
= ADDRESS_REG_WB
;
4546 info
->base
= XEXP (x
, 0);
4547 info
->offset
= NULL_RTX
;
4548 return aarch64_base_register_rtx_p (info
->base
, strict_p
);
4552 info
->type
= ADDRESS_REG_WB
;
4553 info
->base
= XEXP (x
, 0);
4554 if (GET_CODE (XEXP (x
, 1)) == PLUS
4555 && CONST_INT_P (XEXP (XEXP (x
, 1), 1))
4556 && rtx_equal_p (XEXP (XEXP (x
, 1), 0), info
->base
)
4557 && aarch64_base_register_rtx_p (info
->base
, strict_p
))
4559 HOST_WIDE_INT offset
;
4560 info
->offset
= XEXP (XEXP (x
, 1), 1);
4561 offset
= INTVAL (info
->offset
);
4563 /* TImode and TFmode values are allowed in both pairs of X
4564 registers and individual Q registers. The available
4566 X,X: 7-bit signed scaled offset
4567 Q: 9-bit signed offset
4568 We conservatively require an offset representable in either mode.
4570 if (mode
== TImode
|| mode
== TFmode
)
4571 return (aarch64_offset_7bit_signed_scaled_p (mode
, offset
)
4572 && offset_9bit_signed_unscaled_p (mode
, offset
));
4574 if (load_store_pair_p
)
4575 return ((GET_MODE_SIZE (mode
) == 4 || GET_MODE_SIZE (mode
) == 8)
4576 && aarch64_offset_7bit_signed_scaled_p (mode
, offset
));
4578 return offset_9bit_signed_unscaled_p (mode
, offset
);
4585 /* load literal: pc-relative constant pool entry. Only supported
4586 for SI mode or larger. */
4587 info
->type
= ADDRESS_SYMBOLIC
;
4589 if (!load_store_pair_p
&& GET_MODE_SIZE (mode
) >= 4)
4593 split_const (x
, &sym
, &addend
);
4594 return ((GET_CODE (sym
) == LABEL_REF
4595 || (GET_CODE (sym
) == SYMBOL_REF
4596 && CONSTANT_POOL_ADDRESS_P (sym
)
4597 && aarch64_pcrelative_literal_loads
)));
4602 info
->type
= ADDRESS_LO_SUM
;
4603 info
->base
= XEXP (x
, 0);
4604 info
->offset
= XEXP (x
, 1);
4605 if (allow_reg_index_p
4606 && aarch64_base_register_rtx_p (info
->base
, strict_p
))
4609 split_const (info
->offset
, &sym
, &offs
);
4610 if (GET_CODE (sym
) == SYMBOL_REF
4611 && (aarch64_classify_symbol (sym
, offs
) == SYMBOL_SMALL_ABSOLUTE
))
4613 /* The symbol and offset must be aligned to the access size. */
4615 unsigned int ref_size
;
4617 if (CONSTANT_POOL_ADDRESS_P (sym
))
4618 align
= GET_MODE_ALIGNMENT (get_pool_mode (sym
));
4619 else if (TREE_CONSTANT_POOL_ADDRESS_P (sym
))
4621 tree exp
= SYMBOL_REF_DECL (sym
);
4622 align
= TYPE_ALIGN (TREE_TYPE (exp
));
4623 align
= CONSTANT_ALIGNMENT (exp
, align
);
4625 else if (SYMBOL_REF_DECL (sym
))
4626 align
= DECL_ALIGN (SYMBOL_REF_DECL (sym
));
4627 else if (SYMBOL_REF_HAS_BLOCK_INFO_P (sym
)
4628 && SYMBOL_REF_BLOCK (sym
) != NULL
)
4629 align
= SYMBOL_REF_BLOCK (sym
)->alignment
;
4631 align
= BITS_PER_UNIT
;
4633 ref_size
= GET_MODE_SIZE (mode
);
4635 ref_size
= GET_MODE_SIZE (DImode
);
4637 return ((INTVAL (offs
) & (ref_size
- 1)) == 0
4638 && ((align
/ BITS_PER_UNIT
) & (ref_size
- 1)) == 0);
4648 /* Return true if the address X is valid for a PRFM instruction.
4649 STRICT_P is true if we should do strict checking with
4650 aarch64_classify_address. */
4653 aarch64_address_valid_for_prefetch_p (rtx x
, bool strict_p
)
4655 struct aarch64_address_info addr
;
4657 /* PRFM accepts the same addresses as DImode... */
4658 bool res
= aarch64_classify_address (&addr
, x
, DImode
, MEM
, strict_p
);
4662 /* ... except writeback forms. */
4663 return addr
.type
!= ADDRESS_REG_WB
;
4667 aarch64_symbolic_address_p (rtx x
)
4671 split_const (x
, &x
, &offset
);
4672 return GET_CODE (x
) == SYMBOL_REF
|| GET_CODE (x
) == LABEL_REF
;
4675 /* Classify the base of symbolic expression X. */
4677 enum aarch64_symbol_type
4678 aarch64_classify_symbolic_expression (rtx x
)
4682 split_const (x
, &x
, &offset
);
4683 return aarch64_classify_symbol (x
, offset
);
4687 /* Return TRUE if X is a legitimate address for accessing memory in
4690 aarch64_legitimate_address_hook_p (machine_mode mode
, rtx x
, bool strict_p
)
4692 struct aarch64_address_info addr
;
4694 return aarch64_classify_address (&addr
, x
, mode
, MEM
, strict_p
);
4697 /* Return TRUE if X is a legitimate address for accessing memory in
4698 mode MODE. OUTER_CODE will be PARALLEL if this is a load/store
4701 aarch64_legitimate_address_p (machine_mode mode
, rtx x
,
4702 RTX_CODE outer_code
, bool strict_p
)
4704 struct aarch64_address_info addr
;
4706 return aarch64_classify_address (&addr
, x
, mode
, outer_code
, strict_p
);
4709 /* Split an out-of-range address displacement into a base and offset.
4710 Use 4KB range for 1- and 2-byte accesses and a 16KB range otherwise
4711 to increase opportunities for sharing the base address of different sizes.
4712 For unaligned accesses and TI/TF mode use the signed 9-bit range. */
4714 aarch64_legitimize_address_displacement (rtx
*disp
, rtx
*off
, machine_mode mode
)
4716 HOST_WIDE_INT offset
= INTVAL (*disp
);
4717 HOST_WIDE_INT base
= offset
& ~(GET_MODE_SIZE (mode
) < 4 ? 0xfff : 0x3ffc);
4719 if (mode
== TImode
|| mode
== TFmode
4720 || (offset
& (GET_MODE_SIZE (mode
) - 1)) != 0)
4721 base
= (offset
+ 0x100) & ~0x1ff;
4723 *off
= GEN_INT (base
);
4724 *disp
= GEN_INT (offset
- base
);
4728 /* Return the binary representation of floating point constant VALUE in INTVAL.
4729 If the value cannot be converted, return false without setting INTVAL.
4730 The conversion is done in the given MODE. */
4732 aarch64_reinterpret_float_as_int (rtx value
, unsigned HOST_WIDE_INT
*intval
)
4735 /* We make a general exception for 0. */
4736 if (aarch64_float_const_zero_rtx_p (value
))
4742 machine_mode mode
= GET_MODE (value
);
4743 if (GET_CODE (value
) != CONST_DOUBLE
4744 || !SCALAR_FLOAT_MODE_P (mode
)
4745 || GET_MODE_BITSIZE (mode
) > HOST_BITS_PER_WIDE_INT
4746 /* Only support up to DF mode. */
4747 || GET_MODE_BITSIZE (mode
) > GET_MODE_BITSIZE (DFmode
))
4750 unsigned HOST_WIDE_INT ival
= 0;
4753 real_to_target (res
,
4754 CONST_DOUBLE_REAL_VALUE (value
),
4755 REAL_MODE_FORMAT (mode
));
4759 int order
= BYTES_BIG_ENDIAN
? 1 : 0;
4760 ival
= zext_hwi (res
[order
], 32);
4761 ival
|= (zext_hwi (res
[1 - order
], 32) << 32);
4764 ival
= zext_hwi (res
[0], 32);
4770 /* Return TRUE if rtx X is an immediate constant that can be moved using a
4771 single MOV(+MOVK) followed by an FMOV. */
4773 aarch64_float_const_rtx_p (rtx x
)
4775 machine_mode mode
= GET_MODE (x
);
4776 if (mode
== VOIDmode
)
4779 /* Determine whether it's cheaper to write float constants as
4780 mov/movk pairs over ldr/adrp pairs. */
4781 unsigned HOST_WIDE_INT ival
;
4783 if (GET_CODE (x
) == CONST_DOUBLE
4784 && SCALAR_FLOAT_MODE_P (mode
)
4785 && aarch64_reinterpret_float_as_int (x
, &ival
))
4787 scalar_int_mode imode
= (mode
== HFmode
4789 : int_mode_for_mode (mode
).require ());
4790 int num_instr
= aarch64_internal_mov_immediate
4791 (NULL_RTX
, gen_int_mode (ival
, imode
), false, imode
);
4792 return num_instr
< 3;
4798 /* Return TRUE if rtx X is immediate constant 0.0 */
4800 aarch64_float_const_zero_rtx_p (rtx x
)
4802 if (GET_MODE (x
) == VOIDmode
)
4805 if (REAL_VALUE_MINUS_ZERO (*CONST_DOUBLE_REAL_VALUE (x
)))
4806 return !HONOR_SIGNED_ZEROS (GET_MODE (x
));
4807 return real_equal (CONST_DOUBLE_REAL_VALUE (x
), &dconst0
);
4810 /* Return TRUE if rtx X is immediate constant that fits in a single
4811 MOVI immediate operation. */
4813 aarch64_can_const_movi_rtx_p (rtx x
, machine_mode mode
)
4819 scalar_int_mode imode
;
4820 unsigned HOST_WIDE_INT ival
;
4822 if (GET_CODE (x
) == CONST_DOUBLE
4823 && SCALAR_FLOAT_MODE_P (mode
))
4825 if (!aarch64_reinterpret_float_as_int (x
, &ival
))
4828 /* We make a general exception for 0. */
4829 if (aarch64_float_const_zero_rtx_p (x
))
4832 imode
= int_mode_for_mode (mode
).require ();
4834 else if (GET_CODE (x
) == CONST_INT
4835 && is_a
<scalar_int_mode
> (mode
, &imode
))
4840 /* use a 64 bit mode for everything except for DI/DF mode, where we use
4841 a 128 bit vector mode. */
4842 int width
= GET_MODE_BITSIZE (imode
) == 64 ? 128 : 64;
4844 vmode
= aarch64_simd_container_mode (imode
, width
);
4845 rtx v_op
= aarch64_simd_gen_const_vector_dup (vmode
, ival
);
4847 return aarch64_simd_valid_immediate (v_op
, vmode
, false, NULL
);
4851 /* Return the fixed registers used for condition codes. */
4854 aarch64_fixed_condition_code_regs (unsigned int *p1
, unsigned int *p2
)
4857 *p2
= INVALID_REGNUM
;
4861 /* This function is used by the call expanders of the machine description.
4862 RESULT is the register in which the result is returned. It's NULL for
4863 "call" and "sibcall".
4864 MEM is the location of the function call.
4865 SIBCALL indicates whether this function call is normal call or sibling call.
4866 It will generate different pattern accordingly. */
4869 aarch64_expand_call (rtx result
, rtx mem
, bool sibcall
)
4871 rtx call
, callee
, tmp
;
4875 gcc_assert (MEM_P (mem
));
4876 callee
= XEXP (mem
, 0);
4877 mode
= GET_MODE (callee
);
4878 gcc_assert (mode
== Pmode
);
4880 /* Decide if we should generate indirect calls by loading the
4881 address of the callee into a register before performing
4882 the branch-and-link. */
4883 if (SYMBOL_REF_P (callee
)
4884 ? (aarch64_is_long_call_p (callee
)
4885 || aarch64_is_noplt_call_p (callee
))
4887 XEXP (mem
, 0) = force_reg (mode
, callee
);
4889 call
= gen_rtx_CALL (VOIDmode
, mem
, const0_rtx
);
4891 if (result
!= NULL_RTX
)
4892 call
= gen_rtx_SET (result
, call
);
4897 tmp
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (Pmode
, LR_REGNUM
));
4899 vec
= gen_rtvec (2, call
, tmp
);
4900 call
= gen_rtx_PARALLEL (VOIDmode
, vec
);
4902 aarch64_emit_call_insn (call
);
4905 /* Emit call insn with PAT and do aarch64-specific handling. */
4908 aarch64_emit_call_insn (rtx pat
)
4910 rtx insn
= emit_call_insn (pat
);
4912 rtx
*fusage
= &CALL_INSN_FUNCTION_USAGE (insn
);
4913 clobber_reg (fusage
, gen_rtx_REG (word_mode
, IP0_REGNUM
));
4914 clobber_reg (fusage
, gen_rtx_REG (word_mode
, IP1_REGNUM
));
4918 aarch64_select_cc_mode (RTX_CODE code
, rtx x
, rtx y
)
4920 /* All floating point compares return CCFP if it is an equality
4921 comparison, and CCFPE otherwise. */
4922 if (GET_MODE_CLASS (GET_MODE (x
)) == MODE_FLOAT
)
4949 /* Equality comparisons of short modes against zero can be performed
4950 using the TST instruction with the appropriate bitmask. */
4951 if (y
== const0_rtx
&& REG_P (x
)
4952 && (code
== EQ
|| code
== NE
)
4953 && (GET_MODE (x
) == HImode
|| GET_MODE (x
) == QImode
))
4956 /* Similarly, comparisons of zero_extends from shorter modes can
4957 be performed using an ANDS with an immediate mask. */
4958 if (y
== const0_rtx
&& GET_CODE (x
) == ZERO_EXTEND
4959 && (GET_MODE (x
) == SImode
|| GET_MODE (x
) == DImode
)
4960 && (GET_MODE (XEXP (x
, 0)) == HImode
|| GET_MODE (XEXP (x
, 0)) == QImode
)
4961 && (code
== EQ
|| code
== NE
))
4964 if ((GET_MODE (x
) == SImode
|| GET_MODE (x
) == DImode
)
4966 && (code
== EQ
|| code
== NE
|| code
== LT
|| code
== GE
)
4967 && (GET_CODE (x
) == PLUS
|| GET_CODE (x
) == MINUS
|| GET_CODE (x
) == AND
4968 || GET_CODE (x
) == NEG
4969 || (GET_CODE (x
) == ZERO_EXTRACT
&& CONST_INT_P (XEXP (x
, 1))
4970 && CONST_INT_P (XEXP (x
, 2)))))
4973 /* A compare with a shifted operand. Because of canonicalization,
4974 the comparison will have to be swapped when we emit the assembly
4976 if ((GET_MODE (x
) == SImode
|| GET_MODE (x
) == DImode
)
4977 && (REG_P (y
) || GET_CODE (y
) == SUBREG
|| y
== const0_rtx
)
4978 && (GET_CODE (x
) == ASHIFT
|| GET_CODE (x
) == ASHIFTRT
4979 || GET_CODE (x
) == LSHIFTRT
4980 || GET_CODE (x
) == ZERO_EXTEND
|| GET_CODE (x
) == SIGN_EXTEND
))
4983 /* Similarly for a negated operand, but we can only do this for
4985 if ((GET_MODE (x
) == SImode
|| GET_MODE (x
) == DImode
)
4986 && (REG_P (y
) || GET_CODE (y
) == SUBREG
)
4987 && (code
== EQ
|| code
== NE
)
4988 && GET_CODE (x
) == NEG
)
4991 /* A test for unsigned overflow. */
4992 if ((GET_MODE (x
) == DImode
|| GET_MODE (x
) == TImode
)
4994 && GET_CODE (x
) == PLUS
4995 && GET_CODE (y
) == ZERO_EXTEND
)
4998 /* For everything else, return CCmode. */
5003 aarch64_get_condition_code_1 (machine_mode
, enum rtx_code
);
5006 aarch64_get_condition_code (rtx x
)
5008 machine_mode mode
= GET_MODE (XEXP (x
, 0));
5009 enum rtx_code comp_code
= GET_CODE (x
);
5011 if (GET_MODE_CLASS (mode
) != MODE_CC
)
5012 mode
= SELECT_CC_MODE (comp_code
, XEXP (x
, 0), XEXP (x
, 1));
5013 return aarch64_get_condition_code_1 (mode
, comp_code
);
5017 aarch64_get_condition_code_1 (machine_mode mode
, enum rtx_code comp_code
)
5025 case GE
: return AARCH64_GE
;
5026 case GT
: return AARCH64_GT
;
5027 case LE
: return AARCH64_LS
;
5028 case LT
: return AARCH64_MI
;
5029 case NE
: return AARCH64_NE
;
5030 case EQ
: return AARCH64_EQ
;
5031 case ORDERED
: return AARCH64_VC
;
5032 case UNORDERED
: return AARCH64_VS
;
5033 case UNLT
: return AARCH64_LT
;
5034 case UNLE
: return AARCH64_LE
;
5035 case UNGT
: return AARCH64_HI
;
5036 case UNGE
: return AARCH64_PL
;
5044 case NE
: return AARCH64_NE
;
5045 case EQ
: return AARCH64_EQ
;
5046 case GE
: return AARCH64_GE
;
5047 case GT
: return AARCH64_GT
;
5048 case LE
: return AARCH64_LE
;
5049 case LT
: return AARCH64_LT
;
5050 case GEU
: return AARCH64_CS
;
5051 case GTU
: return AARCH64_HI
;
5052 case LEU
: return AARCH64_LS
;
5053 case LTU
: return AARCH64_CC
;
5061 case NE
: return AARCH64_NE
;
5062 case EQ
: return AARCH64_EQ
;
5063 case GE
: return AARCH64_LE
;
5064 case GT
: return AARCH64_LT
;
5065 case LE
: return AARCH64_GE
;
5066 case LT
: return AARCH64_GT
;
5067 case GEU
: return AARCH64_LS
;
5068 case GTU
: return AARCH64_CC
;
5069 case LEU
: return AARCH64_CS
;
5070 case LTU
: return AARCH64_HI
;
5078 case NE
: return AARCH64_NE
;
5079 case EQ
: return AARCH64_EQ
;
5080 case GE
: return AARCH64_PL
;
5081 case LT
: return AARCH64_MI
;
5089 case NE
: return AARCH64_NE
;
5090 case EQ
: return AARCH64_EQ
;
5098 case NE
: return AARCH64_CS
;
5099 case EQ
: return AARCH64_CC
;
5112 aarch64_const_vec_all_same_in_range_p (rtx x
,
5113 HOST_WIDE_INT minval
,
5114 HOST_WIDE_INT maxval
)
5116 HOST_WIDE_INT firstval
;
5119 if (GET_CODE (x
) != CONST_VECTOR
5120 || GET_MODE_CLASS (GET_MODE (x
)) != MODE_VECTOR_INT
)
5123 firstval
= INTVAL (CONST_VECTOR_ELT (x
, 0));
5124 if (firstval
< minval
|| firstval
> maxval
)
5127 count
= CONST_VECTOR_NUNITS (x
);
5128 for (i
= 1; i
< count
; i
++)
5129 if (INTVAL (CONST_VECTOR_ELT (x
, i
)) != firstval
)
5136 aarch64_const_vec_all_same_int_p (rtx x
, HOST_WIDE_INT val
)
5138 return aarch64_const_vec_all_same_in_range_p (x
, val
, val
);
5143 #define AARCH64_CC_V 1
5144 #define AARCH64_CC_C (1 << 1)
5145 #define AARCH64_CC_Z (1 << 2)
5146 #define AARCH64_CC_N (1 << 3)
5148 /* N Z C V flags for ccmp. Indexed by AARCH64_COND_CODE. */
5149 static const int aarch64_nzcv_codes
[] =
5151 0, /* EQ, Z == 1. */
5152 AARCH64_CC_Z
, /* NE, Z == 0. */
5153 0, /* CS, C == 1. */
5154 AARCH64_CC_C
, /* CC, C == 0. */
5155 0, /* MI, N == 1. */
5156 AARCH64_CC_N
, /* PL, N == 0. */
5157 0, /* VS, V == 1. */
5158 AARCH64_CC_V
, /* VC, V == 0. */
5159 0, /* HI, C ==1 && Z == 0. */
5160 AARCH64_CC_C
, /* LS, !(C == 1 && Z == 0). */
5161 AARCH64_CC_V
, /* GE, N == V. */
5162 0, /* LT, N != V. */
5163 AARCH64_CC_Z
, /* GT, Z == 0 && N == V. */
5164 0, /* LE, !(Z == 0 && N == V). */
5169 /* Print operand X to file F in a target specific manner according to CODE.
5170 The acceptable formatting commands given by CODE are:
5171 'c': An integer or symbol address without a preceding #
5173 'e': Print the sign/zero-extend size as a character 8->b,
5175 'p': Prints N such that 2^N == X (X must be power of 2 and
5177 'P': Print the number of non-zero bits in X (a const_int).
5178 'H': Print the higher numbered register of a pair (TImode)
5180 'm': Print a condition (eq, ne, etc).
5181 'M': Same as 'm', but invert condition.
5182 'b/h/s/d/q': Print a scalar FP/SIMD register name.
5183 'S/T/U/V': Print a FP/SIMD register name for a register list.
5184 The register printed is the FP/SIMD register name
5185 of X + 0/1/2/3 for S/T/U/V.
5186 'R': Print a scalar FP/SIMD register name + 1.
5187 'X': Print bottom 16 bits of integer constant in hex.
5188 'w/x': Print a general register name or the zero register
5190 '0': Print a normal operand, if it's a general register,
5191 then we assume DImode.
5192 'k': Print NZCV for conditional compare instructions.
5193 'A': Output address constant representing the first
5194 argument of X, specifying a relocation offset
5196 'L': Output constant address specified by X
5197 with a relocation offset if appropriate.
5198 'G': Prints address of X, specifying a PC relative
5199 relocation mode if appropriate. */
5202 aarch64_print_operand (FILE *f
, rtx x
, int code
)
5207 switch (GET_CODE (x
))
5210 fprintf (f
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
));
5214 output_addr_const (f
, x
);
5218 if (GET_CODE (XEXP (x
, 0)) == PLUS
5219 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == SYMBOL_REF
)
5221 output_addr_const (f
, x
);
5227 output_operand_lossage ("Unsupported operand for code '%c'", code
);
5235 if (!CONST_INT_P (x
)
5236 || (n
= exact_log2 (INTVAL (x
) & ~7)) <= 0)
5238 output_operand_lossage ("invalid operand for '%%%c'", code
);
5254 output_operand_lossage ("invalid operand for '%%%c'", code
);
5264 if (!CONST_INT_P (x
) || (n
= exact_log2 (INTVAL (x
))) < 0)
5266 output_operand_lossage ("invalid operand for '%%%c'", code
);
5270 asm_fprintf (f
, "%d", n
);
5275 if (!CONST_INT_P (x
))
5277 output_operand_lossage ("invalid operand for '%%%c'", code
);
5281 asm_fprintf (f
, "%u", popcount_hwi (INTVAL (x
)));
5285 if (!REG_P (x
) || !GP_REGNUM_P (REGNO (x
) + 1))
5287 output_operand_lossage ("invalid operand for '%%%c'", code
);
5291 asm_fprintf (f
, "%s", reg_names
[REGNO (x
) + 1]);
5298 /* CONST_TRUE_RTX means al/nv (al is the default, don't print it). */
5299 if (x
== const_true_rtx
)
5306 if (!COMPARISON_P (x
))
5308 output_operand_lossage ("invalid operand for '%%%c'", code
);
5312 cond_code
= aarch64_get_condition_code (x
);
5313 gcc_assert (cond_code
>= 0);
5315 cond_code
= AARCH64_INVERSE_CONDITION_CODE (cond_code
);
5316 fputs (aarch64_condition_codes
[cond_code
], f
);
5325 if (!REG_P (x
) || !FP_REGNUM_P (REGNO (x
)))
5327 output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code
);
5330 asm_fprintf (f
, "%c%d", code
, REGNO (x
) - V0_REGNUM
);
5337 if (!REG_P (x
) || !FP_REGNUM_P (REGNO (x
)))
5339 output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code
);
5342 asm_fprintf (f
, "v%d", REGNO (x
) - V0_REGNUM
+ (code
- 'S'));
5346 if (!REG_P (x
) || !FP_REGNUM_P (REGNO (x
)))
5348 output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code
);
5351 asm_fprintf (f
, "q%d", REGNO (x
) - V0_REGNUM
+ 1);
5355 if (!CONST_INT_P (x
))
5357 output_operand_lossage ("invalid operand for '%%%c'", code
);
5360 asm_fprintf (f
, "0x%wx", UINTVAL (x
) & 0xffff);
5366 || (CONST_DOUBLE_P (x
) && aarch64_float_const_zero_rtx_p (x
)))
5368 asm_fprintf (f
, "%czr", code
);
5372 if (REG_P (x
) && GP_REGNUM_P (REGNO (x
)))
5374 asm_fprintf (f
, "%c%d", code
, REGNO (x
) - R0_REGNUM
);
5378 if (REG_P (x
) && REGNO (x
) == SP_REGNUM
)
5380 asm_fprintf (f
, "%ssp", code
== 'w' ? "w" : "");
5389 output_operand_lossage ("missing operand");
5393 switch (GET_CODE (x
))
5396 asm_fprintf (f
, "%s", reg_names
[REGNO (x
)]);
5400 output_address (GET_MODE (x
), XEXP (x
, 0));
5401 /* Check all memory references are Pmode - even with ILP32. */
5402 gcc_assert (GET_MODE (XEXP (x
, 0)) == Pmode
);
5408 output_addr_const (asm_out_file
, x
);
5412 asm_fprintf (f
, "%wd", INTVAL (x
));
5416 if (GET_MODE_CLASS (GET_MODE (x
)) == MODE_VECTOR_INT
)
5419 aarch64_const_vec_all_same_in_range_p (x
,
5421 HOST_WIDE_INT_MAX
));
5422 asm_fprintf (f
, "%wd", INTVAL (CONST_VECTOR_ELT (x
, 0)));
5424 else if (aarch64_simd_imm_zero_p (x
, GET_MODE (x
)))
5433 /* Since we define TARGET_SUPPORTS_WIDE_INT we shouldn't ever
5434 be getting CONST_DOUBLEs holding integers. */
5435 gcc_assert (GET_MODE (x
) != VOIDmode
);
5436 if (aarch64_float_const_zero_rtx_p (x
))
5441 else if (aarch64_float_const_representable_p (x
))
5444 char float_buf
[buf_size
] = {'\0'};
5445 real_to_decimal_for_mode (float_buf
,
5446 CONST_DOUBLE_REAL_VALUE (x
),
5449 asm_fprintf (asm_out_file
, "%s", float_buf
);
5453 output_operand_lossage ("invalid constant");
5456 output_operand_lossage ("invalid operand");
5462 if (GET_CODE (x
) == HIGH
)
5465 switch (aarch64_classify_symbolic_expression (x
))
5467 case SYMBOL_SMALL_GOT_4G
:
5468 asm_fprintf (asm_out_file
, ":got:");
5471 case SYMBOL_SMALL_TLSGD
:
5472 asm_fprintf (asm_out_file
, ":tlsgd:");
5475 case SYMBOL_SMALL_TLSDESC
:
5476 asm_fprintf (asm_out_file
, ":tlsdesc:");
5479 case SYMBOL_SMALL_TLSIE
:
5480 asm_fprintf (asm_out_file
, ":gottprel:");
5483 case SYMBOL_TLSLE24
:
5484 asm_fprintf (asm_out_file
, ":tprel:");
5487 case SYMBOL_TINY_GOT
:
5494 output_addr_const (asm_out_file
, x
);
5498 switch (aarch64_classify_symbolic_expression (x
))
5500 case SYMBOL_SMALL_GOT_4G
:
5501 asm_fprintf (asm_out_file
, ":lo12:");
5504 case SYMBOL_SMALL_TLSGD
:
5505 asm_fprintf (asm_out_file
, ":tlsgd_lo12:");
5508 case SYMBOL_SMALL_TLSDESC
:
5509 asm_fprintf (asm_out_file
, ":tlsdesc_lo12:");
5512 case SYMBOL_SMALL_TLSIE
:
5513 asm_fprintf (asm_out_file
, ":gottprel_lo12:");
5516 case SYMBOL_TLSLE12
:
5517 asm_fprintf (asm_out_file
, ":tprel_lo12:");
5520 case SYMBOL_TLSLE24
:
5521 asm_fprintf (asm_out_file
, ":tprel_lo12_nc:");
5524 case SYMBOL_TINY_GOT
:
5525 asm_fprintf (asm_out_file
, ":got:");
5528 case SYMBOL_TINY_TLSIE
:
5529 asm_fprintf (asm_out_file
, ":gottprel:");
5535 output_addr_const (asm_out_file
, x
);
5539 switch (aarch64_classify_symbolic_expression (x
))
5541 case SYMBOL_TLSLE24
:
5542 asm_fprintf (asm_out_file
, ":tprel_hi12:");
5547 output_addr_const (asm_out_file
, x
);
5552 HOST_WIDE_INT cond_code
;
5554 if (!CONST_INT_P (x
))
5556 output_operand_lossage ("invalid operand for '%%%c'", code
);
5560 cond_code
= INTVAL (x
);
5561 gcc_assert (cond_code
>= 0 && cond_code
<= AARCH64_NV
);
5562 asm_fprintf (f
, "%d", aarch64_nzcv_codes
[cond_code
]);
5567 output_operand_lossage ("invalid operand prefix '%%%c'", code
);
5573 aarch64_print_operand_address (FILE *f
, machine_mode mode
, rtx x
)
5575 struct aarch64_address_info addr
;
5577 if (aarch64_classify_address (&addr
, x
, mode
, MEM
, true))
5580 case ADDRESS_REG_IMM
:
5581 if (addr
.offset
== const0_rtx
)
5582 asm_fprintf (f
, "[%s]", reg_names
[REGNO (addr
.base
)]);
5584 asm_fprintf (f
, "[%s, %wd]", reg_names
[REGNO (addr
.base
)],
5585 INTVAL (addr
.offset
));
5588 case ADDRESS_REG_REG
:
5589 if (addr
.shift
== 0)
5590 asm_fprintf (f
, "[%s, %s]", reg_names
[REGNO (addr
.base
)],
5591 reg_names
[REGNO (addr
.offset
)]);
5593 asm_fprintf (f
, "[%s, %s, lsl %u]", reg_names
[REGNO (addr
.base
)],
5594 reg_names
[REGNO (addr
.offset
)], addr
.shift
);
5597 case ADDRESS_REG_UXTW
:
5598 if (addr
.shift
== 0)
5599 asm_fprintf (f
, "[%s, w%d, uxtw]", reg_names
[REGNO (addr
.base
)],
5600 REGNO (addr
.offset
) - R0_REGNUM
);
5602 asm_fprintf (f
, "[%s, w%d, uxtw %u]", reg_names
[REGNO (addr
.base
)],
5603 REGNO (addr
.offset
) - R0_REGNUM
, addr
.shift
);
5606 case ADDRESS_REG_SXTW
:
5607 if (addr
.shift
== 0)
5608 asm_fprintf (f
, "[%s, w%d, sxtw]", reg_names
[REGNO (addr
.base
)],
5609 REGNO (addr
.offset
) - R0_REGNUM
);
5611 asm_fprintf (f
, "[%s, w%d, sxtw %u]", reg_names
[REGNO (addr
.base
)],
5612 REGNO (addr
.offset
) - R0_REGNUM
, addr
.shift
);
5615 case ADDRESS_REG_WB
:
5616 switch (GET_CODE (x
))
5619 asm_fprintf (f
, "[%s, %d]!", reg_names
[REGNO (addr
.base
)],
5620 GET_MODE_SIZE (mode
));
5623 asm_fprintf (f
, "[%s], %d", reg_names
[REGNO (addr
.base
)],
5624 GET_MODE_SIZE (mode
));
5627 asm_fprintf (f
, "[%s, -%d]!", reg_names
[REGNO (addr
.base
)],
5628 GET_MODE_SIZE (mode
));
5631 asm_fprintf (f
, "[%s], -%d", reg_names
[REGNO (addr
.base
)],
5632 GET_MODE_SIZE (mode
));
5635 asm_fprintf (f
, "[%s, %wd]!", reg_names
[REGNO (addr
.base
)],
5636 INTVAL (addr
.offset
));
5639 asm_fprintf (f
, "[%s], %wd", reg_names
[REGNO (addr
.base
)],
5640 INTVAL (addr
.offset
));
5647 case ADDRESS_LO_SUM
:
5648 asm_fprintf (f
, "[%s, #:lo12:", reg_names
[REGNO (addr
.base
)]);
5649 output_addr_const (f
, addr
.offset
);
5650 asm_fprintf (f
, "]");
5653 case ADDRESS_SYMBOLIC
:
5657 output_addr_const (f
, x
);
5661 aarch64_label_mentioned_p (rtx x
)
5666 if (GET_CODE (x
) == LABEL_REF
)
5669 /* UNSPEC_TLS entries for a symbol include a LABEL_REF for the
5670 referencing instruction, but they are constant offsets, not
5672 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLS
)
5675 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5676 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5682 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
5683 if (aarch64_label_mentioned_p (XVECEXP (x
, i
, j
)))
5686 else if (fmt
[i
] == 'e' && aarch64_label_mentioned_p (XEXP (x
, i
)))
5693 /* Implement REGNO_REG_CLASS. */
5696 aarch64_regno_regclass (unsigned regno
)
5698 if (GP_REGNUM_P (regno
))
5699 return GENERAL_REGS
;
5701 if (regno
== SP_REGNUM
)
5704 if (regno
== FRAME_POINTER_REGNUM
5705 || regno
== ARG_POINTER_REGNUM
)
5706 return POINTER_REGS
;
5708 if (FP_REGNUM_P (regno
))
5709 return FP_LO_REGNUM_P (regno
) ? FP_LO_REGS
: FP_REGS
;
5715 aarch64_legitimize_address (rtx x
, rtx
/* orig_x */, machine_mode mode
)
5717 /* Try to split X+CONST into Y=X+(CONST & ~mask), Y+(CONST&mask),
5718 where mask is selected by alignment and size of the offset.
5719 We try to pick as large a range for the offset as possible to
5720 maximize the chance of a CSE. However, for aligned addresses
5721 we limit the range to 4k so that structures with different sized
5722 elements are likely to use the same base. We need to be careful
5723 not to split a CONST for some forms of address expression, otherwise
5724 it will generate sub-optimal code. */
5726 if (GET_CODE (x
) == PLUS
&& CONST_INT_P (XEXP (x
, 1)))
5728 rtx base
= XEXP (x
, 0);
5729 rtx offset_rtx
= XEXP (x
, 1);
5730 HOST_WIDE_INT offset
= INTVAL (offset_rtx
);
5732 if (GET_CODE (base
) == PLUS
)
5734 rtx op0
= XEXP (base
, 0);
5735 rtx op1
= XEXP (base
, 1);
5737 /* Force any scaling into a temp for CSE. */
5738 op0
= force_reg (Pmode
, op0
);
5739 op1
= force_reg (Pmode
, op1
);
5741 /* Let the pointer register be in op0. */
5742 if (REG_POINTER (op1
))
5743 std::swap (op0
, op1
);
5745 /* If the pointer is virtual or frame related, then we know that
5746 virtual register instantiation or register elimination is going
5747 to apply a second constant. We want the two constants folded
5748 together easily. Therefore, emit as (OP0 + CONST) + OP1. */
5749 if (virt_or_elim_regno_p (REGNO (op0
)))
5751 base
= expand_binop (Pmode
, add_optab
, op0
, offset_rtx
,
5752 NULL_RTX
, true, OPTAB_DIRECT
);
5753 return gen_rtx_PLUS (Pmode
, base
, op1
);
5756 /* Otherwise, in order to encourage CSE (and thence loop strength
5757 reduce) scaled addresses, emit as (OP0 + OP1) + CONST. */
5758 base
= expand_binop (Pmode
, add_optab
, op0
, op1
,
5759 NULL_RTX
, true, OPTAB_DIRECT
);
5760 x
= gen_rtx_PLUS (Pmode
, base
, offset_rtx
);
5763 /* Does it look like we'll need a 16-byte load/store-pair operation? */
5764 HOST_WIDE_INT base_offset
;
5765 if (GET_MODE_SIZE (mode
) > 16)
5766 base_offset
= (offset
+ 0x400) & ~0x7f0;
5767 /* For offsets aren't a multiple of the access size, the limit is
5769 else if (offset
& (GET_MODE_SIZE (mode
) - 1))
5771 base_offset
= (offset
+ 0x100) & ~0x1ff;
5773 /* BLKmode typically uses LDP of X-registers. */
5774 if (mode
== BLKmode
)
5775 base_offset
= (offset
+ 512) & ~0x3ff;
5777 /* Small negative offsets are supported. */
5778 else if (IN_RANGE (offset
, -256, 0))
5780 else if (mode
== TImode
|| mode
== TFmode
)
5781 base_offset
= (offset
+ 0x100) & ~0x1ff;
5782 /* Use 12-bit offset by access size. */
5784 base_offset
= offset
& (~0xfff * GET_MODE_SIZE (mode
));
5786 if (base_offset
!= 0)
5788 base
= plus_constant (Pmode
, base
, base_offset
);
5789 base
= force_operand (base
, NULL_RTX
);
5790 return plus_constant (Pmode
, base
, offset
- base_offset
);
5797 /* Return the reload icode required for a constant pool in mode. */
5798 static enum insn_code
5799 aarch64_constant_pool_reload_icode (machine_mode mode
)
5804 return CODE_FOR_aarch64_reload_movcpsfdi
;
5807 return CODE_FOR_aarch64_reload_movcpdfdi
;
5810 return CODE_FOR_aarch64_reload_movcptfdi
;
5813 return CODE_FOR_aarch64_reload_movcpv8qidi
;
5816 return CODE_FOR_aarch64_reload_movcpv16qidi
;
5819 return CODE_FOR_aarch64_reload_movcpv4hidi
;
5822 return CODE_FOR_aarch64_reload_movcpv8hidi
;
5825 return CODE_FOR_aarch64_reload_movcpv2sidi
;
5828 return CODE_FOR_aarch64_reload_movcpv4sidi
;
5831 return CODE_FOR_aarch64_reload_movcpv2didi
;
5834 return CODE_FOR_aarch64_reload_movcpv2dfdi
;
5843 aarch64_secondary_reload (bool in_p ATTRIBUTE_UNUSED
, rtx x
,
5846 secondary_reload_info
*sri
)
5849 /* If we have to disable direct literal pool loads and stores because the
5850 function is too big, then we need a scratch register. */
5851 if (MEM_P (x
) && GET_CODE (x
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (x
)
5852 && (SCALAR_FLOAT_MODE_P (GET_MODE (x
))
5853 || targetm
.vector_mode_supported_p (GET_MODE (x
)))
5854 && !aarch64_pcrelative_literal_loads
)
5856 sri
->icode
= aarch64_constant_pool_reload_icode (mode
);
5860 /* Without the TARGET_SIMD instructions we cannot move a Q register
5861 to a Q register directly. We need a scratch. */
5862 if (REG_P (x
) && (mode
== TFmode
|| mode
== TImode
) && mode
== GET_MODE (x
)
5863 && FP_REGNUM_P (REGNO (x
)) && !TARGET_SIMD
5864 && reg_class_subset_p (rclass
, FP_REGS
))
5867 sri
->icode
= CODE_FOR_aarch64_reload_movtf
;
5868 else if (mode
== TImode
)
5869 sri
->icode
= CODE_FOR_aarch64_reload_movti
;
5873 /* A TFmode or TImode memory access should be handled via an FP_REGS
5874 because AArch64 has richer addressing modes for LDR/STR instructions
5875 than LDP/STP instructions. */
5876 if (TARGET_FLOAT
&& rclass
== GENERAL_REGS
5877 && GET_MODE_SIZE (mode
) == 16 && MEM_P (x
))
5880 if (rclass
== FP_REGS
&& (mode
== TImode
|| mode
== TFmode
) && CONSTANT_P(x
))
5881 return GENERAL_REGS
;
5887 aarch64_can_eliminate (const int from
, const int to
)
5889 /* If we need a frame pointer, we must eliminate FRAME_POINTER_REGNUM into
5890 HARD_FRAME_POINTER_REGNUM and not into STACK_POINTER_REGNUM. */
5892 if (frame_pointer_needed
)
5894 if (from
== ARG_POINTER_REGNUM
&& to
== HARD_FRAME_POINTER_REGNUM
)
5896 if (from
== ARG_POINTER_REGNUM
&& to
== STACK_POINTER_REGNUM
)
5898 if (from
== FRAME_POINTER_REGNUM
&& to
== STACK_POINTER_REGNUM
5899 && !cfun
->calls_alloca
)
5901 if (from
== FRAME_POINTER_REGNUM
&& to
== HARD_FRAME_POINTER_REGNUM
)
5908 /* If we decided that we didn't need a leaf frame pointer but then used
5909 LR in the function, then we'll want a frame pointer after all, so
5910 prevent this elimination to ensure a frame pointer is used. */
5911 if (to
== STACK_POINTER_REGNUM
5912 && flag_omit_leaf_frame_pointer
5913 && df_regs_ever_live_p (LR_REGNUM
))
5921 aarch64_initial_elimination_offset (unsigned from
, unsigned to
)
5923 aarch64_layout_frame ();
5925 if (to
== HARD_FRAME_POINTER_REGNUM
)
5927 if (from
== ARG_POINTER_REGNUM
)
5928 return cfun
->machine
->frame
.hard_fp_offset
;
5930 if (from
== FRAME_POINTER_REGNUM
)
5931 return cfun
->machine
->frame
.hard_fp_offset
5932 - cfun
->machine
->frame
.locals_offset
;
5935 if (to
== STACK_POINTER_REGNUM
)
5937 if (from
== FRAME_POINTER_REGNUM
)
5938 return cfun
->machine
->frame
.frame_size
5939 - cfun
->machine
->frame
.locals_offset
;
5942 return cfun
->machine
->frame
.frame_size
;
5945 /* Implement RETURN_ADDR_RTX. We do not support moving back to a
5949 aarch64_return_addr (int count
, rtx frame ATTRIBUTE_UNUSED
)
5953 return get_hard_reg_initial_val (Pmode
, LR_REGNUM
);
5958 aarch64_asm_trampoline_template (FILE *f
)
5962 asm_fprintf (f
, "\tldr\tw%d, .+16\n", IP1_REGNUM
- R0_REGNUM
);
5963 asm_fprintf (f
, "\tldr\tw%d, .+16\n", STATIC_CHAIN_REGNUM
- R0_REGNUM
);
5967 asm_fprintf (f
, "\tldr\t%s, .+16\n", reg_names
[IP1_REGNUM
]);
5968 asm_fprintf (f
, "\tldr\t%s, .+20\n", reg_names
[STATIC_CHAIN_REGNUM
]);
5970 asm_fprintf (f
, "\tbr\t%s\n", reg_names
[IP1_REGNUM
]);
5971 assemble_aligned_integer (4, const0_rtx
);
5972 assemble_aligned_integer (POINTER_BYTES
, const0_rtx
);
5973 assemble_aligned_integer (POINTER_BYTES
, const0_rtx
);
5977 aarch64_trampoline_init (rtx m_tramp
, tree fndecl
, rtx chain_value
)
5979 rtx fnaddr
, mem
, a_tramp
;
5980 const int tramp_code_sz
= 16;
5982 /* Don't need to copy the trailing D-words, we fill those in below. */
5983 emit_block_move (m_tramp
, assemble_trampoline_template (),
5984 GEN_INT (tramp_code_sz
), BLOCK_OP_NORMAL
);
5985 mem
= adjust_address (m_tramp
, ptr_mode
, tramp_code_sz
);
5986 fnaddr
= XEXP (DECL_RTL (fndecl
), 0);
5987 if (GET_MODE (fnaddr
) != ptr_mode
)
5988 fnaddr
= convert_memory_address (ptr_mode
, fnaddr
);
5989 emit_move_insn (mem
, fnaddr
);
5991 mem
= adjust_address (m_tramp
, ptr_mode
, tramp_code_sz
+ POINTER_BYTES
);
5992 emit_move_insn (mem
, chain_value
);
5994 /* XXX We should really define a "clear_cache" pattern and use
5995 gen_clear_cache(). */
5996 a_tramp
= XEXP (m_tramp
, 0);
5997 emit_library_call (gen_rtx_SYMBOL_REF (Pmode
, "__clear_cache"),
5998 LCT_NORMAL
, VOIDmode
, a_tramp
, ptr_mode
,
5999 plus_constant (ptr_mode
, a_tramp
, TRAMPOLINE_SIZE
),
6003 static unsigned char
6004 aarch64_class_max_nregs (reg_class_t regclass
, machine_mode mode
)
6008 case CALLER_SAVE_REGS
:
6015 aarch64_vector_mode_p (mode
)
6016 ? (GET_MODE_SIZE (mode
) + UNITS_PER_VREG
- 1) / UNITS_PER_VREG
6017 : (GET_MODE_SIZE (mode
) + UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
6031 aarch64_preferred_reload_class (rtx x
, reg_class_t regclass
)
6033 if (regclass
== POINTER_REGS
)
6034 return GENERAL_REGS
;
6036 if (regclass
== STACK_REG
)
6039 && reg_class_subset_p (REGNO_REG_CLASS (REGNO (x
)), POINTER_REGS
))
6045 /* Register eliminiation can result in a request for
6046 SP+constant->FP_REGS. We cannot support such operations which
6047 use SP as source and an FP_REG as destination, so reject out
6049 if (! reg_class_subset_p (regclass
, GENERAL_REGS
) && GET_CODE (x
) == PLUS
)
6051 rtx lhs
= XEXP (x
, 0);
6053 /* Look through a possible SUBREG introduced by ILP32. */
6054 if (GET_CODE (lhs
) == SUBREG
)
6055 lhs
= SUBREG_REG (lhs
);
6057 gcc_assert (REG_P (lhs
));
6058 gcc_assert (reg_class_subset_p (REGNO_REG_CLASS (REGNO (lhs
)),
6067 aarch64_asm_output_labelref (FILE* f
, const char *name
)
6069 asm_fprintf (f
, "%U%s", name
);
6073 aarch64_elf_asm_constructor (rtx symbol
, int priority
)
6075 if (priority
== DEFAULT_INIT_PRIORITY
)
6076 default_ctor_section_asm_out_constructor (symbol
, priority
);
6080 /* While priority is known to be in range [0, 65535], so 18 bytes
6081 would be enough, the compiler might not know that. To avoid
6082 -Wformat-truncation false positive, use a larger size. */
6084 snprintf (buf
, sizeof (buf
), ".init_array.%.5u", priority
);
6085 s
= get_section (buf
, SECTION_WRITE
, NULL
);
6086 switch_to_section (s
);
6087 assemble_align (POINTER_SIZE
);
6088 assemble_aligned_integer (POINTER_BYTES
, symbol
);
6093 aarch64_elf_asm_destructor (rtx symbol
, int priority
)
6095 if (priority
== DEFAULT_INIT_PRIORITY
)
6096 default_dtor_section_asm_out_destructor (symbol
, priority
);
6100 /* While priority is known to be in range [0, 65535], so 18 bytes
6101 would be enough, the compiler might not know that. To avoid
6102 -Wformat-truncation false positive, use a larger size. */
6104 snprintf (buf
, sizeof (buf
), ".fini_array.%.5u", priority
);
6105 s
= get_section (buf
, SECTION_WRITE
, NULL
);
6106 switch_to_section (s
);
6107 assemble_align (POINTER_SIZE
);
6108 assemble_aligned_integer (POINTER_BYTES
, symbol
);
6113 aarch64_output_casesi (rtx
*operands
)
6117 rtx diff_vec
= PATTERN (NEXT_INSN (as_a
<rtx_insn
*> (operands
[2])));
6119 static const char *const patterns
[4][2] =
6122 "ldrb\t%w3, [%0,%w1,uxtw]",
6123 "add\t%3, %4, %w3, sxtb #2"
6126 "ldrh\t%w3, [%0,%w1,uxtw #1]",
6127 "add\t%3, %4, %w3, sxth #2"
6130 "ldr\t%w3, [%0,%w1,uxtw #2]",
6131 "add\t%3, %4, %w3, sxtw #2"
6133 /* We assume that DImode is only generated when not optimizing and
6134 that we don't really need 64-bit address offsets. That would
6135 imply an object file with 8GB of code in a single function! */
6137 "ldr\t%w3, [%0,%w1,uxtw #2]",
6138 "add\t%3, %4, %w3, sxtw #2"
6142 gcc_assert (GET_CODE (diff_vec
) == ADDR_DIFF_VEC
);
6144 scalar_int_mode mode
= as_a
<scalar_int_mode
> (GET_MODE (diff_vec
));
6145 index
= exact_log2 (GET_MODE_SIZE (mode
));
6147 gcc_assert (index
>= 0 && index
<= 3);
6149 /* Need to implement table size reduction, by chaning the code below. */
6150 output_asm_insn (patterns
[index
][0], operands
);
6151 ASM_GENERATE_INTERNAL_LABEL (label
, "Lrtx", CODE_LABEL_NUMBER (operands
[2]));
6152 snprintf (buf
, sizeof (buf
),
6153 "adr\t%%4, %s", targetm
.strip_name_encoding (label
));
6154 output_asm_insn (buf
, operands
);
6155 output_asm_insn (patterns
[index
][1], operands
);
6156 output_asm_insn ("br\t%3", operands
);
6157 assemble_label (asm_out_file
, label
);
6162 /* Return size in bits of an arithmetic operand which is shifted/scaled and
6163 masked such that it is suitable for a UXTB, UXTH, or UXTW extend
6167 aarch64_uxt_size (int shift
, HOST_WIDE_INT mask
)
6169 if (shift
>= 0 && shift
<= 3)
6172 for (size
= 8; size
<= 32; size
*= 2)
6174 HOST_WIDE_INT bits
= ((HOST_WIDE_INT
)1U << size
) - 1;
6175 if (mask
== bits
<< shift
)
6182 /* Constant pools are per function only when PC relative
6183 literal loads are true or we are in the large memory
6187 aarch64_can_use_per_function_literal_pools_p (void)
6189 return (aarch64_pcrelative_literal_loads
6190 || aarch64_cmodel
== AARCH64_CMODEL_LARGE
);
6194 aarch64_use_blocks_for_constant_p (machine_mode
, const_rtx
)
6196 /* Fixme:: In an ideal world this would work similar
6197 to the logic in aarch64_select_rtx_section but this
6198 breaks bootstrap in gcc go. For now we workaround
6199 this by returning false here. */
6203 /* Select appropriate section for constants depending
6204 on where we place literal pools. */
6207 aarch64_select_rtx_section (machine_mode mode
,
6209 unsigned HOST_WIDE_INT align
)
6211 if (aarch64_can_use_per_function_literal_pools_p ())
6212 return function_section (current_function_decl
);
6214 return default_elf_select_rtx_section (mode
, x
, align
);
6217 /* Implement ASM_OUTPUT_POOL_EPILOGUE. */
6219 aarch64_asm_output_pool_epilogue (FILE *f
, const char *, tree
,
6220 HOST_WIDE_INT offset
)
6222 /* When using per-function literal pools, we must ensure that any code
6223 section is aligned to the minimal instruction length, lest we get
6224 errors from the assembler re "unaligned instructions". */
6225 if ((offset
& 3) && aarch64_can_use_per_function_literal_pools_p ())
6226 ASM_OUTPUT_ALIGN (f
, 2);
6231 /* Helper function for rtx cost calculation. Strip a shift expression
6232 from X. Returns the inner operand if successful, or the original
6233 expression on failure. */
6235 aarch64_strip_shift (rtx x
)
6239 /* We accept both ROTATERT and ROTATE: since the RHS must be a constant
6240 we can convert both to ROR during final output. */
6241 if ((GET_CODE (op
) == ASHIFT
6242 || GET_CODE (op
) == ASHIFTRT
6243 || GET_CODE (op
) == LSHIFTRT
6244 || GET_CODE (op
) == ROTATERT
6245 || GET_CODE (op
) == ROTATE
)
6246 && CONST_INT_P (XEXP (op
, 1)))
6247 return XEXP (op
, 0);
6249 if (GET_CODE (op
) == MULT
6250 && CONST_INT_P (XEXP (op
, 1))
6251 && ((unsigned) exact_log2 (INTVAL (XEXP (op
, 1)))) < 64)
6252 return XEXP (op
, 0);
6257 /* Helper function for rtx cost calculation. Strip an extend
6258 expression from X. Returns the inner operand if successful, or the
6259 original expression on failure. We deal with a number of possible
6260 canonicalization variations here. If STRIP_SHIFT is true, then
6261 we can strip off a shift also. */
6263 aarch64_strip_extend (rtx x
, bool strip_shift
)
6265 scalar_int_mode mode
;
6268 if (!is_a
<scalar_int_mode
> (GET_MODE (op
), &mode
))
6271 /* Zero and sign extraction of a widened value. */
6272 if ((GET_CODE (op
) == ZERO_EXTRACT
|| GET_CODE (op
) == SIGN_EXTRACT
)
6273 && XEXP (op
, 2) == const0_rtx
6274 && GET_CODE (XEXP (op
, 0)) == MULT
6275 && aarch64_is_extend_from_extract (mode
, XEXP (XEXP (op
, 0), 1),
6277 return XEXP (XEXP (op
, 0), 0);
6279 /* It can also be represented (for zero-extend) as an AND with an
6281 if (GET_CODE (op
) == AND
6282 && GET_CODE (XEXP (op
, 0)) == MULT
6283 && CONST_INT_P (XEXP (XEXP (op
, 0), 1))
6284 && CONST_INT_P (XEXP (op
, 1))
6285 && aarch64_uxt_size (exact_log2 (INTVAL (XEXP (XEXP (op
, 0), 1))),
6286 INTVAL (XEXP (op
, 1))) != 0)
6287 return XEXP (XEXP (op
, 0), 0);
6289 /* Now handle extended register, as this may also have an optional
6290 left shift by 1..4. */
6292 && GET_CODE (op
) == ASHIFT
6293 && CONST_INT_P (XEXP (op
, 1))
6294 && ((unsigned HOST_WIDE_INT
) INTVAL (XEXP (op
, 1))) <= 4)
6297 if (GET_CODE (op
) == ZERO_EXTEND
6298 || GET_CODE (op
) == SIGN_EXTEND
)
6307 /* Return true iff CODE is a shift supported in combination
6308 with arithmetic instructions. */
6311 aarch64_shift_p (enum rtx_code code
)
6313 return code
== ASHIFT
|| code
== ASHIFTRT
|| code
== LSHIFTRT
;
6317 /* Return true iff X is a cheap shift without a sign extend. */
6320 aarch64_cheap_mult_shift_p (rtx x
)
6327 if (!(aarch64_tune_params
.extra_tuning_flags
6328 & AARCH64_EXTRA_TUNE_CHEAP_SHIFT_EXTEND
))
6331 if (GET_CODE (op0
) == SIGN_EXTEND
)
6334 if (GET_CODE (x
) == ASHIFT
&& CONST_INT_P (op1
)
6335 && UINTVAL (op1
) <= 4)
6338 if (GET_CODE (x
) != MULT
|| !CONST_INT_P (op1
))
6341 HOST_WIDE_INT l2
= exact_log2 (INTVAL (op1
));
6343 if (l2
> 0 && l2
<= 4)
6349 /* Helper function for rtx cost calculation. Calculate the cost of
6350 a MULT or ASHIFT, which may be part of a compound PLUS/MINUS rtx.
6351 Return the calculated cost of the expression, recursing manually in to
6352 operands where needed. */
6355 aarch64_rtx_mult_cost (rtx x
, enum rtx_code code
, int outer
, bool speed
)
6358 const struct cpu_cost_table
*extra_cost
6359 = aarch64_tune_params
.insn_extra_cost
;
6361 bool compound_p
= (outer
== PLUS
|| outer
== MINUS
);
6362 machine_mode mode
= GET_MODE (x
);
6364 gcc_checking_assert (code
== MULT
);
6369 if (VECTOR_MODE_P (mode
))
6370 mode
= GET_MODE_INNER (mode
);
6372 /* Integer multiply/fma. */
6373 if (GET_MODE_CLASS (mode
) == MODE_INT
)
6375 /* The multiply will be canonicalized as a shift, cost it as such. */
6376 if (aarch64_shift_p (GET_CODE (x
))
6377 || (CONST_INT_P (op1
)
6378 && exact_log2 (INTVAL (op1
)) > 0))
6380 bool is_extend
= GET_CODE (op0
) == ZERO_EXTEND
6381 || GET_CODE (op0
) == SIGN_EXTEND
;
6386 /* If the shift is considered cheap,
6387 then don't add any cost. */
6388 if (aarch64_cheap_mult_shift_p (x
))
6390 else if (REG_P (op1
))
6391 /* ARITH + shift-by-register. */
6392 cost
+= extra_cost
->alu
.arith_shift_reg
;
6394 /* ARITH + extended register. We don't have a cost field
6395 for ARITH+EXTEND+SHIFT, so use extend_arith here. */
6396 cost
+= extra_cost
->alu
.extend_arith
;
6398 /* ARITH + shift-by-immediate. */
6399 cost
+= extra_cost
->alu
.arith_shift
;
6402 /* LSL (immediate). */
6403 cost
+= extra_cost
->alu
.shift
;
6406 /* Strip extends as we will have costed them in the case above. */
6408 op0
= aarch64_strip_extend (op0
, true);
6410 cost
+= rtx_cost (op0
, VOIDmode
, code
, 0, speed
);
6415 /* MNEG or [US]MNEGL. Extract the NEG operand and indicate that it's a
6416 compound and let the below cases handle it. After all, MNEG is a
6417 special-case alias of MSUB. */
6418 if (GET_CODE (op0
) == NEG
)
6420 op0
= XEXP (op0
, 0);
6424 /* Integer multiplies or FMAs have zero/sign extending variants. */
6425 if ((GET_CODE (op0
) == ZERO_EXTEND
6426 && GET_CODE (op1
) == ZERO_EXTEND
)
6427 || (GET_CODE (op0
) == SIGN_EXTEND
6428 && GET_CODE (op1
) == SIGN_EXTEND
))
6430 cost
+= rtx_cost (XEXP (op0
, 0), VOIDmode
, MULT
, 0, speed
);
6431 cost
+= rtx_cost (XEXP (op1
, 0), VOIDmode
, MULT
, 1, speed
);
6436 /* SMADDL/UMADDL/UMSUBL/SMSUBL. */
6437 cost
+= extra_cost
->mult
[0].extend_add
;
6439 /* MUL/SMULL/UMULL. */
6440 cost
+= extra_cost
->mult
[0].extend
;
6446 /* This is either an integer multiply or a MADD. In both cases
6447 we want to recurse and cost the operands. */
6448 cost
+= rtx_cost (op0
, mode
, MULT
, 0, speed
);
6449 cost
+= rtx_cost (op1
, mode
, MULT
, 1, speed
);
6455 cost
+= extra_cost
->mult
[mode
== DImode
].add
;
6458 cost
+= extra_cost
->mult
[mode
== DImode
].simple
;
6467 /* Floating-point FMA/FMUL can also support negations of the
6468 operands, unless the rounding mode is upward or downward in
6469 which case FNMUL is different than FMUL with operand negation. */
6470 bool neg0
= GET_CODE (op0
) == NEG
;
6471 bool neg1
= GET_CODE (op1
) == NEG
;
6472 if (compound_p
|| !flag_rounding_math
|| (neg0
&& neg1
))
6475 op0
= XEXP (op0
, 0);
6477 op1
= XEXP (op1
, 0);
6481 /* FMADD/FNMADD/FNMSUB/FMSUB. */
6482 cost
+= extra_cost
->fp
[mode
== DFmode
].fma
;
6485 cost
+= extra_cost
->fp
[mode
== DFmode
].mult
;
6488 cost
+= rtx_cost (op0
, mode
, MULT
, 0, speed
);
6489 cost
+= rtx_cost (op1
, mode
, MULT
, 1, speed
);
6495 aarch64_address_cost (rtx x
,
6497 addr_space_t as ATTRIBUTE_UNUSED
,
6500 enum rtx_code c
= GET_CODE (x
);
6501 const struct cpu_addrcost_table
*addr_cost
= aarch64_tune_params
.addr_cost
;
6502 struct aarch64_address_info info
;
6506 if (!aarch64_classify_address (&info
, x
, mode
, c
, false))
6508 if (GET_CODE (x
) == CONST
|| GET_CODE (x
) == SYMBOL_REF
)
6510 /* This is a CONST or SYMBOL ref which will be split
6511 in a different way depending on the code model in use.
6512 Cost it through the generic infrastructure. */
6513 int cost_symbol_ref
= rtx_cost (x
, Pmode
, MEM
, 1, speed
);
6514 /* Divide through by the cost of one instruction to
6515 bring it to the same units as the address costs. */
6516 cost_symbol_ref
/= COSTS_N_INSNS (1);
6517 /* The cost is then the cost of preparing the address,
6518 followed by an immediate (possibly 0) offset. */
6519 return cost_symbol_ref
+ addr_cost
->imm_offset
;
6523 /* This is most likely a jump table from a case
6525 return addr_cost
->register_offset
;
6531 case ADDRESS_LO_SUM
:
6532 case ADDRESS_SYMBOLIC
:
6533 case ADDRESS_REG_IMM
:
6534 cost
+= addr_cost
->imm_offset
;
6537 case ADDRESS_REG_WB
:
6538 if (c
== PRE_INC
|| c
== PRE_DEC
|| c
== PRE_MODIFY
)
6539 cost
+= addr_cost
->pre_modify
;
6540 else if (c
== POST_INC
|| c
== POST_DEC
|| c
== POST_MODIFY
)
6541 cost
+= addr_cost
->post_modify
;
6547 case ADDRESS_REG_REG
:
6548 cost
+= addr_cost
->register_offset
;
6551 case ADDRESS_REG_SXTW
:
6552 cost
+= addr_cost
->register_sextend
;
6555 case ADDRESS_REG_UXTW
:
6556 cost
+= addr_cost
->register_zextend
;
6566 /* For the sake of calculating the cost of the shifted register
6567 component, we can treat same sized modes in the same way. */
6568 switch (GET_MODE_BITSIZE (mode
))
6571 cost
+= addr_cost
->addr_scale_costs
.hi
;
6575 cost
+= addr_cost
->addr_scale_costs
.si
;
6579 cost
+= addr_cost
->addr_scale_costs
.di
;
6582 /* We can't tell, or this is a 128-bit vector. */
6584 cost
+= addr_cost
->addr_scale_costs
.ti
;
6592 /* Return the cost of a branch. If SPEED_P is true then the compiler is
6593 optimizing for speed. If PREDICTABLE_P is true then the branch is predicted
6597 aarch64_branch_cost (bool speed_p
, bool predictable_p
)
6599 /* When optimizing for speed, use the cost of unpredictable branches. */
6600 const struct cpu_branch_cost
*branch_costs
=
6601 aarch64_tune_params
.branch_costs
;
6603 if (!speed_p
|| predictable_p
)
6604 return branch_costs
->predictable
;
6606 return branch_costs
->unpredictable
;
6609 /* Return true if the RTX X in mode MODE is a zero or sign extract
6610 usable in an ADD or SUB (extended register) instruction. */
6612 aarch64_rtx_arith_op_extract_p (rtx x
, scalar_int_mode mode
)
6614 /* Catch add with a sign extract.
6615 This is add_<optab><mode>_multp2. */
6616 if (GET_CODE (x
) == SIGN_EXTRACT
6617 || GET_CODE (x
) == ZERO_EXTRACT
)
6619 rtx op0
= XEXP (x
, 0);
6620 rtx op1
= XEXP (x
, 1);
6621 rtx op2
= XEXP (x
, 2);
6623 if (GET_CODE (op0
) == MULT
6624 && CONST_INT_P (op1
)
6625 && op2
== const0_rtx
6626 && CONST_INT_P (XEXP (op0
, 1))
6627 && aarch64_is_extend_from_extract (mode
,
6634 /* The simple case <ARITH>, XD, XN, XM, [us]xt.
6636 else if (GET_CODE (x
) == SIGN_EXTEND
6637 || GET_CODE (x
) == ZERO_EXTEND
)
6638 return REG_P (XEXP (x
, 0));
6644 aarch64_frint_unspec_p (unsigned int u
)
6662 /* Return true iff X is an rtx that will match an extr instruction
6663 i.e. as described in the *extr<mode>5_insn family of patterns.
6664 OP0 and OP1 will be set to the operands of the shifts involved
6665 on success and will be NULL_RTX otherwise. */
6668 aarch64_extr_rtx_p (rtx x
, rtx
*res_op0
, rtx
*res_op1
)
6671 scalar_int_mode mode
;
6672 if (!is_a
<scalar_int_mode
> (GET_MODE (x
), &mode
))
6675 *res_op0
= NULL_RTX
;
6676 *res_op1
= NULL_RTX
;
6678 if (GET_CODE (x
) != IOR
)
6684 if ((GET_CODE (op0
) == ASHIFT
&& GET_CODE (op1
) == LSHIFTRT
)
6685 || (GET_CODE (op1
) == ASHIFT
&& GET_CODE (op0
) == LSHIFTRT
))
6687 /* Canonicalise locally to ashift in op0, lshiftrt in op1. */
6688 if (GET_CODE (op1
) == ASHIFT
)
6689 std::swap (op0
, op1
);
6691 if (!CONST_INT_P (XEXP (op0
, 1)) || !CONST_INT_P (XEXP (op1
, 1)))
6694 unsigned HOST_WIDE_INT shft_amnt_0
= UINTVAL (XEXP (op0
, 1));
6695 unsigned HOST_WIDE_INT shft_amnt_1
= UINTVAL (XEXP (op1
, 1));
6697 if (shft_amnt_0
< GET_MODE_BITSIZE (mode
)
6698 && shft_amnt_0
+ shft_amnt_1
== GET_MODE_BITSIZE (mode
))
6700 *res_op0
= XEXP (op0
, 0);
6701 *res_op1
= XEXP (op1
, 0);
6709 /* Calculate the cost of calculating (if_then_else (OP0) (OP1) (OP2)),
6710 storing it in *COST. Result is true if the total cost of the operation
6711 has now been calculated. */
6713 aarch64_if_then_else_costs (rtx op0
, rtx op1
, rtx op2
, int *cost
, bool speed
)
6717 enum rtx_code cmpcode
;
6719 if (COMPARISON_P (op0
))
6721 inner
= XEXP (op0
, 0);
6722 comparator
= XEXP (op0
, 1);
6723 cmpcode
= GET_CODE (op0
);
6728 comparator
= const0_rtx
;
6732 if (GET_CODE (op1
) == PC
|| GET_CODE (op2
) == PC
)
6734 /* Conditional branch. */
6735 if (GET_MODE_CLASS (GET_MODE (inner
)) == MODE_CC
)
6739 if (cmpcode
== NE
|| cmpcode
== EQ
)
6741 if (comparator
== const0_rtx
)
6743 /* TBZ/TBNZ/CBZ/CBNZ. */
6744 if (GET_CODE (inner
) == ZERO_EXTRACT
)
6746 *cost
+= rtx_cost (XEXP (inner
, 0), VOIDmode
,
6747 ZERO_EXTRACT
, 0, speed
);
6750 *cost
+= rtx_cost (inner
, VOIDmode
, cmpcode
, 0, speed
);
6755 else if (cmpcode
== LT
|| cmpcode
== GE
)
6758 if (comparator
== const0_rtx
)
6763 else if (GET_MODE_CLASS (GET_MODE (inner
)) == MODE_CC
)
6766 if (GET_CODE (op1
) == COMPARE
)
6768 /* Increase cost of CCMP reg, 0, imm, CC to prefer CMP reg, 0. */
6769 if (XEXP (op1
, 1) == const0_rtx
)
6773 machine_mode mode
= GET_MODE (XEXP (op1
, 0));
6774 const struct cpu_cost_table
*extra_cost
6775 = aarch64_tune_params
.insn_extra_cost
;
6777 if (GET_MODE_CLASS (mode
) == MODE_INT
)
6778 *cost
+= extra_cost
->alu
.arith
;
6780 *cost
+= extra_cost
->fp
[mode
== DFmode
].compare
;
6785 /* It's a conditional operation based on the status flags,
6786 so it must be some flavor of CSEL. */
6788 /* CSNEG, CSINV, and CSINC are handled for free as part of CSEL. */
6789 if (GET_CODE (op1
) == NEG
6790 || GET_CODE (op1
) == NOT
6791 || (GET_CODE (op1
) == PLUS
&& XEXP (op1
, 1) == const1_rtx
))
6792 op1
= XEXP (op1
, 0);
6793 else if (GET_CODE (op1
) == ZERO_EXTEND
&& GET_CODE (op2
) == ZERO_EXTEND
)
6795 /* CSEL with zero-extension (*cmovdi_insn_uxtw). */
6796 op1
= XEXP (op1
, 0);
6797 op2
= XEXP (op2
, 0);
6800 *cost
+= rtx_cost (op1
, VOIDmode
, IF_THEN_ELSE
, 1, speed
);
6801 *cost
+= rtx_cost (op2
, VOIDmode
, IF_THEN_ELSE
, 2, speed
);
6805 /* We don't know what this is, cost all operands. */
6809 /* Check whether X is a bitfield operation of the form shift + extend that
6810 maps down to a UBFIZ/SBFIZ/UBFX/SBFX instruction. If so, return the
6811 operand to which the bitfield operation is applied. Otherwise return
6815 aarch64_extend_bitfield_pattern_p (rtx x
)
6817 rtx_code outer_code
= GET_CODE (x
);
6818 machine_mode outer_mode
= GET_MODE (x
);
6820 if (outer_code
!= ZERO_EXTEND
&& outer_code
!= SIGN_EXTEND
6821 && outer_mode
!= SImode
&& outer_mode
!= DImode
)
6824 rtx inner
= XEXP (x
, 0);
6825 rtx_code inner_code
= GET_CODE (inner
);
6826 machine_mode inner_mode
= GET_MODE (inner
);
6832 if (CONST_INT_P (XEXP (inner
, 1))
6833 && (inner_mode
== QImode
|| inner_mode
== HImode
))
6834 op
= XEXP (inner
, 0);
6837 if (outer_code
== ZERO_EXTEND
&& CONST_INT_P (XEXP (inner
, 1))
6838 && (inner_mode
== QImode
|| inner_mode
== HImode
))
6839 op
= XEXP (inner
, 0);
6842 if (outer_code
== SIGN_EXTEND
&& CONST_INT_P (XEXP (inner
, 1))
6843 && (inner_mode
== QImode
|| inner_mode
== HImode
))
6844 op
= XEXP (inner
, 0);
6853 /* Return true if the mask and a shift amount from an RTX of the form
6854 (x << SHFT_AMNT) & MASK are valid to combine into a UBFIZ instruction of
6855 mode MODE. See the *andim_ashift<mode>_bfiz pattern. */
6858 aarch64_mask_and_shift_for_ubfiz_p (scalar_int_mode mode
, rtx mask
,
6861 return CONST_INT_P (mask
) && CONST_INT_P (shft_amnt
)
6862 && INTVAL (shft_amnt
) < GET_MODE_BITSIZE (mode
)
6863 && exact_log2 ((INTVAL (mask
) >> INTVAL (shft_amnt
)) + 1) >= 0
6864 && (INTVAL (mask
) & ((1 << INTVAL (shft_amnt
)) - 1)) == 0;
6867 /* Calculate the cost of calculating X, storing it in *COST. Result
6868 is true if the total cost of the operation has now been calculated. */
6870 aarch64_rtx_costs (rtx x
, machine_mode mode
, int outer ATTRIBUTE_UNUSED
,
6871 int param ATTRIBUTE_UNUSED
, int *cost
, bool speed
)
6874 const struct cpu_cost_table
*extra_cost
6875 = aarch64_tune_params
.insn_extra_cost
;
6876 int code
= GET_CODE (x
);
6877 scalar_int_mode int_mode
;
6879 /* By default, assume that everything has equivalent cost to the
6880 cheapest instruction. Any additional costs are applied as a delta
6881 above this default. */
6882 *cost
= COSTS_N_INSNS (1);
6887 /* The cost depends entirely on the operands to SET. */
6892 switch (GET_CODE (op0
))
6897 rtx address
= XEXP (op0
, 0);
6898 if (VECTOR_MODE_P (mode
))
6899 *cost
+= extra_cost
->ldst
.storev
;
6900 else if (GET_MODE_CLASS (mode
) == MODE_INT
)
6901 *cost
+= extra_cost
->ldst
.store
;
6902 else if (mode
== SFmode
)
6903 *cost
+= extra_cost
->ldst
.storef
;
6904 else if (mode
== DFmode
)
6905 *cost
+= extra_cost
->ldst
.stored
;
6908 COSTS_N_INSNS (aarch64_address_cost (address
, mode
,
6912 *cost
+= rtx_cost (op1
, mode
, SET
, 1, speed
);
6916 if (! REG_P (SUBREG_REG (op0
)))
6917 *cost
+= rtx_cost (SUBREG_REG (op0
), VOIDmode
, SET
, 0, speed
);
6921 /* The cost is one per vector-register copied. */
6922 if (VECTOR_MODE_P (GET_MODE (op0
)) && REG_P (op1
))
6924 int n_minus_1
= (GET_MODE_SIZE (GET_MODE (op0
)) - 1)
6925 / GET_MODE_SIZE (V4SImode
);
6926 *cost
= COSTS_N_INSNS (n_minus_1
+ 1);
6928 /* const0_rtx is in general free, but we will use an
6929 instruction to set a register to 0. */
6930 else if (REG_P (op1
) || op1
== const0_rtx
)
6932 /* The cost is 1 per register copied. */
6933 int n_minus_1
= (GET_MODE_SIZE (GET_MODE (op0
)) - 1)
6935 *cost
= COSTS_N_INSNS (n_minus_1
+ 1);
6938 /* Cost is just the cost of the RHS of the set. */
6939 *cost
+= rtx_cost (op1
, mode
, SET
, 1, speed
);
6944 /* Bit-field insertion. Strip any redundant widening of
6945 the RHS to meet the width of the target. */
6946 if (GET_CODE (op1
) == SUBREG
)
6947 op1
= SUBREG_REG (op1
);
6948 if ((GET_CODE (op1
) == ZERO_EXTEND
6949 || GET_CODE (op1
) == SIGN_EXTEND
)
6950 && CONST_INT_P (XEXP (op0
, 1))
6951 && is_a
<scalar_int_mode
> (GET_MODE (XEXP (op1
, 0)), &int_mode
)
6952 && GET_MODE_BITSIZE (int_mode
) >= INTVAL (XEXP (op0
, 1)))
6953 op1
= XEXP (op1
, 0);
6955 if (CONST_INT_P (op1
))
6957 /* MOV immediate is assumed to always be cheap. */
6958 *cost
= COSTS_N_INSNS (1);
6964 *cost
+= extra_cost
->alu
.bfi
;
6965 *cost
+= rtx_cost (op1
, VOIDmode
, (enum rtx_code
) code
, 1, speed
);
6971 /* We can't make sense of this, assume default cost. */
6972 *cost
= COSTS_N_INSNS (1);
6978 /* If an instruction can incorporate a constant within the
6979 instruction, the instruction's expression avoids calling
6980 rtx_cost() on the constant. If rtx_cost() is called on a
6981 constant, then it is usually because the constant must be
6982 moved into a register by one or more instructions.
6984 The exception is constant 0, which can be expressed
6985 as XZR/WZR and is therefore free. The exception to this is
6986 if we have (set (reg) (const0_rtx)) in which case we must cost
6987 the move. However, we can catch that when we cost the SET, so
6988 we don't need to consider that here. */
6989 if (x
== const0_rtx
)
6993 /* To an approximation, building any other constant is
6994 proportionally expensive to the number of instructions
6995 required to build that constant. This is true whether we
6996 are compiling for SPEED or otherwise. */
6997 if (!is_a
<scalar_int_mode
> (mode
, &int_mode
))
6998 int_mode
= word_mode
;
6999 *cost
= COSTS_N_INSNS (aarch64_internal_mov_immediate
7000 (NULL_RTX
, x
, false, int_mode
));
7006 /* First determine number of instructions to do the move
7007 as an integer constant. */
7008 if (!aarch64_float_const_representable_p (x
)
7009 && !aarch64_can_const_movi_rtx_p (x
, mode
)
7010 && aarch64_float_const_rtx_p (x
))
7012 unsigned HOST_WIDE_INT ival
;
7013 bool succeed
= aarch64_reinterpret_float_as_int (x
, &ival
);
7014 gcc_assert (succeed
);
7016 scalar_int_mode imode
= (mode
== HFmode
7018 : int_mode_for_mode (mode
).require ());
7019 int ncost
= aarch64_internal_mov_immediate
7020 (NULL_RTX
, gen_int_mode (ival
, imode
), false, imode
);
7021 *cost
+= COSTS_N_INSNS (ncost
);
7027 /* mov[df,sf]_aarch64. */
7028 if (aarch64_float_const_representable_p (x
))
7029 /* FMOV (scalar immediate). */
7030 *cost
+= extra_cost
->fp
[mode
== DFmode
].fpconst
;
7031 else if (!aarch64_float_const_zero_rtx_p (x
))
7033 /* This will be a load from memory. */
7035 *cost
+= extra_cost
->ldst
.loadd
;
7037 *cost
+= extra_cost
->ldst
.loadf
;
7040 /* Otherwise this is +0.0. We get this using MOVI d0, #0
7041 or MOV v0.s[0], wzr - neither of which are modeled by the
7042 cost tables. Just use the default cost. */
7052 /* For loads we want the base cost of a load, plus an
7053 approximation for the additional cost of the addressing
7055 rtx address
= XEXP (x
, 0);
7056 if (VECTOR_MODE_P (mode
))
7057 *cost
+= extra_cost
->ldst
.loadv
;
7058 else if (GET_MODE_CLASS (mode
) == MODE_INT
)
7059 *cost
+= extra_cost
->ldst
.load
;
7060 else if (mode
== SFmode
)
7061 *cost
+= extra_cost
->ldst
.loadf
;
7062 else if (mode
== DFmode
)
7063 *cost
+= extra_cost
->ldst
.loadd
;
7066 COSTS_N_INSNS (aarch64_address_cost (address
, mode
,
7075 if (VECTOR_MODE_P (mode
))
7080 *cost
+= extra_cost
->vect
.alu
;
7085 if (GET_MODE_CLASS (mode
) == MODE_INT
)
7087 if (GET_RTX_CLASS (GET_CODE (op0
)) == RTX_COMPARE
7088 || GET_RTX_CLASS (GET_CODE (op0
)) == RTX_COMM_COMPARE
)
7091 *cost
+= rtx_cost (XEXP (op0
, 0), VOIDmode
, NEG
, 0, speed
);
7095 /* Cost this as SUB wzr, X. */
7096 op0
= CONST0_RTX (mode
);
7101 if (GET_MODE_CLASS (mode
) == MODE_FLOAT
)
7103 /* Support (neg(fma...)) as a single instruction only if
7104 sign of zeros is unimportant. This matches the decision
7105 making in aarch64.md. */
7106 if (GET_CODE (op0
) == FMA
&& !HONOR_SIGNED_ZEROS (GET_MODE (op0
)))
7109 *cost
= rtx_cost (op0
, mode
, NEG
, 0, speed
);
7112 if (GET_CODE (op0
) == MULT
)
7115 *cost
= rtx_cost (op0
, mode
, NEG
, 0, speed
);
7120 *cost
+= extra_cost
->fp
[mode
== DFmode
].neg
;
7130 if (VECTOR_MODE_P (mode
))
7131 *cost
+= extra_cost
->vect
.alu
;
7133 *cost
+= extra_cost
->alu
.clz
;
7142 if (op1
== const0_rtx
7143 && GET_CODE (op0
) == AND
)
7146 mode
= GET_MODE (op0
);
7150 if (GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
7152 /* TODO: A write to the CC flags possibly costs extra, this
7153 needs encoding in the cost tables. */
7155 mode
= GET_MODE (op0
);
7157 if (GET_CODE (op0
) == AND
)
7163 if (GET_CODE (op0
) == PLUS
)
7165 /* ADDS (and CMN alias). */
7170 if (GET_CODE (op0
) == MINUS
)
7177 if (GET_CODE (op0
) == ZERO_EXTRACT
&& op1
== const0_rtx
7178 && GET_MODE (x
) == CC_NZmode
&& CONST_INT_P (XEXP (op0
, 1))
7179 && CONST_INT_P (XEXP (op0
, 2)))
7181 /* COMPARE of ZERO_EXTRACT form of TST-immediate.
7182 Handle it here directly rather than going to cost_logic
7183 since we know the immediate generated for the TST is valid
7184 so we can avoid creating an intermediate rtx for it only
7185 for costing purposes. */
7187 *cost
+= extra_cost
->alu
.logical
;
7189 *cost
+= rtx_cost (XEXP (op0
, 0), GET_MODE (op0
),
7190 ZERO_EXTRACT
, 0, speed
);
7194 if (GET_CODE (op1
) == NEG
)
7198 *cost
+= extra_cost
->alu
.arith
;
7200 *cost
+= rtx_cost (op0
, mode
, COMPARE
, 0, speed
);
7201 *cost
+= rtx_cost (XEXP (op1
, 0), mode
, NEG
, 1, speed
);
7207 Compare can freely swap the order of operands, and
7208 canonicalization puts the more complex operation first.
7209 But the integer MINUS logic expects the shift/extend
7210 operation in op1. */
7212 || (GET_CODE (op0
) == SUBREG
&& REG_P (SUBREG_REG (op0
)))))
7220 if (GET_MODE_CLASS (GET_MODE (op0
)) == MODE_FLOAT
)
7224 *cost
+= extra_cost
->fp
[mode
== DFmode
].compare
;
7226 if (CONST_DOUBLE_P (op1
) && aarch64_float_const_zero_rtx_p (op1
))
7228 *cost
+= rtx_cost (op0
, VOIDmode
, COMPARE
, 0, speed
);
7229 /* FCMP supports constant 0.0 for no extra cost. */
7235 if (VECTOR_MODE_P (mode
))
7237 /* Vector compare. */
7239 *cost
+= extra_cost
->vect
.alu
;
7241 if (aarch64_float_const_zero_rtx_p (op1
))
7243 /* Vector cm (eq|ge|gt|lt|le) supports constant 0.0 for no extra
7257 *cost
+= rtx_cost (op0
, mode
, MINUS
, 0, speed
);
7259 /* Detect valid immediates. */
7260 if ((GET_MODE_CLASS (mode
) == MODE_INT
7261 || (GET_MODE_CLASS (mode
) == MODE_CC
7262 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
))
7263 && CONST_INT_P (op1
)
7264 && aarch64_uimm12_shift (INTVAL (op1
)))
7267 /* SUB(S) (immediate). */
7268 *cost
+= extra_cost
->alu
.arith
;
7272 /* Look for SUB (extended register). */
7273 if (is_a
<scalar_int_mode
> (mode
, &int_mode
)
7274 && aarch64_rtx_arith_op_extract_p (op1
, int_mode
))
7277 *cost
+= extra_cost
->alu
.extend_arith
;
7279 op1
= aarch64_strip_extend (op1
, true);
7280 *cost
+= rtx_cost (op1
, VOIDmode
,
7281 (enum rtx_code
) GET_CODE (op1
), 0, speed
);
7285 rtx new_op1
= aarch64_strip_extend (op1
, false);
7287 /* Cost this as an FMA-alike operation. */
7288 if ((GET_CODE (new_op1
) == MULT
7289 || aarch64_shift_p (GET_CODE (new_op1
)))
7292 *cost
+= aarch64_rtx_mult_cost (new_op1
, MULT
,
7293 (enum rtx_code
) code
,
7298 *cost
+= rtx_cost (new_op1
, VOIDmode
, MINUS
, 1, speed
);
7302 if (VECTOR_MODE_P (mode
))
7305 *cost
+= extra_cost
->vect
.alu
;
7307 else if (GET_MODE_CLASS (mode
) == MODE_INT
)
7310 *cost
+= extra_cost
->alu
.arith
;
7312 else if (GET_MODE_CLASS (mode
) == MODE_FLOAT
)
7315 *cost
+= extra_cost
->fp
[mode
== DFmode
].addsub
;
7329 if (GET_RTX_CLASS (GET_CODE (op0
)) == RTX_COMPARE
7330 || GET_RTX_CLASS (GET_CODE (op0
)) == RTX_COMM_COMPARE
)
7333 *cost
+= rtx_cost (XEXP (op0
, 0), mode
, PLUS
, 0, speed
);
7334 *cost
+= rtx_cost (op1
, mode
, PLUS
, 1, speed
);
7338 if (GET_MODE_CLASS (mode
) == MODE_INT
7339 && CONST_INT_P (op1
)
7340 && aarch64_uimm12_shift (INTVAL (op1
)))
7342 *cost
+= rtx_cost (op0
, mode
, PLUS
, 0, speed
);
7345 /* ADD (immediate). */
7346 *cost
+= extra_cost
->alu
.arith
;
7350 *cost
+= rtx_cost (op1
, mode
, PLUS
, 1, speed
);
7352 /* Look for ADD (extended register). */
7353 if (is_a
<scalar_int_mode
> (mode
, &int_mode
)
7354 && aarch64_rtx_arith_op_extract_p (op0
, int_mode
))
7357 *cost
+= extra_cost
->alu
.extend_arith
;
7359 op0
= aarch64_strip_extend (op0
, true);
7360 *cost
+= rtx_cost (op0
, VOIDmode
,
7361 (enum rtx_code
) GET_CODE (op0
), 0, speed
);
7365 /* Strip any extend, leave shifts behind as we will
7366 cost them through mult_cost. */
7367 new_op0
= aarch64_strip_extend (op0
, false);
7369 if (GET_CODE (new_op0
) == MULT
7370 || aarch64_shift_p (GET_CODE (new_op0
)))
7372 *cost
+= aarch64_rtx_mult_cost (new_op0
, MULT
, PLUS
,
7377 *cost
+= rtx_cost (new_op0
, VOIDmode
, PLUS
, 0, speed
);
7381 if (VECTOR_MODE_P (mode
))
7384 *cost
+= extra_cost
->vect
.alu
;
7386 else if (GET_MODE_CLASS (mode
) == MODE_INT
)
7389 *cost
+= extra_cost
->alu
.arith
;
7391 else if (GET_MODE_CLASS (mode
) == MODE_FLOAT
)
7394 *cost
+= extra_cost
->fp
[mode
== DFmode
].addsub
;
7401 *cost
= COSTS_N_INSNS (1);
7405 if (VECTOR_MODE_P (mode
))
7406 *cost
+= extra_cost
->vect
.alu
;
7408 *cost
+= extra_cost
->alu
.rev
;
7413 if (aarch_rev16_p (x
))
7415 *cost
= COSTS_N_INSNS (1);
7419 if (VECTOR_MODE_P (mode
))
7420 *cost
+= extra_cost
->vect
.alu
;
7422 *cost
+= extra_cost
->alu
.rev
;
7427 if (aarch64_extr_rtx_p (x
, &op0
, &op1
))
7429 *cost
+= rtx_cost (op0
, mode
, IOR
, 0, speed
);
7430 *cost
+= rtx_cost (op1
, mode
, IOR
, 1, speed
);
7432 *cost
+= extra_cost
->alu
.shift
;
7443 if (VECTOR_MODE_P (mode
))
7446 *cost
+= extra_cost
->vect
.alu
;
7451 && GET_CODE (op0
) == MULT
7452 && CONST_INT_P (XEXP (op0
, 1))
7453 && CONST_INT_P (op1
)
7454 && aarch64_uxt_size (exact_log2 (INTVAL (XEXP (op0
, 1))),
7457 /* This is a UBFM/SBFM. */
7458 *cost
+= rtx_cost (XEXP (op0
, 0), mode
, ZERO_EXTRACT
, 0, speed
);
7460 *cost
+= extra_cost
->alu
.bfx
;
7464 if (is_int_mode (mode
, &int_mode
))
7466 if (CONST_INT_P (op1
))
7468 /* We have a mask + shift version of a UBFIZ
7469 i.e. the *andim_ashift<mode>_bfiz pattern. */
7470 if (GET_CODE (op0
) == ASHIFT
7471 && aarch64_mask_and_shift_for_ubfiz_p (int_mode
, op1
,
7474 *cost
+= rtx_cost (XEXP (op0
, 0), int_mode
,
7475 (enum rtx_code
) code
, 0, speed
);
7477 *cost
+= extra_cost
->alu
.bfx
;
7481 else if (aarch64_bitmask_imm (INTVAL (op1
), int_mode
))
7483 /* We possibly get the immediate for free, this is not
7485 *cost
+= rtx_cost (op0
, int_mode
,
7486 (enum rtx_code
) code
, 0, speed
);
7488 *cost
+= extra_cost
->alu
.logical
;
7497 /* Handle ORN, EON, or BIC. */
7498 if (GET_CODE (op0
) == NOT
)
7499 op0
= XEXP (op0
, 0);
7501 new_op0
= aarch64_strip_shift (op0
);
7503 /* If we had a shift on op0 then this is a logical-shift-
7504 by-register/immediate operation. Otherwise, this is just
7505 a logical operation. */
7510 /* Shift by immediate. */
7511 if (CONST_INT_P (XEXP (op0
, 1)))
7512 *cost
+= extra_cost
->alu
.log_shift
;
7514 *cost
+= extra_cost
->alu
.log_shift_reg
;
7517 *cost
+= extra_cost
->alu
.logical
;
7520 /* In both cases we want to cost both operands. */
7521 *cost
+= rtx_cost (new_op0
, int_mode
, (enum rtx_code
) code
,
7523 *cost
+= rtx_cost (op1
, int_mode
, (enum rtx_code
) code
,
7533 op0
= aarch64_strip_shift (x
);
7535 if (VECTOR_MODE_P (mode
))
7538 *cost
+= extra_cost
->vect
.alu
;
7542 /* MVN-shifted-reg. */
7545 *cost
+= rtx_cost (op0
, mode
, (enum rtx_code
) code
, 0, speed
);
7548 *cost
+= extra_cost
->alu
.log_shift
;
7552 /* EON can have two forms: (xor (not a) b) but also (not (xor a b)).
7553 Handle the second form here taking care that 'a' in the above can
7555 else if (GET_CODE (op0
) == XOR
)
7557 rtx newop0
= XEXP (op0
, 0);
7558 rtx newop1
= XEXP (op0
, 1);
7559 rtx op0_stripped
= aarch64_strip_shift (newop0
);
7561 *cost
+= rtx_cost (newop1
, mode
, (enum rtx_code
) code
, 1, speed
);
7562 *cost
+= rtx_cost (op0_stripped
, mode
, XOR
, 0, speed
);
7566 if (op0_stripped
!= newop0
)
7567 *cost
+= extra_cost
->alu
.log_shift
;
7569 *cost
+= extra_cost
->alu
.logical
;
7576 *cost
+= extra_cost
->alu
.logical
;
7583 /* If a value is written in SI mode, then zero extended to DI
7584 mode, the operation will in general be free as a write to
7585 a 'w' register implicitly zeroes the upper bits of an 'x'
7586 register. However, if this is
7588 (set (reg) (zero_extend (reg)))
7590 we must cost the explicit register move. */
7592 && GET_MODE (op0
) == SImode
7595 int op_cost
= rtx_cost (op0
, VOIDmode
, ZERO_EXTEND
, 0, speed
);
7597 /* If OP_COST is non-zero, then the cost of the zero extend
7598 is effectively the cost of the inner operation. Otherwise
7599 we have a MOV instruction and we take the cost from the MOV
7600 itself. This is true independently of whether we are
7601 optimizing for space or time. */
7607 else if (MEM_P (op0
))
7609 /* All loads can zero extend to any size for free. */
7610 *cost
= rtx_cost (op0
, VOIDmode
, ZERO_EXTEND
, param
, speed
);
7614 op0
= aarch64_extend_bitfield_pattern_p (x
);
7617 *cost
+= rtx_cost (op0
, mode
, ZERO_EXTEND
, 0, speed
);
7619 *cost
+= extra_cost
->alu
.bfx
;
7625 if (VECTOR_MODE_P (mode
))
7628 *cost
+= extra_cost
->vect
.alu
;
7632 /* We generate an AND instead of UXTB/UXTH. */
7633 *cost
+= extra_cost
->alu
.logical
;
7639 if (MEM_P (XEXP (x
, 0)))
7644 rtx address
= XEXP (XEXP (x
, 0), 0);
7645 *cost
+= extra_cost
->ldst
.load_sign_extend
;
7648 COSTS_N_INSNS (aarch64_address_cost (address
, mode
,
7654 op0
= aarch64_extend_bitfield_pattern_p (x
);
7657 *cost
+= rtx_cost (op0
, mode
, SIGN_EXTEND
, 0, speed
);
7659 *cost
+= extra_cost
->alu
.bfx
;
7665 if (VECTOR_MODE_P (mode
))
7666 *cost
+= extra_cost
->vect
.alu
;
7668 *cost
+= extra_cost
->alu
.extend
;
7676 if (CONST_INT_P (op1
))
7680 if (VECTOR_MODE_P (mode
))
7682 /* Vector shift (immediate). */
7683 *cost
+= extra_cost
->vect
.alu
;
7687 /* LSL (immediate), UBMF, UBFIZ and friends. These are all
7689 *cost
+= extra_cost
->alu
.shift
;
7693 /* We can incorporate zero/sign extend for free. */
7694 if (GET_CODE (op0
) == ZERO_EXTEND
7695 || GET_CODE (op0
) == SIGN_EXTEND
)
7696 op0
= XEXP (op0
, 0);
7698 *cost
+= rtx_cost (op0
, VOIDmode
, ASHIFT
, 0, speed
);
7703 if (VECTOR_MODE_P (mode
))
7706 /* Vector shift (register). */
7707 *cost
+= extra_cost
->vect
.alu
;
7713 *cost
+= extra_cost
->alu
.shift_reg
;
7715 if (GET_CODE (op1
) == AND
&& REG_P (XEXP (op1
, 0))
7716 && CONST_INT_P (XEXP (op1
, 1))
7717 && INTVAL (XEXP (op1
, 1)) == GET_MODE_BITSIZE (mode
) - 1)
7719 *cost
+= rtx_cost (op0
, mode
, (rtx_code
) code
, 0, speed
);
7720 /* We already demanded XEXP (op1, 0) to be REG_P, so
7721 don't recurse into it. */
7725 return false; /* All arguments need to be in registers. */
7735 if (CONST_INT_P (op1
))
7737 /* ASR (immediate) and friends. */
7740 if (VECTOR_MODE_P (mode
))
7741 *cost
+= extra_cost
->vect
.alu
;
7743 *cost
+= extra_cost
->alu
.shift
;
7746 *cost
+= rtx_cost (op0
, mode
, (enum rtx_code
) code
, 0, speed
);
7751 if (VECTOR_MODE_P (mode
))
7754 /* Vector shift (register). */
7755 *cost
+= extra_cost
->vect
.alu
;
7760 /* ASR (register) and friends. */
7761 *cost
+= extra_cost
->alu
.shift_reg
;
7763 if (GET_CODE (op1
) == AND
&& REG_P (XEXP (op1
, 0))
7764 && CONST_INT_P (XEXP (op1
, 1))
7765 && INTVAL (XEXP (op1
, 1)) == GET_MODE_BITSIZE (mode
) - 1)
7767 *cost
+= rtx_cost (op0
, mode
, (rtx_code
) code
, 0, speed
);
7768 /* We already demanded XEXP (op1, 0) to be REG_P, so
7769 don't recurse into it. */
7773 return false; /* All arguments need to be in registers. */
7778 if (aarch64_cmodel
== AARCH64_CMODEL_LARGE
7779 || aarch64_cmodel
== AARCH64_CMODEL_SMALL_SPIC
)
7783 *cost
+= extra_cost
->ldst
.load
;
7785 else if (aarch64_cmodel
== AARCH64_CMODEL_SMALL
7786 || aarch64_cmodel
== AARCH64_CMODEL_SMALL_PIC
)
7788 /* ADRP, followed by ADD. */
7789 *cost
+= COSTS_N_INSNS (1);
7791 *cost
+= 2 * extra_cost
->alu
.arith
;
7793 else if (aarch64_cmodel
== AARCH64_CMODEL_TINY
7794 || aarch64_cmodel
== AARCH64_CMODEL_TINY_PIC
)
7798 *cost
+= extra_cost
->alu
.arith
;
7803 /* One extra load instruction, after accessing the GOT. */
7804 *cost
+= COSTS_N_INSNS (1);
7806 *cost
+= extra_cost
->ldst
.load
;
7812 /* ADRP/ADD (immediate). */
7814 *cost
+= extra_cost
->alu
.arith
;
7822 if (VECTOR_MODE_P (mode
))
7823 *cost
+= extra_cost
->vect
.alu
;
7825 *cost
+= extra_cost
->alu
.bfx
;
7828 /* We can trust that the immediates used will be correct (there
7829 are no by-register forms), so we need only cost op0. */
7830 *cost
+= rtx_cost (XEXP (x
, 0), VOIDmode
, (enum rtx_code
) code
, 0, speed
);
7834 *cost
+= aarch64_rtx_mult_cost (x
, MULT
, 0, speed
);
7835 /* aarch64_rtx_mult_cost always handles recursion to its
7840 /* We can expand signed mod by power of 2 using a NEGS, two parallel
7841 ANDs and a CSNEG. Assume here that CSNEG is the same as the cost of
7842 an unconditional negate. This case should only ever be reached through
7843 the set_smod_pow2_cheap check in expmed.c. */
7844 if (CONST_INT_P (XEXP (x
, 1))
7845 && exact_log2 (INTVAL (XEXP (x
, 1))) > 0
7846 && (mode
== SImode
|| mode
== DImode
))
7848 /* We expand to 4 instructions. Reset the baseline. */
7849 *cost
= COSTS_N_INSNS (4);
7852 *cost
+= 2 * extra_cost
->alu
.logical
7853 + 2 * extra_cost
->alu
.arith
;
7862 /* Slighly prefer UMOD over SMOD. */
7863 if (VECTOR_MODE_P (mode
))
7864 *cost
+= extra_cost
->vect
.alu
;
7865 else if (GET_MODE_CLASS (mode
) == MODE_INT
)
7866 *cost
+= (extra_cost
->mult
[mode
== DImode
].add
7867 + extra_cost
->mult
[mode
== DImode
].idiv
7868 + (code
== MOD
? 1 : 0));
7870 return false; /* All arguments need to be in registers. */
7877 if (VECTOR_MODE_P (mode
))
7878 *cost
+= extra_cost
->vect
.alu
;
7879 else if (GET_MODE_CLASS (mode
) == MODE_INT
)
7880 /* There is no integer SQRT, so only DIV and UDIV can get
7882 *cost
+= (extra_cost
->mult
[mode
== DImode
].idiv
7883 /* Slighly prefer UDIV over SDIV. */
7884 + (code
== DIV
? 1 : 0));
7886 *cost
+= extra_cost
->fp
[mode
== DFmode
].div
;
7888 return false; /* All arguments need to be in registers. */
7891 return aarch64_if_then_else_costs (XEXP (x
, 0), XEXP (x
, 1),
7892 XEXP (x
, 2), cost
, speed
);
7905 return false; /* All arguments must be in registers. */
7914 if (VECTOR_MODE_P (mode
))
7915 *cost
+= extra_cost
->vect
.alu
;
7917 *cost
+= extra_cost
->fp
[mode
== DFmode
].fma
;
7920 /* FMSUB, FNMADD, and FNMSUB are free. */
7921 if (GET_CODE (op0
) == NEG
)
7922 op0
= XEXP (op0
, 0);
7924 if (GET_CODE (op2
) == NEG
)
7925 op2
= XEXP (op2
, 0);
7927 /* aarch64_fnma4_elt_to_64v2df has the NEG as operand 1,
7928 and the by-element operand as operand 0. */
7929 if (GET_CODE (op1
) == NEG
)
7930 op1
= XEXP (op1
, 0);
7932 /* Catch vector-by-element operations. The by-element operand can
7933 either be (vec_duplicate (vec_select (x))) or just
7934 (vec_select (x)), depending on whether we are multiplying by
7935 a vector or a scalar.
7937 Canonicalization is not very good in these cases, FMA4 will put the
7938 by-element operand as operand 0, FNMA4 will have it as operand 1. */
7939 if (GET_CODE (op0
) == VEC_DUPLICATE
)
7940 op0
= XEXP (op0
, 0);
7941 else if (GET_CODE (op1
) == VEC_DUPLICATE
)
7942 op1
= XEXP (op1
, 0);
7944 if (GET_CODE (op0
) == VEC_SELECT
)
7945 op0
= XEXP (op0
, 0);
7946 else if (GET_CODE (op1
) == VEC_SELECT
)
7947 op1
= XEXP (op1
, 0);
7949 /* If the remaining parameters are not registers,
7950 get the cost to put them into registers. */
7951 *cost
+= rtx_cost (op0
, mode
, FMA
, 0, speed
);
7952 *cost
+= rtx_cost (op1
, mode
, FMA
, 1, speed
);
7953 *cost
+= rtx_cost (op2
, mode
, FMA
, 2, speed
);
7957 case UNSIGNED_FLOAT
:
7959 *cost
+= extra_cost
->fp
[mode
== DFmode
].fromint
;
7965 if (VECTOR_MODE_P (mode
))
7967 /*Vector truncate. */
7968 *cost
+= extra_cost
->vect
.alu
;
7971 *cost
+= extra_cost
->fp
[mode
== DFmode
].widen
;
7975 case FLOAT_TRUNCATE
:
7978 if (VECTOR_MODE_P (mode
))
7980 /*Vector conversion. */
7981 *cost
+= extra_cost
->vect
.alu
;
7984 *cost
+= extra_cost
->fp
[mode
== DFmode
].narrow
;
7991 /* Strip the rounding part. They will all be implemented
7992 by the fcvt* family of instructions anyway. */
7993 if (GET_CODE (x
) == UNSPEC
)
7995 unsigned int uns_code
= XINT (x
, 1);
7997 if (uns_code
== UNSPEC_FRINTA
7998 || uns_code
== UNSPEC_FRINTM
7999 || uns_code
== UNSPEC_FRINTN
8000 || uns_code
== UNSPEC_FRINTP
8001 || uns_code
== UNSPEC_FRINTZ
)
8002 x
= XVECEXP (x
, 0, 0);
8007 if (VECTOR_MODE_P (mode
))
8008 *cost
+= extra_cost
->vect
.alu
;
8010 *cost
+= extra_cost
->fp
[GET_MODE (x
) == DFmode
].toint
;
8013 /* We can combine fmul by a power of 2 followed by a fcvt into a single
8014 fixed-point fcvt. */
8015 if (GET_CODE (x
) == MULT
8016 && ((VECTOR_MODE_P (mode
)
8017 && aarch64_vec_fpconst_pow_of_2 (XEXP (x
, 1)) > 0)
8018 || aarch64_fpconst_pow_of_2 (XEXP (x
, 1)) > 0))
8020 *cost
+= rtx_cost (XEXP (x
, 0), VOIDmode
, (rtx_code
) code
,
8025 *cost
+= rtx_cost (x
, VOIDmode
, (enum rtx_code
) code
, 0, speed
);
8029 if (VECTOR_MODE_P (mode
))
8033 *cost
+= extra_cost
->vect
.alu
;
8035 else if (GET_MODE_CLASS (mode
) == MODE_FLOAT
)
8039 /* FABD, which is analogous to FADD. */
8040 if (GET_CODE (op0
) == MINUS
)
8042 *cost
+= rtx_cost (XEXP (op0
, 0), mode
, MINUS
, 0, speed
);
8043 *cost
+= rtx_cost (XEXP (op0
, 1), mode
, MINUS
, 1, speed
);
8045 *cost
+= extra_cost
->fp
[mode
== DFmode
].addsub
;
8049 /* Simple FABS is analogous to FNEG. */
8051 *cost
+= extra_cost
->fp
[mode
== DFmode
].neg
;
8055 /* Integer ABS will either be split to
8056 two arithmetic instructions, or will be an ABS
8057 (scalar), which we don't model. */
8058 *cost
= COSTS_N_INSNS (2);
8060 *cost
+= 2 * extra_cost
->alu
.arith
;
8068 if (VECTOR_MODE_P (mode
))
8069 *cost
+= extra_cost
->vect
.alu
;
8072 /* FMAXNM/FMINNM/FMAX/FMIN.
8073 TODO: This may not be accurate for all implementations, but
8074 we do not model this in the cost tables. */
8075 *cost
+= extra_cost
->fp
[mode
== DFmode
].addsub
;
8081 /* The floating point round to integer frint* instructions. */
8082 if (aarch64_frint_unspec_p (XINT (x
, 1)))
8085 *cost
+= extra_cost
->fp
[mode
== DFmode
].roundint
;
8090 if (XINT (x
, 1) == UNSPEC_RBIT
)
8093 *cost
+= extra_cost
->alu
.rev
;
8101 /* Decompose <su>muldi3_highpart. */
8102 if (/* (truncate:DI */
8105 && GET_MODE (XEXP (x
, 0)) == TImode
8106 && GET_CODE (XEXP (x
, 0)) == LSHIFTRT
8108 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == MULT
8109 /* (ANY_EXTEND:TI (reg:DI))
8110 (ANY_EXTEND:TI (reg:DI))) */
8111 && ((GET_CODE (XEXP (XEXP (XEXP (x
, 0), 0), 0)) == ZERO_EXTEND
8112 && GET_CODE (XEXP (XEXP (XEXP (x
, 0), 0), 1)) == ZERO_EXTEND
)
8113 || (GET_CODE (XEXP (XEXP (XEXP (x
, 0), 0), 0)) == SIGN_EXTEND
8114 && GET_CODE (XEXP (XEXP (XEXP (x
, 0), 0), 1)) == SIGN_EXTEND
))
8115 && GET_MODE (XEXP (XEXP (XEXP (XEXP (x
, 0), 0), 0), 0)) == DImode
8116 && GET_MODE (XEXP (XEXP (XEXP (XEXP (x
, 0), 0), 1), 0)) == DImode
8117 /* (const_int 64) */
8118 && CONST_INT_P (XEXP (XEXP (x
, 0), 1))
8119 && UINTVAL (XEXP (XEXP (x
, 0), 1)) == 64)
8123 *cost
+= extra_cost
->mult
[mode
== DImode
].extend
;
8124 *cost
+= rtx_cost (XEXP (XEXP (XEXP (XEXP (x
, 0), 0), 0), 0),
8125 mode
, MULT
, 0, speed
);
8126 *cost
+= rtx_cost (XEXP (XEXP (XEXP (XEXP (x
, 0), 0), 1), 0),
8127 mode
, MULT
, 1, speed
);
8137 && flag_aarch64_verbose_cost
)
8139 "\nFailed to cost RTX. Assuming default cost.\n");
8144 /* Wrapper around aarch64_rtx_costs, dumps the partial, or total cost
8145 calculated for X. This cost is stored in *COST. Returns true
8146 if the total cost of X was calculated. */
8148 aarch64_rtx_costs_wrapper (rtx x
, machine_mode mode
, int outer
,
8149 int param
, int *cost
, bool speed
)
8151 bool result
= aarch64_rtx_costs (x
, mode
, outer
, param
, cost
, speed
);
8154 && flag_aarch64_verbose_cost
)
8156 print_rtl_single (dump_file
, x
);
8157 fprintf (dump_file
, "\n%s cost: %d (%s)\n",
8158 speed
? "Hot" : "Cold",
8159 *cost
, result
? "final" : "partial");
8166 aarch64_register_move_cost (machine_mode mode
,
8167 reg_class_t from_i
, reg_class_t to_i
)
8169 enum reg_class from
= (enum reg_class
) from_i
;
8170 enum reg_class to
= (enum reg_class
) to_i
;
8171 const struct cpu_regmove_cost
*regmove_cost
8172 = aarch64_tune_params
.regmove_cost
;
8174 /* Caller save and pointer regs are equivalent to GENERAL_REGS. */
8175 if (to
== CALLER_SAVE_REGS
|| to
== POINTER_REGS
)
8178 if (from
== CALLER_SAVE_REGS
|| from
== POINTER_REGS
)
8179 from
= GENERAL_REGS
;
8181 /* Moving between GPR and stack cost is the same as GP2GP. */
8182 if ((from
== GENERAL_REGS
&& to
== STACK_REG
)
8183 || (to
== GENERAL_REGS
&& from
== STACK_REG
))
8184 return regmove_cost
->GP2GP
;
8186 /* To/From the stack register, we move via the gprs. */
8187 if (to
== STACK_REG
|| from
== STACK_REG
)
8188 return aarch64_register_move_cost (mode
, from
, GENERAL_REGS
)
8189 + aarch64_register_move_cost (mode
, GENERAL_REGS
, to
);
8191 if (GET_MODE_SIZE (mode
) == 16)
8193 /* 128-bit operations on general registers require 2 instructions. */
8194 if (from
== GENERAL_REGS
&& to
== GENERAL_REGS
)
8195 return regmove_cost
->GP2GP
* 2;
8196 else if (from
== GENERAL_REGS
)
8197 return regmove_cost
->GP2FP
* 2;
8198 else if (to
== GENERAL_REGS
)
8199 return regmove_cost
->FP2GP
* 2;
8201 /* When AdvSIMD instructions are disabled it is not possible to move
8202 a 128-bit value directly between Q registers. This is handled in
8203 secondary reload. A general register is used as a scratch to move
8204 the upper DI value and the lower DI value is moved directly,
8205 hence the cost is the sum of three moves. */
8207 return regmove_cost
->GP2FP
+ regmove_cost
->FP2GP
+ regmove_cost
->FP2FP
;
8209 return regmove_cost
->FP2FP
;
8212 if (from
== GENERAL_REGS
&& to
== GENERAL_REGS
)
8213 return regmove_cost
->GP2GP
;
8214 else if (from
== GENERAL_REGS
)
8215 return regmove_cost
->GP2FP
;
8216 else if (to
== GENERAL_REGS
)
8217 return regmove_cost
->FP2GP
;
8219 return regmove_cost
->FP2FP
;
8223 aarch64_memory_move_cost (machine_mode mode ATTRIBUTE_UNUSED
,
8224 reg_class_t rclass ATTRIBUTE_UNUSED
,
8225 bool in ATTRIBUTE_UNUSED
)
8227 return aarch64_tune_params
.memmov_cost
;
8230 /* Return true if it is safe and beneficial to use the approximate rsqrt optabs
8231 to optimize 1.0/sqrt. */
8234 use_rsqrt_p (machine_mode mode
)
8236 return (!flag_trapping_math
8237 && flag_unsafe_math_optimizations
8238 && ((aarch64_tune_params
.approx_modes
->recip_sqrt
8239 & AARCH64_APPROX_MODE (mode
))
8240 || flag_mrecip_low_precision_sqrt
));
8243 /* Function to decide when to use the approximate reciprocal square root
8247 aarch64_builtin_reciprocal (tree fndecl
)
8249 machine_mode mode
= TYPE_MODE (TREE_TYPE (fndecl
));
8251 if (!use_rsqrt_p (mode
))
8253 return aarch64_builtin_rsqrt (DECL_FUNCTION_CODE (fndecl
));
8256 typedef rtx (*rsqrte_type
) (rtx
, rtx
);
8258 /* Select reciprocal square root initial estimate insn depending on machine
8262 get_rsqrte_type (machine_mode mode
)
8266 case E_DFmode
: return gen_aarch64_rsqrtedf
;
8267 case E_SFmode
: return gen_aarch64_rsqrtesf
;
8268 case E_V2DFmode
: return gen_aarch64_rsqrtev2df
;
8269 case E_V2SFmode
: return gen_aarch64_rsqrtev2sf
;
8270 case E_V4SFmode
: return gen_aarch64_rsqrtev4sf
;
8271 default: gcc_unreachable ();
8275 typedef rtx (*rsqrts_type
) (rtx
, rtx
, rtx
);
8277 /* Select reciprocal square root series step insn depending on machine mode. */
8280 get_rsqrts_type (machine_mode mode
)
8284 case E_DFmode
: return gen_aarch64_rsqrtsdf
;
8285 case E_SFmode
: return gen_aarch64_rsqrtssf
;
8286 case E_V2DFmode
: return gen_aarch64_rsqrtsv2df
;
8287 case E_V2SFmode
: return gen_aarch64_rsqrtsv2sf
;
8288 case E_V4SFmode
: return gen_aarch64_rsqrtsv4sf
;
8289 default: gcc_unreachable ();
8293 /* Emit instruction sequence to compute either the approximate square root
8294 or its approximate reciprocal, depending on the flag RECP, and return
8295 whether the sequence was emitted or not. */
8298 aarch64_emit_approx_sqrt (rtx dst
, rtx src
, bool recp
)
8300 machine_mode mode
= GET_MODE (dst
);
8302 if (GET_MODE_INNER (mode
) == HFmode
)
8310 if (!(flag_mlow_precision_sqrt
8311 || (aarch64_tune_params
.approx_modes
->sqrt
8312 & AARCH64_APPROX_MODE (mode
))))
8315 if (flag_finite_math_only
8316 || flag_trapping_math
8317 || !flag_unsafe_math_optimizations
8318 || optimize_function_for_size_p (cfun
))
8322 /* Caller assumes we cannot fail. */
8323 gcc_assert (use_rsqrt_p (mode
));
8325 machine_mode mmsk
= mode_for_int_vector (mode
).require ();
8326 rtx xmsk
= gen_reg_rtx (mmsk
);
8328 /* When calculating the approximate square root, compare the
8329 argument with 0.0 and create a mask. */
8330 emit_insn (gen_rtx_SET (xmsk
,
8332 gen_rtx_EQ (mmsk
, src
,
8333 CONST0_RTX (mode
)))));
8335 /* Estimate the approximate reciprocal square root. */
8336 rtx xdst
= gen_reg_rtx (mode
);
8337 emit_insn ((*get_rsqrte_type (mode
)) (xdst
, src
));
8339 /* Iterate over the series twice for SF and thrice for DF. */
8340 int iterations
= (GET_MODE_INNER (mode
) == DFmode
) ? 3 : 2;
8342 /* Optionally iterate over the series once less for faster performance
8343 while sacrificing the accuracy. */
8344 if ((recp
&& flag_mrecip_low_precision_sqrt
)
8345 || (!recp
&& flag_mlow_precision_sqrt
))
8348 /* Iterate over the series to calculate the approximate reciprocal square
8350 rtx x1
= gen_reg_rtx (mode
);
8351 while (iterations
--)
8353 rtx x2
= gen_reg_rtx (mode
);
8354 emit_set_insn (x2
, gen_rtx_MULT (mode
, xdst
, xdst
));
8356 emit_insn ((*get_rsqrts_type (mode
)) (x1
, src
, x2
));
8359 emit_set_insn (xdst
, gen_rtx_MULT (mode
, xdst
, x1
));
8364 /* Qualify the approximate reciprocal square root when the argument is
8365 0.0 by squashing the intermediary result to 0.0. */
8366 rtx xtmp
= gen_reg_rtx (mmsk
);
8367 emit_set_insn (xtmp
, gen_rtx_AND (mmsk
, gen_rtx_NOT (mmsk
, xmsk
),
8368 gen_rtx_SUBREG (mmsk
, xdst
, 0)));
8369 emit_move_insn (xdst
, gen_rtx_SUBREG (mode
, xtmp
, 0));
8371 /* Calculate the approximate square root. */
8372 emit_set_insn (xdst
, gen_rtx_MULT (mode
, xdst
, src
));
8375 /* Finalize the approximation. */
8376 emit_set_insn (dst
, gen_rtx_MULT (mode
, xdst
, x1
));
8381 typedef rtx (*recpe_type
) (rtx
, rtx
);
8383 /* Select reciprocal initial estimate insn depending on machine mode. */
8386 get_recpe_type (machine_mode mode
)
8390 case E_SFmode
: return (gen_aarch64_frecpesf
);
8391 case E_V2SFmode
: return (gen_aarch64_frecpev2sf
);
8392 case E_V4SFmode
: return (gen_aarch64_frecpev4sf
);
8393 case E_DFmode
: return (gen_aarch64_frecpedf
);
8394 case E_V2DFmode
: return (gen_aarch64_frecpev2df
);
8395 default: gcc_unreachable ();
8399 typedef rtx (*recps_type
) (rtx
, rtx
, rtx
);
8401 /* Select reciprocal series step insn depending on machine mode. */
8404 get_recps_type (machine_mode mode
)
8408 case E_SFmode
: return (gen_aarch64_frecpssf
);
8409 case E_V2SFmode
: return (gen_aarch64_frecpsv2sf
);
8410 case E_V4SFmode
: return (gen_aarch64_frecpsv4sf
);
8411 case E_DFmode
: return (gen_aarch64_frecpsdf
);
8412 case E_V2DFmode
: return (gen_aarch64_frecpsv2df
);
8413 default: gcc_unreachable ();
8417 /* Emit the instruction sequence to compute the approximation for the division
8418 of NUM by DEN in QUO and return whether the sequence was emitted or not. */
8421 aarch64_emit_approx_div (rtx quo
, rtx num
, rtx den
)
8423 machine_mode mode
= GET_MODE (quo
);
8425 if (GET_MODE_INNER (mode
) == HFmode
)
8428 bool use_approx_division_p
= (flag_mlow_precision_div
8429 || (aarch64_tune_params
.approx_modes
->division
8430 & AARCH64_APPROX_MODE (mode
)));
8432 if (!flag_finite_math_only
8433 || flag_trapping_math
8434 || !flag_unsafe_math_optimizations
8435 || optimize_function_for_size_p (cfun
)
8436 || !use_approx_division_p
)
8439 /* Estimate the approximate reciprocal. */
8440 rtx xrcp
= gen_reg_rtx (mode
);
8441 emit_insn ((*get_recpe_type (mode
)) (xrcp
, den
));
8443 /* Iterate over the series twice for SF and thrice for DF. */
8444 int iterations
= (GET_MODE_INNER (mode
) == DFmode
) ? 3 : 2;
8446 /* Optionally iterate over the series once less for faster performance,
8447 while sacrificing the accuracy. */
8448 if (flag_mlow_precision_div
)
8451 /* Iterate over the series to calculate the approximate reciprocal. */
8452 rtx xtmp
= gen_reg_rtx (mode
);
8453 while (iterations
--)
8455 emit_insn ((*get_recps_type (mode
)) (xtmp
, xrcp
, den
));
8458 emit_set_insn (xrcp
, gen_rtx_MULT (mode
, xrcp
, xtmp
));
8461 if (num
!= CONST1_RTX (mode
))
8463 /* As the approximate reciprocal of DEN is already calculated, only
8464 calculate the approximate division when NUM is not 1.0. */
8465 rtx xnum
= force_reg (mode
, num
);
8466 emit_set_insn (xrcp
, gen_rtx_MULT (mode
, xrcp
, xnum
));
8469 /* Finalize the approximation. */
8470 emit_set_insn (quo
, gen_rtx_MULT (mode
, xrcp
, xtmp
));
8474 /* Return the number of instructions that can be issued per cycle. */
8476 aarch64_sched_issue_rate (void)
8478 return aarch64_tune_params
.issue_rate
;
8482 aarch64_sched_first_cycle_multipass_dfa_lookahead (void)
8484 int issue_rate
= aarch64_sched_issue_rate ();
8486 return issue_rate
> 1 && !sched_fusion
? issue_rate
: 0;
8490 /* Implement TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD as
8491 autopref_multipass_dfa_lookahead_guard from haifa-sched.c. It only
8492 has an effect if PARAM_SCHED_AUTOPREF_QUEUE_DEPTH > 0. */
8495 aarch64_first_cycle_multipass_dfa_lookahead_guard (rtx_insn
*insn
,
8498 return autopref_multipass_dfa_lookahead_guard (insn
, ready_index
);
8502 /* Vectorizer cost model target hooks. */
8504 /* Implement targetm.vectorize.builtin_vectorization_cost. */
8506 aarch64_builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost
,
8508 int misalign ATTRIBUTE_UNUSED
)
8511 const cpu_vector_cost
*costs
= aarch64_tune_params
.vec_costs
;
8514 if (vectype
!= NULL
)
8515 fp
= FLOAT_TYPE_P (vectype
);
8517 switch (type_of_cost
)
8520 return fp
? costs
->scalar_fp_stmt_cost
: costs
->scalar_int_stmt_cost
;
8523 return costs
->scalar_load_cost
;
8526 return costs
->scalar_store_cost
;
8529 return fp
? costs
->vec_fp_stmt_cost
: costs
->vec_int_stmt_cost
;
8532 return costs
->vec_align_load_cost
;
8535 return costs
->vec_store_cost
;
8538 return costs
->vec_to_scalar_cost
;
8541 return costs
->scalar_to_vec_cost
;
8543 case unaligned_load
:
8544 return costs
->vec_unalign_load_cost
;
8546 case unaligned_store
:
8547 return costs
->vec_unalign_store_cost
;
8549 case cond_branch_taken
:
8550 return costs
->cond_taken_branch_cost
;
8552 case cond_branch_not_taken
:
8553 return costs
->cond_not_taken_branch_cost
;
8556 return costs
->vec_permute_cost
;
8558 case vec_promote_demote
:
8559 return fp
? costs
->vec_fp_stmt_cost
: costs
->vec_int_stmt_cost
;
8562 elements
= TYPE_VECTOR_SUBPARTS (vectype
);
8563 return elements
/ 2 + 1;
8570 /* Implement targetm.vectorize.add_stmt_cost. */
8572 aarch64_add_stmt_cost (void *data
, int count
, enum vect_cost_for_stmt kind
,
8573 struct _stmt_vec_info
*stmt_info
, int misalign
,
8574 enum vect_cost_model_location where
)
8576 unsigned *cost
= (unsigned *) data
;
8577 unsigned retval
= 0;
8579 if (flag_vect_cost_model
)
8581 tree vectype
= stmt_info
? stmt_vectype (stmt_info
) : NULL_TREE
;
8583 aarch64_builtin_vectorization_cost (kind
, vectype
, misalign
);
8585 /* Statements in an inner loop relative to the loop being
8586 vectorized are weighted more heavily. The value here is
8587 arbitrary and could potentially be improved with analysis. */
8588 if (where
== vect_body
&& stmt_info
&& stmt_in_inner_loop_p (stmt_info
))
8589 count
*= 50; /* FIXME */
8591 retval
= (unsigned) (count
* stmt_cost
);
8592 cost
[where
] += retval
;
8598 static void initialize_aarch64_code_model (struct gcc_options
*);
8600 /* Parse the TO_PARSE string and put the architecture struct that it
8601 selects into RES and the architectural features into ISA_FLAGS.
8602 Return an aarch64_parse_opt_result describing the parse result.
8603 If there is an error parsing, RES and ISA_FLAGS are left unchanged. */
8605 static enum aarch64_parse_opt_result
8606 aarch64_parse_arch (const char *to_parse
, const struct processor
**res
,
8607 unsigned long *isa_flags
)
8610 const struct processor
*arch
;
8611 char *str
= (char *) alloca (strlen (to_parse
) + 1);
8614 strcpy (str
, to_parse
);
8616 ext
= strchr (str
, '+');
8624 return AARCH64_PARSE_MISSING_ARG
;
8627 /* Loop through the list of supported ARCHes to find a match. */
8628 for (arch
= all_architectures
; arch
->name
!= NULL
; arch
++)
8630 if (strlen (arch
->name
) == len
&& strncmp (arch
->name
, str
, len
) == 0)
8632 unsigned long isa_temp
= arch
->flags
;
8636 /* TO_PARSE string contains at least one extension. */
8637 enum aarch64_parse_opt_result ext_res
8638 = aarch64_parse_extension (ext
, &isa_temp
);
8640 if (ext_res
!= AARCH64_PARSE_OK
)
8643 /* Extension parsing was successful. Confirm the result
8644 arch and ISA flags. */
8646 *isa_flags
= isa_temp
;
8647 return AARCH64_PARSE_OK
;
8651 /* ARCH name not found in list. */
8652 return AARCH64_PARSE_INVALID_ARG
;
8655 /* Parse the TO_PARSE string and put the result tuning in RES and the
8656 architecture flags in ISA_FLAGS. Return an aarch64_parse_opt_result
8657 describing the parse result. If there is an error parsing, RES and
8658 ISA_FLAGS are left unchanged. */
8660 static enum aarch64_parse_opt_result
8661 aarch64_parse_cpu (const char *to_parse
, const struct processor
**res
,
8662 unsigned long *isa_flags
)
8665 const struct processor
*cpu
;
8666 char *str
= (char *) alloca (strlen (to_parse
) + 1);
8669 strcpy (str
, to_parse
);
8671 ext
= strchr (str
, '+');
8679 return AARCH64_PARSE_MISSING_ARG
;
8682 /* Loop through the list of supported CPUs to find a match. */
8683 for (cpu
= all_cores
; cpu
->name
!= NULL
; cpu
++)
8685 if (strlen (cpu
->name
) == len
&& strncmp (cpu
->name
, str
, len
) == 0)
8687 unsigned long isa_temp
= cpu
->flags
;
8692 /* TO_PARSE string contains at least one extension. */
8693 enum aarch64_parse_opt_result ext_res
8694 = aarch64_parse_extension (ext
, &isa_temp
);
8696 if (ext_res
!= AARCH64_PARSE_OK
)
8699 /* Extension parsing was successfull. Confirm the result
8700 cpu and ISA flags. */
8702 *isa_flags
= isa_temp
;
8703 return AARCH64_PARSE_OK
;
8707 /* CPU name not found in list. */
8708 return AARCH64_PARSE_INVALID_ARG
;
8711 /* Parse the TO_PARSE string and put the cpu it selects into RES.
8712 Return an aarch64_parse_opt_result describing the parse result.
8713 If the parsing fails the RES does not change. */
8715 static enum aarch64_parse_opt_result
8716 aarch64_parse_tune (const char *to_parse
, const struct processor
**res
)
8718 const struct processor
*cpu
;
8719 char *str
= (char *) alloca (strlen (to_parse
) + 1);
8721 strcpy (str
, to_parse
);
8723 /* Loop through the list of supported CPUs to find a match. */
8724 for (cpu
= all_cores
; cpu
->name
!= NULL
; cpu
++)
8726 if (strcmp (cpu
->name
, str
) == 0)
8729 return AARCH64_PARSE_OK
;
8733 /* CPU name not found in list. */
8734 return AARCH64_PARSE_INVALID_ARG
;
8737 /* Parse TOKEN, which has length LENGTH to see if it is an option
8738 described in FLAG. If it is, return the index bit for that fusion type.
8739 If not, error (printing OPTION_NAME) and return zero. */
8742 aarch64_parse_one_option_token (const char *token
,
8744 const struct aarch64_flag_desc
*flag
,
8745 const char *option_name
)
8747 for (; flag
->name
!= NULL
; flag
++)
8749 if (length
== strlen (flag
->name
)
8750 && !strncmp (flag
->name
, token
, length
))
8754 error ("unknown flag passed in -moverride=%s (%s)", option_name
, token
);
8758 /* Parse OPTION which is a comma-separated list of flags to enable.
8759 FLAGS gives the list of flags we understand, INITIAL_STATE gives any
8760 default state we inherit from the CPU tuning structures. OPTION_NAME
8761 gives the top-level option we are parsing in the -moverride string,
8762 for use in error messages. */
8765 aarch64_parse_boolean_options (const char *option
,
8766 const struct aarch64_flag_desc
*flags
,
8767 unsigned int initial_state
,
8768 const char *option_name
)
8770 const char separator
= '.';
8771 const char* specs
= option
;
8772 const char* ntoken
= option
;
8773 unsigned int found_flags
= initial_state
;
8775 while ((ntoken
= strchr (specs
, separator
)))
8777 size_t token_length
= ntoken
- specs
;
8778 unsigned token_ops
= aarch64_parse_one_option_token (specs
,
8782 /* If we find "none" (or, for simplicity's sake, an error) anywhere
8783 in the token stream, reset the supported operations. So:
8785 adrp+add.cmp+branch.none.adrp+add
8787 would have the result of turning on only adrp+add fusion. */
8791 found_flags
|= token_ops
;
8795 /* We ended with a comma, print something. */
8798 error ("%s string ill-formed\n", option_name
);
8802 /* We still have one more token to parse. */
8803 size_t token_length
= strlen (specs
);
8804 unsigned token_ops
= aarch64_parse_one_option_token (specs
,
8811 found_flags
|= token_ops
;
8815 /* Support for overriding instruction fusion. */
8818 aarch64_parse_fuse_string (const char *fuse_string
,
8819 struct tune_params
*tune
)
8821 tune
->fusible_ops
= aarch64_parse_boolean_options (fuse_string
,
8822 aarch64_fusible_pairs
,
8827 /* Support for overriding other tuning flags. */
8830 aarch64_parse_tune_string (const char *tune_string
,
8831 struct tune_params
*tune
)
8833 tune
->extra_tuning_flags
8834 = aarch64_parse_boolean_options (tune_string
,
8835 aarch64_tuning_flags
,
8836 tune
->extra_tuning_flags
,
8840 /* Parse TOKEN, which has length LENGTH to see if it is a tuning option
8841 we understand. If it is, extract the option string and handoff to
8842 the appropriate function. */
8845 aarch64_parse_one_override_token (const char* token
,
8847 struct tune_params
*tune
)
8849 const struct aarch64_tuning_override_function
*fn
8850 = aarch64_tuning_override_functions
;
8852 const char *option_part
= strchr (token
, '=');
8855 error ("tuning string missing in option (%s)", token
);
8859 /* Get the length of the option name. */
8860 length
= option_part
- token
;
8861 /* Skip the '=' to get to the option string. */
8864 for (; fn
->name
!= NULL
; fn
++)
8866 if (!strncmp (fn
->name
, token
, length
))
8868 fn
->parse_override (option_part
, tune
);
8873 error ("unknown tuning option (%s)",token
);
8877 /* A checking mechanism for the implementation of the tls size. */
8880 initialize_aarch64_tls_size (struct gcc_options
*opts
)
8882 if (aarch64_tls_size
== 0)
8883 aarch64_tls_size
= 24;
8885 switch (opts
->x_aarch64_cmodel_var
)
8887 case AARCH64_CMODEL_TINY
:
8888 /* Both the default and maximum TLS size allowed under tiny is 1M which
8889 needs two instructions to address, so we clamp the size to 24. */
8890 if (aarch64_tls_size
> 24)
8891 aarch64_tls_size
= 24;
8893 case AARCH64_CMODEL_SMALL
:
8894 /* The maximum TLS size allowed under small is 4G. */
8895 if (aarch64_tls_size
> 32)
8896 aarch64_tls_size
= 32;
8898 case AARCH64_CMODEL_LARGE
:
8899 /* The maximum TLS size allowed under large is 16E.
8900 FIXME: 16E should be 64bit, we only support 48bit offset now. */
8901 if (aarch64_tls_size
> 48)
8902 aarch64_tls_size
= 48;
8911 /* Parse STRING looking for options in the format:
8912 string :: option:string
8913 option :: name=substring
8915 substring :: defined by option. */
8918 aarch64_parse_override_string (const char* input_string
,
8919 struct tune_params
* tune
)
8921 const char separator
= ':';
8922 size_t string_length
= strlen (input_string
) + 1;
8923 char *string_root
= (char *) xmalloc (sizeof (*string_root
) * string_length
);
8924 char *string
= string_root
;
8925 strncpy (string
, input_string
, string_length
);
8926 string
[string_length
- 1] = '\0';
8928 char* ntoken
= string
;
8930 while ((ntoken
= strchr (string
, separator
)))
8932 size_t token_length
= ntoken
- string
;
8933 /* Make this substring look like a string. */
8935 aarch64_parse_one_override_token (string
, token_length
, tune
);
8939 /* One last option to parse. */
8940 aarch64_parse_one_override_token (string
, strlen (string
), tune
);
8946 aarch64_override_options_after_change_1 (struct gcc_options
*opts
)
8948 /* The logic here is that if we are disabling all frame pointer generation
8949 then we do not need to disable leaf frame pointer generation as a
8950 separate operation. But if we are *only* disabling leaf frame pointer
8951 generation then we set flag_omit_frame_pointer to true, but in
8952 aarch64_frame_pointer_required we return false only for leaf functions.
8954 PR 70044: We have to be careful about being called multiple times for the
8955 same function. Once we have decided to set flag_omit_frame_pointer just
8956 so that we can omit leaf frame pointers, we must then not interpret a
8957 second call as meaning that all frame pointer generation should be
8958 omitted. We do this by setting flag_omit_frame_pointer to a special,
8960 if (opts
->x_flag_omit_frame_pointer
== 2)
8961 opts
->x_flag_omit_frame_pointer
= 0;
8963 if (opts
->x_flag_omit_frame_pointer
)
8964 opts
->x_flag_omit_leaf_frame_pointer
= false;
8965 else if (opts
->x_flag_omit_leaf_frame_pointer
)
8966 opts
->x_flag_omit_frame_pointer
= 2;
8968 /* If not optimizing for size, set the default
8969 alignment to what the target wants. */
8970 if (!opts
->x_optimize_size
)
8972 if (opts
->x_align_loops
<= 0)
8973 opts
->x_align_loops
= aarch64_tune_params
.loop_align
;
8974 if (opts
->x_align_jumps
<= 0)
8975 opts
->x_align_jumps
= aarch64_tune_params
.jump_align
;
8976 if (opts
->x_align_functions
<= 0)
8977 opts
->x_align_functions
= aarch64_tune_params
.function_align
;
8980 /* We default to no pc-relative literal loads. */
8982 aarch64_pcrelative_literal_loads
= false;
8984 /* If -mpc-relative-literal-loads is set on the command line, this
8985 implies that the user asked for PC relative literal loads. */
8986 if (opts
->x_pcrelative_literal_loads
== 1)
8987 aarch64_pcrelative_literal_loads
= true;
8989 /* This is PR70113. When building the Linux kernel with
8990 CONFIG_ARM64_ERRATUM_843419, support for relocations
8991 R_AARCH64_ADR_PREL_PG_HI21 and R_AARCH64_ADR_PREL_PG_HI21_NC is
8992 removed from the kernel to avoid loading objects with possibly
8993 offending sequences. Without -mpc-relative-literal-loads we would
8994 generate such relocations, preventing the kernel build from
8996 if (opts
->x_pcrelative_literal_loads
== 2
8997 && TARGET_FIX_ERR_A53_843419
)
8998 aarch64_pcrelative_literal_loads
= true;
9000 /* In the tiny memory model it makes no sense to disallow PC relative
9001 literal pool loads. */
9002 if (aarch64_cmodel
== AARCH64_CMODEL_TINY
9003 || aarch64_cmodel
== AARCH64_CMODEL_TINY_PIC
)
9004 aarch64_pcrelative_literal_loads
= true;
9006 /* When enabling the lower precision Newton series for the square root, also
9007 enable it for the reciprocal square root, since the latter is an
9008 intermediary step for the former. */
9009 if (flag_mlow_precision_sqrt
)
9010 flag_mrecip_low_precision_sqrt
= true;
9013 /* 'Unpack' up the internal tuning structs and update the options
9014 in OPTS. The caller must have set up selected_tune and selected_arch
9015 as all the other target-specific codegen decisions are
9016 derived from them. */
9019 aarch64_override_options_internal (struct gcc_options
*opts
)
9021 aarch64_tune_flags
= selected_tune
->flags
;
9022 aarch64_tune
= selected_tune
->sched_core
;
9023 /* Make a copy of the tuning parameters attached to the core, which
9024 we may later overwrite. */
9025 aarch64_tune_params
= *(selected_tune
->tune
);
9026 aarch64_architecture_version
= selected_arch
->architecture_version
;
9028 if (opts
->x_aarch64_override_tune_string
)
9029 aarch64_parse_override_string (opts
->x_aarch64_override_tune_string
,
9030 &aarch64_tune_params
);
9032 /* This target defaults to strict volatile bitfields. */
9033 if (opts
->x_flag_strict_volatile_bitfields
< 0 && abi_version_at_least (2))
9034 opts
->x_flag_strict_volatile_bitfields
= 1;
9036 initialize_aarch64_code_model (opts
);
9037 initialize_aarch64_tls_size (opts
);
9039 int queue_depth
= 0;
9040 switch (aarch64_tune_params
.autoprefetcher_model
)
9042 case tune_params::AUTOPREFETCHER_OFF
:
9045 case tune_params::AUTOPREFETCHER_WEAK
:
9048 case tune_params::AUTOPREFETCHER_STRONG
:
9049 queue_depth
= max_insn_queue_index
+ 1;
9055 /* We don't mind passing in global_options_set here as we don't use
9056 the *options_set structs anyway. */
9057 maybe_set_param_value (PARAM_SCHED_AUTOPREF_QUEUE_DEPTH
,
9059 opts
->x_param_values
,
9060 global_options_set
.x_param_values
);
9062 /* Set up parameters to be used in prefetching algorithm. Do not
9063 override the defaults unless we are tuning for a core we have
9064 researched values for. */
9065 if (aarch64_tune_params
.prefetch
->num_slots
> 0)
9066 maybe_set_param_value (PARAM_SIMULTANEOUS_PREFETCHES
,
9067 aarch64_tune_params
.prefetch
->num_slots
,
9068 opts
->x_param_values
,
9069 global_options_set
.x_param_values
);
9070 if (aarch64_tune_params
.prefetch
->l1_cache_size
>= 0)
9071 maybe_set_param_value (PARAM_L1_CACHE_SIZE
,
9072 aarch64_tune_params
.prefetch
->l1_cache_size
,
9073 opts
->x_param_values
,
9074 global_options_set
.x_param_values
);
9075 if (aarch64_tune_params
.prefetch
->l1_cache_line_size
>= 0)
9076 maybe_set_param_value (PARAM_L1_CACHE_LINE_SIZE
,
9077 aarch64_tune_params
.prefetch
->l1_cache_line_size
,
9078 opts
->x_param_values
,
9079 global_options_set
.x_param_values
);
9080 if (aarch64_tune_params
.prefetch
->l2_cache_size
>= 0)
9081 maybe_set_param_value (PARAM_L2_CACHE_SIZE
,
9082 aarch64_tune_params
.prefetch
->l2_cache_size
,
9083 opts
->x_param_values
,
9084 global_options_set
.x_param_values
);
9086 /* Enable sw prefetching at specified optimization level for
9087 CPUS that have prefetch. Lower optimization level threshold by 1
9088 when profiling is enabled. */
9089 if (opts
->x_flag_prefetch_loop_arrays
< 0
9090 && !opts
->x_optimize_size
9091 && aarch64_tune_params
.prefetch
->default_opt_level
>= 0
9092 && opts
->x_optimize
>= aarch64_tune_params
.prefetch
->default_opt_level
)
9093 opts
->x_flag_prefetch_loop_arrays
= 1;
9095 aarch64_override_options_after_change_1 (opts
);
9098 /* Print a hint with a suggestion for a core or architecture name that
9099 most closely resembles what the user passed in STR. ARCH is true if
9100 the user is asking for an architecture name. ARCH is false if the user
9101 is asking for a core name. */
9104 aarch64_print_hint_for_core_or_arch (const char *str
, bool arch
)
9106 auto_vec
<const char *> candidates
;
9107 const struct processor
*entry
= arch
? all_architectures
: all_cores
;
9108 for (; entry
->name
!= NULL
; entry
++)
9109 candidates
.safe_push (entry
->name
);
9111 const char *hint
= candidates_list_and_hint (str
, s
, candidates
);
9113 inform (input_location
, "valid arguments are: %s;"
9114 " did you mean %qs?", s
, hint
);
9118 /* Print a hint with a suggestion for a core name that most closely resembles
9119 what the user passed in STR. */
9122 aarch64_print_hint_for_core (const char *str
)
9124 aarch64_print_hint_for_core_or_arch (str
, false);
9127 /* Print a hint with a suggestion for an architecture name that most closely
9128 resembles what the user passed in STR. */
9131 aarch64_print_hint_for_arch (const char *str
)
9133 aarch64_print_hint_for_core_or_arch (str
, true);
9136 /* Validate a command-line -mcpu option. Parse the cpu and extensions (if any)
9137 specified in STR and throw errors if appropriate. Put the results if
9138 they are valid in RES and ISA_FLAGS. Return whether the option is
9142 aarch64_validate_mcpu (const char *str
, const struct processor
**res
,
9143 unsigned long *isa_flags
)
9145 enum aarch64_parse_opt_result parse_res
9146 = aarch64_parse_cpu (str
, res
, isa_flags
);
9148 if (parse_res
== AARCH64_PARSE_OK
)
9153 case AARCH64_PARSE_MISSING_ARG
:
9154 error ("missing cpu name in %<-mcpu=%s%>", str
);
9156 case AARCH64_PARSE_INVALID_ARG
:
9157 error ("unknown value %qs for -mcpu", str
);
9158 aarch64_print_hint_for_core (str
);
9160 case AARCH64_PARSE_INVALID_FEATURE
:
9161 error ("invalid feature modifier in %<-mcpu=%s%>", str
);
9170 /* Validate a command-line -march option. Parse the arch and extensions
9171 (if any) specified in STR and throw errors if appropriate. Put the
9172 results, if they are valid, in RES and ISA_FLAGS. Return whether the
9176 aarch64_validate_march (const char *str
, const struct processor
**res
,
9177 unsigned long *isa_flags
)
9179 enum aarch64_parse_opt_result parse_res
9180 = aarch64_parse_arch (str
, res
, isa_flags
);
9182 if (parse_res
== AARCH64_PARSE_OK
)
9187 case AARCH64_PARSE_MISSING_ARG
:
9188 error ("missing arch name in %<-march=%s%>", str
);
9190 case AARCH64_PARSE_INVALID_ARG
:
9191 error ("unknown value %qs for -march", str
);
9192 aarch64_print_hint_for_arch (str
);
9194 case AARCH64_PARSE_INVALID_FEATURE
:
9195 error ("invalid feature modifier in %<-march=%s%>", str
);
9204 /* Validate a command-line -mtune option. Parse the cpu
9205 specified in STR and throw errors if appropriate. Put the
9206 result, if it is valid, in RES. Return whether the option is
9210 aarch64_validate_mtune (const char *str
, const struct processor
**res
)
9212 enum aarch64_parse_opt_result parse_res
9213 = aarch64_parse_tune (str
, res
);
9215 if (parse_res
== AARCH64_PARSE_OK
)
9220 case AARCH64_PARSE_MISSING_ARG
:
9221 error ("missing cpu name in %<-mtune=%s%>", str
);
9223 case AARCH64_PARSE_INVALID_ARG
:
9224 error ("unknown value %qs for -mtune", str
);
9225 aarch64_print_hint_for_core (str
);
9233 /* Return the CPU corresponding to the enum CPU.
9234 If it doesn't specify a cpu, return the default. */
9236 static const struct processor
*
9237 aarch64_get_tune_cpu (enum aarch64_processor cpu
)
9239 if (cpu
!= aarch64_none
)
9240 return &all_cores
[cpu
];
9242 /* The & 0x3f is to extract the bottom 6 bits that encode the
9243 default cpu as selected by the --with-cpu GCC configure option
9245 ???: The whole TARGET_CPU_DEFAULT and AARCH64_CPU_DEFAULT_FLAGS
9246 flags mechanism should be reworked to make it more sane. */
9247 return &all_cores
[TARGET_CPU_DEFAULT
& 0x3f];
9250 /* Return the architecture corresponding to the enum ARCH.
9251 If it doesn't specify a valid architecture, return the default. */
9253 static const struct processor
*
9254 aarch64_get_arch (enum aarch64_arch arch
)
9256 if (arch
!= aarch64_no_arch
)
9257 return &all_architectures
[arch
];
9259 const struct processor
*cpu
= &all_cores
[TARGET_CPU_DEFAULT
& 0x3f];
9261 return &all_architectures
[cpu
->arch
];
9264 /* Implement TARGET_OPTION_OVERRIDE. This is called once in the beginning
9265 and is used to parse the -m{cpu,tune,arch} strings and setup the initial
9266 tuning structs. In particular it must set selected_tune and
9267 aarch64_isa_flags that define the available ISA features and tuning
9268 decisions. It must also set selected_arch as this will be used to
9269 output the .arch asm tags for each function. */
9272 aarch64_override_options (void)
9274 unsigned long cpu_isa
= 0;
9275 unsigned long arch_isa
= 0;
9276 aarch64_isa_flags
= 0;
9278 bool valid_cpu
= true;
9279 bool valid_tune
= true;
9280 bool valid_arch
= true;
9282 selected_cpu
= NULL
;
9283 selected_arch
= NULL
;
9284 selected_tune
= NULL
;
9286 /* -mcpu=CPU is shorthand for -march=ARCH_FOR_CPU, -mtune=CPU.
9287 If either of -march or -mtune is given, they override their
9288 respective component of -mcpu. */
9289 if (aarch64_cpu_string
)
9290 valid_cpu
= aarch64_validate_mcpu (aarch64_cpu_string
, &selected_cpu
,
9293 if (aarch64_arch_string
)
9294 valid_arch
= aarch64_validate_march (aarch64_arch_string
, &selected_arch
,
9297 if (aarch64_tune_string
)
9298 valid_tune
= aarch64_validate_mtune (aarch64_tune_string
, &selected_tune
);
9300 /* If the user did not specify a processor, choose the default
9301 one for them. This will be the CPU set during configuration using
9302 --with-cpu, otherwise it is "generic". */
9307 selected_cpu
= &all_cores
[selected_arch
->ident
];
9308 aarch64_isa_flags
= arch_isa
;
9309 explicit_arch
= selected_arch
->arch
;
9313 /* Get default configure-time CPU. */
9314 selected_cpu
= aarch64_get_tune_cpu (aarch64_none
);
9315 aarch64_isa_flags
= TARGET_CPU_DEFAULT
>> 6;
9319 explicit_tune_core
= selected_tune
->ident
;
9321 /* If both -mcpu and -march are specified check that they are architecturally
9322 compatible, warn if they're not and prefer the -march ISA flags. */
9323 else if (selected_arch
)
9325 if (selected_arch
->arch
!= selected_cpu
->arch
)
9327 warning (0, "switch -mcpu=%s conflicts with -march=%s switch",
9328 all_architectures
[selected_cpu
->arch
].name
,
9329 selected_arch
->name
);
9331 aarch64_isa_flags
= arch_isa
;
9332 explicit_arch
= selected_arch
->arch
;
9333 explicit_tune_core
= selected_tune
? selected_tune
->ident
9334 : selected_cpu
->ident
;
9338 /* -mcpu but no -march. */
9339 aarch64_isa_flags
= cpu_isa
;
9340 explicit_tune_core
= selected_tune
? selected_tune
->ident
9341 : selected_cpu
->ident
;
9342 gcc_assert (selected_cpu
);
9343 selected_arch
= &all_architectures
[selected_cpu
->arch
];
9344 explicit_arch
= selected_arch
->arch
;
9347 /* Set the arch as well as we will need it when outputing
9348 the .arch directive in assembly. */
9351 gcc_assert (selected_cpu
);
9352 selected_arch
= &all_architectures
[selected_cpu
->arch
];
9356 selected_tune
= selected_cpu
;
9358 #ifndef HAVE_AS_MABI_OPTION
9359 /* The compiler may have been configured with 2.23.* binutils, which does
9360 not have support for ILP32. */
9362 error ("Assembler does not support -mabi=ilp32");
9365 if (aarch64_ra_sign_scope
!= AARCH64_FUNCTION_NONE
&& TARGET_ILP32
)
9366 sorry ("Return address signing is only supported for -mabi=lp64");
9368 /* Make sure we properly set up the explicit options. */
9369 if ((aarch64_cpu_string
&& valid_cpu
)
9370 || (aarch64_tune_string
&& valid_tune
))
9371 gcc_assert (explicit_tune_core
!= aarch64_none
);
9373 if ((aarch64_cpu_string
&& valid_cpu
)
9374 || (aarch64_arch_string
&& valid_arch
))
9375 gcc_assert (explicit_arch
!= aarch64_no_arch
);
9377 aarch64_override_options_internal (&global_options
);
9379 /* Save these options as the default ones in case we push and pop them later
9380 while processing functions with potential target attributes. */
9381 target_option_default_node
= target_option_current_node
9382 = build_target_option_node (&global_options
);
9385 /* Implement targetm.override_options_after_change. */
9388 aarch64_override_options_after_change (void)
9390 aarch64_override_options_after_change_1 (&global_options
);
9393 static struct machine_function
*
9394 aarch64_init_machine_status (void)
9396 struct machine_function
*machine
;
9397 machine
= ggc_cleared_alloc
<machine_function
> ();
9402 aarch64_init_expanders (void)
9404 init_machine_status
= aarch64_init_machine_status
;
9407 /* A checking mechanism for the implementation of the various code models. */
9409 initialize_aarch64_code_model (struct gcc_options
*opts
)
9411 if (opts
->x_flag_pic
)
9413 switch (opts
->x_aarch64_cmodel_var
)
9415 case AARCH64_CMODEL_TINY
:
9416 aarch64_cmodel
= AARCH64_CMODEL_TINY_PIC
;
9418 case AARCH64_CMODEL_SMALL
:
9419 #ifdef HAVE_AS_SMALL_PIC_RELOCS
9420 aarch64_cmodel
= (flag_pic
== 2
9421 ? AARCH64_CMODEL_SMALL_PIC
9422 : AARCH64_CMODEL_SMALL_SPIC
);
9424 aarch64_cmodel
= AARCH64_CMODEL_SMALL_PIC
;
9427 case AARCH64_CMODEL_LARGE
:
9428 sorry ("code model %qs with -f%s", "large",
9429 opts
->x_flag_pic
> 1 ? "PIC" : "pic");
9436 aarch64_cmodel
= opts
->x_aarch64_cmodel_var
;
9439 /* Implement TARGET_OPTION_SAVE. */
9442 aarch64_option_save (struct cl_target_option
*ptr
, struct gcc_options
*opts
)
9444 ptr
->x_aarch64_override_tune_string
= opts
->x_aarch64_override_tune_string
;
9447 /* Implements TARGET_OPTION_RESTORE. Restore the backend codegen decisions
9448 using the information saved in PTR. */
9451 aarch64_option_restore (struct gcc_options
*opts
, struct cl_target_option
*ptr
)
9453 opts
->x_explicit_tune_core
= ptr
->x_explicit_tune_core
;
9454 selected_tune
= aarch64_get_tune_cpu (ptr
->x_explicit_tune_core
);
9455 opts
->x_explicit_arch
= ptr
->x_explicit_arch
;
9456 selected_arch
= aarch64_get_arch (ptr
->x_explicit_arch
);
9457 opts
->x_aarch64_override_tune_string
= ptr
->x_aarch64_override_tune_string
;
9459 aarch64_override_options_internal (opts
);
9462 /* Implement TARGET_OPTION_PRINT. */
9465 aarch64_option_print (FILE *file
, int indent
, struct cl_target_option
*ptr
)
9467 const struct processor
*cpu
9468 = aarch64_get_tune_cpu (ptr
->x_explicit_tune_core
);
9469 unsigned long isa_flags
= ptr
->x_aarch64_isa_flags
;
9470 const struct processor
*arch
= aarch64_get_arch (ptr
->x_explicit_arch
);
9471 std::string extension
9472 = aarch64_get_extension_string_for_isa_flags (isa_flags
, arch
->flags
);
9474 fprintf (file
, "%*sselected tune = %s\n", indent
, "", cpu
->name
);
9475 fprintf (file
, "%*sselected arch = %s%s\n", indent
, "",
9476 arch
->name
, extension
.c_str ());
9479 static GTY(()) tree aarch64_previous_fndecl
;
9482 aarch64_reset_previous_fndecl (void)
9484 aarch64_previous_fndecl
= NULL
;
9487 /* Restore or save the TREE_TARGET_GLOBALS from or to NEW_TREE.
9488 Used by aarch64_set_current_function and aarch64_pragma_target_parse to
9489 make sure optab availability predicates are recomputed when necessary. */
9492 aarch64_save_restore_target_globals (tree new_tree
)
9494 if (TREE_TARGET_GLOBALS (new_tree
))
9495 restore_target_globals (TREE_TARGET_GLOBALS (new_tree
));
9496 else if (new_tree
== target_option_default_node
)
9497 restore_target_globals (&default_target_globals
);
9499 TREE_TARGET_GLOBALS (new_tree
) = save_target_globals_default_opts ();
9502 /* Implement TARGET_SET_CURRENT_FUNCTION. Unpack the codegen decisions
9503 like tuning and ISA features from the DECL_FUNCTION_SPECIFIC_TARGET
9504 of the function, if such exists. This function may be called multiple
9505 times on a single function so use aarch64_previous_fndecl to avoid
9506 setting up identical state. */
9509 aarch64_set_current_function (tree fndecl
)
9511 if (!fndecl
|| fndecl
== aarch64_previous_fndecl
)
9514 tree old_tree
= (aarch64_previous_fndecl
9515 ? DECL_FUNCTION_SPECIFIC_TARGET (aarch64_previous_fndecl
)
9518 tree new_tree
= DECL_FUNCTION_SPECIFIC_TARGET (fndecl
);
9520 /* If current function has no attributes but the previous one did,
9521 use the default node. */
9522 if (!new_tree
&& old_tree
)
9523 new_tree
= target_option_default_node
;
9525 /* If nothing to do, return. #pragma GCC reset or #pragma GCC pop to
9526 the default have been handled by aarch64_save_restore_target_globals from
9527 aarch64_pragma_target_parse. */
9528 if (old_tree
== new_tree
)
9531 aarch64_previous_fndecl
= fndecl
;
9533 /* First set the target options. */
9534 cl_target_option_restore (&global_options
, TREE_TARGET_OPTION (new_tree
));
9536 aarch64_save_restore_target_globals (new_tree
);
9539 /* Enum describing the various ways we can handle attributes.
9540 In many cases we can reuse the generic option handling machinery. */
9542 enum aarch64_attr_opt_type
9544 aarch64_attr_mask
, /* Attribute should set a bit in target_flags. */
9545 aarch64_attr_bool
, /* Attribute sets or unsets a boolean variable. */
9546 aarch64_attr_enum
, /* Attribute sets an enum variable. */
9547 aarch64_attr_custom
/* Attribute requires a custom handling function. */
9550 /* All the information needed to handle a target attribute.
9551 NAME is the name of the attribute.
9552 ATTR_TYPE specifies the type of behavior of the attribute as described
9553 in the definition of enum aarch64_attr_opt_type.
9554 ALLOW_NEG is true if the attribute supports a "no-" form.
9555 HANDLER is the function that takes the attribute string and whether
9556 it is a pragma or attribute and handles the option. It is needed only
9557 when the ATTR_TYPE is aarch64_attr_custom.
9558 OPT_NUM is the enum specifying the option that the attribute modifies.
9559 This is needed for attributes that mirror the behavior of a command-line
9560 option, that is it has ATTR_TYPE aarch64_attr_mask, aarch64_attr_bool or
9561 aarch64_attr_enum. */
9563 struct aarch64_attribute_info
9566 enum aarch64_attr_opt_type attr_type
;
9568 bool (*handler
) (const char *, const char *);
9569 enum opt_code opt_num
;
9572 /* Handle the ARCH_STR argument to the arch= target attribute.
9573 PRAGMA_OR_ATTR is used in potential error messages. */
9576 aarch64_handle_attr_arch (const char *str
, const char *pragma_or_attr
)
9578 const struct processor
*tmp_arch
= NULL
;
9579 enum aarch64_parse_opt_result parse_res
9580 = aarch64_parse_arch (str
, &tmp_arch
, &aarch64_isa_flags
);
9582 if (parse_res
== AARCH64_PARSE_OK
)
9584 gcc_assert (tmp_arch
);
9585 selected_arch
= tmp_arch
;
9586 explicit_arch
= selected_arch
->arch
;
9592 case AARCH64_PARSE_MISSING_ARG
:
9593 error ("missing architecture name in 'arch' target %s", pragma_or_attr
);
9595 case AARCH64_PARSE_INVALID_ARG
:
9596 error ("unknown value %qs for 'arch' target %s", str
, pragma_or_attr
);
9597 aarch64_print_hint_for_arch (str
);
9599 case AARCH64_PARSE_INVALID_FEATURE
:
9600 error ("invalid feature modifier %qs for 'arch' target %s",
9601 str
, pragma_or_attr
);
9610 /* Handle the argument CPU_STR to the cpu= target attribute.
9611 PRAGMA_OR_ATTR is used in potential error messages. */
9614 aarch64_handle_attr_cpu (const char *str
, const char *pragma_or_attr
)
9616 const struct processor
*tmp_cpu
= NULL
;
9617 enum aarch64_parse_opt_result parse_res
9618 = aarch64_parse_cpu (str
, &tmp_cpu
, &aarch64_isa_flags
);
9620 if (parse_res
== AARCH64_PARSE_OK
)
9622 gcc_assert (tmp_cpu
);
9623 selected_tune
= tmp_cpu
;
9624 explicit_tune_core
= selected_tune
->ident
;
9626 selected_arch
= &all_architectures
[tmp_cpu
->arch
];
9627 explicit_arch
= selected_arch
->arch
;
9633 case AARCH64_PARSE_MISSING_ARG
:
9634 error ("missing cpu name in 'cpu' target %s", pragma_or_attr
);
9636 case AARCH64_PARSE_INVALID_ARG
:
9637 error ("unknown value %qs for 'cpu' target %s", str
, pragma_or_attr
);
9638 aarch64_print_hint_for_core (str
);
9640 case AARCH64_PARSE_INVALID_FEATURE
:
9641 error ("invalid feature modifier %qs for 'cpu' target %s",
9642 str
, pragma_or_attr
);
9651 /* Handle the argument STR to the tune= target attribute.
9652 PRAGMA_OR_ATTR is used in potential error messages. */
9655 aarch64_handle_attr_tune (const char *str
, const char *pragma_or_attr
)
9657 const struct processor
*tmp_tune
= NULL
;
9658 enum aarch64_parse_opt_result parse_res
9659 = aarch64_parse_tune (str
, &tmp_tune
);
9661 if (parse_res
== AARCH64_PARSE_OK
)
9663 gcc_assert (tmp_tune
);
9664 selected_tune
= tmp_tune
;
9665 explicit_tune_core
= selected_tune
->ident
;
9671 case AARCH64_PARSE_INVALID_ARG
:
9672 error ("unknown value %qs for 'tune' target %s", str
, pragma_or_attr
);
9673 aarch64_print_hint_for_core (str
);
9682 /* Parse an architecture extensions target attribute string specified in STR.
9683 For example "+fp+nosimd". Show any errors if needed. Return TRUE
9684 if successful. Update aarch64_isa_flags to reflect the ISA features
9686 PRAGMA_OR_ATTR is used in potential error messages. */
9689 aarch64_handle_attr_isa_flags (char *str
, const char *pragma_or_attr
)
9691 enum aarch64_parse_opt_result parse_res
;
9692 unsigned long isa_flags
= aarch64_isa_flags
;
9694 /* We allow "+nothing" in the beginning to clear out all architectural
9695 features if the user wants to handpick specific features. */
9696 if (strncmp ("+nothing", str
, 8) == 0)
9702 parse_res
= aarch64_parse_extension (str
, &isa_flags
);
9704 if (parse_res
== AARCH64_PARSE_OK
)
9706 aarch64_isa_flags
= isa_flags
;
9712 case AARCH64_PARSE_MISSING_ARG
:
9713 error ("missing feature modifier in target %s %qs",
9714 pragma_or_attr
, str
);
9717 case AARCH64_PARSE_INVALID_FEATURE
:
9718 error ("invalid feature modifier in target %s %qs",
9719 pragma_or_attr
, str
);
9729 /* The target attributes that we support. On top of these we also support just
9730 ISA extensions, like __attribute__ ((target ("+crc"))), but that case is
9731 handled explicitly in aarch64_process_one_target_attr. */
9733 static const struct aarch64_attribute_info aarch64_attributes
[] =
9735 { "general-regs-only", aarch64_attr_mask
, false, NULL
,
9736 OPT_mgeneral_regs_only
},
9737 { "fix-cortex-a53-835769", aarch64_attr_bool
, true, NULL
,
9738 OPT_mfix_cortex_a53_835769
},
9739 { "fix-cortex-a53-843419", aarch64_attr_bool
, true, NULL
,
9740 OPT_mfix_cortex_a53_843419
},
9741 { "cmodel", aarch64_attr_enum
, false, NULL
, OPT_mcmodel_
},
9742 { "strict-align", aarch64_attr_mask
, false, NULL
, OPT_mstrict_align
},
9743 { "omit-leaf-frame-pointer", aarch64_attr_bool
, true, NULL
,
9744 OPT_momit_leaf_frame_pointer
},
9745 { "tls-dialect", aarch64_attr_enum
, false, NULL
, OPT_mtls_dialect_
},
9746 { "arch", aarch64_attr_custom
, false, aarch64_handle_attr_arch
,
9748 { "cpu", aarch64_attr_custom
, false, aarch64_handle_attr_cpu
, OPT_mcpu_
},
9749 { "tune", aarch64_attr_custom
, false, aarch64_handle_attr_tune
,
9751 { "sign-return-address", aarch64_attr_enum
, false, NULL
,
9752 OPT_msign_return_address_
},
9753 { NULL
, aarch64_attr_custom
, false, NULL
, OPT____
}
9756 /* Parse ARG_STR which contains the definition of one target attribute.
9757 Show appropriate errors if any or return true if the attribute is valid.
9758 PRAGMA_OR_ATTR holds the string to use in error messages about whether
9759 we're processing a target attribute or pragma. */
9762 aarch64_process_one_target_attr (char *arg_str
, const char* pragma_or_attr
)
9764 bool invert
= false;
9766 size_t len
= strlen (arg_str
);
9770 error ("malformed target %s", pragma_or_attr
);
9774 char *str_to_check
= (char *) alloca (len
+ 1);
9775 strcpy (str_to_check
, arg_str
);
9777 /* Skip leading whitespace. */
9778 while (*str_to_check
== ' ' || *str_to_check
== '\t')
9781 /* We have something like __attribute__ ((target ("+fp+nosimd"))).
9782 It is easier to detect and handle it explicitly here rather than going
9783 through the machinery for the rest of the target attributes in this
9785 if (*str_to_check
== '+')
9786 return aarch64_handle_attr_isa_flags (str_to_check
, pragma_or_attr
);
9788 if (len
> 3 && strncmp (str_to_check
, "no-", 3) == 0)
9793 char *arg
= strchr (str_to_check
, '=');
9795 /* If we found opt=foo then terminate STR_TO_CHECK at the '='
9796 and point ARG to "foo". */
9802 const struct aarch64_attribute_info
*p_attr
;
9804 for (p_attr
= aarch64_attributes
; p_attr
->name
; p_attr
++)
9806 /* If the names don't match up, or the user has given an argument
9807 to an attribute that doesn't accept one, or didn't give an argument
9808 to an attribute that expects one, fail to match. */
9809 if (strcmp (str_to_check
, p_attr
->name
) != 0)
9813 bool attr_need_arg_p
= p_attr
->attr_type
== aarch64_attr_custom
9814 || p_attr
->attr_type
== aarch64_attr_enum
;
9816 if (attr_need_arg_p
^ (arg
!= NULL
))
9818 error ("target %s %qs does not accept an argument",
9819 pragma_or_attr
, str_to_check
);
9823 /* If the name matches but the attribute does not allow "no-" versions
9824 then we can't match. */
9825 if (invert
&& !p_attr
->allow_neg
)
9827 error ("target %s %qs does not allow a negated form",
9828 pragma_or_attr
, str_to_check
);
9832 switch (p_attr
->attr_type
)
9834 /* Has a custom handler registered.
9835 For example, cpu=, arch=, tune=. */
9836 case aarch64_attr_custom
:
9837 gcc_assert (p_attr
->handler
);
9838 if (!p_attr
->handler (arg
, pragma_or_attr
))
9842 /* Either set or unset a boolean option. */
9843 case aarch64_attr_bool
:
9845 struct cl_decoded_option decoded
;
9847 generate_option (p_attr
->opt_num
, NULL
, !invert
,
9848 CL_TARGET
, &decoded
);
9849 aarch64_handle_option (&global_options
, &global_options_set
,
9850 &decoded
, input_location
);
9853 /* Set or unset a bit in the target_flags. aarch64_handle_option
9854 should know what mask to apply given the option number. */
9855 case aarch64_attr_mask
:
9857 struct cl_decoded_option decoded
;
9858 /* We only need to specify the option number.
9859 aarch64_handle_option will know which mask to apply. */
9860 decoded
.opt_index
= p_attr
->opt_num
;
9861 decoded
.value
= !invert
;
9862 aarch64_handle_option (&global_options
, &global_options_set
,
9863 &decoded
, input_location
);
9866 /* Use the option setting machinery to set an option to an enum. */
9867 case aarch64_attr_enum
:
9872 valid
= opt_enum_arg_to_value (p_attr
->opt_num
, arg
,
9876 set_option (&global_options
, NULL
, p_attr
->opt_num
, value
,
9877 NULL
, DK_UNSPECIFIED
, input_location
,
9882 error ("target %s %s=%s is not valid",
9883 pragma_or_attr
, str_to_check
, arg
);
9892 /* If we reached here we either have found an attribute and validated
9893 it or didn't match any. If we matched an attribute but its arguments
9894 were malformed we will have returned false already. */
9898 /* Count how many times the character C appears in
9899 NULL-terminated string STR. */
9902 num_occurences_in_str (char c
, char *str
)
9904 unsigned int res
= 0;
9905 while (*str
!= '\0')
9916 /* Parse the tree in ARGS that contains the target attribute information
9917 and update the global target options space. PRAGMA_OR_ATTR is a string
9918 to be used in error messages, specifying whether this is processing
9919 a target attribute or a target pragma. */
9922 aarch64_process_target_attr (tree args
, const char* pragma_or_attr
)
9924 if (TREE_CODE (args
) == TREE_LIST
)
9928 tree head
= TREE_VALUE (args
);
9931 if (!aarch64_process_target_attr (head
, pragma_or_attr
))
9934 args
= TREE_CHAIN (args
);
9940 if (TREE_CODE (args
) != STRING_CST
)
9942 error ("attribute %<target%> argument not a string");
9946 size_t len
= strlen (TREE_STRING_POINTER (args
));
9947 char *str_to_check
= (char *) alloca (len
+ 1);
9948 strcpy (str_to_check
, TREE_STRING_POINTER (args
));
9952 error ("malformed target %s value", pragma_or_attr
);
9956 /* Used to catch empty spaces between commas i.e.
9957 attribute ((target ("attr1,,attr2"))). */
9958 unsigned int num_commas
= num_occurences_in_str (',', str_to_check
);
9960 /* Handle multiple target attributes separated by ','. */
9961 char *token
= strtok (str_to_check
, ",");
9963 unsigned int num_attrs
= 0;
9967 if (!aarch64_process_one_target_attr (token
, pragma_or_attr
))
9969 error ("target %s %qs is invalid", pragma_or_attr
, token
);
9973 token
= strtok (NULL
, ",");
9976 if (num_attrs
!= num_commas
+ 1)
9978 error ("malformed target %s list %qs",
9979 pragma_or_attr
, TREE_STRING_POINTER (args
));
9986 /* Implement TARGET_OPTION_VALID_ATTRIBUTE_P. This is used to
9987 process attribute ((target ("..."))). */
9990 aarch64_option_valid_attribute_p (tree fndecl
, tree
, tree args
, int)
9992 struct cl_target_option cur_target
;
9995 tree new_target
, new_optimize
;
9996 tree existing_target
= DECL_FUNCTION_SPECIFIC_TARGET (fndecl
);
9998 /* If what we're processing is the current pragma string then the
9999 target option node is already stored in target_option_current_node
10000 by aarch64_pragma_target_parse in aarch64-c.c. Use that to avoid
10001 having to re-parse the string. This is especially useful to keep
10002 arm_neon.h compile times down since that header contains a lot
10003 of intrinsics enclosed in pragmas. */
10004 if (!existing_target
&& args
== current_target_pragma
)
10006 DECL_FUNCTION_SPECIFIC_TARGET (fndecl
) = target_option_current_node
;
10009 tree func_optimize
= DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl
);
10011 old_optimize
= build_optimization_node (&global_options
);
10012 func_optimize
= DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl
);
10014 /* If the function changed the optimization levels as well as setting
10015 target options, start with the optimizations specified. */
10016 if (func_optimize
&& func_optimize
!= old_optimize
)
10017 cl_optimization_restore (&global_options
,
10018 TREE_OPTIMIZATION (func_optimize
));
10020 /* Save the current target options to restore at the end. */
10021 cl_target_option_save (&cur_target
, &global_options
);
10023 /* If fndecl already has some target attributes applied to it, unpack
10024 them so that we add this attribute on top of them, rather than
10025 overwriting them. */
10026 if (existing_target
)
10028 struct cl_target_option
*existing_options
10029 = TREE_TARGET_OPTION (existing_target
);
10031 if (existing_options
)
10032 cl_target_option_restore (&global_options
, existing_options
);
10035 cl_target_option_restore (&global_options
,
10036 TREE_TARGET_OPTION (target_option_current_node
));
10039 ret
= aarch64_process_target_attr (args
, "attribute");
10041 /* Set up any additional state. */
10044 aarch64_override_options_internal (&global_options
);
10045 /* Initialize SIMD builtins if we haven't already.
10046 Set current_target_pragma to NULL for the duration so that
10047 the builtin initialization code doesn't try to tag the functions
10048 being built with the attributes specified by any current pragma, thus
10049 going into an infinite recursion. */
10052 tree saved_current_target_pragma
= current_target_pragma
;
10053 current_target_pragma
= NULL
;
10054 aarch64_init_simd_builtins ();
10055 current_target_pragma
= saved_current_target_pragma
;
10057 new_target
= build_target_option_node (&global_options
);
10062 new_optimize
= build_optimization_node (&global_options
);
10066 DECL_FUNCTION_SPECIFIC_TARGET (fndecl
) = new_target
;
10068 if (old_optimize
!= new_optimize
)
10069 DECL_FUNCTION_SPECIFIC_OPTIMIZATION (fndecl
) = new_optimize
;
10072 cl_target_option_restore (&global_options
, &cur_target
);
10074 if (old_optimize
!= new_optimize
)
10075 cl_optimization_restore (&global_options
,
10076 TREE_OPTIMIZATION (old_optimize
));
10080 /* Helper for aarch64_can_inline_p. In the case where CALLER and CALLEE are
10081 tri-bool options (yes, no, don't care) and the default value is
10082 DEF, determine whether to reject inlining. */
10085 aarch64_tribools_ok_for_inlining_p (int caller
, int callee
,
10086 int dont_care
, int def
)
10088 /* If the callee doesn't care, always allow inlining. */
10089 if (callee
== dont_care
)
10092 /* If the caller doesn't care, always allow inlining. */
10093 if (caller
== dont_care
)
10096 /* Otherwise, allow inlining if either the callee and caller values
10097 agree, or if the callee is using the default value. */
10098 return (callee
== caller
|| callee
== def
);
10101 /* Implement TARGET_CAN_INLINE_P. Decide whether it is valid
10102 to inline CALLEE into CALLER based on target-specific info.
10103 Make sure that the caller and callee have compatible architectural
10104 features. Then go through the other possible target attributes
10105 and see if they can block inlining. Try not to reject always_inline
10106 callees unless they are incompatible architecturally. */
10109 aarch64_can_inline_p (tree caller
, tree callee
)
10111 tree caller_tree
= DECL_FUNCTION_SPECIFIC_TARGET (caller
);
10112 tree callee_tree
= DECL_FUNCTION_SPECIFIC_TARGET (callee
);
10114 /* If callee has no option attributes, then it is ok to inline. */
10118 struct cl_target_option
*caller_opts
10119 = TREE_TARGET_OPTION (caller_tree
? caller_tree
10120 : target_option_default_node
);
10122 struct cl_target_option
*callee_opts
= TREE_TARGET_OPTION (callee_tree
);
10125 /* Callee's ISA flags should be a subset of the caller's. */
10126 if ((caller_opts
->x_aarch64_isa_flags
& callee_opts
->x_aarch64_isa_flags
)
10127 != callee_opts
->x_aarch64_isa_flags
)
10130 /* Allow non-strict aligned functions inlining into strict
10132 if ((TARGET_STRICT_ALIGN_P (caller_opts
->x_target_flags
)
10133 != TARGET_STRICT_ALIGN_P (callee_opts
->x_target_flags
))
10134 && !(!TARGET_STRICT_ALIGN_P (callee_opts
->x_target_flags
)
10135 && TARGET_STRICT_ALIGN_P (caller_opts
->x_target_flags
)))
10138 bool always_inline
= lookup_attribute ("always_inline",
10139 DECL_ATTRIBUTES (callee
));
10141 /* If the architectural features match up and the callee is always_inline
10142 then the other attributes don't matter. */
10146 if (caller_opts
->x_aarch64_cmodel_var
10147 != callee_opts
->x_aarch64_cmodel_var
)
10150 if (caller_opts
->x_aarch64_tls_dialect
10151 != callee_opts
->x_aarch64_tls_dialect
)
10154 /* Honour explicit requests to workaround errata. */
10155 if (!aarch64_tribools_ok_for_inlining_p (
10156 caller_opts
->x_aarch64_fix_a53_err835769
,
10157 callee_opts
->x_aarch64_fix_a53_err835769
,
10158 2, TARGET_FIX_ERR_A53_835769_DEFAULT
))
10161 if (!aarch64_tribools_ok_for_inlining_p (
10162 caller_opts
->x_aarch64_fix_a53_err843419
,
10163 callee_opts
->x_aarch64_fix_a53_err843419
,
10164 2, TARGET_FIX_ERR_A53_843419
))
10167 /* If the user explicitly specified -momit-leaf-frame-pointer for the
10168 caller and calle and they don't match up, reject inlining. */
10169 if (!aarch64_tribools_ok_for_inlining_p (
10170 caller_opts
->x_flag_omit_leaf_frame_pointer
,
10171 callee_opts
->x_flag_omit_leaf_frame_pointer
,
10175 /* If the callee has specific tuning overrides, respect them. */
10176 if (callee_opts
->x_aarch64_override_tune_string
!= NULL
10177 && caller_opts
->x_aarch64_override_tune_string
== NULL
)
10180 /* If the user specified tuning override strings for the
10181 caller and callee and they don't match up, reject inlining.
10182 We just do a string compare here, we don't analyze the meaning
10183 of the string, as it would be too costly for little gain. */
10184 if (callee_opts
->x_aarch64_override_tune_string
10185 && caller_opts
->x_aarch64_override_tune_string
10186 && (strcmp (callee_opts
->x_aarch64_override_tune_string
,
10187 caller_opts
->x_aarch64_override_tune_string
) != 0))
10193 /* Return true if SYMBOL_REF X binds locally. */
10196 aarch64_symbol_binds_local_p (const_rtx x
)
10198 return (SYMBOL_REF_DECL (x
)
10199 ? targetm
.binds_local_p (SYMBOL_REF_DECL (x
))
10200 : SYMBOL_REF_LOCAL_P (x
));
10203 /* Return true if SYMBOL_REF X is thread local */
10205 aarch64_tls_symbol_p (rtx x
)
10207 if (! TARGET_HAVE_TLS
)
10210 if (GET_CODE (x
) != SYMBOL_REF
)
10213 return SYMBOL_REF_TLS_MODEL (x
) != 0;
10216 /* Classify a TLS symbol into one of the TLS kinds. */
10217 enum aarch64_symbol_type
10218 aarch64_classify_tls_symbol (rtx x
)
10220 enum tls_model tls_kind
= tls_symbolic_operand_type (x
);
10224 case TLS_MODEL_GLOBAL_DYNAMIC
:
10225 case TLS_MODEL_LOCAL_DYNAMIC
:
10226 return TARGET_TLS_DESC
? SYMBOL_SMALL_TLSDESC
: SYMBOL_SMALL_TLSGD
;
10228 case TLS_MODEL_INITIAL_EXEC
:
10229 switch (aarch64_cmodel
)
10231 case AARCH64_CMODEL_TINY
:
10232 case AARCH64_CMODEL_TINY_PIC
:
10233 return SYMBOL_TINY_TLSIE
;
10235 return SYMBOL_SMALL_TLSIE
;
10238 case TLS_MODEL_LOCAL_EXEC
:
10239 if (aarch64_tls_size
== 12)
10240 return SYMBOL_TLSLE12
;
10241 else if (aarch64_tls_size
== 24)
10242 return SYMBOL_TLSLE24
;
10243 else if (aarch64_tls_size
== 32)
10244 return SYMBOL_TLSLE32
;
10245 else if (aarch64_tls_size
== 48)
10246 return SYMBOL_TLSLE48
;
10248 gcc_unreachable ();
10250 case TLS_MODEL_EMULATED
:
10251 case TLS_MODEL_NONE
:
10252 return SYMBOL_FORCE_TO_MEM
;
10255 gcc_unreachable ();
10259 /* Return the method that should be used to access SYMBOL_REF or
10262 enum aarch64_symbol_type
10263 aarch64_classify_symbol (rtx x
, rtx offset
)
10265 if (GET_CODE (x
) == LABEL_REF
)
10267 switch (aarch64_cmodel
)
10269 case AARCH64_CMODEL_LARGE
:
10270 return SYMBOL_FORCE_TO_MEM
;
10272 case AARCH64_CMODEL_TINY_PIC
:
10273 case AARCH64_CMODEL_TINY
:
10274 return SYMBOL_TINY_ABSOLUTE
;
10276 case AARCH64_CMODEL_SMALL_SPIC
:
10277 case AARCH64_CMODEL_SMALL_PIC
:
10278 case AARCH64_CMODEL_SMALL
:
10279 return SYMBOL_SMALL_ABSOLUTE
;
10282 gcc_unreachable ();
10286 if (GET_CODE (x
) == SYMBOL_REF
)
10288 if (aarch64_tls_symbol_p (x
))
10289 return aarch64_classify_tls_symbol (x
);
10291 switch (aarch64_cmodel
)
10293 case AARCH64_CMODEL_TINY
:
10294 /* When we retrieve symbol + offset address, we have to make sure
10295 the offset does not cause overflow of the final address. But
10296 we have no way of knowing the address of symbol at compile time
10297 so we can't accurately say if the distance between the PC and
10298 symbol + offset is outside the addressible range of +/-1M in the
10299 TINY code model. So we rely on images not being greater than
10300 1M and cap the offset at 1M and anything beyond 1M will have to
10301 be loaded using an alternative mechanism. Furthermore if the
10302 symbol is a weak reference to something that isn't known to
10303 resolve to a symbol in this module, then force to memory. */
10304 if ((SYMBOL_REF_WEAK (x
)
10305 && !aarch64_symbol_binds_local_p (x
))
10306 || INTVAL (offset
) < -1048575 || INTVAL (offset
) > 1048575)
10307 return SYMBOL_FORCE_TO_MEM
;
10308 return SYMBOL_TINY_ABSOLUTE
;
10310 case AARCH64_CMODEL_SMALL
:
10311 /* Same reasoning as the tiny code model, but the offset cap here is
10313 if ((SYMBOL_REF_WEAK (x
)
10314 && !aarch64_symbol_binds_local_p (x
))
10315 || !IN_RANGE (INTVAL (offset
), HOST_WIDE_INT_C (-4294967263),
10316 HOST_WIDE_INT_C (4294967264)))
10317 return SYMBOL_FORCE_TO_MEM
;
10318 return SYMBOL_SMALL_ABSOLUTE
;
10320 case AARCH64_CMODEL_TINY_PIC
:
10321 if (!aarch64_symbol_binds_local_p (x
))
10322 return SYMBOL_TINY_GOT
;
10323 return SYMBOL_TINY_ABSOLUTE
;
10325 case AARCH64_CMODEL_SMALL_SPIC
:
10326 case AARCH64_CMODEL_SMALL_PIC
:
10327 if (!aarch64_symbol_binds_local_p (x
))
10328 return (aarch64_cmodel
== AARCH64_CMODEL_SMALL_SPIC
10329 ? SYMBOL_SMALL_GOT_28K
: SYMBOL_SMALL_GOT_4G
);
10330 return SYMBOL_SMALL_ABSOLUTE
;
10332 case AARCH64_CMODEL_LARGE
:
10333 /* This is alright even in PIC code as the constant
10334 pool reference is always PC relative and within
10335 the same translation unit. */
10336 if (!aarch64_pcrelative_literal_loads
&& CONSTANT_POOL_ADDRESS_P (x
))
10337 return SYMBOL_SMALL_ABSOLUTE
;
10339 return SYMBOL_FORCE_TO_MEM
;
10342 gcc_unreachable ();
10346 /* By default push everything into the constant pool. */
10347 return SYMBOL_FORCE_TO_MEM
;
10351 aarch64_constant_address_p (rtx x
)
10353 return (CONSTANT_P (x
) && memory_address_p (DImode
, x
));
10357 aarch64_legitimate_pic_operand_p (rtx x
)
10359 if (GET_CODE (x
) == SYMBOL_REF
10360 || (GET_CODE (x
) == CONST
10361 && GET_CODE (XEXP (x
, 0)) == PLUS
10362 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == SYMBOL_REF
))
10368 /* Return true if X holds either a quarter-precision or
10369 floating-point +0.0 constant. */
10371 aarch64_valid_floating_const (rtx x
)
10373 if (!CONST_DOUBLE_P (x
))
10376 /* This call determines which constants can be used in mov<mode>
10377 as integer moves instead of constant loads. */
10378 if (aarch64_float_const_rtx_p (x
))
10381 return aarch64_float_const_representable_p (x
);
10385 aarch64_legitimate_constant_p (machine_mode mode
, rtx x
)
10387 /* Do not allow vector struct mode constants. We could support
10388 0 and -1 easily, but they need support in aarch64-simd.md. */
10389 if (TARGET_SIMD
&& aarch64_vect_struct_mode_p (mode
))
10392 /* For these cases we never want to use a literal load.
10393 As such we have to prevent the compiler from forcing these
10395 if ((GET_CODE (x
) == CONST_VECTOR
10396 && aarch64_simd_valid_immediate (x
, mode
, false, NULL
))
10398 || aarch64_valid_floating_const (x
)
10399 || aarch64_can_const_movi_rtx_p (x
, mode
)
10400 || aarch64_float_const_rtx_p (x
))
10401 return !targetm
.cannot_force_const_mem (mode
, x
);
10403 if (GET_CODE (x
) == HIGH
10404 && aarch64_valid_symref (XEXP (x
, 0), GET_MODE (XEXP (x
, 0))))
10407 /* Treat symbols as constants. Avoid TLS symbols as they are complex,
10408 so spilling them is better than rematerialization. */
10409 if (SYMBOL_REF_P (x
) && !SYMBOL_REF_TLS_MODEL (x
))
10412 return aarch64_constant_address_p (x
);
10416 aarch64_load_tp (rtx target
)
10419 || GET_MODE (target
) != Pmode
10420 || !register_operand (target
, Pmode
))
10421 target
= gen_reg_rtx (Pmode
);
10423 /* Can return in any reg. */
10424 emit_insn (gen_aarch64_load_tp_hard (target
));
10428 /* On AAPCS systems, this is the "struct __va_list". */
10429 static GTY(()) tree va_list_type
;
10431 /* Implement TARGET_BUILD_BUILTIN_VA_LIST.
10432 Return the type to use as __builtin_va_list.
10434 AAPCS64 \S 7.1.4 requires that va_list be a typedef for a type defined as:
10446 aarch64_build_builtin_va_list (void)
10449 tree f_stack
, f_grtop
, f_vrtop
, f_groff
, f_vroff
;
10451 /* Create the type. */
10452 va_list_type
= lang_hooks
.types
.make_type (RECORD_TYPE
);
10453 /* Give it the required name. */
10454 va_list_name
= build_decl (BUILTINS_LOCATION
,
10456 get_identifier ("__va_list"),
10458 DECL_ARTIFICIAL (va_list_name
) = 1;
10459 TYPE_NAME (va_list_type
) = va_list_name
;
10460 TYPE_STUB_DECL (va_list_type
) = va_list_name
;
10462 /* Create the fields. */
10463 f_stack
= build_decl (BUILTINS_LOCATION
,
10464 FIELD_DECL
, get_identifier ("__stack"),
10466 f_grtop
= build_decl (BUILTINS_LOCATION
,
10467 FIELD_DECL
, get_identifier ("__gr_top"),
10469 f_vrtop
= build_decl (BUILTINS_LOCATION
,
10470 FIELD_DECL
, get_identifier ("__vr_top"),
10472 f_groff
= build_decl (BUILTINS_LOCATION
,
10473 FIELD_DECL
, get_identifier ("__gr_offs"),
10474 integer_type_node
);
10475 f_vroff
= build_decl (BUILTINS_LOCATION
,
10476 FIELD_DECL
, get_identifier ("__vr_offs"),
10477 integer_type_node
);
10479 /* Tell tree-stdarg pass about our internal offset fields.
10480 NOTE: va_list_gpr/fpr_counter_field are only used for tree comparision
10481 purpose to identify whether the code is updating va_list internal
10482 offset fields through irregular way. */
10483 va_list_gpr_counter_field
= f_groff
;
10484 va_list_fpr_counter_field
= f_vroff
;
10486 DECL_ARTIFICIAL (f_stack
) = 1;
10487 DECL_ARTIFICIAL (f_grtop
) = 1;
10488 DECL_ARTIFICIAL (f_vrtop
) = 1;
10489 DECL_ARTIFICIAL (f_groff
) = 1;
10490 DECL_ARTIFICIAL (f_vroff
) = 1;
10492 DECL_FIELD_CONTEXT (f_stack
) = va_list_type
;
10493 DECL_FIELD_CONTEXT (f_grtop
) = va_list_type
;
10494 DECL_FIELD_CONTEXT (f_vrtop
) = va_list_type
;
10495 DECL_FIELD_CONTEXT (f_groff
) = va_list_type
;
10496 DECL_FIELD_CONTEXT (f_vroff
) = va_list_type
;
10498 TYPE_FIELDS (va_list_type
) = f_stack
;
10499 DECL_CHAIN (f_stack
) = f_grtop
;
10500 DECL_CHAIN (f_grtop
) = f_vrtop
;
10501 DECL_CHAIN (f_vrtop
) = f_groff
;
10502 DECL_CHAIN (f_groff
) = f_vroff
;
10504 /* Compute its layout. */
10505 layout_type (va_list_type
);
10507 return va_list_type
;
10510 /* Implement TARGET_EXPAND_BUILTIN_VA_START. */
10512 aarch64_expand_builtin_va_start (tree valist
, rtx nextarg ATTRIBUTE_UNUSED
)
10514 const CUMULATIVE_ARGS
*cum
;
10515 tree f_stack
, f_grtop
, f_vrtop
, f_groff
, f_vroff
;
10516 tree stack
, grtop
, vrtop
, groff
, vroff
;
10518 int gr_save_area_size
= cfun
->va_list_gpr_size
;
10519 int vr_save_area_size
= cfun
->va_list_fpr_size
;
10522 cum
= &crtl
->args
.info
;
10523 if (cfun
->va_list_gpr_size
)
10524 gr_save_area_size
= MIN ((NUM_ARG_REGS
- cum
->aapcs_ncrn
) * UNITS_PER_WORD
,
10525 cfun
->va_list_gpr_size
);
10526 if (cfun
->va_list_fpr_size
)
10527 vr_save_area_size
= MIN ((NUM_FP_ARG_REGS
- cum
->aapcs_nvrn
)
10528 * UNITS_PER_VREG
, cfun
->va_list_fpr_size
);
10532 gcc_assert (cum
->aapcs_nvrn
== 0);
10533 vr_save_area_size
= 0;
10536 f_stack
= TYPE_FIELDS (va_list_type_node
);
10537 f_grtop
= DECL_CHAIN (f_stack
);
10538 f_vrtop
= DECL_CHAIN (f_grtop
);
10539 f_groff
= DECL_CHAIN (f_vrtop
);
10540 f_vroff
= DECL_CHAIN (f_groff
);
10542 stack
= build3 (COMPONENT_REF
, TREE_TYPE (f_stack
), valist
, f_stack
,
10544 grtop
= build3 (COMPONENT_REF
, TREE_TYPE (f_grtop
), valist
, f_grtop
,
10546 vrtop
= build3 (COMPONENT_REF
, TREE_TYPE (f_vrtop
), valist
, f_vrtop
,
10548 groff
= build3 (COMPONENT_REF
, TREE_TYPE (f_groff
), valist
, f_groff
,
10550 vroff
= build3 (COMPONENT_REF
, TREE_TYPE (f_vroff
), valist
, f_vroff
,
10553 /* Emit code to initialize STACK, which points to the next varargs stack
10554 argument. CUM->AAPCS_STACK_SIZE gives the number of stack words used
10555 by named arguments. STACK is 8-byte aligned. */
10556 t
= make_tree (TREE_TYPE (stack
), virtual_incoming_args_rtx
);
10557 if (cum
->aapcs_stack_size
> 0)
10558 t
= fold_build_pointer_plus_hwi (t
, cum
->aapcs_stack_size
* UNITS_PER_WORD
);
10559 t
= build2 (MODIFY_EXPR
, TREE_TYPE (stack
), stack
, t
);
10560 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
10562 /* Emit code to initialize GRTOP, the top of the GR save area.
10563 virtual_incoming_args_rtx should have been 16 byte aligned. */
10564 t
= make_tree (TREE_TYPE (grtop
), virtual_incoming_args_rtx
);
10565 t
= build2 (MODIFY_EXPR
, TREE_TYPE (grtop
), grtop
, t
);
10566 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
10568 /* Emit code to initialize VRTOP, the top of the VR save area.
10569 This address is gr_save_area_bytes below GRTOP, rounded
10570 down to the next 16-byte boundary. */
10571 t
= make_tree (TREE_TYPE (vrtop
), virtual_incoming_args_rtx
);
10572 vr_offset
= ROUND_UP (gr_save_area_size
,
10573 STACK_BOUNDARY
/ BITS_PER_UNIT
);
10576 t
= fold_build_pointer_plus_hwi (t
, -vr_offset
);
10577 t
= build2 (MODIFY_EXPR
, TREE_TYPE (vrtop
), vrtop
, t
);
10578 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
10580 /* Emit code to initialize GROFF, the offset from GRTOP of the
10581 next GPR argument. */
10582 t
= build2 (MODIFY_EXPR
, TREE_TYPE (groff
), groff
,
10583 build_int_cst (TREE_TYPE (groff
), -gr_save_area_size
));
10584 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
10586 /* Likewise emit code to initialize VROFF, the offset from FTOP
10587 of the next VR argument. */
10588 t
= build2 (MODIFY_EXPR
, TREE_TYPE (vroff
), vroff
,
10589 build_int_cst (TREE_TYPE (vroff
), -vr_save_area_size
));
10590 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
10593 /* Implement TARGET_GIMPLIFY_VA_ARG_EXPR. */
10596 aarch64_gimplify_va_arg_expr (tree valist
, tree type
, gimple_seq
*pre_p
,
10597 gimple_seq
*post_p ATTRIBUTE_UNUSED
)
10601 bool is_ha
; /* is HFA or HVA. */
10602 bool dw_align
; /* double-word align. */
10603 machine_mode ag_mode
= VOIDmode
;
10607 tree f_stack
, f_grtop
, f_vrtop
, f_groff
, f_vroff
;
10608 tree stack
, f_top
, f_off
, off
, arg
, roundup
, on_stack
;
10609 HOST_WIDE_INT size
, rsize
, adjust
, align
;
10610 tree t
, u
, cond1
, cond2
;
10612 indirect_p
= pass_by_reference (NULL
, TYPE_MODE (type
), type
, false);
10614 type
= build_pointer_type (type
);
10616 mode
= TYPE_MODE (type
);
10618 f_stack
= TYPE_FIELDS (va_list_type_node
);
10619 f_grtop
= DECL_CHAIN (f_stack
);
10620 f_vrtop
= DECL_CHAIN (f_grtop
);
10621 f_groff
= DECL_CHAIN (f_vrtop
);
10622 f_vroff
= DECL_CHAIN (f_groff
);
10624 stack
= build3 (COMPONENT_REF
, TREE_TYPE (f_stack
), unshare_expr (valist
),
10625 f_stack
, NULL_TREE
);
10626 size
= int_size_in_bytes (type
);
10627 align
= aarch64_function_arg_alignment (mode
, type
) / BITS_PER_UNIT
;
10631 if (aarch64_vfp_is_call_or_return_candidate (mode
,
10637 /* TYPE passed in fp/simd registers. */
10639 aarch64_err_no_fpadvsimd (mode
, "varargs");
10641 f_top
= build3 (COMPONENT_REF
, TREE_TYPE (f_vrtop
),
10642 unshare_expr (valist
), f_vrtop
, NULL_TREE
);
10643 f_off
= build3 (COMPONENT_REF
, TREE_TYPE (f_vroff
),
10644 unshare_expr (valist
), f_vroff
, NULL_TREE
);
10646 rsize
= nregs
* UNITS_PER_VREG
;
10650 if (BYTES_BIG_ENDIAN
&& GET_MODE_SIZE (ag_mode
) < UNITS_PER_VREG
)
10651 adjust
= UNITS_PER_VREG
- GET_MODE_SIZE (ag_mode
);
10653 else if (BLOCK_REG_PADDING (mode
, type
, 1) == PAD_DOWNWARD
10654 && size
< UNITS_PER_VREG
)
10656 adjust
= UNITS_PER_VREG
- size
;
10661 /* TYPE passed in general registers. */
10662 f_top
= build3 (COMPONENT_REF
, TREE_TYPE (f_grtop
),
10663 unshare_expr (valist
), f_grtop
, NULL_TREE
);
10664 f_off
= build3 (COMPONENT_REF
, TREE_TYPE (f_groff
),
10665 unshare_expr (valist
), f_groff
, NULL_TREE
);
10666 rsize
= ROUND_UP (size
, UNITS_PER_WORD
);
10667 nregs
= rsize
/ UNITS_PER_WORD
;
10672 if (BLOCK_REG_PADDING (mode
, type
, 1) == PAD_DOWNWARD
10673 && size
< UNITS_PER_WORD
)
10675 adjust
= UNITS_PER_WORD
- size
;
10679 /* Get a local temporary for the field value. */
10680 off
= get_initialized_tmp_var (f_off
, pre_p
, NULL
);
10682 /* Emit code to branch if off >= 0. */
10683 t
= build2 (GE_EXPR
, boolean_type_node
, off
,
10684 build_int_cst (TREE_TYPE (off
), 0));
10685 cond1
= build3 (COND_EXPR
, ptr_type_node
, t
, NULL_TREE
, NULL_TREE
);
10689 /* Emit: offs = (offs + 15) & -16. */
10690 t
= build2 (PLUS_EXPR
, TREE_TYPE (off
), off
,
10691 build_int_cst (TREE_TYPE (off
), 15));
10692 t
= build2 (BIT_AND_EXPR
, TREE_TYPE (off
), t
,
10693 build_int_cst (TREE_TYPE (off
), -16));
10694 roundup
= build2 (MODIFY_EXPR
, TREE_TYPE (off
), off
, t
);
10699 /* Update ap.__[g|v]r_offs */
10700 t
= build2 (PLUS_EXPR
, TREE_TYPE (off
), off
,
10701 build_int_cst (TREE_TYPE (off
), rsize
));
10702 t
= build2 (MODIFY_EXPR
, TREE_TYPE (f_off
), unshare_expr (f_off
), t
);
10706 t
= build2 (COMPOUND_EXPR
, TREE_TYPE (t
), roundup
, t
);
10708 /* [cond2] if (ap.__[g|v]r_offs > 0) */
10709 u
= build2 (GT_EXPR
, boolean_type_node
, unshare_expr (f_off
),
10710 build_int_cst (TREE_TYPE (f_off
), 0));
10711 cond2
= build3 (COND_EXPR
, ptr_type_node
, u
, NULL_TREE
, NULL_TREE
);
10713 /* String up: make sure the assignment happens before the use. */
10714 t
= build2 (COMPOUND_EXPR
, TREE_TYPE (cond2
), t
, cond2
);
10715 COND_EXPR_ELSE (cond1
) = t
;
10717 /* Prepare the trees handling the argument that is passed on the stack;
10718 the top level node will store in ON_STACK. */
10719 arg
= get_initialized_tmp_var (stack
, pre_p
, NULL
);
10722 /* if (alignof(type) > 8) (arg = arg + 15) & -16; */
10723 t
= fold_convert (intDI_type_node
, arg
);
10724 t
= build2 (PLUS_EXPR
, TREE_TYPE (t
), t
,
10725 build_int_cst (TREE_TYPE (t
), 15));
10726 t
= build2 (BIT_AND_EXPR
, TREE_TYPE (t
), t
,
10727 build_int_cst (TREE_TYPE (t
), -16));
10728 t
= fold_convert (TREE_TYPE (arg
), t
);
10729 roundup
= build2 (MODIFY_EXPR
, TREE_TYPE (arg
), arg
, t
);
10733 /* Advance ap.__stack */
10734 t
= fold_convert (intDI_type_node
, arg
);
10735 t
= build2 (PLUS_EXPR
, TREE_TYPE (t
), t
,
10736 build_int_cst (TREE_TYPE (t
), size
+ 7));
10737 t
= build2 (BIT_AND_EXPR
, TREE_TYPE (t
), t
,
10738 build_int_cst (TREE_TYPE (t
), -8));
10739 t
= fold_convert (TREE_TYPE (arg
), t
);
10740 t
= build2 (MODIFY_EXPR
, TREE_TYPE (stack
), unshare_expr (stack
), t
);
10741 /* String up roundup and advance. */
10743 t
= build2 (COMPOUND_EXPR
, TREE_TYPE (t
), roundup
, t
);
10744 /* String up with arg */
10745 on_stack
= build2 (COMPOUND_EXPR
, TREE_TYPE (arg
), t
, arg
);
10746 /* Big-endianness related address adjustment. */
10747 if (BLOCK_REG_PADDING (mode
, type
, 1) == PAD_DOWNWARD
10748 && size
< UNITS_PER_WORD
)
10750 t
= build2 (POINTER_PLUS_EXPR
, TREE_TYPE (arg
), arg
,
10751 size_int (UNITS_PER_WORD
- size
));
10752 on_stack
= build2 (COMPOUND_EXPR
, TREE_TYPE (arg
), on_stack
, t
);
10755 COND_EXPR_THEN (cond1
) = unshare_expr (on_stack
);
10756 COND_EXPR_THEN (cond2
) = unshare_expr (on_stack
);
10758 /* Adjustment to OFFSET in the case of BIG_ENDIAN. */
10761 t
= build2 (PREINCREMENT_EXPR
, TREE_TYPE (off
), off
,
10762 build_int_cst (TREE_TYPE (off
), adjust
));
10764 t
= fold_convert (sizetype
, t
);
10765 t
= build2 (POINTER_PLUS_EXPR
, TREE_TYPE (f_top
), f_top
, t
);
10769 /* type ha; // treat as "struct {ftype field[n];}"
10770 ... [computing offs]
10771 for (i = 0; i <nregs; ++i, offs += 16)
10772 ha.field[i] = *((ftype *)(ap.__vr_top + offs));
10775 tree tmp_ha
, field_t
, field_ptr_t
;
10777 /* Declare a local variable. */
10778 tmp_ha
= create_tmp_var_raw (type
, "ha");
10779 gimple_add_tmp_var (tmp_ha
);
10781 /* Establish the base type. */
10785 field_t
= float_type_node
;
10786 field_ptr_t
= float_ptr_type_node
;
10789 field_t
= double_type_node
;
10790 field_ptr_t
= double_ptr_type_node
;
10793 field_t
= long_double_type_node
;
10794 field_ptr_t
= long_double_ptr_type_node
;
10797 field_t
= aarch64_fp16_type_node
;
10798 field_ptr_t
= aarch64_fp16_ptr_type_node
;
10803 tree innertype
= make_signed_type (GET_MODE_PRECISION (SImode
));
10804 field_t
= build_vector_type_for_mode (innertype
, ag_mode
);
10805 field_ptr_t
= build_pointer_type (field_t
);
10812 /* *(field_ptr_t)&ha = *((field_ptr_t)vr_saved_area */
10813 tmp_ha
= build1 (ADDR_EXPR
, field_ptr_t
, tmp_ha
);
10815 t
= fold_convert (field_ptr_t
, addr
);
10816 t
= build2 (MODIFY_EXPR
, field_t
,
10817 build1 (INDIRECT_REF
, field_t
, tmp_ha
),
10818 build1 (INDIRECT_REF
, field_t
, t
));
10820 /* ha.field[i] = *((field_ptr_t)vr_saved_area + i) */
10821 for (i
= 1; i
< nregs
; ++i
)
10823 addr
= fold_build_pointer_plus_hwi (addr
, UNITS_PER_VREG
);
10824 u
= fold_convert (field_ptr_t
, addr
);
10825 u
= build2 (MODIFY_EXPR
, field_t
,
10826 build2 (MEM_REF
, field_t
, tmp_ha
,
10827 build_int_cst (field_ptr_t
,
10829 int_size_in_bytes (field_t
)))),
10830 build1 (INDIRECT_REF
, field_t
, u
));
10831 t
= build2 (COMPOUND_EXPR
, TREE_TYPE (t
), t
, u
);
10834 u
= fold_convert (TREE_TYPE (f_top
), tmp_ha
);
10835 t
= build2 (COMPOUND_EXPR
, TREE_TYPE (f_top
), t
, u
);
10838 COND_EXPR_ELSE (cond2
) = t
;
10839 addr
= fold_convert (build_pointer_type (type
), cond1
);
10840 addr
= build_va_arg_indirect_ref (addr
);
10843 addr
= build_va_arg_indirect_ref (addr
);
10848 /* Implement TARGET_SETUP_INCOMING_VARARGS. */
10851 aarch64_setup_incoming_varargs (cumulative_args_t cum_v
, machine_mode mode
,
10852 tree type
, int *pretend_size ATTRIBUTE_UNUSED
,
10855 CUMULATIVE_ARGS
*cum
= get_cumulative_args (cum_v
);
10856 CUMULATIVE_ARGS local_cum
;
10857 int gr_saved
= cfun
->va_list_gpr_size
;
10858 int vr_saved
= cfun
->va_list_fpr_size
;
10860 /* The caller has advanced CUM up to, but not beyond, the last named
10861 argument. Advance a local copy of CUM past the last "real" named
10862 argument, to find out how many registers are left over. */
10864 aarch64_function_arg_advance (pack_cumulative_args(&local_cum
), mode
, type
, true);
10866 /* Found out how many registers we need to save.
10867 Honor tree-stdvar analysis results. */
10868 if (cfun
->va_list_gpr_size
)
10869 gr_saved
= MIN (NUM_ARG_REGS
- local_cum
.aapcs_ncrn
,
10870 cfun
->va_list_gpr_size
/ UNITS_PER_WORD
);
10871 if (cfun
->va_list_fpr_size
)
10872 vr_saved
= MIN (NUM_FP_ARG_REGS
- local_cum
.aapcs_nvrn
,
10873 cfun
->va_list_fpr_size
/ UNITS_PER_VREG
);
10877 gcc_assert (local_cum
.aapcs_nvrn
== 0);
10887 /* virtual_incoming_args_rtx should have been 16-byte aligned. */
10888 ptr
= plus_constant (Pmode
, virtual_incoming_args_rtx
,
10889 - gr_saved
* UNITS_PER_WORD
);
10890 mem
= gen_frame_mem (BLKmode
, ptr
);
10891 set_mem_alias_set (mem
, get_varargs_alias_set ());
10893 move_block_from_reg (local_cum
.aapcs_ncrn
+ R0_REGNUM
,
10898 /* We can't use move_block_from_reg, because it will use
10899 the wrong mode, storing D regs only. */
10900 machine_mode mode
= TImode
;
10901 int off
, i
, vr_start
;
10903 /* Set OFF to the offset from virtual_incoming_args_rtx of
10904 the first vector register. The VR save area lies below
10905 the GR one, and is aligned to 16 bytes. */
10906 off
= -ROUND_UP (gr_saved
* UNITS_PER_WORD
,
10907 STACK_BOUNDARY
/ BITS_PER_UNIT
);
10908 off
-= vr_saved
* UNITS_PER_VREG
;
10910 vr_start
= V0_REGNUM
+ local_cum
.aapcs_nvrn
;
10911 for (i
= 0; i
< vr_saved
; ++i
)
10915 ptr
= plus_constant (Pmode
, virtual_incoming_args_rtx
, off
);
10916 mem
= gen_frame_mem (mode
, ptr
);
10917 set_mem_alias_set (mem
, get_varargs_alias_set ());
10918 aarch64_emit_move (mem
, gen_rtx_REG (mode
, vr_start
+ i
));
10919 off
+= UNITS_PER_VREG
;
10924 /* We don't save the size into *PRETEND_SIZE because we want to avoid
10925 any complication of having crtl->args.pretend_args_size changed. */
10926 cfun
->machine
->frame
.saved_varargs_size
10927 = (ROUND_UP (gr_saved
* UNITS_PER_WORD
,
10928 STACK_BOUNDARY
/ BITS_PER_UNIT
)
10929 + vr_saved
* UNITS_PER_VREG
);
10933 aarch64_conditional_register_usage (void)
10938 for (i
= V0_REGNUM
; i
<= V31_REGNUM
; i
++)
10941 call_used_regs
[i
] = 1;
10946 /* Walk down the type tree of TYPE counting consecutive base elements.
10947 If *MODEP is VOIDmode, then set it to the first valid floating point
10948 type. If a non-floating point type is found, or if a floating point
10949 type that doesn't match a non-VOIDmode *MODEP is found, then return -1,
10950 otherwise return the count in the sub-tree. */
10952 aapcs_vfp_sub_candidate (const_tree type
, machine_mode
*modep
)
10955 HOST_WIDE_INT size
;
10957 switch (TREE_CODE (type
))
10960 mode
= TYPE_MODE (type
);
10961 if (mode
!= DFmode
&& mode
!= SFmode
10962 && mode
!= TFmode
&& mode
!= HFmode
)
10965 if (*modep
== VOIDmode
)
10968 if (*modep
== mode
)
10974 mode
= TYPE_MODE (TREE_TYPE (type
));
10975 if (mode
!= DFmode
&& mode
!= SFmode
10976 && mode
!= TFmode
&& mode
!= HFmode
)
10979 if (*modep
== VOIDmode
)
10982 if (*modep
== mode
)
10988 /* Use V2SImode and V4SImode as representatives of all 64-bit
10989 and 128-bit vector types. */
10990 size
= int_size_in_bytes (type
);
11003 if (*modep
== VOIDmode
)
11006 /* Vector modes are considered to be opaque: two vectors are
11007 equivalent for the purposes of being homogeneous aggregates
11008 if they are the same size. */
11009 if (*modep
== mode
)
11017 tree index
= TYPE_DOMAIN (type
);
11019 /* Can't handle incomplete types nor sizes that are not
11021 if (!COMPLETE_TYPE_P (type
)
11022 || TREE_CODE (TYPE_SIZE (type
)) != INTEGER_CST
)
11025 count
= aapcs_vfp_sub_candidate (TREE_TYPE (type
), modep
);
11028 || !TYPE_MAX_VALUE (index
)
11029 || !tree_fits_uhwi_p (TYPE_MAX_VALUE (index
))
11030 || !TYPE_MIN_VALUE (index
)
11031 || !tree_fits_uhwi_p (TYPE_MIN_VALUE (index
))
11035 count
*= (1 + tree_to_uhwi (TYPE_MAX_VALUE (index
))
11036 - tree_to_uhwi (TYPE_MIN_VALUE (index
)));
11038 /* There must be no padding. */
11039 if (wi::ne_p (TYPE_SIZE (type
), count
* GET_MODE_BITSIZE (*modep
)))
11051 /* Can't handle incomplete types nor sizes that are not
11053 if (!COMPLETE_TYPE_P (type
)
11054 || TREE_CODE (TYPE_SIZE (type
)) != INTEGER_CST
)
11057 for (field
= TYPE_FIELDS (type
); field
; field
= TREE_CHAIN (field
))
11059 if (TREE_CODE (field
) != FIELD_DECL
)
11062 sub_count
= aapcs_vfp_sub_candidate (TREE_TYPE (field
), modep
);
11065 count
+= sub_count
;
11068 /* There must be no padding. */
11069 if (wi::ne_p (TYPE_SIZE (type
), count
* GET_MODE_BITSIZE (*modep
)))
11076 case QUAL_UNION_TYPE
:
11078 /* These aren't very interesting except in a degenerate case. */
11083 /* Can't handle incomplete types nor sizes that are not
11085 if (!COMPLETE_TYPE_P (type
)
11086 || TREE_CODE (TYPE_SIZE (type
)) != INTEGER_CST
)
11089 for (field
= TYPE_FIELDS (type
); field
; field
= TREE_CHAIN (field
))
11091 if (TREE_CODE (field
) != FIELD_DECL
)
11094 sub_count
= aapcs_vfp_sub_candidate (TREE_TYPE (field
), modep
);
11097 count
= count
> sub_count
? count
: sub_count
;
11100 /* There must be no padding. */
11101 if (wi::ne_p (TYPE_SIZE (type
), count
* GET_MODE_BITSIZE (*modep
)))
11114 /* Return TRUE if the type, as described by TYPE and MODE, is a short vector
11115 type as described in AAPCS64 \S 4.1.2.
11117 See the comment above aarch64_composite_type_p for the notes on MODE. */
11120 aarch64_short_vector_p (const_tree type
,
11123 HOST_WIDE_INT size
= -1;
11125 if (type
&& TREE_CODE (type
) == VECTOR_TYPE
)
11126 size
= int_size_in_bytes (type
);
11127 else if (GET_MODE_CLASS (mode
) == MODE_VECTOR_INT
11128 || GET_MODE_CLASS (mode
) == MODE_VECTOR_FLOAT
)
11129 size
= GET_MODE_SIZE (mode
);
11131 return (size
== 8 || size
== 16);
11134 /* Return TRUE if the type, as described by TYPE and MODE, is a composite
11135 type as described in AAPCS64 \S 4.3. This includes aggregate, union and
11136 array types. The C99 floating-point complex types are also considered
11137 as composite types, according to AAPCS64 \S 7.1.1. The complex integer
11138 types, which are GCC extensions and out of the scope of AAPCS64, are
11139 treated as composite types here as well.
11141 Note that MODE itself is not sufficient in determining whether a type
11142 is such a composite type or not. This is because
11143 stor-layout.c:compute_record_mode may have already changed the MODE
11144 (BLKmode) of a RECORD_TYPE TYPE to some other mode. For example, a
11145 structure with only one field may have its MODE set to the mode of the
11146 field. Also an integer mode whose size matches the size of the
11147 RECORD_TYPE type may be used to substitute the original mode
11148 (i.e. BLKmode) in certain circumstances. In other words, MODE cannot be
11149 solely relied on. */
11152 aarch64_composite_type_p (const_tree type
,
11155 if (aarch64_short_vector_p (type
, mode
))
11158 if (type
&& (AGGREGATE_TYPE_P (type
) || TREE_CODE (type
) == COMPLEX_TYPE
))
11161 if (mode
== BLKmode
11162 || GET_MODE_CLASS (mode
) == MODE_COMPLEX_FLOAT
11163 || GET_MODE_CLASS (mode
) == MODE_COMPLEX_INT
)
11169 /* Return TRUE if an argument, whose type is described by TYPE and MODE,
11170 shall be passed or returned in simd/fp register(s) (providing these
11171 parameter passing registers are available).
11173 Upon successful return, *COUNT returns the number of needed registers,
11174 *BASE_MODE returns the mode of the individual register and when IS_HAF
11175 is not NULL, *IS_HA indicates whether or not the argument is a homogeneous
11176 floating-point aggregate or a homogeneous short-vector aggregate. */
11179 aarch64_vfp_is_call_or_return_candidate (machine_mode mode
,
11181 machine_mode
*base_mode
,
11185 machine_mode new_mode
= VOIDmode
;
11186 bool composite_p
= aarch64_composite_type_p (type
, mode
);
11188 if (is_ha
!= NULL
) *is_ha
= false;
11190 if ((!composite_p
&& GET_MODE_CLASS (mode
) == MODE_FLOAT
)
11191 || aarch64_short_vector_p (type
, mode
))
11196 else if (GET_MODE_CLASS (mode
) == MODE_COMPLEX_FLOAT
)
11198 if (is_ha
!= NULL
) *is_ha
= true;
11200 new_mode
= GET_MODE_INNER (mode
);
11202 else if (type
&& composite_p
)
11204 int ag_count
= aapcs_vfp_sub_candidate (type
, &new_mode
);
11206 if (ag_count
> 0 && ag_count
<= HA_MAX_NUM_FLDS
)
11208 if (is_ha
!= NULL
) *is_ha
= true;
11217 *base_mode
= new_mode
;
11221 /* Implement TARGET_STRUCT_VALUE_RTX. */
11224 aarch64_struct_value_rtx (tree fndecl ATTRIBUTE_UNUSED
,
11225 int incoming ATTRIBUTE_UNUSED
)
11227 return gen_rtx_REG (Pmode
, AARCH64_STRUCT_VALUE_REGNUM
);
11230 /* Implements target hook vector_mode_supported_p. */
11232 aarch64_vector_mode_supported_p (machine_mode mode
)
11235 && (mode
== V4SImode
|| mode
== V8HImode
11236 || mode
== V16QImode
|| mode
== V2DImode
11237 || mode
== V2SImode
|| mode
== V4HImode
11238 || mode
== V8QImode
|| mode
== V2SFmode
11239 || mode
== V4SFmode
|| mode
== V2DFmode
11240 || mode
== V4HFmode
|| mode
== V8HFmode
11241 || mode
== V1DFmode
))
11247 /* Return appropriate SIMD container
11248 for MODE within a vector of WIDTH bits. */
11249 static machine_mode
11250 aarch64_simd_container_mode (scalar_mode mode
, unsigned width
)
11252 gcc_assert (width
== 64 || width
== 128);
11295 /* Return 128-bit container as the preferred SIMD mode for MODE. */
11296 static machine_mode
11297 aarch64_preferred_simd_mode (scalar_mode mode
)
11299 return aarch64_simd_container_mode (mode
, 128);
11302 /* Return the bitmask of possible vector sizes for the vectorizer
11303 to iterate over. */
11304 static unsigned int
11305 aarch64_autovectorize_vector_sizes (void)
11310 /* Implement TARGET_MANGLE_TYPE. */
11312 static const char *
11313 aarch64_mangle_type (const_tree type
)
11315 /* The AArch64 ABI documents say that "__va_list" has to be
11316 managled as if it is in the "std" namespace. */
11317 if (lang_hooks
.types_compatible_p (CONST_CAST_TREE (type
), va_list_type
))
11318 return "St9__va_list";
11320 /* Half-precision float. */
11321 if (TREE_CODE (type
) == REAL_TYPE
&& TYPE_PRECISION (type
) == 16)
11324 /* Mangle AArch64-specific internal types. TYPE_NAME is non-NULL_TREE for
11326 if (TYPE_NAME (type
) != NULL
)
11327 return aarch64_mangle_builtin_type (type
);
11329 /* Use the default mangling. */
11333 /* Find the first rtx_insn before insn that will generate an assembly
11337 aarch64_prev_real_insn (rtx_insn
*insn
)
11344 insn
= prev_real_insn (insn
);
11346 while (insn
&& recog_memoized (insn
) < 0);
11352 is_madd_op (enum attr_type t1
)
11355 /* A number of these may be AArch32 only. */
11356 enum attr_type mlatypes
[] = {
11357 TYPE_MLA
, TYPE_MLAS
, TYPE_SMLAD
, TYPE_SMLADX
, TYPE_SMLAL
, TYPE_SMLALD
,
11358 TYPE_SMLALS
, TYPE_SMLALXY
, TYPE_SMLAWX
, TYPE_SMLAWY
, TYPE_SMLAXY
,
11359 TYPE_SMMLA
, TYPE_UMLAL
, TYPE_UMLALS
,TYPE_SMLSD
, TYPE_SMLSDX
, TYPE_SMLSLD
11362 for (i
= 0; i
< sizeof (mlatypes
) / sizeof (enum attr_type
); i
++)
11364 if (t1
== mlatypes
[i
])
11371 /* Check if there is a register dependency between a load and the insn
11372 for which we hold recog_data. */
11375 dep_between_memop_and_curr (rtx memop
)
11380 gcc_assert (GET_CODE (memop
) == SET
);
11382 if (!REG_P (SET_DEST (memop
)))
11385 load_reg
= SET_DEST (memop
);
11386 for (opno
= 1; opno
< recog_data
.n_operands
; opno
++)
11388 rtx operand
= recog_data
.operand
[opno
];
11389 if (REG_P (operand
)
11390 && reg_overlap_mentioned_p (load_reg
, operand
))
11398 /* When working around the Cortex-A53 erratum 835769,
11399 given rtx_insn INSN, return true if it is a 64-bit multiply-accumulate
11400 instruction and has a preceding memory instruction such that a NOP
11401 should be inserted between them. */
11404 aarch64_madd_needs_nop (rtx_insn
* insn
)
11406 enum attr_type attr_type
;
11410 if (!TARGET_FIX_ERR_A53_835769
)
11413 if (!INSN_P (insn
) || recog_memoized (insn
) < 0)
11416 attr_type
= get_attr_type (insn
);
11417 if (!is_madd_op (attr_type
))
11420 prev
= aarch64_prev_real_insn (insn
);
11421 /* aarch64_prev_real_insn can call recog_memoized on insns other than INSN.
11422 Restore recog state to INSN to avoid state corruption. */
11423 extract_constrain_insn_cached (insn
);
11425 if (!prev
|| !contains_mem_rtx_p (PATTERN (prev
)))
11428 body
= single_set (prev
);
11430 /* If the previous insn is a memory op and there is no dependency between
11431 it and the DImode madd, emit a NOP between them. If body is NULL then we
11432 have a complex memory operation, probably a load/store pair.
11433 Be conservative for now and emit a NOP. */
11434 if (GET_MODE (recog_data
.operand
[0]) == DImode
11435 && (!body
|| !dep_between_memop_and_curr (body
)))
11443 /* Implement FINAL_PRESCAN_INSN. */
11446 aarch64_final_prescan_insn (rtx_insn
*insn
)
11448 if (aarch64_madd_needs_nop (insn
))
11449 fprintf (asm_out_file
, "\tnop // between mem op and mult-accumulate\n");
11453 /* Return the equivalent letter for size. */
11455 sizetochar (int size
)
11459 case 64: return 'd';
11460 case 32: return 's';
11461 case 16: return 'h';
11462 case 8 : return 'b';
11463 default: gcc_unreachable ();
11467 /* Return true iff x is a uniform vector of floating-point
11468 constants, and the constant can be represented in
11469 quarter-precision form. Note, as aarch64_float_const_representable
11470 rejects both +0.0 and -0.0, we will also reject +0.0 and -0.0. */
11472 aarch64_vect_float_const_representable_p (rtx x
)
11475 return (GET_MODE_CLASS (GET_MODE (x
)) == MODE_VECTOR_FLOAT
11476 && const_vec_duplicate_p (x
, &elt
)
11477 && aarch64_float_const_representable_p (elt
));
11480 /* Return true for valid and false for invalid. */
11482 aarch64_simd_valid_immediate (rtx op
, machine_mode mode
, bool inverse
,
11483 struct simd_immediate_info
*info
)
11485 #define CHECK(STRIDE, ELSIZE, CLASS, TEST, SHIFT, NEG) \
11487 for (i = 0; i < idx; i += (STRIDE)) \
11492 immtype = (CLASS); \
11493 elsize = (ELSIZE); \
11494 eshift = (SHIFT); \
11499 unsigned int i
, elsize
= 0, idx
= 0, n_elts
= CONST_VECTOR_NUNITS (op
);
11500 unsigned int innersize
= GET_MODE_UNIT_SIZE (mode
);
11501 unsigned char bytes
[16];
11502 int immtype
= -1, matches
;
11503 unsigned int invmask
= inverse
? 0xff : 0;
11506 if (GET_MODE_CLASS (mode
) == MODE_VECTOR_FLOAT
)
11508 if (! (aarch64_simd_imm_zero_p (op
, mode
)
11509 || aarch64_vect_float_const_representable_p (op
)))
11514 rtx elt
= CONST_VECTOR_ELT (op
, 0);
11515 scalar_float_mode elt_mode
11516 = as_a
<scalar_float_mode
> (GET_MODE (elt
));
11519 info
->element_width
= GET_MODE_BITSIZE (elt_mode
);
11527 /* Splat vector constant out into a byte vector. */
11528 for (i
= 0; i
< n_elts
; i
++)
11530 /* The vector is provided in gcc endian-neutral fashion. For aarch64_be,
11531 it must be laid out in the vector register in reverse order. */
11532 rtx el
= CONST_VECTOR_ELT (op
, BYTES_BIG_ENDIAN
? (n_elts
- 1 - i
) : i
);
11533 unsigned HOST_WIDE_INT elpart
;
11535 gcc_assert (CONST_INT_P (el
));
11536 elpart
= INTVAL (el
);
11538 for (unsigned int byte
= 0; byte
< innersize
; byte
++)
11540 bytes
[idx
++] = (elpart
& 0xff) ^ invmask
;
11541 elpart
>>= BITS_PER_UNIT
;
11546 /* Sanity check. */
11547 gcc_assert (idx
== GET_MODE_SIZE (mode
));
11551 CHECK (4, 32, 0, bytes
[i
] == bytes
[0] && bytes
[i
+ 1] == 0
11552 && bytes
[i
+ 2] == 0 && bytes
[i
+ 3] == 0, 0, 0);
11554 CHECK (4, 32, 1, bytes
[i
] == 0 && bytes
[i
+ 1] == bytes
[1]
11555 && bytes
[i
+ 2] == 0 && bytes
[i
+ 3] == 0, 8, 0);
11557 CHECK (4, 32, 2, bytes
[i
] == 0 && bytes
[i
+ 1] == 0
11558 && bytes
[i
+ 2] == bytes
[2] && bytes
[i
+ 3] == 0, 16, 0);
11560 CHECK (4, 32, 3, bytes
[i
] == 0 && bytes
[i
+ 1] == 0
11561 && bytes
[i
+ 2] == 0 && bytes
[i
+ 3] == bytes
[3], 24, 0);
11563 CHECK (2, 16, 4, bytes
[i
] == bytes
[0] && bytes
[i
+ 1] == 0, 0, 0);
11565 CHECK (2, 16, 5, bytes
[i
] == 0 && bytes
[i
+ 1] == bytes
[1], 8, 0);
11567 CHECK (4, 32, 6, bytes
[i
] == bytes
[0] && bytes
[i
+ 1] == 0xff
11568 && bytes
[i
+ 2] == 0xff && bytes
[i
+ 3] == 0xff, 0, 1);
11570 CHECK (4, 32, 7, bytes
[i
] == 0xff && bytes
[i
+ 1] == bytes
[1]
11571 && bytes
[i
+ 2] == 0xff && bytes
[i
+ 3] == 0xff, 8, 1);
11573 CHECK (4, 32, 8, bytes
[i
] == 0xff && bytes
[i
+ 1] == 0xff
11574 && bytes
[i
+ 2] == bytes
[2] && bytes
[i
+ 3] == 0xff, 16, 1);
11576 CHECK (4, 32, 9, bytes
[i
] == 0xff && bytes
[i
+ 1] == 0xff
11577 && bytes
[i
+ 2] == 0xff && bytes
[i
+ 3] == bytes
[3], 24, 1);
11579 CHECK (2, 16, 10, bytes
[i
] == bytes
[0] && bytes
[i
+ 1] == 0xff, 0, 1);
11581 CHECK (2, 16, 11, bytes
[i
] == 0xff && bytes
[i
+ 1] == bytes
[1], 8, 1);
11583 CHECK (4, 32, 12, bytes
[i
] == 0xff && bytes
[i
+ 1] == bytes
[1]
11584 && bytes
[i
+ 2] == 0 && bytes
[i
+ 3] == 0, 8, 0);
11586 CHECK (4, 32, 13, bytes
[i
] == 0 && bytes
[i
+ 1] == bytes
[1]
11587 && bytes
[i
+ 2] == 0xff && bytes
[i
+ 3] == 0xff, 8, 1);
11589 CHECK (4, 32, 14, bytes
[i
] == 0xff && bytes
[i
+ 1] == 0xff
11590 && bytes
[i
+ 2] == bytes
[2] && bytes
[i
+ 3] == 0, 16, 0);
11592 CHECK (4, 32, 15, bytes
[i
] == 0 && bytes
[i
+ 1] == 0
11593 && bytes
[i
+ 2] == bytes
[2] && bytes
[i
+ 3] == 0xff, 16, 1);
11595 CHECK (1, 8, 16, bytes
[i
] == bytes
[0], 0, 0);
11597 CHECK (1, 64, 17, (bytes
[i
] == 0 || bytes
[i
] == 0xff)
11598 && bytes
[i
] == bytes
[(i
+ 8) % idx
], 0, 0);
11607 info
->element_width
= elsize
;
11608 info
->mvn
= emvn
!= 0;
11609 info
->shift
= eshift
;
11611 unsigned HOST_WIDE_INT imm
= 0;
11613 if (immtype
>= 12 && immtype
<= 15)
11616 /* Un-invert bytes of recognized vector, if necessary. */
11618 for (i
= 0; i
< idx
; i
++)
11619 bytes
[i
] ^= invmask
;
11623 /* FIXME: Broken on 32-bit H_W_I hosts. */
11624 gcc_assert (sizeof (HOST_WIDE_INT
) == 8);
11626 for (i
= 0; i
< 8; i
++)
11627 imm
|= (unsigned HOST_WIDE_INT
) (bytes
[i
] ? 0xff : 0)
11628 << (i
* BITS_PER_UNIT
);
11631 info
->value
= GEN_INT (imm
);
11635 for (i
= 0; i
< elsize
/ BITS_PER_UNIT
; i
++)
11636 imm
|= (unsigned HOST_WIDE_INT
) bytes
[i
] << (i
* BITS_PER_UNIT
);
11638 /* Construct 'abcdefgh' because the assembler cannot handle
11639 generic constants. */
11642 imm
= (imm
>> info
->shift
) & 0xff;
11643 info
->value
= GEN_INT (imm
);
11651 /* Check of immediate shift constants are within range. */
11653 aarch64_simd_shift_imm_p (rtx x
, machine_mode mode
, bool left
)
11655 int bit_width
= GET_MODE_UNIT_SIZE (mode
) * BITS_PER_UNIT
;
11657 return aarch64_const_vec_all_same_in_range_p (x
, 0, bit_width
- 1);
11659 return aarch64_const_vec_all_same_in_range_p (x
, 1, bit_width
);
11662 /* Return true if X is a uniform vector where all elements
11663 are either the floating-point constant 0.0 or the
11664 integer constant 0. */
11666 aarch64_simd_imm_zero_p (rtx x
, machine_mode mode
)
11668 return x
== CONST0_RTX (mode
);
11672 /* Return the bitmask CONST_INT to select the bits required by a zero extract
11673 operation of width WIDTH at bit position POS. */
11676 aarch64_mask_from_zextract_ops (rtx width
, rtx pos
)
11678 gcc_assert (CONST_INT_P (width
));
11679 gcc_assert (CONST_INT_P (pos
));
11681 unsigned HOST_WIDE_INT mask
11682 = ((unsigned HOST_WIDE_INT
) 1 << UINTVAL (width
)) - 1;
11683 return GEN_INT (mask
<< UINTVAL (pos
));
11687 aarch64_mov_operand_p (rtx x
, machine_mode mode
)
11689 if (GET_CODE (x
) == HIGH
11690 && aarch64_valid_symref (XEXP (x
, 0), GET_MODE (XEXP (x
, 0))))
11693 if (CONST_INT_P (x
))
11696 if (GET_CODE (x
) == SYMBOL_REF
&& mode
== DImode
&& CONSTANT_ADDRESS_P (x
))
11699 return aarch64_classify_symbolic_expression (x
)
11700 == SYMBOL_TINY_ABSOLUTE
;
11703 /* Return a const_int vector of VAL. */
11705 aarch64_simd_gen_const_vector_dup (machine_mode mode
, HOST_WIDE_INT val
)
11707 int nunits
= GET_MODE_NUNITS (mode
);
11708 rtvec v
= rtvec_alloc (nunits
);
11711 rtx cache
= GEN_INT (val
);
11713 for (i
=0; i
< nunits
; i
++)
11714 RTVEC_ELT (v
, i
) = cache
;
11716 return gen_rtx_CONST_VECTOR (mode
, v
);
11719 /* Check OP is a legal scalar immediate for the MOVI instruction. */
11722 aarch64_simd_scalar_immediate_valid_for_move (rtx op
, scalar_int_mode mode
)
11724 machine_mode vmode
;
11726 vmode
= aarch64_preferred_simd_mode (mode
);
11727 rtx op_v
= aarch64_simd_gen_const_vector_dup (vmode
, INTVAL (op
));
11728 return aarch64_simd_valid_immediate (op_v
, vmode
, false, NULL
);
11731 /* Construct and return a PARALLEL RTX vector with elements numbering the
11732 lanes of either the high (HIGH == TRUE) or low (HIGH == FALSE) half of
11733 the vector - from the perspective of the architecture. This does not
11734 line up with GCC's perspective on lane numbers, so we end up with
11735 different masks depending on our target endian-ness. The diagram
11736 below may help. We must draw the distinction when building masks
11737 which select one half of the vector. An instruction selecting
11738 architectural low-lanes for a big-endian target, must be described using
11739 a mask selecting GCC high-lanes.
11741 Big-Endian Little-Endian
11743 GCC 0 1 2 3 3 2 1 0
11744 | x | x | x | x | | x | x | x | x |
11745 Architecture 3 2 1 0 3 2 1 0
11747 Low Mask: { 2, 3 } { 0, 1 }
11748 High Mask: { 0, 1 } { 2, 3 }
11752 aarch64_simd_vect_par_cnst_half (machine_mode mode
, bool high
)
11754 int nunits
= GET_MODE_NUNITS (mode
);
11755 rtvec v
= rtvec_alloc (nunits
/ 2);
11756 int high_base
= nunits
/ 2;
11762 if (BYTES_BIG_ENDIAN
)
11763 base
= high
? low_base
: high_base
;
11765 base
= high
? high_base
: low_base
;
11767 for (i
= 0; i
< nunits
/ 2; i
++)
11768 RTVEC_ELT (v
, i
) = GEN_INT (base
+ i
);
11770 t1
= gen_rtx_PARALLEL (mode
, v
);
11774 /* Check OP for validity as a PARALLEL RTX vector with elements
11775 numbering the lanes of either the high (HIGH == TRUE) or low lanes,
11776 from the perspective of the architecture. See the diagram above
11777 aarch64_simd_vect_par_cnst_half for more details. */
11780 aarch64_simd_check_vect_par_cnst_half (rtx op
, machine_mode mode
,
11783 rtx ideal
= aarch64_simd_vect_par_cnst_half (mode
, high
);
11784 HOST_WIDE_INT count_op
= XVECLEN (op
, 0);
11785 HOST_WIDE_INT count_ideal
= XVECLEN (ideal
, 0);
11788 if (!VECTOR_MODE_P (mode
))
11791 if (count_op
!= count_ideal
)
11794 for (i
= 0; i
< count_ideal
; i
++)
11796 rtx elt_op
= XVECEXP (op
, 0, i
);
11797 rtx elt_ideal
= XVECEXP (ideal
, 0, i
);
11799 if (!CONST_INT_P (elt_op
)
11800 || INTVAL (elt_ideal
) != INTVAL (elt_op
))
11806 /* Bounds-check lanes. Ensure OPERAND lies between LOW (inclusive) and
11807 HIGH (exclusive). */
11809 aarch64_simd_lane_bounds (rtx operand
, HOST_WIDE_INT low
, HOST_WIDE_INT high
,
11812 HOST_WIDE_INT lane
;
11813 gcc_assert (CONST_INT_P (operand
));
11814 lane
= INTVAL (operand
);
11816 if (lane
< low
|| lane
>= high
)
11819 error ("%Klane %wd out of range %wd - %wd", exp
, lane
, low
, high
- 1);
11821 error ("lane %wd out of range %wd - %wd", lane
, low
, high
- 1);
11825 /* Return TRUE if OP is a valid vector addressing mode. */
11827 aarch64_simd_mem_operand_p (rtx op
)
11829 return MEM_P (op
) && (GET_CODE (XEXP (op
, 0)) == POST_INC
11830 || REG_P (XEXP (op
, 0)));
11833 /* Emit a register copy from operand to operand, taking care not to
11834 early-clobber source registers in the process.
11836 COUNT is the number of components into which the copy needs to be
11839 aarch64_simd_emit_reg_reg_move (rtx
*operands
, machine_mode mode
,
11840 unsigned int count
)
11843 int rdest
= REGNO (operands
[0]);
11844 int rsrc
= REGNO (operands
[1]);
11846 if (!reg_overlap_mentioned_p (operands
[0], operands
[1])
11848 for (i
= 0; i
< count
; i
++)
11849 emit_move_insn (gen_rtx_REG (mode
, rdest
+ i
),
11850 gen_rtx_REG (mode
, rsrc
+ i
));
11852 for (i
= 0; i
< count
; i
++)
11853 emit_move_insn (gen_rtx_REG (mode
, rdest
+ count
- i
- 1),
11854 gen_rtx_REG (mode
, rsrc
+ count
- i
- 1));
11857 /* Compute and return the length of aarch64_simd_reglist<mode>, where <mode> is
11858 one of VSTRUCT modes: OI, CI, or XI. */
11860 aarch64_simd_attr_length_rglist (machine_mode mode
)
11862 return (GET_MODE_SIZE (mode
) / UNITS_PER_VREG
) * 4;
11865 /* Implement target hook TARGET_VECTOR_ALIGNMENT. The AAPCS64 sets the maximum
11866 alignment of a vector to 128 bits. */
11867 static HOST_WIDE_INT
11868 aarch64_simd_vector_alignment (const_tree type
)
11870 HOST_WIDE_INT align
= tree_to_shwi (TYPE_SIZE (type
));
11871 return MIN (align
, 128);
11874 /* Implement target hook TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE. */
11876 aarch64_simd_vector_alignment_reachable (const_tree type
, bool is_packed
)
11881 /* We guarantee alignment for vectors up to 128-bits. */
11882 if (tree_int_cst_compare (TYPE_SIZE (type
),
11883 bitsize_int (BIGGEST_ALIGNMENT
)) > 0)
11886 /* Vectors whose size is <= BIGGEST_ALIGNMENT are naturally aligned. */
11890 /* Return true if the vector misalignment factor is supported by the
11893 aarch64_builtin_support_vector_misalignment (machine_mode mode
,
11894 const_tree type
, int misalignment
,
11897 if (TARGET_SIMD
&& STRICT_ALIGNMENT
)
11899 /* Return if movmisalign pattern is not supported for this mode. */
11900 if (optab_handler (movmisalign_optab
, mode
) == CODE_FOR_nothing
)
11903 if (misalignment
== -1)
11905 /* Misalignment factor is unknown at compile time but we know
11906 it's word aligned. */
11907 if (aarch64_simd_vector_alignment_reachable (type
, is_packed
))
11909 int element_size
= TREE_INT_CST_LOW (TYPE_SIZE (type
));
11911 if (element_size
!= 64)
11917 return default_builtin_support_vector_misalignment (mode
, type
, misalignment
,
11921 /* If VALS is a vector constant that can be loaded into a register
11922 using DUP, generate instructions to do so and return an RTX to
11923 assign to the register. Otherwise return NULL_RTX. */
11925 aarch64_simd_dup_constant (rtx vals
)
11927 machine_mode mode
= GET_MODE (vals
);
11928 machine_mode inner_mode
= GET_MODE_INNER (mode
);
11931 if (!const_vec_duplicate_p (vals
, &x
))
11934 /* We can load this constant by using DUP and a constant in a
11935 single ARM register. This will be cheaper than a vector
11937 x
= copy_to_mode_reg (inner_mode
, x
);
11938 return gen_rtx_VEC_DUPLICATE (mode
, x
);
11942 /* Generate code to load VALS, which is a PARALLEL containing only
11943 constants (for vec_init) or CONST_VECTOR, efficiently into a
11944 register. Returns an RTX to copy into the register, or NULL_RTX
11945 for a PARALLEL that can not be converted into a CONST_VECTOR. */
11947 aarch64_simd_make_constant (rtx vals
)
11949 machine_mode mode
= GET_MODE (vals
);
11951 rtx const_vec
= NULL_RTX
;
11952 int n_elts
= GET_MODE_NUNITS (mode
);
11956 if (GET_CODE (vals
) == CONST_VECTOR
)
11958 else if (GET_CODE (vals
) == PARALLEL
)
11960 /* A CONST_VECTOR must contain only CONST_INTs and
11961 CONST_DOUBLEs, but CONSTANT_P allows more (e.g. SYMBOL_REF).
11962 Only store valid constants in a CONST_VECTOR. */
11963 for (i
= 0; i
< n_elts
; ++i
)
11965 rtx x
= XVECEXP (vals
, 0, i
);
11966 if (CONST_INT_P (x
) || CONST_DOUBLE_P (x
))
11969 if (n_const
== n_elts
)
11970 const_vec
= gen_rtx_CONST_VECTOR (mode
, XVEC (vals
, 0));
11973 gcc_unreachable ();
11975 if (const_vec
!= NULL_RTX
11976 && aarch64_simd_valid_immediate (const_vec
, mode
, false, NULL
))
11977 /* Load using MOVI/MVNI. */
11979 else if ((const_dup
= aarch64_simd_dup_constant (vals
)) != NULL_RTX
)
11980 /* Loaded using DUP. */
11982 else if (const_vec
!= NULL_RTX
)
11983 /* Load from constant pool. We can not take advantage of single-cycle
11984 LD1 because we need a PC-relative addressing mode. */
11987 /* A PARALLEL containing something not valid inside CONST_VECTOR.
11988 We can not construct an initializer. */
11992 /* Expand a vector initialisation sequence, such that TARGET is
11993 initialised to contain VALS. */
11996 aarch64_expand_vector_init (rtx target
, rtx vals
)
11998 machine_mode mode
= GET_MODE (target
);
11999 scalar_mode inner_mode
= GET_MODE_INNER (mode
);
12000 /* The number of vector elements. */
12001 int n_elts
= GET_MODE_NUNITS (mode
);
12002 /* The number of vector elements which are not constant. */
12004 rtx any_const
= NULL_RTX
;
12005 /* The first element of vals. */
12006 rtx v0
= XVECEXP (vals
, 0, 0);
12007 bool all_same
= true;
12009 /* Count the number of variable elements to initialise. */
12010 for (int i
= 0; i
< n_elts
; ++i
)
12012 rtx x
= XVECEXP (vals
, 0, i
);
12013 if (!(CONST_INT_P (x
) || CONST_DOUBLE_P (x
)))
12018 all_same
&= rtx_equal_p (x
, v0
);
12021 /* No variable elements, hand off to aarch64_simd_make_constant which knows
12022 how best to handle this. */
12025 rtx constant
= aarch64_simd_make_constant (vals
);
12026 if (constant
!= NULL_RTX
)
12028 emit_move_insn (target
, constant
);
12033 /* Splat a single non-constant element if we can. */
12036 rtx x
= copy_to_mode_reg (inner_mode
, v0
);
12037 aarch64_emit_move (target
, gen_rtx_VEC_DUPLICATE (mode
, x
));
12041 enum insn_code icode
= optab_handler (vec_set_optab
, mode
);
12042 gcc_assert (icode
!= CODE_FOR_nothing
);
12044 /* If there are only variable elements, try to optimize
12045 the insertion using dup for the most common element
12046 followed by insertions. */
12048 /* The algorithm will fill matches[*][0] with the earliest matching element,
12049 and matches[X][1] with the count of duplicate elements (if X is the
12050 earliest element which has duplicates). */
12052 if (n_var
== n_elts
&& n_elts
<= 16)
12054 int matches
[16][2] = {0};
12055 for (int i
= 0; i
< n_elts
; i
++)
12057 for (int j
= 0; j
<= i
; j
++)
12059 if (rtx_equal_p (XVECEXP (vals
, 0, i
), XVECEXP (vals
, 0, j
)))
12067 int maxelement
= 0;
12069 for (int i
= 0; i
< n_elts
; i
++)
12070 if (matches
[i
][1] > maxv
)
12073 maxv
= matches
[i
][1];
12076 /* Create a duplicate of the most common element. */
12077 rtx x
= copy_to_mode_reg (inner_mode
, XVECEXP (vals
, 0, maxelement
));
12078 aarch64_emit_move (target
, gen_rtx_VEC_DUPLICATE (mode
, x
));
12080 /* Insert the rest. */
12081 for (int i
= 0; i
< n_elts
; i
++)
12083 rtx x
= XVECEXP (vals
, 0, i
);
12084 if (matches
[i
][0] == maxelement
)
12086 x
= copy_to_mode_reg (inner_mode
, x
);
12087 emit_insn (GEN_FCN (icode
) (target
, x
, GEN_INT (i
)));
12092 /* Initialise a vector which is part-variable. We want to first try
12093 to build those lanes which are constant in the most efficient way we
12095 if (n_var
!= n_elts
)
12097 rtx copy
= copy_rtx (vals
);
12099 /* Load constant part of vector. We really don't care what goes into the
12100 parts we will overwrite, but we're more likely to be able to load the
12101 constant efficiently if it has fewer, larger, repeating parts
12102 (see aarch64_simd_valid_immediate). */
12103 for (int i
= 0; i
< n_elts
; i
++)
12105 rtx x
= XVECEXP (vals
, 0, i
);
12106 if (CONST_INT_P (x
) || CONST_DOUBLE_P (x
))
12108 rtx subst
= any_const
;
12109 for (int bit
= n_elts
/ 2; bit
> 0; bit
/= 2)
12111 /* Look in the copied vector, as more elements are const. */
12112 rtx test
= XVECEXP (copy
, 0, i
^ bit
);
12113 if (CONST_INT_P (test
) || CONST_DOUBLE_P (test
))
12119 XVECEXP (copy
, 0, i
) = subst
;
12121 aarch64_expand_vector_init (target
, copy
);
12124 /* Insert the variable lanes directly. */
12125 for (int i
= 0; i
< n_elts
; i
++)
12127 rtx x
= XVECEXP (vals
, 0, i
);
12128 if (CONST_INT_P (x
) || CONST_DOUBLE_P (x
))
12130 x
= copy_to_mode_reg (inner_mode
, x
);
12131 emit_insn (GEN_FCN (icode
) (target
, x
, GEN_INT (i
)));
12135 static unsigned HOST_WIDE_INT
12136 aarch64_shift_truncation_mask (machine_mode mode
)
12139 (!SHIFT_COUNT_TRUNCATED
12140 || aarch64_vector_mode_supported_p (mode
)
12141 || aarch64_vect_struct_mode_p (mode
)) ? 0 : (GET_MODE_BITSIZE (mode
) - 1);
12144 /* Select a format to encode pointers in exception handling data. */
12146 aarch64_asm_preferred_eh_data_format (int code ATTRIBUTE_UNUSED
, int global
)
12149 switch (aarch64_cmodel
)
12151 case AARCH64_CMODEL_TINY
:
12152 case AARCH64_CMODEL_TINY_PIC
:
12153 case AARCH64_CMODEL_SMALL
:
12154 case AARCH64_CMODEL_SMALL_PIC
:
12155 case AARCH64_CMODEL_SMALL_SPIC
:
12156 /* text+got+data < 4Gb. 4-byte signed relocs are sufficient
12158 type
= DW_EH_PE_sdata4
;
12161 /* No assumptions here. 8-byte relocs required. */
12162 type
= DW_EH_PE_sdata8
;
12165 return (global
? DW_EH_PE_indirect
: 0) | DW_EH_PE_pcrel
| type
;
12168 /* The last .arch and .tune assembly strings that we printed. */
12169 static std::string aarch64_last_printed_arch_string
;
12170 static std::string aarch64_last_printed_tune_string
;
12172 /* Implement ASM_DECLARE_FUNCTION_NAME. Output the ISA features used
12173 by the function fndecl. */
12176 aarch64_declare_function_name (FILE *stream
, const char* name
,
12179 tree target_parts
= DECL_FUNCTION_SPECIFIC_TARGET (fndecl
);
12181 struct cl_target_option
*targ_options
;
12183 targ_options
= TREE_TARGET_OPTION (target_parts
);
12185 targ_options
= TREE_TARGET_OPTION (target_option_current_node
);
12186 gcc_assert (targ_options
);
12188 const struct processor
*this_arch
12189 = aarch64_get_arch (targ_options
->x_explicit_arch
);
12191 unsigned long isa_flags
= targ_options
->x_aarch64_isa_flags
;
12192 std::string extension
12193 = aarch64_get_extension_string_for_isa_flags (isa_flags
,
12195 /* Only update the assembler .arch string if it is distinct from the last
12196 such string we printed. */
12197 std::string to_print
= this_arch
->name
+ extension
;
12198 if (to_print
!= aarch64_last_printed_arch_string
)
12200 asm_fprintf (asm_out_file
, "\t.arch %s\n", to_print
.c_str ());
12201 aarch64_last_printed_arch_string
= to_print
;
12204 /* Print the cpu name we're tuning for in the comments, might be
12205 useful to readers of the generated asm. Do it only when it changes
12206 from function to function and verbose assembly is requested. */
12207 const struct processor
*this_tune
12208 = aarch64_get_tune_cpu (targ_options
->x_explicit_tune_core
);
12210 if (flag_debug_asm
&& aarch64_last_printed_tune_string
!= this_tune
->name
)
12212 asm_fprintf (asm_out_file
, "\t" ASM_COMMENT_START
".tune %s\n",
12214 aarch64_last_printed_tune_string
= this_tune
->name
;
12217 /* Don't forget the type directive for ELF. */
12218 ASM_OUTPUT_TYPE_DIRECTIVE (stream
, name
, "function");
12219 ASM_OUTPUT_LABEL (stream
, name
);
12222 /* Implements TARGET_ASM_FILE_START. Output the assembly header. */
12225 aarch64_start_file (void)
12227 struct cl_target_option
*default_options
12228 = TREE_TARGET_OPTION (target_option_default_node
);
12230 const struct processor
*default_arch
12231 = aarch64_get_arch (default_options
->x_explicit_arch
);
12232 unsigned long default_isa_flags
= default_options
->x_aarch64_isa_flags
;
12233 std::string extension
12234 = aarch64_get_extension_string_for_isa_flags (default_isa_flags
,
12235 default_arch
->flags
);
12237 aarch64_last_printed_arch_string
= default_arch
->name
+ extension
;
12238 aarch64_last_printed_tune_string
= "";
12239 asm_fprintf (asm_out_file
, "\t.arch %s\n",
12240 aarch64_last_printed_arch_string
.c_str ());
12242 default_file_start ();
12245 /* Emit load exclusive. */
12248 aarch64_emit_load_exclusive (machine_mode mode
, rtx rval
,
12249 rtx mem
, rtx model_rtx
)
12251 rtx (*gen
) (rtx
, rtx
, rtx
);
12255 case E_QImode
: gen
= gen_aarch64_load_exclusiveqi
; break;
12256 case E_HImode
: gen
= gen_aarch64_load_exclusivehi
; break;
12257 case E_SImode
: gen
= gen_aarch64_load_exclusivesi
; break;
12258 case E_DImode
: gen
= gen_aarch64_load_exclusivedi
; break;
12260 gcc_unreachable ();
12263 emit_insn (gen (rval
, mem
, model_rtx
));
12266 /* Emit store exclusive. */
12269 aarch64_emit_store_exclusive (machine_mode mode
, rtx bval
,
12270 rtx rval
, rtx mem
, rtx model_rtx
)
12272 rtx (*gen
) (rtx
, rtx
, rtx
, rtx
);
12276 case E_QImode
: gen
= gen_aarch64_store_exclusiveqi
; break;
12277 case E_HImode
: gen
= gen_aarch64_store_exclusivehi
; break;
12278 case E_SImode
: gen
= gen_aarch64_store_exclusivesi
; break;
12279 case E_DImode
: gen
= gen_aarch64_store_exclusivedi
; break;
12281 gcc_unreachable ();
12284 emit_insn (gen (bval
, rval
, mem
, model_rtx
));
12287 /* Mark the previous jump instruction as unlikely. */
12290 aarch64_emit_unlikely_jump (rtx insn
)
12292 rtx_insn
*jump
= emit_jump_insn (insn
);
12293 add_reg_br_prob_note (jump
, profile_probability::very_unlikely ());
12296 /* Expand a compare and swap pattern. */
12299 aarch64_expand_compare_and_swap (rtx operands
[])
12301 rtx bval
, rval
, mem
, oldval
, newval
, is_weak
, mod_s
, mod_f
, x
;
12302 machine_mode mode
, cmp_mode
;
12303 typedef rtx (*gen_cas_fn
) (rtx
, rtx
, rtx
, rtx
, rtx
, rtx
, rtx
);
12306 const gen_cas_fn split_cas
[] =
12308 gen_aarch64_compare_and_swapqi
,
12309 gen_aarch64_compare_and_swaphi
,
12310 gen_aarch64_compare_and_swapsi
,
12311 gen_aarch64_compare_and_swapdi
12313 const gen_cas_fn atomic_cas
[] =
12315 gen_aarch64_compare_and_swapqi_lse
,
12316 gen_aarch64_compare_and_swaphi_lse
,
12317 gen_aarch64_compare_and_swapsi_lse
,
12318 gen_aarch64_compare_and_swapdi_lse
12321 bval
= operands
[0];
12322 rval
= operands
[1];
12324 oldval
= operands
[3];
12325 newval
= operands
[4];
12326 is_weak
= operands
[5];
12327 mod_s
= operands
[6];
12328 mod_f
= operands
[7];
12329 mode
= GET_MODE (mem
);
12332 /* Normally the succ memory model must be stronger than fail, but in the
12333 unlikely event of fail being ACQUIRE and succ being RELEASE we need to
12334 promote succ to ACQ_REL so that we don't lose the acquire semantics. */
12336 if (is_mm_acquire (memmodel_from_int (INTVAL (mod_f
)))
12337 && is_mm_release (memmodel_from_int (INTVAL (mod_s
))))
12338 mod_s
= GEN_INT (MEMMODEL_ACQ_REL
);
12344 /* For short modes, we're going to perform the comparison in SImode,
12345 so do the zero-extension now. */
12347 rval
= gen_reg_rtx (SImode
);
12348 oldval
= convert_modes (SImode
, mode
, oldval
, true);
12349 /* Fall through. */
12353 /* Force the value into a register if needed. */
12354 if (!aarch64_plus_operand (oldval
, mode
))
12355 oldval
= force_reg (cmp_mode
, oldval
);
12359 gcc_unreachable ();
12364 case E_QImode
: idx
= 0; break;
12365 case E_HImode
: idx
= 1; break;
12366 case E_SImode
: idx
= 2; break;
12367 case E_DImode
: idx
= 3; break;
12369 gcc_unreachable ();
12372 gen
= atomic_cas
[idx
];
12374 gen
= split_cas
[idx
];
12376 emit_insn (gen (rval
, mem
, oldval
, newval
, is_weak
, mod_s
, mod_f
));
12378 if (mode
== QImode
|| mode
== HImode
)
12379 emit_move_insn (operands
[1], gen_lowpart (mode
, rval
));
12381 x
= gen_rtx_REG (CCmode
, CC_REGNUM
);
12382 x
= gen_rtx_EQ (SImode
, x
, const0_rtx
);
12383 emit_insn (gen_rtx_SET (bval
, x
));
12386 /* Test whether the target supports using a atomic load-operate instruction.
12387 CODE is the operation and AFTER is TRUE if the data in memory after the
12388 operation should be returned and FALSE if the data before the operation
12389 should be returned. Returns FALSE if the operation isn't supported by the
12393 aarch64_atomic_ldop_supported_p (enum rtx_code code
)
12412 /* Emit a barrier, that is appropriate for memory model MODEL, at the end of a
12413 sequence implementing an atomic operation. */
12416 aarch64_emit_post_barrier (enum memmodel model
)
12418 const enum memmodel base_model
= memmodel_base (model
);
12420 if (is_mm_sync (model
)
12421 && (base_model
== MEMMODEL_ACQUIRE
12422 || base_model
== MEMMODEL_ACQ_REL
12423 || base_model
== MEMMODEL_SEQ_CST
))
12425 emit_insn (gen_mem_thread_fence (GEN_INT (MEMMODEL_SEQ_CST
)));
12429 /* Emit an atomic compare-and-swap operation. RVAL is the destination register
12430 for the data in memory. EXPECTED is the value expected to be in memory.
12431 DESIRED is the value to store to memory. MEM is the memory location. MODEL
12432 is the memory ordering to use. */
12435 aarch64_gen_atomic_cas (rtx rval
, rtx mem
,
12436 rtx expected
, rtx desired
,
12439 rtx (*gen
) (rtx
, rtx
, rtx
, rtx
);
12442 mode
= GET_MODE (mem
);
12446 case E_QImode
: gen
= gen_aarch64_atomic_casqi
; break;
12447 case E_HImode
: gen
= gen_aarch64_atomic_cashi
; break;
12448 case E_SImode
: gen
= gen_aarch64_atomic_cassi
; break;
12449 case E_DImode
: gen
= gen_aarch64_atomic_casdi
; break;
12451 gcc_unreachable ();
12454 /* Move the expected value into the CAS destination register. */
12455 emit_insn (gen_rtx_SET (rval
, expected
));
12457 /* Emit the CAS. */
12458 emit_insn (gen (rval
, mem
, desired
, model
));
12460 /* Compare the expected value with the value loaded by the CAS, to establish
12461 whether the swap was made. */
12462 aarch64_gen_compare_reg (EQ
, rval
, expected
);
12465 /* Split a compare and swap pattern. */
12468 aarch64_split_compare_and_swap (rtx operands
[])
12470 rtx rval
, mem
, oldval
, newval
, scratch
;
12473 rtx_code_label
*label1
, *label2
;
12475 enum memmodel model
;
12478 rval
= operands
[0];
12480 oldval
= operands
[2];
12481 newval
= operands
[3];
12482 is_weak
= (operands
[4] != const0_rtx
);
12483 model_rtx
= operands
[5];
12484 scratch
= operands
[7];
12485 mode
= GET_MODE (mem
);
12486 model
= memmodel_from_int (INTVAL (model_rtx
));
12488 /* When OLDVAL is zero and we want the strong version we can emit a tighter
12491 LD[A]XR rval, [mem]
12493 ST[L]XR scratch, newval, [mem]
12494 CBNZ scratch, .label1
12497 bool strong_zero_p
= !is_weak
&& oldval
== const0_rtx
;
12502 label1
= gen_label_rtx ();
12503 emit_label (label1
);
12505 label2
= gen_label_rtx ();
12507 /* The initial load can be relaxed for a __sync operation since a final
12508 barrier will be emitted to stop code hoisting. */
12509 if (is_mm_sync (model
))
12510 aarch64_emit_load_exclusive (mode
, rval
, mem
,
12511 GEN_INT (MEMMODEL_RELAXED
));
12513 aarch64_emit_load_exclusive (mode
, rval
, mem
, model_rtx
);
12517 x
= gen_rtx_NE (VOIDmode
, rval
, const0_rtx
);
12518 x
= gen_rtx_IF_THEN_ELSE (VOIDmode
, x
,
12519 gen_rtx_LABEL_REF (Pmode
, label2
), pc_rtx
);
12520 aarch64_emit_unlikely_jump (gen_rtx_SET (pc_rtx
, x
));
12524 cond
= aarch64_gen_compare_reg (NE
, rval
, oldval
);
12525 x
= gen_rtx_NE (VOIDmode
, cond
, const0_rtx
);
12526 x
= gen_rtx_IF_THEN_ELSE (VOIDmode
, x
,
12527 gen_rtx_LABEL_REF (Pmode
, label2
), pc_rtx
);
12528 aarch64_emit_unlikely_jump (gen_rtx_SET (pc_rtx
, x
));
12531 aarch64_emit_store_exclusive (mode
, scratch
, mem
, newval
, model_rtx
);
12535 x
= gen_rtx_NE (VOIDmode
, scratch
, const0_rtx
);
12536 x
= gen_rtx_IF_THEN_ELSE (VOIDmode
, x
,
12537 gen_rtx_LABEL_REF (Pmode
, label1
), pc_rtx
);
12538 aarch64_emit_unlikely_jump (gen_rtx_SET (pc_rtx
, x
));
12542 cond
= gen_rtx_REG (CCmode
, CC_REGNUM
);
12543 x
= gen_rtx_COMPARE (CCmode
, scratch
, const0_rtx
);
12544 emit_insn (gen_rtx_SET (cond
, x
));
12547 emit_label (label2
);
12548 /* If we used a CBNZ in the exchange loop emit an explicit compare with RVAL
12549 to set the condition flags. If this is not used it will be removed by
12553 cond
= gen_rtx_REG (CCmode
, CC_REGNUM
);
12554 x
= gen_rtx_COMPARE (CCmode
, rval
, const0_rtx
);
12555 emit_insn (gen_rtx_SET (cond
, x
));
12557 /* Emit any final barrier needed for a __sync operation. */
12558 if (is_mm_sync (model
))
12559 aarch64_emit_post_barrier (model
);
12562 /* Emit a BIC instruction. */
12565 aarch64_emit_bic (machine_mode mode
, rtx dst
, rtx s1
, rtx s2
, int shift
)
12567 rtx shift_rtx
= GEN_INT (shift
);
12568 rtx (*gen
) (rtx
, rtx
, rtx
, rtx
);
12572 case E_SImode
: gen
= gen_and_one_cmpl_lshrsi3
; break;
12573 case E_DImode
: gen
= gen_and_one_cmpl_lshrdi3
; break;
12575 gcc_unreachable ();
12578 emit_insn (gen (dst
, s2
, shift_rtx
, s1
));
12581 /* Emit an atomic swap. */
12584 aarch64_emit_atomic_swap (machine_mode mode
, rtx dst
, rtx value
,
12585 rtx mem
, rtx model
)
12587 rtx (*gen
) (rtx
, rtx
, rtx
, rtx
);
12591 case E_QImode
: gen
= gen_aarch64_atomic_swpqi
; break;
12592 case E_HImode
: gen
= gen_aarch64_atomic_swphi
; break;
12593 case E_SImode
: gen
= gen_aarch64_atomic_swpsi
; break;
12594 case E_DImode
: gen
= gen_aarch64_atomic_swpdi
; break;
12596 gcc_unreachable ();
12599 emit_insn (gen (dst
, mem
, value
, model
));
12602 /* Operations supported by aarch64_emit_atomic_load_op. */
12604 enum aarch64_atomic_load_op_code
12606 AARCH64_LDOP_PLUS
, /* A + B */
12607 AARCH64_LDOP_XOR
, /* A ^ B */
12608 AARCH64_LDOP_OR
, /* A | B */
12609 AARCH64_LDOP_BIC
/* A & ~B */
12612 /* Emit an atomic load-operate. */
12615 aarch64_emit_atomic_load_op (enum aarch64_atomic_load_op_code code
,
12616 machine_mode mode
, rtx dst
, rtx src
,
12617 rtx mem
, rtx model
)
12619 typedef rtx (*aarch64_atomic_load_op_fn
) (rtx
, rtx
, rtx
, rtx
);
12620 const aarch64_atomic_load_op_fn plus
[] =
12622 gen_aarch64_atomic_loadaddqi
,
12623 gen_aarch64_atomic_loadaddhi
,
12624 gen_aarch64_atomic_loadaddsi
,
12625 gen_aarch64_atomic_loadadddi
12627 const aarch64_atomic_load_op_fn eor
[] =
12629 gen_aarch64_atomic_loadeorqi
,
12630 gen_aarch64_atomic_loadeorhi
,
12631 gen_aarch64_atomic_loadeorsi
,
12632 gen_aarch64_atomic_loadeordi
12634 const aarch64_atomic_load_op_fn ior
[] =
12636 gen_aarch64_atomic_loadsetqi
,
12637 gen_aarch64_atomic_loadsethi
,
12638 gen_aarch64_atomic_loadsetsi
,
12639 gen_aarch64_atomic_loadsetdi
12641 const aarch64_atomic_load_op_fn bic
[] =
12643 gen_aarch64_atomic_loadclrqi
,
12644 gen_aarch64_atomic_loadclrhi
,
12645 gen_aarch64_atomic_loadclrsi
,
12646 gen_aarch64_atomic_loadclrdi
12648 aarch64_atomic_load_op_fn gen
;
12653 case E_QImode
: idx
= 0; break;
12654 case E_HImode
: idx
= 1; break;
12655 case E_SImode
: idx
= 2; break;
12656 case E_DImode
: idx
= 3; break;
12658 gcc_unreachable ();
12663 case AARCH64_LDOP_PLUS
: gen
= plus
[idx
]; break;
12664 case AARCH64_LDOP_XOR
: gen
= eor
[idx
]; break;
12665 case AARCH64_LDOP_OR
: gen
= ior
[idx
]; break;
12666 case AARCH64_LDOP_BIC
: gen
= bic
[idx
]; break;
12668 gcc_unreachable ();
12671 emit_insn (gen (dst
, mem
, src
, model
));
12674 /* Emit an atomic load+operate. CODE is the operation. OUT_DATA is the
12675 location to store the data read from memory. OUT_RESULT is the location to
12676 store the result of the operation. MEM is the memory location to read and
12677 modify. MODEL_RTX is the memory ordering to use. VALUE is the second
12678 operand for the operation. Either OUT_DATA or OUT_RESULT, but not both, can
12682 aarch64_gen_atomic_ldop (enum rtx_code code
, rtx out_data
, rtx out_result
,
12683 rtx mem
, rtx value
, rtx model_rtx
)
12685 machine_mode mode
= GET_MODE (mem
);
12686 machine_mode wmode
= (mode
== DImode
? DImode
: SImode
);
12687 const bool short_mode
= (mode
< SImode
);
12688 aarch64_atomic_load_op_code ldop_code
;
12693 out_data
= gen_lowpart (mode
, out_data
);
12696 out_result
= gen_lowpart (mode
, out_result
);
12698 /* Make sure the value is in a register, putting it into a destination
12699 register if it needs to be manipulated. */
12700 if (!register_operand (value
, mode
)
12701 || code
== AND
|| code
== MINUS
)
12703 src
= out_result
? out_result
: out_data
;
12704 emit_move_insn (src
, gen_lowpart (mode
, value
));
12708 gcc_assert (register_operand (src
, mode
));
12710 /* Preprocess the data for the operation as necessary. If the operation is
12711 a SET then emit a swap instruction and finish. */
12715 aarch64_emit_atomic_swap (mode
, out_data
, src
, mem
, model_rtx
);
12719 /* Negate the value and treat it as a PLUS. */
12723 /* Resize the value if necessary. */
12725 src
= gen_lowpart (wmode
, src
);
12727 neg_src
= gen_rtx_NEG (wmode
, src
);
12728 emit_insn (gen_rtx_SET (src
, neg_src
));
12731 src
= gen_lowpart (mode
, src
);
12733 /* Fall-through. */
12735 ldop_code
= AARCH64_LDOP_PLUS
;
12739 ldop_code
= AARCH64_LDOP_OR
;
12743 ldop_code
= AARCH64_LDOP_XOR
;
12750 /* Resize the value if necessary. */
12752 src
= gen_lowpart (wmode
, src
);
12754 not_src
= gen_rtx_NOT (wmode
, src
);
12755 emit_insn (gen_rtx_SET (src
, not_src
));
12758 src
= gen_lowpart (mode
, src
);
12760 ldop_code
= AARCH64_LDOP_BIC
;
12764 /* The operation can't be done with atomic instructions. */
12765 gcc_unreachable ();
12768 aarch64_emit_atomic_load_op (ldop_code
, mode
, out_data
, src
, mem
, model_rtx
);
12770 /* If necessary, calculate the data in memory after the update by redoing the
12771 operation from values in registers. */
12777 src
= gen_lowpart (wmode
, src
);
12778 out_data
= gen_lowpart (wmode
, out_data
);
12779 out_result
= gen_lowpart (wmode
, out_result
);
12788 x
= gen_rtx_PLUS (wmode
, out_data
, src
);
12791 x
= gen_rtx_IOR (wmode
, out_data
, src
);
12794 x
= gen_rtx_XOR (wmode
, out_data
, src
);
12797 aarch64_emit_bic (wmode
, out_result
, out_data
, src
, 0);
12800 gcc_unreachable ();
12803 emit_set_insn (out_result
, x
);
12808 /* Split an atomic operation. */
12811 aarch64_split_atomic_op (enum rtx_code code
, rtx old_out
, rtx new_out
, rtx mem
,
12812 rtx value
, rtx model_rtx
, rtx cond
)
12814 machine_mode mode
= GET_MODE (mem
);
12815 machine_mode wmode
= (mode
== DImode
? DImode
: SImode
);
12816 const enum memmodel model
= memmodel_from_int (INTVAL (model_rtx
));
12817 const bool is_sync
= is_mm_sync (model
);
12818 rtx_code_label
*label
;
12821 /* Split the atomic operation into a sequence. */
12822 label
= gen_label_rtx ();
12823 emit_label (label
);
12826 new_out
= gen_lowpart (wmode
, new_out
);
12828 old_out
= gen_lowpart (wmode
, old_out
);
12831 value
= simplify_gen_subreg (wmode
, value
, mode
, 0);
12833 /* The initial load can be relaxed for a __sync operation since a final
12834 barrier will be emitted to stop code hoisting. */
12836 aarch64_emit_load_exclusive (mode
, old_out
, mem
,
12837 GEN_INT (MEMMODEL_RELAXED
));
12839 aarch64_emit_load_exclusive (mode
, old_out
, mem
, model_rtx
);
12848 x
= gen_rtx_AND (wmode
, old_out
, value
);
12849 emit_insn (gen_rtx_SET (new_out
, x
));
12850 x
= gen_rtx_NOT (wmode
, new_out
);
12851 emit_insn (gen_rtx_SET (new_out
, x
));
12855 if (CONST_INT_P (value
))
12857 value
= GEN_INT (-INTVAL (value
));
12860 /* Fall through. */
12863 x
= gen_rtx_fmt_ee (code
, wmode
, old_out
, value
);
12864 emit_insn (gen_rtx_SET (new_out
, x
));
12868 aarch64_emit_store_exclusive (mode
, cond
, mem
,
12869 gen_lowpart (mode
, new_out
), model_rtx
);
12871 x
= gen_rtx_NE (VOIDmode
, cond
, const0_rtx
);
12872 x
= gen_rtx_IF_THEN_ELSE (VOIDmode
, x
,
12873 gen_rtx_LABEL_REF (Pmode
, label
), pc_rtx
);
12874 aarch64_emit_unlikely_jump (gen_rtx_SET (pc_rtx
, x
));
12876 /* Emit any final barrier needed for a __sync operation. */
12878 aarch64_emit_post_barrier (model
);
12882 aarch64_init_libfuncs (void)
12884 /* Half-precision float operations. The compiler handles all operations
12885 with NULL libfuncs by converting to SFmode. */
12888 set_conv_libfunc (trunc_optab
, HFmode
, SFmode
, "__gnu_f2h_ieee");
12889 set_conv_libfunc (sext_optab
, SFmode
, HFmode
, "__gnu_h2f_ieee");
12892 set_optab_libfunc (add_optab
, HFmode
, NULL
);
12893 set_optab_libfunc (sdiv_optab
, HFmode
, NULL
);
12894 set_optab_libfunc (smul_optab
, HFmode
, NULL
);
12895 set_optab_libfunc (neg_optab
, HFmode
, NULL
);
12896 set_optab_libfunc (sub_optab
, HFmode
, NULL
);
12899 set_optab_libfunc (eq_optab
, HFmode
, NULL
);
12900 set_optab_libfunc (ne_optab
, HFmode
, NULL
);
12901 set_optab_libfunc (lt_optab
, HFmode
, NULL
);
12902 set_optab_libfunc (le_optab
, HFmode
, NULL
);
12903 set_optab_libfunc (ge_optab
, HFmode
, NULL
);
12904 set_optab_libfunc (gt_optab
, HFmode
, NULL
);
12905 set_optab_libfunc (unord_optab
, HFmode
, NULL
);
12908 /* Target hook for c_mode_for_suffix. */
12909 static machine_mode
12910 aarch64_c_mode_for_suffix (char suffix
)
12918 /* We can only represent floating point constants which will fit in
12919 "quarter-precision" values. These values are characterised by
12920 a sign bit, a 4-bit mantissa and a 3-bit exponent. And are given
12923 (-1)^s * (n/16) * 2^r
12926 's' is the sign bit.
12927 'n' is an integer in the range 16 <= n <= 31.
12928 'r' is an integer in the range -3 <= r <= 4. */
12930 /* Return true iff X can be represented by a quarter-precision
12931 floating point immediate operand X. Note, we cannot represent 0.0. */
12933 aarch64_float_const_representable_p (rtx x
)
12935 /* This represents our current view of how many bits
12936 make up the mantissa. */
12937 int point_pos
= 2 * HOST_BITS_PER_WIDE_INT
- 1;
12939 unsigned HOST_WIDE_INT mantissa
, mask
;
12940 REAL_VALUE_TYPE r
, m
;
12943 if (!CONST_DOUBLE_P (x
))
12946 /* We don't support HFmode constants yet. */
12947 if (GET_MODE (x
) == VOIDmode
|| GET_MODE (x
) == HFmode
)
12950 r
= *CONST_DOUBLE_REAL_VALUE (x
);
12952 /* We cannot represent infinities, NaNs or +/-zero. We won't
12953 know if we have +zero until we analyse the mantissa, but we
12954 can reject the other invalid values. */
12955 if (REAL_VALUE_ISINF (r
) || REAL_VALUE_ISNAN (r
)
12956 || REAL_VALUE_MINUS_ZERO (r
))
12959 /* Extract exponent. */
12960 r
= real_value_abs (&r
);
12961 exponent
= REAL_EXP (&r
);
12963 /* For the mantissa, we expand into two HOST_WIDE_INTS, apart from the
12964 highest (sign) bit, with a fixed binary point at bit point_pos.
12965 m1 holds the low part of the mantissa, m2 the high part.
12966 WARNING: If we ever have a representation using more than 2 * H_W_I - 1
12967 bits for the mantissa, this can fail (low bits will be lost). */
12968 real_ldexp (&m
, &r
, point_pos
- exponent
);
12969 wide_int w
= real_to_integer (&m
, &fail
, HOST_BITS_PER_WIDE_INT
* 2);
12971 /* If the low part of the mantissa has bits set we cannot represent
12973 if (w
.ulow () != 0)
12975 /* We have rejected the lower HOST_WIDE_INT, so update our
12976 understanding of how many bits lie in the mantissa and
12977 look only at the high HOST_WIDE_INT. */
12978 mantissa
= w
.elt (1);
12979 point_pos
-= HOST_BITS_PER_WIDE_INT
;
12981 /* We can only represent values with a mantissa of the form 1.xxxx. */
12982 mask
= ((unsigned HOST_WIDE_INT
)1 << (point_pos
- 5)) - 1;
12983 if ((mantissa
& mask
) != 0)
12986 /* Having filtered unrepresentable values, we may now remove all
12987 but the highest 5 bits. */
12988 mantissa
>>= point_pos
- 5;
12990 /* We cannot represent the value 0.0, so reject it. This is handled
12995 /* Then, as bit 4 is always set, we can mask it off, leaving
12996 the mantissa in the range [0, 15]. */
12997 mantissa
&= ~(1 << 4);
12998 gcc_assert (mantissa
<= 15);
13000 /* GCC internally does not use IEEE754-like encoding (where normalized
13001 significands are in the range [1, 2). GCC uses [0.5, 1) (see real.c).
13002 Our mantissa values are shifted 4 places to the left relative to
13003 normalized IEEE754 so we must modify the exponent returned by REAL_EXP
13004 by 5 places to correct for GCC's representation. */
13005 exponent
= 5 - exponent
;
13007 return (exponent
>= 0 && exponent
<= 7);
13011 aarch64_output_simd_mov_immediate (rtx const_vector
,
13016 static char templ
[40];
13017 const char *mnemonic
;
13018 const char *shift_op
;
13019 unsigned int lane_count
= 0;
13022 struct simd_immediate_info info
= { NULL_RTX
, 0, 0, false, false };
13024 /* This will return true to show const_vector is legal for use as either
13025 a AdvSIMD MOVI instruction (or, implicitly, MVNI) immediate. It will
13026 also update INFO to show how the immediate should be generated. */
13027 is_valid
= aarch64_simd_valid_immediate (const_vector
, mode
, false, &info
);
13028 gcc_assert (is_valid
);
13030 element_char
= sizetochar (info
.element_width
);
13031 lane_count
= width
/ info
.element_width
;
13033 mode
= GET_MODE_INNER (mode
);
13034 if (GET_MODE_CLASS (mode
) == MODE_FLOAT
)
13036 gcc_assert (info
.shift
== 0 && ! info
.mvn
);
13037 /* For FP zero change it to a CONST_INT 0 and use the integer SIMD
13038 move immediate path. */
13039 if (aarch64_float_const_zero_rtx_p (info
.value
))
13040 info
.value
= GEN_INT (0);
13043 const unsigned int buf_size
= 20;
13044 char float_buf
[buf_size
] = {'\0'};
13045 real_to_decimal_for_mode (float_buf
,
13046 CONST_DOUBLE_REAL_VALUE (info
.value
),
13047 buf_size
, buf_size
, 1, mode
);
13049 if (lane_count
== 1)
13050 snprintf (templ
, sizeof (templ
), "fmov\t%%d0, %s", float_buf
);
13052 snprintf (templ
, sizeof (templ
), "fmov\t%%0.%d%c, %s",
13053 lane_count
, element_char
, float_buf
);
13058 mnemonic
= info
.mvn
? "mvni" : "movi";
13059 shift_op
= info
.msl
? "msl" : "lsl";
13061 gcc_assert (CONST_INT_P (info
.value
));
13062 if (lane_count
== 1)
13063 snprintf (templ
, sizeof (templ
), "%s\t%%d0, " HOST_WIDE_INT_PRINT_HEX
,
13064 mnemonic
, UINTVAL (info
.value
));
13065 else if (info
.shift
)
13066 snprintf (templ
, sizeof (templ
), "%s\t%%0.%d%c, " HOST_WIDE_INT_PRINT_HEX
13067 ", %s %d", mnemonic
, lane_count
, element_char
,
13068 UINTVAL (info
.value
), shift_op
, info
.shift
);
13070 snprintf (templ
, sizeof (templ
), "%s\t%%0.%d%c, " HOST_WIDE_INT_PRINT_HEX
,
13071 mnemonic
, lane_count
, element_char
, UINTVAL (info
.value
));
13076 aarch64_output_scalar_simd_mov_immediate (rtx immediate
, scalar_int_mode mode
)
13079 /* If a floating point number was passed and we desire to use it in an
13080 integer mode do the conversion to integer. */
13081 if (CONST_DOUBLE_P (immediate
) && GET_MODE_CLASS (mode
) == MODE_INT
)
13083 unsigned HOST_WIDE_INT ival
;
13084 if (!aarch64_reinterpret_float_as_int (immediate
, &ival
))
13085 gcc_unreachable ();
13086 immediate
= gen_int_mode (ival
, mode
);
13089 machine_mode vmode
;
13090 /* use a 64 bit mode for everything except for DI/DF mode, where we use
13091 a 128 bit vector mode. */
13092 int width
= GET_MODE_BITSIZE (mode
) == 64 ? 128 : 64;
13094 vmode
= aarch64_simd_container_mode (mode
, width
);
13095 rtx v_op
= aarch64_simd_gen_const_vector_dup (vmode
, INTVAL (immediate
));
13096 return aarch64_output_simd_mov_immediate (v_op
, vmode
, width
);
13099 /* Split operands into moves from op[1] + op[2] into op[0]. */
13102 aarch64_split_combinev16qi (rtx operands
[3])
13104 unsigned int dest
= REGNO (operands
[0]);
13105 unsigned int src1
= REGNO (operands
[1]);
13106 unsigned int src2
= REGNO (operands
[2]);
13107 machine_mode halfmode
= GET_MODE (operands
[1]);
13108 unsigned int halfregs
= REG_NREGS (operands
[1]);
13109 rtx destlo
, desthi
;
13111 gcc_assert (halfmode
== V16QImode
);
13113 if (src1
== dest
&& src2
== dest
+ halfregs
)
13115 /* No-op move. Can't split to nothing; emit something. */
13116 emit_note (NOTE_INSN_DELETED
);
13120 /* Preserve register attributes for variable tracking. */
13121 destlo
= gen_rtx_REG_offset (operands
[0], halfmode
, dest
, 0);
13122 desthi
= gen_rtx_REG_offset (operands
[0], halfmode
, dest
+ halfregs
,
13123 GET_MODE_SIZE (halfmode
));
13125 /* Special case of reversed high/low parts. */
13126 if (reg_overlap_mentioned_p (operands
[2], destlo
)
13127 && reg_overlap_mentioned_p (operands
[1], desthi
))
13129 emit_insn (gen_xorv16qi3 (operands
[1], operands
[1], operands
[2]));
13130 emit_insn (gen_xorv16qi3 (operands
[2], operands
[1], operands
[2]));
13131 emit_insn (gen_xorv16qi3 (operands
[1], operands
[1], operands
[2]));
13133 else if (!reg_overlap_mentioned_p (operands
[2], destlo
))
13135 /* Try to avoid unnecessary moves if part of the result
13136 is in the right place already. */
13138 emit_move_insn (destlo
, operands
[1]);
13139 if (src2
!= dest
+ halfregs
)
13140 emit_move_insn (desthi
, operands
[2]);
13144 if (src2
!= dest
+ halfregs
)
13145 emit_move_insn (desthi
, operands
[2]);
13147 emit_move_insn (destlo
, operands
[1]);
13151 /* vec_perm support. */
13153 #define MAX_VECT_LEN 16
13155 struct expand_vec_perm_d
13157 rtx target
, op0
, op1
;
13158 unsigned char perm
[MAX_VECT_LEN
];
13159 machine_mode vmode
;
13160 unsigned char nelt
;
13165 /* Generate a variable permutation. */
13168 aarch64_expand_vec_perm_1 (rtx target
, rtx op0
, rtx op1
, rtx sel
)
13170 machine_mode vmode
= GET_MODE (target
);
13171 bool one_vector_p
= rtx_equal_p (op0
, op1
);
13173 gcc_checking_assert (vmode
== V8QImode
|| vmode
== V16QImode
);
13174 gcc_checking_assert (GET_MODE (op0
) == vmode
);
13175 gcc_checking_assert (GET_MODE (op1
) == vmode
);
13176 gcc_checking_assert (GET_MODE (sel
) == vmode
);
13177 gcc_checking_assert (TARGET_SIMD
);
13181 if (vmode
== V8QImode
)
13183 /* Expand the argument to a V16QI mode by duplicating it. */
13184 rtx pair
= gen_reg_rtx (V16QImode
);
13185 emit_insn (gen_aarch64_combinev8qi (pair
, op0
, op0
));
13186 emit_insn (gen_aarch64_tbl1v8qi (target
, pair
, sel
));
13190 emit_insn (gen_aarch64_tbl1v16qi (target
, op0
, sel
));
13197 if (vmode
== V8QImode
)
13199 pair
= gen_reg_rtx (V16QImode
);
13200 emit_insn (gen_aarch64_combinev8qi (pair
, op0
, op1
));
13201 emit_insn (gen_aarch64_tbl1v8qi (target
, pair
, sel
));
13205 pair
= gen_reg_rtx (OImode
);
13206 emit_insn (gen_aarch64_combinev16qi (pair
, op0
, op1
));
13207 emit_insn (gen_aarch64_tbl2v16qi (target
, pair
, sel
));
13213 aarch64_expand_vec_perm (rtx target
, rtx op0
, rtx op1
, rtx sel
)
13215 machine_mode vmode
= GET_MODE (target
);
13216 unsigned int nelt
= GET_MODE_NUNITS (vmode
);
13217 bool one_vector_p
= rtx_equal_p (op0
, op1
);
13220 /* The TBL instruction does not use a modulo index, so we must take care
13221 of that ourselves. */
13222 mask
= aarch64_simd_gen_const_vector_dup (vmode
,
13223 one_vector_p
? nelt
- 1 : 2 * nelt
- 1);
13224 sel
= expand_simple_binop (vmode
, AND
, sel
, mask
, NULL
, 0, OPTAB_LIB_WIDEN
);
13226 /* For big-endian, we also need to reverse the index within the vector
13227 (but not which vector). */
13228 if (BYTES_BIG_ENDIAN
)
13230 /* If one_vector_p, mask is a vector of (nelt - 1)'s already. */
13232 mask
= aarch64_simd_gen_const_vector_dup (vmode
, nelt
- 1);
13233 sel
= expand_simple_binop (vmode
, XOR
, sel
, mask
,
13234 NULL
, 0, OPTAB_LIB_WIDEN
);
13236 aarch64_expand_vec_perm_1 (target
, op0
, op1
, sel
);
13239 /* Recognize patterns suitable for the TRN instructions. */
13241 aarch64_evpc_trn (struct expand_vec_perm_d
*d
)
13243 unsigned int i
, odd
, mask
, nelt
= d
->nelt
;
13244 rtx out
, in0
, in1
, x
;
13245 rtx (*gen
) (rtx
, rtx
, rtx
);
13246 machine_mode vmode
= d
->vmode
;
13248 if (GET_MODE_UNIT_SIZE (vmode
) > 8)
13251 /* Note that these are little-endian tests.
13252 We correct for big-endian later. */
13253 if (d
->perm
[0] == 0)
13255 else if (d
->perm
[0] == 1)
13259 mask
= (d
->one_vector_p
? nelt
- 1 : 2 * nelt
- 1);
13261 for (i
= 0; i
< nelt
; i
+= 2)
13263 if (d
->perm
[i
] != i
+ odd
)
13265 if (d
->perm
[i
+ 1] != ((i
+ nelt
+ odd
) & mask
))
13275 if (BYTES_BIG_ENDIAN
)
13277 x
= in0
, in0
= in1
, in1
= x
;
13286 case E_V16QImode
: gen
= gen_aarch64_trn2v16qi
; break;
13287 case E_V8QImode
: gen
= gen_aarch64_trn2v8qi
; break;
13288 case E_V8HImode
: gen
= gen_aarch64_trn2v8hi
; break;
13289 case E_V4HImode
: gen
= gen_aarch64_trn2v4hi
; break;
13290 case E_V4SImode
: gen
= gen_aarch64_trn2v4si
; break;
13291 case E_V2SImode
: gen
= gen_aarch64_trn2v2si
; break;
13292 case E_V2DImode
: gen
= gen_aarch64_trn2v2di
; break;
13293 case E_V4HFmode
: gen
= gen_aarch64_trn2v4hf
; break;
13294 case E_V8HFmode
: gen
= gen_aarch64_trn2v8hf
; break;
13295 case E_V4SFmode
: gen
= gen_aarch64_trn2v4sf
; break;
13296 case E_V2SFmode
: gen
= gen_aarch64_trn2v2sf
; break;
13297 case E_V2DFmode
: gen
= gen_aarch64_trn2v2df
; break;
13306 case E_V16QImode
: gen
= gen_aarch64_trn1v16qi
; break;
13307 case E_V8QImode
: gen
= gen_aarch64_trn1v8qi
; break;
13308 case E_V8HImode
: gen
= gen_aarch64_trn1v8hi
; break;
13309 case E_V4HImode
: gen
= gen_aarch64_trn1v4hi
; break;
13310 case E_V4SImode
: gen
= gen_aarch64_trn1v4si
; break;
13311 case E_V2SImode
: gen
= gen_aarch64_trn1v2si
; break;
13312 case E_V2DImode
: gen
= gen_aarch64_trn1v2di
; break;
13313 case E_V4HFmode
: gen
= gen_aarch64_trn1v4hf
; break;
13314 case E_V8HFmode
: gen
= gen_aarch64_trn1v8hf
; break;
13315 case E_V4SFmode
: gen
= gen_aarch64_trn1v4sf
; break;
13316 case E_V2SFmode
: gen
= gen_aarch64_trn1v2sf
; break;
13317 case E_V2DFmode
: gen
= gen_aarch64_trn1v2df
; break;
13323 emit_insn (gen (out
, in0
, in1
));
13327 /* Recognize patterns suitable for the UZP instructions. */
13329 aarch64_evpc_uzp (struct expand_vec_perm_d
*d
)
13331 unsigned int i
, odd
, mask
, nelt
= d
->nelt
;
13332 rtx out
, in0
, in1
, x
;
13333 rtx (*gen
) (rtx
, rtx
, rtx
);
13334 machine_mode vmode
= d
->vmode
;
13336 if (GET_MODE_UNIT_SIZE (vmode
) > 8)
13339 /* Note that these are little-endian tests.
13340 We correct for big-endian later. */
13341 if (d
->perm
[0] == 0)
13343 else if (d
->perm
[0] == 1)
13347 mask
= (d
->one_vector_p
? nelt
- 1 : 2 * nelt
- 1);
13349 for (i
= 0; i
< nelt
; i
++)
13351 unsigned elt
= (i
* 2 + odd
) & mask
;
13352 if (d
->perm
[i
] != elt
)
13362 if (BYTES_BIG_ENDIAN
)
13364 x
= in0
, in0
= in1
, in1
= x
;
13373 case E_V16QImode
: gen
= gen_aarch64_uzp2v16qi
; break;
13374 case E_V8QImode
: gen
= gen_aarch64_uzp2v8qi
; break;
13375 case E_V8HImode
: gen
= gen_aarch64_uzp2v8hi
; break;
13376 case E_V4HImode
: gen
= gen_aarch64_uzp2v4hi
; break;
13377 case E_V4SImode
: gen
= gen_aarch64_uzp2v4si
; break;
13378 case E_V2SImode
: gen
= gen_aarch64_uzp2v2si
; break;
13379 case E_V2DImode
: gen
= gen_aarch64_uzp2v2di
; break;
13380 case E_V4HFmode
: gen
= gen_aarch64_uzp2v4hf
; break;
13381 case E_V8HFmode
: gen
= gen_aarch64_uzp2v8hf
; break;
13382 case E_V4SFmode
: gen
= gen_aarch64_uzp2v4sf
; break;
13383 case E_V2SFmode
: gen
= gen_aarch64_uzp2v2sf
; break;
13384 case E_V2DFmode
: gen
= gen_aarch64_uzp2v2df
; break;
13393 case E_V16QImode
: gen
= gen_aarch64_uzp1v16qi
; break;
13394 case E_V8QImode
: gen
= gen_aarch64_uzp1v8qi
; break;
13395 case E_V8HImode
: gen
= gen_aarch64_uzp1v8hi
; break;
13396 case E_V4HImode
: gen
= gen_aarch64_uzp1v4hi
; break;
13397 case E_V4SImode
: gen
= gen_aarch64_uzp1v4si
; break;
13398 case E_V2SImode
: gen
= gen_aarch64_uzp1v2si
; break;
13399 case E_V2DImode
: gen
= gen_aarch64_uzp1v2di
; break;
13400 case E_V4HFmode
: gen
= gen_aarch64_uzp1v4hf
; break;
13401 case E_V8HFmode
: gen
= gen_aarch64_uzp1v8hf
; break;
13402 case E_V4SFmode
: gen
= gen_aarch64_uzp1v4sf
; break;
13403 case E_V2SFmode
: gen
= gen_aarch64_uzp1v2sf
; break;
13404 case E_V2DFmode
: gen
= gen_aarch64_uzp1v2df
; break;
13410 emit_insn (gen (out
, in0
, in1
));
13414 /* Recognize patterns suitable for the ZIP instructions. */
13416 aarch64_evpc_zip (struct expand_vec_perm_d
*d
)
13418 unsigned int i
, high
, mask
, nelt
= d
->nelt
;
13419 rtx out
, in0
, in1
, x
;
13420 rtx (*gen
) (rtx
, rtx
, rtx
);
13421 machine_mode vmode
= d
->vmode
;
13423 if (GET_MODE_UNIT_SIZE (vmode
) > 8)
13426 /* Note that these are little-endian tests.
13427 We correct for big-endian later. */
13429 if (d
->perm
[0] == high
)
13432 else if (d
->perm
[0] == 0)
13436 mask
= (d
->one_vector_p
? nelt
- 1 : 2 * nelt
- 1);
13438 for (i
= 0; i
< nelt
/ 2; i
++)
13440 unsigned elt
= (i
+ high
) & mask
;
13441 if (d
->perm
[i
* 2] != elt
)
13443 elt
= (elt
+ nelt
) & mask
;
13444 if (d
->perm
[i
* 2 + 1] != elt
)
13454 if (BYTES_BIG_ENDIAN
)
13456 x
= in0
, in0
= in1
, in1
= x
;
13465 case E_V16QImode
: gen
= gen_aarch64_zip2v16qi
; break;
13466 case E_V8QImode
: gen
= gen_aarch64_zip2v8qi
; break;
13467 case E_V8HImode
: gen
= gen_aarch64_zip2v8hi
; break;
13468 case E_V4HImode
: gen
= gen_aarch64_zip2v4hi
; break;
13469 case E_V4SImode
: gen
= gen_aarch64_zip2v4si
; break;
13470 case E_V2SImode
: gen
= gen_aarch64_zip2v2si
; break;
13471 case E_V2DImode
: gen
= gen_aarch64_zip2v2di
; break;
13472 case E_V4HFmode
: gen
= gen_aarch64_zip2v4hf
; break;
13473 case E_V8HFmode
: gen
= gen_aarch64_zip2v8hf
; break;
13474 case E_V4SFmode
: gen
= gen_aarch64_zip2v4sf
; break;
13475 case E_V2SFmode
: gen
= gen_aarch64_zip2v2sf
; break;
13476 case E_V2DFmode
: gen
= gen_aarch64_zip2v2df
; break;
13485 case E_V16QImode
: gen
= gen_aarch64_zip1v16qi
; break;
13486 case E_V8QImode
: gen
= gen_aarch64_zip1v8qi
; break;
13487 case E_V8HImode
: gen
= gen_aarch64_zip1v8hi
; break;
13488 case E_V4HImode
: gen
= gen_aarch64_zip1v4hi
; break;
13489 case E_V4SImode
: gen
= gen_aarch64_zip1v4si
; break;
13490 case E_V2SImode
: gen
= gen_aarch64_zip1v2si
; break;
13491 case E_V2DImode
: gen
= gen_aarch64_zip1v2di
; break;
13492 case E_V4HFmode
: gen
= gen_aarch64_zip1v4hf
; break;
13493 case E_V8HFmode
: gen
= gen_aarch64_zip1v8hf
; break;
13494 case E_V4SFmode
: gen
= gen_aarch64_zip1v4sf
; break;
13495 case E_V2SFmode
: gen
= gen_aarch64_zip1v2sf
; break;
13496 case E_V2DFmode
: gen
= gen_aarch64_zip1v2df
; break;
13502 emit_insn (gen (out
, in0
, in1
));
13506 /* Recognize patterns for the EXT insn. */
13509 aarch64_evpc_ext (struct expand_vec_perm_d
*d
)
13511 unsigned int i
, nelt
= d
->nelt
;
13512 rtx (*gen
) (rtx
, rtx
, rtx
, rtx
);
13515 unsigned int location
= d
->perm
[0]; /* Always < nelt. */
13517 /* Check if the extracted indices are increasing by one. */
13518 for (i
= 1; i
< nelt
; i
++)
13520 unsigned int required
= location
+ i
;
13521 if (d
->one_vector_p
)
13523 /* We'll pass the same vector in twice, so allow indices to wrap. */
13524 required
&= (nelt
- 1);
13526 if (d
->perm
[i
] != required
)
13532 case E_V16QImode
: gen
= gen_aarch64_extv16qi
; break;
13533 case E_V8QImode
: gen
= gen_aarch64_extv8qi
; break;
13534 case E_V4HImode
: gen
= gen_aarch64_extv4hi
; break;
13535 case E_V8HImode
: gen
= gen_aarch64_extv8hi
; break;
13536 case E_V2SImode
: gen
= gen_aarch64_extv2si
; break;
13537 case E_V4SImode
: gen
= gen_aarch64_extv4si
; break;
13538 case E_V4HFmode
: gen
= gen_aarch64_extv4hf
; break;
13539 case E_V8HFmode
: gen
= gen_aarch64_extv8hf
; break;
13540 case E_V2SFmode
: gen
= gen_aarch64_extv2sf
; break;
13541 case E_V4SFmode
: gen
= gen_aarch64_extv4sf
; break;
13542 case E_V2DImode
: gen
= gen_aarch64_extv2di
; break;
13543 case E_V2DFmode
: gen
= gen_aarch64_extv2df
; break;
13552 /* The case where (location == 0) is a no-op for both big- and little-endian,
13553 and is removed by the mid-end at optimization levels -O1 and higher. */
13555 if (BYTES_BIG_ENDIAN
&& (location
!= 0))
13557 /* After setup, we want the high elements of the first vector (stored
13558 at the LSB end of the register), and the low elements of the second
13559 vector (stored at the MSB end of the register). So swap. */
13560 std::swap (d
->op0
, d
->op1
);
13561 /* location != 0 (above), so safe to assume (nelt - location) < nelt. */
13562 location
= nelt
- location
;
13565 offset
= GEN_INT (location
);
13566 emit_insn (gen (d
->target
, d
->op0
, d
->op1
, offset
));
13570 /* Recognize patterns for the REV insns. */
13573 aarch64_evpc_rev (struct expand_vec_perm_d
*d
)
13575 unsigned int i
, j
, diff
, nelt
= d
->nelt
;
13576 rtx (*gen
) (rtx
, rtx
);
13578 if (!d
->one_vector_p
)
13587 case E_V16QImode
: gen
= gen_aarch64_rev64v16qi
; break;
13588 case E_V8QImode
: gen
= gen_aarch64_rev64v8qi
; break;
13596 case E_V16QImode
: gen
= gen_aarch64_rev32v16qi
; break;
13597 case E_V8QImode
: gen
= gen_aarch64_rev32v8qi
; break;
13598 case E_V8HImode
: gen
= gen_aarch64_rev64v8hi
; break;
13599 case E_V4HImode
: gen
= gen_aarch64_rev64v4hi
; break;
13607 case E_V16QImode
: gen
= gen_aarch64_rev16v16qi
; break;
13608 case E_V8QImode
: gen
= gen_aarch64_rev16v8qi
; break;
13609 case E_V8HImode
: gen
= gen_aarch64_rev32v8hi
; break;
13610 case E_V4HImode
: gen
= gen_aarch64_rev32v4hi
; break;
13611 case E_V4SImode
: gen
= gen_aarch64_rev64v4si
; break;
13612 case E_V2SImode
: gen
= gen_aarch64_rev64v2si
; break;
13613 case E_V4SFmode
: gen
= gen_aarch64_rev64v4sf
; break;
13614 case E_V2SFmode
: gen
= gen_aarch64_rev64v2sf
; break;
13615 case E_V8HFmode
: gen
= gen_aarch64_rev64v8hf
; break;
13616 case E_V4HFmode
: gen
= gen_aarch64_rev64v4hf
; break;
13625 for (i
= 0; i
< nelt
; i
+= diff
+ 1)
13626 for (j
= 0; j
<= diff
; j
+= 1)
13628 /* This is guaranteed to be true as the value of diff
13629 is 7, 3, 1 and we should have enough elements in the
13630 queue to generate this. Getting a vector mask with a
13631 value of diff other than these values implies that
13632 something is wrong by the time we get here. */
13633 gcc_assert (i
+ j
< nelt
);
13634 if (d
->perm
[i
+ j
] != i
+ diff
- j
)
13642 emit_insn (gen (d
->target
, d
->op0
));
13647 aarch64_evpc_dup (struct expand_vec_perm_d
*d
)
13649 rtx (*gen
) (rtx
, rtx
, rtx
);
13650 rtx out
= d
->target
;
13652 machine_mode vmode
= d
->vmode
;
13653 unsigned int i
, elt
, nelt
= d
->nelt
;
13657 for (i
= 1; i
< nelt
; i
++)
13659 if (elt
!= d
->perm
[i
])
13663 /* The generic preparation in aarch64_expand_vec_perm_const_1
13664 swaps the operand order and the permute indices if it finds
13665 d->perm[0] to be in the second operand. Thus, we can always
13666 use d->op0 and need not do any extra arithmetic to get the
13667 correct lane number. */
13669 lane
= GEN_INT (elt
); /* The pattern corrects for big-endian. */
13673 case E_V16QImode
: gen
= gen_aarch64_dup_lanev16qi
; break;
13674 case E_V8QImode
: gen
= gen_aarch64_dup_lanev8qi
; break;
13675 case E_V8HImode
: gen
= gen_aarch64_dup_lanev8hi
; break;
13676 case E_V4HImode
: gen
= gen_aarch64_dup_lanev4hi
; break;
13677 case E_V4SImode
: gen
= gen_aarch64_dup_lanev4si
; break;
13678 case E_V2SImode
: gen
= gen_aarch64_dup_lanev2si
; break;
13679 case E_V2DImode
: gen
= gen_aarch64_dup_lanev2di
; break;
13680 case E_V8HFmode
: gen
= gen_aarch64_dup_lanev8hf
; break;
13681 case E_V4HFmode
: gen
= gen_aarch64_dup_lanev4hf
; break;
13682 case E_V4SFmode
: gen
= gen_aarch64_dup_lanev4sf
; break;
13683 case E_V2SFmode
: gen
= gen_aarch64_dup_lanev2sf
; break;
13684 case E_V2DFmode
: gen
= gen_aarch64_dup_lanev2df
; break;
13689 emit_insn (gen (out
, in0
, lane
));
13694 aarch64_evpc_tbl (struct expand_vec_perm_d
*d
)
13696 rtx rperm
[MAX_VECT_LEN
], sel
;
13697 machine_mode vmode
= d
->vmode
;
13698 unsigned int i
, nelt
= d
->nelt
;
13703 /* Generic code will try constant permutation twice. Once with the
13704 original mode and again with the elements lowered to QImode.
13705 So wait and don't do the selector expansion ourselves. */
13706 if (vmode
!= V8QImode
&& vmode
!= V16QImode
)
13709 for (i
= 0; i
< nelt
; ++i
)
13711 int nunits
= GET_MODE_NUNITS (vmode
);
13713 /* If big-endian and two vectors we end up with a weird mixed-endian
13714 mode on NEON. Reverse the index within each word but not the word
13716 rperm
[i
] = GEN_INT (BYTES_BIG_ENDIAN
? d
->perm
[i
] ^ (nunits
- 1)
13719 sel
= gen_rtx_CONST_VECTOR (vmode
, gen_rtvec_v (nelt
, rperm
));
13720 sel
= force_reg (vmode
, sel
);
13722 aarch64_expand_vec_perm_1 (d
->target
, d
->op0
, d
->op1
, sel
);
13727 aarch64_expand_vec_perm_const_1 (struct expand_vec_perm_d
*d
)
13729 /* The pattern matching functions above are written to look for a small
13730 number to begin the sequence (0, 1, N/2). If we begin with an index
13731 from the second operand, we can swap the operands. */
13732 if (d
->perm
[0] >= d
->nelt
)
13734 unsigned i
, nelt
= d
->nelt
;
13736 gcc_assert (nelt
== (nelt
& -nelt
));
13737 for (i
= 0; i
< nelt
; ++i
)
13738 d
->perm
[i
] ^= nelt
; /* Keep the same index, but in the other vector. */
13740 std::swap (d
->op0
, d
->op1
);
13745 if (aarch64_evpc_rev (d
))
13747 else if (aarch64_evpc_ext (d
))
13749 else if (aarch64_evpc_dup (d
))
13751 else if (aarch64_evpc_zip (d
))
13753 else if (aarch64_evpc_uzp (d
))
13755 else if (aarch64_evpc_trn (d
))
13757 return aarch64_evpc_tbl (d
);
13762 /* Expand a vec_perm_const pattern. */
13765 aarch64_expand_vec_perm_const (rtx target
, rtx op0
, rtx op1
, rtx sel
)
13767 struct expand_vec_perm_d d
;
13768 int i
, nelt
, which
;
13774 d
.vmode
= GET_MODE (target
);
13775 gcc_assert (VECTOR_MODE_P (d
.vmode
));
13776 d
.nelt
= nelt
= GET_MODE_NUNITS (d
.vmode
);
13777 d
.testing_p
= false;
13779 for (i
= which
= 0; i
< nelt
; ++i
)
13781 rtx e
= XVECEXP (sel
, 0, i
);
13782 int ei
= INTVAL (e
) & (2 * nelt
- 1);
13783 which
|= (ei
< nelt
? 1 : 2);
13790 gcc_unreachable ();
13793 d
.one_vector_p
= false;
13794 if (!rtx_equal_p (op0
, op1
))
13797 /* The elements of PERM do not suggest that only the first operand
13798 is used, but both operands are identical. Allow easier matching
13799 of the permutation by folding the permutation into the single
13801 /* Fall Through. */
13803 for (i
= 0; i
< nelt
; ++i
)
13804 d
.perm
[i
] &= nelt
- 1;
13806 d
.one_vector_p
= true;
13811 d
.one_vector_p
= true;
13815 return aarch64_expand_vec_perm_const_1 (&d
);
13819 aarch64_vectorize_vec_perm_const_ok (machine_mode vmode
,
13820 const unsigned char *sel
)
13822 struct expand_vec_perm_d d
;
13823 unsigned int i
, nelt
, which
;
13827 d
.nelt
= nelt
= GET_MODE_NUNITS (d
.vmode
);
13828 d
.testing_p
= true;
13829 memcpy (d
.perm
, sel
, nelt
);
13831 /* Calculate whether all elements are in one vector. */
13832 for (i
= which
= 0; i
< nelt
; ++i
)
13834 unsigned char e
= d
.perm
[i
];
13835 gcc_assert (e
< 2 * nelt
);
13836 which
|= (e
< nelt
? 1 : 2);
13839 /* If all elements are from the second vector, reindex as if from the
13842 for (i
= 0; i
< nelt
; ++i
)
13845 /* Check whether the mask can be applied to a single vector. */
13846 d
.one_vector_p
= (which
!= 3);
13848 d
.target
= gen_raw_REG (d
.vmode
, LAST_VIRTUAL_REGISTER
+ 1);
13849 d
.op1
= d
.op0
= gen_raw_REG (d
.vmode
, LAST_VIRTUAL_REGISTER
+ 2);
13850 if (!d
.one_vector_p
)
13851 d
.op1
= gen_raw_REG (d
.vmode
, LAST_VIRTUAL_REGISTER
+ 3);
13854 ret
= aarch64_expand_vec_perm_const_1 (&d
);
13861 aarch64_reverse_mask (machine_mode mode
)
13863 /* We have to reverse each vector because we dont have
13864 a permuted load that can reverse-load according to ABI rules. */
13866 rtvec v
= rtvec_alloc (16);
13868 int nunits
= GET_MODE_NUNITS (mode
);
13869 int usize
= GET_MODE_UNIT_SIZE (mode
);
13871 gcc_assert (BYTES_BIG_ENDIAN
);
13872 gcc_assert (AARCH64_VALID_SIMD_QREG_MODE (mode
));
13874 for (i
= 0; i
< nunits
; i
++)
13875 for (j
= 0; j
< usize
; j
++)
13876 RTVEC_ELT (v
, i
* usize
+ j
) = GEN_INT ((i
+ 1) * usize
- 1 - j
);
13877 mask
= gen_rtx_CONST_VECTOR (V16QImode
, v
);
13878 return force_reg (V16QImode
, mask
);
13881 /* Implement TARGET_MODES_TIEABLE_P. In principle we should always return
13882 true. However due to issues with register allocation it is preferable
13883 to avoid tieing integer scalar and FP scalar modes. Executing integer
13884 operations in general registers is better than treating them as scalar
13885 vector operations. This reduces latency and avoids redundant int<->FP
13886 moves. So tie modes if they are either the same class, or vector modes
13887 with other vector modes, vector structs or any scalar mode. */
13890 aarch64_modes_tieable_p (machine_mode mode1
, machine_mode mode2
)
13892 if (GET_MODE_CLASS (mode1
) == GET_MODE_CLASS (mode2
))
13895 /* We specifically want to allow elements of "structure" modes to
13896 be tieable to the structure. This more general condition allows
13897 other rarer situations too. */
13898 if (aarch64_vector_mode_p (mode1
) && aarch64_vector_mode_p (mode2
))
13901 /* Also allow any scalar modes with vectors. */
13902 if (aarch64_vector_mode_supported_p (mode1
)
13903 || aarch64_vector_mode_supported_p (mode2
))
13909 /* Return a new RTX holding the result of moving POINTER forward by
13913 aarch64_move_pointer (rtx pointer
, int amount
)
13915 rtx next
= plus_constant (Pmode
, XEXP (pointer
, 0), amount
);
13917 return adjust_automodify_address (pointer
, GET_MODE (pointer
),
13921 /* Return a new RTX holding the result of moving POINTER forward by the
13922 size of the mode it points to. */
13925 aarch64_progress_pointer (rtx pointer
)
13927 HOST_WIDE_INT amount
= GET_MODE_SIZE (GET_MODE (pointer
));
13929 return aarch64_move_pointer (pointer
, amount
);
13932 /* Copy one MODE sized block from SRC to DST, then progress SRC and DST by
13936 aarch64_copy_one_block_and_progress_pointers (rtx
*src
, rtx
*dst
,
13939 rtx reg
= gen_reg_rtx (mode
);
13941 /* "Cast" the pointers to the correct mode. */
13942 *src
= adjust_address (*src
, mode
, 0);
13943 *dst
= adjust_address (*dst
, mode
, 0);
13944 /* Emit the memcpy. */
13945 emit_move_insn (reg
, *src
);
13946 emit_move_insn (*dst
, reg
);
13947 /* Move the pointers forward. */
13948 *src
= aarch64_progress_pointer (*src
);
13949 *dst
= aarch64_progress_pointer (*dst
);
13952 /* Expand movmem, as if from a __builtin_memcpy. Return true if
13953 we succeed, otherwise return false. */
13956 aarch64_expand_movmem (rtx
*operands
)
13959 rtx dst
= operands
[0];
13960 rtx src
= operands
[1];
13962 bool speed_p
= !optimize_function_for_size_p (cfun
);
13964 /* When optimizing for size, give a better estimate of the length of a
13965 memcpy call, but use the default otherwise. */
13966 unsigned int max_instructions
= (speed_p
? 15 : AARCH64_CALL_RATIO
) / 2;
13968 /* We can't do anything smart if the amount to copy is not constant. */
13969 if (!CONST_INT_P (operands
[2]))
13972 n
= UINTVAL (operands
[2]);
13974 /* Try to keep the number of instructions low. For cases below 16 bytes we
13975 need to make at most two moves. For cases above 16 bytes it will be one
13976 move for each 16 byte chunk, then at most two additional moves. */
13977 if (((n
/ 16) + (n
% 16 ? 2 : 0)) > max_instructions
)
13980 base
= copy_to_mode_reg (Pmode
, XEXP (dst
, 0));
13981 dst
= adjust_automodify_address (dst
, VOIDmode
, base
, 0);
13983 base
= copy_to_mode_reg (Pmode
, XEXP (src
, 0));
13984 src
= adjust_automodify_address (src
, VOIDmode
, base
, 0);
13986 /* Simple cases. Copy 0-3 bytes, as (if applicable) a 2-byte, then a
13992 aarch64_copy_one_block_and_progress_pointers (&src
, &dst
, HImode
);
13997 aarch64_copy_one_block_and_progress_pointers (&src
, &dst
, QImode
);
14002 /* Copy 4-8 bytes. First a 4-byte chunk, then (if applicable) a second
14003 4-byte chunk, partially overlapping with the previously copied chunk. */
14006 aarch64_copy_one_block_and_progress_pointers (&src
, &dst
, SImode
);
14012 src
= aarch64_move_pointer (src
, move
);
14013 dst
= aarch64_move_pointer (dst
, move
);
14014 aarch64_copy_one_block_and_progress_pointers (&src
, &dst
, SImode
);
14019 /* Copy more than 8 bytes. Copy chunks of 16 bytes until we run out of
14020 them, then (if applicable) an 8-byte chunk. */
14025 aarch64_copy_one_block_and_progress_pointers (&src
, &dst
, TImode
);
14030 aarch64_copy_one_block_and_progress_pointers (&src
, &dst
, DImode
);
14035 /* Finish the final bytes of the copy. We can always do this in one
14036 instruction. We either copy the exact amount we need, or partially
14037 overlap with the previous chunk we copied and copy 8-bytes. */
14041 aarch64_copy_one_block_and_progress_pointers (&src
, &dst
, QImode
);
14043 aarch64_copy_one_block_and_progress_pointers (&src
, &dst
, HImode
);
14045 aarch64_copy_one_block_and_progress_pointers (&src
, &dst
, SImode
);
14050 src
= aarch64_move_pointer (src
, -1);
14051 dst
= aarch64_move_pointer (dst
, -1);
14052 aarch64_copy_one_block_and_progress_pointers (&src
, &dst
, SImode
);
14058 src
= aarch64_move_pointer (src
, move
);
14059 dst
= aarch64_move_pointer (dst
, move
);
14060 aarch64_copy_one_block_and_progress_pointers (&src
, &dst
, DImode
);
14067 /* Split a DImode store of a CONST_INT SRC to MEM DST as two
14068 SImode stores. Handle the case when the constant has identical
14069 bottom and top halves. This is beneficial when the two stores can be
14070 merged into an STP and we avoid synthesising potentially expensive
14071 immediates twice. Return true if such a split is possible. */
14074 aarch64_split_dimode_const_store (rtx dst
, rtx src
)
14076 rtx lo
= gen_lowpart (SImode
, src
);
14077 rtx hi
= gen_highpart_mode (SImode
, DImode
, src
);
14079 bool size_p
= optimize_function_for_size_p (cfun
);
14081 if (!rtx_equal_p (lo
, hi
))
14084 unsigned int orig_cost
14085 = aarch64_internal_mov_immediate (NULL_RTX
, src
, false, DImode
);
14086 unsigned int lo_cost
14087 = aarch64_internal_mov_immediate (NULL_RTX
, lo
, false, SImode
);
14089 /* We want to transform:
14091 MOVK x1, 0x140, lsl 16
14092 MOVK x1, 0xc0da, lsl 32
14093 MOVK x1, 0x140, lsl 48
14097 MOVK w1, 0x140, lsl 16
14099 So we want to perform this only when we save two instructions
14100 or more. When optimizing for size, however, accept any code size
14102 if (size_p
&& orig_cost
<= lo_cost
)
14106 && (orig_cost
<= lo_cost
+ 1))
14109 rtx mem_lo
= adjust_address (dst
, SImode
, 0);
14110 if (!aarch64_mem_pair_operand (mem_lo
, SImode
))
14113 rtx tmp_reg
= gen_reg_rtx (SImode
);
14114 aarch64_expand_mov_immediate (tmp_reg
, lo
);
14115 rtx mem_hi
= aarch64_move_pointer (mem_lo
, GET_MODE_SIZE (SImode
));
14116 /* Don't emit an explicit store pair as this may not be always profitable.
14117 Let the sched-fusion logic decide whether to merge them. */
14118 emit_move_insn (mem_lo
, tmp_reg
);
14119 emit_move_insn (mem_hi
, tmp_reg
);
14124 /* Implement the TARGET_ASAN_SHADOW_OFFSET hook. */
14126 static unsigned HOST_WIDE_INT
14127 aarch64_asan_shadow_offset (void)
14129 return (HOST_WIDE_INT_1
<< 36);
14133 aarch64_use_by_pieces_infrastructure_p (unsigned HOST_WIDE_INT size
,
14134 unsigned int align
,
14135 enum by_pieces_operation op
,
14138 /* STORE_BY_PIECES can be used when copying a constant string, but
14139 in that case each 64-bit chunk takes 5 insns instead of 2 (LDR/STR).
14140 For now we always fail this and let the move_by_pieces code copy
14141 the string from read-only memory. */
14142 if (op
== STORE_BY_PIECES
)
14145 return default_use_by_pieces_infrastructure_p (size
, align
, op
, speed_p
);
14149 aarch64_gen_ccmp_first (rtx_insn
**prep_seq
, rtx_insn
**gen_seq
,
14150 int code
, tree treeop0
, tree treeop1
)
14152 machine_mode op_mode
, cmp_mode
, cc_mode
= CCmode
;
14154 int unsignedp
= TYPE_UNSIGNED (TREE_TYPE (treeop0
));
14156 struct expand_operand ops
[4];
14159 expand_operands (treeop0
, treeop1
, NULL_RTX
, &op0
, &op1
, EXPAND_NORMAL
);
14161 op_mode
= GET_MODE (op0
);
14162 if (op_mode
== VOIDmode
)
14163 op_mode
= GET_MODE (op1
);
14171 icode
= CODE_FOR_cmpsi
;
14176 icode
= CODE_FOR_cmpdi
;
14181 cc_mode
= aarch64_select_cc_mode ((rtx_code
) code
, op0
, op1
);
14182 icode
= cc_mode
== CCFPEmode
? CODE_FOR_fcmpesf
: CODE_FOR_fcmpsf
;
14187 cc_mode
= aarch64_select_cc_mode ((rtx_code
) code
, op0
, op1
);
14188 icode
= cc_mode
== CCFPEmode
? CODE_FOR_fcmpedf
: CODE_FOR_fcmpdf
;
14196 op0
= prepare_operand (icode
, op0
, 0, op_mode
, cmp_mode
, unsignedp
);
14197 op1
= prepare_operand (icode
, op1
, 1, op_mode
, cmp_mode
, unsignedp
);
14203 *prep_seq
= get_insns ();
14206 create_fixed_operand (&ops
[0], op0
);
14207 create_fixed_operand (&ops
[1], op1
);
14210 if (!maybe_expand_insn (icode
, 2, ops
))
14215 *gen_seq
= get_insns ();
14218 return gen_rtx_fmt_ee ((rtx_code
) code
, cc_mode
,
14219 gen_rtx_REG (cc_mode
, CC_REGNUM
), const0_rtx
);
14223 aarch64_gen_ccmp_next (rtx_insn
**prep_seq
, rtx_insn
**gen_seq
, rtx prev
,
14224 int cmp_code
, tree treeop0
, tree treeop1
, int bit_code
)
14226 rtx op0
, op1
, target
;
14227 machine_mode op_mode
, cmp_mode
, cc_mode
= CCmode
;
14228 int unsignedp
= TYPE_UNSIGNED (TREE_TYPE (treeop0
));
14230 struct expand_operand ops
[6];
14233 push_to_sequence (*prep_seq
);
14234 expand_operands (treeop0
, treeop1
, NULL_RTX
, &op0
, &op1
, EXPAND_NORMAL
);
14236 op_mode
= GET_MODE (op0
);
14237 if (op_mode
== VOIDmode
)
14238 op_mode
= GET_MODE (op1
);
14246 icode
= CODE_FOR_ccmpsi
;
14251 icode
= CODE_FOR_ccmpdi
;
14256 cc_mode
= aarch64_select_cc_mode ((rtx_code
) cmp_code
, op0
, op1
);
14257 icode
= cc_mode
== CCFPEmode
? CODE_FOR_fccmpesf
: CODE_FOR_fccmpsf
;
14262 cc_mode
= aarch64_select_cc_mode ((rtx_code
) cmp_code
, op0
, op1
);
14263 icode
= cc_mode
== CCFPEmode
? CODE_FOR_fccmpedf
: CODE_FOR_fccmpdf
;
14271 op0
= prepare_operand (icode
, op0
, 2, op_mode
, cmp_mode
, unsignedp
);
14272 op1
= prepare_operand (icode
, op1
, 3, op_mode
, cmp_mode
, unsignedp
);
14278 *prep_seq
= get_insns ();
14281 target
= gen_rtx_REG (cc_mode
, CC_REGNUM
);
14282 aarch64_cond
= aarch64_get_condition_code_1 (cc_mode
, (rtx_code
) cmp_code
);
14284 if (bit_code
!= AND
)
14286 prev
= gen_rtx_fmt_ee (REVERSE_CONDITION (GET_CODE (prev
),
14287 GET_MODE (XEXP (prev
, 0))),
14288 VOIDmode
, XEXP (prev
, 0), const0_rtx
);
14289 aarch64_cond
= AARCH64_INVERSE_CONDITION_CODE (aarch64_cond
);
14292 create_fixed_operand (&ops
[0], XEXP (prev
, 0));
14293 create_fixed_operand (&ops
[1], target
);
14294 create_fixed_operand (&ops
[2], op0
);
14295 create_fixed_operand (&ops
[3], op1
);
14296 create_fixed_operand (&ops
[4], prev
);
14297 create_fixed_operand (&ops
[5], GEN_INT (aarch64_cond
));
14299 push_to_sequence (*gen_seq
);
14300 if (!maybe_expand_insn (icode
, 6, ops
))
14306 *gen_seq
= get_insns ();
14309 return gen_rtx_fmt_ee ((rtx_code
) cmp_code
, VOIDmode
, target
, const0_rtx
);
14312 #undef TARGET_GEN_CCMP_FIRST
14313 #define TARGET_GEN_CCMP_FIRST aarch64_gen_ccmp_first
14315 #undef TARGET_GEN_CCMP_NEXT
14316 #define TARGET_GEN_CCMP_NEXT aarch64_gen_ccmp_next
14318 /* Implement TARGET_SCHED_MACRO_FUSION_P. Return true if target supports
14319 instruction fusion of some sort. */
14322 aarch64_macro_fusion_p (void)
14324 return aarch64_tune_params
.fusible_ops
!= AARCH64_FUSE_NOTHING
;
14328 /* Implement TARGET_SCHED_MACRO_FUSION_PAIR_P. Return true if PREV and CURR
14329 should be kept together during scheduling. */
14332 aarch_macro_fusion_pair_p (rtx_insn
*prev
, rtx_insn
*curr
)
14335 rtx prev_set
= single_set (prev
);
14336 rtx curr_set
= single_set (curr
);
14337 /* prev and curr are simple SET insns i.e. no flag setting or branching. */
14338 bool simple_sets_p
= prev_set
&& curr_set
&& !any_condjump_p (curr
);
14340 if (!aarch64_macro_fusion_p ())
14343 if (simple_sets_p
&& aarch64_fusion_enabled_p (AARCH64_FUSE_MOV_MOVK
))
14345 /* We are trying to match:
14346 prev (mov) == (set (reg r0) (const_int imm16))
14347 curr (movk) == (set (zero_extract (reg r0)
14350 (const_int imm16_1)) */
14352 set_dest
= SET_DEST (curr_set
);
14354 if (GET_CODE (set_dest
) == ZERO_EXTRACT
14355 && CONST_INT_P (SET_SRC (curr_set
))
14356 && CONST_INT_P (SET_SRC (prev_set
))
14357 && CONST_INT_P (XEXP (set_dest
, 2))
14358 && INTVAL (XEXP (set_dest
, 2)) == 16
14359 && REG_P (XEXP (set_dest
, 0))
14360 && REG_P (SET_DEST (prev_set
))
14361 && REGNO (XEXP (set_dest
, 0)) == REGNO (SET_DEST (prev_set
)))
14367 if (simple_sets_p
&& aarch64_fusion_enabled_p (AARCH64_FUSE_ADRP_ADD
))
14370 /* We're trying to match:
14371 prev (adrp) == (set (reg r1)
14372 (high (symbol_ref ("SYM"))))
14373 curr (add) == (set (reg r0)
14375 (symbol_ref ("SYM"))))
14376 Note that r0 need not necessarily be the same as r1, especially
14377 during pre-regalloc scheduling. */
14379 if (satisfies_constraint_Ush (SET_SRC (prev_set
))
14380 && REG_P (SET_DEST (prev_set
)) && REG_P (SET_DEST (curr_set
)))
14382 if (GET_CODE (SET_SRC (curr_set
)) == LO_SUM
14383 && REG_P (XEXP (SET_SRC (curr_set
), 0))
14384 && REGNO (XEXP (SET_SRC (curr_set
), 0))
14385 == REGNO (SET_DEST (prev_set
))
14386 && rtx_equal_p (XEXP (SET_SRC (prev_set
), 0),
14387 XEXP (SET_SRC (curr_set
), 1)))
14392 if (simple_sets_p
&& aarch64_fusion_enabled_p (AARCH64_FUSE_MOVK_MOVK
))
14395 /* We're trying to match:
14396 prev (movk) == (set (zero_extract (reg r0)
14399 (const_int imm16_1))
14400 curr (movk) == (set (zero_extract (reg r0)
14403 (const_int imm16_2)) */
14405 if (GET_CODE (SET_DEST (prev_set
)) == ZERO_EXTRACT
14406 && GET_CODE (SET_DEST (curr_set
)) == ZERO_EXTRACT
14407 && REG_P (XEXP (SET_DEST (prev_set
), 0))
14408 && REG_P (XEXP (SET_DEST (curr_set
), 0))
14409 && REGNO (XEXP (SET_DEST (prev_set
), 0))
14410 == REGNO (XEXP (SET_DEST (curr_set
), 0))
14411 && CONST_INT_P (XEXP (SET_DEST (prev_set
), 2))
14412 && CONST_INT_P (XEXP (SET_DEST (curr_set
), 2))
14413 && INTVAL (XEXP (SET_DEST (prev_set
), 2)) == 32
14414 && INTVAL (XEXP (SET_DEST (curr_set
), 2)) == 48
14415 && CONST_INT_P (SET_SRC (prev_set
))
14416 && CONST_INT_P (SET_SRC (curr_set
)))
14420 if (simple_sets_p
&& aarch64_fusion_enabled_p (AARCH64_FUSE_ADRP_LDR
))
14422 /* We're trying to match:
14423 prev (adrp) == (set (reg r0)
14424 (high (symbol_ref ("SYM"))))
14425 curr (ldr) == (set (reg r1)
14426 (mem (lo_sum (reg r0)
14427 (symbol_ref ("SYM")))))
14429 curr (ldr) == (set (reg r1)
14432 (symbol_ref ("SYM")))))) */
14433 if (satisfies_constraint_Ush (SET_SRC (prev_set
))
14434 && REG_P (SET_DEST (prev_set
)) && REG_P (SET_DEST (curr_set
)))
14436 rtx curr_src
= SET_SRC (curr_set
);
14438 if (GET_CODE (curr_src
) == ZERO_EXTEND
)
14439 curr_src
= XEXP (curr_src
, 0);
14441 if (MEM_P (curr_src
) && GET_CODE (XEXP (curr_src
, 0)) == LO_SUM
14442 && REG_P (XEXP (XEXP (curr_src
, 0), 0))
14443 && REGNO (XEXP (XEXP (curr_src
, 0), 0))
14444 == REGNO (SET_DEST (prev_set
))
14445 && rtx_equal_p (XEXP (XEXP (curr_src
, 0), 1),
14446 XEXP (SET_SRC (prev_set
), 0)))
14451 if (aarch64_fusion_enabled_p (AARCH64_FUSE_AES_AESMC
)
14452 && aarch_crypto_can_dual_issue (prev
, curr
))
14455 if (aarch64_fusion_enabled_p (AARCH64_FUSE_CMP_BRANCH
)
14456 && any_condjump_p (curr
))
14458 enum attr_type prev_type
= get_attr_type (prev
);
14460 unsigned int condreg1
, condreg2
;
14462 aarch64_fixed_condition_code_regs (&condreg1
, &condreg2
);
14463 cc_reg_1
= gen_rtx_REG (CCmode
, condreg1
);
14465 if (reg_referenced_p (cc_reg_1
, PATTERN (curr
))
14467 && modified_in_p (cc_reg_1
, prev
))
14469 /* FIXME: this misses some which is considered simple arthematic
14470 instructions for ThunderX. Simple shifts are missed here. */
14471 if (prev_type
== TYPE_ALUS_SREG
14472 || prev_type
== TYPE_ALUS_IMM
14473 || prev_type
== TYPE_LOGICS_REG
14474 || prev_type
== TYPE_LOGICS_IMM
)
14481 && aarch64_fusion_enabled_p (AARCH64_FUSE_ALU_BRANCH
)
14482 && any_condjump_p (curr
))
14484 /* We're trying to match:
14485 prev (alu_insn) == (set (r0) plus ((r0) (r1/imm)))
14486 curr (cbz) == (set (pc) (if_then_else (eq/ne) (r0)
14488 (label_ref ("SYM"))
14490 if (SET_DEST (curr_set
) == (pc_rtx
)
14491 && GET_CODE (SET_SRC (curr_set
)) == IF_THEN_ELSE
14492 && REG_P (XEXP (XEXP (SET_SRC (curr_set
), 0), 0))
14493 && REG_P (SET_DEST (prev_set
))
14494 && REGNO (SET_DEST (prev_set
))
14495 == REGNO (XEXP (XEXP (SET_SRC (curr_set
), 0), 0)))
14497 /* Fuse ALU operations followed by conditional branch instruction. */
14498 switch (get_attr_type (prev
))
14501 case TYPE_ALU_SREG
:
14504 case TYPE_ADCS_REG
:
14505 case TYPE_ADCS_IMM
:
14506 case TYPE_LOGIC_REG
:
14507 case TYPE_LOGIC_IMM
:
14511 case TYPE_SHIFT_REG
:
14512 case TYPE_SHIFT_IMM
:
14527 /* Return true iff the instruction fusion described by OP is enabled. */
14530 aarch64_fusion_enabled_p (enum aarch64_fusion_pairs op
)
14532 return (aarch64_tune_params
.fusible_ops
& op
) != 0;
14535 /* If MEM is in the form of [base+offset], extract the two parts
14536 of address and set to BASE and OFFSET, otherwise return false
14537 after clearing BASE and OFFSET. */
14540 extract_base_offset_in_addr (rtx mem
, rtx
*base
, rtx
*offset
)
14544 gcc_assert (MEM_P (mem
));
14546 addr
= XEXP (mem
, 0);
14551 *offset
= const0_rtx
;
14555 if (GET_CODE (addr
) == PLUS
14556 && REG_P (XEXP (addr
, 0)) && CONST_INT_P (XEXP (addr
, 1)))
14558 *base
= XEXP (addr
, 0);
14559 *offset
= XEXP (addr
, 1);
14564 *offset
= NULL_RTX
;
14569 /* Types for scheduling fusion. */
14570 enum sched_fusion_type
14572 SCHED_FUSION_NONE
= 0,
14573 SCHED_FUSION_LD_SIGN_EXTEND
,
14574 SCHED_FUSION_LD_ZERO_EXTEND
,
14580 /* If INSN is a load or store of address in the form of [base+offset],
14581 extract the two parts and set to BASE and OFFSET. Return scheduling
14582 fusion type this INSN is. */
14584 static enum sched_fusion_type
14585 fusion_load_store (rtx_insn
*insn
, rtx
*base
, rtx
*offset
)
14588 enum sched_fusion_type fusion
= SCHED_FUSION_LD
;
14590 gcc_assert (INSN_P (insn
));
14591 x
= PATTERN (insn
);
14592 if (GET_CODE (x
) != SET
)
14593 return SCHED_FUSION_NONE
;
14596 dest
= SET_DEST (x
);
14598 machine_mode dest_mode
= GET_MODE (dest
);
14600 if (!aarch64_mode_valid_for_sched_fusion_p (dest_mode
))
14601 return SCHED_FUSION_NONE
;
14603 if (GET_CODE (src
) == SIGN_EXTEND
)
14605 fusion
= SCHED_FUSION_LD_SIGN_EXTEND
;
14606 src
= XEXP (src
, 0);
14607 if (GET_CODE (src
) != MEM
|| GET_MODE (src
) != SImode
)
14608 return SCHED_FUSION_NONE
;
14610 else if (GET_CODE (src
) == ZERO_EXTEND
)
14612 fusion
= SCHED_FUSION_LD_ZERO_EXTEND
;
14613 src
= XEXP (src
, 0);
14614 if (GET_CODE (src
) != MEM
|| GET_MODE (src
) != SImode
)
14615 return SCHED_FUSION_NONE
;
14618 if (GET_CODE (src
) == MEM
&& REG_P (dest
))
14619 extract_base_offset_in_addr (src
, base
, offset
);
14620 else if (GET_CODE (dest
) == MEM
&& (REG_P (src
) || src
== const0_rtx
))
14622 fusion
= SCHED_FUSION_ST
;
14623 extract_base_offset_in_addr (dest
, base
, offset
);
14626 return SCHED_FUSION_NONE
;
14628 if (*base
== NULL_RTX
|| *offset
== NULL_RTX
)
14629 fusion
= SCHED_FUSION_NONE
;
14634 /* Implement the TARGET_SCHED_FUSION_PRIORITY hook.
14636 Currently we only support to fuse ldr or str instructions, so FUSION_PRI
14637 and PRI are only calculated for these instructions. For other instruction,
14638 FUSION_PRI and PRI are simply set to MAX_PRI - 1. In the future, other
14639 type instruction fusion can be added by returning different priorities.
14641 It's important that irrelevant instructions get the largest FUSION_PRI. */
14644 aarch64_sched_fusion_priority (rtx_insn
*insn
, int max_pri
,
14645 int *fusion_pri
, int *pri
)
14649 enum sched_fusion_type fusion
;
14651 gcc_assert (INSN_P (insn
));
14654 fusion
= fusion_load_store (insn
, &base
, &offset
);
14655 if (fusion
== SCHED_FUSION_NONE
)
14662 /* Set FUSION_PRI according to fusion type and base register. */
14663 *fusion_pri
= tmp
- fusion
* FIRST_PSEUDO_REGISTER
- REGNO (base
);
14665 /* Calculate PRI. */
14668 /* INSN with smaller offset goes first. */
14669 off_val
= (int)(INTVAL (offset
));
14671 tmp
-= (off_val
& 0xfffff);
14673 tmp
+= ((- off_val
) & 0xfffff);
14679 /* Implement the TARGET_SCHED_ADJUST_PRIORITY hook.
14680 Adjust priority of sha1h instructions so they are scheduled before
14681 other SHA1 instructions. */
14684 aarch64_sched_adjust_priority (rtx_insn
*insn
, int priority
)
14686 rtx x
= PATTERN (insn
);
14688 if (GET_CODE (x
) == SET
)
14692 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_SHA1H
)
14693 return priority
+ 10;
14699 /* Given OPERANDS of consecutive load/store, check if we can merge
14700 them into ldp/stp. LOAD is true if they are load instructions.
14701 MODE is the mode of memory operands. */
14704 aarch64_operands_ok_for_ldpstp (rtx
*operands
, bool load
,
14707 HOST_WIDE_INT offval_1
, offval_2
, msize
;
14708 enum reg_class rclass_1
, rclass_2
;
14709 rtx mem_1
, mem_2
, reg_1
, reg_2
, base_1
, base_2
, offset_1
, offset_2
;
14713 mem_1
= operands
[1];
14714 mem_2
= operands
[3];
14715 reg_1
= operands
[0];
14716 reg_2
= operands
[2];
14717 gcc_assert (REG_P (reg_1
) && REG_P (reg_2
));
14718 if (REGNO (reg_1
) == REGNO (reg_2
))
14723 mem_1
= operands
[0];
14724 mem_2
= operands
[2];
14725 reg_1
= operands
[1];
14726 reg_2
= operands
[3];
14729 /* The mems cannot be volatile. */
14730 if (MEM_VOLATILE_P (mem_1
) || MEM_VOLATILE_P (mem_2
))
14733 /* If we have SImode and slow unaligned ldp,
14734 check the alignment to be at least 8 byte. */
14736 && (aarch64_tune_params
.extra_tuning_flags
14737 & AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW
)
14739 && MEM_ALIGN (mem_1
) < 8 * BITS_PER_UNIT
)
14742 /* Check if the addresses are in the form of [base+offset]. */
14743 extract_base_offset_in_addr (mem_1
, &base_1
, &offset_1
);
14744 if (base_1
== NULL_RTX
|| offset_1
== NULL_RTX
)
14746 extract_base_offset_in_addr (mem_2
, &base_2
, &offset_2
);
14747 if (base_2
== NULL_RTX
|| offset_2
== NULL_RTX
)
14750 /* Check if the bases are same. */
14751 if (!rtx_equal_p (base_1
, base_2
))
14754 offval_1
= INTVAL (offset_1
);
14755 offval_2
= INTVAL (offset_2
);
14756 msize
= GET_MODE_SIZE (mode
);
14757 /* Check if the offsets are consecutive. */
14758 if (offval_1
!= (offval_2
+ msize
) && offval_2
!= (offval_1
+ msize
))
14761 /* Check if the addresses are clobbered by load. */
14764 if (reg_mentioned_p (reg_1
, mem_1
))
14767 /* In increasing order, the last load can clobber the address. */
14768 if (offval_1
> offval_2
&& reg_mentioned_p (reg_2
, mem_2
))
14772 if (REG_P (reg_1
) && FP_REGNUM_P (REGNO (reg_1
)))
14773 rclass_1
= FP_REGS
;
14775 rclass_1
= GENERAL_REGS
;
14777 if (REG_P (reg_2
) && FP_REGNUM_P (REGNO (reg_2
)))
14778 rclass_2
= FP_REGS
;
14780 rclass_2
= GENERAL_REGS
;
14782 /* Check if the registers are of same class. */
14783 if (rclass_1
!= rclass_2
)
14789 /* Given OPERANDS of consecutive load/store, check if we can merge
14790 them into ldp/stp by adjusting the offset. LOAD is true if they
14791 are load instructions. MODE is the mode of memory operands.
14793 Given below consecutive stores:
14795 str w1, [xb, 0x100]
14796 str w1, [xb, 0x104]
14797 str w1, [xb, 0x108]
14798 str w1, [xb, 0x10c]
14800 Though the offsets are out of the range supported by stp, we can
14801 still pair them after adjusting the offset, like:
14803 add scratch, xb, 0x100
14804 stp w1, w1, [scratch]
14805 stp w1, w1, [scratch, 0x8]
14807 The peephole patterns detecting this opportunity should guarantee
14808 the scratch register is avaliable. */
14811 aarch64_operands_adjust_ok_for_ldpstp (rtx
*operands
, bool load
,
14814 enum reg_class rclass_1
, rclass_2
, rclass_3
, rclass_4
;
14815 HOST_WIDE_INT offval_1
, offval_2
, offval_3
, offval_4
, msize
;
14816 rtx mem_1
, mem_2
, mem_3
, mem_4
, reg_1
, reg_2
, reg_3
, reg_4
;
14817 rtx base_1
, base_2
, base_3
, base_4
, offset_1
, offset_2
, offset_3
, offset_4
;
14821 reg_1
= operands
[0];
14822 mem_1
= operands
[1];
14823 reg_2
= operands
[2];
14824 mem_2
= operands
[3];
14825 reg_3
= operands
[4];
14826 mem_3
= operands
[5];
14827 reg_4
= operands
[6];
14828 mem_4
= operands
[7];
14829 gcc_assert (REG_P (reg_1
) && REG_P (reg_2
)
14830 && REG_P (reg_3
) && REG_P (reg_4
));
14831 if (REGNO (reg_1
) == REGNO (reg_2
) || REGNO (reg_3
) == REGNO (reg_4
))
14836 mem_1
= operands
[0];
14837 reg_1
= operands
[1];
14838 mem_2
= operands
[2];
14839 reg_2
= operands
[3];
14840 mem_3
= operands
[4];
14841 reg_3
= operands
[5];
14842 mem_4
= operands
[6];
14843 reg_4
= operands
[7];
14845 /* Skip if memory operand is by itslef valid for ldp/stp. */
14846 if (!MEM_P (mem_1
) || aarch64_mem_pair_operand (mem_1
, mode
))
14849 /* The mems cannot be volatile. */
14850 if (MEM_VOLATILE_P (mem_1
) || MEM_VOLATILE_P (mem_2
)
14851 || MEM_VOLATILE_P (mem_3
) ||MEM_VOLATILE_P (mem_4
))
14854 /* Check if the addresses are in the form of [base+offset]. */
14855 extract_base_offset_in_addr (mem_1
, &base_1
, &offset_1
);
14856 if (base_1
== NULL_RTX
|| offset_1
== NULL_RTX
)
14858 extract_base_offset_in_addr (mem_2
, &base_2
, &offset_2
);
14859 if (base_2
== NULL_RTX
|| offset_2
== NULL_RTX
)
14861 extract_base_offset_in_addr (mem_3
, &base_3
, &offset_3
);
14862 if (base_3
== NULL_RTX
|| offset_3
== NULL_RTX
)
14864 extract_base_offset_in_addr (mem_4
, &base_4
, &offset_4
);
14865 if (base_4
== NULL_RTX
|| offset_4
== NULL_RTX
)
14868 /* Check if the bases are same. */
14869 if (!rtx_equal_p (base_1
, base_2
)
14870 || !rtx_equal_p (base_2
, base_3
)
14871 || !rtx_equal_p (base_3
, base_4
))
14874 offval_1
= INTVAL (offset_1
);
14875 offval_2
= INTVAL (offset_2
);
14876 offval_3
= INTVAL (offset_3
);
14877 offval_4
= INTVAL (offset_4
);
14878 msize
= GET_MODE_SIZE (mode
);
14879 /* Check if the offsets are consecutive. */
14880 if ((offval_1
!= (offval_2
+ msize
)
14881 || offval_1
!= (offval_3
+ msize
* 2)
14882 || offval_1
!= (offval_4
+ msize
* 3))
14883 && (offval_4
!= (offval_3
+ msize
)
14884 || offval_4
!= (offval_2
+ msize
* 2)
14885 || offval_4
!= (offval_1
+ msize
* 3)))
14888 /* Check if the addresses are clobbered by load. */
14891 if (reg_mentioned_p (reg_1
, mem_1
)
14892 || reg_mentioned_p (reg_2
, mem_2
)
14893 || reg_mentioned_p (reg_3
, mem_3
))
14896 /* In increasing order, the last load can clobber the address. */
14897 if (offval_1
> offval_2
&& reg_mentioned_p (reg_4
, mem_4
))
14901 /* If we have SImode and slow unaligned ldp,
14902 check the alignment to be at least 8 byte. */
14904 && (aarch64_tune_params
.extra_tuning_flags
14905 & AARCH64_EXTRA_TUNE_SLOW_UNALIGNED_LDPW
)
14907 && MEM_ALIGN (mem_1
) < 8 * BITS_PER_UNIT
)
14910 if (REG_P (reg_1
) && FP_REGNUM_P (REGNO (reg_1
)))
14911 rclass_1
= FP_REGS
;
14913 rclass_1
= GENERAL_REGS
;
14915 if (REG_P (reg_2
) && FP_REGNUM_P (REGNO (reg_2
)))
14916 rclass_2
= FP_REGS
;
14918 rclass_2
= GENERAL_REGS
;
14920 if (REG_P (reg_3
) && FP_REGNUM_P (REGNO (reg_3
)))
14921 rclass_3
= FP_REGS
;
14923 rclass_3
= GENERAL_REGS
;
14925 if (REG_P (reg_4
) && FP_REGNUM_P (REGNO (reg_4
)))
14926 rclass_4
= FP_REGS
;
14928 rclass_4
= GENERAL_REGS
;
14930 /* Check if the registers are of same class. */
14931 if (rclass_1
!= rclass_2
|| rclass_2
!= rclass_3
|| rclass_3
!= rclass_4
)
14937 /* Given OPERANDS of consecutive load/store, this function pairs them
14938 into ldp/stp after adjusting the offset. It depends on the fact
14939 that addresses of load/store instructions are in increasing order.
14940 MODE is the mode of memory operands. CODE is the rtl operator
14941 which should be applied to all memory operands, it's SIGN_EXTEND,
14942 ZERO_EXTEND or UNKNOWN. */
14945 aarch64_gen_adjusted_ldpstp (rtx
*operands
, bool load
,
14946 scalar_mode mode
, RTX_CODE code
)
14948 rtx base
, offset
, t1
, t2
;
14949 rtx mem_1
, mem_2
, mem_3
, mem_4
;
14950 HOST_WIDE_INT off_val
, abs_off
, adj_off
, new_off
, stp_off_limit
, msize
;
14954 mem_1
= operands
[1];
14955 mem_2
= operands
[3];
14956 mem_3
= operands
[5];
14957 mem_4
= operands
[7];
14961 mem_1
= operands
[0];
14962 mem_2
= operands
[2];
14963 mem_3
= operands
[4];
14964 mem_4
= operands
[6];
14965 gcc_assert (code
== UNKNOWN
);
14968 extract_base_offset_in_addr (mem_1
, &base
, &offset
);
14969 gcc_assert (base
!= NULL_RTX
&& offset
!= NULL_RTX
);
14971 /* Adjust offset thus it can fit in ldp/stp instruction. */
14972 msize
= GET_MODE_SIZE (mode
);
14973 stp_off_limit
= msize
* 0x40;
14974 off_val
= INTVAL (offset
);
14975 abs_off
= (off_val
< 0) ? -off_val
: off_val
;
14976 new_off
= abs_off
% stp_off_limit
;
14977 adj_off
= abs_off
- new_off
;
14979 /* Further adjust to make sure all offsets are OK. */
14980 if ((new_off
+ msize
* 2) >= stp_off_limit
)
14982 adj_off
+= stp_off_limit
;
14983 new_off
-= stp_off_limit
;
14986 /* Make sure the adjustment can be done with ADD/SUB instructions. */
14987 if (adj_off
>= 0x1000)
14992 adj_off
= -adj_off
;
14993 new_off
= -new_off
;
14996 /* Create new memory references. */
14997 mem_1
= change_address (mem_1
, VOIDmode
,
14998 plus_constant (DImode
, operands
[8], new_off
));
15000 /* Check if the adjusted address is OK for ldp/stp. */
15001 if (!aarch64_mem_pair_operand (mem_1
, mode
))
15004 msize
= GET_MODE_SIZE (mode
);
15005 mem_2
= change_address (mem_2
, VOIDmode
,
15006 plus_constant (DImode
,
15009 mem_3
= change_address (mem_3
, VOIDmode
,
15010 plus_constant (DImode
,
15012 new_off
+ msize
* 2));
15013 mem_4
= change_address (mem_4
, VOIDmode
,
15014 plus_constant (DImode
,
15016 new_off
+ msize
* 3));
15018 if (code
== ZERO_EXTEND
)
15020 mem_1
= gen_rtx_ZERO_EXTEND (DImode
, mem_1
);
15021 mem_2
= gen_rtx_ZERO_EXTEND (DImode
, mem_2
);
15022 mem_3
= gen_rtx_ZERO_EXTEND (DImode
, mem_3
);
15023 mem_4
= gen_rtx_ZERO_EXTEND (DImode
, mem_4
);
15025 else if (code
== SIGN_EXTEND
)
15027 mem_1
= gen_rtx_SIGN_EXTEND (DImode
, mem_1
);
15028 mem_2
= gen_rtx_SIGN_EXTEND (DImode
, mem_2
);
15029 mem_3
= gen_rtx_SIGN_EXTEND (DImode
, mem_3
);
15030 mem_4
= gen_rtx_SIGN_EXTEND (DImode
, mem_4
);
15035 operands
[1] = mem_1
;
15036 operands
[3] = mem_2
;
15037 operands
[5] = mem_3
;
15038 operands
[7] = mem_4
;
15042 operands
[0] = mem_1
;
15043 operands
[2] = mem_2
;
15044 operands
[4] = mem_3
;
15045 operands
[6] = mem_4
;
15048 /* Emit adjusting instruction. */
15049 emit_insn (gen_rtx_SET (operands
[8], plus_constant (DImode
, base
, adj_off
)));
15050 /* Emit ldp/stp instructions. */
15051 t1
= gen_rtx_SET (operands
[0], operands
[1]);
15052 t2
= gen_rtx_SET (operands
[2], operands
[3]);
15053 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, t1
, t2
)));
15054 t1
= gen_rtx_SET (operands
[4], operands
[5]);
15055 t2
= gen_rtx_SET (operands
[6], operands
[7]);
15056 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, t1
, t2
)));
15060 /* Return 1 if pseudo register should be created and used to hold
15061 GOT address for PIC code. */
15064 aarch64_use_pseudo_pic_reg (void)
15066 return aarch64_cmodel
== AARCH64_CMODEL_SMALL_SPIC
;
15069 /* Implement TARGET_UNSPEC_MAY_TRAP_P. */
15072 aarch64_unspec_may_trap_p (const_rtx x
, unsigned flags
)
15074 switch (XINT (x
, 1))
15076 case UNSPEC_GOTSMALLPIC
:
15077 case UNSPEC_GOTSMALLPIC28K
:
15078 case UNSPEC_GOTTINYPIC
:
15084 return default_unspec_may_trap_p (x
, flags
);
15088 /* If X is a positive CONST_DOUBLE with a value that is a power of 2
15089 return the log2 of that value. Otherwise return -1. */
15092 aarch64_fpconst_pow_of_2 (rtx x
)
15094 const REAL_VALUE_TYPE
*r
;
15096 if (!CONST_DOUBLE_P (x
))
15099 r
= CONST_DOUBLE_REAL_VALUE (x
);
15101 if (REAL_VALUE_NEGATIVE (*r
)
15102 || REAL_VALUE_ISNAN (*r
)
15103 || REAL_VALUE_ISINF (*r
)
15104 || !real_isinteger (r
, DFmode
))
15107 return exact_log2 (real_to_integer (r
));
15110 /* If X is a vector of equal CONST_DOUBLE values and that value is
15111 Y, return the aarch64_fpconst_pow_of_2 of Y. Otherwise return -1. */
15114 aarch64_vec_fpconst_pow_of_2 (rtx x
)
15116 if (GET_CODE (x
) != CONST_VECTOR
)
15119 if (GET_MODE_CLASS (GET_MODE (x
)) != MODE_VECTOR_FLOAT
)
15122 int firstval
= aarch64_fpconst_pow_of_2 (CONST_VECTOR_ELT (x
, 0));
15126 for (int i
= 1; i
< CONST_VECTOR_NUNITS (x
); i
++)
15127 if (aarch64_fpconst_pow_of_2 (CONST_VECTOR_ELT (x
, i
)) != firstval
)
15133 /* Implement TARGET_PROMOTED_TYPE to promote 16-bit floating point types
15136 __fp16 always promotes through this hook.
15137 _Float16 may promote if TARGET_FLT_EVAL_METHOD is 16, but we do that
15138 through the generic excess precision logic rather than here. */
15141 aarch64_promoted_type (const_tree t
)
15143 if (SCALAR_FLOAT_TYPE_P (t
)
15144 && TYPE_MAIN_VARIANT (t
) == aarch64_fp16_type_node
)
15145 return float_type_node
;
15150 /* Implement the TARGET_OPTAB_SUPPORTED_P hook. */
15153 aarch64_optab_supported_p (int op
, machine_mode mode1
, machine_mode
,
15154 optimization_type opt_type
)
15159 return opt_type
== OPTIMIZE_FOR_SPEED
&& use_rsqrt_p (mode1
);
15166 /* Implement TARGET_LIBGCC_FLOATING_POINT_MODE_SUPPORTED_P - return TRUE
15167 if MODE is HFmode, and punt to the generic implementation otherwise. */
15170 aarch64_libgcc_floating_mode_supported_p (scalar_float_mode mode
)
15172 return (mode
== HFmode
15174 : default_libgcc_floating_mode_supported_p (mode
));
15177 /* Implement TARGET_SCALAR_MODE_SUPPORTED_P - return TRUE
15178 if MODE is HFmode, and punt to the generic implementation otherwise. */
15181 aarch64_scalar_mode_supported_p (scalar_mode mode
)
15183 return (mode
== HFmode
15185 : default_scalar_mode_supported_p (mode
));
15188 /* Set the value of FLT_EVAL_METHOD.
15189 ISO/IEC TS 18661-3 defines two values that we'd like to make use of:
15191 0: evaluate all operations and constants, whose semantic type has at
15192 most the range and precision of type float, to the range and
15193 precision of float; evaluate all other operations and constants to
15194 the range and precision of the semantic type;
15196 N, where _FloatN is a supported interchange floating type
15197 evaluate all operations and constants, whose semantic type has at
15198 most the range and precision of _FloatN type, to the range and
15199 precision of the _FloatN type; evaluate all other operations and
15200 constants to the range and precision of the semantic type;
15202 If we have the ARMv8.2-A extensions then we support _Float16 in native
15203 precision, so we should set this to 16. Otherwise, we support the type,
15204 but want to evaluate expressions in float precision, so set this to
15207 static enum flt_eval_method
15208 aarch64_excess_precision (enum excess_precision_type type
)
15212 case EXCESS_PRECISION_TYPE_FAST
:
15213 case EXCESS_PRECISION_TYPE_STANDARD
:
15214 /* We can calculate either in 16-bit range and precision or
15215 32-bit range and precision. Make that decision based on whether
15216 we have native support for the ARMv8.2-A 16-bit floating-point
15217 instructions or not. */
15218 return (TARGET_FP_F16INST
15219 ? FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16
15220 : FLT_EVAL_METHOD_PROMOTE_TO_FLOAT
);
15221 case EXCESS_PRECISION_TYPE_IMPLICIT
:
15222 return FLT_EVAL_METHOD_PROMOTE_TO_FLOAT16
;
15224 gcc_unreachable ();
15226 return FLT_EVAL_METHOD_UNPREDICTABLE
;
15229 /* Implement TARGET_SCHED_CAN_SPECULATE_INSN. Return true if INSN can be
15230 scheduled for speculative execution. Reject the long-running division
15231 and square-root instructions. */
15234 aarch64_sched_can_speculate_insn (rtx_insn
*insn
)
15236 switch (get_attr_type (insn
))
15244 case TYPE_NEON_FP_SQRT_S
:
15245 case TYPE_NEON_FP_SQRT_D
:
15246 case TYPE_NEON_FP_SQRT_S_Q
:
15247 case TYPE_NEON_FP_SQRT_D_Q
:
15248 case TYPE_NEON_FP_DIV_S
:
15249 case TYPE_NEON_FP_DIV_D
:
15250 case TYPE_NEON_FP_DIV_S_Q
:
15251 case TYPE_NEON_FP_DIV_D_Q
:
15258 /* Target-specific selftests. */
15262 namespace selftest
{
15264 /* Selftest for the RTL loader.
15265 Verify that the RTL loader copes with a dump from
15266 print_rtx_function. This is essentially just a test that class
15267 function_reader can handle a real dump, but it also verifies
15268 that lookup_reg_by_dump_name correctly handles hard regs.
15269 The presence of hard reg names in the dump means that the test is
15270 target-specific, hence it is in this file. */
15273 aarch64_test_loading_full_dump ()
15275 rtl_dump_test
t (SELFTEST_LOCATION
, locate_file ("aarch64/times-two.rtl"));
15277 ASSERT_STREQ ("times_two", IDENTIFIER_POINTER (DECL_NAME (cfun
->decl
)));
15279 rtx_insn
*insn_1
= get_insn_by_uid (1);
15280 ASSERT_EQ (NOTE
, GET_CODE (insn_1
));
15282 rtx_insn
*insn_15
= get_insn_by_uid (15);
15283 ASSERT_EQ (INSN
, GET_CODE (insn_15
));
15284 ASSERT_EQ (USE
, GET_CODE (PATTERN (insn_15
)));
15286 /* Verify crtl->return_rtx. */
15287 ASSERT_EQ (REG
, GET_CODE (crtl
->return_rtx
));
15288 ASSERT_EQ (0, REGNO (crtl
->return_rtx
));
15289 ASSERT_EQ (SImode
, GET_MODE (crtl
->return_rtx
));
15292 /* Run all target-specific selftests. */
15295 aarch64_run_selftests (void)
15297 aarch64_test_loading_full_dump ();
15300 } // namespace selftest
15302 #endif /* #if CHECKING_P */
15304 #undef TARGET_ADDRESS_COST
15305 #define TARGET_ADDRESS_COST aarch64_address_cost
15307 /* This hook will determines whether unnamed bitfields affect the alignment
15308 of the containing structure. The hook returns true if the structure
15309 should inherit the alignment requirements of an unnamed bitfield's
15311 #undef TARGET_ALIGN_ANON_BITFIELD
15312 #define TARGET_ALIGN_ANON_BITFIELD hook_bool_void_true
15314 #undef TARGET_ASM_ALIGNED_DI_OP
15315 #define TARGET_ASM_ALIGNED_DI_OP "\t.xword\t"
15317 #undef TARGET_ASM_ALIGNED_HI_OP
15318 #define TARGET_ASM_ALIGNED_HI_OP "\t.hword\t"
15320 #undef TARGET_ASM_ALIGNED_SI_OP
15321 #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t"
15323 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
15324 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK \
15325 hook_bool_const_tree_hwi_hwi_const_tree_true
15327 #undef TARGET_ASM_FILE_START
15328 #define TARGET_ASM_FILE_START aarch64_start_file
15330 #undef TARGET_ASM_OUTPUT_MI_THUNK
15331 #define TARGET_ASM_OUTPUT_MI_THUNK aarch64_output_mi_thunk
15333 #undef TARGET_ASM_SELECT_RTX_SECTION
15334 #define TARGET_ASM_SELECT_RTX_SECTION aarch64_select_rtx_section
15336 #undef TARGET_ASM_TRAMPOLINE_TEMPLATE
15337 #define TARGET_ASM_TRAMPOLINE_TEMPLATE aarch64_asm_trampoline_template
15339 #undef TARGET_BUILD_BUILTIN_VA_LIST
15340 #define TARGET_BUILD_BUILTIN_VA_LIST aarch64_build_builtin_va_list
15342 #undef TARGET_CALLEE_COPIES
15343 #define TARGET_CALLEE_COPIES hook_bool_CUMULATIVE_ARGS_mode_tree_bool_false
15345 #undef TARGET_CAN_ELIMINATE
15346 #define TARGET_CAN_ELIMINATE aarch64_can_eliminate
15348 #undef TARGET_CAN_INLINE_P
15349 #define TARGET_CAN_INLINE_P aarch64_can_inline_p
15351 #undef TARGET_CANNOT_FORCE_CONST_MEM
15352 #define TARGET_CANNOT_FORCE_CONST_MEM aarch64_cannot_force_const_mem
15354 #undef TARGET_CASE_VALUES_THRESHOLD
15355 #define TARGET_CASE_VALUES_THRESHOLD aarch64_case_values_threshold
15357 #undef TARGET_CONDITIONAL_REGISTER_USAGE
15358 #define TARGET_CONDITIONAL_REGISTER_USAGE aarch64_conditional_register_usage
15360 /* Only the least significant bit is used for initialization guard
15362 #undef TARGET_CXX_GUARD_MASK_BIT
15363 #define TARGET_CXX_GUARD_MASK_BIT hook_bool_void_true
15365 #undef TARGET_C_MODE_FOR_SUFFIX
15366 #define TARGET_C_MODE_FOR_SUFFIX aarch64_c_mode_for_suffix
15368 #ifdef TARGET_BIG_ENDIAN_DEFAULT
15369 #undef TARGET_DEFAULT_TARGET_FLAGS
15370 #define TARGET_DEFAULT_TARGET_FLAGS (MASK_BIG_END)
15373 #undef TARGET_CLASS_MAX_NREGS
15374 #define TARGET_CLASS_MAX_NREGS aarch64_class_max_nregs
15376 #undef TARGET_BUILTIN_DECL
15377 #define TARGET_BUILTIN_DECL aarch64_builtin_decl
15379 #undef TARGET_BUILTIN_RECIPROCAL
15380 #define TARGET_BUILTIN_RECIPROCAL aarch64_builtin_reciprocal
15382 #undef TARGET_C_EXCESS_PRECISION
15383 #define TARGET_C_EXCESS_PRECISION aarch64_excess_precision
15385 #undef TARGET_EXPAND_BUILTIN
15386 #define TARGET_EXPAND_BUILTIN aarch64_expand_builtin
15388 #undef TARGET_EXPAND_BUILTIN_VA_START
15389 #define TARGET_EXPAND_BUILTIN_VA_START aarch64_expand_builtin_va_start
15391 #undef TARGET_FOLD_BUILTIN
15392 #define TARGET_FOLD_BUILTIN aarch64_fold_builtin
15394 #undef TARGET_FUNCTION_ARG
15395 #define TARGET_FUNCTION_ARG aarch64_function_arg
15397 #undef TARGET_FUNCTION_ARG_ADVANCE
15398 #define TARGET_FUNCTION_ARG_ADVANCE aarch64_function_arg_advance
15400 #undef TARGET_FUNCTION_ARG_BOUNDARY
15401 #define TARGET_FUNCTION_ARG_BOUNDARY aarch64_function_arg_boundary
15403 #undef TARGET_FUNCTION_ARG_PADDING
15404 #define TARGET_FUNCTION_ARG_PADDING aarch64_function_arg_padding
15406 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
15407 #define TARGET_FUNCTION_OK_FOR_SIBCALL aarch64_function_ok_for_sibcall
15409 #undef TARGET_FUNCTION_VALUE
15410 #define TARGET_FUNCTION_VALUE aarch64_function_value
15412 #undef TARGET_FUNCTION_VALUE_REGNO_P
15413 #define TARGET_FUNCTION_VALUE_REGNO_P aarch64_function_value_regno_p
15415 #undef TARGET_FRAME_POINTER_REQUIRED
15416 #define TARGET_FRAME_POINTER_REQUIRED aarch64_frame_pointer_required
15418 #undef TARGET_GIMPLE_FOLD_BUILTIN
15419 #define TARGET_GIMPLE_FOLD_BUILTIN aarch64_gimple_fold_builtin
15421 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
15422 #define TARGET_GIMPLIFY_VA_ARG_EXPR aarch64_gimplify_va_arg_expr
15424 #undef TARGET_INIT_BUILTINS
15425 #define TARGET_INIT_BUILTINS aarch64_init_builtins
15427 #undef TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS
15428 #define TARGET_IRA_CHANGE_PSEUDO_ALLOCNO_CLASS \
15429 aarch64_ira_change_pseudo_allocno_class
15431 #undef TARGET_LEGITIMATE_ADDRESS_P
15432 #define TARGET_LEGITIMATE_ADDRESS_P aarch64_legitimate_address_hook_p
15434 #undef TARGET_LEGITIMATE_CONSTANT_P
15435 #define TARGET_LEGITIMATE_CONSTANT_P aarch64_legitimate_constant_p
15437 #undef TARGET_LEGITIMIZE_ADDRESS_DISPLACEMENT
15438 #define TARGET_LEGITIMIZE_ADDRESS_DISPLACEMENT \
15439 aarch64_legitimize_address_displacement
15441 #undef TARGET_LIBGCC_CMP_RETURN_MODE
15442 #define TARGET_LIBGCC_CMP_RETURN_MODE aarch64_libgcc_cmp_return_mode
15444 #undef TARGET_LIBGCC_FLOATING_MODE_SUPPORTED_P
15445 #define TARGET_LIBGCC_FLOATING_MODE_SUPPORTED_P \
15446 aarch64_libgcc_floating_mode_supported_p
15448 #undef TARGET_MANGLE_TYPE
15449 #define TARGET_MANGLE_TYPE aarch64_mangle_type
15451 #undef TARGET_MEMORY_MOVE_COST
15452 #define TARGET_MEMORY_MOVE_COST aarch64_memory_move_cost
15454 #undef TARGET_MIN_DIVISIONS_FOR_RECIP_MUL
15455 #define TARGET_MIN_DIVISIONS_FOR_RECIP_MUL aarch64_min_divisions_for_recip_mul
15457 #undef TARGET_MUST_PASS_IN_STACK
15458 #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
15460 /* This target hook should return true if accesses to volatile bitfields
15461 should use the narrowest mode possible. It should return false if these
15462 accesses should use the bitfield container type. */
15463 #undef TARGET_NARROW_VOLATILE_BITFIELD
15464 #define TARGET_NARROW_VOLATILE_BITFIELD hook_bool_void_false
15466 #undef TARGET_OPTION_OVERRIDE
15467 #define TARGET_OPTION_OVERRIDE aarch64_override_options
15469 #undef TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE
15470 #define TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE \
15471 aarch64_override_options_after_change
15473 #undef TARGET_OPTION_SAVE
15474 #define TARGET_OPTION_SAVE aarch64_option_save
15476 #undef TARGET_OPTION_RESTORE
15477 #define TARGET_OPTION_RESTORE aarch64_option_restore
15479 #undef TARGET_OPTION_PRINT
15480 #define TARGET_OPTION_PRINT aarch64_option_print
15482 #undef TARGET_OPTION_VALID_ATTRIBUTE_P
15483 #define TARGET_OPTION_VALID_ATTRIBUTE_P aarch64_option_valid_attribute_p
15485 #undef TARGET_SET_CURRENT_FUNCTION
15486 #define TARGET_SET_CURRENT_FUNCTION aarch64_set_current_function
15488 #undef TARGET_PASS_BY_REFERENCE
15489 #define TARGET_PASS_BY_REFERENCE aarch64_pass_by_reference
15491 #undef TARGET_PREFERRED_RELOAD_CLASS
15492 #define TARGET_PREFERRED_RELOAD_CLASS aarch64_preferred_reload_class
15494 #undef TARGET_SCHED_REASSOCIATION_WIDTH
15495 #define TARGET_SCHED_REASSOCIATION_WIDTH aarch64_reassociation_width
15497 #undef TARGET_PROMOTED_TYPE
15498 #define TARGET_PROMOTED_TYPE aarch64_promoted_type
15500 #undef TARGET_SECONDARY_RELOAD
15501 #define TARGET_SECONDARY_RELOAD aarch64_secondary_reload
15503 #undef TARGET_SHIFT_TRUNCATION_MASK
15504 #define TARGET_SHIFT_TRUNCATION_MASK aarch64_shift_truncation_mask
15506 #undef TARGET_SETUP_INCOMING_VARARGS
15507 #define TARGET_SETUP_INCOMING_VARARGS aarch64_setup_incoming_varargs
15509 #undef TARGET_STRUCT_VALUE_RTX
15510 #define TARGET_STRUCT_VALUE_RTX aarch64_struct_value_rtx
15512 #undef TARGET_REGISTER_MOVE_COST
15513 #define TARGET_REGISTER_MOVE_COST aarch64_register_move_cost
15515 #undef TARGET_RETURN_IN_MEMORY
15516 #define TARGET_RETURN_IN_MEMORY aarch64_return_in_memory
15518 #undef TARGET_RETURN_IN_MSB
15519 #define TARGET_RETURN_IN_MSB aarch64_return_in_msb
15521 #undef TARGET_RTX_COSTS
15522 #define TARGET_RTX_COSTS aarch64_rtx_costs_wrapper
15524 #undef TARGET_SCALAR_MODE_SUPPORTED_P
15525 #define TARGET_SCALAR_MODE_SUPPORTED_P aarch64_scalar_mode_supported_p
15527 #undef TARGET_SCHED_ISSUE_RATE
15528 #define TARGET_SCHED_ISSUE_RATE aarch64_sched_issue_rate
15530 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
15531 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD \
15532 aarch64_sched_first_cycle_multipass_dfa_lookahead
15534 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD
15535 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD \
15536 aarch64_first_cycle_multipass_dfa_lookahead_guard
15538 #undef TARGET_SHRINK_WRAP_GET_SEPARATE_COMPONENTS
15539 #define TARGET_SHRINK_WRAP_GET_SEPARATE_COMPONENTS \
15540 aarch64_get_separate_components
15542 #undef TARGET_SHRINK_WRAP_COMPONENTS_FOR_BB
15543 #define TARGET_SHRINK_WRAP_COMPONENTS_FOR_BB \
15544 aarch64_components_for_bb
15546 #undef TARGET_SHRINK_WRAP_DISQUALIFY_COMPONENTS
15547 #define TARGET_SHRINK_WRAP_DISQUALIFY_COMPONENTS \
15548 aarch64_disqualify_components
15550 #undef TARGET_SHRINK_WRAP_EMIT_PROLOGUE_COMPONENTS
15551 #define TARGET_SHRINK_WRAP_EMIT_PROLOGUE_COMPONENTS \
15552 aarch64_emit_prologue_components
15554 #undef TARGET_SHRINK_WRAP_EMIT_EPILOGUE_COMPONENTS
15555 #define TARGET_SHRINK_WRAP_EMIT_EPILOGUE_COMPONENTS \
15556 aarch64_emit_epilogue_components
15558 #undef TARGET_SHRINK_WRAP_SET_HANDLED_COMPONENTS
15559 #define TARGET_SHRINK_WRAP_SET_HANDLED_COMPONENTS \
15560 aarch64_set_handled_components
15562 #undef TARGET_TRAMPOLINE_INIT
15563 #define TARGET_TRAMPOLINE_INIT aarch64_trampoline_init
15565 #undef TARGET_USE_BLOCKS_FOR_CONSTANT_P
15566 #define TARGET_USE_BLOCKS_FOR_CONSTANT_P aarch64_use_blocks_for_constant_p
15568 #undef TARGET_VECTOR_MODE_SUPPORTED_P
15569 #define TARGET_VECTOR_MODE_SUPPORTED_P aarch64_vector_mode_supported_p
15571 #undef TARGET_VECTORIZE_SUPPORT_VECTOR_MISALIGNMENT
15572 #define TARGET_VECTORIZE_SUPPORT_VECTOR_MISALIGNMENT \
15573 aarch64_builtin_support_vector_misalignment
15575 #undef TARGET_ARRAY_MODE_SUPPORTED_P
15576 #define TARGET_ARRAY_MODE_SUPPORTED_P aarch64_array_mode_supported_p
15578 #undef TARGET_VECTORIZE_ADD_STMT_COST
15579 #define TARGET_VECTORIZE_ADD_STMT_COST aarch64_add_stmt_cost
15581 #undef TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST
15582 #define TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST \
15583 aarch64_builtin_vectorization_cost
15585 #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE
15586 #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE aarch64_preferred_simd_mode
15588 #undef TARGET_VECTORIZE_BUILTINS
15589 #define TARGET_VECTORIZE_BUILTINS
15591 #undef TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION
15592 #define TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION \
15593 aarch64_builtin_vectorized_function
15595 #undef TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES
15596 #define TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES \
15597 aarch64_autovectorize_vector_sizes
15599 #undef TARGET_ATOMIC_ASSIGN_EXPAND_FENV
15600 #define TARGET_ATOMIC_ASSIGN_EXPAND_FENV \
15601 aarch64_atomic_assign_expand_fenv
15603 /* Section anchor support. */
15605 #undef TARGET_MIN_ANCHOR_OFFSET
15606 #define TARGET_MIN_ANCHOR_OFFSET -256
15608 /* Limit the maximum anchor offset to 4k-1, since that's the limit for a
15609 byte offset; we can do much more for larger data types, but have no way
15610 to determine the size of the access. We assume accesses are aligned. */
15611 #undef TARGET_MAX_ANCHOR_OFFSET
15612 #define TARGET_MAX_ANCHOR_OFFSET 4095
15614 #undef TARGET_VECTOR_ALIGNMENT
15615 #define TARGET_VECTOR_ALIGNMENT aarch64_simd_vector_alignment
15617 #undef TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE
15618 #define TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE \
15619 aarch64_simd_vector_alignment_reachable
15621 /* vec_perm support. */
15623 #undef TARGET_VECTORIZE_VEC_PERM_CONST_OK
15624 #define TARGET_VECTORIZE_VEC_PERM_CONST_OK \
15625 aarch64_vectorize_vec_perm_const_ok
15627 #undef TARGET_INIT_LIBFUNCS
15628 #define TARGET_INIT_LIBFUNCS aarch64_init_libfuncs
15630 #undef TARGET_FIXED_CONDITION_CODE_REGS
15631 #define TARGET_FIXED_CONDITION_CODE_REGS aarch64_fixed_condition_code_regs
15633 #undef TARGET_FLAGS_REGNUM
15634 #define TARGET_FLAGS_REGNUM CC_REGNUM
15636 #undef TARGET_CALL_FUSAGE_CONTAINS_NON_CALLEE_CLOBBERS
15637 #define TARGET_CALL_FUSAGE_CONTAINS_NON_CALLEE_CLOBBERS true
15639 #undef TARGET_ASAN_SHADOW_OFFSET
15640 #define TARGET_ASAN_SHADOW_OFFSET aarch64_asan_shadow_offset
15642 #undef TARGET_LEGITIMIZE_ADDRESS
15643 #define TARGET_LEGITIMIZE_ADDRESS aarch64_legitimize_address
15645 #undef TARGET_USE_BY_PIECES_INFRASTRUCTURE_P
15646 #define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \
15647 aarch64_use_by_pieces_infrastructure_p
15649 #undef TARGET_SCHED_CAN_SPECULATE_INSN
15650 #define TARGET_SCHED_CAN_SPECULATE_INSN aarch64_sched_can_speculate_insn
15652 #undef TARGET_CAN_USE_DOLOOP_P
15653 #define TARGET_CAN_USE_DOLOOP_P can_use_doloop_if_innermost
15655 #undef TARGET_SCHED_ADJUST_PRIORITY
15656 #define TARGET_SCHED_ADJUST_PRIORITY aarch64_sched_adjust_priority
15658 #undef TARGET_SCHED_MACRO_FUSION_P
15659 #define TARGET_SCHED_MACRO_FUSION_P aarch64_macro_fusion_p
15661 #undef TARGET_SCHED_MACRO_FUSION_PAIR_P
15662 #define TARGET_SCHED_MACRO_FUSION_PAIR_P aarch_macro_fusion_pair_p
15664 #undef TARGET_SCHED_FUSION_PRIORITY
15665 #define TARGET_SCHED_FUSION_PRIORITY aarch64_sched_fusion_priority
15667 #undef TARGET_UNSPEC_MAY_TRAP_P
15668 #define TARGET_UNSPEC_MAY_TRAP_P aarch64_unspec_may_trap_p
15670 #undef TARGET_USE_PSEUDO_PIC_REG
15671 #define TARGET_USE_PSEUDO_PIC_REG aarch64_use_pseudo_pic_reg
15673 #undef TARGET_PRINT_OPERAND
15674 #define TARGET_PRINT_OPERAND aarch64_print_operand
15676 #undef TARGET_PRINT_OPERAND_ADDRESS
15677 #define TARGET_PRINT_OPERAND_ADDRESS aarch64_print_operand_address
15679 #undef TARGET_OPTAB_SUPPORTED_P
15680 #define TARGET_OPTAB_SUPPORTED_P aarch64_optab_supported_p
15682 #undef TARGET_OMIT_STRUCT_RETURN_REG
15683 #define TARGET_OMIT_STRUCT_RETURN_REG true
15685 /* The architecture reserves bits 0 and 1 so use bit 2 for descriptors. */
15686 #undef TARGET_CUSTOM_FUNCTION_DESCRIPTORS
15687 #define TARGET_CUSTOM_FUNCTION_DESCRIPTORS 4
15689 #undef TARGET_HARD_REGNO_MODE_OK
15690 #define TARGET_HARD_REGNO_MODE_OK aarch64_hard_regno_mode_ok
15692 #undef TARGET_MODES_TIEABLE_P
15693 #define TARGET_MODES_TIEABLE_P aarch64_modes_tieable_p
15695 #undef TARGET_HARD_REGNO_CALL_PART_CLOBBERED
15696 #define TARGET_HARD_REGNO_CALL_PART_CLOBBERED \
15697 aarch64_hard_regno_call_part_clobbered
15700 #undef TARGET_RUN_TARGET_SELFTESTS
15701 #define TARGET_RUN_TARGET_SELFTESTS selftest::aarch64_run_selftests
15702 #endif /* #if CHECKING_P */
15704 struct gcc_target targetm
= TARGET_INITIALIZER
;
15706 #include "gt-aarch64.h"