| 1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173 |
- // Code generated by command: go run blake2s_amd64_asm.go -out ../blake2s_amd64.s -pkg blake2s. DO NOT EDIT.
- //go:build amd64 && gc && !purego
- #include "textflag.h"
- // func hashBlocksSSE2(h *[8]uint32, c *[2]uint32, flag uint32, blocks []byte)
- // Requires: SSE2
- TEXT ·hashBlocksSSE2(SB), $672-48
- MOVQ h+0(FP), AX
- MOVQ c+8(FP), BX
- MOVL flag+16(FP), CX
- MOVQ blocks_base+24(FP), SI
- MOVQ blocks_len+32(FP), DX
- MOVQ SP, BP
- ADDQ $0x0f, BP
- ANDQ $-16, BP
- MOVQ (BX), R9
- MOVQ R9, (BP)
- MOVQ CX, 8(BP)
- MOVOU (AX), X0
- MOVOU 16(AX), X1
- MOVOU iv0<>+0(SB), X2
- MOVOU iv1<>+0(SB), X3
- MOVOU counter<>+0(SB), X12
- MOVOU rol16<>+0(SB), X13
- MOVOU rol8<>+0(SB), X14
- MOVO (BP), X15
- loop:
- MOVO X0, X4
- MOVO X1, X5
- MOVO X2, X6
- MOVO X3, X7
- PADDQ X12, X15
- PXOR X15, X7
- MOVQ (SI), R8
- MOVQ 8(SI), R9
- MOVQ 16(SI), R10
- MOVQ 24(SI), R11
- MOVQ 32(SI), R12
- MOVQ 40(SI), R13
- MOVQ 48(SI), R14
- MOVQ 56(SI), R15
- MOVL R8, 16(BP)
- MOVL R8, 116(BP)
- MOVL R8, 164(BP)
- MOVL R8, 264(BP)
- MOVL R8, 288(BP)
- MOVL R8, 344(BP)
- MOVL R8, 432(BP)
- MOVL R8, 512(BP)
- MOVL R8, 540(BP)
- MOVL R8, 652(BP)
- SHRQ $0x20, R8
- MOVL R8, 32(BP)
- MOVL R8, 112(BP)
- MOVL R8, 200(BP)
- MOVL R8, 228(BP)
- MOVL R8, 320(BP)
- MOVL R8, 380(BP)
- MOVL R8, 404(BP)
- MOVL R8, 488(BP)
- MOVL R8, 568(BP)
- MOVL R8, 604(BP)
- MOVL R9, 20(BP)
- MOVL R9, 132(BP)
- MOVL R9, 168(BP)
- MOVL R9, 240(BP)
- MOVL R9, 280(BP)
- MOVL R9, 336(BP)
- MOVL R9, 456(BP)
- MOVL R9, 508(BP)
- MOVL R9, 576(BP)
- MOVL R9, 608(BP)
- SHRQ $0x20, R9
- MOVL R9, 36(BP)
- MOVL R9, 140(BP)
- MOVL R9, 180(BP)
- MOVL R9, 212(BP)
- MOVL R9, 316(BP)
- MOVL R9, 364(BP)
- MOVL R9, 452(BP)
- MOVL R9, 476(BP)
- MOVL R9, 552(BP)
- MOVL R9, 632(BP)
- MOVL R10, 24(BP)
- MOVL R10, 84(BP)
- MOVL R10, 204(BP)
- MOVL R10, 248(BP)
- MOVL R10, 296(BP)
- MOVL R10, 368(BP)
- MOVL R10, 412(BP)
- MOVL R10, 516(BP)
- MOVL R10, 584(BP)
- MOVL R10, 612(BP)
- SHRQ $0x20, R10
- MOVL R10, 40(BP)
- MOVL R10, 124(BP)
- MOVL R10, 152(BP)
- MOVL R10, 244(BP)
- MOVL R10, 276(BP)
- MOVL R10, 388(BP)
- MOVL R10, 416(BP)
- MOVL R10, 496(BP)
- MOVL R10, 588(BP)
- MOVL R10, 620(BP)
- MOVL R11, 28(BP)
- MOVL R11, 108(BP)
- MOVL R11, 196(BP)
- MOVL R11, 256(BP)
- MOVL R11, 312(BP)
- MOVL R11, 340(BP)
- MOVL R11, 436(BP)
- MOVL R11, 520(BP)
- MOVL R11, 528(BP)
- MOVL R11, 616(BP)
- SHRQ $0x20, R11
- MOVL R11, 44(BP)
- MOVL R11, 136(BP)
- MOVL R11, 184(BP)
- MOVL R11, 208(BP)
- MOVL R11, 292(BP)
- MOVL R11, 372(BP)
- MOVL R11, 448(BP)
- MOVL R11, 468(BP)
- MOVL R11, 580(BP)
- MOVL R11, 600(BP)
- MOVL R12, 48(BP)
- MOVL R12, 100(BP)
- MOVL R12, 160(BP)
- MOVL R12, 268(BP)
- MOVL R12, 328(BP)
- MOVL R12, 348(BP)
- MOVL R12, 444(BP)
- MOVL R12, 504(BP)
- MOVL R12, 556(BP)
- MOVL R12, 596(BP)
- SHRQ $0x20, R12
- MOVL R12, 64(BP)
- MOVL R12, 88(BP)
- MOVL R12, 188(BP)
- MOVL R12, 224(BP)
- MOVL R12, 272(BP)
- MOVL R12, 396(BP)
- MOVL R12, 440(BP)
- MOVL R12, 492(BP)
- MOVL R12, 548(BP)
- MOVL R12, 628(BP)
- MOVL R13, 52(BP)
- MOVL R13, 96(BP)
- MOVL R13, 176(BP)
- MOVL R13, 260(BP)
- MOVL R13, 284(BP)
- MOVL R13, 356(BP)
- MOVL R13, 428(BP)
- MOVL R13, 524(BP)
- MOVL R13, 572(BP)
- MOVL R13, 592(BP)
- SHRQ $0x20, R13
- MOVL R13, 68(BP)
- MOVL R13, 120(BP)
- MOVL R13, 144(BP)
- MOVL R13, 220(BP)
- MOVL R13, 308(BP)
- MOVL R13, 360(BP)
- MOVL R13, 460(BP)
- MOVL R13, 480(BP)
- MOVL R13, 536(BP)
- MOVL R13, 640(BP)
- MOVL R14, 56(BP)
- MOVL R14, 128(BP)
- MOVL R14, 148(BP)
- MOVL R14, 232(BP)
- MOVL R14, 324(BP)
- MOVL R14, 352(BP)
- MOVL R14, 400(BP)
- MOVL R14, 472(BP)
- MOVL R14, 560(BP)
- MOVL R14, 648(BP)
- SHRQ $0x20, R14
- MOVL R14, 72(BP)
- MOVL R14, 92(BP)
- MOVL R14, 172(BP)
- MOVL R14, 216(BP)
- MOVL R14, 332(BP)
- MOVL R14, 384(BP)
- MOVL R14, 424(BP)
- MOVL R14, 464(BP)
- MOVL R14, 564(BP)
- MOVL R14, 636(BP)
- MOVL R15, 60(BP)
- MOVL R15, 80(BP)
- MOVL R15, 192(BP)
- MOVL R15, 236(BP)
- MOVL R15, 304(BP)
- MOVL R15, 392(BP)
- MOVL R15, 408(BP)
- MOVL R15, 484(BP)
- MOVL R15, 532(BP)
- MOVL R15, 644(BP)
- SHRQ $0x20, R15
- MOVL R15, 76(BP)
- MOVL R15, 104(BP)
- MOVL R15, 156(BP)
- MOVL R15, 252(BP)
- MOVL R15, 300(BP)
- MOVL R15, 376(BP)
- MOVL R15, 420(BP)
- MOVL R15, 500(BP)
- MOVL R15, 544(BP)
- MOVL R15, 624(BP)
- PADDL 16(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 32(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 48(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 64(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 80(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 96(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 112(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 128(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 144(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 160(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 176(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 192(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 208(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 224(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 240(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 256(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 272(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 288(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 304(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 320(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 336(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 352(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 368(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 384(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 400(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 416(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 432(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 448(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 464(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 480(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 496(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 512(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 528(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 544(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 560(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 576(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 592(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 608(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 624(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x10, X8
- PSRLL $0x10, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 640(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- MOVO X7, X8
- PSLLL $0x18, X8
- PSRLL $0x08, X7
- PXOR X8, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PXOR X4, X0
- PXOR X5, X1
- PXOR X6, X0
- PXOR X7, X1
- LEAQ 64(SI), SI
- SUBQ $0x40, DX
- JNE loop
- MOVO X15, (BP)
- MOVQ (BP), R9
- MOVQ R9, (BX)
- MOVOU X0, (AX)
- MOVOU X1, 16(AX)
- RET
- DATA iv0<>+0(SB)/4, $0x6a09e667
- DATA iv0<>+4(SB)/4, $0xbb67ae85
- DATA iv0<>+8(SB)/4, $0x3c6ef372
- DATA iv0<>+12(SB)/4, $0xa54ff53a
- GLOBL iv0<>(SB), RODATA|NOPTR, $16
- DATA iv1<>+0(SB)/4, $0x510e527f
- DATA iv1<>+4(SB)/4, $0x9b05688c
- DATA iv1<>+8(SB)/4, $0x1f83d9ab
- DATA iv1<>+12(SB)/4, $0x5be0cd19
- GLOBL iv1<>(SB), RODATA|NOPTR, $16
- DATA counter<>+0(SB)/8, $0x0000000000000040
- DATA counter<>+8(SB)/8, $0x0000000000000000
- GLOBL counter<>(SB), RODATA|NOPTR, $16
- DATA rol16<>+0(SB)/8, $0x0504070601000302
- DATA rol16<>+8(SB)/8, $0x0d0c0f0e09080b0a
- GLOBL rol16<>(SB), RODATA|NOPTR, $16
- DATA rol8<>+0(SB)/8, $0x0407060500030201
- DATA rol8<>+8(SB)/8, $0x0c0f0e0d080b0a09
- GLOBL rol8<>(SB), RODATA|NOPTR, $16
- // func hashBlocksSSSE3(h *[8]uint32, c *[2]uint32, flag uint32, blocks []byte)
- // Requires: SSE2, SSSE3
- TEXT ·hashBlocksSSSE3(SB), $672-48
- MOVQ h+0(FP), AX
- MOVQ c+8(FP), BX
- MOVL flag+16(FP), CX
- MOVQ blocks_base+24(FP), SI
- MOVQ blocks_len+32(FP), DX
- MOVQ SP, BP
- ADDQ $0x0f, BP
- ANDQ $-16, BP
- MOVQ (BX), R9
- MOVQ R9, (BP)
- MOVQ CX, 8(BP)
- MOVOU (AX), X0
- MOVOU 16(AX), X1
- MOVOU iv0<>+0(SB), X2
- MOVOU iv1<>+0(SB), X3
- MOVOU counter<>+0(SB), X12
- MOVOU rol16<>+0(SB), X13
- MOVOU rol8<>+0(SB), X14
- MOVO (BP), X15
- loop:
- MOVO X0, X4
- MOVO X1, X5
- MOVO X2, X6
- MOVO X3, X7
- PADDQ X12, X15
- PXOR X15, X7
- MOVQ (SI), R8
- MOVQ 8(SI), R9
- MOVQ 16(SI), R10
- MOVQ 24(SI), R11
- MOVQ 32(SI), R12
- MOVQ 40(SI), R13
- MOVQ 48(SI), R14
- MOVQ 56(SI), R15
- MOVL R8, 16(BP)
- MOVL R8, 116(BP)
- MOVL R8, 164(BP)
- MOVL R8, 264(BP)
- MOVL R8, 288(BP)
- MOVL R8, 344(BP)
- MOVL R8, 432(BP)
- MOVL R8, 512(BP)
- MOVL R8, 540(BP)
- MOVL R8, 652(BP)
- SHRQ $0x20, R8
- MOVL R8, 32(BP)
- MOVL R8, 112(BP)
- MOVL R8, 200(BP)
- MOVL R8, 228(BP)
- MOVL R8, 320(BP)
- MOVL R8, 380(BP)
- MOVL R8, 404(BP)
- MOVL R8, 488(BP)
- MOVL R8, 568(BP)
- MOVL R8, 604(BP)
- MOVL R9, 20(BP)
- MOVL R9, 132(BP)
- MOVL R9, 168(BP)
- MOVL R9, 240(BP)
- MOVL R9, 280(BP)
- MOVL R9, 336(BP)
- MOVL R9, 456(BP)
- MOVL R9, 508(BP)
- MOVL R9, 576(BP)
- MOVL R9, 608(BP)
- SHRQ $0x20, R9
- MOVL R9, 36(BP)
- MOVL R9, 140(BP)
- MOVL R9, 180(BP)
- MOVL R9, 212(BP)
- MOVL R9, 316(BP)
- MOVL R9, 364(BP)
- MOVL R9, 452(BP)
- MOVL R9, 476(BP)
- MOVL R9, 552(BP)
- MOVL R9, 632(BP)
- MOVL R10, 24(BP)
- MOVL R10, 84(BP)
- MOVL R10, 204(BP)
- MOVL R10, 248(BP)
- MOVL R10, 296(BP)
- MOVL R10, 368(BP)
- MOVL R10, 412(BP)
- MOVL R10, 516(BP)
- MOVL R10, 584(BP)
- MOVL R10, 612(BP)
- SHRQ $0x20, R10
- MOVL R10, 40(BP)
- MOVL R10, 124(BP)
- MOVL R10, 152(BP)
- MOVL R10, 244(BP)
- MOVL R10, 276(BP)
- MOVL R10, 388(BP)
- MOVL R10, 416(BP)
- MOVL R10, 496(BP)
- MOVL R10, 588(BP)
- MOVL R10, 620(BP)
- MOVL R11, 28(BP)
- MOVL R11, 108(BP)
- MOVL R11, 196(BP)
- MOVL R11, 256(BP)
- MOVL R11, 312(BP)
- MOVL R11, 340(BP)
- MOVL R11, 436(BP)
- MOVL R11, 520(BP)
- MOVL R11, 528(BP)
- MOVL R11, 616(BP)
- SHRQ $0x20, R11
- MOVL R11, 44(BP)
- MOVL R11, 136(BP)
- MOVL R11, 184(BP)
- MOVL R11, 208(BP)
- MOVL R11, 292(BP)
- MOVL R11, 372(BP)
- MOVL R11, 448(BP)
- MOVL R11, 468(BP)
- MOVL R11, 580(BP)
- MOVL R11, 600(BP)
- MOVL R12, 48(BP)
- MOVL R12, 100(BP)
- MOVL R12, 160(BP)
- MOVL R12, 268(BP)
- MOVL R12, 328(BP)
- MOVL R12, 348(BP)
- MOVL R12, 444(BP)
- MOVL R12, 504(BP)
- MOVL R12, 556(BP)
- MOVL R12, 596(BP)
- SHRQ $0x20, R12
- MOVL R12, 64(BP)
- MOVL R12, 88(BP)
- MOVL R12, 188(BP)
- MOVL R12, 224(BP)
- MOVL R12, 272(BP)
- MOVL R12, 396(BP)
- MOVL R12, 440(BP)
- MOVL R12, 492(BP)
- MOVL R12, 548(BP)
- MOVL R12, 628(BP)
- MOVL R13, 52(BP)
- MOVL R13, 96(BP)
- MOVL R13, 176(BP)
- MOVL R13, 260(BP)
- MOVL R13, 284(BP)
- MOVL R13, 356(BP)
- MOVL R13, 428(BP)
- MOVL R13, 524(BP)
- MOVL R13, 572(BP)
- MOVL R13, 592(BP)
- SHRQ $0x20, R13
- MOVL R13, 68(BP)
- MOVL R13, 120(BP)
- MOVL R13, 144(BP)
- MOVL R13, 220(BP)
- MOVL R13, 308(BP)
- MOVL R13, 360(BP)
- MOVL R13, 460(BP)
- MOVL R13, 480(BP)
- MOVL R13, 536(BP)
- MOVL R13, 640(BP)
- MOVL R14, 56(BP)
- MOVL R14, 128(BP)
- MOVL R14, 148(BP)
- MOVL R14, 232(BP)
- MOVL R14, 324(BP)
- MOVL R14, 352(BP)
- MOVL R14, 400(BP)
- MOVL R14, 472(BP)
- MOVL R14, 560(BP)
- MOVL R14, 648(BP)
- SHRQ $0x20, R14
- MOVL R14, 72(BP)
- MOVL R14, 92(BP)
- MOVL R14, 172(BP)
- MOVL R14, 216(BP)
- MOVL R14, 332(BP)
- MOVL R14, 384(BP)
- MOVL R14, 424(BP)
- MOVL R14, 464(BP)
- MOVL R14, 564(BP)
- MOVL R14, 636(BP)
- MOVL R15, 60(BP)
- MOVL R15, 80(BP)
- MOVL R15, 192(BP)
- MOVL R15, 236(BP)
- MOVL R15, 304(BP)
- MOVL R15, 392(BP)
- MOVL R15, 408(BP)
- MOVL R15, 484(BP)
- MOVL R15, 532(BP)
- MOVL R15, 644(BP)
- SHRQ $0x20, R15
- MOVL R15, 76(BP)
- MOVL R15, 104(BP)
- MOVL R15, 156(BP)
- MOVL R15, 252(BP)
- MOVL R15, 300(BP)
- MOVL R15, 376(BP)
- MOVL R15, 420(BP)
- MOVL R15, 500(BP)
- MOVL R15, 544(BP)
- MOVL R15, 624(BP)
- PADDL 16(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 32(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 48(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 64(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 80(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 96(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 112(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 128(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 144(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 160(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 176(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 192(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 208(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 224(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 240(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 256(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 272(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 288(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 304(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 320(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 336(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 352(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 368(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 384(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 400(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 416(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 432(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 448(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 464(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 480(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 496(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 512(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 528(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 544(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 560(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 576(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PADDL 592(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 608(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL 624(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL 640(BP), X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PXOR X4, X0
- PXOR X5, X1
- PXOR X6, X0
- PXOR X7, X1
- LEAQ 64(SI), SI
- SUBQ $0x40, DX
- JNE loop
- MOVO X15, (BP)
- MOVQ (BP), R9
- MOVQ R9, (BX)
- MOVOU X0, (AX)
- MOVOU X1, 16(AX)
- RET
- // func hashBlocksSSE4(h *[8]uint32, c *[2]uint32, flag uint32, blocks []byte)
- // Requires: SSE2, SSE4.1, SSSE3
- TEXT ·hashBlocksSSE4(SB), $32-48
- MOVQ h+0(FP), AX
- MOVQ c+8(FP), BX
- MOVL flag+16(FP), CX
- MOVQ blocks_base+24(FP), SI
- MOVQ blocks_len+32(FP), DX
- MOVQ SP, BP
- ADDQ $0x0f, BP
- ANDQ $-16, BP
- MOVQ (BX), R9
- MOVQ R9, (BP)
- MOVQ CX, 8(BP)
- MOVOU (AX), X0
- MOVOU 16(AX), X1
- MOVOU iv0<>+0(SB), X2
- MOVOU iv1<>+0(SB), X3
- MOVOU counter<>+0(SB), X12
- MOVOU rol16<>+0(SB), X13
- MOVOU rol8<>+0(SB), X14
- MOVO (BP), X15
- loop:
- MOVO X0, X4
- MOVO X1, X5
- MOVO X2, X6
- MOVO X3, X7
- PADDQ X12, X15
- PXOR X15, X7
- MOVL (SI), X8
- PINSRD $0x01, 8(SI), X8
- PINSRD $0x02, 16(SI), X8
- PINSRD $0x03, 24(SI), X8
- MOVL 4(SI), X9
- PINSRD $0x01, 12(SI), X9
- PINSRD $0x02, 20(SI), X9
- PINSRD $0x03, 28(SI), X9
- MOVL 32(SI), X10
- PINSRD $0x01, 40(SI), X10
- PINSRD $0x02, 48(SI), X10
- PINSRD $0x03, 56(SI), X10
- MOVL 36(SI), X11
- PINSRD $0x01, 44(SI), X11
- PINSRD $0x02, 52(SI), X11
- PINSRD $0x03, 60(SI), X11
- PADDL X8, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X9, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL X10, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X11, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- MOVL 56(SI), X8
- PINSRD $0x01, 16(SI), X8
- PINSRD $0x02, 36(SI), X8
- PINSRD $0x03, 52(SI), X8
- MOVL 40(SI), X9
- PINSRD $0x01, 32(SI), X9
- PINSRD $0x02, 60(SI), X9
- PINSRD $0x03, 24(SI), X9
- MOVL 4(SI), X10
- PINSRD $0x01, (SI), X10
- PINSRD $0x02, 44(SI), X10
- PINSRD $0x03, 20(SI), X10
- MOVL 48(SI), X11
- PINSRD $0x01, 8(SI), X11
- PINSRD $0x02, 28(SI), X11
- PINSRD $0x03, 12(SI), X11
- PADDL X8, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X9, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL X10, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X11, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- MOVL 44(SI), X8
- PINSRD $0x01, 48(SI), X8
- PINSRD $0x02, 20(SI), X8
- PINSRD $0x03, 60(SI), X8
- MOVL 32(SI), X9
- PINSRD $0x01, (SI), X9
- PINSRD $0x02, 8(SI), X9
- PINSRD $0x03, 52(SI), X9
- MOVL 40(SI), X10
- PINSRD $0x01, 12(SI), X10
- PINSRD $0x02, 28(SI), X10
- PINSRD $0x03, 36(SI), X10
- MOVL 56(SI), X11
- PINSRD $0x01, 24(SI), X11
- PINSRD $0x02, 4(SI), X11
- PINSRD $0x03, 16(SI), X11
- PADDL X8, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X9, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL X10, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X11, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- MOVL 28(SI), X8
- PINSRD $0x01, 12(SI), X8
- PINSRD $0x02, 52(SI), X8
- PINSRD $0x03, 44(SI), X8
- MOVL 36(SI), X9
- PINSRD $0x01, 4(SI), X9
- PINSRD $0x02, 48(SI), X9
- PINSRD $0x03, 56(SI), X9
- MOVL 8(SI), X10
- PINSRD $0x01, 20(SI), X10
- PINSRD $0x02, 16(SI), X10
- PINSRD $0x03, 60(SI), X10
- MOVL 24(SI), X11
- PINSRD $0x01, 40(SI), X11
- PINSRD $0x02, (SI), X11
- PINSRD $0x03, 32(SI), X11
- PADDL X8, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X9, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL X10, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X11, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- MOVL 36(SI), X8
- PINSRD $0x01, 20(SI), X8
- PINSRD $0x02, 8(SI), X8
- PINSRD $0x03, 40(SI), X8
- MOVL (SI), X9
- PINSRD $0x01, 28(SI), X9
- PINSRD $0x02, 16(SI), X9
- PINSRD $0x03, 60(SI), X9
- MOVL 56(SI), X10
- PINSRD $0x01, 44(SI), X10
- PINSRD $0x02, 24(SI), X10
- PINSRD $0x03, 12(SI), X10
- MOVL 4(SI), X11
- PINSRD $0x01, 48(SI), X11
- PINSRD $0x02, 32(SI), X11
- PINSRD $0x03, 52(SI), X11
- PADDL X8, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X9, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL X10, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X11, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- MOVL 8(SI), X8
- PINSRD $0x01, 24(SI), X8
- PINSRD $0x02, (SI), X8
- PINSRD $0x03, 32(SI), X8
- MOVL 48(SI), X9
- PINSRD $0x01, 40(SI), X9
- PINSRD $0x02, 44(SI), X9
- PINSRD $0x03, 12(SI), X9
- MOVL 16(SI), X10
- PINSRD $0x01, 28(SI), X10
- PINSRD $0x02, 60(SI), X10
- PINSRD $0x03, 4(SI), X10
- MOVL 52(SI), X11
- PINSRD $0x01, 20(SI), X11
- PINSRD $0x02, 56(SI), X11
- PINSRD $0x03, 36(SI), X11
- PADDL X8, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X9, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL X10, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X11, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- MOVL 48(SI), X8
- PINSRD $0x01, 4(SI), X8
- PINSRD $0x02, 56(SI), X8
- PINSRD $0x03, 16(SI), X8
- MOVL 20(SI), X9
- PINSRD $0x01, 60(SI), X9
- PINSRD $0x02, 52(SI), X9
- PINSRD $0x03, 40(SI), X9
- MOVL (SI), X10
- PINSRD $0x01, 24(SI), X10
- PINSRD $0x02, 36(SI), X10
- PINSRD $0x03, 32(SI), X10
- MOVL 28(SI), X11
- PINSRD $0x01, 12(SI), X11
- PINSRD $0x02, 8(SI), X11
- PINSRD $0x03, 44(SI), X11
- PADDL X8, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X9, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL X10, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X11, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- MOVL 52(SI), X8
- PINSRD $0x01, 28(SI), X8
- PINSRD $0x02, 48(SI), X8
- PINSRD $0x03, 12(SI), X8
- MOVL 44(SI), X9
- PINSRD $0x01, 56(SI), X9
- PINSRD $0x02, 4(SI), X9
- PINSRD $0x03, 36(SI), X9
- MOVL 20(SI), X10
- PINSRD $0x01, 60(SI), X10
- PINSRD $0x02, 32(SI), X10
- PINSRD $0x03, 8(SI), X10
- MOVL (SI), X11
- PINSRD $0x01, 16(SI), X11
- PINSRD $0x02, 24(SI), X11
- PINSRD $0x03, 40(SI), X11
- PADDL X8, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X9, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL X10, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X11, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- MOVL 24(SI), X8
- PINSRD $0x01, 56(SI), X8
- PINSRD $0x02, 44(SI), X8
- PINSRD $0x03, (SI), X8
- MOVL 60(SI), X9
- PINSRD $0x01, 36(SI), X9
- PINSRD $0x02, 12(SI), X9
- PINSRD $0x03, 32(SI), X9
- MOVL 48(SI), X10
- PINSRD $0x01, 52(SI), X10
- PINSRD $0x02, 4(SI), X10
- PINSRD $0x03, 40(SI), X10
- MOVL 8(SI), X11
- PINSRD $0x01, 28(SI), X11
- PINSRD $0x02, 16(SI), X11
- PINSRD $0x03, 20(SI), X11
- PADDL X8, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X9, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL X10, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X11, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- MOVL 40(SI), X8
- PINSRD $0x01, 32(SI), X8
- PINSRD $0x02, 28(SI), X8
- PINSRD $0x03, 4(SI), X8
- MOVL 8(SI), X9
- PINSRD $0x01, 16(SI), X9
- PINSRD $0x02, 24(SI), X9
- PINSRD $0x03, 20(SI), X9
- MOVL 60(SI), X10
- PINSRD $0x01, 36(SI), X10
- PINSRD $0x02, 12(SI), X10
- PINSRD $0x03, 52(SI), X10
- MOVL 44(SI), X11
- PINSRD $0x01, 56(SI), X11
- PINSRD $0x02, 48(SI), X11
- PINSRD $0x03, (SI), X11
- PADDL X8, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X9, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X5, X5
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X7, X7
- PADDL X10, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X13, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x14, X8
- PSRLL $0x0c, X5
- PXOR X8, X5
- PADDL X11, X4
- PADDL X5, X4
- PXOR X4, X7
- PSHUFB X14, X7
- PADDL X7, X6
- PXOR X6, X5
- MOVO X5, X8
- PSLLL $0x19, X8
- PSRLL $0x07, X5
- PXOR X8, X5
- PSHUFL $0x39, X7, X7
- PSHUFL $0x4e, X6, X6
- PSHUFL $0x93, X5, X5
- PXOR X4, X0
- PXOR X5, X1
- PXOR X6, X0
- PXOR X7, X1
- LEAQ 64(SI), SI
- SUBQ $0x40, DX
- JNE loop
- MOVO X15, (BP)
- MOVQ (BP), R9
- MOVQ R9, (BX)
- MOVOU X0, (AX)
- MOVOU X1, 16(AX)
- RET
|