12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267226822692270227122722273227422752276227722782279228022812282228322842285228622872288228922902291229222932294229522962297229822992300230123022303230423052306230723082309231023112312231323142315231623172318231923202321232223232324232523262327232823292330233123322333233423352336233723382339234023412342234323442345234623472348234923502351235223532354235523562357235823592360236123622363236423652366236723682369237023712372237323742375237623772378237923802381238223832384238523862387238823892390239123922393239423952396239723982399240024012402240324042405240624072408240924102411241224132414241524162417241824192420242124222423242424252426242724282429243024312432243324342435243624372438243924402441244224432444244524462447244824492450245124522453245424552456245724582459246024612462246324642465246624672468246924702471247224732474247524762477247824792480248124822483248424852486248724882489249024912492249324942495249624972498249925002501250225032504250525062507250825092510251125122513251425152516251725182519252025212522252325242525252625272528252925302531253225332534253525362537253825392540254125422543254425452546254725482549255025512552255325542555255625572558255925602561256225632564256525662567256825692570257125722573257425752576257725782579258025812582258325842585258625872588258925902591259225932594259525962597259825992600260126022603260426052606260726082609261026112612261326142615261626172618261926202621262226232624262526262627262826292630263126322633263426352636263726382639264026412642264326442645264626472648264926502651265226532654265526562657265826592660266126622663266426652666266726682669267026712672267326742675267626772678267926802681268226832684268526862687268826892690269126922693269426952696269726982699270027012702270327042705270627072708270927102711271227132714271527162717271827192720272127222723272427252726272727282729273027312732273327342735273627372738273927402741274227432744274527462747274827492750275127522753275427552756275727582759276027612762276327642765276627672768276927702771277227732774277527762777277827792780278127822783278427852786278727882789279027912792279327942795279627972798279928002801280228032804280528062807280828092810281128122813281428152816281728182819282028212822282328242825282628272828282928302831283228332834283528362837283828392840284128422843284428452846284728482849285028512852285328542855285628572858285928602861286228632864286528662867286828692870287128722873287428752876287728782879288028812882288328842885288628872888288928902891289228932894289528962897289828992900290129022903290429052906290729082909291029112912291329142915291629172918291929202921292229232924292529262927292829292930293129322933293429352936293729382939294029412942294329442945294629472948294929502951295229532954295529562957295829592960296129622963296429652966296729682969297029712972297329742975297629772978297929802981298229832984298529862987298829892990299129922993299429952996299729982999300030013002300330043005300630073008300930103011301230133014301530163017301830193020302130223023302430253026302730283029303030313032303330343035303630373038303930403041304230433044304530463047304830493050305130523053305430553056305730583059306030613062306330643065306630673068306930703071307230733074307530763077307830793080308130823083308430853086308730883089309030913092309330943095309630973098309931003101310231033104310531063107310831093110311131123113311431153116311731183119312031213122312331243125312631273128312931303131313231333134313531363137313831393140314131423143314431453146314731483149315031513152315331543155315631573158315931603161316231633164316531663167316831693170317131723173317431753176317731783179318031813182318331843185318631873188318931903191319231933194319531963197319831993200320132023203320432053206320732083209321032113212321332143215321632173218321932203221322232233224322532263227322832293230323132323233323432353236323732383239324032413242324332443245324632473248324932503251325232533254325532563257325832593260326132623263326432653266326732683269327032713272327332743275327632773278327932803281328232833284328532863287328832893290329132923293329432953296329732983299330033013302330333043305330633073308330933103311331233133314331533163317331833193320332133223323332433253326332733283329333033313332333333343335333633373338333933403341334233433344334533463347334833493350335133523353335433553356335733583359336033613362336333643365336633673368336933703371337233733374337533763377337833793380338133823383338433853386338733883389339033913392339333943395339633973398339934003401340234033404340534063407340834093410341134123413341434153416341734183419342034213422342334243425342634273428342934303431343234333434343534363437343834393440344134423443344434453446344734483449345034513452345334543455345634573458345934603461346234633464346534663467346834693470347134723473347434753476347734783479348034813482348334843485348634873488348934903491349234933494349534963497349834993500350135023503350435053506350735083509351035113512351335143515351635173518351935203521352235233524352535263527352835293530353135323533353435353536353735383539354035413542354335443545354635473548354935503551355235533554355535563557355835593560356135623563356435653566356735683569357035713572357335743575357635773578357935803581358235833584358535863587358835893590359135923593359435953596359735983599360036013602360336043605360636073608360936103611361236133614361536163617361836193620362136223623362436253626362736283629363036313632363336343635363636373638363936403641364236433644364536463647364836493650365136523653365436553656365736583659366036613662366336643665366636673668366936703671367236733674367536763677367836793680368136823683368436853686368736883689369036913692369336943695369636973698369937003701370237033704370537063707370837093710371137123713371437153716371737183719372037213722372337243725372637273728372937303731373237333734373537363737373837393740374137423743374437453746374737483749375037513752375337543755375637573758375937603761376237633764376537663767376837693770377137723773377437753776377737783779378037813782378337843785378637873788378937903791379237933794379537963797379837993800380138023803380438053806380738083809381038113812381338143815381638173818381938203821382238233824382538263827382838293830383138323833383438353836383738383839384038413842384338443845384638473848384938503851385238533854385538563857385838593860386138623863386438653866386738683869387038713872387338743875387638773878387938803881388238833884388538863887388838893890389138923893389438953896389738983899390039013902390339043905390639073908390939103911391239133914391539163917391839193920392139223923392439253926392739283929393039313932393339343935393639373938393939403941394239433944394539463947394839493950395139523953395439553956395739583959396039613962396339643965396639673968396939703971397239733974397539763977397839793980398139823983398439853986398739883989399039913992399339943995399639973998399940004001400240034004400540064007400840094010401140124013401440154016401740184019402040214022402340244025402640274028402940304031 |
- ; 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_bn_mul_mont_gather5
- ALIGN 64
- GFp_bn_mul_mont_gather5:
- mov QWORD[8+rsp],rdi ;WIN64 prologue
- mov QWORD[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_GFp_bn_mul_mont_gather5:
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- mov r8,QWORD[40+rsp]
- mov r9,QWORD[48+rsp]
- mov r9d,r9d
- mov rax,rsp
- test r9d,7
- jnz NEAR $L$mul_enter
- lea r11,[GFp_ia32cap_P]
- mov r11d,DWORD[8+r11]
- jmp NEAR $L$mul4x_enter
- ALIGN 16
- $L$mul_enter:
- movd xmm5,DWORD[56+rsp]
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- neg r9
- mov r11,rsp
- lea r10,[((-280))+r9*8+rsp]
- neg r9
- and r10,-1024
- sub r11,r10
- and r11,-4096
- lea rsp,[r11*1+r10]
- mov r11,QWORD[rsp]
- cmp rsp,r10
- ja NEAR $L$mul_page_walk
- jmp NEAR $L$mul_page_walk_done
- $L$mul_page_walk:
- lea rsp,[((-4096))+rsp]
- mov r11,QWORD[rsp]
- cmp rsp,r10
- ja NEAR $L$mul_page_walk
- $L$mul_page_walk_done:
- lea r10,[$L$inc]
- mov QWORD[8+r9*8+rsp],rax
- $L$mul_body:
- lea r12,[128+rdx]
- movdqa xmm0,XMMWORD[r10]
- movdqa xmm1,XMMWORD[16+r10]
- lea r10,[((24-112))+r9*8+rsp]
- and r10,-16
- pshufd xmm5,xmm5,0
- movdqa xmm4,xmm1
- movdqa xmm2,xmm1
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- DB 0x67
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[112+r10],xmm0
- movdqa xmm0,xmm4
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[128+r10],xmm1
- movdqa xmm1,xmm4
- paddd xmm0,xmm3
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[144+r10],xmm2
- movdqa xmm2,xmm4
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- movdqa XMMWORD[160+r10],xmm3
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[176+r10],xmm0
- movdqa xmm0,xmm4
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[192+r10],xmm1
- movdqa xmm1,xmm4
- paddd xmm0,xmm3
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[208+r10],xmm2
- movdqa xmm2,xmm4
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- movdqa XMMWORD[224+r10],xmm3
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[240+r10],xmm0
- movdqa xmm0,xmm4
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[256+r10],xmm1
- movdqa xmm1,xmm4
- paddd xmm0,xmm3
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[272+r10],xmm2
- movdqa xmm2,xmm4
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- movdqa XMMWORD[288+r10],xmm3
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[304+r10],xmm0
- paddd xmm3,xmm2
- DB 0x67
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[320+r10],xmm1
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[336+r10],xmm2
- pand xmm0,XMMWORD[64+r12]
- pand xmm1,XMMWORD[80+r12]
- pand xmm2,XMMWORD[96+r12]
- movdqa XMMWORD[352+r10],xmm3
- pand xmm3,XMMWORD[112+r12]
- por xmm0,xmm2
- por xmm1,xmm3
- movdqa xmm4,XMMWORD[((-128))+r12]
- movdqa xmm5,XMMWORD[((-112))+r12]
- movdqa xmm2,XMMWORD[((-96))+r12]
- pand xmm4,XMMWORD[112+r10]
- movdqa xmm3,XMMWORD[((-80))+r12]
- pand xmm5,XMMWORD[128+r10]
- por xmm0,xmm4
- pand xmm2,XMMWORD[144+r10]
- por xmm1,xmm5
- pand xmm3,XMMWORD[160+r10]
- por xmm0,xmm2
- por xmm1,xmm3
- movdqa xmm4,XMMWORD[((-64))+r12]
- movdqa xmm5,XMMWORD[((-48))+r12]
- movdqa xmm2,XMMWORD[((-32))+r12]
- pand xmm4,XMMWORD[176+r10]
- movdqa xmm3,XMMWORD[((-16))+r12]
- pand xmm5,XMMWORD[192+r10]
- por xmm0,xmm4
- pand xmm2,XMMWORD[208+r10]
- por xmm1,xmm5
- pand xmm3,XMMWORD[224+r10]
- por xmm0,xmm2
- por xmm1,xmm3
- movdqa xmm4,XMMWORD[r12]
- movdqa xmm5,XMMWORD[16+r12]
- movdqa xmm2,XMMWORD[32+r12]
- pand xmm4,XMMWORD[240+r10]
- movdqa xmm3,XMMWORD[48+r12]
- pand xmm5,XMMWORD[256+r10]
- por xmm0,xmm4
- pand xmm2,XMMWORD[272+r10]
- por xmm1,xmm5
- pand xmm3,XMMWORD[288+r10]
- por xmm0,xmm2
- por xmm1,xmm3
- por xmm0,xmm1
- pshufd xmm1,xmm0,0x4e
- por xmm0,xmm1
- lea r12,[256+r12]
- DB 102,72,15,126,195
- mov r8,QWORD[r8]
- mov rax,QWORD[rsi]
- xor r14,r14
- xor r15,r15
- mov rbp,r8
- mul rbx
- mov r10,rax
- mov rax,QWORD[rcx]
- imul rbp,r10
- mov r11,rdx
- mul rbp
- add r10,rax
- mov rax,QWORD[8+rsi]
- adc rdx,0
- mov r13,rdx
- lea r15,[1+r15]
- jmp NEAR $L$1st_enter
- ALIGN 16
- $L$1st:
- add r13,rax
- mov rax,QWORD[r15*8+rsi]
- adc rdx,0
- add r13,r11
- mov r11,r10
- adc rdx,0
- mov QWORD[((-16))+r15*8+rsp],r13
- mov r13,rdx
- $L$1st_enter:
- mul rbx
- add r11,rax
- mov rax,QWORD[r15*8+rcx]
- adc rdx,0
- lea r15,[1+r15]
- mov r10,rdx
- mul rbp
- cmp r15,r9
- jne NEAR $L$1st
- add r13,rax
- adc rdx,0
- add r13,r11
- adc rdx,0
- mov QWORD[((-16))+r9*8+rsp],r13
- mov r13,rdx
- mov r11,r10
- xor rdx,rdx
- add r13,r11
- adc rdx,0
- mov QWORD[((-8))+r9*8+rsp],r13
- mov QWORD[r9*8+rsp],rdx
- lea r14,[1+r14]
- jmp NEAR $L$outer
- ALIGN 16
- $L$outer:
- lea rdx,[((24+128))+r9*8+rsp]
- and rdx,-16
- pxor xmm4,xmm4
- pxor xmm5,xmm5
- movdqa xmm0,XMMWORD[((-128))+r12]
- movdqa xmm1,XMMWORD[((-112))+r12]
- movdqa xmm2,XMMWORD[((-96))+r12]
- movdqa xmm3,XMMWORD[((-80))+r12]
- pand xmm0,XMMWORD[((-128))+rdx]
- pand xmm1,XMMWORD[((-112))+rdx]
- por xmm4,xmm0
- pand xmm2,XMMWORD[((-96))+rdx]
- por xmm5,xmm1
- pand xmm3,XMMWORD[((-80))+rdx]
- por xmm4,xmm2
- por xmm5,xmm3
- movdqa xmm0,XMMWORD[((-64))+r12]
- movdqa xmm1,XMMWORD[((-48))+r12]
- movdqa xmm2,XMMWORD[((-32))+r12]
- movdqa xmm3,XMMWORD[((-16))+r12]
- pand xmm0,XMMWORD[((-64))+rdx]
- pand xmm1,XMMWORD[((-48))+rdx]
- por xmm4,xmm0
- pand xmm2,XMMWORD[((-32))+rdx]
- por xmm5,xmm1
- pand xmm3,XMMWORD[((-16))+rdx]
- por xmm4,xmm2
- por xmm5,xmm3
- movdqa xmm0,XMMWORD[r12]
- movdqa xmm1,XMMWORD[16+r12]
- movdqa xmm2,XMMWORD[32+r12]
- movdqa xmm3,XMMWORD[48+r12]
- pand xmm0,XMMWORD[rdx]
- pand xmm1,XMMWORD[16+rdx]
- por xmm4,xmm0
- pand xmm2,XMMWORD[32+rdx]
- por xmm5,xmm1
- pand xmm3,XMMWORD[48+rdx]
- por xmm4,xmm2
- por xmm5,xmm3
- movdqa xmm0,XMMWORD[64+r12]
- movdqa xmm1,XMMWORD[80+r12]
- movdqa xmm2,XMMWORD[96+r12]
- movdqa xmm3,XMMWORD[112+r12]
- pand xmm0,XMMWORD[64+rdx]
- pand xmm1,XMMWORD[80+rdx]
- por xmm4,xmm0
- pand xmm2,XMMWORD[96+rdx]
- por xmm5,xmm1
- pand xmm3,XMMWORD[112+rdx]
- por xmm4,xmm2
- por xmm5,xmm3
- por xmm4,xmm5
- pshufd xmm0,xmm4,0x4e
- por xmm0,xmm4
- lea r12,[256+r12]
- mov rax,QWORD[rsi]
- DB 102,72,15,126,195
- xor r15,r15
- mov rbp,r8
- mov r10,QWORD[rsp]
- mul rbx
- add r10,rax
- mov rax,QWORD[rcx]
- adc rdx,0
- imul rbp,r10
- mov r11,rdx
- mul rbp
- add r10,rax
- mov rax,QWORD[8+rsi]
- adc rdx,0
- mov r10,QWORD[8+rsp]
- mov r13,rdx
- lea r15,[1+r15]
- jmp NEAR $L$inner_enter
- ALIGN 16
- $L$inner:
- add r13,rax
- mov rax,QWORD[r15*8+rsi]
- adc rdx,0
- add r13,r10
- mov r10,QWORD[r15*8+rsp]
- adc rdx,0
- mov QWORD[((-16))+r15*8+rsp],r13
- mov r13,rdx
- $L$inner_enter:
- mul rbx
- add r11,rax
- mov rax,QWORD[r15*8+rcx]
- adc rdx,0
- add r10,r11
- mov r11,rdx
- adc r11,0
- lea r15,[1+r15]
- mul rbp
- cmp r15,r9
- jne NEAR $L$inner
- add r13,rax
- adc rdx,0
- add r13,r10
- mov r10,QWORD[r9*8+rsp]
- adc rdx,0
- mov QWORD[((-16))+r9*8+rsp],r13
- mov r13,rdx
- xor rdx,rdx
- add r13,r11
- adc rdx,0
- add r13,r10
- adc rdx,0
- mov QWORD[((-8))+r9*8+rsp],r13
- mov QWORD[r9*8+rsp],rdx
- lea r14,[1+r14]
- cmp r14,r9
- jb NEAR $L$outer
- xor r14,r14
- mov rax,QWORD[rsp]
- lea rsi,[rsp]
- mov r15,r9
- jmp NEAR $L$sub
- ALIGN 16
- $L$sub: sbb rax,QWORD[r14*8+rcx]
- mov QWORD[r14*8+rdi],rax
- mov rax,QWORD[8+r14*8+rsi]
- lea r14,[1+r14]
- dec r15
- jnz NEAR $L$sub
- sbb rax,0
- mov rbx,-1
- xor rbx,rax
- xor r14,r14
- mov r15,r9
- $L$copy:
- mov rcx,QWORD[r14*8+rdi]
- mov rdx,QWORD[r14*8+rsp]
- and rcx,rbx
- and rdx,rax
- mov QWORD[r14*8+rsp],r14
- or rdx,rcx
- mov QWORD[r14*8+rdi],rdx
- lea r14,[1+r14]
- sub r15,1
- jnz NEAR $L$copy
- mov rsi,QWORD[8+r9*8+rsp]
- mov rax,1
- 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]
- $L$mul_epilogue:
- mov rdi,QWORD[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_GFp_bn_mul_mont_gather5:
- ALIGN 32
- bn_mul4x_mont_gather5:
- mov QWORD[8+rsp],rdi ;WIN64 prologue
- mov QWORD[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_bn_mul4x_mont_gather5:
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- mov r8,QWORD[40+rsp]
- mov r9,QWORD[48+rsp]
- DB 0x67
- mov rax,rsp
- $L$mul4x_enter:
- and r11d,0x80108
- cmp r11d,0x80108
- je NEAR $L$mulx4x_enter
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- $L$mul4x_prologue:
- DB 0x67
- shl r9d,3
- lea r10,[r9*2+r9]
- neg r9
- lea r11,[((-320))+r9*2+rsp]
- mov rbp,rsp
- sub r11,rdi
- and r11,4095
- cmp r10,r11
- jb NEAR $L$mul4xsp_alt
- sub rbp,r11
- lea rbp,[((-320))+r9*2+rbp]
- jmp NEAR $L$mul4xsp_done
- ALIGN 32
- $L$mul4xsp_alt:
- lea r10,[((4096-320))+r9*2]
- lea rbp,[((-320))+r9*2+rbp]
- sub r11,r10
- mov r10,0
- cmovc r11,r10
- sub rbp,r11
- $L$mul4xsp_done:
- and rbp,-64
- mov r11,rsp
- sub r11,rbp
- and r11,-4096
- lea rsp,[rbp*1+r11]
- mov r10,QWORD[rsp]
- cmp rsp,rbp
- ja NEAR $L$mul4x_page_walk
- jmp NEAR $L$mul4x_page_walk_done
- $L$mul4x_page_walk:
- lea rsp,[((-4096))+rsp]
- mov r10,QWORD[rsp]
- cmp rsp,rbp
- ja NEAR $L$mul4x_page_walk
- $L$mul4x_page_walk_done:
- neg r9
- mov QWORD[40+rsp],rax
- $L$mul4x_body:
- call mul4x_internal
- mov rsi,QWORD[40+rsp]
- mov rax,1
- 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]
- $L$mul4x_epilogue:
- mov rdi,QWORD[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_bn_mul4x_mont_gather5:
- ALIGN 32
- mul4x_internal:
- shl r9,5
- movd xmm5,DWORD[56+rax]
- lea rax,[$L$inc]
- lea r13,[128+r9*1+rdx]
- shr r9,5
- movdqa xmm0,XMMWORD[rax]
- movdqa xmm1,XMMWORD[16+rax]
- lea r10,[((88-112))+r9*1+rsp]
- lea r12,[128+rdx]
- pshufd xmm5,xmm5,0
- movdqa xmm4,xmm1
- DB 0x67,0x67
- movdqa xmm2,xmm1
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- DB 0x67
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[112+r10],xmm0
- movdqa xmm0,xmm4
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[128+r10],xmm1
- movdqa xmm1,xmm4
- paddd xmm0,xmm3
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[144+r10],xmm2
- movdqa xmm2,xmm4
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- movdqa XMMWORD[160+r10],xmm3
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[176+r10],xmm0
- movdqa xmm0,xmm4
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[192+r10],xmm1
- movdqa xmm1,xmm4
- paddd xmm0,xmm3
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[208+r10],xmm2
- movdqa xmm2,xmm4
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- movdqa XMMWORD[224+r10],xmm3
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[240+r10],xmm0
- movdqa xmm0,xmm4
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[256+r10],xmm1
- movdqa xmm1,xmm4
- paddd xmm0,xmm3
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[272+r10],xmm2
- movdqa xmm2,xmm4
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- movdqa XMMWORD[288+r10],xmm3
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[304+r10],xmm0
- paddd xmm3,xmm2
- DB 0x67
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[320+r10],xmm1
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[336+r10],xmm2
- pand xmm0,XMMWORD[64+r12]
- pand xmm1,XMMWORD[80+r12]
- pand xmm2,XMMWORD[96+r12]
- movdqa XMMWORD[352+r10],xmm3
- pand xmm3,XMMWORD[112+r12]
- por xmm0,xmm2
- por xmm1,xmm3
- movdqa xmm4,XMMWORD[((-128))+r12]
- movdqa xmm5,XMMWORD[((-112))+r12]
- movdqa xmm2,XMMWORD[((-96))+r12]
- pand xmm4,XMMWORD[112+r10]
- movdqa xmm3,XMMWORD[((-80))+r12]
- pand xmm5,XMMWORD[128+r10]
- por xmm0,xmm4
- pand xmm2,XMMWORD[144+r10]
- por xmm1,xmm5
- pand xmm3,XMMWORD[160+r10]
- por xmm0,xmm2
- por xmm1,xmm3
- movdqa xmm4,XMMWORD[((-64))+r12]
- movdqa xmm5,XMMWORD[((-48))+r12]
- movdqa xmm2,XMMWORD[((-32))+r12]
- pand xmm4,XMMWORD[176+r10]
- movdqa xmm3,XMMWORD[((-16))+r12]
- pand xmm5,XMMWORD[192+r10]
- por xmm0,xmm4
- pand xmm2,XMMWORD[208+r10]
- por xmm1,xmm5
- pand xmm3,XMMWORD[224+r10]
- por xmm0,xmm2
- por xmm1,xmm3
- movdqa xmm4,XMMWORD[r12]
- movdqa xmm5,XMMWORD[16+r12]
- movdqa xmm2,XMMWORD[32+r12]
- pand xmm4,XMMWORD[240+r10]
- movdqa xmm3,XMMWORD[48+r12]
- pand xmm5,XMMWORD[256+r10]
- por xmm0,xmm4
- pand xmm2,XMMWORD[272+r10]
- por xmm1,xmm5
- pand xmm3,XMMWORD[288+r10]
- por xmm0,xmm2
- por xmm1,xmm3
- por xmm0,xmm1
- pshufd xmm1,xmm0,0x4e
- por xmm0,xmm1
- lea r12,[256+r12]
- DB 102,72,15,126,195
- mov QWORD[((16+8))+rsp],r13
- mov QWORD[((56+8))+rsp],rdi
- mov r8,QWORD[r8]
- mov rax,QWORD[rsi]
- lea rsi,[r9*1+rsi]
- neg r9
- mov rbp,r8
- mul rbx
- mov r10,rax
- mov rax,QWORD[rcx]
- imul rbp,r10
- lea r14,[((64+8))+rsp]
- mov r11,rdx
- mul rbp
- add r10,rax
- mov rax,QWORD[8+r9*1+rsi]
- adc rdx,0
- mov rdi,rdx
- mul rbx
- add r11,rax
- mov rax,QWORD[8+rcx]
- adc rdx,0
- mov r10,rdx
- mul rbp
- add rdi,rax
- mov rax,QWORD[16+r9*1+rsi]
- adc rdx,0
- add rdi,r11
- lea r15,[32+r9]
- lea rcx,[32+rcx]
- adc rdx,0
- mov QWORD[r14],rdi
- mov r13,rdx
- jmp NEAR $L$1st4x
- ALIGN 32
- $L$1st4x:
- mul rbx
- add r10,rax
- mov rax,QWORD[((-16))+rcx]
- lea r14,[32+r14]
- adc rdx,0
- mov r11,rdx
- mul rbp
- add r13,rax
- mov rax,QWORD[((-8))+r15*1+rsi]
- adc rdx,0
- add r13,r10
- adc rdx,0
- mov QWORD[((-24))+r14],r13
- mov rdi,rdx
- mul rbx
- add r11,rax
- mov rax,QWORD[((-8))+rcx]
- adc rdx,0
- mov r10,rdx
- mul rbp
- add rdi,rax
- mov rax,QWORD[r15*1+rsi]
- adc rdx,0
- add rdi,r11
- adc rdx,0
- mov QWORD[((-16))+r14],rdi
- mov r13,rdx
- mul rbx
- add r10,rax
- mov rax,QWORD[rcx]
- adc rdx,0
- mov r11,rdx
- mul rbp
- add r13,rax
- mov rax,QWORD[8+r15*1+rsi]
- adc rdx,0
- add r13,r10
- adc rdx,0
- mov QWORD[((-8))+r14],r13
- mov rdi,rdx
- mul rbx
- add r11,rax
- mov rax,QWORD[8+rcx]
- adc rdx,0
- mov r10,rdx
- mul rbp
- add rdi,rax
- mov rax,QWORD[16+r15*1+rsi]
- adc rdx,0
- add rdi,r11
- lea rcx,[32+rcx]
- adc rdx,0
- mov QWORD[r14],rdi
- mov r13,rdx
- add r15,32
- jnz NEAR $L$1st4x
- mul rbx
- add r10,rax
- mov rax,QWORD[((-16))+rcx]
- lea r14,[32+r14]
- adc rdx,0
- mov r11,rdx
- mul rbp
- add r13,rax
- mov rax,QWORD[((-8))+rsi]
- adc rdx,0
- add r13,r10
- adc rdx,0
- mov QWORD[((-24))+r14],r13
- mov rdi,rdx
- mul rbx
- add r11,rax
- mov rax,QWORD[((-8))+rcx]
- adc rdx,0
- mov r10,rdx
- mul rbp
- add rdi,rax
- mov rax,QWORD[r9*1+rsi]
- adc rdx,0
- add rdi,r11
- adc rdx,0
- mov QWORD[((-16))+r14],rdi
- mov r13,rdx
- lea rcx,[r9*1+rcx]
- xor rdi,rdi
- add r13,r10
- adc rdi,0
- mov QWORD[((-8))+r14],r13
- jmp NEAR $L$outer4x
- ALIGN 32
- $L$outer4x:
- lea rdx,[((16+128))+r14]
- pxor xmm4,xmm4
- pxor xmm5,xmm5
- movdqa xmm0,XMMWORD[((-128))+r12]
- movdqa xmm1,XMMWORD[((-112))+r12]
- movdqa xmm2,XMMWORD[((-96))+r12]
- movdqa xmm3,XMMWORD[((-80))+r12]
- pand xmm0,XMMWORD[((-128))+rdx]
- pand xmm1,XMMWORD[((-112))+rdx]
- por xmm4,xmm0
- pand xmm2,XMMWORD[((-96))+rdx]
- por xmm5,xmm1
- pand xmm3,XMMWORD[((-80))+rdx]
- por xmm4,xmm2
- por xmm5,xmm3
- movdqa xmm0,XMMWORD[((-64))+r12]
- movdqa xmm1,XMMWORD[((-48))+r12]
- movdqa xmm2,XMMWORD[((-32))+r12]
- movdqa xmm3,XMMWORD[((-16))+r12]
- pand xmm0,XMMWORD[((-64))+rdx]
- pand xmm1,XMMWORD[((-48))+rdx]
- por xmm4,xmm0
- pand xmm2,XMMWORD[((-32))+rdx]
- por xmm5,xmm1
- pand xmm3,XMMWORD[((-16))+rdx]
- por xmm4,xmm2
- por xmm5,xmm3
- movdqa xmm0,XMMWORD[r12]
- movdqa xmm1,XMMWORD[16+r12]
- movdqa xmm2,XMMWORD[32+r12]
- movdqa xmm3,XMMWORD[48+r12]
- pand xmm0,XMMWORD[rdx]
- pand xmm1,XMMWORD[16+rdx]
- por xmm4,xmm0
- pand xmm2,XMMWORD[32+rdx]
- por xmm5,xmm1
- pand xmm3,XMMWORD[48+rdx]
- por xmm4,xmm2
- por xmm5,xmm3
- movdqa xmm0,XMMWORD[64+r12]
- movdqa xmm1,XMMWORD[80+r12]
- movdqa xmm2,XMMWORD[96+r12]
- movdqa xmm3,XMMWORD[112+r12]
- pand xmm0,XMMWORD[64+rdx]
- pand xmm1,XMMWORD[80+rdx]
- por xmm4,xmm0
- pand xmm2,XMMWORD[96+rdx]
- por xmm5,xmm1
- pand xmm3,XMMWORD[112+rdx]
- por xmm4,xmm2
- por xmm5,xmm3
- por xmm4,xmm5
- pshufd xmm0,xmm4,0x4e
- por xmm0,xmm4
- lea r12,[256+r12]
- DB 102,72,15,126,195
- mov r10,QWORD[r9*1+r14]
- mov rbp,r8
- mul rbx
- add r10,rax
- mov rax,QWORD[rcx]
- adc rdx,0
- imul rbp,r10
- mov r11,rdx
- mov QWORD[r14],rdi
- lea r14,[r9*1+r14]
- mul rbp
- add r10,rax
- mov rax,QWORD[8+r9*1+rsi]
- adc rdx,0
- mov rdi,rdx
- mul rbx
- add r11,rax
- mov rax,QWORD[8+rcx]
- adc rdx,0
- add r11,QWORD[8+r14]
- adc rdx,0
- mov r10,rdx
- mul rbp
- add rdi,rax
- mov rax,QWORD[16+r9*1+rsi]
- adc rdx,0
- add rdi,r11
- lea r15,[32+r9]
- lea rcx,[32+rcx]
- adc rdx,0
- mov r13,rdx
- jmp NEAR $L$inner4x
- ALIGN 32
- $L$inner4x:
- mul rbx
- add r10,rax
- mov rax,QWORD[((-16))+rcx]
- adc rdx,0
- add r10,QWORD[16+r14]
- lea r14,[32+r14]
- adc rdx,0
- mov r11,rdx
- mul rbp
- add r13,rax
- mov rax,QWORD[((-8))+r15*1+rsi]
- adc rdx,0
- add r13,r10
- adc rdx,0
- mov QWORD[((-32))+r14],rdi
- mov rdi,rdx
- mul rbx
- add r11,rax
- mov rax,QWORD[((-8))+rcx]
- adc rdx,0
- add r11,QWORD[((-8))+r14]
- adc rdx,0
- mov r10,rdx
- mul rbp
- add rdi,rax
- mov rax,QWORD[r15*1+rsi]
- adc rdx,0
- add rdi,r11
- adc rdx,0
- mov QWORD[((-24))+r14],r13
- mov r13,rdx
- mul rbx
- add r10,rax
- mov rax,QWORD[rcx]
- adc rdx,0
- add r10,QWORD[r14]
- adc rdx,0
- mov r11,rdx
- mul rbp
- add r13,rax
- mov rax,QWORD[8+r15*1+rsi]
- adc rdx,0
- add r13,r10
- adc rdx,0
- mov QWORD[((-16))+r14],rdi
- mov rdi,rdx
- mul rbx
- add r11,rax
- mov rax,QWORD[8+rcx]
- adc rdx,0
- add r11,QWORD[8+r14]
- adc rdx,0
- mov r10,rdx
- mul rbp
- add rdi,rax
- mov rax,QWORD[16+r15*1+rsi]
- adc rdx,0
- add rdi,r11
- lea rcx,[32+rcx]
- adc rdx,0
- mov QWORD[((-8))+r14],r13
- mov r13,rdx
- add r15,32
- jnz NEAR $L$inner4x
- mul rbx
- add r10,rax
- mov rax,QWORD[((-16))+rcx]
- adc rdx,0
- add r10,QWORD[16+r14]
- lea r14,[32+r14]
- adc rdx,0
- mov r11,rdx
- mul rbp
- add r13,rax
- mov rax,QWORD[((-8))+rsi]
- adc rdx,0
- add r13,r10
- adc rdx,0
- mov QWORD[((-32))+r14],rdi
- mov rdi,rdx
- mul rbx
- add r11,rax
- mov rax,rbp
- mov rbp,QWORD[((-8))+rcx]
- adc rdx,0
- add r11,QWORD[((-8))+r14]
- adc rdx,0
- mov r10,rdx
- mul rbp
- add rdi,rax
- mov rax,QWORD[r9*1+rsi]
- adc rdx,0
- add rdi,r11
- adc rdx,0
- mov QWORD[((-24))+r14],r13
- mov r13,rdx
- mov QWORD[((-16))+r14],rdi
- lea rcx,[r9*1+rcx]
- xor rdi,rdi
- add r13,r10
- adc rdi,0
- add r13,QWORD[r14]
- adc rdi,0
- mov QWORD[((-8))+r14],r13
- cmp r12,QWORD[((16+8))+rsp]
- jb NEAR $L$outer4x
- xor rax,rax
- sub rbp,r13
- adc r15,r15
- or rdi,r15
- sub rax,rdi
- lea rbx,[r9*1+r14]
- mov r12,QWORD[rcx]
- lea rbp,[rcx]
- mov rcx,r9
- sar rcx,3+2
- mov rdi,QWORD[((56+8))+rsp]
- dec r12
- xor r10,r10
- mov r13,QWORD[8+rbp]
- mov r14,QWORD[16+rbp]
- mov r15,QWORD[24+rbp]
- jmp NEAR $L$sqr4x_sub_entry
- global GFp_bn_power5
- ALIGN 32
- GFp_bn_power5:
- mov QWORD[8+rsp],rdi ;WIN64 prologue
- mov QWORD[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_GFp_bn_power5:
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- mov r8,QWORD[40+rsp]
- mov r9,QWORD[48+rsp]
- mov rax,rsp
- lea r11,[GFp_ia32cap_P]
- mov r11d,DWORD[8+r11]
- and r11d,0x80108
- cmp r11d,0x80108
- je NEAR $L$powerx5_enter
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- $L$power5_prologue:
- shl r9d,3
- lea r10d,[r9*2+r9]
- neg r9
- mov r8,QWORD[r8]
- lea r11,[((-320))+r9*2+rsp]
- mov rbp,rsp
- sub r11,rdi
- and r11,4095
- cmp r10,r11
- jb NEAR $L$pwr_sp_alt
- sub rbp,r11
- lea rbp,[((-320))+r9*2+rbp]
- jmp NEAR $L$pwr_sp_done
- ALIGN 32
- $L$pwr_sp_alt:
- lea r10,[((4096-320))+r9*2]
- lea rbp,[((-320))+r9*2+rbp]
- sub r11,r10
- mov r10,0
- cmovc r11,r10
- sub rbp,r11
- $L$pwr_sp_done:
- and rbp,-64
- mov r11,rsp
- sub r11,rbp
- and r11,-4096
- lea rsp,[rbp*1+r11]
- mov r10,QWORD[rsp]
- cmp rsp,rbp
- ja NEAR $L$pwr_page_walk
- jmp NEAR $L$pwr_page_walk_done
- $L$pwr_page_walk:
- lea rsp,[((-4096))+rsp]
- mov r10,QWORD[rsp]
- cmp rsp,rbp
- ja NEAR $L$pwr_page_walk
- $L$pwr_page_walk_done:
- mov r10,r9
- neg r9
- mov QWORD[32+rsp],r8
- mov QWORD[40+rsp],rax
- $L$power5_body:
- DB 102,72,15,110,207
- DB 102,72,15,110,209
- DB 102,73,15,110,218
- DB 102,72,15,110,226
- call __bn_sqr8x_internal
- call __bn_post4x_internal
- call __bn_sqr8x_internal
- call __bn_post4x_internal
- call __bn_sqr8x_internal
- call __bn_post4x_internal
- call __bn_sqr8x_internal
- call __bn_post4x_internal
- call __bn_sqr8x_internal
- call __bn_post4x_internal
- DB 102,72,15,126,209
- DB 102,72,15,126,226
- mov rdi,rsi
- mov rax,QWORD[40+rsp]
- lea r8,[32+rsp]
- call mul4x_internal
- mov rsi,QWORD[40+rsp]
- mov rax,1
- 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]
- $L$power5_epilogue:
- mov rdi,QWORD[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_GFp_bn_power5:
- global GFp_bn_sqr8x_internal
- ALIGN 32
- GFp_bn_sqr8x_internal:
- __bn_sqr8x_internal:
- lea rbp,[32+r10]
- lea rsi,[r9*1+rsi]
- mov rcx,r9
- mov r14,QWORD[((-32))+rbp*1+rsi]
- lea rdi,[((48+8))+r9*2+rsp]
- mov rax,QWORD[((-24))+rbp*1+rsi]
- lea rdi,[((-32))+rbp*1+rdi]
- mov rbx,QWORD[((-16))+rbp*1+rsi]
- mov r15,rax
- mul r14
- mov r10,rax
- mov rax,rbx
- mov r11,rdx
- mov QWORD[((-24))+rbp*1+rdi],r10
- mul r14
- add r11,rax
- mov rax,rbx
- adc rdx,0
- mov QWORD[((-16))+rbp*1+rdi],r11
- mov r10,rdx
- mov rbx,QWORD[((-8))+rbp*1+rsi]
- mul r15
- mov r12,rax
- mov rax,rbx
- mov r13,rdx
- lea rcx,[rbp]
- mul r14
- add r10,rax
- mov rax,rbx
- mov r11,rdx
- adc r11,0
- add r10,r12
- adc r11,0
- mov QWORD[((-8))+rcx*1+rdi],r10
- jmp NEAR $L$sqr4x_1st
- ALIGN 32
- $L$sqr4x_1st:
- mov rbx,QWORD[rcx*1+rsi]
- mul r15
- add r13,rax
- mov rax,rbx
- mov r12,rdx
- adc r12,0
- mul r14
- add r11,rax
- mov rax,rbx
- mov rbx,QWORD[8+rcx*1+rsi]
- mov r10,rdx
- adc r10,0
- add r11,r13
- adc r10,0
- mul r15
- add r12,rax
- mov rax,rbx
- mov QWORD[rcx*1+rdi],r11
- mov r13,rdx
- adc r13,0
- mul r14
- add r10,rax
- mov rax,rbx
- mov rbx,QWORD[16+rcx*1+rsi]
- mov r11,rdx
- adc r11,0
- add r10,r12
- adc r11,0
- mul r15
- add r13,rax
- mov rax,rbx
- mov QWORD[8+rcx*1+rdi],r10
- mov r12,rdx
- adc r12,0
- mul r14
- add r11,rax
- mov rax,rbx
- mov rbx,QWORD[24+rcx*1+rsi]
- mov r10,rdx
- adc r10,0
- add r11,r13
- adc r10,0
- mul r15
- add r12,rax
- mov rax,rbx
- mov QWORD[16+rcx*1+rdi],r11
- mov r13,rdx
- adc r13,0
- lea rcx,[32+rcx]
- mul r14
- add r10,rax
- mov rax,rbx
- mov r11,rdx
- adc r11,0
- add r10,r12
- adc r11,0
- mov QWORD[((-8))+rcx*1+rdi],r10
- cmp rcx,0
- jne NEAR $L$sqr4x_1st
- mul r15
- add r13,rax
- lea rbp,[16+rbp]
- adc rdx,0
- add r13,r11
- adc rdx,0
- mov QWORD[rdi],r13
- mov r12,rdx
- mov QWORD[8+rdi],rdx
- jmp NEAR $L$sqr4x_outer
- ALIGN 32
- $L$sqr4x_outer:
- mov r14,QWORD[((-32))+rbp*1+rsi]
- lea rdi,[((48+8))+r9*2+rsp]
- mov rax,QWORD[((-24))+rbp*1+rsi]
- lea rdi,[((-32))+rbp*1+rdi]
- mov rbx,QWORD[((-16))+rbp*1+rsi]
- mov r15,rax
- mul r14
- mov r10,QWORD[((-24))+rbp*1+rdi]
- add r10,rax
- mov rax,rbx
- adc rdx,0
- mov QWORD[((-24))+rbp*1+rdi],r10
- mov r11,rdx
- mul r14
- add r11,rax
- mov rax,rbx
- adc rdx,0
- add r11,QWORD[((-16))+rbp*1+rdi]
- mov r10,rdx
- adc r10,0
- mov QWORD[((-16))+rbp*1+rdi],r11
- xor r12,r12
- mov rbx,QWORD[((-8))+rbp*1+rsi]
- mul r15
- add r12,rax
- mov rax,rbx
- adc rdx,0
- add r12,QWORD[((-8))+rbp*1+rdi]
- mov r13,rdx
- adc r13,0
- mul r14
- add r10,rax
- mov rax,rbx
- adc rdx,0
- add r10,r12
- mov r11,rdx
- adc r11,0
- mov QWORD[((-8))+rbp*1+rdi],r10
- lea rcx,[rbp]
- jmp NEAR $L$sqr4x_inner
- ALIGN 32
- $L$sqr4x_inner:
- mov rbx,QWORD[rcx*1+rsi]
- mul r15
- add r13,rax
- mov rax,rbx
- mov r12,rdx
- adc r12,0
- add r13,QWORD[rcx*1+rdi]
- adc r12,0
- DB 0x67
- mul r14
- add r11,rax
- mov rax,rbx
- mov rbx,QWORD[8+rcx*1+rsi]
- mov r10,rdx
- adc r10,0
- add r11,r13
- adc r10,0
- mul r15
- add r12,rax
- mov QWORD[rcx*1+rdi],r11
- mov rax,rbx
- mov r13,rdx
- adc r13,0
- add r12,QWORD[8+rcx*1+rdi]
- lea rcx,[16+rcx]
- adc r13,0
- mul r14
- add r10,rax
- mov rax,rbx
- adc rdx,0
- add r10,r12
- mov r11,rdx
- adc r11,0
- mov QWORD[((-8))+rcx*1+rdi],r10
- cmp rcx,0
- jne NEAR $L$sqr4x_inner
- DB 0x67
- mul r15
- add r13,rax
- adc rdx,0
- add r13,r11
- adc rdx,0
- mov QWORD[rdi],r13
- mov r12,rdx
- mov QWORD[8+rdi],rdx
- add rbp,16
- jnz NEAR $L$sqr4x_outer
- mov r14,QWORD[((-32))+rsi]
- lea rdi,[((48+8))+r9*2+rsp]
- mov rax,QWORD[((-24))+rsi]
- lea rdi,[((-32))+rbp*1+rdi]
- mov rbx,QWORD[((-16))+rsi]
- mov r15,rax
- mul r14
- add r10,rax
- mov rax,rbx
- mov r11,rdx
- adc r11,0
- mul r14
- add r11,rax
- mov rax,rbx
- mov QWORD[((-24))+rdi],r10
- mov r10,rdx
- adc r10,0
- add r11,r13
- mov rbx,QWORD[((-8))+rsi]
- adc r10,0
- mul r15
- add r12,rax
- mov rax,rbx
- mov QWORD[((-16))+rdi],r11
- mov r13,rdx
- adc r13,0
- mul r14
- add r10,rax
- mov rax,rbx
- mov r11,rdx
- adc r11,0
- add r10,r12
- adc r11,0
- mov QWORD[((-8))+rdi],r10
- mul r15
- add r13,rax
- mov rax,QWORD[((-16))+rsi]
- adc rdx,0
- add r13,r11
- adc rdx,0
- mov QWORD[rdi],r13
- mov r12,rdx
- mov QWORD[8+rdi],rdx
- mul rbx
- add rbp,16
- xor r14,r14
- sub rbp,r9
- xor r15,r15
- add rax,r12
- adc rdx,0
- mov QWORD[8+rdi],rax
- mov QWORD[16+rdi],rdx
- mov QWORD[24+rdi],r15
- mov rax,QWORD[((-16))+rbp*1+rsi]
- lea rdi,[((48+8))+rsp]
- xor r10,r10
- mov r11,QWORD[8+rdi]
- lea r12,[r10*2+r14]
- shr r10,63
- lea r13,[r11*2+rcx]
- shr r11,63
- or r13,r10
- mov r10,QWORD[16+rdi]
- mov r14,r11
- mul rax
- neg r15
- mov r11,QWORD[24+rdi]
- adc r12,rax
- mov rax,QWORD[((-8))+rbp*1+rsi]
- mov QWORD[rdi],r12
- adc r13,rdx
- lea rbx,[r10*2+r14]
- mov QWORD[8+rdi],r13
- sbb r15,r15
- shr r10,63
- lea r8,[r11*2+rcx]
- shr r11,63
- or r8,r10
- mov r10,QWORD[32+rdi]
- mov r14,r11
- mul rax
- neg r15
- mov r11,QWORD[40+rdi]
- adc rbx,rax
- mov rax,QWORD[rbp*1+rsi]
- mov QWORD[16+rdi],rbx
- adc r8,rdx
- lea rbp,[16+rbp]
- mov QWORD[24+rdi],r8
- sbb r15,r15
- lea rdi,[64+rdi]
- jmp NEAR $L$sqr4x_shift_n_add
- ALIGN 32
- $L$sqr4x_shift_n_add:
- lea r12,[r10*2+r14]
- shr r10,63
- lea r13,[r11*2+rcx]
- shr r11,63
- or r13,r10
- mov r10,QWORD[((-16))+rdi]
- mov r14,r11
- mul rax
- neg r15
- mov r11,QWORD[((-8))+rdi]
- adc r12,rax
- mov rax,QWORD[((-8))+rbp*1+rsi]
- mov QWORD[((-32))+rdi],r12
- adc r13,rdx
- lea rbx,[r10*2+r14]
- mov QWORD[((-24))+rdi],r13
- sbb r15,r15
- shr r10,63
- lea r8,[r11*2+rcx]
- shr r11,63
- or r8,r10
- mov r10,QWORD[rdi]
- mov r14,r11
- mul rax
- neg r15
- mov r11,QWORD[8+rdi]
- adc rbx,rax
- mov rax,QWORD[rbp*1+rsi]
- mov QWORD[((-16))+rdi],rbx
- adc r8,rdx
- lea r12,[r10*2+r14]
- mov QWORD[((-8))+rdi],r8
- sbb r15,r15
- shr r10,63
- lea r13,[r11*2+rcx]
- shr r11,63
- or r13,r10
- mov r10,QWORD[16+rdi]
- mov r14,r11
- mul rax
- neg r15
- mov r11,QWORD[24+rdi]
- adc r12,rax
- mov rax,QWORD[8+rbp*1+rsi]
- mov QWORD[rdi],r12
- adc r13,rdx
- lea rbx,[r10*2+r14]
- mov QWORD[8+rdi],r13
- sbb r15,r15
- shr r10,63
- lea r8,[r11*2+rcx]
- shr r11,63
- or r8,r10
- mov r10,QWORD[32+rdi]
- mov r14,r11
- mul rax
- neg r15
- mov r11,QWORD[40+rdi]
- adc rbx,rax
- mov rax,QWORD[16+rbp*1+rsi]
- mov QWORD[16+rdi],rbx
- adc r8,rdx
- mov QWORD[24+rdi],r8
- sbb r15,r15
- lea rdi,[64+rdi]
- add rbp,32
- jnz NEAR $L$sqr4x_shift_n_add
- lea r12,[r10*2+r14]
- DB 0x67
- shr r10,63
- lea r13,[r11*2+rcx]
- shr r11,63
- or r13,r10
- mov r10,QWORD[((-16))+rdi]
- mov r14,r11
- mul rax
- neg r15
- mov r11,QWORD[((-8))+rdi]
- adc r12,rax
- mov rax,QWORD[((-8))+rsi]
- mov QWORD[((-32))+rdi],r12
- adc r13,rdx
- lea rbx,[r10*2+r14]
- mov QWORD[((-24))+rdi],r13
- sbb r15,r15
- shr r10,63
- lea r8,[r11*2+rcx]
- shr r11,63
- or r8,r10
- mul rax
- neg r15
- adc rbx,rax
- adc r8,rdx
- mov QWORD[((-16))+rdi],rbx
- mov QWORD[((-8))+rdi],r8
- DB 102,72,15,126,213
- __bn_sqr8x_reduction:
- xor rax,rax
- lea rcx,[rbp*1+r9]
- lea rdx,[((48+8))+r9*2+rsp]
- mov QWORD[((0+8))+rsp],rcx
- lea rdi,[((48+8))+r9*1+rsp]
- mov QWORD[((8+8))+rsp],rdx
- neg r9
- jmp NEAR $L$8x_reduction_loop
- ALIGN 32
- $L$8x_reduction_loop:
- lea rdi,[r9*1+rdi]
- DB 0x66
- mov rbx,QWORD[rdi]
- mov r9,QWORD[8+rdi]
- mov r10,QWORD[16+rdi]
- mov r11,QWORD[24+rdi]
- mov r12,QWORD[32+rdi]
- mov r13,QWORD[40+rdi]
- mov r14,QWORD[48+rdi]
- mov r15,QWORD[56+rdi]
- mov QWORD[rdx],rax
- lea rdi,[64+rdi]
- DB 0x67
- mov r8,rbx
- imul rbx,QWORD[((32+8))+rsp]
- mov rax,QWORD[rbp]
- mov ecx,8
- jmp NEAR $L$8x_reduce
- ALIGN 32
- $L$8x_reduce:
- mul rbx
- mov rax,QWORD[8+rbp]
- neg r8
- mov r8,rdx
- adc r8,0
- mul rbx
- add r9,rax
- mov rax,QWORD[16+rbp]
- adc rdx,0
- add r8,r9
- mov QWORD[((48-8+8))+rcx*8+rsp],rbx
- mov r9,rdx
- adc r9,0
- mul rbx
- add r10,rax
- mov rax,QWORD[24+rbp]
- adc rdx,0
- add r9,r10
- mov rsi,QWORD[((32+8))+rsp]
- mov r10,rdx
- adc r10,0
- mul rbx
- add r11,rax
- mov rax,QWORD[32+rbp]
- adc rdx,0
- imul rsi,r8
- add r10,r11
- mov r11,rdx
- adc r11,0
- mul rbx
- add r12,rax
- mov rax,QWORD[40+rbp]
- adc rdx,0
- add r11,r12
- mov r12,rdx
- adc r12,0
- mul rbx
- add r13,rax
- mov rax,QWORD[48+rbp]
- adc rdx,0
- add r12,r13
- mov r13,rdx
- adc r13,0
- mul rbx
- add r14,rax
- mov rax,QWORD[56+rbp]
- adc rdx,0
- add r13,r14
- mov r14,rdx
- adc r14,0
- mul rbx
- mov rbx,rsi
- add r15,rax
- mov rax,QWORD[rbp]
- adc rdx,0
- add r14,r15
- mov r15,rdx
- adc r15,0
- dec ecx
- jnz NEAR $L$8x_reduce
- lea rbp,[64+rbp]
- xor rax,rax
- mov rdx,QWORD[((8+8))+rsp]
- cmp rbp,QWORD[((0+8))+rsp]
- jae NEAR $L$8x_no_tail
- DB 0x66
- add r8,QWORD[rdi]
- adc r9,QWORD[8+rdi]
- adc r10,QWORD[16+rdi]
- adc r11,QWORD[24+rdi]
- adc r12,QWORD[32+rdi]
- adc r13,QWORD[40+rdi]
- adc r14,QWORD[48+rdi]
- adc r15,QWORD[56+rdi]
- sbb rsi,rsi
- mov rbx,QWORD[((48+56+8))+rsp]
- mov ecx,8
- mov rax,QWORD[rbp]
- jmp NEAR $L$8x_tail
- ALIGN 32
- $L$8x_tail:
- mul rbx
- add r8,rax
- mov rax,QWORD[8+rbp]
- mov QWORD[rdi],r8
- mov r8,rdx
- adc r8,0
- mul rbx
- add r9,rax
- mov rax,QWORD[16+rbp]
- adc rdx,0
- add r8,r9
- lea rdi,[8+rdi]
- mov r9,rdx
- adc r9,0
- mul rbx
- add r10,rax
- mov rax,QWORD[24+rbp]
- adc rdx,0
- add r9,r10
- mov r10,rdx
- adc r10,0
- mul rbx
- add r11,rax
- mov rax,QWORD[32+rbp]
- adc rdx,0
- add r10,r11
- mov r11,rdx
- adc r11,0
- mul rbx
- add r12,rax
- mov rax,QWORD[40+rbp]
- adc rdx,0
- add r11,r12
- mov r12,rdx
- adc r12,0
- mul rbx
- add r13,rax
- mov rax,QWORD[48+rbp]
- adc rdx,0
- add r12,r13
- mov r13,rdx
- adc r13,0
- mul rbx
- add r14,rax
- mov rax,QWORD[56+rbp]
- adc rdx,0
- add r13,r14
- mov r14,rdx
- adc r14,0
- mul rbx
- mov rbx,QWORD[((48-16+8))+rcx*8+rsp]
- add r15,rax
- adc rdx,0
- add r14,r15
- mov rax,QWORD[rbp]
- mov r15,rdx
- adc r15,0
- dec ecx
- jnz NEAR $L$8x_tail
- lea rbp,[64+rbp]
- mov rdx,QWORD[((8+8))+rsp]
- cmp rbp,QWORD[((0+8))+rsp]
- jae NEAR $L$8x_tail_done
- mov rbx,QWORD[((48+56+8))+rsp]
- neg rsi
- mov rax,QWORD[rbp]
- adc r8,QWORD[rdi]
- adc r9,QWORD[8+rdi]
- adc r10,QWORD[16+rdi]
- adc r11,QWORD[24+rdi]
- adc r12,QWORD[32+rdi]
- adc r13,QWORD[40+rdi]
- adc r14,QWORD[48+rdi]
- adc r15,QWORD[56+rdi]
- sbb rsi,rsi
- mov ecx,8
- jmp NEAR $L$8x_tail
- ALIGN 32
- $L$8x_tail_done:
- xor rax,rax
- add r8,QWORD[rdx]
- adc r9,0
- adc r10,0
- adc r11,0
- adc r12,0
- adc r13,0
- adc r14,0
- adc r15,0
- adc rax,0
- neg rsi
- $L$8x_no_tail:
- adc r8,QWORD[rdi]
- adc r9,QWORD[8+rdi]
- adc r10,QWORD[16+rdi]
- adc r11,QWORD[24+rdi]
- adc r12,QWORD[32+rdi]
- adc r13,QWORD[40+rdi]
- adc r14,QWORD[48+rdi]
- adc r15,QWORD[56+rdi]
- adc rax,0
- mov rcx,QWORD[((-8))+rbp]
- xor rsi,rsi
- DB 102,72,15,126,213
- mov QWORD[rdi],r8
- mov QWORD[8+rdi],r9
- DB 102,73,15,126,217
- mov QWORD[16+rdi],r10
- mov QWORD[24+rdi],r11
- mov QWORD[32+rdi],r12
- mov QWORD[40+rdi],r13
- mov QWORD[48+rdi],r14
- mov QWORD[56+rdi],r15
- lea rdi,[64+rdi]
- cmp rdi,rdx
- jb NEAR $L$8x_reduction_loop
- DB 0F3h,0C3h ;repret
- ALIGN 32
- __bn_post4x_internal:
- mov r12,QWORD[rbp]
- lea rbx,[r9*1+rdi]
- mov rcx,r9
- DB 102,72,15,126,207
- neg rax
- DB 102,72,15,126,206
- sar rcx,3+2
- dec r12
- xor r10,r10
- mov r13,QWORD[8+rbp]
- mov r14,QWORD[16+rbp]
- mov r15,QWORD[24+rbp]
- jmp NEAR $L$sqr4x_sub_entry
- ALIGN 16
- $L$sqr4x_sub:
- mov r12,QWORD[rbp]
- mov r13,QWORD[8+rbp]
- mov r14,QWORD[16+rbp]
- mov r15,QWORD[24+rbp]
- $L$sqr4x_sub_entry:
- lea rbp,[32+rbp]
- not r12
- not r13
- not r14
- not r15
- and r12,rax
- and r13,rax
- and r14,rax
- and r15,rax
- neg r10
- adc r12,QWORD[rbx]
- adc r13,QWORD[8+rbx]
- adc r14,QWORD[16+rbx]
- adc r15,QWORD[24+rbx]
- mov QWORD[rdi],r12
- lea rbx,[32+rbx]
- mov QWORD[8+rdi],r13
- sbb r10,r10
- mov QWORD[16+rdi],r14
- mov QWORD[24+rdi],r15
- lea rdi,[32+rdi]
- inc rcx
- jnz NEAR $L$sqr4x_sub
- mov r10,r9
- neg r9
- DB 0F3h,0C3h ;repret
- global GFp_bn_from_montgomery
- ALIGN 32
- GFp_bn_from_montgomery:
- test DWORD[48+rsp],7
- jz NEAR bn_from_mont8x
- xor eax,eax
- DB 0F3h,0C3h ;repret
- ALIGN 32
- bn_from_mont8x:
- mov QWORD[8+rsp],rdi ;WIN64 prologue
- mov QWORD[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_bn_from_mont8x:
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- mov r8,QWORD[40+rsp]
- mov r9,QWORD[48+rsp]
- DB 0x67
- mov rax,rsp
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- $L$from_prologue:
- shl r9d,3
- lea r10,[r9*2+r9]
- neg r9
- mov r8,QWORD[r8]
- lea r11,[((-320))+r9*2+rsp]
- mov rbp,rsp
- sub r11,rdi
- and r11,4095
- cmp r10,r11
- jb NEAR $L$from_sp_alt
- sub rbp,r11
- lea rbp,[((-320))+r9*2+rbp]
- jmp NEAR $L$from_sp_done
- ALIGN 32
- $L$from_sp_alt:
- lea r10,[((4096-320))+r9*2]
- lea rbp,[((-320))+r9*2+rbp]
- sub r11,r10
- mov r10,0
- cmovc r11,r10
- sub rbp,r11
- $L$from_sp_done:
- and rbp,-64
- mov r11,rsp
- sub r11,rbp
- and r11,-4096
- lea rsp,[rbp*1+r11]
- mov r10,QWORD[rsp]
- cmp rsp,rbp
- ja NEAR $L$from_page_walk
- jmp NEAR $L$from_page_walk_done
- $L$from_page_walk:
- lea rsp,[((-4096))+rsp]
- mov r10,QWORD[rsp]
- cmp rsp,rbp
- ja NEAR $L$from_page_walk
- $L$from_page_walk_done:
- mov r10,r9
- neg r9
- mov QWORD[32+rsp],r8
- mov QWORD[40+rsp],rax
- $L$from_body:
- mov r11,r9
- lea rax,[48+rsp]
- pxor xmm0,xmm0
- jmp NEAR $L$mul_by_1
- ALIGN 32
- $L$mul_by_1:
- movdqu xmm1,XMMWORD[rsi]
- movdqu xmm2,XMMWORD[16+rsi]
- movdqu xmm3,XMMWORD[32+rsi]
- movdqa XMMWORD[r9*1+rax],xmm0
- movdqu xmm4,XMMWORD[48+rsi]
- movdqa XMMWORD[16+r9*1+rax],xmm0
- DB 0x48,0x8d,0xb6,0x40,0x00,0x00,0x00
- movdqa XMMWORD[rax],xmm1
- movdqa XMMWORD[32+r9*1+rax],xmm0
- movdqa XMMWORD[16+rax],xmm2
- movdqa XMMWORD[48+r9*1+rax],xmm0
- movdqa XMMWORD[32+rax],xmm3
- movdqa XMMWORD[48+rax],xmm4
- lea rax,[64+rax]
- sub r11,64
- jnz NEAR $L$mul_by_1
- DB 102,72,15,110,207
- DB 102,72,15,110,209
- DB 0x67
- mov rbp,rcx
- DB 102,73,15,110,218
- lea r11,[GFp_ia32cap_P]
- mov r11d,DWORD[8+r11]
- and r11d,0x80108
- cmp r11d,0x80108
- jne NEAR $L$from_mont_nox
- lea rdi,[r9*1+rax]
- call __bn_sqrx8x_reduction
- call __bn_postx4x_internal
- pxor xmm0,xmm0
- lea rax,[48+rsp]
- jmp NEAR $L$from_mont_zero
- ALIGN 32
- $L$from_mont_nox:
- call __bn_sqr8x_reduction
- call __bn_post4x_internal
- pxor xmm0,xmm0
- lea rax,[48+rsp]
- jmp NEAR $L$from_mont_zero
- ALIGN 32
- $L$from_mont_zero:
- mov rsi,QWORD[40+rsp]
- movdqa XMMWORD[rax],xmm0
- movdqa XMMWORD[16+rax],xmm0
- movdqa XMMWORD[32+rax],xmm0
- movdqa XMMWORD[48+rax],xmm0
- lea rax,[64+rax]
- sub r9,32
- jnz NEAR $L$from_mont_zero
- mov rax,1
- 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]
- $L$from_epilogue:
- mov rdi,QWORD[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_bn_from_mont8x:
- ALIGN 32
- bn_mulx4x_mont_gather5:
- mov QWORD[8+rsp],rdi ;WIN64 prologue
- mov QWORD[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_bn_mulx4x_mont_gather5:
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- mov r8,QWORD[40+rsp]
- mov r9,QWORD[48+rsp]
- mov rax,rsp
- $L$mulx4x_enter:
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- $L$mulx4x_prologue:
- shl r9d,3
- lea r10,[r9*2+r9]
- neg r9
- mov r8,QWORD[r8]
- lea r11,[((-320))+r9*2+rsp]
- mov rbp,rsp
- sub r11,rdi
- and r11,4095
- cmp r10,r11
- jb NEAR $L$mulx4xsp_alt
- sub rbp,r11
- lea rbp,[((-320))+r9*2+rbp]
- jmp NEAR $L$mulx4xsp_done
- $L$mulx4xsp_alt:
- lea r10,[((4096-320))+r9*2]
- lea rbp,[((-320))+r9*2+rbp]
- sub r11,r10
- mov r10,0
- cmovc r11,r10
- sub rbp,r11
- $L$mulx4xsp_done:
- and rbp,-64
- mov r11,rsp
- sub r11,rbp
- and r11,-4096
- lea rsp,[rbp*1+r11]
- mov r10,QWORD[rsp]
- cmp rsp,rbp
- ja NEAR $L$mulx4x_page_walk
- jmp NEAR $L$mulx4x_page_walk_done
- $L$mulx4x_page_walk:
- lea rsp,[((-4096))+rsp]
- mov r10,QWORD[rsp]
- cmp rsp,rbp
- ja NEAR $L$mulx4x_page_walk
- $L$mulx4x_page_walk_done:
- mov QWORD[32+rsp],r8
- mov QWORD[40+rsp],rax
- $L$mulx4x_body:
- call mulx4x_internal
- mov rsi,QWORD[40+rsp]
- mov rax,1
- 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]
- $L$mulx4x_epilogue:
- mov rdi,QWORD[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_bn_mulx4x_mont_gather5:
- ALIGN 32
- mulx4x_internal:
- mov QWORD[8+rsp],r9
- mov r10,r9
- neg r9
- shl r9,5
- neg r10
- lea r13,[128+r9*1+rdx]
- shr r9,5+5
- movd xmm5,DWORD[56+rax]
- sub r9,1
- lea rax,[$L$inc]
- mov QWORD[((16+8))+rsp],r13
- mov QWORD[((24+8))+rsp],r9
- mov QWORD[((56+8))+rsp],rdi
- movdqa xmm0,XMMWORD[rax]
- movdqa xmm1,XMMWORD[16+rax]
- lea r10,[((88-112))+r10*1+rsp]
- lea rdi,[128+rdx]
- pshufd xmm5,xmm5,0
- movdqa xmm4,xmm1
- DB 0x67
- movdqa xmm2,xmm1
- DB 0x67
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[112+r10],xmm0
- movdqa xmm0,xmm4
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[128+r10],xmm1
- movdqa xmm1,xmm4
- paddd xmm0,xmm3
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[144+r10],xmm2
- movdqa xmm2,xmm4
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- movdqa XMMWORD[160+r10],xmm3
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[176+r10],xmm0
- movdqa xmm0,xmm4
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[192+r10],xmm1
- movdqa xmm1,xmm4
- paddd xmm0,xmm3
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[208+r10],xmm2
- movdqa xmm2,xmm4
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- movdqa XMMWORD[224+r10],xmm3
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[240+r10],xmm0
- movdqa xmm0,xmm4
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[256+r10],xmm1
- movdqa xmm1,xmm4
- paddd xmm0,xmm3
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[272+r10],xmm2
- movdqa xmm2,xmm4
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- movdqa XMMWORD[288+r10],xmm3
- movdqa xmm3,xmm4
- DB 0x67
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[304+r10],xmm0
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[320+r10],xmm1
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[336+r10],xmm2
- pand xmm0,XMMWORD[64+rdi]
- pand xmm1,XMMWORD[80+rdi]
- pand xmm2,XMMWORD[96+rdi]
- movdqa XMMWORD[352+r10],xmm3
- pand xmm3,XMMWORD[112+rdi]
- por xmm0,xmm2
- por xmm1,xmm3
- movdqa xmm4,XMMWORD[((-128))+rdi]
- movdqa xmm5,XMMWORD[((-112))+rdi]
- movdqa xmm2,XMMWORD[((-96))+rdi]
- pand xmm4,XMMWORD[112+r10]
- movdqa xmm3,XMMWORD[((-80))+rdi]
- pand xmm5,XMMWORD[128+r10]
- por xmm0,xmm4
- pand xmm2,XMMWORD[144+r10]
- por xmm1,xmm5
- pand xmm3,XMMWORD[160+r10]
- por xmm0,xmm2
- por xmm1,xmm3
- movdqa xmm4,XMMWORD[((-64))+rdi]
- movdqa xmm5,XMMWORD[((-48))+rdi]
- movdqa xmm2,XMMWORD[((-32))+rdi]
- pand xmm4,XMMWORD[176+r10]
- movdqa xmm3,XMMWORD[((-16))+rdi]
- pand xmm5,XMMWORD[192+r10]
- por xmm0,xmm4
- pand xmm2,XMMWORD[208+r10]
- por xmm1,xmm5
- pand xmm3,XMMWORD[224+r10]
- por xmm0,xmm2
- por xmm1,xmm3
- movdqa xmm4,XMMWORD[rdi]
- movdqa xmm5,XMMWORD[16+rdi]
- movdqa xmm2,XMMWORD[32+rdi]
- pand xmm4,XMMWORD[240+r10]
- movdqa xmm3,XMMWORD[48+rdi]
- pand xmm5,XMMWORD[256+r10]
- por xmm0,xmm4
- pand xmm2,XMMWORD[272+r10]
- por xmm1,xmm5
- pand xmm3,XMMWORD[288+r10]
- por xmm0,xmm2
- por xmm1,xmm3
- pxor xmm0,xmm1
- pshufd xmm1,xmm0,0x4e
- por xmm0,xmm1
- lea rdi,[256+rdi]
- DB 102,72,15,126,194
- lea rbx,[((64+32+8))+rsp]
- mov r9,rdx
- mulx rax,r8,QWORD[rsi]
- mulx r12,r11,QWORD[8+rsi]
- add r11,rax
- mulx r13,rax,QWORD[16+rsi]
- adc r12,rax
- adc r13,0
- mulx r14,rax,QWORD[24+rsi]
- mov r15,r8
- imul r8,QWORD[((32+8))+rsp]
- xor rbp,rbp
- mov rdx,r8
- mov QWORD[((8+8))+rsp],rdi
- lea rsi,[32+rsi]
- adcx r13,rax
- adcx r14,rbp
- mulx r10,rax,QWORD[rcx]
- adcx r15,rax
- adox r10,r11
- mulx r11,rax,QWORD[8+rcx]
- adcx r10,rax
- adox r11,r12
- mulx r12,rax,QWORD[16+rcx]
- mov rdi,QWORD[((24+8))+rsp]
- mov QWORD[((-32))+rbx],r10
- adcx r11,rax
- adox r12,r13
- mulx r15,rax,QWORD[24+rcx]
- mov rdx,r9
- mov QWORD[((-24))+rbx],r11
- adcx r12,rax
- adox r15,rbp
- lea rcx,[32+rcx]
- mov QWORD[((-16))+rbx],r12
- jmp NEAR $L$mulx4x_1st
- ALIGN 32
- $L$mulx4x_1st:
- adcx r15,rbp
- mulx rax,r10,QWORD[rsi]
- adcx r10,r14
- mulx r14,r11,QWORD[8+rsi]
- adcx r11,rax
- mulx rax,r12,QWORD[16+rsi]
- adcx r12,r14
- mulx r14,r13,QWORD[24+rsi]
- DB 0x67,0x67
- mov rdx,r8
- adcx r13,rax
- adcx r14,rbp
- lea rsi,[32+rsi]
- lea rbx,[32+rbx]
- adox r10,r15
- mulx r15,rax,QWORD[rcx]
- adcx r10,rax
- adox r11,r15
- mulx r15,rax,QWORD[8+rcx]
- adcx r11,rax
- adox r12,r15
- mulx r15,rax,QWORD[16+rcx]
- mov QWORD[((-40))+rbx],r10
- adcx r12,rax
- mov QWORD[((-32))+rbx],r11
- adox r13,r15
- mulx r15,rax,QWORD[24+rcx]
- mov rdx,r9
- mov QWORD[((-24))+rbx],r12
- adcx r13,rax
- adox r15,rbp
- lea rcx,[32+rcx]
- mov QWORD[((-16))+rbx],r13
- dec rdi
- jnz NEAR $L$mulx4x_1st
- mov rax,QWORD[8+rsp]
- adc r15,rbp
- lea rsi,[rax*1+rsi]
- add r14,r15
- mov rdi,QWORD[((8+8))+rsp]
- adc rbp,rbp
- mov QWORD[((-8))+rbx],r14
- jmp NEAR $L$mulx4x_outer
- ALIGN 32
- $L$mulx4x_outer:
- lea r10,[((16-256))+rbx]
- pxor xmm4,xmm4
- DB 0x67,0x67
- pxor xmm5,xmm5
- movdqa xmm0,XMMWORD[((-128))+rdi]
- movdqa xmm1,XMMWORD[((-112))+rdi]
- movdqa xmm2,XMMWORD[((-96))+rdi]
- pand xmm0,XMMWORD[256+r10]
- movdqa xmm3,XMMWORD[((-80))+rdi]
- pand xmm1,XMMWORD[272+r10]
- por xmm4,xmm0
- pand xmm2,XMMWORD[288+r10]
- por xmm5,xmm1
- pand xmm3,XMMWORD[304+r10]
- por xmm4,xmm2
- por xmm5,xmm3
- movdqa xmm0,XMMWORD[((-64))+rdi]
- movdqa xmm1,XMMWORD[((-48))+rdi]
- movdqa xmm2,XMMWORD[((-32))+rdi]
- pand xmm0,XMMWORD[320+r10]
- movdqa xmm3,XMMWORD[((-16))+rdi]
- pand xmm1,XMMWORD[336+r10]
- por xmm4,xmm0
- pand xmm2,XMMWORD[352+r10]
- por xmm5,xmm1
- pand xmm3,XMMWORD[368+r10]
- por xmm4,xmm2
- por xmm5,xmm3
- movdqa xmm0,XMMWORD[rdi]
- movdqa xmm1,XMMWORD[16+rdi]
- movdqa xmm2,XMMWORD[32+rdi]
- pand xmm0,XMMWORD[384+r10]
- movdqa xmm3,XMMWORD[48+rdi]
- pand xmm1,XMMWORD[400+r10]
- por xmm4,xmm0
- pand xmm2,XMMWORD[416+r10]
- por xmm5,xmm1
- pand xmm3,XMMWORD[432+r10]
- por xmm4,xmm2
- por xmm5,xmm3
- movdqa xmm0,XMMWORD[64+rdi]
- movdqa xmm1,XMMWORD[80+rdi]
- movdqa xmm2,XMMWORD[96+rdi]
- pand xmm0,XMMWORD[448+r10]
- movdqa xmm3,XMMWORD[112+rdi]
- pand xmm1,XMMWORD[464+r10]
- por xmm4,xmm0
- pand xmm2,XMMWORD[480+r10]
- por xmm5,xmm1
- pand xmm3,XMMWORD[496+r10]
- por xmm4,xmm2
- por xmm5,xmm3
- por xmm4,xmm5
- pshufd xmm0,xmm4,0x4e
- por xmm0,xmm4
- lea rdi,[256+rdi]
- DB 102,72,15,126,194
- mov QWORD[rbx],rbp
- lea rbx,[32+rax*1+rbx]
- mulx r11,r8,QWORD[rsi]
- xor rbp,rbp
- mov r9,rdx
- mulx r12,r14,QWORD[8+rsi]
- adox r8,QWORD[((-32))+rbx]
- adcx r11,r14
- mulx r13,r15,QWORD[16+rsi]
- adox r11,QWORD[((-24))+rbx]
- adcx r12,r15
- mulx r14,rdx,QWORD[24+rsi]
- adox r12,QWORD[((-16))+rbx]
- adcx r13,rdx
- lea rcx,[rax*1+rcx]
- lea rsi,[32+rsi]
- adox r13,QWORD[((-8))+rbx]
- adcx r14,rbp
- adox r14,rbp
- mov r15,r8
- imul r8,QWORD[((32+8))+rsp]
- mov rdx,r8
- xor rbp,rbp
- mov QWORD[((8+8))+rsp],rdi
- mulx r10,rax,QWORD[rcx]
- adcx r15,rax
- adox r10,r11
- mulx r11,rax,QWORD[8+rcx]
- adcx r10,rax
- adox r11,r12
- mulx r12,rax,QWORD[16+rcx]
- adcx r11,rax
- adox r12,r13
- mulx r15,rax,QWORD[24+rcx]
- mov rdx,r9
- mov rdi,QWORD[((24+8))+rsp]
- mov QWORD[((-32))+rbx],r10
- adcx r12,rax
- mov QWORD[((-24))+rbx],r11
- adox r15,rbp
- mov QWORD[((-16))+rbx],r12
- lea rcx,[32+rcx]
- jmp NEAR $L$mulx4x_inner
- ALIGN 32
- $L$mulx4x_inner:
- mulx rax,r10,QWORD[rsi]
- adcx r15,rbp
- adox r10,r14
- mulx r14,r11,QWORD[8+rsi]
- adcx r10,QWORD[rbx]
- adox r11,rax
- mulx rax,r12,QWORD[16+rsi]
- adcx r11,QWORD[8+rbx]
- adox r12,r14
- mulx r14,r13,QWORD[24+rsi]
- mov rdx,r8
- adcx r12,QWORD[16+rbx]
- adox r13,rax
- adcx r13,QWORD[24+rbx]
- adox r14,rbp
- lea rsi,[32+rsi]
- lea rbx,[32+rbx]
- adcx r14,rbp
- adox r10,r15
- mulx r15,rax,QWORD[rcx]
- adcx r10,rax
- adox r11,r15
- mulx r15,rax,QWORD[8+rcx]
- adcx r11,rax
- adox r12,r15
- mulx r15,rax,QWORD[16+rcx]
- mov QWORD[((-40))+rbx],r10
- adcx r12,rax
- adox r13,r15
- mov QWORD[((-32))+rbx],r11
- mulx r15,rax,QWORD[24+rcx]
- mov rdx,r9
- lea rcx,[32+rcx]
- mov QWORD[((-24))+rbx],r12
- adcx r13,rax
- adox r15,rbp
- mov QWORD[((-16))+rbx],r13
- dec rdi
- jnz NEAR $L$mulx4x_inner
- mov rax,QWORD[((0+8))+rsp]
- adc r15,rbp
- sub rdi,QWORD[rbx]
- mov rdi,QWORD[((8+8))+rsp]
- mov r10,QWORD[((16+8))+rsp]
- adc r14,r15
- lea rsi,[rax*1+rsi]
- adc rbp,rbp
- mov QWORD[((-8))+rbx],r14
- cmp rdi,r10
- jb NEAR $L$mulx4x_outer
- mov r10,QWORD[((-8))+rcx]
- mov r8,rbp
- mov r12,QWORD[rax*1+rcx]
- lea rbp,[rax*1+rcx]
- mov rcx,rax
- lea rdi,[rax*1+rbx]
- xor eax,eax
- xor r15,r15
- sub r10,r14
- adc r15,r15
- or r8,r15
- sar rcx,3+2
- sub rax,r8
- mov rdx,QWORD[((56+8))+rsp]
- dec r12
- mov r13,QWORD[8+rbp]
- xor r8,r8
- mov r14,QWORD[16+rbp]
- mov r15,QWORD[24+rbp]
- jmp NEAR $L$sqrx4x_sub_entry
- ALIGN 32
- bn_powerx5:
- mov QWORD[8+rsp],rdi ;WIN64 prologue
- mov QWORD[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_bn_powerx5:
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- mov r8,QWORD[40+rsp]
- mov r9,QWORD[48+rsp]
- mov rax,rsp
- $L$powerx5_enter:
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- $L$powerx5_prologue:
- shl r9d,3
- lea r10,[r9*2+r9]
- neg r9
- mov r8,QWORD[r8]
- lea r11,[((-320))+r9*2+rsp]
- mov rbp,rsp
- sub r11,rdi
- and r11,4095
- cmp r10,r11
- jb NEAR $L$pwrx_sp_alt
- sub rbp,r11
- lea rbp,[((-320))+r9*2+rbp]
- jmp NEAR $L$pwrx_sp_done
- ALIGN 32
- $L$pwrx_sp_alt:
- lea r10,[((4096-320))+r9*2]
- lea rbp,[((-320))+r9*2+rbp]
- sub r11,r10
- mov r10,0
- cmovc r11,r10
- sub rbp,r11
- $L$pwrx_sp_done:
- and rbp,-64
- mov r11,rsp
- sub r11,rbp
- and r11,-4096
- lea rsp,[rbp*1+r11]
- mov r10,QWORD[rsp]
- cmp rsp,rbp
- ja NEAR $L$pwrx_page_walk
- jmp NEAR $L$pwrx_page_walk_done
- $L$pwrx_page_walk:
- lea rsp,[((-4096))+rsp]
- mov r10,QWORD[rsp]
- cmp rsp,rbp
- ja NEAR $L$pwrx_page_walk
- $L$pwrx_page_walk_done:
- mov r10,r9
- neg r9
- pxor xmm0,xmm0
- DB 102,72,15,110,207
- DB 102,72,15,110,209
- DB 102,73,15,110,218
- DB 102,72,15,110,226
- mov QWORD[32+rsp],r8
- mov QWORD[40+rsp],rax
- $L$powerx5_body:
- call __bn_sqrx8x_internal
- call __bn_postx4x_internal
- call __bn_sqrx8x_internal
- call __bn_postx4x_internal
- call __bn_sqrx8x_internal
- call __bn_postx4x_internal
- call __bn_sqrx8x_internal
- call __bn_postx4x_internal
- call __bn_sqrx8x_internal
- call __bn_postx4x_internal
- mov r9,r10
- mov rdi,rsi
- DB 102,72,15,126,209
- DB 102,72,15,126,226
- mov rax,QWORD[40+rsp]
- call mulx4x_internal
- mov rsi,QWORD[40+rsp]
- mov rax,1
- 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]
- $L$powerx5_epilogue:
- mov rdi,QWORD[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_bn_powerx5:
- global GFp_bn_sqrx8x_internal
- ALIGN 32
- GFp_bn_sqrx8x_internal:
- __bn_sqrx8x_internal:
- lea rdi,[((48+8))+rsp]
- lea rbp,[r9*1+rsi]
- mov QWORD[((0+8))+rsp],r9
- mov QWORD[((8+8))+rsp],rbp
- jmp NEAR $L$sqr8x_zero_start
- ALIGN 32
- DB 0x66,0x66,0x66,0x2e,0x0f,0x1f,0x84,0x00,0x00,0x00,0x00,0x00
- $L$sqrx8x_zero:
- DB 0x3e
- movdqa XMMWORD[rdi],xmm0
- movdqa XMMWORD[16+rdi],xmm0
- movdqa XMMWORD[32+rdi],xmm0
- movdqa XMMWORD[48+rdi],xmm0
- $L$sqr8x_zero_start:
- movdqa XMMWORD[64+rdi],xmm0
- movdqa XMMWORD[80+rdi],xmm0
- movdqa XMMWORD[96+rdi],xmm0
- movdqa XMMWORD[112+rdi],xmm0
- lea rdi,[128+rdi]
- sub r9,64
- jnz NEAR $L$sqrx8x_zero
- mov rdx,QWORD[rsi]
- xor r10,r10
- xor r11,r11
- xor r12,r12
- xor r13,r13
- xor r14,r14
- xor r15,r15
- lea rdi,[((48+8))+rsp]
- xor rbp,rbp
- jmp NEAR $L$sqrx8x_outer_loop
- ALIGN 32
- $L$sqrx8x_outer_loop:
- mulx rax,r8,QWORD[8+rsi]
- adcx r8,r9
- adox r10,rax
- mulx rax,r9,QWORD[16+rsi]
- adcx r9,r10
- adox r11,rax
- DB 0xc4,0xe2,0xab,0xf6,0x86,0x18,0x00,0x00,0x00
- adcx r10,r11
- adox r12,rax
- DB 0xc4,0xe2,0xa3,0xf6,0x86,0x20,0x00,0x00,0x00
- adcx r11,r12
- adox r13,rax
- mulx rax,r12,QWORD[40+rsi]
- adcx r12,r13
- adox r14,rax
- mulx rax,r13,QWORD[48+rsi]
- adcx r13,r14
- adox rax,r15
- mulx r15,r14,QWORD[56+rsi]
- mov rdx,QWORD[8+rsi]
- adcx r14,rax
- adox r15,rbp
- adc r15,QWORD[64+rdi]
- mov QWORD[8+rdi],r8
- mov QWORD[16+rdi],r9
- sbb rcx,rcx
- xor rbp,rbp
- mulx rbx,r8,QWORD[16+rsi]
- mulx rax,r9,QWORD[24+rsi]
- adcx r8,r10
- adox r9,rbx
- mulx rbx,r10,QWORD[32+rsi]
- adcx r9,r11
- adox r10,rax
- DB 0xc4,0xe2,0xa3,0xf6,0x86,0x28,0x00,0x00,0x00
- adcx r10,r12
- adox r11,rbx
- DB 0xc4,0xe2,0x9b,0xf6,0x9e,0x30,0x00,0x00,0x00
- adcx r11,r13
- adox r12,r14
- DB 0xc4,0x62,0x93,0xf6,0xb6,0x38,0x00,0x00,0x00
- mov rdx,QWORD[16+rsi]
- adcx r12,rax
- adox r13,rbx
- adcx r13,r15
- adox r14,rbp
- adcx r14,rbp
- mov QWORD[24+rdi],r8
- mov QWORD[32+rdi],r9
- mulx rbx,r8,QWORD[24+rsi]
- mulx rax,r9,QWORD[32+rsi]
- adcx r8,r10
- adox r9,rbx
- mulx rbx,r10,QWORD[40+rsi]
- adcx r9,r11
- adox r10,rax
- DB 0xc4,0xe2,0xa3,0xf6,0x86,0x30,0x00,0x00,0x00
- adcx r10,r12
- adox r11,r13
- DB 0xc4,0x62,0x9b,0xf6,0xae,0x38,0x00,0x00,0x00
- DB 0x3e
- mov rdx,QWORD[24+rsi]
- adcx r11,rbx
- adox r12,rax
- adcx r12,r14
- mov QWORD[40+rdi],r8
- mov QWORD[48+rdi],r9
- mulx rax,r8,QWORD[32+rsi]
- adox r13,rbp
- adcx r13,rbp
- mulx rbx,r9,QWORD[40+rsi]
- adcx r8,r10
- adox r9,rax
- mulx rax,r10,QWORD[48+rsi]
- adcx r9,r11
- adox r10,r12
- mulx r12,r11,QWORD[56+rsi]
- mov rdx,QWORD[32+rsi]
- mov r14,QWORD[40+rsi]
- adcx r10,rbx
- adox r11,rax
- mov r15,QWORD[48+rsi]
- adcx r11,r13
- adox r12,rbp
- adcx r12,rbp
- mov QWORD[56+rdi],r8
- mov QWORD[64+rdi],r9
- mulx rax,r9,r14
- mov r8,QWORD[56+rsi]
- adcx r9,r10
- mulx rbx,r10,r15
- adox r10,rax
- adcx r10,r11
- mulx rax,r11,r8
- mov rdx,r14
- adox r11,rbx
- adcx r11,r12
- adcx rax,rbp
- mulx rbx,r14,r15
- mulx r13,r12,r8
- mov rdx,r15
- lea rsi,[64+rsi]
- adcx r11,r14
- adox r12,rbx
- adcx r12,rax
- adox r13,rbp
- DB 0x67,0x67
- mulx r14,r8,r8
- adcx r13,r8
- adcx r14,rbp
- cmp rsi,QWORD[((8+8))+rsp]
- je NEAR $L$sqrx8x_outer_break
- neg rcx
- mov rcx,-8
- mov r15,rbp
- mov r8,QWORD[64+rdi]
- adcx r9,QWORD[72+rdi]
- adcx r10,QWORD[80+rdi]
- adcx r11,QWORD[88+rdi]
- adc r12,QWORD[96+rdi]
- adc r13,QWORD[104+rdi]
- adc r14,QWORD[112+rdi]
- adc r15,QWORD[120+rdi]
- lea rbp,[rsi]
- lea rdi,[128+rdi]
- sbb rax,rax
- mov rdx,QWORD[((-64))+rsi]
- mov QWORD[((16+8))+rsp],rax
- mov QWORD[((24+8))+rsp],rdi
- xor eax,eax
- jmp NEAR $L$sqrx8x_loop
- ALIGN 32
- $L$sqrx8x_loop:
- mov rbx,r8
- mulx r8,rax,QWORD[rbp]
- adcx rbx,rax
- adox r8,r9
- mulx r9,rax,QWORD[8+rbp]
- adcx r8,rax
- adox r9,r10
- mulx r10,rax,QWORD[16+rbp]
- adcx r9,rax
- adox r10,r11
- mulx r11,rax,QWORD[24+rbp]
- adcx r10,rax
- adox r11,r12
- DB 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00
- adcx r11,rax
- adox r12,r13
- mulx r13,rax,QWORD[40+rbp]
- adcx r12,rax
- adox r13,r14
- mulx r14,rax,QWORD[48+rbp]
- mov QWORD[rcx*8+rdi],rbx
- mov ebx,0
- adcx r13,rax
- adox r14,r15
- DB 0xc4,0x62,0xfb,0xf6,0xbd,0x38,0x00,0x00,0x00
- mov rdx,QWORD[8+rcx*8+rsi]
- adcx r14,rax
- adox r15,rbx
- adcx r15,rbx
- DB 0x67
- inc rcx
- jnz NEAR $L$sqrx8x_loop
- lea rbp,[64+rbp]
- mov rcx,-8
- cmp rbp,QWORD[((8+8))+rsp]
- je NEAR $L$sqrx8x_break
- sub rbx,QWORD[((16+8))+rsp]
- DB 0x66
- mov rdx,QWORD[((-64))+rsi]
- adcx r8,QWORD[rdi]
- adcx r9,QWORD[8+rdi]
- adc r10,QWORD[16+rdi]
- adc r11,QWORD[24+rdi]
- adc r12,QWORD[32+rdi]
- adc r13,QWORD[40+rdi]
- adc r14,QWORD[48+rdi]
- adc r15,QWORD[56+rdi]
- lea rdi,[64+rdi]
- DB 0x67
- sbb rax,rax
- xor ebx,ebx
- mov QWORD[((16+8))+rsp],rax
- jmp NEAR $L$sqrx8x_loop
- ALIGN 32
- $L$sqrx8x_break:
- xor rbp,rbp
- sub rbx,QWORD[((16+8))+rsp]
- adcx r8,rbp
- mov rcx,QWORD[((24+8))+rsp]
- adcx r9,rbp
- mov rdx,QWORD[rsi]
- adc r10,0
- mov QWORD[rdi],r8
- adc r11,0
- adc r12,0
- adc r13,0
- adc r14,0
- adc r15,0
- cmp rdi,rcx
- je NEAR $L$sqrx8x_outer_loop
- mov QWORD[8+rdi],r9
- mov r9,QWORD[8+rcx]
- mov QWORD[16+rdi],r10
- mov r10,QWORD[16+rcx]
- mov QWORD[24+rdi],r11
- mov r11,QWORD[24+rcx]
- mov QWORD[32+rdi],r12
- mov r12,QWORD[32+rcx]
- mov QWORD[40+rdi],r13
- mov r13,QWORD[40+rcx]
- mov QWORD[48+rdi],r14
- mov r14,QWORD[48+rcx]
- mov QWORD[56+rdi],r15
- mov r15,QWORD[56+rcx]
- mov rdi,rcx
- jmp NEAR $L$sqrx8x_outer_loop
- ALIGN 32
- $L$sqrx8x_outer_break:
- mov QWORD[72+rdi],r9
- DB 102,72,15,126,217
- mov QWORD[80+rdi],r10
- mov QWORD[88+rdi],r11
- mov QWORD[96+rdi],r12
- mov QWORD[104+rdi],r13
- mov QWORD[112+rdi],r14
- lea rdi,[((48+8))+rsp]
- mov rdx,QWORD[rcx*1+rsi]
- mov r11,QWORD[8+rdi]
- xor r10,r10
- mov r9,QWORD[((0+8))+rsp]
- adox r11,r11
- mov r12,QWORD[16+rdi]
- mov r13,QWORD[24+rdi]
- ALIGN 32
- $L$sqrx4x_shift_n_add:
- mulx rbx,rax,rdx
- adox r12,r12
- adcx rax,r10
- DB 0x48,0x8b,0x94,0x0e,0x08,0x00,0x00,0x00
- DB 0x4c,0x8b,0x97,0x20,0x00,0x00,0x00
- adox r13,r13
- adcx rbx,r11
- mov r11,QWORD[40+rdi]
- mov QWORD[rdi],rax
- mov QWORD[8+rdi],rbx
- mulx rbx,rax,rdx
- adox r10,r10
- adcx rax,r12
- mov rdx,QWORD[16+rcx*1+rsi]
- mov r12,QWORD[48+rdi]
- adox r11,r11
- adcx rbx,r13
- mov r13,QWORD[56+rdi]
- mov QWORD[16+rdi],rax
- mov QWORD[24+rdi],rbx
- mulx rbx,rax,rdx
- adox r12,r12
- adcx rax,r10
- mov rdx,QWORD[24+rcx*1+rsi]
- lea rcx,[32+rcx]
- mov r10,QWORD[64+rdi]
- adox r13,r13
- adcx rbx,r11
- mov r11,QWORD[72+rdi]
- mov QWORD[32+rdi],rax
- mov QWORD[40+rdi],rbx
- mulx rbx,rax,rdx
- adox r10,r10
- adcx rax,r12
- jrcxz $L$sqrx4x_shift_n_add_break
- DB 0x48,0x8b,0x94,0x0e,0x00,0x00,0x00,0x00
- adox r11,r11
- adcx rbx,r13
- mov r12,QWORD[80+rdi]
- mov r13,QWORD[88+rdi]
- mov QWORD[48+rdi],rax
- mov QWORD[56+rdi],rbx
- lea rdi,[64+rdi]
- nop
- jmp NEAR $L$sqrx4x_shift_n_add
- ALIGN 32
- $L$sqrx4x_shift_n_add_break:
- adcx rbx,r13
- mov QWORD[48+rdi],rax
- mov QWORD[56+rdi],rbx
- lea rdi,[64+rdi]
- DB 102,72,15,126,213
- __bn_sqrx8x_reduction:
- xor eax,eax
- mov rbx,QWORD[((32+8))+rsp]
- mov rdx,QWORD[((48+8))+rsp]
- lea rcx,[((-64))+r9*1+rbp]
- mov QWORD[((0+8))+rsp],rcx
- mov QWORD[((8+8))+rsp],rdi
- lea rdi,[((48+8))+rsp]
- jmp NEAR $L$sqrx8x_reduction_loop
- ALIGN 32
- $L$sqrx8x_reduction_loop:
- mov r9,QWORD[8+rdi]
- mov r10,QWORD[16+rdi]
- mov r11,QWORD[24+rdi]
- mov r12,QWORD[32+rdi]
- mov r8,rdx
- imul rdx,rbx
- mov r13,QWORD[40+rdi]
- mov r14,QWORD[48+rdi]
- mov r15,QWORD[56+rdi]
- mov QWORD[((24+8))+rsp],rax
- lea rdi,[64+rdi]
- xor rsi,rsi
- mov rcx,-8
- jmp NEAR $L$sqrx8x_reduce
- ALIGN 32
- $L$sqrx8x_reduce:
- mov rbx,r8
- mulx r8,rax,QWORD[rbp]
- adcx rax,rbx
- adox r8,r9
- mulx r9,rbx,QWORD[8+rbp]
- adcx r8,rbx
- adox r9,r10
- mulx r10,rbx,QWORD[16+rbp]
- adcx r9,rbx
- adox r10,r11
- mulx r11,rbx,QWORD[24+rbp]
- adcx r10,rbx
- adox r11,r12
- DB 0xc4,0x62,0xe3,0xf6,0xa5,0x20,0x00,0x00,0x00
- mov rax,rdx
- mov rdx,r8
- adcx r11,rbx
- adox r12,r13
- mulx rdx,rbx,QWORD[((32+8))+rsp]
- mov rdx,rax
- mov QWORD[((64+48+8))+rcx*8+rsp],rax
- mulx r13,rax,QWORD[40+rbp]
- adcx r12,rax
- adox r13,r14
- mulx r14,rax,QWORD[48+rbp]
- adcx r13,rax
- adox r14,r15
- mulx r15,rax,QWORD[56+rbp]
- mov rdx,rbx
- adcx r14,rax
- adox r15,rsi
- adcx r15,rsi
- DB 0x67,0x67,0x67
- inc rcx
- jnz NEAR $L$sqrx8x_reduce
- mov rax,rsi
- cmp rbp,QWORD[((0+8))+rsp]
- jae NEAR $L$sqrx8x_no_tail
- mov rdx,QWORD[((48+8))+rsp]
- add r8,QWORD[rdi]
- lea rbp,[64+rbp]
- mov rcx,-8
- adcx r9,QWORD[8+rdi]
- adcx r10,QWORD[16+rdi]
- adc r11,QWORD[24+rdi]
- adc r12,QWORD[32+rdi]
- adc r13,QWORD[40+rdi]
- adc r14,QWORD[48+rdi]
- adc r15,QWORD[56+rdi]
- lea rdi,[64+rdi]
- sbb rax,rax
- xor rsi,rsi
- mov QWORD[((16+8))+rsp],rax
- jmp NEAR $L$sqrx8x_tail
- ALIGN 32
- $L$sqrx8x_tail:
- mov rbx,r8
- mulx r8,rax,QWORD[rbp]
- adcx rbx,rax
- adox r8,r9
- mulx r9,rax,QWORD[8+rbp]
- adcx r8,rax
- adox r9,r10
- mulx r10,rax,QWORD[16+rbp]
- adcx r9,rax
- adox r10,r11
- mulx r11,rax,QWORD[24+rbp]
- adcx r10,rax
- adox r11,r12
- DB 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00
- adcx r11,rax
- adox r12,r13
- mulx r13,rax,QWORD[40+rbp]
- adcx r12,rax
- adox r13,r14
- mulx r14,rax,QWORD[48+rbp]
- adcx r13,rax
- adox r14,r15
- mulx r15,rax,QWORD[56+rbp]
- mov rdx,QWORD[((72+48+8))+rcx*8+rsp]
- adcx r14,rax
- adox r15,rsi
- mov QWORD[rcx*8+rdi],rbx
- mov rbx,r8
- adcx r15,rsi
- inc rcx
- jnz NEAR $L$sqrx8x_tail
- cmp rbp,QWORD[((0+8))+rsp]
- jae NEAR $L$sqrx8x_tail_done
- sub rsi,QWORD[((16+8))+rsp]
- mov rdx,QWORD[((48+8))+rsp]
- lea rbp,[64+rbp]
- adc r8,QWORD[rdi]
- adc r9,QWORD[8+rdi]
- adc r10,QWORD[16+rdi]
- adc r11,QWORD[24+rdi]
- adc r12,QWORD[32+rdi]
- adc r13,QWORD[40+rdi]
- adc r14,QWORD[48+rdi]
- adc r15,QWORD[56+rdi]
- lea rdi,[64+rdi]
- sbb rax,rax
- sub rcx,8
- xor rsi,rsi
- mov QWORD[((16+8))+rsp],rax
- jmp NEAR $L$sqrx8x_tail
- ALIGN 32
- $L$sqrx8x_tail_done:
- xor rax,rax
- add r8,QWORD[((24+8))+rsp]
- adc r9,0
- adc r10,0
- adc r11,0
- adc r12,0
- adc r13,0
- adc r14,0
- adc r15,0
- adc rax,0
- sub rsi,QWORD[((16+8))+rsp]
- $L$sqrx8x_no_tail:
- adc r8,QWORD[rdi]
- DB 102,72,15,126,217
- adc r9,QWORD[8+rdi]
- mov rsi,QWORD[56+rbp]
- DB 102,72,15,126,213
- adc r10,QWORD[16+rdi]
- adc r11,QWORD[24+rdi]
- adc r12,QWORD[32+rdi]
- adc r13,QWORD[40+rdi]
- adc r14,QWORD[48+rdi]
- adc r15,QWORD[56+rdi]
- adc rax,0
- mov rbx,QWORD[((32+8))+rsp]
- mov rdx,QWORD[64+rcx*1+rdi]
- mov QWORD[rdi],r8
- lea r8,[64+rdi]
- mov QWORD[8+rdi],r9
- mov QWORD[16+rdi],r10
- mov QWORD[24+rdi],r11
- mov QWORD[32+rdi],r12
- mov QWORD[40+rdi],r13
- mov QWORD[48+rdi],r14
- mov QWORD[56+rdi],r15
- lea rdi,[64+rcx*1+rdi]
- cmp r8,QWORD[((8+8))+rsp]
- jb NEAR $L$sqrx8x_reduction_loop
- DB 0F3h,0C3h ;repret
- ALIGN 32
- __bn_postx4x_internal:
- mov r12,QWORD[rbp]
- mov r10,rcx
- mov r9,rcx
- neg rax
- sar rcx,3+2
- DB 102,72,15,126,202
- DB 102,72,15,126,206
- dec r12
- mov r13,QWORD[8+rbp]
- xor r8,r8
- mov r14,QWORD[16+rbp]
- mov r15,QWORD[24+rbp]
- jmp NEAR $L$sqrx4x_sub_entry
- ALIGN 16
- $L$sqrx4x_sub:
- mov r12,QWORD[rbp]
- mov r13,QWORD[8+rbp]
- mov r14,QWORD[16+rbp]
- mov r15,QWORD[24+rbp]
- $L$sqrx4x_sub_entry:
- andn r12,r12,rax
- lea rbp,[32+rbp]
- andn r13,r13,rax
- andn r14,r14,rax
- andn r15,r15,rax
- neg r8
- adc r12,QWORD[rdi]
- adc r13,QWORD[8+rdi]
- adc r14,QWORD[16+rdi]
- adc r15,QWORD[24+rdi]
- mov QWORD[rdx],r12
- lea rdi,[32+rdi]
- mov QWORD[8+rdx],r13
- sbb r8,r8
- mov QWORD[16+rdx],r14
- mov QWORD[24+rdx],r15
- lea rdx,[32+rdx]
- inc rcx
- jnz NEAR $L$sqrx4x_sub
- neg r9
- DB 0F3h,0C3h ;repret
- global GFp_bn_scatter5
- ALIGN 16
- GFp_bn_scatter5:
- cmp edx,0
- jz NEAR $L$scatter_epilogue
- lea r8,[r9*8+r8]
- $L$scatter:
- mov rax,QWORD[rcx]
- lea rcx,[8+rcx]
- mov QWORD[r8],rax
- lea r8,[256+r8]
- sub edx,1
- jnz NEAR $L$scatter
- $L$scatter_epilogue:
- DB 0F3h,0C3h ;repret
- global GFp_bn_gather5
- ALIGN 32
- GFp_bn_gather5:
- $L$SEH_begin_GFp_bn_gather5:
- DB 0x4c,0x8d,0x14,0x24
- DB 0x48,0x81,0xec,0x08,0x01,0x00,0x00
- lea rax,[$L$inc]
- and rsp,-16
- movd xmm5,r9d
- movdqa xmm0,XMMWORD[rax]
- movdqa xmm1,XMMWORD[16+rax]
- lea r11,[128+r8]
- lea rax,[128+rsp]
- pshufd xmm5,xmm5,0
- movdqa xmm4,xmm1
- movdqa xmm2,xmm1
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[(-128)+rax],xmm0
- movdqa xmm0,xmm4
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[(-112)+rax],xmm1
- movdqa xmm1,xmm4
- paddd xmm0,xmm3
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[(-96)+rax],xmm2
- movdqa xmm2,xmm4
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- movdqa XMMWORD[(-80)+rax],xmm3
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[(-64)+rax],xmm0
- movdqa xmm0,xmm4
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[(-48)+rax],xmm1
- movdqa xmm1,xmm4
- paddd xmm0,xmm3
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[(-32)+rax],xmm2
- movdqa xmm2,xmm4
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- movdqa XMMWORD[(-16)+rax],xmm3
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[rax],xmm0
- movdqa xmm0,xmm4
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[16+rax],xmm1
- movdqa xmm1,xmm4
- paddd xmm0,xmm3
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[32+rax],xmm2
- movdqa xmm2,xmm4
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm5
- movdqa XMMWORD[48+rax],xmm3
- movdqa xmm3,xmm4
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm5
- movdqa XMMWORD[64+rax],xmm0
- movdqa xmm0,xmm4
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm5
- movdqa XMMWORD[80+rax],xmm1
- movdqa xmm1,xmm4
- paddd xmm0,xmm3
- pcmpeqd xmm3,xmm5
- movdqa XMMWORD[96+rax],xmm2
- movdqa xmm2,xmm4
- movdqa XMMWORD[112+rax],xmm3
- jmp NEAR $L$gather
- ALIGN 32
- $L$gather:
- pxor xmm4,xmm4
- pxor xmm5,xmm5
- movdqa xmm0,XMMWORD[((-128))+r11]
- movdqa xmm1,XMMWORD[((-112))+r11]
- movdqa xmm2,XMMWORD[((-96))+r11]
- pand xmm0,XMMWORD[((-128))+rax]
- movdqa xmm3,XMMWORD[((-80))+r11]
- pand xmm1,XMMWORD[((-112))+rax]
- por xmm4,xmm0
- pand xmm2,XMMWORD[((-96))+rax]
- por xmm5,xmm1
- pand xmm3,XMMWORD[((-80))+rax]
- por xmm4,xmm2
- por xmm5,xmm3
- movdqa xmm0,XMMWORD[((-64))+r11]
- movdqa xmm1,XMMWORD[((-48))+r11]
- movdqa xmm2,XMMWORD[((-32))+r11]
- pand xmm0,XMMWORD[((-64))+rax]
- movdqa xmm3,XMMWORD[((-16))+r11]
- pand xmm1,XMMWORD[((-48))+rax]
- por xmm4,xmm0
- pand xmm2,XMMWORD[((-32))+rax]
- por xmm5,xmm1
- pand xmm3,XMMWORD[((-16))+rax]
- por xmm4,xmm2
- por xmm5,xmm3
- movdqa xmm0,XMMWORD[r11]
- movdqa xmm1,XMMWORD[16+r11]
- movdqa xmm2,XMMWORD[32+r11]
- pand xmm0,XMMWORD[rax]
- movdqa xmm3,XMMWORD[48+r11]
- pand xmm1,XMMWORD[16+rax]
- por xmm4,xmm0
- pand xmm2,XMMWORD[32+rax]
- por xmm5,xmm1
- pand xmm3,XMMWORD[48+rax]
- por xmm4,xmm2
- por xmm5,xmm3
- movdqa xmm0,XMMWORD[64+r11]
- movdqa xmm1,XMMWORD[80+r11]
- movdqa xmm2,XMMWORD[96+r11]
- pand xmm0,XMMWORD[64+rax]
- movdqa xmm3,XMMWORD[112+r11]
- pand xmm1,XMMWORD[80+rax]
- por xmm4,xmm0
- pand xmm2,XMMWORD[96+rax]
- por xmm5,xmm1
- pand xmm3,XMMWORD[112+rax]
- por xmm4,xmm2
- por xmm5,xmm3
- por xmm4,xmm5
- lea r11,[256+r11]
- pshufd xmm0,xmm4,0x4e
- por xmm0,xmm4
- movq QWORD[rcx],xmm0
- lea rcx,[8+rcx]
- sub edx,1
- jnz NEAR $L$gather
- lea rsp,[r10]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_GFp_bn_gather5:
- ALIGN 64
- $L$inc:
- DD 0,0,1,1
- DD 2,2,2,2
- DB 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
- DB 112,108,105,99,97,116,105,111,110,32,119,105,116,104,32,115
- DB 99,97,116,116,101,114,47,103,97,116,104,101,114,32,102,111
- DB 114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79
- DB 71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111
- DB 112,101,110,115,115,108,46,111,114,103,62,0
- EXTERN __imp_RtlVirtualUnwind
- ALIGN 16
- mul_handler:
- push rsi
- push rdi
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- pushfq
- sub rsp,64
- mov rax,QWORD[120+r8]
- mov rbx,QWORD[248+r8]
- mov rsi,QWORD[8+r9]
- mov r11,QWORD[56+r9]
- mov r10d,DWORD[r11]
- lea r10,[r10*1+rsi]
- cmp rbx,r10
- jb NEAR $L$common_seh_tail
- mov r10d,DWORD[4+r11]
- lea r10,[r10*1+rsi]
- cmp rbx,r10
- jb NEAR $L$common_pop_regs
- mov rax,QWORD[152+r8]
- mov r10d,DWORD[8+r11]
- lea r10,[r10*1+rsi]
- cmp rbx,r10
- jae NEAR $L$common_seh_tail
- lea r10,[$L$mul_epilogue]
- cmp rbx,r10
- ja NEAR $L$body_40
- mov r10,QWORD[192+r8]
- mov rax,QWORD[8+r10*8+rax]
- jmp NEAR $L$common_pop_regs
- $L$body_40:
- mov rax,QWORD[40+rax]
- $L$common_pop_regs:
- 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
- $L$common_seh_tail:
- mov rdi,QWORD[8+rax]
- mov rsi,QWORD[16+rax]
- mov QWORD[152+r8],rax
- mov QWORD[168+r8],rsi
- mov QWORD[176+r8],rdi
- mov rdi,QWORD[40+r9]
- mov rsi,r8
- mov ecx,154
- DD 0xa548f3fc
- mov rsi,r9
- xor rcx,rcx
- mov rdx,QWORD[8+rsi]
- mov r8,QWORD[rsi]
- mov r9,QWORD[16+rsi]
- mov r10,QWORD[40+rsi]
- lea r11,[56+rsi]
- lea r12,[24+rsi]
- mov QWORD[32+rsp],r10
- mov QWORD[40+rsp],r11
- mov QWORD[48+rsp],r12
- mov QWORD[56+rsp],rcx
- call QWORD[__imp_RtlVirtualUnwind]
- mov eax,1
- add rsp,64
- popfq
- pop r15
- pop r14
- pop r13
- pop r12
- pop rbp
- pop rbx
- pop rdi
- pop rsi
- DB 0F3h,0C3h ;repret
- section .pdata rdata align=4
- ALIGN 4
- DD $L$SEH_begin_GFp_bn_mul_mont_gather5 wrt ..imagebase
- DD $L$SEH_end_GFp_bn_mul_mont_gather5 wrt ..imagebase
- DD $L$SEH_info_GFp_bn_mul_mont_gather5 wrt ..imagebase
- DD $L$SEH_begin_bn_mul4x_mont_gather5 wrt ..imagebase
- DD $L$SEH_end_bn_mul4x_mont_gather5 wrt ..imagebase
- DD $L$SEH_info_bn_mul4x_mont_gather5 wrt ..imagebase
- DD $L$SEH_begin_GFp_bn_power5 wrt ..imagebase
- DD $L$SEH_end_GFp_bn_power5 wrt ..imagebase
- DD $L$SEH_info_GFp_bn_power5 wrt ..imagebase
- DD $L$SEH_begin_bn_from_mont8x wrt ..imagebase
- DD $L$SEH_end_bn_from_mont8x wrt ..imagebase
- DD $L$SEH_info_bn_from_mont8x wrt ..imagebase
- DD $L$SEH_begin_bn_mulx4x_mont_gather5 wrt ..imagebase
- DD $L$SEH_end_bn_mulx4x_mont_gather5 wrt ..imagebase
- DD $L$SEH_info_bn_mulx4x_mont_gather5 wrt ..imagebase
- DD $L$SEH_begin_bn_powerx5 wrt ..imagebase
- DD $L$SEH_end_bn_powerx5 wrt ..imagebase
- DD $L$SEH_info_GFp_bn_powerx5 wrt ..imagebase
- DD $L$SEH_begin_GFp_bn_gather5 wrt ..imagebase
- DD $L$SEH_end_GFp_bn_gather5 wrt ..imagebase
- DD $L$SEH_info_GFp_bn_gather5 wrt ..imagebase
- section .xdata rdata align=8
- ALIGN 8
- $L$SEH_info_GFp_bn_mul_mont_gather5:
- DB 9,0,0,0
- DD mul_handler wrt ..imagebase
- DD $L$mul_body wrt ..imagebase,$L$mul_body wrt ..imagebase,$L$mul_epilogue wrt ..imagebase
- ALIGN 8
- $L$SEH_info_bn_mul4x_mont_gather5:
- DB 9,0,0,0
- DD mul_handler wrt ..imagebase
- DD $L$mul4x_prologue wrt ..imagebase,$L$mul4x_body wrt ..imagebase,$L$mul4x_epilogue wrt ..imagebase
- ALIGN 8
- $L$SEH_info_GFp_bn_power5:
- DB 9,0,0,0
- DD mul_handler wrt ..imagebase
- DD $L$power5_prologue wrt ..imagebase,$L$power5_body wrt ..imagebase,$L$power5_epilogue wrt ..imagebase
- ALIGN 8
- $L$SEH_info_bn_from_mont8x:
- DB 9,0,0,0
- DD mul_handler wrt ..imagebase
- DD $L$from_prologue wrt ..imagebase,$L$from_body wrt ..imagebase,$L$from_epilogue wrt ..imagebase
- ALIGN 8
- $L$SEH_info_bn_mulx4x_mont_gather5:
- DB 9,0,0,0
- DD mul_handler wrt ..imagebase
- DD $L$mulx4x_prologue wrt ..imagebase,$L$mulx4x_body wrt ..imagebase,$L$mulx4x_epilogue wrt ..imagebase
- ALIGN 8
- $L$SEH_info_GFp_bn_powerx5:
- DB 9,0,0,0
- DD mul_handler wrt ..imagebase
- DD $L$powerx5_prologue wrt ..imagebase,$L$powerx5_body wrt ..imagebase,$L$powerx5_epilogue wrt ..imagebase
- ALIGN 8
- $L$SEH_info_GFp_bn_gather5:
- DB 0x01,0x0b,0x03,0x0a
- DB 0x0b,0x01,0x21,0x00
- DB 0x04,0xa3,0x00,0x00
- ALIGN 8
|