12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253125412551256125712581259126012611262126312641265126612671268126912701271127212731274127512761277127812791280128112821283128412851286128712881289129012911292129312941295129612971298129913001301130213031304130513061307130813091310131113121313131413151316131713181319132013211322132313241325132613271328132913301331133213331334133513361337133813391340134113421343134413451346134713481349135013511352135313541355135613571358135913601361136213631364136513661367136813691370137113721373137413751376137713781379138013811382138313841385138613871388138913901391139213931394139513961397139813991400140114021403140414051406140714081409141014111412141314141415141614171418141914201421142214231424142514261427142814291430143114321433143414351436143714381439144014411442144314441445144614471448144914501451145214531454145514561457145814591460146114621463146414651466146714681469147014711472147314741475147614771478147914801481148214831484148514861487148814891490149114921493149414951496149714981499150015011502150315041505150615071508150915101511151215131514151515161517151815191520152115221523152415251526152715281529153015311532153315341535153615371538153915401541154215431544154515461547154815491550155115521553155415551556155715581559156015611562156315641565156615671568156915701571157215731574157515761577157815791580158115821583158415851586158715881589159015911592159315941595159615971598159916001601160216031604160516061607160816091610161116121613161416151616161716181619162016211622162316241625162616271628162916301631163216331634163516361637163816391640164116421643164416451646164716481649165016511652165316541655165616571658165916601661166216631664166516661667166816691670167116721673167416751676167716781679168016811682168316841685168616871688168916901691169216931694169516961697169816991700170117021703170417051706170717081709171017111712171317141715171617171718171917201721172217231724172517261727172817291730173117321733173417351736173717381739174017411742174317441745174617471748174917501751175217531754175517561757175817591760176117621763176417651766176717681769177017711772177317741775177617771778177917801781178217831784178517861787178817891790179117921793179417951796179717981799180018011802180318041805180618071808180918101811181218131814181518161817181818191820182118221823182418251826182718281829183018311832183318341835183618371838183918401841184218431844184518461847184818491850185118521853185418551856185718581859186018611862186318641865186618671868186918701871187218731874187518761877187818791880188118821883188418851886188718881889189018911892189318941895189618971898189919001901190219031904190519061907190819091910191119121913191419151916191719181919192019211922192319241925192619271928192919301931193219331934193519361937193819391940194119421943194419451946194719481949195019511952195319541955195619571958195919601961196219631964196519661967196819691970197119721973197419751976197719781979198019811982198319841985198619871988198919901991199219931994199519961997199819992000200120022003200420052006200720082009201020112012201320142015201620172018201920202021202220232024202520262027202820292030203120322033203420352036203720382039204020412042204320442045204620472048204920502051205220532054205520562057205820592060206120622063206420652066206720682069207020712072207320742075207620772078207920802081208220832084208520862087208820892090209120922093209420952096209720982099210021012102210321042105210621072108210921102111211221132114211521162117211821192120212121222123212421252126212721282129213021312132213321342135213621372138213921402141214221432144214521462147214821492150215121522153215421552156215721582159216021612162216321642165216621672168216921702171217221732174217521762177217821792180218121822183218421852186218721882189219021912192219321942195219621972198219922002201220222032204220522062207220822092210221122122213221422152216221722182219222022212222222322242225222622272228222922302231223222332234223522362237223822392240224122422243224422452246224722482249225022512252225322542255225622572258225922602261226222632264226522662267 |
- OPTION DOTNAME
- .text$ SEGMENT ALIGN(256) 'CODE'
- EXTERN OPENSSL_ia32cap_P:NEAR
- PUBLIC rsaz_512_sqr
- ALIGN 32
- rsaz_512_sqr PROC PUBLIC
- mov QWORD PTR[8+rsp],rdi ;WIN64 prologue
- mov QWORD PTR[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_rsaz_512_sqr::
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- mov r8,QWORD PTR[40+rsp]
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- sub rsp,128+24
- $L$sqr_body::
- DB 102,72,15,110,202
- mov rdx,QWORD PTR[rsi]
- mov rax,QWORD PTR[8+rsi]
- mov QWORD PTR[128+rsp],rcx
- mov r11d,080100h
- and r11d,DWORD PTR[((OPENSSL_ia32cap_P+8))]
- cmp r11d,080100h
- je $L$oop_sqrx
- jmp $L$oop_sqr
- ALIGN 32
- $L$oop_sqr::
- mov DWORD PTR[((128+8))+rsp],r8d
- mov rbx,rdx
- mov rbp,rax
- mul rdx
- mov r8,rax
- mov rax,QWORD PTR[16+rsi]
- mov r9,rdx
- mul rbx
- add r9,rax
- mov rax,QWORD PTR[24+rsi]
- mov r10,rdx
- adc r10,0
- mul rbx
- add r10,rax
- mov rax,QWORD PTR[32+rsi]
- mov r11,rdx
- adc r11,0
- mul rbx
- add r11,rax
- mov rax,QWORD PTR[40+rsi]
- mov r12,rdx
- adc r12,0
- mul rbx
- add r12,rax
- mov rax,QWORD PTR[48+rsi]
- mov r13,rdx
- adc r13,0
- mul rbx
- add r13,rax
- mov rax,QWORD PTR[56+rsi]
- mov r14,rdx
- adc r14,0
- mul rbx
- add r14,rax
- mov rax,rbx
- adc rdx,0
- xor rcx,rcx
- add r8,r8
- mov r15,rdx
- adc rcx,0
- mul rax
- add rdx,r8
- adc rcx,0
- mov QWORD PTR[rsp],rax
- mov QWORD PTR[8+rsp],rdx
- mov rax,QWORD PTR[16+rsi]
- mul rbp
- add r10,rax
- mov rax,QWORD PTR[24+rsi]
- mov rbx,rdx
- adc rbx,0
- mul rbp
- add r11,rax
- mov rax,QWORD PTR[32+rsi]
- adc rdx,0
- add r11,rbx
- mov rbx,rdx
- adc rbx,0
- mul rbp
- add r12,rax
- mov rax,QWORD PTR[40+rsi]
- adc rdx,0
- add r12,rbx
- mov rbx,rdx
- adc rbx,0
- mul rbp
- add r13,rax
- mov rax,QWORD PTR[48+rsi]
- adc rdx,0
- add r13,rbx
- mov rbx,rdx
- adc rbx,0
- mul rbp
- add r14,rax
- mov rax,QWORD PTR[56+rsi]
- adc rdx,0
- add r14,rbx
- mov rbx,rdx
- adc rbx,0
- mul rbp
- add r15,rax
- mov rax,rbp
- adc rdx,0
- add r15,rbx
- adc rdx,0
- xor rbx,rbx
- add r9,r9
- mov r8,rdx
- adc r10,r10
- adc rbx,0
- mul rax
- add rax,rcx
- mov rbp,QWORD PTR[16+rsi]
- add r9,rax
- mov rax,QWORD PTR[24+rsi]
- adc r10,rdx
- adc rbx,0
- mov QWORD PTR[16+rsp],r9
- mov QWORD PTR[24+rsp],r10
- mul rbp
- add r12,rax
- mov rax,QWORD PTR[32+rsi]
- mov rcx,rdx
- adc rcx,0
- mul rbp
- add r13,rax
- mov rax,QWORD PTR[40+rsi]
- adc rdx,0
- add r13,rcx
- mov rcx,rdx
- adc rcx,0
- mul rbp
- add r14,rax
- mov rax,QWORD PTR[48+rsi]
- adc rdx,0
- add r14,rcx
- mov rcx,rdx
- adc rcx,0
- mul rbp
- add r15,rax
- mov rax,QWORD PTR[56+rsi]
- adc rdx,0
- add r15,rcx
- mov rcx,rdx
- adc rcx,0
- mul rbp
- add r8,rax
- mov rax,rbp
- adc rdx,0
- add r8,rcx
- adc rdx,0
- xor rcx,rcx
- add r11,r11
- mov r9,rdx
- adc r12,r12
- adc rcx,0
- mul rax
- add rax,rbx
- mov r10,QWORD PTR[24+rsi]
- add r11,rax
- mov rax,QWORD PTR[32+rsi]
- adc r12,rdx
- adc rcx,0
- mov QWORD PTR[32+rsp],r11
- mov QWORD PTR[40+rsp],r12
- mov r11,rax
- mul r10
- add r14,rax
- mov rax,QWORD PTR[40+rsi]
- mov rbx,rdx
- adc rbx,0
- mov r12,rax
- mul r10
- add r15,rax
- mov rax,QWORD PTR[48+rsi]
- adc rdx,0
- add r15,rbx
- mov rbx,rdx
- adc rbx,0
- mov rbp,rax
- mul r10
- add r8,rax
- mov rax,QWORD PTR[56+rsi]
- adc rdx,0
- add r8,rbx
- mov rbx,rdx
- adc rbx,0
- mul r10
- add r9,rax
- mov rax,r10
- adc rdx,0
- add r9,rbx
- adc rdx,0
- xor rbx,rbx
- add r13,r13
- mov r10,rdx
- adc r14,r14
- adc rbx,0
- mul rax
- add rax,rcx
- add r13,rax
- mov rax,r12
- adc r14,rdx
- adc rbx,0
- mov QWORD PTR[48+rsp],r13
- mov QWORD PTR[56+rsp],r14
- mul r11
- add r8,rax
- mov rax,rbp
- mov rcx,rdx
- adc rcx,0
- mul r11
- add r9,rax
- mov rax,QWORD PTR[56+rsi]
- adc rdx,0
- add r9,rcx
- mov rcx,rdx
- adc rcx,0
- mov r14,rax
- mul r11
- add r10,rax
- mov rax,r11
- adc rdx,0
- add r10,rcx
- adc rdx,0
- xor rcx,rcx
- add r15,r15
- mov r11,rdx
- adc r8,r8
- adc rcx,0
- mul rax
- add rax,rbx
- add r15,rax
- mov rax,rbp
- adc r8,rdx
- adc rcx,0
- mov QWORD PTR[64+rsp],r15
- mov QWORD PTR[72+rsp],r8
- mul r12
- add r10,rax
- mov rax,r14
- mov rbx,rdx
- adc rbx,0
- mul r12
- add r11,rax
- mov rax,r12
- adc rdx,0
- add r11,rbx
- adc rdx,0
- xor rbx,rbx
- add r9,r9
- mov r12,rdx
- adc r10,r10
- adc rbx,0
- mul rax
- add rax,rcx
- add r9,rax
- mov rax,r14
- adc r10,rdx
- adc rbx,0
- mov QWORD PTR[80+rsp],r9
- mov QWORD PTR[88+rsp],r10
- mul rbp
- add r12,rax
- mov rax,rbp
- adc rdx,0
- xor rcx,rcx
- add r11,r11
- mov r13,rdx
- adc r12,r12
- adc rcx,0
- mul rax
- add rax,rbx
- add r11,rax
- mov rax,r14
- adc r12,rdx
- adc rcx,0
- mov QWORD PTR[96+rsp],r11
- mov QWORD PTR[104+rsp],r12
- xor rbx,rbx
- add r13,r13
- adc rbx,0
- mul rax
- add rax,rcx
- add rax,r13
- adc rdx,rbx
- mov r8,QWORD PTR[rsp]
- mov r9,QWORD PTR[8+rsp]
- mov r10,QWORD PTR[16+rsp]
- mov r11,QWORD PTR[24+rsp]
- mov r12,QWORD PTR[32+rsp]
- mov r13,QWORD PTR[40+rsp]
- mov r14,QWORD PTR[48+rsp]
- mov r15,QWORD PTR[56+rsp]
- DB 102,72,15,126,205
- mov QWORD PTR[112+rsp],rax
- mov QWORD PTR[120+rsp],rdx
- call __rsaz_512_reduce
- add r8,QWORD PTR[64+rsp]
- adc r9,QWORD PTR[72+rsp]
- adc r10,QWORD PTR[80+rsp]
- adc r11,QWORD PTR[88+rsp]
- adc r12,QWORD PTR[96+rsp]
- adc r13,QWORD PTR[104+rsp]
- adc r14,QWORD PTR[112+rsp]
- adc r15,QWORD PTR[120+rsp]
- sbb rcx,rcx
- call __rsaz_512_subtract
- mov rdx,r8
- mov rax,r9
- mov r8d,DWORD PTR[((128+8))+rsp]
- mov rsi,rdi
- dec r8d
- jnz $L$oop_sqr
- jmp $L$sqr_tail
- ALIGN 32
- $L$oop_sqrx::
- mov DWORD PTR[((128+8))+rsp],r8d
- DB 102,72,15,110,199
- mulx r9,r8,rax
- mov rbx,rax
- mulx r10,rcx,QWORD PTR[16+rsi]
- xor rbp,rbp
- mulx r11,rax,QWORD PTR[24+rsi]
- adcx r9,rcx
- DB 0c4h,062h,0f3h,0f6h,0a6h,020h,000h,000h,000h
- adcx r10,rax
- DB 0c4h,062h,0fbh,0f6h,0aeh,028h,000h,000h,000h
- adcx r11,rcx
- mulx r14,rcx,QWORD PTR[48+rsi]
- adcx r12,rax
- adcx r13,rcx
- mulx r15,rax,QWORD PTR[56+rsi]
- adcx r14,rax
- adcx r15,rbp
- mulx rdi,rax,rdx
- mov rdx,rbx
- xor rcx,rcx
- adox r8,r8
- adcx r8,rdi
- adox rcx,rbp
- adcx rcx,rbp
- mov QWORD PTR[rsp],rax
- mov QWORD PTR[8+rsp],r8
- DB 0c4h,0e2h,0fbh,0f6h,09eh,010h,000h,000h,000h
- adox r10,rax
- adcx r11,rbx
- mulx r8,rdi,QWORD PTR[24+rsi]
- adox r11,rdi
- DB 066h
- adcx r12,r8
- mulx rbx,rax,QWORD PTR[32+rsi]
- adox r12,rax
- adcx r13,rbx
- mulx r8,rdi,QWORD PTR[40+rsi]
- adox r13,rdi
- adcx r14,r8
- DB 0c4h,0e2h,0fbh,0f6h,09eh,030h,000h,000h,000h
- adox r14,rax
- adcx r15,rbx
- DB 0c4h,062h,0c3h,0f6h,086h,038h,000h,000h,000h
- adox r15,rdi
- adcx r8,rbp
- mulx rdi,rax,rdx
- adox r8,rbp
- DB 048h,08bh,096h,010h,000h,000h,000h
- xor rbx,rbx
- adox r9,r9
- adcx rax,rcx
- adox r10,r10
- adcx r9,rax
- adox rbx,rbp
- adcx r10,rdi
- adcx rbx,rbp
- mov QWORD PTR[16+rsp],r9
- DB 04ch,089h,094h,024h,018h,000h,000h,000h
- mulx r9,rdi,QWORD PTR[24+rsi]
- adox r12,rdi
- adcx r13,r9
- mulx rcx,rax,QWORD PTR[32+rsi]
- adox r13,rax
- adcx r14,rcx
- DB 0c4h,062h,0c3h,0f6h,08eh,028h,000h,000h,000h
- adox r14,rdi
- adcx r15,r9
- DB 0c4h,0e2h,0fbh,0f6h,08eh,030h,000h,000h,000h
- adox r15,rax
- adcx r8,rcx
- mulx r9,rdi,QWORD PTR[56+rsi]
- adox r8,rdi
- adcx r9,rbp
- mulx rdi,rax,rdx
- adox r9,rbp
- mov rdx,QWORD PTR[24+rsi]
- xor rcx,rcx
- adox r11,r11
- adcx rax,rbx
- adox r12,r12
- adcx r11,rax
- adox rcx,rbp
- adcx r12,rdi
- adcx rcx,rbp
- mov QWORD PTR[32+rsp],r11
- mov QWORD PTR[40+rsp],r12
- mulx rbx,rax,QWORD PTR[32+rsi]
- adox r14,rax
- adcx r15,rbx
- mulx r10,rdi,QWORD PTR[40+rsi]
- adox r15,rdi
- adcx r8,r10
- mulx rbx,rax,QWORD PTR[48+rsi]
- adox r8,rax
- adcx r9,rbx
- mulx r10,rdi,QWORD PTR[56+rsi]
- adox r9,rdi
- adcx r10,rbp
- mulx rdi,rax,rdx
- adox r10,rbp
- mov rdx,QWORD PTR[32+rsi]
- xor rbx,rbx
- adox r13,r13
- adcx rax,rcx
- adox r14,r14
- adcx r13,rax
- adox rbx,rbp
- adcx r14,rdi
- adcx rbx,rbp
- mov QWORD PTR[48+rsp],r13
- mov QWORD PTR[56+rsp],r14
- mulx r11,rdi,QWORD PTR[40+rsi]
- adox r8,rdi
- adcx r9,r11
- mulx rcx,rax,QWORD PTR[48+rsi]
- adox r9,rax
- adcx r10,rcx
- mulx r11,rdi,QWORD PTR[56+rsi]
- adox r10,rdi
- adcx r11,rbp
- mulx rdi,rax,rdx
- mov rdx,QWORD PTR[40+rsi]
- adox r11,rbp
- xor rcx,rcx
- adox r15,r15
- adcx rax,rbx
- adox r8,r8
- adcx r15,rax
- adox rcx,rbp
- adcx r8,rdi
- adcx rcx,rbp
- mov QWORD PTR[64+rsp],r15
- mov QWORD PTR[72+rsp],r8
- DB 0c4h,0e2h,0fbh,0f6h,09eh,030h,000h,000h,000h
- adox r10,rax
- adcx r11,rbx
- DB 0c4h,062h,0c3h,0f6h,0a6h,038h,000h,000h,000h
- adox r11,rdi
- adcx r12,rbp
- mulx rdi,rax,rdx
- adox r12,rbp
- mov rdx,QWORD PTR[48+rsi]
- xor rbx,rbx
- adox r9,r9
- adcx rax,rcx
- adox r10,r10
- adcx r9,rax
- adcx r10,rdi
- adox rbx,rbp
- adcx rbx,rbp
- mov QWORD PTR[80+rsp],r9
- mov QWORD PTR[88+rsp],r10
- DB 0c4h,062h,0fbh,0f6h,0aeh,038h,000h,000h,000h
- adox r12,rax
- adox r13,rbp
- mulx rdi,rax,rdx
- xor rcx,rcx
- mov rdx,QWORD PTR[56+rsi]
- adox r11,r11
- adcx rax,rbx
- adox r12,r12
- adcx r11,rax
- adox rcx,rbp
- adcx r12,rdi
- adcx rcx,rbp
- DB 04ch,089h,09ch,024h,060h,000h,000h,000h
- DB 04ch,089h,0a4h,024h,068h,000h,000h,000h
- mulx rdx,rax,rdx
- xor rbx,rbx
- adox r13,r13
- adcx rax,rcx
- adox rbx,rbp
- adcx rax,r13
- adcx rbx,rdx
- DB 102,72,15,126,199
- DB 102,72,15,126,205
- mov rdx,QWORD PTR[128+rsp]
- mov r8,QWORD PTR[rsp]
- mov r9,QWORD PTR[8+rsp]
- mov r10,QWORD PTR[16+rsp]
- mov r11,QWORD PTR[24+rsp]
- mov r12,QWORD PTR[32+rsp]
- mov r13,QWORD PTR[40+rsp]
- mov r14,QWORD PTR[48+rsp]
- mov r15,QWORD PTR[56+rsp]
- mov QWORD PTR[112+rsp],rax
- mov QWORD PTR[120+rsp],rbx
- call __rsaz_512_reducex
- add r8,QWORD PTR[64+rsp]
- adc r9,QWORD PTR[72+rsp]
- adc r10,QWORD PTR[80+rsp]
- adc r11,QWORD PTR[88+rsp]
- adc r12,QWORD PTR[96+rsp]
- adc r13,QWORD PTR[104+rsp]
- adc r14,QWORD PTR[112+rsp]
- adc r15,QWORD PTR[120+rsp]
- sbb rcx,rcx
- call __rsaz_512_subtract
- mov rdx,r8
- mov rax,r9
- mov r8d,DWORD PTR[((128+8))+rsp]
- mov rsi,rdi
- dec r8d
- jnz $L$oop_sqrx
- $L$sqr_tail::
- lea rax,QWORD PTR[((128+24+48))+rsp]
- mov r15,QWORD PTR[((-48))+rax]
- mov r14,QWORD PTR[((-40))+rax]
- mov r13,QWORD PTR[((-32))+rax]
- mov r12,QWORD PTR[((-24))+rax]
- mov rbp,QWORD PTR[((-16))+rax]
- mov rbx,QWORD PTR[((-8))+rax]
- lea rsp,QWORD PTR[rax]
- $L$sqr_epilogue::
- mov rdi,QWORD PTR[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD PTR[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_rsaz_512_sqr::
- rsaz_512_sqr ENDP
- PUBLIC rsaz_512_mul
- ALIGN 32
- rsaz_512_mul PROC PUBLIC
- mov QWORD PTR[8+rsp],rdi ;WIN64 prologue
- mov QWORD PTR[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_rsaz_512_mul::
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- mov r8,QWORD PTR[40+rsp]
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- sub rsp,128+24
- $L$mul_body::
- DB 102,72,15,110,199
- DB 102,72,15,110,201
- mov QWORD PTR[128+rsp],r8
- mov r11d,080100h
- and r11d,DWORD PTR[((OPENSSL_ia32cap_P+8))]
- cmp r11d,080100h
- je $L$mulx
- mov rbx,QWORD PTR[rdx]
- mov rbp,rdx
- call __rsaz_512_mul
- DB 102,72,15,126,199
- DB 102,72,15,126,205
- mov r8,QWORD PTR[rsp]
- mov r9,QWORD PTR[8+rsp]
- mov r10,QWORD PTR[16+rsp]
- mov r11,QWORD PTR[24+rsp]
- mov r12,QWORD PTR[32+rsp]
- mov r13,QWORD PTR[40+rsp]
- mov r14,QWORD PTR[48+rsp]
- mov r15,QWORD PTR[56+rsp]
- call __rsaz_512_reduce
- jmp $L$mul_tail
- ALIGN 32
- $L$mulx::
- mov rbp,rdx
- mov rdx,QWORD PTR[rdx]
- call __rsaz_512_mulx
- DB 102,72,15,126,199
- DB 102,72,15,126,205
- mov rdx,QWORD PTR[128+rsp]
- mov r8,QWORD PTR[rsp]
- mov r9,QWORD PTR[8+rsp]
- mov r10,QWORD PTR[16+rsp]
- mov r11,QWORD PTR[24+rsp]
- mov r12,QWORD PTR[32+rsp]
- mov r13,QWORD PTR[40+rsp]
- mov r14,QWORD PTR[48+rsp]
- mov r15,QWORD PTR[56+rsp]
- call __rsaz_512_reducex
- $L$mul_tail::
- add r8,QWORD PTR[64+rsp]
- adc r9,QWORD PTR[72+rsp]
- adc r10,QWORD PTR[80+rsp]
- adc r11,QWORD PTR[88+rsp]
- adc r12,QWORD PTR[96+rsp]
- adc r13,QWORD PTR[104+rsp]
- adc r14,QWORD PTR[112+rsp]
- adc r15,QWORD PTR[120+rsp]
- sbb rcx,rcx
- call __rsaz_512_subtract
- lea rax,QWORD PTR[((128+24+48))+rsp]
- mov r15,QWORD PTR[((-48))+rax]
- mov r14,QWORD PTR[((-40))+rax]
- mov r13,QWORD PTR[((-32))+rax]
- mov r12,QWORD PTR[((-24))+rax]
- mov rbp,QWORD PTR[((-16))+rax]
- mov rbx,QWORD PTR[((-8))+rax]
- lea rsp,QWORD PTR[rax]
- $L$mul_epilogue::
- mov rdi,QWORD PTR[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD PTR[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_rsaz_512_mul::
- rsaz_512_mul ENDP
- PUBLIC rsaz_512_mul_gather4
- ALIGN 32
- rsaz_512_mul_gather4 PROC PUBLIC
- mov QWORD PTR[8+rsp],rdi ;WIN64 prologue
- mov QWORD PTR[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_rsaz_512_mul_gather4::
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- mov r8,QWORD PTR[40+rsp]
- mov r9,QWORD PTR[48+rsp]
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- sub rsp,328
- movaps XMMWORD PTR[160+rsp],xmm6
- movaps XMMWORD PTR[176+rsp],xmm7
- movaps XMMWORD PTR[192+rsp],xmm8
- movaps XMMWORD PTR[208+rsp],xmm9
- movaps XMMWORD PTR[224+rsp],xmm10
- movaps XMMWORD PTR[240+rsp],xmm11
- movaps XMMWORD PTR[256+rsp],xmm12
- movaps XMMWORD PTR[272+rsp],xmm13
- movaps XMMWORD PTR[288+rsp],xmm14
- movaps XMMWORD PTR[304+rsp],xmm15
- $L$mul_gather4_body::
- movd xmm8,r9d
- movdqa xmm1,XMMWORD PTR[(($L$inc+16))]
- movdqa xmm0,XMMWORD PTR[$L$inc]
- pshufd xmm8,xmm8,0
- movdqa xmm7,xmm1
- movdqa xmm2,xmm1
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm8
- movdqa xmm3,xmm7
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm8
- movdqa xmm4,xmm7
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm8
- movdqa xmm5,xmm7
- paddd xmm4,xmm3
- pcmpeqd xmm3,xmm8
- movdqa xmm6,xmm7
- paddd xmm5,xmm4
- pcmpeqd xmm4,xmm8
- paddd xmm6,xmm5
- pcmpeqd xmm5,xmm8
- paddd xmm7,xmm6
- pcmpeqd xmm6,xmm8
- pcmpeqd xmm7,xmm8
- movdqa xmm8,XMMWORD PTR[rdx]
- movdqa xmm9,XMMWORD PTR[16+rdx]
- movdqa xmm10,XMMWORD PTR[32+rdx]
- movdqa xmm11,XMMWORD PTR[48+rdx]
- pand xmm8,xmm0
- movdqa xmm12,XMMWORD PTR[64+rdx]
- pand xmm9,xmm1
- movdqa xmm13,XMMWORD PTR[80+rdx]
- pand xmm10,xmm2
- movdqa xmm14,XMMWORD PTR[96+rdx]
- pand xmm11,xmm3
- movdqa xmm15,XMMWORD PTR[112+rdx]
- lea rbp,QWORD PTR[128+rdx]
- pand xmm12,xmm4
- pand xmm13,xmm5
- pand xmm14,xmm6
- pand xmm15,xmm7
- por xmm8,xmm10
- por xmm9,xmm11
- por xmm8,xmm12
- por xmm9,xmm13
- por xmm8,xmm14
- por xmm9,xmm15
- por xmm8,xmm9
- pshufd xmm9,xmm8,04eh
- por xmm8,xmm9
- mov r11d,080100h
- and r11d,DWORD PTR[((OPENSSL_ia32cap_P+8))]
- cmp r11d,080100h
- je $L$mulx_gather
- DB 102,76,15,126,195
- mov QWORD PTR[128+rsp],r8
- mov QWORD PTR[((128+8))+rsp],rdi
- mov QWORD PTR[((128+16))+rsp],rcx
- mov rax,QWORD PTR[rsi]
- mov rcx,QWORD PTR[8+rsi]
- mul rbx
- mov QWORD PTR[rsp],rax
- mov rax,rcx
- mov r8,rdx
- mul rbx
- add r8,rax
- mov rax,QWORD PTR[16+rsi]
- mov r9,rdx
- adc r9,0
- mul rbx
- add r9,rax
- mov rax,QWORD PTR[24+rsi]
- mov r10,rdx
- adc r10,0
- mul rbx
- add r10,rax
- mov rax,QWORD PTR[32+rsi]
- mov r11,rdx
- adc r11,0
- mul rbx
- add r11,rax
- mov rax,QWORD PTR[40+rsi]
- mov r12,rdx
- adc r12,0
- mul rbx
- add r12,rax
- mov rax,QWORD PTR[48+rsi]
- mov r13,rdx
- adc r13,0
- mul rbx
- add r13,rax
- mov rax,QWORD PTR[56+rsi]
- mov r14,rdx
- adc r14,0
- mul rbx
- add r14,rax
- mov rax,QWORD PTR[rsi]
- mov r15,rdx
- adc r15,0
- lea rdi,QWORD PTR[8+rsp]
- mov ecx,7
- jmp $L$oop_mul_gather
- ALIGN 32
- $L$oop_mul_gather::
- movdqa xmm8,XMMWORD PTR[rbp]
- movdqa xmm9,XMMWORD PTR[16+rbp]
- movdqa xmm10,XMMWORD PTR[32+rbp]
- movdqa xmm11,XMMWORD PTR[48+rbp]
- pand xmm8,xmm0
- movdqa xmm12,XMMWORD PTR[64+rbp]
- pand xmm9,xmm1
- movdqa xmm13,XMMWORD PTR[80+rbp]
- pand xmm10,xmm2
- movdqa xmm14,XMMWORD PTR[96+rbp]
- pand xmm11,xmm3
- movdqa xmm15,XMMWORD PTR[112+rbp]
- lea rbp,QWORD PTR[128+rbp]
- pand xmm12,xmm4
- pand xmm13,xmm5
- pand xmm14,xmm6
- pand xmm15,xmm7
- por xmm8,xmm10
- por xmm9,xmm11
- por xmm8,xmm12
- por xmm9,xmm13
- por xmm8,xmm14
- por xmm9,xmm15
- por xmm8,xmm9
- pshufd xmm9,xmm8,04eh
- por xmm8,xmm9
- DB 102,76,15,126,195
- mul rbx
- add r8,rax
- mov rax,QWORD PTR[8+rsi]
- mov QWORD PTR[rdi],r8
- mov r8,rdx
- adc r8,0
- mul rbx
- add r9,rax
- mov rax,QWORD PTR[16+rsi]
- adc rdx,0
- add r8,r9
- mov r9,rdx
- adc r9,0
- mul rbx
- add r10,rax
- mov rax,QWORD PTR[24+rsi]
- adc rdx,0
- add r9,r10
- mov r10,rdx
- adc r10,0
- mul rbx
- add r11,rax
- mov rax,QWORD PTR[32+rsi]
- adc rdx,0
- add r10,r11
- mov r11,rdx
- adc r11,0
- mul rbx
- add r12,rax
- mov rax,QWORD PTR[40+rsi]
- adc rdx,0
- add r11,r12
- mov r12,rdx
- adc r12,0
- mul rbx
- add r13,rax
- mov rax,QWORD PTR[48+rsi]
- adc rdx,0
- add r12,r13
- mov r13,rdx
- adc r13,0
- mul rbx
- add r14,rax
- mov rax,QWORD PTR[56+rsi]
- adc rdx,0
- add r13,r14
- mov r14,rdx
- adc r14,0
- mul rbx
- add r15,rax
- mov rax,QWORD PTR[rsi]
- adc rdx,0
- add r14,r15
- mov r15,rdx
- adc r15,0
- lea rdi,QWORD PTR[8+rdi]
- dec ecx
- jnz $L$oop_mul_gather
- mov QWORD PTR[rdi],r8
- mov QWORD PTR[8+rdi],r9
- mov QWORD PTR[16+rdi],r10
- mov QWORD PTR[24+rdi],r11
- mov QWORD PTR[32+rdi],r12
- mov QWORD PTR[40+rdi],r13
- mov QWORD PTR[48+rdi],r14
- mov QWORD PTR[56+rdi],r15
- mov rdi,QWORD PTR[((128+8))+rsp]
- mov rbp,QWORD PTR[((128+16))+rsp]
- mov r8,QWORD PTR[rsp]
- mov r9,QWORD PTR[8+rsp]
- mov r10,QWORD PTR[16+rsp]
- mov r11,QWORD PTR[24+rsp]
- mov r12,QWORD PTR[32+rsp]
- mov r13,QWORD PTR[40+rsp]
- mov r14,QWORD PTR[48+rsp]
- mov r15,QWORD PTR[56+rsp]
- call __rsaz_512_reduce
- jmp $L$mul_gather_tail
- ALIGN 32
- $L$mulx_gather::
- DB 102,76,15,126,194
- mov QWORD PTR[128+rsp],r8
- mov QWORD PTR[((128+8))+rsp],rdi
- mov QWORD PTR[((128+16))+rsp],rcx
- mulx r8,rbx,QWORD PTR[rsi]
- mov QWORD PTR[rsp],rbx
- xor edi,edi
- mulx r9,rax,QWORD PTR[8+rsi]
- mulx r10,rbx,QWORD PTR[16+rsi]
- adcx r8,rax
- mulx r11,rax,QWORD PTR[24+rsi]
- adcx r9,rbx
- mulx r12,rbx,QWORD PTR[32+rsi]
- adcx r10,rax
- mulx r13,rax,QWORD PTR[40+rsi]
- adcx r11,rbx
- mulx r14,rbx,QWORD PTR[48+rsi]
- adcx r12,rax
- mulx r15,rax,QWORD PTR[56+rsi]
- adcx r13,rbx
- adcx r14,rax
- DB 067h
- mov rbx,r8
- adcx r15,rdi
- mov rcx,-7
- jmp $L$oop_mulx_gather
- ALIGN 32
- $L$oop_mulx_gather::
- movdqa xmm8,XMMWORD PTR[rbp]
- movdqa xmm9,XMMWORD PTR[16+rbp]
- movdqa xmm10,XMMWORD PTR[32+rbp]
- movdqa xmm11,XMMWORD PTR[48+rbp]
- pand xmm8,xmm0
- movdqa xmm12,XMMWORD PTR[64+rbp]
- pand xmm9,xmm1
- movdqa xmm13,XMMWORD PTR[80+rbp]
- pand xmm10,xmm2
- movdqa xmm14,XMMWORD PTR[96+rbp]
- pand xmm11,xmm3
- movdqa xmm15,XMMWORD PTR[112+rbp]
- lea rbp,QWORD PTR[128+rbp]
- pand xmm12,xmm4
- pand xmm13,xmm5
- pand xmm14,xmm6
- pand xmm15,xmm7
- por xmm8,xmm10
- por xmm9,xmm11
- por xmm8,xmm12
- por xmm9,xmm13
- por xmm8,xmm14
- por xmm9,xmm15
- por xmm8,xmm9
- pshufd xmm9,xmm8,04eh
- por xmm8,xmm9
- DB 102,76,15,126,194
- DB 0c4h,062h,0fbh,0f6h,086h,000h,000h,000h,000h
- adcx rbx,rax
- adox r8,r9
- mulx r9,rax,QWORD PTR[8+rsi]
- adcx r8,rax
- adox r9,r10
- mulx r10,rax,QWORD PTR[16+rsi]
- adcx r9,rax
- adox r10,r11
- DB 0c4h,062h,0fbh,0f6h,09eh,018h,000h,000h,000h
- adcx r10,rax
- adox r11,r12
- mulx r12,rax,QWORD PTR[32+rsi]
- adcx r11,rax
- adox r12,r13
- mulx r13,rax,QWORD PTR[40+rsi]
- adcx r12,rax
- adox r13,r14
- DB 0c4h,062h,0fbh,0f6h,0b6h,030h,000h,000h,000h
- adcx r13,rax
- DB 067h
- adox r14,r15
- mulx r15,rax,QWORD PTR[56+rsi]
- mov QWORD PTR[64+rcx*8+rsp],rbx
- adcx r14,rax
- adox r15,rdi
- mov rbx,r8
- adcx r15,rdi
- inc rcx
- jnz $L$oop_mulx_gather
- mov QWORD PTR[64+rsp],r8
- mov QWORD PTR[((64+8))+rsp],r9
- mov QWORD PTR[((64+16))+rsp],r10
- mov QWORD PTR[((64+24))+rsp],r11
- mov QWORD PTR[((64+32))+rsp],r12
- mov QWORD PTR[((64+40))+rsp],r13
- mov QWORD PTR[((64+48))+rsp],r14
- mov QWORD PTR[((64+56))+rsp],r15
- mov rdx,QWORD PTR[128+rsp]
- mov rdi,QWORD PTR[((128+8))+rsp]
- mov rbp,QWORD PTR[((128+16))+rsp]
- mov r8,QWORD PTR[rsp]
- mov r9,QWORD PTR[8+rsp]
- mov r10,QWORD PTR[16+rsp]
- mov r11,QWORD PTR[24+rsp]
- mov r12,QWORD PTR[32+rsp]
- mov r13,QWORD PTR[40+rsp]
- mov r14,QWORD PTR[48+rsp]
- mov r15,QWORD PTR[56+rsp]
- call __rsaz_512_reducex
- $L$mul_gather_tail::
- add r8,QWORD PTR[64+rsp]
- adc r9,QWORD PTR[72+rsp]
- adc r10,QWORD PTR[80+rsp]
- adc r11,QWORD PTR[88+rsp]
- adc r12,QWORD PTR[96+rsp]
- adc r13,QWORD PTR[104+rsp]
- adc r14,QWORD PTR[112+rsp]
- adc r15,QWORD PTR[120+rsp]
- sbb rcx,rcx
- call __rsaz_512_subtract
- lea rax,QWORD PTR[((128+24+48))+rsp]
- movaps xmm6,XMMWORD PTR[((160-200))+rax]
- movaps xmm7,XMMWORD PTR[((176-200))+rax]
- movaps xmm8,XMMWORD PTR[((192-200))+rax]
- movaps xmm9,XMMWORD PTR[((208-200))+rax]
- movaps xmm10,XMMWORD PTR[((224-200))+rax]
- movaps xmm11,XMMWORD PTR[((240-200))+rax]
- movaps xmm12,XMMWORD PTR[((256-200))+rax]
- movaps xmm13,XMMWORD PTR[((272-200))+rax]
- movaps xmm14,XMMWORD PTR[((288-200))+rax]
- movaps xmm15,XMMWORD PTR[((304-200))+rax]
- lea rax,QWORD PTR[176+rax]
- mov r15,QWORD PTR[((-48))+rax]
- mov r14,QWORD PTR[((-40))+rax]
- mov r13,QWORD PTR[((-32))+rax]
- mov r12,QWORD PTR[((-24))+rax]
- mov rbp,QWORD PTR[((-16))+rax]
- mov rbx,QWORD PTR[((-8))+rax]
- lea rsp,QWORD PTR[rax]
- $L$mul_gather4_epilogue::
- mov rdi,QWORD PTR[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD PTR[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_rsaz_512_mul_gather4::
- rsaz_512_mul_gather4 ENDP
- PUBLIC rsaz_512_mul_scatter4
- ALIGN 32
- rsaz_512_mul_scatter4 PROC PUBLIC
- mov QWORD PTR[8+rsp],rdi ;WIN64 prologue
- mov QWORD PTR[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_rsaz_512_mul_scatter4::
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- mov r8,QWORD PTR[40+rsp]
- mov r9,QWORD PTR[48+rsp]
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- mov r9d,r9d
- sub rsp,128+24
- $L$mul_scatter4_body::
- lea r8,QWORD PTR[r9*8+r8]
- DB 102,72,15,110,199
- DB 102,72,15,110,202
- DB 102,73,15,110,208
- mov QWORD PTR[128+rsp],rcx
- mov rbp,rdi
- mov r11d,080100h
- and r11d,DWORD PTR[((OPENSSL_ia32cap_P+8))]
- cmp r11d,080100h
- je $L$mulx_scatter
- mov rbx,QWORD PTR[rdi]
- call __rsaz_512_mul
- DB 102,72,15,126,199
- DB 102,72,15,126,205
- mov r8,QWORD PTR[rsp]
- mov r9,QWORD PTR[8+rsp]
- mov r10,QWORD PTR[16+rsp]
- mov r11,QWORD PTR[24+rsp]
- mov r12,QWORD PTR[32+rsp]
- mov r13,QWORD PTR[40+rsp]
- mov r14,QWORD PTR[48+rsp]
- mov r15,QWORD PTR[56+rsp]
- call __rsaz_512_reduce
- jmp $L$mul_scatter_tail
- ALIGN 32
- $L$mulx_scatter::
- mov rdx,QWORD PTR[rdi]
- call __rsaz_512_mulx
- DB 102,72,15,126,199
- DB 102,72,15,126,205
- mov rdx,QWORD PTR[128+rsp]
- mov r8,QWORD PTR[rsp]
- mov r9,QWORD PTR[8+rsp]
- mov r10,QWORD PTR[16+rsp]
- mov r11,QWORD PTR[24+rsp]
- mov r12,QWORD PTR[32+rsp]
- mov r13,QWORD PTR[40+rsp]
- mov r14,QWORD PTR[48+rsp]
- mov r15,QWORD PTR[56+rsp]
- call __rsaz_512_reducex
- $L$mul_scatter_tail::
- add r8,QWORD PTR[64+rsp]
- adc r9,QWORD PTR[72+rsp]
- adc r10,QWORD PTR[80+rsp]
- adc r11,QWORD PTR[88+rsp]
- adc r12,QWORD PTR[96+rsp]
- adc r13,QWORD PTR[104+rsp]
- adc r14,QWORD PTR[112+rsp]
- adc r15,QWORD PTR[120+rsp]
- DB 102,72,15,126,214
- sbb rcx,rcx
- call __rsaz_512_subtract
- mov QWORD PTR[rsi],r8
- mov QWORD PTR[128+rsi],r9
- mov QWORD PTR[256+rsi],r10
- mov QWORD PTR[384+rsi],r11
- mov QWORD PTR[512+rsi],r12
- mov QWORD PTR[640+rsi],r13
- mov QWORD PTR[768+rsi],r14
- mov QWORD PTR[896+rsi],r15
- lea rax,QWORD PTR[((128+24+48))+rsp]
- mov r15,QWORD PTR[((-48))+rax]
- mov r14,QWORD PTR[((-40))+rax]
- mov r13,QWORD PTR[((-32))+rax]
- mov r12,QWORD PTR[((-24))+rax]
- mov rbp,QWORD PTR[((-16))+rax]
- mov rbx,QWORD PTR[((-8))+rax]
- lea rsp,QWORD PTR[rax]
- $L$mul_scatter4_epilogue::
- mov rdi,QWORD PTR[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD PTR[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_rsaz_512_mul_scatter4::
- rsaz_512_mul_scatter4 ENDP
- PUBLIC rsaz_512_mul_by_one
- ALIGN 32
- rsaz_512_mul_by_one PROC PUBLIC
- mov QWORD PTR[8+rsp],rdi ;WIN64 prologue
- mov QWORD PTR[16+rsp],rsi
- mov rax,rsp
- $L$SEH_begin_rsaz_512_mul_by_one::
- mov rdi,rcx
- mov rsi,rdx
- mov rdx,r8
- mov rcx,r9
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- sub rsp,128+24
- $L$mul_by_one_body::
- mov eax,DWORD PTR[((OPENSSL_ia32cap_P+8))]
- mov rbp,rdx
- mov QWORD PTR[128+rsp],rcx
- mov r8,QWORD PTR[rsi]
- pxor xmm0,xmm0
- mov r9,QWORD PTR[8+rsi]
- mov r10,QWORD PTR[16+rsi]
- mov r11,QWORD PTR[24+rsi]
- mov r12,QWORD PTR[32+rsi]
- mov r13,QWORD PTR[40+rsi]
- mov r14,QWORD PTR[48+rsi]
- mov r15,QWORD PTR[56+rsi]
- movdqa XMMWORD PTR[rsp],xmm0
- movdqa XMMWORD PTR[16+rsp],xmm0
- movdqa XMMWORD PTR[32+rsp],xmm0
- movdqa XMMWORD PTR[48+rsp],xmm0
- movdqa XMMWORD PTR[64+rsp],xmm0
- movdqa XMMWORD PTR[80+rsp],xmm0
- movdqa XMMWORD PTR[96+rsp],xmm0
- and eax,080100h
- cmp eax,080100h
- je $L$by_one_callx
- call __rsaz_512_reduce
- jmp $L$by_one_tail
- ALIGN 32
- $L$by_one_callx::
- mov rdx,QWORD PTR[128+rsp]
- call __rsaz_512_reducex
- $L$by_one_tail::
- mov QWORD PTR[rdi],r8
- mov QWORD PTR[8+rdi],r9
- mov QWORD PTR[16+rdi],r10
- mov QWORD PTR[24+rdi],r11
- mov QWORD PTR[32+rdi],r12
- mov QWORD PTR[40+rdi],r13
- mov QWORD PTR[48+rdi],r14
- mov QWORD PTR[56+rdi],r15
- lea rax,QWORD PTR[((128+24+48))+rsp]
- mov r15,QWORD PTR[((-48))+rax]
- mov r14,QWORD PTR[((-40))+rax]
- mov r13,QWORD PTR[((-32))+rax]
- mov r12,QWORD PTR[((-24))+rax]
- mov rbp,QWORD PTR[((-16))+rax]
- mov rbx,QWORD PTR[((-8))+rax]
- lea rsp,QWORD PTR[rax]
- $L$mul_by_one_epilogue::
- mov rdi,QWORD PTR[8+rsp] ;WIN64 epilogue
- mov rsi,QWORD PTR[16+rsp]
- DB 0F3h,0C3h ;repret
- $L$SEH_end_rsaz_512_mul_by_one::
- rsaz_512_mul_by_one ENDP
- ALIGN 32
- __rsaz_512_reduce PROC PRIVATE
- mov rbx,r8
- imul rbx,QWORD PTR[((128+8))+rsp]
- mov rax,QWORD PTR[rbp]
- mov ecx,8
- jmp $L$reduction_loop
- ALIGN 32
- $L$reduction_loop::
- mul rbx
- mov rax,QWORD PTR[8+rbp]
- neg r8
- mov r8,rdx
- adc r8,0
- mul rbx
- add r9,rax
- mov rax,QWORD PTR[16+rbp]
- adc rdx,0
- add r8,r9
- mov r9,rdx
- adc r9,0
- mul rbx
- add r10,rax
- mov rax,QWORD PTR[24+rbp]
- adc rdx,0
- add r9,r10
- mov r10,rdx
- adc r10,0
- mul rbx
- add r11,rax
- mov rax,QWORD PTR[32+rbp]
- adc rdx,0
- add r10,r11
- mov rsi,QWORD PTR[((128+8))+rsp]
- adc rdx,0
- mov r11,rdx
- mul rbx
- add r12,rax
- mov rax,QWORD PTR[40+rbp]
- adc rdx,0
- imul rsi,r8
- add r11,r12
- mov r12,rdx
- adc r12,0
- mul rbx
- add r13,rax
- mov rax,QWORD PTR[48+rbp]
- adc rdx,0
- add r12,r13
- mov r13,rdx
- adc r13,0
- mul rbx
- add r14,rax
- mov rax,QWORD PTR[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 PTR[rbp]
- adc rdx,0
- add r14,r15
- mov r15,rdx
- adc r15,0
- dec ecx
- jne $L$reduction_loop
- DB 0F3h,0C3h ;repret
- __rsaz_512_reduce ENDP
- ALIGN 32
- __rsaz_512_reducex PROC PRIVATE
- imul rdx,r8
- xor rsi,rsi
- mov ecx,8
- jmp $L$reduction_loopx
- ALIGN 32
- $L$reduction_loopx::
- mov rbx,r8
- mulx r8,rax,QWORD PTR[rbp]
- adcx rax,rbx
- adox r8,r9
- mulx r9,rax,QWORD PTR[8+rbp]
- adcx r8,rax
- adox r9,r10
- mulx r10,rbx,QWORD PTR[16+rbp]
- adcx r9,rbx
- adox r10,r11
- mulx r11,rbx,QWORD PTR[24+rbp]
- adcx r10,rbx
- adox r11,r12
- DB 0c4h,062h,0e3h,0f6h,0a5h,020h,000h,000h,000h
- mov rax,rdx
- mov rdx,r8
- adcx r11,rbx
- adox r12,r13
- mulx rdx,rbx,QWORD PTR[((128+8))+rsp]
- mov rdx,rax
- mulx r13,rax,QWORD PTR[40+rbp]
- adcx r12,rax
- adox r13,r14
- DB 0c4h,062h,0fbh,0f6h,0b5h,030h,000h,000h,000h
- adcx r13,rax
- adox r14,r15
- mulx r15,rax,QWORD PTR[56+rbp]
- mov rdx,rbx
- adcx r14,rax
- adox r15,rsi
- adcx r15,rsi
- dec ecx
- jne $L$reduction_loopx
- DB 0F3h,0C3h ;repret
- __rsaz_512_reducex ENDP
- ALIGN 32
- __rsaz_512_subtract PROC PRIVATE
- mov QWORD PTR[rdi],r8
- mov QWORD PTR[8+rdi],r9
- mov QWORD PTR[16+rdi],r10
- mov QWORD PTR[24+rdi],r11
- mov QWORD PTR[32+rdi],r12
- mov QWORD PTR[40+rdi],r13
- mov QWORD PTR[48+rdi],r14
- mov QWORD PTR[56+rdi],r15
- mov r8,QWORD PTR[rbp]
- mov r9,QWORD PTR[8+rbp]
- neg r8
- not r9
- and r8,rcx
- mov r10,QWORD PTR[16+rbp]
- and r9,rcx
- not r10
- mov r11,QWORD PTR[24+rbp]
- and r10,rcx
- not r11
- mov r12,QWORD PTR[32+rbp]
- and r11,rcx
- not r12
- mov r13,QWORD PTR[40+rbp]
- and r12,rcx
- not r13
- mov r14,QWORD PTR[48+rbp]
- and r13,rcx
- not r14
- mov r15,QWORD PTR[56+rbp]
- and r14,rcx
- not r15
- and r15,rcx
- add r8,QWORD PTR[rdi]
- adc r9,QWORD PTR[8+rdi]
- adc r10,QWORD PTR[16+rdi]
- adc r11,QWORD PTR[24+rdi]
- adc r12,QWORD PTR[32+rdi]
- adc r13,QWORD PTR[40+rdi]
- adc r14,QWORD PTR[48+rdi]
- adc r15,QWORD PTR[56+rdi]
- mov QWORD PTR[rdi],r8
- mov QWORD PTR[8+rdi],r9
- mov QWORD PTR[16+rdi],r10
- mov QWORD PTR[24+rdi],r11
- mov QWORD PTR[32+rdi],r12
- mov QWORD PTR[40+rdi],r13
- mov QWORD PTR[48+rdi],r14
- mov QWORD PTR[56+rdi],r15
- DB 0F3h,0C3h ;repret
- __rsaz_512_subtract ENDP
- ALIGN 32
- __rsaz_512_mul PROC PRIVATE
- lea rdi,QWORD PTR[8+rsp]
- mov rax,QWORD PTR[rsi]
- mul rbx
- mov QWORD PTR[rdi],rax
- mov rax,QWORD PTR[8+rsi]
- mov r8,rdx
- mul rbx
- add r8,rax
- mov rax,QWORD PTR[16+rsi]
- mov r9,rdx
- adc r9,0
- mul rbx
- add r9,rax
- mov rax,QWORD PTR[24+rsi]
- mov r10,rdx
- adc r10,0
- mul rbx
- add r10,rax
- mov rax,QWORD PTR[32+rsi]
- mov r11,rdx
- adc r11,0
- mul rbx
- add r11,rax
- mov rax,QWORD PTR[40+rsi]
- mov r12,rdx
- adc r12,0
- mul rbx
- add r12,rax
- mov rax,QWORD PTR[48+rsi]
- mov r13,rdx
- adc r13,0
- mul rbx
- add r13,rax
- mov rax,QWORD PTR[56+rsi]
- mov r14,rdx
- adc r14,0
- mul rbx
- add r14,rax
- mov rax,QWORD PTR[rsi]
- mov r15,rdx
- adc r15,0
- lea rbp,QWORD PTR[8+rbp]
- lea rdi,QWORD PTR[8+rdi]
- mov ecx,7
- jmp $L$oop_mul
- ALIGN 32
- $L$oop_mul::
- mov rbx,QWORD PTR[rbp]
- mul rbx
- add r8,rax
- mov rax,QWORD PTR[8+rsi]
- mov QWORD PTR[rdi],r8
- mov r8,rdx
- adc r8,0
- mul rbx
- add r9,rax
- mov rax,QWORD PTR[16+rsi]
- adc rdx,0
- add r8,r9
- mov r9,rdx
- adc r9,0
- mul rbx
- add r10,rax
- mov rax,QWORD PTR[24+rsi]
- adc rdx,0
- add r9,r10
- mov r10,rdx
- adc r10,0
- mul rbx
- add r11,rax
- mov rax,QWORD PTR[32+rsi]
- adc rdx,0
- add r10,r11
- mov r11,rdx
- adc r11,0
- mul rbx
- add r12,rax
- mov rax,QWORD PTR[40+rsi]
- adc rdx,0
- add r11,r12
- mov r12,rdx
- adc r12,0
- mul rbx
- add r13,rax
- mov rax,QWORD PTR[48+rsi]
- adc rdx,0
- add r12,r13
- mov r13,rdx
- adc r13,0
- mul rbx
- add r14,rax
- mov rax,QWORD PTR[56+rsi]
- adc rdx,0
- add r13,r14
- mov r14,rdx
- lea rbp,QWORD PTR[8+rbp]
- adc r14,0
- mul rbx
- add r15,rax
- mov rax,QWORD PTR[rsi]
- adc rdx,0
- add r14,r15
- mov r15,rdx
- adc r15,0
- lea rdi,QWORD PTR[8+rdi]
- dec ecx
- jnz $L$oop_mul
- mov QWORD PTR[rdi],r8
- mov QWORD PTR[8+rdi],r9
- mov QWORD PTR[16+rdi],r10
- mov QWORD PTR[24+rdi],r11
- mov QWORD PTR[32+rdi],r12
- mov QWORD PTR[40+rdi],r13
- mov QWORD PTR[48+rdi],r14
- mov QWORD PTR[56+rdi],r15
- DB 0F3h,0C3h ;repret
- __rsaz_512_mul ENDP
- ALIGN 32
- __rsaz_512_mulx PROC PRIVATE
- mulx r8,rbx,QWORD PTR[rsi]
- mov rcx,-6
- mulx r9,rax,QWORD PTR[8+rsi]
- mov QWORD PTR[8+rsp],rbx
- mulx r10,rbx,QWORD PTR[16+rsi]
- adc r8,rax
- mulx r11,rax,QWORD PTR[24+rsi]
- adc r9,rbx
- mulx r12,rbx,QWORD PTR[32+rsi]
- adc r10,rax
- mulx r13,rax,QWORD PTR[40+rsi]
- adc r11,rbx
- mulx r14,rbx,QWORD PTR[48+rsi]
- adc r12,rax
- mulx r15,rax,QWORD PTR[56+rsi]
- mov rdx,QWORD PTR[8+rbp]
- adc r13,rbx
- adc r14,rax
- adc r15,0
- xor rdi,rdi
- jmp $L$oop_mulx
- ALIGN 32
- $L$oop_mulx::
- mov rbx,r8
- mulx r8,rax,QWORD PTR[rsi]
- adcx rbx,rax
- adox r8,r9
- mulx r9,rax,QWORD PTR[8+rsi]
- adcx r8,rax
- adox r9,r10
- mulx r10,rax,QWORD PTR[16+rsi]
- adcx r9,rax
- adox r10,r11
- mulx r11,rax,QWORD PTR[24+rsi]
- adcx r10,rax
- adox r11,r12
- DB 03eh,0c4h,062h,0fbh,0f6h,0a6h,020h,000h,000h,000h
- adcx r11,rax
- adox r12,r13
- mulx r13,rax,QWORD PTR[40+rsi]
- adcx r12,rax
- adox r13,r14
- mulx r14,rax,QWORD PTR[48+rsi]
- adcx r13,rax
- adox r14,r15
- mulx r15,rax,QWORD PTR[56+rsi]
- mov rdx,QWORD PTR[64+rcx*8+rbp]
- mov QWORD PTR[((8+64-8))+rcx*8+rsp],rbx
- adcx r14,rax
- adox r15,rdi
- adcx r15,rdi
- inc rcx
- jnz $L$oop_mulx
- mov rbx,r8
- mulx r8,rax,QWORD PTR[rsi]
- adcx rbx,rax
- adox r8,r9
- DB 0c4h,062h,0fbh,0f6h,08eh,008h,000h,000h,000h
- adcx r8,rax
- adox r9,r10
- DB 0c4h,062h,0fbh,0f6h,096h,010h,000h,000h,000h
- adcx r9,rax
- adox r10,r11
- mulx r11,rax,QWORD PTR[24+rsi]
- adcx r10,rax
- adox r11,r12
- mulx r12,rax,QWORD PTR[32+rsi]
- adcx r11,rax
- adox r12,r13
- mulx r13,rax,QWORD PTR[40+rsi]
- adcx r12,rax
- adox r13,r14
- DB 0c4h,062h,0fbh,0f6h,0b6h,030h,000h,000h,000h
- adcx r13,rax
- adox r14,r15
- DB 0c4h,062h,0fbh,0f6h,0beh,038h,000h,000h,000h
- adcx r14,rax
- adox r15,rdi
- adcx r15,rdi
- mov QWORD PTR[((8+64-8))+rsp],rbx
- mov QWORD PTR[((8+64))+rsp],r8
- mov QWORD PTR[((8+64+8))+rsp],r9
- mov QWORD PTR[((8+64+16))+rsp],r10
- mov QWORD PTR[((8+64+24))+rsp],r11
- mov QWORD PTR[((8+64+32))+rsp],r12
- mov QWORD PTR[((8+64+40))+rsp],r13
- mov QWORD PTR[((8+64+48))+rsp],r14
- mov QWORD PTR[((8+64+56))+rsp],r15
- DB 0F3h,0C3h ;repret
- __rsaz_512_mulx ENDP
- PUBLIC rsaz_512_scatter4
- ALIGN 16
- rsaz_512_scatter4 PROC PUBLIC
- lea rcx,QWORD PTR[r8*8+rcx]
- mov r9d,8
- jmp $L$oop_scatter
- ALIGN 16
- $L$oop_scatter::
- mov rax,QWORD PTR[rdx]
- lea rdx,QWORD PTR[8+rdx]
- mov QWORD PTR[rcx],rax
- lea rcx,QWORD PTR[128+rcx]
- dec r9d
- jnz $L$oop_scatter
- DB 0F3h,0C3h ;repret
- rsaz_512_scatter4 ENDP
- PUBLIC rsaz_512_gather4
- ALIGN 16
- rsaz_512_gather4 PROC PUBLIC
- $L$SEH_begin_rsaz_512_gather4::
- DB 048h,081h,0ech,0a8h,000h,000h,000h
- DB 00fh,029h,034h,024h
- DB 00fh,029h,07ch,024h,010h
- DB 044h,00fh,029h,044h,024h,020h
- DB 044h,00fh,029h,04ch,024h,030h
- DB 044h,00fh,029h,054h,024h,040h
- DB 044h,00fh,029h,05ch,024h,050h
- DB 044h,00fh,029h,064h,024h,060h
- DB 044h,00fh,029h,06ch,024h,070h
- DB 044h,00fh,029h,0b4h,024h,080h,0,0,0
- DB 044h,00fh,029h,0bch,024h,090h,0,0,0
- movd xmm8,r8d
- movdqa xmm1,XMMWORD PTR[(($L$inc+16))]
- movdqa xmm0,XMMWORD PTR[$L$inc]
- pshufd xmm8,xmm8,0
- movdqa xmm7,xmm1
- movdqa xmm2,xmm1
- paddd xmm1,xmm0
- pcmpeqd xmm0,xmm8
- movdqa xmm3,xmm7
- paddd xmm2,xmm1
- pcmpeqd xmm1,xmm8
- movdqa xmm4,xmm7
- paddd xmm3,xmm2
- pcmpeqd xmm2,xmm8
- movdqa xmm5,xmm7
- paddd xmm4,xmm3
- pcmpeqd xmm3,xmm8
- movdqa xmm6,xmm7
- paddd xmm5,xmm4
- pcmpeqd xmm4,xmm8
- paddd xmm6,xmm5
- pcmpeqd xmm5,xmm8
- paddd xmm7,xmm6
- pcmpeqd xmm6,xmm8
- pcmpeqd xmm7,xmm8
- mov r9d,8
- jmp $L$oop_gather
- ALIGN 16
- $L$oop_gather::
- movdqa xmm8,XMMWORD PTR[rdx]
- movdqa xmm9,XMMWORD PTR[16+rdx]
- movdqa xmm10,XMMWORD PTR[32+rdx]
- movdqa xmm11,XMMWORD PTR[48+rdx]
- pand xmm8,xmm0
- movdqa xmm12,XMMWORD PTR[64+rdx]
- pand xmm9,xmm1
- movdqa xmm13,XMMWORD PTR[80+rdx]
- pand xmm10,xmm2
- movdqa xmm14,XMMWORD PTR[96+rdx]
- pand xmm11,xmm3
- movdqa xmm15,XMMWORD PTR[112+rdx]
- lea rdx,QWORD PTR[128+rdx]
- pand xmm12,xmm4
- pand xmm13,xmm5
- pand xmm14,xmm6
- pand xmm15,xmm7
- por xmm8,xmm10
- por xmm9,xmm11
- por xmm8,xmm12
- por xmm9,xmm13
- por xmm8,xmm14
- por xmm9,xmm15
- por xmm8,xmm9
- pshufd xmm9,xmm8,04eh
- por xmm8,xmm9
- movq QWORD PTR[rcx],xmm8
- lea rcx,QWORD PTR[8+rcx]
- dec r9d
- jnz $L$oop_gather
- movaps xmm6,XMMWORD PTR[rsp]
- movaps xmm7,XMMWORD PTR[16+rsp]
- movaps xmm8,XMMWORD PTR[32+rsp]
- movaps xmm9,XMMWORD PTR[48+rsp]
- movaps xmm10,XMMWORD PTR[64+rsp]
- movaps xmm11,XMMWORD PTR[80+rsp]
- movaps xmm12,XMMWORD PTR[96+rsp]
- movaps xmm13,XMMWORD PTR[112+rsp]
- movaps xmm14,XMMWORD PTR[128+rsp]
- movaps xmm15,XMMWORD PTR[144+rsp]
- add rsp,0a8h
- DB 0F3h,0C3h ;repret
- $L$SEH_end_rsaz_512_gather4::
- rsaz_512_gather4 ENDP
- ALIGN 64
- $L$inc::
- DD 0,0,1,1
- DD 2,2,2,2
- EXTERN __imp_RtlVirtualUnwind:NEAR
- ALIGN 16
- se_handler PROC PRIVATE
- push rsi
- push rdi
- push rbx
- push rbp
- push r12
- push r13
- push r14
- push r15
- pushfq
- sub rsp,64
- mov rax,QWORD PTR[120+r8]
- mov rbx,QWORD PTR[248+r8]
- mov rsi,QWORD PTR[8+r9]
- mov r11,QWORD PTR[56+r9]
- mov r10d,DWORD PTR[r11]
- lea r10,QWORD PTR[r10*1+rsi]
- cmp rbx,r10
- jb $L$common_seh_tail
- mov rax,QWORD PTR[152+r8]
- mov r10d,DWORD PTR[4+r11]
- lea r10,QWORD PTR[r10*1+rsi]
- cmp rbx,r10
- jae $L$common_seh_tail
- lea rax,QWORD PTR[((128+24+48))+rax]
- lea rbx,QWORD PTR[$L$mul_gather4_epilogue]
- cmp rbx,r10
- jne $L$se_not_in_mul_gather4
- lea rax,QWORD PTR[176+rax]
- lea rsi,QWORD PTR[((-48-168))+rax]
- lea rdi,QWORD PTR[512+r8]
- mov ecx,20
- DD 0a548f3fch
- $L$se_not_in_mul_gather4::
- mov rbx,QWORD PTR[((-8))+rax]
- mov rbp,QWORD PTR[((-16))+rax]
- mov r12,QWORD PTR[((-24))+rax]
- mov r13,QWORD PTR[((-32))+rax]
- mov r14,QWORD PTR[((-40))+rax]
- mov r15,QWORD PTR[((-48))+rax]
- mov QWORD PTR[144+r8],rbx
- mov QWORD PTR[160+r8],rbp
- mov QWORD PTR[216+r8],r12
- mov QWORD PTR[224+r8],r13
- mov QWORD PTR[232+r8],r14
- mov QWORD PTR[240+r8],r15
- $L$common_seh_tail::
- mov rdi,QWORD PTR[8+rax]
- mov rsi,QWORD PTR[16+rax]
- mov QWORD PTR[152+r8],rax
- mov QWORD PTR[168+r8],rsi
- mov QWORD PTR[176+r8],rdi
- mov rdi,QWORD PTR[40+r9]
- mov rsi,r8
- mov ecx,154
- DD 0a548f3fch
- mov rsi,r9
- xor rcx,rcx
- mov rdx,QWORD PTR[8+rsi]
- mov r8,QWORD PTR[rsi]
- mov r9,QWORD PTR[16+rsi]
- mov r10,QWORD PTR[40+rsi]
- lea r11,QWORD PTR[56+rsi]
- lea r12,QWORD PTR[24+rsi]
- mov QWORD PTR[32+rsp],r10
- mov QWORD PTR[40+rsp],r11
- mov QWORD PTR[48+rsp],r12
- mov QWORD PTR[56+rsp],rcx
- call QWORD PTR[__imp_RtlVirtualUnwind]
- mov eax,1
- add rsp,64
- popfq
- pop r15
- pop r14
- pop r13
- pop r12
- pop rbp
- pop rbx
- pop rdi
- pop rsi
- DB 0F3h,0C3h ;repret
- se_handler ENDP
- .text$ ENDS
- .pdata SEGMENT READONLY ALIGN(4)
- ALIGN 4
- DD imagerel $L$SEH_begin_rsaz_512_sqr
- DD imagerel $L$SEH_end_rsaz_512_sqr
- DD imagerel $L$SEH_info_rsaz_512_sqr
- DD imagerel $L$SEH_begin_rsaz_512_mul
- DD imagerel $L$SEH_end_rsaz_512_mul
- DD imagerel $L$SEH_info_rsaz_512_mul
- DD imagerel $L$SEH_begin_rsaz_512_mul_gather4
- DD imagerel $L$SEH_end_rsaz_512_mul_gather4
- DD imagerel $L$SEH_info_rsaz_512_mul_gather4
- DD imagerel $L$SEH_begin_rsaz_512_mul_scatter4
- DD imagerel $L$SEH_end_rsaz_512_mul_scatter4
- DD imagerel $L$SEH_info_rsaz_512_mul_scatter4
- DD imagerel $L$SEH_begin_rsaz_512_mul_by_one
- DD imagerel $L$SEH_end_rsaz_512_mul_by_one
- DD imagerel $L$SEH_info_rsaz_512_mul_by_one
- DD imagerel $L$SEH_begin_rsaz_512_gather4
- DD imagerel $L$SEH_end_rsaz_512_gather4
- DD imagerel $L$SEH_info_rsaz_512_gather4
- .pdata ENDS
- .xdata SEGMENT READONLY ALIGN(8)
- ALIGN 8
- $L$SEH_info_rsaz_512_sqr::
- DB 9,0,0,0
- DD imagerel se_handler
- DD imagerel $L$sqr_body,imagerel $L$sqr_epilogue
- $L$SEH_info_rsaz_512_mul::
- DB 9,0,0,0
- DD imagerel se_handler
- DD imagerel $L$mul_body,imagerel $L$mul_epilogue
- $L$SEH_info_rsaz_512_mul_gather4::
- DB 9,0,0,0
- DD imagerel se_handler
- DD imagerel $L$mul_gather4_body,imagerel $L$mul_gather4_epilogue
- $L$SEH_info_rsaz_512_mul_scatter4::
- DB 9,0,0,0
- DD imagerel se_handler
- DD imagerel $L$mul_scatter4_body,imagerel $L$mul_scatter4_epilogue
- $L$SEH_info_rsaz_512_mul_by_one::
- DB 9,0,0,0
- DD imagerel se_handler
- DD imagerel $L$mul_by_one_body,imagerel $L$mul_by_one_epilogue
- $L$SEH_info_rsaz_512_gather4::
- DB 001h,046h,016h,000h
- DB 046h,0f8h,009h,000h
- DB 03dh,0e8h,008h,000h
- DB 034h,0d8h,007h,000h
- DB 02eh,0c8h,006h,000h
- DB 028h,0b8h,005h,000h
- DB 022h,0a8h,004h,000h
- DB 01ch,098h,003h,000h
- DB 016h,088h,002h,000h
- DB 010h,078h,001h,000h
- DB 00bh,068h,000h,000h
- DB 007h,001h,015h,000h
- .xdata ENDS
- END
|