123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023 |
- IF @Version LT 800
- ECHO MASM version 8.00 or later is strongly recommended.
- ENDIF
- .686
- .XMM
- IF @Version LT 800
- XMMWORD STRUCT 16
- DQ 2 dup (?)
- XMMWORD ENDS
- ENDIF
- .MODEL FLAT
- OPTION DOTNAME
- IF @Version LT 800
- .text$ SEGMENT PAGE 'CODE'
- ELSE
- .text$ SEGMENT ALIGN(64) 'CODE'
- ENDIF
- ALIGN 16
- _ChaCha20_ctr32 PROC PUBLIC
- $L_ChaCha20_ctr32_begin::
- push ebp
- push ebx
- push esi
- push edi
- xor eax,eax
- cmp eax,DWORD PTR 28[esp]
- je $L000no_data
- call $Lpic_point
- $Lpic_point::
- pop eax
- lea ebp,DWORD PTR _OPENSSL_ia32cap_P
- test DWORD PTR [ebp],16777216
- jz $L001x86
- test DWORD PTR 4[ebp],512
- jz $L001x86
- jmp $Lssse3_shortcut
- $L001x86:
- mov esi,DWORD PTR 32[esp]
- mov edi,DWORD PTR 36[esp]
- sub esp,132
- mov eax,DWORD PTR [esi]
- mov ebx,DWORD PTR 4[esi]
- mov ecx,DWORD PTR 8[esi]
- mov edx,DWORD PTR 12[esi]
- mov DWORD PTR 80[esp],eax
- mov DWORD PTR 84[esp],ebx
- mov DWORD PTR 88[esp],ecx
- mov DWORD PTR 92[esp],edx
- mov eax,DWORD PTR 16[esi]
- mov ebx,DWORD PTR 20[esi]
- mov ecx,DWORD PTR 24[esi]
- mov edx,DWORD PTR 28[esi]
- mov DWORD PTR 96[esp],eax
- mov DWORD PTR 100[esp],ebx
- mov DWORD PTR 104[esp],ecx
- mov DWORD PTR 108[esp],edx
- mov eax,DWORD PTR [edi]
- mov ebx,DWORD PTR 4[edi]
- mov ecx,DWORD PTR 8[edi]
- mov edx,DWORD PTR 12[edi]
- sub eax,1
- mov DWORD PTR 112[esp],eax
- mov DWORD PTR 116[esp],ebx
- mov DWORD PTR 120[esp],ecx
- mov DWORD PTR 124[esp],edx
- jmp $L002entry
- ALIGN 16
- $L003outer_loop:
- mov DWORD PTR 156[esp],ebx
- mov DWORD PTR 152[esp],eax
- mov DWORD PTR 160[esp],ecx
- $L002entry:
- mov eax,1634760805
- mov DWORD PTR 4[esp],857760878
- mov DWORD PTR 8[esp],2036477234
- mov DWORD PTR 12[esp],1797285236
- mov ebx,DWORD PTR 84[esp]
- mov ebp,DWORD PTR 88[esp]
- mov ecx,DWORD PTR 104[esp]
- mov esi,DWORD PTR 108[esp]
- mov edx,DWORD PTR 116[esp]
- mov edi,DWORD PTR 120[esp]
- mov DWORD PTR 20[esp],ebx
- mov DWORD PTR 24[esp],ebp
- mov DWORD PTR 40[esp],ecx
- mov DWORD PTR 44[esp],esi
- mov DWORD PTR 52[esp],edx
- mov DWORD PTR 56[esp],edi
- mov ebx,DWORD PTR 92[esp]
- mov edi,DWORD PTR 124[esp]
- mov edx,DWORD PTR 112[esp]
- mov ebp,DWORD PTR 80[esp]
- mov ecx,DWORD PTR 96[esp]
- mov esi,DWORD PTR 100[esp]
- add edx,1
- mov DWORD PTR 28[esp],ebx
- mov DWORD PTR 60[esp],edi
- mov DWORD PTR 112[esp],edx
- mov ebx,10
- jmp $L004loop
- ALIGN 16
- $L004loop:
- add eax,ebp
- mov DWORD PTR 128[esp],ebx
- mov ebx,ebp
- xor edx,eax
- rol edx,16
- add ecx,edx
- xor ebx,ecx
- mov edi,DWORD PTR 52[esp]
- rol ebx,12
- mov ebp,DWORD PTR 20[esp]
- add eax,ebx
- xor edx,eax
- mov DWORD PTR [esp],eax
- rol edx,8
- mov eax,DWORD PTR 4[esp]
- add ecx,edx
- mov DWORD PTR 48[esp],edx
- xor ebx,ecx
- add eax,ebp
- rol ebx,7
- xor edi,eax
- mov DWORD PTR 32[esp],ecx
- rol edi,16
- mov DWORD PTR 16[esp],ebx
- add esi,edi
- mov ecx,DWORD PTR 40[esp]
- xor ebp,esi
- mov edx,DWORD PTR 56[esp]
- rol ebp,12
- mov ebx,DWORD PTR 24[esp]
- add eax,ebp
- xor edi,eax
- mov DWORD PTR 4[esp],eax
- rol edi,8
- mov eax,DWORD PTR 8[esp]
- add esi,edi
- mov DWORD PTR 52[esp],edi
- xor ebp,esi
- add eax,ebx
- rol ebp,7
- xor edx,eax
- mov DWORD PTR 36[esp],esi
- rol edx,16
- mov DWORD PTR 20[esp],ebp
- add ecx,edx
- mov esi,DWORD PTR 44[esp]
- xor ebx,ecx
- mov edi,DWORD PTR 60[esp]
- rol ebx,12
- mov ebp,DWORD PTR 28[esp]
- add eax,ebx
- xor edx,eax
- mov DWORD PTR 8[esp],eax
- rol edx,8
- mov eax,DWORD PTR 12[esp]
- add ecx,edx
- mov DWORD PTR 56[esp],edx
- xor ebx,ecx
- add eax,ebp
- rol ebx,7
- xor edi,eax
- rol edi,16
- mov DWORD PTR 24[esp],ebx
- add esi,edi
- xor ebp,esi
- rol ebp,12
- mov ebx,DWORD PTR 20[esp]
- add eax,ebp
- xor edi,eax
- mov DWORD PTR 12[esp],eax
- rol edi,8
- mov eax,DWORD PTR [esp]
- add esi,edi
- mov edx,edi
- xor ebp,esi
- add eax,ebx
- rol ebp,7
- xor edx,eax
- rol edx,16
- mov DWORD PTR 28[esp],ebp
- add ecx,edx
- xor ebx,ecx
- mov edi,DWORD PTR 48[esp]
- rol ebx,12
- mov ebp,DWORD PTR 24[esp]
- add eax,ebx
- xor edx,eax
- mov DWORD PTR [esp],eax
- rol edx,8
- mov eax,DWORD PTR 4[esp]
- add ecx,edx
- mov DWORD PTR 60[esp],edx
- xor ebx,ecx
- add eax,ebp
- rol ebx,7
- xor edi,eax
- mov DWORD PTR 40[esp],ecx
- rol edi,16
- mov DWORD PTR 20[esp],ebx
- add esi,edi
- mov ecx,DWORD PTR 32[esp]
- xor ebp,esi
- mov edx,DWORD PTR 52[esp]
- rol ebp,12
- mov ebx,DWORD PTR 28[esp]
- add eax,ebp
- xor edi,eax
- mov DWORD PTR 4[esp],eax
- rol edi,8
- mov eax,DWORD PTR 8[esp]
- add esi,edi
- mov DWORD PTR 48[esp],edi
- xor ebp,esi
- add eax,ebx
- rol ebp,7
- xor edx,eax
- mov DWORD PTR 44[esp],esi
- rol edx,16
- mov DWORD PTR 24[esp],ebp
- add ecx,edx
- mov esi,DWORD PTR 36[esp]
- xor ebx,ecx
- mov edi,DWORD PTR 56[esp]
- rol ebx,12
- mov ebp,DWORD PTR 16[esp]
- add eax,ebx
- xor edx,eax
- mov DWORD PTR 8[esp],eax
- rol edx,8
- mov eax,DWORD PTR 12[esp]
- add ecx,edx
- mov DWORD PTR 52[esp],edx
- xor ebx,ecx
- add eax,ebp
- rol ebx,7
- xor edi,eax
- rol edi,16
- mov DWORD PTR 28[esp],ebx
- add esi,edi
- xor ebp,esi
- mov edx,DWORD PTR 48[esp]
- rol ebp,12
- mov ebx,DWORD PTR 128[esp]
- add eax,ebp
- xor edi,eax
- mov DWORD PTR 12[esp],eax
- rol edi,8
- mov eax,DWORD PTR [esp]
- add esi,edi
- mov DWORD PTR 56[esp],edi
- xor ebp,esi
- rol ebp,7
- dec ebx
- jnz $L004loop
- mov ebx,DWORD PTR 160[esp]
- add eax,1634760805
- add ebp,DWORD PTR 80[esp]
- add ecx,DWORD PTR 96[esp]
- add esi,DWORD PTR 100[esp]
- cmp ebx,64
- jb $L005tail
- mov ebx,DWORD PTR 156[esp]
- add edx,DWORD PTR 112[esp]
- add edi,DWORD PTR 120[esp]
- xor eax,DWORD PTR [ebx]
- xor ebp,DWORD PTR 16[ebx]
- mov DWORD PTR [esp],eax
- mov eax,DWORD PTR 152[esp]
- xor ecx,DWORD PTR 32[ebx]
- xor esi,DWORD PTR 36[ebx]
- xor edx,DWORD PTR 48[ebx]
- xor edi,DWORD PTR 56[ebx]
- mov DWORD PTR 16[eax],ebp
- mov DWORD PTR 32[eax],ecx
- mov DWORD PTR 36[eax],esi
- mov DWORD PTR 48[eax],edx
- mov DWORD PTR 56[eax],edi
- mov ebp,DWORD PTR 4[esp]
- mov ecx,DWORD PTR 8[esp]
- mov esi,DWORD PTR 12[esp]
- mov edx,DWORD PTR 20[esp]
- mov edi,DWORD PTR 24[esp]
- add ebp,857760878
- add ecx,2036477234
- add esi,1797285236
- add edx,DWORD PTR 84[esp]
- add edi,DWORD PTR 88[esp]
- xor ebp,DWORD PTR 4[ebx]
- xor ecx,DWORD PTR 8[ebx]
- xor esi,DWORD PTR 12[ebx]
- xor edx,DWORD PTR 20[ebx]
- xor edi,DWORD PTR 24[ebx]
- mov DWORD PTR 4[eax],ebp
- mov DWORD PTR 8[eax],ecx
- mov DWORD PTR 12[eax],esi
- mov DWORD PTR 20[eax],edx
- mov DWORD PTR 24[eax],edi
- mov ebp,DWORD PTR 28[esp]
- mov ecx,DWORD PTR 40[esp]
- mov esi,DWORD PTR 44[esp]
- mov edx,DWORD PTR 52[esp]
- mov edi,DWORD PTR 60[esp]
- add ebp,DWORD PTR 92[esp]
- add ecx,DWORD PTR 104[esp]
- add esi,DWORD PTR 108[esp]
- add edx,DWORD PTR 116[esp]
- add edi,DWORD PTR 124[esp]
- xor ebp,DWORD PTR 28[ebx]
- xor ecx,DWORD PTR 40[ebx]
- xor esi,DWORD PTR 44[ebx]
- xor edx,DWORD PTR 52[ebx]
- xor edi,DWORD PTR 60[ebx]
- lea ebx,DWORD PTR 64[ebx]
- mov DWORD PTR 28[eax],ebp
- mov ebp,DWORD PTR [esp]
- mov DWORD PTR 40[eax],ecx
- mov ecx,DWORD PTR 160[esp]
- mov DWORD PTR 44[eax],esi
- mov DWORD PTR 52[eax],edx
- mov DWORD PTR 60[eax],edi
- mov DWORD PTR [eax],ebp
- lea eax,DWORD PTR 64[eax]
- sub ecx,64
- jnz $L003outer_loop
- jmp $L006done
- $L005tail:
- add edx,DWORD PTR 112[esp]
- add edi,DWORD PTR 120[esp]
- mov DWORD PTR [esp],eax
- mov DWORD PTR 16[esp],ebp
- mov DWORD PTR 32[esp],ecx
- mov DWORD PTR 36[esp],esi
- mov DWORD PTR 48[esp],edx
- mov DWORD PTR 56[esp],edi
- mov ebp,DWORD PTR 4[esp]
- mov ecx,DWORD PTR 8[esp]
- mov esi,DWORD PTR 12[esp]
- mov edx,DWORD PTR 20[esp]
- mov edi,DWORD PTR 24[esp]
- add ebp,857760878
- add ecx,2036477234
- add esi,1797285236
- add edx,DWORD PTR 84[esp]
- add edi,DWORD PTR 88[esp]
- mov DWORD PTR 4[esp],ebp
- mov DWORD PTR 8[esp],ecx
- mov DWORD PTR 12[esp],esi
- mov DWORD PTR 20[esp],edx
- mov DWORD PTR 24[esp],edi
- mov ebp,DWORD PTR 28[esp]
- mov ecx,DWORD PTR 40[esp]
- mov esi,DWORD PTR 44[esp]
- mov edx,DWORD PTR 52[esp]
- mov edi,DWORD PTR 60[esp]
- add ebp,DWORD PTR 92[esp]
- add ecx,DWORD PTR 104[esp]
- add esi,DWORD PTR 108[esp]
- add edx,DWORD PTR 116[esp]
- add edi,DWORD PTR 124[esp]
- mov DWORD PTR 28[esp],ebp
- mov ebp,DWORD PTR 156[esp]
- mov DWORD PTR 40[esp],ecx
- mov ecx,DWORD PTR 152[esp]
- mov DWORD PTR 44[esp],esi
- xor esi,esi
- mov DWORD PTR 52[esp],edx
- mov DWORD PTR 60[esp],edi
- xor eax,eax
- xor edx,edx
- $L007tail_loop:
- mov al,BYTE PTR [ebp*1+esi]
- mov dl,BYTE PTR [esi*1+esp]
- lea esi,DWORD PTR 1[esi]
- xor al,dl
- mov BYTE PTR [esi*1+ecx-1],al
- dec ebx
- jnz $L007tail_loop
- $L006done:
- add esp,132
- $L000no_data:
- pop edi
- pop esi
- pop ebx
- pop ebp
- ret
- _ChaCha20_ctr32 ENDP
- ALIGN 16
- _ChaCha20_ssse3 PROC PUBLIC
- $L_ChaCha20_ssse3_begin::
- push ebp
- push ebx
- push esi
- push edi
- $Lssse3_shortcut::
- test DWORD PTR 4[ebp],2048
- jnz $Lxop_shortcut
- mov edi,DWORD PTR 20[esp]
- mov esi,DWORD PTR 24[esp]
- mov ecx,DWORD PTR 28[esp]
- mov edx,DWORD PTR 32[esp]
- mov ebx,DWORD PTR 36[esp]
- mov ebp,esp
- sub esp,524
- and esp,-64
- mov DWORD PTR 512[esp],ebp
- lea eax,DWORD PTR ($Lssse3_data-$Lpic_point)[eax]
- movdqu xmm3,XMMWORD PTR [ebx]
- $L0081x:
- movdqa xmm0,XMMWORD PTR 32[eax]
- movdqu xmm1,XMMWORD PTR [edx]
- movdqu xmm2,XMMWORD PTR 16[edx]
- movdqa xmm6,XMMWORD PTR [eax]
- movdqa xmm7,XMMWORD PTR 16[eax]
- mov DWORD PTR 48[esp],ebp
- movdqa XMMWORD PTR [esp],xmm0
- movdqa XMMWORD PTR 16[esp],xmm1
- movdqa XMMWORD PTR 32[esp],xmm2
- movdqa XMMWORD PTR 48[esp],xmm3
- mov edx,10
- jmp $L009loop1x
- ALIGN 16
- $L010outer1x:
- movdqa xmm3,XMMWORD PTR 80[eax]
- movdqa xmm0,XMMWORD PTR [esp]
- movdqa xmm1,XMMWORD PTR 16[esp]
- movdqa xmm2,XMMWORD PTR 32[esp]
- paddd xmm3,XMMWORD PTR 48[esp]
- mov edx,10
- movdqa XMMWORD PTR 48[esp],xmm3
- jmp $L009loop1x
- ALIGN 16
- $L009loop1x:
- 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 $L009loop1x
- paddd xmm0,XMMWORD PTR [esp]
- paddd xmm1,XMMWORD PTR 16[esp]
- paddd xmm2,XMMWORD PTR 32[esp]
- paddd xmm3,XMMWORD PTR 48[esp]
- cmp ecx,64
- jb $L011tail
- movdqu xmm4,XMMWORD PTR [esi]
- movdqu xmm5,XMMWORD PTR 16[esi]
- pxor xmm0,xmm4
- movdqu xmm4,XMMWORD PTR 32[esi]
- pxor xmm1,xmm5
- movdqu xmm5,XMMWORD PTR 48[esi]
- pxor xmm2,xmm4
- pxor xmm3,xmm5
- lea esi,DWORD PTR 64[esi]
- movdqu XMMWORD PTR [edi],xmm0
- movdqu XMMWORD PTR 16[edi],xmm1
- movdqu XMMWORD PTR 32[edi],xmm2
- movdqu XMMWORD PTR 48[edi],xmm3
- lea edi,DWORD PTR 64[edi]
- sub ecx,64
- jnz $L010outer1x
- jmp $L012done
- $L011tail:
- movdqa XMMWORD PTR [esp],xmm0
- movdqa XMMWORD PTR 16[esp],xmm1
- movdqa XMMWORD PTR 32[esp],xmm2
- movdqa XMMWORD PTR 48[esp],xmm3
- xor eax,eax
- xor edx,edx
- xor ebp,ebp
- $L013tail_loop:
- mov al,BYTE PTR [ebp*1+esp]
- mov dl,BYTE PTR [ebp*1+esi]
- lea ebp,DWORD PTR 1[ebp]
- xor al,dl
- mov BYTE PTR [ebp*1+edi-1],al
- dec ecx
- jnz $L013tail_loop
- $L012done:
- mov esp,DWORD PTR 512[esp]
- pop edi
- pop esi
- pop ebx
- pop ebp
- ret
- _ChaCha20_ssse3 ENDP
- ALIGN 64
- $Lssse3_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
- ALIGN 16
- _ChaCha20_xop PROC PUBLIC
- $L_ChaCha20_xop_begin::
- push ebp
- push ebx
- push esi
- push edi
- $Lxop_shortcut::
- mov edi,DWORD PTR 20[esp]
- mov esi,DWORD PTR 24[esp]
- mov ecx,DWORD PTR 28[esp]
- mov edx,DWORD PTR 32[esp]
- mov ebx,DWORD PTR 36[esp]
- vzeroupper
- mov ebp,esp
- sub esp,524
- and esp,-64
- mov DWORD PTR 512[esp],ebp
- lea eax,DWORD PTR ($Lssse3_data-$Lpic_point)[eax]
- vmovdqu xmm3,XMMWORD PTR [ebx]
- cmp ecx,256
- jb $L0141x
- mov DWORD PTR 516[esp],edx
- mov DWORD PTR 520[esp],ebx
- sub ecx,256
- lea ebp,DWORD PTR 384[esp]
- vmovdqu xmm7,XMMWORD PTR [edx]
- vpshufd xmm0,xmm3,0
- vpshufd xmm1,xmm3,85
- vpshufd xmm2,xmm3,170
- vpshufd xmm3,xmm3,255
- vpaddd xmm0,xmm0,XMMWORD PTR 48[eax]
- vpshufd xmm4,xmm7,0
- vpshufd xmm5,xmm7,85
- vpsubd xmm0,xmm0,XMMWORD PTR 64[eax]
- vpshufd xmm6,xmm7,170
- vpshufd xmm7,xmm7,255
- vmovdqa XMMWORD PTR 64[ebp],xmm0
- vmovdqa XMMWORD PTR 80[ebp],xmm1
- vmovdqa XMMWORD PTR 96[ebp],xmm2
- vmovdqa XMMWORD PTR 112[ebp],xmm3
- vmovdqu xmm3,XMMWORD PTR 16[edx]
- vmovdqa XMMWORD PTR [ebp-64],xmm4
- vmovdqa XMMWORD PTR [ebp-48],xmm5
- vmovdqa XMMWORD PTR [ebp-32],xmm6
- vmovdqa XMMWORD PTR [ebp-16],xmm7
- vmovdqa xmm7,XMMWORD PTR 32[eax]
- lea ebx,DWORD PTR 128[esp]
- vpshufd xmm0,xmm3,0
- vpshufd xmm1,xmm3,85
- vpshufd xmm2,xmm3,170
- vpshufd xmm3,xmm3,255
- vpshufd xmm4,xmm7,0
- vpshufd xmm5,xmm7,85
- vpshufd xmm6,xmm7,170
- vpshufd xmm7,xmm7,255
- vmovdqa XMMWORD PTR [ebp],xmm0
- vmovdqa XMMWORD PTR 16[ebp],xmm1
- vmovdqa XMMWORD PTR 32[ebp],xmm2
- vmovdqa XMMWORD PTR 48[ebp],xmm3
- vmovdqa XMMWORD PTR [ebp-128],xmm4
- vmovdqa XMMWORD PTR [ebp-112],xmm5
- vmovdqa XMMWORD PTR [ebp-96],xmm6
- vmovdqa XMMWORD PTR [ebp-80],xmm7
- lea esi,DWORD PTR 128[esi]
- lea edi,DWORD PTR 128[edi]
- jmp $L015outer_loop
- ALIGN 32
- $L015outer_loop:
- vmovdqa xmm1,XMMWORD PTR [ebp-112]
- vmovdqa xmm2,XMMWORD PTR [ebp-96]
- vmovdqa xmm3,XMMWORD PTR [ebp-80]
- vmovdqa xmm5,XMMWORD PTR [ebp-48]
- vmovdqa xmm6,XMMWORD PTR [ebp-32]
- vmovdqa xmm7,XMMWORD PTR [ebp-16]
- vmovdqa XMMWORD PTR [ebx-112],xmm1
- vmovdqa XMMWORD PTR [ebx-96],xmm2
- vmovdqa XMMWORD PTR [ebx-80],xmm3
- vmovdqa XMMWORD PTR [ebx-48],xmm5
- vmovdqa XMMWORD PTR [ebx-32],xmm6
- vmovdqa XMMWORD PTR [ebx-16],xmm7
- vmovdqa xmm2,XMMWORD PTR 32[ebp]
- vmovdqa xmm3,XMMWORD PTR 48[ebp]
- vmovdqa xmm4,XMMWORD PTR 64[ebp]
- vmovdqa xmm5,XMMWORD PTR 80[ebp]
- vmovdqa xmm6,XMMWORD PTR 96[ebp]
- vmovdqa xmm7,XMMWORD PTR 112[ebp]
- vpaddd xmm4,xmm4,XMMWORD PTR 64[eax]
- vmovdqa XMMWORD PTR 32[ebx],xmm2
- vmovdqa XMMWORD PTR 48[ebx],xmm3
- vmovdqa XMMWORD PTR 64[ebx],xmm4
- vmovdqa XMMWORD PTR 80[ebx],xmm5
- vmovdqa XMMWORD PTR 96[ebx],xmm6
- vmovdqa XMMWORD PTR 112[ebx],xmm7
- vmovdqa XMMWORD PTR 64[ebp],xmm4
- vmovdqa xmm0,XMMWORD PTR [ebp-128]
- vmovdqa xmm6,xmm4
- vmovdqa xmm3,XMMWORD PTR [ebp-64]
- vmovdqa xmm4,XMMWORD PTR [ebp]
- vmovdqa xmm5,XMMWORD PTR 16[ebp]
- mov edx,10
- nop
- ALIGN 32
- $L016loop:
- vpaddd xmm0,xmm0,xmm3
- vpxor xmm6,xmm6,xmm0
- DB 143,232,120,194,246,16
- vpaddd xmm4,xmm4,xmm6
- vpxor xmm2,xmm3,xmm4
- vmovdqa xmm1,XMMWORD PTR [ebx-112]
- DB 143,232,120,194,210,12
- vmovdqa xmm3,XMMWORD PTR [ebx-48]
- vpaddd xmm0,xmm0,xmm2
- vmovdqa xmm7,XMMWORD PTR 80[ebx]
- vpxor xmm6,xmm6,xmm0
- vpaddd xmm1,xmm1,xmm3
- DB 143,232,120,194,246,8
- vmovdqa XMMWORD PTR [ebx-128],xmm0
- vpaddd xmm4,xmm4,xmm6
- vmovdqa XMMWORD PTR 64[ebx],xmm6
- vpxor xmm2,xmm2,xmm4
- vpxor xmm7,xmm7,xmm1
- DB 143,232,120,194,210,7
- vmovdqa XMMWORD PTR [ebx],xmm4
- DB 143,232,120,194,255,16
- vmovdqa XMMWORD PTR [ebx-64],xmm2
- vpaddd xmm5,xmm5,xmm7
- vmovdqa xmm4,XMMWORD PTR 32[ebx]
- vpxor xmm3,xmm3,xmm5
- vmovdqa xmm0,XMMWORD PTR [ebx-96]
- DB 143,232,120,194,219,12
- vmovdqa xmm2,XMMWORD PTR [ebx-32]
- vpaddd xmm1,xmm1,xmm3
- vmovdqa xmm6,XMMWORD PTR 96[ebx]
- vpxor xmm7,xmm7,xmm1
- vpaddd xmm0,xmm0,xmm2
- DB 143,232,120,194,255,8
- vmovdqa XMMWORD PTR [ebx-112],xmm1
- vpaddd xmm5,xmm5,xmm7
- vmovdqa XMMWORD PTR 80[ebx],xmm7
- vpxor xmm3,xmm3,xmm5
- vpxor xmm6,xmm6,xmm0
- DB 143,232,120,194,219,7
- vmovdqa XMMWORD PTR 16[ebx],xmm5
- DB 143,232,120,194,246,16
- vmovdqa XMMWORD PTR [ebx-48],xmm3
- vpaddd xmm4,xmm4,xmm6
- vmovdqa xmm5,XMMWORD PTR 48[ebx]
- vpxor xmm2,xmm2,xmm4
- vmovdqa xmm1,XMMWORD PTR [ebx-80]
- DB 143,232,120,194,210,12
- vmovdqa xmm3,XMMWORD PTR [ebx-16]
- vpaddd xmm0,xmm0,xmm2
- vmovdqa xmm7,XMMWORD PTR 112[ebx]
- vpxor xmm6,xmm6,xmm0
- vpaddd xmm1,xmm1,xmm3
- DB 143,232,120,194,246,8
- vmovdqa XMMWORD PTR [ebx-96],xmm0
- vpaddd xmm4,xmm4,xmm6
- vmovdqa XMMWORD PTR 96[ebx],xmm6
- vpxor xmm2,xmm2,xmm4
- vpxor xmm7,xmm7,xmm1
- DB 143,232,120,194,210,7
- DB 143,232,120,194,255,16
- vmovdqa XMMWORD PTR [ebx-32],xmm2
- vpaddd xmm5,xmm5,xmm7
- vpxor xmm3,xmm3,xmm5
- vmovdqa xmm0,XMMWORD PTR [ebx-128]
- DB 143,232,120,194,219,12
- vmovdqa xmm2,XMMWORD PTR [ebx-48]
- vpaddd xmm1,xmm1,xmm3
- vpxor xmm7,xmm7,xmm1
- vpaddd xmm0,xmm0,xmm2
- DB 143,232,120,194,255,8
- vmovdqa XMMWORD PTR [ebx-80],xmm1
- vpaddd xmm5,xmm5,xmm7
- vpxor xmm3,xmm3,xmm5
- vpxor xmm6,xmm7,xmm0
- DB 143,232,120,194,219,7
- DB 143,232,120,194,246,16
- vmovdqa XMMWORD PTR [ebx-16],xmm3
- vpaddd xmm4,xmm4,xmm6
- vpxor xmm2,xmm2,xmm4
- vmovdqa xmm1,XMMWORD PTR [ebx-112]
- DB 143,232,120,194,210,12
- vmovdqa xmm3,XMMWORD PTR [ebx-32]
- vpaddd xmm0,xmm0,xmm2
- vmovdqa xmm7,XMMWORD PTR 64[ebx]
- vpxor xmm6,xmm6,xmm0
- vpaddd xmm1,xmm1,xmm3
- DB 143,232,120,194,246,8
- vmovdqa XMMWORD PTR [ebx-128],xmm0
- vpaddd xmm4,xmm4,xmm6
- vmovdqa XMMWORD PTR 112[ebx],xmm6
- vpxor xmm2,xmm2,xmm4
- vpxor xmm7,xmm7,xmm1
- DB 143,232,120,194,210,7
- vmovdqa XMMWORD PTR 32[ebx],xmm4
- DB 143,232,120,194,255,16
- vmovdqa XMMWORD PTR [ebx-48],xmm2
- vpaddd xmm5,xmm5,xmm7
- vmovdqa xmm4,XMMWORD PTR [ebx]
- vpxor xmm3,xmm3,xmm5
- vmovdqa xmm0,XMMWORD PTR [ebx-96]
- DB 143,232,120,194,219,12
- vmovdqa xmm2,XMMWORD PTR [ebx-16]
- vpaddd xmm1,xmm1,xmm3
- vmovdqa xmm6,XMMWORD PTR 80[ebx]
- vpxor xmm7,xmm7,xmm1
- vpaddd xmm0,xmm0,xmm2
- DB 143,232,120,194,255,8
- vmovdqa XMMWORD PTR [ebx-112],xmm1
- vpaddd xmm5,xmm5,xmm7
- vmovdqa XMMWORD PTR 64[ebx],xmm7
- vpxor xmm3,xmm3,xmm5
- vpxor xmm6,xmm6,xmm0
- DB 143,232,120,194,219,7
- vmovdqa XMMWORD PTR 48[ebx],xmm5
- DB 143,232,120,194,246,16
- vmovdqa XMMWORD PTR [ebx-32],xmm3
- vpaddd xmm4,xmm4,xmm6
- vmovdqa xmm5,XMMWORD PTR 16[ebx]
- vpxor xmm2,xmm2,xmm4
- vmovdqa xmm1,XMMWORD PTR [ebx-80]
- DB 143,232,120,194,210,12
- vmovdqa xmm3,XMMWORD PTR [ebx-64]
- vpaddd xmm0,xmm0,xmm2
- vmovdqa xmm7,XMMWORD PTR 96[ebx]
- vpxor xmm6,xmm6,xmm0
- vpaddd xmm1,xmm1,xmm3
- DB 143,232,120,194,246,8
- vmovdqa XMMWORD PTR [ebx-96],xmm0
- vpaddd xmm4,xmm4,xmm6
- vmovdqa XMMWORD PTR 80[ebx],xmm6
- vpxor xmm2,xmm2,xmm4
- vpxor xmm7,xmm7,xmm1
- DB 143,232,120,194,210,7
- DB 143,232,120,194,255,16
- vmovdqa XMMWORD PTR [ebx-16],xmm2
- vpaddd xmm5,xmm5,xmm7
- vpxor xmm3,xmm3,xmm5
- vmovdqa xmm0,XMMWORD PTR [ebx-128]
- DB 143,232,120,194,219,12
- vpaddd xmm1,xmm1,xmm3
- vmovdqa xmm6,XMMWORD PTR 64[ebx]
- vpxor xmm7,xmm7,xmm1
- DB 143,232,120,194,255,8
- vmovdqa XMMWORD PTR [ebx-80],xmm1
- vpaddd xmm5,xmm5,xmm7
- vmovdqa XMMWORD PTR 96[ebx],xmm7
- vpxor xmm3,xmm3,xmm5
- DB 143,232,120,194,219,7
- dec edx
- jnz $L016loop
- vmovdqa XMMWORD PTR [ebx-64],xmm3
- vmovdqa XMMWORD PTR [ebx],xmm4
- vmovdqa XMMWORD PTR 16[ebx],xmm5
- vmovdqa XMMWORD PTR 64[ebx],xmm6
- vmovdqa XMMWORD PTR 96[ebx],xmm7
- vmovdqa xmm1,XMMWORD PTR [ebx-112]
- vmovdqa xmm2,XMMWORD PTR [ebx-96]
- vmovdqa xmm3,XMMWORD PTR [ebx-80]
- vpaddd xmm0,xmm0,XMMWORD PTR [ebp-128]
- vpaddd xmm1,xmm1,XMMWORD PTR [ebp-112]
- vpaddd xmm2,xmm2,XMMWORD PTR [ebp-96]
- vpaddd xmm3,xmm3,XMMWORD PTR [ebp-80]
- vpunpckldq xmm6,xmm0,xmm1
- vpunpckldq xmm7,xmm2,xmm3
- vpunpckhdq xmm0,xmm0,xmm1
- vpunpckhdq xmm2,xmm2,xmm3
- vpunpcklqdq xmm1,xmm6,xmm7
- vpunpckhqdq xmm6,xmm6,xmm7
- vpunpcklqdq xmm7,xmm0,xmm2
- vpunpckhqdq xmm3,xmm0,xmm2
- vpxor xmm4,xmm1,XMMWORD PTR [esi-128]
- vpxor xmm5,xmm6,XMMWORD PTR [esi-64]
- vpxor xmm6,xmm7,XMMWORD PTR [esi]
- vpxor xmm7,xmm3,XMMWORD PTR 64[esi]
- lea esi,QWORD PTR 16[esi]
- vmovdqa xmm0,XMMWORD PTR [ebx-64]
- vmovdqa xmm1,XMMWORD PTR [ebx-48]
- vmovdqa xmm2,XMMWORD PTR [ebx-32]
- vmovdqa xmm3,XMMWORD PTR [ebx-16]
- vmovdqu XMMWORD PTR [edi-128],xmm4
- vmovdqu XMMWORD PTR [edi-64],xmm5
- vmovdqu XMMWORD PTR [edi],xmm6
- vmovdqu XMMWORD PTR 64[edi],xmm7
- lea edi,QWORD PTR 16[edi]
- vpaddd xmm0,xmm0,XMMWORD PTR [ebp-64]
- vpaddd xmm1,xmm1,XMMWORD PTR [ebp-48]
- vpaddd xmm2,xmm2,XMMWORD PTR [ebp-32]
- vpaddd xmm3,xmm3,XMMWORD PTR [ebp-16]
- vpunpckldq xmm6,xmm0,xmm1
- vpunpckldq xmm7,xmm2,xmm3
- vpunpckhdq xmm0,xmm0,xmm1
- vpunpckhdq xmm2,xmm2,xmm3
- vpunpcklqdq xmm1,xmm6,xmm7
- vpunpckhqdq xmm6,xmm6,xmm7
- vpunpcklqdq xmm7,xmm0,xmm2
- vpunpckhqdq xmm3,xmm0,xmm2
- vpxor xmm4,xmm1,XMMWORD PTR [esi-128]
- vpxor xmm5,xmm6,XMMWORD PTR [esi-64]
- vpxor xmm6,xmm7,XMMWORD PTR [esi]
- vpxor xmm7,xmm3,XMMWORD PTR 64[esi]
- lea esi,QWORD PTR 16[esi]
- vmovdqa xmm0,XMMWORD PTR [ebx]
- vmovdqa xmm1,XMMWORD PTR 16[ebx]
- vmovdqa xmm2,XMMWORD PTR 32[ebx]
- vmovdqa xmm3,XMMWORD PTR 48[ebx]
- vmovdqu XMMWORD PTR [edi-128],xmm4
- vmovdqu XMMWORD PTR [edi-64],xmm5
- vmovdqu XMMWORD PTR [edi],xmm6
- vmovdqu XMMWORD PTR 64[edi],xmm7
- lea edi,QWORD PTR 16[edi]
- vpaddd xmm0,xmm0,XMMWORD PTR [ebp]
- vpaddd xmm1,xmm1,XMMWORD PTR 16[ebp]
- vpaddd xmm2,xmm2,XMMWORD PTR 32[ebp]
- vpaddd xmm3,xmm3,XMMWORD PTR 48[ebp]
- vpunpckldq xmm6,xmm0,xmm1
- vpunpckldq xmm7,xmm2,xmm3
- vpunpckhdq xmm0,xmm0,xmm1
- vpunpckhdq xmm2,xmm2,xmm3
- vpunpcklqdq xmm1,xmm6,xmm7
- vpunpckhqdq xmm6,xmm6,xmm7
- vpunpcklqdq xmm7,xmm0,xmm2
- vpunpckhqdq xmm3,xmm0,xmm2
- vpxor xmm4,xmm1,XMMWORD PTR [esi-128]
- vpxor xmm5,xmm6,XMMWORD PTR [esi-64]
- vpxor xmm6,xmm7,XMMWORD PTR [esi]
- vpxor xmm7,xmm3,XMMWORD PTR 64[esi]
- lea esi,QWORD PTR 16[esi]
- vmovdqa xmm0,XMMWORD PTR 64[ebx]
- vmovdqa xmm1,XMMWORD PTR 80[ebx]
- vmovdqa xmm2,XMMWORD PTR 96[ebx]
- vmovdqa xmm3,XMMWORD PTR 112[ebx]
- vmovdqu XMMWORD PTR [edi-128],xmm4
- vmovdqu XMMWORD PTR [edi-64],xmm5
- vmovdqu XMMWORD PTR [edi],xmm6
- vmovdqu XMMWORD PTR 64[edi],xmm7
- lea edi,QWORD PTR 16[edi]
- vpaddd xmm0,xmm0,XMMWORD PTR 64[ebp]
- vpaddd xmm1,xmm1,XMMWORD PTR 80[ebp]
- vpaddd xmm2,xmm2,XMMWORD PTR 96[ebp]
- vpaddd xmm3,xmm3,XMMWORD PTR 112[ebp]
- vpunpckldq xmm6,xmm0,xmm1
- vpunpckldq xmm7,xmm2,xmm3
- vpunpckhdq xmm0,xmm0,xmm1
- vpunpckhdq xmm2,xmm2,xmm3
- vpunpcklqdq xmm1,xmm6,xmm7
- vpunpckhqdq xmm6,xmm6,xmm7
- vpunpcklqdq xmm7,xmm0,xmm2
- vpunpckhqdq xmm3,xmm0,xmm2
- vpxor xmm4,xmm1,XMMWORD PTR [esi-128]
- vpxor xmm5,xmm6,XMMWORD PTR [esi-64]
- vpxor xmm6,xmm7,XMMWORD PTR [esi]
- vpxor xmm7,xmm3,XMMWORD PTR 64[esi]
- lea esi,QWORD PTR 208[esi]
- vmovdqu XMMWORD PTR [edi-128],xmm4
- vmovdqu XMMWORD PTR [edi-64],xmm5
- vmovdqu XMMWORD PTR [edi],xmm6
- vmovdqu XMMWORD PTR 64[edi],xmm7
- lea edi,QWORD PTR 208[edi]
- sub ecx,256
- jnc $L015outer_loop
- add ecx,256
- jz $L017done
- mov ebx,DWORD PTR 520[esp]
- lea esi,DWORD PTR [esi-128]
- mov edx,DWORD PTR 516[esp]
- lea edi,DWORD PTR [edi-128]
- vmovd xmm2,DWORD PTR 64[ebp]
- vmovdqu xmm3,XMMWORD PTR [ebx]
- vpaddd xmm2,xmm2,XMMWORD PTR 96[eax]
- vpand xmm3,xmm3,XMMWORD PTR 112[eax]
- vpor xmm3,xmm3,xmm2
- $L0141x:
- vmovdqa xmm0,XMMWORD PTR 32[eax]
- vmovdqu xmm1,XMMWORD PTR [edx]
- vmovdqu xmm2,XMMWORD PTR 16[edx]
- vmovdqa xmm6,XMMWORD PTR [eax]
- vmovdqa xmm7,XMMWORD PTR 16[eax]
- mov DWORD PTR 48[esp],ebp
- vmovdqa XMMWORD PTR [esp],xmm0
- vmovdqa XMMWORD PTR 16[esp],xmm1
- vmovdqa XMMWORD PTR 32[esp],xmm2
- vmovdqa XMMWORD PTR 48[esp],xmm3
- mov edx,10
- jmp $L018loop1x
- ALIGN 16
- $L019outer1x:
- vmovdqa xmm3,XMMWORD PTR 80[eax]
- vmovdqa xmm0,XMMWORD PTR [esp]
- vmovdqa xmm1,XMMWORD PTR 16[esp]
- vmovdqa xmm2,XMMWORD PTR 32[esp]
- vpaddd xmm3,xmm3,XMMWORD PTR 48[esp]
- mov edx,10
- vmovdqa XMMWORD PTR 48[esp],xmm3
- jmp $L018loop1x
- ALIGN 16
- $L018loop1x:
- vpaddd xmm0,xmm0,xmm1
- vpxor xmm3,xmm3,xmm0
- DB 143,232,120,194,219,16
- vpaddd xmm2,xmm2,xmm3
- vpxor xmm1,xmm1,xmm2
- DB 143,232,120,194,201,12
- vpaddd xmm0,xmm0,xmm1
- vpxor xmm3,xmm3,xmm0
- DB 143,232,120,194,219,8
- vpaddd xmm2,xmm2,xmm3
- vpxor xmm1,xmm1,xmm2
- DB 143,232,120,194,201,7
- vpshufd xmm2,xmm2,78
- vpshufd xmm1,xmm1,57
- vpshufd xmm3,xmm3,147
- vpaddd xmm0,xmm0,xmm1
- vpxor xmm3,xmm3,xmm0
- DB 143,232,120,194,219,16
- vpaddd xmm2,xmm2,xmm3
- vpxor xmm1,xmm1,xmm2
- DB 143,232,120,194,201,12
- vpaddd xmm0,xmm0,xmm1
- vpxor xmm3,xmm3,xmm0
- DB 143,232,120,194,219,8
- vpaddd xmm2,xmm2,xmm3
- vpxor xmm1,xmm1,xmm2
- DB 143,232,120,194,201,7
- vpshufd xmm2,xmm2,78
- vpshufd xmm1,xmm1,147
- vpshufd xmm3,xmm3,57
- dec edx
- jnz $L018loop1x
- vpaddd xmm0,xmm0,XMMWORD PTR [esp]
- vpaddd xmm1,xmm1,XMMWORD PTR 16[esp]
- vpaddd xmm2,xmm2,XMMWORD PTR 32[esp]
- vpaddd xmm3,xmm3,XMMWORD PTR 48[esp]
- cmp ecx,64
- jb $L020tail
- vpxor xmm0,xmm0,XMMWORD PTR [esi]
- vpxor xmm1,xmm1,XMMWORD PTR 16[esi]
- vpxor xmm2,xmm2,XMMWORD PTR 32[esi]
- vpxor xmm3,xmm3,XMMWORD PTR 48[esi]
- lea esi,DWORD PTR 64[esi]
- vmovdqu XMMWORD PTR [edi],xmm0
- vmovdqu XMMWORD PTR 16[edi],xmm1
- vmovdqu XMMWORD PTR 32[edi],xmm2
- vmovdqu XMMWORD PTR 48[edi],xmm3
- lea edi,DWORD PTR 64[edi]
- sub ecx,64
- jnz $L019outer1x
- jmp $L017done
- $L020tail:
- vmovdqa XMMWORD PTR [esp],xmm0
- vmovdqa XMMWORD PTR 16[esp],xmm1
- vmovdqa XMMWORD PTR 32[esp],xmm2
- vmovdqa XMMWORD PTR 48[esp],xmm3
- xor eax,eax
- xor edx,edx
- xor ebp,ebp
- $L021tail_loop:
- mov al,BYTE PTR [ebp*1+esp]
- mov dl,BYTE PTR [ebp*1+esi]
- lea ebp,DWORD PTR 1[ebp]
- xor al,dl
- mov BYTE PTR [ebp*1+edi-1],al
- dec ecx
- jnz $L021tail_loop
- $L017done:
- vzeroupper
- mov esp,DWORD PTR 512[esp]
- pop edi
- pop esi
- pop ebx
- pop ebp
- ret
- _ChaCha20_xop ENDP
- .text$ ENDS
- .bss SEGMENT 'BSS'
- COMM _OPENSSL_ia32cap_P:DWORD:4
- .bss ENDS
- END
|