@@ -367,7 +367,7 @@ $L$ChaCha20_ssse3:
367
367
ja NEAR $L$ChaCha20_4x
368
368
369
369
$L$do_sse3_after_all:
370
- sub rsp,64+40
370
+ sub rsp,64+168
371
371
movaps XMMWORD[(-40)+r9],xmm6
372
372
movaps XMMWORD[(-24)+r9],xmm7
373
373
$L$ssse3_body:
@@ -2378,9 +2378,17 @@ $L$ChaCha20_avx512:
2378
2378
cmp rdx,512
2379
2379
ja NEAR $L$ChaCha20_16x
2380
2380
2381
- sub rsp,64+40
2382
- movaps XMMWORD[(-40)+r9],xmm6
2383
- movaps XMMWORD[(-24)+r9],xmm7
2381
+ sub rsp,64+168
2382
+ movaps XMMWORD[(-168)+r9],xmm6
2383
+ movaps XMMWORD[(-152)+r9],xmm7
2384
+ movaps XMMWORD[(-136)+r9],xmm8
2385
+ movaps XMMWORD[(-120)+r9],xmm9
2386
+ movaps XMMWORD[(-104)+r9],xmm10
2387
+ movaps XMMWORD[(-88)+r9],xmm11
2388
+ movaps XMMWORD[(-72)+r9],xmm12
2389
+ movaps XMMWORD[(-56)+r9],xmm13
2390
+ movaps XMMWORD[(-40)+r9],xmm14
2391
+ movaps XMMWORD[(-24)+r9],xmm15
2384
2392
$L$avx512_body:
2385
2393
vbroadcasti32x4 zmm0,ZMMWORD[$L$sigma]
2386
2394
vbroadcasti32x4 zmm1,ZMMWORD[rcx]
@@ -2560,8 +2568,16 @@ $L$oop_tail_avx512:
2560
2568
2561
2569
$L$done_avx512:
2562
2570
vzeroall
2563
- movaps xmm6,XMMWORD[((-40))+r9]
2564
- movaps xmm7,XMMWORD[((-24))+r9]
2571
+ movaps xmm6,XMMWORD[((-168))+r9]
2572
+ movaps xmm7,XMMWORD[((-152))+r9]
2573
+ movaps xmm8,XMMWORD[((-136))+r9]
2574
+ movaps xmm9,XMMWORD[((-120))+r9]
2575
+ movaps xmm10,XMMWORD[((-104))+r9]
2576
+ movaps xmm11,XMMWORD[((-88))+r9]
2577
+ movaps xmm12,XMMWORD[((-72))+r9]
2578
+ movaps xmm13,XMMWORD[((-56))+r9]
2579
+ movaps xmm14,XMMWORD[((-40))+r9]
2580
+ movaps xmm15,XMMWORD[((-24))+r9]
2565
2581
lea rsp,[r9]
2566
2582
2567
2583
$L$avx512_epilogue:
@@ -2591,9 +2607,17 @@ $L$ChaCha20_avx512vl:
2591
2607
cmp rdx,128
2592
2608
ja NEAR $L$ChaCha20_8xvl
2593
2609
2594
- sub rsp,64+40
2595
- movaps XMMWORD[(-40)+r9],xmm6
2596
- movaps XMMWORD[(-24)+r9],xmm7
2610
+ sub rsp,64+168
2611
+ movaps XMMWORD[(-168)+r9],xmm6
2612
+ movaps XMMWORD[(-152)+r9],xmm7
2613
+ movaps XMMWORD[(-136)+r9],xmm8
2614
+ movaps XMMWORD[(-120)+r9],xmm9
2615
+ movaps XMMWORD[(-104)+r9],xmm10
2616
+ movaps XMMWORD[(-88)+r9],xmm11
2617
+ movaps XMMWORD[(-72)+r9],xmm12
2618
+ movaps XMMWORD[(-56)+r9],xmm13
2619
+ movaps XMMWORD[(-40)+r9],xmm14
2620
+ movaps XMMWORD[(-24)+r9],xmm15
2597
2621
$L$avx512vl_body:
2598
2622
vbroadcasti128 ymm0,XMMWORD[$L$sigma]
2599
2623
vbroadcasti128 ymm1,XMMWORD[rcx]
@@ -2730,8 +2754,16 @@ $L$oop_tail_avx512vl:
2730
2754
2731
2755
$L$done_avx512vl:
2732
2756
vzeroall
2733
- movaps xmm6,XMMWORD[((-40))+r9]
2734
- movaps xmm7,XMMWORD[((-24))+r9]
2757
+ movaps xmm6,XMMWORD[((-168))+r9]
2758
+ movaps xmm7,XMMWORD[((-152))+r9]
2759
+ movaps xmm8,XMMWORD[((-136))+r9]
2760
+ movaps xmm9,XMMWORD[((-120))+r9]
2761
+ movaps xmm10,XMMWORD[((-104))+r9]
2762
+ movaps xmm11,XMMWORD[((-88))+r9]
2763
+ movaps xmm12,XMMWORD[((-72))+r9]
2764
+ movaps xmm13,XMMWORD[((-56))+r9]
2765
+ movaps xmm14,XMMWORD[((-40))+r9]
2766
+ movaps xmm15,XMMWORD[((-24))+r9]
2735
2767
lea rsp,[r9]
2736
2768
2737
2769
$L$avx512vl_epilogue:
0 commit comments