Source release 14.0.0

This commit is contained in:
John W. Bruce
2018-05-16 17:35:40 -07:00
parent 31381a1311
commit 3ab70cec4e
2053 changed files with 1585838 additions and 4614 deletions

View File

@@ -0,0 +1,977 @@
%ifidn __OUTPUT_FORMAT__,obj
section code use32 class=code align=64
%elifidn __OUTPUT_FORMAT__,win32
%ifdef __YASM_VERSION_ID__
%if __YASM_VERSION_ID__ < 01010000h
%error yasm version 1.1.0 or later needed.
%endif
; Yasm automatically includes .00 and complains about redefining it.
; https://www.tortall.net/projects/yasm/manual/html/objfmt-win32-safeseh.html
%else
$@feat.00 equ 1
%endif
section .text code align=64
%else
section .text code
%endif
global _ChaCha20_ctr32
align 16
_ChaCha20_ctr32:
L$_ChaCha20_ctr32_begin:
push ebp
push ebx
push esi
push edi
xor eax,eax
cmp eax,DWORD [28+esp]
je NEAR L$000no_data
call L$pic_point
L$pic_point:
pop eax
lea ebp,[_OPENSSL_ia32cap_P]
test DWORD [ebp],16777216
jz NEAR L$001x86
test DWORD [4+ebp],512
jz NEAR L$001x86
jmp NEAR L$ssse3_shortcut
L$001x86:
mov esi,DWORD [32+esp]
mov edi,DWORD [36+esp]
sub esp,132
mov eax,DWORD [esi]
mov ebx,DWORD [4+esi]
mov ecx,DWORD [8+esi]
mov edx,DWORD [12+esi]
mov DWORD [80+esp],eax
mov DWORD [84+esp],ebx
mov DWORD [88+esp],ecx
mov DWORD [92+esp],edx
mov eax,DWORD [16+esi]
mov ebx,DWORD [20+esi]
mov ecx,DWORD [24+esi]
mov edx,DWORD [28+esi]
mov DWORD [96+esp],eax
mov DWORD [100+esp],ebx
mov DWORD [104+esp],ecx
mov DWORD [108+esp],edx
mov eax,DWORD [edi]
mov ebx,DWORD [4+edi]
mov ecx,DWORD [8+edi]
mov edx,DWORD [12+edi]
sub eax,1
mov DWORD [112+esp],eax
mov DWORD [116+esp],ebx
mov DWORD [120+esp],ecx
mov DWORD [124+esp],edx
jmp NEAR L$002entry
align 16
L$003outer_loop:
mov DWORD [156+esp],ebx
mov DWORD [152+esp],eax
mov DWORD [160+esp],ecx
L$002entry:
mov eax,1634760805
mov DWORD [4+esp],857760878
mov DWORD [8+esp],2036477234
mov DWORD [12+esp],1797285236
mov ebx,DWORD [84+esp]
mov ebp,DWORD [88+esp]
mov ecx,DWORD [104+esp]
mov esi,DWORD [108+esp]
mov edx,DWORD [116+esp]
mov edi,DWORD [120+esp]
mov DWORD [20+esp],ebx
mov DWORD [24+esp],ebp
mov DWORD [40+esp],ecx
mov DWORD [44+esp],esi
mov DWORD [52+esp],edx
mov DWORD [56+esp],edi
mov ebx,DWORD [92+esp]
mov edi,DWORD [124+esp]
mov edx,DWORD [112+esp]
mov ebp,DWORD [80+esp]
mov ecx,DWORD [96+esp]
mov esi,DWORD [100+esp]
add edx,1
mov DWORD [28+esp],ebx
mov DWORD [60+esp],edi
mov DWORD [112+esp],edx
mov ebx,10
jmp NEAR L$004loop
align 16
L$004loop:
add eax,ebp
mov DWORD [128+esp],ebx
mov ebx,ebp
xor edx,eax
rol edx,16
add ecx,edx
xor ebx,ecx
mov edi,DWORD [52+esp]
rol ebx,12
mov ebp,DWORD [20+esp]
add eax,ebx
xor edx,eax
mov DWORD [esp],eax
rol edx,8
mov eax,DWORD [4+esp]
add ecx,edx
mov DWORD [48+esp],edx
xor ebx,ecx
add eax,ebp
rol ebx,7
xor edi,eax
mov DWORD [32+esp],ecx
rol edi,16
mov DWORD [16+esp],ebx
add esi,edi
mov ecx,DWORD [40+esp]
xor ebp,esi
mov edx,DWORD [56+esp]
rol ebp,12
mov ebx,DWORD [24+esp]
add eax,ebp
xor edi,eax
mov DWORD [4+esp],eax
rol edi,8
mov eax,DWORD [8+esp]
add esi,edi
mov DWORD [52+esp],edi
xor ebp,esi
add eax,ebx
rol ebp,7
xor edx,eax
mov DWORD [36+esp],esi
rol edx,16
mov DWORD [20+esp],ebp
add ecx,edx
mov esi,DWORD [44+esp]
xor ebx,ecx
mov edi,DWORD [60+esp]
rol ebx,12
mov ebp,DWORD [28+esp]
add eax,ebx
xor edx,eax
mov DWORD [8+esp],eax
rol edx,8
mov eax,DWORD [12+esp]
add ecx,edx
mov DWORD [56+esp],edx
xor ebx,ecx
add eax,ebp
rol ebx,7
xor edi,eax
rol edi,16
mov DWORD [24+esp],ebx
add esi,edi
xor ebp,esi
rol ebp,12
mov ebx,DWORD [20+esp]
add eax,ebp
xor edi,eax
mov DWORD [12+esp],eax
rol edi,8
mov eax,DWORD [esp]
add esi,edi
mov edx,edi
xor ebp,esi
add eax,ebx
rol ebp,7
xor edx,eax
rol edx,16
mov DWORD [28+esp],ebp
add ecx,edx
xor ebx,ecx
mov edi,DWORD [48+esp]
rol ebx,12
mov ebp,DWORD [24+esp]
add eax,ebx
xor edx,eax
mov DWORD [esp],eax
rol edx,8
mov eax,DWORD [4+esp]
add ecx,edx
mov DWORD [60+esp],edx
xor ebx,ecx
add eax,ebp
rol ebx,7
xor edi,eax
mov DWORD [40+esp],ecx
rol edi,16
mov DWORD [20+esp],ebx
add esi,edi
mov ecx,DWORD [32+esp]
xor ebp,esi
mov edx,DWORD [52+esp]
rol ebp,12
mov ebx,DWORD [28+esp]
add eax,ebp
xor edi,eax
mov DWORD [4+esp],eax
rol edi,8
mov eax,DWORD [8+esp]
add esi,edi
mov DWORD [48+esp],edi
xor ebp,esi
add eax,ebx
rol ebp,7
xor edx,eax
mov DWORD [44+esp],esi
rol edx,16
mov DWORD [24+esp],ebp
add ecx,edx
mov esi,DWORD [36+esp]
xor ebx,ecx
mov edi,DWORD [56+esp]
rol ebx,12
mov ebp,DWORD [16+esp]
add eax,ebx
xor edx,eax
mov DWORD [8+esp],eax
rol edx,8
mov eax,DWORD [12+esp]
add ecx,edx
mov DWORD [52+esp],edx
xor ebx,ecx
add eax,ebp
rol ebx,7
xor edi,eax
rol edi,16
mov DWORD [28+esp],ebx
add esi,edi
xor ebp,esi
mov edx,DWORD [48+esp]
rol ebp,12
mov ebx,DWORD [128+esp]
add eax,ebp
xor edi,eax
mov DWORD [12+esp],eax
rol edi,8
mov eax,DWORD [esp]
add esi,edi
mov DWORD [56+esp],edi
xor ebp,esi
rol ebp,7
dec ebx
jnz NEAR L$004loop
mov ebx,DWORD [160+esp]
add eax,1634760805
add ebp,DWORD [80+esp]
add ecx,DWORD [96+esp]
add esi,DWORD [100+esp]
cmp ebx,64
jb NEAR L$005tail
mov ebx,DWORD [156+esp]
add edx,DWORD [112+esp]
add edi,DWORD [120+esp]
xor eax,DWORD [ebx]
xor ebp,DWORD [16+ebx]
mov DWORD [esp],eax
mov eax,DWORD [152+esp]
xor ecx,DWORD [32+ebx]
xor esi,DWORD [36+ebx]
xor edx,DWORD [48+ebx]
xor edi,DWORD [56+ebx]
mov DWORD [16+eax],ebp
mov DWORD [32+eax],ecx
mov DWORD [36+eax],esi
mov DWORD [48+eax],edx
mov DWORD [56+eax],edi
mov ebp,DWORD [4+esp]
mov ecx,DWORD [8+esp]
mov esi,DWORD [12+esp]
mov edx,DWORD [20+esp]
mov edi,DWORD [24+esp]
add ebp,857760878
add ecx,2036477234
add esi,1797285236
add edx,DWORD [84+esp]
add edi,DWORD [88+esp]
xor ebp,DWORD [4+ebx]
xor ecx,DWORD [8+ebx]
xor esi,DWORD [12+ebx]
xor edx,DWORD [20+ebx]
xor edi,DWORD [24+ebx]
mov DWORD [4+eax],ebp
mov DWORD [8+eax],ecx
mov DWORD [12+eax],esi
mov DWORD [20+eax],edx
mov DWORD [24+eax],edi
mov ebp,DWORD [28+esp]
mov ecx,DWORD [40+esp]
mov esi,DWORD [44+esp]
mov edx,DWORD [52+esp]
mov edi,DWORD [60+esp]
add ebp,DWORD [92+esp]
add ecx,DWORD [104+esp]
add esi,DWORD [108+esp]
add edx,DWORD [116+esp]
add edi,DWORD [124+esp]
xor ebp,DWORD [28+ebx]
xor ecx,DWORD [40+ebx]
xor esi,DWORD [44+ebx]
xor edx,DWORD [52+ebx]
xor edi,DWORD [60+ebx]
lea ebx,[64+ebx]
mov DWORD [28+eax],ebp
mov ebp,DWORD [esp]
mov DWORD [40+eax],ecx
mov ecx,DWORD [160+esp]
mov DWORD [44+eax],esi
mov DWORD [52+eax],edx
mov DWORD [60+eax],edi
mov DWORD [eax],ebp
lea eax,[64+eax]
sub ecx,64
jnz NEAR L$003outer_loop
jmp NEAR L$006done
L$005tail:
add edx,DWORD [112+esp]
add edi,DWORD [120+esp]
mov DWORD [esp],eax
mov DWORD [16+esp],ebp
mov DWORD [32+esp],ecx
mov DWORD [36+esp],esi
mov DWORD [48+esp],edx
mov DWORD [56+esp],edi
mov ebp,DWORD [4+esp]
mov ecx,DWORD [8+esp]
mov esi,DWORD [12+esp]
mov edx,DWORD [20+esp]
mov edi,DWORD [24+esp]
add ebp,857760878
add ecx,2036477234
add esi,1797285236
add edx,DWORD [84+esp]
add edi,DWORD [88+esp]
mov DWORD [4+esp],ebp
mov DWORD [8+esp],ecx
mov DWORD [12+esp],esi
mov DWORD [20+esp],edx
mov DWORD [24+esp],edi
mov ebp,DWORD [28+esp]
mov ecx,DWORD [40+esp]
mov esi,DWORD [44+esp]
mov edx,DWORD [52+esp]
mov edi,DWORD [60+esp]
add ebp,DWORD [92+esp]
add ecx,DWORD [104+esp]
add esi,DWORD [108+esp]
add edx,DWORD [116+esp]
add edi,DWORD [124+esp]
mov DWORD [28+esp],ebp
mov ebp,DWORD [156+esp]
mov DWORD [40+esp],ecx
mov ecx,DWORD [152+esp]
mov DWORD [44+esp],esi
xor esi,esi
mov DWORD [52+esp],edx
mov DWORD [60+esp],edi
xor eax,eax
xor edx,edx
L$007tail_loop:
mov al,BYTE [ebp*1+esi]
mov dl,BYTE [esi*1+esp]
lea esi,[1+esi]
xor al,dl
mov BYTE [esi*1+ecx-1],al
dec ebx
jnz NEAR L$007tail_loop
L$006done:
add esp,132
L$000no_data:
pop edi
pop esi
pop ebx
pop ebp
ret
global _ChaCha20_ssse3
align 16
_ChaCha20_ssse3:
L$_ChaCha20_ssse3_begin:
push ebp
push ebx
push esi
push edi
L$ssse3_shortcut:
mov edi,DWORD [20+esp]
mov esi,DWORD [24+esp]
mov ecx,DWORD [28+esp]
mov edx,DWORD [32+esp]
mov ebx,DWORD [36+esp]
mov ebp,esp
sub esp,524
and esp,-64
mov DWORD [512+esp],ebp
lea eax,[(L$ssse3_data-L$pic_point)+eax]
movdqu xmm3,[ebx]
cmp ecx,256
jb NEAR L$0081x
mov DWORD [516+esp],edx
mov DWORD [520+esp],ebx
sub ecx,256
lea ebp,[384+esp]
movdqu xmm7,[edx]
pshufd xmm0,xmm3,0
pshufd xmm1,xmm3,85
pshufd xmm2,xmm3,170
pshufd xmm3,xmm3,255
paddd xmm0,[48+eax]
pshufd xmm4,xmm7,0
pshufd xmm5,xmm7,85
psubd xmm0,[64+eax]
pshufd xmm6,xmm7,170
pshufd xmm7,xmm7,255
movdqa [64+ebp],xmm0
movdqa [80+ebp],xmm1
movdqa [96+ebp],xmm2
movdqa [112+ebp],xmm3
movdqu xmm3,[16+edx]
movdqa [ebp-64],xmm4
movdqa [ebp-48],xmm5
movdqa [ebp-32],xmm6
movdqa [ebp-16],xmm7
movdqa xmm7,[32+eax]
lea ebx,[128+esp]
pshufd xmm0,xmm3,0
pshufd xmm1,xmm3,85
pshufd xmm2,xmm3,170
pshufd xmm3,xmm3,255
pshufd xmm4,xmm7,0
pshufd xmm5,xmm7,85
pshufd xmm6,xmm7,170
pshufd xmm7,xmm7,255
movdqa [ebp],xmm0
movdqa [16+ebp],xmm1
movdqa [32+ebp],xmm2
movdqa [48+ebp],xmm3
movdqa [ebp-128],xmm4
movdqa [ebp-112],xmm5
movdqa [ebp-96],xmm6
movdqa [ebp-80],xmm7
lea esi,[128+esi]
lea edi,[128+edi]
jmp NEAR L$009outer_loop
align 16
L$009outer_loop:
movdqa xmm1,[ebp-112]
movdqa xmm2,[ebp-96]
movdqa xmm3,[ebp-80]
movdqa xmm5,[ebp-48]
movdqa xmm6,[ebp-32]
movdqa xmm7,[ebp-16]
movdqa [ebx-112],xmm1
movdqa [ebx-96],xmm2
movdqa [ebx-80],xmm3
movdqa [ebx-48],xmm5
movdqa [ebx-32],xmm6
movdqa [ebx-16],xmm7
movdqa xmm2,[32+ebp]
movdqa xmm3,[48+ebp]
movdqa xmm4,[64+ebp]
movdqa xmm5,[80+ebp]
movdqa xmm6,[96+ebp]
movdqa xmm7,[112+ebp]
paddd xmm4,[64+eax]
movdqa [32+ebx],xmm2
movdqa [48+ebx],xmm3
movdqa [64+ebx],xmm4
movdqa [80+ebx],xmm5
movdqa [96+ebx],xmm6
movdqa [112+ebx],xmm7
movdqa [64+ebp],xmm4
movdqa xmm0,[ebp-128]
movdqa xmm6,xmm4
movdqa xmm3,[ebp-64]
movdqa xmm4,[ebp]
movdqa xmm5,[16+ebp]
mov edx,10
nop
align 16
L$010loop:
paddd xmm0,xmm3
movdqa xmm2,xmm3
pxor xmm6,xmm0
pshufb xmm6,[eax]
paddd xmm4,xmm6
pxor xmm2,xmm4
movdqa xmm3,[ebx-48]
movdqa xmm1,xmm2
pslld xmm2,12
psrld xmm1,20
por xmm2,xmm1
movdqa xmm1,[ebx-112]
paddd xmm0,xmm2
movdqa xmm7,[80+ebx]
pxor xmm6,xmm0
movdqa [ebx-128],xmm0
pshufb xmm6,[16+eax]
paddd xmm4,xmm6
movdqa [64+ebx],xmm6
pxor xmm2,xmm4
paddd xmm1,xmm3
movdqa xmm0,xmm2
pslld xmm2,7
psrld xmm0,25
pxor xmm7,xmm1
por xmm2,xmm0
movdqa [ebx],xmm4
pshufb xmm7,[eax]
movdqa [ebx-64],xmm2
paddd xmm5,xmm7
movdqa xmm4,[32+ebx]
pxor xmm3,xmm5
movdqa xmm2,[ebx-32]
movdqa xmm0,xmm3
pslld xmm3,12
psrld xmm0,20
por xmm3,xmm0
movdqa xmm0,[ebx-96]
paddd xmm1,xmm3
movdqa xmm6,[96+ebx]
pxor xmm7,xmm1
movdqa [ebx-112],xmm1
pshufb xmm7,[16+eax]
paddd xmm5,xmm7
movdqa [80+ebx],xmm7
pxor xmm3,xmm5
paddd xmm0,xmm2
movdqa xmm1,xmm3
pslld xmm3,7
psrld xmm1,25
pxor xmm6,xmm0
por xmm3,xmm1
movdqa [16+ebx],xmm5
pshufb xmm6,[eax]
movdqa [ebx-48],xmm3
paddd xmm4,xmm6
movdqa xmm5,[48+ebx]
pxor xmm2,xmm4
movdqa xmm3,[ebx-16]
movdqa xmm1,xmm2
pslld xmm2,12
psrld xmm1,20
por xmm2,xmm1
movdqa xmm1,[ebx-80]
paddd xmm0,xmm2
movdqa xmm7,[112+ebx]
pxor xmm6,xmm0
movdqa [ebx-96],xmm0
pshufb xmm6,[16+eax]
paddd xmm4,xmm6
movdqa [96+ebx],xmm6
pxor xmm2,xmm4
paddd xmm1,xmm3
movdqa xmm0,xmm2
pslld xmm2,7
psrld xmm0,25
pxor xmm7,xmm1
por xmm2,xmm0
pshufb xmm7,[eax]
movdqa [ebx-32],xmm2
paddd xmm5,xmm7
pxor xmm3,xmm5
movdqa xmm2,[ebx-48]
movdqa xmm0,xmm3
pslld xmm3,12
psrld xmm0,20
por xmm3,xmm0
movdqa xmm0,[ebx-128]
paddd xmm1,xmm3
pxor xmm7,xmm1
movdqa [ebx-80],xmm1
pshufb xmm7,[16+eax]
paddd xmm5,xmm7
movdqa xmm6,xmm7
pxor xmm3,xmm5
paddd xmm0,xmm2
movdqa xmm1,xmm3
pslld xmm3,7
psrld xmm1,25
pxor xmm6,xmm0
por xmm3,xmm1
pshufb xmm6,[eax]
movdqa [ebx-16],xmm3
paddd xmm4,xmm6
pxor xmm2,xmm4
movdqa xmm3,[ebx-32]
movdqa xmm1,xmm2
pslld xmm2,12
psrld xmm1,20
por xmm2,xmm1
movdqa xmm1,[ebx-112]
paddd xmm0,xmm2
movdqa xmm7,[64+ebx]
pxor xmm6,xmm0
movdqa [ebx-128],xmm0
pshufb xmm6,[16+eax]
paddd xmm4,xmm6
movdqa [112+ebx],xmm6
pxor xmm2,xmm4
paddd xmm1,xmm3
movdqa xmm0,xmm2
pslld xmm2,7
psrld xmm0,25
pxor xmm7,xmm1
por xmm2,xmm0
movdqa [32+ebx],xmm4
pshufb xmm7,[eax]
movdqa [ebx-48],xmm2
paddd xmm5,xmm7
movdqa xmm4,[ebx]
pxor xmm3,xmm5
movdqa xmm2,[ebx-16]
movdqa xmm0,xmm3
pslld xmm3,12
psrld xmm0,20
por xmm3,xmm0
movdqa xmm0,[ebx-96]
paddd xmm1,xmm3
movdqa xmm6,[80+ebx]
pxor xmm7,xmm1
movdqa [ebx-112],xmm1
pshufb xmm7,[16+eax]
paddd xmm5,xmm7
movdqa [64+ebx],xmm7
pxor xmm3,xmm5
paddd xmm0,xmm2
movdqa xmm1,xmm3
pslld xmm3,7
psrld xmm1,25
pxor xmm6,xmm0
por xmm3,xmm1
movdqa [48+ebx],xmm5
pshufb xmm6,[eax]
movdqa [ebx-32],xmm3
paddd xmm4,xmm6
movdqa xmm5,[16+ebx]
pxor xmm2,xmm4
movdqa xmm3,[ebx-64]
movdqa xmm1,xmm2
pslld xmm2,12
psrld xmm1,20
por xmm2,xmm1
movdqa xmm1,[ebx-80]
paddd xmm0,xmm2
movdqa xmm7,[96+ebx]
pxor xmm6,xmm0
movdqa [ebx-96],xmm0
pshufb xmm6,[16+eax]
paddd xmm4,xmm6
movdqa [80+ebx],xmm6
pxor xmm2,xmm4
paddd xmm1,xmm3
movdqa xmm0,xmm2
pslld xmm2,7
psrld xmm0,25
pxor xmm7,xmm1
por xmm2,xmm0
pshufb xmm7,[eax]
movdqa [ebx-16],xmm2
paddd xmm5,xmm7
pxor xmm3,xmm5
movdqa xmm0,xmm3
pslld xmm3,12
psrld xmm0,20
por xmm3,xmm0
movdqa xmm0,[ebx-128]
paddd xmm1,xmm3
movdqa xmm6,[64+ebx]
pxor xmm7,xmm1
movdqa [ebx-80],xmm1
pshufb xmm7,[16+eax]
paddd xmm5,xmm7
movdqa [96+ebx],xmm7
pxor xmm3,xmm5
movdqa xmm1,xmm3
pslld xmm3,7
psrld xmm1,25
por xmm3,xmm1
dec edx
jnz NEAR L$010loop
movdqa [ebx-64],xmm3
movdqa [ebx],xmm4
movdqa [16+ebx],xmm5
movdqa [64+ebx],xmm6
movdqa [96+ebx],xmm7
movdqa xmm1,[ebx-112]
movdqa xmm2,[ebx-96]
movdqa xmm3,[ebx-80]
paddd xmm0,[ebp-128]
paddd xmm1,[ebp-112]
paddd xmm2,[ebp-96]
paddd xmm3,[ebp-80]
movdqa xmm6,xmm0
punpckldq xmm0,xmm1
movdqa xmm7,xmm2
punpckldq xmm2,xmm3
punpckhdq xmm6,xmm1
punpckhdq xmm7,xmm3
movdqa xmm1,xmm0
punpcklqdq xmm0,xmm2
movdqa xmm3,xmm6
punpcklqdq xmm6,xmm7
punpckhqdq xmm1,xmm2
punpckhqdq xmm3,xmm7
movdqu xmm4,[esi-128]
movdqu xmm5,[esi-64]
movdqu xmm2,[esi]
movdqu xmm7,[64+esi]
lea esi,[16+esi]
pxor xmm4,xmm0
movdqa xmm0,[ebx-64]
pxor xmm5,xmm1
movdqa xmm1,[ebx-48]
pxor xmm6,xmm2
movdqa xmm2,[ebx-32]
pxor xmm7,xmm3
movdqa xmm3,[ebx-16]
movdqu [edi-128],xmm4
movdqu [edi-64],xmm5
movdqu [edi],xmm6
movdqu [64+edi],xmm7
lea edi,[16+edi]
paddd xmm0,[ebp-64]
paddd xmm1,[ebp-48]
paddd xmm2,[ebp-32]
paddd xmm3,[ebp-16]
movdqa xmm6,xmm0
punpckldq xmm0,xmm1
movdqa xmm7,xmm2
punpckldq xmm2,xmm3
punpckhdq xmm6,xmm1
punpckhdq xmm7,xmm3
movdqa xmm1,xmm0
punpcklqdq xmm0,xmm2
movdqa xmm3,xmm6
punpcklqdq xmm6,xmm7
punpckhqdq xmm1,xmm2
punpckhqdq xmm3,xmm7
movdqu xmm4,[esi-128]
movdqu xmm5,[esi-64]
movdqu xmm2,[esi]
movdqu xmm7,[64+esi]
lea esi,[16+esi]
pxor xmm4,xmm0
movdqa xmm0,[ebx]
pxor xmm5,xmm1
movdqa xmm1,[16+ebx]
pxor xmm6,xmm2
movdqa xmm2,[32+ebx]
pxor xmm7,xmm3
movdqa xmm3,[48+ebx]
movdqu [edi-128],xmm4
movdqu [edi-64],xmm5
movdqu [edi],xmm6
movdqu [64+edi],xmm7
lea edi,[16+edi]
paddd xmm0,[ebp]
paddd xmm1,[16+ebp]
paddd xmm2,[32+ebp]
paddd xmm3,[48+ebp]
movdqa xmm6,xmm0
punpckldq xmm0,xmm1
movdqa xmm7,xmm2
punpckldq xmm2,xmm3
punpckhdq xmm6,xmm1
punpckhdq xmm7,xmm3
movdqa xmm1,xmm0
punpcklqdq xmm0,xmm2
movdqa xmm3,xmm6
punpcklqdq xmm6,xmm7
punpckhqdq xmm1,xmm2
punpckhqdq xmm3,xmm7
movdqu xmm4,[esi-128]
movdqu xmm5,[esi-64]
movdqu xmm2,[esi]
movdqu xmm7,[64+esi]
lea esi,[16+esi]
pxor xmm4,xmm0
movdqa xmm0,[64+ebx]
pxor xmm5,xmm1
movdqa xmm1,[80+ebx]
pxor xmm6,xmm2
movdqa xmm2,[96+ebx]
pxor xmm7,xmm3
movdqa xmm3,[112+ebx]
movdqu [edi-128],xmm4
movdqu [edi-64],xmm5
movdqu [edi],xmm6
movdqu [64+edi],xmm7
lea edi,[16+edi]
paddd xmm0,[64+ebp]
paddd xmm1,[80+ebp]
paddd xmm2,[96+ebp]
paddd xmm3,[112+ebp]
movdqa xmm6,xmm0
punpckldq xmm0,xmm1
movdqa xmm7,xmm2
punpckldq xmm2,xmm3
punpckhdq xmm6,xmm1
punpckhdq xmm7,xmm3
movdqa xmm1,xmm0
punpcklqdq xmm0,xmm2
movdqa xmm3,xmm6
punpcklqdq xmm6,xmm7
punpckhqdq xmm1,xmm2
punpckhqdq xmm3,xmm7
movdqu xmm4,[esi-128]
movdqu xmm5,[esi-64]
movdqu xmm2,[esi]
movdqu xmm7,[64+esi]
lea esi,[208+esi]
pxor xmm4,xmm0
pxor xmm5,xmm1
pxor xmm6,xmm2
pxor xmm7,xmm3
movdqu [edi-128],xmm4
movdqu [edi-64],xmm5
movdqu [edi],xmm6
movdqu [64+edi],xmm7
lea edi,[208+edi]
sub ecx,256
jnc NEAR L$009outer_loop
add ecx,256
jz NEAR L$011done
mov ebx,DWORD [520+esp]
lea esi,[esi-128]
mov edx,DWORD [516+esp]
lea edi,[edi-128]
movd xmm2,DWORD [64+ebp]
movdqu xmm3,[ebx]
paddd xmm2,[96+eax]
pand xmm3,[112+eax]
por xmm3,xmm2
L$0081x:
movdqa xmm0,[32+eax]
movdqu xmm1,[edx]
movdqu xmm2,[16+edx]
movdqa xmm6,[eax]
movdqa xmm7,[16+eax]
mov DWORD [48+esp],ebp
movdqa [esp],xmm0
movdqa [16+esp],xmm1
movdqa [32+esp],xmm2
movdqa [48+esp],xmm3
mov edx,10
jmp NEAR L$012loop1x
align 16
L$013outer1x:
movdqa xmm3,[80+eax]
movdqa xmm0,[esp]
movdqa xmm1,[16+esp]
movdqa xmm2,[32+esp]
paddd xmm3,[48+esp]
mov edx,10
movdqa [48+esp],xmm3
jmp NEAR L$012loop1x
align 16
L$012loop1x:
paddd xmm0,xmm1
pxor xmm3,xmm0
db 102,15,56,0,222
paddd xmm2,xmm3
pxor xmm1,xmm2
movdqa xmm4,xmm1
psrld xmm1,20
pslld xmm4,12
por xmm1,xmm4
paddd xmm0,xmm1
pxor xmm3,xmm0
db 102,15,56,0,223
paddd xmm2,xmm3
pxor xmm1,xmm2
movdqa xmm4,xmm1
psrld xmm1,25
pslld xmm4,7
por xmm1,xmm4
pshufd xmm2,xmm2,78
pshufd xmm1,xmm1,57
pshufd xmm3,xmm3,147
nop
paddd xmm0,xmm1
pxor xmm3,xmm0
db 102,15,56,0,222
paddd xmm2,xmm3
pxor xmm1,xmm2
movdqa xmm4,xmm1
psrld xmm1,20
pslld xmm4,12
por xmm1,xmm4
paddd xmm0,xmm1
pxor xmm3,xmm0
db 102,15,56,0,223
paddd xmm2,xmm3
pxor xmm1,xmm2
movdqa xmm4,xmm1
psrld xmm1,25
pslld xmm4,7
por xmm1,xmm4
pshufd xmm2,xmm2,78
pshufd xmm1,xmm1,147
pshufd xmm3,xmm3,57
dec edx
jnz NEAR L$012loop1x
paddd xmm0,[esp]
paddd xmm1,[16+esp]
paddd xmm2,[32+esp]
paddd xmm3,[48+esp]
cmp ecx,64
jb NEAR L$014tail
movdqu xmm4,[esi]
movdqu xmm5,[16+esi]
pxor xmm0,xmm4
movdqu xmm4,[32+esi]
pxor xmm1,xmm5
movdqu xmm5,[48+esi]
pxor xmm2,xmm4
pxor xmm3,xmm5
lea esi,[64+esi]
movdqu [edi],xmm0
movdqu [16+edi],xmm1
movdqu [32+edi],xmm2
movdqu [48+edi],xmm3
lea edi,[64+edi]
sub ecx,64
jnz NEAR L$013outer1x
jmp NEAR L$011done
L$014tail:
movdqa [esp],xmm0
movdqa [16+esp],xmm1
movdqa [32+esp],xmm2
movdqa [48+esp],xmm3
xor eax,eax
xor edx,edx
xor ebp,ebp
L$015tail_loop:
mov al,BYTE [ebp*1+esp]
mov dl,BYTE [ebp*1+esi]
lea ebp,[1+ebp]
xor al,dl
mov BYTE [ebp*1+edi-1],al
dec ecx
jnz NEAR L$015tail_loop
L$011done:
mov esp,DWORD [512+esp]
pop edi
pop esi
pop ebx
pop ebp
ret
align 64
L$ssse3_data:
db 2,3,0,1,6,7,4,5,10,11,8,9,14,15,12,13
db 3,0,1,2,7,4,5,6,11,8,9,10,15,12,13,14
dd 1634760805,857760878,2036477234,1797285236
dd 0,1,2,3
dd 4,4,4,4
dd 1,0,0,0
dd 4,0,0,0
dd 0,-1,-1,-1
align 64
db 67,104,97,67,104,97,50,48,32,102,111,114,32,120,56,54
db 44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32
db 60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111
db 114,103,62,0
segment .bss
common _OPENSSL_ia32cap_P 16

File diff suppressed because it is too large Load Diff

File diff suppressed because it is too large Load Diff

File diff suppressed because it is too large Load Diff

File diff suppressed because it is too large Load Diff

File diff suppressed because it is too large Load Diff

View File

@@ -0,0 +1,691 @@
%ifidn __OUTPUT_FORMAT__,obj
section code use32 class=code align=64
%elifidn __OUTPUT_FORMAT__,win32
%ifdef __YASM_VERSION_ID__
%if __YASM_VERSION_ID__ < 01010000h
%error yasm version 1.1.0 or later needed.
%endif
; Yasm automatically includes .00 and complains about redefining it.
; https://www.tortall.net/projects/yasm/manual/html/objfmt-win32-safeseh.html
%else
$@feat.00 equ 1
%endif
section .text code align=64
%else
section .text code
%endif
global _md5_block_asm_data_order
align 16
_md5_block_asm_data_order:
L$_md5_block_asm_data_order_begin:
push esi
push edi
mov edi,DWORD [12+esp]
mov esi,DWORD [16+esp]
mov ecx,DWORD [20+esp]
push ebp
shl ecx,6
push ebx
add ecx,esi
sub ecx,64
mov eax,DWORD [edi]
push ecx
mov ebx,DWORD [4+edi]
mov ecx,DWORD [8+edi]
mov edx,DWORD [12+edi]
L$000start:
;
; R0 section
mov edi,ecx
mov ebp,DWORD [esi]
; R0 0
xor edi,edx
and edi,ebx
lea eax,[3614090360+ebp*1+eax]
xor edi,edx
add eax,edi
mov edi,ebx
rol eax,7
mov ebp,DWORD [4+esi]
add eax,ebx
; R0 1
xor edi,ecx
and edi,eax
lea edx,[3905402710+ebp*1+edx]
xor edi,ecx
add edx,edi
mov edi,eax
rol edx,12
mov ebp,DWORD [8+esi]
add edx,eax
; R0 2
xor edi,ebx
and edi,edx
lea ecx,[606105819+ebp*1+ecx]
xor edi,ebx
add ecx,edi
mov edi,edx
rol ecx,17
mov ebp,DWORD [12+esi]
add ecx,edx
; R0 3
xor edi,eax
and edi,ecx
lea ebx,[3250441966+ebp*1+ebx]
xor edi,eax
add ebx,edi
mov edi,ecx
rol ebx,22
mov ebp,DWORD [16+esi]
add ebx,ecx
; R0 4
xor edi,edx
and edi,ebx
lea eax,[4118548399+ebp*1+eax]
xor edi,edx
add eax,edi
mov edi,ebx
rol eax,7
mov ebp,DWORD [20+esi]
add eax,ebx
; R0 5
xor edi,ecx
and edi,eax
lea edx,[1200080426+ebp*1+edx]
xor edi,ecx
add edx,edi
mov edi,eax
rol edx,12
mov ebp,DWORD [24+esi]
add edx,eax
; R0 6
xor edi,ebx
and edi,edx
lea ecx,[2821735955+ebp*1+ecx]
xor edi,ebx
add ecx,edi
mov edi,edx
rol ecx,17
mov ebp,DWORD [28+esi]
add ecx,edx
; R0 7
xor edi,eax
and edi,ecx
lea ebx,[4249261313+ebp*1+ebx]
xor edi,eax
add ebx,edi
mov edi,ecx
rol ebx,22
mov ebp,DWORD [32+esi]
add ebx,ecx
; R0 8
xor edi,edx
and edi,ebx
lea eax,[1770035416+ebp*1+eax]
xor edi,edx
add eax,edi
mov edi,ebx
rol eax,7
mov ebp,DWORD [36+esi]
add eax,ebx
; R0 9
xor edi,ecx
and edi,eax
lea edx,[2336552879+ebp*1+edx]
xor edi,ecx
add edx,edi
mov edi,eax
rol edx,12
mov ebp,DWORD [40+esi]
add edx,eax
; R0 10
xor edi,ebx
and edi,edx
lea ecx,[4294925233+ebp*1+ecx]
xor edi,ebx
add ecx,edi
mov edi,edx
rol ecx,17
mov ebp,DWORD [44+esi]
add ecx,edx
; R0 11
xor edi,eax
and edi,ecx
lea ebx,[2304563134+ebp*1+ebx]
xor edi,eax
add ebx,edi
mov edi,ecx
rol ebx,22
mov ebp,DWORD [48+esi]
add ebx,ecx
; R0 12
xor edi,edx
and edi,ebx
lea eax,[1804603682+ebp*1+eax]
xor edi,edx
add eax,edi
mov edi,ebx
rol eax,7
mov ebp,DWORD [52+esi]
add eax,ebx
; R0 13
xor edi,ecx
and edi,eax
lea edx,[4254626195+ebp*1+edx]
xor edi,ecx
add edx,edi
mov edi,eax
rol edx,12
mov ebp,DWORD [56+esi]
add edx,eax
; R0 14
xor edi,ebx
and edi,edx
lea ecx,[2792965006+ebp*1+ecx]
xor edi,ebx
add ecx,edi
mov edi,edx
rol ecx,17
mov ebp,DWORD [60+esi]
add ecx,edx
; R0 15
xor edi,eax
and edi,ecx
lea ebx,[1236535329+ebp*1+ebx]
xor edi,eax
add ebx,edi
mov edi,ecx
rol ebx,22
mov ebp,DWORD [4+esi]
add ebx,ecx
;
; R1 section
; R1 16
lea eax,[4129170786+ebp*1+eax]
xor edi,ebx
and edi,edx
mov ebp,DWORD [24+esi]
xor edi,ecx
add eax,edi
mov edi,ebx
rol eax,5
add eax,ebx
; R1 17
lea edx,[3225465664+ebp*1+edx]
xor edi,eax
and edi,ecx
mov ebp,DWORD [44+esi]
xor edi,ebx
add edx,edi
mov edi,eax
rol edx,9
add edx,eax
; R1 18
lea ecx,[643717713+ebp*1+ecx]
xor edi,edx
and edi,ebx
mov ebp,DWORD [esi]
xor edi,eax
add ecx,edi
mov edi,edx
rol ecx,14
add ecx,edx
; R1 19
lea ebx,[3921069994+ebp*1+ebx]
xor edi,ecx
and edi,eax
mov ebp,DWORD [20+esi]
xor edi,edx
add ebx,edi
mov edi,ecx
rol ebx,20
add ebx,ecx
; R1 20
lea eax,[3593408605+ebp*1+eax]
xor edi,ebx
and edi,edx
mov ebp,DWORD [40+esi]
xor edi,ecx
add eax,edi
mov edi,ebx
rol eax,5
add eax,ebx
; R1 21
lea edx,[38016083+ebp*1+edx]
xor edi,eax
and edi,ecx
mov ebp,DWORD [60+esi]
xor edi,ebx
add edx,edi
mov edi,eax
rol edx,9
add edx,eax
; R1 22
lea ecx,[3634488961+ebp*1+ecx]
xor edi,edx
and edi,ebx
mov ebp,DWORD [16+esi]
xor edi,eax
add ecx,edi
mov edi,edx
rol ecx,14
add ecx,edx
; R1 23
lea ebx,[3889429448+ebp*1+ebx]
xor edi,ecx
and edi,eax
mov ebp,DWORD [36+esi]
xor edi,edx
add ebx,edi
mov edi,ecx
rol ebx,20
add ebx,ecx
; R1 24
lea eax,[568446438+ebp*1+eax]
xor edi,ebx
and edi,edx
mov ebp,DWORD [56+esi]
xor edi,ecx
add eax,edi
mov edi,ebx
rol eax,5
add eax,ebx
; R1 25
lea edx,[3275163606+ebp*1+edx]
xor edi,eax
and edi,ecx
mov ebp,DWORD [12+esi]
xor edi,ebx
add edx,edi
mov edi,eax
rol edx,9
add edx,eax
; R1 26
lea ecx,[4107603335+ebp*1+ecx]
xor edi,edx
and edi,ebx
mov ebp,DWORD [32+esi]
xor edi,eax
add ecx,edi
mov edi,edx
rol ecx,14
add ecx,edx
; R1 27
lea ebx,[1163531501+ebp*1+ebx]
xor edi,ecx
and edi,eax
mov ebp,DWORD [52+esi]
xor edi,edx
add ebx,edi
mov edi,ecx
rol ebx,20
add ebx,ecx
; R1 28
lea eax,[2850285829+ebp*1+eax]
xor edi,ebx
and edi,edx
mov ebp,DWORD [8+esi]
xor edi,ecx
add eax,edi
mov edi,ebx
rol eax,5
add eax,ebx
; R1 29
lea edx,[4243563512+ebp*1+edx]
xor edi,eax
and edi,ecx
mov ebp,DWORD [28+esi]
xor edi,ebx
add edx,edi
mov edi,eax
rol edx,9
add edx,eax
; R1 30
lea ecx,[1735328473+ebp*1+ecx]
xor edi,edx
and edi,ebx
mov ebp,DWORD [48+esi]
xor edi,eax
add ecx,edi
mov edi,edx
rol ecx,14
add ecx,edx
; R1 31
lea ebx,[2368359562+ebp*1+ebx]
xor edi,ecx
and edi,eax
mov ebp,DWORD [20+esi]
xor edi,edx
add ebx,edi
mov edi,ecx
rol ebx,20
add ebx,ecx
;
; R2 section
; R2 32
xor edi,edx
xor edi,ebx
lea eax,[4294588738+ebp*1+eax]
add eax,edi
rol eax,4
mov ebp,DWORD [32+esi]
mov edi,ebx
; R2 33
lea edx,[2272392833+ebp*1+edx]
add eax,ebx
xor edi,ecx
xor edi,eax
mov ebp,DWORD [44+esi]
add edx,edi
mov edi,eax
rol edx,11
add edx,eax
; R2 34
xor edi,ebx
xor edi,edx
lea ecx,[1839030562+ebp*1+ecx]
add ecx,edi
rol ecx,16
mov ebp,DWORD [56+esi]
mov edi,edx
; R2 35
lea ebx,[4259657740+ebp*1+ebx]
add ecx,edx
xor edi,eax
xor edi,ecx
mov ebp,DWORD [4+esi]
add ebx,edi
mov edi,ecx
rol ebx,23
add ebx,ecx
; R2 36
xor edi,edx
xor edi,ebx
lea eax,[2763975236+ebp*1+eax]
add eax,edi
rol eax,4
mov ebp,DWORD [16+esi]
mov edi,ebx
; R2 37
lea edx,[1272893353+ebp*1+edx]
add eax,ebx
xor edi,ecx
xor edi,eax
mov ebp,DWORD [28+esi]
add edx,edi
mov edi,eax
rol edx,11
add edx,eax
; R2 38
xor edi,ebx
xor edi,edx
lea ecx,[4139469664+ebp*1+ecx]
add ecx,edi
rol ecx,16
mov ebp,DWORD [40+esi]
mov edi,edx
; R2 39
lea ebx,[3200236656+ebp*1+ebx]
add ecx,edx
xor edi,eax
xor edi,ecx
mov ebp,DWORD [52+esi]
add ebx,edi
mov edi,ecx
rol ebx,23
add ebx,ecx
; R2 40
xor edi,edx
xor edi,ebx
lea eax,[681279174+ebp*1+eax]
add eax,edi
rol eax,4
mov ebp,DWORD [esi]
mov edi,ebx
; R2 41
lea edx,[3936430074+ebp*1+edx]
add eax,ebx
xor edi,ecx
xor edi,eax
mov ebp,DWORD [12+esi]
add edx,edi
mov edi,eax
rol edx,11
add edx,eax
; R2 42
xor edi,ebx
xor edi,edx
lea ecx,[3572445317+ebp*1+ecx]
add ecx,edi
rol ecx,16
mov ebp,DWORD [24+esi]
mov edi,edx
; R2 43
lea ebx,[76029189+ebp*1+ebx]
add ecx,edx
xor edi,eax
xor edi,ecx
mov ebp,DWORD [36+esi]
add ebx,edi
mov edi,ecx
rol ebx,23
add ebx,ecx
; R2 44
xor edi,edx
xor edi,ebx
lea eax,[3654602809+ebp*1+eax]
add eax,edi
rol eax,4
mov ebp,DWORD [48+esi]
mov edi,ebx
; R2 45
lea edx,[3873151461+ebp*1+edx]
add eax,ebx
xor edi,ecx
xor edi,eax
mov ebp,DWORD [60+esi]
add edx,edi
mov edi,eax
rol edx,11
add edx,eax
; R2 46
xor edi,ebx
xor edi,edx
lea ecx,[530742520+ebp*1+ecx]
add ecx,edi
rol ecx,16
mov ebp,DWORD [8+esi]
mov edi,edx
; R2 47
lea ebx,[3299628645+ebp*1+ebx]
add ecx,edx
xor edi,eax
xor edi,ecx
mov ebp,DWORD [esi]
add ebx,edi
mov edi,-1
rol ebx,23
add ebx,ecx
;
; R3 section
; R3 48
xor edi,edx
or edi,ebx
lea eax,[4096336452+ebp*1+eax]
xor edi,ecx
mov ebp,DWORD [28+esi]
add eax,edi
mov edi,-1
rol eax,6
xor edi,ecx
add eax,ebx
; R3 49
or edi,eax
lea edx,[1126891415+ebp*1+edx]
xor edi,ebx
mov ebp,DWORD [56+esi]
add edx,edi
mov edi,-1
rol edx,10
xor edi,ebx
add edx,eax
; R3 50
or edi,edx
lea ecx,[2878612391+ebp*1+ecx]
xor edi,eax
mov ebp,DWORD [20+esi]
add ecx,edi
mov edi,-1
rol ecx,15
xor edi,eax
add ecx,edx
; R3 51
or edi,ecx
lea ebx,[4237533241+ebp*1+ebx]
xor edi,edx
mov ebp,DWORD [48+esi]
add ebx,edi
mov edi,-1
rol ebx,21
xor edi,edx
add ebx,ecx
; R3 52
or edi,ebx
lea eax,[1700485571+ebp*1+eax]
xor edi,ecx
mov ebp,DWORD [12+esi]
add eax,edi
mov edi,-1
rol eax,6
xor edi,ecx
add eax,ebx
; R3 53
or edi,eax
lea edx,[2399980690+ebp*1+edx]
xor edi,ebx
mov ebp,DWORD [40+esi]
add edx,edi
mov edi,-1
rol edx,10
xor edi,ebx
add edx,eax
; R3 54
or edi,edx
lea ecx,[4293915773+ebp*1+ecx]
xor edi,eax
mov ebp,DWORD [4+esi]
add ecx,edi
mov edi,-1
rol ecx,15
xor edi,eax
add ecx,edx
; R3 55
or edi,ecx
lea ebx,[2240044497+ebp*1+ebx]
xor edi,edx
mov ebp,DWORD [32+esi]
add ebx,edi
mov edi,-1
rol ebx,21
xor edi,edx
add ebx,ecx
; R3 56
or edi,ebx
lea eax,[1873313359+ebp*1+eax]
xor edi,ecx
mov ebp,DWORD [60+esi]
add eax,edi
mov edi,-1
rol eax,6
xor edi,ecx
add eax,ebx
; R3 57
or edi,eax
lea edx,[4264355552+ebp*1+edx]
xor edi,ebx
mov ebp,DWORD [24+esi]
add edx,edi
mov edi,-1
rol edx,10
xor edi,ebx
add edx,eax
; R3 58
or edi,edx
lea ecx,[2734768916+ebp*1+ecx]
xor edi,eax
mov ebp,DWORD [52+esi]
add ecx,edi
mov edi,-1
rol ecx,15
xor edi,eax
add ecx,edx
; R3 59
or edi,ecx
lea ebx,[1309151649+ebp*1+ebx]
xor edi,edx
mov ebp,DWORD [16+esi]
add ebx,edi
mov edi,-1
rol ebx,21
xor edi,edx
add ebx,ecx
; R3 60
or edi,ebx
lea eax,[4149444226+ebp*1+eax]
xor edi,ecx
mov ebp,DWORD [44+esi]
add eax,edi
mov edi,-1
rol eax,6
xor edi,ecx
add eax,ebx
; R3 61
or edi,eax
lea edx,[3174756917+ebp*1+edx]
xor edi,ebx
mov ebp,DWORD [8+esi]
add edx,edi
mov edi,-1
rol edx,10
xor edi,ebx
add edx,eax
; R3 62
or edi,edx
lea ecx,[718787259+ebp*1+ecx]
xor edi,eax
mov ebp,DWORD [36+esi]
add ecx,edi
mov edi,-1
rol ecx,15
xor edi,eax
add ecx,edx
; R3 63
or edi,ecx
lea ebx,[3951481745+ebp*1+ebx]
xor edi,edx
mov ebp,DWORD [24+esp]
add ebx,edi
add esi,64
rol ebx,21
mov edi,DWORD [ebp]
add ebx,ecx
add eax,edi
mov edi,DWORD [4+ebp]
add ebx,edi
mov edi,DWORD [8+ebp]
add ecx,edi
mov edi,DWORD [12+ebp]
add edx,edi
mov DWORD [ebp],eax
mov DWORD [4+ebp],ebx
mov edi,DWORD [esp]
mov DWORD [8+ebp],ecx
mov DWORD [12+ebp],edx
cmp edi,esi
jae NEAR L$000start
pop eax
pop ebx
pop ebp
pop edi
pop esi
ret

File diff suppressed because it is too large Load Diff

File diff suppressed because it is too large Load Diff

File diff suppressed because it is too large Load Diff

View File

@@ -0,0 +1,649 @@
%ifidn __OUTPUT_FORMAT__,obj
section code use32 class=code align=64
%elifidn __OUTPUT_FORMAT__,win32
%ifdef __YASM_VERSION_ID__
%if __YASM_VERSION_ID__ < 01010000h
%error yasm version 1.1.0 or later needed.
%endif
; Yasm automatically includes .00 and complains about redefining it.
; https://www.tortall.net/projects/yasm/manual/html/objfmt-win32-safeseh.html
%else
$@feat.00 equ 1
%endif
section .text code align=64
%else
section .text code
%endif
align 64
L$_vpaes_consts:
dd 218628480,235210255,168496130,67568393
dd 252381056,17041926,33884169,51187212
dd 252645135,252645135,252645135,252645135
dd 1512730624,3266504856,1377990664,3401244816
dd 830229760,1275146365,2969422977,3447763452
dd 3411033600,2979783055,338359620,2782886510
dd 4209124096,907596821,221174255,1006095553
dd 191964160,3799684038,3164090317,1589111125
dd 182528256,1777043520,2877432650,3265356744
dd 1874708224,3503451415,3305285752,363511674
dd 1606117888,3487855781,1093350906,2384367825
dd 197121,67569157,134941193,202313229
dd 67569157,134941193,202313229,197121
dd 134941193,202313229,197121,67569157
dd 202313229,197121,67569157,134941193
dd 33619971,100992007,168364043,235736079
dd 235736079,33619971,100992007,168364043
dd 168364043,235736079,33619971,100992007
dd 100992007,168364043,235736079,33619971
dd 50462976,117835012,185207048,252579084
dd 252314880,51251460,117574920,184942860
dd 184682752,252054788,50987272,118359308
dd 118099200,185467140,251790600,50727180
dd 2946363062,528716217,1300004225,1881839624
dd 1532713819,1532713819,1532713819,1532713819
dd 3602276352,4288629033,3737020424,4153884961
dd 1354558464,32357713,2958822624,3775749553
dd 1201988352,132424512,1572796698,503232858
dd 2213177600,1597421020,4103937655,675398315
dd 2749646592,4273543773,1511898873,121693092
dd 3040248576,1103263732,2871565598,1608280554
dd 2236667136,2588920351,482954393,64377734
dd 3069987328,291237287,2117370568,3650299247
dd 533321216,3573750986,2572112006,1401264716
dd 1339849704,2721158661,548607111,3445553514
dd 2128193280,3054596040,2183486460,1257083700
dd 655635200,1165381986,3923443150,2344132524
dd 190078720,256924420,290342170,357187870
dd 1610966272,2263057382,4103205268,309794674
dd 2592527872,2233205587,1335446729,3402964816
dd 3973531904,3225098121,3002836325,1918774430
dd 3870401024,2102906079,2284471353,4117666579
dd 617007872,1021508343,366931923,691083277
dd 2528395776,3491914898,2968704004,1613121270
dd 3445188352,3247741094,844474987,4093578302
dd 651481088,1190302358,1689581232,574775300
dd 4289380608,206939853,2555985458,2489840491
dd 2130264064,327674451,3566485037,3349835193
dd 2470714624,316102159,3636825756,3393945945
db 86,101,99,116,111,114,32,80,101,114,109,117,116,97,116,105
db 111,110,32,65,69,83,32,102,111,114,32,120,56,54,47,83
db 83,83,69,51,44,32,77,105,107,101,32,72,97,109,98,117
db 114,103,32,40,83,116,97,110,102,111,114,100,32,85,110,105
db 118,101,114,115,105,116,121,41,0
align 64
align 16
__vpaes_preheat:
add ebp,DWORD [esp]
movdqa xmm7,[ebp-48]
movdqa xmm6,[ebp-16]
ret
align 16
__vpaes_encrypt_core:
mov ecx,16
mov eax,DWORD [240+edx]
movdqa xmm1,xmm6
movdqa xmm2,[ebp]
pandn xmm1,xmm0
pand xmm0,xmm6
movdqu xmm5,[edx]
db 102,15,56,0,208
movdqa xmm0,[16+ebp]
pxor xmm2,xmm5
psrld xmm1,4
add edx,16
db 102,15,56,0,193
lea ebx,[192+ebp]
pxor xmm0,xmm2
jmp NEAR L$000enc_entry
align 16
L$001enc_loop:
movdqa xmm4,[32+ebp]
movdqa xmm0,[48+ebp]
db 102,15,56,0,226
db 102,15,56,0,195
pxor xmm4,xmm5
movdqa xmm5,[64+ebp]
pxor xmm0,xmm4
movdqa xmm1,[ecx*1+ebx-64]
db 102,15,56,0,234
movdqa xmm2,[80+ebp]
movdqa xmm4,[ecx*1+ebx]
db 102,15,56,0,211
movdqa xmm3,xmm0
pxor xmm2,xmm5
db 102,15,56,0,193
add edx,16
pxor xmm0,xmm2
db 102,15,56,0,220
add ecx,16
pxor xmm3,xmm0
db 102,15,56,0,193
and ecx,48
sub eax,1
pxor xmm0,xmm3
L$000enc_entry:
movdqa xmm1,xmm6
movdqa xmm5,[ebp-32]
pandn xmm1,xmm0
psrld xmm1,4
pand xmm0,xmm6
db 102,15,56,0,232
movdqa xmm3,xmm7
pxor xmm0,xmm1
db 102,15,56,0,217
movdqa xmm4,xmm7
pxor xmm3,xmm5
db 102,15,56,0,224
movdqa xmm2,xmm7
pxor xmm4,xmm5
db 102,15,56,0,211
movdqa xmm3,xmm7
pxor xmm2,xmm0
db 102,15,56,0,220
movdqu xmm5,[edx]
pxor xmm3,xmm1
jnz NEAR L$001enc_loop
movdqa xmm4,[96+ebp]
movdqa xmm0,[112+ebp]
db 102,15,56,0,226
pxor xmm4,xmm5
db 102,15,56,0,195
movdqa xmm1,[64+ecx*1+ebx]
pxor xmm0,xmm4
db 102,15,56,0,193
ret
align 16
__vpaes_decrypt_core:
lea ebx,[608+ebp]
mov eax,DWORD [240+edx]
movdqa xmm1,xmm6
movdqa xmm2,[ebx-64]
pandn xmm1,xmm0
mov ecx,eax
psrld xmm1,4
movdqu xmm5,[edx]
shl ecx,4
pand xmm0,xmm6
db 102,15,56,0,208
movdqa xmm0,[ebx-48]
xor ecx,48
db 102,15,56,0,193
and ecx,48
pxor xmm2,xmm5
movdqa xmm5,[176+ebp]
pxor xmm0,xmm2
add edx,16
lea ecx,[ecx*1+ebx-352]
jmp NEAR L$002dec_entry
align 16
L$003dec_loop:
movdqa xmm4,[ebx-32]
movdqa xmm1,[ebx-16]
db 102,15,56,0,226
db 102,15,56,0,203
pxor xmm0,xmm4
movdqa xmm4,[ebx]
pxor xmm0,xmm1
movdqa xmm1,[16+ebx]
db 102,15,56,0,226
db 102,15,56,0,197
db 102,15,56,0,203
pxor xmm0,xmm4
movdqa xmm4,[32+ebx]
pxor xmm0,xmm1
movdqa xmm1,[48+ebx]
db 102,15,56,0,226
db 102,15,56,0,197
db 102,15,56,0,203
pxor xmm0,xmm4
movdqa xmm4,[64+ebx]
pxor xmm0,xmm1
movdqa xmm1,[80+ebx]
db 102,15,56,0,226
db 102,15,56,0,197
db 102,15,56,0,203
pxor xmm0,xmm4
add edx,16
db 102,15,58,15,237,12
pxor xmm0,xmm1
sub eax,1
L$002dec_entry:
movdqa xmm1,xmm6
movdqa xmm2,[ebp-32]
pandn xmm1,xmm0
pand xmm0,xmm6
psrld xmm1,4
db 102,15,56,0,208
movdqa xmm3,xmm7
pxor xmm0,xmm1
db 102,15,56,0,217
movdqa xmm4,xmm7
pxor xmm3,xmm2
db 102,15,56,0,224
pxor xmm4,xmm2
movdqa xmm2,xmm7
db 102,15,56,0,211
movdqa xmm3,xmm7
pxor xmm2,xmm0
db 102,15,56,0,220
movdqu xmm0,[edx]
pxor xmm3,xmm1
jnz NEAR L$003dec_loop
movdqa xmm4,[96+ebx]
db 102,15,56,0,226
pxor xmm4,xmm0
movdqa xmm0,[112+ebx]
movdqa xmm2,[ecx]
db 102,15,56,0,195
pxor xmm0,xmm4
db 102,15,56,0,194
ret
align 16
__vpaes_schedule_core:
add ebp,DWORD [esp]
movdqu xmm0,[esi]
movdqa xmm2,[320+ebp]
movdqa xmm3,xmm0
lea ebx,[ebp]
movdqa [4+esp],xmm2
call __vpaes_schedule_transform
movdqa xmm7,xmm0
test edi,edi
jnz NEAR L$004schedule_am_decrypting
movdqu [edx],xmm0
jmp NEAR L$005schedule_go
L$004schedule_am_decrypting:
movdqa xmm1,[256+ecx*1+ebp]
db 102,15,56,0,217
movdqu [edx],xmm3
xor ecx,48
L$005schedule_go:
cmp eax,192
ja NEAR L$006schedule_256
je NEAR L$007schedule_192
L$008schedule_128:
mov eax,10
L$009loop_schedule_128:
call __vpaes_schedule_round
dec eax
jz NEAR L$010schedule_mangle_last
call __vpaes_schedule_mangle
jmp NEAR L$009loop_schedule_128
align 16
L$007schedule_192:
movdqu xmm0,[8+esi]
call __vpaes_schedule_transform
movdqa xmm6,xmm0
pxor xmm4,xmm4
movhlps xmm6,xmm4
mov eax,4
L$011loop_schedule_192:
call __vpaes_schedule_round
db 102,15,58,15,198,8
call __vpaes_schedule_mangle
call __vpaes_schedule_192_smear
call __vpaes_schedule_mangle
call __vpaes_schedule_round
dec eax
jz NEAR L$010schedule_mangle_last
call __vpaes_schedule_mangle
call __vpaes_schedule_192_smear
jmp NEAR L$011loop_schedule_192
align 16
L$006schedule_256:
movdqu xmm0,[16+esi]
call __vpaes_schedule_transform
mov eax,7
L$012loop_schedule_256:
call __vpaes_schedule_mangle
movdqa xmm6,xmm0
call __vpaes_schedule_round
dec eax
jz NEAR L$010schedule_mangle_last
call __vpaes_schedule_mangle
pshufd xmm0,xmm0,255
movdqa [20+esp],xmm7
movdqa xmm7,xmm6
call L$_vpaes_schedule_low_round
movdqa xmm7,[20+esp]
jmp NEAR L$012loop_schedule_256
align 16
L$010schedule_mangle_last:
lea ebx,[384+ebp]
test edi,edi
jnz NEAR L$013schedule_mangle_last_dec
movdqa xmm1,[256+ecx*1+ebp]
db 102,15,56,0,193
lea ebx,[352+ebp]
add edx,32
L$013schedule_mangle_last_dec:
add edx,-16
pxor xmm0,[336+ebp]
call __vpaes_schedule_transform
movdqu [edx],xmm0
pxor xmm0,xmm0
pxor xmm1,xmm1
pxor xmm2,xmm2
pxor xmm3,xmm3
pxor xmm4,xmm4
pxor xmm5,xmm5
pxor xmm6,xmm6
pxor xmm7,xmm7
ret
align 16
__vpaes_schedule_192_smear:
pshufd xmm1,xmm6,128
pshufd xmm0,xmm7,254
pxor xmm6,xmm1
pxor xmm1,xmm1
pxor xmm6,xmm0
movdqa xmm0,xmm6
movhlps xmm6,xmm1
ret
align 16
__vpaes_schedule_round:
movdqa xmm2,[8+esp]
pxor xmm1,xmm1
db 102,15,58,15,202,15
db 102,15,58,15,210,15
pxor xmm7,xmm1
pshufd xmm0,xmm0,255
db 102,15,58,15,192,1
movdqa [8+esp],xmm2
L$_vpaes_schedule_low_round:
movdqa xmm1,xmm7
pslldq xmm7,4
pxor xmm7,xmm1
movdqa xmm1,xmm7
pslldq xmm7,8
pxor xmm7,xmm1
pxor xmm7,[336+ebp]
movdqa xmm4,[ebp-16]
movdqa xmm5,[ebp-48]
movdqa xmm1,xmm4
pandn xmm1,xmm0
psrld xmm1,4
pand xmm0,xmm4
movdqa xmm2,[ebp-32]
db 102,15,56,0,208
pxor xmm0,xmm1
movdqa xmm3,xmm5
db 102,15,56,0,217
pxor xmm3,xmm2
movdqa xmm4,xmm5
db 102,15,56,0,224
pxor xmm4,xmm2
movdqa xmm2,xmm5
db 102,15,56,0,211
pxor xmm2,xmm0
movdqa xmm3,xmm5
db 102,15,56,0,220
pxor xmm3,xmm1
movdqa xmm4,[32+ebp]
db 102,15,56,0,226
movdqa xmm0,[48+ebp]
db 102,15,56,0,195
pxor xmm0,xmm4
pxor xmm0,xmm7
movdqa xmm7,xmm0
ret
align 16
__vpaes_schedule_transform:
movdqa xmm2,[ebp-16]
movdqa xmm1,xmm2
pandn xmm1,xmm0
psrld xmm1,4
pand xmm0,xmm2
movdqa xmm2,[ebx]
db 102,15,56,0,208
movdqa xmm0,[16+ebx]
db 102,15,56,0,193
pxor xmm0,xmm2
ret
align 16
__vpaes_schedule_mangle:
movdqa xmm4,xmm0
movdqa xmm5,[128+ebp]
test edi,edi
jnz NEAR L$014schedule_mangle_dec
add edx,16
pxor xmm4,[336+ebp]
db 102,15,56,0,229
movdqa xmm3,xmm4
db 102,15,56,0,229
pxor xmm3,xmm4
db 102,15,56,0,229
pxor xmm3,xmm4
jmp NEAR L$015schedule_mangle_both
align 16
L$014schedule_mangle_dec:
movdqa xmm2,[ebp-16]
lea esi,[416+ebp]
movdqa xmm1,xmm2
pandn xmm1,xmm4
psrld xmm1,4
pand xmm4,xmm2
movdqa xmm2,[esi]
db 102,15,56,0,212
movdqa xmm3,[16+esi]
db 102,15,56,0,217
pxor xmm3,xmm2
db 102,15,56,0,221
movdqa xmm2,[32+esi]
db 102,15,56,0,212
pxor xmm2,xmm3
movdqa xmm3,[48+esi]
db 102,15,56,0,217
pxor xmm3,xmm2
db 102,15,56,0,221
movdqa xmm2,[64+esi]
db 102,15,56,0,212
pxor xmm2,xmm3
movdqa xmm3,[80+esi]
db 102,15,56,0,217
pxor xmm3,xmm2
db 102,15,56,0,221
movdqa xmm2,[96+esi]
db 102,15,56,0,212
pxor xmm2,xmm3
movdqa xmm3,[112+esi]
db 102,15,56,0,217
pxor xmm3,xmm2
add edx,-16
L$015schedule_mangle_both:
movdqa xmm1,[256+ecx*1+ebp]
db 102,15,56,0,217
add ecx,-16
and ecx,48
movdqu [edx],xmm3
ret
global _vpaes_set_encrypt_key
align 16
_vpaes_set_encrypt_key:
L$_vpaes_set_encrypt_key_begin:
push ebp
push ebx
push esi
push edi
mov esi,DWORD [20+esp]
lea ebx,[esp-56]
mov eax,DWORD [24+esp]
and ebx,-16
mov edx,DWORD [28+esp]
xchg ebx,esp
mov DWORD [48+esp],ebx
mov ebx,eax
shr ebx,5
add ebx,5
mov DWORD [240+edx],ebx
mov ecx,48
mov edi,0
lea ebp,[(L$_vpaes_consts+0x30-L$016pic_point)]
call __vpaes_schedule_core
L$016pic_point:
mov esp,DWORD [48+esp]
xor eax,eax
pop edi
pop esi
pop ebx
pop ebp
ret
global _vpaes_set_decrypt_key
align 16
_vpaes_set_decrypt_key:
L$_vpaes_set_decrypt_key_begin:
push ebp
push ebx
push esi
push edi
mov esi,DWORD [20+esp]
lea ebx,[esp-56]
mov eax,DWORD [24+esp]
and ebx,-16
mov edx,DWORD [28+esp]
xchg ebx,esp
mov DWORD [48+esp],ebx
mov ebx,eax
shr ebx,5
add ebx,5
mov DWORD [240+edx],ebx
shl ebx,4
lea edx,[16+ebx*1+edx]
mov edi,1
mov ecx,eax
shr ecx,1
and ecx,32
xor ecx,32
lea ebp,[(L$_vpaes_consts+0x30-L$017pic_point)]
call __vpaes_schedule_core
L$017pic_point:
mov esp,DWORD [48+esp]
xor eax,eax
pop edi
pop esi
pop ebx
pop ebp
ret
global _vpaes_encrypt
align 16
_vpaes_encrypt:
L$_vpaes_encrypt_begin:
push ebp
push ebx
push esi
push edi
lea ebp,[(L$_vpaes_consts+0x30-L$018pic_point)]
call __vpaes_preheat
L$018pic_point:
mov esi,DWORD [20+esp]
lea ebx,[esp-56]
mov edi,DWORD [24+esp]
and ebx,-16
mov edx,DWORD [28+esp]
xchg ebx,esp
mov DWORD [48+esp],ebx
movdqu xmm0,[esi]
call __vpaes_encrypt_core
movdqu [edi],xmm0
mov esp,DWORD [48+esp]
pop edi
pop esi
pop ebx
pop ebp
ret
global _vpaes_decrypt
align 16
_vpaes_decrypt:
L$_vpaes_decrypt_begin:
push ebp
push ebx
push esi
push edi
lea ebp,[(L$_vpaes_consts+0x30-L$019pic_point)]
call __vpaes_preheat
L$019pic_point:
mov esi,DWORD [20+esp]
lea ebx,[esp-56]
mov edi,DWORD [24+esp]
and ebx,-16
mov edx,DWORD [28+esp]
xchg ebx,esp
mov DWORD [48+esp],ebx
movdqu xmm0,[esi]
call __vpaes_decrypt_core
movdqu [edi],xmm0
mov esp,DWORD [48+esp]
pop edi
pop esi
pop ebx
pop ebp
ret
global _vpaes_cbc_encrypt
align 16
_vpaes_cbc_encrypt:
L$_vpaes_cbc_encrypt_begin:
push ebp
push ebx
push esi
push edi
mov esi,DWORD [20+esp]
mov edi,DWORD [24+esp]
mov eax,DWORD [28+esp]
mov edx,DWORD [32+esp]
sub eax,16
jc NEAR L$020cbc_abort
lea ebx,[esp-56]
mov ebp,DWORD [36+esp]
and ebx,-16
mov ecx,DWORD [40+esp]
xchg ebx,esp
movdqu xmm1,[ebp]
sub edi,esi
mov DWORD [48+esp],ebx
mov DWORD [esp],edi
mov DWORD [4+esp],edx
mov DWORD [8+esp],ebp
mov edi,eax
lea ebp,[(L$_vpaes_consts+0x30-L$021pic_point)]
call __vpaes_preheat
L$021pic_point:
cmp ecx,0
je NEAR L$022cbc_dec_loop
jmp NEAR L$023cbc_enc_loop
align 16
L$023cbc_enc_loop:
movdqu xmm0,[esi]
pxor xmm0,xmm1
call __vpaes_encrypt_core
mov ebx,DWORD [esp]
mov edx,DWORD [4+esp]
movdqa xmm1,xmm0
movdqu [esi*1+ebx],xmm0
lea esi,[16+esi]
sub edi,16
jnc NEAR L$023cbc_enc_loop
jmp NEAR L$024cbc_done
align 16
L$022cbc_dec_loop:
movdqu xmm0,[esi]
movdqa [16+esp],xmm1
movdqa [32+esp],xmm0
call __vpaes_decrypt_core
mov ebx,DWORD [esp]
mov edx,DWORD [4+esp]
pxor xmm0,[16+esp]
movdqa xmm1,[32+esp]
movdqu [esi*1+ebx],xmm0
lea esi,[16+esi]
sub edi,16
jnc NEAR L$022cbc_dec_loop
L$024cbc_done:
mov ebx,DWORD [8+esp]
mov esp,DWORD [48+esp]
movdqu [ebx],xmm1
L$020cbc_abort:
pop edi
pop esi
pop ebx
pop ebp
ret

View File

@@ -0,0 +1,485 @@
%ifidn __OUTPUT_FORMAT__,obj
section code use32 class=code align=64
%elifidn __OUTPUT_FORMAT__,win32
%ifdef __YASM_VERSION_ID__
%if __YASM_VERSION_ID__ < 01010000h
%error yasm version 1.1.0 or later needed.
%endif
; Yasm automatically includes .00 and complains about redefining it.
; https://www.tortall.net/projects/yasm/manual/html/objfmt-win32-safeseh.html
%else
$@feat.00 equ 1
%endif
section .text code align=64
%else
section .text code
%endif
;extern _OPENSSL_ia32cap_P
global _bn_mul_mont
align 16
_bn_mul_mont:
L$_bn_mul_mont_begin:
push ebp
push ebx
push esi
push edi
xor eax,eax
mov edi,DWORD [40+esp]
cmp edi,4
jl NEAR L$000just_leave
lea esi,[20+esp]
lea edx,[24+esp]
add edi,2
neg edi
lea ebp,[edi*4+esp-32]
neg edi
mov eax,ebp
sub eax,edx
and eax,2047
sub ebp,eax
xor edx,ebp
and edx,2048
xor edx,2048
sub ebp,edx
and ebp,-64
mov eax,esp
sub eax,ebp
and eax,-4096
mov edx,esp
lea esp,[eax*1+ebp]
mov eax,DWORD [esp]
cmp esp,ebp
ja NEAR L$001page_walk
jmp NEAR L$002page_walk_done
align 16
L$001page_walk:
lea esp,[esp-4096]
mov eax,DWORD [esp]
cmp esp,ebp
ja NEAR L$001page_walk
L$002page_walk_done:
mov eax,DWORD [esi]
mov ebx,DWORD [4+esi]
mov ecx,DWORD [8+esi]
mov ebp,DWORD [12+esi]
mov esi,DWORD [16+esi]
mov esi,DWORD [esi]
mov DWORD [4+esp],eax
mov DWORD [8+esp],ebx
mov DWORD [12+esp],ecx
mov DWORD [16+esp],ebp
mov DWORD [20+esp],esi
lea ebx,[edi-3]
mov DWORD [24+esp],edx
lea eax,[_OPENSSL_ia32cap_P]
bt DWORD [eax],26
jnc NEAR L$003non_sse2
mov eax,-1
movd mm7,eax
mov esi,DWORD [8+esp]
mov edi,DWORD [12+esp]
mov ebp,DWORD [16+esp]
xor edx,edx
xor ecx,ecx
movd mm4,DWORD [edi]
movd mm5,DWORD [esi]
movd mm3,DWORD [ebp]
pmuludq mm5,mm4
movq mm2,mm5
movq mm0,mm5
pand mm0,mm7
pmuludq mm5,[20+esp]
pmuludq mm3,mm5
paddq mm3,mm0
movd mm1,DWORD [4+ebp]
movd mm0,DWORD [4+esi]
psrlq mm2,32
psrlq mm3,32
inc ecx
align 16
L$0041st:
pmuludq mm0,mm4
pmuludq mm1,mm5
paddq mm2,mm0
paddq mm3,mm1
movq mm0,mm2
pand mm0,mm7
movd mm1,DWORD [4+ecx*4+ebp]
paddq mm3,mm0
movd mm0,DWORD [4+ecx*4+esi]
psrlq mm2,32
movd DWORD [28+ecx*4+esp],mm3
psrlq mm3,32
lea ecx,[1+ecx]
cmp ecx,ebx
jl NEAR L$0041st
pmuludq mm0,mm4
pmuludq mm1,mm5
paddq mm2,mm0
paddq mm3,mm1
movq mm0,mm2
pand mm0,mm7
paddq mm3,mm0
movd DWORD [28+ecx*4+esp],mm3
psrlq mm2,32
psrlq mm3,32
paddq mm3,mm2
movq [32+ebx*4+esp],mm3
inc edx
L$005outer:
xor ecx,ecx
movd mm4,DWORD [edx*4+edi]
movd mm5,DWORD [esi]
movd mm6,DWORD [32+esp]
movd mm3,DWORD [ebp]
pmuludq mm5,mm4
paddq mm5,mm6
movq mm0,mm5
movq mm2,mm5
pand mm0,mm7
pmuludq mm5,[20+esp]
pmuludq mm3,mm5
paddq mm3,mm0
movd mm6,DWORD [36+esp]
movd mm1,DWORD [4+ebp]
movd mm0,DWORD [4+esi]
psrlq mm2,32
psrlq mm3,32
paddq mm2,mm6
inc ecx
dec ebx
L$006inner:
pmuludq mm0,mm4
pmuludq mm1,mm5
paddq mm2,mm0
paddq mm3,mm1
movq mm0,mm2
movd mm6,DWORD [36+ecx*4+esp]
pand mm0,mm7
movd mm1,DWORD [4+ecx*4+ebp]
paddq mm3,mm0
movd mm0,DWORD [4+ecx*4+esi]
psrlq mm2,32
movd DWORD [28+ecx*4+esp],mm3
psrlq mm3,32
paddq mm2,mm6
dec ebx
lea ecx,[1+ecx]
jnz NEAR L$006inner
mov ebx,ecx
pmuludq mm0,mm4
pmuludq mm1,mm5
paddq mm2,mm0
paddq mm3,mm1
movq mm0,mm2
pand mm0,mm7
paddq mm3,mm0
movd DWORD [28+ecx*4+esp],mm3
psrlq mm2,32
psrlq mm3,32
movd mm6,DWORD [36+ebx*4+esp]
paddq mm3,mm2
paddq mm3,mm6
movq [32+ebx*4+esp],mm3
lea edx,[1+edx]
cmp edx,ebx
jle NEAR L$005outer
emms
jmp NEAR L$007common_tail
align 16
L$003non_sse2:
mov esi,DWORD [8+esp]
lea ebp,[1+ebx]
mov edi,DWORD [12+esp]
xor ecx,ecx
mov edx,esi
and ebp,1
sub edx,edi
lea eax,[4+ebx*4+edi]
or ebp,edx
mov edi,DWORD [edi]
jz NEAR L$008bn_sqr_mont
mov DWORD [28+esp],eax
mov eax,DWORD [esi]
xor edx,edx
align 16
L$009mull:
mov ebp,edx
mul edi
add ebp,eax
lea ecx,[1+ecx]
adc edx,0
mov eax,DWORD [ecx*4+esi]
cmp ecx,ebx
mov DWORD [28+ecx*4+esp],ebp
jl NEAR L$009mull
mov ebp,edx
mul edi
mov edi,DWORD [20+esp]
add eax,ebp
mov esi,DWORD [16+esp]
adc edx,0
imul edi,DWORD [32+esp]
mov DWORD [32+ebx*4+esp],eax
xor ecx,ecx
mov DWORD [36+ebx*4+esp],edx
mov DWORD [40+ebx*4+esp],ecx
mov eax,DWORD [esi]
mul edi
add eax,DWORD [32+esp]
mov eax,DWORD [4+esi]
adc edx,0
inc ecx
jmp NEAR L$0102ndmadd
align 16
L$0111stmadd:
mov ebp,edx
mul edi
add ebp,DWORD [32+ecx*4+esp]
lea ecx,[1+ecx]
adc edx,0
add ebp,eax
mov eax,DWORD [ecx*4+esi]
adc edx,0
cmp ecx,ebx
mov DWORD [28+ecx*4+esp],ebp
jl NEAR L$0111stmadd
mov ebp,edx
mul edi
add eax,DWORD [32+ebx*4+esp]
mov edi,DWORD [20+esp]
adc edx,0
mov esi,DWORD [16+esp]
add ebp,eax
adc edx,0
imul edi,DWORD [32+esp]
xor ecx,ecx
add edx,DWORD [36+ebx*4+esp]
mov DWORD [32+ebx*4+esp],ebp
adc ecx,0
mov eax,DWORD [esi]
mov DWORD [36+ebx*4+esp],edx
mov DWORD [40+ebx*4+esp],ecx
mul edi
add eax,DWORD [32+esp]
mov eax,DWORD [4+esi]
adc edx,0
mov ecx,1
align 16
L$0102ndmadd:
mov ebp,edx
mul edi
add ebp,DWORD [32+ecx*4+esp]
lea ecx,[1+ecx]
adc edx,0
add ebp,eax
mov eax,DWORD [ecx*4+esi]
adc edx,0
cmp ecx,ebx
mov DWORD [24+ecx*4+esp],ebp
jl NEAR L$0102ndmadd
mov ebp,edx
mul edi
add ebp,DWORD [32+ebx*4+esp]
adc edx,0
add ebp,eax
adc edx,0
mov DWORD [28+ebx*4+esp],ebp
xor eax,eax
mov ecx,DWORD [12+esp]
add edx,DWORD [36+ebx*4+esp]
adc eax,DWORD [40+ebx*4+esp]
lea ecx,[4+ecx]
mov DWORD [32+ebx*4+esp],edx
cmp ecx,DWORD [28+esp]
mov DWORD [36+ebx*4+esp],eax
je NEAR L$007common_tail
mov edi,DWORD [ecx]
mov esi,DWORD [8+esp]
mov DWORD [12+esp],ecx
xor ecx,ecx
xor edx,edx
mov eax,DWORD [esi]
jmp NEAR L$0111stmadd
align 16
L$008bn_sqr_mont:
mov DWORD [esp],ebx
mov DWORD [12+esp],ecx
mov eax,edi
mul edi
mov DWORD [32+esp],eax
mov ebx,edx
shr edx,1
and ebx,1
inc ecx
align 16
L$012sqr:
mov eax,DWORD [ecx*4+esi]
mov ebp,edx
mul edi
add eax,ebp
lea ecx,[1+ecx]
adc edx,0
lea ebp,[eax*2+ebx]
shr eax,31
cmp ecx,DWORD [esp]
mov ebx,eax
mov DWORD [28+ecx*4+esp],ebp
jl NEAR L$012sqr
mov eax,DWORD [ecx*4+esi]
mov ebp,edx
mul edi
add eax,ebp
mov edi,DWORD [20+esp]
adc edx,0
mov esi,DWORD [16+esp]
lea ebp,[eax*2+ebx]
imul edi,DWORD [32+esp]
shr eax,31
mov DWORD [32+ecx*4+esp],ebp
lea ebp,[edx*2+eax]
mov eax,DWORD [esi]
shr edx,31
mov DWORD [36+ecx*4+esp],ebp
mov DWORD [40+ecx*4+esp],edx
mul edi
add eax,DWORD [32+esp]
mov ebx,ecx
adc edx,0
mov eax,DWORD [4+esi]
mov ecx,1
align 16
L$0133rdmadd:
mov ebp,edx
mul edi
add ebp,DWORD [32+ecx*4+esp]
adc edx,0
add ebp,eax
mov eax,DWORD [4+ecx*4+esi]
adc edx,0
mov DWORD [28+ecx*4+esp],ebp
mov ebp,edx
mul edi
add ebp,DWORD [36+ecx*4+esp]
lea ecx,[2+ecx]
adc edx,0
add ebp,eax
mov eax,DWORD [ecx*4+esi]
adc edx,0
cmp ecx,ebx
mov DWORD [24+ecx*4+esp],ebp
jl NEAR L$0133rdmadd
mov ebp,edx
mul edi
add ebp,DWORD [32+ebx*4+esp]
adc edx,0
add ebp,eax
adc edx,0
mov DWORD [28+ebx*4+esp],ebp
mov ecx,DWORD [12+esp]
xor eax,eax
mov esi,DWORD [8+esp]
add edx,DWORD [36+ebx*4+esp]
adc eax,DWORD [40+ebx*4+esp]
mov DWORD [32+ebx*4+esp],edx
cmp ecx,ebx
mov DWORD [36+ebx*4+esp],eax
je NEAR L$007common_tail
mov edi,DWORD [4+ecx*4+esi]
lea ecx,[1+ecx]
mov eax,edi
mov DWORD [12+esp],ecx
mul edi
add eax,DWORD [32+ecx*4+esp]
adc edx,0
mov DWORD [32+ecx*4+esp],eax
xor ebp,ebp
cmp ecx,ebx
lea ecx,[1+ecx]
je NEAR L$014sqrlast
mov ebx,edx
shr edx,1
and ebx,1
align 16
L$015sqradd:
mov eax,DWORD [ecx*4+esi]
mov ebp,edx
mul edi
add eax,ebp
lea ebp,[eax*1+eax]
adc edx,0
shr eax,31
add ebp,DWORD [32+ecx*4+esp]
lea ecx,[1+ecx]
adc eax,0
add ebp,ebx
adc eax,0
cmp ecx,DWORD [esp]
mov DWORD [28+ecx*4+esp],ebp
mov ebx,eax
jle NEAR L$015sqradd
mov ebp,edx
add edx,edx
shr ebp,31
add edx,ebx
adc ebp,0
L$014sqrlast:
mov edi,DWORD [20+esp]
mov esi,DWORD [16+esp]
imul edi,DWORD [32+esp]
add edx,DWORD [32+ecx*4+esp]
mov eax,DWORD [esi]
adc ebp,0
mov DWORD [32+ecx*4+esp],edx
mov DWORD [36+ecx*4+esp],ebp
mul edi
add eax,DWORD [32+esp]
lea ebx,[ecx-1]
adc edx,0
mov ecx,1
mov eax,DWORD [4+esi]
jmp NEAR L$0133rdmadd
align 16
L$007common_tail:
mov ebp,DWORD [16+esp]
mov edi,DWORD [4+esp]
lea esi,[32+esp]
mov eax,DWORD [esi]
mov ecx,ebx
xor edx,edx
align 16
L$016sub:
sbb eax,DWORD [edx*4+ebp]
mov DWORD [edx*4+edi],eax
dec ecx
mov eax,DWORD [4+edx*4+esi]
lea edx,[1+edx]
jge NEAR L$016sub
sbb eax,0
and esi,eax
not eax
mov ebp,edi
and ebp,eax
or esi,ebp
align 16
L$017copy:
mov eax,DWORD [ebx*4+esi]
mov DWORD [ebx*4+edi],eax
mov DWORD [32+ebx*4+esp],ecx
dec ebx
jge NEAR L$017copy
mov esp,DWORD [24+esp]
mov eax,1
L$000just_leave:
pop edi
pop esi
pop ebx
pop ebp
ret
db 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
db 112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
db 54,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121
db 32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
db 111,114,103,62,0
segment .bss
common _OPENSSL_ia32cap_P 16