summaryrefslogtreecommitdiffstats
path: root/win-x86/crypto/sha
diff options
context:
space:
mode:
Diffstat (limited to 'win-x86/crypto/sha')
-rw-r--r--win-x86/crypto/sha/sha1-586.asm182
-rw-r--r--win-x86/crypto/sha/sha256-586.asm236
2 files changed, 394 insertions, 24 deletions
diff --git a/win-x86/crypto/sha/sha1-586.asm b/win-x86/crypto/sha/sha1-586.asm
index 43bf964..e24449d 100644
--- a/win-x86/crypto/sha/sha1-586.asm
+++ b/win-x86/crypto/sha/sha1-586.asm
@@ -35,6 +35,8 @@ L$000pic_point:
mov ecx,DWORD [8+esi]
test eax,16777216
jz NEAR L$001x86
+ test ecx,536870912
+ jnz NEAR L$shaext_shortcut
jmp NEAR L$ssse3_shortcut
align 16
L$001x86:
@@ -1403,7 +1405,7 @@ L$002loop:
pop ebp
ret
align 16
-__sha1_block_data_order_ssse3:
+__sha1_block_data_order_shaext:
push ebp
push ebx
push esi
@@ -1412,6 +1414,174 @@ __sha1_block_data_order_ssse3:
L$003pic_point:
pop ebp
lea ebp,[(L$K_XX_XX-L$003pic_point)+ebp]
+L$shaext_shortcut:
+ mov edi,DWORD [20+esp]
+ mov ebx,esp
+ mov esi,DWORD [24+esp]
+ mov ecx,DWORD [28+esp]
+ sub esp,32
+ movdqu xmm0,[edi]
+ movd xmm1,DWORD [16+edi]
+ and esp,-32
+ movdqa xmm3,[80+ebp]
+ movdqu xmm4,[esi]
+ pshufd xmm0,xmm0,27
+ movdqu xmm5,[16+esi]
+ pshufd xmm1,xmm1,27
+ movdqu xmm6,[32+esi]
+db 102,15,56,0,227
+ movdqu xmm7,[48+esi]
+db 102,15,56,0,235
+db 102,15,56,0,243
+db 102,15,56,0,251
+ jmp NEAR L$004loop_shaext
+align 16
+L$004loop_shaext:
+ dec ecx
+ lea eax,[64+esi]
+ movdqa [esp],xmm1
+ paddd xmm1,xmm4
+ cmovne esi,eax
+ movdqa [16+esp],xmm0
+db 15,56,201,229
+ movdqa xmm2,xmm0
+db 15,58,204,193,0
+db 15,56,200,213
+ pxor xmm4,xmm6
+db 15,56,201,238
+db 15,56,202,231
+ movdqa xmm1,xmm0
+db 15,58,204,194,0
+db 15,56,200,206
+ pxor xmm5,xmm7
+db 15,56,202,236
+db 15,56,201,247
+ movdqa xmm2,xmm0
+db 15,58,204,193,0
+db 15,56,200,215
+ pxor xmm6,xmm4
+db 15,56,201,252
+db 15,56,202,245
+ movdqa xmm1,xmm0
+db 15,58,204,194,0
+db 15,56,200,204
+ pxor xmm7,xmm5
+db 15,56,202,254
+db 15,56,201,229
+ movdqa xmm2,xmm0
+db 15,58,204,193,0
+db 15,56,200,213
+ pxor xmm4,xmm6
+db 15,56,201,238
+db 15,56,202,231
+ movdqa xmm1,xmm0
+db 15,58,204,194,1
+db 15,56,200,206
+ pxor xmm5,xmm7
+db 15,56,202,236
+db 15,56,201,247
+ movdqa xmm2,xmm0
+db 15,58,204,193,1
+db 15,56,200,215
+ pxor xmm6,xmm4
+db 15,56,201,252
+db 15,56,202,245
+ movdqa xmm1,xmm0
+db 15,58,204,194,1
+db 15,56,200,204
+ pxor xmm7,xmm5
+db 15,56,202,254
+db 15,56,201,229
+ movdqa xmm2,xmm0
+db 15,58,204,193,1
+db 15,56,200,213
+ pxor xmm4,xmm6
+db 15,56,201,238
+db 15,56,202,231
+ movdqa xmm1,xmm0
+db 15,58,204,194,1
+db 15,56,200,206
+ pxor xmm5,xmm7
+db 15,56,202,236
+db 15,56,201,247
+ movdqa xmm2,xmm0
+db 15,58,204,193,2
+db 15,56,200,215
+ pxor xmm6,xmm4
+db 15,56,201,252
+db 15,56,202,245
+ movdqa xmm1,xmm0
+db 15,58,204,194,2
+db 15,56,200,204
+ pxor xmm7,xmm5
+db 15,56,202,254
+db 15,56,201,229
+ movdqa xmm2,xmm0
+db 15,58,204,193,2
+db 15,56,200,213
+ pxor xmm4,xmm6
+db 15,56,201,238
+db 15,56,202,231
+ movdqa xmm1,xmm0
+db 15,58,204,194,2
+db 15,56,200,206
+ pxor xmm5,xmm7
+db 15,56,202,236
+db 15,56,201,247
+ movdqa xmm2,xmm0
+db 15,58,204,193,2
+db 15,56,200,215
+ pxor xmm6,xmm4
+db 15,56,201,252
+db 15,56,202,245
+ movdqa xmm1,xmm0
+db 15,58,204,194,3
+db 15,56,200,204
+ pxor xmm7,xmm5
+db 15,56,202,254
+ movdqu xmm4,[esi]
+ movdqa xmm2,xmm0
+db 15,58,204,193,3
+db 15,56,200,213
+ movdqu xmm5,[16+esi]
+db 102,15,56,0,227
+ movdqa xmm1,xmm0
+db 15,58,204,194,3
+db 15,56,200,206
+ movdqu xmm6,[32+esi]
+db 102,15,56,0,235
+ movdqa xmm2,xmm0
+db 15,58,204,193,3
+db 15,56,200,215
+ movdqu xmm7,[48+esi]
+db 102,15,56,0,243
+ movdqa xmm1,xmm0
+db 15,58,204,194,3
+ movdqa xmm2,[esp]
+db 102,15,56,0,251
+db 15,56,200,202
+ paddd xmm0,[16+esp]
+ jnz NEAR L$004loop_shaext
+ pshufd xmm0,xmm0,27
+ pshufd xmm1,xmm1,27
+ movdqu [edi],xmm0
+ movd DWORD [16+edi],xmm1
+ mov esp,ebx
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+align 16
+__sha1_block_data_order_ssse3:
+ push ebp
+ push ebx
+ push esi
+ push edi
+ call L$005pic_point
+L$005pic_point:
+ pop ebp
+ lea ebp,[(L$K_XX_XX-L$005pic_point)+ebp]
L$ssse3_shortcut:
movdqa xmm7,[ebp]
movdqa xmm0,[16+ebp]
@@ -1464,9 +1634,9 @@ db 102,15,56,0,222
xor ebp,edx
pshufd xmm4,xmm0,238
and esi,ebp
- jmp NEAR L$004loop
+ jmp NEAR L$006loop
align 16
-L$004loop:
+L$006loop:
ror ebx,2
xor esi,edx
mov ebp,eax
@@ -2369,7 +2539,7 @@ L$004loop:
add ecx,edx
mov ebp,DWORD [196+esp]
cmp ebp,DWORD [200+esp]
- je NEAR L$005done
+ je NEAR L$007done
movdqa xmm7,[160+esp]
movdqa xmm6,[176+esp]
movdqu xmm0,[ebp]
@@ -2504,9 +2674,9 @@ db 102,15,56,0,222
pshufd xmm4,xmm0,238
and esi,ebx
mov ebx,ebp
- jmp NEAR L$004loop
+ jmp NEAR L$006loop
align 16
-L$005done:
+L$007done:
add ebx,DWORD [16+esp]
xor esi,edi
mov ebp,ecx
diff --git a/win-x86/crypto/sha/sha256-586.asm b/win-x86/crypto/sha/sha256-586.asm
index d03558c..fe36bc5 100644
--- a/win-x86/crypto/sha/sha256-586.asm
+++ b/win-x86/crypto/sha/sha256-586.asm
@@ -49,15 +49,17 @@ L$000pic_point:
jz NEAR L$003no_xmm
and ecx,1073741824
and ebx,268435968
+ test edx,536870912
+ jnz NEAR L$004shaext
or ecx,ebx
and ecx,1342177280
cmp ecx,1342177280
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 +131,7 @@ 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 +169,11 @@ 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 +228,7 @@ 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 +272,7 @@ 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 +289,9 @@ 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 +3171,7 @@ 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 +3179,205 @@ L$008grand_loop:
pop ebp
ret
align 32
-L$004SSSE3:
+L$004shaext:
+ sub esp,32
+ movdqu xmm1,[esi]
+ lea ebp,[128+ebp]
+ movdqu xmm2,[16+esi]
+ movdqa xmm7,[128+ebp]
+ pshufd xmm0,xmm1,27
+ pshufd xmm1,xmm1,177
+ pshufd xmm2,xmm2,27
+db 102,15,58,15,202,8
+ punpcklqdq xmm2,xmm0
+ jmp NEAR L$010loop_shaext
+align 16
+L$010loop_shaext:
+ movdqu xmm3,[edi]
+ movdqu xmm4,[16+edi]
+ movdqu xmm5,[32+edi]
+db 102,15,56,0,223
+ movdqu xmm6,[48+edi]
+ movdqa [16+esp],xmm2
+ movdqa xmm0,[ebp-128]
+ paddd xmm0,xmm3
+db 102,15,56,0,231
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ nop
+ movdqa [esp],xmm1
+db 15,56,203,202
+ movdqa xmm0,[ebp-112]
+ paddd xmm0,xmm4
+db 102,15,56,0,239
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ lea edi,[64+edi]
+db 15,56,204,220
+db 15,56,203,202
+ movdqa xmm0,[ebp-96]
+ paddd xmm0,xmm5
+db 102,15,56,0,247
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ movdqa xmm7,xmm6
+db 102,15,58,15,253,4
+ nop
+ paddd xmm3,xmm7
+db 15,56,204,229
+db 15,56,203,202
+ movdqa xmm0,[ebp-80]
+ paddd xmm0,xmm6
+db 15,56,205,222
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ movdqa xmm7,xmm3
+db 102,15,58,15,254,4
+ nop
+ paddd xmm4,xmm7
+db 15,56,204,238
+db 15,56,203,202
+ movdqa xmm0,[ebp-64]
+ paddd xmm0,xmm3
+db 15,56,205,227
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ movdqa xmm7,xmm4
+db 102,15,58,15,251,4
+ nop
+ paddd xmm5,xmm7
+db 15,56,204,243
+db 15,56,203,202
+ movdqa xmm0,[ebp-48]
+ paddd xmm0,xmm4
+db 15,56,205,236
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ movdqa xmm7,xmm5
+db 102,15,58,15,252,4
+ nop
+ paddd xmm6,xmm7
+db 15,56,204,220
+db 15,56,203,202
+ movdqa xmm0,[ebp-32]
+ paddd xmm0,xmm5
+db 15,56,205,245
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ movdqa xmm7,xmm6
+db 102,15,58,15,253,4
+ nop
+ paddd xmm3,xmm7
+db 15,56,204,229
+db 15,56,203,202
+ movdqa xmm0,[ebp-16]
+ paddd xmm0,xmm6
+db 15,56,205,222
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ movdqa xmm7,xmm3
+db 102,15,58,15,254,4
+ nop
+ paddd xmm4,xmm7
+db 15,56,204,238
+db 15,56,203,202
+ movdqa xmm0,[ebp]
+ paddd xmm0,xmm3
+db 15,56,205,227
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ movdqa xmm7,xmm4
+db 102,15,58,15,251,4
+ nop
+ paddd xmm5,xmm7
+db 15,56,204,243
+db 15,56,203,202
+ movdqa xmm0,[16+ebp]
+ paddd xmm0,xmm4
+db 15,56,205,236
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ movdqa xmm7,xmm5
+db 102,15,58,15,252,4
+ nop
+ paddd xmm6,xmm7
+db 15,56,204,220
+db 15,56,203,202
+ movdqa xmm0,[32+ebp]
+ paddd xmm0,xmm5
+db 15,56,205,245
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ movdqa xmm7,xmm6
+db 102,15,58,15,253,4
+ nop
+ paddd xmm3,xmm7
+db 15,56,204,229
+db 15,56,203,202
+ movdqa xmm0,[48+ebp]
+ paddd xmm0,xmm6
+db 15,56,205,222
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ movdqa xmm7,xmm3
+db 102,15,58,15,254,4
+ nop
+ paddd xmm4,xmm7
+db 15,56,204,238
+db 15,56,203,202
+ movdqa xmm0,[64+ebp]
+ paddd xmm0,xmm3
+db 15,56,205,227
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ movdqa xmm7,xmm4
+db 102,15,58,15,251,4
+ nop
+ paddd xmm5,xmm7
+db 15,56,204,243
+db 15,56,203,202
+ movdqa xmm0,[80+ebp]
+ paddd xmm0,xmm4
+db 15,56,205,236
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ movdqa xmm7,xmm5
+db 102,15,58,15,252,4
+db 15,56,203,202
+ paddd xmm6,xmm7
+ movdqa xmm0,[96+ebp]
+ paddd xmm0,xmm5
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+db 15,56,205,245
+ movdqa xmm7,[128+ebp]
+db 15,56,203,202
+ movdqa xmm0,[112+ebp]
+ paddd xmm0,xmm6
+ nop
+db 15,56,203,209
+ pshufd xmm0,xmm0,14
+ cmp eax,edi
+ nop
+db 15,56,203,202
+ paddd xmm2,[16+esp]
+ paddd xmm1,[esp]
+ jnz NEAR L$010loop_shaext
+ pshufd xmm2,xmm2,177
+ pshufd xmm7,xmm1,27
+ pshufd xmm1,xmm1,177
+ punpckhqdq xmm1,xmm2
+db 102,15,58,15,215,8
+ mov esp,DWORD [44+esp]
+ movdqu [esi],xmm1
+ movdqu [16+esi],xmm2
+ pop edi
+ pop esi
+ pop ebx
+ pop ebp
+ ret
+align 32
+L$005SSSE3:
lea esp,[esp-96]
mov eax,DWORD [esi]
mov ebx,DWORD [4+esi]
@@ -3196,9 +3396,9 @@ L$004SSSE3:
mov DWORD [24+esp],ecx
mov DWORD [28+esp],esi
movdqa xmm7,[256+ebp]
- jmp NEAR L$009grand_ssse3
+ jmp NEAR L$011grand_ssse3
align 16
-L$009grand_ssse3:
+L$011grand_ssse3:
movdqu xmm0,[edi]
movdqu xmm1,[16+edi]
movdqu xmm2,[32+edi]
@@ -3221,9 +3421,9 @@ 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$012ssse3_00_47
align 16
-L$010ssse3_00_47:
+L$012ssse3_00_47:
add ebp,64
mov ecx,edx
movdqa xmm4,xmm1
@@ -3866,7 +4066,7 @@ 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$012ssse3_00_47
mov ecx,edx
ror edx,14
mov esi,DWORD [20+esp]
@@ -4380,7 +4580,7 @@ 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$011grand_ssse3
mov esp,DWORD [108+esp]
pop edi
pop esi