1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789 |
- .machine "any"
- .abiversion 2
- .text
- .globl bn_mul_mont_int
- .type bn_mul_mont_int,@function
- .align 5
- bn_mul_mont_int:
- .localentry bn_mul_mont_int,0
- mr 9,3
- li 3,0
- slwi 8,8,3
- li 12,-4096
- addi 3,8,352
- subf 3,3,1
- and 3,3,12
- subf 3,1,3
- mr 12,1
- srwi 8,8,3
- stdux 1,1,3
- std 20,-96(12)
- std 21,-88(12)
- std 22,-80(12)
- std 23,-72(12)
- std 24,-64(12)
- std 25,-56(12)
- std 26,-48(12)
- std 27,-40(12)
- std 28,-32(12)
- std 29,-24(12)
- std 30,-16(12)
- std 31,-8(12)
- ld 7,0(7)
- addi 8,8,-2
- ld 23,0(5)
- ld 10,0(4)
- addi 22,1,64
- mulld 25,10,23
- mulhdu 26,10,23
- ld 10,8(4)
- ld 11,0(6)
- mulld 24,25,7
- mulld 29,10,23
- mulhdu 30,10,23
- mulld 27,11,24
- mulhdu 28,11,24
- ld 11,8(6)
- addc 27,27,25
- addze 28,28
- mulld 31,11,24
- mulhdu 0,11,24
- mtctr 8
- li 21,16
- .align 4
- .L1st:
- ldx 10,4,21
- addc 25,29,26
- ldx 11,6,21
- addze 26,30
- mulld 29,10,23
- addc 27,31,28
- mulhdu 30,10,23
- addze 28,0
- mulld 31,11,24
- addc 27,27,25
- mulhdu 0,11,24
- addze 28,28
- std 27,0(22)
- addi 21,21,8
- addi 22,22,8
- bdnz .L1st
- addc 25,29,26
- addze 26,30
- addc 27,31,28
- addze 28,0
- addc 27,27,25
- addze 28,28
- std 27,0(22)
- li 3,0
- addc 28,28,26
- addze 3,3
- std 28,8(22)
- li 20,8
- .align 4
- .Louter:
- ldx 23,5,20
- ld 10,0(4)
- addi 22,1,64
- ld 12,64(1)
- mulld 25,10,23
- mulhdu 26,10,23
- ld 10,8(4)
- ld 11,0(6)
- addc 25,25,12
- mulld 29,10,23
- addze 26,26
- mulld 24,25,7
- mulhdu 30,10,23
- mulld 27,11,24
- mulhdu 28,11,24
- ld 11,8(6)
- addc 27,27,25
- mulld 31,11,24
- addze 28,28
- mulhdu 0,11,24
- mtctr 8
- li 21,16
- .align 4
- .Linner:
- ldx 10,4,21
- addc 25,29,26
- ld 12,8(22)
- addze 26,30
- ldx 11,6,21
- addc 27,31,28
- mulld 29,10,23
- addze 28,0
- mulhdu 30,10,23
- addc 25,25,12
- mulld 31,11,24
- addze 26,26
- mulhdu 0,11,24
- addc 27,27,25
- addi 21,21,8
- addze 28,28
- std 27,0(22)
- addi 22,22,8
- bdnz .Linner
- ld 12,8(22)
- addc 25,29,26
- addze 26,30
- addc 25,25,12
- addze 26,26
- addc 27,31,28
- addze 28,0
- addc 27,27,25
- addze 28,28
- std 27,0(22)
- addic 3,3,-1
- li 3,0
- adde 28,28,26
- addze 3,3
- std 28,8(22)
- slwi 12,8,3
- cmpld 20,12
- addi 20,20,8
- ble .Louter
- addi 8,8,2
- subfc 21,21,21
- addi 22,1,64
- mtctr 8
- .align 4
- .Lsub: ldx 12,22,21
- ldx 11,6,21
- subfe 10,11,12
- stdx 10,9,21
- addi 21,21,8
- bdnz .Lsub
- li 21,0
- mtctr 8
- subfe 3,21,3
- .align 4
- .Lcopy:
- ldx 12,22,21
- ldx 10,9,21
- and 12,12,3
- andc 10,10,3
- stdx 21,22,21
- or 10,10,12
- stdx 10,9,21
- addi 21,21,8
- bdnz .Lcopy
- ld 12,0(1)
- li 3,1
- ld 20,-96(12)
- ld 21,-88(12)
- ld 22,-80(12)
- ld 23,-72(12)
- ld 24,-64(12)
- ld 25,-56(12)
- ld 26,-48(12)
- ld 27,-40(12)
- ld 28,-32(12)
- ld 29,-24(12)
- ld 30,-16(12)
- ld 31,-8(12)
- mr 1,12
- blr
- .long 0
- .byte 0,12,4,0,0x80,12,6,0
- .long 0
- .size bn_mul_mont_int,.-bn_mul_mont_int
- .globl bn_mul4x_mont_int
- .type bn_mul4x_mont_int,@function
- .align 5
- bn_mul4x_mont_int:
- .localentry bn_mul4x_mont_int,0
- andi. 0,8,7
- bne .Lmul4x_do
- cmpld 4,5
- bne .Lmul4x_do
- b .Lsqr8x_do
- .Lmul4x_do:
- slwi 8,8,3
- mr 9,1
- li 10,-32*8
- sub 10,10,8
- stdux 1,1,10
- std 14,-8*18(9)
- std 15,-8*17(9)
- std 16,-8*16(9)
- std 17,-8*15(9)
- std 18,-8*14(9)
- std 19,-8*13(9)
- std 20,-8*12(9)
- std 21,-8*11(9)
- std 22,-8*10(9)
- std 23,-8*9(9)
- std 24,-8*8(9)
- std 25,-8*7(9)
- std 26,-8*6(9)
- std 27,-8*5(9)
- std 28,-8*4(9)
- std 29,-8*3(9)
- std 30,-8*2(9)
- std 31,-8*1(9)
- subi 4,4,8
- subi 6,6,8
- subi 3,3,8
- ld 7,0(7)
- add 14,5,8
- add 30,4,8
- subi 14,14,8*4
- ld 27,8*0(5)
- li 22,0
- ld 9,8*1(4)
- li 23,0
- ld 10,8*2(4)
- li 24,0
- ld 11,8*3(4)
- li 25,0
- ldu 12,8*4(4)
- ld 18,8*1(6)
- ld 19,8*2(6)
- ld 20,8*3(6)
- ldu 21,8*4(6)
- std 3,8*6(1)
- std 14,8*7(1)
- li 3,0
- addic 29,1,8*7
- li 31,0
- li 0,0
- b .Loop_mul4x_1st_reduction
- .align 5
- .Loop_mul4x_1st_reduction:
- mulld 14,9,27
- addze 3,3
- mulld 15,10,27
- addi 31,31,8
- mulld 16,11,27
- andi. 31,31,8*4-1
- mulld 17,12,27
- addc 22,22,14
- mulhdu 14,9,27
- adde 23,23,15
- mulhdu 15,10,27
- adde 24,24,16
- mulld 28,22,7
- adde 25,25,17
- mulhdu 16,11,27
- addze 26,0
- mulhdu 17,12,27
- ldx 27,5,31
- addc 23,23,14
- stdu 28,8(29)
- adde 24,24,15
- mulld 15,19,28
- adde 25,25,16
- mulld 16,20,28
- adde 26,26,17
- mulld 17,21,28
- addic 22,22,-1
- mulhdu 14,18,28
- adde 22,23,15
- mulhdu 15,19,28
- adde 23,24,16
- mulhdu 16,20,28
- adde 24,25,17
- mulhdu 17,21,28
- adde 25,26,3
- addze 3,0
- addc 22,22,14
- adde 23,23,15
- adde 24,24,16
- adde 25,25,17
- bne .Loop_mul4x_1st_reduction
- cmpld 30,4
- beq .Lmul4x4_post_condition
- ld 9,8*1(4)
- ld 10,8*2(4)
- ld 11,8*3(4)
- ldu 12,8*4(4)
- ld 28,8*8(1)
- ld 18,8*1(6)
- ld 19,8*2(6)
- ld 20,8*3(6)
- ldu 21,8*4(6)
- b .Loop_mul4x_1st_tail
- .align 5
- .Loop_mul4x_1st_tail:
- mulld 14,9,27
- addze 3,3
- mulld 15,10,27
- addi 31,31,8
- mulld 16,11,27
- andi. 31,31,8*4-1
- mulld 17,12,27
- addc 22,22,14
- mulhdu 14,9,27
- adde 23,23,15
- mulhdu 15,10,27
- adde 24,24,16
- mulhdu 16,11,27
- adde 25,25,17
- mulhdu 17,12,27
- addze 26,0
- ldx 27,5,31
- addc 23,23,14
- mulld 14,18,28
- adde 24,24,15
- mulld 15,19,28
- adde 25,25,16
- mulld 16,20,28
- adde 26,26,17
- mulld 17,21,28
- addc 22,22,14
- mulhdu 14,18,28
- adde 23,23,15
- mulhdu 15,19,28
- adde 24,24,16
- mulhdu 16,20,28
- adde 25,25,17
- adde 26,26,3
- mulhdu 17,21,28
- addze 3,0
- addi 28,1,8*8
- ldx 28,28,31
- stdu 22,8(29)
- addc 22,23,14
- adde 23,24,15
- adde 24,25,16
- adde 25,26,17
- bne .Loop_mul4x_1st_tail
- sub 15,30,8
- cmpld 30,4
- beq .Lmul4x_proceed
- ld 9,8*1(4)
- ld 10,8*2(4)
- ld 11,8*3(4)
- ldu 12,8*4(4)
- ld 18,8*1(6)
- ld 19,8*2(6)
- ld 20,8*3(6)
- ldu 21,8*4(6)
- b .Loop_mul4x_1st_tail
- .align 5
- .Lmul4x_proceed:
- ldu 27,8*4(5)
- addze 3,3
- ld 9,8*1(15)
- ld 10,8*2(15)
- ld 11,8*3(15)
- ld 12,8*4(15)
- addi 4,15,8*4
- sub 6,6,8
- std 22,8*1(29)
- std 23,8*2(29)
- std 24,8*3(29)
- std 25,8*4(29)
- std 3,8*5(29)
- ld 22,8*12(1)
- ld 23,8*13(1)
- ld 24,8*14(1)
- ld 25,8*15(1)
- ld 18,8*1(6)
- ld 19,8*2(6)
- ld 20,8*3(6)
- ldu 21,8*4(6)
- addic 29,1,8*7
- li 3,0
- b .Loop_mul4x_reduction
- .align 5
- .Loop_mul4x_reduction:
- mulld 14,9,27
- addze 3,3
- mulld 15,10,27
- addi 31,31,8
- mulld 16,11,27
- andi. 31,31,8*4-1
- mulld 17,12,27
- addc 22,22,14
- mulhdu 14,9,27
- adde 23,23,15
- mulhdu 15,10,27
- adde 24,24,16
- mulld 28,22,7
- adde 25,25,17
- mulhdu 16,11,27
- addze 26,0
- mulhdu 17,12,27
- ldx 27,5,31
- addc 23,23,14
- stdu 28,8(29)
- adde 24,24,15
- mulld 15,19,28
- adde 25,25,16
- mulld 16,20,28
- adde 26,26,17
- mulld 17,21,28
- addic 22,22,-1
- mulhdu 14,18,28
- adde 22,23,15
- mulhdu 15,19,28
- adde 23,24,16
- mulhdu 16,20,28
- adde 24,25,17
- mulhdu 17,21,28
- adde 25,26,3
- addze 3,0
- addc 22,22,14
- adde 23,23,15
- adde 24,24,16
- adde 25,25,17
- bne .Loop_mul4x_reduction
- ld 14,8*5(29)
- addze 3,3
- ld 15,8*6(29)
- ld 16,8*7(29)
- ld 17,8*8(29)
- ld 9,8*1(4)
- ld 10,8*2(4)
- ld 11,8*3(4)
- ldu 12,8*4(4)
- addc 22,22,14
- adde 23,23,15
- adde 24,24,16
- adde 25,25,17
- ld 28,8*8(1)
- ld 18,8*1(6)
- ld 19,8*2(6)
- ld 20,8*3(6)
- ldu 21,8*4(6)
- b .Loop_mul4x_tail
- .align 5
- .Loop_mul4x_tail:
- mulld 14,9,27
- addze 3,3
- mulld 15,10,27
- addi 31,31,8
- mulld 16,11,27
- andi. 31,31,8*4-1
- mulld 17,12,27
- addc 22,22,14
- mulhdu 14,9,27
- adde 23,23,15
- mulhdu 15,10,27
- adde 24,24,16
- mulhdu 16,11,27
- adde 25,25,17
- mulhdu 17,12,27
- addze 26,0
- ldx 27,5,31
- addc 23,23,14
- mulld 14,18,28
- adde 24,24,15
- mulld 15,19,28
- adde 25,25,16
- mulld 16,20,28
- adde 26,26,17
- mulld 17,21,28
- addc 22,22,14
- mulhdu 14,18,28
- adde 23,23,15
- mulhdu 15,19,28
- adde 24,24,16
- mulhdu 16,20,28
- adde 25,25,17
- mulhdu 17,21,28
- adde 26,26,3
- addi 28,1,8*8
- ldx 28,28,31
- addze 3,0
- stdu 22,8(29)
- addc 22,23,14
- adde 23,24,15
- adde 24,25,16
- adde 25,26,17
- bne .Loop_mul4x_tail
- ld 14,8*5(29)
- sub 15,6,8
- addze 3,3
- cmpld 30,4
- beq .Loop_mul4x_break
- ld 15,8*6(29)
- ld 16,8*7(29)
- ld 17,8*8(29)
- ld 9,8*1(4)
- ld 10,8*2(4)
- ld 11,8*3(4)
- ldu 12,8*4(4)
- addc 22,22,14
- adde 23,23,15
- adde 24,24,16
- adde 25,25,17
- ld 18,8*1(6)
- ld 19,8*2(6)
- ld 20,8*3(6)
- ldu 21,8*4(6)
- b .Loop_mul4x_tail
- .align 5
- .Loop_mul4x_break:
- ld 16,8*6(1)
- ld 17,8*7(1)
- addc 9,22,14
- ld 22,8*12(1)
- addze 10,23
- ld 23,8*13(1)
- addze 11,24
- ld 24,8*14(1)
- addze 12,25
- ld 25,8*15(1)
- addze 3,3
- std 9,8*1(29)
- sub 4,30,8
- std 10,8*2(29)
- std 11,8*3(29)
- std 12,8*4(29)
- std 3,8*5(29)
- ld 18,8*1(15)
- ld 19,8*2(15)
- ld 20,8*3(15)
- ld 21,8*4(15)
- addi 6,15,8*4
- cmpld 5,17
- beq .Lmul4x_post
- ldu 27,8*4(5)
- ld 9,8*1(4)
- ld 10,8*2(4)
- ld 11,8*3(4)
- ldu 12,8*4(4)
- li 3,0
- addic 29,1,8*7
- b .Loop_mul4x_reduction
- .align 5
- .Lmul4x_post:
- srwi 31,8,5
- mr 5,16
- subi 31,31,1
- mr 30,16
- subfc 14,18,22
- addi 29,1,8*15
- subfe 15,19,23
- mtctr 31
- .Lmul4x_sub:
- ld 18,8*1(6)
- ld 22,8*1(29)
- subfe 16,20,24
- ld 19,8*2(6)
- ld 23,8*2(29)
- subfe 17,21,25
- ld 20,8*3(6)
- ld 24,8*3(29)
- ldu 21,8*4(6)
- ldu 25,8*4(29)
- std 14,8*1(5)
- std 15,8*2(5)
- subfe 14,18,22
- std 16,8*3(5)
- stdu 17,8*4(5)
- subfe 15,19,23
- bdnz .Lmul4x_sub
- ld 9,8*1(30)
- std 14,8*1(5)
- ld 14,8*12(1)
- subfe 16,20,24
- ld 10,8*2(30)
- std 15,8*2(5)
- ld 15,8*13(1)
- subfe 17,21,25
- subfe 3,0,3
- addi 29,1,8*12
- ld 11,8*3(30)
- std 16,8*3(5)
- ld 16,8*14(1)
- ld 12,8*4(30)
- std 17,8*4(5)
- ld 17,8*15(1)
- mtctr 31
- .Lmul4x_cond_copy:
- and 14,14,3
- andc 9,9,3
- std 0,8*0(29)
- and 15,15,3
- andc 10,10,3
- std 0,8*1(29)
- and 16,16,3
- andc 11,11,3
- std 0,8*2(29)
- and 17,17,3
- andc 12,12,3
- std 0,8*3(29)
- or 22,14,9
- ld 9,8*5(30)
- ld 14,8*4(29)
- or 23,15,10
- ld 10,8*6(30)
- ld 15,8*5(29)
- or 24,16,11
- ld 11,8*7(30)
- ld 16,8*6(29)
- or 25,17,12
- ld 12,8*8(30)
- ld 17,8*7(29)
- addi 29,29,8*4
- std 22,8*1(30)
- std 23,8*2(30)
- std 24,8*3(30)
- stdu 25,8*4(30)
- bdnz .Lmul4x_cond_copy
- ld 5,0(1)
- and 14,14,3
- andc 9,9,3
- std 0,8*0(29)
- and 15,15,3
- andc 10,10,3
- std 0,8*1(29)
- and 16,16,3
- andc 11,11,3
- std 0,8*2(29)
- and 17,17,3
- andc 12,12,3
- std 0,8*3(29)
- or 22,14,9
- or 23,15,10
- std 0,8*4(29)
- or 24,16,11
- or 25,17,12
- std 22,8*1(30)
- std 23,8*2(30)
- std 24,8*3(30)
- std 25,8*4(30)
- b .Lmul4x_done
- .align 4
- .Lmul4x4_post_condition:
- ld 4,8*6(1)
- ld 5,0(1)
- addze 3,3
- subfc 9,18,22
- subfe 10,19,23
- subfe 11,20,24
- subfe 12,21,25
- subfe 3,0,3
- and 18,18,3
- and 19,19,3
- addc 9,9,18
- and 20,20,3
- adde 10,10,19
- and 21,21,3
- adde 11,11,20
- adde 12,12,21
- std 9,8*1(4)
- std 10,8*2(4)
- std 11,8*3(4)
- std 12,8*4(4)
- .Lmul4x_done:
- std 0,8*8(1)
- std 0,8*9(1)
- std 0,8*10(1)
- std 0,8*11(1)
- li 3,1
- ld 14,-8*18(5)
- ld 15,-8*17(5)
- ld 16,-8*16(5)
- ld 17,-8*15(5)
- ld 18,-8*14(5)
- ld 19,-8*13(5)
- ld 20,-8*12(5)
- ld 21,-8*11(5)
- ld 22,-8*10(5)
- ld 23,-8*9(5)
- ld 24,-8*8(5)
- ld 25,-8*7(5)
- ld 26,-8*6(5)
- ld 27,-8*5(5)
- ld 28,-8*4(5)
- ld 29,-8*3(5)
- ld 30,-8*2(5)
- ld 31,-8*1(5)
- mr 1,5
- blr
- .long 0
- .byte 0,12,4,0x20,0x80,18,6,0
- .long 0
- .size bn_mul4x_mont_int,.-bn_mul4x_mont_int
- .align 5
- __bn_sqr8x_mont:
- .Lsqr8x_do:
- mr 9,1
- slwi 10,8,4
- li 11,-32*8
- sub 10,11,10
- slwi 8,8,3
- stdux 1,1,10
- std 14,-8*18(9)
- std 15,-8*17(9)
- std 16,-8*16(9)
- std 17,-8*15(9)
- std 18,-8*14(9)
- std 19,-8*13(9)
- std 20,-8*12(9)
- std 21,-8*11(9)
- std 22,-8*10(9)
- std 23,-8*9(9)
- std 24,-8*8(9)
- std 25,-8*7(9)
- std 26,-8*6(9)
- std 27,-8*5(9)
- std 28,-8*4(9)
- std 29,-8*3(9)
- std 30,-8*2(9)
- std 31,-8*1(9)
- subi 4,4,8
- subi 18,6,8
- subi 3,3,8
- ld 7,0(7)
- li 0,0
- add 6,4,8
- ld 9,8*1(4)
- ld 10,8*2(4)
- li 23,0
- ld 11,8*3(4)
- li 24,0
- ld 12,8*4(4)
- li 25,0
- ld 14,8*5(4)
- li 26,0
- ld 15,8*6(4)
- li 27,0
- ld 16,8*7(4)
- li 28,0
- ldu 17,8*8(4)
- li 29,0
- addi 5,1,8*11
- subic. 30,8,8*8
- b .Lsqr8x_zero_start
- .align 5
- .Lsqr8x_zero:
- subic. 30,30,8*8
- std 0,8*1(5)
- std 0,8*2(5)
- std 0,8*3(5)
- std 0,8*4(5)
- std 0,8*5(5)
- std 0,8*6(5)
- std 0,8*7(5)
- std 0,8*8(5)
- .Lsqr8x_zero_start:
- std 0,8*9(5)
- std 0,8*10(5)
- std 0,8*11(5)
- std 0,8*12(5)
- std 0,8*13(5)
- std 0,8*14(5)
- std 0,8*15(5)
- stdu 0,8*16(5)
- bne .Lsqr8x_zero
- std 3,8*6(1)
- std 18,8*7(1)
- std 7,8*8(1)
- std 5,8*9(1)
- std 0,8*10(1)
- addi 5,1,8*11
- .align 5
- .Lsqr8x_outer_loop:
- mulld 18,10,9
- mulld 19,11,9
- mulld 20,12,9
- mulld 21,14,9
- addc 23,23,18
- mulld 18,15,9
- adde 24,24,19
- mulld 19,16,9
- adde 25,25,20
- mulld 20,17,9
- adde 26,26,21
- mulhdu 21,10,9
- adde 27,27,18
- mulhdu 18,11,9
- adde 28,28,19
- mulhdu 19,12,9
- adde 29,29,20
- mulhdu 20,14,9
- std 22,8*1(5)
- addze 22,0
- std 23,8*2(5)
- addc 24,24,21
- mulhdu 21,15,9
- adde 25,25,18
- mulhdu 18,16,9
- adde 26,26,19
- mulhdu 19,17,9
- adde 27,27,20
- mulld 20,11,10
- adde 28,28,21
- mulld 21,12,10
- adde 29,29,18
- mulld 18,14,10
- adde 22,22,19
- mulld 19,15,10
- addc 25,25,20
- mulld 20,16,10
- adde 26,26,21
- mulld 21,17,10
- adde 27,27,18
- mulhdu 18,11,10
- adde 28,28,19
- mulhdu 19,12,10
- adde 29,29,20
- mulhdu 20,14,10
- adde 22,22,21
- mulhdu 21,15,10
- std 24,8*3(5)
- addze 23,0
- std 25,8*4(5)
- addc 26,26,18
- mulhdu 18,16,10
- adde 27,27,19
- mulhdu 19,17,10
- adde 28,28,20
- mulld 20,12,11
- adde 29,29,21
- mulld 21,14,11
- adde 22,22,18
- mulld 18,15,11
- adde 23,23,19
- mulld 19,16,11
- addc 27,27,20
- mulld 20,17,11
- adde 28,28,21
- mulhdu 21,12,11
- adde 29,29,18
- mulhdu 18,14,11
- adde 22,22,19
- mulhdu 19,15,11
- adde 23,23,20
- mulhdu 20,16,11
- std 26,8*5(5)
- addze 24,0
- std 27,8*6(5)
- addc 28,28,21
- mulhdu 21,17,11
- adde 29,29,18
- mulld 18,14,12
- adde 22,22,19
- mulld 19,15,12
- adde 23,23,20
- mulld 20,16,12
- adde 24,24,21
- mulld 21,17,12
- addc 29,29,18
- mulhdu 18,14,12
- adde 22,22,19
- mulhdu 19,15,12
- adde 23,23,20
- mulhdu 20,16,12
- adde 24,24,21
- mulhdu 21,17,12
- std 28,8*7(5)
- addze 25,0
- stdu 29,8*8(5)
- addc 22,22,18
- mulld 18,15,14
- adde 23,23,19
- mulld 19,16,14
- adde 24,24,20
- mulld 20,17,14
- adde 25,25,21
- mulhdu 21,15,14
- addc 23,23,18
- mulhdu 18,16,14
- adde 24,24,19
- mulhdu 19,17,14
- adde 25,25,20
- mulld 20,16,15
- addze 26,0
- addc 24,24,21
- mulld 21,17,15
- adde 25,25,18
- mulhdu 18,16,15
- adde 26,26,19
- mulhdu 19,17,15
- addc 25,25,20
- mulld 20,17,16
- adde 26,26,21
- mulhdu 21,17,16
- addze 27,0
- addc 26,26,18
- cmpld 6,4
- adde 27,27,19
- addc 27,27,20
- sub 18,6,8
- addze 28,0
- add 28,28,21
- beq .Lsqr8x_outer_break
- mr 7,9
- ld 9,8*1(5)
- ld 10,8*2(5)
- ld 11,8*3(5)
- ld 12,8*4(5)
- ld 14,8*5(5)
- ld 15,8*6(5)
- ld 16,8*7(5)
- ld 17,8*8(5)
- addc 22,22,9
- ld 9,8*1(4)
- adde 23,23,10
- ld 10,8*2(4)
- adde 24,24,11
- ld 11,8*3(4)
- adde 25,25,12
- ld 12,8*4(4)
- adde 26,26,14
- ld 14,8*5(4)
- adde 27,27,15
- ld 15,8*6(4)
- adde 28,28,16
- ld 16,8*7(4)
- subi 3,4,8*7
- addze 29,17
- ldu 17,8*8(4)
- li 30,0
- b .Lsqr8x_mul
- .align 5
- .Lsqr8x_mul:
- mulld 18,9,7
- addze 31,0
- mulld 19,10,7
- addi 30,30,8
- mulld 20,11,7
- andi. 30,30,8*8-1
- mulld 21,12,7
- addc 22,22,18
- mulld 18,14,7
- adde 23,23,19
- mulld 19,15,7
- adde 24,24,20
- mulld 20,16,7
- adde 25,25,21
- mulld 21,17,7
- adde 26,26,18
- mulhdu 18,9,7
- adde 27,27,19
- mulhdu 19,10,7
- adde 28,28,20
- mulhdu 20,11,7
- adde 29,29,21
- mulhdu 21,12,7
- addze 31,31
- stdu 22,8(5)
- addc 22,23,18
- mulhdu 18,14,7
- adde 23,24,19
- mulhdu 19,15,7
- adde 24,25,20
- mulhdu 20,16,7
- adde 25,26,21
- mulhdu 21,17,7
- ldx 7,3,30
- adde 26,27,18
- adde 27,28,19
- adde 28,29,20
- adde 29,31,21
- bne .Lsqr8x_mul
- cmpld 4,6
- beq .Lsqr8x_break
- ld 9,8*1(5)
- ld 10,8*2(5)
- ld 11,8*3(5)
- ld 12,8*4(5)
- ld 14,8*5(5)
- ld 15,8*6(5)
- ld 16,8*7(5)
- ld 17,8*8(5)
- addc 22,22,9
- ld 9,8*1(4)
- adde 23,23,10
- ld 10,8*2(4)
- adde 24,24,11
- ld 11,8*3(4)
- adde 25,25,12
- ld 12,8*4(4)
- adde 26,26,14
- ld 14,8*5(4)
- adde 27,27,15
- ld 15,8*6(4)
- adde 28,28,16
- ld 16,8*7(4)
- adde 29,29,17
- ldu 17,8*8(4)
- b .Lsqr8x_mul
- .align 5
- .Lsqr8x_break:
- ld 9,8*8(3)
- addi 4,3,8*15
- ld 10,8*9(3)
- sub. 18,6,4
- ld 11,8*10(3)
- sub 19,5,18
- ld 12,8*11(3)
- ld 14,8*12(3)
- ld 15,8*13(3)
- ld 16,8*14(3)
- ld 17,8*15(3)
- beq .Lsqr8x_outer_loop
- std 22,8*1(5)
- ld 22,8*1(19)
- std 23,8*2(5)
- ld 23,8*2(19)
- std 24,8*3(5)
- ld 24,8*3(19)
- std 25,8*4(5)
- ld 25,8*4(19)
- std 26,8*5(5)
- ld 26,8*5(19)
- std 27,8*6(5)
- ld 27,8*6(19)
- std 28,8*7(5)
- ld 28,8*7(19)
- std 29,8*8(5)
- ld 29,8*8(19)
- mr 5,19
- b .Lsqr8x_outer_loop
- .align 5
- .Lsqr8x_outer_break:
- ld 10,8*1(18)
- ld 12,8*2(18)
- ld 15,8*3(18)
- ld 17,8*4(18)
- addi 4,18,8*4
- ld 19,8*13(1)
- ld 20,8*14(1)
- ld 21,8*15(1)
- ld 18,8*16(1)
- std 22,8*1(5)
- srwi 30,8,5
- std 23,8*2(5)
- subi 30,30,1
- std 24,8*3(5)
- std 25,8*4(5)
- std 26,8*5(5)
- std 27,8*6(5)
- std 28,8*7(5)
- addi 5,1,8*11
- mulld 22,10,10
- mulhdu 10,10,10
- add 23,19,19
- srdi 19,19,64-1
- mulld 11,12,12
- mulhdu 12,12,12
- addc 23,23,10
- add 24,20,20
- srdi 20,20,64-1
- add 25,21,21
- srdi 21,21,64-1
- or 24,24,19
- mtctr 30
- .Lsqr4x_shift_n_add:
- mulld 14,15,15
- mulhdu 15,15,15
- ld 19,8*6(5)
- ld 10,8*1(4)
- adde 24,24,11
- add 26,18,18
- srdi 18,18,64-1
- or 25,25,20
- ld 20,8*7(5)
- adde 25,25,12
- ld 12,8*2(4)
- add 27,19,19
- srdi 19,19,64-1
- or 26,26,21
- ld 21,8*8(5)
- mulld 16,17,17
- mulhdu 17,17,17
- adde 26,26,14
- add 28,20,20
- srdi 20,20,64-1
- or 27,27,18
- ld 18,8*9(5)
- adde 27,27,15
- ld 15,8*3(4)
- add 29,21,21
- srdi 21,21,64-1
- or 28,28,19
- ld 19,8*10(5)
- mulld 9,10,10
- mulhdu 10,10,10
- adde 28,28,16
- std 22,8*1(5)
- add 22,18,18
- srdi 18,18,64-1
- or 29,29,20
- ld 20,8*11(5)
- adde 29,29,17
- ldu 17,8*4(4)
- std 23,8*2(5)
- add 23,19,19
- srdi 19,19,64-1
- or 22,22,21
- ld 21,8*12(5)
- mulld 11,12,12
- mulhdu 12,12,12
- adde 22,22,9
- std 24,8*3(5)
- add 24,20,20
- srdi 20,20,64-1
- or 23,23,18
- ld 18,8*13(5)
- adde 23,23,10
- std 25,8*4(5)
- std 26,8*5(5)
- std 27,8*6(5)
- std 28,8*7(5)
- stdu 29,8*8(5)
- add 25,21,21
- srdi 21,21,64-1
- or 24,24,19
- bdnz .Lsqr4x_shift_n_add
- ld 4,8*7(1)
- ld 7,8*8(1)
- mulld 14,15,15
- mulhdu 15,15,15
- std 22,8*1(5)
- ld 22,8*12(1)
- ld 19,8*6(5)
- adde 24,24,11
- add 26,18,18
- srdi 18,18,64-1
- or 25,25,20
- ld 20,8*7(5)
- adde 25,25,12
- add 27,19,19
- srdi 19,19,64-1
- or 26,26,21
- mulld 16,17,17
- mulhdu 17,17,17
- adde 26,26,14
- add 28,20,20
- srdi 20,20,64-1
- or 27,27,18
- std 23,8*2(5)
- ld 23,8*13(1)
- adde 27,27,15
- or 28,28,19
- ld 9,8*1(4)
- ld 10,8*2(4)
- adde 28,28,16
- ld 11,8*3(4)
- ld 12,8*4(4)
- adde 29,17,20
- ld 14,8*5(4)
- ld 15,8*6(4)
- mulld 31,7,22
- li 30,8
- ld 16,8*7(4)
- add 6,4,8
- ldu 17,8*8(4)
- std 24,8*3(5)
- ld 24,8*14(1)
- std 25,8*4(5)
- ld 25,8*15(1)
- std 26,8*5(5)
- ld 26,8*16(1)
- std 27,8*6(5)
- ld 27,8*17(1)
- std 28,8*7(5)
- ld 28,8*18(1)
- std 29,8*8(5)
- ld 29,8*19(1)
- addi 5,1,8*11
- mtctr 30
- b .Lsqr8x_reduction
- .align 5
- .Lsqr8x_reduction:
- mulld 19,10,31
- mulld 20,11,31
- stdu 31,8(5)
- mulld 21,12,31
- addic 22,22,-1
- mulld 18,14,31
- adde 22,23,19
- mulld 19,15,31
- adde 23,24,20
- mulld 20,16,31
- adde 24,25,21
- mulld 21,17,31
- adde 25,26,18
- mulhdu 18,9,31
- adde 26,27,19
- mulhdu 19,10,31
- adde 27,28,20
- mulhdu 20,11,31
- adde 28,29,21
- mulhdu 21,12,31
- addze 29,0
- addc 22,22,18
- mulhdu 18,14,31
- adde 23,23,19
- mulhdu 19,15,31
- adde 24,24,20
- mulhdu 20,16,31
- adde 25,25,21
- mulhdu 21,17,31
- mulld 31,7,22
- adde 26,26,18
- adde 27,27,19
- adde 28,28,20
- adde 29,29,21
- bdnz .Lsqr8x_reduction
- ld 18,8*1(5)
- ld 19,8*2(5)
- ld 20,8*3(5)
- ld 21,8*4(5)
- subi 3,5,8*7
- cmpld 6,4
- addc 22,22,18
- ld 18,8*5(5)
- adde 23,23,19
- ld 19,8*6(5)
- adde 24,24,20
- ld 20,8*7(5)
- adde 25,25,21
- ld 21,8*8(5)
- adde 26,26,18
- adde 27,27,19
- adde 28,28,20
- adde 29,29,21
- beq .Lsqr8x8_post_condition
- ld 7,8*0(3)
- ld 9,8*1(4)
- ld 10,8*2(4)
- ld 11,8*3(4)
- ld 12,8*4(4)
- ld 14,8*5(4)
- ld 15,8*6(4)
- ld 16,8*7(4)
- ldu 17,8*8(4)
- li 30,0
- .align 5
- .Lsqr8x_tail:
- mulld 18,9,7
- addze 31,0
- mulld 19,10,7
- addi 30,30,8
- mulld 20,11,7
- andi. 30,30,8*8-1
- mulld 21,12,7
- addc 22,22,18
- mulld 18,14,7
- adde 23,23,19
- mulld 19,15,7
- adde 24,24,20
- mulld 20,16,7
- adde 25,25,21
- mulld 21,17,7
- adde 26,26,18
- mulhdu 18,9,7
- adde 27,27,19
- mulhdu 19,10,7
- adde 28,28,20
- mulhdu 20,11,7
- adde 29,29,21
- mulhdu 21,12,7
- addze 31,31
- stdu 22,8(5)
- addc 22,23,18
- mulhdu 18,14,7
- adde 23,24,19
- mulhdu 19,15,7
- adde 24,25,20
- mulhdu 20,16,7
- adde 25,26,21
- mulhdu 21,17,7
- ldx 7,3,30
- adde 26,27,18
- adde 27,28,19
- adde 28,29,20
- adde 29,31,21
- bne .Lsqr8x_tail
- ld 9,8*1(5)
- ld 31,8*10(1)
- cmpld 6,4
- ld 10,8*2(5)
- sub 20,6,8
- ld 11,8*3(5)
- ld 12,8*4(5)
- ld 14,8*5(5)
- ld 15,8*6(5)
- ld 16,8*7(5)
- ld 17,8*8(5)
- beq .Lsqr8x_tail_break
- addc 22,22,9
- ld 9,8*1(4)
- adde 23,23,10
- ld 10,8*2(4)
- adde 24,24,11
- ld 11,8*3(4)
- adde 25,25,12
- ld 12,8*4(4)
- adde 26,26,14
- ld 14,8*5(4)
- adde 27,27,15
- ld 15,8*6(4)
- adde 28,28,16
- ld 16,8*7(4)
- adde 29,29,17
- ldu 17,8*8(4)
- b .Lsqr8x_tail
- .align 5
- .Lsqr8x_tail_break:
- ld 7,8*8(1)
- ld 21,8*9(1)
- addi 30,5,8*8
- addic 31,31,-1
- adde 18,22,9
- ld 22,8*8(3)
- ld 9,8*1(20)
- adde 19,23,10
- ld 23,8*9(3)
- ld 10,8*2(20)
- adde 24,24,11
- ld 11,8*3(20)
- adde 25,25,12
- ld 12,8*4(20)
- adde 26,26,14
- ld 14,8*5(20)
- adde 27,27,15
- ld 15,8*6(20)
- adde 28,28,16
- ld 16,8*7(20)
- adde 29,29,17
- ld 17,8*8(20)
- addi 4,20,8*8
- addze 20,0
- mulld 31,7,22
- std 18,8*1(5)
- cmpld 30,21
- std 19,8*2(5)
- li 30,8
- std 24,8*3(5)
- ld 24,8*10(3)
- std 25,8*4(5)
- ld 25,8*11(3)
- std 26,8*5(5)
- ld 26,8*12(3)
- std 27,8*6(5)
- ld 27,8*13(3)
- std 28,8*7(5)
- ld 28,8*14(3)
- std 29,8*8(5)
- ld 29,8*15(3)
- std 20,8*10(1)
- addi 5,3,8*7
- mtctr 30
- bne .Lsqr8x_reduction
- ld 3,8*6(1)
- srwi 30,8,6
- mr 7,5
- addi 5,5,8*8
- subi 30,30,1
- subfc 18,9,22
- subfe 19,10,23
- mr 31,20
- mr 6,3
- mtctr 30
- b .Lsqr8x_sub
- .align 5
- .Lsqr8x_sub:
- ld 9,8*1(4)
- ld 22,8*1(5)
- ld 10,8*2(4)
- ld 23,8*2(5)
- subfe 20,11,24
- ld 11,8*3(4)
- ld 24,8*3(5)
- subfe 21,12,25
- ld 12,8*4(4)
- ld 25,8*4(5)
- std 18,8*1(3)
- subfe 18,14,26
- ld 14,8*5(4)
- ld 26,8*5(5)
- std 19,8*2(3)
- subfe 19,15,27
- ld 15,8*6(4)
- ld 27,8*6(5)
- std 20,8*3(3)
- subfe 20,16,28
- ld 16,8*7(4)
- ld 28,8*7(5)
- std 21,8*4(3)
- subfe 21,17,29
- ldu 17,8*8(4)
- ldu 29,8*8(5)
- std 18,8*5(3)
- subfe 18,9,22
- std 19,8*6(3)
- subfe 19,10,23
- std 20,8*7(3)
- stdu 21,8*8(3)
- bdnz .Lsqr8x_sub
- srwi 30,8,5
- ld 9,8*1(6)
- ld 22,8*1(7)
- subi 30,30,1
- ld 10,8*2(6)
- ld 23,8*2(7)
- subfe 20,11,24
- ld 11,8*3(6)
- ld 24,8*3(7)
- subfe 21,12,25
- ld 12,8*4(6)
- ldu 25,8*4(7)
- std 18,8*1(3)
- subfe 18,14,26
- std 19,8*2(3)
- subfe 19,15,27
- std 20,8*3(3)
- subfe 20,16,28
- std 21,8*4(3)
- subfe 21,17,29
- std 18,8*5(3)
- subfe 31,0,31
- std 19,8*6(3)
- std 20,8*7(3)
- std 21,8*8(3)
- addi 5,1,8*11
- mtctr 30
- .Lsqr4x_cond_copy:
- andc 9,9,31
- std 0,-8*3(7)
- and 22,22,31
- std 0,-8*2(7)
- andc 10,10,31
- std 0,-8*1(7)
- and 23,23,31
- std 0,-8*0(7)
- andc 11,11,31
- std 0,8*1(5)
- and 24,24,31
- std 0,8*2(5)
- andc 12,12,31
- std 0,8*3(5)
- and 25,25,31
- stdu 0,8*4(5)
- or 18,9,22
- ld 9,8*5(6)
- ld 22,8*1(7)
- or 19,10,23
- ld 10,8*6(6)
- ld 23,8*2(7)
- or 20,11,24
- ld 11,8*7(6)
- ld 24,8*3(7)
- or 21,12,25
- ld 12,8*8(6)
- ldu 25,8*4(7)
- std 18,8*1(6)
- std 19,8*2(6)
- std 20,8*3(6)
- stdu 21,8*4(6)
- bdnz .Lsqr4x_cond_copy
- ld 4,0(1)
- andc 9,9,31
- and 22,22,31
- andc 10,10,31
- and 23,23,31
- andc 11,11,31
- and 24,24,31
- andc 12,12,31
- and 25,25,31
- or 18,9,22
- or 19,10,23
- or 20,11,24
- or 21,12,25
- std 18,8*1(6)
- std 19,8*2(6)
- std 20,8*3(6)
- std 21,8*4(6)
- b .Lsqr8x_done
- .align 5
- .Lsqr8x8_post_condition:
- ld 3,8*6(1)
- ld 4,0(1)
- addze 31,0
- subfc 22,9,22
- subfe 23,10,23
- std 0,8*12(1)
- std 0,8*13(1)
- subfe 24,11,24
- std 0,8*14(1)
- std 0,8*15(1)
- subfe 25,12,25
- std 0,8*16(1)
- std 0,8*17(1)
- subfe 26,14,26
- std 0,8*18(1)
- std 0,8*19(1)
- subfe 27,15,27
- std 0,8*20(1)
- std 0,8*21(1)
- subfe 28,16,28
- std 0,8*22(1)
- std 0,8*23(1)
- subfe 29,17,29
- std 0,8*24(1)
- std 0,8*25(1)
- subfe 31,0,31
- std 0,8*26(1)
- std 0,8*27(1)
- and 9,9,31
- and 10,10,31
- addc 22,22,9
- and 11,11,31
- adde 23,23,10
- and 12,12,31
- adde 24,24,11
- and 14,14,31
- adde 25,25,12
- and 15,15,31
- adde 26,26,14
- and 16,16,31
- adde 27,27,15
- and 17,17,31
- adde 28,28,16
- adde 29,29,17
- std 22,8*1(3)
- std 23,8*2(3)
- std 24,8*3(3)
- std 25,8*4(3)
- std 26,8*5(3)
- std 27,8*6(3)
- std 28,8*7(3)
- std 29,8*8(3)
- .Lsqr8x_done:
- std 0,8*8(1)
- std 0,8*10(1)
- ld 14,-8*18(4)
- li 3,1
- ld 15,-8*17(4)
- ld 16,-8*16(4)
- ld 17,-8*15(4)
- ld 18,-8*14(4)
- ld 19,-8*13(4)
- ld 20,-8*12(4)
- ld 21,-8*11(4)
- ld 22,-8*10(4)
- ld 23,-8*9(4)
- ld 24,-8*8(4)
- ld 25,-8*7(4)
- ld 26,-8*6(4)
- ld 27,-8*5(4)
- ld 28,-8*4(4)
- ld 29,-8*3(4)
- ld 30,-8*2(4)
- ld 31,-8*1(4)
- mr 1,4
- blr
- .long 0
- .byte 0,12,4,0x20,0x80,18,6,0
- .long 0
- .size __bn_sqr8x_mont,.-__bn_sqr8x_mont
- .byte 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105,112,108,105,99,97,116,105,111,110,32,102,111,114,32,80,80,67,44,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0
- .align 2
|