123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824 |
- // Copyright © 2021 Jeffrey H. Johnson <trnsz@pobox.com>.
- // Copyright © 2021 Filippo Valsorda.
- // Copyright © 2012 The Go Authors.
- //
- // All rights reserved.
- // Use of this source code is governed by the BSD-style
- // license that can be found in the LICENSE file.
- // This code was translated into a form compatible with
- // Go's 6a from the public domain sources in SUPERCOP:
- // https://bench.cr.yp.to/supercop.html
- // +build amd64,gc,!purego
- #include "textflag.h"
- #define REDMASK51 0x0007FFFFFFFFFFFF
- // These constants cannot be encoded in non-MOVQ immediates.
- // We access them directly from memory instead.
- DATA ·_121666_213(SB)/8, $996687872
- GLOBL ·_121666_213(SB), 8, $8
- DATA ·_2P0(SB)/8, $0xFFFFFFFFFFFDA
- GLOBL ·_2P0(SB), 8, $8
- DATA ·_2P1234(SB)/8, $0xFFFFFFFFFFFFE
- GLOBL ·_2P1234(SB), 8, $8
- // func freeze(inout *[5]uint64)
- TEXT ·freeze(SB), 7, $0-8
- MOVQ inout+0(FP), DI
- MOVQ 0(DI), SI
- MOVQ 8(DI), DX
- MOVQ 16(DI), CX
- MOVQ 24(DI), R8
- MOVQ 32(DI), R9
- MOVQ $REDMASK51, AX
- MOVQ AX, R10
- SUBQ $18, R10
- MOVQ $3, R11
- REDUCELOOP:
- MOVQ SI, R12
- SHRQ $51, R12
- ANDQ AX, SI
- ADDQ R12, DX
- MOVQ DX, R12
- SHRQ $51, R12
- ANDQ AX, DX
- ADDQ R12, CX
- MOVQ CX, R12
- SHRQ $51, R12
- ANDQ AX, CX
- ADDQ R12, R8
- MOVQ R8, R12
- SHRQ $51, R12
- ANDQ AX, R8
- ADDQ R12, R9
- MOVQ R9, R12
- SHRQ $51, R12
- ANDQ AX, R9
- IMUL3Q $19, R12, R12
- ADDQ R12, SI
- SUBQ $1, R11
- JA REDUCELOOP
- MOVQ $1, R12
- CMPQ R10, SI
- CMOVQLT R11, R12
- CMPQ AX, DX
- CMOVQNE R11, R12
- CMPQ AX, CX
- CMOVQNE R11, R12
- CMPQ AX, R8
- CMOVQNE R11, R12
- CMPQ AX, R9
- CMOVQNE R11, R12
- NEGQ R12
- ANDQ R12, AX
- ANDQ R12, R10
- SUBQ R10, SI
- SUBQ AX, DX
- SUBQ AX, CX
- SUBQ AX, R8
- SUBQ AX, R9
- MOVQ SI, 0(DI)
- MOVQ DX, 8(DI)
- MOVQ CX, 16(DI)
- MOVQ R8, 24(DI)
- MOVQ R9, 32(DI)
- RET
- // func ladderstep(inout *[5][5]uint64)
- TEXT ·ladderstep(SB), 0, $296-8
- MOVQ inout+0(FP), DI
- MOVQ 40(DI), SI
- MOVQ 48(DI), DX
- MOVQ 56(DI), CX
- MOVQ 64(DI), R8
- MOVQ 72(DI), R9
- MOVQ SI, AX
- MOVQ DX, R10
- MOVQ CX, R11
- MOVQ R8, R12
- MOVQ R9, R13
- ADDQ ·_2P0(SB), AX
- ADDQ ·_2P1234(SB), R10
- ADDQ ·_2P1234(SB), R11
- ADDQ ·_2P1234(SB), R12
- ADDQ ·_2P1234(SB), R13
- ADDQ 80(DI), SI
- ADDQ 88(DI), DX
- ADDQ 96(DI), CX
- ADDQ 104(DI), R8
- ADDQ 112(DI), R9
- SUBQ 80(DI), AX
- SUBQ 88(DI), R10
- SUBQ 96(DI), R11
- SUBQ 104(DI), R12
- SUBQ 112(DI), R13
- MOVQ SI, 0(SP)
- MOVQ DX, 8(SP)
- MOVQ CX, 16(SP)
- MOVQ R8, 24(SP)
- MOVQ R9, 32(SP)
- MOVQ AX, 40(SP)
- MOVQ R10, 48(SP)
- MOVQ R11, 56(SP)
- MOVQ R12, 64(SP)
- MOVQ R13, 72(SP)
- MOVQ 40(SP), AX
- MULQ 40(SP)
- MOVQ AX, SI
- MOVQ DX, CX
- MOVQ 40(SP), AX
- SHLQ $1, AX
- MULQ 48(SP)
- MOVQ AX, R8
- MOVQ DX, R9
- MOVQ 40(SP), AX
- SHLQ $1, AX
- MULQ 56(SP)
- MOVQ AX, R10
- MOVQ DX, R11
- MOVQ 40(SP), AX
- SHLQ $1, AX
- MULQ 64(SP)
- MOVQ AX, R12
- MOVQ DX, R13
- MOVQ 40(SP), AX
- SHLQ $1, AX
- MULQ 72(SP)
- MOVQ AX, R14
- MOVQ DX, R15
- MOVQ 48(SP), AX
- MULQ 48(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 48(SP), AX
- SHLQ $1, AX
- MULQ 56(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 48(SP), AX
- SHLQ $1, AX
- MULQ 64(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 48(SP), DX
- IMUL3Q $38, DX, AX
- MULQ 72(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 56(SP), AX
- MULQ 56(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 56(SP), DX
- IMUL3Q $38, DX, AX
- MULQ 64(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 56(SP), DX
- IMUL3Q $38, DX, AX
- MULQ 72(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 64(SP), DX
- IMUL3Q $19, DX, AX
- MULQ 64(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 64(SP), DX
- IMUL3Q $38, DX, AX
- MULQ 72(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 72(SP), DX
- IMUL3Q $19, DX, AX
- MULQ 72(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ $REDMASK51, DX
- SHLQ $13, SI, CX
- ANDQ DX, SI
- SHLQ $13, R8, R9
- ANDQ DX, R8
- ADDQ CX, R8
- SHLQ $13, R10, R11
- ANDQ DX, R10
- ADDQ R9, R10
- SHLQ $13, R12, R13
- ANDQ DX, R12
- ADDQ R11, R12
- SHLQ $13, R14, R15
- ANDQ DX, R14
- ADDQ R13, R14
- IMUL3Q $19, R15, CX
- ADDQ CX, SI
- MOVQ SI, CX
- SHRQ $51, CX
- ADDQ R8, CX
- ANDQ DX, SI
- MOVQ CX, R8
- SHRQ $51, CX
- ADDQ R10, CX
- ANDQ DX, R8
- MOVQ CX, R9
- SHRQ $51, CX
- ADDQ R12, CX
- ANDQ DX, R9
- MOVQ CX, AX
- SHRQ $51, CX
- ADDQ R14, CX
- ANDQ DX, AX
- MOVQ CX, R10
- SHRQ $51, CX
- IMUL3Q $19, CX, CX
- ADDQ CX, SI
- ANDQ DX, R10
- MOVQ SI, 80(SP)
- MOVQ R8, 88(SP)
- MOVQ R9, 96(SP)
- MOVQ AX, 104(SP)
- MOVQ R10, 112(SP)
- MOVQ 0(SP), AX
- MULQ 0(SP)
- MOVQ AX, SI
- MOVQ DX, CX
- MOVQ 0(SP), AX
- SHLQ $1, AX
- MULQ 8(SP)
- MOVQ AX, R8
- MOVQ DX, R9
- MOVQ 0(SP), AX
- SHLQ $1, AX
- MULQ 16(SP)
- MOVQ AX, R10
- MOVQ DX, R11
- MOVQ 0(SP), AX
- SHLQ $1, AX
- MULQ 24(SP)
- MOVQ AX, R12
- MOVQ DX, R13
- MOVQ 0(SP), AX
- SHLQ $1, AX
- MULQ 32(SP)
- MOVQ AX, R14
- MOVQ DX, R15
- MOVQ 8(SP), AX
- MULQ 8(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 8(SP), AX
- SHLQ $1, AX
- MULQ 16(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 8(SP), AX
- SHLQ $1, AX
- MULQ 24(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 8(SP), DX
- IMUL3Q $38, DX, AX
- MULQ 32(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 16(SP), AX
- MULQ 16(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 16(SP), DX
- IMUL3Q $38, DX, AX
- MULQ 24(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 16(SP), DX
- IMUL3Q $38, DX, AX
- MULQ 32(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 24(SP), DX
- IMUL3Q $19, DX, AX
- MULQ 24(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 24(SP), DX
- IMUL3Q $38, DX, AX
- MULQ 32(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 32(SP), DX
- IMUL3Q $19, DX, AX
- MULQ 32(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ $REDMASK51, DX
- SHLQ $13, SI, CX
- ANDQ DX, SI
- SHLQ $13, R8, R9
- ANDQ DX, R8
- ADDQ CX, R8
- SHLQ $13, R10, R11
- ANDQ DX, R10
- ADDQ R9, R10
- SHLQ $13, R12, R13
- ANDQ DX, R12
- ADDQ R11, R12
- SHLQ $13, R14, R15
- ANDQ DX, R14
- ADDQ R13, R14
- IMUL3Q $19, R15, CX
- ADDQ CX, SI
- MOVQ SI, CX
- SHRQ $51, CX
- ADDQ R8, CX
- ANDQ DX, SI
- MOVQ CX, R8
- SHRQ $51, CX
- ADDQ R10, CX
- ANDQ DX, R8
- MOVQ CX, R9
- SHRQ $51, CX
- ADDQ R12, CX
- ANDQ DX, R9
- MOVQ CX, AX
- SHRQ $51, CX
- ADDQ R14, CX
- ANDQ DX, AX
- MOVQ CX, R10
- SHRQ $51, CX
- IMUL3Q $19, CX, CX
- ADDQ CX, SI
- ANDQ DX, R10
- MOVQ SI, 120(SP)
- MOVQ R8, 128(SP)
- MOVQ R9, 136(SP)
- MOVQ AX, 144(SP)
- MOVQ R10, 152(SP)
- MOVQ SI, SI
- MOVQ R8, DX
- MOVQ R9, CX
- MOVQ AX, R8
- MOVQ R10, R9
- ADDQ ·_2P0(SB), SI
- ADDQ ·_2P1234(SB), DX
- ADDQ ·_2P1234(SB), CX
- ADDQ ·_2P1234(SB), R8
- ADDQ ·_2P1234(SB), R9
- SUBQ 80(SP), SI
- SUBQ 88(SP), DX
- SUBQ 96(SP), CX
- SUBQ 104(SP), R8
- SUBQ 112(SP), R9
- MOVQ SI, 160(SP)
- MOVQ DX, 168(SP)
- MOVQ CX, 176(SP)
- MOVQ R8, 184(SP)
- MOVQ R9, 192(SP)
- MOVQ 120(DI), SI
- MOVQ 128(DI), DX
- MOVQ 136(DI), CX
- MOVQ 144(DI), R8
- MOVQ 152(DI), R9
- MOVQ SI, AX
- MOVQ DX, R10
- MOVQ CX, R11
- MOVQ R8, R12
- MOVQ R9, R13
- ADDQ ·_2P0(SB), AX
- ADDQ ·_2P1234(SB), R10
- ADDQ ·_2P1234(SB), R11
- ADDQ ·_2P1234(SB), R12
- ADDQ ·_2P1234(SB), R13
- ADDQ 160(DI), SI
- ADDQ 168(DI), DX
- ADDQ 176(DI), CX
- ADDQ 184(DI), R8
- ADDQ 192(DI), R9
- SUBQ 160(DI), AX
- SUBQ 168(DI), R10
- SUBQ 176(DI), R11
- SUBQ 184(DI), R12
- SUBQ 192(DI), R13
- MOVQ SI, 200(SP)
- MOVQ DX, 208(SP)
- MOVQ CX, 216(SP)
- MOVQ R8, 224(SP)
- MOVQ R9, 232(SP)
- MOVQ AX, 240(SP)
- MOVQ R10, 248(SP)
- MOVQ R11, 256(SP)
- MOVQ R12, 264(SP)
- MOVQ R13, 272(SP)
- MOVQ 224(SP), SI
- IMUL3Q $19, SI, AX
- MOVQ AX, 280(SP)
- MULQ 56(SP)
- MOVQ AX, SI
- MOVQ DX, CX
- MOVQ 232(SP), DX
- IMUL3Q $19, DX, AX
- MOVQ AX, 288(SP)
- MULQ 48(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 200(SP), AX
- MULQ 40(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 200(SP), AX
- MULQ 48(SP)
- MOVQ AX, R8
- MOVQ DX, R9
- MOVQ 200(SP), AX
- MULQ 56(SP)
- MOVQ AX, R10
- MOVQ DX, R11
- MOVQ 200(SP), AX
- MULQ 64(SP)
- MOVQ AX, R12
- MOVQ DX, R13
- MOVQ 200(SP), AX
- MULQ 72(SP)
- MOVQ AX, R14
- MOVQ DX, R15
- MOVQ 208(SP), AX
- MULQ 40(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 208(SP), AX
- MULQ 48(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 208(SP), AX
- MULQ 56(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 208(SP), AX
- MULQ 64(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 208(SP), DX
- IMUL3Q $19, DX, AX
- MULQ 72(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 216(SP), AX
- MULQ 40(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 216(SP), AX
- MULQ 48(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 216(SP), AX
- MULQ 56(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 216(SP), DX
- IMUL3Q $19, DX, AX
- MULQ 64(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 216(SP), DX
- IMUL3Q $19, DX, AX
- MULQ 72(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 224(SP), AX
- MULQ 40(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 224(SP), AX
- MULQ 48(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 280(SP), AX
- MULQ 64(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 280(SP), AX
- MULQ 72(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 232(SP), AX
- MULQ 40(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 288(SP), AX
- MULQ 56(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 288(SP), AX
- MULQ 64(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 288(SP), AX
- MULQ 72(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ $REDMASK51, DX
- SHLQ $13, SI, CX
- ANDQ DX, SI
- SHLQ $13, R8, R9
- ANDQ DX, R8
- ADDQ CX, R8
- SHLQ $13, R10, R11
- ANDQ DX, R10
- ADDQ R9, R10
- SHLQ $13, R12, R13
- ANDQ DX, R12
- ADDQ R11, R12
- SHLQ $13, R14, R15
- ANDQ DX, R14
- ADDQ R13, R14
- IMUL3Q $19, R15, CX
- ADDQ CX, SI
- MOVQ SI, CX
- SHRQ $51, CX
- ADDQ R8, CX
- MOVQ CX, R8
- SHRQ $51, CX
- ANDQ DX, SI
- ADDQ R10, CX
- MOVQ CX, R9
- SHRQ $51, CX
- ANDQ DX, R8
- ADDQ R12, CX
- MOVQ CX, AX
- SHRQ $51, CX
- ANDQ DX, R9
- ADDQ R14, CX
- MOVQ CX, R10
- SHRQ $51, CX
- ANDQ DX, AX
- IMUL3Q $19, CX, CX
- ADDQ CX, SI
- ANDQ DX, R10
- MOVQ SI, 40(SP)
- MOVQ R8, 48(SP)
- MOVQ R9, 56(SP)
- MOVQ AX, 64(SP)
- MOVQ R10, 72(SP)
- MOVQ 264(SP), SI
- IMUL3Q $19, SI, AX
- MOVQ AX, 200(SP)
- MULQ 16(SP)
- MOVQ AX, SI
- MOVQ DX, CX
- MOVQ 272(SP), DX
- IMUL3Q $19, DX, AX
- MOVQ AX, 208(SP)
- MULQ 8(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 240(SP), AX
- MULQ 0(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 240(SP), AX
- MULQ 8(SP)
- MOVQ AX, R8
- MOVQ DX, R9
- MOVQ 240(SP), AX
- MULQ 16(SP)
- MOVQ AX, R10
- MOVQ DX, R11
- MOVQ 240(SP), AX
- MULQ 24(SP)
- MOVQ AX, R12
- MOVQ DX, R13
- MOVQ 240(SP), AX
- MULQ 32(SP)
- MOVQ AX, R14
- MOVQ DX, R15
- MOVQ 248(SP), AX
- MULQ 0(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 248(SP), AX
- MULQ 8(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 248(SP), AX
- MULQ 16(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 248(SP), AX
- MULQ 24(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 248(SP), DX
- IMUL3Q $19, DX, AX
- MULQ 32(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 256(SP), AX
- MULQ 0(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 256(SP), AX
- MULQ 8(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 256(SP), AX
- MULQ 16(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 256(SP), DX
- IMUL3Q $19, DX, AX
- MULQ 24(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 256(SP), DX
- IMUL3Q $19, DX, AX
- MULQ 32(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 264(SP), AX
- MULQ 0(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 264(SP), AX
- MULQ 8(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 200(SP), AX
- MULQ 24(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 200(SP), AX
- MULQ 32(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 272(SP), AX
- MULQ 0(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 208(SP), AX
- MULQ 16(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 208(SP), AX
- MULQ 24(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 208(SP), AX
- MULQ 32(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ $REDMASK51, DX
- SHLQ $13, SI, CX
- ANDQ DX, SI
- SHLQ $13, R8, R9
- ANDQ DX, R8
- ADDQ CX, R8
- SHLQ $13, R10, R11
- ANDQ DX, R10
- ADDQ R9, R10
- SHLQ $13, R12, R13
- ANDQ DX, R12
- ADDQ R11, R12
- SHLQ $13, R14, R15
- ANDQ DX, R14
- ADDQ R13, R14
- IMUL3Q $19, R15, CX
- ADDQ CX, SI
- MOVQ SI, CX
- SHRQ $51, CX
- ADDQ R8, CX
- MOVQ CX, R8
- SHRQ $51, CX
- ANDQ DX, SI
- ADDQ R10, CX
- MOVQ CX, R9
- SHRQ $51, CX
- ANDQ DX, R8
- ADDQ R12, CX
- MOVQ CX, AX
- SHRQ $51, CX
- ANDQ DX, R9
- ADDQ R14, CX
- MOVQ CX, R10
- SHRQ $51, CX
- ANDQ DX, AX
- IMUL3Q $19, CX, CX
- ADDQ CX, SI
- ANDQ DX, R10
- MOVQ SI, DX
- MOVQ R8, CX
- MOVQ R9, R11
- MOVQ AX, R12
- MOVQ R10, R13
- ADDQ ·_2P0(SB), DX
- ADDQ ·_2P1234(SB), CX
- ADDQ ·_2P1234(SB), R11
- ADDQ ·_2P1234(SB), R12
- ADDQ ·_2P1234(SB), R13
- ADDQ 40(SP), SI
- ADDQ 48(SP), R8
- ADDQ 56(SP), R9
- ADDQ 64(SP), AX
- ADDQ 72(SP), R10
- SUBQ 40(SP), DX
- SUBQ 48(SP), CX
- SUBQ 56(SP), R11
- SUBQ 64(SP), R12
- SUBQ 72(SP), R13
- MOVQ SI, 120(DI)
- MOVQ R8, 128(DI)
- MOVQ R9, 136(DI)
- MOVQ AX, 144(DI)
- MOVQ R10, 152(DI)
- MOVQ DX, 160(DI)
- MOVQ CX, 168(DI)
- MOVQ R11, 176(DI)
- MOVQ R12, 184(DI)
- MOVQ R13, 192(DI)
- MOVQ 120(DI), AX
- MULQ 120(DI)
- MOVQ AX, SI
- MOVQ DX, CX
- MOVQ 120(DI), AX
- SHLQ $1, AX
- MULQ 128(DI)
- MOVQ AX, R8
- MOVQ DX, R9
- MOVQ 120(DI), AX
- SHLQ $1, AX
- MULQ 136(DI)
- MOVQ AX, R10
- MOVQ DX, R11
- MOVQ 120(DI), AX
- SHLQ $1, AX
- MULQ 144(DI)
- MOVQ AX, R12
- MOVQ DX, R13
- MOVQ 120(DI), AX
- SHLQ $1, AX
- MULQ 152(DI)
- MOVQ AX, R14
- MOVQ DX, R15
- MOVQ 128(DI), AX
- MULQ 128(DI)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 128(DI), AX
- SHLQ $1, AX
- MULQ 136(DI)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 128(DI), AX
- SHLQ $1, AX
- MULQ 144(DI)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 128(DI), DX
- IMUL3Q $38, DX, AX
- MULQ 152(DI)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 136(DI), AX
- MULQ 136(DI)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 136(DI), DX
- IMUL3Q $38, DX, AX
- MULQ 144(DI)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 136(DI), DX
- IMUL3Q $38, DX, AX
- MULQ 152(DI)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 144(DI), DX
- IMUL3Q $19, DX, AX
- MULQ 144(DI)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 144(DI), DX
- IMUL3Q $38, DX, AX
- MULQ 152(DI)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 152(DI), DX
- IMUL3Q $19, DX, AX
- MULQ 152(DI)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ $REDMASK51, DX
- SHLQ $13, SI, CX
- ANDQ DX, SI
- SHLQ $13, R8, R9
- ANDQ DX, R8
- ADDQ CX, R8
- SHLQ $13, R10, R11
- ANDQ DX, R10
- ADDQ R9, R10
- SHLQ $13, R12, R13
- ANDQ DX, R12
- ADDQ R11, R12
- SHLQ $13, R14, R15
- ANDQ DX, R14
- ADDQ R13, R14
- IMUL3Q $19, R15, CX
- ADDQ CX, SI
- MOVQ SI, CX
- SHRQ $51, CX
- ADDQ R8, CX
- ANDQ DX, SI
- MOVQ CX, R8
- SHRQ $51, CX
- ADDQ R10, CX
- ANDQ DX, R8
- MOVQ CX, R9
- SHRQ $51, CX
- ADDQ R12, CX
- ANDQ DX, R9
- MOVQ CX, AX
- SHRQ $51, CX
- ADDQ R14, CX
- ANDQ DX, AX
- MOVQ CX, R10
- SHRQ $51, CX
- IMUL3Q $19, CX, CX
- ADDQ CX, SI
- ANDQ DX, R10
- MOVQ SI, 120(DI)
- MOVQ R8, 128(DI)
- MOVQ R9, 136(DI)
- MOVQ AX, 144(DI)
- MOVQ R10, 152(DI)
- MOVQ 160(DI), AX
- MULQ 160(DI)
- MOVQ AX, SI
- MOVQ DX, CX
- MOVQ 160(DI), AX
- SHLQ $1, AX
- MULQ 168(DI)
- MOVQ AX, R8
- MOVQ DX, R9
- MOVQ 160(DI), AX
- SHLQ $1, AX
- MULQ 176(DI)
- MOVQ AX, R10
- MOVQ DX, R11
- MOVQ 160(DI), AX
- SHLQ $1, AX
- MULQ 184(DI)
- MOVQ AX, R12
- MOVQ DX, R13
- MOVQ 160(DI), AX
- SHLQ $1, AX
- MULQ 192(DI)
- MOVQ AX, R14
- MOVQ DX, R15
- MOVQ 168(DI), AX
- MULQ 168(DI)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 168(DI), AX
- SHLQ $1, AX
- MULQ 176(DI)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 168(DI), AX
- SHLQ $1, AX
- MULQ 184(DI)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 168(DI), DX
- IMUL3Q $38, DX, AX
- MULQ 192(DI)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 176(DI), AX
- MULQ 176(DI)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 176(DI), DX
- IMUL3Q $38, DX, AX
- MULQ 184(DI)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 176(DI), DX
- IMUL3Q $38, DX, AX
- MULQ 192(DI)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 184(DI), DX
- IMUL3Q $19, DX, AX
- MULQ 184(DI)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 184(DI), DX
- IMUL3Q $38, DX, AX
- MULQ 192(DI)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 192(DI), DX
- IMUL3Q $19, DX, AX
- MULQ 192(DI)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ $REDMASK51, DX
- SHLQ $13, SI, CX
- ANDQ DX, SI
- SHLQ $13, R8, R9
- ANDQ DX, R8
- ADDQ CX, R8
- SHLQ $13, R10, R11
- ANDQ DX, R10
- ADDQ R9, R10
- SHLQ $13, R12, R13
- ANDQ DX, R12
- ADDQ R11, R12
- SHLQ $13, R14, R15
- ANDQ DX, R14
- ADDQ R13, R14
- IMUL3Q $19, R15, CX
- ADDQ CX, SI
- MOVQ SI, CX
- SHRQ $51, CX
- ADDQ R8, CX
- ANDQ DX, SI
- MOVQ CX, R8
- SHRQ $51, CX
- ADDQ R10, CX
- ANDQ DX, R8
- MOVQ CX, R9
- SHRQ $51, CX
- ADDQ R12, CX
- ANDQ DX, R9
- MOVQ CX, AX
- SHRQ $51, CX
- ADDQ R14, CX
- ANDQ DX, AX
- MOVQ CX, R10
- SHRQ $51, CX
- IMUL3Q $19, CX, CX
- ADDQ CX, SI
- ANDQ DX, R10
- MOVQ SI, 160(DI)
- MOVQ R8, 168(DI)
- MOVQ R9, 176(DI)
- MOVQ AX, 184(DI)
- MOVQ R10, 192(DI)
- MOVQ 184(DI), SI
- IMUL3Q $19, SI, AX
- MOVQ AX, 0(SP)
- MULQ 16(DI)
- MOVQ AX, SI
- MOVQ DX, CX
- MOVQ 192(DI), DX
- IMUL3Q $19, DX, AX
- MOVQ AX, 8(SP)
- MULQ 8(DI)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 160(DI), AX
- MULQ 0(DI)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 160(DI), AX
- MULQ 8(DI)
- MOVQ AX, R8
- MOVQ DX, R9
- MOVQ 160(DI), AX
- MULQ 16(DI)
- MOVQ AX, R10
- MOVQ DX, R11
- MOVQ 160(DI), AX
- MULQ 24(DI)
- MOVQ AX, R12
- MOVQ DX, R13
- MOVQ 160(DI), AX
- MULQ 32(DI)
- MOVQ AX, R14
- MOVQ DX, R15
- MOVQ 168(DI), AX
- MULQ 0(DI)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 168(DI), AX
- MULQ 8(DI)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 168(DI), AX
- MULQ 16(DI)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 168(DI), AX
- MULQ 24(DI)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 168(DI), DX
- IMUL3Q $19, DX, AX
- MULQ 32(DI)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 176(DI), AX
- MULQ 0(DI)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 176(DI), AX
- MULQ 8(DI)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 176(DI), AX
- MULQ 16(DI)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 176(DI), DX
- IMUL3Q $19, DX, AX
- MULQ 24(DI)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 176(DI), DX
- IMUL3Q $19, DX, AX
- MULQ 32(DI)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 184(DI), AX
- MULQ 0(DI)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 184(DI), AX
- MULQ 8(DI)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 0(SP), AX
- MULQ 24(DI)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 0(SP), AX
- MULQ 32(DI)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 192(DI), AX
- MULQ 0(DI)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 8(SP), AX
- MULQ 16(DI)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 8(SP), AX
- MULQ 24(DI)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 8(SP), AX
- MULQ 32(DI)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ $REDMASK51, DX
- SHLQ $13, SI, CX
- ANDQ DX, SI
- SHLQ $13, R8, R9
- ANDQ DX, R8
- ADDQ CX, R8
- SHLQ $13, R10, R11
- ANDQ DX, R10
- ADDQ R9, R10
- SHLQ $13, R12, R13
- ANDQ DX, R12
- ADDQ R11, R12
- SHLQ $13, R14, R15
- ANDQ DX, R14
- ADDQ R13, R14
- IMUL3Q $19, R15, CX
- ADDQ CX, SI
- MOVQ SI, CX
- SHRQ $51, CX
- ADDQ R8, CX
- MOVQ CX, R8
- SHRQ $51, CX
- ANDQ DX, SI
- ADDQ R10, CX
- MOVQ CX, R9
- SHRQ $51, CX
- ANDQ DX, R8
- ADDQ R12, CX
- MOVQ CX, AX
- SHRQ $51, CX
- ANDQ DX, R9
- ADDQ R14, CX
- MOVQ CX, R10
- SHRQ $51, CX
- ANDQ DX, AX
- IMUL3Q $19, CX, CX
- ADDQ CX, SI
- ANDQ DX, R10
- MOVQ SI, 160(DI)
- MOVQ R8, 168(DI)
- MOVQ R9, 176(DI)
- MOVQ AX, 184(DI)
- MOVQ R10, 192(DI)
- MOVQ 144(SP), SI
- IMUL3Q $19, SI, AX
- MOVQ AX, 0(SP)
- MULQ 96(SP)
- MOVQ AX, SI
- MOVQ DX, CX
- MOVQ 152(SP), DX
- IMUL3Q $19, DX, AX
- MOVQ AX, 8(SP)
- MULQ 88(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 120(SP), AX
- MULQ 80(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 120(SP), AX
- MULQ 88(SP)
- MOVQ AX, R8
- MOVQ DX, R9
- MOVQ 120(SP), AX
- MULQ 96(SP)
- MOVQ AX, R10
- MOVQ DX, R11
- MOVQ 120(SP), AX
- MULQ 104(SP)
- MOVQ AX, R12
- MOVQ DX, R13
- MOVQ 120(SP), AX
- MULQ 112(SP)
- MOVQ AX, R14
- MOVQ DX, R15
- MOVQ 128(SP), AX
- MULQ 80(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 128(SP), AX
- MULQ 88(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 128(SP), AX
- MULQ 96(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 128(SP), AX
- MULQ 104(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 128(SP), DX
- IMUL3Q $19, DX, AX
- MULQ 112(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 136(SP), AX
- MULQ 80(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 136(SP), AX
- MULQ 88(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 136(SP), AX
- MULQ 96(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 136(SP), DX
- IMUL3Q $19, DX, AX
- MULQ 104(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 136(SP), DX
- IMUL3Q $19, DX, AX
- MULQ 112(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 144(SP), AX
- MULQ 80(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 144(SP), AX
- MULQ 88(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 0(SP), AX
- MULQ 104(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 0(SP), AX
- MULQ 112(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 152(SP), AX
- MULQ 80(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 8(SP), AX
- MULQ 96(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 8(SP), AX
- MULQ 104(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 8(SP), AX
- MULQ 112(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ $REDMASK51, DX
- SHLQ $13, SI, CX
- ANDQ DX, SI
- SHLQ $13, R8, R9
- ANDQ DX, R8
- ADDQ CX, R8
- SHLQ $13, R10, R11
- ANDQ DX, R10
- ADDQ R9, R10
- SHLQ $13, R12, R13
- ANDQ DX, R12
- ADDQ R11, R12
- SHLQ $13, R14, R15
- ANDQ DX, R14
- ADDQ R13, R14
- IMUL3Q $19, R15, CX
- ADDQ CX, SI
- MOVQ SI, CX
- SHRQ $51, CX
- ADDQ R8, CX
- MOVQ CX, R8
- SHRQ $51, CX
- ANDQ DX, SI
- ADDQ R10, CX
- MOVQ CX, R9
- SHRQ $51, CX
- ANDQ DX, R8
- ADDQ R12, CX
- MOVQ CX, AX
- SHRQ $51, CX
- ANDQ DX, R9
- ADDQ R14, CX
- MOVQ CX, R10
- SHRQ $51, CX
- ANDQ DX, AX
- IMUL3Q $19, CX, CX
- ADDQ CX, SI
- ANDQ DX, R10
- MOVQ SI, 40(DI)
- MOVQ R8, 48(DI)
- MOVQ R9, 56(DI)
- MOVQ AX, 64(DI)
- MOVQ R10, 72(DI)
- MOVQ 160(SP), AX
- MULQ ·_121666_213(SB)
- SHRQ $13, AX
- MOVQ AX, SI
- MOVQ DX, CX
- MOVQ 168(SP), AX
- MULQ ·_121666_213(SB)
- SHRQ $13, AX
- ADDQ AX, CX
- MOVQ DX, R8
- MOVQ 176(SP), AX
- MULQ ·_121666_213(SB)
- SHRQ $13, AX
- ADDQ AX, R8
- MOVQ DX, R9
- MOVQ 184(SP), AX
- MULQ ·_121666_213(SB)
- SHRQ $13, AX
- ADDQ AX, R9
- MOVQ DX, R10
- MOVQ 192(SP), AX
- MULQ ·_121666_213(SB)
- SHRQ $13, AX
- ADDQ AX, R10
- IMUL3Q $19, DX, DX
- ADDQ DX, SI
- ADDQ 80(SP), SI
- ADDQ 88(SP), CX
- ADDQ 96(SP), R8
- ADDQ 104(SP), R9
- ADDQ 112(SP), R10
- MOVQ SI, 80(DI)
- MOVQ CX, 88(DI)
- MOVQ R8, 96(DI)
- MOVQ R9, 104(DI)
- MOVQ R10, 112(DI)
- MOVQ 104(DI), SI
- IMUL3Q $19, SI, AX
- MOVQ AX, 0(SP)
- MULQ 176(SP)
- MOVQ AX, SI
- MOVQ DX, CX
- MOVQ 112(DI), DX
- IMUL3Q $19, DX, AX
- MOVQ AX, 8(SP)
- MULQ 168(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 80(DI), AX
- MULQ 160(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 80(DI), AX
- MULQ 168(SP)
- MOVQ AX, R8
- MOVQ DX, R9
- MOVQ 80(DI), AX
- MULQ 176(SP)
- MOVQ AX, R10
- MOVQ DX, R11
- MOVQ 80(DI), AX
- MULQ 184(SP)
- MOVQ AX, R12
- MOVQ DX, R13
- MOVQ 80(DI), AX
- MULQ 192(SP)
- MOVQ AX, R14
- MOVQ DX, R15
- MOVQ 88(DI), AX
- MULQ 160(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 88(DI), AX
- MULQ 168(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 88(DI), AX
- MULQ 176(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 88(DI), AX
- MULQ 184(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 88(DI), DX
- IMUL3Q $19, DX, AX
- MULQ 192(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 96(DI), AX
- MULQ 160(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 96(DI), AX
- MULQ 168(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 96(DI), AX
- MULQ 176(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 96(DI), DX
- IMUL3Q $19, DX, AX
- MULQ 184(SP)
- ADDQ AX, SI
- ADCQ DX, CX
- MOVQ 96(DI), DX
- IMUL3Q $19, DX, AX
- MULQ 192(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 104(DI), AX
- MULQ 160(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 104(DI), AX
- MULQ 168(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 0(SP), AX
- MULQ 184(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 0(SP), AX
- MULQ 192(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 112(DI), AX
- MULQ 160(SP)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 8(SP), AX
- MULQ 176(SP)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 8(SP), AX
- MULQ 184(SP)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 8(SP), AX
- MULQ 192(SP)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ $REDMASK51, DX
- SHLQ $13, SI, CX
- ANDQ DX, SI
- SHLQ $13, R8, R9
- ANDQ DX, R8
- ADDQ CX, R8
- SHLQ $13, R10, R11
- ANDQ DX, R10
- ADDQ R9, R10
- SHLQ $13, R12, R13
- ANDQ DX, R12
- ADDQ R11, R12
- SHLQ $13, R14, R15
- ANDQ DX, R14
- ADDQ R13, R14
- IMUL3Q $19, R15, CX
- ADDQ CX, SI
- MOVQ SI, CX
- SHRQ $51, CX
- ADDQ R8, CX
- MOVQ CX, R8
- SHRQ $51, CX
- ANDQ DX, SI
- ADDQ R10, CX
- MOVQ CX, R9
- SHRQ $51, CX
- ANDQ DX, R8
- ADDQ R12, CX
- MOVQ CX, AX
- SHRQ $51, CX
- ANDQ DX, R9
- ADDQ R14, CX
- MOVQ CX, R10
- SHRQ $51, CX
- ANDQ DX, AX
- IMUL3Q $19, CX, CX
- ADDQ CX, SI
- ANDQ DX, R10
- MOVQ SI, 80(DI)
- MOVQ R8, 88(DI)
- MOVQ R9, 96(DI)
- MOVQ AX, 104(DI)
- MOVQ R10, 112(DI)
- RET
- // func cswap(inout *[4][5]uint64, v uint64)
- TEXT ·cswap(SB), 7, $0
- MOVQ inout+0(FP), DI
- MOVQ v+8(FP), SI
- SUBQ $1, SI
- NOTQ SI
- MOVQ SI, X15
- PSHUFD $0x44, X15, X15
- MOVOU 0(DI), X0
- MOVOU 16(DI), X2
- MOVOU 32(DI), X4
- MOVOU 48(DI), X6
- MOVOU 64(DI), X8
- MOVOU 80(DI), X1
- MOVOU 96(DI), X3
- MOVOU 112(DI), X5
- MOVOU 128(DI), X7
- MOVOU 144(DI), X9
- MOVO X1, X10
- MOVO X3, X11
- MOVO X5, X12
- MOVO X7, X13
- MOVO X9, X14
- PXOR X0, X10
- PXOR X2, X11
- PXOR X4, X12
- PXOR X6, X13
- PXOR X8, X14
- PAND X15, X10
- PAND X15, X11
- PAND X15, X12
- PAND X15, X13
- PAND X15, X14
- PXOR X10, X0
- PXOR X10, X1
- PXOR X11, X2
- PXOR X11, X3
- PXOR X12, X4
- PXOR X12, X5
- PXOR X13, X6
- PXOR X13, X7
- PXOR X14, X8
- PXOR X14, X9
- MOVOU X0, 0(DI)
- MOVOU X2, 16(DI)
- MOVOU X4, 32(DI)
- MOVOU X6, 48(DI)
- MOVOU X8, 64(DI)
- MOVOU X1, 80(DI)
- MOVOU X3, 96(DI)
- MOVOU X5, 112(DI)
- MOVOU X7, 128(DI)
- MOVOU X9, 144(DI)
- RET
- // func mul(outp *uint64, xp *uint64, yp *uint64)
- TEXT ·mul(SB), NOSPLIT, $0
- MOVQ outp+0(FP), DI
- MOVQ xp+8(FP), BX
- MOVQ yp+16(FP), CX
- // Calculate r0
- MOVQ 0(BX), AX // rax <-- x0
- MULQ 0(CX) // rdx, rax <-- x0*y0
- MOVQ AX, SI // r00 = rax
- MOVQ DX, BP // r01 = rdx
- MOVQ 8(BX), DX // rdx <-- x1
- IMUL3Q $19, DX, AX // rax <-- x1*19
- MULQ 32(CX) // rdx, rax <-- x1_19*y4
- ADDQ AX, SI // r00 += rax
- ADCQ DX, BP // r01 += rdx
- MOVQ 16(BX), DX // rdx <-- x2
- IMUL3Q $19, DX, AX // rax <-- x2*19
- MULQ 24(CX) // rdx, rax <-- x2_19*y3
- ADDQ AX, SI // r00 += rax
- ADCQ DX, BP // r01 += rdx
- MOVQ 24(BX), DX // rdx <-- x3
- IMUL3Q $19, DX, AX // rax <-- x3*19
- MULQ 16(CX) // rdx, rax <-- x3_19 * y2
- ADDQ AX, SI // r00 += rax
- ADCQ DX, BP // r01 += rdx
- MOVQ 32(BX), DX // rdx <-- x4
- IMUL3Q $19, DX, AX // rax <-- x4*19
- MULQ 8(CX) // rdx rax <-- x4_19*y1
- ADDQ AX, SI // r00 += rax
- ADCQ DX, BP // r01 += rdx
- // Calculate r1
- MOVQ 0(BX), AX
- MULQ 8(CX)
- MOVQ AX, R8 // r10
- MOVQ DX, R9 // r11
- MOVQ 8(BX), AX
- MULQ 0(CX)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 16(BX), DX
- IMUL3Q $19, DX, AX
- MULQ 32(CX)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 24(BX), DX
- IMUL3Q $19, DX, AX
- MULQ 24(CX)
- ADDQ AX, R8
- ADCQ DX, R9
- MOVQ 32(BX), DX
- IMUL3Q $19, DX, AX
- MULQ 16(CX)
- ADDQ AX, R8
- ADCQ DX, R9
- // Calculate r2
- MOVQ 0(BX), AX
- MULQ 16(CX)
- MOVQ AX, R10 // r20
- MOVQ DX, R11 // r21
- MOVQ 8(BX), AX
- MULQ 8(CX)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 16(BX), AX
- MULQ 0(CX)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 24(BX), DX
- IMUL3Q $19, DX, AX
- MULQ 32(CX)
- ADDQ AX, R10
- ADCQ DX, R11
- MOVQ 32(BX), DX
- IMUL3Q $19, DX, AX
- MULQ 24(CX)
- ADDQ AX, R10
- ADCQ DX, R11
- // Calculate r3
- MOVQ 0(BX), AX
- MULQ 24(CX)
- MOVQ AX, R12 // r30
- MOVQ DX, R13 // r31
- MOVQ 8(BX), AX
- MULQ 16(CX)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 16(BX), AX
- MULQ 8(CX)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 24(BX), AX
- MULQ 0(CX)
- ADDQ AX, R12
- ADCQ DX, R13
- MOVQ 32(BX), DX
- IMUL3Q $19, DX, AX
- MULQ 32(CX)
- ADDQ AX, R12
- ADCQ DX, R13
- // Calculate r4
- MOVQ 0(BX), AX
- MULQ 32(CX)
- MOVQ AX, R14 // r40
- MOVQ DX, R15 // r41
- MOVQ 8(BX), AX
- MULQ 24(CX)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 16(BX), AX
- MULQ 16(CX)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 24(BX), AX
- MULQ 8(CX)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ 32(BX), AX
- MULQ 0(CX)
- ADDQ AX, R14
- ADCQ DX, R15
- MOVQ $2251799813685247, AX // (1<<51) - 1
- SHLQ $13, SI, BP // r01 = shld with r00
- ANDQ AX, SI // r00 &= mask51
- SHLQ $13, R8, R9 // r11 = shld with r10
- ANDQ AX, R8 // r10 &= mask51
- ADDQ BP, R8 // r10 += r01
- SHLQ $13, R10, R11 // r21 = shld with r20
- ANDQ AX, R10 // r20 &= mask51
- ADDQ R9, R10 // r20 += r11
- SHLQ $13, R12, R13 // r31 = shld with r30
- ANDQ AX, R12 // r30 &= mask51
- ADDQ R11, R12 // r30 += r21
- SHLQ $13, R14, R15 // r41 = shld with r40
- ANDQ AX, R14 // r40 &= mask51
- ADDQ R13, R14 // r40 += r31
- IMUL3Q $19, R15, R15 // r41 = r41*19
- ADDQ R15, SI // r00 += r41
- MOVQ SI, DX // rdx <-- r00
- SHRQ $51, DX // rdx <-- r00 >> 51
- ADDQ DX, R8 // r10 += r00 >> 51
- MOVQ R8, DX // rdx <-- r10
- SHRQ $51, DX // rdx <-- r10 >> 51
- ANDQ AX, SI // r00 &= mask51
- ADDQ DX, R10 // r20 += r10 >> 51
- MOVQ R10, DX // rdx <-- r20
- SHRQ $51, DX // rdx <-- r20 >> 51
- ANDQ AX, R8 // r10 &= mask51
- ADDQ DX, R12 // r30 += r20 >> 51
- MOVQ R12, DX // rdx <-- r30
- SHRQ $51, DX // rdx <-- r30 >> 51
- ANDQ AX, R10 // r20 &= mask51
- ADDQ DX, R14 // r40 += r30 >> 51
- MOVQ R14, DX // rdx <-- r40
- SHRQ $51, DX // rdx <-- r40 >> 51
- ANDQ AX, R12 // r30 &= mask51
- IMUL3Q $19, DX, DX // rdx <-- (r40 >> 51) * 19
- ADDQ DX, SI // r00 += (r40 >> 51) *19
- ANDQ AX, R14 // r40 &= mask51
- MOVQ SI, 0(DI)
- MOVQ R8, 8(DI)
- MOVQ R10, 16(DI)
- MOVQ R12, 24(DI)
- MOVQ R14, 32(DI)
- RET
- // func square(outp *uint64, xp *uint64)
- TEXT ·square(SB), NOSPLIT, $0
- MOVQ outp+0(FP), DI
- MOVQ xp+8(FP), SI
- // r0 = x0*x0 + x1*38*x4 + x2*38*x3
- MOVQ 0(SI), AX
- MULQ 0(SI)
- MOVQ AX, CX // r00
- MOVQ DX, R8 // r01
- MOVQ 8(SI), DX
- IMUL3Q $38, DX, AX
- MULQ 32(SI)
- ADDQ AX, CX
- ADCQ DX, R8
- MOVQ 16(SI), DX
- IMUL3Q $38, DX, AX
- MULQ 24(SI)
- ADDQ AX, CX
- ADCQ DX, R8
- // r1 = x0*2*x1 + x2*38*x4 + x3*19*x3
- MOVQ 0(SI), AX
- SHLQ $1, AX
- MULQ 8(SI)
- MOVQ AX, R9 // r10
- MOVQ DX, R10 // r11
- MOVQ 16(SI), DX
- IMUL3Q $38, DX, AX
- MULQ 32(SI)
- ADDQ AX, R9
- ADCQ DX, R10
- MOVQ 24(SI), DX
- IMUL3Q $19, DX, AX
- MULQ 24(SI)
- ADDQ AX, R9
- ADCQ DX, R10
- // r2 = x0*2*x2 + x1*x1 + x3*38*x4
- MOVQ 0(SI), AX
- SHLQ $1, AX
- MULQ 16(SI)
- MOVQ AX, R11 // r20
- MOVQ DX, R12 // r21
- MOVQ 8(SI), AX
- MULQ 8(SI)
- ADDQ AX, R11
- ADCQ DX, R12
- MOVQ 24(SI), DX
- IMUL3Q $38, DX, AX
- MULQ 32(SI)
- ADDQ AX, R11
- ADCQ DX, R12
- // r3 = x0*2*x3 + x1*2*x2 + x4*19*x4
- MOVQ 0(SI), AX
- SHLQ $1, AX
- MULQ 24(SI)
- MOVQ AX, R13 // r30
- MOVQ DX, R14 // r31
- MOVQ 8(SI), AX
- SHLQ $1, AX
- MULQ 16(SI)
- ADDQ AX, R13
- ADCQ DX, R14
- MOVQ 32(SI), DX
- IMUL3Q $19, DX, AX
- MULQ 32(SI)
- ADDQ AX, R13
- ADCQ DX, R14
- // r4 = x0*2*x4 + x1*2*x3 + x2*x2
- MOVQ 0(SI), AX
- SHLQ $1, AX
- MULQ 32(SI)
- MOVQ AX, R15 // r40
- MOVQ DX, BX // r41
- MOVQ 8(SI), AX
- SHLQ $1, AX
- MULQ 24(SI)
- ADDQ AX, R15
- ADCQ DX, BX
- MOVQ 16(SI), AX
- MULQ 16(SI)
- ADDQ AX, R15
- ADCQ DX, BX
- // Reduce
- MOVQ $2251799813685247, AX // (1<<51) - 1
- SHLQ $13, CX, R8 // r01 = shld with r00
- ANDQ AX, CX // r00 &= mask51
- SHLQ $13, R9, R10 // r11 = shld with r10
- ANDQ AX, R9 // r10 &= mask51
- ADDQ R8, R9 // r10 += r01
- SHLQ $13, R11, R12 // r21 = shld with r20
- ANDQ AX, R11 // r20 &= mask51
- ADDQ R10, R11 // r20 += r11
- SHLQ $13, R13, R14 // r31 = shld with r30
- ANDQ AX, R13 // r30 &= mask51
- ADDQ R12, R13 // r30 += r21
- SHLQ $13, R15, BX // r41 = shld with r40
- ANDQ AX, R15 // r40 &= mask51
- ADDQ R14, R15 // r40 += r31
- IMUL3Q $19, BX, DX // r41 = r41*19
- ADDQ DX, CX // r00 += r41
- MOVQ CX, DX // rdx <-- r00
- SHRQ $51, DX // rdx <-- r00 >> 51
- ADDQ DX, R9 // r10 += r00 >> 51
- MOVQ R9, DX // rdx <-- r10
- SHRQ $51, DX // rdx <-- r10 >> 51
- ANDQ AX, CX // r00 &= mask51
- ADDQ DX, R11 // r20 += r10 >> 51
- MOVQ R11, DX // rdx <-- r20
- SHRQ $51, DX // rdx <-- r20 >> 51
- ANDQ AX, R9 // r10 &= mask51
- ADDQ DX, R13 // r30 += r20 >> 51
- MOVQ R13, DX // rdx <-- r30
- SHRQ $51, DX // rdx <-- r30 >> 51
- ANDQ AX, R11 // r20 &= mask51
- ADDQ DX, R15 // r40 += r30 >> 51
- MOVQ R15, DX // rdx <-- r40
- SHRQ $51, DX // rdx <-- r40 >> 51
- ANDQ AX, R13 // r30 &= mask51
- IMUL3Q $19, DX, DX // rdx <-- (r40 >> 51) * 19
- ADDQ DX, CX // r00 += (r40 >> 51) *19
- ANDQ AX, R15 // r40 &= mask51
- MOVQ CX, 0(DI)
- MOVQ R9, 8(DI)
- MOVQ R11, 16(DI)
- MOVQ R13, 24(DI)
- MOVQ R15, 32(DI)
- RET
|