Enable instruction fusion of dependent AESE; AESMC and AESD; AESIMC pairs.
This can give up to 2x speedup on many AArch64 implementations. Also model the crypto instructions on Cortex-A57 according to the Optimization Guide. gcc/ * config/aarch64/aarch64.c (cortexa53_tunings): Enable AES fusion. (cortexa57_tunings): Likewise. (cortexa72_tunings): Likewise. (arch_macro_fusion_pair_p): Add support for AES fusion. * config/aarch64/aarch64-fusion-pairs.def: Add AES_AESMC entry. * config/arm/aarch-common.c (aarch_crypto_can_dual_issue): Allow virtual registers before reload so early scheduling works. * config/arm/cortex-a57.md (cortex_a57_crypto_simple): Use correct latency and pipeline. (cortex_a57_crypto_complex): Likewise. (cortex_a57_crypto_xor): Likewise. (define_bypass): Add AES bypass. From-SVN: r233268
This commit is contained in:
parent
24a179f835
commit
00a8574af2
|
@ -1,3 +1,18 @@
|
|||
2016-02-10 Wilco Dijkstra <wdijkstr@arm.com>
|
||||
|
||||
* config/aarch64/aarch64.c (cortexa53_tunings): Enable AES fusion.
|
||||
(cortexa57_tunings): Likewise.
|
||||
(cortexa72_tunings): Likewise.
|
||||
(arch_macro_fusion_pair_p): Add support for AES fusion.
|
||||
* config/aarch64/aarch64-fusion-pairs.def: Add AES_AESMC entry.
|
||||
* config/arm/aarch-common.c (aarch_crypto_can_dual_issue):
|
||||
Allow virtual registers before reload so early scheduling works.
|
||||
* config/arm/cortex-a57.md (cortex_a57_crypto_simple): Use
|
||||
correct latency and pipeline.
|
||||
(cortex_a57_crypto_complex): Likewise.
|
||||
(cortex_a57_crypto_xor): Likewise.
|
||||
(define_bypass): Add AES bypass.
|
||||
|
||||
2016-02-10 Richard Biener <rguenther@suse.de>
|
||||
|
||||
PR tree-optimization/69726
|
||||
|
|
|
@ -33,4 +33,5 @@ AARCH64_FUSION_PAIR ("adrp+add", ADRP_ADD)
|
|||
AARCH64_FUSION_PAIR ("movk+movk", MOVK_MOVK)
|
||||
AARCH64_FUSION_PAIR ("adrp+ldr", ADRP_LDR)
|
||||
AARCH64_FUSION_PAIR ("cmp+branch", CMP_BRANCH)
|
||||
AARCH64_FUSION_PAIR ("aes+aesmc", AES_AESMC)
|
||||
|
||||
|
|
|
@ -451,7 +451,7 @@ static const struct tune_params cortexa53_tunings =
|
|||
&generic_branch_cost,
|
||||
4, /* memmov_cost */
|
||||
2, /* issue_rate */
|
||||
(AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD
|
||||
(AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD
|
||||
| AARCH64_FUSE_MOVK_MOVK | AARCH64_FUSE_ADRP_LDR), /* fusible_ops */
|
||||
8, /* function_align. */
|
||||
8, /* jump_align. */
|
||||
|
@ -476,7 +476,7 @@ static const struct tune_params cortexa57_tunings =
|
|||
&cortexa57_branch_cost,
|
||||
4, /* memmov_cost */
|
||||
3, /* issue_rate */
|
||||
(AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD
|
||||
(AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD
|
||||
| AARCH64_FUSE_MOVK_MOVK), /* fusible_ops */
|
||||
16, /* function_align. */
|
||||
8, /* jump_align. */
|
||||
|
@ -502,7 +502,7 @@ static const struct tune_params cortexa72_tunings =
|
|||
&generic_branch_cost,
|
||||
4, /* memmov_cost */
|
||||
3, /* issue_rate */
|
||||
(AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD
|
||||
(AARCH64_FUSE_AES_AESMC | AARCH64_FUSE_MOV_MOVK | AARCH64_FUSE_ADRP_ADD
|
||||
| AARCH64_FUSE_MOVK_MOVK), /* fusible_ops */
|
||||
16, /* function_align. */
|
||||
8, /* jump_align. */
|
||||
|
@ -13328,6 +13328,10 @@ aarch_macro_fusion_pair_p (rtx_insn *prev, rtx_insn *curr)
|
|||
}
|
||||
}
|
||||
|
||||
if ((aarch64_tune_params.fusible_ops & AARCH64_FUSE_AES_AESMC)
|
||||
&& aarch_crypto_can_dual_issue (prev, curr))
|
||||
return true;
|
||||
|
||||
if ((aarch64_tune_params.fusible_ops & AARCH64_FUSE_CMP_BRANCH)
|
||||
&& any_condjump_p (curr))
|
||||
{
|
||||
|
|
|
@ -58,8 +58,11 @@ aarch_crypto_can_dual_issue (rtx_insn *producer_insn, rtx_insn *consumer_insn)
|
|||
{
|
||||
unsigned int regno = REGNO (SET_DEST (producer_set));
|
||||
|
||||
return REGNO (SET_DEST (consumer_set)) == regno
|
||||
&& REGNO (XVECEXP (consumer_src, 0, 0)) == regno;
|
||||
/* Before reload the registers are virtual, so the destination of
|
||||
consumer_set doesn't need to match. */
|
||||
|
||||
return (REGNO (SET_DEST (consumer_set)) == regno || !reload_completed)
|
||||
&& REGNO (XVECEXP (consumer_src, 0, 0)) == regno;
|
||||
}
|
||||
|
||||
return 0;
|
||||
|
|
|
@ -747,20 +747,20 @@
|
|||
neon_fp_sqrt_s_q, neon_fp_sqrt_d_q"))
|
||||
"ca57_cx2_block*3")
|
||||
|
||||
(define_insn_reservation "cortex_a57_crypto_simple" 4
|
||||
(define_insn_reservation "cortex_a57_crypto_simple" 3
|
||||
(and (eq_attr "tune" "cortexa57")
|
||||
(eq_attr "type" "crypto_aese,crypto_aesmc,crypto_sha1_fast,crypto_sha256_fast"))
|
||||
"ca57_cx2")
|
||||
"ca57_cx1")
|
||||
|
||||
(define_insn_reservation "cortex_a57_crypto_complex" 7
|
||||
(define_insn_reservation "cortex_a57_crypto_complex" 6
|
||||
(and (eq_attr "tune" "cortexa57")
|
||||
(eq_attr "type" "crypto_sha1_slow,crypto_sha256_slow"))
|
||||
"ca57_cx2+(ca57_cx2_issue,ca57_cx2)")
|
||||
"ca57_cx1*2")
|
||||
|
||||
(define_insn_reservation "cortex_a57_crypto_xor" 7
|
||||
(define_insn_reservation "cortex_a57_crypto_xor" 6
|
||||
(and (eq_attr "tune" "cortexa57")
|
||||
(eq_attr "type" "crypto_sha1_xor"))
|
||||
"(ca57_cx1+ca57_cx2)")
|
||||
"(ca57_cx1*2)|(ca57_cx2*2)")
|
||||
|
||||
;; We lie with calls. They take up all issue slots, but are otherwise
|
||||
;; not harmful.
|
||||
|
@ -797,3 +797,8 @@
|
|||
(define_bypass 1 "cortex_a57_*"
|
||||
"cortex_a57_call,cortex_a57_branch")
|
||||
|
||||
;; AESE+AESMC and AESD+AESIMC pairs forward with zero latency
|
||||
(define_bypass 0 "cortex_a57_crypto_simple"
|
||||
"cortex_a57_crypto_simple"
|
||||
"aarch_crypto_can_dual_issue")
|
||||
|
||||
|
|
Loading…
Reference in New Issue