Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(1379)

Unified Diff: third_party/boringssl/win-x86_64/crypto/sha/sha1-x86_64.asm

Issue 2219933002: Land BoringSSL roll on master (Closed) Base URL: git@github.com:dart-lang/sdk.git@master
Patch Set: Created 4 years, 4 months ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View side-by-side diff with in-line comments
Download patch
Index: third_party/boringssl/win-x86_64/crypto/sha/sha1-x86_64.asm
diff --git a/third_party/boringssl/win-x86_64/crypto/sha/sha1-x86_64.asm b/third_party/boringssl/win-x86_64/crypto/sha/sha1-x86_64.asm
index 0f5361a1b54d5e5dab0a18bb790020c572a771f5..168f78db3f491345fbaf5e296298db00e13c559d 100644
--- a/third_party/boringssl/win-x86_64/crypto/sha/sha1-x86_64.asm
+++ b/third_party/boringssl/win-x86_64/crypto/sha/sha1-x86_64.asm
@@ -24,6 +24,11 @@ $L$SEH_begin_sha1_block_data_order:
mov r10d,DWORD[((OPENSSL_ia32cap_P+8))]
test r8d,512
jz NEAR $L$ialu
+ and r8d,268435456
+ and r9d,1073741824
+ or r8d,r9d
+ cmp r8d,1342177280
+ je NEAR _avx_shortcut
jmp NEAR _ssse3_shortcut
ALIGN 16
@@ -2445,6 +2450,1146 @@ $L$epilogue_ssse3:
mov rsi,QWORD[16+rsp]
DB 0F3h,0C3h ;repret
$L$SEH_end_sha1_block_data_order_ssse3:
+
+ALIGN 16
+sha1_block_data_order_avx:
+ mov QWORD[8+rsp],rdi ;WIN64 prologue
+ mov QWORD[16+rsp],rsi
+ mov rax,rsp
+$L$SEH_begin_sha1_block_data_order_avx:
+ mov rdi,rcx
+ mov rsi,rdx
+ mov rdx,r8
+
+
+_avx_shortcut:
+ mov rax,rsp
+ push rbx
+ push rbp
+ push r12
+ push r13
+ push r14
+ lea rsp,[((-160))+rsp]
+ vzeroupper
+ vmovaps XMMWORD[(-40-96)+rax],xmm6
+ vmovaps XMMWORD[(-40-80)+rax],xmm7
+ vmovaps XMMWORD[(-40-64)+rax],xmm8
+ vmovaps XMMWORD[(-40-48)+rax],xmm9
+ vmovaps XMMWORD[(-40-32)+rax],xmm10
+ vmovaps XMMWORD[(-40-16)+rax],xmm11
+$L$prologue_avx:
+ mov r14,rax
+ and rsp,-64
+ mov r8,rdi
+ mov r9,rsi
+ mov r10,rdx
+
+ shl r10,6
+ add r10,r9
+ lea r11,[((K_XX_XX+64))]
+
+ mov eax,DWORD[r8]
+ mov ebx,DWORD[4+r8]
+ mov ecx,DWORD[8+r8]
+ mov edx,DWORD[12+r8]
+ mov esi,ebx
+ mov ebp,DWORD[16+r8]
+ mov edi,ecx
+ xor edi,edx
+ and esi,edi
+
+ vmovdqa xmm6,XMMWORD[64+r11]
+ vmovdqa xmm11,XMMWORD[((-64))+r11]
+ vmovdqu xmm0,XMMWORD[r9]
+ vmovdqu xmm1,XMMWORD[16+r9]
+ vmovdqu xmm2,XMMWORD[32+r9]
+ vmovdqu xmm3,XMMWORD[48+r9]
+ vpshufb xmm0,xmm0,xmm6
+ add r9,64
+ vpshufb xmm1,xmm1,xmm6
+ vpshufb xmm2,xmm2,xmm6
+ vpshufb xmm3,xmm3,xmm6
+ vpaddd xmm4,xmm0,xmm11
+ vpaddd xmm5,xmm1,xmm11
+ vpaddd xmm6,xmm2,xmm11
+ vmovdqa XMMWORD[rsp],xmm4
+ vmovdqa XMMWORD[16+rsp],xmm5
+ vmovdqa XMMWORD[32+rsp],xmm6
+ jmp NEAR $L$oop_avx
+ALIGN 16
+$L$oop_avx:
+ shrd ebx,ebx,2
+ xor esi,edx
+ vpalignr xmm4,xmm1,xmm0,8
+ mov edi,eax
+ add ebp,DWORD[rsp]
+ vpaddd xmm9,xmm11,xmm3
+ xor ebx,ecx
+ shld eax,eax,5
+ vpsrldq xmm8,xmm3,4
+ add ebp,esi
+ and edi,ebx
+ vpxor xmm4,xmm4,xmm0
+ xor ebx,ecx
+ add ebp,eax
+ vpxor xmm8,xmm8,xmm2
+ shrd eax,eax,7
+ xor edi,ecx
+ mov esi,ebp
+ add edx,DWORD[4+rsp]
+ vpxor xmm4,xmm4,xmm8
+ xor eax,ebx
+ shld ebp,ebp,5
+ vmovdqa XMMWORD[48+rsp],xmm9
+ add edx,edi
+ and esi,eax
+ vpsrld xmm8,xmm4,31
+ xor eax,ebx
+ add edx,ebp
+ shrd ebp,ebp,7
+ xor esi,ebx
+ vpslldq xmm10,xmm4,12
+ vpaddd xmm4,xmm4,xmm4
+ mov edi,edx
+ add ecx,DWORD[8+rsp]
+ xor ebp,eax
+ shld edx,edx,5
+ vpsrld xmm9,xmm10,30
+ vpor xmm4,xmm4,xmm8
+ add ecx,esi
+ and edi,ebp
+ xor ebp,eax
+ add ecx,edx
+ vpslld xmm10,xmm10,2
+ vpxor xmm4,xmm4,xmm9
+ shrd edx,edx,7
+ xor edi,eax
+ mov esi,ecx
+ add ebx,DWORD[12+rsp]
+ vpxor xmm4,xmm4,xmm10
+ xor edx,ebp
+ shld ecx,ecx,5
+ add ebx,edi
+ and esi,edx
+ xor edx,ebp
+ add ebx,ecx
+ shrd ecx,ecx,7
+ xor esi,ebp
+ vpalignr xmm5,xmm2,xmm1,8
+ mov edi,ebx
+ add eax,DWORD[16+rsp]
+ vpaddd xmm9,xmm11,xmm4
+ xor ecx,edx
+ shld ebx,ebx,5
+ vpsrldq xmm8,xmm4,4
+ add eax,esi
+ and edi,ecx
+ vpxor xmm5,xmm5,xmm1
+ xor ecx,edx
+ add eax,ebx
+ vpxor xmm8,xmm8,xmm3
+ shrd ebx,ebx,7
+ xor edi,edx
+ mov esi,eax
+ add ebp,DWORD[20+rsp]
+ vpxor xmm5,xmm5,xmm8
+ xor ebx,ecx
+ shld eax,eax,5
+ vmovdqa XMMWORD[rsp],xmm9
+ add ebp,edi
+ and esi,ebx
+ vpsrld xmm8,xmm5,31
+ xor ebx,ecx
+ add ebp,eax
+ shrd eax,eax,7
+ xor esi,ecx
+ vpslldq xmm10,xmm5,12
+ vpaddd xmm5,xmm5,xmm5
+ mov edi,ebp
+ add edx,DWORD[24+rsp]
+ xor eax,ebx
+ shld ebp,ebp,5
+ vpsrld xmm9,xmm10,30
+ vpor xmm5,xmm5,xmm8
+ add edx,esi
+ and edi,eax
+ xor eax,ebx
+ add edx,ebp
+ vpslld xmm10,xmm10,2
+ vpxor xmm5,xmm5,xmm9
+ shrd ebp,ebp,7
+ xor edi,ebx
+ mov esi,edx
+ add ecx,DWORD[28+rsp]
+ vpxor xmm5,xmm5,xmm10
+ xor ebp,eax
+ shld edx,edx,5
+ vmovdqa xmm11,XMMWORD[((-32))+r11]
+ add ecx,edi
+ and esi,ebp
+ xor ebp,eax
+ add ecx,edx
+ shrd edx,edx,7
+ xor esi,eax
+ vpalignr xmm6,xmm3,xmm2,8
+ mov edi,ecx
+ add ebx,DWORD[32+rsp]
+ vpaddd xmm9,xmm11,xmm5
+ xor edx,ebp
+ shld ecx,ecx,5
+ vpsrldq xmm8,xmm5,4
+ add ebx,esi
+ and edi,edx
+ vpxor xmm6,xmm6,xmm2
+ xor edx,ebp
+ add ebx,ecx
+ vpxor xmm8,xmm8,xmm4
+ shrd ecx,ecx,7
+ xor edi,ebp
+ mov esi,ebx
+ add eax,DWORD[36+rsp]
+ vpxor xmm6,xmm6,xmm8
+ xor ecx,edx
+ shld ebx,ebx,5
+ vmovdqa XMMWORD[16+rsp],xmm9
+ add eax,edi
+ and esi,ecx
+ vpsrld xmm8,xmm6,31
+ xor ecx,edx
+ add eax,ebx
+ shrd ebx,ebx,7
+ xor esi,edx
+ vpslldq xmm10,xmm6,12
+ vpaddd xmm6,xmm6,xmm6
+ mov edi,eax
+ add ebp,DWORD[40+rsp]
+ xor ebx,ecx
+ shld eax,eax,5
+ vpsrld xmm9,xmm10,30
+ vpor xmm6,xmm6,xmm8
+ add ebp,esi
+ and edi,ebx
+ xor ebx,ecx
+ add ebp,eax
+ vpslld xmm10,xmm10,2
+ vpxor xmm6,xmm6,xmm9
+ shrd eax,eax,7
+ xor edi,ecx
+ mov esi,ebp
+ add edx,DWORD[44+rsp]
+ vpxor xmm6,xmm6,xmm10
+ xor eax,ebx
+ shld ebp,ebp,5
+ add edx,edi
+ and esi,eax
+ xor eax,ebx
+ add edx,ebp
+ shrd ebp,ebp,7
+ xor esi,ebx
+ vpalignr xmm7,xmm4,xmm3,8
+ mov edi,edx
+ add ecx,DWORD[48+rsp]
+ vpaddd xmm9,xmm11,xmm6
+ xor ebp,eax
+ shld edx,edx,5
+ vpsrldq xmm8,xmm6,4
+ add ecx,esi
+ and edi,ebp
+ vpxor xmm7,xmm7,xmm3
+ xor ebp,eax
+ add ecx,edx
+ vpxor xmm8,xmm8,xmm5
+ shrd edx,edx,7
+ xor edi,eax
+ mov esi,ecx
+ add ebx,DWORD[52+rsp]
+ vpxor xmm7,xmm7,xmm8
+ xor edx,ebp
+ shld ecx,ecx,5
+ vmovdqa XMMWORD[32+rsp],xmm9
+ add ebx,edi
+ and esi,edx
+ vpsrld xmm8,xmm7,31
+ xor edx,ebp
+ add ebx,ecx
+ shrd ecx,ecx,7
+ xor esi,ebp
+ vpslldq xmm10,xmm7,12
+ vpaddd xmm7,xmm7,xmm7
+ mov edi,ebx
+ add eax,DWORD[56+rsp]
+ xor ecx,edx
+ shld ebx,ebx,5
+ vpsrld xmm9,xmm10,30
+ vpor xmm7,xmm7,xmm8
+ add eax,esi
+ and edi,ecx
+ xor ecx,edx
+ add eax,ebx
+ vpslld xmm10,xmm10,2
+ vpxor xmm7,xmm7,xmm9
+ shrd ebx,ebx,7
+ xor edi,edx
+ mov esi,eax
+ add ebp,DWORD[60+rsp]
+ vpxor xmm7,xmm7,xmm10
+ xor ebx,ecx
+ shld eax,eax,5
+ add ebp,edi
+ and esi,ebx
+ xor ebx,ecx
+ add ebp,eax
+ vpalignr xmm8,xmm7,xmm6,8
+ vpxor xmm0,xmm0,xmm4
+ shrd eax,eax,7
+ xor esi,ecx
+ mov edi,ebp
+ add edx,DWORD[rsp]
+ vpxor xmm0,xmm0,xmm1
+ xor eax,ebx
+ shld ebp,ebp,5
+ vpaddd xmm9,xmm11,xmm7
+ add edx,esi
+ and edi,eax
+ vpxor xmm0,xmm0,xmm8
+ xor eax,ebx
+ add edx,ebp
+ shrd ebp,ebp,7
+ xor edi,ebx
+ vpsrld xmm8,xmm0,30
+ vmovdqa XMMWORD[48+rsp],xmm9
+ mov esi,edx
+ add ecx,DWORD[4+rsp]
+ xor ebp,eax
+ shld edx,edx,5
+ vpslld xmm0,xmm0,2
+ add ecx,edi
+ and esi,ebp
+ xor ebp,eax
+ add ecx,edx
+ shrd edx,edx,7
+ xor esi,eax
+ mov edi,ecx
+ add ebx,DWORD[8+rsp]
+ vpor xmm0,xmm0,xmm8
+ xor edx,ebp
+ shld ecx,ecx,5
+ add ebx,esi
+ and edi,edx
+ xor edx,ebp
+ add ebx,ecx
+ add eax,DWORD[12+rsp]
+ xor edi,ebp
+ mov esi,ebx
+ shld ebx,ebx,5
+ add eax,edi
+ xor esi,edx
+ shrd ecx,ecx,7
+ add eax,ebx
+ vpalignr xmm8,xmm0,xmm7,8
+ vpxor xmm1,xmm1,xmm5
+ add ebp,DWORD[16+rsp]
+ xor esi,ecx
+ mov edi,eax
+ shld eax,eax,5
+ vpxor xmm1,xmm1,xmm2
+ add ebp,esi
+ xor edi,ecx
+ vpaddd xmm9,xmm11,xmm0
+ shrd ebx,ebx,7
+ add ebp,eax
+ vpxor xmm1,xmm1,xmm8
+ add edx,DWORD[20+rsp]
+ xor edi,ebx
+ mov esi,ebp
+ shld ebp,ebp,5
+ vpsrld xmm8,xmm1,30
+ vmovdqa XMMWORD[rsp],xmm9
+ add edx,edi
+ xor esi,ebx
+ shrd eax,eax,7
+ add edx,ebp
+ vpslld xmm1,xmm1,2
+ add ecx,DWORD[24+rsp]
+ xor esi,eax
+ mov edi,edx
+ shld edx,edx,5
+ add ecx,esi
+ xor edi,eax
+ shrd ebp,ebp,7
+ add ecx,edx
+ vpor xmm1,xmm1,xmm8
+ add ebx,DWORD[28+rsp]
+ xor edi,ebp
+ mov esi,ecx
+ shld ecx,ecx,5
+ add ebx,edi
+ xor esi,ebp
+ shrd edx,edx,7
+ add ebx,ecx
+ vpalignr xmm8,xmm1,xmm0,8
+ vpxor xmm2,xmm2,xmm6
+ add eax,DWORD[32+rsp]
+ xor esi,edx
+ mov edi,ebx
+ shld ebx,ebx,5
+ vpxor xmm2,xmm2,xmm3
+ add eax,esi
+ xor edi,edx
+ vpaddd xmm9,xmm11,xmm1
+ vmovdqa xmm11,XMMWORD[r11]
+ shrd ecx,ecx,7
+ add eax,ebx
+ vpxor xmm2,xmm2,xmm8
+ add ebp,DWORD[36+rsp]
+ xor edi,ecx
+ mov esi,eax
+ shld eax,eax,5
+ vpsrld xmm8,xmm2,30
+ vmovdqa XMMWORD[16+rsp],xmm9
+ add ebp,edi
+ xor esi,ecx
+ shrd ebx,ebx,7
+ add ebp,eax
+ vpslld xmm2,xmm2,2
+ add edx,DWORD[40+rsp]
+ xor esi,ebx
+ mov edi,ebp
+ shld ebp,ebp,5
+ add edx,esi
+ xor edi,ebx
+ shrd eax,eax,7
+ add edx,ebp
+ vpor xmm2,xmm2,xmm8
+ add ecx,DWORD[44+rsp]
+ xor edi,eax
+ mov esi,edx
+ shld edx,edx,5
+ add ecx,edi
+ xor esi,eax
+ shrd ebp,ebp,7
+ add ecx,edx
+ vpalignr xmm8,xmm2,xmm1,8
+ vpxor xmm3,xmm3,xmm7
+ add ebx,DWORD[48+rsp]
+ xor esi,ebp
+ mov edi,ecx
+ shld ecx,ecx,5
+ vpxor xmm3,xmm3,xmm4
+ add ebx,esi
+ xor edi,ebp
+ vpaddd xmm9,xmm11,xmm2
+ shrd edx,edx,7
+ add ebx,ecx
+ vpxor xmm3,xmm3,xmm8
+ add eax,DWORD[52+rsp]
+ xor edi,edx
+ mov esi,ebx
+ shld ebx,ebx,5
+ vpsrld xmm8,xmm3,30
+ vmovdqa XMMWORD[32+rsp],xmm9
+ add eax,edi
+ xor esi,edx
+ shrd ecx,ecx,7
+ add eax,ebx
+ vpslld xmm3,xmm3,2
+ add ebp,DWORD[56+rsp]
+ xor esi,ecx
+ mov edi,eax
+ shld eax,eax,5
+ add ebp,esi
+ xor edi,ecx
+ shrd ebx,ebx,7
+ add ebp,eax
+ vpor xmm3,xmm3,xmm8
+ add edx,DWORD[60+rsp]
+ xor edi,ebx
+ mov esi,ebp
+ shld ebp,ebp,5
+ add edx,edi
+ xor esi,ebx
+ shrd eax,eax,7
+ add edx,ebp
+ vpalignr xmm8,xmm3,xmm2,8
+ vpxor xmm4,xmm4,xmm0
+ add ecx,DWORD[rsp]
+ xor esi,eax
+ mov edi,edx
+ shld edx,edx,5
+ vpxor xmm4,xmm4,xmm5
+ add ecx,esi
+ xor edi,eax
+ vpaddd xmm9,xmm11,xmm3
+ shrd ebp,ebp,7
+ add ecx,edx
+ vpxor xmm4,xmm4,xmm8
+ add ebx,DWORD[4+rsp]
+ xor edi,ebp
+ mov esi,ecx
+ shld ecx,ecx,5
+ vpsrld xmm8,xmm4,30
+ vmovdqa XMMWORD[48+rsp],xmm9
+ add ebx,edi
+ xor esi,ebp
+ shrd edx,edx,7
+ add ebx,ecx
+ vpslld xmm4,xmm4,2
+ add eax,DWORD[8+rsp]
+ xor esi,edx
+ mov edi,ebx
+ shld ebx,ebx,5
+ add eax,esi
+ xor edi,edx
+ shrd ecx,ecx,7
+ add eax,ebx
+ vpor xmm4,xmm4,xmm8
+ add ebp,DWORD[12+rsp]
+ xor edi,ecx
+ mov esi,eax
+ shld eax,eax,5
+ add ebp,edi
+ xor esi,ecx
+ shrd ebx,ebx,7
+ add ebp,eax
+ vpalignr xmm8,xmm4,xmm3,8
+ vpxor xmm5,xmm5,xmm1
+ add edx,DWORD[16+rsp]
+ xor esi,ebx
+ mov edi,ebp
+ shld ebp,ebp,5
+ vpxor xmm5,xmm5,xmm6
+ add edx,esi
+ xor edi,ebx
+ vpaddd xmm9,xmm11,xmm4
+ shrd eax,eax,7
+ add edx,ebp
+ vpxor xmm5,xmm5,xmm8
+ add ecx,DWORD[20+rsp]
+ xor edi,eax
+ mov esi,edx
+ shld edx,edx,5
+ vpsrld xmm8,xmm5,30
+ vmovdqa XMMWORD[rsp],xmm9
+ add ecx,edi
+ xor esi,eax
+ shrd ebp,ebp,7
+ add ecx,edx
+ vpslld xmm5,xmm5,2
+ add ebx,DWORD[24+rsp]
+ xor esi,ebp
+ mov edi,ecx
+ shld ecx,ecx,5
+ add ebx,esi
+ xor edi,ebp
+ shrd edx,edx,7
+ add ebx,ecx
+ vpor xmm5,xmm5,xmm8
+ add eax,DWORD[28+rsp]
+ shrd ecx,ecx,7
+ mov esi,ebx
+ xor edi,edx
+ shld ebx,ebx,5
+ add eax,edi
+ xor esi,ecx
+ xor ecx,edx
+ add eax,ebx
+ vpalignr xmm8,xmm5,xmm4,8
+ vpxor xmm6,xmm6,xmm2
+ add ebp,DWORD[32+rsp]
+ and esi,ecx
+ xor ecx,edx
+ shrd ebx,ebx,7
+ vpxor xmm6,xmm6,xmm7
+ mov edi,eax
+ xor esi,ecx
+ vpaddd xmm9,xmm11,xmm5
+ shld eax,eax,5
+ add ebp,esi
+ vpxor xmm6,xmm6,xmm8
+ xor edi,ebx
+ xor ebx,ecx
+ add ebp,eax
+ add edx,DWORD[36+rsp]
+ vpsrld xmm8,xmm6,30
+ vmovdqa XMMWORD[16+rsp],xmm9
+ and edi,ebx
+ xor ebx,ecx
+ shrd eax,eax,7
+ mov esi,ebp
+ vpslld xmm6,xmm6,2
+ xor edi,ebx
+ shld ebp,ebp,5
+ add edx,edi
+ xor esi,eax
+ xor eax,ebx
+ add edx,ebp
+ add ecx,DWORD[40+rsp]
+ and esi,eax
+ vpor xmm6,xmm6,xmm8
+ xor eax,ebx
+ shrd ebp,ebp,7
+ mov edi,edx
+ xor esi,eax
+ shld edx,edx,5
+ add ecx,esi
+ xor edi,ebp
+ xor ebp,eax
+ add ecx,edx
+ add ebx,DWORD[44+rsp]
+ and edi,ebp
+ xor ebp,eax
+ shrd edx,edx,7
+ mov esi,ecx
+ xor edi,ebp
+ shld ecx,ecx,5
+ add ebx,edi
+ xor esi,edx
+ xor edx,ebp
+ add ebx,ecx
+ vpalignr xmm8,xmm6,xmm5,8
+ vpxor xmm7,xmm7,xmm3
+ add eax,DWORD[48+rsp]
+ and esi,edx
+ xor edx,ebp
+ shrd ecx,ecx,7
+ vpxor xmm7,xmm7,xmm0
+ mov edi,ebx
+ xor esi,edx
+ vpaddd xmm9,xmm11,xmm6
+ vmovdqa xmm11,XMMWORD[32+r11]
+ shld ebx,ebx,5
+ add eax,esi
+ vpxor xmm7,xmm7,xmm8
+ xor edi,ecx
+ xor ecx,edx
+ add eax,ebx
+ add ebp,DWORD[52+rsp]
+ vpsrld xmm8,xmm7,30
+ vmovdqa XMMWORD[32+rsp],xmm9
+ and edi,ecx
+ xor ecx,edx
+ shrd ebx,ebx,7
+ mov esi,eax
+ vpslld xmm7,xmm7,2
+ xor edi,ecx
+ shld eax,eax,5
+ add ebp,edi
+ xor esi,ebx
+ xor ebx,ecx
+ add ebp,eax
+ add edx,DWORD[56+rsp]
+ and esi,ebx
+ vpor xmm7,xmm7,xmm8
+ xor ebx,ecx
+ shrd eax,eax,7
+ mov edi,ebp
+ xor esi,ebx
+ shld ebp,ebp,5
+ add edx,esi
+ xor edi,eax
+ xor eax,ebx
+ add edx,ebp
+ add ecx,DWORD[60+rsp]
+ and edi,eax
+ xor eax,ebx
+ shrd ebp,ebp,7
+ mov esi,edx
+ xor edi,eax
+ shld edx,edx,5
+ add ecx,edi
+ xor esi,ebp
+ xor ebp,eax
+ add ecx,edx
+ vpalignr xmm8,xmm7,xmm6,8
+ vpxor xmm0,xmm0,xmm4
+ add ebx,DWORD[rsp]
+ and esi,ebp
+ xor ebp,eax
+ shrd edx,edx,7
+ vpxor xmm0,xmm0,xmm1
+ mov edi,ecx
+ xor esi,ebp
+ vpaddd xmm9,xmm11,xmm7
+ shld ecx,ecx,5
+ add ebx,esi
+ vpxor xmm0,xmm0,xmm8
+ xor edi,edx
+ xor edx,ebp
+ add ebx,ecx
+ add eax,DWORD[4+rsp]
+ vpsrld xmm8,xmm0,30
+ vmovdqa XMMWORD[48+rsp],xmm9
+ and edi,edx
+ xor edx,ebp
+ shrd ecx,ecx,7
+ mov esi,ebx
+ vpslld xmm0,xmm0,2
+ xor edi,edx
+ shld ebx,ebx,5
+ add eax,edi
+ xor esi,ecx
+ xor ecx,edx
+ add eax,ebx
+ add ebp,DWORD[8+rsp]
+ and esi,ecx
+ vpor xmm0,xmm0,xmm8
+ xor ecx,edx
+ shrd ebx,ebx,7
+ mov edi,eax
+ xor esi,ecx
+ shld eax,eax,5
+ add ebp,esi
+ xor edi,ebx
+ xor ebx,ecx
+ add ebp,eax
+ add edx,DWORD[12+rsp]
+ and edi,ebx
+ xor ebx,ecx
+ shrd eax,eax,7
+ mov esi,ebp
+ xor edi,ebx
+ shld ebp,ebp,5
+ add edx,edi
+ xor esi,eax
+ xor eax,ebx
+ add edx,ebp
+ vpalignr xmm8,xmm0,xmm7,8
+ vpxor xmm1,xmm1,xmm5
+ add ecx,DWORD[16+rsp]
+ and esi,eax
+ xor eax,ebx
+ shrd ebp,ebp,7
+ vpxor xmm1,xmm1,xmm2
+ mov edi,edx
+ xor esi,eax
+ vpaddd xmm9,xmm11,xmm0
+ shld edx,edx,5
+ add ecx,esi
+ vpxor xmm1,xmm1,xmm8
+ xor edi,ebp
+ xor ebp,eax
+ add ecx,edx
+ add ebx,DWORD[20+rsp]
+ vpsrld xmm8,xmm1,30
+ vmovdqa XMMWORD[rsp],xmm9
+ and edi,ebp
+ xor ebp,eax
+ shrd edx,edx,7
+ mov esi,ecx
+ vpslld xmm1,xmm1,2
+ xor edi,ebp
+ shld ecx,ecx,5
+ add ebx,edi
+ xor esi,edx
+ xor edx,ebp
+ add ebx,ecx
+ add eax,DWORD[24+rsp]
+ and esi,edx
+ vpor xmm1,xmm1,xmm8
+ xor edx,ebp
+ shrd ecx,ecx,7
+ mov edi,ebx
+ xor esi,edx
+ shld ebx,ebx,5
+ add eax,esi
+ xor edi,ecx
+ xor ecx,edx
+ add eax,ebx
+ add ebp,DWORD[28+rsp]
+ and edi,ecx
+ xor ecx,edx
+ shrd ebx,ebx,7
+ mov esi,eax
+ xor edi,ecx
+ shld eax,eax,5
+ add ebp,edi
+ xor esi,ebx
+ xor ebx,ecx
+ add ebp,eax
+ vpalignr xmm8,xmm1,xmm0,8
+ vpxor xmm2,xmm2,xmm6
+ add edx,DWORD[32+rsp]
+ and esi,ebx
+ xor ebx,ecx
+ shrd eax,eax,7
+ vpxor xmm2,xmm2,xmm3
+ mov edi,ebp
+ xor esi,ebx
+ vpaddd xmm9,xmm11,xmm1
+ shld ebp,ebp,5
+ add edx,esi
+ vpxor xmm2,xmm2,xmm8
+ xor edi,eax
+ xor eax,ebx
+ add edx,ebp
+ add ecx,DWORD[36+rsp]
+ vpsrld xmm8,xmm2,30
+ vmovdqa XMMWORD[16+rsp],xmm9
+ and edi,eax
+ xor eax,ebx
+ shrd ebp,ebp,7
+ mov esi,edx
+ vpslld xmm2,xmm2,2
+ xor edi,eax
+ shld edx,edx,5
+ add ecx,edi
+ xor esi,ebp
+ xor ebp,eax
+ add ecx,edx
+ add ebx,DWORD[40+rsp]
+ and esi,ebp
+ vpor xmm2,xmm2,xmm8
+ xor ebp,eax
+ shrd edx,edx,7
+ mov edi,ecx
+ xor esi,ebp
+ shld ecx,ecx,5
+ add ebx,esi
+ xor edi,edx
+ xor edx,ebp
+ add ebx,ecx
+ add eax,DWORD[44+rsp]
+ and edi,edx
+ xor edx,ebp
+ shrd ecx,ecx,7
+ mov esi,ebx
+ xor edi,edx
+ shld ebx,ebx,5
+ add eax,edi
+ xor esi,edx
+ add eax,ebx
+ vpalignr xmm8,xmm2,xmm1,8
+ vpxor xmm3,xmm3,xmm7
+ add ebp,DWORD[48+rsp]
+ xor esi,ecx
+ mov edi,eax
+ shld eax,eax,5
+ vpxor xmm3,xmm3,xmm4
+ add ebp,esi
+ xor edi,ecx
+ vpaddd xmm9,xmm11,xmm2
+ shrd ebx,ebx,7
+ add ebp,eax
+ vpxor xmm3,xmm3,xmm8
+ add edx,DWORD[52+rsp]
+ xor edi,ebx
+ mov esi,ebp
+ shld ebp,ebp,5
+ vpsrld xmm8,xmm3,30
+ vmovdqa XMMWORD[32+rsp],xmm9
+ add edx,edi
+ xor esi,ebx
+ shrd eax,eax,7
+ add edx,ebp
+ vpslld xmm3,xmm3,2
+ add ecx,DWORD[56+rsp]
+ xor esi,eax
+ mov edi,edx
+ shld edx,edx,5
+ add ecx,esi
+ xor edi,eax
+ shrd ebp,ebp,7
+ add ecx,edx
+ vpor xmm3,xmm3,xmm8
+ add ebx,DWORD[60+rsp]
+ xor edi,ebp
+ mov esi,ecx
+ shld ecx,ecx,5
+ add ebx,edi
+ xor esi,ebp
+ shrd edx,edx,7
+ add ebx,ecx
+ add eax,DWORD[rsp]
+ vpaddd xmm9,xmm11,xmm3
+ xor esi,edx
+ mov edi,ebx
+ shld ebx,ebx,5
+ add eax,esi
+ vmovdqa XMMWORD[48+rsp],xmm9
+ xor edi,edx
+ shrd ecx,ecx,7
+ add eax,ebx
+ add ebp,DWORD[4+rsp]
+ xor edi,ecx
+ mov esi,eax
+ shld eax,eax,5
+ add ebp,edi
+ xor esi,ecx
+ shrd ebx,ebx,7
+ add ebp,eax
+ add edx,DWORD[8+rsp]
+ xor esi,ebx
+ mov edi,ebp
+ shld ebp,ebp,5
+ add edx,esi
+ xor edi,ebx
+ shrd eax,eax,7
+ add edx,ebp
+ add ecx,DWORD[12+rsp]
+ xor edi,eax
+ mov esi,edx
+ shld edx,edx,5
+ add ecx,edi
+ xor esi,eax
+ shrd ebp,ebp,7
+ add ecx,edx
+ cmp r9,r10
+ je NEAR $L$done_avx
+ vmovdqa xmm6,XMMWORD[64+r11]
+ vmovdqa xmm11,XMMWORD[((-64))+r11]
+ vmovdqu xmm0,XMMWORD[r9]
+ vmovdqu xmm1,XMMWORD[16+r9]
+ vmovdqu xmm2,XMMWORD[32+r9]
+ vmovdqu xmm3,XMMWORD[48+r9]
+ vpshufb xmm0,xmm0,xmm6
+ add r9,64
+ add ebx,DWORD[16+rsp]
+ xor esi,ebp
+ vpshufb xmm1,xmm1,xmm6
+ mov edi,ecx
+ shld ecx,ecx,5
+ vpaddd xmm4,xmm0,xmm11
+ add ebx,esi
+ xor edi,ebp
+ shrd edx,edx,7
+ add ebx,ecx
+ vmovdqa XMMWORD[rsp],xmm4
+ add eax,DWORD[20+rsp]
+ xor edi,edx
+ mov esi,ebx
+ shld ebx,ebx,5
+ add eax,edi
+ xor esi,edx
+ shrd ecx,ecx,7
+ add eax,ebx
+ add ebp,DWORD[24+rsp]
+ xor esi,ecx
+ mov edi,eax
+ shld eax,eax,5
+ add ebp,esi
+ xor edi,ecx
+ shrd ebx,ebx,7
+ add ebp,eax
+ add edx,DWORD[28+rsp]
+ xor edi,ebx
+ mov esi,ebp
+ shld ebp,ebp,5
+ add edx,edi
+ xor esi,ebx
+ shrd eax,eax,7
+ add edx,ebp
+ add ecx,DWORD[32+rsp]
+ xor esi,eax
+ vpshufb xmm2,xmm2,xmm6
+ mov edi,edx
+ shld edx,edx,5
+ vpaddd xmm5,xmm1,xmm11
+ add ecx,esi
+ xor edi,eax
+ shrd ebp,ebp,7
+ add ecx,edx
+ vmovdqa XMMWORD[16+rsp],xmm5
+ add ebx,DWORD[36+rsp]
+ xor edi,ebp
+ mov esi,ecx
+ shld ecx,ecx,5
+ add ebx,edi
+ xor esi,ebp
+ shrd edx,edx,7
+ add ebx,ecx
+ add eax,DWORD[40+rsp]
+ xor esi,edx
+ mov edi,ebx
+ shld ebx,ebx,5
+ add eax,esi
+ xor edi,edx
+ shrd ecx,ecx,7
+ add eax,ebx
+ add ebp,DWORD[44+rsp]
+ xor edi,ecx
+ mov esi,eax
+ shld eax,eax,5
+ add ebp,edi
+ xor esi,ecx
+ shrd ebx,ebx,7
+ add ebp,eax
+ add edx,DWORD[48+rsp]
+ xor esi,ebx
+ vpshufb xmm3,xmm3,xmm6
+ mov edi,ebp
+ shld ebp,ebp,5
+ vpaddd xmm6,xmm2,xmm11
+ add edx,esi
+ xor edi,ebx
+ shrd eax,eax,7
+ add edx,ebp
+ vmovdqa XMMWORD[32+rsp],xmm6
+ add ecx,DWORD[52+rsp]
+ xor edi,eax
+ mov esi,edx
+ shld edx,edx,5
+ add ecx,edi
+ xor esi,eax
+ shrd ebp,ebp,7
+ add ecx,edx
+ add ebx,DWORD[56+rsp]
+ xor esi,ebp
+ mov edi,ecx
+ shld ecx,ecx,5
+ add ebx,esi
+ xor edi,ebp
+ shrd edx,edx,7
+ add ebx,ecx
+ add eax,DWORD[60+rsp]
+ xor edi,edx
+ mov esi,ebx
+ shld ebx,ebx,5
+ add eax,edi
+ shrd ecx,ecx,7
+ add eax,ebx
+ add eax,DWORD[r8]
+ add esi,DWORD[4+r8]
+ add ecx,DWORD[8+r8]
+ add edx,DWORD[12+r8]
+ mov DWORD[r8],eax
+ add ebp,DWORD[16+r8]
+ mov DWORD[4+r8],esi
+ mov ebx,esi
+ mov DWORD[8+r8],ecx
+ mov edi,ecx
+ mov DWORD[12+r8],edx
+ xor edi,edx
+ mov DWORD[16+r8],ebp
+ and esi,edi
+ jmp NEAR $L$oop_avx
+
+ALIGN 16
+$L$done_avx:
+ add ebx,DWORD[16+rsp]
+ xor esi,ebp
+ mov edi,ecx
+ shld ecx,ecx,5
+ add ebx,esi
+ xor edi,ebp
+ shrd edx,edx,7
+ add ebx,ecx
+ add eax,DWORD[20+rsp]
+ xor edi,edx
+ mov esi,ebx
+ shld ebx,ebx,5
+ add eax,edi
+ xor esi,edx
+ shrd ecx,ecx,7
+ add eax,ebx
+ add ebp,DWORD[24+rsp]
+ xor esi,ecx
+ mov edi,eax
+ shld eax,eax,5
+ add ebp,esi
+ xor edi,ecx
+ shrd ebx,ebx,7
+ add ebp,eax
+ add edx,DWORD[28+rsp]
+ xor edi,ebx
+ mov esi,ebp
+ shld ebp,ebp,5
+ add edx,edi
+ xor esi,ebx
+ shrd eax,eax,7
+ add edx,ebp
+ add ecx,DWORD[32+rsp]
+ xor esi,eax
+ mov edi,edx
+ shld edx,edx,5
+ add ecx,esi
+ xor edi,eax
+ shrd ebp,ebp,7
+ add ecx,edx
+ add ebx,DWORD[36+rsp]
+ xor edi,ebp
+ mov esi,ecx
+ shld ecx,ecx,5
+ add ebx,edi
+ xor esi,ebp
+ shrd edx,edx,7
+ add ebx,ecx
+ add eax,DWORD[40+rsp]
+ xor esi,edx
+ mov edi,ebx
+ shld ebx,ebx,5
+ add eax,esi
+ xor edi,edx
+ shrd ecx,ecx,7
+ add eax,ebx
+ add ebp,DWORD[44+rsp]
+ xor edi,ecx
+ mov esi,eax
+ shld eax,eax,5
+ add ebp,edi
+ xor esi,ecx
+ shrd ebx,ebx,7
+ add ebp,eax
+ add edx,DWORD[48+rsp]
+ xor esi,ebx
+ mov edi,ebp
+ shld ebp,ebp,5
+ add edx,esi
+ xor edi,ebx
+ shrd eax,eax,7
+ add edx,ebp
+ add ecx,DWORD[52+rsp]
+ xor edi,eax
+ mov esi,edx
+ shld edx,edx,5
+ add ecx,edi
+ xor esi,eax
+ shrd ebp,ebp,7
+ add ecx,edx
+ add ebx,DWORD[56+rsp]
+ xor esi,ebp
+ mov edi,ecx
+ shld ecx,ecx,5
+ add ebx,esi
+ xor edi,ebp
+ shrd edx,edx,7
+ add ebx,ecx
+ add eax,DWORD[60+rsp]
+ xor edi,edx
+ mov esi,ebx
+ shld ebx,ebx,5
+ add eax,edi
+ shrd ecx,ecx,7
+ add eax,ebx
+ vzeroupper
+
+ add eax,DWORD[r8]
+ add esi,DWORD[4+r8]
+ add ecx,DWORD[8+r8]
+ mov DWORD[r8],eax
+ add edx,DWORD[12+r8]
+ mov DWORD[4+r8],esi
+ add ebp,DWORD[16+r8]
+ mov DWORD[8+r8],ecx
+ mov DWORD[12+r8],edx
+ mov DWORD[16+r8],ebp
+ movaps xmm6,XMMWORD[((-40-96))+r14]
+ movaps xmm7,XMMWORD[((-40-80))+r14]
+ movaps xmm8,XMMWORD[((-40-64))+r14]
+ movaps xmm9,XMMWORD[((-40-48))+r14]
+ movaps xmm10,XMMWORD[((-40-32))+r14]
+ movaps xmm11,XMMWORD[((-40-16))+r14]
+ lea rsi,[r14]
+ mov r14,QWORD[((-40))+rsi]
+ mov r13,QWORD[((-32))+rsi]
+ mov r12,QWORD[((-24))+rsi]
+ mov rbp,QWORD[((-16))+rsi]
+ mov rbx,QWORD[((-8))+rsi]
+ lea rsp,[rsi]
+$L$epilogue_avx:
+ mov rdi,QWORD[8+rsp] ;WIN64 epilogue
+ mov rsi,QWORD[16+rsp]
+ DB 0F3h,0C3h ;repret
+$L$SEH_end_sha1_block_data_order_avx:
ALIGN 64
K_XX_XX:
DD 0x5a827999,0x5a827999,0x5a827999,0x5a827999
@@ -2605,6 +3750,9 @@ ALIGN 4
DD $L$SEH_begin_sha1_block_data_order_ssse3 wrt ..imagebase
DD $L$SEH_end_sha1_block_data_order_ssse3 wrt ..imagebase
DD $L$SEH_info_sha1_block_data_order_ssse3 wrt ..imagebase
+ DD $L$SEH_begin_sha1_block_data_order_avx wrt ..imagebase
+ DD $L$SEH_end_sha1_block_data_order_avx wrt ..imagebase
+ DD $L$SEH_info_sha1_block_data_order_avx wrt ..imagebase
section .xdata rdata align=8
ALIGN 8
$L$SEH_info_sha1_block_data_order:
@@ -2614,3 +3762,7 @@ $L$SEH_info_sha1_block_data_order_ssse3:
DB 9,0,0,0
DD ssse3_handler wrt ..imagebase
DD $L$prologue_ssse3 wrt ..imagebase,$L$epilogue_ssse3 wrt ..imagebase
+$L$SEH_info_sha1_block_data_order_avx:
+DB 9,0,0,0
+ DD ssse3_handler wrt ..imagebase
+ DD $L$prologue_avx wrt ..imagebase,$L$epilogue_avx wrt ..imagebase

Powered by Google App Engine
This is Rietveld 408576698