1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173217421752176217721782179218021812182218321842185218621872188218921902191219221932194219521962197219821992200220122022203220422052206220722082209221022112212221322142215221622172218221922202221222222232224222522262227222822292230223122322233223422352236223722382239224022412242224322442245224622472248224922502251225222532254225522562257225822592260226122622263226422652266226722682269227022712272227322742275227622772278227922802281228222832284228522862287228822892290229122922293229422952296229722982299230023012302230323042305230623072308230923102311231223132314231523162317231823192320232123222323232423252326232723282329233023312332233323342335233623372338233923402341234223432344234523462347234823492350235123522353235423552356235723582359236023612362236323642365236623672368236923702371237223732374237523762377237823792380238123822383238423852386238723882389239023912392239323942395239623972398239924002401240224032404240524062407240824092410241124122413241424152416241724182419242024212422242324242425242624272428242924302431243224332434243524362437243824392440244124422443244424452446244724482449245024512452245324542455245624572458245924602461246224632464246524662467246824692470247124722473247424752476247724782479248024812482248324842485248624872488248924902491249224932494249524962497249824992500250125022503250425052506250725082509251025112512251325142515251625172518251925202521252225232524252525262527252825292530253125322533253425352536253725382539254025412542254325442545254625472548254925502551255225532554255525562557255825592560256125622563256425652566256725682569257025712572257325742575257625772578257925802581258225832584258525862587258825892590259125922593259425952596259725982599260026012602260326042605260626072608260926102611261226132614261526162617261826192620262126222623262426252626262726282629263026312632263326342635263626372638263926402641264226432644264526462647264826492650265126522653265426552656265726582659266026612662266326642665266626672668266926702671267226732674267526762677267826792680268126822683268426852686268726882689269026912692269326942695269626972698269927002701270227032704270527062707270827092710271127122713271427152716271727182719272027212722272327242725272627272728272927302731273227332734273527362737273827392740274127422743274427452746274727482749275027512752275327542755275627572758275927602761276227632764276527662767276827692770277127722773277427752776277727782779278027812782278327842785278627872788278927902791279227932794279527962797279827992800280128022803280428052806280728082809281028112812281328142815281628172818281928202821282228232824282528262827282828292830283128322833283428352836283728382839284028412842284328442845284628472848284928502851285228532854285528562857285828592860286128622863286428652866286728682869287028712872287328742875287628772878287928802881288228832884288528862887288828892890289128922893289428952896289728982899290029012902290329042905290629072908290929102911291229132914291529162917291829192920292129222923292429252926292729282929293029312932293329342935293629372938293929402941294229432944294529462947294829492950295129522953295429552956295729582959296029612962296329642965296629672968296929702971297229732974297529762977297829792980298129822983298429852986298729882989299029912992299329942995299629972998299930003001300230033004300530063007300830093010301130123013301430153016301730183019302030213022302330243025302630273028302930303031303230333034303530363037303830393040304130423043304430453046304730483049305030513052305330543055305630573058305930603061306230633064306530663067306830693070307130723073307430753076307730783079308030813082308330843085308630873088308930903091309230933094309530963097309830993100310131023103310431053106310731083109311031113112311331143115311631173118311931203121312231233124312531263127312831293130313131323133313431353136313731383139314031413142314331443145314631473148314931503151315231533154315531563157315831593160316131623163316431653166316731683169317031713172317331743175317631773178317931803181318231833184318531863187318831893190319131923193319431953196319731983199320032013202320332043205320632073208320932103211321232133214321532163217321832193220322132223223322432253226322732283229323032313232323332343235323632373238323932403241324232433244324532463247324832493250325132523253325432553256325732583259326032613262326332643265326632673268326932703271327232733274327532763277327832793280328132823283328432853286328732883289329032913292329332943295329632973298329933003301330233033304330533063307330833093310331133123313331433153316331733183319332033213322332333243325332633273328332933303331333233333334333533363337333833393340334133423343334433453346334733483349335033513352335333543355335633573358335933603361336233633364336533663367336833693370337133723373337433753376337733783379338033813382338333843385338633873388338933903391339233933394339533963397339833993400340134023403340434053406340734083409341034113412341334143415341634173418341934203421342234233424342534263427342834293430343134323433343434353436343734383439344034413442344334443445344634473448344934503451345234533454345534563457345834593460346134623463346434653466346734683469347034713472347334743475347634773478347934803481348234833484348534863487348834893490349134923493349434953496349734983499350035013502350335043505350635073508350935103511351235133514351535163517351835193520352135223523352435253526352735283529353035313532353335343535353635373538353935403541354235433544354535463547354835493550355135523553355435553556355735583559356035613562356335643565356635673568356935703571357235733574357535763577357835793580358135823583358435853586358735883589359035913592359335943595359635973598359936003601360236033604360536063607360836093610361136123613361436153616361736183619362036213622362336243625362636273628362936303631363236333634363536363637363836393640364136423643364436453646364736483649365036513652365336543655365636573658365936603661366236633664366536663667366836693670367136723673367436753676367736783679368036813682368336843685368636873688368936903691369236933694369536963697369836993700370137023703370437053706370737083709371037113712371337143715371637173718371937203721372237233724372537263727372837293730373137323733373437353736373737383739374037413742374337443745374637473748374937503751375237533754375537563757375837593760376137623763376437653766376737683769377037713772377337743775377637773778377937803781378237833784378537863787378837893790379137923793379437953796379737983799380038013802380338043805380638073808380938103811381238133814381538163817381838193820382138223823382438253826382738283829383038313832383338343835383638373838383938403841384238433844384538463847384838493850385138523853385438553856385738583859386038613862386338643865386638673868386938703871387238733874387538763877387838793880388138823883388438853886388738883889389038913892389338943895389638973898389939003901390239033904390539063907390839093910391139123913391439153916391739183919392039213922392339243925392639273928392939303931393239333934393539363937393839393940394139423943394439453946394739483949395039513952395339543955395639573958395939603961396239633964396539663967396839693970397139723973397439753976397739783979398039813982398339843985398639873988398939903991399239933994399539963997399839994000400140024003400440054006400740084009401040114012401340144015401640174018401940204021402240234024402540264027402840294030403140324033403440354036403740384039404040414042404340444045404640474048404940504051405240534054405540564057405840594060406140624063406440654066406740684069407040714072407340744075407640774078407940804081408240834084408540864087408840894090409140924093409440954096409740984099410041014102410341044105410641074108410941104111411241134114411541164117411841194120412141224123412441254126412741284129413041314132413341344135413641374138 |
- ; This file is generated from a similarly-named Perl script in the BoringSSL
- ; source tree. Do not edit by hand.
- default rel
- %define XMMWORD
- %define YMMWORD
- %define ZMMWORD
- section .text code align=64
- EXTERN GFp_ia32cap_P
- global GFp_sha256_block_data_order
- ALIGN 16
- GFp_sha256_block_data_order:
- mov QWORD[8+rsp],rdi ;WIN64 prologue
- mov QWORD[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_GFp_sha256_block_data_order:
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- lea r11,[GFp_ia32cap_P]
- mov r9d,DWORD[r11]
- mov r10d,DWORD[4+r11]
- mov r11d,DWORD[8+r11]
- and r9d,1073741824
- and r10d,268435968
- or r10d,r9d
- cmp r10d,1342177792
- je NEAR $L$avx_shortcut
- test r10d,512
- jnz NEAR $L$ssse3_shortcut
- mov rax,rsp
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- shl rdx,4
- sub rsp,16*4+4*8
- lea rdx,[rdx*4+rsi]
- and rsp,-64
- mov QWORD[((64+0))+rsp],rdi
- mov QWORD[((64+8))+rsp],rsi
- mov QWORD[((64+16))+rsp],rdx
- mov QWORD[88+rsp],rax
- $L$prologue:
- mov eax,DWORD[rdi]
- mov ebx,DWORD[4+rdi]
- mov ecx,DWORD[8+rdi]
- mov edx,DWORD[12+rdi]
- mov r8d,DWORD[16+rdi]
- mov r9d,DWORD[20+rdi]
- mov r10d,DWORD[24+rdi]
- mov r11d,DWORD[28+rdi]
- jmp NEAR $L$loop
- ALIGN 16
- $L$loop:
- mov edi,ebx
- lea rbp,[K256]
- xor edi,ecx
- mov r12d,DWORD[rsi]
- mov r13d,r8d
- mov r14d,eax
- bswap r12d
- ror r13d,14
- mov r15d,r9d
- xor r13d,r8d
- ror r14d,9
- xor r15d,r10d
- mov DWORD[rsp],r12d
- xor r14d,eax
- and r15d,r8d
- ror r13d,5
- add r12d,r11d
- xor r15d,r10d
- ror r14d,11
- xor r13d,r8d
- add r12d,r15d
- mov r15d,eax
- add r12d,DWORD[rbp]
- xor r14d,eax
- xor r15d,ebx
- ror r13d,6
- mov r11d,ebx
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor r11d,edi
- add edx,r12d
- add r11d,r12d
- lea rbp,[4+rbp]
- add r11d,r14d
- mov r12d,DWORD[4+rsi]
- mov r13d,edx
- mov r14d,r11d
- bswap r12d
- ror r13d,14
- mov edi,r8d
- xor r13d,edx
- ror r14d,9
- xor edi,r9d
- mov DWORD[4+rsp],r12d
- xor r14d,r11d
- and edi,edx
- ror r13d,5
- add r12d,r10d
- xor edi,r9d
- ror r14d,11
- xor r13d,edx
- add r12d,edi
- mov edi,r11d
- add r12d,DWORD[rbp]
- xor r14d,r11d
- xor edi,eax
- ror r13d,6
- mov r10d,eax
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor r10d,r15d
- add ecx,r12d
- add r10d,r12d
- lea rbp,[4+rbp]
- add r10d,r14d
- mov r12d,DWORD[8+rsi]
- mov r13d,ecx
- mov r14d,r10d
- bswap r12d
- ror r13d,14
- mov r15d,edx
- xor r13d,ecx
- ror r14d,9
- xor r15d,r8d
- mov DWORD[8+rsp],r12d
- xor r14d,r10d
- and r15d,ecx
- ror r13d,5
- add r12d,r9d
- xor r15d,r8d
- ror r14d,11
- xor r13d,ecx
- add r12d,r15d
- mov r15d,r10d
- add r12d,DWORD[rbp]
- xor r14d,r10d
- xor r15d,r11d
- ror r13d,6
- mov r9d,r11d
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor r9d,edi
- add ebx,r12d
- add r9d,r12d
- lea rbp,[4+rbp]
- add r9d,r14d
- mov r12d,DWORD[12+rsi]
- mov r13d,ebx
- mov r14d,r9d
- bswap r12d
- ror r13d,14
- mov edi,ecx
- xor r13d,ebx
- ror r14d,9
- xor edi,edx
- mov DWORD[12+rsp],r12d
- xor r14d,r9d
- and edi,ebx
- ror r13d,5
- add r12d,r8d
- xor edi,edx
- ror r14d,11
- xor r13d,ebx
- add r12d,edi
- mov edi,r9d
- add r12d,DWORD[rbp]
- xor r14d,r9d
- xor edi,r10d
- ror r13d,6
- mov r8d,r10d
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor r8d,r15d
- add eax,r12d
- add r8d,r12d
- lea rbp,[20+rbp]
- add r8d,r14d
- mov r12d,DWORD[16+rsi]
- mov r13d,eax
- mov r14d,r8d
- bswap r12d
- ror r13d,14
- mov r15d,ebx
- xor r13d,eax
- ror r14d,9
- xor r15d,ecx
- mov DWORD[16+rsp],r12d
- xor r14d,r8d
- and r15d,eax
- ror r13d,5
- add r12d,edx
- xor r15d,ecx
- ror r14d,11
- xor r13d,eax
- add r12d,r15d
- mov r15d,r8d
- add r12d,DWORD[rbp]
- xor r14d,r8d
- xor r15d,r9d
- ror r13d,6
- mov edx,r9d
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor edx,edi
- add r11d,r12d
- add edx,r12d
- lea rbp,[4+rbp]
- add edx,r14d
- mov r12d,DWORD[20+rsi]
- mov r13d,r11d
- mov r14d,edx
- bswap r12d
- ror r13d,14
- mov edi,eax
- xor r13d,r11d
- ror r14d,9
- xor edi,ebx
- mov DWORD[20+rsp],r12d
- xor r14d,edx
- and edi,r11d
- ror r13d,5
- add r12d,ecx
- xor edi,ebx
- ror r14d,11
- xor r13d,r11d
- add r12d,edi
- mov edi,edx
- add r12d,DWORD[rbp]
- xor r14d,edx
- xor edi,r8d
- ror r13d,6
- mov ecx,r8d
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor ecx,r15d
- add r10d,r12d
- add ecx,r12d
- lea rbp,[4+rbp]
- add ecx,r14d
- mov r12d,DWORD[24+rsi]
- mov r13d,r10d
- mov r14d,ecx
- bswap r12d
- ror r13d,14
- mov r15d,r11d
- xor r13d,r10d
- ror r14d,9
- xor r15d,eax
- mov DWORD[24+rsp],r12d
- xor r14d,ecx
- and r15d,r10d
- ror r13d,5
- add r12d,ebx
- xor r15d,eax
- ror r14d,11
- xor r13d,r10d
- add r12d,r15d
- mov r15d,ecx
- add r12d,DWORD[rbp]
- xor r14d,ecx
- xor r15d,edx
- ror r13d,6
- mov ebx,edx
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor ebx,edi
- add r9d,r12d
- add ebx,r12d
- lea rbp,[4+rbp]
- add ebx,r14d
- mov r12d,DWORD[28+rsi]
- mov r13d,r9d
- mov r14d,ebx
- bswap r12d
- ror r13d,14
- mov edi,r10d
- xor r13d,r9d
- ror r14d,9
- xor edi,r11d
- mov DWORD[28+rsp],r12d
- xor r14d,ebx
- and edi,r9d
- ror r13d,5
- add r12d,eax
- xor edi,r11d
- ror r14d,11
- xor r13d,r9d
- add r12d,edi
- mov edi,ebx
- add r12d,DWORD[rbp]
- xor r14d,ebx
- xor edi,ecx
- ror r13d,6
- mov eax,ecx
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor eax,r15d
- add r8d,r12d
- add eax,r12d
- lea rbp,[20+rbp]
- add eax,r14d
- mov r12d,DWORD[32+rsi]
- mov r13d,r8d
- mov r14d,eax
- bswap r12d
- ror r13d,14
- mov r15d,r9d
- xor r13d,r8d
- ror r14d,9
- xor r15d,r10d
- mov DWORD[32+rsp],r12d
- xor r14d,eax
- and r15d,r8d
- ror r13d,5
- add r12d,r11d
- xor r15d,r10d
- ror r14d,11
- xor r13d,r8d
- add r12d,r15d
- mov r15d,eax
- add r12d,DWORD[rbp]
- xor r14d,eax
- xor r15d,ebx
- ror r13d,6
- mov r11d,ebx
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor r11d,edi
- add edx,r12d
- add r11d,r12d
- lea rbp,[4+rbp]
- add r11d,r14d
- mov r12d,DWORD[36+rsi]
- mov r13d,edx
- mov r14d,r11d
- bswap r12d
- ror r13d,14
- mov edi,r8d
- xor r13d,edx
- ror r14d,9
- xor edi,r9d
- mov DWORD[36+rsp],r12d
- xor r14d,r11d
- and edi,edx
- ror r13d,5
- add r12d,r10d
- xor edi,r9d
- ror r14d,11
- xor r13d,edx
- add r12d,edi
- mov edi,r11d
- add r12d,DWORD[rbp]
- xor r14d,r11d
- xor edi,eax
- ror r13d,6
- mov r10d,eax
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor r10d,r15d
- add ecx,r12d
- add r10d,r12d
- lea rbp,[4+rbp]
- add r10d,r14d
- mov r12d,DWORD[40+rsi]
- mov r13d,ecx
- mov r14d,r10d
- bswap r12d
- ror r13d,14
- mov r15d,edx
- xor r13d,ecx
- ror r14d,9
- xor r15d,r8d
- mov DWORD[40+rsp],r12d
- xor r14d,r10d
- and r15d,ecx
- ror r13d,5
- add r12d,r9d
- xor r15d,r8d
- ror r14d,11
- xor r13d,ecx
- add r12d,r15d
- mov r15d,r10d
- add r12d,DWORD[rbp]
- xor r14d,r10d
- xor r15d,r11d
- ror r13d,6
- mov r9d,r11d
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor r9d,edi
- add ebx,r12d
- add r9d,r12d
- lea rbp,[4+rbp]
- add r9d,r14d
- mov r12d,DWORD[44+rsi]
- mov r13d,ebx
- mov r14d,r9d
- bswap r12d
- ror r13d,14
- mov edi,ecx
- xor r13d,ebx
- ror r14d,9
- xor edi,edx
- mov DWORD[44+rsp],r12d
- xor r14d,r9d
- and edi,ebx
- ror r13d,5
- add r12d,r8d
- xor edi,edx
- ror r14d,11
- xor r13d,ebx
- add r12d,edi
- mov edi,r9d
- add r12d,DWORD[rbp]
- xor r14d,r9d
- xor edi,r10d
- ror r13d,6
- mov r8d,r10d
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor r8d,r15d
- add eax,r12d
- add r8d,r12d
- lea rbp,[20+rbp]
- add r8d,r14d
- mov r12d,DWORD[48+rsi]
- mov r13d,eax
- mov r14d,r8d
- bswap r12d
- ror r13d,14
- mov r15d,ebx
- xor r13d,eax
- ror r14d,9
- xor r15d,ecx
- mov DWORD[48+rsp],r12d
- xor r14d,r8d
- and r15d,eax
- ror r13d,5
- add r12d,edx
- xor r15d,ecx
- ror r14d,11
- xor r13d,eax
- add r12d,r15d
- mov r15d,r8d
- add r12d,DWORD[rbp]
- xor r14d,r8d
- xor r15d,r9d
- ror r13d,6
- mov edx,r9d
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor edx,edi
- add r11d,r12d
- add edx,r12d
- lea rbp,[4+rbp]
- add edx,r14d
- mov r12d,DWORD[52+rsi]
- mov r13d,r11d
- mov r14d,edx
- bswap r12d
- ror r13d,14
- mov edi,eax
- xor r13d,r11d
- ror r14d,9
- xor edi,ebx
- mov DWORD[52+rsp],r12d
- xor r14d,edx
- and edi,r11d
- ror r13d,5
- add r12d,ecx
- xor edi,ebx
- ror r14d,11
- xor r13d,r11d
- add r12d,edi
- mov edi,edx
- add r12d,DWORD[rbp]
- xor r14d,edx
- xor edi,r8d
- ror r13d,6
- mov ecx,r8d
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor ecx,r15d
- add r10d,r12d
- add ecx,r12d
- lea rbp,[4+rbp]
- add ecx,r14d
- mov r12d,DWORD[56+rsi]
- mov r13d,r10d
- mov r14d,ecx
- bswap r12d
- ror r13d,14
- mov r15d,r11d
- xor r13d,r10d
- ror r14d,9
- xor r15d,eax
- mov DWORD[56+rsp],r12d
- xor r14d,ecx
- and r15d,r10d
- ror r13d,5
- add r12d,ebx
- xor r15d,eax
- ror r14d,11
- xor r13d,r10d
- add r12d,r15d
- mov r15d,ecx
- add r12d,DWORD[rbp]
- xor r14d,ecx
- xor r15d,edx
- ror r13d,6
- mov ebx,edx
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor ebx,edi
- add r9d,r12d
- add ebx,r12d
- lea rbp,[4+rbp]
- add ebx,r14d
- mov r12d,DWORD[60+rsi]
- mov r13d,r9d
- mov r14d,ebx
- bswap r12d
- ror r13d,14
- mov edi,r10d
- xor r13d,r9d
- ror r14d,9
- xor edi,r11d
- mov DWORD[60+rsp],r12d
- xor r14d,ebx
- and edi,r9d
- ror r13d,5
- add r12d,eax
- xor edi,r11d
- ror r14d,11
- xor r13d,r9d
- add r12d,edi
- mov edi,ebx
- add r12d,DWORD[rbp]
- xor r14d,ebx
- xor edi,ecx
- ror r13d,6
- mov eax,ecx
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor eax,r15d
- add r8d,r12d
- add eax,r12d
- lea rbp,[20+rbp]
- jmp NEAR $L$rounds_16_xx
- ALIGN 16
- $L$rounds_16_xx:
- mov r13d,DWORD[4+rsp]
- mov r15d,DWORD[56+rsp]
- mov r12d,r13d
- ror r13d,11
- add eax,r14d
- mov r14d,r15d
- ror r15d,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor r15d,r14d
- shr r14d,10
- ror r15d,17
- xor r12d,r13d
- xor r15d,r14d
- add r12d,DWORD[36+rsp]
- add r12d,DWORD[rsp]
- mov r13d,r8d
- add r12d,r15d
- mov r14d,eax
- ror r13d,14
- mov r15d,r9d
- xor r13d,r8d
- ror r14d,9
- xor r15d,r10d
- mov DWORD[rsp],r12d
- xor r14d,eax
- and r15d,r8d
- ror r13d,5
- add r12d,r11d
- xor r15d,r10d
- ror r14d,11
- xor r13d,r8d
- add r12d,r15d
- mov r15d,eax
- add r12d,DWORD[rbp]
- xor r14d,eax
- xor r15d,ebx
- ror r13d,6
- mov r11d,ebx
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor r11d,edi
- add edx,r12d
- add r11d,r12d
- lea rbp,[4+rbp]
- mov r13d,DWORD[8+rsp]
- mov edi,DWORD[60+rsp]
- mov r12d,r13d
- ror r13d,11
- add r11d,r14d
- mov r14d,edi
- ror edi,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor edi,r14d
- shr r14d,10
- ror edi,17
- xor r12d,r13d
- xor edi,r14d
- add r12d,DWORD[40+rsp]
- add r12d,DWORD[4+rsp]
- mov r13d,edx
- add r12d,edi
- mov r14d,r11d
- ror r13d,14
- mov edi,r8d
- xor r13d,edx
- ror r14d,9
- xor edi,r9d
- mov DWORD[4+rsp],r12d
- xor r14d,r11d
- and edi,edx
- ror r13d,5
- add r12d,r10d
- xor edi,r9d
- ror r14d,11
- xor r13d,edx
- add r12d,edi
- mov edi,r11d
- add r12d,DWORD[rbp]
- xor r14d,r11d
- xor edi,eax
- ror r13d,6
- mov r10d,eax
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor r10d,r15d
- add ecx,r12d
- add r10d,r12d
- lea rbp,[4+rbp]
- mov r13d,DWORD[12+rsp]
- mov r15d,DWORD[rsp]
- mov r12d,r13d
- ror r13d,11
- add r10d,r14d
- mov r14d,r15d
- ror r15d,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor r15d,r14d
- shr r14d,10
- ror r15d,17
- xor r12d,r13d
- xor r15d,r14d
- add r12d,DWORD[44+rsp]
- add r12d,DWORD[8+rsp]
- mov r13d,ecx
- add r12d,r15d
- mov r14d,r10d
- ror r13d,14
- mov r15d,edx
- xor r13d,ecx
- ror r14d,9
- xor r15d,r8d
- mov DWORD[8+rsp],r12d
- xor r14d,r10d
- and r15d,ecx
- ror r13d,5
- add r12d,r9d
- xor r15d,r8d
- ror r14d,11
- xor r13d,ecx
- add r12d,r15d
- mov r15d,r10d
- add r12d,DWORD[rbp]
- xor r14d,r10d
- xor r15d,r11d
- ror r13d,6
- mov r9d,r11d
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor r9d,edi
- add ebx,r12d
- add r9d,r12d
- lea rbp,[4+rbp]
- mov r13d,DWORD[16+rsp]
- mov edi,DWORD[4+rsp]
- mov r12d,r13d
- ror r13d,11
- add r9d,r14d
- mov r14d,edi
- ror edi,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor edi,r14d
- shr r14d,10
- ror edi,17
- xor r12d,r13d
- xor edi,r14d
- add r12d,DWORD[48+rsp]
- add r12d,DWORD[12+rsp]
- mov r13d,ebx
- add r12d,edi
- mov r14d,r9d
- ror r13d,14
- mov edi,ecx
- xor r13d,ebx
- ror r14d,9
- xor edi,edx
- mov DWORD[12+rsp],r12d
- xor r14d,r9d
- and edi,ebx
- ror r13d,5
- add r12d,r8d
- xor edi,edx
- ror r14d,11
- xor r13d,ebx
- add r12d,edi
- mov edi,r9d
- add r12d,DWORD[rbp]
- xor r14d,r9d
- xor edi,r10d
- ror r13d,6
- mov r8d,r10d
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor r8d,r15d
- add eax,r12d
- add r8d,r12d
- lea rbp,[20+rbp]
- mov r13d,DWORD[20+rsp]
- mov r15d,DWORD[8+rsp]
- mov r12d,r13d
- ror r13d,11
- add r8d,r14d
- mov r14d,r15d
- ror r15d,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor r15d,r14d
- shr r14d,10
- ror r15d,17
- xor r12d,r13d
- xor r15d,r14d
- add r12d,DWORD[52+rsp]
- add r12d,DWORD[16+rsp]
- mov r13d,eax
- add r12d,r15d
- mov r14d,r8d
- ror r13d,14
- mov r15d,ebx
- xor r13d,eax
- ror r14d,9
- xor r15d,ecx
- mov DWORD[16+rsp],r12d
- xor r14d,r8d
- and r15d,eax
- ror r13d,5
- add r12d,edx
- xor r15d,ecx
- ror r14d,11
- xor r13d,eax
- add r12d,r15d
- mov r15d,r8d
- add r12d,DWORD[rbp]
- xor r14d,r8d
- xor r15d,r9d
- ror r13d,6
- mov edx,r9d
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor edx,edi
- add r11d,r12d
- add edx,r12d
- lea rbp,[4+rbp]
- mov r13d,DWORD[24+rsp]
- mov edi,DWORD[12+rsp]
- mov r12d,r13d
- ror r13d,11
- add edx,r14d
- mov r14d,edi
- ror edi,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor edi,r14d
- shr r14d,10
- ror edi,17
- xor r12d,r13d
- xor edi,r14d
- add r12d,DWORD[56+rsp]
- add r12d,DWORD[20+rsp]
- mov r13d,r11d
- add r12d,edi
- mov r14d,edx
- ror r13d,14
- mov edi,eax
- xor r13d,r11d
- ror r14d,9
- xor edi,ebx
- mov DWORD[20+rsp],r12d
- xor r14d,edx
- and edi,r11d
- ror r13d,5
- add r12d,ecx
- xor edi,ebx
- ror r14d,11
- xor r13d,r11d
- add r12d,edi
- mov edi,edx
- add r12d,DWORD[rbp]
- xor r14d,edx
- xor edi,r8d
- ror r13d,6
- mov ecx,r8d
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor ecx,r15d
- add r10d,r12d
- add ecx,r12d
- lea rbp,[4+rbp]
- mov r13d,DWORD[28+rsp]
- mov r15d,DWORD[16+rsp]
- mov r12d,r13d
- ror r13d,11
- add ecx,r14d
- mov r14d,r15d
- ror r15d,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor r15d,r14d
- shr r14d,10
- ror r15d,17
- xor r12d,r13d
- xor r15d,r14d
- add r12d,DWORD[60+rsp]
- add r12d,DWORD[24+rsp]
- mov r13d,r10d
- add r12d,r15d
- mov r14d,ecx
- ror r13d,14
- mov r15d,r11d
- xor r13d,r10d
- ror r14d,9
- xor r15d,eax
- mov DWORD[24+rsp],r12d
- xor r14d,ecx
- and r15d,r10d
- ror r13d,5
- add r12d,ebx
- xor r15d,eax
- ror r14d,11
- xor r13d,r10d
- add r12d,r15d
- mov r15d,ecx
- add r12d,DWORD[rbp]
- xor r14d,ecx
- xor r15d,edx
- ror r13d,6
- mov ebx,edx
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor ebx,edi
- add r9d,r12d
- add ebx,r12d
- lea rbp,[4+rbp]
- mov r13d,DWORD[32+rsp]
- mov edi,DWORD[20+rsp]
- mov r12d,r13d
- ror r13d,11
- add ebx,r14d
- mov r14d,edi
- ror edi,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor edi,r14d
- shr r14d,10
- ror edi,17
- xor r12d,r13d
- xor edi,r14d
- add r12d,DWORD[rsp]
- add r12d,DWORD[28+rsp]
- mov r13d,r9d
- add r12d,edi
- mov r14d,ebx
- ror r13d,14
- mov edi,r10d
- xor r13d,r9d
- ror r14d,9
- xor edi,r11d
- mov DWORD[28+rsp],r12d
- xor r14d,ebx
- and edi,r9d
- ror r13d,5
- add r12d,eax
- xor edi,r11d
- ror r14d,11
- xor r13d,r9d
- add r12d,edi
- mov edi,ebx
- add r12d,DWORD[rbp]
- xor r14d,ebx
- xor edi,ecx
- ror r13d,6
- mov eax,ecx
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor eax,r15d
- add r8d,r12d
- add eax,r12d
- lea rbp,[20+rbp]
- mov r13d,DWORD[36+rsp]
- mov r15d,DWORD[24+rsp]
- mov r12d,r13d
- ror r13d,11
- add eax,r14d
- mov r14d,r15d
- ror r15d,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor r15d,r14d
- shr r14d,10
- ror r15d,17
- xor r12d,r13d
- xor r15d,r14d
- add r12d,DWORD[4+rsp]
- add r12d,DWORD[32+rsp]
- mov r13d,r8d
- add r12d,r15d
- mov r14d,eax
- ror r13d,14
- mov r15d,r9d
- xor r13d,r8d
- ror r14d,9
- xor r15d,r10d
- mov DWORD[32+rsp],r12d
- xor r14d,eax
- and r15d,r8d
- ror r13d,5
- add r12d,r11d
- xor r15d,r10d
- ror r14d,11
- xor r13d,r8d
- add r12d,r15d
- mov r15d,eax
- add r12d,DWORD[rbp]
- xor r14d,eax
- xor r15d,ebx
- ror r13d,6
- mov r11d,ebx
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor r11d,edi
- add edx,r12d
- add r11d,r12d
- lea rbp,[4+rbp]
- mov r13d,DWORD[40+rsp]
- mov edi,DWORD[28+rsp]
- mov r12d,r13d
- ror r13d,11
- add r11d,r14d
- mov r14d,edi
- ror edi,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor edi,r14d
- shr r14d,10
- ror edi,17
- xor r12d,r13d
- xor edi,r14d
- add r12d,DWORD[8+rsp]
- add r12d,DWORD[36+rsp]
- mov r13d,edx
- add r12d,edi
- mov r14d,r11d
- ror r13d,14
- mov edi,r8d
- xor r13d,edx
- ror r14d,9
- xor edi,r9d
- mov DWORD[36+rsp],r12d
- xor r14d,r11d
- and edi,edx
- ror r13d,5
- add r12d,r10d
- xor edi,r9d
- ror r14d,11
- xor r13d,edx
- add r12d,edi
- mov edi,r11d
- add r12d,DWORD[rbp]
- xor r14d,r11d
- xor edi,eax
- ror r13d,6
- mov r10d,eax
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor r10d,r15d
- add ecx,r12d
- add r10d,r12d
- lea rbp,[4+rbp]
- mov r13d,DWORD[44+rsp]
- mov r15d,DWORD[32+rsp]
- mov r12d,r13d
- ror r13d,11
- add r10d,r14d
- mov r14d,r15d
- ror r15d,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor r15d,r14d
- shr r14d,10
- ror r15d,17
- xor r12d,r13d
- xor r15d,r14d
- add r12d,DWORD[12+rsp]
- add r12d,DWORD[40+rsp]
- mov r13d,ecx
- add r12d,r15d
- mov r14d,r10d
- ror r13d,14
- mov r15d,edx
- xor r13d,ecx
- ror r14d,9
- xor r15d,r8d
- mov DWORD[40+rsp],r12d
- xor r14d,r10d
- and r15d,ecx
- ror r13d,5
- add r12d,r9d
- xor r15d,r8d
- ror r14d,11
- xor r13d,ecx
- add r12d,r15d
- mov r15d,r10d
- add r12d,DWORD[rbp]
- xor r14d,r10d
- xor r15d,r11d
- ror r13d,6
- mov r9d,r11d
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor r9d,edi
- add ebx,r12d
- add r9d,r12d
- lea rbp,[4+rbp]
- mov r13d,DWORD[48+rsp]
- mov edi,DWORD[36+rsp]
- mov r12d,r13d
- ror r13d,11
- add r9d,r14d
- mov r14d,edi
- ror edi,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor edi,r14d
- shr r14d,10
- ror edi,17
- xor r12d,r13d
- xor edi,r14d
- add r12d,DWORD[16+rsp]
- add r12d,DWORD[44+rsp]
- mov r13d,ebx
- add r12d,edi
- mov r14d,r9d
- ror r13d,14
- mov edi,ecx
- xor r13d,ebx
- ror r14d,9
- xor edi,edx
- mov DWORD[44+rsp],r12d
- xor r14d,r9d
- and edi,ebx
- ror r13d,5
- add r12d,r8d
- xor edi,edx
- ror r14d,11
- xor r13d,ebx
- add r12d,edi
- mov edi,r9d
- add r12d,DWORD[rbp]
- xor r14d,r9d
- xor edi,r10d
- ror r13d,6
- mov r8d,r10d
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor r8d,r15d
- add eax,r12d
- add r8d,r12d
- lea rbp,[20+rbp]
- mov r13d,DWORD[52+rsp]
- mov r15d,DWORD[40+rsp]
- mov r12d,r13d
- ror r13d,11
- add r8d,r14d
- mov r14d,r15d
- ror r15d,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor r15d,r14d
- shr r14d,10
- ror r15d,17
- xor r12d,r13d
- xor r15d,r14d
- add r12d,DWORD[20+rsp]
- add r12d,DWORD[48+rsp]
- mov r13d,eax
- add r12d,r15d
- mov r14d,r8d
- ror r13d,14
- mov r15d,ebx
- xor r13d,eax
- ror r14d,9
- xor r15d,ecx
- mov DWORD[48+rsp],r12d
- xor r14d,r8d
- and r15d,eax
- ror r13d,5
- add r12d,edx
- xor r15d,ecx
- ror r14d,11
- xor r13d,eax
- add r12d,r15d
- mov r15d,r8d
- add r12d,DWORD[rbp]
- xor r14d,r8d
- xor r15d,r9d
- ror r13d,6
- mov edx,r9d
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor edx,edi
- add r11d,r12d
- add edx,r12d
- lea rbp,[4+rbp]
- mov r13d,DWORD[56+rsp]
- mov edi,DWORD[44+rsp]
- mov r12d,r13d
- ror r13d,11
- add edx,r14d
- mov r14d,edi
- ror edi,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor edi,r14d
- shr r14d,10
- ror edi,17
- xor r12d,r13d
- xor edi,r14d
- add r12d,DWORD[24+rsp]
- add r12d,DWORD[52+rsp]
- mov r13d,r11d
- add r12d,edi
- mov r14d,edx
- ror r13d,14
- mov edi,eax
- xor r13d,r11d
- ror r14d,9
- xor edi,ebx
- mov DWORD[52+rsp],r12d
- xor r14d,edx
- and edi,r11d
- ror r13d,5
- add r12d,ecx
- xor edi,ebx
- ror r14d,11
- xor r13d,r11d
- add r12d,edi
- mov edi,edx
- add r12d,DWORD[rbp]
- xor r14d,edx
- xor edi,r8d
- ror r13d,6
- mov ecx,r8d
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor ecx,r15d
- add r10d,r12d
- add ecx,r12d
- lea rbp,[4+rbp]
- mov r13d,DWORD[60+rsp]
- mov r15d,DWORD[48+rsp]
- mov r12d,r13d
- ror r13d,11
- add ecx,r14d
- mov r14d,r15d
- ror r15d,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor r15d,r14d
- shr r14d,10
- ror r15d,17
- xor r12d,r13d
- xor r15d,r14d
- add r12d,DWORD[28+rsp]
- add r12d,DWORD[56+rsp]
- mov r13d,r10d
- add r12d,r15d
- mov r14d,ecx
- ror r13d,14
- mov r15d,r11d
- xor r13d,r10d
- ror r14d,9
- xor r15d,eax
- mov DWORD[56+rsp],r12d
- xor r14d,ecx
- and r15d,r10d
- ror r13d,5
- add r12d,ebx
- xor r15d,eax
- ror r14d,11
- xor r13d,r10d
- add r12d,r15d
- mov r15d,ecx
- add r12d,DWORD[rbp]
- xor r14d,ecx
- xor r15d,edx
- ror r13d,6
- mov ebx,edx
- and edi,r15d
- ror r14d,2
- add r12d,r13d
- xor ebx,edi
- add r9d,r12d
- add ebx,r12d
- lea rbp,[4+rbp]
- mov r13d,DWORD[rsp]
- mov edi,DWORD[52+rsp]
- mov r12d,r13d
- ror r13d,11
- add ebx,r14d
- mov r14d,edi
- ror edi,2
- xor r13d,r12d
- shr r12d,3
- ror r13d,7
- xor edi,r14d
- shr r14d,10
- ror edi,17
- xor r12d,r13d
- xor edi,r14d
- add r12d,DWORD[32+rsp]
- add r12d,DWORD[60+rsp]
- mov r13d,r9d
- add r12d,edi
- mov r14d,ebx
- ror r13d,14
- mov edi,r10d
- xor r13d,r9d
- ror r14d,9
- xor edi,r11d
- mov DWORD[60+rsp],r12d
- xor r14d,ebx
- and edi,r9d
- ror r13d,5
- add r12d,eax
- xor edi,r11d
- ror r14d,11
- xor r13d,r9d
- add r12d,edi
- mov edi,ebx
- add r12d,DWORD[rbp]
- xor r14d,ebx
- xor edi,ecx
- ror r13d,6
- mov eax,ecx
- and r15d,edi
- ror r14d,2
- add r12d,r13d
- xor eax,r15d
- add r8d,r12d
- add eax,r12d
- lea rbp,[20+rbp]
- cmp BYTE[3+rbp],0
- jnz NEAR $L$rounds_16_xx
- mov rdi,QWORD[((64+0))+rsp]
- add eax,r14d
- lea rsi,[64+rsi]
- add eax,DWORD[rdi]
- add ebx,DWORD[4+rdi]
- add ecx,DWORD[8+rdi]
- add edx,DWORD[12+rdi]
- add r8d,DWORD[16+rdi]
- add r9d,DWORD[20+rdi]
- add r10d,DWORD[24+rdi]
- add r11d,DWORD[28+rdi]
- cmp rsi,QWORD[((64+16))+rsp]
- mov DWORD[rdi],eax
- mov DWORD[4+rdi],ebx
- mov DWORD[8+rdi],ecx
- mov DWORD[12+rdi],edx
- mov DWORD[16+rdi],r8d
- mov DWORD[20+rdi],r9d
- mov DWORD[24+rdi],r10d
- mov DWORD[28+rdi],r11d
- jb NEAR $L$loop
- mov rsi,QWORD[88+rsp]
- mov r15,QWORD[((-48))+rsi]
- mov r14,QWORD[((-40))+rsi]
- mov r13,QWORD[((-32))+rsi]
- mov r12,QWORD[((-24))+rsi]
- mov rbp,QWORD[((-16))+rsi]
- mov rbx,QWORD[((-8))+rsi]
- lea rsp,[rsi]
- $L$epilogue:
- mov rdi,QWORD[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_GFp_sha256_block_data_order:
- ALIGN 64
- K256:
- DD 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
- DD 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
- DD 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
- DD 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
- DD 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
- DD 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
- DD 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
- DD 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
- DD 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
- DD 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
- DD 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
- DD 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
- DD 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
- DD 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
- DD 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
- DD 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
- DD 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
- DD 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
- DD 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
- DD 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
- DD 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
- DD 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
- DD 0xd192e819,0xd6990624,0xf40e3585,0x106aa070
- DD 0xd192e819,0xd6990624,0xf40e3585,0x106aa070
- DD 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
- DD 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
- DD 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
- DD 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
- DD 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
- DD 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
- DD 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
- DD 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
- DD 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f
- DD 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f
- DD 0x03020100,0x0b0a0908,0xffffffff,0xffffffff
- DD 0x03020100,0x0b0a0908,0xffffffff,0xffffffff
- DD 0xffffffff,0xffffffff,0x03020100,0x0b0a0908
- DD 0xffffffff,0xffffffff,0x03020100,0x0b0a0908
- DB 83,72,65,50,53,54,32,98,108,111,99,107,32,116,114,97
- DB 110,115,102,111,114,109,32,102,111,114,32,120,56,54,95,54
- DB 52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121
- DB 32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
- DB 111,114,103,62,0
- ALIGN 64
- GFp_sha256_block_data_order_ssse3:
- mov QWORD[8+rsp],rdi ;WIN64 prologue
- mov QWORD[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_GFp_sha256_block_data_order_ssse3:
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- $L$ssse3_shortcut:
- mov rax,rsp
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- shl rdx,4
- sub rsp,160
- lea rdx,[rdx*4+rsi]
- and rsp,-64
- mov QWORD[((64+0))+rsp],rdi
- mov QWORD[((64+8))+rsp],rsi
- mov QWORD[((64+16))+rsp],rdx
- mov QWORD[88+rsp],rax
- movaps XMMWORD[(64+32)+rsp],xmm6
- movaps XMMWORD[(64+48)+rsp],xmm7
- movaps XMMWORD[(64+64)+rsp],xmm8
- movaps XMMWORD[(64+80)+rsp],xmm9
- $L$prologue_ssse3:
- mov eax,DWORD[rdi]
- mov ebx,DWORD[4+rdi]
- mov ecx,DWORD[8+rdi]
- mov edx,DWORD[12+rdi]
- mov r8d,DWORD[16+rdi]
- mov r9d,DWORD[20+rdi]
- mov r10d,DWORD[24+rdi]
- mov r11d,DWORD[28+rdi]
- jmp NEAR $L$loop_ssse3
- ALIGN 16
- $L$loop_ssse3:
- movdqa xmm7,XMMWORD[((K256+512))]
- movdqu xmm0,XMMWORD[rsi]
- movdqu xmm1,XMMWORD[16+rsi]
- movdqu xmm2,XMMWORD[32+rsi]
- DB 102,15,56,0,199
- movdqu xmm3,XMMWORD[48+rsi]
- lea rbp,[K256]
- DB 102,15,56,0,207
- movdqa xmm4,XMMWORD[rbp]
- movdqa xmm5,XMMWORD[32+rbp]
- DB 102,15,56,0,215
- paddd xmm4,xmm0
- movdqa xmm6,XMMWORD[64+rbp]
- DB 102,15,56,0,223
- movdqa xmm7,XMMWORD[96+rbp]
- paddd xmm5,xmm1
- paddd xmm6,xmm2
- paddd xmm7,xmm3
- movdqa XMMWORD[rsp],xmm4
- mov r14d,eax
- movdqa XMMWORD[16+rsp],xmm5
- mov edi,ebx
- movdqa XMMWORD[32+rsp],xmm6
- xor edi,ecx
- movdqa XMMWORD[48+rsp],xmm7
- mov r13d,r8d
- jmp NEAR $L$ssse3_00_47
- ALIGN 16
- $L$ssse3_00_47:
- sub rbp,-128
- ror r13d,14
- movdqa xmm4,xmm1
- mov eax,r14d
- mov r12d,r9d
- movdqa xmm7,xmm3
- ror r14d,9
- xor r13d,r8d
- xor r12d,r10d
- ror r13d,5
- xor r14d,eax
- DB 102,15,58,15,224,4
- and r12d,r8d
- xor r13d,r8d
- DB 102,15,58,15,250,4
- add r11d,DWORD[rsp]
- mov r15d,eax
- xor r12d,r10d
- ror r14d,11
- movdqa xmm5,xmm4
- xor r15d,ebx
- add r11d,r12d
- movdqa xmm6,xmm4
- ror r13d,6
- and edi,r15d
- psrld xmm4,3
- xor r14d,eax
- add r11d,r13d
- xor edi,ebx
- paddd xmm0,xmm7
- ror r14d,2
- add edx,r11d
- psrld xmm6,7
- add r11d,edi
- mov r13d,edx
- pshufd xmm7,xmm3,250
- add r14d,r11d
- ror r13d,14
- pslld xmm5,14
- mov r11d,r14d
- mov r12d,r8d
- pxor xmm4,xmm6
- ror r14d,9
- xor r13d,edx
- xor r12d,r9d
- ror r13d,5
- psrld xmm6,11
- xor r14d,r11d
- pxor xmm4,xmm5
- and r12d,edx
- xor r13d,edx
- pslld xmm5,11
- add r10d,DWORD[4+rsp]
- mov edi,r11d
- pxor xmm4,xmm6
- xor r12d,r9d
- ror r14d,11
- movdqa xmm6,xmm7
- xor edi,eax
- add r10d,r12d
- pxor xmm4,xmm5
- ror r13d,6
- and r15d,edi
- xor r14d,r11d
- psrld xmm7,10
- add r10d,r13d
- xor r15d,eax
- paddd xmm0,xmm4
- ror r14d,2
- add ecx,r10d
- psrlq xmm6,17
- add r10d,r15d
- mov r13d,ecx
- add r14d,r10d
- pxor xmm7,xmm6
- ror r13d,14
- mov r10d,r14d
- mov r12d,edx
- ror r14d,9
- psrlq xmm6,2
- xor r13d,ecx
- xor r12d,r8d
- pxor xmm7,xmm6
- ror r13d,5
- xor r14d,r10d
- and r12d,ecx
- pshufd xmm7,xmm7,128
- xor r13d,ecx
- add r9d,DWORD[8+rsp]
- mov r15d,r10d
- psrldq xmm7,8
- xor r12d,r8d
- ror r14d,11
- xor r15d,r11d
- add r9d,r12d
- ror r13d,6
- paddd xmm0,xmm7
- and edi,r15d
- xor r14d,r10d
- add r9d,r13d
- pshufd xmm7,xmm0,80
- xor edi,r11d
- ror r14d,2
- add ebx,r9d
- movdqa xmm6,xmm7
- add r9d,edi
- mov r13d,ebx
- psrld xmm7,10
- add r14d,r9d
- ror r13d,14
- psrlq xmm6,17
- mov r9d,r14d
- mov r12d,ecx
- pxor xmm7,xmm6
- ror r14d,9
- xor r13d,ebx
- xor r12d,edx
- ror r13d,5
- xor r14d,r9d
- psrlq xmm6,2
- and r12d,ebx
- xor r13d,ebx
- add r8d,DWORD[12+rsp]
- pxor xmm7,xmm6
- mov edi,r9d
- xor r12d,edx
- ror r14d,11
- pshufd xmm7,xmm7,8
- xor edi,r10d
- add r8d,r12d
- movdqa xmm6,XMMWORD[rbp]
- ror r13d,6
- and r15d,edi
- pslldq xmm7,8
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- paddd xmm0,xmm7
- ror r14d,2
- add eax,r8d
- add r8d,r15d
- paddd xmm6,xmm0
- mov r13d,eax
- add r14d,r8d
- movdqa XMMWORD[rsp],xmm6
- ror r13d,14
- movdqa xmm4,xmm2
- mov r8d,r14d
- mov r12d,ebx
- movdqa xmm7,xmm0
- ror r14d,9
- xor r13d,eax
- xor r12d,ecx
- ror r13d,5
- xor r14d,r8d
- DB 102,15,58,15,225,4
- and r12d,eax
- xor r13d,eax
- DB 102,15,58,15,251,4
- add edx,DWORD[16+rsp]
- mov r15d,r8d
- xor r12d,ecx
- ror r14d,11
- movdqa xmm5,xmm4
- xor r15d,r9d
- add edx,r12d
- movdqa xmm6,xmm4
- ror r13d,6
- and edi,r15d
- psrld xmm4,3
- xor r14d,r8d
- add edx,r13d
- xor edi,r9d
- paddd xmm1,xmm7
- ror r14d,2
- add r11d,edx
- psrld xmm6,7
- add edx,edi
- mov r13d,r11d
- pshufd xmm7,xmm0,250
- add r14d,edx
- ror r13d,14
- pslld xmm5,14
- mov edx,r14d
- mov r12d,eax
- pxor xmm4,xmm6
- ror r14d,9
- xor r13d,r11d
- xor r12d,ebx
- ror r13d,5
- psrld xmm6,11
- xor r14d,edx
- pxor xmm4,xmm5
- and r12d,r11d
- xor r13d,r11d
- pslld xmm5,11
- add ecx,DWORD[20+rsp]
- mov edi,edx
- pxor xmm4,xmm6
- xor r12d,ebx
- ror r14d,11
- movdqa xmm6,xmm7
- xor edi,r8d
- add ecx,r12d
- pxor xmm4,xmm5
- ror r13d,6
- and r15d,edi
- xor r14d,edx
- psrld xmm7,10
- add ecx,r13d
- xor r15d,r8d
- paddd xmm1,xmm4
- ror r14d,2
- add r10d,ecx
- psrlq xmm6,17
- add ecx,r15d
- mov r13d,r10d
- add r14d,ecx
- pxor xmm7,xmm6
- ror r13d,14
- mov ecx,r14d
- mov r12d,r11d
- ror r14d,9
- psrlq xmm6,2
- xor r13d,r10d
- xor r12d,eax
- pxor xmm7,xmm6
- ror r13d,5
- xor r14d,ecx
- and r12d,r10d
- pshufd xmm7,xmm7,128
- xor r13d,r10d
- add ebx,DWORD[24+rsp]
- mov r15d,ecx
- psrldq xmm7,8
- xor r12d,eax
- ror r14d,11
- xor r15d,edx
- add ebx,r12d
- ror r13d,6
- paddd xmm1,xmm7
- and edi,r15d
- xor r14d,ecx
- add ebx,r13d
- pshufd xmm7,xmm1,80
- xor edi,edx
- ror r14d,2
- add r9d,ebx
- movdqa xmm6,xmm7
- add ebx,edi
- mov r13d,r9d
- psrld xmm7,10
- add r14d,ebx
- ror r13d,14
- psrlq xmm6,17
- mov ebx,r14d
- mov r12d,r10d
- pxor xmm7,xmm6
- ror r14d,9
- xor r13d,r9d
- xor r12d,r11d
- ror r13d,5
- xor r14d,ebx
- psrlq xmm6,2
- and r12d,r9d
- xor r13d,r9d
- add eax,DWORD[28+rsp]
- pxor xmm7,xmm6
- mov edi,ebx
- xor r12d,r11d
- ror r14d,11
- pshufd xmm7,xmm7,8
- xor edi,ecx
- add eax,r12d
- movdqa xmm6,XMMWORD[32+rbp]
- ror r13d,6
- and r15d,edi
- pslldq xmm7,8
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- paddd xmm1,xmm7
- ror r14d,2
- add r8d,eax
- add eax,r15d
- paddd xmm6,xmm1
- mov r13d,r8d
- add r14d,eax
- movdqa XMMWORD[16+rsp],xmm6
- ror r13d,14
- movdqa xmm4,xmm3
- mov eax,r14d
- mov r12d,r9d
- movdqa xmm7,xmm1
- ror r14d,9
- xor r13d,r8d
- xor r12d,r10d
- ror r13d,5
- xor r14d,eax
- DB 102,15,58,15,226,4
- and r12d,r8d
- xor r13d,r8d
- DB 102,15,58,15,248,4
- add r11d,DWORD[32+rsp]
- mov r15d,eax
- xor r12d,r10d
- ror r14d,11
- movdqa xmm5,xmm4
- xor r15d,ebx
- add r11d,r12d
- movdqa xmm6,xmm4
- ror r13d,6
- and edi,r15d
- psrld xmm4,3
- xor r14d,eax
- add r11d,r13d
- xor edi,ebx
- paddd xmm2,xmm7
- ror r14d,2
- add edx,r11d
- psrld xmm6,7
- add r11d,edi
- mov r13d,edx
- pshufd xmm7,xmm1,250
- add r14d,r11d
- ror r13d,14
- pslld xmm5,14
- mov r11d,r14d
- mov r12d,r8d
- pxor xmm4,xmm6
- ror r14d,9
- xor r13d,edx
- xor r12d,r9d
- ror r13d,5
- psrld xmm6,11
- xor r14d,r11d
- pxor xmm4,xmm5
- and r12d,edx
- xor r13d,edx
- pslld xmm5,11
- add r10d,DWORD[36+rsp]
- mov edi,r11d
- pxor xmm4,xmm6
- xor r12d,r9d
- ror r14d,11
- movdqa xmm6,xmm7
- xor edi,eax
- add r10d,r12d
- pxor xmm4,xmm5
- ror r13d,6
- and r15d,edi
- xor r14d,r11d
- psrld xmm7,10
- add r10d,r13d
- xor r15d,eax
- paddd xmm2,xmm4
- ror r14d,2
- add ecx,r10d
- psrlq xmm6,17
- add r10d,r15d
- mov r13d,ecx
- add r14d,r10d
- pxor xmm7,xmm6
- ror r13d,14
- mov r10d,r14d
- mov r12d,edx
- ror r14d,9
- psrlq xmm6,2
- xor r13d,ecx
- xor r12d,r8d
- pxor xmm7,xmm6
- ror r13d,5
- xor r14d,r10d
- and r12d,ecx
- pshufd xmm7,xmm7,128
- xor r13d,ecx
- add r9d,DWORD[40+rsp]
- mov r15d,r10d
- psrldq xmm7,8
- xor r12d,r8d
- ror r14d,11
- xor r15d,r11d
- add r9d,r12d
- ror r13d,6
- paddd xmm2,xmm7
- and edi,r15d
- xor r14d,r10d
- add r9d,r13d
- pshufd xmm7,xmm2,80
- xor edi,r11d
- ror r14d,2
- add ebx,r9d
- movdqa xmm6,xmm7
- add r9d,edi
- mov r13d,ebx
- psrld xmm7,10
- add r14d,r9d
- ror r13d,14
- psrlq xmm6,17
- mov r9d,r14d
- mov r12d,ecx
- pxor xmm7,xmm6
- ror r14d,9
- xor r13d,ebx
- xor r12d,edx
- ror r13d,5
- xor r14d,r9d
- psrlq xmm6,2
- and r12d,ebx
- xor r13d,ebx
- add r8d,DWORD[44+rsp]
- pxor xmm7,xmm6
- mov edi,r9d
- xor r12d,edx
- ror r14d,11
- pshufd xmm7,xmm7,8
- xor edi,r10d
- add r8d,r12d
- movdqa xmm6,XMMWORD[64+rbp]
- ror r13d,6
- and r15d,edi
- pslldq xmm7,8
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- paddd xmm2,xmm7
- ror r14d,2
- add eax,r8d
- add r8d,r15d
- paddd xmm6,xmm2
- mov r13d,eax
- add r14d,r8d
- movdqa XMMWORD[32+rsp],xmm6
- ror r13d,14
- movdqa xmm4,xmm0
- mov r8d,r14d
- mov r12d,ebx
- movdqa xmm7,xmm2
- ror r14d,9
- xor r13d,eax
- xor r12d,ecx
- ror r13d,5
- xor r14d,r8d
- DB 102,15,58,15,227,4
- and r12d,eax
- xor r13d,eax
- DB 102,15,58,15,249,4
- add edx,DWORD[48+rsp]
- mov r15d,r8d
- xor r12d,ecx
- ror r14d,11
- movdqa xmm5,xmm4
- xor r15d,r9d
- add edx,r12d
- movdqa xmm6,xmm4
- ror r13d,6
- and edi,r15d
- psrld xmm4,3
- xor r14d,r8d
- add edx,r13d
- xor edi,r9d
- paddd xmm3,xmm7
- ror r14d,2
- add r11d,edx
- psrld xmm6,7
- add edx,edi
- mov r13d,r11d
- pshufd xmm7,xmm2,250
- add r14d,edx
- ror r13d,14
- pslld xmm5,14
- mov edx,r14d
- mov r12d,eax
- pxor xmm4,xmm6
- ror r14d,9
- xor r13d,r11d
- xor r12d,ebx
- ror r13d,5
- psrld xmm6,11
- xor r14d,edx
- pxor xmm4,xmm5
- and r12d,r11d
- xor r13d,r11d
- pslld xmm5,11
- add ecx,DWORD[52+rsp]
- mov edi,edx
- pxor xmm4,xmm6
- xor r12d,ebx
- ror r14d,11
- movdqa xmm6,xmm7
- xor edi,r8d
- add ecx,r12d
- pxor xmm4,xmm5
- ror r13d,6
- and r15d,edi
- xor r14d,edx
- psrld xmm7,10
- add ecx,r13d
- xor r15d,r8d
- paddd xmm3,xmm4
- ror r14d,2
- add r10d,ecx
- psrlq xmm6,17
- add ecx,r15d
- mov r13d,r10d
- add r14d,ecx
- pxor xmm7,xmm6
- ror r13d,14
- mov ecx,r14d
- mov r12d,r11d
- ror r14d,9
- psrlq xmm6,2
- xor r13d,r10d
- xor r12d,eax
- pxor xmm7,xmm6
- ror r13d,5
- xor r14d,ecx
- and r12d,r10d
- pshufd xmm7,xmm7,128
- xor r13d,r10d
- add ebx,DWORD[56+rsp]
- mov r15d,ecx
- psrldq xmm7,8
- xor r12d,eax
- ror r14d,11
- xor r15d,edx
- add ebx,r12d
- ror r13d,6
- paddd xmm3,xmm7
- and edi,r15d
- xor r14d,ecx
- add ebx,r13d
- pshufd xmm7,xmm3,80
- xor edi,edx
- ror r14d,2
- add r9d,ebx
- movdqa xmm6,xmm7
- add ebx,edi
- mov r13d,r9d
- psrld xmm7,10
- add r14d,ebx
- ror r13d,14
- psrlq xmm6,17
- mov ebx,r14d
- mov r12d,r10d
- pxor xmm7,xmm6
- ror r14d,9
- xor r13d,r9d
- xor r12d,r11d
- ror r13d,5
- xor r14d,ebx
- psrlq xmm6,2
- and r12d,r9d
- xor r13d,r9d
- add eax,DWORD[60+rsp]
- pxor xmm7,xmm6
- mov edi,ebx
- xor r12d,r11d
- ror r14d,11
- pshufd xmm7,xmm7,8
- xor edi,ecx
- add eax,r12d
- movdqa xmm6,XMMWORD[96+rbp]
- ror r13d,6
- and r15d,edi
- pslldq xmm7,8
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- paddd xmm3,xmm7
- ror r14d,2
- add r8d,eax
- add eax,r15d
- paddd xmm6,xmm3
- mov r13d,r8d
- add r14d,eax
- movdqa XMMWORD[48+rsp],xmm6
- cmp BYTE[131+rbp],0
- jne NEAR $L$ssse3_00_47
- ror r13d,14
- mov eax,r14d
- mov r12d,r9d
- ror r14d,9
- xor r13d,r8d
- xor r12d,r10d
- ror r13d,5
- xor r14d,eax
- and r12d,r8d
- xor r13d,r8d
- add r11d,DWORD[rsp]
- mov r15d,eax
- xor r12d,r10d
- ror r14d,11
- xor r15d,ebx
- add r11d,r12d
- ror r13d,6
- and edi,r15d
- xor r14d,eax
- add r11d,r13d
- xor edi,ebx
- ror r14d,2
- add edx,r11d
- add r11d,edi
- mov r13d,edx
- add r14d,r11d
- ror r13d,14
- mov r11d,r14d
- mov r12d,r8d
- ror r14d,9
- xor r13d,edx
- xor r12d,r9d
- ror r13d,5
- xor r14d,r11d
- and r12d,edx
- xor r13d,edx
- add r10d,DWORD[4+rsp]
- mov edi,r11d
- xor r12d,r9d
- ror r14d,11
- xor edi,eax
- add r10d,r12d
- ror r13d,6
- and r15d,edi
- xor r14d,r11d
- add r10d,r13d
- xor r15d,eax
- ror r14d,2
- add ecx,r10d
- add r10d,r15d
- mov r13d,ecx
- add r14d,r10d
- ror r13d,14
- mov r10d,r14d
- mov r12d,edx
- ror r14d,9
- xor r13d,ecx
- xor r12d,r8d
- ror r13d,5
- xor r14d,r10d
- and r12d,ecx
- xor r13d,ecx
- add r9d,DWORD[8+rsp]
- mov r15d,r10d
- xor r12d,r8d
- ror r14d,11
- xor r15d,r11d
- add r9d,r12d
- ror r13d,6
- and edi,r15d
- xor r14d,r10d
- add r9d,r13d
- xor edi,r11d
- ror r14d,2
- add ebx,r9d
- add r9d,edi
- mov r13d,ebx
- add r14d,r9d
- ror r13d,14
- mov r9d,r14d
- mov r12d,ecx
- ror r14d,9
- xor r13d,ebx
- xor r12d,edx
- ror r13d,5
- xor r14d,r9d
- and r12d,ebx
- xor r13d,ebx
- add r8d,DWORD[12+rsp]
- mov edi,r9d
- xor r12d,edx
- ror r14d,11
- xor edi,r10d
- add r8d,r12d
- ror r13d,6
- and r15d,edi
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- ror r14d,2
- add eax,r8d
- add r8d,r15d
- mov r13d,eax
- add r14d,r8d
- ror r13d,14
- mov r8d,r14d
- mov r12d,ebx
- ror r14d,9
- xor r13d,eax
- xor r12d,ecx
- ror r13d,5
- xor r14d,r8d
- and r12d,eax
- xor r13d,eax
- add edx,DWORD[16+rsp]
- mov r15d,r8d
- xor r12d,ecx
- ror r14d,11
- xor r15d,r9d
- add edx,r12d
- ror r13d,6
- and edi,r15d
- xor r14d,r8d
- add edx,r13d
- xor edi,r9d
- ror r14d,2
- add r11d,edx
- add edx,edi
- mov r13d,r11d
- add r14d,edx
- ror r13d,14
- mov edx,r14d
- mov r12d,eax
- ror r14d,9
- xor r13d,r11d
- xor r12d,ebx
- ror r13d,5
- xor r14d,edx
- and r12d,r11d
- xor r13d,r11d
- add ecx,DWORD[20+rsp]
- mov edi,edx
- xor r12d,ebx
- ror r14d,11
- xor edi,r8d
- add ecx,r12d
- ror r13d,6
- and r15d,edi
- xor r14d,edx
- add ecx,r13d
- xor r15d,r8d
- ror r14d,2
- add r10d,ecx
- add ecx,r15d
- mov r13d,r10d
- add r14d,ecx
- ror r13d,14
- mov ecx,r14d
- mov r12d,r11d
- ror r14d,9
- xor r13d,r10d
- xor r12d,eax
- ror r13d,5
- xor r14d,ecx
- and r12d,r10d
- xor r13d,r10d
- add ebx,DWORD[24+rsp]
- mov r15d,ecx
- xor r12d,eax
- ror r14d,11
- xor r15d,edx
- add ebx,r12d
- ror r13d,6
- and edi,r15d
- xor r14d,ecx
- add ebx,r13d
- xor edi,edx
- ror r14d,2
- add r9d,ebx
- add ebx,edi
- mov r13d,r9d
- add r14d,ebx
- ror r13d,14
- mov ebx,r14d
- mov r12d,r10d
- ror r14d,9
- xor r13d,r9d
- xor r12d,r11d
- ror r13d,5
- xor r14d,ebx
- and r12d,r9d
- xor r13d,r9d
- add eax,DWORD[28+rsp]
- mov edi,ebx
- xor r12d,r11d
- ror r14d,11
- xor edi,ecx
- add eax,r12d
- ror r13d,6
- and r15d,edi
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- ror r14d,2
- add r8d,eax
- add eax,r15d
- mov r13d,r8d
- add r14d,eax
- ror r13d,14
- mov eax,r14d
- mov r12d,r9d
- ror r14d,9
- xor r13d,r8d
- xor r12d,r10d
- ror r13d,5
- xor r14d,eax
- and r12d,r8d
- xor r13d,r8d
- add r11d,DWORD[32+rsp]
- mov r15d,eax
- xor r12d,r10d
- ror r14d,11
- xor r15d,ebx
- add r11d,r12d
- ror r13d,6
- and edi,r15d
- xor r14d,eax
- add r11d,r13d
- xor edi,ebx
- ror r14d,2
- add edx,r11d
- add r11d,edi
- mov r13d,edx
- add r14d,r11d
- ror r13d,14
- mov r11d,r14d
- mov r12d,r8d
- ror r14d,9
- xor r13d,edx
- xor r12d,r9d
- ror r13d,5
- xor r14d,r11d
- and r12d,edx
- xor r13d,edx
- add r10d,DWORD[36+rsp]
- mov edi,r11d
- xor r12d,r9d
- ror r14d,11
- xor edi,eax
- add r10d,r12d
- ror r13d,6
- and r15d,edi
- xor r14d,r11d
- add r10d,r13d
- xor r15d,eax
- ror r14d,2
- add ecx,r10d
- add r10d,r15d
- mov r13d,ecx
- add r14d,r10d
- ror r13d,14
- mov r10d,r14d
- mov r12d,edx
- ror r14d,9
- xor r13d,ecx
- xor r12d,r8d
- ror r13d,5
- xor r14d,r10d
- and r12d,ecx
- xor r13d,ecx
- add r9d,DWORD[40+rsp]
- mov r15d,r10d
- xor r12d,r8d
- ror r14d,11
- xor r15d,r11d
- add r9d,r12d
- ror r13d,6
- and edi,r15d
- xor r14d,r10d
- add r9d,r13d
- xor edi,r11d
- ror r14d,2
- add ebx,r9d
- add r9d,edi
- mov r13d,ebx
- add r14d,r9d
- ror r13d,14
- mov r9d,r14d
- mov r12d,ecx
- ror r14d,9
- xor r13d,ebx
- xor r12d,edx
- ror r13d,5
- xor r14d,r9d
- and r12d,ebx
- xor r13d,ebx
- add r8d,DWORD[44+rsp]
- mov edi,r9d
- xor r12d,edx
- ror r14d,11
- xor edi,r10d
- add r8d,r12d
- ror r13d,6
- and r15d,edi
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- ror r14d,2
- add eax,r8d
- add r8d,r15d
- mov r13d,eax
- add r14d,r8d
- ror r13d,14
- mov r8d,r14d
- mov r12d,ebx
- ror r14d,9
- xor r13d,eax
- xor r12d,ecx
- ror r13d,5
- xor r14d,r8d
- and r12d,eax
- xor r13d,eax
- add edx,DWORD[48+rsp]
- mov r15d,r8d
- xor r12d,ecx
- ror r14d,11
- xor r15d,r9d
- add edx,r12d
- ror r13d,6
- and edi,r15d
- xor r14d,r8d
- add edx,r13d
- xor edi,r9d
- ror r14d,2
- add r11d,edx
- add edx,edi
- mov r13d,r11d
- add r14d,edx
- ror r13d,14
- mov edx,r14d
- mov r12d,eax
- ror r14d,9
- xor r13d,r11d
- xor r12d,ebx
- ror r13d,5
- xor r14d,edx
- and r12d,r11d
- xor r13d,r11d
- add ecx,DWORD[52+rsp]
- mov edi,edx
- xor r12d,ebx
- ror r14d,11
- xor edi,r8d
- add ecx,r12d
- ror r13d,6
- and r15d,edi
- xor r14d,edx
- add ecx,r13d
- xor r15d,r8d
- ror r14d,2
- add r10d,ecx
- add ecx,r15d
- mov r13d,r10d
- add r14d,ecx
- ror r13d,14
- mov ecx,r14d
- mov r12d,r11d
- ror r14d,9
- xor r13d,r10d
- xor r12d,eax
- ror r13d,5
- xor r14d,ecx
- and r12d,r10d
- xor r13d,r10d
- add ebx,DWORD[56+rsp]
- mov r15d,ecx
- xor r12d,eax
- ror r14d,11
- xor r15d,edx
- add ebx,r12d
- ror r13d,6
- and edi,r15d
- xor r14d,ecx
- add ebx,r13d
- xor edi,edx
- ror r14d,2
- add r9d,ebx
- add ebx,edi
- mov r13d,r9d
- add r14d,ebx
- ror r13d,14
- mov ebx,r14d
- mov r12d,r10d
- ror r14d,9
- xor r13d,r9d
- xor r12d,r11d
- ror r13d,5
- xor r14d,ebx
- and r12d,r9d
- xor r13d,r9d
- add eax,DWORD[60+rsp]
- mov edi,ebx
- xor r12d,r11d
- ror r14d,11
- xor edi,ecx
- add eax,r12d
- ror r13d,6
- and r15d,edi
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- ror r14d,2
- add r8d,eax
- add eax,r15d
- mov r13d,r8d
- add r14d,eax
- mov rdi,QWORD[((64+0))+rsp]
- mov eax,r14d
- add eax,DWORD[rdi]
- lea rsi,[64+rsi]
- add ebx,DWORD[4+rdi]
- add ecx,DWORD[8+rdi]
- add edx,DWORD[12+rdi]
- add r8d,DWORD[16+rdi]
- add r9d,DWORD[20+rdi]
- add r10d,DWORD[24+rdi]
- add r11d,DWORD[28+rdi]
- cmp rsi,QWORD[((64+16))+rsp]
- mov DWORD[rdi],eax
- mov DWORD[4+rdi],ebx
- mov DWORD[8+rdi],ecx
- mov DWORD[12+rdi],edx
- mov DWORD[16+rdi],r8d
- mov DWORD[20+rdi],r9d
- mov DWORD[24+rdi],r10d
- mov DWORD[28+rdi],r11d
- jb NEAR $L$loop_ssse3
- mov rsi,QWORD[88+rsp]
- movaps xmm6,XMMWORD[((64+32))+rsp]
- movaps xmm7,XMMWORD[((64+48))+rsp]
- movaps xmm8,XMMWORD[((64+64))+rsp]
- movaps xmm9,XMMWORD[((64+80))+rsp]
- mov r15,QWORD[((-48))+rsi]
- mov r14,QWORD[((-40))+rsi]
- mov r13,QWORD[((-32))+rsi]
- mov r12,QWORD[((-24))+rsi]
- mov rbp,QWORD[((-16))+rsi]
- mov rbx,QWORD[((-8))+rsi]
- lea rsp,[rsi]
- $L$epilogue_ssse3:
- mov rdi,QWORD[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_GFp_sha256_block_data_order_ssse3:
- ALIGN 64
- GFp_sha256_block_data_order_avx:
- mov QWORD[8+rsp],rdi ;WIN64 prologue
- mov QWORD[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_GFp_sha256_block_data_order_avx:
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- $L$avx_shortcut:
- mov rax,rsp
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- shl rdx,4
- sub rsp,160
- lea rdx,[rdx*4+rsi]
- and rsp,-64
- mov QWORD[((64+0))+rsp],rdi
- mov QWORD[((64+8))+rsp],rsi
- mov QWORD[((64+16))+rsp],rdx
- mov QWORD[88+rsp],rax
- movaps XMMWORD[(64+32)+rsp],xmm6
- movaps XMMWORD[(64+48)+rsp],xmm7
- movaps XMMWORD[(64+64)+rsp],xmm8
- movaps XMMWORD[(64+80)+rsp],xmm9
- $L$prologue_avx:
- vzeroupper
- mov eax,DWORD[rdi]
- mov ebx,DWORD[4+rdi]
- mov ecx,DWORD[8+rdi]
- mov edx,DWORD[12+rdi]
- mov r8d,DWORD[16+rdi]
- mov r9d,DWORD[20+rdi]
- mov r10d,DWORD[24+rdi]
- mov r11d,DWORD[28+rdi]
- vmovdqa xmm8,XMMWORD[((K256+512+32))]
- vmovdqa xmm9,XMMWORD[((K256+512+64))]
- jmp NEAR $L$loop_avx
- ALIGN 16
- $L$loop_avx:
- vmovdqa xmm7,XMMWORD[((K256+512))]
- vmovdqu xmm0,XMMWORD[rsi]
- vmovdqu xmm1,XMMWORD[16+rsi]
- vmovdqu xmm2,XMMWORD[32+rsi]
- vmovdqu xmm3,XMMWORD[48+rsi]
- vpshufb xmm0,xmm0,xmm7
- lea rbp,[K256]
- vpshufb xmm1,xmm1,xmm7
- vpshufb xmm2,xmm2,xmm7
- vpaddd xmm4,xmm0,XMMWORD[rbp]
- vpshufb xmm3,xmm3,xmm7
- vpaddd xmm5,xmm1,XMMWORD[32+rbp]
- vpaddd xmm6,xmm2,XMMWORD[64+rbp]
- vpaddd xmm7,xmm3,XMMWORD[96+rbp]
- vmovdqa XMMWORD[rsp],xmm4
- mov r14d,eax
- vmovdqa XMMWORD[16+rsp],xmm5
- mov edi,ebx
- vmovdqa XMMWORD[32+rsp],xmm6
- xor edi,ecx
- vmovdqa XMMWORD[48+rsp],xmm7
- mov r13d,r8d
- jmp NEAR $L$avx_00_47
- ALIGN 16
- $L$avx_00_47:
- sub rbp,-128
- vpalignr xmm4,xmm1,xmm0,4
- shrd r13d,r13d,14
- mov eax,r14d
- mov r12d,r9d
- vpalignr xmm7,xmm3,xmm2,4
- shrd r14d,r14d,9
- xor r13d,r8d
- xor r12d,r10d
- vpsrld xmm6,xmm4,7
- shrd r13d,r13d,5
- xor r14d,eax
- and r12d,r8d
- vpaddd xmm0,xmm0,xmm7
- xor r13d,r8d
- add r11d,DWORD[rsp]
- mov r15d,eax
- vpsrld xmm7,xmm4,3
- xor r12d,r10d
- shrd r14d,r14d,11
- xor r15d,ebx
- vpslld xmm5,xmm4,14
- add r11d,r12d
- shrd r13d,r13d,6
- and edi,r15d
- vpxor xmm4,xmm7,xmm6
- xor r14d,eax
- add r11d,r13d
- xor edi,ebx
- vpshufd xmm7,xmm3,250
- shrd r14d,r14d,2
- add edx,r11d
- add r11d,edi
- vpsrld xmm6,xmm6,11
- mov r13d,edx
- add r14d,r11d
- shrd r13d,r13d,14
- vpxor xmm4,xmm4,xmm5
- mov r11d,r14d
- mov r12d,r8d
- shrd r14d,r14d,9
- vpslld xmm5,xmm5,11
- xor r13d,edx
- xor r12d,r9d
- shrd r13d,r13d,5
- vpxor xmm4,xmm4,xmm6
- xor r14d,r11d
- and r12d,edx
- xor r13d,edx
- vpsrld xmm6,xmm7,10
- add r10d,DWORD[4+rsp]
- mov edi,r11d
- xor r12d,r9d
- vpxor xmm4,xmm4,xmm5
- shrd r14d,r14d,11
- xor edi,eax
- add r10d,r12d
- vpsrlq xmm7,xmm7,17
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,r11d
- vpaddd xmm0,xmm0,xmm4
- add r10d,r13d
- xor r15d,eax
- shrd r14d,r14d,2
- vpxor xmm6,xmm6,xmm7
- add ecx,r10d
- 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
- shrd r14d,r14d,9
- xor r13d,ecx
- vpshufb xmm6,xmm6,xmm8
- xor r12d,r8d
- shrd r13d,r13d,5
- xor r14d,r10d
- vpaddd xmm0,xmm0,xmm6
- and r12d,ecx
- xor r13d,ecx
- add r9d,DWORD[8+rsp]
- vpshufd xmm7,xmm0,80
- mov r15d,r10d
- xor r12d,r8d
- shrd r14d,r14d,11
- vpsrld xmm6,xmm7,10
- xor r15d,r11d
- add r9d,r12d
- shrd r13d,r13d,6
- vpsrlq xmm7,xmm7,17
- and edi,r15d
- xor r14d,r10d
- add r9d,r13d
- vpxor xmm6,xmm6,xmm7
- xor edi,r11d
- shrd r14d,r14d,2
- add ebx,r9d
- vpsrlq xmm7,xmm7,2
- add r9d,edi
- mov r13d,ebx
- add r14d,r9d
- vpxor xmm6,xmm6,xmm7
- shrd r13d,r13d,14
- mov r9d,r14d
- mov r12d,ecx
- vpshufb xmm6,xmm6,xmm9
- shrd r14d,r14d,9
- xor r13d,ebx
- xor r12d,edx
- vpaddd xmm0,xmm0,xmm6
- shrd r13d,r13d,5
- xor r14d,r9d
- and r12d,ebx
- vpaddd xmm6,xmm0,XMMWORD[rbp]
- xor r13d,ebx
- add r8d,DWORD[12+rsp]
- mov edi,r9d
- xor r12d,edx
- shrd r14d,r14d,11
- xor edi,r10d
- add r8d,r12d
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- shrd r14d,r14d,2
- add eax,r8d
- add r8d,r15d
- mov r13d,eax
- add r14d,r8d
- vmovdqa XMMWORD[rsp],xmm6
- vpalignr xmm4,xmm2,xmm1,4
- shrd r13d,r13d,14
- mov r8d,r14d
- mov r12d,ebx
- vpalignr xmm7,xmm0,xmm3,4
- shrd r14d,r14d,9
- xor r13d,eax
- xor r12d,ecx
- vpsrld xmm6,xmm4,7
- shrd r13d,r13d,5
- xor r14d,r8d
- and r12d,eax
- vpaddd xmm1,xmm1,xmm7
- xor r13d,eax
- add edx,DWORD[16+rsp]
- mov r15d,r8d
- vpsrld xmm7,xmm4,3
- xor r12d,ecx
- shrd r14d,r14d,11
- xor r15d,r9d
- vpslld xmm5,xmm4,14
- add edx,r12d
- shrd r13d,r13d,6
- and edi,r15d
- vpxor xmm4,xmm7,xmm6
- xor r14d,r8d
- add edx,r13d
- xor edi,r9d
- vpshufd xmm7,xmm0,250
- shrd r14d,r14d,2
- add r11d,edx
- add edx,edi
- vpsrld xmm6,xmm6,11
- mov r13d,r11d
- add r14d,edx
- shrd r13d,r13d,14
- vpxor xmm4,xmm4,xmm5
- mov edx,r14d
- mov r12d,eax
- shrd r14d,r14d,9
- vpslld xmm5,xmm5,11
- xor r13d,r11d
- xor r12d,ebx
- shrd r13d,r13d,5
- vpxor xmm4,xmm4,xmm6
- xor r14d,edx
- and r12d,r11d
- xor r13d,r11d
- vpsrld xmm6,xmm7,10
- add ecx,DWORD[20+rsp]
- mov edi,edx
- xor r12d,ebx
- vpxor xmm4,xmm4,xmm5
- shrd r14d,r14d,11
- xor edi,r8d
- add ecx,r12d
- vpsrlq xmm7,xmm7,17
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,edx
- vpaddd xmm1,xmm1,xmm4
- add ecx,r13d
- xor r15d,r8d
- shrd r14d,r14d,2
- vpxor xmm6,xmm6,xmm7
- add r10d,ecx
- 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
- shrd r14d,r14d,9
- xor r13d,r10d
- vpshufb xmm6,xmm6,xmm8
- xor r12d,eax
- shrd r13d,r13d,5
- xor r14d,ecx
- vpaddd xmm1,xmm1,xmm6
- and r12d,r10d
- xor r13d,r10d
- add ebx,DWORD[24+rsp]
- vpshufd xmm7,xmm1,80
- mov r15d,ecx
- xor r12d,eax
- shrd r14d,r14d,11
- vpsrld xmm6,xmm7,10
- xor r15d,edx
- add ebx,r12d
- shrd r13d,r13d,6
- vpsrlq xmm7,xmm7,17
- and edi,r15d
- xor r14d,ecx
- add ebx,r13d
- vpxor xmm6,xmm6,xmm7
- xor edi,edx
- shrd r14d,r14d,2
- add r9d,ebx
- vpsrlq xmm7,xmm7,2
- add ebx,edi
- mov r13d,r9d
- add r14d,ebx
- vpxor xmm6,xmm6,xmm7
- shrd r13d,r13d,14
- mov ebx,r14d
- mov r12d,r10d
- vpshufb xmm6,xmm6,xmm9
- shrd r14d,r14d,9
- xor r13d,r9d
- xor r12d,r11d
- vpaddd xmm1,xmm1,xmm6
- shrd r13d,r13d,5
- xor r14d,ebx
- and r12d,r9d
- vpaddd xmm6,xmm1,XMMWORD[32+rbp]
- xor r13d,r9d
- add eax,DWORD[28+rsp]
- mov edi,ebx
- xor r12d,r11d
- shrd r14d,r14d,11
- xor edi,ecx
- add eax,r12d
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- shrd r14d,r14d,2
- add r8d,eax
- add eax,r15d
- mov r13d,r8d
- add r14d,eax
- vmovdqa XMMWORD[16+rsp],xmm6
- vpalignr xmm4,xmm3,xmm2,4
- shrd r13d,r13d,14
- mov eax,r14d
- mov r12d,r9d
- vpalignr xmm7,xmm1,xmm0,4
- shrd r14d,r14d,9
- xor r13d,r8d
- xor r12d,r10d
- vpsrld xmm6,xmm4,7
- shrd r13d,r13d,5
- xor r14d,eax
- and r12d,r8d
- vpaddd xmm2,xmm2,xmm7
- xor r13d,r8d
- add r11d,DWORD[32+rsp]
- mov r15d,eax
- vpsrld xmm7,xmm4,3
- xor r12d,r10d
- shrd r14d,r14d,11
- xor r15d,ebx
- vpslld xmm5,xmm4,14
- add r11d,r12d
- shrd r13d,r13d,6
- and edi,r15d
- vpxor xmm4,xmm7,xmm6
- xor r14d,eax
- add r11d,r13d
- xor edi,ebx
- vpshufd xmm7,xmm1,250
- shrd r14d,r14d,2
- add edx,r11d
- add r11d,edi
- vpsrld xmm6,xmm6,11
- mov r13d,edx
- add r14d,r11d
- shrd r13d,r13d,14
- vpxor xmm4,xmm4,xmm5
- mov r11d,r14d
- mov r12d,r8d
- shrd r14d,r14d,9
- vpslld xmm5,xmm5,11
- xor r13d,edx
- xor r12d,r9d
- shrd r13d,r13d,5
- vpxor xmm4,xmm4,xmm6
- xor r14d,r11d
- and r12d,edx
- xor r13d,edx
- vpsrld xmm6,xmm7,10
- add r10d,DWORD[36+rsp]
- mov edi,r11d
- xor r12d,r9d
- vpxor xmm4,xmm4,xmm5
- shrd r14d,r14d,11
- xor edi,eax
- add r10d,r12d
- vpsrlq xmm7,xmm7,17
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,r11d
- vpaddd xmm2,xmm2,xmm4
- add r10d,r13d
- xor r15d,eax
- shrd r14d,r14d,2
- vpxor xmm6,xmm6,xmm7
- add ecx,r10d
- 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
- shrd r14d,r14d,9
- xor r13d,ecx
- vpshufb xmm6,xmm6,xmm8
- xor r12d,r8d
- shrd r13d,r13d,5
- xor r14d,r10d
- vpaddd xmm2,xmm2,xmm6
- and r12d,ecx
- xor r13d,ecx
- add r9d,DWORD[40+rsp]
- vpshufd xmm7,xmm2,80
- mov r15d,r10d
- xor r12d,r8d
- shrd r14d,r14d,11
- vpsrld xmm6,xmm7,10
- xor r15d,r11d
- add r9d,r12d
- shrd r13d,r13d,6
- vpsrlq xmm7,xmm7,17
- and edi,r15d
- xor r14d,r10d
- add r9d,r13d
- vpxor xmm6,xmm6,xmm7
- xor edi,r11d
- shrd r14d,r14d,2
- add ebx,r9d
- vpsrlq xmm7,xmm7,2
- add r9d,edi
- mov r13d,ebx
- add r14d,r9d
- vpxor xmm6,xmm6,xmm7
- shrd r13d,r13d,14
- mov r9d,r14d
- mov r12d,ecx
- vpshufb xmm6,xmm6,xmm9
- shrd r14d,r14d,9
- xor r13d,ebx
- xor r12d,edx
- vpaddd xmm2,xmm2,xmm6
- shrd r13d,r13d,5
- xor r14d,r9d
- and r12d,ebx
- vpaddd xmm6,xmm2,XMMWORD[64+rbp]
- xor r13d,ebx
- add r8d,DWORD[44+rsp]
- mov edi,r9d
- xor r12d,edx
- shrd r14d,r14d,11
- xor edi,r10d
- add r8d,r12d
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- shrd r14d,r14d,2
- add eax,r8d
- add r8d,r15d
- mov r13d,eax
- add r14d,r8d
- vmovdqa XMMWORD[32+rsp],xmm6
- vpalignr xmm4,xmm0,xmm3,4
- shrd r13d,r13d,14
- mov r8d,r14d
- mov r12d,ebx
- vpalignr xmm7,xmm2,xmm1,4
- shrd r14d,r14d,9
- xor r13d,eax
- xor r12d,ecx
- vpsrld xmm6,xmm4,7
- shrd r13d,r13d,5
- xor r14d,r8d
- and r12d,eax
- vpaddd xmm3,xmm3,xmm7
- xor r13d,eax
- add edx,DWORD[48+rsp]
- mov r15d,r8d
- vpsrld xmm7,xmm4,3
- xor r12d,ecx
- shrd r14d,r14d,11
- xor r15d,r9d
- vpslld xmm5,xmm4,14
- add edx,r12d
- shrd r13d,r13d,6
- and edi,r15d
- vpxor xmm4,xmm7,xmm6
- xor r14d,r8d
- add edx,r13d
- xor edi,r9d
- vpshufd xmm7,xmm2,250
- shrd r14d,r14d,2
- add r11d,edx
- add edx,edi
- vpsrld xmm6,xmm6,11
- mov r13d,r11d
- add r14d,edx
- shrd r13d,r13d,14
- vpxor xmm4,xmm4,xmm5
- mov edx,r14d
- mov r12d,eax
- shrd r14d,r14d,9
- vpslld xmm5,xmm5,11
- xor r13d,r11d
- xor r12d,ebx
- shrd r13d,r13d,5
- vpxor xmm4,xmm4,xmm6
- xor r14d,edx
- and r12d,r11d
- xor r13d,r11d
- vpsrld xmm6,xmm7,10
- add ecx,DWORD[52+rsp]
- mov edi,edx
- xor r12d,ebx
- vpxor xmm4,xmm4,xmm5
- shrd r14d,r14d,11
- xor edi,r8d
- add ecx,r12d
- vpsrlq xmm7,xmm7,17
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,edx
- vpaddd xmm3,xmm3,xmm4
- add ecx,r13d
- xor r15d,r8d
- shrd r14d,r14d,2
- vpxor xmm6,xmm6,xmm7
- add r10d,ecx
- 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
- shrd r14d,r14d,9
- xor r13d,r10d
- vpshufb xmm6,xmm6,xmm8
- xor r12d,eax
- shrd r13d,r13d,5
- xor r14d,ecx
- vpaddd xmm3,xmm3,xmm6
- and r12d,r10d
- xor r13d,r10d
- add ebx,DWORD[56+rsp]
- vpshufd xmm7,xmm3,80
- mov r15d,ecx
- xor r12d,eax
- shrd r14d,r14d,11
- vpsrld xmm6,xmm7,10
- xor r15d,edx
- add ebx,r12d
- shrd r13d,r13d,6
- vpsrlq xmm7,xmm7,17
- and edi,r15d
- xor r14d,ecx
- add ebx,r13d
- vpxor xmm6,xmm6,xmm7
- xor edi,edx
- shrd r14d,r14d,2
- add r9d,ebx
- vpsrlq xmm7,xmm7,2
- add ebx,edi
- mov r13d,r9d
- add r14d,ebx
- vpxor xmm6,xmm6,xmm7
- shrd r13d,r13d,14
- mov ebx,r14d
- mov r12d,r10d
- vpshufb xmm6,xmm6,xmm9
- shrd r14d,r14d,9
- xor r13d,r9d
- xor r12d,r11d
- vpaddd xmm3,xmm3,xmm6
- shrd r13d,r13d,5
- xor r14d,ebx
- and r12d,r9d
- vpaddd xmm6,xmm3,XMMWORD[96+rbp]
- xor r13d,r9d
- add eax,DWORD[60+rsp]
- mov edi,ebx
- xor r12d,r11d
- shrd r14d,r14d,11
- xor edi,ecx
- add eax,r12d
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- shrd r14d,r14d,2
- add r8d,eax
- add eax,r15d
- mov r13d,r8d
- add r14d,eax
- vmovdqa XMMWORD[48+rsp],xmm6
- cmp BYTE[131+rbp],0
- jne NEAR $L$avx_00_47
- shrd r13d,r13d,14
- mov eax,r14d
- mov r12d,r9d
- shrd r14d,r14d,9
- xor r13d,r8d
- xor r12d,r10d
- shrd r13d,r13d,5
- xor r14d,eax
- and r12d,r8d
- xor r13d,r8d
- add r11d,DWORD[rsp]
- mov r15d,eax
- xor r12d,r10d
- shrd r14d,r14d,11
- xor r15d,ebx
- add r11d,r12d
- shrd r13d,r13d,6
- and edi,r15d
- xor r14d,eax
- add r11d,r13d
- xor edi,ebx
- shrd r14d,r14d,2
- add edx,r11d
- add r11d,edi
- mov r13d,edx
- add r14d,r11d
- shrd r13d,r13d,14
- mov r11d,r14d
- mov r12d,r8d
- shrd r14d,r14d,9
- xor r13d,edx
- xor r12d,r9d
- shrd r13d,r13d,5
- xor r14d,r11d
- and r12d,edx
- xor r13d,edx
- add r10d,DWORD[4+rsp]
- mov edi,r11d
- xor r12d,r9d
- shrd r14d,r14d,11
- xor edi,eax
- add r10d,r12d
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,r11d
- add r10d,r13d
- xor r15d,eax
- shrd r14d,r14d,2
- add ecx,r10d
- add r10d,r15d
- mov r13d,ecx
- add r14d,r10d
- shrd r13d,r13d,14
- mov r10d,r14d
- mov r12d,edx
- shrd r14d,r14d,9
- xor r13d,ecx
- xor r12d,r8d
- shrd r13d,r13d,5
- xor r14d,r10d
- and r12d,ecx
- xor r13d,ecx
- add r9d,DWORD[8+rsp]
- mov r15d,r10d
- xor r12d,r8d
- shrd r14d,r14d,11
- xor r15d,r11d
- add r9d,r12d
- shrd r13d,r13d,6
- and edi,r15d
- xor r14d,r10d
- add r9d,r13d
- xor edi,r11d
- shrd r14d,r14d,2
- add ebx,r9d
- add r9d,edi
- mov r13d,ebx
- add r14d,r9d
- shrd r13d,r13d,14
- mov r9d,r14d
- mov r12d,ecx
- shrd r14d,r14d,9
- xor r13d,ebx
- xor r12d,edx
- shrd r13d,r13d,5
- xor r14d,r9d
- and r12d,ebx
- xor r13d,ebx
- add r8d,DWORD[12+rsp]
- mov edi,r9d
- xor r12d,edx
- shrd r14d,r14d,11
- xor edi,r10d
- add r8d,r12d
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- shrd r14d,r14d,2
- add eax,r8d
- add r8d,r15d
- mov r13d,eax
- add r14d,r8d
- shrd r13d,r13d,14
- mov r8d,r14d
- mov r12d,ebx
- shrd r14d,r14d,9
- xor r13d,eax
- xor r12d,ecx
- shrd r13d,r13d,5
- xor r14d,r8d
- and r12d,eax
- xor r13d,eax
- add edx,DWORD[16+rsp]
- mov r15d,r8d
- xor r12d,ecx
- shrd r14d,r14d,11
- xor r15d,r9d
- add edx,r12d
- shrd r13d,r13d,6
- and edi,r15d
- xor r14d,r8d
- add edx,r13d
- xor edi,r9d
- shrd r14d,r14d,2
- add r11d,edx
- add edx,edi
- mov r13d,r11d
- add r14d,edx
- shrd r13d,r13d,14
- mov edx,r14d
- mov r12d,eax
- shrd r14d,r14d,9
- xor r13d,r11d
- xor r12d,ebx
- shrd r13d,r13d,5
- xor r14d,edx
- and r12d,r11d
- xor r13d,r11d
- add ecx,DWORD[20+rsp]
- mov edi,edx
- xor r12d,ebx
- shrd r14d,r14d,11
- xor edi,r8d
- add ecx,r12d
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,edx
- add ecx,r13d
- xor r15d,r8d
- shrd r14d,r14d,2
- add r10d,ecx
- add ecx,r15d
- mov r13d,r10d
- add r14d,ecx
- shrd r13d,r13d,14
- mov ecx,r14d
- mov r12d,r11d
- shrd r14d,r14d,9
- xor r13d,r10d
- xor r12d,eax
- shrd r13d,r13d,5
- xor r14d,ecx
- and r12d,r10d
- xor r13d,r10d
- add ebx,DWORD[24+rsp]
- mov r15d,ecx
- xor r12d,eax
- shrd r14d,r14d,11
- xor r15d,edx
- add ebx,r12d
- shrd r13d,r13d,6
- and edi,r15d
- xor r14d,ecx
- add ebx,r13d
- xor edi,edx
- shrd r14d,r14d,2
- add r9d,ebx
- add ebx,edi
- mov r13d,r9d
- add r14d,ebx
- shrd r13d,r13d,14
- mov ebx,r14d
- mov r12d,r10d
- shrd r14d,r14d,9
- xor r13d,r9d
- xor r12d,r11d
- shrd r13d,r13d,5
- xor r14d,ebx
- and r12d,r9d
- xor r13d,r9d
- add eax,DWORD[28+rsp]
- mov edi,ebx
- xor r12d,r11d
- shrd r14d,r14d,11
- xor edi,ecx
- add eax,r12d
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- shrd r14d,r14d,2
- add r8d,eax
- add eax,r15d
- mov r13d,r8d
- add r14d,eax
- shrd r13d,r13d,14
- mov eax,r14d
- mov r12d,r9d
- shrd r14d,r14d,9
- xor r13d,r8d
- xor r12d,r10d
- shrd r13d,r13d,5
- xor r14d,eax
- and r12d,r8d
- xor r13d,r8d
- add r11d,DWORD[32+rsp]
- mov r15d,eax
- xor r12d,r10d
- shrd r14d,r14d,11
- xor r15d,ebx
- add r11d,r12d
- shrd r13d,r13d,6
- and edi,r15d
- xor r14d,eax
- add r11d,r13d
- xor edi,ebx
- shrd r14d,r14d,2
- add edx,r11d
- add r11d,edi
- mov r13d,edx
- add r14d,r11d
- shrd r13d,r13d,14
- mov r11d,r14d
- mov r12d,r8d
- shrd r14d,r14d,9
- xor r13d,edx
- xor r12d,r9d
- shrd r13d,r13d,5
- xor r14d,r11d
- and r12d,edx
- xor r13d,edx
- add r10d,DWORD[36+rsp]
- mov edi,r11d
- xor r12d,r9d
- shrd r14d,r14d,11
- xor edi,eax
- add r10d,r12d
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,r11d
- add r10d,r13d
- xor r15d,eax
- shrd r14d,r14d,2
- add ecx,r10d
- add r10d,r15d
- mov r13d,ecx
- add r14d,r10d
- shrd r13d,r13d,14
- mov r10d,r14d
- mov r12d,edx
- shrd r14d,r14d,9
- xor r13d,ecx
- xor r12d,r8d
- shrd r13d,r13d,5
- xor r14d,r10d
- and r12d,ecx
- xor r13d,ecx
- add r9d,DWORD[40+rsp]
- mov r15d,r10d
- xor r12d,r8d
- shrd r14d,r14d,11
- xor r15d,r11d
- add r9d,r12d
- shrd r13d,r13d,6
- and edi,r15d
- xor r14d,r10d
- add r9d,r13d
- xor edi,r11d
- shrd r14d,r14d,2
- add ebx,r9d
- add r9d,edi
- mov r13d,ebx
- add r14d,r9d
- shrd r13d,r13d,14
- mov r9d,r14d
- mov r12d,ecx
- shrd r14d,r14d,9
- xor r13d,ebx
- xor r12d,edx
- shrd r13d,r13d,5
- xor r14d,r9d
- and r12d,ebx
- xor r13d,ebx
- add r8d,DWORD[44+rsp]
- mov edi,r9d
- xor r12d,edx
- shrd r14d,r14d,11
- xor edi,r10d
- add r8d,r12d
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,r9d
- add r8d,r13d
- xor r15d,r10d
- shrd r14d,r14d,2
- add eax,r8d
- add r8d,r15d
- mov r13d,eax
- add r14d,r8d
- shrd r13d,r13d,14
- mov r8d,r14d
- mov r12d,ebx
- shrd r14d,r14d,9
- xor r13d,eax
- xor r12d,ecx
- shrd r13d,r13d,5
- xor r14d,r8d
- and r12d,eax
- xor r13d,eax
- add edx,DWORD[48+rsp]
- mov r15d,r8d
- xor r12d,ecx
- shrd r14d,r14d,11
- xor r15d,r9d
- add edx,r12d
- shrd r13d,r13d,6
- and edi,r15d
- xor r14d,r8d
- add edx,r13d
- xor edi,r9d
- shrd r14d,r14d,2
- add r11d,edx
- add edx,edi
- mov r13d,r11d
- add r14d,edx
- shrd r13d,r13d,14
- mov edx,r14d
- mov r12d,eax
- shrd r14d,r14d,9
- xor r13d,r11d
- xor r12d,ebx
- shrd r13d,r13d,5
- xor r14d,edx
- and r12d,r11d
- xor r13d,r11d
- add ecx,DWORD[52+rsp]
- mov edi,edx
- xor r12d,ebx
- shrd r14d,r14d,11
- xor edi,r8d
- add ecx,r12d
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,edx
- add ecx,r13d
- xor r15d,r8d
- shrd r14d,r14d,2
- add r10d,ecx
- add ecx,r15d
- mov r13d,r10d
- add r14d,ecx
- shrd r13d,r13d,14
- mov ecx,r14d
- mov r12d,r11d
- shrd r14d,r14d,9
- xor r13d,r10d
- xor r12d,eax
- shrd r13d,r13d,5
- xor r14d,ecx
- and r12d,r10d
- xor r13d,r10d
- add ebx,DWORD[56+rsp]
- mov r15d,ecx
- xor r12d,eax
- shrd r14d,r14d,11
- xor r15d,edx
- add ebx,r12d
- shrd r13d,r13d,6
- and edi,r15d
- xor r14d,ecx
- add ebx,r13d
- xor edi,edx
- shrd r14d,r14d,2
- add r9d,ebx
- add ebx,edi
- mov r13d,r9d
- add r14d,ebx
- shrd r13d,r13d,14
- mov ebx,r14d
- mov r12d,r10d
- shrd r14d,r14d,9
- xor r13d,r9d
- xor r12d,r11d
- shrd r13d,r13d,5
- xor r14d,ebx
- and r12d,r9d
- xor r13d,r9d
- add eax,DWORD[60+rsp]
- mov edi,ebx
- xor r12d,r11d
- shrd r14d,r14d,11
- xor edi,ecx
- add eax,r12d
- shrd r13d,r13d,6
- and r15d,edi
- xor r14d,ebx
- add eax,r13d
- xor r15d,ecx
- shrd r14d,r14d,2
- add r8d,eax
- add eax,r15d
- mov r13d,r8d
- add r14d,eax
- mov rdi,QWORD[((64+0))+rsp]
- mov eax,r14d
- add eax,DWORD[rdi]
- lea rsi,[64+rsi]
- add ebx,DWORD[4+rdi]
- add ecx,DWORD[8+rdi]
- add edx,DWORD[12+rdi]
- add r8d,DWORD[16+rdi]
- add r9d,DWORD[20+rdi]
- add r10d,DWORD[24+rdi]
- add r11d,DWORD[28+rdi]
- cmp rsi,QWORD[((64+16))+rsp]
- mov DWORD[rdi],eax
- mov DWORD[4+rdi],ebx
- mov DWORD[8+rdi],ecx
- mov DWORD[12+rdi],edx
- mov DWORD[16+rdi],r8d
- mov DWORD[20+rdi],r9d
- mov DWORD[24+rdi],r10d
- mov DWORD[28+rdi],r11d
- jb NEAR $L$loop_avx
- mov rsi,QWORD[88+rsp]
- vzeroupper
- movaps xmm6,XMMWORD[((64+32))+rsp]
- movaps xmm7,XMMWORD[((64+48))+rsp]
- movaps xmm8,XMMWORD[((64+64))+rsp]
- movaps xmm9,XMMWORD[((64+80))+rsp]
- mov r15,QWORD[((-48))+rsi]
- mov r14,QWORD[((-40))+rsi]
- mov r13,QWORD[((-32))+rsi]
- mov r12,QWORD[((-24))+rsi]
- mov rbp,QWORD[((-16))+rsi]
- mov rbx,QWORD[((-8))+rsi]
- lea rsp,[rsi]
- $L$epilogue_avx:
- mov rdi,QWORD[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_GFp_sha256_block_data_order_avx:
- EXTERN __imp_RtlVirtualUnwind
- ALIGN 16
- se_handler:
- push rsi
- push rdi
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- pushfq
- sub rsp,64
- mov rax,QWORD[120+r8]
- mov rbx,QWORD[248+r8]
- mov rsi,QWORD[8+r9]
- mov r11,QWORD[56+r9]
- mov r10d,DWORD[r11]
- lea r10,[r10*1+rsi]
- cmp rbx,r10
- jb NEAR $L$in_prologue
- mov rax,QWORD[152+r8]
- mov r10d,DWORD[4+r11]
- lea r10,[r10*1+rsi]
- cmp rbx,r10
- jae NEAR $L$in_prologue
- mov rsi,rax
- mov rax,QWORD[((64+24))+rax]
- mov rbx,QWORD[((-8))+rax]
- mov rbp,QWORD[((-16))+rax]
- mov r12,QWORD[((-24))+rax]
- mov r13,QWORD[((-32))+rax]
- mov r14,QWORD[((-40))+rax]
- mov r15,QWORD[((-48))+rax]
- mov QWORD[144+r8],rbx
- mov QWORD[160+r8],rbp
- mov QWORD[216+r8],r12
- mov QWORD[224+r8],r13
- mov QWORD[232+r8],r14
- mov QWORD[240+r8],r15
- lea r10,[$L$epilogue]
- cmp rbx,r10
- jb NEAR $L$in_prologue
- lea rsi,[((64+32))+rsi]
- lea rdi,[512+r8]
- mov ecx,8
- DD 0xa548f3fc
- $L$in_prologue:
- mov rdi,QWORD[8+rax]
- mov rsi,QWORD[16+rax]
- mov QWORD[152+r8],rax
- mov QWORD[168+r8],rsi
- mov QWORD[176+r8],rdi
- mov rdi,QWORD[40+r9]
- mov rsi,r8
- mov ecx,154
- DD 0xa548f3fc
- mov rsi,r9
- xor rcx,rcx
- mov rdx,QWORD[8+rsi]
- mov r8,QWORD[rsi]
- mov r9,QWORD[16+rsi]
- mov r10,QWORD[40+rsi]
- lea r11,[56+rsi]
- lea r12,[24+rsi]
- mov QWORD[32+rsp],r10
- mov QWORD[40+rsp],r11
- mov QWORD[48+rsp],r12
- mov QWORD[56+rsp],rcx
- call QWORD[__imp_RtlVirtualUnwind]
- mov eax,1
- add rsp,64
- popfq
- pop r15
- pop r14
- pop r13
- pop r12
- pop rbp
- pop rbx
- pop rdi
- pop rsi
- DB 0F3h,0C3h ;repret
- section .pdata rdata align=4
- ALIGN 4
- DD $L$SEH_begin_GFp_sha256_block_data_order wrt ..imagebase
- DD $L$SEH_end_GFp_sha256_block_data_order wrt ..imagebase
- DD $L$SEH_info_GFp_sha256_block_data_order wrt ..imagebase
- DD $L$SEH_begin_GFp_sha256_block_data_order_ssse3 wrt ..imagebase
- DD $L$SEH_end_GFp_sha256_block_data_order_ssse3 wrt ..imagebase
- DD $L$SEH_info_GFp_sha256_block_data_order_ssse3 wrt ..imagebase
- DD $L$SEH_begin_GFp_sha256_block_data_order_avx wrt ..imagebase
- DD $L$SEH_end_GFp_sha256_block_data_order_avx wrt ..imagebase
- DD $L$SEH_info_GFp_sha256_block_data_order_avx wrt ..imagebase
- section .xdata rdata align=8
- ALIGN 8
- $L$SEH_info_GFp_sha256_block_data_order:
- DB 9,0,0,0
- DD se_handler wrt ..imagebase
- DD $L$prologue wrt ..imagebase,$L$epilogue wrt ..imagebase
- $L$SEH_info_GFp_sha256_block_data_order_ssse3:
- DB 9,0,0,0
- DD se_handler wrt ..imagebase
- DD $L$prologue_ssse3 wrt ..imagebase,$L$epilogue_ssse3 wrt ..imagebase
- $L$SEH_info_GFp_sha256_block_data_order_avx:
- DB 9,0,0,0
- DD se_handler wrt ..imagebase
- DD $L$prologue_avx wrt ..imagebase,$L$epilogue_avx wrt ..imagebase
|