diff options
Diffstat (limited to 'win-x86/crypto/sha/sha256-586.asm')
-rw-r--r-- | win-x86/crypto/sha/sha256-586.asm | 1218 |
1 files changed, 1200 insertions, 18 deletions
diff --git a/win-x86/crypto/sha/sha256-586.asm b/win-x86/crypto/sha/sha256-586.asm index d03558c..3e7cfcc 100644 --- a/win-x86/crypto/sha/sha256-586.asm +++ b/win-x86/crypto/sha/sha256-586.asm @@ -52,12 +52,13 @@ L$000pic_point: or ecx,ebx and ecx,1342177280 cmp ecx,1342177280 + je NEAR L$004AVX test ebx,512 - jnz NEAR L$004SSSE3 + jnz NEAR L$005SSSE3 L$003no_xmm: sub eax,edi cmp eax,256 - jae NEAR L$005unrolled + jae NEAR L$006unrolled jmp NEAR L$002loop align 16 L$002loop: @@ -129,7 +130,7 @@ L$002loop: mov DWORD [28+esp],ecx mov DWORD [32+esp],edi align 16 -L$00600_15: +L$00700_15: mov ecx,edx mov esi,DWORD [24+esp] ror ecx,14 @@ -167,11 +168,11 @@ L$00600_15: add ebp,4 add eax,ebx cmp esi,3248222580 - jne NEAR L$00600_15 + jne NEAR L$00700_15 mov ecx,DWORD [156+esp] - jmp NEAR L$00716_63 + jmp NEAR L$00816_63 align 16 -L$00716_63: +L$00816_63: mov ebx,ecx mov esi,DWORD [104+esp] ror ecx,11 @@ -226,7 +227,7 @@ L$00716_63: add ebp,4 add eax,ebx cmp esi,3329325298 - jne NEAR L$00716_63 + jne NEAR L$00816_63 mov esi,DWORD [356+esp] mov ebx,DWORD [8+esp] mov ecx,DWORD [16+esp] @@ -270,7 +271,7 @@ db 67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97 db 112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103 db 62,0 align 16 -L$005unrolled: +L$006unrolled: lea esp,[esp-96] mov eax,DWORD [esi] mov ebp,DWORD [4+esi] @@ -287,9 +288,9 @@ L$005unrolled: mov DWORD [20+esp],ebx mov DWORD [24+esp],ecx mov DWORD [28+esp],esi - jmp NEAR L$008grand_loop + jmp NEAR L$009grand_loop align 16 -L$008grand_loop: +L$009grand_loop: mov ebx,DWORD [edi] mov ecx,DWORD [4+edi] bswap ebx @@ -3169,7 +3170,7 @@ L$008grand_loop: mov DWORD [24+esp],ebx mov DWORD [28+esp],ecx cmp edi,DWORD [104+esp] - jb NEAR L$008grand_loop + jb NEAR L$009grand_loop mov esp,DWORD [108+esp] pop edi pop esi @@ -3177,7 +3178,7 @@ L$008grand_loop: pop ebp ret align 32 -L$004SSSE3: +L$005SSSE3: lea esp,[esp-96] mov eax,DWORD [esi] mov ebx,DWORD [4+esi] @@ -3196,9 +3197,9 @@ L$004SSSE3: mov DWORD [24+esp],ecx mov DWORD [28+esp],esi movdqa xmm7,[256+ebp] - jmp NEAR L$009grand_ssse3 + jmp NEAR L$010grand_ssse3 align 16 -L$009grand_ssse3: +L$010grand_ssse3: movdqu xmm0,[edi] movdqu xmm1,[16+edi] movdqu xmm2,[32+edi] @@ -3221,9 +3222,9 @@ db 102,15,56,0,223 paddd xmm7,xmm3 movdqa [64+esp],xmm6 movdqa [80+esp],xmm7 - jmp NEAR L$010ssse3_00_47 + jmp NEAR L$011ssse3_00_47 align 16 -L$010ssse3_00_47: +L$011ssse3_00_47: add ebp,64 mov ecx,edx movdqa xmm4,xmm1 @@ -3866,7 +3867,7 @@ db 102,15,58,15,249,4 add eax,ecx movdqa [80+esp],xmm6 cmp DWORD [64+ebp],66051 - jne NEAR L$010ssse3_00_47 + jne NEAR L$011ssse3_00_47 mov ecx,edx ror edx,14 mov esi,DWORD [20+esp] @@ -4380,12 +4381,1193 @@ db 102,15,58,15,249,4 movdqa xmm7,[64+ebp] sub ebp,192 cmp edi,DWORD [104+esp] - jb NEAR L$009grand_ssse3 + jb NEAR L$010grand_ssse3 mov esp,DWORD [108+esp] pop edi pop esi pop ebx pop ebp ret +align 32 +L$004AVX: + lea esp,[esp-96] + vzeroall + mov eax,DWORD [esi] + mov ebx,DWORD [4+esi] + mov ecx,DWORD [8+esi] + mov edi,DWORD [12+esi] + mov DWORD [4+esp],ebx + xor ebx,ecx + mov DWORD [8+esp],ecx + mov DWORD [12+esp],edi + mov edx,DWORD [16+esi] + mov edi,DWORD [20+esi] + mov ecx,DWORD [24+esi] + mov esi,DWORD [28+esi] + mov DWORD [20+esp],edi + mov edi,DWORD [100+esp] + mov DWORD [24+esp],ecx + mov DWORD [28+esp],esi + vmovdqa xmm7,[256+ebp] + jmp NEAR L$012grand_avx +align 32 +L$012grand_avx: + vmovdqu xmm0,[edi] + vmovdqu xmm1,[16+edi] + vmovdqu xmm2,[32+edi] + vmovdqu xmm3,[48+edi] + add edi,64 + vpshufb xmm0,xmm0,xmm7 + mov DWORD [100+esp],edi + vpshufb xmm1,xmm1,xmm7 + vpshufb xmm2,xmm2,xmm7 + vpaddd xmm4,xmm0,[ebp] + vpshufb xmm3,xmm3,xmm7 + vpaddd xmm5,xmm1,[16+ebp] + vpaddd xmm6,xmm2,[32+ebp] + vpaddd xmm7,xmm3,[48+ebp] + vmovdqa [32+esp],xmm4 + vmovdqa [48+esp],xmm5 + vmovdqa [64+esp],xmm6 + vmovdqa [80+esp],xmm7 + jmp NEAR L$013avx_00_47 +align 16 +L$013avx_00_47: + add ebp,64 + vpalignr xmm4,xmm1,xmm0,4 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [20+esp] + vpalignr xmm7,xmm3,xmm2,4 + xor edx,ecx + mov edi,DWORD [24+esp] + xor esi,edi + vpsrld xmm6,xmm4,7 + shrd edx,edx,5 + and esi,ecx + mov DWORD [16+esp],ecx + vpaddd xmm0,xmm0,xmm7 + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + vpsrld xmm7,xmm4,3 + mov ecx,eax + add edx,edi + mov edi,DWORD [4+esp] + vpslld xmm5,xmm4,14 + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [esp],eax + vpxor xmm4,xmm7,xmm6 + xor ecx,eax + xor eax,edi + add edx,DWORD [28+esp] + vpshufd xmm7,xmm3,250 + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + vpsrld xmm6,xmm6,11 + add edx,DWORD [32+esp] + xor ebx,edi + shrd ecx,ecx,2 + vpxor xmm4,xmm4,xmm5 + add ebx,edx + add edx,DWORD [12+esp] + add ebx,ecx + vpslld xmm5,xmm5,11 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [16+esp] + vpxor xmm4,xmm4,xmm6 + xor edx,ecx + mov edi,DWORD [20+esp] + xor esi,edi + vpsrld xmm6,xmm7,10 + shrd edx,edx,5 + and esi,ecx + mov DWORD [12+esp],ecx + vpxor xmm4,xmm4,xmm5 + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + vpsrlq xmm5,xmm7,17 + mov ecx,ebx + add edx,edi + mov edi,DWORD [esp] + vpaddd xmm0,xmm0,xmm4 + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [28+esp],ebx + vpxor xmm6,xmm6,xmm5 + xor ecx,ebx + xor ebx,edi + add edx,DWORD [24+esp] + vpsrlq xmm7,xmm7,19 + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + vpxor xmm6,xmm6,xmm7 + add edx,DWORD [36+esp] + xor eax,edi + shrd ecx,ecx,2 + vpshufd xmm7,xmm6,132 + add eax,edx + add edx,DWORD [8+esp] + add eax,ecx + vpsrldq xmm7,xmm7,8 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [12+esp] + vpaddd xmm0,xmm0,xmm7 + xor edx,ecx + mov edi,DWORD [16+esp] + xor esi,edi + vpshufd xmm7,xmm0,80 + shrd edx,edx,5 + and esi,ecx + mov DWORD [8+esp],ecx + vpsrld xmm6,xmm7,10 + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + vpsrlq xmm5,xmm7,17 + mov ecx,eax + add edx,edi + mov edi,DWORD [28+esp] + vpxor xmm6,xmm6,xmm5 + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [24+esp],eax + vpsrlq xmm7,xmm7,19 + xor ecx,eax + xor eax,edi + add edx,DWORD [20+esp] + vpxor xmm6,xmm6,xmm7 + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + vpshufd xmm7,xmm6,232 + add edx,DWORD [40+esp] + xor ebx,edi + shrd ecx,ecx,2 + vpslldq xmm7,xmm7,8 + add ebx,edx + add edx,DWORD [4+esp] + add ebx,ecx + vpaddd xmm0,xmm0,xmm7 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [8+esp] + vpaddd xmm6,xmm0,[ebp] + xor edx,ecx + mov edi,DWORD [12+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [4+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,ebx + add edx,edi + mov edi,DWORD [24+esp] + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [20+esp],ebx + xor ecx,ebx + xor ebx,edi + add edx,DWORD [16+esp] + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + add edx,DWORD [44+esp] + xor eax,edi + shrd ecx,ecx,2 + add eax,edx + add edx,DWORD [esp] + add eax,ecx + vmovdqa [32+esp],xmm6 + vpalignr xmm4,xmm2,xmm1,4 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [4+esp] + vpalignr xmm7,xmm0,xmm3,4 + xor edx,ecx + mov edi,DWORD [8+esp] + xor esi,edi + vpsrld xmm6,xmm4,7 + shrd edx,edx,5 + and esi,ecx + mov DWORD [esp],ecx + vpaddd xmm1,xmm1,xmm7 + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + vpsrld xmm7,xmm4,3 + mov ecx,eax + add edx,edi + mov edi,DWORD [20+esp] + vpslld xmm5,xmm4,14 + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [16+esp],eax + vpxor xmm4,xmm7,xmm6 + xor ecx,eax + xor eax,edi + add edx,DWORD [12+esp] + vpshufd xmm7,xmm0,250 + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + vpsrld xmm6,xmm6,11 + add edx,DWORD [48+esp] + xor ebx,edi + shrd ecx,ecx,2 + vpxor xmm4,xmm4,xmm5 + add ebx,edx + add edx,DWORD [28+esp] + add ebx,ecx + vpslld xmm5,xmm5,11 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [esp] + vpxor xmm4,xmm4,xmm6 + xor edx,ecx + mov edi,DWORD [4+esp] + xor esi,edi + vpsrld xmm6,xmm7,10 + shrd edx,edx,5 + and esi,ecx + mov DWORD [28+esp],ecx + vpxor xmm4,xmm4,xmm5 + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + vpsrlq xmm5,xmm7,17 + mov ecx,ebx + add edx,edi + mov edi,DWORD [16+esp] + vpaddd xmm1,xmm1,xmm4 + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [12+esp],ebx + vpxor xmm6,xmm6,xmm5 + xor ecx,ebx + xor ebx,edi + add edx,DWORD [8+esp] + vpsrlq xmm7,xmm7,19 + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + vpxor xmm6,xmm6,xmm7 + add edx,DWORD [52+esp] + xor eax,edi + shrd ecx,ecx,2 + vpshufd xmm7,xmm6,132 + add eax,edx + add edx,DWORD [24+esp] + add eax,ecx + vpsrldq xmm7,xmm7,8 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [28+esp] + vpaddd xmm1,xmm1,xmm7 + xor edx,ecx + mov edi,DWORD [esp] + xor esi,edi + vpshufd xmm7,xmm1,80 + shrd edx,edx,5 + and esi,ecx + mov DWORD [24+esp],ecx + vpsrld xmm6,xmm7,10 + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + vpsrlq xmm5,xmm7,17 + mov ecx,eax + add edx,edi + mov edi,DWORD [12+esp] + vpxor xmm6,xmm6,xmm5 + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [8+esp],eax + vpsrlq xmm7,xmm7,19 + xor ecx,eax + xor eax,edi + add edx,DWORD [4+esp] + vpxor xmm6,xmm6,xmm7 + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + vpshufd xmm7,xmm6,232 + add edx,DWORD [56+esp] + xor ebx,edi + shrd ecx,ecx,2 + vpslldq xmm7,xmm7,8 + add ebx,edx + add edx,DWORD [20+esp] + add ebx,ecx + vpaddd xmm1,xmm1,xmm7 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [24+esp] + vpaddd xmm6,xmm1,[16+ebp] + xor edx,ecx + mov edi,DWORD [28+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [20+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,ebx + add edx,edi + mov edi,DWORD [8+esp] + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [4+esp],ebx + xor ecx,ebx + xor ebx,edi + add edx,DWORD [esp] + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + add edx,DWORD [60+esp] + xor eax,edi + shrd ecx,ecx,2 + add eax,edx + add edx,DWORD [16+esp] + add eax,ecx + vmovdqa [48+esp],xmm6 + vpalignr xmm4,xmm3,xmm2,4 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [20+esp] + vpalignr xmm7,xmm1,xmm0,4 + xor edx,ecx + mov edi,DWORD [24+esp] + xor esi,edi + vpsrld xmm6,xmm4,7 + shrd edx,edx,5 + and esi,ecx + mov DWORD [16+esp],ecx + vpaddd xmm2,xmm2,xmm7 + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + vpsrld xmm7,xmm4,3 + mov ecx,eax + add edx,edi + mov edi,DWORD [4+esp] + vpslld xmm5,xmm4,14 + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [esp],eax + vpxor xmm4,xmm7,xmm6 + xor ecx,eax + xor eax,edi + add edx,DWORD [28+esp] + vpshufd xmm7,xmm1,250 + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + vpsrld xmm6,xmm6,11 + add edx,DWORD [64+esp] + xor ebx,edi + shrd ecx,ecx,2 + vpxor xmm4,xmm4,xmm5 + add ebx,edx + add edx,DWORD [12+esp] + add ebx,ecx + vpslld xmm5,xmm5,11 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [16+esp] + vpxor xmm4,xmm4,xmm6 + xor edx,ecx + mov edi,DWORD [20+esp] + xor esi,edi + vpsrld xmm6,xmm7,10 + shrd edx,edx,5 + and esi,ecx + mov DWORD [12+esp],ecx + vpxor xmm4,xmm4,xmm5 + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + vpsrlq xmm5,xmm7,17 + mov ecx,ebx + add edx,edi + mov edi,DWORD [esp] + vpaddd xmm2,xmm2,xmm4 + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [28+esp],ebx + vpxor xmm6,xmm6,xmm5 + xor ecx,ebx + xor ebx,edi + add edx,DWORD [24+esp] + vpsrlq xmm7,xmm7,19 + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + vpxor xmm6,xmm6,xmm7 + add edx,DWORD [68+esp] + xor eax,edi + shrd ecx,ecx,2 + vpshufd xmm7,xmm6,132 + add eax,edx + add edx,DWORD [8+esp] + add eax,ecx + vpsrldq xmm7,xmm7,8 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [12+esp] + vpaddd xmm2,xmm2,xmm7 + xor edx,ecx + mov edi,DWORD [16+esp] + xor esi,edi + vpshufd xmm7,xmm2,80 + shrd edx,edx,5 + and esi,ecx + mov DWORD [8+esp],ecx + vpsrld xmm6,xmm7,10 + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + vpsrlq xmm5,xmm7,17 + mov ecx,eax + add edx,edi + mov edi,DWORD [28+esp] + vpxor xmm6,xmm6,xmm5 + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [24+esp],eax + vpsrlq xmm7,xmm7,19 + xor ecx,eax + xor eax,edi + add edx,DWORD [20+esp] + vpxor xmm6,xmm6,xmm7 + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + vpshufd xmm7,xmm6,232 + add edx,DWORD [72+esp] + xor ebx,edi + shrd ecx,ecx,2 + vpslldq xmm7,xmm7,8 + add ebx,edx + add edx,DWORD [4+esp] + add ebx,ecx + vpaddd xmm2,xmm2,xmm7 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [8+esp] + vpaddd xmm6,xmm2,[32+ebp] + xor edx,ecx + mov edi,DWORD [12+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [4+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,ebx + add edx,edi + mov edi,DWORD [24+esp] + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [20+esp],ebx + xor ecx,ebx + xor ebx,edi + add edx,DWORD [16+esp] + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + add edx,DWORD [76+esp] + xor eax,edi + shrd ecx,ecx,2 + add eax,edx + add edx,DWORD [esp] + add eax,ecx + vmovdqa [64+esp],xmm6 + vpalignr xmm4,xmm0,xmm3,4 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [4+esp] + vpalignr xmm7,xmm2,xmm1,4 + xor edx,ecx + mov edi,DWORD [8+esp] + xor esi,edi + vpsrld xmm6,xmm4,7 + shrd edx,edx,5 + and esi,ecx + mov DWORD [esp],ecx + vpaddd xmm3,xmm3,xmm7 + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + vpsrld xmm7,xmm4,3 + mov ecx,eax + add edx,edi + mov edi,DWORD [20+esp] + vpslld xmm5,xmm4,14 + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [16+esp],eax + vpxor xmm4,xmm7,xmm6 + xor ecx,eax + xor eax,edi + add edx,DWORD [12+esp] + vpshufd xmm7,xmm2,250 + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + vpsrld xmm6,xmm6,11 + add edx,DWORD [80+esp] + xor ebx,edi + shrd ecx,ecx,2 + vpxor xmm4,xmm4,xmm5 + add ebx,edx + add edx,DWORD [28+esp] + add ebx,ecx + vpslld xmm5,xmm5,11 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [esp] + vpxor xmm4,xmm4,xmm6 + xor edx,ecx + mov edi,DWORD [4+esp] + xor esi,edi + vpsrld xmm6,xmm7,10 + shrd edx,edx,5 + and esi,ecx + mov DWORD [28+esp],ecx + vpxor xmm4,xmm4,xmm5 + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + vpsrlq xmm5,xmm7,17 + mov ecx,ebx + add edx,edi + mov edi,DWORD [16+esp] + vpaddd xmm3,xmm3,xmm4 + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [12+esp],ebx + vpxor xmm6,xmm6,xmm5 + xor ecx,ebx + xor ebx,edi + add edx,DWORD [8+esp] + vpsrlq xmm7,xmm7,19 + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + vpxor xmm6,xmm6,xmm7 + add edx,DWORD [84+esp] + xor eax,edi + shrd ecx,ecx,2 + vpshufd xmm7,xmm6,132 + add eax,edx + add edx,DWORD [24+esp] + add eax,ecx + vpsrldq xmm7,xmm7,8 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [28+esp] + vpaddd xmm3,xmm3,xmm7 + xor edx,ecx + mov edi,DWORD [esp] + xor esi,edi + vpshufd xmm7,xmm3,80 + shrd edx,edx,5 + and esi,ecx + mov DWORD [24+esp],ecx + vpsrld xmm6,xmm7,10 + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + vpsrlq xmm5,xmm7,17 + mov ecx,eax + add edx,edi + mov edi,DWORD [12+esp] + vpxor xmm6,xmm6,xmm5 + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [8+esp],eax + vpsrlq xmm7,xmm7,19 + xor ecx,eax + xor eax,edi + add edx,DWORD [4+esp] + vpxor xmm6,xmm6,xmm7 + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + vpshufd xmm7,xmm6,232 + add edx,DWORD [88+esp] + xor ebx,edi + shrd ecx,ecx,2 + vpslldq xmm7,xmm7,8 + add ebx,edx + add edx,DWORD [20+esp] + add ebx,ecx + vpaddd xmm3,xmm3,xmm7 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [24+esp] + vpaddd xmm6,xmm3,[48+ebp] + xor edx,ecx + mov edi,DWORD [28+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [20+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,ebx + add edx,edi + mov edi,DWORD [8+esp] + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [4+esp],ebx + xor ecx,ebx + xor ebx,edi + add edx,DWORD [esp] + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + add edx,DWORD [92+esp] + xor eax,edi + shrd ecx,ecx,2 + add eax,edx + add edx,DWORD [16+esp] + add eax,ecx + vmovdqa [80+esp],xmm6 + cmp DWORD [64+ebp],66051 + jne NEAR L$013avx_00_47 + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [20+esp] + xor edx,ecx + mov edi,DWORD [24+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [16+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,eax + add edx,edi + mov edi,DWORD [4+esp] + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [esp],eax + xor ecx,eax + xor eax,edi + add edx,DWORD [28+esp] + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + add edx,DWORD [32+esp] + xor ebx,edi + shrd ecx,ecx,2 + add ebx,edx + add edx,DWORD [12+esp] + add ebx,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [16+esp] + xor edx,ecx + mov edi,DWORD [20+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [12+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,ebx + add edx,edi + mov edi,DWORD [esp] + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [28+esp],ebx + xor ecx,ebx + xor ebx,edi + add edx,DWORD [24+esp] + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + add edx,DWORD [36+esp] + xor eax,edi + shrd ecx,ecx,2 + add eax,edx + add edx,DWORD [8+esp] + add eax,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [12+esp] + xor edx,ecx + mov edi,DWORD [16+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [8+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,eax + add edx,edi + mov edi,DWORD [28+esp] + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [24+esp],eax + xor ecx,eax + xor eax,edi + add edx,DWORD [20+esp] + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + add edx,DWORD [40+esp] + xor ebx,edi + shrd ecx,ecx,2 + add ebx,edx + add edx,DWORD [4+esp] + add ebx,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [8+esp] + xor edx,ecx + mov edi,DWORD [12+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [4+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,ebx + add edx,edi + mov edi,DWORD [24+esp] + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [20+esp],ebx + xor ecx,ebx + xor ebx,edi + add edx,DWORD [16+esp] + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + add edx,DWORD [44+esp] + xor eax,edi + shrd ecx,ecx,2 + add eax,edx + add edx,DWORD [esp] + add eax,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [4+esp] + xor edx,ecx + mov edi,DWORD [8+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,eax + add edx,edi + mov edi,DWORD [20+esp] + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [16+esp],eax + xor ecx,eax + xor eax,edi + add edx,DWORD [12+esp] + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + add edx,DWORD [48+esp] + xor ebx,edi + shrd ecx,ecx,2 + add ebx,edx + add edx,DWORD [28+esp] + add ebx,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [esp] + xor edx,ecx + mov edi,DWORD [4+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [28+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,ebx + add edx,edi + mov edi,DWORD [16+esp] + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [12+esp],ebx + xor ecx,ebx + xor ebx,edi + add edx,DWORD [8+esp] + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + add edx,DWORD [52+esp] + xor eax,edi + shrd ecx,ecx,2 + add eax,edx + add edx,DWORD [24+esp] + add eax,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [28+esp] + xor edx,ecx + mov edi,DWORD [esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [24+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,eax + add edx,edi + mov edi,DWORD [12+esp] + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [8+esp],eax + xor ecx,eax + xor eax,edi + add edx,DWORD [4+esp] + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + add edx,DWORD [56+esp] + xor ebx,edi + shrd ecx,ecx,2 + add ebx,edx + add edx,DWORD [20+esp] + add ebx,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [24+esp] + xor edx,ecx + mov edi,DWORD [28+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [20+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,ebx + add edx,edi + mov edi,DWORD [8+esp] + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [4+esp],ebx + xor ecx,ebx + xor ebx,edi + add edx,DWORD [esp] + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + add edx,DWORD [60+esp] + xor eax,edi + shrd ecx,ecx,2 + add eax,edx + add edx,DWORD [16+esp] + add eax,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [20+esp] + xor edx,ecx + mov edi,DWORD [24+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [16+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,eax + add edx,edi + mov edi,DWORD [4+esp] + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [esp],eax + xor ecx,eax + xor eax,edi + add edx,DWORD [28+esp] + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + add edx,DWORD [64+esp] + xor ebx,edi + shrd ecx,ecx,2 + add ebx,edx + add edx,DWORD [12+esp] + add ebx,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [16+esp] + xor edx,ecx + mov edi,DWORD [20+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [12+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,ebx + add edx,edi + mov edi,DWORD [esp] + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [28+esp],ebx + xor ecx,ebx + xor ebx,edi + add edx,DWORD [24+esp] + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + add edx,DWORD [68+esp] + xor eax,edi + shrd ecx,ecx,2 + add eax,edx + add edx,DWORD [8+esp] + add eax,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [12+esp] + xor edx,ecx + mov edi,DWORD [16+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [8+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,eax + add edx,edi + mov edi,DWORD [28+esp] + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [24+esp],eax + xor ecx,eax + xor eax,edi + add edx,DWORD [20+esp] + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + add edx,DWORD [72+esp] + xor ebx,edi + shrd ecx,ecx,2 + add ebx,edx + add edx,DWORD [4+esp] + add ebx,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [8+esp] + xor edx,ecx + mov edi,DWORD [12+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [4+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,ebx + add edx,edi + mov edi,DWORD [24+esp] + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [20+esp],ebx + xor ecx,ebx + xor ebx,edi + add edx,DWORD [16+esp] + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + add edx,DWORD [76+esp] + xor eax,edi + shrd ecx,ecx,2 + add eax,edx + add edx,DWORD [esp] + add eax,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [4+esp] + xor edx,ecx + mov edi,DWORD [8+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,eax + add edx,edi + mov edi,DWORD [20+esp] + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [16+esp],eax + xor ecx,eax + xor eax,edi + add edx,DWORD [12+esp] + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + add edx,DWORD [80+esp] + xor ebx,edi + shrd ecx,ecx,2 + add ebx,edx + add edx,DWORD [28+esp] + add ebx,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [esp] + xor edx,ecx + mov edi,DWORD [4+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [28+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,ebx + add edx,edi + mov edi,DWORD [16+esp] + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [12+esp],ebx + xor ecx,ebx + xor ebx,edi + add edx,DWORD [8+esp] + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + add edx,DWORD [84+esp] + xor eax,edi + shrd ecx,ecx,2 + add eax,edx + add edx,DWORD [24+esp] + add eax,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [28+esp] + xor edx,ecx + mov edi,DWORD [esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [24+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,eax + add edx,edi + mov edi,DWORD [12+esp] + mov esi,eax + shrd ecx,ecx,9 + mov DWORD [8+esp],eax + xor ecx,eax + xor eax,edi + add edx,DWORD [4+esp] + shrd ecx,ecx,11 + and ebx,eax + xor ecx,esi + add edx,DWORD [88+esp] + xor ebx,edi + shrd ecx,ecx,2 + add ebx,edx + add edx,DWORD [20+esp] + add ebx,ecx + mov ecx,edx + shrd edx,edx,14 + mov esi,DWORD [24+esp] + xor edx,ecx + mov edi,DWORD [28+esp] + xor esi,edi + shrd edx,edx,5 + and esi,ecx + mov DWORD [20+esp],ecx + xor edx,ecx + xor edi,esi + shrd edx,edx,6 + mov ecx,ebx + add edx,edi + mov edi,DWORD [8+esp] + mov esi,ebx + shrd ecx,ecx,9 + mov DWORD [4+esp],ebx + xor ecx,ebx + xor ebx,edi + add edx,DWORD [esp] + shrd ecx,ecx,11 + and eax,ebx + xor ecx,esi + add edx,DWORD [92+esp] + xor eax,edi + shrd ecx,ecx,2 + add eax,edx + add edx,DWORD [16+esp] + add eax,ecx + mov esi,DWORD [96+esp] + xor ebx,edi + mov ecx,DWORD [12+esp] + add eax,DWORD [esi] + add ebx,DWORD [4+esi] + add edi,DWORD [8+esi] + add ecx,DWORD [12+esi] + mov DWORD [esi],eax + mov DWORD [4+esi],ebx + mov DWORD [8+esi],edi + mov DWORD [12+esi],ecx + mov DWORD [4+esp],ebx + xor ebx,edi + mov DWORD [8+esp],edi + mov DWORD [12+esp],ecx + mov edi,DWORD [20+esp] + mov ecx,DWORD [24+esp] + add edx,DWORD [16+esi] + add edi,DWORD [20+esi] + add ecx,DWORD [24+esi] + mov DWORD [16+esi],edx + mov DWORD [20+esi],edi + mov DWORD [20+esp],edi + mov edi,DWORD [28+esp] + mov DWORD [24+esi],ecx + add edi,DWORD [28+esi] + mov DWORD [24+esp],ecx + mov DWORD [28+esi],edi + mov DWORD [28+esp],edi + mov edi,DWORD [100+esp] + vmovdqa xmm7,[64+ebp] + sub ebp,192 + cmp edi,DWORD [104+esp] + jb NEAR L$012grand_avx + mov esp,DWORD [108+esp] + vzeroall + pop edi + pop esi + pop ebx + pop ebp + ret segment .bss common _OPENSSL_ia32cap_P 16 |