123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310231123122313231423152316231723182319232023212322232323242325232623272328232923302331233223332334233523362337233823392340234123422343234423452346234723482349235023512352235323542355235623572358235923602361236223632364236523662367236823692370237123722373237423752376237723782379238023812382238323842385238623872388238923902391239223932394239523962397239823992400240124022403240424052406240724082409241024112412241324142415241624172418241924202421242224232424242524262427242824292430243124322433243424352436243724382439244024412442244324442445244624472448244924502451245224532454245524562457245824592460246124622463246424652466246724682469247024712472247324742475247624772478247924802481248224832484248524862487248824892490249124922493249424952496249724982499250025012502250325042505250625072508250925102511251225132514251525162517251825192520252125222523252425252526252725282529253025312532253325342535253625372538253925402541254225432544254525462547254825492550255125522553255425552556255725582559256025612562256325642565256625672568256925702571257225732574257525762577257825792580258125822583258425852586258725882589259025912592259325942595259625972598259926002601260226032604260526062607260826092610261126122613261426152616261726182619262026212622262326242625262626272628262926302631263226332634263526362637263826392640264126422643264426452646264726482649265026512652265326542655265626572658265926602661266226632664266526662667266826692670267126722673267426752676267726782679268026812682268326842685268626872688268926902691269226932694269526962697269826992700270127022703270427052706270727082709271027112712271327142715271627172718271927202721272227232724272527262727272827292730273127322733273427352736273727382739274027412742274327442745274627472748274927502751275227532754275527562757275827592760276127622763276427652766276727682769277027712772277327742775277627772778277927802781278227832784278527862787278827892790279127922793279427952796279727982799280028012802280328042805280628072808280928102811281228132814281528162817281828192820282128222823282428252826282728282829283028312832283328342835283628372838283928402841284228432844284528462847284828492850285128522853285428552856285728582859286028612862286328642865286628672868286928702871287228732874287528762877287828792880288128822883288428852886288728882889289028912892289328942895289628972898289929002901290229032904290529062907290829092910291129122913291429152916291729182919292029212922292329242925292629272928292929302931293229332934293529362937293829392940294129422943294429452946294729482949295029512952295329542955295629572958295929602961296229632964296529662967296829692970297129722973297429752976297729782979298029812982298329842985298629872988298929902991299229932994299529962997299829993000300130023003300430053006300730083009301030113012301330143015301630173018301930203021302230233024302530263027302830293030303130323033303430353036303730383039304030413042304330443045304630473048304930503051305230533054305530563057305830593060306130623063306430653066306730683069307030713072307330743075307630773078307930803081308230833084308530863087308830893090309130923093309430953096309730983099310031013102310331043105310631073108310931103111311231133114311531163117311831193120312131223123312431253126312731283129313031313132313331343135313631373138313931403141314231433144314531463147314831493150315131523153315431553156315731583159316031613162316331643165316631673168316931703171317231733174317531763177317831793180318131823183318431853186318731883189319031913192319331943195319631973198319932003201320232033204320532063207320832093210321132123213321432153216321732183219322032213222322332243225322632273228322932303231323232333234323532363237323832393240324132423243324432453246324732483249325032513252325332543255325632573258325932603261326232633264326532663267326832693270327132723273327432753276327732783279328032813282328332843285328632873288328932903291329232933294329532963297329832993300330133023303330433053306330733083309331033113312331333143315331633173318331933203321332233233324332533263327332833293330333133323333333433353336333733383339334033413342334333443345334633473348334933503351335233533354335533563357335833593360336133623363336433653366336733683369337033713372337333743375337633773378337933803381338233833384338533863387338833893390339133923393339433953396339733983399340034013402340334043405340634073408340934103411341234133414341534163417341834193420342134223423342434253426342734283429343034313432343334343435343634373438343934403441344234433444344534463447344834493450345134523453345434553456345734583459346034613462346334643465346634673468346934703471347234733474347534763477347834793480348134823483348434853486348734883489349034913492349334943495349634973498349935003501350235033504350535063507350835093510351135123513351435153516351735183519352035213522352335243525352635273528352935303531353235333534353535363537353835393540354135423543354435453546354735483549355035513552355335543555355635573558355935603561356235633564356535663567356835693570357135723573357435753576357735783579358035813582358335843585358635873588358935903591359235933594359535963597359835993600360136023603360436053606360736083609361036113612361336143615361636173618361936203621362236233624362536263627362836293630363136323633363436353636363736383639364036413642364336443645364636473648364936503651365236533654365536563657365836593660366136623663366436653666366736683669367036713672367336743675367636773678367936803681368236833684368536863687368836893690369136923693369436953696369736983699370037013702370337043705370637073708370937103711371237133714371537163717371837193720372137223723372437253726372737283729373037313732373337343735373637373738373937403741374237433744374537463747374837493750375137523753375437553756375737583759376037613762376337643765376637673768376937703771377237733774377537763777377837793780378137823783378437853786378737883789379037913792379337943795379637973798379938003801380238033804380538063807380838093810381138123813381438153816381738183819382038213822382338243825382638273828382938303831383238333834383538363837383838393840384138423843384438453846384738483849385038513852385338543855385638573858385938603861386238633864386538663867386838693870387138723873387438753876387738783879388038813882388338843885388638873888388938903891389238933894389538963897389838993900390139023903390439053906390739083909391039113912391339143915391639173918391939203921392239233924392539263927392839293930393139323933393439353936393739383939394039413942394339443945394639473948394939503951395239533954395539563957395839593960396139623963396439653966396739683969397039713972397339743975397639773978397939803981398239833984398539863987398839893990399139923993399439953996399739983999400040014002400340044005400640074008400940104011401240134014401540164017401840194020402140224023402440254026402740284029403040314032403340344035403640374038403940404041404240434044404540464047404840494050405140524053405440554056405740584059406040614062406340644065406640674068406940704071407240734074407540764077407840794080408140824083408440854086408740884089409040914092409340944095409640974098409941004101410241034104410541064107410841094110411141124113411441154116411741184119412041214122412341244125412641274128412941304131413241334134413541364137413841394140414141424143414441454146414741484149415041514152415341544155415641574158415941604161416241634164416541664167416841694170417141724173417441754176417741784179418041814182418341844185418641874188418941904191419241934194419541964197419841994200420142024203420442054206420742084209421042114212421342144215421642174218421942204221422242234224422542264227422842294230423142324233423442354236423742384239424042414242424342444245424642474248424942504251425242534254425542564257425842594260426142624263426442654266426742684269427042714272427342744275427642774278427942804281428242834284428542864287428842894290429142924293429442954296429742984299430043014302430343044305430643074308430943104311431243134314431543164317431843194320432143224323432443254326432743284329433043314332433343344335433643374338433943404341434243434344434543464347434843494350435143524353435443554356435743584359436043614362436343644365436643674368436943704371437243734374437543764377437843794380438143824383438443854386438743884389439043914392439343944395439643974398439944004401440244034404440544064407440844094410441144124413441444154416441744184419442044214422442344244425442644274428442944304431443244334434443544364437443844394440444144424443444444454446444744484449445044514452445344544455445644574458445944604461446244634464446544664467446844694470447144724473447444754476447744784479448044814482448344844485448644874488448944904491449244934494449544964497449844994500450145024503450445054506450745084509451045114512451345144515451645174518451945204521452245234524452545264527452845294530453145324533453445354536453745384539454045414542454345444545454645474548454945504551455245534554455545564557455845594560456145624563456445654566456745684569457045714572457345744575457645774578457945804581458245834584458545864587458845894590459145924593459445954596459745984599460046014602460346044605460646074608460946104611461246134614461546164617461846194620462146224623462446254626462746284629463046314632463346344635463646374638463946404641464246434644464546464647464846494650465146524653465446554656465746584659466046614662466346644665466646674668466946704671467246734674467546764677467846794680468146824683468446854686468746884689469046914692469346944695469646974698469947004701470247034704470547064707470847094710471147124713 |
- OPTION DOTNAME
- .text$ SEGMENT ALIGN(256) 'CODE'
- EXTERN OPENSSL_ia32cap_P:NEAR
- PUBLIC aesni_cbc_sha256_enc
- ALIGN 16
- aesni_cbc_sha256_enc PROC PUBLIC
- lea r11,QWORD PTR[OPENSSL_ia32cap_P]
- mov eax,1
- cmp rcx,0
- je $L$probe
- mov eax,DWORD PTR[r11]
- mov r10,QWORD PTR[4+r11]
- bt r10,61
- jc aesni_cbc_sha256_enc_shaext
- mov r11,r10
- shr r11,32
- test r10d,2048
- jnz aesni_cbc_sha256_enc_xop
- and r11d,296
- cmp r11d,296
- je aesni_cbc_sha256_enc_avx2
- and r10d,268435456
- jnz aesni_cbc_sha256_enc_avx
- ud2
- xor eax,eax
- cmp rcx,0
- je $L$probe
- ud2
- $L$probe::
- DB 0F3h,0C3h ;repret
- aesni_cbc_sha256_enc ENDP
- ALIGN 64
- K256::
- DD 0428a2f98h,071374491h,0b5c0fbcfh,0e9b5dba5h
- DD 0428a2f98h,071374491h,0b5c0fbcfh,0e9b5dba5h
- DD 03956c25bh,059f111f1h,0923f82a4h,0ab1c5ed5h
- DD 03956c25bh,059f111f1h,0923f82a4h,0ab1c5ed5h
- DD 0d807aa98h,012835b01h,0243185beh,0550c7dc3h
- DD 0d807aa98h,012835b01h,0243185beh,0550c7dc3h
- DD 072be5d74h,080deb1feh,09bdc06a7h,0c19bf174h
- DD 072be5d74h,080deb1feh,09bdc06a7h,0c19bf174h
- DD 0e49b69c1h,0efbe4786h,00fc19dc6h,0240ca1cch
- DD 0e49b69c1h,0efbe4786h,00fc19dc6h,0240ca1cch
- DD 02de92c6fh,04a7484aah,05cb0a9dch,076f988dah
- DD 02de92c6fh,04a7484aah,05cb0a9dch,076f988dah
- DD 0983e5152h,0a831c66dh,0b00327c8h,0bf597fc7h
- DD 0983e5152h,0a831c66dh,0b00327c8h,0bf597fc7h
- DD 0c6e00bf3h,0d5a79147h,006ca6351h,014292967h
- DD 0c6e00bf3h,0d5a79147h,006ca6351h,014292967h
- DD 027b70a85h,02e1b2138h,04d2c6dfch,053380d13h
- DD 027b70a85h,02e1b2138h,04d2c6dfch,053380d13h
- DD 0650a7354h,0766a0abbh,081c2c92eh,092722c85h
- DD 0650a7354h,0766a0abbh,081c2c92eh,092722c85h
- DD 0a2bfe8a1h,0a81a664bh,0c24b8b70h,0c76c51a3h
- DD 0a2bfe8a1h,0a81a664bh,0c24b8b70h,0c76c51a3h
- DD 0d192e819h,0d6990624h,0f40e3585h,0106aa070h
- DD 0d192e819h,0d6990624h,0f40e3585h,0106aa070h
- DD 019a4c116h,01e376c08h,02748774ch,034b0bcb5h
- DD 019a4c116h,01e376c08h,02748774ch,034b0bcb5h
- DD 0391c0cb3h,04ed8aa4ah,05b9cca4fh,0682e6ff3h
- DD 0391c0cb3h,04ed8aa4ah,05b9cca4fh,0682e6ff3h
- DD 0748f82eeh,078a5636fh,084c87814h,08cc70208h
- DD 0748f82eeh,078a5636fh,084c87814h,08cc70208h
- DD 090befffah,0a4506cebh,0bef9a3f7h,0c67178f2h
- DD 090befffah,0a4506cebh,0bef9a3f7h,0c67178f2h
- DD 000010203h,004050607h,008090a0bh,00c0d0e0fh
- DD 000010203h,004050607h,008090a0bh,00c0d0e0fh
- DD 0,0,0,0,0,0,0,0,-1,-1,-1,-1
- DD 0,0,0,0,0,0,0,0
- DB 65,69,83,78,73,45,67,66,67,43,83,72,65,50,53,54
- DB 32,115,116,105,116,99,104,32,102,111,114,32,120,56,54,95
- DB 54,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98
- DB 121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108
- DB 46,111,114,103,62,0
- ALIGN 64
- ALIGN 64
- aesni_cbc_sha256_enc_xop PROC PRIVATE
- mov QWORD PTR[8+rsp],rdi ;WIN64 prologue
- mov QWORD PTR[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_aesni_cbc_sha256_enc_xop::
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- mov r8,QWORD PTR[40+rsp]
- mov r9,QWORD PTR[48+rsp]
- $L$xop_shortcut::
- mov r10,QWORD PTR[56+rsp]
- mov rax,rsp
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- sub rsp,288
- and rsp,-64
- shl rdx,6
- sub rsi,rdi
- sub r10,rdi
- add rdx,rdi
- mov QWORD PTR[((64+8))+rsp],rsi
- mov QWORD PTR[((64+16))+rsp],rdx
- mov QWORD PTR[((64+32))+rsp],r8
- mov QWORD PTR[((64+40))+rsp],r9
- mov QWORD PTR[((64+48))+rsp],r10
- mov QWORD PTR[120+rsp],rax
- movaps XMMWORD PTR[128+rsp],xmm6
- movaps XMMWORD PTR[144+rsp],xmm7
- movaps XMMWORD PTR[160+rsp],xmm8
- movaps XMMWORD PTR[176+rsp],xmm9
- movaps XMMWORD PTR[192+rsp],xmm10
- movaps XMMWORD PTR[208+rsp],xmm11
- movaps XMMWORD PTR[224+rsp],xmm12
- movaps XMMWORD PTR[240+rsp],xmm13
- movaps XMMWORD PTR[256+rsp],xmm14
- movaps XMMWORD PTR[272+rsp],xmm15
- $L$prologue_xop::
- vzeroall
- mov r12,rdi
- lea rdi,QWORD PTR[128+rcx]
- lea r13,QWORD PTR[((K256+544))]
- mov r14d,DWORD PTR[((240-128))+rdi]
- mov r15,r9
- mov rsi,r10
- vmovdqu xmm8,XMMWORD PTR[r8]
- sub r14,9
- mov eax,DWORD PTR[r15]
- mov ebx,DWORD PTR[4+r15]
- mov ecx,DWORD PTR[8+r15]
- mov edx,DWORD PTR[12+r15]
- mov r8d,DWORD PTR[16+r15]
- mov r9d,DWORD PTR[20+r15]
- mov r10d,DWORD PTR[24+r15]
- mov r11d,DWORD PTR[28+r15]
- vmovdqa xmm14,XMMWORD PTR[r14*8+r13]
- vmovdqa xmm13,XMMWORD PTR[16+r14*8+r13]
- vmovdqa xmm12,XMMWORD PTR[32+r14*8+r13]
- vmovdqu xmm10,XMMWORD PTR[((0-128))+rdi]
- jmp $L$loop_xop
- ALIGN 16
- $L$loop_xop::
- vmovdqa xmm7,XMMWORD PTR[((K256+512))]
- vmovdqu xmm0,XMMWORD PTR[r12*1+rsi]
- vmovdqu xmm1,XMMWORD PTR[16+r12*1+rsi]
- vmovdqu xmm2,XMMWORD PTR[32+r12*1+rsi]
- vmovdqu xmm3,XMMWORD PTR[48+r12*1+rsi]
- vpshufb xmm0,xmm0,xmm7
- lea rbp,QWORD PTR[K256]
- vpshufb xmm1,xmm1,xmm7
- vpshufb xmm2,xmm2,xmm7
- vpaddd xmm4,xmm0,XMMWORD PTR[rbp]
- vpshufb xmm3,xmm3,xmm7
- vpaddd xmm5,xmm1,XMMWORD PTR[32+rbp]
- vpaddd xmm6,xmm2,XMMWORD PTR[64+rbp]
- vpaddd xmm7,xmm3,XMMWORD PTR[96+rbp]
- vmovdqa XMMWORD PTR[rsp],xmm4
- mov r14d,eax
- vmovdqa XMMWORD PTR[16+rsp],xmm5
- mov esi,ebx
- vmovdqa XMMWORD PTR[32+rsp],xmm6
- xor esi,ecx
- vmovdqa XMMWORD PTR[48+rsp],xmm7
- mov r13d,r8d
- jmp $L$xop_00_47
- ALIGN 16
- $L$xop_00_47::
- sub rbp,-16*2*4
- vmovdqu xmm9,XMMWORD PTR[r12]
- mov QWORD PTR[((64+0))+rsp],r12
- vpalignr xmm4,xmm1,xmm0,4
- ror r13d,14
- mov eax,r14d
- vpalignr xmm7,xmm3,xmm2,4
- mov r12d,r9d
- xor r13d,r8d
- DB 143,232,120,194,236,14
- ror r14d,9
- xor r12d,r10d
- vpsrld xmm4,xmm4,3
- ror r13d,5
- xor r14d,eax
- vpaddd xmm0,xmm0,xmm7
- and r12d,r8d
- vpxor xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((16-128))+rdi]
- xor r13d,r8d
- add r11d,DWORD PTR[rsp]
- mov r15d,eax
- DB 143,232,120,194,245,11
- ror r14d,11
- xor r12d,r10d
- vpxor xmm4,xmm4,xmm5
- xor r15d,ebx
- ror r13d,6
- add r11d,r12d
- and esi,r15d
- DB 143,232,120,194,251,13
- xor r14d,eax
- add r11d,r13d
- vpxor xmm4,xmm4,xmm6
- xor esi,ebx
- add edx,r11d
- vpsrld xmm6,xmm3,10
- ror r14d,2
- add r11d,esi
- vpaddd xmm0,xmm0,xmm4
- mov r13d,edx
- add r14d,r11d
- DB 143,232,120,194,239,2
- ror r13d,14
- mov r11d,r14d
- vpxor xmm7,xmm7,xmm6
- mov r12d,r8d
- xor r13d,edx
- ror r14d,9
- xor r12d,r9d
- vpxor xmm7,xmm7,xmm5
- ror r13d,5
- xor r14d,r11d
- and r12d,edx
- vpxor xmm9,xmm9,xmm8
- xor r13d,edx
- vpsrldq xmm7,xmm7,8
- add r10d,DWORD PTR[4+rsp]
- mov esi,r11d
- ror r14d,11
- xor r12d,r9d
- vpaddd xmm0,xmm0,xmm7
- xor esi,eax
- ror r13d,6
- add r10d,r12d
- and r15d,esi
- DB 143,232,120,194,248,13
- xor r14d,r11d
- add r10d,r13d
- vpsrld xmm6,xmm0,10
- xor r15d,eax
- add ecx,r10d
- DB 143,232,120,194,239,2
- ror r14d,2
- add r10d,r15d
- vpxor xmm7,xmm7,xmm6
- mov r13d,ecx
- add r14d,r10d
- ror r13d,14
- mov r10d,r14d
- vpxor xmm7,xmm7,xmm5
- mov r12d,edx
- xor r13d,ecx
- ror r14d,9
- xor r12d,r8d
- vpslldq xmm7,xmm7,8
- ror r13d,5
- xor r14d,r10d
- and r12d,ecx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((32-128))+rdi]
- xor r13d,ecx
- vpaddd xmm0,xmm0,xmm7
- add r9d,DWORD PTR[8+rsp]
- mov r15d,r10d
- ror r14d,11
- xor r12d,r8d
- vpaddd xmm6,xmm0,XMMWORD PTR[rbp]
- xor r15d,r11d
- ror r13d,6
- add r9d,r12d
- and esi,r15d
- xor r14d,r10d
- add r9d,r13d
- xor esi,r11d
- add ebx,r9d
- ror r14d,2
- add r9d,esi
- mov r13d,ebx
- add r14d,r9d
- ror r13d,14
- mov r9d,r14d
- mov r12d,ecx
- xor r13d,ebx
- ror r14d,9
- xor r12d,edx
- ror r13d,5
- xor r14d,r9d
- and r12d,ebx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((48-128))+rdi]
- xor r13d,ebx
- add r8d,DWORD PTR[12+rsp]
- mov esi,r9d
- ror r14d,11
- xor r12d,edx
- xor esi,r10d
- ror r13d,6
- add r8d,r12d
- and r15d,esi
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- add eax,r8d
- ror r14d,2
- add r8d,r15d
- mov r13d,eax
- add r14d,r8d
- vmovdqa XMMWORD PTR[rsp],xmm6
- vpalignr xmm4,xmm2,xmm1,4
- ror r13d,14
- mov r8d,r14d
- vpalignr xmm7,xmm0,xmm3,4
- mov r12d,ebx
- xor r13d,eax
- DB 143,232,120,194,236,14
- ror r14d,9
- xor r12d,ecx
- vpsrld xmm4,xmm4,3
- ror r13d,5
- xor r14d,r8d
- vpaddd xmm1,xmm1,xmm7
- and r12d,eax
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((64-128))+rdi]
- xor r13d,eax
- add edx,DWORD PTR[16+rsp]
- mov r15d,r8d
- DB 143,232,120,194,245,11
- ror r14d,11
- xor r12d,ecx
- vpxor xmm4,xmm4,xmm5
- xor r15d,r9d
- ror r13d,6
- add edx,r12d
- and esi,r15d
- DB 143,232,120,194,248,13
- xor r14d,r8d
- add edx,r13d
- vpxor xmm4,xmm4,xmm6
- xor esi,r9d
- add r11d,edx
- vpsrld xmm6,xmm0,10
- ror r14d,2
- add edx,esi
- vpaddd xmm1,xmm1,xmm4
- mov r13d,r11d
- add r14d,edx
- DB 143,232,120,194,239,2
- ror r13d,14
- mov edx,r14d
- vpxor xmm7,xmm7,xmm6
- mov r12d,eax
- xor r13d,r11d
- ror r14d,9
- xor r12d,ebx
- vpxor xmm7,xmm7,xmm5
- ror r13d,5
- xor r14d,edx
- and r12d,r11d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((80-128))+rdi]
- xor r13d,r11d
- vpsrldq xmm7,xmm7,8
- add ecx,DWORD PTR[20+rsp]
- mov esi,edx
- ror r14d,11
- xor r12d,ebx
- vpaddd xmm1,xmm1,xmm7
- xor esi,r8d
- ror r13d,6
- add ecx,r12d
- and r15d,esi
- DB 143,232,120,194,249,13
- xor r14d,edx
- add ecx,r13d
- vpsrld xmm6,xmm1,10
- xor r15d,r8d
- add r10d,ecx
- DB 143,232,120,194,239,2
- ror r14d,2
- add ecx,r15d
- vpxor xmm7,xmm7,xmm6
- mov r13d,r10d
- add r14d,ecx
- ror r13d,14
- mov ecx,r14d
- vpxor xmm7,xmm7,xmm5
- mov r12d,r11d
- xor r13d,r10d
- ror r14d,9
- xor r12d,eax
- vpslldq xmm7,xmm7,8
- ror r13d,5
- xor r14d,ecx
- and r12d,r10d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((96-128))+rdi]
- xor r13d,r10d
- vpaddd xmm1,xmm1,xmm7
- add ebx,DWORD PTR[24+rsp]
- mov r15d,ecx
- ror r14d,11
- xor r12d,eax
- vpaddd xmm6,xmm1,XMMWORD PTR[32+rbp]
- xor r15d,edx
- ror r13d,6
- add ebx,r12d
- and esi,r15d
- xor r14d,ecx
- add ebx,r13d
- xor esi,edx
- add r9d,ebx
- ror r14d,2
- add ebx,esi
- mov r13d,r9d
- add r14d,ebx
- ror r13d,14
- mov ebx,r14d
- mov r12d,r10d
- xor r13d,r9d
- ror r14d,9
- xor r12d,r11d
- ror r13d,5
- xor r14d,ebx
- and r12d,r9d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((112-128))+rdi]
- xor r13d,r9d
- add eax,DWORD PTR[28+rsp]
- mov esi,ebx
- ror r14d,11
- xor r12d,r11d
- xor esi,ecx
- ror r13d,6
- add eax,r12d
- and r15d,esi
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- add r8d,eax
- ror r14d,2
- add eax,r15d
- mov r13d,r8d
- add r14d,eax
- vmovdqa XMMWORD PTR[16+rsp],xmm6
- vpalignr xmm4,xmm3,xmm2,4
- ror r13d,14
- mov eax,r14d
- vpalignr xmm7,xmm1,xmm0,4
- mov r12d,r9d
- xor r13d,r8d
- DB 143,232,120,194,236,14
- ror r14d,9
- xor r12d,r10d
- vpsrld xmm4,xmm4,3
- ror r13d,5
- xor r14d,eax
- vpaddd xmm2,xmm2,xmm7
- and r12d,r8d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((128-128))+rdi]
- xor r13d,r8d
- add r11d,DWORD PTR[32+rsp]
- mov r15d,eax
- DB 143,232,120,194,245,11
- ror r14d,11
- xor r12d,r10d
- vpxor xmm4,xmm4,xmm5
- xor r15d,ebx
- ror r13d,6
- add r11d,r12d
- and esi,r15d
- DB 143,232,120,194,249,13
- xor r14d,eax
- add r11d,r13d
- vpxor xmm4,xmm4,xmm6
- xor esi,ebx
- add edx,r11d
- vpsrld xmm6,xmm1,10
- ror r14d,2
- add r11d,esi
- vpaddd xmm2,xmm2,xmm4
- mov r13d,edx
- add r14d,r11d
- DB 143,232,120,194,239,2
- ror r13d,14
- mov r11d,r14d
- vpxor xmm7,xmm7,xmm6
- mov r12d,r8d
- xor r13d,edx
- ror r14d,9
- xor r12d,r9d
- vpxor xmm7,xmm7,xmm5
- ror r13d,5
- xor r14d,r11d
- and r12d,edx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((144-128))+rdi]
- xor r13d,edx
- vpsrldq xmm7,xmm7,8
- add r10d,DWORD PTR[36+rsp]
- mov esi,r11d
- ror r14d,11
- xor r12d,r9d
- vpaddd xmm2,xmm2,xmm7
- xor esi,eax
- ror r13d,6
- add r10d,r12d
- and r15d,esi
- DB 143,232,120,194,250,13
- xor r14d,r11d
- add r10d,r13d
- vpsrld xmm6,xmm2,10
- xor r15d,eax
- add ecx,r10d
- DB 143,232,120,194,239,2
- ror r14d,2
- add r10d,r15d
- vpxor xmm7,xmm7,xmm6
- mov r13d,ecx
- add r14d,r10d
- ror r13d,14
- mov r10d,r14d
- vpxor xmm7,xmm7,xmm5
- mov r12d,edx
- xor r13d,ecx
- ror r14d,9
- xor r12d,r8d
- vpslldq xmm7,xmm7,8
- ror r13d,5
- xor r14d,r10d
- and r12d,ecx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((160-128))+rdi]
- xor r13d,ecx
- vpaddd xmm2,xmm2,xmm7
- add r9d,DWORD PTR[40+rsp]
- mov r15d,r10d
- ror r14d,11
- xor r12d,r8d
- vpaddd xmm6,xmm2,XMMWORD PTR[64+rbp]
- xor r15d,r11d
- ror r13d,6
- add r9d,r12d
- and esi,r15d
- xor r14d,r10d
- add r9d,r13d
- xor esi,r11d
- add ebx,r9d
- ror r14d,2
- add r9d,esi
- mov r13d,ebx
- add r14d,r9d
- ror r13d,14
- mov r9d,r14d
- mov r12d,ecx
- xor r13d,ebx
- ror r14d,9
- xor r12d,edx
- ror r13d,5
- xor r14d,r9d
- and r12d,ebx
- vaesenclast xmm11,xmm9,xmm10
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((176-128))+rdi]
- xor r13d,ebx
- add r8d,DWORD PTR[44+rsp]
- mov esi,r9d
- ror r14d,11
- xor r12d,edx
- xor esi,r10d
- ror r13d,6
- add r8d,r12d
- and r15d,esi
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- add eax,r8d
- ror r14d,2
- add r8d,r15d
- mov r13d,eax
- add r14d,r8d
- vmovdqa XMMWORD PTR[32+rsp],xmm6
- vpalignr xmm4,xmm0,xmm3,4
- ror r13d,14
- mov r8d,r14d
- vpalignr xmm7,xmm2,xmm1,4
- mov r12d,ebx
- xor r13d,eax
- DB 143,232,120,194,236,14
- ror r14d,9
- xor r12d,ecx
- vpsrld xmm4,xmm4,3
- ror r13d,5
- xor r14d,r8d
- vpaddd xmm3,xmm3,xmm7
- and r12d,eax
- vpand xmm8,xmm11,xmm12
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((192-128))+rdi]
- xor r13d,eax
- add edx,DWORD PTR[48+rsp]
- mov r15d,r8d
- DB 143,232,120,194,245,11
- ror r14d,11
- xor r12d,ecx
- vpxor xmm4,xmm4,xmm5
- xor r15d,r9d
- ror r13d,6
- add edx,r12d
- and esi,r15d
- DB 143,232,120,194,250,13
- xor r14d,r8d
- add edx,r13d
- vpxor xmm4,xmm4,xmm6
- xor esi,r9d
- add r11d,edx
- vpsrld xmm6,xmm2,10
- ror r14d,2
- add edx,esi
- vpaddd xmm3,xmm3,xmm4
- mov r13d,r11d
- add r14d,edx
- DB 143,232,120,194,239,2
- ror r13d,14
- mov edx,r14d
- vpxor xmm7,xmm7,xmm6
- mov r12d,eax
- xor r13d,r11d
- ror r14d,9
- xor r12d,ebx
- vpxor xmm7,xmm7,xmm5
- ror r13d,5
- xor r14d,edx
- and r12d,r11d
- vaesenclast xmm11,xmm9,xmm10
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((208-128))+rdi]
- xor r13d,r11d
- vpsrldq xmm7,xmm7,8
- add ecx,DWORD PTR[52+rsp]
- mov esi,edx
- ror r14d,11
- xor r12d,ebx
- vpaddd xmm3,xmm3,xmm7
- xor esi,r8d
- ror r13d,6
- add ecx,r12d
- and r15d,esi
- DB 143,232,120,194,251,13
- xor r14d,edx
- add ecx,r13d
- vpsrld xmm6,xmm3,10
- xor r15d,r8d
- add r10d,ecx
- DB 143,232,120,194,239,2
- ror r14d,2
- add ecx,r15d
- vpxor xmm7,xmm7,xmm6
- mov r13d,r10d
- add r14d,ecx
- ror r13d,14
- mov ecx,r14d
- vpxor xmm7,xmm7,xmm5
- mov r12d,r11d
- xor r13d,r10d
- ror r14d,9
- xor r12d,eax
- vpslldq xmm7,xmm7,8
- ror r13d,5
- xor r14d,ecx
- and r12d,r10d
- vpand xmm11,xmm11,xmm13
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((224-128))+rdi]
- xor r13d,r10d
- vpaddd xmm3,xmm3,xmm7
- add ebx,DWORD PTR[56+rsp]
- mov r15d,ecx
- ror r14d,11
- xor r12d,eax
- vpaddd xmm6,xmm3,XMMWORD PTR[96+rbp]
- xor r15d,edx
- ror r13d,6
- add ebx,r12d
- and esi,r15d
- xor r14d,ecx
- add ebx,r13d
- xor esi,edx
- add r9d,ebx
- ror r14d,2
- add ebx,esi
- mov r13d,r9d
- add r14d,ebx
- ror r13d,14
- mov ebx,r14d
- mov r12d,r10d
- xor r13d,r9d
- ror r14d,9
- xor r12d,r11d
- ror r13d,5
- xor r14d,ebx
- and r12d,r9d
- vpor xmm8,xmm8,xmm11
- vaesenclast xmm11,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((0-128))+rdi]
- xor r13d,r9d
- add eax,DWORD PTR[60+rsp]
- mov esi,ebx
- ror r14d,11
- xor r12d,r11d
- xor esi,ecx
- ror r13d,6
- add eax,r12d
- and r15d,esi
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- add r8d,eax
- ror r14d,2
- add eax,r15d
- mov r13d,r8d
- add r14d,eax
- vmovdqa XMMWORD PTR[48+rsp],xmm6
- mov r12,QWORD PTR[((64+0))+rsp]
- vpand xmm11,xmm11,xmm14
- mov r15,QWORD PTR[((64+8))+rsp]
- vpor xmm8,xmm8,xmm11
- vmovdqu XMMWORD PTR[r12*1+r15],xmm8
- lea r12,QWORD PTR[16+r12]
- cmp BYTE PTR[131+rbp],0
- jne $L$xop_00_47
- vmovdqu xmm9,XMMWORD PTR[r12]
- mov QWORD PTR[((64+0))+rsp],r12
- ror r13d,14
- mov eax,r14d
- mov r12d,r9d
- xor r13d,r8d
- ror r14d,9
- xor r12d,r10d
- ror r13d,5
- xor r14d,eax
- and r12d,r8d
- vpxor xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((16-128))+rdi]
- xor r13d,r8d
- add r11d,DWORD PTR[rsp]
- mov r15d,eax
- ror r14d,11
- xor r12d,r10d
- xor r15d,ebx
- ror r13d,6
- add r11d,r12d
- and esi,r15d
- xor r14d,eax
- add r11d,r13d
- xor esi,ebx
- add edx,r11d
- ror r14d,2
- add r11d,esi
- mov r13d,edx
- add r14d,r11d
- ror r13d,14
- mov r11d,r14d
- mov r12d,r8d
- xor r13d,edx
- ror r14d,9
- xor r12d,r9d
- ror r13d,5
- xor r14d,r11d
- and r12d,edx
- vpxor xmm9,xmm9,xmm8
- xor r13d,edx
- add r10d,DWORD PTR[4+rsp]
- mov esi,r11d
- ror r14d,11
- xor r12d,r9d
- xor esi,eax
- ror r13d,6
- add r10d,r12d
- and r15d,esi
- xor r14d,r11d
- add r10d,r13d
- xor r15d,eax
- add ecx,r10d
- ror r14d,2
- add r10d,r15d
- mov r13d,ecx
- add r14d,r10d
- ror r13d,14
- mov r10d,r14d
- mov r12d,edx
- xor r13d,ecx
- ror r14d,9
- xor r12d,r8d
- ror r13d,5
- xor r14d,r10d
- and r12d,ecx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((32-128))+rdi]
- xor r13d,ecx
- add r9d,DWORD PTR[8+rsp]
- mov r15d,r10d
- ror r14d,11
- xor r12d,r8d
- xor r15d,r11d
- ror r13d,6
- add r9d,r12d
- and esi,r15d
- xor r14d,r10d
- add r9d,r13d
- xor esi,r11d
- add ebx,r9d
- ror r14d,2
- add r9d,esi
- mov r13d,ebx
- add r14d,r9d
- ror r13d,14
- mov r9d,r14d
- mov r12d,ecx
- xor r13d,ebx
- ror r14d,9
- xor r12d,edx
- ror r13d,5
- xor r14d,r9d
- and r12d,ebx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((48-128))+rdi]
- xor r13d,ebx
- add r8d,DWORD PTR[12+rsp]
- mov esi,r9d
- ror r14d,11
- xor r12d,edx
- xor esi,r10d
- ror r13d,6
- add r8d,r12d
- and r15d,esi
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- add eax,r8d
- ror r14d,2
- add r8d,r15d
- mov r13d,eax
- add r14d,r8d
- ror r13d,14
- mov r8d,r14d
- mov r12d,ebx
- xor r13d,eax
- ror r14d,9
- xor r12d,ecx
- ror r13d,5
- xor r14d,r8d
- and r12d,eax
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((64-128))+rdi]
- xor r13d,eax
- add edx,DWORD PTR[16+rsp]
- mov r15d,r8d
- ror r14d,11
- xor r12d,ecx
- xor r15d,r9d
- ror r13d,6
- add edx,r12d
- and esi,r15d
- xor r14d,r8d
- add edx,r13d
- xor esi,r9d
- add r11d,edx
- ror r14d,2
- add edx,esi
- mov r13d,r11d
- add r14d,edx
- ror r13d,14
- mov edx,r14d
- mov r12d,eax
- xor r13d,r11d
- ror r14d,9
- xor r12d,ebx
- ror r13d,5
- xor r14d,edx
- and r12d,r11d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((80-128))+rdi]
- xor r13d,r11d
- add ecx,DWORD PTR[20+rsp]
- mov esi,edx
- ror r14d,11
- xor r12d,ebx
- xor esi,r8d
- ror r13d,6
- add ecx,r12d
- and r15d,esi
- xor r14d,edx
- add ecx,r13d
- xor r15d,r8d
- add r10d,ecx
- ror r14d,2
- add ecx,r15d
- mov r13d,r10d
- add r14d,ecx
- ror r13d,14
- mov ecx,r14d
- mov r12d,r11d
- xor r13d,r10d
- ror r14d,9
- xor r12d,eax
- ror r13d,5
- xor r14d,ecx
- and r12d,r10d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((96-128))+rdi]
- xor r13d,r10d
- add ebx,DWORD PTR[24+rsp]
- mov r15d,ecx
- ror r14d,11
- xor r12d,eax
- xor r15d,edx
- ror r13d,6
- add ebx,r12d
- and esi,r15d
- xor r14d,ecx
- add ebx,r13d
- xor esi,edx
- add r9d,ebx
- ror r14d,2
- add ebx,esi
- mov r13d,r9d
- add r14d,ebx
- ror r13d,14
- mov ebx,r14d
- mov r12d,r10d
- xor r13d,r9d
- ror r14d,9
- xor r12d,r11d
- ror r13d,5
- xor r14d,ebx
- and r12d,r9d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((112-128))+rdi]
- xor r13d,r9d
- add eax,DWORD PTR[28+rsp]
- mov esi,ebx
- ror r14d,11
- xor r12d,r11d
- xor esi,ecx
- ror r13d,6
- add eax,r12d
- and r15d,esi
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- add r8d,eax
- ror r14d,2
- add eax,r15d
- mov r13d,r8d
- add r14d,eax
- ror r13d,14
- mov eax,r14d
- mov r12d,r9d
- xor r13d,r8d
- ror r14d,9
- xor r12d,r10d
- ror r13d,5
- xor r14d,eax
- and r12d,r8d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((128-128))+rdi]
- xor r13d,r8d
- add r11d,DWORD PTR[32+rsp]
- mov r15d,eax
- ror r14d,11
- xor r12d,r10d
- xor r15d,ebx
- ror r13d,6
- add r11d,r12d
- and esi,r15d
- xor r14d,eax
- add r11d,r13d
- xor esi,ebx
- add edx,r11d
- ror r14d,2
- add r11d,esi
- mov r13d,edx
- add r14d,r11d
- ror r13d,14
- mov r11d,r14d
- mov r12d,r8d
- xor r13d,edx
- ror r14d,9
- xor r12d,r9d
- ror r13d,5
- xor r14d,r11d
- and r12d,edx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((144-128))+rdi]
- xor r13d,edx
- add r10d,DWORD PTR[36+rsp]
- mov esi,r11d
- ror r14d,11
- xor r12d,r9d
- xor esi,eax
- ror r13d,6
- add r10d,r12d
- and r15d,esi
- xor r14d,r11d
- add r10d,r13d
- xor r15d,eax
- add ecx,r10d
- ror r14d,2
- add r10d,r15d
- mov r13d,ecx
- add r14d,r10d
- ror r13d,14
- mov r10d,r14d
- mov r12d,edx
- xor r13d,ecx
- ror r14d,9
- xor r12d,r8d
- ror r13d,5
- xor r14d,r10d
- and r12d,ecx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((160-128))+rdi]
- xor r13d,ecx
- add r9d,DWORD PTR[40+rsp]
- mov r15d,r10d
- ror r14d,11
- xor r12d,r8d
- xor r15d,r11d
- ror r13d,6
- add r9d,r12d
- and esi,r15d
- xor r14d,r10d
- add r9d,r13d
- xor esi,r11d
- add ebx,r9d
- ror r14d,2
- add r9d,esi
- mov r13d,ebx
- add r14d,r9d
- ror r13d,14
- mov r9d,r14d
- mov r12d,ecx
- xor r13d,ebx
- ror r14d,9
- xor r12d,edx
- ror r13d,5
- xor r14d,r9d
- and r12d,ebx
- vaesenclast xmm11,xmm9,xmm10
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((176-128))+rdi]
- xor r13d,ebx
- add r8d,DWORD PTR[44+rsp]
- mov esi,r9d
- ror r14d,11
- xor r12d,edx
- xor esi,r10d
- ror r13d,6
- add r8d,r12d
- and r15d,esi
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- add eax,r8d
- ror r14d,2
- add r8d,r15d
- mov r13d,eax
- add r14d,r8d
- ror r13d,14
- mov r8d,r14d
- mov r12d,ebx
- xor r13d,eax
- ror r14d,9
- xor r12d,ecx
- ror r13d,5
- xor r14d,r8d
- and r12d,eax
- vpand xmm8,xmm11,xmm12
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((192-128))+rdi]
- xor r13d,eax
- add edx,DWORD PTR[48+rsp]
- mov r15d,r8d
- ror r14d,11
- xor r12d,ecx
- xor r15d,r9d
- ror r13d,6
- add edx,r12d
- and esi,r15d
- xor r14d,r8d
- add edx,r13d
- xor esi,r9d
- add r11d,edx
- ror r14d,2
- add edx,esi
- mov r13d,r11d
- add r14d,edx
- ror r13d,14
- mov edx,r14d
- mov r12d,eax
- xor r13d,r11d
- ror r14d,9
- xor r12d,ebx
- ror r13d,5
- xor r14d,edx
- and r12d,r11d
- vaesenclast xmm11,xmm9,xmm10
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((208-128))+rdi]
- xor r13d,r11d
- add ecx,DWORD PTR[52+rsp]
- mov esi,edx
- ror r14d,11
- xor r12d,ebx
- xor esi,r8d
- ror r13d,6
- add ecx,r12d
- and r15d,esi
- xor r14d,edx
- add ecx,r13d
- xor r15d,r8d
- add r10d,ecx
- ror r14d,2
- add ecx,r15d
- mov r13d,r10d
- add r14d,ecx
- ror r13d,14
- mov ecx,r14d
- mov r12d,r11d
- xor r13d,r10d
- ror r14d,9
- xor r12d,eax
- ror r13d,5
- xor r14d,ecx
- and r12d,r10d
- vpand xmm11,xmm11,xmm13
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((224-128))+rdi]
- xor r13d,r10d
- add ebx,DWORD PTR[56+rsp]
- mov r15d,ecx
- ror r14d,11
- xor r12d,eax
- xor r15d,edx
- ror r13d,6
- add ebx,r12d
- and esi,r15d
- xor r14d,ecx
- add ebx,r13d
- xor esi,edx
- add r9d,ebx
- ror r14d,2
- add ebx,esi
- mov r13d,r9d
- add r14d,ebx
- ror r13d,14
- mov ebx,r14d
- mov r12d,r10d
- xor r13d,r9d
- ror r14d,9
- xor r12d,r11d
- ror r13d,5
- xor r14d,ebx
- and r12d,r9d
- vpor xmm8,xmm8,xmm11
- vaesenclast xmm11,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((0-128))+rdi]
- xor r13d,r9d
- add eax,DWORD PTR[60+rsp]
- mov esi,ebx
- ror r14d,11
- xor r12d,r11d
- xor esi,ecx
- ror r13d,6
- add eax,r12d
- and r15d,esi
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- add r8d,eax
- ror r14d,2
- add eax,r15d
- mov r13d,r8d
- add r14d,eax
- mov r12,QWORD PTR[((64+0))+rsp]
- mov r13,QWORD PTR[((64+8))+rsp]
- mov r15,QWORD PTR[((64+40))+rsp]
- mov rsi,QWORD PTR[((64+48))+rsp]
- vpand xmm11,xmm11,xmm14
- mov eax,r14d
- vpor xmm8,xmm8,xmm11
- vmovdqu XMMWORD PTR[r13*1+r12],xmm8
- lea r12,QWORD PTR[16+r12]
- add eax,DWORD PTR[r15]
- add ebx,DWORD PTR[4+r15]
- add ecx,DWORD PTR[8+r15]
- add edx,DWORD PTR[12+r15]
- add r8d,DWORD PTR[16+r15]
- add r9d,DWORD PTR[20+r15]
- add r10d,DWORD PTR[24+r15]
- add r11d,DWORD PTR[28+r15]
- cmp r12,QWORD PTR[((64+16))+rsp]
- mov DWORD PTR[r15],eax
- mov DWORD PTR[4+r15],ebx
- mov DWORD PTR[8+r15],ecx
- mov DWORD PTR[12+r15],edx
- mov DWORD PTR[16+r15],r8d
- mov DWORD PTR[20+r15],r9d
- mov DWORD PTR[24+r15],r10d
- mov DWORD PTR[28+r15],r11d
- jb $L$loop_xop
- mov r8,QWORD PTR[((64+32))+rsp]
- mov rsi,QWORD PTR[120+rsp]
- vmovdqu XMMWORD PTR[r8],xmm8
- vzeroall
- movaps xmm6,XMMWORD PTR[128+rsp]
- movaps xmm7,XMMWORD PTR[144+rsp]
- movaps xmm8,XMMWORD PTR[160+rsp]
- movaps xmm9,XMMWORD PTR[176+rsp]
- movaps xmm10,XMMWORD PTR[192+rsp]
- movaps xmm11,XMMWORD PTR[208+rsp]
- movaps xmm12,XMMWORD PTR[224+rsp]
- movaps xmm13,XMMWORD PTR[240+rsp]
- movaps xmm14,XMMWORD PTR[256+rsp]
- movaps xmm15,XMMWORD PTR[272+rsp]
- mov r15,QWORD PTR[((-48))+rsi]
- mov r14,QWORD PTR[((-40))+rsi]
- mov r13,QWORD PTR[((-32))+rsi]
- mov r12,QWORD PTR[((-24))+rsi]
- mov rbp,QWORD PTR[((-16))+rsi]
- mov rbx,QWORD PTR[((-8))+rsi]
- lea rsp,QWORD PTR[rsi]
- $L$epilogue_xop::
- mov rdi,QWORD PTR[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD PTR[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_aesni_cbc_sha256_enc_xop::
- aesni_cbc_sha256_enc_xop ENDP
- ALIGN 64
- aesni_cbc_sha256_enc_avx PROC PRIVATE
- mov QWORD PTR[8+rsp],rdi ;WIN64 prologue
- mov QWORD PTR[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_aesni_cbc_sha256_enc_avx::
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- mov r8,QWORD PTR[40+rsp]
- mov r9,QWORD PTR[48+rsp]
- $L$avx_shortcut::
- mov r10,QWORD PTR[56+rsp]
- mov rax,rsp
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- sub rsp,288
- and rsp,-64
- shl rdx,6
- sub rsi,rdi
- sub r10,rdi
- add rdx,rdi
- mov QWORD PTR[((64+8))+rsp],rsi
- mov QWORD PTR[((64+16))+rsp],rdx
- mov QWORD PTR[((64+32))+rsp],r8
- mov QWORD PTR[((64+40))+rsp],r9
- mov QWORD PTR[((64+48))+rsp],r10
- mov QWORD PTR[120+rsp],rax
- movaps XMMWORD PTR[128+rsp],xmm6
- movaps XMMWORD PTR[144+rsp],xmm7
- movaps XMMWORD PTR[160+rsp],xmm8
- movaps XMMWORD PTR[176+rsp],xmm9
- movaps XMMWORD PTR[192+rsp],xmm10
- movaps XMMWORD PTR[208+rsp],xmm11
- movaps XMMWORD PTR[224+rsp],xmm12
- movaps XMMWORD PTR[240+rsp],xmm13
- movaps XMMWORD PTR[256+rsp],xmm14
- movaps XMMWORD PTR[272+rsp],xmm15
- $L$prologue_avx::
- vzeroall
- mov r12,rdi
- lea rdi,QWORD PTR[128+rcx]
- lea r13,QWORD PTR[((K256+544))]
- mov r14d,DWORD PTR[((240-128))+rdi]
- mov r15,r9
- mov rsi,r10
- vmovdqu xmm8,XMMWORD PTR[r8]
- sub r14,9
- mov eax,DWORD PTR[r15]
- mov ebx,DWORD PTR[4+r15]
- mov ecx,DWORD PTR[8+r15]
- mov edx,DWORD PTR[12+r15]
- mov r8d,DWORD PTR[16+r15]
- mov r9d,DWORD PTR[20+r15]
- mov r10d,DWORD PTR[24+r15]
- mov r11d,DWORD PTR[28+r15]
- vmovdqa xmm14,XMMWORD PTR[r14*8+r13]
- vmovdqa xmm13,XMMWORD PTR[16+r14*8+r13]
- vmovdqa xmm12,XMMWORD PTR[32+r14*8+r13]
- vmovdqu xmm10,XMMWORD PTR[((0-128))+rdi]
- jmp $L$loop_avx
- ALIGN 16
- $L$loop_avx::
- vmovdqa xmm7,XMMWORD PTR[((K256+512))]
- vmovdqu xmm0,XMMWORD PTR[r12*1+rsi]
- vmovdqu xmm1,XMMWORD PTR[16+r12*1+rsi]
- vmovdqu xmm2,XMMWORD PTR[32+r12*1+rsi]
- vmovdqu xmm3,XMMWORD PTR[48+r12*1+rsi]
- vpshufb xmm0,xmm0,xmm7
- lea rbp,QWORD PTR[K256]
- vpshufb xmm1,xmm1,xmm7
- vpshufb xmm2,xmm2,xmm7
- vpaddd xmm4,xmm0,XMMWORD PTR[rbp]
- vpshufb xmm3,xmm3,xmm7
- vpaddd xmm5,xmm1,XMMWORD PTR[32+rbp]
- vpaddd xmm6,xmm2,XMMWORD PTR[64+rbp]
- vpaddd xmm7,xmm3,XMMWORD PTR[96+rbp]
- vmovdqa XMMWORD PTR[rsp],xmm4
- mov r14d,eax
- vmovdqa XMMWORD PTR[16+rsp],xmm5
- mov esi,ebx
- vmovdqa XMMWORD PTR[32+rsp],xmm6
- xor esi,ecx
- vmovdqa XMMWORD PTR[48+rsp],xmm7
- mov r13d,r8d
- jmp $L$avx_00_47
- ALIGN 16
- $L$avx_00_47::
- sub rbp,-16*2*4
- vmovdqu xmm9,XMMWORD PTR[r12]
- mov QWORD PTR[((64+0))+rsp],r12
- vpalignr xmm4,xmm1,xmm0,4
- shrd r13d,r13d,14
- mov eax,r14d
- mov r12d,r9d
- vpalignr xmm7,xmm3,xmm2,4
- xor r13d,r8d
- shrd r14d,r14d,9
- xor r12d,r10d
- vpsrld xmm6,xmm4,7
- shrd r13d,r13d,5
- xor r14d,eax
- and r12d,r8d
- vpaddd xmm0,xmm0,xmm7
- vpxor xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((16-128))+rdi]
- xor r13d,r8d
- add r11d,DWORD PTR[rsp]
- mov r15d,eax
- vpsrld xmm7,xmm4,3
- shrd r14d,r14d,11
- xor r12d,r10d
- xor r15d,ebx
- vpslld xmm5,xmm4,14
- shrd r13d,r13d,6
- add r11d,r12d
- and esi,r15d
- vpxor xmm4,xmm7,xmm6
- xor r14d,eax
- add r11d,r13d
- xor esi,ebx
- vpshufd xmm7,xmm3,250
- add edx,r11d
- shrd r14d,r14d,2
- add r11d,esi
- vpsrld xmm6,xmm6,11
- mov r13d,edx
- add r14d,r11d
- shrd r13d,r13d,14
- vpxor xmm4,xmm4,xmm5
- mov r11d,r14d
- mov r12d,r8d
- xor r13d,edx
- vpslld xmm5,xmm5,11
- shrd r14d,r14d,9
- xor r12d,r9d
- shrd r13d,r13d,5
- vpxor xmm4,xmm4,xmm6
- xor r14d,r11d
- and r12d,edx
- vpxor xmm9,xmm9,xmm8
- xor r13d,edx
- vpsrld xmm6,xmm7,10
- add r10d,DWORD PTR[4+rsp]
- mov esi,r11d
- shrd r14d,r14d,11
- vpxor xmm4,xmm4,xmm5
- xor r12d,r9d
- xor esi,eax
- shrd r13d,r13d,6
- vpsrlq xmm7,xmm7,17
- add r10d,r12d
- and r15d,esi
- xor r14d,r11d
- vpaddd xmm0,xmm0,xmm4
- add r10d,r13d
- xor r15d,eax
- add ecx,r10d
- vpxor xmm6,xmm6,xmm7
- shrd r14d,r14d,2
- add r10d,r15d
- mov r13d,ecx
- vpsrlq xmm7,xmm7,2
- add r14d,r10d
- shrd r13d,r13d,14
- mov r10d,r14d
- vpxor xmm6,xmm6,xmm7
- mov r12d,edx
- xor r13d,ecx
- shrd r14d,r14d,9
- vpshufd xmm6,xmm6,132
- xor r12d,r8d
- shrd r13d,r13d,5
- xor r14d,r10d
- vpsrldq xmm6,xmm6,8
- and r12d,ecx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((32-128))+rdi]
- xor r13d,ecx
- add r9d,DWORD PTR[8+rsp]
- vpaddd xmm0,xmm0,xmm6
- mov r15d,r10d
- shrd r14d,r14d,11
- xor r12d,r8d
- vpshufd xmm7,xmm0,80
- xor r15d,r11d
- shrd r13d,r13d,6
- add r9d,r12d
- vpsrld xmm6,xmm7,10
- and esi,r15d
- xor r14d,r10d
- add r9d,r13d
- vpsrlq xmm7,xmm7,17
- xor esi,r11d
- add ebx,r9d
- shrd r14d,r14d,2
- vpxor xmm6,xmm6,xmm7
- add r9d,esi
- mov r13d,ebx
- add r14d,r9d
- vpsrlq xmm7,xmm7,2
- shrd r13d,r13d,14
- mov r9d,r14d
- mov r12d,ecx
- vpxor xmm6,xmm6,xmm7
- xor r13d,ebx
- shrd r14d,r14d,9
- xor r12d,edx
- vpshufd xmm6,xmm6,232
- shrd r13d,r13d,5
- xor r14d,r9d
- and r12d,ebx
- vpslldq xmm6,xmm6,8
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((48-128))+rdi]
- xor r13d,ebx
- add r8d,DWORD PTR[12+rsp]
- mov esi,r9d
- vpaddd xmm0,xmm0,xmm6
- shrd r14d,r14d,11
- xor r12d,edx
- xor esi,r10d
- vpaddd xmm6,xmm0,XMMWORD PTR[rbp]
- shrd r13d,r13d,6
- add r8d,r12d
- and r15d,esi
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- add eax,r8d
- shrd r14d,r14d,2
- add r8d,r15d
- mov r13d,eax
- add r14d,r8d
- vmovdqa XMMWORD PTR[rsp],xmm6
- vpalignr xmm4,xmm2,xmm1,4
- shrd r13d,r13d,14
- mov r8d,r14d
- mov r12d,ebx
- vpalignr xmm7,xmm0,xmm3,4
- xor r13d,eax
- shrd r14d,r14d,9
- xor r12d,ecx
- vpsrld xmm6,xmm4,7
- shrd r13d,r13d,5
- xor r14d,r8d
- and r12d,eax
- vpaddd xmm1,xmm1,xmm7
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((64-128))+rdi]
- xor r13d,eax
- add edx,DWORD PTR[16+rsp]
- mov r15d,r8d
- vpsrld xmm7,xmm4,3
- shrd r14d,r14d,11
- xor r12d,ecx
- xor r15d,r9d
- vpslld xmm5,xmm4,14
- shrd r13d,r13d,6
- add edx,r12d
- and esi,r15d
- vpxor xmm4,xmm7,xmm6
- xor r14d,r8d
- add edx,r13d
- xor esi,r9d
- vpshufd xmm7,xmm0,250
- add r11d,edx
- shrd r14d,r14d,2
- add edx,esi
- vpsrld xmm6,xmm6,11
- mov r13d,r11d
- add r14d,edx
- shrd r13d,r13d,14
- vpxor xmm4,xmm4,xmm5
- mov edx,r14d
- mov r12d,eax
- xor r13d,r11d
- vpslld xmm5,xmm5,11
- shrd r14d,r14d,9
- xor r12d,ebx
- shrd r13d,r13d,5
- vpxor xmm4,xmm4,xmm6
- xor r14d,edx
- and r12d,r11d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((80-128))+rdi]
- xor r13d,r11d
- vpsrld xmm6,xmm7,10
- add ecx,DWORD PTR[20+rsp]
- mov esi,edx
- shrd r14d,r14d,11
- vpxor xmm4,xmm4,xmm5
- xor r12d,ebx
- xor esi,r8d
- shrd r13d,r13d,6
- vpsrlq xmm7,xmm7,17
- add ecx,r12d
- and r15d,esi
- xor r14d,edx
- vpaddd xmm1,xmm1,xmm4
- add ecx,r13d
- xor r15d,r8d
- add r10d,ecx
- vpxor xmm6,xmm6,xmm7
- shrd r14d,r14d,2
- add ecx,r15d
- mov r13d,r10d
- vpsrlq xmm7,xmm7,2
- add r14d,ecx
- shrd r13d,r13d,14
- mov ecx,r14d
- vpxor xmm6,xmm6,xmm7
- mov r12d,r11d
- xor r13d,r10d
- shrd r14d,r14d,9
- vpshufd xmm6,xmm6,132
- xor r12d,eax
- shrd r13d,r13d,5
- xor r14d,ecx
- vpsrldq xmm6,xmm6,8
- and r12d,r10d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((96-128))+rdi]
- xor r13d,r10d
- add ebx,DWORD PTR[24+rsp]
- vpaddd xmm1,xmm1,xmm6
- mov r15d,ecx
- shrd r14d,r14d,11
- xor r12d,eax
- vpshufd xmm7,xmm1,80
- xor r15d,edx
- shrd r13d,r13d,6
- add ebx,r12d
- vpsrld xmm6,xmm7,10
- and esi,r15d
- xor r14d,ecx
- add ebx,r13d
- vpsrlq xmm7,xmm7,17
- xor esi,edx
- add r9d,ebx
- shrd r14d,r14d,2
- vpxor xmm6,xmm6,xmm7
- add ebx,esi
- mov r13d,r9d
- add r14d,ebx
- vpsrlq xmm7,xmm7,2
- shrd r13d,r13d,14
- mov ebx,r14d
- mov r12d,r10d
- vpxor xmm6,xmm6,xmm7
- xor r13d,r9d
- shrd r14d,r14d,9
- xor r12d,r11d
- vpshufd xmm6,xmm6,232
- shrd r13d,r13d,5
- xor r14d,ebx
- and r12d,r9d
- vpslldq xmm6,xmm6,8
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((112-128))+rdi]
- xor r13d,r9d
- add eax,DWORD PTR[28+rsp]
- mov esi,ebx
- vpaddd xmm1,xmm1,xmm6
- shrd r14d,r14d,11
- xor r12d,r11d
- xor esi,ecx
- vpaddd xmm6,xmm1,XMMWORD PTR[32+rbp]
- shrd r13d,r13d,6
- add eax,r12d
- and r15d,esi
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- add r8d,eax
- shrd r14d,r14d,2
- add eax,r15d
- mov r13d,r8d
- add r14d,eax
- vmovdqa XMMWORD PTR[16+rsp],xmm6
- vpalignr xmm4,xmm3,xmm2,4
- shrd r13d,r13d,14
- mov eax,r14d
- mov r12d,r9d
- vpalignr xmm7,xmm1,xmm0,4
- xor r13d,r8d
- shrd r14d,r14d,9
- xor r12d,r10d
- vpsrld xmm6,xmm4,7
- shrd r13d,r13d,5
- xor r14d,eax
- and r12d,r8d
- vpaddd xmm2,xmm2,xmm7
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((128-128))+rdi]
- xor r13d,r8d
- add r11d,DWORD PTR[32+rsp]
- mov r15d,eax
- vpsrld xmm7,xmm4,3
- shrd r14d,r14d,11
- xor r12d,r10d
- xor r15d,ebx
- vpslld xmm5,xmm4,14
- shrd r13d,r13d,6
- add r11d,r12d
- and esi,r15d
- vpxor xmm4,xmm7,xmm6
- xor r14d,eax
- add r11d,r13d
- xor esi,ebx
- vpshufd xmm7,xmm1,250
- add edx,r11d
- shrd r14d,r14d,2
- add r11d,esi
- vpsrld xmm6,xmm6,11
- mov r13d,edx
- add r14d,r11d
- shrd r13d,r13d,14
- vpxor xmm4,xmm4,xmm5
- mov r11d,r14d
- mov r12d,r8d
- xor r13d,edx
- vpslld xmm5,xmm5,11
- shrd r14d,r14d,9
- xor r12d,r9d
- shrd r13d,r13d,5
- vpxor xmm4,xmm4,xmm6
- xor r14d,r11d
- and r12d,edx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((144-128))+rdi]
- xor r13d,edx
- vpsrld xmm6,xmm7,10
- add r10d,DWORD PTR[36+rsp]
- mov esi,r11d
- shrd r14d,r14d,11
- vpxor xmm4,xmm4,xmm5
- xor r12d,r9d
- xor esi,eax
- shrd r13d,r13d,6
- vpsrlq xmm7,xmm7,17
- add r10d,r12d
- and r15d,esi
- xor r14d,r11d
- vpaddd xmm2,xmm2,xmm4
- add r10d,r13d
- xor r15d,eax
- add ecx,r10d
- vpxor xmm6,xmm6,xmm7
- shrd r14d,r14d,2
- add r10d,r15d
- mov r13d,ecx
- vpsrlq xmm7,xmm7,2
- add r14d,r10d
- shrd r13d,r13d,14
- mov r10d,r14d
- vpxor xmm6,xmm6,xmm7
- mov r12d,edx
- xor r13d,ecx
- shrd r14d,r14d,9
- vpshufd xmm6,xmm6,132
- xor r12d,r8d
- shrd r13d,r13d,5
- xor r14d,r10d
- vpsrldq xmm6,xmm6,8
- and r12d,ecx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((160-128))+rdi]
- xor r13d,ecx
- add r9d,DWORD PTR[40+rsp]
- vpaddd xmm2,xmm2,xmm6
- mov r15d,r10d
- shrd r14d,r14d,11
- xor r12d,r8d
- vpshufd xmm7,xmm2,80
- xor r15d,r11d
- shrd r13d,r13d,6
- add r9d,r12d
- vpsrld xmm6,xmm7,10
- and esi,r15d
- xor r14d,r10d
- add r9d,r13d
- vpsrlq xmm7,xmm7,17
- xor esi,r11d
- add ebx,r9d
- shrd r14d,r14d,2
- vpxor xmm6,xmm6,xmm7
- add r9d,esi
- mov r13d,ebx
- add r14d,r9d
- vpsrlq xmm7,xmm7,2
- shrd r13d,r13d,14
- mov r9d,r14d
- mov r12d,ecx
- vpxor xmm6,xmm6,xmm7
- xor r13d,ebx
- shrd r14d,r14d,9
- xor r12d,edx
- vpshufd xmm6,xmm6,232
- shrd r13d,r13d,5
- xor r14d,r9d
- and r12d,ebx
- vpslldq xmm6,xmm6,8
- vaesenclast xmm11,xmm9,xmm10
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((176-128))+rdi]
- xor r13d,ebx
- add r8d,DWORD PTR[44+rsp]
- mov esi,r9d
- vpaddd xmm2,xmm2,xmm6
- shrd r14d,r14d,11
- xor r12d,edx
- xor esi,r10d
- vpaddd xmm6,xmm2,XMMWORD PTR[64+rbp]
- shrd r13d,r13d,6
- add r8d,r12d
- and r15d,esi
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- add eax,r8d
- shrd r14d,r14d,2
- add r8d,r15d
- mov r13d,eax
- add r14d,r8d
- vmovdqa XMMWORD PTR[32+rsp],xmm6
- vpalignr xmm4,xmm0,xmm3,4
- shrd r13d,r13d,14
- mov r8d,r14d
- mov r12d,ebx
- vpalignr xmm7,xmm2,xmm1,4
- xor r13d,eax
- shrd r14d,r14d,9
- xor r12d,ecx
- vpsrld xmm6,xmm4,7
- shrd r13d,r13d,5
- xor r14d,r8d
- and r12d,eax
- vpaddd xmm3,xmm3,xmm7
- vpand xmm8,xmm11,xmm12
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((192-128))+rdi]
- xor r13d,eax
- add edx,DWORD PTR[48+rsp]
- mov r15d,r8d
- vpsrld xmm7,xmm4,3
- shrd r14d,r14d,11
- xor r12d,ecx
- xor r15d,r9d
- vpslld xmm5,xmm4,14
- shrd r13d,r13d,6
- add edx,r12d
- and esi,r15d
- vpxor xmm4,xmm7,xmm6
- xor r14d,r8d
- add edx,r13d
- xor esi,r9d
- vpshufd xmm7,xmm2,250
- add r11d,edx
- shrd r14d,r14d,2
- add edx,esi
- vpsrld xmm6,xmm6,11
- mov r13d,r11d
- add r14d,edx
- shrd r13d,r13d,14
- vpxor xmm4,xmm4,xmm5
- mov edx,r14d
- mov r12d,eax
- xor r13d,r11d
- vpslld xmm5,xmm5,11
- shrd r14d,r14d,9
- xor r12d,ebx
- shrd r13d,r13d,5
- vpxor xmm4,xmm4,xmm6
- xor r14d,edx
- and r12d,r11d
- vaesenclast xmm11,xmm9,xmm10
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((208-128))+rdi]
- xor r13d,r11d
- vpsrld xmm6,xmm7,10
- add ecx,DWORD PTR[52+rsp]
- mov esi,edx
- shrd r14d,r14d,11
- vpxor xmm4,xmm4,xmm5
- xor r12d,ebx
- xor esi,r8d
- shrd r13d,r13d,6
- vpsrlq xmm7,xmm7,17
- add ecx,r12d
- and r15d,esi
- xor r14d,edx
- vpaddd xmm3,xmm3,xmm4
- add ecx,r13d
- xor r15d,r8d
- add r10d,ecx
- vpxor xmm6,xmm6,xmm7
- shrd r14d,r14d,2
- add ecx,r15d
- mov r13d,r10d
- vpsrlq xmm7,xmm7,2
- add r14d,ecx
- shrd r13d,r13d,14
- mov ecx,r14d
- vpxor xmm6,xmm6,xmm7
- mov r12d,r11d
- xor r13d,r10d
- shrd r14d,r14d,9
- vpshufd xmm6,xmm6,132
- xor r12d,eax
- shrd r13d,r13d,5
- xor r14d,ecx
- vpsrldq xmm6,xmm6,8
- and r12d,r10d
- vpand xmm11,xmm11,xmm13
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((224-128))+rdi]
- xor r13d,r10d
- add ebx,DWORD PTR[56+rsp]
- vpaddd xmm3,xmm3,xmm6
- mov r15d,ecx
- shrd r14d,r14d,11
- xor r12d,eax
- vpshufd xmm7,xmm3,80
- xor r15d,edx
- shrd r13d,r13d,6
- add ebx,r12d
- vpsrld xmm6,xmm7,10
- and esi,r15d
- xor r14d,ecx
- add ebx,r13d
- vpsrlq xmm7,xmm7,17
- xor esi,edx
- add r9d,ebx
- shrd r14d,r14d,2
- vpxor xmm6,xmm6,xmm7
- add ebx,esi
- mov r13d,r9d
- add r14d,ebx
- vpsrlq xmm7,xmm7,2
- shrd r13d,r13d,14
- mov ebx,r14d
- mov r12d,r10d
- vpxor xmm6,xmm6,xmm7
- xor r13d,r9d
- shrd r14d,r14d,9
- xor r12d,r11d
- vpshufd xmm6,xmm6,232
- shrd r13d,r13d,5
- xor r14d,ebx
- and r12d,r9d
- vpslldq xmm6,xmm6,8
- vpor xmm8,xmm8,xmm11
- vaesenclast xmm11,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((0-128))+rdi]
- xor r13d,r9d
- add eax,DWORD PTR[60+rsp]
- mov esi,ebx
- vpaddd xmm3,xmm3,xmm6
- shrd r14d,r14d,11
- xor r12d,r11d
- xor esi,ecx
- vpaddd xmm6,xmm3,XMMWORD PTR[96+rbp]
- shrd r13d,r13d,6
- add eax,r12d
- and r15d,esi
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- add r8d,eax
- shrd r14d,r14d,2
- add eax,r15d
- mov r13d,r8d
- add r14d,eax
- vmovdqa XMMWORD PTR[48+rsp],xmm6
- mov r12,QWORD PTR[((64+0))+rsp]
- vpand xmm11,xmm11,xmm14
- mov r15,QWORD PTR[((64+8))+rsp]
- vpor xmm8,xmm8,xmm11
- vmovdqu XMMWORD PTR[r12*1+r15],xmm8
- lea r12,QWORD PTR[16+r12]
- cmp BYTE PTR[131+rbp],0
- jne $L$avx_00_47
- vmovdqu xmm9,XMMWORD PTR[r12]
- mov QWORD PTR[((64+0))+rsp],r12
- shrd r13d,r13d,14
- mov eax,r14d
- mov r12d,r9d
- xor r13d,r8d
- shrd r14d,r14d,9
- xor r12d,r10d
- shrd r13d,r13d,5
- xor r14d,eax
- and r12d,r8d
- vpxor xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((16-128))+rdi]
- xor r13d,r8d
- add r11d,DWORD PTR[rsp]
- mov r15d,eax
- shrd r14d,r14d,11
- xor r12d,r10d
- xor r15d,ebx
- shrd r13d,r13d,6
- add r11d,r12d
- and esi,r15d
- xor r14d,eax
- add r11d,r13d
- xor esi,ebx
- add edx,r11d
- shrd r14d,r14d,2
- add r11d,esi
- mov r13d,edx
- add r14d,r11d
- shrd r13d,r13d,14
- mov r11d,r14d
- mov r12d,r8d
- xor r13d,edx
- shrd r14d,r14d,9
- xor r12d,r9d
- shrd r13d,r13d,5
- xor r14d,r11d
- and r12d,edx
- vpxor xmm9,xmm9,xmm8
- xor r13d,edx
- add r10d,DWORD PTR[4+rsp]
- mov esi,r11d
- shrd r14d,r14d,11
- xor r12d,r9d
- xor esi,eax
- shrd r13d,r13d,6
- add r10d,r12d
- and r15d,esi
- xor r14d,r11d
- add r10d,r13d
- xor r15d,eax
- add ecx,r10d
- shrd r14d,r14d,2
- add r10d,r15d
- mov r13d,ecx
- add r14d,r10d
- shrd r13d,r13d,14
- mov r10d,r14d
- mov r12d,edx
- xor r13d,ecx
- shrd r14d,r14d,9
- xor r12d,r8d
- shrd r13d,r13d,5
- xor r14d,r10d
- and r12d,ecx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((32-128))+rdi]
- xor r13d,ecx
- add r9d,DWORD PTR[8+rsp]
- mov r15d,r10d
- shrd r14d,r14d,11
- xor r12d,r8d
- xor r15d,r11d
- shrd r13d,r13d,6
- add r9d,r12d
- and esi,r15d
- xor r14d,r10d
- add r9d,r13d
- xor esi,r11d
- add ebx,r9d
- shrd r14d,r14d,2
- add r9d,esi
- mov r13d,ebx
- add r14d,r9d
- shrd r13d,r13d,14
- mov r9d,r14d
- mov r12d,ecx
- xor r13d,ebx
- shrd r14d,r14d,9
- xor r12d,edx
- shrd r13d,r13d,5
- xor r14d,r9d
- and r12d,ebx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((48-128))+rdi]
- xor r13d,ebx
- add r8d,DWORD PTR[12+rsp]
- mov esi,r9d
- shrd r14d,r14d,11
- xor r12d,edx
- xor esi,r10d
- shrd r13d,r13d,6
- add r8d,r12d
- and r15d,esi
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- add eax,r8d
- shrd r14d,r14d,2
- add r8d,r15d
- mov r13d,eax
- add r14d,r8d
- shrd r13d,r13d,14
- mov r8d,r14d
- mov r12d,ebx
- xor r13d,eax
- shrd r14d,r14d,9
- xor r12d,ecx
- shrd r13d,r13d,5
- xor r14d,r8d
- and r12d,eax
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((64-128))+rdi]
- xor r13d,eax
- add edx,DWORD PTR[16+rsp]
- mov r15d,r8d
- shrd r14d,r14d,11
- xor r12d,ecx
- xor r15d,r9d
- shrd r13d,r13d,6
- add edx,r12d
- and esi,r15d
- xor r14d,r8d
- add edx,r13d
- xor esi,r9d
- add r11d,edx
- shrd r14d,r14d,2
- add edx,esi
- mov r13d,r11d
- add r14d,edx
- shrd r13d,r13d,14
- mov edx,r14d
- mov r12d,eax
- xor r13d,r11d
- shrd r14d,r14d,9
- xor r12d,ebx
- shrd r13d,r13d,5
- xor r14d,edx
- and r12d,r11d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((80-128))+rdi]
- xor r13d,r11d
- add ecx,DWORD PTR[20+rsp]
- mov esi,edx
- shrd r14d,r14d,11
- xor r12d,ebx
- xor esi,r8d
- shrd r13d,r13d,6
- add ecx,r12d
- and r15d,esi
- xor r14d,edx
- add ecx,r13d
- xor r15d,r8d
- add r10d,ecx
- shrd r14d,r14d,2
- add ecx,r15d
- mov r13d,r10d
- add r14d,ecx
- shrd r13d,r13d,14
- mov ecx,r14d
- mov r12d,r11d
- xor r13d,r10d
- shrd r14d,r14d,9
- xor r12d,eax
- shrd r13d,r13d,5
- xor r14d,ecx
- and r12d,r10d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((96-128))+rdi]
- xor r13d,r10d
- add ebx,DWORD PTR[24+rsp]
- mov r15d,ecx
- shrd r14d,r14d,11
- xor r12d,eax
- xor r15d,edx
- shrd r13d,r13d,6
- add ebx,r12d
- and esi,r15d
- xor r14d,ecx
- add ebx,r13d
- xor esi,edx
- add r9d,ebx
- shrd r14d,r14d,2
- add ebx,esi
- mov r13d,r9d
- add r14d,ebx
- shrd r13d,r13d,14
- mov ebx,r14d
- mov r12d,r10d
- xor r13d,r9d
- shrd r14d,r14d,9
- xor r12d,r11d
- shrd r13d,r13d,5
- xor r14d,ebx
- and r12d,r9d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((112-128))+rdi]
- xor r13d,r9d
- add eax,DWORD PTR[28+rsp]
- mov esi,ebx
- shrd r14d,r14d,11
- xor r12d,r11d
- xor esi,ecx
- shrd r13d,r13d,6
- add eax,r12d
- and r15d,esi
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- add r8d,eax
- shrd r14d,r14d,2
- add eax,r15d
- mov r13d,r8d
- add r14d,eax
- shrd r13d,r13d,14
- mov eax,r14d
- mov r12d,r9d
- xor r13d,r8d
- shrd r14d,r14d,9
- xor r12d,r10d
- shrd r13d,r13d,5
- xor r14d,eax
- and r12d,r8d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((128-128))+rdi]
- xor r13d,r8d
- add r11d,DWORD PTR[32+rsp]
- mov r15d,eax
- shrd r14d,r14d,11
- xor r12d,r10d
- xor r15d,ebx
- shrd r13d,r13d,6
- add r11d,r12d
- and esi,r15d
- xor r14d,eax
- add r11d,r13d
- xor esi,ebx
- add edx,r11d
- shrd r14d,r14d,2
- add r11d,esi
- mov r13d,edx
- add r14d,r11d
- shrd r13d,r13d,14
- mov r11d,r14d
- mov r12d,r8d
- xor r13d,edx
- shrd r14d,r14d,9
- xor r12d,r9d
- shrd r13d,r13d,5
- xor r14d,r11d
- and r12d,edx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((144-128))+rdi]
- xor r13d,edx
- add r10d,DWORD PTR[36+rsp]
- mov esi,r11d
- shrd r14d,r14d,11
- xor r12d,r9d
- xor esi,eax
- shrd r13d,r13d,6
- add r10d,r12d
- and r15d,esi
- xor r14d,r11d
- add r10d,r13d
- xor r15d,eax
- add ecx,r10d
- shrd r14d,r14d,2
- add r10d,r15d
- mov r13d,ecx
- add r14d,r10d
- shrd r13d,r13d,14
- mov r10d,r14d
- mov r12d,edx
- xor r13d,ecx
- shrd r14d,r14d,9
- xor r12d,r8d
- shrd r13d,r13d,5
- xor r14d,r10d
- and r12d,ecx
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((160-128))+rdi]
- xor r13d,ecx
- add r9d,DWORD PTR[40+rsp]
- mov r15d,r10d
- shrd r14d,r14d,11
- xor r12d,r8d
- xor r15d,r11d
- shrd r13d,r13d,6
- add r9d,r12d
- and esi,r15d
- xor r14d,r10d
- add r9d,r13d
- xor esi,r11d
- add ebx,r9d
- shrd r14d,r14d,2
- add r9d,esi
- mov r13d,ebx
- add r14d,r9d
- shrd r13d,r13d,14
- mov r9d,r14d
- mov r12d,ecx
- xor r13d,ebx
- shrd r14d,r14d,9
- xor r12d,edx
- shrd r13d,r13d,5
- xor r14d,r9d
- and r12d,ebx
- vaesenclast xmm11,xmm9,xmm10
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((176-128))+rdi]
- xor r13d,ebx
- add r8d,DWORD PTR[44+rsp]
- mov esi,r9d
- shrd r14d,r14d,11
- xor r12d,edx
- xor esi,r10d
- shrd r13d,r13d,6
- add r8d,r12d
- and r15d,esi
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- add eax,r8d
- shrd r14d,r14d,2
- add r8d,r15d
- mov r13d,eax
- add r14d,r8d
- shrd r13d,r13d,14
- mov r8d,r14d
- mov r12d,ebx
- xor r13d,eax
- shrd r14d,r14d,9
- xor r12d,ecx
- shrd r13d,r13d,5
- xor r14d,r8d
- and r12d,eax
- vpand xmm8,xmm11,xmm12
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((192-128))+rdi]
- xor r13d,eax
- add edx,DWORD PTR[48+rsp]
- mov r15d,r8d
- shrd r14d,r14d,11
- xor r12d,ecx
- xor r15d,r9d
- shrd r13d,r13d,6
- add edx,r12d
- and esi,r15d
- xor r14d,r8d
- add edx,r13d
- xor esi,r9d
- add r11d,edx
- shrd r14d,r14d,2
- add edx,esi
- mov r13d,r11d
- add r14d,edx
- shrd r13d,r13d,14
- mov edx,r14d
- mov r12d,eax
- xor r13d,r11d
- shrd r14d,r14d,9
- xor r12d,ebx
- shrd r13d,r13d,5
- xor r14d,edx
- and r12d,r11d
- vaesenclast xmm11,xmm9,xmm10
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((208-128))+rdi]
- xor r13d,r11d
- add ecx,DWORD PTR[52+rsp]
- mov esi,edx
- shrd r14d,r14d,11
- xor r12d,ebx
- xor esi,r8d
- shrd r13d,r13d,6
- add ecx,r12d
- and r15d,esi
- xor r14d,edx
- add ecx,r13d
- xor r15d,r8d
- add r10d,ecx
- shrd r14d,r14d,2
- add ecx,r15d
- mov r13d,r10d
- add r14d,ecx
- shrd r13d,r13d,14
- mov ecx,r14d
- mov r12d,r11d
- xor r13d,r10d
- shrd r14d,r14d,9
- xor r12d,eax
- shrd r13d,r13d,5
- xor r14d,ecx
- and r12d,r10d
- vpand xmm11,xmm11,xmm13
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((224-128))+rdi]
- xor r13d,r10d
- add ebx,DWORD PTR[56+rsp]
- mov r15d,ecx
- shrd r14d,r14d,11
- xor r12d,eax
- xor r15d,edx
- shrd r13d,r13d,6
- add ebx,r12d
- and esi,r15d
- xor r14d,ecx
- add ebx,r13d
- xor esi,edx
- add r9d,ebx
- shrd r14d,r14d,2
- add ebx,esi
- mov r13d,r9d
- add r14d,ebx
- shrd r13d,r13d,14
- mov ebx,r14d
- mov r12d,r10d
- xor r13d,r9d
- shrd r14d,r14d,9
- xor r12d,r11d
- shrd r13d,r13d,5
- xor r14d,ebx
- and r12d,r9d
- vpor xmm8,xmm8,xmm11
- vaesenclast xmm11,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((0-128))+rdi]
- xor r13d,r9d
- add eax,DWORD PTR[60+rsp]
- mov esi,ebx
- shrd r14d,r14d,11
- xor r12d,r11d
- xor esi,ecx
- shrd r13d,r13d,6
- add eax,r12d
- and r15d,esi
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- add r8d,eax
- shrd r14d,r14d,2
- add eax,r15d
- mov r13d,r8d
- add r14d,eax
- mov r12,QWORD PTR[((64+0))+rsp]
- mov r13,QWORD PTR[((64+8))+rsp]
- mov r15,QWORD PTR[((64+40))+rsp]
- mov rsi,QWORD PTR[((64+48))+rsp]
- vpand xmm11,xmm11,xmm14
- mov eax,r14d
- vpor xmm8,xmm8,xmm11
- vmovdqu XMMWORD PTR[r13*1+r12],xmm8
- lea r12,QWORD PTR[16+r12]
- add eax,DWORD PTR[r15]
- add ebx,DWORD PTR[4+r15]
- add ecx,DWORD PTR[8+r15]
- add edx,DWORD PTR[12+r15]
- add r8d,DWORD PTR[16+r15]
- add r9d,DWORD PTR[20+r15]
- add r10d,DWORD PTR[24+r15]
- add r11d,DWORD PTR[28+r15]
- cmp r12,QWORD PTR[((64+16))+rsp]
- mov DWORD PTR[r15],eax
- mov DWORD PTR[4+r15],ebx
- mov DWORD PTR[8+r15],ecx
- mov DWORD PTR[12+r15],edx
- mov DWORD PTR[16+r15],r8d
- mov DWORD PTR[20+r15],r9d
- mov DWORD PTR[24+r15],r10d
- mov DWORD PTR[28+r15],r11d
- jb $L$loop_avx
- mov r8,QWORD PTR[((64+32))+rsp]
- mov rsi,QWORD PTR[120+rsp]
- vmovdqu XMMWORD PTR[r8],xmm8
- vzeroall
- movaps xmm6,XMMWORD PTR[128+rsp]
- movaps xmm7,XMMWORD PTR[144+rsp]
- movaps xmm8,XMMWORD PTR[160+rsp]
- movaps xmm9,XMMWORD PTR[176+rsp]
- movaps xmm10,XMMWORD PTR[192+rsp]
- movaps xmm11,XMMWORD PTR[208+rsp]
- movaps xmm12,XMMWORD PTR[224+rsp]
- movaps xmm13,XMMWORD PTR[240+rsp]
- movaps xmm14,XMMWORD PTR[256+rsp]
- movaps xmm15,XMMWORD PTR[272+rsp]
- mov r15,QWORD PTR[((-48))+rsi]
- mov r14,QWORD PTR[((-40))+rsi]
- mov r13,QWORD PTR[((-32))+rsi]
- mov r12,QWORD PTR[((-24))+rsi]
- mov rbp,QWORD PTR[((-16))+rsi]
- mov rbx,QWORD PTR[((-8))+rsi]
- lea rsp,QWORD PTR[rsi]
- $L$epilogue_avx::
- mov rdi,QWORD PTR[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD PTR[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_aesni_cbc_sha256_enc_avx::
- aesni_cbc_sha256_enc_avx ENDP
- ALIGN 64
- aesni_cbc_sha256_enc_avx2 PROC PRIVATE
- mov QWORD PTR[8+rsp],rdi ;WIN64 prologue
- mov QWORD PTR[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_aesni_cbc_sha256_enc_avx2::
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- mov r8,QWORD PTR[40+rsp]
- mov r9,QWORD PTR[48+rsp]
- $L$avx2_shortcut::
- mov r10,QWORD PTR[56+rsp]
- mov rax,rsp
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- sub rsp,736
- and rsp,-256*4
- add rsp,448
- shl rdx,6
- sub rsi,rdi
- sub r10,rdi
- add rdx,rdi
- mov QWORD PTR[((64+16))+rsp],rdx
- mov QWORD PTR[((64+32))+rsp],r8
- mov QWORD PTR[((64+40))+rsp],r9
- mov QWORD PTR[((64+48))+rsp],r10
- mov QWORD PTR[120+rsp],rax
- movaps XMMWORD PTR[128+rsp],xmm6
- movaps XMMWORD PTR[144+rsp],xmm7
- movaps XMMWORD PTR[160+rsp],xmm8
- movaps XMMWORD PTR[176+rsp],xmm9
- movaps XMMWORD PTR[192+rsp],xmm10
- movaps XMMWORD PTR[208+rsp],xmm11
- movaps XMMWORD PTR[224+rsp],xmm12
- movaps XMMWORD PTR[240+rsp],xmm13
- movaps XMMWORD PTR[256+rsp],xmm14
- movaps XMMWORD PTR[272+rsp],xmm15
- $L$prologue_avx2::
- vzeroall
- mov r13,rdi
- vpinsrq xmm15,xmm15,rsi,1
- lea rdi,QWORD PTR[128+rcx]
- lea r12,QWORD PTR[((K256+544))]
- mov r14d,DWORD PTR[((240-128))+rdi]
- mov r15,r9
- mov rsi,r10
- vmovdqu xmm8,XMMWORD PTR[r8]
- lea r14,QWORD PTR[((-9))+r14]
- vmovdqa xmm14,XMMWORD PTR[r14*8+r12]
- vmovdqa xmm13,XMMWORD PTR[16+r14*8+r12]
- vmovdqa xmm12,XMMWORD PTR[32+r14*8+r12]
- sub r13,-16*4
- mov eax,DWORD PTR[r15]
- lea r12,QWORD PTR[r13*1+rsi]
- mov ebx,DWORD PTR[4+r15]
- cmp r13,rdx
- mov ecx,DWORD PTR[8+r15]
- cmove r12,rsp
- mov edx,DWORD PTR[12+r15]
- mov r8d,DWORD PTR[16+r15]
- mov r9d,DWORD PTR[20+r15]
- mov r10d,DWORD PTR[24+r15]
- mov r11d,DWORD PTR[28+r15]
- vmovdqu xmm10,XMMWORD PTR[((0-128))+rdi]
- jmp $L$oop_avx2
- ALIGN 16
- $L$oop_avx2::
- vmovdqa ymm7,YMMWORD PTR[((K256+512))]
- vmovdqu xmm0,XMMWORD PTR[((-64+0))+r13*1+rsi]
- vmovdqu xmm1,XMMWORD PTR[((-64+16))+r13*1+rsi]
- vmovdqu xmm2,XMMWORD PTR[((-64+32))+r13*1+rsi]
- vmovdqu xmm3,XMMWORD PTR[((-64+48))+r13*1+rsi]
- vinserti128 ymm0,ymm0,XMMWORD PTR[r12],1
- vinserti128 ymm1,ymm1,XMMWORD PTR[16+r12],1
- vpshufb ymm0,ymm0,ymm7
- vinserti128 ymm2,ymm2,XMMWORD PTR[32+r12],1
- vpshufb ymm1,ymm1,ymm7
- vinserti128 ymm3,ymm3,XMMWORD PTR[48+r12],1
- lea rbp,QWORD PTR[K256]
- vpshufb ymm2,ymm2,ymm7
- lea r13,QWORD PTR[((-64))+r13]
- vpaddd ymm4,ymm0,YMMWORD PTR[rbp]
- vpshufb ymm3,ymm3,ymm7
- vpaddd ymm5,ymm1,YMMWORD PTR[32+rbp]
- vpaddd ymm6,ymm2,YMMWORD PTR[64+rbp]
- vpaddd ymm7,ymm3,YMMWORD PTR[96+rbp]
- vmovdqa YMMWORD PTR[rsp],ymm4
- xor r14d,r14d
- vmovdqa YMMWORD PTR[32+rsp],ymm5
- lea rsp,QWORD PTR[((-64))+rsp]
- mov esi,ebx
- vmovdqa YMMWORD PTR[rsp],ymm6
- xor esi,ecx
- vmovdqa YMMWORD PTR[32+rsp],ymm7
- mov r12d,r9d
- sub rbp,-16*2*4
- jmp $L$avx2_00_47
- ALIGN 16
- $L$avx2_00_47::
- vmovdqu xmm9,XMMWORD PTR[r13]
- vpinsrq xmm15,xmm15,r13,0
- lea rsp,QWORD PTR[((-64))+rsp]
- vpalignr ymm4,ymm1,ymm0,4
- add r11d,DWORD PTR[((0+128))+rsp]
- and r12d,r8d
- rorx r13d,r8d,25
- vpalignr ymm7,ymm3,ymm2,4
- rorx r15d,r8d,11
- lea eax,DWORD PTR[r14*1+rax]
- lea r11d,DWORD PTR[r12*1+r11]
- vpsrld ymm6,ymm4,7
- andn r12d,r8d,r10d
- xor r13d,r15d
- rorx r14d,r8d,6
- vpaddd ymm0,ymm0,ymm7
- lea r11d,DWORD PTR[r12*1+r11]
- xor r13d,r14d
- mov r15d,eax
- vpsrld ymm7,ymm4,3
- rorx r12d,eax,22
- lea r11d,DWORD PTR[r13*1+r11]
- xor r15d,ebx
- vpslld ymm5,ymm4,14
- rorx r14d,eax,13
- rorx r13d,eax,2
- lea edx,DWORD PTR[r11*1+rdx]
- vpxor ymm4,ymm7,ymm6
- and esi,r15d
- vpxor xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((16-128))+rdi]
- xor r14d,r12d
- xor esi,ebx
- vpshufd ymm7,ymm3,250
- xor r14d,r13d
- lea r11d,DWORD PTR[rsi*1+r11]
- mov r12d,r8d
- vpsrld ymm6,ymm6,11
- add r10d,DWORD PTR[((4+128))+rsp]
- and r12d,edx
- rorx r13d,edx,25
- vpxor ymm4,ymm4,ymm5
- rorx esi,edx,11
- lea r11d,DWORD PTR[r14*1+r11]
- lea r10d,DWORD PTR[r12*1+r10]
- vpslld ymm5,ymm5,11
- andn r12d,edx,r9d
- xor r13d,esi
- rorx r14d,edx,6
- vpxor ymm4,ymm4,ymm6
- lea r10d,DWORD PTR[r12*1+r10]
- xor r13d,r14d
- mov esi,r11d
- vpsrld ymm6,ymm7,10
- rorx r12d,r11d,22
- lea r10d,DWORD PTR[r13*1+r10]
- xor esi,eax
- vpxor ymm4,ymm4,ymm5
- rorx r14d,r11d,13
- rorx r13d,r11d,2
- lea ecx,DWORD PTR[r10*1+rcx]
- vpsrlq ymm7,ymm7,17
- and r15d,esi
- vpxor xmm9,xmm9,xmm8
- xor r14d,r12d
- xor r15d,eax
- vpaddd ymm0,ymm0,ymm4
- xor r14d,r13d
- lea r10d,DWORD PTR[r15*1+r10]
- mov r12d,edx
- vpxor ymm6,ymm6,ymm7
- add r9d,DWORD PTR[((8+128))+rsp]
- and r12d,ecx
- rorx r13d,ecx,25
- vpsrlq ymm7,ymm7,2
- rorx r15d,ecx,11
- lea r10d,DWORD PTR[r14*1+r10]
- lea r9d,DWORD PTR[r12*1+r9]
- vpxor ymm6,ymm6,ymm7
- andn r12d,ecx,r8d
- xor r13d,r15d
- rorx r14d,ecx,6
- vpshufd ymm6,ymm6,132
- lea r9d,DWORD PTR[r12*1+r9]
- xor r13d,r14d
- mov r15d,r10d
- vpsrldq ymm6,ymm6,8
- rorx r12d,r10d,22
- lea r9d,DWORD PTR[r13*1+r9]
- xor r15d,r11d
- vpaddd ymm0,ymm0,ymm6
- rorx r14d,r10d,13
- rorx r13d,r10d,2
- lea ebx,DWORD PTR[r9*1+rbx]
- vpshufd ymm7,ymm0,80
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((32-128))+rdi]
- xor r14d,r12d
- xor esi,r11d
- vpsrld ymm6,ymm7,10
- xor r14d,r13d
- lea r9d,DWORD PTR[rsi*1+r9]
- mov r12d,ecx
- vpsrlq ymm7,ymm7,17
- add r8d,DWORD PTR[((12+128))+rsp]
- and r12d,ebx
- rorx r13d,ebx,25
- vpxor ymm6,ymm6,ymm7
- rorx esi,ebx,11
- lea r9d,DWORD PTR[r14*1+r9]
- lea r8d,DWORD PTR[r12*1+r8]
- vpsrlq ymm7,ymm7,2
- andn r12d,ebx,edx
- xor r13d,esi
- rorx r14d,ebx,6
- vpxor ymm6,ymm6,ymm7
- lea r8d,DWORD PTR[r12*1+r8]
- xor r13d,r14d
- mov esi,r9d
- vpshufd ymm6,ymm6,232
- rorx r12d,r9d,22
- lea r8d,DWORD PTR[r13*1+r8]
- xor esi,r10d
- vpslldq ymm6,ymm6,8
- rorx r14d,r9d,13
- rorx r13d,r9d,2
- lea eax,DWORD PTR[r8*1+rax]
- vpaddd ymm0,ymm0,ymm6
- and r15d,esi
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((48-128))+rdi]
- xor r14d,r12d
- xor r15d,r10d
- vpaddd ymm6,ymm0,YMMWORD PTR[rbp]
- xor r14d,r13d
- lea r8d,DWORD PTR[r15*1+r8]
- mov r12d,ebx
- vmovdqa YMMWORD PTR[rsp],ymm6
- vpalignr ymm4,ymm2,ymm1,4
- add edx,DWORD PTR[((32+128))+rsp]
- and r12d,eax
- rorx r13d,eax,25
- vpalignr ymm7,ymm0,ymm3,4
- rorx r15d,eax,11
- lea r8d,DWORD PTR[r14*1+r8]
- lea edx,DWORD PTR[r12*1+rdx]
- vpsrld ymm6,ymm4,7
- andn r12d,eax,ecx
- xor r13d,r15d
- rorx r14d,eax,6
- vpaddd ymm1,ymm1,ymm7
- lea edx,DWORD PTR[r12*1+rdx]
- xor r13d,r14d
- mov r15d,r8d
- vpsrld ymm7,ymm4,3
- rorx r12d,r8d,22
- lea edx,DWORD PTR[r13*1+rdx]
- xor r15d,r9d
- vpslld ymm5,ymm4,14
- rorx r14d,r8d,13
- rorx r13d,r8d,2
- lea r11d,DWORD PTR[rdx*1+r11]
- vpxor ymm4,ymm7,ymm6
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((64-128))+rdi]
- xor r14d,r12d
- xor esi,r9d
- vpshufd ymm7,ymm0,250
- xor r14d,r13d
- lea edx,DWORD PTR[rsi*1+rdx]
- mov r12d,eax
- vpsrld ymm6,ymm6,11
- add ecx,DWORD PTR[((36+128))+rsp]
- and r12d,r11d
- rorx r13d,r11d,25
- vpxor ymm4,ymm4,ymm5
- rorx esi,r11d,11
- lea edx,DWORD PTR[r14*1+rdx]
- lea ecx,DWORD PTR[r12*1+rcx]
- vpslld ymm5,ymm5,11
- andn r12d,r11d,ebx
- xor r13d,esi
- rorx r14d,r11d,6
- vpxor ymm4,ymm4,ymm6
- lea ecx,DWORD PTR[r12*1+rcx]
- xor r13d,r14d
- mov esi,edx
- vpsrld ymm6,ymm7,10
- rorx r12d,edx,22
- lea ecx,DWORD PTR[r13*1+rcx]
- xor esi,r8d
- vpxor ymm4,ymm4,ymm5
- rorx r14d,edx,13
- rorx r13d,edx,2
- lea r10d,DWORD PTR[rcx*1+r10]
- vpsrlq ymm7,ymm7,17
- and r15d,esi
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((80-128))+rdi]
- xor r14d,r12d
- xor r15d,r8d
- vpaddd ymm1,ymm1,ymm4
- xor r14d,r13d
- lea ecx,DWORD PTR[r15*1+rcx]
- mov r12d,r11d
- vpxor ymm6,ymm6,ymm7
- add ebx,DWORD PTR[((40+128))+rsp]
- and r12d,r10d
- rorx r13d,r10d,25
- vpsrlq ymm7,ymm7,2
- rorx r15d,r10d,11
- lea ecx,DWORD PTR[r14*1+rcx]
- lea ebx,DWORD PTR[r12*1+rbx]
- vpxor ymm6,ymm6,ymm7
- andn r12d,r10d,eax
- xor r13d,r15d
- rorx r14d,r10d,6
- vpshufd ymm6,ymm6,132
- lea ebx,DWORD PTR[r12*1+rbx]
- xor r13d,r14d
- mov r15d,ecx
- vpsrldq ymm6,ymm6,8
- rorx r12d,ecx,22
- lea ebx,DWORD PTR[r13*1+rbx]
- xor r15d,edx
- vpaddd ymm1,ymm1,ymm6
- rorx r14d,ecx,13
- rorx r13d,ecx,2
- lea r9d,DWORD PTR[rbx*1+r9]
- vpshufd ymm7,ymm1,80
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((96-128))+rdi]
- xor r14d,r12d
- xor esi,edx
- vpsrld ymm6,ymm7,10
- xor r14d,r13d
- lea ebx,DWORD PTR[rsi*1+rbx]
- mov r12d,r10d
- vpsrlq ymm7,ymm7,17
- add eax,DWORD PTR[((44+128))+rsp]
- and r12d,r9d
- rorx r13d,r9d,25
- vpxor ymm6,ymm6,ymm7
- rorx esi,r9d,11
- lea ebx,DWORD PTR[r14*1+rbx]
- lea eax,DWORD PTR[r12*1+rax]
- vpsrlq ymm7,ymm7,2
- andn r12d,r9d,r11d
- xor r13d,esi
- rorx r14d,r9d,6
- vpxor ymm6,ymm6,ymm7
- lea eax,DWORD PTR[r12*1+rax]
- xor r13d,r14d
- mov esi,ebx
- vpshufd ymm6,ymm6,232
- rorx r12d,ebx,22
- lea eax,DWORD PTR[r13*1+rax]
- xor esi,ecx
- vpslldq ymm6,ymm6,8
- rorx r14d,ebx,13
- rorx r13d,ebx,2
- lea r8d,DWORD PTR[rax*1+r8]
- vpaddd ymm1,ymm1,ymm6
- and r15d,esi
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((112-128))+rdi]
- xor r14d,r12d
- xor r15d,ecx
- vpaddd ymm6,ymm1,YMMWORD PTR[32+rbp]
- xor r14d,r13d
- lea eax,DWORD PTR[r15*1+rax]
- mov r12d,r9d
- vmovdqa YMMWORD PTR[32+rsp],ymm6
- lea rsp,QWORD PTR[((-64))+rsp]
- vpalignr ymm4,ymm3,ymm2,4
- add r11d,DWORD PTR[((0+128))+rsp]
- and r12d,r8d
- rorx r13d,r8d,25
- vpalignr ymm7,ymm1,ymm0,4
- rorx r15d,r8d,11
- lea eax,DWORD PTR[r14*1+rax]
- lea r11d,DWORD PTR[r12*1+r11]
- vpsrld ymm6,ymm4,7
- andn r12d,r8d,r10d
- xor r13d,r15d
- rorx r14d,r8d,6
- vpaddd ymm2,ymm2,ymm7
- lea r11d,DWORD PTR[r12*1+r11]
- xor r13d,r14d
- mov r15d,eax
- vpsrld ymm7,ymm4,3
- rorx r12d,eax,22
- lea r11d,DWORD PTR[r13*1+r11]
- xor r15d,ebx
- vpslld ymm5,ymm4,14
- rorx r14d,eax,13
- rorx r13d,eax,2
- lea edx,DWORD PTR[r11*1+rdx]
- vpxor ymm4,ymm7,ymm6
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((128-128))+rdi]
- xor r14d,r12d
- xor esi,ebx
- vpshufd ymm7,ymm1,250
- xor r14d,r13d
- lea r11d,DWORD PTR[rsi*1+r11]
- mov r12d,r8d
- vpsrld ymm6,ymm6,11
- add r10d,DWORD PTR[((4+128))+rsp]
- and r12d,edx
- rorx r13d,edx,25
- vpxor ymm4,ymm4,ymm5
- rorx esi,edx,11
- lea r11d,DWORD PTR[r14*1+r11]
- lea r10d,DWORD PTR[r12*1+r10]
- vpslld ymm5,ymm5,11
- andn r12d,edx,r9d
- xor r13d,esi
- rorx r14d,edx,6
- vpxor ymm4,ymm4,ymm6
- lea r10d,DWORD PTR[r12*1+r10]
- xor r13d,r14d
- mov esi,r11d
- vpsrld ymm6,ymm7,10
- rorx r12d,r11d,22
- lea r10d,DWORD PTR[r13*1+r10]
- xor esi,eax
- vpxor ymm4,ymm4,ymm5
- rorx r14d,r11d,13
- rorx r13d,r11d,2
- lea ecx,DWORD PTR[r10*1+rcx]
- vpsrlq ymm7,ymm7,17
- and r15d,esi
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((144-128))+rdi]
- xor r14d,r12d
- xor r15d,eax
- vpaddd ymm2,ymm2,ymm4
- xor r14d,r13d
- lea r10d,DWORD PTR[r15*1+r10]
- mov r12d,edx
- vpxor ymm6,ymm6,ymm7
- add r9d,DWORD PTR[((8+128))+rsp]
- and r12d,ecx
- rorx r13d,ecx,25
- vpsrlq ymm7,ymm7,2
- rorx r15d,ecx,11
- lea r10d,DWORD PTR[r14*1+r10]
- lea r9d,DWORD PTR[r12*1+r9]
- vpxor ymm6,ymm6,ymm7
- andn r12d,ecx,r8d
- xor r13d,r15d
- rorx r14d,ecx,6
- vpshufd ymm6,ymm6,132
- lea r9d,DWORD PTR[r12*1+r9]
- xor r13d,r14d
- mov r15d,r10d
- vpsrldq ymm6,ymm6,8
- rorx r12d,r10d,22
- lea r9d,DWORD PTR[r13*1+r9]
- xor r15d,r11d
- vpaddd ymm2,ymm2,ymm6
- rorx r14d,r10d,13
- rorx r13d,r10d,2
- lea ebx,DWORD PTR[r9*1+rbx]
- vpshufd ymm7,ymm2,80
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((160-128))+rdi]
- xor r14d,r12d
- xor esi,r11d
- vpsrld ymm6,ymm7,10
- xor r14d,r13d
- lea r9d,DWORD PTR[rsi*1+r9]
- mov r12d,ecx
- vpsrlq ymm7,ymm7,17
- add r8d,DWORD PTR[((12+128))+rsp]
- and r12d,ebx
- rorx r13d,ebx,25
- vpxor ymm6,ymm6,ymm7
- rorx esi,ebx,11
- lea r9d,DWORD PTR[r14*1+r9]
- lea r8d,DWORD PTR[r12*1+r8]
- vpsrlq ymm7,ymm7,2
- andn r12d,ebx,edx
- xor r13d,esi
- rorx r14d,ebx,6
- vpxor ymm6,ymm6,ymm7
- lea r8d,DWORD PTR[r12*1+r8]
- xor r13d,r14d
- mov esi,r9d
- vpshufd ymm6,ymm6,232
- rorx r12d,r9d,22
- lea r8d,DWORD PTR[r13*1+r8]
- xor esi,r10d
- vpslldq ymm6,ymm6,8
- rorx r14d,r9d,13
- rorx r13d,r9d,2
- lea eax,DWORD PTR[r8*1+rax]
- vpaddd ymm2,ymm2,ymm6
- and r15d,esi
- vaesenclast xmm11,xmm9,xmm10
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((176-128))+rdi]
- xor r14d,r12d
- xor r15d,r10d
- vpaddd ymm6,ymm2,YMMWORD PTR[64+rbp]
- xor r14d,r13d
- lea r8d,DWORD PTR[r15*1+r8]
- mov r12d,ebx
- vmovdqa YMMWORD PTR[rsp],ymm6
- vpalignr ymm4,ymm0,ymm3,4
- add edx,DWORD PTR[((32+128))+rsp]
- and r12d,eax
- rorx r13d,eax,25
- vpalignr ymm7,ymm2,ymm1,4
- rorx r15d,eax,11
- lea r8d,DWORD PTR[r14*1+r8]
- lea edx,DWORD PTR[r12*1+rdx]
- vpsrld ymm6,ymm4,7
- andn r12d,eax,ecx
- xor r13d,r15d
- rorx r14d,eax,6
- vpaddd ymm3,ymm3,ymm7
- lea edx,DWORD PTR[r12*1+rdx]
- xor r13d,r14d
- mov r15d,r8d
- vpsrld ymm7,ymm4,3
- rorx r12d,r8d,22
- lea edx,DWORD PTR[r13*1+rdx]
- xor r15d,r9d
- vpslld ymm5,ymm4,14
- rorx r14d,r8d,13
- rorx r13d,r8d,2
- lea r11d,DWORD PTR[rdx*1+r11]
- vpxor ymm4,ymm7,ymm6
- and esi,r15d
- vpand xmm8,xmm11,xmm12
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((192-128))+rdi]
- xor r14d,r12d
- xor esi,r9d
- vpshufd ymm7,ymm2,250
- xor r14d,r13d
- lea edx,DWORD PTR[rsi*1+rdx]
- mov r12d,eax
- vpsrld ymm6,ymm6,11
- add ecx,DWORD PTR[((36+128))+rsp]
- and r12d,r11d
- rorx r13d,r11d,25
- vpxor ymm4,ymm4,ymm5
- rorx esi,r11d,11
- lea edx,DWORD PTR[r14*1+rdx]
- lea ecx,DWORD PTR[r12*1+rcx]
- vpslld ymm5,ymm5,11
- andn r12d,r11d,ebx
- xor r13d,esi
- rorx r14d,r11d,6
- vpxor ymm4,ymm4,ymm6
- lea ecx,DWORD PTR[r12*1+rcx]
- xor r13d,r14d
- mov esi,edx
- vpsrld ymm6,ymm7,10
- rorx r12d,edx,22
- lea ecx,DWORD PTR[r13*1+rcx]
- xor esi,r8d
- vpxor ymm4,ymm4,ymm5
- rorx r14d,edx,13
- rorx r13d,edx,2
- lea r10d,DWORD PTR[rcx*1+r10]
- vpsrlq ymm7,ymm7,17
- and r15d,esi
- vaesenclast xmm11,xmm9,xmm10
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((208-128))+rdi]
- xor r14d,r12d
- xor r15d,r8d
- vpaddd ymm3,ymm3,ymm4
- xor r14d,r13d
- lea ecx,DWORD PTR[r15*1+rcx]
- mov r12d,r11d
- vpxor ymm6,ymm6,ymm7
- add ebx,DWORD PTR[((40+128))+rsp]
- and r12d,r10d
- rorx r13d,r10d,25
- vpsrlq ymm7,ymm7,2
- rorx r15d,r10d,11
- lea ecx,DWORD PTR[r14*1+rcx]
- lea ebx,DWORD PTR[r12*1+rbx]
- vpxor ymm6,ymm6,ymm7
- andn r12d,r10d,eax
- xor r13d,r15d
- rorx r14d,r10d,6
- vpshufd ymm6,ymm6,132
- lea ebx,DWORD PTR[r12*1+rbx]
- xor r13d,r14d
- mov r15d,ecx
- vpsrldq ymm6,ymm6,8
- rorx r12d,ecx,22
- lea ebx,DWORD PTR[r13*1+rbx]
- xor r15d,edx
- vpaddd ymm3,ymm3,ymm6
- rorx r14d,ecx,13
- rorx r13d,ecx,2
- lea r9d,DWORD PTR[rbx*1+r9]
- vpshufd ymm7,ymm3,80
- and esi,r15d
- vpand xmm11,xmm11,xmm13
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((224-128))+rdi]
- xor r14d,r12d
- xor esi,edx
- vpsrld ymm6,ymm7,10
- xor r14d,r13d
- lea ebx,DWORD PTR[rsi*1+rbx]
- mov r12d,r10d
- vpsrlq ymm7,ymm7,17
- add eax,DWORD PTR[((44+128))+rsp]
- and r12d,r9d
- rorx r13d,r9d,25
- vpxor ymm6,ymm6,ymm7
- rorx esi,r9d,11
- lea ebx,DWORD PTR[r14*1+rbx]
- lea eax,DWORD PTR[r12*1+rax]
- vpsrlq ymm7,ymm7,2
- andn r12d,r9d,r11d
- xor r13d,esi
- rorx r14d,r9d,6
- vpxor ymm6,ymm6,ymm7
- lea eax,DWORD PTR[r12*1+rax]
- xor r13d,r14d
- mov esi,ebx
- vpshufd ymm6,ymm6,232
- rorx r12d,ebx,22
- lea eax,DWORD PTR[r13*1+rax]
- xor esi,ecx
- vpslldq ymm6,ymm6,8
- rorx r14d,ebx,13
- rorx r13d,ebx,2
- lea r8d,DWORD PTR[rax*1+r8]
- vpaddd ymm3,ymm3,ymm6
- and r15d,esi
- vpor xmm8,xmm8,xmm11
- vaesenclast xmm11,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((0-128))+rdi]
- xor r14d,r12d
- xor r15d,ecx
- vpaddd ymm6,ymm3,YMMWORD PTR[96+rbp]
- xor r14d,r13d
- lea eax,DWORD PTR[r15*1+rax]
- mov r12d,r9d
- vmovdqa YMMWORD PTR[32+rsp],ymm6
- vmovq r13,xmm15
- vpextrq r15,xmm15,1
- vpand xmm11,xmm11,xmm14
- vpor xmm8,xmm8,xmm11
- vmovdqu XMMWORD PTR[r13*1+r15],xmm8
- lea r13,QWORD PTR[16+r13]
- lea rbp,QWORD PTR[128+rbp]
- cmp BYTE PTR[3+rbp],0
- jne $L$avx2_00_47
- vmovdqu xmm9,XMMWORD PTR[r13]
- vpinsrq xmm15,xmm15,r13,0
- add r11d,DWORD PTR[((0+64))+rsp]
- and r12d,r8d
- rorx r13d,r8d,25
- rorx r15d,r8d,11
- lea eax,DWORD PTR[r14*1+rax]
- lea r11d,DWORD PTR[r12*1+r11]
- andn r12d,r8d,r10d
- xor r13d,r15d
- rorx r14d,r8d,6
- lea r11d,DWORD PTR[r12*1+r11]
- xor r13d,r14d
- mov r15d,eax
- rorx r12d,eax,22
- lea r11d,DWORD PTR[r13*1+r11]
- xor r15d,ebx
- rorx r14d,eax,13
- rorx r13d,eax,2
- lea edx,DWORD PTR[r11*1+rdx]
- and esi,r15d
- vpxor xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((16-128))+rdi]
- xor r14d,r12d
- xor esi,ebx
- xor r14d,r13d
- lea r11d,DWORD PTR[rsi*1+r11]
- mov r12d,r8d
- add r10d,DWORD PTR[((4+64))+rsp]
- and r12d,edx
- rorx r13d,edx,25
- rorx esi,edx,11
- lea r11d,DWORD PTR[r14*1+r11]
- lea r10d,DWORD PTR[r12*1+r10]
- andn r12d,edx,r9d
- xor r13d,esi
- rorx r14d,edx,6
- lea r10d,DWORD PTR[r12*1+r10]
- xor r13d,r14d
- mov esi,r11d
- rorx r12d,r11d,22
- lea r10d,DWORD PTR[r13*1+r10]
- xor esi,eax
- rorx r14d,r11d,13
- rorx r13d,r11d,2
- lea ecx,DWORD PTR[r10*1+rcx]
- and r15d,esi
- vpxor xmm9,xmm9,xmm8
- xor r14d,r12d
- xor r15d,eax
- xor r14d,r13d
- lea r10d,DWORD PTR[r15*1+r10]
- mov r12d,edx
- add r9d,DWORD PTR[((8+64))+rsp]
- and r12d,ecx
- rorx r13d,ecx,25
- rorx r15d,ecx,11
- lea r10d,DWORD PTR[r14*1+r10]
- lea r9d,DWORD PTR[r12*1+r9]
- andn r12d,ecx,r8d
- xor r13d,r15d
- rorx r14d,ecx,6
- lea r9d,DWORD PTR[r12*1+r9]
- xor r13d,r14d
- mov r15d,r10d
- rorx r12d,r10d,22
- lea r9d,DWORD PTR[r13*1+r9]
- xor r15d,r11d
- rorx r14d,r10d,13
- rorx r13d,r10d,2
- lea ebx,DWORD PTR[r9*1+rbx]
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((32-128))+rdi]
- xor r14d,r12d
- xor esi,r11d
- xor r14d,r13d
- lea r9d,DWORD PTR[rsi*1+r9]
- mov r12d,ecx
- add r8d,DWORD PTR[((12+64))+rsp]
- and r12d,ebx
- rorx r13d,ebx,25
- rorx esi,ebx,11
- lea r9d,DWORD PTR[r14*1+r9]
- lea r8d,DWORD PTR[r12*1+r8]
- andn r12d,ebx,edx
- xor r13d,esi
- rorx r14d,ebx,6
- lea r8d,DWORD PTR[r12*1+r8]
- xor r13d,r14d
- mov esi,r9d
- rorx r12d,r9d,22
- lea r8d,DWORD PTR[r13*1+r8]
- xor esi,r10d
- rorx r14d,r9d,13
- rorx r13d,r9d,2
- lea eax,DWORD PTR[r8*1+rax]
- and r15d,esi
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((48-128))+rdi]
- xor r14d,r12d
- xor r15d,r10d
- xor r14d,r13d
- lea r8d,DWORD PTR[r15*1+r8]
- mov r12d,ebx
- add edx,DWORD PTR[((32+64))+rsp]
- and r12d,eax
- rorx r13d,eax,25
- rorx r15d,eax,11
- lea r8d,DWORD PTR[r14*1+r8]
- lea edx,DWORD PTR[r12*1+rdx]
- andn r12d,eax,ecx
- xor r13d,r15d
- rorx r14d,eax,6
- lea edx,DWORD PTR[r12*1+rdx]
- xor r13d,r14d
- mov r15d,r8d
- rorx r12d,r8d,22
- lea edx,DWORD PTR[r13*1+rdx]
- xor r15d,r9d
- rorx r14d,r8d,13
- rorx r13d,r8d,2
- lea r11d,DWORD PTR[rdx*1+r11]
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((64-128))+rdi]
- xor r14d,r12d
- xor esi,r9d
- xor r14d,r13d
- lea edx,DWORD PTR[rsi*1+rdx]
- mov r12d,eax
- add ecx,DWORD PTR[((36+64))+rsp]
- and r12d,r11d
- rorx r13d,r11d,25
- rorx esi,r11d,11
- lea edx,DWORD PTR[r14*1+rdx]
- lea ecx,DWORD PTR[r12*1+rcx]
- andn r12d,r11d,ebx
- xor r13d,esi
- rorx r14d,r11d,6
- lea ecx,DWORD PTR[r12*1+rcx]
- xor r13d,r14d
- mov esi,edx
- rorx r12d,edx,22
- lea ecx,DWORD PTR[r13*1+rcx]
- xor esi,r8d
- rorx r14d,edx,13
- rorx r13d,edx,2
- lea r10d,DWORD PTR[rcx*1+r10]
- and r15d,esi
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((80-128))+rdi]
- xor r14d,r12d
- xor r15d,r8d
- xor r14d,r13d
- lea ecx,DWORD PTR[r15*1+rcx]
- mov r12d,r11d
- add ebx,DWORD PTR[((40+64))+rsp]
- and r12d,r10d
- rorx r13d,r10d,25
- rorx r15d,r10d,11
- lea ecx,DWORD PTR[r14*1+rcx]
- lea ebx,DWORD PTR[r12*1+rbx]
- andn r12d,r10d,eax
- xor r13d,r15d
- rorx r14d,r10d,6
- lea ebx,DWORD PTR[r12*1+rbx]
- xor r13d,r14d
- mov r15d,ecx
- rorx r12d,ecx,22
- lea ebx,DWORD PTR[r13*1+rbx]
- xor r15d,edx
- rorx r14d,ecx,13
- rorx r13d,ecx,2
- lea r9d,DWORD PTR[rbx*1+r9]
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((96-128))+rdi]
- xor r14d,r12d
- xor esi,edx
- xor r14d,r13d
- lea ebx,DWORD PTR[rsi*1+rbx]
- mov r12d,r10d
- add eax,DWORD PTR[((44+64))+rsp]
- and r12d,r9d
- rorx r13d,r9d,25
- rorx esi,r9d,11
- lea ebx,DWORD PTR[r14*1+rbx]
- lea eax,DWORD PTR[r12*1+rax]
- andn r12d,r9d,r11d
- xor r13d,esi
- rorx r14d,r9d,6
- lea eax,DWORD PTR[r12*1+rax]
- xor r13d,r14d
- mov esi,ebx
- rorx r12d,ebx,22
- lea eax,DWORD PTR[r13*1+rax]
- xor esi,ecx
- rorx r14d,ebx,13
- rorx r13d,ebx,2
- lea r8d,DWORD PTR[rax*1+r8]
- and r15d,esi
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((112-128))+rdi]
- xor r14d,r12d
- xor r15d,ecx
- xor r14d,r13d
- lea eax,DWORD PTR[r15*1+rax]
- mov r12d,r9d
- add r11d,DWORD PTR[rsp]
- and r12d,r8d
- rorx r13d,r8d,25
- rorx r15d,r8d,11
- lea eax,DWORD PTR[r14*1+rax]
- lea r11d,DWORD PTR[r12*1+r11]
- andn r12d,r8d,r10d
- xor r13d,r15d
- rorx r14d,r8d,6
- lea r11d,DWORD PTR[r12*1+r11]
- xor r13d,r14d
- mov r15d,eax
- rorx r12d,eax,22
- lea r11d,DWORD PTR[r13*1+r11]
- xor r15d,ebx
- rorx r14d,eax,13
- rorx r13d,eax,2
- lea edx,DWORD PTR[r11*1+rdx]
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((128-128))+rdi]
- xor r14d,r12d
- xor esi,ebx
- xor r14d,r13d
- lea r11d,DWORD PTR[rsi*1+r11]
- mov r12d,r8d
- add r10d,DWORD PTR[4+rsp]
- and r12d,edx
- rorx r13d,edx,25
- rorx esi,edx,11
- lea r11d,DWORD PTR[r14*1+r11]
- lea r10d,DWORD PTR[r12*1+r10]
- andn r12d,edx,r9d
- xor r13d,esi
- rorx r14d,edx,6
- lea r10d,DWORD PTR[r12*1+r10]
- xor r13d,r14d
- mov esi,r11d
- rorx r12d,r11d,22
- lea r10d,DWORD PTR[r13*1+r10]
- xor esi,eax
- rorx r14d,r11d,13
- rorx r13d,r11d,2
- lea ecx,DWORD PTR[r10*1+rcx]
- and r15d,esi
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((144-128))+rdi]
- xor r14d,r12d
- xor r15d,eax
- xor r14d,r13d
- lea r10d,DWORD PTR[r15*1+r10]
- mov r12d,edx
- add r9d,DWORD PTR[8+rsp]
- and r12d,ecx
- rorx r13d,ecx,25
- rorx r15d,ecx,11
- lea r10d,DWORD PTR[r14*1+r10]
- lea r9d,DWORD PTR[r12*1+r9]
- andn r12d,ecx,r8d
- xor r13d,r15d
- rorx r14d,ecx,6
- lea r9d,DWORD PTR[r12*1+r9]
- xor r13d,r14d
- mov r15d,r10d
- rorx r12d,r10d,22
- lea r9d,DWORD PTR[r13*1+r9]
- xor r15d,r11d
- rorx r14d,r10d,13
- rorx r13d,r10d,2
- lea ebx,DWORD PTR[r9*1+rbx]
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((160-128))+rdi]
- xor r14d,r12d
- xor esi,r11d
- xor r14d,r13d
- lea r9d,DWORD PTR[rsi*1+r9]
- mov r12d,ecx
- add r8d,DWORD PTR[12+rsp]
- and r12d,ebx
- rorx r13d,ebx,25
- rorx esi,ebx,11
- lea r9d,DWORD PTR[r14*1+r9]
- lea r8d,DWORD PTR[r12*1+r8]
- andn r12d,ebx,edx
- xor r13d,esi
- rorx r14d,ebx,6
- lea r8d,DWORD PTR[r12*1+r8]
- xor r13d,r14d
- mov esi,r9d
- rorx r12d,r9d,22
- lea r8d,DWORD PTR[r13*1+r8]
- xor esi,r10d
- rorx r14d,r9d,13
- rorx r13d,r9d,2
- lea eax,DWORD PTR[r8*1+rax]
- and r15d,esi
- vaesenclast xmm11,xmm9,xmm10
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((176-128))+rdi]
- xor r14d,r12d
- xor r15d,r10d
- xor r14d,r13d
- lea r8d,DWORD PTR[r15*1+r8]
- mov r12d,ebx
- add edx,DWORD PTR[32+rsp]
- and r12d,eax
- rorx r13d,eax,25
- rorx r15d,eax,11
- lea r8d,DWORD PTR[r14*1+r8]
- lea edx,DWORD PTR[r12*1+rdx]
- andn r12d,eax,ecx
- xor r13d,r15d
- rorx r14d,eax,6
- lea edx,DWORD PTR[r12*1+rdx]
- xor r13d,r14d
- mov r15d,r8d
- rorx r12d,r8d,22
- lea edx,DWORD PTR[r13*1+rdx]
- xor r15d,r9d
- rorx r14d,r8d,13
- rorx r13d,r8d,2
- lea r11d,DWORD PTR[rdx*1+r11]
- and esi,r15d
- vpand xmm8,xmm11,xmm12
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((192-128))+rdi]
- xor r14d,r12d
- xor esi,r9d
- xor r14d,r13d
- lea edx,DWORD PTR[rsi*1+rdx]
- mov r12d,eax
- add ecx,DWORD PTR[36+rsp]
- and r12d,r11d
- rorx r13d,r11d,25
- rorx esi,r11d,11
- lea edx,DWORD PTR[r14*1+rdx]
- lea ecx,DWORD PTR[r12*1+rcx]
- andn r12d,r11d,ebx
- xor r13d,esi
- rorx r14d,r11d,6
- lea ecx,DWORD PTR[r12*1+rcx]
- xor r13d,r14d
- mov esi,edx
- rorx r12d,edx,22
- lea ecx,DWORD PTR[r13*1+rcx]
- xor esi,r8d
- rorx r14d,edx,13
- rorx r13d,edx,2
- lea r10d,DWORD PTR[rcx*1+r10]
- and r15d,esi
- vaesenclast xmm11,xmm9,xmm10
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((208-128))+rdi]
- xor r14d,r12d
- xor r15d,r8d
- xor r14d,r13d
- lea ecx,DWORD PTR[r15*1+rcx]
- mov r12d,r11d
- add ebx,DWORD PTR[40+rsp]
- and r12d,r10d
- rorx r13d,r10d,25
- rorx r15d,r10d,11
- lea ecx,DWORD PTR[r14*1+rcx]
- lea ebx,DWORD PTR[r12*1+rbx]
- andn r12d,r10d,eax
- xor r13d,r15d
- rorx r14d,r10d,6
- lea ebx,DWORD PTR[r12*1+rbx]
- xor r13d,r14d
- mov r15d,ecx
- rorx r12d,ecx,22
- lea ebx,DWORD PTR[r13*1+rbx]
- xor r15d,edx
- rorx r14d,ecx,13
- rorx r13d,ecx,2
- lea r9d,DWORD PTR[rbx*1+r9]
- and esi,r15d
- vpand xmm11,xmm11,xmm13
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((224-128))+rdi]
- xor r14d,r12d
- xor esi,edx
- xor r14d,r13d
- lea ebx,DWORD PTR[rsi*1+rbx]
- mov r12d,r10d
- add eax,DWORD PTR[44+rsp]
- and r12d,r9d
- rorx r13d,r9d,25
- rorx esi,r9d,11
- lea ebx,DWORD PTR[r14*1+rbx]
- lea eax,DWORD PTR[r12*1+rax]
- andn r12d,r9d,r11d
- xor r13d,esi
- rorx r14d,r9d,6
- lea eax,DWORD PTR[r12*1+rax]
- xor r13d,r14d
- mov esi,ebx
- rorx r12d,ebx,22
- lea eax,DWORD PTR[r13*1+rax]
- xor esi,ecx
- rorx r14d,ebx,13
- rorx r13d,ebx,2
- lea r8d,DWORD PTR[rax*1+r8]
- and r15d,esi
- vpor xmm8,xmm8,xmm11
- vaesenclast xmm11,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((0-128))+rdi]
- xor r14d,r12d
- xor r15d,ecx
- xor r14d,r13d
- lea eax,DWORD PTR[r15*1+rax]
- mov r12d,r9d
- vpextrq r12,xmm15,1
- vmovq r13,xmm15
- mov r15,QWORD PTR[552+rsp]
- add eax,r14d
- lea rbp,QWORD PTR[448+rsp]
- vpand xmm11,xmm11,xmm14
- vpor xmm8,xmm8,xmm11
- vmovdqu XMMWORD PTR[r13*1+r12],xmm8
- lea r13,QWORD PTR[16+r13]
- add eax,DWORD PTR[r15]
- add ebx,DWORD PTR[4+r15]
- add ecx,DWORD PTR[8+r15]
- add edx,DWORD PTR[12+r15]
- add r8d,DWORD PTR[16+r15]
- add r9d,DWORD PTR[20+r15]
- add r10d,DWORD PTR[24+r15]
- add r11d,DWORD PTR[28+r15]
- mov DWORD PTR[r15],eax
- mov DWORD PTR[4+r15],ebx
- mov DWORD PTR[8+r15],ecx
- mov DWORD PTR[12+r15],edx
- mov DWORD PTR[16+r15],r8d
- mov DWORD PTR[20+r15],r9d
- mov DWORD PTR[24+r15],r10d
- mov DWORD PTR[28+r15],r11d
- cmp r13,QWORD PTR[80+rbp]
- je $L$done_avx2
- xor r14d,r14d
- mov esi,ebx
- mov r12d,r9d
- xor esi,ecx
- jmp $L$ower_avx2
- ALIGN 16
- $L$ower_avx2::
- vmovdqu xmm9,XMMWORD PTR[r13]
- vpinsrq xmm15,xmm15,r13,0
- add r11d,DWORD PTR[((0+16))+rbp]
- and r12d,r8d
- rorx r13d,r8d,25
- rorx r15d,r8d,11
- lea eax,DWORD PTR[r14*1+rax]
- lea r11d,DWORD PTR[r12*1+r11]
- andn r12d,r8d,r10d
- xor r13d,r15d
- rorx r14d,r8d,6
- lea r11d,DWORD PTR[r12*1+r11]
- xor r13d,r14d
- mov r15d,eax
- rorx r12d,eax,22
- lea r11d,DWORD PTR[r13*1+r11]
- xor r15d,ebx
- rorx r14d,eax,13
- rorx r13d,eax,2
- lea edx,DWORD PTR[r11*1+rdx]
- and esi,r15d
- vpxor xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((16-128))+rdi]
- xor r14d,r12d
- xor esi,ebx
- xor r14d,r13d
- lea r11d,DWORD PTR[rsi*1+r11]
- mov r12d,r8d
- add r10d,DWORD PTR[((4+16))+rbp]
- and r12d,edx
- rorx r13d,edx,25
- rorx esi,edx,11
- lea r11d,DWORD PTR[r14*1+r11]
- lea r10d,DWORD PTR[r12*1+r10]
- andn r12d,edx,r9d
- xor r13d,esi
- rorx r14d,edx,6
- lea r10d,DWORD PTR[r12*1+r10]
- xor r13d,r14d
- mov esi,r11d
- rorx r12d,r11d,22
- lea r10d,DWORD PTR[r13*1+r10]
- xor esi,eax
- rorx r14d,r11d,13
- rorx r13d,r11d,2
- lea ecx,DWORD PTR[r10*1+rcx]
- and r15d,esi
- vpxor xmm9,xmm9,xmm8
- xor r14d,r12d
- xor r15d,eax
- xor r14d,r13d
- lea r10d,DWORD PTR[r15*1+r10]
- mov r12d,edx
- add r9d,DWORD PTR[((8+16))+rbp]
- and r12d,ecx
- rorx r13d,ecx,25
- rorx r15d,ecx,11
- lea r10d,DWORD PTR[r14*1+r10]
- lea r9d,DWORD PTR[r12*1+r9]
- andn r12d,ecx,r8d
- xor r13d,r15d
- rorx r14d,ecx,6
- lea r9d,DWORD PTR[r12*1+r9]
- xor r13d,r14d
- mov r15d,r10d
- rorx r12d,r10d,22
- lea r9d,DWORD PTR[r13*1+r9]
- xor r15d,r11d
- rorx r14d,r10d,13
- rorx r13d,r10d,2
- lea ebx,DWORD PTR[r9*1+rbx]
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((32-128))+rdi]
- xor r14d,r12d
- xor esi,r11d
- xor r14d,r13d
- lea r9d,DWORD PTR[rsi*1+r9]
- mov r12d,ecx
- add r8d,DWORD PTR[((12+16))+rbp]
- and r12d,ebx
- rorx r13d,ebx,25
- rorx esi,ebx,11
- lea r9d,DWORD PTR[r14*1+r9]
- lea r8d,DWORD PTR[r12*1+r8]
- andn r12d,ebx,edx
- xor r13d,esi
- rorx r14d,ebx,6
- lea r8d,DWORD PTR[r12*1+r8]
- xor r13d,r14d
- mov esi,r9d
- rorx r12d,r9d,22
- lea r8d,DWORD PTR[r13*1+r8]
- xor esi,r10d
- rorx r14d,r9d,13
- rorx r13d,r9d,2
- lea eax,DWORD PTR[r8*1+rax]
- and r15d,esi
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((48-128))+rdi]
- xor r14d,r12d
- xor r15d,r10d
- xor r14d,r13d
- lea r8d,DWORD PTR[r15*1+r8]
- mov r12d,ebx
- add edx,DWORD PTR[((32+16))+rbp]
- and r12d,eax
- rorx r13d,eax,25
- rorx r15d,eax,11
- lea r8d,DWORD PTR[r14*1+r8]
- lea edx,DWORD PTR[r12*1+rdx]
- andn r12d,eax,ecx
- xor r13d,r15d
- rorx r14d,eax,6
- lea edx,DWORD PTR[r12*1+rdx]
- xor r13d,r14d
- mov r15d,r8d
- rorx r12d,r8d,22
- lea edx,DWORD PTR[r13*1+rdx]
- xor r15d,r9d
- rorx r14d,r8d,13
- rorx r13d,r8d,2
- lea r11d,DWORD PTR[rdx*1+r11]
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((64-128))+rdi]
- xor r14d,r12d
- xor esi,r9d
- xor r14d,r13d
- lea edx,DWORD PTR[rsi*1+rdx]
- mov r12d,eax
- add ecx,DWORD PTR[((36+16))+rbp]
- and r12d,r11d
- rorx r13d,r11d,25
- rorx esi,r11d,11
- lea edx,DWORD PTR[r14*1+rdx]
- lea ecx,DWORD PTR[r12*1+rcx]
- andn r12d,r11d,ebx
- xor r13d,esi
- rorx r14d,r11d,6
- lea ecx,DWORD PTR[r12*1+rcx]
- xor r13d,r14d
- mov esi,edx
- rorx r12d,edx,22
- lea ecx,DWORD PTR[r13*1+rcx]
- xor esi,r8d
- rorx r14d,edx,13
- rorx r13d,edx,2
- lea r10d,DWORD PTR[rcx*1+r10]
- and r15d,esi
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((80-128))+rdi]
- xor r14d,r12d
- xor r15d,r8d
- xor r14d,r13d
- lea ecx,DWORD PTR[r15*1+rcx]
- mov r12d,r11d
- add ebx,DWORD PTR[((40+16))+rbp]
- and r12d,r10d
- rorx r13d,r10d,25
- rorx r15d,r10d,11
- lea ecx,DWORD PTR[r14*1+rcx]
- lea ebx,DWORD PTR[r12*1+rbx]
- andn r12d,r10d,eax
- xor r13d,r15d
- rorx r14d,r10d,6
- lea ebx,DWORD PTR[r12*1+rbx]
- xor r13d,r14d
- mov r15d,ecx
- rorx r12d,ecx,22
- lea ebx,DWORD PTR[r13*1+rbx]
- xor r15d,edx
- rorx r14d,ecx,13
- rorx r13d,ecx,2
- lea r9d,DWORD PTR[rbx*1+r9]
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((96-128))+rdi]
- xor r14d,r12d
- xor esi,edx
- xor r14d,r13d
- lea ebx,DWORD PTR[rsi*1+rbx]
- mov r12d,r10d
- add eax,DWORD PTR[((44+16))+rbp]
- and r12d,r9d
- rorx r13d,r9d,25
- rorx esi,r9d,11
- lea ebx,DWORD PTR[r14*1+rbx]
- lea eax,DWORD PTR[r12*1+rax]
- andn r12d,r9d,r11d
- xor r13d,esi
- rorx r14d,r9d,6
- lea eax,DWORD PTR[r12*1+rax]
- xor r13d,r14d
- mov esi,ebx
- rorx r12d,ebx,22
- lea eax,DWORD PTR[r13*1+rax]
- xor esi,ecx
- rorx r14d,ebx,13
- rorx r13d,ebx,2
- lea r8d,DWORD PTR[rax*1+r8]
- and r15d,esi
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((112-128))+rdi]
- xor r14d,r12d
- xor r15d,ecx
- xor r14d,r13d
- lea eax,DWORD PTR[r15*1+rax]
- mov r12d,r9d
- lea rbp,QWORD PTR[((-64))+rbp]
- add r11d,DWORD PTR[((0+16))+rbp]
- and r12d,r8d
- rorx r13d,r8d,25
- rorx r15d,r8d,11
- lea eax,DWORD PTR[r14*1+rax]
- lea r11d,DWORD PTR[r12*1+r11]
- andn r12d,r8d,r10d
- xor r13d,r15d
- rorx r14d,r8d,6
- lea r11d,DWORD PTR[r12*1+r11]
- xor r13d,r14d
- mov r15d,eax
- rorx r12d,eax,22
- lea r11d,DWORD PTR[r13*1+r11]
- xor r15d,ebx
- rorx r14d,eax,13
- rorx r13d,eax,2
- lea edx,DWORD PTR[r11*1+rdx]
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((128-128))+rdi]
- xor r14d,r12d
- xor esi,ebx
- xor r14d,r13d
- lea r11d,DWORD PTR[rsi*1+r11]
- mov r12d,r8d
- add r10d,DWORD PTR[((4+16))+rbp]
- and r12d,edx
- rorx r13d,edx,25
- rorx esi,edx,11
- lea r11d,DWORD PTR[r14*1+r11]
- lea r10d,DWORD PTR[r12*1+r10]
- andn r12d,edx,r9d
- xor r13d,esi
- rorx r14d,edx,6
- lea r10d,DWORD PTR[r12*1+r10]
- xor r13d,r14d
- mov esi,r11d
- rorx r12d,r11d,22
- lea r10d,DWORD PTR[r13*1+r10]
- xor esi,eax
- rorx r14d,r11d,13
- rorx r13d,r11d,2
- lea ecx,DWORD PTR[r10*1+rcx]
- and r15d,esi
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((144-128))+rdi]
- xor r14d,r12d
- xor r15d,eax
- xor r14d,r13d
- lea r10d,DWORD PTR[r15*1+r10]
- mov r12d,edx
- add r9d,DWORD PTR[((8+16))+rbp]
- and r12d,ecx
- rorx r13d,ecx,25
- rorx r15d,ecx,11
- lea r10d,DWORD PTR[r14*1+r10]
- lea r9d,DWORD PTR[r12*1+r9]
- andn r12d,ecx,r8d
- xor r13d,r15d
- rorx r14d,ecx,6
- lea r9d,DWORD PTR[r12*1+r9]
- xor r13d,r14d
- mov r15d,r10d
- rorx r12d,r10d,22
- lea r9d,DWORD PTR[r13*1+r9]
- xor r15d,r11d
- rorx r14d,r10d,13
- rorx r13d,r10d,2
- lea ebx,DWORD PTR[r9*1+rbx]
- and esi,r15d
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((160-128))+rdi]
- xor r14d,r12d
- xor esi,r11d
- xor r14d,r13d
- lea r9d,DWORD PTR[rsi*1+r9]
- mov r12d,ecx
- add r8d,DWORD PTR[((12+16))+rbp]
- and r12d,ebx
- rorx r13d,ebx,25
- rorx esi,ebx,11
- lea r9d,DWORD PTR[r14*1+r9]
- lea r8d,DWORD PTR[r12*1+r8]
- andn r12d,ebx,edx
- xor r13d,esi
- rorx r14d,ebx,6
- lea r8d,DWORD PTR[r12*1+r8]
- xor r13d,r14d
- mov esi,r9d
- rorx r12d,r9d,22
- lea r8d,DWORD PTR[r13*1+r8]
- xor esi,r10d
- rorx r14d,r9d,13
- rorx r13d,r9d,2
- lea eax,DWORD PTR[r8*1+rax]
- and r15d,esi
- vaesenclast xmm11,xmm9,xmm10
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((176-128))+rdi]
- xor r14d,r12d
- xor r15d,r10d
- xor r14d,r13d
- lea r8d,DWORD PTR[r15*1+r8]
- mov r12d,ebx
- add edx,DWORD PTR[((32+16))+rbp]
- and r12d,eax
- rorx r13d,eax,25
- rorx r15d,eax,11
- lea r8d,DWORD PTR[r14*1+r8]
- lea edx,DWORD PTR[r12*1+rdx]
- andn r12d,eax,ecx
- xor r13d,r15d
- rorx r14d,eax,6
- lea edx,DWORD PTR[r12*1+rdx]
- xor r13d,r14d
- mov r15d,r8d
- rorx r12d,r8d,22
- lea edx,DWORD PTR[r13*1+rdx]
- xor r15d,r9d
- rorx r14d,r8d,13
- rorx r13d,r8d,2
- lea r11d,DWORD PTR[rdx*1+r11]
- and esi,r15d
- vpand xmm8,xmm11,xmm12
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((192-128))+rdi]
- xor r14d,r12d
- xor esi,r9d
- xor r14d,r13d
- lea edx,DWORD PTR[rsi*1+rdx]
- mov r12d,eax
- add ecx,DWORD PTR[((36+16))+rbp]
- and r12d,r11d
- rorx r13d,r11d,25
- rorx esi,r11d,11
- lea edx,DWORD PTR[r14*1+rdx]
- lea ecx,DWORD PTR[r12*1+rcx]
- andn r12d,r11d,ebx
- xor r13d,esi
- rorx r14d,r11d,6
- lea ecx,DWORD PTR[r12*1+rcx]
- xor r13d,r14d
- mov esi,edx
- rorx r12d,edx,22
- lea ecx,DWORD PTR[r13*1+rcx]
- xor esi,r8d
- rorx r14d,edx,13
- rorx r13d,edx,2
- lea r10d,DWORD PTR[rcx*1+r10]
- and r15d,esi
- vaesenclast xmm11,xmm9,xmm10
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((208-128))+rdi]
- xor r14d,r12d
- xor r15d,r8d
- xor r14d,r13d
- lea ecx,DWORD PTR[r15*1+rcx]
- mov r12d,r11d
- add ebx,DWORD PTR[((40+16))+rbp]
- and r12d,r10d
- rorx r13d,r10d,25
- rorx r15d,r10d,11
- lea ecx,DWORD PTR[r14*1+rcx]
- lea ebx,DWORD PTR[r12*1+rbx]
- andn r12d,r10d,eax
- xor r13d,r15d
- rorx r14d,r10d,6
- lea ebx,DWORD PTR[r12*1+rbx]
- xor r13d,r14d
- mov r15d,ecx
- rorx r12d,ecx,22
- lea ebx,DWORD PTR[r13*1+rbx]
- xor r15d,edx
- rorx r14d,ecx,13
- rorx r13d,ecx,2
- lea r9d,DWORD PTR[rbx*1+r9]
- and esi,r15d
- vpand xmm11,xmm11,xmm13
- vaesenc xmm9,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((224-128))+rdi]
- xor r14d,r12d
- xor esi,edx
- xor r14d,r13d
- lea ebx,DWORD PTR[rsi*1+rbx]
- mov r12d,r10d
- add eax,DWORD PTR[((44+16))+rbp]
- and r12d,r9d
- rorx r13d,r9d,25
- rorx esi,r9d,11
- lea ebx,DWORD PTR[r14*1+rbx]
- lea eax,DWORD PTR[r12*1+rax]
- andn r12d,r9d,r11d
- xor r13d,esi
- rorx r14d,r9d,6
- lea eax,DWORD PTR[r12*1+rax]
- xor r13d,r14d
- mov esi,ebx
- rorx r12d,ebx,22
- lea eax,DWORD PTR[r13*1+rax]
- xor esi,ecx
- rorx r14d,ebx,13
- rorx r13d,ebx,2
- lea r8d,DWORD PTR[rax*1+r8]
- and r15d,esi
- vpor xmm8,xmm8,xmm11
- vaesenclast xmm11,xmm9,xmm10
- vmovdqu xmm10,XMMWORD PTR[((0-128))+rdi]
- xor r14d,r12d
- xor r15d,ecx
- xor r14d,r13d
- lea eax,DWORD PTR[r15*1+rax]
- mov r12d,r9d
- vmovq r13,xmm15
- vpextrq r15,xmm15,1
- vpand xmm11,xmm11,xmm14
- vpor xmm8,xmm8,xmm11
- lea rbp,QWORD PTR[((-64))+rbp]
- vmovdqu XMMWORD PTR[r13*1+r15],xmm8
- lea r13,QWORD PTR[16+r13]
- cmp rbp,rsp
- jae $L$ower_avx2
- mov r15,QWORD PTR[552+rsp]
- lea r13,QWORD PTR[64+r13]
- mov rsi,QWORD PTR[560+rsp]
- add eax,r14d
- lea rsp,QWORD PTR[448+rsp]
- add eax,DWORD PTR[r15]
- add ebx,DWORD PTR[4+r15]
- add ecx,DWORD PTR[8+r15]
- add edx,DWORD PTR[12+r15]
- add r8d,DWORD PTR[16+r15]
- add r9d,DWORD PTR[20+r15]
- add r10d,DWORD PTR[24+r15]
- lea r12,QWORD PTR[r13*1+rsi]
- add r11d,DWORD PTR[28+r15]
- cmp r13,QWORD PTR[((64+16))+rsp]
- mov DWORD PTR[r15],eax
- cmove r12,rsp
- mov DWORD PTR[4+r15],ebx
- mov DWORD PTR[8+r15],ecx
- mov DWORD PTR[12+r15],edx
- mov DWORD PTR[16+r15],r8d
- mov DWORD PTR[20+r15],r9d
- mov DWORD PTR[24+r15],r10d
- mov DWORD PTR[28+r15],r11d
- jbe $L$oop_avx2
- lea rbp,QWORD PTR[rsp]
- $L$done_avx2::
- mov r8,QWORD PTR[((64+32))+rbp]
- mov rsi,QWORD PTR[((64+56))+rbp]
- vmovdqu XMMWORD PTR[r8],xmm8
- vzeroall
- movaps xmm6,XMMWORD PTR[128+rbp]
- movaps xmm7,XMMWORD PTR[144+rbp]
- movaps xmm8,XMMWORD PTR[160+rbp]
- movaps xmm9,XMMWORD PTR[176+rbp]
- movaps xmm10,XMMWORD PTR[192+rbp]
- movaps xmm11,XMMWORD PTR[208+rbp]
- movaps xmm12,XMMWORD PTR[224+rbp]
- movaps xmm13,XMMWORD PTR[240+rbp]
- movaps xmm14,XMMWORD PTR[256+rbp]
- movaps xmm15,XMMWORD PTR[272+rbp]
- mov r15,QWORD PTR[((-48))+rsi]
- mov r14,QWORD PTR[((-40))+rsi]
- mov r13,QWORD PTR[((-32))+rsi]
- mov r12,QWORD PTR[((-24))+rsi]
- mov rbp,QWORD PTR[((-16))+rsi]
- mov rbx,QWORD PTR[((-8))+rsi]
- lea rsp,QWORD PTR[rsi]
- $L$epilogue_avx2::
- mov rdi,QWORD PTR[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD PTR[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_aesni_cbc_sha256_enc_avx2::
- aesni_cbc_sha256_enc_avx2 ENDP
- ALIGN 32
- aesni_cbc_sha256_enc_shaext PROC PRIVATE
- mov QWORD PTR[8+rsp],rdi ;WIN64 prologue
- mov QWORD PTR[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_aesni_cbc_sha256_enc_shaext::
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- mov r8,QWORD PTR[40+rsp]
- mov r9,QWORD PTR[48+rsp]
- mov r10,QWORD PTR[56+rsp]
- lea rsp,QWORD PTR[((-168))+rsp]
- movaps XMMWORD PTR[(-8-160)+rax],xmm6
- movaps XMMWORD PTR[(-8-144)+rax],xmm7
- movaps XMMWORD PTR[(-8-128)+rax],xmm8
- movaps XMMWORD PTR[(-8-112)+rax],xmm9
- movaps XMMWORD PTR[(-8-96)+rax],xmm10
- movaps XMMWORD PTR[(-8-80)+rax],xmm11
- movaps XMMWORD PTR[(-8-64)+rax],xmm12
- movaps XMMWORD PTR[(-8-48)+rax],xmm13
- movaps XMMWORD PTR[(-8-32)+rax],xmm14
- movaps XMMWORD PTR[(-8-16)+rax],xmm15
- $L$prologue_shaext::
- lea rax,QWORD PTR[((K256+128))]
- movdqu xmm1,XMMWORD PTR[r9]
- movdqu xmm2,XMMWORD PTR[16+r9]
- movdqa xmm3,XMMWORD PTR[((512-128))+rax]
- mov r11d,DWORD PTR[240+rcx]
- sub rsi,rdi
- movups xmm15,XMMWORD PTR[rcx]
- movups xmm6,XMMWORD PTR[r8]
- movups xmm4,XMMWORD PTR[16+rcx]
- lea rcx,QWORD PTR[112+rcx]
- pshufd xmm0,xmm1,01bh
- pshufd xmm1,xmm1,0b1h
- pshufd xmm2,xmm2,01bh
- movdqa xmm7,xmm3
- DB 102,15,58,15,202,8
- punpcklqdq xmm2,xmm0
- jmp $L$oop_shaext
- ALIGN 16
- $L$oop_shaext::
- movdqu xmm10,XMMWORD PTR[r10]
- movdqu xmm11,XMMWORD PTR[16+r10]
- movdqu xmm12,XMMWORD PTR[32+r10]
- DB 102,68,15,56,0,211
- movdqu xmm13,XMMWORD PTR[48+r10]
- movdqa xmm0,XMMWORD PTR[((0-128))+rax]
- paddd xmm0,xmm10
- DB 102,68,15,56,0,219
- movdqa xmm9,xmm2
- movdqa xmm8,xmm1
- movups xmm14,XMMWORD PTR[rdi]
- xorps xmm14,xmm15
- xorps xmm6,xmm14
- movups xmm5,XMMWORD PTR[((-80))+rcx]
- aesenc xmm6,xmm4
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movups xmm4,XMMWORD PTR[((-64))+rcx]
- aesenc xmm6,xmm5
- DB 15,56,203,202
- movdqa xmm0,XMMWORD PTR[((32-128))+rax]
- paddd xmm0,xmm11
- DB 102,68,15,56,0,227
- lea r10,QWORD PTR[64+r10]
- movups xmm5,XMMWORD PTR[((-48))+rcx]
- aesenc xmm6,xmm4
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movups xmm4,XMMWORD PTR[((-32))+rcx]
- aesenc xmm6,xmm5
- DB 15,56,203,202
- movdqa xmm0,XMMWORD PTR[((64-128))+rax]
- paddd xmm0,xmm12
- DB 102,68,15,56,0,235
- DB 69,15,56,204,211
- movups xmm5,XMMWORD PTR[((-16))+rcx]
- aesenc xmm6,xmm4
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movdqa xmm3,xmm13
- DB 102,65,15,58,15,220,4
- paddd xmm10,xmm3
- movups xmm4,XMMWORD PTR[rcx]
- aesenc xmm6,xmm5
- DB 15,56,203,202
- movdqa xmm0,XMMWORD PTR[((96-128))+rax]
- paddd xmm0,xmm13
- DB 69,15,56,205,213
- DB 69,15,56,204,220
- movups xmm5,XMMWORD PTR[16+rcx]
- aesenc xmm6,xmm4
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movups xmm4,XMMWORD PTR[32+rcx]
- aesenc xmm6,xmm5
- movdqa xmm3,xmm10
- DB 102,65,15,58,15,221,4
- paddd xmm11,xmm3
- DB 15,56,203,202
- movdqa xmm0,XMMWORD PTR[((128-128))+rax]
- paddd xmm0,xmm10
- DB 69,15,56,205,218
- DB 69,15,56,204,229
- movups xmm5,XMMWORD PTR[48+rcx]
- aesenc xmm6,xmm4
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movdqa xmm3,xmm11
- DB 102,65,15,58,15,218,4
- paddd xmm12,xmm3
- cmp r11d,11
- jb $L$aesenclast1
- movups xmm4,XMMWORD PTR[64+rcx]
- aesenc xmm6,xmm5
- movups xmm5,XMMWORD PTR[80+rcx]
- aesenc xmm6,xmm4
- je $L$aesenclast1
- movups xmm4,XMMWORD PTR[96+rcx]
- aesenc xmm6,xmm5
- movups xmm5,XMMWORD PTR[112+rcx]
- aesenc xmm6,xmm4
- $L$aesenclast1::
- aesenclast xmm6,xmm5
- movups xmm4,XMMWORD PTR[((16-112))+rcx]
- nop
- DB 15,56,203,202
- movups xmm14,XMMWORD PTR[16+rdi]
- xorps xmm14,xmm15
- movups XMMWORD PTR[rdi*1+rsi],xmm6
- xorps xmm6,xmm14
- movups xmm5,XMMWORD PTR[((-80))+rcx]
- aesenc xmm6,xmm4
- movdqa xmm0,XMMWORD PTR[((160-128))+rax]
- paddd xmm0,xmm11
- DB 69,15,56,205,227
- DB 69,15,56,204,234
- movups xmm4,XMMWORD PTR[((-64))+rcx]
- aesenc xmm6,xmm5
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movdqa xmm3,xmm12
- DB 102,65,15,58,15,219,4
- paddd xmm13,xmm3
- movups xmm5,XMMWORD PTR[((-48))+rcx]
- aesenc xmm6,xmm4
- DB 15,56,203,202
- movdqa xmm0,XMMWORD PTR[((192-128))+rax]
- paddd xmm0,xmm12
- DB 69,15,56,205,236
- DB 69,15,56,204,211
- movups xmm4,XMMWORD PTR[((-32))+rcx]
- aesenc xmm6,xmm5
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movdqa xmm3,xmm13
- DB 102,65,15,58,15,220,4
- paddd xmm10,xmm3
- movups xmm5,XMMWORD PTR[((-16))+rcx]
- aesenc xmm6,xmm4
- DB 15,56,203,202
- movdqa xmm0,XMMWORD PTR[((224-128))+rax]
- paddd xmm0,xmm13
- DB 69,15,56,205,213
- DB 69,15,56,204,220
- movups xmm4,XMMWORD PTR[rcx]
- aesenc xmm6,xmm5
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movdqa xmm3,xmm10
- DB 102,65,15,58,15,221,4
- paddd xmm11,xmm3
- movups xmm5,XMMWORD PTR[16+rcx]
- aesenc xmm6,xmm4
- DB 15,56,203,202
- movdqa xmm0,XMMWORD PTR[((256-128))+rax]
- paddd xmm0,xmm10
- DB 69,15,56,205,218
- DB 69,15,56,204,229
- movups xmm4,XMMWORD PTR[32+rcx]
- aesenc xmm6,xmm5
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movdqa xmm3,xmm11
- DB 102,65,15,58,15,218,4
- paddd xmm12,xmm3
- movups xmm5,XMMWORD PTR[48+rcx]
- aesenc xmm6,xmm4
- cmp r11d,11
- jb $L$aesenclast2
- movups xmm4,XMMWORD PTR[64+rcx]
- aesenc xmm6,xmm5
- movups xmm5,XMMWORD PTR[80+rcx]
- aesenc xmm6,xmm4
- je $L$aesenclast2
- movups xmm4,XMMWORD PTR[96+rcx]
- aesenc xmm6,xmm5
- movups xmm5,XMMWORD PTR[112+rcx]
- aesenc xmm6,xmm4
- $L$aesenclast2::
- aesenclast xmm6,xmm5
- movups xmm4,XMMWORD PTR[((16-112))+rcx]
- nop
- DB 15,56,203,202
- movups xmm14,XMMWORD PTR[32+rdi]
- xorps xmm14,xmm15
- movups XMMWORD PTR[16+rdi*1+rsi],xmm6
- xorps xmm6,xmm14
- movups xmm5,XMMWORD PTR[((-80))+rcx]
- aesenc xmm6,xmm4
- movdqa xmm0,XMMWORD PTR[((288-128))+rax]
- paddd xmm0,xmm11
- DB 69,15,56,205,227
- DB 69,15,56,204,234
- movups xmm4,XMMWORD PTR[((-64))+rcx]
- aesenc xmm6,xmm5
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movdqa xmm3,xmm12
- DB 102,65,15,58,15,219,4
- paddd xmm13,xmm3
- movups xmm5,XMMWORD PTR[((-48))+rcx]
- aesenc xmm6,xmm4
- DB 15,56,203,202
- movdqa xmm0,XMMWORD PTR[((320-128))+rax]
- paddd xmm0,xmm12
- DB 69,15,56,205,236
- DB 69,15,56,204,211
- movups xmm4,XMMWORD PTR[((-32))+rcx]
- aesenc xmm6,xmm5
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movdqa xmm3,xmm13
- DB 102,65,15,58,15,220,4
- paddd xmm10,xmm3
- movups xmm5,XMMWORD PTR[((-16))+rcx]
- aesenc xmm6,xmm4
- DB 15,56,203,202
- movdqa xmm0,XMMWORD PTR[((352-128))+rax]
- paddd xmm0,xmm13
- DB 69,15,56,205,213
- DB 69,15,56,204,220
- movups xmm4,XMMWORD PTR[rcx]
- aesenc xmm6,xmm5
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movdqa xmm3,xmm10
- DB 102,65,15,58,15,221,4
- paddd xmm11,xmm3
- movups xmm5,XMMWORD PTR[16+rcx]
- aesenc xmm6,xmm4
- DB 15,56,203,202
- movdqa xmm0,XMMWORD PTR[((384-128))+rax]
- paddd xmm0,xmm10
- DB 69,15,56,205,218
- DB 69,15,56,204,229
- movups xmm4,XMMWORD PTR[32+rcx]
- aesenc xmm6,xmm5
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movdqa xmm3,xmm11
- DB 102,65,15,58,15,218,4
- paddd xmm12,xmm3
- movups xmm5,XMMWORD PTR[48+rcx]
- aesenc xmm6,xmm4
- DB 15,56,203,202
- movdqa xmm0,XMMWORD PTR[((416-128))+rax]
- paddd xmm0,xmm11
- DB 69,15,56,205,227
- DB 69,15,56,204,234
- cmp r11d,11
- jb $L$aesenclast3
- movups xmm4,XMMWORD PTR[64+rcx]
- aesenc xmm6,xmm5
- movups xmm5,XMMWORD PTR[80+rcx]
- aesenc xmm6,xmm4
- je $L$aesenclast3
- movups xmm4,XMMWORD PTR[96+rcx]
- aesenc xmm6,xmm5
- movups xmm5,XMMWORD PTR[112+rcx]
- aesenc xmm6,xmm4
- $L$aesenclast3::
- aesenclast xmm6,xmm5
- movups xmm4,XMMWORD PTR[((16-112))+rcx]
- nop
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movdqa xmm3,xmm12
- DB 102,65,15,58,15,219,4
- paddd xmm13,xmm3
- movups xmm14,XMMWORD PTR[48+rdi]
- xorps xmm14,xmm15
- movups XMMWORD PTR[32+rdi*1+rsi],xmm6
- xorps xmm6,xmm14
- movups xmm5,XMMWORD PTR[((-80))+rcx]
- aesenc xmm6,xmm4
- movups xmm4,XMMWORD PTR[((-64))+rcx]
- aesenc xmm6,xmm5
- DB 15,56,203,202
- movdqa xmm0,XMMWORD PTR[((448-128))+rax]
- paddd xmm0,xmm12
- DB 69,15,56,205,236
- movdqa xmm3,xmm7
- movups xmm5,XMMWORD PTR[((-48))+rcx]
- aesenc xmm6,xmm4
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movups xmm4,XMMWORD PTR[((-32))+rcx]
- aesenc xmm6,xmm5
- DB 15,56,203,202
- movdqa xmm0,XMMWORD PTR[((480-128))+rax]
- paddd xmm0,xmm13
- movups xmm5,XMMWORD PTR[((-16))+rcx]
- aesenc xmm6,xmm4
- movups xmm4,XMMWORD PTR[rcx]
- aesenc xmm6,xmm5
- DB 15,56,203,209
- pshufd xmm0,xmm0,00eh
- movups xmm5,XMMWORD PTR[16+rcx]
- aesenc xmm6,xmm4
- DB 15,56,203,202
- movups xmm4,XMMWORD PTR[32+rcx]
- aesenc xmm6,xmm5
- movups xmm5,XMMWORD PTR[48+rcx]
- aesenc xmm6,xmm4
- cmp r11d,11
- jb $L$aesenclast4
- movups xmm4,XMMWORD PTR[64+rcx]
- aesenc xmm6,xmm5
- movups xmm5,XMMWORD PTR[80+rcx]
- aesenc xmm6,xmm4
- je $L$aesenclast4
- movups xmm4,XMMWORD PTR[96+rcx]
- aesenc xmm6,xmm5
- movups xmm5,XMMWORD PTR[112+rcx]
- aesenc xmm6,xmm4
- $L$aesenclast4::
- aesenclast xmm6,xmm5
- movups xmm4,XMMWORD PTR[((16-112))+rcx]
- nop
- paddd xmm2,xmm9
- paddd xmm1,xmm8
- dec rdx
- movups XMMWORD PTR[48+rdi*1+rsi],xmm6
- lea rdi,QWORD PTR[64+rdi]
- jnz $L$oop_shaext
- pshufd xmm2,xmm2,0b1h
- pshufd xmm3,xmm1,01bh
- pshufd xmm1,xmm1,0b1h
- punpckhqdq xmm1,xmm2
- DB 102,15,58,15,211,8
- movups XMMWORD PTR[r8],xmm6
- movdqu XMMWORD PTR[r9],xmm1
- movdqu XMMWORD PTR[16+r9],xmm2
- movaps xmm6,XMMWORD PTR[rsp]
- movaps xmm7,XMMWORD PTR[16+rsp]
- movaps xmm8,XMMWORD PTR[32+rsp]
- movaps xmm9,XMMWORD PTR[48+rsp]
- movaps xmm10,XMMWORD PTR[64+rsp]
- movaps xmm11,XMMWORD PTR[80+rsp]
- movaps xmm12,XMMWORD PTR[96+rsp]
- movaps xmm13,XMMWORD PTR[112+rsp]
- movaps xmm14,XMMWORD PTR[128+rsp]
- movaps xmm15,XMMWORD PTR[144+rsp]
- lea rsp,QWORD PTR[((8+160))+rsp]
- $L$epilogue_shaext::
- mov rdi,QWORD PTR[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD PTR[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_aesni_cbc_sha256_enc_shaext::
- aesni_cbc_sha256_enc_shaext ENDP
- EXTERN __imp_RtlVirtualUnwind:NEAR
- ALIGN 16
- se_handler PROC PRIVATE
- push rsi
- push rdi
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- pushfq
- sub rsp,64
- mov rax,QWORD PTR[120+r8]
- mov rbx,QWORD PTR[248+r8]
- mov rsi,QWORD PTR[8+r9]
- mov r11,QWORD PTR[56+r9]
- mov r10d,DWORD PTR[r11]
- lea r10,QWORD PTR[r10*1+rsi]
- cmp rbx,r10
- jb $L$in_prologue
- mov rax,QWORD PTR[152+r8]
- mov r10d,DWORD PTR[4+r11]
- lea r10,QWORD PTR[r10*1+rsi]
- cmp rbx,r10
- jae $L$in_prologue
- lea r10,QWORD PTR[aesni_cbc_sha256_enc_shaext]
- cmp rbx,r10
- jb $L$not_in_shaext
- lea rsi,QWORD PTR[rax]
- lea rdi,QWORD PTR[512+r8]
- mov ecx,20
- DD 0a548f3fch
- lea rax,QWORD PTR[168+rax]
- jmp $L$in_prologue
- $L$not_in_shaext::
- lea r10,QWORD PTR[$L$avx2_shortcut]
- cmp rbx,r10
- jb $L$not_in_avx2
- and rax,-256*4
- add rax,448
- $L$not_in_avx2::
- mov rsi,rax
- mov rax,QWORD PTR[((64+56))+rax]
- mov rbx,QWORD PTR[((-8))+rax]
- mov rbp,QWORD PTR[((-16))+rax]
- mov r12,QWORD PTR[((-24))+rax]
- mov r13,QWORD PTR[((-32))+rax]
- mov r14,QWORD PTR[((-40))+rax]
- mov r15,QWORD PTR[((-48))+rax]
- mov QWORD PTR[144+r8],rbx
- mov QWORD PTR[160+r8],rbp
- mov QWORD PTR[216+r8],r12
- mov QWORD PTR[224+r8],r13
- mov QWORD PTR[232+r8],r14
- mov QWORD PTR[240+r8],r15
- lea rsi,QWORD PTR[((64+64))+rsi]
- lea rdi,QWORD PTR[512+r8]
- mov ecx,20
- DD 0a548f3fch
- $L$in_prologue::
- mov rdi,QWORD PTR[8+rax]
- mov rsi,QWORD PTR[16+rax]
- mov QWORD PTR[152+r8],rax
- mov QWORD PTR[168+r8],rsi
- mov QWORD PTR[176+r8],rdi
- mov rdi,QWORD PTR[40+r9]
- mov rsi,r8
- mov ecx,154
- DD 0a548f3fch
- mov rsi,r9
- xor rcx,rcx
- mov rdx,QWORD PTR[8+rsi]
- mov r8,QWORD PTR[rsi]
- mov r9,QWORD PTR[16+rsi]
- mov r10,QWORD PTR[40+rsi]
- lea r11,QWORD PTR[56+rsi]
- lea r12,QWORD PTR[24+rsi]
- mov QWORD PTR[32+rsp],r10
- mov QWORD PTR[40+rsp],r11
- mov QWORD PTR[48+rsp],r12
- mov QWORD PTR[56+rsp],rcx
- call QWORD PTR[__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
- se_handler ENDP
- .text$ ENDS
- .pdata SEGMENT READONLY ALIGN(4)
- DD imagerel $L$SEH_begin_aesni_cbc_sha256_enc_xop
- DD imagerel $L$SEH_end_aesni_cbc_sha256_enc_xop
- DD imagerel $L$SEH_info_aesni_cbc_sha256_enc_xop
- DD imagerel $L$SEH_begin_aesni_cbc_sha256_enc_avx
- DD imagerel $L$SEH_end_aesni_cbc_sha256_enc_avx
- DD imagerel $L$SEH_info_aesni_cbc_sha256_enc_avx
- DD imagerel $L$SEH_begin_aesni_cbc_sha256_enc_avx2
- DD imagerel $L$SEH_end_aesni_cbc_sha256_enc_avx2
- DD imagerel $L$SEH_info_aesni_cbc_sha256_enc_avx2
- DD imagerel $L$SEH_begin_aesni_cbc_sha256_enc_shaext
- DD imagerel $L$SEH_end_aesni_cbc_sha256_enc_shaext
- DD imagerel $L$SEH_info_aesni_cbc_sha256_enc_shaext
- .pdata ENDS
- .xdata SEGMENT READONLY ALIGN(8)
- ALIGN 8
- $L$SEH_info_aesni_cbc_sha256_enc_xop::
- DB 9,0,0,0
- DD imagerel se_handler
- DD imagerel $L$prologue_xop,imagerel $L$epilogue_xop
- $L$SEH_info_aesni_cbc_sha256_enc_avx::
- DB 9,0,0,0
- DD imagerel se_handler
- DD imagerel $L$prologue_avx,imagerel $L$epilogue_avx
- $L$SEH_info_aesni_cbc_sha256_enc_avx2::
- DB 9,0,0,0
- DD imagerel se_handler
- DD imagerel $L$prologue_avx2,imagerel $L$epilogue_avx2
- $L$SEH_info_aesni_cbc_sha256_enc_shaext::
- DB 9,0,0,0
- DD imagerel se_handler
- DD imagerel $L$prologue_shaext,imagerel $L$epilogue_shaext
- .xdata ENDS
- END
|