| 12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136113711381139114011411142114311441145114611471148114911501151115211531154115511561157115811591160116111621163116411651166116711681169117011711172117311741175117611771178117911801181118211831184118511861187118811891190119111921193119411951196119711981199120012011202120312041205120612071208120912101211121212131214121512161217121812191220122112221223122412251226122712281229123012311232123312341235123612371238123912401241124212431244124512461247124812491250125112521253 |
- %ifidn __OUTPUT_FORMAT__,obj
- section code use32 class=code align=256
- %elifidn __OUTPUT_FORMAT__,win32
- [email protected] equ 1
- section .text code align=256
- %else
- section .text code
- %endif
- global _bn_mul_comba8
- align 16
- _bn_mul_comba8:
- L$_bn_mul_comba8_begin:
- push esi
- mov esi,DWORD [12+esp]
- push edi
- mov edi,DWORD [20+esp]
- push ebp
- push ebx
- xor ebx,ebx
- mov eax,DWORD [esi]
- xor ecx,ecx
- mov edx,DWORD [edi]
- ; ################## Calculate word 0
- xor ebp,ebp
- ; mul a[0]*b[0]
- mul edx
- add ebx,eax
- mov eax,DWORD [20+esp]
- adc ecx,edx
- mov edx,DWORD [edi]
- adc ebp,0
- mov DWORD [eax],ebx
- mov eax,DWORD [4+esi]
- ; saved r[0]
- ; ################## Calculate word 1
- xor ebx,ebx
- ; mul a[1]*b[0]
- mul edx
- add ecx,eax
- mov eax,DWORD [esi]
- adc ebp,edx
- mov edx,DWORD [4+edi]
- adc ebx,0
- ; mul a[0]*b[1]
- mul edx
- add ecx,eax
- mov eax,DWORD [20+esp]
- adc ebp,edx
- mov edx,DWORD [edi]
- adc ebx,0
- mov DWORD [4+eax],ecx
- mov eax,DWORD [8+esi]
- ; saved r[1]
- ; ################## Calculate word 2
- xor ecx,ecx
- ; mul a[2]*b[0]
- mul edx
- add ebp,eax
- mov eax,DWORD [4+esi]
- adc ebx,edx
- mov edx,DWORD [4+edi]
- adc ecx,0
- ; mul a[1]*b[1]
- mul edx
- add ebp,eax
- mov eax,DWORD [esi]
- adc ebx,edx
- mov edx,DWORD [8+edi]
- adc ecx,0
- ; mul a[0]*b[2]
- mul edx
- add ebp,eax
- mov eax,DWORD [20+esp]
- adc ebx,edx
- mov edx,DWORD [edi]
- adc ecx,0
- mov DWORD [8+eax],ebp
- mov eax,DWORD [12+esi]
- ; saved r[2]
- ; ################## Calculate word 3
- xor ebp,ebp
- ; mul a[3]*b[0]
- mul edx
- add ebx,eax
- mov eax,DWORD [8+esi]
- adc ecx,edx
- mov edx,DWORD [4+edi]
- adc ebp,0
- ; mul a[2]*b[1]
- mul edx
- add ebx,eax
- mov eax,DWORD [4+esi]
- adc ecx,edx
- mov edx,DWORD [8+edi]
- adc ebp,0
- ; mul a[1]*b[2]
- mul edx
- add ebx,eax
- mov eax,DWORD [esi]
- adc ecx,edx
- mov edx,DWORD [12+edi]
- adc ebp,0
- ; mul a[0]*b[3]
- mul edx
- add ebx,eax
- mov eax,DWORD [20+esp]
- adc ecx,edx
- mov edx,DWORD [edi]
- adc ebp,0
- mov DWORD [12+eax],ebx
- mov eax,DWORD [16+esi]
- ; saved r[3]
- ; ################## Calculate word 4
- xor ebx,ebx
- ; mul a[4]*b[0]
- mul edx
- add ecx,eax
- mov eax,DWORD [12+esi]
- adc ebp,edx
- mov edx,DWORD [4+edi]
- adc ebx,0
- ; mul a[3]*b[1]
- mul edx
- add ecx,eax
- mov eax,DWORD [8+esi]
- adc ebp,edx
- mov edx,DWORD [8+edi]
- adc ebx,0
- ; mul a[2]*b[2]
- mul edx
- add ecx,eax
- mov eax,DWORD [4+esi]
- adc ebp,edx
- mov edx,DWORD [12+edi]
- adc ebx,0
- ; mul a[1]*b[3]
- mul edx
- add ecx,eax
- mov eax,DWORD [esi]
- adc ebp,edx
- mov edx,DWORD [16+edi]
- adc ebx,0
- ; mul a[0]*b[4]
- mul edx
- add ecx,eax
- mov eax,DWORD [20+esp]
- adc ebp,edx
- mov edx,DWORD [edi]
- adc ebx,0
- mov DWORD [16+eax],ecx
- mov eax,DWORD [20+esi]
- ; saved r[4]
- ; ################## Calculate word 5
- xor ecx,ecx
- ; mul a[5]*b[0]
- mul edx
- add ebp,eax
- mov eax,DWORD [16+esi]
- adc ebx,edx
- mov edx,DWORD [4+edi]
- adc ecx,0
- ; mul a[4]*b[1]
- mul edx
- add ebp,eax
- mov eax,DWORD [12+esi]
- adc ebx,edx
- mov edx,DWORD [8+edi]
- adc ecx,0
- ; mul a[3]*b[2]
- mul edx
- add ebp,eax
- mov eax,DWORD [8+esi]
- adc ebx,edx
- mov edx,DWORD [12+edi]
- adc ecx,0
- ; mul a[2]*b[3]
- mul edx
- add ebp,eax
- mov eax,DWORD [4+esi]
- adc ebx,edx
- mov edx,DWORD [16+edi]
- adc ecx,0
- ; mul a[1]*b[4]
- mul edx
- add ebp,eax
- mov eax,DWORD [esi]
- adc ebx,edx
- mov edx,DWORD [20+edi]
- adc ecx,0
- ; mul a[0]*b[5]
- mul edx
- add ebp,eax
- mov eax,DWORD [20+esp]
- adc ebx,edx
- mov edx,DWORD [edi]
- adc ecx,0
- mov DWORD [20+eax],ebp
- mov eax,DWORD [24+esi]
- ; saved r[5]
- ; ################## Calculate word 6
- xor ebp,ebp
- ; mul a[6]*b[0]
- mul edx
- add ebx,eax
- mov eax,DWORD [20+esi]
- adc ecx,edx
- mov edx,DWORD [4+edi]
- adc ebp,0
- ; mul a[5]*b[1]
- mul edx
- add ebx,eax
- mov eax,DWORD [16+esi]
- adc ecx,edx
- mov edx,DWORD [8+edi]
- adc ebp,0
- ; mul a[4]*b[2]
- mul edx
- add ebx,eax
- mov eax,DWORD [12+esi]
- adc ecx,edx
- mov edx,DWORD [12+edi]
- adc ebp,0
- ; mul a[3]*b[3]
- mul edx
- add ebx,eax
- mov eax,DWORD [8+esi]
- adc ecx,edx
- mov edx,DWORD [16+edi]
- adc ebp,0
- ; mul a[2]*b[4]
- mul edx
- add ebx,eax
- mov eax,DWORD [4+esi]
- adc ecx,edx
- mov edx,DWORD [20+edi]
- adc ebp,0
- ; mul a[1]*b[5]
- mul edx
- add ebx,eax
- mov eax,DWORD [esi]
- adc ecx,edx
- mov edx,DWORD [24+edi]
- adc ebp,0
- ; mul a[0]*b[6]
- mul edx
- add ebx,eax
- mov eax,DWORD [20+esp]
- adc ecx,edx
- mov edx,DWORD [edi]
- adc ebp,0
- mov DWORD [24+eax],ebx
- mov eax,DWORD [28+esi]
- ; saved r[6]
- ; ################## Calculate word 7
- xor ebx,ebx
- ; mul a[7]*b[0]
- mul edx
- add ecx,eax
- mov eax,DWORD [24+esi]
- adc ebp,edx
- mov edx,DWORD [4+edi]
- adc ebx,0
- ; mul a[6]*b[1]
- mul edx
- add ecx,eax
- mov eax,DWORD [20+esi]
- adc ebp,edx
- mov edx,DWORD [8+edi]
- adc ebx,0
- ; mul a[5]*b[2]
- mul edx
- add ecx,eax
- mov eax,DWORD [16+esi]
- adc ebp,edx
- mov edx,DWORD [12+edi]
- adc ebx,0
- ; mul a[4]*b[3]
- mul edx
- add ecx,eax
- mov eax,DWORD [12+esi]
- adc ebp,edx
- mov edx,DWORD [16+edi]
- adc ebx,0
- ; mul a[3]*b[4]
- mul edx
- add ecx,eax
- mov eax,DWORD [8+esi]
- adc ebp,edx
- mov edx,DWORD [20+edi]
- adc ebx,0
- ; mul a[2]*b[5]
- mul edx
- add ecx,eax
- mov eax,DWORD [4+esi]
- adc ebp,edx
- mov edx,DWORD [24+edi]
- adc ebx,0
- ; mul a[1]*b[6]
- mul edx
- add ecx,eax
- mov eax,DWORD [esi]
- adc ebp,edx
- mov edx,DWORD [28+edi]
- adc ebx,0
- ; mul a[0]*b[7]
- mul edx
- add ecx,eax
- mov eax,DWORD [20+esp]
- adc ebp,edx
- mov edx,DWORD [4+edi]
- adc ebx,0
- mov DWORD [28+eax],ecx
- mov eax,DWORD [28+esi]
- ; saved r[7]
- ; ################## Calculate word 8
- xor ecx,ecx
- ; mul a[7]*b[1]
- mul edx
- add ebp,eax
- mov eax,DWORD [24+esi]
- adc ebx,edx
- mov edx,DWORD [8+edi]
- adc ecx,0
- ; mul a[6]*b[2]
- mul edx
- add ebp,eax
- mov eax,DWORD [20+esi]
- adc ebx,edx
- mov edx,DWORD [12+edi]
- adc ecx,0
- ; mul a[5]*b[3]
- mul edx
- add ebp,eax
- mov eax,DWORD [16+esi]
- adc ebx,edx
- mov edx,DWORD [16+edi]
- adc ecx,0
- ; mul a[4]*b[4]
- mul edx
- add ebp,eax
- mov eax,DWORD [12+esi]
- adc ebx,edx
- mov edx,DWORD [20+edi]
- adc ecx,0
- ; mul a[3]*b[5]
- mul edx
- add ebp,eax
- mov eax,DWORD [8+esi]
- adc ebx,edx
- mov edx,DWORD [24+edi]
- adc ecx,0
- ; mul a[2]*b[6]
- mul edx
- add ebp,eax
- mov eax,DWORD [4+esi]
- adc ebx,edx
- mov edx,DWORD [28+edi]
- adc ecx,0
- ; mul a[1]*b[7]
- mul edx
- add ebp,eax
- mov eax,DWORD [20+esp]
- adc ebx,edx
- mov edx,DWORD [8+edi]
- adc ecx,0
- mov DWORD [32+eax],ebp
- mov eax,DWORD [28+esi]
- ; saved r[8]
- ; ################## Calculate word 9
- xor ebp,ebp
- ; mul a[7]*b[2]
- mul edx
- add ebx,eax
- mov eax,DWORD [24+esi]
- adc ecx,edx
- mov edx,DWORD [12+edi]
- adc ebp,0
- ; mul a[6]*b[3]
- mul edx
- add ebx,eax
- mov eax,DWORD [20+esi]
- adc ecx,edx
- mov edx,DWORD [16+edi]
- adc ebp,0
- ; mul a[5]*b[4]
- mul edx
- add ebx,eax
- mov eax,DWORD [16+esi]
- adc ecx,edx
- mov edx,DWORD [20+edi]
- adc ebp,0
- ; mul a[4]*b[5]
- mul edx
- add ebx,eax
- mov eax,DWORD [12+esi]
- adc ecx,edx
- mov edx,DWORD [24+edi]
- adc ebp,0
- ; mul a[3]*b[6]
- mul edx
- add ebx,eax
- mov eax,DWORD [8+esi]
- adc ecx,edx
- mov edx,DWORD [28+edi]
- adc ebp,0
- ; mul a[2]*b[7]
- mul edx
- add ebx,eax
- mov eax,DWORD [20+esp]
- adc ecx,edx
- mov edx,DWORD [12+edi]
- adc ebp,0
- mov DWORD [36+eax],ebx
- mov eax,DWORD [28+esi]
- ; saved r[9]
- ; ################## Calculate word 10
- xor ebx,ebx
- ; mul a[7]*b[3]
- mul edx
- add ecx,eax
- mov eax,DWORD [24+esi]
- adc ebp,edx
- mov edx,DWORD [16+edi]
- adc ebx,0
- ; mul a[6]*b[4]
- mul edx
- add ecx,eax
- mov eax,DWORD [20+esi]
- adc ebp,edx
- mov edx,DWORD [20+edi]
- adc ebx,0
- ; mul a[5]*b[5]
- mul edx
- add ecx,eax
- mov eax,DWORD [16+esi]
- adc ebp,edx
- mov edx,DWORD [24+edi]
- adc ebx,0
- ; mul a[4]*b[6]
- mul edx
- add ecx,eax
- mov eax,DWORD [12+esi]
- adc ebp,edx
- mov edx,DWORD [28+edi]
- adc ebx,0
- ; mul a[3]*b[7]
- mul edx
- add ecx,eax
- mov eax,DWORD [20+esp]
- adc ebp,edx
- mov edx,DWORD [16+edi]
- adc ebx,0
- mov DWORD [40+eax],ecx
- mov eax,DWORD [28+esi]
- ; saved r[10]
- ; ################## Calculate word 11
- xor ecx,ecx
- ; mul a[7]*b[4]
- mul edx
- add ebp,eax
- mov eax,DWORD [24+esi]
- adc ebx,edx
- mov edx,DWORD [20+edi]
- adc ecx,0
- ; mul a[6]*b[5]
- mul edx
- add ebp,eax
- mov eax,DWORD [20+esi]
- adc ebx,edx
- mov edx,DWORD [24+edi]
- adc ecx,0
- ; mul a[5]*b[6]
- mul edx
- add ebp,eax
- mov eax,DWORD [16+esi]
- adc ebx,edx
- mov edx,DWORD [28+edi]
- adc ecx,0
- ; mul a[4]*b[7]
- mul edx
- add ebp,eax
- mov eax,DWORD [20+esp]
- adc ebx,edx
- mov edx,DWORD [20+edi]
- adc ecx,0
- mov DWORD [44+eax],ebp
- mov eax,DWORD [28+esi]
- ; saved r[11]
- ; ################## Calculate word 12
- xor ebp,ebp
- ; mul a[7]*b[5]
- mul edx
- add ebx,eax
- mov eax,DWORD [24+esi]
- adc ecx,edx
- mov edx,DWORD [24+edi]
- adc ebp,0
- ; mul a[6]*b[6]
- mul edx
- add ebx,eax
- mov eax,DWORD [20+esi]
- adc ecx,edx
- mov edx,DWORD [28+edi]
- adc ebp,0
- ; mul a[5]*b[7]
- mul edx
- add ebx,eax
- mov eax,DWORD [20+esp]
- adc ecx,edx
- mov edx,DWORD [24+edi]
- adc ebp,0
- mov DWORD [48+eax],ebx
- mov eax,DWORD [28+esi]
- ; saved r[12]
- ; ################## Calculate word 13
- xor ebx,ebx
- ; mul a[7]*b[6]
- mul edx
- add ecx,eax
- mov eax,DWORD [24+esi]
- adc ebp,edx
- mov edx,DWORD [28+edi]
- adc ebx,0
- ; mul a[6]*b[7]
- mul edx
- add ecx,eax
- mov eax,DWORD [20+esp]
- adc ebp,edx
- mov edx,DWORD [28+edi]
- adc ebx,0
- mov DWORD [52+eax],ecx
- mov eax,DWORD [28+esi]
- ; saved r[13]
- ; ################## Calculate word 14
- xor ecx,ecx
- ; mul a[7]*b[7]
- mul edx
- add ebp,eax
- mov eax,DWORD [20+esp]
- adc ebx,edx
- adc ecx,0
- mov DWORD [56+eax],ebp
- ; saved r[14]
- ; save r[15]
- mov DWORD [60+eax],ebx
- pop ebx
- pop ebp
- pop edi
- pop esi
- ret
- global _bn_mul_comba4
- align 16
- _bn_mul_comba4:
- L$_bn_mul_comba4_begin:
- push esi
- mov esi,DWORD [12+esp]
- push edi
- mov edi,DWORD [20+esp]
- push ebp
- push ebx
- xor ebx,ebx
- mov eax,DWORD [esi]
- xor ecx,ecx
- mov edx,DWORD [edi]
- ; ################## Calculate word 0
- xor ebp,ebp
- ; mul a[0]*b[0]
- mul edx
- add ebx,eax
- mov eax,DWORD [20+esp]
- adc ecx,edx
- mov edx,DWORD [edi]
- adc ebp,0
- mov DWORD [eax],ebx
- mov eax,DWORD [4+esi]
- ; saved r[0]
- ; ################## Calculate word 1
- xor ebx,ebx
- ; mul a[1]*b[0]
- mul edx
- add ecx,eax
- mov eax,DWORD [esi]
- adc ebp,edx
- mov edx,DWORD [4+edi]
- adc ebx,0
- ; mul a[0]*b[1]
- mul edx
- add ecx,eax
- mov eax,DWORD [20+esp]
- adc ebp,edx
- mov edx,DWORD [edi]
- adc ebx,0
- mov DWORD [4+eax],ecx
- mov eax,DWORD [8+esi]
- ; saved r[1]
- ; ################## Calculate word 2
- xor ecx,ecx
- ; mul a[2]*b[0]
- mul edx
- add ebp,eax
- mov eax,DWORD [4+esi]
- adc ebx,edx
- mov edx,DWORD [4+edi]
- adc ecx,0
- ; mul a[1]*b[1]
- mul edx
- add ebp,eax
- mov eax,DWORD [esi]
- adc ebx,edx
- mov edx,DWORD [8+edi]
- adc ecx,0
- ; mul a[0]*b[2]
- mul edx
- add ebp,eax
- mov eax,DWORD [20+esp]
- adc ebx,edx
- mov edx,DWORD [edi]
- adc ecx,0
- mov DWORD [8+eax],ebp
- mov eax,DWORD [12+esi]
- ; saved r[2]
- ; ################## Calculate word 3
- xor ebp,ebp
- ; mul a[3]*b[0]
- mul edx
- add ebx,eax
- mov eax,DWORD [8+esi]
- adc ecx,edx
- mov edx,DWORD [4+edi]
- adc ebp,0
- ; mul a[2]*b[1]
- mul edx
- add ebx,eax
- mov eax,DWORD [4+esi]
- adc ecx,edx
- mov edx,DWORD [8+edi]
- adc ebp,0
- ; mul a[1]*b[2]
- mul edx
- add ebx,eax
- mov eax,DWORD [esi]
- adc ecx,edx
- mov edx,DWORD [12+edi]
- adc ebp,0
- ; mul a[0]*b[3]
- mul edx
- add ebx,eax
- mov eax,DWORD [20+esp]
- adc ecx,edx
- mov edx,DWORD [4+edi]
- adc ebp,0
- mov DWORD [12+eax],ebx
- mov eax,DWORD [12+esi]
- ; saved r[3]
- ; ################## Calculate word 4
- xor ebx,ebx
- ; mul a[3]*b[1]
- mul edx
- add ecx,eax
- mov eax,DWORD [8+esi]
- adc ebp,edx
- mov edx,DWORD [8+edi]
- adc ebx,0
- ; mul a[2]*b[2]
- mul edx
- add ecx,eax
- mov eax,DWORD [4+esi]
- adc ebp,edx
- mov edx,DWORD [12+edi]
- adc ebx,0
- ; mul a[1]*b[3]
- mul edx
- add ecx,eax
- mov eax,DWORD [20+esp]
- adc ebp,edx
- mov edx,DWORD [8+edi]
- adc ebx,0
- mov DWORD [16+eax],ecx
- mov eax,DWORD [12+esi]
- ; saved r[4]
- ; ################## Calculate word 5
- xor ecx,ecx
- ; mul a[3]*b[2]
- mul edx
- add ebp,eax
- mov eax,DWORD [8+esi]
- adc ebx,edx
- mov edx,DWORD [12+edi]
- adc ecx,0
- ; mul a[2]*b[3]
- mul edx
- add ebp,eax
- mov eax,DWORD [20+esp]
- adc ebx,edx
- mov edx,DWORD [12+edi]
- adc ecx,0
- mov DWORD [20+eax],ebp
- mov eax,DWORD [12+esi]
- ; saved r[5]
- ; ################## Calculate word 6
- xor ebp,ebp
- ; mul a[3]*b[3]
- mul edx
- add ebx,eax
- mov eax,DWORD [20+esp]
- adc ecx,edx
- adc ebp,0
- mov DWORD [24+eax],ebx
- ; saved r[6]
- ; save r[7]
- mov DWORD [28+eax],ecx
- pop ebx
- pop ebp
- pop edi
- pop esi
- ret
- global _bn_sqr_comba8
- align 16
- _bn_sqr_comba8:
- L$_bn_sqr_comba8_begin:
- push esi
- push edi
- push ebp
- push ebx
- mov edi,DWORD [20+esp]
- mov esi,DWORD [24+esp]
- xor ebx,ebx
- xor ecx,ecx
- mov eax,DWORD [esi]
- ; ############### Calculate word 0
- xor ebp,ebp
- ; sqr a[0]*a[0]
- mul eax
- add ebx,eax
- adc ecx,edx
- mov edx,DWORD [esi]
- adc ebp,0
- mov DWORD [edi],ebx
- mov eax,DWORD [4+esi]
- ; saved r[0]
- ; ############### Calculate word 1
- xor ebx,ebx
- ; sqr a[1]*a[0]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebx,0
- add ecx,eax
- adc ebp,edx
- mov eax,DWORD [8+esi]
- adc ebx,0
- mov DWORD [4+edi],ecx
- mov edx,DWORD [esi]
- ; saved r[1]
- ; ############### Calculate word 2
- xor ecx,ecx
- ; sqr a[2]*a[0]
- mul edx
- add eax,eax
- adc edx,edx
- adc ecx,0
- add ebp,eax
- adc ebx,edx
- mov eax,DWORD [4+esi]
- adc ecx,0
- ; sqr a[1]*a[1]
- mul eax
- add ebp,eax
- adc ebx,edx
- mov edx,DWORD [esi]
- adc ecx,0
- mov DWORD [8+edi],ebp
- mov eax,DWORD [12+esi]
- ; saved r[2]
- ; ############### Calculate word 3
- xor ebp,ebp
- ; sqr a[3]*a[0]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebp,0
- add ebx,eax
- adc ecx,edx
- mov eax,DWORD [8+esi]
- adc ebp,0
- mov edx,DWORD [4+esi]
- ; sqr a[2]*a[1]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebp,0
- add ebx,eax
- adc ecx,edx
- mov eax,DWORD [16+esi]
- adc ebp,0
- mov DWORD [12+edi],ebx
- mov edx,DWORD [esi]
- ; saved r[3]
- ; ############### Calculate word 4
- xor ebx,ebx
- ; sqr a[4]*a[0]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebx,0
- add ecx,eax
- adc ebp,edx
- mov eax,DWORD [12+esi]
- adc ebx,0
- mov edx,DWORD [4+esi]
- ; sqr a[3]*a[1]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebx,0
- add ecx,eax
- adc ebp,edx
- mov eax,DWORD [8+esi]
- adc ebx,0
- ; sqr a[2]*a[2]
- mul eax
- add ecx,eax
- adc ebp,edx
- mov edx,DWORD [esi]
- adc ebx,0
- mov DWORD [16+edi],ecx
- mov eax,DWORD [20+esi]
- ; saved r[4]
- ; ############### Calculate word 5
- xor ecx,ecx
- ; sqr a[5]*a[0]
- mul edx
- add eax,eax
- adc edx,edx
- adc ecx,0
- add ebp,eax
- adc ebx,edx
- mov eax,DWORD [16+esi]
- adc ecx,0
- mov edx,DWORD [4+esi]
- ; sqr a[4]*a[1]
- mul edx
- add eax,eax
- adc edx,edx
- adc ecx,0
- add ebp,eax
- adc ebx,edx
- mov eax,DWORD [12+esi]
- adc ecx,0
- mov edx,DWORD [8+esi]
- ; sqr a[3]*a[2]
- mul edx
- add eax,eax
- adc edx,edx
- adc ecx,0
- add ebp,eax
- adc ebx,edx
- mov eax,DWORD [24+esi]
- adc ecx,0
- mov DWORD [20+edi],ebp
- mov edx,DWORD [esi]
- ; saved r[5]
- ; ############### Calculate word 6
- xor ebp,ebp
- ; sqr a[6]*a[0]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebp,0
- add ebx,eax
- adc ecx,edx
- mov eax,DWORD [20+esi]
- adc ebp,0
- mov edx,DWORD [4+esi]
- ; sqr a[5]*a[1]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebp,0
- add ebx,eax
- adc ecx,edx
- mov eax,DWORD [16+esi]
- adc ebp,0
- mov edx,DWORD [8+esi]
- ; sqr a[4]*a[2]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebp,0
- add ebx,eax
- adc ecx,edx
- mov eax,DWORD [12+esi]
- adc ebp,0
- ; sqr a[3]*a[3]
- mul eax
- add ebx,eax
- adc ecx,edx
- mov edx,DWORD [esi]
- adc ebp,0
- mov DWORD [24+edi],ebx
- mov eax,DWORD [28+esi]
- ; saved r[6]
- ; ############### Calculate word 7
- xor ebx,ebx
- ; sqr a[7]*a[0]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebx,0
- add ecx,eax
- adc ebp,edx
- mov eax,DWORD [24+esi]
- adc ebx,0
- mov edx,DWORD [4+esi]
- ; sqr a[6]*a[1]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebx,0
- add ecx,eax
- adc ebp,edx
- mov eax,DWORD [20+esi]
- adc ebx,0
- mov edx,DWORD [8+esi]
- ; sqr a[5]*a[2]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebx,0
- add ecx,eax
- adc ebp,edx
- mov eax,DWORD [16+esi]
- adc ebx,0
- mov edx,DWORD [12+esi]
- ; sqr a[4]*a[3]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebx,0
- add ecx,eax
- adc ebp,edx
- mov eax,DWORD [28+esi]
- adc ebx,0
- mov DWORD [28+edi],ecx
- mov edx,DWORD [4+esi]
- ; saved r[7]
- ; ############### Calculate word 8
- xor ecx,ecx
- ; sqr a[7]*a[1]
- mul edx
- add eax,eax
- adc edx,edx
- adc ecx,0
- add ebp,eax
- adc ebx,edx
- mov eax,DWORD [24+esi]
- adc ecx,0
- mov edx,DWORD [8+esi]
- ; sqr a[6]*a[2]
- mul edx
- add eax,eax
- adc edx,edx
- adc ecx,0
- add ebp,eax
- adc ebx,edx
- mov eax,DWORD [20+esi]
- adc ecx,0
- mov edx,DWORD [12+esi]
- ; sqr a[5]*a[3]
- mul edx
- add eax,eax
- adc edx,edx
- adc ecx,0
- add ebp,eax
- adc ebx,edx
- mov eax,DWORD [16+esi]
- adc ecx,0
- ; sqr a[4]*a[4]
- mul eax
- add ebp,eax
- adc ebx,edx
- mov edx,DWORD [8+esi]
- adc ecx,0
- mov DWORD [32+edi],ebp
- mov eax,DWORD [28+esi]
- ; saved r[8]
- ; ############### Calculate word 9
- xor ebp,ebp
- ; sqr a[7]*a[2]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebp,0
- add ebx,eax
- adc ecx,edx
- mov eax,DWORD [24+esi]
- adc ebp,0
- mov edx,DWORD [12+esi]
- ; sqr a[6]*a[3]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebp,0
- add ebx,eax
- adc ecx,edx
- mov eax,DWORD [20+esi]
- adc ebp,0
- mov edx,DWORD [16+esi]
- ; sqr a[5]*a[4]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebp,0
- add ebx,eax
- adc ecx,edx
- mov eax,DWORD [28+esi]
- adc ebp,0
- mov DWORD [36+edi],ebx
- mov edx,DWORD [12+esi]
- ; saved r[9]
- ; ############### Calculate word 10
- xor ebx,ebx
- ; sqr a[7]*a[3]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebx,0
- add ecx,eax
- adc ebp,edx
- mov eax,DWORD [24+esi]
- adc ebx,0
- mov edx,DWORD [16+esi]
- ; sqr a[6]*a[4]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebx,0
- add ecx,eax
- adc ebp,edx
- mov eax,DWORD [20+esi]
- adc ebx,0
- ; sqr a[5]*a[5]
- mul eax
- add ecx,eax
- adc ebp,edx
- mov edx,DWORD [16+esi]
- adc ebx,0
- mov DWORD [40+edi],ecx
- mov eax,DWORD [28+esi]
- ; saved r[10]
- ; ############### Calculate word 11
- xor ecx,ecx
- ; sqr a[7]*a[4]
- mul edx
- add eax,eax
- adc edx,edx
- adc ecx,0
- add ebp,eax
- adc ebx,edx
- mov eax,DWORD [24+esi]
- adc ecx,0
- mov edx,DWORD [20+esi]
- ; sqr a[6]*a[5]
- mul edx
- add eax,eax
- adc edx,edx
- adc ecx,0
- add ebp,eax
- adc ebx,edx
- mov eax,DWORD [28+esi]
- adc ecx,0
- mov DWORD [44+edi],ebp
- mov edx,DWORD [20+esi]
- ; saved r[11]
- ; ############### Calculate word 12
- xor ebp,ebp
- ; sqr a[7]*a[5]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebp,0
- add ebx,eax
- adc ecx,edx
- mov eax,DWORD [24+esi]
- adc ebp,0
- ; sqr a[6]*a[6]
- mul eax
- add ebx,eax
- adc ecx,edx
- mov edx,DWORD [24+esi]
- adc ebp,0
- mov DWORD [48+edi],ebx
- mov eax,DWORD [28+esi]
- ; saved r[12]
- ; ############### Calculate word 13
- xor ebx,ebx
- ; sqr a[7]*a[6]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebx,0
- add ecx,eax
- adc ebp,edx
- mov eax,DWORD [28+esi]
- adc ebx,0
- mov DWORD [52+edi],ecx
- ; saved r[13]
- ; ############### Calculate word 14
- xor ecx,ecx
- ; sqr a[7]*a[7]
- mul eax
- add ebp,eax
- adc ebx,edx
- adc ecx,0
- mov DWORD [56+edi],ebp
- ; saved r[14]
- mov DWORD [60+edi],ebx
- pop ebx
- pop ebp
- pop edi
- pop esi
- ret
- global _bn_sqr_comba4
- align 16
- _bn_sqr_comba4:
- L$_bn_sqr_comba4_begin:
- push esi
- push edi
- push ebp
- push ebx
- mov edi,DWORD [20+esp]
- mov esi,DWORD [24+esp]
- xor ebx,ebx
- xor ecx,ecx
- mov eax,DWORD [esi]
- ; ############### Calculate word 0
- xor ebp,ebp
- ; sqr a[0]*a[0]
- mul eax
- add ebx,eax
- adc ecx,edx
- mov edx,DWORD [esi]
- adc ebp,0
- mov DWORD [edi],ebx
- mov eax,DWORD [4+esi]
- ; saved r[0]
- ; ############### Calculate word 1
- xor ebx,ebx
- ; sqr a[1]*a[0]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebx,0
- add ecx,eax
- adc ebp,edx
- mov eax,DWORD [8+esi]
- adc ebx,0
- mov DWORD [4+edi],ecx
- mov edx,DWORD [esi]
- ; saved r[1]
- ; ############### Calculate word 2
- xor ecx,ecx
- ; sqr a[2]*a[0]
- mul edx
- add eax,eax
- adc edx,edx
- adc ecx,0
- add ebp,eax
- adc ebx,edx
- mov eax,DWORD [4+esi]
- adc ecx,0
- ; sqr a[1]*a[1]
- mul eax
- add ebp,eax
- adc ebx,edx
- mov edx,DWORD [esi]
- adc ecx,0
- mov DWORD [8+edi],ebp
- mov eax,DWORD [12+esi]
- ; saved r[2]
- ; ############### Calculate word 3
- xor ebp,ebp
- ; sqr a[3]*a[0]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebp,0
- add ebx,eax
- adc ecx,edx
- mov eax,DWORD [8+esi]
- adc ebp,0
- mov edx,DWORD [4+esi]
- ; sqr a[2]*a[1]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebp,0
- add ebx,eax
- adc ecx,edx
- mov eax,DWORD [12+esi]
- adc ebp,0
- mov DWORD [12+edi],ebx
- mov edx,DWORD [4+esi]
- ; saved r[3]
- ; ############### Calculate word 4
- xor ebx,ebx
- ; sqr a[3]*a[1]
- mul edx
- add eax,eax
- adc edx,edx
- adc ebx,0
- add ecx,eax
- adc ebp,edx
- mov eax,DWORD [8+esi]
- adc ebx,0
- ; sqr a[2]*a[2]
- mul eax
- add ecx,eax
- adc ebp,edx
- mov edx,DWORD [8+esi]
- adc ebx,0
- mov DWORD [16+edi],ecx
- mov eax,DWORD [12+esi]
- ; saved r[4]
- ; ############### Calculate word 5
- xor ecx,ecx
- ; sqr a[3]*a[2]
- mul edx
- add eax,eax
- adc edx,edx
- adc ecx,0
- add ebp,eax
- adc ebx,edx
- mov eax,DWORD [12+esi]
- adc ecx,0
- mov DWORD [20+edi],ebp
- ; saved r[5]
- ; ############### Calculate word 6
- xor ebp,ebp
- ; sqr a[3]*a[3]
- mul eax
- add ebx,eax
- adc ecx,edx
- adc ebp,0
- mov DWORD [24+edi],ebx
- ; saved r[6]
- mov DWORD [28+edi],ecx
- pop ebx
- pop ebp
- pop edi
- pop esi
- ret
|