mirror of
https://github.com/CloverHackyColor/CloverBootloader.git
synced 2025-01-25 21:31:32 +01:00
3269 lines
91 KiB
NASM
3269 lines
91 KiB
NASM
default rel
|
|
%define XMMWORD
|
|
%define YMMWORD
|
|
%define ZMMWORD
|
|
section .text code align=64
|
|
|
|
EXTERN OPENSSL_ia32cap_P
|
|
|
|
global aesni_cbc_sha1_enc
|
|
|
|
ALIGN 32
|
|
aesni_cbc_sha1_enc:
|
|
|
|
|
|
mov r10d,DWORD[((OPENSSL_ia32cap_P+0))]
|
|
mov r11,QWORD[((OPENSSL_ia32cap_P+4))]
|
|
bt r11,61
|
|
jc NEAR aesni_cbc_sha1_enc_shaext
|
|
and r11d,268435456
|
|
and r10d,1073741824
|
|
or r10d,r11d
|
|
cmp r10d,1342177280
|
|
je NEAR aesni_cbc_sha1_enc_avx
|
|
jmp NEAR aesni_cbc_sha1_enc_ssse3
|
|
DB 0F3h,0C3h ;repret
|
|
|
|
|
|
|
|
ALIGN 32
|
|
aesni_cbc_sha1_enc_ssse3:
|
|
mov QWORD[8+rsp],rdi ;WIN64 prologue
|
|
mov QWORD[16+rsp],rsi
|
|
mov rax,rsp
|
|
$L$SEH_begin_aesni_cbc_sha1_enc_ssse3:
|
|
mov rdi,rcx
|
|
mov rsi,rdx
|
|
mov rdx,r8
|
|
mov rcx,r9
|
|
mov r8,QWORD[40+rsp]
|
|
mov r9,QWORD[48+rsp]
|
|
|
|
|
|
|
|
mov r10,QWORD[56+rsp]
|
|
|
|
|
|
push rbx
|
|
|
|
push rbp
|
|
|
|
push r12
|
|
|
|
push r13
|
|
|
|
push r14
|
|
|
|
push r15
|
|
|
|
lea rsp,[((-264))+rsp]
|
|
|
|
|
|
|
|
movaps XMMWORD[(96+0)+rsp],xmm6
|
|
movaps XMMWORD[(96+16)+rsp],xmm7
|
|
movaps XMMWORD[(96+32)+rsp],xmm8
|
|
movaps XMMWORD[(96+48)+rsp],xmm9
|
|
movaps XMMWORD[(96+64)+rsp],xmm10
|
|
movaps XMMWORD[(96+80)+rsp],xmm11
|
|
movaps XMMWORD[(96+96)+rsp],xmm12
|
|
movaps XMMWORD[(96+112)+rsp],xmm13
|
|
movaps XMMWORD[(96+128)+rsp],xmm14
|
|
movaps XMMWORD[(96+144)+rsp],xmm15
|
|
$L$prologue_ssse3:
|
|
mov r12,rdi
|
|
mov r13,rsi
|
|
mov r14,rdx
|
|
lea r15,[112+rcx]
|
|
movdqu xmm2,XMMWORD[r8]
|
|
mov QWORD[88+rsp],r8
|
|
shl r14,6
|
|
sub r13,r12
|
|
mov r8d,DWORD[((240-112))+r15]
|
|
add r14,r10
|
|
|
|
lea r11,[K_XX_XX]
|
|
mov eax,DWORD[r9]
|
|
mov ebx,DWORD[4+r9]
|
|
mov ecx,DWORD[8+r9]
|
|
mov edx,DWORD[12+r9]
|
|
mov esi,ebx
|
|
mov ebp,DWORD[16+r9]
|
|
mov edi,ecx
|
|
xor edi,edx
|
|
and esi,edi
|
|
|
|
movdqa xmm3,XMMWORD[64+r11]
|
|
movdqa xmm13,XMMWORD[r11]
|
|
movdqu xmm4,XMMWORD[r10]
|
|
movdqu xmm5,XMMWORD[16+r10]
|
|
movdqu xmm6,XMMWORD[32+r10]
|
|
movdqu xmm7,XMMWORD[48+r10]
|
|
DB 102,15,56,0,227
|
|
DB 102,15,56,0,235
|
|
DB 102,15,56,0,243
|
|
add r10,64
|
|
paddd xmm4,xmm13
|
|
DB 102,15,56,0,251
|
|
paddd xmm5,xmm13
|
|
paddd xmm6,xmm13
|
|
movdqa XMMWORD[rsp],xmm4
|
|
psubd xmm4,xmm13
|
|
movdqa XMMWORD[16+rsp],xmm5
|
|
psubd xmm5,xmm13
|
|
movdqa XMMWORD[32+rsp],xmm6
|
|
psubd xmm6,xmm13
|
|
movups xmm15,XMMWORD[((-112))+r15]
|
|
movups xmm0,XMMWORD[((16-112))+r15]
|
|
jmp NEAR $L$oop_ssse3
|
|
ALIGN 32
|
|
$L$oop_ssse3:
|
|
ror ebx,2
|
|
movups xmm14,XMMWORD[r12]
|
|
xorps xmm14,xmm15
|
|
xorps xmm2,xmm14
|
|
movups xmm1,XMMWORD[((-80))+r15]
|
|
DB 102,15,56,220,208
|
|
pshufd xmm8,xmm4,238
|
|
xor esi,edx
|
|
movdqa xmm12,xmm7
|
|
paddd xmm13,xmm7
|
|
mov edi,eax
|
|
add ebp,DWORD[rsp]
|
|
punpcklqdq xmm8,xmm5
|
|
xor ebx,ecx
|
|
rol eax,5
|
|
add ebp,esi
|
|
psrldq xmm12,4
|
|
and edi,ebx
|
|
xor ebx,ecx
|
|
pxor xmm8,xmm4
|
|
add ebp,eax
|
|
ror eax,7
|
|
pxor xmm12,xmm6
|
|
xor edi,ecx
|
|
mov esi,ebp
|
|
add edx,DWORD[4+rsp]
|
|
pxor xmm8,xmm12
|
|
xor eax,ebx
|
|
rol ebp,5
|
|
movdqa XMMWORD[48+rsp],xmm13
|
|
add edx,edi
|
|
movups xmm0,XMMWORD[((-64))+r15]
|
|
DB 102,15,56,220,209
|
|
and esi,eax
|
|
movdqa xmm3,xmm8
|
|
xor eax,ebx
|
|
add edx,ebp
|
|
ror ebp,7
|
|
movdqa xmm12,xmm8
|
|
xor esi,ebx
|
|
pslldq xmm3,12
|
|
paddd xmm8,xmm8
|
|
mov edi,edx
|
|
add ecx,DWORD[8+rsp]
|
|
psrld xmm12,31
|
|
xor ebp,eax
|
|
rol edx,5
|
|
add ecx,esi
|
|
movdqa xmm13,xmm3
|
|
and edi,ebp
|
|
xor ebp,eax
|
|
psrld xmm3,30
|
|
add ecx,edx
|
|
ror edx,7
|
|
por xmm8,xmm12
|
|
xor edi,eax
|
|
mov esi,ecx
|
|
add ebx,DWORD[12+rsp]
|
|
movups xmm1,XMMWORD[((-48))+r15]
|
|
DB 102,15,56,220,208
|
|
pslld xmm13,2
|
|
pxor xmm8,xmm3
|
|
xor edx,ebp
|
|
movdqa xmm3,XMMWORD[r11]
|
|
rol ecx,5
|
|
add ebx,edi
|
|
and esi,edx
|
|
pxor xmm8,xmm13
|
|
xor edx,ebp
|
|
add ebx,ecx
|
|
ror ecx,7
|
|
pshufd xmm9,xmm5,238
|
|
xor esi,ebp
|
|
movdqa xmm13,xmm8
|
|
paddd xmm3,xmm8
|
|
mov edi,ebx
|
|
add eax,DWORD[16+rsp]
|
|
punpcklqdq xmm9,xmm6
|
|
xor ecx,edx
|
|
rol ebx,5
|
|
add eax,esi
|
|
psrldq xmm13,4
|
|
and edi,ecx
|
|
xor ecx,edx
|
|
pxor xmm9,xmm5
|
|
add eax,ebx
|
|
ror ebx,7
|
|
movups xmm0,XMMWORD[((-32))+r15]
|
|
DB 102,15,56,220,209
|
|
pxor xmm13,xmm7
|
|
xor edi,edx
|
|
mov esi,eax
|
|
add ebp,DWORD[20+rsp]
|
|
pxor xmm9,xmm13
|
|
xor ebx,ecx
|
|
rol eax,5
|
|
movdqa XMMWORD[rsp],xmm3
|
|
add ebp,edi
|
|
and esi,ebx
|
|
movdqa xmm12,xmm9
|
|
xor ebx,ecx
|
|
add ebp,eax
|
|
ror eax,7
|
|
movdqa xmm13,xmm9
|
|
xor esi,ecx
|
|
pslldq xmm12,12
|
|
paddd xmm9,xmm9
|
|
mov edi,ebp
|
|
add edx,DWORD[24+rsp]
|
|
psrld xmm13,31
|
|
xor eax,ebx
|
|
rol ebp,5
|
|
add edx,esi
|
|
movups xmm1,XMMWORD[((-16))+r15]
|
|
DB 102,15,56,220,208
|
|
movdqa xmm3,xmm12
|
|
and edi,eax
|
|
xor eax,ebx
|
|
psrld xmm12,30
|
|
add edx,ebp
|
|
ror ebp,7
|
|
por xmm9,xmm13
|
|
xor edi,ebx
|
|
mov esi,edx
|
|
add ecx,DWORD[28+rsp]
|
|
pslld xmm3,2
|
|
pxor xmm9,xmm12
|
|
xor ebp,eax
|
|
movdqa xmm12,XMMWORD[16+r11]
|
|
rol edx,5
|
|
add ecx,edi
|
|
and esi,ebp
|
|
pxor xmm9,xmm3
|
|
xor ebp,eax
|
|
add ecx,edx
|
|
ror edx,7
|
|
pshufd xmm10,xmm6,238
|
|
xor esi,eax
|
|
movdqa xmm3,xmm9
|
|
paddd xmm12,xmm9
|
|
mov edi,ecx
|
|
add ebx,DWORD[32+rsp]
|
|
movups xmm0,XMMWORD[r15]
|
|
DB 102,15,56,220,209
|
|
punpcklqdq xmm10,xmm7
|
|
xor edx,ebp
|
|
rol ecx,5
|
|
add ebx,esi
|
|
psrldq xmm3,4
|
|
and edi,edx
|
|
xor edx,ebp
|
|
pxor xmm10,xmm6
|
|
add ebx,ecx
|
|
ror ecx,7
|
|
pxor xmm3,xmm8
|
|
xor edi,ebp
|
|
mov esi,ebx
|
|
add eax,DWORD[36+rsp]
|
|
pxor xmm10,xmm3
|
|
xor ecx,edx
|
|
rol ebx,5
|
|
movdqa XMMWORD[16+rsp],xmm12
|
|
add eax,edi
|
|
and esi,ecx
|
|
movdqa xmm13,xmm10
|
|
xor ecx,edx
|
|
add eax,ebx
|
|
ror ebx,7
|
|
movups xmm1,XMMWORD[16+r15]
|
|
DB 102,15,56,220,208
|
|
movdqa xmm3,xmm10
|
|
xor esi,edx
|
|
pslldq xmm13,12
|
|
paddd xmm10,xmm10
|
|
mov edi,eax
|
|
add ebp,DWORD[40+rsp]
|
|
psrld xmm3,31
|
|
xor ebx,ecx
|
|
rol eax,5
|
|
add ebp,esi
|
|
movdqa xmm12,xmm13
|
|
and edi,ebx
|
|
xor ebx,ecx
|
|
psrld xmm13,30
|
|
add ebp,eax
|
|
ror eax,7
|
|
por xmm10,xmm3
|
|
xor edi,ecx
|
|
mov esi,ebp
|
|
add edx,DWORD[44+rsp]
|
|
pslld xmm12,2
|
|
pxor xmm10,xmm13
|
|
xor eax,ebx
|
|
movdqa xmm13,XMMWORD[16+r11]
|
|
rol ebp,5
|
|
add edx,edi
|
|
movups xmm0,XMMWORD[32+r15]
|
|
DB 102,15,56,220,209
|
|
and esi,eax
|
|
pxor xmm10,xmm12
|
|
xor eax,ebx
|
|
add edx,ebp
|
|
ror ebp,7
|
|
pshufd xmm11,xmm7,238
|
|
xor esi,ebx
|
|
movdqa xmm12,xmm10
|
|
paddd xmm13,xmm10
|
|
mov edi,edx
|
|
add ecx,DWORD[48+rsp]
|
|
punpcklqdq xmm11,xmm8
|
|
xor ebp,eax
|
|
rol edx,5
|
|
add ecx,esi
|
|
psrldq xmm12,4
|
|
and edi,ebp
|
|
xor ebp,eax
|
|
pxor xmm11,xmm7
|
|
add ecx,edx
|
|
ror edx,7
|
|
pxor xmm12,xmm9
|
|
xor edi,eax
|
|
mov esi,ecx
|
|
add ebx,DWORD[52+rsp]
|
|
movups xmm1,XMMWORD[48+r15]
|
|
DB 102,15,56,220,208
|
|
pxor xmm11,xmm12
|
|
xor edx,ebp
|
|
rol ecx,5
|
|
movdqa XMMWORD[32+rsp],xmm13
|
|
add ebx,edi
|
|
and esi,edx
|
|
movdqa xmm3,xmm11
|
|
xor edx,ebp
|
|
add ebx,ecx
|
|
ror ecx,7
|
|
movdqa xmm12,xmm11
|
|
xor esi,ebp
|
|
pslldq xmm3,12
|
|
paddd xmm11,xmm11
|
|
mov edi,ebx
|
|
add eax,DWORD[56+rsp]
|
|
psrld xmm12,31
|
|
xor ecx,edx
|
|
rol ebx,5
|
|
add eax,esi
|
|
movdqa xmm13,xmm3
|
|
and edi,ecx
|
|
xor ecx,edx
|
|
psrld xmm3,30
|
|
add eax,ebx
|
|
ror ebx,7
|
|
cmp r8d,11
|
|
jb NEAR $L$aesenclast1
|
|
movups xmm0,XMMWORD[64+r15]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[80+r15]
|
|
DB 102,15,56,220,208
|
|
je NEAR $L$aesenclast1
|
|
movups xmm0,XMMWORD[96+r15]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[112+r15]
|
|
DB 102,15,56,220,208
|
|
$L$aesenclast1:
|
|
DB 102,15,56,221,209
|
|
movups xmm0,XMMWORD[((16-112))+r15]
|
|
por xmm11,xmm12
|
|
xor edi,edx
|
|
mov esi,eax
|
|
add ebp,DWORD[60+rsp]
|
|
pslld xmm13,2
|
|
pxor xmm11,xmm3
|
|
xor ebx,ecx
|
|
movdqa xmm3,XMMWORD[16+r11]
|
|
rol eax,5
|
|
add ebp,edi
|
|
and esi,ebx
|
|
pxor xmm11,xmm13
|
|
pshufd xmm13,xmm10,238
|
|
xor ebx,ecx
|
|
add ebp,eax
|
|
ror eax,7
|
|
pxor xmm4,xmm8
|
|
xor esi,ecx
|
|
mov edi,ebp
|
|
add edx,DWORD[rsp]
|
|
punpcklqdq xmm13,xmm11
|
|
xor eax,ebx
|
|
rol ebp,5
|
|
pxor xmm4,xmm5
|
|
add edx,esi
|
|
movups xmm14,XMMWORD[16+r12]
|
|
xorps xmm14,xmm15
|
|
movups XMMWORD[r13*1+r12],xmm2
|
|
xorps xmm2,xmm14
|
|
movups xmm1,XMMWORD[((-80))+r15]
|
|
DB 102,15,56,220,208
|
|
and edi,eax
|
|
movdqa xmm12,xmm3
|
|
xor eax,ebx
|
|
paddd xmm3,xmm11
|
|
add edx,ebp
|
|
pxor xmm4,xmm13
|
|
ror ebp,7
|
|
xor edi,ebx
|
|
mov esi,edx
|
|
add ecx,DWORD[4+rsp]
|
|
movdqa xmm13,xmm4
|
|
xor ebp,eax
|
|
rol edx,5
|
|
movdqa XMMWORD[48+rsp],xmm3
|
|
add ecx,edi
|
|
and esi,ebp
|
|
xor ebp,eax
|
|
pslld xmm4,2
|
|
add ecx,edx
|
|
ror edx,7
|
|
psrld xmm13,30
|
|
xor esi,eax
|
|
mov edi,ecx
|
|
add ebx,DWORD[8+rsp]
|
|
movups xmm0,XMMWORD[((-64))+r15]
|
|
DB 102,15,56,220,209
|
|
por xmm4,xmm13
|
|
xor edx,ebp
|
|
rol ecx,5
|
|
pshufd xmm3,xmm11,238
|
|
add ebx,esi
|
|
and edi,edx
|
|
xor edx,ebp
|
|
add ebx,ecx
|
|
add eax,DWORD[12+rsp]
|
|
xor edi,ebp
|
|
mov esi,ebx
|
|
rol ebx,5
|
|
add eax,edi
|
|
xor esi,edx
|
|
ror ecx,7
|
|
add eax,ebx
|
|
pxor xmm5,xmm9
|
|
add ebp,DWORD[16+rsp]
|
|
movups xmm1,XMMWORD[((-48))+r15]
|
|
DB 102,15,56,220,208
|
|
xor esi,ecx
|
|
punpcklqdq xmm3,xmm4
|
|
mov edi,eax
|
|
rol eax,5
|
|
pxor xmm5,xmm6
|
|
add ebp,esi
|
|
xor edi,ecx
|
|
movdqa xmm13,xmm12
|
|
ror ebx,7
|
|
paddd xmm12,xmm4
|
|
add ebp,eax
|
|
pxor xmm5,xmm3
|
|
add edx,DWORD[20+rsp]
|
|
xor edi,ebx
|
|
mov esi,ebp
|
|
rol ebp,5
|
|
movdqa xmm3,xmm5
|
|
add edx,edi
|
|
xor esi,ebx
|
|
movdqa XMMWORD[rsp],xmm12
|
|
ror eax,7
|
|
add edx,ebp
|
|
add ecx,DWORD[24+rsp]
|
|
pslld xmm5,2
|
|
xor esi,eax
|
|
mov edi,edx
|
|
psrld xmm3,30
|
|
rol edx,5
|
|
add ecx,esi
|
|
movups xmm0,XMMWORD[((-32))+r15]
|
|
DB 102,15,56,220,209
|
|
xor edi,eax
|
|
ror ebp,7
|
|
por xmm5,xmm3
|
|
add ecx,edx
|
|
add ebx,DWORD[28+rsp]
|
|
pshufd xmm12,xmm4,238
|
|
xor edi,ebp
|
|
mov esi,ecx
|
|
rol ecx,5
|
|
add ebx,edi
|
|
xor esi,ebp
|
|
ror edx,7
|
|
add ebx,ecx
|
|
pxor xmm6,xmm10
|
|
add eax,DWORD[32+rsp]
|
|
xor esi,edx
|
|
punpcklqdq xmm12,xmm5
|
|
mov edi,ebx
|
|
rol ebx,5
|
|
pxor xmm6,xmm7
|
|
add eax,esi
|
|
xor edi,edx
|
|
movdqa xmm3,XMMWORD[32+r11]
|
|
ror ecx,7
|
|
paddd xmm13,xmm5
|
|
add eax,ebx
|
|
pxor xmm6,xmm12
|
|
add ebp,DWORD[36+rsp]
|
|
movups xmm1,XMMWORD[((-16))+r15]
|
|
DB 102,15,56,220,208
|
|
xor edi,ecx
|
|
mov esi,eax
|
|
rol eax,5
|
|
movdqa xmm12,xmm6
|
|
add ebp,edi
|
|
xor esi,ecx
|
|
movdqa XMMWORD[16+rsp],xmm13
|
|
ror ebx,7
|
|
add ebp,eax
|
|
add edx,DWORD[40+rsp]
|
|
pslld xmm6,2
|
|
xor esi,ebx
|
|
mov edi,ebp
|
|
psrld xmm12,30
|
|
rol ebp,5
|
|
add edx,esi
|
|
xor edi,ebx
|
|
ror eax,7
|
|
por xmm6,xmm12
|
|
add edx,ebp
|
|
add ecx,DWORD[44+rsp]
|
|
pshufd xmm13,xmm5,238
|
|
xor edi,eax
|
|
mov esi,edx
|
|
rol edx,5
|
|
add ecx,edi
|
|
movups xmm0,XMMWORD[r15]
|
|
DB 102,15,56,220,209
|
|
xor esi,eax
|
|
ror ebp,7
|
|
add ecx,edx
|
|
pxor xmm7,xmm11
|
|
add ebx,DWORD[48+rsp]
|
|
xor esi,ebp
|
|
punpcklqdq xmm13,xmm6
|
|
mov edi,ecx
|
|
rol ecx,5
|
|
pxor xmm7,xmm8
|
|
add ebx,esi
|
|
xor edi,ebp
|
|
movdqa xmm12,xmm3
|
|
ror edx,7
|
|
paddd xmm3,xmm6
|
|
add ebx,ecx
|
|
pxor xmm7,xmm13
|
|
add eax,DWORD[52+rsp]
|
|
xor edi,edx
|
|
mov esi,ebx
|
|
rol ebx,5
|
|
movdqa xmm13,xmm7
|
|
add eax,edi
|
|
xor esi,edx
|
|
movdqa XMMWORD[32+rsp],xmm3
|
|
ror ecx,7
|
|
add eax,ebx
|
|
add ebp,DWORD[56+rsp]
|
|
movups xmm1,XMMWORD[16+r15]
|
|
DB 102,15,56,220,208
|
|
pslld xmm7,2
|
|
xor esi,ecx
|
|
mov edi,eax
|
|
psrld xmm13,30
|
|
rol eax,5
|
|
add ebp,esi
|
|
xor edi,ecx
|
|
ror ebx,7
|
|
por xmm7,xmm13
|
|
add ebp,eax
|
|
add edx,DWORD[60+rsp]
|
|
pshufd xmm3,xmm6,238
|
|
xor edi,ebx
|
|
mov esi,ebp
|
|
rol ebp,5
|
|
add edx,edi
|
|
xor esi,ebx
|
|
ror eax,7
|
|
add edx,ebp
|
|
pxor xmm8,xmm4
|
|
add ecx,DWORD[rsp]
|
|
xor esi,eax
|
|
punpcklqdq xmm3,xmm7
|
|
mov edi,edx
|
|
rol edx,5
|
|
pxor xmm8,xmm9
|
|
add ecx,esi
|
|
movups xmm0,XMMWORD[32+r15]
|
|
DB 102,15,56,220,209
|
|
xor edi,eax
|
|
movdqa xmm13,xmm12
|
|
ror ebp,7
|
|
paddd xmm12,xmm7
|
|
add ecx,edx
|
|
pxor xmm8,xmm3
|
|
add ebx,DWORD[4+rsp]
|
|
xor edi,ebp
|
|
mov esi,ecx
|
|
rol ecx,5
|
|
movdqa xmm3,xmm8
|
|
add ebx,edi
|
|
xor esi,ebp
|
|
movdqa XMMWORD[48+rsp],xmm12
|
|
ror edx,7
|
|
add ebx,ecx
|
|
add eax,DWORD[8+rsp]
|
|
pslld xmm8,2
|
|
xor esi,edx
|
|
mov edi,ebx
|
|
psrld xmm3,30
|
|
rol ebx,5
|
|
add eax,esi
|
|
xor edi,edx
|
|
ror ecx,7
|
|
por xmm8,xmm3
|
|
add eax,ebx
|
|
add ebp,DWORD[12+rsp]
|
|
movups xmm1,XMMWORD[48+r15]
|
|
DB 102,15,56,220,208
|
|
pshufd xmm12,xmm7,238
|
|
xor edi,ecx
|
|
mov esi,eax
|
|
rol eax,5
|
|
add ebp,edi
|
|
xor esi,ecx
|
|
ror ebx,7
|
|
add ebp,eax
|
|
pxor xmm9,xmm5
|
|
add edx,DWORD[16+rsp]
|
|
xor esi,ebx
|
|
punpcklqdq xmm12,xmm8
|
|
mov edi,ebp
|
|
rol ebp,5
|
|
pxor xmm9,xmm10
|
|
add edx,esi
|
|
xor edi,ebx
|
|
movdqa xmm3,xmm13
|
|
ror eax,7
|
|
paddd xmm13,xmm8
|
|
add edx,ebp
|
|
pxor xmm9,xmm12
|
|
add ecx,DWORD[20+rsp]
|
|
xor edi,eax
|
|
mov esi,edx
|
|
rol edx,5
|
|
movdqa xmm12,xmm9
|
|
add ecx,edi
|
|
cmp r8d,11
|
|
jb NEAR $L$aesenclast2
|
|
movups xmm0,XMMWORD[64+r15]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[80+r15]
|
|
DB 102,15,56,220,208
|
|
je NEAR $L$aesenclast2
|
|
movups xmm0,XMMWORD[96+r15]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[112+r15]
|
|
DB 102,15,56,220,208
|
|
$L$aesenclast2:
|
|
DB 102,15,56,221,209
|
|
movups xmm0,XMMWORD[((16-112))+r15]
|
|
xor esi,eax
|
|
movdqa XMMWORD[rsp],xmm13
|
|
ror ebp,7
|
|
add ecx,edx
|
|
add ebx,DWORD[24+rsp]
|
|
pslld xmm9,2
|
|
xor esi,ebp
|
|
mov edi,ecx
|
|
psrld xmm12,30
|
|
rol ecx,5
|
|
add ebx,esi
|
|
xor edi,ebp
|
|
ror edx,7
|
|
por xmm9,xmm12
|
|
add ebx,ecx
|
|
add eax,DWORD[28+rsp]
|
|
pshufd xmm13,xmm8,238
|
|
ror ecx,7
|
|
mov esi,ebx
|
|
xor edi,edx
|
|
rol ebx,5
|
|
add eax,edi
|
|
xor esi,ecx
|
|
xor ecx,edx
|
|
add eax,ebx
|
|
pxor xmm10,xmm6
|
|
add ebp,DWORD[32+rsp]
|
|
movups xmm14,XMMWORD[32+r12]
|
|
xorps xmm14,xmm15
|
|
movups XMMWORD[16+r12*1+r13],xmm2
|
|
xorps xmm2,xmm14
|
|
movups xmm1,XMMWORD[((-80))+r15]
|
|
DB 102,15,56,220,208
|
|
and esi,ecx
|
|
xor ecx,edx
|
|
ror ebx,7
|
|
punpcklqdq xmm13,xmm9
|
|
mov edi,eax
|
|
xor esi,ecx
|
|
pxor xmm10,xmm11
|
|
rol eax,5
|
|
add ebp,esi
|
|
movdqa xmm12,xmm3
|
|
xor edi,ebx
|
|
paddd xmm3,xmm9
|
|
xor ebx,ecx
|
|
pxor xmm10,xmm13
|
|
add ebp,eax
|
|
add edx,DWORD[36+rsp]
|
|
and edi,ebx
|
|
xor ebx,ecx
|
|
ror eax,7
|
|
movdqa xmm13,xmm10
|
|
mov esi,ebp
|
|
xor edi,ebx
|
|
movdqa XMMWORD[16+rsp],xmm3
|
|
rol ebp,5
|
|
add edx,edi
|
|
movups xmm0,XMMWORD[((-64))+r15]
|
|
DB 102,15,56,220,209
|
|
xor esi,eax
|
|
pslld xmm10,2
|
|
xor eax,ebx
|
|
add edx,ebp
|
|
psrld xmm13,30
|
|
add ecx,DWORD[40+rsp]
|
|
and esi,eax
|
|
xor eax,ebx
|
|
por xmm10,xmm13
|
|
ror ebp,7
|
|
mov edi,edx
|
|
xor esi,eax
|
|
rol edx,5
|
|
pshufd xmm3,xmm9,238
|
|
add ecx,esi
|
|
xor edi,ebp
|
|
xor ebp,eax
|
|
add ecx,edx
|
|
add ebx,DWORD[44+rsp]
|
|
and edi,ebp
|
|
xor ebp,eax
|
|
ror edx,7
|
|
movups xmm1,XMMWORD[((-48))+r15]
|
|
DB 102,15,56,220,208
|
|
mov esi,ecx
|
|
xor edi,ebp
|
|
rol ecx,5
|
|
add ebx,edi
|
|
xor esi,edx
|
|
xor edx,ebp
|
|
add ebx,ecx
|
|
pxor xmm11,xmm7
|
|
add eax,DWORD[48+rsp]
|
|
and esi,edx
|
|
xor edx,ebp
|
|
ror ecx,7
|
|
punpcklqdq xmm3,xmm10
|
|
mov edi,ebx
|
|
xor esi,edx
|
|
pxor xmm11,xmm4
|
|
rol ebx,5
|
|
add eax,esi
|
|
movdqa xmm13,XMMWORD[48+r11]
|
|
xor edi,ecx
|
|
paddd xmm12,xmm10
|
|
xor ecx,edx
|
|
pxor xmm11,xmm3
|
|
add eax,ebx
|
|
add ebp,DWORD[52+rsp]
|
|
movups xmm0,XMMWORD[((-32))+r15]
|
|
DB 102,15,56,220,209
|
|
and edi,ecx
|
|
xor ecx,edx
|
|
ror ebx,7
|
|
movdqa xmm3,xmm11
|
|
mov esi,eax
|
|
xor edi,ecx
|
|
movdqa XMMWORD[32+rsp],xmm12
|
|
rol eax,5
|
|
add ebp,edi
|
|
xor esi,ebx
|
|
pslld xmm11,2
|
|
xor ebx,ecx
|
|
add ebp,eax
|
|
psrld xmm3,30
|
|
add edx,DWORD[56+rsp]
|
|
and esi,ebx
|
|
xor ebx,ecx
|
|
por xmm11,xmm3
|
|
ror eax,7
|
|
mov edi,ebp
|
|
xor esi,ebx
|
|
rol ebp,5
|
|
pshufd xmm12,xmm10,238
|
|
add edx,esi
|
|
movups xmm1,XMMWORD[((-16))+r15]
|
|
DB 102,15,56,220,208
|
|
xor edi,eax
|
|
xor eax,ebx
|
|
add edx,ebp
|
|
add ecx,DWORD[60+rsp]
|
|
and edi,eax
|
|
xor eax,ebx
|
|
ror ebp,7
|
|
mov esi,edx
|
|
xor edi,eax
|
|
rol edx,5
|
|
add ecx,edi
|
|
xor esi,ebp
|
|
xor ebp,eax
|
|
add ecx,edx
|
|
pxor xmm4,xmm8
|
|
add ebx,DWORD[rsp]
|
|
and esi,ebp
|
|
xor ebp,eax
|
|
ror edx,7
|
|
movups xmm0,XMMWORD[r15]
|
|
DB 102,15,56,220,209
|
|
punpcklqdq xmm12,xmm11
|
|
mov edi,ecx
|
|
xor esi,ebp
|
|
pxor xmm4,xmm5
|
|
rol ecx,5
|
|
add ebx,esi
|
|
movdqa xmm3,xmm13
|
|
xor edi,edx
|
|
paddd xmm13,xmm11
|
|
xor edx,ebp
|
|
pxor xmm4,xmm12
|
|
add ebx,ecx
|
|
add eax,DWORD[4+rsp]
|
|
and edi,edx
|
|
xor edx,ebp
|
|
ror ecx,7
|
|
movdqa xmm12,xmm4
|
|
mov esi,ebx
|
|
xor edi,edx
|
|
movdqa XMMWORD[48+rsp],xmm13
|
|
rol ebx,5
|
|
add eax,edi
|
|
xor esi,ecx
|
|
pslld xmm4,2
|
|
xor ecx,edx
|
|
add eax,ebx
|
|
psrld xmm12,30
|
|
add ebp,DWORD[8+rsp]
|
|
movups xmm1,XMMWORD[16+r15]
|
|
DB 102,15,56,220,208
|
|
and esi,ecx
|
|
xor ecx,edx
|
|
por xmm4,xmm12
|
|
ror ebx,7
|
|
mov edi,eax
|
|
xor esi,ecx
|
|
rol eax,5
|
|
pshufd xmm13,xmm11,238
|
|
add ebp,esi
|
|
xor edi,ebx
|
|
xor ebx,ecx
|
|
add ebp,eax
|
|
add edx,DWORD[12+rsp]
|
|
and edi,ebx
|
|
xor ebx,ecx
|
|
ror eax,7
|
|
mov esi,ebp
|
|
xor edi,ebx
|
|
rol ebp,5
|
|
add edx,edi
|
|
movups xmm0,XMMWORD[32+r15]
|
|
DB 102,15,56,220,209
|
|
xor esi,eax
|
|
xor eax,ebx
|
|
add edx,ebp
|
|
pxor xmm5,xmm9
|
|
add ecx,DWORD[16+rsp]
|
|
and esi,eax
|
|
xor eax,ebx
|
|
ror ebp,7
|
|
punpcklqdq xmm13,xmm4
|
|
mov edi,edx
|
|
xor esi,eax
|
|
pxor xmm5,xmm6
|
|
rol edx,5
|
|
add ecx,esi
|
|
movdqa xmm12,xmm3
|
|
xor edi,ebp
|
|
paddd xmm3,xmm4
|
|
xor ebp,eax
|
|
pxor xmm5,xmm13
|
|
add ecx,edx
|
|
add ebx,DWORD[20+rsp]
|
|
and edi,ebp
|
|
xor ebp,eax
|
|
ror edx,7
|
|
movups xmm1,XMMWORD[48+r15]
|
|
DB 102,15,56,220,208
|
|
movdqa xmm13,xmm5
|
|
mov esi,ecx
|
|
xor edi,ebp
|
|
movdqa XMMWORD[rsp],xmm3
|
|
rol ecx,5
|
|
add ebx,edi
|
|
xor esi,edx
|
|
pslld xmm5,2
|
|
xor edx,ebp
|
|
add ebx,ecx
|
|
psrld xmm13,30
|
|
add eax,DWORD[24+rsp]
|
|
and esi,edx
|
|
xor edx,ebp
|
|
por xmm5,xmm13
|
|
ror ecx,7
|
|
mov edi,ebx
|
|
xor esi,edx
|
|
rol ebx,5
|
|
pshufd xmm3,xmm4,238
|
|
add eax,esi
|
|
xor edi,ecx
|
|
xor ecx,edx
|
|
add eax,ebx
|
|
add ebp,DWORD[28+rsp]
|
|
cmp r8d,11
|
|
jb NEAR $L$aesenclast3
|
|
movups xmm0,XMMWORD[64+r15]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[80+r15]
|
|
DB 102,15,56,220,208
|
|
je NEAR $L$aesenclast3
|
|
movups xmm0,XMMWORD[96+r15]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[112+r15]
|
|
DB 102,15,56,220,208
|
|
$L$aesenclast3:
|
|
DB 102,15,56,221,209
|
|
movups xmm0,XMMWORD[((16-112))+r15]
|
|
and edi,ecx
|
|
xor ecx,edx
|
|
ror ebx,7
|
|
mov esi,eax
|
|
xor edi,ecx
|
|
rol eax,5
|
|
add ebp,edi
|
|
xor esi,ebx
|
|
xor ebx,ecx
|
|
add ebp,eax
|
|
pxor xmm6,xmm10
|
|
add edx,DWORD[32+rsp]
|
|
and esi,ebx
|
|
xor ebx,ecx
|
|
ror eax,7
|
|
punpcklqdq xmm3,xmm5
|
|
mov edi,ebp
|
|
xor esi,ebx
|
|
pxor xmm6,xmm7
|
|
rol ebp,5
|
|
add edx,esi
|
|
movups xmm14,XMMWORD[48+r12]
|
|
xorps xmm14,xmm15
|
|
movups XMMWORD[32+r12*1+r13],xmm2
|
|
xorps xmm2,xmm14
|
|
movups xmm1,XMMWORD[((-80))+r15]
|
|
DB 102,15,56,220,208
|
|
movdqa xmm13,xmm12
|
|
xor edi,eax
|
|
paddd xmm12,xmm5
|
|
xor eax,ebx
|
|
pxor xmm6,xmm3
|
|
add edx,ebp
|
|
add ecx,DWORD[36+rsp]
|
|
and edi,eax
|
|
xor eax,ebx
|
|
ror ebp,7
|
|
movdqa xmm3,xmm6
|
|
mov esi,edx
|
|
xor edi,eax
|
|
movdqa XMMWORD[16+rsp],xmm12
|
|
rol edx,5
|
|
add ecx,edi
|
|
xor esi,ebp
|
|
pslld xmm6,2
|
|
xor ebp,eax
|
|
add ecx,edx
|
|
psrld xmm3,30
|
|
add ebx,DWORD[40+rsp]
|
|
and esi,ebp
|
|
xor ebp,eax
|
|
por xmm6,xmm3
|
|
ror edx,7
|
|
movups xmm0,XMMWORD[((-64))+r15]
|
|
DB 102,15,56,220,209
|
|
mov edi,ecx
|
|
xor esi,ebp
|
|
rol ecx,5
|
|
pshufd xmm12,xmm5,238
|
|
add ebx,esi
|
|
xor edi,edx
|
|
xor edx,ebp
|
|
add ebx,ecx
|
|
add eax,DWORD[44+rsp]
|
|
and edi,edx
|
|
xor edx,ebp
|
|
ror ecx,7
|
|
mov esi,ebx
|
|
xor edi,edx
|
|
rol ebx,5
|
|
add eax,edi
|
|
xor esi,edx
|
|
add eax,ebx
|
|
pxor xmm7,xmm11
|
|
add ebp,DWORD[48+rsp]
|
|
movups xmm1,XMMWORD[((-48))+r15]
|
|
DB 102,15,56,220,208
|
|
xor esi,ecx
|
|
punpcklqdq xmm12,xmm6
|
|
mov edi,eax
|
|
rol eax,5
|
|
pxor xmm7,xmm8
|
|
add ebp,esi
|
|
xor edi,ecx
|
|
movdqa xmm3,xmm13
|
|
ror ebx,7
|
|
paddd xmm13,xmm6
|
|
add ebp,eax
|
|
pxor xmm7,xmm12
|
|
add edx,DWORD[52+rsp]
|
|
xor edi,ebx
|
|
mov esi,ebp
|
|
rol ebp,5
|
|
movdqa xmm12,xmm7
|
|
add edx,edi
|
|
xor esi,ebx
|
|
movdqa XMMWORD[32+rsp],xmm13
|
|
ror eax,7
|
|
add edx,ebp
|
|
add ecx,DWORD[56+rsp]
|
|
pslld xmm7,2
|
|
xor esi,eax
|
|
mov edi,edx
|
|
psrld xmm12,30
|
|
rol edx,5
|
|
add ecx,esi
|
|
movups xmm0,XMMWORD[((-32))+r15]
|
|
DB 102,15,56,220,209
|
|
xor edi,eax
|
|
ror ebp,7
|
|
por xmm7,xmm12
|
|
add ecx,edx
|
|
add ebx,DWORD[60+rsp]
|
|
xor edi,ebp
|
|
mov esi,ecx
|
|
rol ecx,5
|
|
add ebx,edi
|
|
xor esi,ebp
|
|
ror edx,7
|
|
add ebx,ecx
|
|
add eax,DWORD[rsp]
|
|
xor esi,edx
|
|
mov edi,ebx
|
|
rol ebx,5
|
|
paddd xmm3,xmm7
|
|
add eax,esi
|
|
xor edi,edx
|
|
movdqa XMMWORD[48+rsp],xmm3
|
|
ror ecx,7
|
|
add eax,ebx
|
|
add ebp,DWORD[4+rsp]
|
|
movups xmm1,XMMWORD[((-16))+r15]
|
|
DB 102,15,56,220,208
|
|
xor edi,ecx
|
|
mov esi,eax
|
|
rol eax,5
|
|
add ebp,edi
|
|
xor esi,ecx
|
|
ror ebx,7
|
|
add ebp,eax
|
|
add edx,DWORD[8+rsp]
|
|
xor esi,ebx
|
|
mov edi,ebp
|
|
rol ebp,5
|
|
add edx,esi
|
|
xor edi,ebx
|
|
ror eax,7
|
|
add edx,ebp
|
|
add ecx,DWORD[12+rsp]
|
|
xor edi,eax
|
|
mov esi,edx
|
|
rol edx,5
|
|
add ecx,edi
|
|
movups xmm0,XMMWORD[r15]
|
|
DB 102,15,56,220,209
|
|
xor esi,eax
|
|
ror ebp,7
|
|
add ecx,edx
|
|
cmp r10,r14
|
|
je NEAR $L$done_ssse3
|
|
movdqa xmm3,XMMWORD[64+r11]
|
|
movdqa xmm13,XMMWORD[r11]
|
|
movdqu xmm4,XMMWORD[r10]
|
|
movdqu xmm5,XMMWORD[16+r10]
|
|
movdqu xmm6,XMMWORD[32+r10]
|
|
movdqu xmm7,XMMWORD[48+r10]
|
|
DB 102,15,56,0,227
|
|
add r10,64
|
|
add ebx,DWORD[16+rsp]
|
|
xor esi,ebp
|
|
mov edi,ecx
|
|
DB 102,15,56,0,235
|
|
rol ecx,5
|
|
add ebx,esi
|
|
xor edi,ebp
|
|
ror edx,7
|
|
paddd xmm4,xmm13
|
|
add ebx,ecx
|
|
add eax,DWORD[20+rsp]
|
|
xor edi,edx
|
|
mov esi,ebx
|
|
movdqa XMMWORD[rsp],xmm4
|
|
rol ebx,5
|
|
add eax,edi
|
|
xor esi,edx
|
|
ror ecx,7
|
|
psubd xmm4,xmm13
|
|
add eax,ebx
|
|
add ebp,DWORD[24+rsp]
|
|
movups xmm1,XMMWORD[16+r15]
|
|
DB 102,15,56,220,208
|
|
xor esi,ecx
|
|
mov edi,eax
|
|
rol eax,5
|
|
add ebp,esi
|
|
xor edi,ecx
|
|
ror ebx,7
|
|
add ebp,eax
|
|
add edx,DWORD[28+rsp]
|
|
xor edi,ebx
|
|
mov esi,ebp
|
|
rol ebp,5
|
|
add edx,edi
|
|
xor esi,ebx
|
|
ror eax,7
|
|
add edx,ebp
|
|
add ecx,DWORD[32+rsp]
|
|
xor esi,eax
|
|
mov edi,edx
|
|
DB 102,15,56,0,243
|
|
rol edx,5
|
|
add ecx,esi
|
|
movups xmm0,XMMWORD[32+r15]
|
|
DB 102,15,56,220,209
|
|
xor edi,eax
|
|
ror ebp,7
|
|
paddd xmm5,xmm13
|
|
add ecx,edx
|
|
add ebx,DWORD[36+rsp]
|
|
xor edi,ebp
|
|
mov esi,ecx
|
|
movdqa XMMWORD[16+rsp],xmm5
|
|
rol ecx,5
|
|
add ebx,edi
|
|
xor esi,ebp
|
|
ror edx,7
|
|
psubd xmm5,xmm13
|
|
add ebx,ecx
|
|
add eax,DWORD[40+rsp]
|
|
xor esi,edx
|
|
mov edi,ebx
|
|
rol ebx,5
|
|
add eax,esi
|
|
xor edi,edx
|
|
ror ecx,7
|
|
add eax,ebx
|
|
add ebp,DWORD[44+rsp]
|
|
movups xmm1,XMMWORD[48+r15]
|
|
DB 102,15,56,220,208
|
|
xor edi,ecx
|
|
mov esi,eax
|
|
rol eax,5
|
|
add ebp,edi
|
|
xor esi,ecx
|
|
ror ebx,7
|
|
add ebp,eax
|
|
add edx,DWORD[48+rsp]
|
|
xor esi,ebx
|
|
mov edi,ebp
|
|
DB 102,15,56,0,251
|
|
rol ebp,5
|
|
add edx,esi
|
|
xor edi,ebx
|
|
ror eax,7
|
|
paddd xmm6,xmm13
|
|
add edx,ebp
|
|
add ecx,DWORD[52+rsp]
|
|
xor edi,eax
|
|
mov esi,edx
|
|
movdqa XMMWORD[32+rsp],xmm6
|
|
rol edx,5
|
|
add ecx,edi
|
|
cmp r8d,11
|
|
jb NEAR $L$aesenclast4
|
|
movups xmm0,XMMWORD[64+r15]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[80+r15]
|
|
DB 102,15,56,220,208
|
|
je NEAR $L$aesenclast4
|
|
movups xmm0,XMMWORD[96+r15]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[112+r15]
|
|
DB 102,15,56,220,208
|
|
$L$aesenclast4:
|
|
DB 102,15,56,221,209
|
|
movups xmm0,XMMWORD[((16-112))+r15]
|
|
xor esi,eax
|
|
ror ebp,7
|
|
psubd xmm6,xmm13
|
|
add ecx,edx
|
|
add ebx,DWORD[56+rsp]
|
|
xor esi,ebp
|
|
mov edi,ecx
|
|
rol ecx,5
|
|
add ebx,esi
|
|
xor edi,ebp
|
|
ror edx,7
|
|
add ebx,ecx
|
|
add eax,DWORD[60+rsp]
|
|
xor edi,edx
|
|
mov esi,ebx
|
|
rol ebx,5
|
|
add eax,edi
|
|
ror ecx,7
|
|
add eax,ebx
|
|
movups XMMWORD[48+r12*1+r13],xmm2
|
|
lea r12,[64+r12]
|
|
|
|
add eax,DWORD[r9]
|
|
add esi,DWORD[4+r9]
|
|
add ecx,DWORD[8+r9]
|
|
add edx,DWORD[12+r9]
|
|
mov DWORD[r9],eax
|
|
add ebp,DWORD[16+r9]
|
|
mov DWORD[4+r9],esi
|
|
mov ebx,esi
|
|
mov DWORD[8+r9],ecx
|
|
mov edi,ecx
|
|
mov DWORD[12+r9],edx
|
|
xor edi,edx
|
|
mov DWORD[16+r9],ebp
|
|
and esi,edi
|
|
jmp NEAR $L$oop_ssse3
|
|
|
|
$L$done_ssse3:
|
|
add ebx,DWORD[16+rsp]
|
|
xor esi,ebp
|
|
mov edi,ecx
|
|
rol ecx,5
|
|
add ebx,esi
|
|
xor edi,ebp
|
|
ror edx,7
|
|
add ebx,ecx
|
|
add eax,DWORD[20+rsp]
|
|
xor edi,edx
|
|
mov esi,ebx
|
|
rol ebx,5
|
|
add eax,edi
|
|
xor esi,edx
|
|
ror ecx,7
|
|
add eax,ebx
|
|
add ebp,DWORD[24+rsp]
|
|
movups xmm1,XMMWORD[16+r15]
|
|
DB 102,15,56,220,208
|
|
xor esi,ecx
|
|
mov edi,eax
|
|
rol eax,5
|
|
add ebp,esi
|
|
xor edi,ecx
|
|
ror ebx,7
|
|
add ebp,eax
|
|
add edx,DWORD[28+rsp]
|
|
xor edi,ebx
|
|
mov esi,ebp
|
|
rol ebp,5
|
|
add edx,edi
|
|
xor esi,ebx
|
|
ror eax,7
|
|
add edx,ebp
|
|
add ecx,DWORD[32+rsp]
|
|
xor esi,eax
|
|
mov edi,edx
|
|
rol edx,5
|
|
add ecx,esi
|
|
movups xmm0,XMMWORD[32+r15]
|
|
DB 102,15,56,220,209
|
|
xor edi,eax
|
|
ror ebp,7
|
|
add ecx,edx
|
|
add ebx,DWORD[36+rsp]
|
|
xor edi,ebp
|
|
mov esi,ecx
|
|
rol ecx,5
|
|
add ebx,edi
|
|
xor esi,ebp
|
|
ror edx,7
|
|
add ebx,ecx
|
|
add eax,DWORD[40+rsp]
|
|
xor esi,edx
|
|
mov edi,ebx
|
|
rol ebx,5
|
|
add eax,esi
|
|
xor edi,edx
|
|
ror ecx,7
|
|
add eax,ebx
|
|
add ebp,DWORD[44+rsp]
|
|
movups xmm1,XMMWORD[48+r15]
|
|
DB 102,15,56,220,208
|
|
xor edi,ecx
|
|
mov esi,eax
|
|
rol eax,5
|
|
add ebp,edi
|
|
xor esi,ecx
|
|
ror ebx,7
|
|
add ebp,eax
|
|
add edx,DWORD[48+rsp]
|
|
xor esi,ebx
|
|
mov edi,ebp
|
|
rol ebp,5
|
|
add edx,esi
|
|
xor edi,ebx
|
|
ror eax,7
|
|
add edx,ebp
|
|
add ecx,DWORD[52+rsp]
|
|
xor edi,eax
|
|
mov esi,edx
|
|
rol edx,5
|
|
add ecx,edi
|
|
cmp r8d,11
|
|
jb NEAR $L$aesenclast5
|
|
movups xmm0,XMMWORD[64+r15]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[80+r15]
|
|
DB 102,15,56,220,208
|
|
je NEAR $L$aesenclast5
|
|
movups xmm0,XMMWORD[96+r15]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[112+r15]
|
|
DB 102,15,56,220,208
|
|
$L$aesenclast5:
|
|
DB 102,15,56,221,209
|
|
movups xmm0,XMMWORD[((16-112))+r15]
|
|
xor esi,eax
|
|
ror ebp,7
|
|
add ecx,edx
|
|
add ebx,DWORD[56+rsp]
|
|
xor esi,ebp
|
|
mov edi,ecx
|
|
rol ecx,5
|
|
add ebx,esi
|
|
xor edi,ebp
|
|
ror edx,7
|
|
add ebx,ecx
|
|
add eax,DWORD[60+rsp]
|
|
xor edi,edx
|
|
mov esi,ebx
|
|
rol ebx,5
|
|
add eax,edi
|
|
ror ecx,7
|
|
add eax,ebx
|
|
movups XMMWORD[48+r12*1+r13],xmm2
|
|
mov r8,QWORD[88+rsp]
|
|
|
|
add eax,DWORD[r9]
|
|
add esi,DWORD[4+r9]
|
|
add ecx,DWORD[8+r9]
|
|
mov DWORD[r9],eax
|
|
add edx,DWORD[12+r9]
|
|
mov DWORD[4+r9],esi
|
|
add ebp,DWORD[16+r9]
|
|
mov DWORD[8+r9],ecx
|
|
mov DWORD[12+r9],edx
|
|
mov DWORD[16+r9],ebp
|
|
movups XMMWORD[r8],xmm2
|
|
movaps xmm6,XMMWORD[((96+0))+rsp]
|
|
movaps xmm7,XMMWORD[((96+16))+rsp]
|
|
movaps xmm8,XMMWORD[((96+32))+rsp]
|
|
movaps xmm9,XMMWORD[((96+48))+rsp]
|
|
movaps xmm10,XMMWORD[((96+64))+rsp]
|
|
movaps xmm11,XMMWORD[((96+80))+rsp]
|
|
movaps xmm12,XMMWORD[((96+96))+rsp]
|
|
movaps xmm13,XMMWORD[((96+112))+rsp]
|
|
movaps xmm14,XMMWORD[((96+128))+rsp]
|
|
movaps xmm15,XMMWORD[((96+144))+rsp]
|
|
lea rsi,[264+rsp]
|
|
|
|
mov r15,QWORD[rsi]
|
|
|
|
mov r14,QWORD[8+rsi]
|
|
|
|
mov r13,QWORD[16+rsi]
|
|
|
|
mov r12,QWORD[24+rsi]
|
|
|
|
mov rbp,QWORD[32+rsi]
|
|
|
|
mov rbx,QWORD[40+rsi]
|
|
|
|
lea rsp,[48+rsi]
|
|
|
|
$L$epilogue_ssse3:
|
|
mov rdi,QWORD[8+rsp] ;WIN64 epilogue
|
|
mov rsi,QWORD[16+rsp]
|
|
DB 0F3h,0C3h ;repret
|
|
|
|
$L$SEH_end_aesni_cbc_sha1_enc_ssse3:
|
|
|
|
ALIGN 32
|
|
aesni_cbc_sha1_enc_avx:
|
|
mov QWORD[8+rsp],rdi ;WIN64 prologue
|
|
mov QWORD[16+rsp],rsi
|
|
mov rax,rsp
|
|
$L$SEH_begin_aesni_cbc_sha1_enc_avx:
|
|
mov rdi,rcx
|
|
mov rsi,rdx
|
|
mov rdx,r8
|
|
mov rcx,r9
|
|
mov r8,QWORD[40+rsp]
|
|
mov r9,QWORD[48+rsp]
|
|
|
|
|
|
|
|
mov r10,QWORD[56+rsp]
|
|
|
|
|
|
push rbx
|
|
|
|
push rbp
|
|
|
|
push r12
|
|
|
|
push r13
|
|
|
|
push r14
|
|
|
|
push r15
|
|
|
|
lea rsp,[((-264))+rsp]
|
|
|
|
|
|
|
|
movaps XMMWORD[(96+0)+rsp],xmm6
|
|
movaps XMMWORD[(96+16)+rsp],xmm7
|
|
movaps XMMWORD[(96+32)+rsp],xmm8
|
|
movaps XMMWORD[(96+48)+rsp],xmm9
|
|
movaps XMMWORD[(96+64)+rsp],xmm10
|
|
movaps XMMWORD[(96+80)+rsp],xmm11
|
|
movaps XMMWORD[(96+96)+rsp],xmm12
|
|
movaps XMMWORD[(96+112)+rsp],xmm13
|
|
movaps XMMWORD[(96+128)+rsp],xmm14
|
|
movaps XMMWORD[(96+144)+rsp],xmm15
|
|
$L$prologue_avx:
|
|
vzeroall
|
|
mov r12,rdi
|
|
mov r13,rsi
|
|
mov r14,rdx
|
|
lea r15,[112+rcx]
|
|
vmovdqu xmm12,XMMWORD[r8]
|
|
mov QWORD[88+rsp],r8
|
|
shl r14,6
|
|
sub r13,r12
|
|
mov r8d,DWORD[((240-112))+r15]
|
|
add r14,r10
|
|
|
|
lea r11,[K_XX_XX]
|
|
mov eax,DWORD[r9]
|
|
mov ebx,DWORD[4+r9]
|
|
mov ecx,DWORD[8+r9]
|
|
mov edx,DWORD[12+r9]
|
|
mov esi,ebx
|
|
mov ebp,DWORD[16+r9]
|
|
mov edi,ecx
|
|
xor edi,edx
|
|
and esi,edi
|
|
|
|
vmovdqa xmm6,XMMWORD[64+r11]
|
|
vmovdqa xmm10,XMMWORD[r11]
|
|
vmovdqu xmm0,XMMWORD[r10]
|
|
vmovdqu xmm1,XMMWORD[16+r10]
|
|
vmovdqu xmm2,XMMWORD[32+r10]
|
|
vmovdqu xmm3,XMMWORD[48+r10]
|
|
vpshufb xmm0,xmm0,xmm6
|
|
add r10,64
|
|
vpshufb xmm1,xmm1,xmm6
|
|
vpshufb xmm2,xmm2,xmm6
|
|
vpshufb xmm3,xmm3,xmm6
|
|
vpaddd xmm4,xmm0,xmm10
|
|
vpaddd xmm5,xmm1,xmm10
|
|
vpaddd xmm6,xmm2,xmm10
|
|
vmovdqa XMMWORD[rsp],xmm4
|
|
vmovdqa XMMWORD[16+rsp],xmm5
|
|
vmovdqa XMMWORD[32+rsp],xmm6
|
|
vmovups xmm15,XMMWORD[((-112))+r15]
|
|
vmovups xmm14,XMMWORD[((16-112))+r15]
|
|
jmp NEAR $L$oop_avx
|
|
ALIGN 32
|
|
$L$oop_avx:
|
|
shrd ebx,ebx,2
|
|
vmovdqu xmm13,XMMWORD[r12]
|
|
vpxor xmm13,xmm13,xmm15
|
|
vpxor xmm12,xmm12,xmm13
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[((-80))+r15]
|
|
xor esi,edx
|
|
vpalignr xmm4,xmm1,xmm0,8
|
|
mov edi,eax
|
|
add ebp,DWORD[rsp]
|
|
vpaddd xmm9,xmm10,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
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[((-64))+r15]
|
|
and esi,eax
|
|
vpsrld xmm8,xmm4,31
|
|
xor eax,ebx
|
|
add edx,ebp
|
|
shrd ebp,ebp,7
|
|
xor esi,ebx
|
|
vpslldq xmm9,xmm4,12
|
|
vpaddd xmm4,xmm4,xmm4
|
|
mov edi,edx
|
|
add ecx,DWORD[8+rsp]
|
|
xor ebp,eax
|
|
shld edx,edx,5
|
|
vpor xmm4,xmm4,xmm8
|
|
vpsrld xmm8,xmm9,30
|
|
add ecx,esi
|
|
and edi,ebp
|
|
xor ebp,eax
|
|
add ecx,edx
|
|
vpslld xmm9,xmm9,2
|
|
vpxor xmm4,xmm4,xmm8
|
|
shrd edx,edx,7
|
|
xor edi,eax
|
|
mov esi,ecx
|
|
add ebx,DWORD[12+rsp]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[((-48))+r15]
|
|
vpxor xmm4,xmm4,xmm9
|
|
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,xmm10,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
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[((-32))+r15]
|
|
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 xmm9,xmm5,12
|
|
vpaddd xmm5,xmm5,xmm5
|
|
mov edi,ebp
|
|
add edx,DWORD[24+rsp]
|
|
xor eax,ebx
|
|
shld ebp,ebp,5
|
|
vpor xmm5,xmm5,xmm8
|
|
vpsrld xmm8,xmm9,30
|
|
add edx,esi
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[((-16))+r15]
|
|
and edi,eax
|
|
xor eax,ebx
|
|
add edx,ebp
|
|
vpslld xmm9,xmm9,2
|
|
vpxor xmm5,xmm5,xmm8
|
|
shrd ebp,ebp,7
|
|
xor edi,ebx
|
|
mov esi,edx
|
|
add ecx,DWORD[28+rsp]
|
|
vpxor xmm5,xmm5,xmm9
|
|
xor ebp,eax
|
|
shld edx,edx,5
|
|
vmovdqa xmm10,XMMWORD[16+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]
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[r15]
|
|
vpaddd xmm9,xmm10,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
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[16+r15]
|
|
xor esi,edx
|
|
vpslldq xmm9,xmm6,12
|
|
vpaddd xmm6,xmm6,xmm6
|
|
mov edi,eax
|
|
add ebp,DWORD[40+rsp]
|
|
xor ebx,ecx
|
|
shld eax,eax,5
|
|
vpor xmm6,xmm6,xmm8
|
|
vpsrld xmm8,xmm9,30
|
|
add ebp,esi
|
|
and edi,ebx
|
|
xor ebx,ecx
|
|
add ebp,eax
|
|
vpslld xmm9,xmm9,2
|
|
vpxor xmm6,xmm6,xmm8
|
|
shrd eax,eax,7
|
|
xor edi,ecx
|
|
mov esi,ebp
|
|
add edx,DWORD[44+rsp]
|
|
vpxor xmm6,xmm6,xmm9
|
|
xor eax,ebx
|
|
shld ebp,ebp,5
|
|
add edx,edi
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[32+r15]
|
|
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,xmm10,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]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[48+r15]
|
|
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 xmm9,xmm7,12
|
|
vpaddd xmm7,xmm7,xmm7
|
|
mov edi,ebx
|
|
add eax,DWORD[56+rsp]
|
|
xor ecx,edx
|
|
shld ebx,ebx,5
|
|
vpor xmm7,xmm7,xmm8
|
|
vpsrld xmm8,xmm9,30
|
|
add eax,esi
|
|
and edi,ecx
|
|
xor ecx,edx
|
|
add eax,ebx
|
|
vpslld xmm9,xmm9,2
|
|
vpxor xmm7,xmm7,xmm8
|
|
shrd ebx,ebx,7
|
|
cmp r8d,11
|
|
jb NEAR $L$vaesenclast6
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[64+r15]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[80+r15]
|
|
je NEAR $L$vaesenclast6
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[96+r15]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[112+r15]
|
|
$L$vaesenclast6:
|
|
vaesenclast xmm12,xmm12,xmm15
|
|
vmovups xmm15,XMMWORD[((-112))+r15]
|
|
vmovups xmm14,XMMWORD[((16-112))+r15]
|
|
xor edi,edx
|
|
mov esi,eax
|
|
add ebp,DWORD[60+rsp]
|
|
vpxor xmm7,xmm7,xmm9
|
|
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,xmm10,xmm7
|
|
add edx,esi
|
|
vmovdqu xmm13,XMMWORD[16+r12]
|
|
vpxor xmm13,xmm13,xmm15
|
|
vmovups XMMWORD[r13*1+r12],xmm12
|
|
vpxor xmm12,xmm12,xmm13
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[((-80))+r15]
|
|
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]
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[((-64))+r15]
|
|
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]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[((-48))+r15]
|
|
xor esi,ecx
|
|
mov edi,eax
|
|
shld eax,eax,5
|
|
vpxor xmm1,xmm1,xmm2
|
|
add ebp,esi
|
|
xor edi,ecx
|
|
vpaddd xmm9,xmm10,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
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[((-32))+r15]
|
|
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,xmm10,xmm1
|
|
vmovdqa xmm10,XMMWORD[32+r11]
|
|
shrd ecx,ecx,7
|
|
add eax,ebx
|
|
vpxor xmm2,xmm2,xmm8
|
|
add ebp,DWORD[36+rsp]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[((-16))+r15]
|
|
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
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[r15]
|
|
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,xmm10,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]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[16+r15]
|
|
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
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[32+r15]
|
|
xor edi,eax
|
|
vpaddd xmm9,xmm10,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]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[48+r15]
|
|
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,xmm10,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
|
|
cmp r8d,11
|
|
jb NEAR $L$vaesenclast7
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[64+r15]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[80+r15]
|
|
je NEAR $L$vaesenclast7
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[96+r15]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[112+r15]
|
|
$L$vaesenclast7:
|
|
vaesenclast xmm12,xmm12,xmm15
|
|
vmovups xmm15,XMMWORD[((-112))+r15]
|
|
vmovups xmm14,XMMWORD[((16-112))+r15]
|
|
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]
|
|
vmovdqu xmm13,XMMWORD[32+r12]
|
|
vpxor xmm13,xmm13,xmm15
|
|
vmovups XMMWORD[16+r12*1+r13],xmm12
|
|
vpxor xmm12,xmm12,xmm13
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[((-80))+r15]
|
|
and esi,ecx
|
|
xor ecx,edx
|
|
shrd ebx,ebx,7
|
|
vpxor xmm6,xmm6,xmm7
|
|
mov edi,eax
|
|
xor esi,ecx
|
|
vpaddd xmm9,xmm10,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
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[((-64))+r15]
|
|
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
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[((-48))+r15]
|
|
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,xmm10,xmm6
|
|
vmovdqa xmm10,XMMWORD[48+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]
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[((-32))+r15]
|
|
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
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[((-16))+r15]
|
|
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
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[r15]
|
|
vpxor xmm0,xmm0,xmm1
|
|
mov edi,ecx
|
|
xor esi,ebp
|
|
vpaddd xmm9,xmm10,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]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[16+r15]
|
|
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
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[32+r15]
|
|
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,xmm10,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
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[48+r15]
|
|
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]
|
|
cmp r8d,11
|
|
jb NEAR $L$vaesenclast8
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[64+r15]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[80+r15]
|
|
je NEAR $L$vaesenclast8
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[96+r15]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[112+r15]
|
|
$L$vaesenclast8:
|
|
vaesenclast xmm12,xmm12,xmm15
|
|
vmovups xmm15,XMMWORD[((-112))+r15]
|
|
vmovups xmm14,XMMWORD[((16-112))+r15]
|
|
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,xmm10,xmm1
|
|
shld ebp,ebp,5
|
|
add edx,esi
|
|
vmovdqu xmm13,XMMWORD[48+r12]
|
|
vpxor xmm13,xmm13,xmm15
|
|
vmovups XMMWORD[32+r12*1+r13],xmm12
|
|
vpxor xmm12,xmm12,xmm13
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[((-80))+r15]
|
|
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
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[((-64))+r15]
|
|
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]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[((-48))+r15]
|
|
xor esi,ecx
|
|
mov edi,eax
|
|
shld eax,eax,5
|
|
vpxor xmm3,xmm3,xmm4
|
|
add ebp,esi
|
|
xor edi,ecx
|
|
vpaddd xmm9,xmm10,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
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[((-32))+r15]
|
|
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,xmm10,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]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[((-16))+r15]
|
|
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
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[r15]
|
|
xor esi,eax
|
|
shrd ebp,ebp,7
|
|
add ecx,edx
|
|
cmp r10,r14
|
|
je NEAR $L$done_avx
|
|
vmovdqa xmm9,XMMWORD[64+r11]
|
|
vmovdqa xmm10,XMMWORD[r11]
|
|
vmovdqu xmm0,XMMWORD[r10]
|
|
vmovdqu xmm1,XMMWORD[16+r10]
|
|
vmovdqu xmm2,XMMWORD[32+r10]
|
|
vmovdqu xmm3,XMMWORD[48+r10]
|
|
vpshufb xmm0,xmm0,xmm9
|
|
add r10,64
|
|
add ebx,DWORD[16+rsp]
|
|
xor esi,ebp
|
|
vpshufb xmm1,xmm1,xmm9
|
|
mov edi,ecx
|
|
shld ecx,ecx,5
|
|
vpaddd xmm8,xmm0,xmm10
|
|
add ebx,esi
|
|
xor edi,ebp
|
|
shrd edx,edx,7
|
|
add ebx,ecx
|
|
vmovdqa XMMWORD[rsp],xmm8
|
|
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]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[16+r15]
|
|
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,xmm9
|
|
mov edi,edx
|
|
shld edx,edx,5
|
|
vpaddd xmm8,xmm1,xmm10
|
|
add ecx,esi
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[32+r15]
|
|
xor edi,eax
|
|
shrd ebp,ebp,7
|
|
add ecx,edx
|
|
vmovdqa XMMWORD[16+rsp],xmm8
|
|
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]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[48+r15]
|
|
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,xmm9
|
|
mov edi,ebp
|
|
shld ebp,ebp,5
|
|
vpaddd xmm8,xmm2,xmm10
|
|
add edx,esi
|
|
xor edi,ebx
|
|
shrd eax,eax,7
|
|
add edx,ebp
|
|
vmovdqa XMMWORD[32+rsp],xmm8
|
|
add ecx,DWORD[52+rsp]
|
|
xor edi,eax
|
|
mov esi,edx
|
|
shld edx,edx,5
|
|
add ecx,edi
|
|
cmp r8d,11
|
|
jb NEAR $L$vaesenclast9
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[64+r15]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[80+r15]
|
|
je NEAR $L$vaesenclast9
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[96+r15]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[112+r15]
|
|
$L$vaesenclast9:
|
|
vaesenclast xmm12,xmm12,xmm15
|
|
vmovups xmm15,XMMWORD[((-112))+r15]
|
|
vmovups xmm14,XMMWORD[((16-112))+r15]
|
|
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
|
|
vmovups XMMWORD[48+r12*1+r13],xmm12
|
|
lea r12,[64+r12]
|
|
|
|
add eax,DWORD[r9]
|
|
add esi,DWORD[4+r9]
|
|
add ecx,DWORD[8+r9]
|
|
add edx,DWORD[12+r9]
|
|
mov DWORD[r9],eax
|
|
add ebp,DWORD[16+r9]
|
|
mov DWORD[4+r9],esi
|
|
mov ebx,esi
|
|
mov DWORD[8+r9],ecx
|
|
mov edi,ecx
|
|
mov DWORD[12+r9],edx
|
|
xor edi,edx
|
|
mov DWORD[16+r9],ebp
|
|
and esi,edi
|
|
jmp NEAR $L$oop_avx
|
|
|
|
$L$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]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[16+r15]
|
|
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
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[32+r15]
|
|
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]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[48+r15]
|
|
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
|
|
cmp r8d,11
|
|
jb NEAR $L$vaesenclast10
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[64+r15]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[80+r15]
|
|
je NEAR $L$vaesenclast10
|
|
vaesenc xmm12,xmm12,xmm15
|
|
vmovups xmm14,XMMWORD[96+r15]
|
|
vaesenc xmm12,xmm12,xmm14
|
|
vmovups xmm15,XMMWORD[112+r15]
|
|
$L$vaesenclast10:
|
|
vaesenclast xmm12,xmm12,xmm15
|
|
vmovups xmm15,XMMWORD[((-112))+r15]
|
|
vmovups xmm14,XMMWORD[((16-112))+r15]
|
|
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
|
|
vmovups XMMWORD[48+r12*1+r13],xmm12
|
|
mov r8,QWORD[88+rsp]
|
|
|
|
add eax,DWORD[r9]
|
|
add esi,DWORD[4+r9]
|
|
add ecx,DWORD[8+r9]
|
|
mov DWORD[r9],eax
|
|
add edx,DWORD[12+r9]
|
|
mov DWORD[4+r9],esi
|
|
add ebp,DWORD[16+r9]
|
|
mov DWORD[8+r9],ecx
|
|
mov DWORD[12+r9],edx
|
|
mov DWORD[16+r9],ebp
|
|
vmovups XMMWORD[r8],xmm12
|
|
vzeroall
|
|
movaps xmm6,XMMWORD[((96+0))+rsp]
|
|
movaps xmm7,XMMWORD[((96+16))+rsp]
|
|
movaps xmm8,XMMWORD[((96+32))+rsp]
|
|
movaps xmm9,XMMWORD[((96+48))+rsp]
|
|
movaps xmm10,XMMWORD[((96+64))+rsp]
|
|
movaps xmm11,XMMWORD[((96+80))+rsp]
|
|
movaps xmm12,XMMWORD[((96+96))+rsp]
|
|
movaps xmm13,XMMWORD[((96+112))+rsp]
|
|
movaps xmm14,XMMWORD[((96+128))+rsp]
|
|
movaps xmm15,XMMWORD[((96+144))+rsp]
|
|
lea rsi,[264+rsp]
|
|
|
|
mov r15,QWORD[rsi]
|
|
|
|
mov r14,QWORD[8+rsi]
|
|
|
|
mov r13,QWORD[16+rsi]
|
|
|
|
mov r12,QWORD[24+rsi]
|
|
|
|
mov rbp,QWORD[32+rsi]
|
|
|
|
mov rbx,QWORD[40+rsi]
|
|
|
|
lea rsp,[48+rsi]
|
|
|
|
$L$epilogue_avx:
|
|
mov rdi,QWORD[8+rsp] ;WIN64 epilogue
|
|
mov rsi,QWORD[16+rsp]
|
|
DB 0F3h,0C3h ;repret
|
|
|
|
$L$SEH_end_aesni_cbc_sha1_enc_avx:
|
|
ALIGN 64
|
|
K_XX_XX:
|
|
DD 0x5a827999,0x5a827999,0x5a827999,0x5a827999
|
|
DD 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1
|
|
DD 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc
|
|
DD 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6
|
|
DD 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f
|
|
DB 0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0
|
|
|
|
DB 65,69,83,78,73,45,67,66,67,43,83,72,65,49,32,115
|
|
DB 116,105,116,99,104,32,102,111,114,32,120,56,54,95,54,52
|
|
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
|
|
ALIGN 64
|
|
|
|
ALIGN 32
|
|
aesni_cbc_sha1_enc_shaext:
|
|
mov QWORD[8+rsp],rdi ;WIN64 prologue
|
|
mov QWORD[16+rsp],rsi
|
|
mov rax,rsp
|
|
$L$SEH_begin_aesni_cbc_sha1_enc_shaext:
|
|
mov rdi,rcx
|
|
mov rsi,rdx
|
|
mov rdx,r8
|
|
mov rcx,r9
|
|
mov r8,QWORD[40+rsp]
|
|
mov r9,QWORD[48+rsp]
|
|
|
|
|
|
|
|
mov r10,QWORD[56+rsp]
|
|
lea rsp,[((-168))+rsp]
|
|
movaps XMMWORD[(-8-160)+rax],xmm6
|
|
movaps XMMWORD[(-8-144)+rax],xmm7
|
|
movaps XMMWORD[(-8-128)+rax],xmm8
|
|
movaps XMMWORD[(-8-112)+rax],xmm9
|
|
movaps XMMWORD[(-8-96)+rax],xmm10
|
|
movaps XMMWORD[(-8-80)+rax],xmm11
|
|
movaps XMMWORD[(-8-64)+rax],xmm12
|
|
movaps XMMWORD[(-8-48)+rax],xmm13
|
|
movaps XMMWORD[(-8-32)+rax],xmm14
|
|
movaps XMMWORD[(-8-16)+rax],xmm15
|
|
$L$prologue_shaext:
|
|
movdqu xmm8,XMMWORD[r9]
|
|
movd xmm9,DWORD[16+r9]
|
|
movdqa xmm7,XMMWORD[((K_XX_XX+80))]
|
|
|
|
mov r11d,DWORD[240+rcx]
|
|
sub rsi,rdi
|
|
movups xmm15,XMMWORD[rcx]
|
|
movups xmm2,XMMWORD[r8]
|
|
movups xmm0,XMMWORD[16+rcx]
|
|
lea rcx,[112+rcx]
|
|
|
|
pshufd xmm8,xmm8,27
|
|
pshufd xmm9,xmm9,27
|
|
jmp NEAR $L$oop_shaext
|
|
|
|
ALIGN 16
|
|
$L$oop_shaext:
|
|
movups xmm14,XMMWORD[rdi]
|
|
xorps xmm14,xmm15
|
|
xorps xmm2,xmm14
|
|
movups xmm1,XMMWORD[((-80))+rcx]
|
|
DB 102,15,56,220,208
|
|
movdqu xmm3,XMMWORD[r10]
|
|
movdqa xmm12,xmm9
|
|
DB 102,15,56,0,223
|
|
movdqu xmm4,XMMWORD[16+r10]
|
|
movdqa xmm11,xmm8
|
|
movups xmm0,XMMWORD[((-64))+rcx]
|
|
DB 102,15,56,220,209
|
|
DB 102,15,56,0,231
|
|
|
|
paddd xmm9,xmm3
|
|
movdqu xmm5,XMMWORD[32+r10]
|
|
lea r10,[64+r10]
|
|
pxor xmm3,xmm12
|
|
movups xmm1,XMMWORD[((-48))+rcx]
|
|
DB 102,15,56,220,208
|
|
pxor xmm3,xmm12
|
|
movdqa xmm10,xmm8
|
|
DB 102,15,56,0,239
|
|
DB 69,15,58,204,193,0
|
|
DB 68,15,56,200,212
|
|
movups xmm0,XMMWORD[((-32))+rcx]
|
|
DB 102,15,56,220,209
|
|
DB 15,56,201,220
|
|
movdqu xmm6,XMMWORD[((-16))+r10]
|
|
movdqa xmm9,xmm8
|
|
DB 102,15,56,0,247
|
|
movups xmm1,XMMWORD[((-16))+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 69,15,58,204,194,0
|
|
DB 68,15,56,200,205
|
|
pxor xmm3,xmm5
|
|
DB 15,56,201,229
|
|
movups xmm0,XMMWORD[rcx]
|
|
DB 102,15,56,220,209
|
|
movdqa xmm10,xmm8
|
|
DB 69,15,58,204,193,0
|
|
DB 68,15,56,200,214
|
|
movups xmm1,XMMWORD[16+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 15,56,202,222
|
|
pxor xmm4,xmm6
|
|
DB 15,56,201,238
|
|
movups xmm0,XMMWORD[32+rcx]
|
|
DB 102,15,56,220,209
|
|
movdqa xmm9,xmm8
|
|
DB 69,15,58,204,194,0
|
|
DB 68,15,56,200,203
|
|
movups xmm1,XMMWORD[48+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 15,56,202,227
|
|
pxor xmm5,xmm3
|
|
DB 15,56,201,243
|
|
cmp r11d,11
|
|
jb NEAR $L$aesenclast11
|
|
movups xmm0,XMMWORD[64+rcx]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[80+rcx]
|
|
DB 102,15,56,220,208
|
|
je NEAR $L$aesenclast11
|
|
movups xmm0,XMMWORD[96+rcx]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[112+rcx]
|
|
DB 102,15,56,220,208
|
|
$L$aesenclast11:
|
|
DB 102,15,56,221,209
|
|
movups xmm0,XMMWORD[((16-112))+rcx]
|
|
movdqa xmm10,xmm8
|
|
DB 69,15,58,204,193,0
|
|
DB 68,15,56,200,212
|
|
movups xmm14,XMMWORD[16+rdi]
|
|
xorps xmm14,xmm15
|
|
movups XMMWORD[rdi*1+rsi],xmm2
|
|
xorps xmm2,xmm14
|
|
movups xmm1,XMMWORD[((-80))+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 15,56,202,236
|
|
pxor xmm6,xmm4
|
|
DB 15,56,201,220
|
|
movups xmm0,XMMWORD[((-64))+rcx]
|
|
DB 102,15,56,220,209
|
|
movdqa xmm9,xmm8
|
|
DB 69,15,58,204,194,1
|
|
DB 68,15,56,200,205
|
|
movups xmm1,XMMWORD[((-48))+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 15,56,202,245
|
|
pxor xmm3,xmm5
|
|
DB 15,56,201,229
|
|
movups xmm0,XMMWORD[((-32))+rcx]
|
|
DB 102,15,56,220,209
|
|
movdqa xmm10,xmm8
|
|
DB 69,15,58,204,193,1
|
|
DB 68,15,56,200,214
|
|
movups xmm1,XMMWORD[((-16))+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 15,56,202,222
|
|
pxor xmm4,xmm6
|
|
DB 15,56,201,238
|
|
movups xmm0,XMMWORD[rcx]
|
|
DB 102,15,56,220,209
|
|
movdqa xmm9,xmm8
|
|
DB 69,15,58,204,194,1
|
|
DB 68,15,56,200,203
|
|
movups xmm1,XMMWORD[16+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 15,56,202,227
|
|
pxor xmm5,xmm3
|
|
DB 15,56,201,243
|
|
movups xmm0,XMMWORD[32+rcx]
|
|
DB 102,15,56,220,209
|
|
movdqa xmm10,xmm8
|
|
DB 69,15,58,204,193,1
|
|
DB 68,15,56,200,212
|
|
movups xmm1,XMMWORD[48+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 15,56,202,236
|
|
pxor xmm6,xmm4
|
|
DB 15,56,201,220
|
|
cmp r11d,11
|
|
jb NEAR $L$aesenclast12
|
|
movups xmm0,XMMWORD[64+rcx]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[80+rcx]
|
|
DB 102,15,56,220,208
|
|
je NEAR $L$aesenclast12
|
|
movups xmm0,XMMWORD[96+rcx]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[112+rcx]
|
|
DB 102,15,56,220,208
|
|
$L$aesenclast12:
|
|
DB 102,15,56,221,209
|
|
movups xmm0,XMMWORD[((16-112))+rcx]
|
|
movdqa xmm9,xmm8
|
|
DB 69,15,58,204,194,1
|
|
DB 68,15,56,200,205
|
|
movups xmm14,XMMWORD[32+rdi]
|
|
xorps xmm14,xmm15
|
|
movups XMMWORD[16+rdi*1+rsi],xmm2
|
|
xorps xmm2,xmm14
|
|
movups xmm1,XMMWORD[((-80))+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 15,56,202,245
|
|
pxor xmm3,xmm5
|
|
DB 15,56,201,229
|
|
movups xmm0,XMMWORD[((-64))+rcx]
|
|
DB 102,15,56,220,209
|
|
movdqa xmm10,xmm8
|
|
DB 69,15,58,204,193,2
|
|
DB 68,15,56,200,214
|
|
movups xmm1,XMMWORD[((-48))+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 15,56,202,222
|
|
pxor xmm4,xmm6
|
|
DB 15,56,201,238
|
|
movups xmm0,XMMWORD[((-32))+rcx]
|
|
DB 102,15,56,220,209
|
|
movdqa xmm9,xmm8
|
|
DB 69,15,58,204,194,2
|
|
DB 68,15,56,200,203
|
|
movups xmm1,XMMWORD[((-16))+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 15,56,202,227
|
|
pxor xmm5,xmm3
|
|
DB 15,56,201,243
|
|
movups xmm0,XMMWORD[rcx]
|
|
DB 102,15,56,220,209
|
|
movdqa xmm10,xmm8
|
|
DB 69,15,58,204,193,2
|
|
DB 68,15,56,200,212
|
|
movups xmm1,XMMWORD[16+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 15,56,202,236
|
|
pxor xmm6,xmm4
|
|
DB 15,56,201,220
|
|
movups xmm0,XMMWORD[32+rcx]
|
|
DB 102,15,56,220,209
|
|
movdqa xmm9,xmm8
|
|
DB 69,15,58,204,194,2
|
|
DB 68,15,56,200,205
|
|
movups xmm1,XMMWORD[48+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 15,56,202,245
|
|
pxor xmm3,xmm5
|
|
DB 15,56,201,229
|
|
cmp r11d,11
|
|
jb NEAR $L$aesenclast13
|
|
movups xmm0,XMMWORD[64+rcx]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[80+rcx]
|
|
DB 102,15,56,220,208
|
|
je NEAR $L$aesenclast13
|
|
movups xmm0,XMMWORD[96+rcx]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[112+rcx]
|
|
DB 102,15,56,220,208
|
|
$L$aesenclast13:
|
|
DB 102,15,56,221,209
|
|
movups xmm0,XMMWORD[((16-112))+rcx]
|
|
movdqa xmm10,xmm8
|
|
DB 69,15,58,204,193,2
|
|
DB 68,15,56,200,214
|
|
movups xmm14,XMMWORD[48+rdi]
|
|
xorps xmm14,xmm15
|
|
movups XMMWORD[32+rdi*1+rsi],xmm2
|
|
xorps xmm2,xmm14
|
|
movups xmm1,XMMWORD[((-80))+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 15,56,202,222
|
|
pxor xmm4,xmm6
|
|
DB 15,56,201,238
|
|
movups xmm0,XMMWORD[((-64))+rcx]
|
|
DB 102,15,56,220,209
|
|
movdqa xmm9,xmm8
|
|
DB 69,15,58,204,194,3
|
|
DB 68,15,56,200,203
|
|
movups xmm1,XMMWORD[((-48))+rcx]
|
|
DB 102,15,56,220,208
|
|
DB 15,56,202,227
|
|
pxor xmm5,xmm3
|
|
DB 15,56,201,243
|
|
movups xmm0,XMMWORD[((-32))+rcx]
|
|
DB 102,15,56,220,209
|
|
movdqa xmm10,xmm8
|
|
DB 69,15,58,204,193,3
|
|
DB 68,15,56,200,212
|
|
DB 15,56,202,236
|
|
pxor xmm6,xmm4
|
|
movups xmm1,XMMWORD[((-16))+rcx]
|
|
DB 102,15,56,220,208
|
|
movdqa xmm9,xmm8
|
|
DB 69,15,58,204,194,3
|
|
DB 68,15,56,200,205
|
|
DB 15,56,202,245
|
|
movups xmm0,XMMWORD[rcx]
|
|
DB 102,15,56,220,209
|
|
movdqa xmm5,xmm12
|
|
movdqa xmm10,xmm8
|
|
DB 69,15,58,204,193,3
|
|
DB 68,15,56,200,214
|
|
movups xmm1,XMMWORD[16+rcx]
|
|
DB 102,15,56,220,208
|
|
movdqa xmm9,xmm8
|
|
DB 69,15,58,204,194,3
|
|
DB 68,15,56,200,205
|
|
movups xmm0,XMMWORD[32+rcx]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[48+rcx]
|
|
DB 102,15,56,220,208
|
|
cmp r11d,11
|
|
jb NEAR $L$aesenclast14
|
|
movups xmm0,XMMWORD[64+rcx]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[80+rcx]
|
|
DB 102,15,56,220,208
|
|
je NEAR $L$aesenclast14
|
|
movups xmm0,XMMWORD[96+rcx]
|
|
DB 102,15,56,220,209
|
|
movups xmm1,XMMWORD[112+rcx]
|
|
DB 102,15,56,220,208
|
|
$L$aesenclast14:
|
|
DB 102,15,56,221,209
|
|
movups xmm0,XMMWORD[((16-112))+rcx]
|
|
dec rdx
|
|
|
|
paddd xmm8,xmm11
|
|
movups XMMWORD[48+rdi*1+rsi],xmm2
|
|
lea rdi,[64+rdi]
|
|
jnz NEAR $L$oop_shaext
|
|
|
|
pshufd xmm8,xmm8,27
|
|
pshufd xmm9,xmm9,27
|
|
movups XMMWORD[r8],xmm2
|
|
movdqu XMMWORD[r9],xmm8
|
|
movd DWORD[16+r9],xmm9
|
|
movaps xmm6,XMMWORD[((-8-160))+rax]
|
|
movaps xmm7,XMMWORD[((-8-144))+rax]
|
|
movaps xmm8,XMMWORD[((-8-128))+rax]
|
|
movaps xmm9,XMMWORD[((-8-112))+rax]
|
|
movaps xmm10,XMMWORD[((-8-96))+rax]
|
|
movaps xmm11,XMMWORD[((-8-80))+rax]
|
|
movaps xmm12,XMMWORD[((-8-64))+rax]
|
|
movaps xmm13,XMMWORD[((-8-48))+rax]
|
|
movaps xmm14,XMMWORD[((-8-32))+rax]
|
|
movaps xmm15,XMMWORD[((-8-16))+rax]
|
|
mov rsp,rax
|
|
$L$epilogue_shaext:
|
|
mov rdi,QWORD[8+rsp] ;WIN64 epilogue
|
|
mov rsi,QWORD[16+rsp]
|
|
DB 0F3h,0C3h ;repret
|
|
|
|
$L$SEH_end_aesni_cbc_sha1_enc_shaext:
|
|
EXTERN __imp_RtlVirtualUnwind
|
|
|
|
ALIGN 16
|
|
ssse3_handler:
|
|
push rsi
|
|
push rdi
|
|
push rbx
|
|
push rbp
|
|
push r12
|
|
push r13
|
|
push r14
|
|
push r15
|
|
pushfq
|
|
sub rsp,64
|
|
|
|
mov rax,QWORD[120+r8]
|
|
mov rbx,QWORD[248+r8]
|
|
|
|
mov rsi,QWORD[8+r9]
|
|
mov r11,QWORD[56+r9]
|
|
|
|
mov r10d,DWORD[r11]
|
|
lea r10,[r10*1+rsi]
|
|
cmp rbx,r10
|
|
jb NEAR $L$common_seh_tail
|
|
|
|
mov rax,QWORD[152+r8]
|
|
|
|
mov r10d,DWORD[4+r11]
|
|
lea r10,[r10*1+rsi]
|
|
cmp rbx,r10
|
|
jae NEAR $L$common_seh_tail
|
|
lea r10,[aesni_cbc_sha1_enc_shaext]
|
|
cmp rbx,r10
|
|
jb NEAR $L$seh_no_shaext
|
|
|
|
lea rsi,[rax]
|
|
lea rdi,[512+r8]
|
|
mov ecx,20
|
|
DD 0xa548f3fc
|
|
lea rax,[168+rax]
|
|
jmp NEAR $L$common_seh_tail
|
|
$L$seh_no_shaext:
|
|
lea rsi,[96+rax]
|
|
lea rdi,[512+r8]
|
|
mov ecx,20
|
|
DD 0xa548f3fc
|
|
lea rax,[264+rax]
|
|
|
|
mov r15,QWORD[rax]
|
|
mov r14,QWORD[8+rax]
|
|
mov r13,QWORD[16+rax]
|
|
mov r12,QWORD[24+rax]
|
|
mov rbp,QWORD[32+rax]
|
|
mov rbx,QWORD[40+rax]
|
|
lea rax,[48+rax]
|
|
mov QWORD[144+r8],rbx
|
|
mov QWORD[160+r8],rbp
|
|
mov QWORD[216+r8],r12
|
|
mov QWORD[224+r8],r13
|
|
mov QWORD[232+r8],r14
|
|
mov QWORD[240+r8],r15
|
|
|
|
$L$common_seh_tail:
|
|
mov rdi,QWORD[8+rax]
|
|
mov rsi,QWORD[16+rax]
|
|
mov QWORD[152+r8],rax
|
|
mov QWORD[168+r8],rsi
|
|
mov QWORD[176+r8],rdi
|
|
|
|
mov rdi,QWORD[40+r9]
|
|
mov rsi,r8
|
|
mov ecx,154
|
|
DD 0xa548f3fc
|
|
|
|
mov rsi,r9
|
|
xor rcx,rcx
|
|
mov rdx,QWORD[8+rsi]
|
|
mov r8,QWORD[rsi]
|
|
mov r9,QWORD[16+rsi]
|
|
mov r10,QWORD[40+rsi]
|
|
lea r11,[56+rsi]
|
|
lea r12,[24+rsi]
|
|
mov QWORD[32+rsp],r10
|
|
mov QWORD[40+rsp],r11
|
|
mov QWORD[48+rsp],r12
|
|
mov QWORD[56+rsp],rcx
|
|
call QWORD[__imp_RtlVirtualUnwind]
|
|
|
|
mov eax,1
|
|
add rsp,64
|
|
popfq
|
|
pop r15
|
|
pop r14
|
|
pop r13
|
|
pop r12
|
|
pop rbp
|
|
pop rbx
|
|
pop rdi
|
|
pop rsi
|
|
DB 0F3h,0C3h ;repret
|
|
|
|
|
|
section .pdata rdata align=4
|
|
ALIGN 4
|
|
DD $L$SEH_begin_aesni_cbc_sha1_enc_ssse3 wrt ..imagebase
|
|
DD $L$SEH_end_aesni_cbc_sha1_enc_ssse3 wrt ..imagebase
|
|
DD $L$SEH_info_aesni_cbc_sha1_enc_ssse3 wrt ..imagebase
|
|
DD $L$SEH_begin_aesni_cbc_sha1_enc_avx wrt ..imagebase
|
|
DD $L$SEH_end_aesni_cbc_sha1_enc_avx wrt ..imagebase
|
|
DD $L$SEH_info_aesni_cbc_sha1_enc_avx wrt ..imagebase
|
|
DD $L$SEH_begin_aesni_cbc_sha1_enc_shaext wrt ..imagebase
|
|
DD $L$SEH_end_aesni_cbc_sha1_enc_shaext wrt ..imagebase
|
|
DD $L$SEH_info_aesni_cbc_sha1_enc_shaext wrt ..imagebase
|
|
section .xdata rdata align=8
|
|
ALIGN 8
|
|
$L$SEH_info_aesni_cbc_sha1_enc_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_aesni_cbc_sha1_enc_avx:
|
|
DB 9,0,0,0
|
|
DD ssse3_handler wrt ..imagebase
|
|
DD $L$prologue_avx wrt ..imagebase,$L$epilogue_avx wrt ..imagebase
|
|
$L$SEH_info_aesni_cbc_sha1_enc_shaext:
|
|
DB 9,0,0,0
|
|
DD ssse3_handler wrt ..imagebase
|
|
DD $L$prologue_shaext wrt ..imagebase,$L$epilogue_shaext wrt ..imagebase
|