You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

trsm_kernel_LT.S 180 kB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639164016411642164316441645164616471648164916501651165216531654165516561657165816591660166116621663166416651666166716681669167016711672167316741675167616771678167916801681168216831684168516861687168816891690169116921693169416951696169716981699170017011702170317041705170617071708170917101711171217131714171517161717171817191720172117221723172417251726172717281729173017311732173317341735173617371738173917401741174217431744174517461747174817491750175117521753175417551756175717581759176017611762176317641765176617671768176917701771177217731774177517761777177817791780178117821783178417851786178717881789179017911792179317941795179617971798179918001801180218031804180518061807180818091810181118121813181418151816181718181819182018211822182318241825182618271828182918301831183218331834183518361837183818391840184118421843184418451846184718481849185018511852185318541855185618571858185918601861186218631864186518661867186818691870187118721873187418751876187718781879188018811882188318841885188618871888188918901891189218931894189518961897189818991900190119021903190419051906190719081909191019111912191319141915191619171918191919201921192219231924192519261927192819291930193119321933193419351936193719381939194019411942194319441945194619471948194919501951195219531954195519561957195819591960196119621963196419651966196719681969197019711972197319741975197619771978197919801981198219831984198519861987198819891990199119921993199419951996199719981999200020012002200320042005200620072008200920102011201220132014201520162017201820192020202120222023202420252026202720282029203020312032203320342035203620372038203920402041204220432044204520462047204820492050205120522053205420552056205720582059206020612062206320642065206620672068206920702071207220732074207520762077207820792080208120822083208420852086208720882089209020912092209320942095209620972098209921002101210221032104210521062107210821092110211121122113211421152116211721182119212021212122212321242125212621272128212921302131213221332134213521362137213821392140214121422143214421452146214721482149215021512152215321542155215621572158215921602161216221632164216521662167216821692170217121722173217421752176217721782179218021812182218321842185218621872188218921902191219221932194219521962197219821992200220122022203220422052206220722082209221022112212221322142215221622172218221922202221222222232224222522262227222822292230223122322233223422352236223722382239224022412242224322442245224622472248224922502251225222532254225522562257225822592260226122622263226422652266226722682269227022712272227322742275227622772278227922802281228222832284228522862287228822892290229122922293229422952296229722982299230023012302230323042305230623072308230923102311231223132314231523162317231823192320232123222323232423252326232723282329233023312332233323342335233623372338233923402341234223432344234523462347234823492350235123522353235423552356235723582359236023612362236323642365236623672368236923702371237223732374237523762377237823792380238123822383238423852386238723882389239023912392239323942395239623972398239924002401240224032404240524062407240824092410241124122413241424152416241724182419242024212422242324242425242624272428242924302431243224332434243524362437243824392440244124422443244424452446244724482449245024512452245324542455245624572458245924602461246224632464246524662467246824692470247124722473247424752476247724782479248024812482248324842485248624872488248924902491249224932494249524962497249824992500250125022503250425052506250725082509251025112512251325142515251625172518251925202521252225232524252525262527252825292530253125322533253425352536253725382539254025412542254325442545254625472548254925502551255225532554255525562557255825592560256125622563256425652566256725682569257025712572257325742575257625772578257925802581258225832584258525862587258825892590259125922593259425952596259725982599260026012602260326042605260626072608260926102611261226132614261526162617261826192620262126222623262426252626262726282629263026312632263326342635263626372638263926402641264226432644264526462647264826492650265126522653265426552656265726582659266026612662266326642665266626672668266926702671267226732674267526762677267826792680268126822683268426852686268726882689269026912692269326942695269626972698269927002701270227032704270527062707270827092710271127122713271427152716271727182719272027212722272327242725272627272728272927302731273227332734273527362737273827392740274127422743274427452746274727482749275027512752275327542755275627572758275927602761276227632764276527662767276827692770277127722773277427752776277727782779278027812782278327842785278627872788278927902791279227932794279527962797279827992800280128022803280428052806280728082809281028112812281328142815281628172818281928202821282228232824282528262827282828292830283128322833283428352836283728382839284028412842284328442845284628472848284928502851285228532854285528562857285828592860286128622863286428652866286728682869287028712872287328742875287628772878287928802881288228832884288528862887288828892890289128922893289428952896289728982899290029012902290329042905290629072908290929102911291229132914291529162917291829192920292129222923292429252926292729282929293029312932293329342935293629372938293929402941294229432944294529462947294829492950295129522953295429552956295729582959296029612962296329642965296629672968296929702971297229732974297529762977297829792980298129822983298429852986298729882989299029912992299329942995299629972998299930003001300230033004300530063007300830093010301130123013301430153016301730183019302030213022302330243025302630273028302930303031303230333034303530363037303830393040304130423043304430453046304730483049305030513052305330543055305630573058305930603061306230633064306530663067306830693070307130723073307430753076307730783079308030813082308330843085308630873088308930903091309230933094309530963097309830993100310131023103310431053106310731083109311031113112311331143115311631173118311931203121312231233124312531263127312831293130313131323133313431353136313731383139314031413142314331443145314631473148314931503151315231533154315531563157315831593160316131623163316431653166316731683169317031713172317331743175317631773178317931803181318231833184318531863187318831893190319131923193319431953196319731983199320032013202320332043205320632073208320932103211321232133214321532163217321832193220322132223223322432253226322732283229323032313232323332343235323632373238323932403241324232433244324532463247324832493250325132523253325432553256325732583259326032613262326332643265326632673268326932703271327232733274327532763277327832793280328132823283328432853286328732883289329032913292329332943295329632973298329933003301330233033304330533063307330833093310331133123313331433153316331733183319332033213322332333243325332633273328332933303331333233333334333533363337333833393340334133423343334433453346334733483349335033513352335333543355335633573358335933603361336233633364336533663367336833693370337133723373337433753376337733783379338033813382338333843385338633873388338933903391339233933394339533963397339833993400340134023403340434053406340734083409341034113412341334143415341634173418341934203421342234233424342534263427342834293430343134323433343434353436343734383439344034413442344334443445344634473448344934503451345234533454345534563457345834593460346134623463346434653466346734683469347034713472347334743475347634773478347934803481348234833484348534863487348834893490349134923493349434953496349734983499350035013502350335043505350635073508350935103511351235133514351535163517351835193520352135223523352435253526352735283529353035313532353335343535353635373538353935403541354235433544354535463547354835493550355135523553355435553556355735583559356035613562356335643565356635673568356935703571357235733574357535763577357835793580358135823583358435853586358735883589359035913592359335943595359635973598359936003601360236033604360536063607360836093610361136123613361436153616361736183619362036213622362336243625362636273628362936303631363236333634363536363637363836393640364136423643364436453646364736483649365036513652365336543655365636573658365936603661366236633664366536663667366836693670367136723673367436753676367736783679368036813682368336843685368636873688368936903691369236933694369536963697369836993700370137023703370437053706370737083709371037113712371337143715371637173718371937203721372237233724372537263727372837293730373137323733373437353736373737383739374037413742374337443745374637473748374937503751375237533754375537563757375837593760376137623763376437653766376737683769377037713772377337743775377637773778377937803781378237833784378537863787378837893790379137923793379437953796379737983799380038013802380338043805380638073808380938103811381238133814381538163817381838193820382138223823382438253826382738283829383038313832383338343835383638373838383938403841384238433844384538463847384838493850385138523853385438553856385738583859386038613862386338643865386638673868386938703871387238733874387538763877387838793880388138823883388438853886388738883889389038913892389338943895389638973898389939003901390239033904390539063907390839093910391139123913391439153916391739183919392039213922392339243925392639273928392939303931393239333934393539363937393839393940394139423943394439453946394739483949395039513952395339543955395639573958395939603961396239633964396539663967396839693970397139723973397439753976397739783979398039813982398339843985398639873988398939903991399239933994399539963997399839994000400140024003400440054006400740084009401040114012401340144015401640174018401940204021402240234024402540264027402840294030403140324033403440354036403740384039404040414042404340444045404640474048404940504051405240534054405540564057405840594060406140624063406440654066406740684069407040714072407340744075407640774078407940804081408240834084408540864087408840894090409140924093409440954096409740984099410041014102410341044105410641074108410941104111411241134114411541164117411841194120412141224123412441254126412741284129413041314132413341344135413641374138413941404141414241434144414541464147414841494150415141524153415441554156415741584159416041614162416341644165416641674168416941704171417241734174417541764177417841794180418141824183418441854186418741884189419041914192419341944195419641974198419942004201420242034204420542064207420842094210421142124213421442154216421742184219422042214222422342244225422642274228422942304231423242334234423542364237423842394240424142424243424442454246424742484249425042514252425342544255425642574258425942604261426242634264426542664267426842694270427142724273427442754276427742784279428042814282428342844285428642874288428942904291429242934294429542964297429842994300430143024303430443054306430743084309431043114312431343144315431643174318431943204321432243234324432543264327432843294330433143324333433443354336433743384339434043414342434343444345434643474348434943504351435243534354435543564357435843594360436143624363436443654366436743684369437043714372437343744375437643774378437943804381438243834384438543864387438843894390439143924393439443954396439743984399440044014402440344044405440644074408440944104411441244134414441544164417441844194420442144224423442444254426442744284429443044314432443344344435443644374438443944404441444244434444444544464447444844494450445144524453445444554456445744584459446044614462446344644465446644674468446944704471447244734474447544764477447844794480448144824483448444854486448744884489449044914492449344944495449644974498449945004501450245034504450545064507450845094510451145124513451445154516451745184519452045214522452345244525452645274528452945304531453245334534453545364537453845394540454145424543454445454546454745484549455045514552455345544555455645574558455945604561456245634564456545664567456845694570457145724573457445754576457745784579458045814582458345844585458645874588458945904591459245934594459545964597459845994600460146024603460446054606460746084609461046114612461346144615461646174618461946204621462246234624462546264627462846294630463146324633463446354636463746384639464046414642464346444645464646474648464946504651465246534654465546564657465846594660466146624663466446654666466746684669467046714672467346744675467646774678467946804681468246834684468546864687468846894690469146924693469446954696469746984699470047014702470347044705470647074708470947104711471247134714471547164717471847194720472147224723472447254726472747284729473047314732473347344735473647374738473947404741474247434744474547464747474847494750475147524753475447554756475747584759476047614762476347644765476647674768476947704771477247734774477547764777477847794780478147824783478447854786478747884789479047914792479347944795479647974798479948004801480248034804480548064807480848094810481148124813481448154816481748184819482048214822482348244825482648274828482948304831483248334834483548364837483848394840484148424843484448454846484748484849485048514852485348544855485648574858485948604861486248634864486548664867486848694870487148724873487448754876487748784879488048814882488348844885488648874888488948904891489248934894489548964897489848994900490149024903490449054906490749084909491049114912491349144915491649174918491949204921492249234924492549264927492849294930493149324933493449354936493749384939494049414942494349444945494649474948494949504951495249534954495549564957495849594960496149624963496449654966496749684969497049714972497349744975497649774978497949804981498249834984498549864987498849894990499149924993499449954996499749984999500050015002500350045005500650075008500950105011501250135014501550165017501850195020502150225023502450255026502750285029503050315032503350345035503650375038503950405041504250435044504550465047504850495050505150525053505450555056505750585059506050615062506350645065506650675068506950705071507250735074507550765077507850795080508150825083508450855086508750885089509050915092509350945095509650975098509951005101510251035104510551065107510851095110511151125113511451155116511751185119512051215122512351245125512651275128512951305131513251335134513551365137513851395140514151425143514451455146514751485149515051515152515351545155515651575158515951605161516251635164516551665167516851695170517151725173517451755176517751785179518051815182518351845185518651875188518951905191519251935194519551965197519851995200520152025203520452055206520752085209521052115212521352145215521652175218521952205221522252235224522552265227522852295230523152325233523452355236523752385239524052415242524352445245524652475248524952505251525252535254525552565257525852595260526152625263526452655266526752685269527052715272527352745275527652775278527952805281528252835284528552865287528852895290529152925293529452955296529752985299530053015302530353045305530653075308530953105311531253135314531553165317531853195320532153225323532453255326532753285329533053315332533353345335533653375338533953405341534253435344534553465347534853495350535153525353535453555356535753585359536053615362536353645365536653675368536953705371537253735374537553765377537853795380538153825383538453855386538753885389539053915392539353945395539653975398539954005401540254035404540554065407540854095410541154125413541454155416541754185419542054215422542354245425542654275428542954305431543254335434543554365437543854395440544154425443544454455446544754485449545054515452545354545455545654575458545954605461546254635464546554665467546854695470547154725473547454755476547754785479548054815482548354845485548654875488548954905491549254935494549554965497549854995500550155025503550455055506550755085509551055115512551355145515551655175518551955205521552255235524552555265527552855295530553155325533553455355536553755385539554055415542554355445545554655475548554955505551555255535554555555565557555855595560556155625563556455655566556755685569557055715572557355745575557655775578557955805581558255835584558555865587558855895590559155925593559455955596559755985599560056015602560356045605560656075608560956105611561256135614561556165617561856195620562156225623562456255626562756285629563056315632563356345635563656375638563956405641564256435644564556465647564856495650565156525653565456555656565756585659566056615662566356645665566656675668566956705671567256735674567556765677567856795680568156825683568456855686568756885689569056915692569356945695569656975698569957005701570257035704570557065707570857095710571157125713571457155716571757185719572057215722572357245725572657275728572957305731573257335734573557365737573857395740574157425743574457455746574757485749575057515752575357545755575657575758575957605761576257635764576557665767576857695770577157725773577457755776577757785779578057815782578357845785578657875788578957905791579257935794579557965797579857995800580158025803580458055806580758085809581058115812581358145815581658175818581958205821582258235824582558265827582858295830583158325833583458355836583758385839584058415842584358445845584658475848584958505851585258535854585558565857585858595860586158625863586458655866586758685869587058715872587358745875587658775878587958805881588258835884588558865887588858895890589158925893589458955896589758985899590059015902590359045905590659075908590959105911591259135914591559165917591859195920592159225923592459255926592759285929593059315932593359345935593659375938593959405941594259435944594559465947594859495950595159525953595459555956595759585959596059615962596359645965596659675968596959705971597259735974597559765977597859795980598159825983598459855986598759885989599059915992599359945995599659975998599960006001600260036004600560066007600860096010601160126013601460156016601760186019602060216022602360246025602660276028602960306031603260336034603560366037603860396040604160426043604460456046604760486049605060516052605360546055605660576058605960606061606260636064606560666067606860696070607160726073607460756076607760786079608060816082608360846085608660876088608960906091609260936094609560966097609860996100610161026103610461056106610761086109611061116112611361146115611661176118611961206121612261236124612561266127612861296130613161326133613461356136613761386139614061416142614361446145614661476148614961506151615261536154615561566157615861596160616161626163616461656166616761686169617061716172617361746175617661776178617961806181618261836184618561866187618861896190619161926193619461956196619761986199620062016202620362046205620662076208620962106211621262136214621562166217621862196220622162226223622462256226622762286229623062316232623362346235623662376238623962406241624262436244624562466247624862496250625162526253625462556256625762586259626062616262626362646265626662676268626962706271627262736274627562766277627862796280628162826283628462856286628762886289629062916292629362946295629662976298629963006301630263036304630563066307630863096310631163126313631463156316631763186319632063216322632363246325632663276328632963306331633263336334633563366337633863396340634163426343634463456346634763486349635063516352635363546355635663576358635963606361636263636364636563666367636863696370637163726373637463756376637763786379638063816382638363846385638663876388638963906391639263936394639563966397639863996400640164026403640464056406640764086409641064116412641364146415641664176418641964206421642264236424642564266427642864296430643164326433643464356436643764386439644064416442644364446445644664476448644964506451645264536454645564566457645864596460646164626463646464656466646764686469647064716472647364746475647664776478647964806481648264836484648564866487648864896490649164926493649464956496649764986499650065016502650365046505650665076508650965106511651265136514651565166517651865196520652165226523652465256526652765286529653065316532653365346535653665376538653965406541654265436544654565466547654865496550655165526553655465556556655765586559656065616562656365646565656665676568656965706571657265736574657565766577657865796580658165826583658465856586658765886589659065916592659365946595659665976598659966006601660266036604660566066607660866096610661166126613661466156616661766186619662066216622662366246625662666276628662966306631663266336634663566366637663866396640664166426643664466456646664766486649665066516652665366546655665666576658665966606661666266636664666566666667666866696670667166726673667466756676667766786679668066816682668366846685668666876688668966906691669266936694669566966697669866996700670167026703670467056706670767086709671067116712671367146715671667176718671967206721672267236724672567266727672867296730673167326733673467356736673767386739674067416742674367446745674667476748674967506751675267536754675567566757675867596760676167626763676467656766676767686769677067716772677367746775677667776778677967806781678267836784678567866787678867896790679167926793679467956796679767986799680068016802680368046805680668076808680968106811681268136814681568166817681868196820682168226823682468256826682768286829683068316832683368346835683668376838683968406841684268436844684568466847684868496850685168526853685468556856685768586859686068616862686368646865686668676868686968706871687268736874687568766877687868796880688168826883688468856886688768886889689068916892689368946895689668976898689969006901690269036904690569066907690869096910691169126913691469156916691769186919692069216922692369246925692669276928692969306931693269336934693569366937693869396940694169426943694469456946694769486949695069516952695369546955695669576958695969606961696269636964696569666967696869696970697169726973697469756976697769786979698069816982698369846985698669876988698969906991699269936994699569966997699869997000700170027003700470057006700770087009701070117012701370147015701670177018701970207021702270237024702570267027702870297030703170327033703470357036703770387039704070417042704370447045704670477048704970507051705270537054705570567057705870597060706170627063706470657066706770687069707070717072707370747075707670777078707970807081708270837084708570867087708870897090709170927093709470957096709770987099710071017102710371047105710671077108710971107111711271137114711571167117711871197120712171227123712471257126712771287129713071317132713371347135713671377138713971407141714271437144714571467147714871497150715171527153715471557156715771587159716071617162716371647165716671677168716971707171717271737174717571767177717871797180718171827183718471857186718771887189719071917192719371947195719671977198719972007201720272037204720572067207720872097210721172127213721472157216721772187219722072217222722372247225722672277228722972307231723272337234723572367237723872397240724172427243724472457246724772487249725072517252725372547255725672577258725972607261726272637264726572667267726872697270727172727273727472757276727772787279728072817282728372847285728672877288728972907291729272937294729572967297729872997300730173027303730473057306730773087309731073117312731373147315731673177318731973207321732273237324732573267327732873297330733173327333733473357336733773387339734073417342734373447345734673477348734973507351735273537354735573567357735873597360736173627363736473657366736773687369737073717372737373747375737673777378737973807381738273837384738573867387738873897390739173927393739473957396739773987399740074017402740374047405740674077408740974107411741274137414741574167417741874197420742174227423742474257426742774287429743074317432743374347435743674377438743974407441744274437444744574467447744874497450745174527453745474557456745774587459746074617462746374647465746674677468746974707471747274737474747574767477747874797480748174827483748474857486748774887489749074917492749374947495749674977498749975007501750275037504750575067507750875097510751175127513751475157516751775187519752075217522752375247525752675277528752975307531753275337534753575367537753875397540754175427543754475457546754775487549755075517552755375547555755675577558755975607561756275637564756575667567756875697570757175727573757475757576757775787579758075817582758375847585758675877588758975907591759275937594759575967597759875997600760176027603760476057606760776087609761076117612761376147615761676177618761976207621762276237624762576267627762876297630763176327633763476357636763776387639764076417642764376447645764676477648764976507651765276537654765576567657765876597660766176627663766476657666766776687669767076717672767376747675767676777678767976807681768276837684768576867687768876897690769176927693769476957696769776987699770077017702770377047705770677077708770977107711771277137714771577167717771877197720772177227723772477257726772777287729773077317732773377347735773677377738773977407741774277437744774577467747774877497750775177527753775477557756775777587759776077617762776377647765776677677768776977707771777277737774777577767777777877797780778177827783778477857786778777887789779077917792779377947795779677977798779978007801780278037804780578067807780878097810781178127813781478157816781778187819782078217822782378247825782678277828782978307831783278337834783578367837783878397840784178427843784478457846784778487849785078517852785378547855785678577858785978607861786278637864786578667867786878697870787178727873787478757876787778787879788078817882788378847885788678877888788978907891789278937894789578967897789878997900790179027903790479057906790779087909791079117912791379147915791679177918791979207921792279237924792579267927792879297930793179327933793479357936793779387939794079417942794379447945794679477948794979507951795279537954795579567957795879597960796179627963796479657966796779687969797079717972797379747975797679777978797979807981798279837984798579867987798879897990799179927993799479957996799779987999800080018002800380048005800680078008800980108011801280138014801580168017801880198020802180228023802480258026802780288029803080318032803380348035803680378038803980408041804280438044804580468047804880498050805180528053805480558056805780588059806080618062806380648065806680678068806980708071807280738074807580768077807880798080808180828083808480858086808780888089809080918092809380948095809680978098809981008101810281038104810581068107810881098110811181128113811481158116811781188119812081218122812381248125812681278128812981308131813281338134813581368137813881398140814181428143814481458146814781488149815081518152815381548155815681578158815981608161816281638164816581668167816881698170817181728173817481758176817781788179818081818182818381848185818681878188818981908191819281938194819581968197819881998200820182028203820482058206820782088209821082118212821382148215821682178218821982208221822282238224822582268227822882298230823182328233823482358236823782388239824082418242824382448245824682478248824982508251825282538254825582568257825882598260826182628263826482658266826782688269827082718272827382748275827682778278827982808281828282838284828582868287828882898290829182928293829482958296829782988299830083018302830383048305830683078308830983108311831283138314831583168317831883198320832183228323832483258326832783288329833083318332833383348335833683378338833983408341834283438344834583468347834883498350835183528353835483558356835783588359836083618362836383648365836683678368836983708371837283738374837583768377837883798380838183828383838483858386838783888389839083918392839383948395839683978398839984008401840284038404840584068407840884098410841184128413841484158416841784188419842084218422842384248425842684278428842984308431843284338434843584368437843884398440844184428443844484458446844784488449845084518452845384548455845684578458845984608461846284638464846584668467846884698470847184728473847484758476847784788479848084818482848384848485848684878488848984908491849284938494849584968497849884998500850185028503850485058506850785088509851085118512851385148515851685178518851985208521852285238524852585268527852885298530853185328533853485358536853785388539854085418542854385448545854685478548854985508551855285538554855585568557855885598560856185628563856485658566856785688569857085718572857385748575857685778578857985808581858285838584858585868587858885898590859185928593859485958596859785988599860086018602860386048605860686078608860986108611861286138614861586168617861886198620862186228623862486258626862786288629863086318632863386348635863686378638863986408641864286438644864586468647864886498650865186528653865486558656865786588659866086618662866386648665866686678668866986708671867286738674867586768677867886798680868186828683868486858686868786888689869086918692869386948695869686978698869987008701870287038704870587068707870887098710871187128713871487158716871787188719872087218722872387248725872687278728872987308731873287338734873587368737873887398740874187428743874487458746874787488749875087518752875387548755875687578758875987608761876287638764876587668767876887698770877187728773877487758776877787788779878087818782878387848785878687878788878987908791879287938794879587968797879887998800880188028803880488058806880788088809881088118812881388148815881688178818881988208821882288238824882588268827882888298830883188328833883488358836883788388839884088418842884388448845884688478848884988508851885288538854885588568857885888598860886188628863886488658866886788688869887088718872887388748875887688778878887988808881888288838884888588868887888888898890889188928893889488958896889788988899890089018902890389048905890689078908890989108911891289138914891589168917891889198920892189228923892489258926892789288929893089318932893389348935893689378938893989408941894289438944894589468947894889498950895189528953895489558956895789588959896089618962896389648965896689678968896989708971897289738974897589768977897889798980898189828983898489858986898789888989899089918992899389948995899689978998899990009001900290039004900590069007900890099010901190129013901490159016901790189019902090219022902390249025902690279028902990309031903290339034903590369037903890399040904190429043904490459046904790489049905090519052905390549055905690579058905990609061906290639064906590669067906890699070907190729073907490759076907790789079908090819082908390849085908690879088908990909091909290939094909590969097909890999100910191029103910491059106910791089109911091119112911391149115911691179118911991209121912291239124912591269127912891299130913191329133913491359136913791389139914091419142914391449145914691479148914991509151915291539154915591569157915891599160916191629163916491659166916791689169917091719172917391749175917691779178917991809181918291839184918591869187918891899190919191929193919491959196919791989199920092019202920392049205920692079208920992109211921292139214921592169217921892199220922192229223922492259226922792289229923092319232923392349235923692379238923992409241924292439244924592469247924892499250925192529253925492559256925792589259926092619262926392649265926692679268926992709271927292739274927592769277927892799280928192829283928492859286928792889289929092919292929392949295929692979298929993009301930293039304930593069307930893099310931193129313931493159316931793189319932093219322932393249325932693279328932993309331933293339334933593369337933893399340934193429343934493459346934793489349935093519352935393549355935693579358935993609361936293639364936593669367936893699370937193729373937493759376937793789379938093819382938393849385938693879388938993909391939293939394939593969397939893999400940194029403940494059406940794089409941094119412941394149415941694179418941994209421942294239424942594269427942894299430943194329433943494359436943794389439944094419442944394449445944694479448944994509451945294539454945594569457945894599460946194629463946494659466946794689469947094719472947394749475947694779478947994809481948294839484948594869487948894899490949194929493949494959496949794989499950095019502950395049505950695079508950995109511951295139514951595169517951895199520952195229523952495259526952795289529953095319532953395349535953695379538953995409541954295439544954595469547954895499550955195529553955495559556955795589559956095619562956395649565956695679568956995709571957295739574957595769577957895799580958195829583958495859586958795889589959095919592959395949595959695979598959996009601960296039604960596069607960896099610961196129613961496159616961796189619962096219622962396249625962696279628962996309631963296339634963596369637963896399640964196429643964496459646964796489649965096519652965396549655965696579658965996609661966296639664966596669667966896699670967196729673967496759676967796789679968096819682968396849685968696879688968996909691969296939694969596969697969896999700970197029703970497059706970797089709971097119712971397149715971697179718971997209721972297239724972597269727972897299730973197329733973497359736973797389739974097419742974397449745974697479748974997509751975297539754975597569757975897599760976197629763976497659766976797689769977097719772977397749775977697779778977997809781978297839784978597869787978897899790979197929793979497959796979797989799980098019802980398049805980698079808980998109811981298139814981598169817981898199820982198229823982498259826982798289829983098319832983398349835983698379838983998409841984298439844984598469847984898499850985198529853985498559856985798589859986098619862986398649865986698679868986998709871987298739874987598769877987898799880988198829883988498859886988798889889989098919892989398949895989698979898989999009901990299039904990599069907990899099910991199129913991499159916991799189919992099219922992399249925992699279928992999309931993299339934993599369937993899399940994199429943994499459946994799489949995099519952995399549955995699579958995999609961996299639964996599669967996899699970997199729973997499759976997799789979998099819982998399849985998699879988998999909991999299939994999599969997999899991000010001100021000310004100051000610007100081000910010100111001210013100141001510016100171001810019100201002110022100231002410025100261002710028100291003010031100321003310034100351003610037100381003910040100411004210043100441004510046100471004810049100501005110052100531005410055100561005710058100591006010061100621006310064100651006610067100681006910070100711007210073100741007510076100771007810079100801008110082100831008410085100861008710088100891009010091100921009310094100951009610097100981009910100101011010210103101041010510106101071010810109101101011110112101131011410115101161011710118101191012010121101221012310124101251012610127101281012910130101311013210133101341013510136101371013810139101401014110142101431014410145101461014710148101491015010151101521015310154101551015610157101581015910160101611016210163101641016510166101671016810169101701017110172101731017410175101761017710178101791018010181101821018310184101851018610187101881018910190101911019210193101941019510196101971019810199102001020110202102031020410205102061020710208102091021010211102121021310214102151021610217102181021910220102211022210223102241022510226102271022810229102301023110232102331023410235102361023710238102391024010241102421024310244102451024610247102481024910250102511025210253102541025510256102571025810259102601026110262102631026410265102661026710268102691027010271102721027310274102751027610277102781027910280102811028210283102841028510286102871028810289102901029110292102931029410295102961029710298102991030010301103021030310304103051030610307103081030910310103111031210313103141031510316103171031810319103201032110322103231032410325103261032710328103291033010331103321033310334103351033610337103381033910340103411034210343103441034510346103471034810349103501035110352103531035410355103561035710358103591036010361103621036310364103651036610367103681036910370103711037210373103741037510376103771037810379103801038110382103831038410385103861038710388103891039010391103921039310394103951039610397103981039910400104011040210403104041040510406104071040810409104101041110412104131041410415104161041710418104191042010421104221042310424104251042610427104281042910430104311043210433104341043510436104371043810439104401044110442104431044410445104461044710448104491045010451104521045310454104551045610457104581045910460104611046210463104641046510466104671046810469104701047110472104731047410475104761047710478104791048010481104821048310484104851048610487104881048910490104911049210493104941049510496104971049810499105001050110502105031050410505105061050710508105091051010511105121051310514105151051610517105181051910520105211052210523105241052510526105271052810529105301053110532105331053410535105361053710538105391054010541105421054310544105451054610547105481054910550105511055210553105541055510556105571055810559105601056110562105631056410565105661056710568105691057010571105721057310574105751057610577105781057910580105811058210583105841058510586105871058810589105901059110592105931059410595105961059710598105991060010601106021060310604106051060610607106081060910610106111061210613106141061510616106171061810619106201062110622106231062410625106261062710628106291063010631106321063310634106351063610637106381063910640106411064210643106441064510646106471064810649106501065110652106531065410655106561065710658106591066010661106621066310664106651066610667106681066910670106711067210673106741067510676106771067810679106801068110682106831068410685106861068710688106891069010691106921069310694106951069610697106981069910700107011070210703107041070510706107071070810709107101071110712107131071410715107161071710718107191072010721107221072310724107251072610727107281072910730107311073210733107341073510736107371073810739107401074110742107431074410745107461074710748107491075010751107521075310754107551075610757107581075910760107611076210763107641076510766107671076810769107701077110772107731077410775107761077710778107791078010781107821078310784107851078610787107881078910790107911079210793107941079510796107971079810799108001080110802108031080410805108061080710808108091081010811108121081310814108151081610817108181081910820108211082210823108241082510826108271082810829108301083110832108331083410835108361083710838108391084010841108421084310844108451084610847108481084910850108511085210853108541085510856108571085810859108601086110862108631086410865108661086710868108691087010871108721087310874108751087610877108781087910880108811088210883108841088510886108871088810889108901089110892108931089410895108961089710898108991090010901109021090310904109051090610907109081090910910109111091210913109141091510916109171091810919109201092110922109231092410925109261092710928109291093010931109321093310934109351093610937109381093910940109411094210943109441094510946109471094810949109501095110952109531095410955109561095710958109591096010961109621096310964109651096610967109681096910970109711097210973109741097510976109771097810979109801098110982109831098410985109861098710988109891099010991109921099310994109951099610997109981099911000110011100211003110041100511006110071100811009110101101111012110131101411015110161101711018110191102011021110221102311024110251102611027
  1. /*********************************************************************/
  2. /* Copyright 2009, 2010 The University of Texas at Austin. */
  3. /* All rights reserved. */
  4. /* */
  5. /* Redistribution and use in source and binary forms, with or */
  6. /* without modification, are permitted provided that the following */
  7. /* conditions are met: */
  8. /* */
  9. /* 1. Redistributions of source code must retain the above */
  10. /* copyright notice, this list of conditions and the following */
  11. /* disclaimer. */
  12. /* */
  13. /* 2. Redistributions in binary form must reproduce the above */
  14. /* copyright notice, this list of conditions and the following */
  15. /* disclaimer in the documentation and/or other materials */
  16. /* provided with the distribution. */
  17. /* */
  18. /* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
  19. /* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
  20. /* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
  21. /* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
  22. /* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
  23. /* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
  24. /* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
  25. /* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
  26. /* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
  27. /* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
  28. /* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
  29. /* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
  30. /* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
  31. /* POSSIBILITY OF SUCH DAMAGE. */
  32. /* */
  33. /* The views and conclusions contained in the software and */
  34. /* documentation are those of the authors and should not be */
  35. /* interpreted as representing official policies, either expressed */
  36. /* or implied, of The University of Texas at Austin. */
  37. /*********************************************************************/
  38. #define ASSEMBLER
  39. #include "common.h"
  40. #ifdef DOUBLE
  41. #define PREFETCHSIZE (16 * 8)
  42. #else
  43. #define PREFETCHSIZE (32 * 8)
  44. #endif
  45. #define CPREFETCHSIZE 7
  46. #define CPREFETCH lfetch.excl.nt1
  47. #define M r32
  48. #define N r33
  49. #define K r34
  50. #define A r36
  51. #define B r37
  52. #define C r38
  53. #define LDC r39
  54. #define I r15
  55. #define J r16
  56. #define AOFFSET r17
  57. #define BOFFSET r18
  58. #define TEMP r19
  59. #define L r20
  60. #define C1 r21
  61. #define C2 r22
  62. #define C3 r23
  63. #define C4 r24
  64. #define C5 r25
  65. #define C6 r26
  66. #define C7 r27
  67. #define C8 r28
  68. #define C9 loc0
  69. #define C10 loc1
  70. #define C11 loc2
  71. #define C12 loc3
  72. #define C13 loc4
  73. #define C14 loc5
  74. #define C15 loc6
  75. #define C16 loc7
  76. #define PREA r8
  77. #define PREB r9
  78. #define PREC r10
  79. #define SP r12
  80. #define ARLC r29
  81. #define PR r30
  82. #define ARPFS r31
  83. #define ALPHA f8
  84. #define AORIG loc8
  85. #define KK loc9
  86. #define KK8 loc10
  87. #define OFFSET loc11
  88. #define AOFFSET2 loc12
  89. #define BOFFSET2 loc13
  90. PROLOGUE
  91. .prologue
  92. PROFCODE
  93. { .mmi
  94. .save ar.pfs, ARPFS
  95. alloc ARPFS = ar.pfs, 8, 16, 0, 0
  96. adds r14 = 16, SP
  97. mov ARLC = ar.lc
  98. }
  99. { .mmi
  100. adds r8 = -6 * 16, SP
  101. adds r9 = -5 * 16, SP
  102. adds SP = -6 * 16, SP
  103. }
  104. ;;
  105. { .mmi
  106. ld8 OFFSET = [r14]
  107. mov AOFFSET = A
  108. mov PR = pr
  109. }
  110. ;;
  111. { .mmi
  112. stf.spill [r8] = f16, 32
  113. stf.spill [r9] = f17, 32
  114. shr J = N, 3
  115. }
  116. ;;
  117. { .mmi
  118. stf.spill [r8] = f18, 32
  119. stf.spill [r9] = f19, 32
  120. shladd LDC = LDC, BASE_SHIFT, r0
  121. }
  122. ;;
  123. .body
  124. { .mmi
  125. stf.spill [r8] = f20
  126. stf.spill [r9] = f21
  127. cmp.ge p6, p0 = 0, J
  128. }
  129. { .mib
  130. nop __LINE__
  131. #ifdef RN
  132. sub KK = r0, OFFSET
  133. #else
  134. nop __LINE__
  135. #endif
  136. (p6) br.cond.dpnt .L050
  137. }
  138. ;;
  139. .align 8
  140. .L010:
  141. { .mfi
  142. adds J = -1, J
  143. mov f64 = f0
  144. shr I = M, 3
  145. }
  146. { .mfi
  147. mov C1 = C // coffset1 = c + 0 * ldc
  148. mov f72 = f0
  149. #ifdef LT
  150. mov KK = OFFSET
  151. #else
  152. nop __LINE__
  153. #endif
  154. }
  155. ;;
  156. { .mmf
  157. cmp.eq p6, p7 = 0, I
  158. mov AOFFSET = A
  159. mov f80 = f0
  160. }
  161. { .mmf
  162. add C2 = LDC, C // coffset2 = c + 1 * ldc
  163. shladd C3 = LDC, 1, C // coffset3 = c + 2 * ldc
  164. mov f88 = f0
  165. }
  166. ;;
  167. { .mmf
  168. shladd C5 = LDC, 2, C // coffset5 = c + 4 * ldc
  169. shladd C = LDC, 3, C // coffset += 8 * ldc
  170. mov f96 = f0
  171. }
  172. { .mmf
  173. shladd C4 = LDC, 1, C2
  174. shladd C6 = LDC, 2, C2
  175. mov f104 = f0
  176. }
  177. ;;
  178. { .mfi
  179. shladd C7 = LDC, 2, C3
  180. mov f112 = f0
  181. mov L = KK
  182. }{ .mfb
  183. shladd C8 = LDC, 2, C4
  184. mov f120 = f0
  185. (p6) br.cond.dpnt .L020
  186. }
  187. ;;
  188. .align 16
  189. .L011:
  190. { .mmf
  191. cmp.ne p7, p0 = r0, L
  192. adds BOFFSET = 0 * SIZE, B
  193. mov f65 = f0
  194. }
  195. ;;
  196. { .mmf
  197. (p7) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  198. setf.d f73 = r0
  199. mov f81 = f0
  200. }
  201. ;;
  202. { .mmf
  203. (p7) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  204. setf.d f119 = r0
  205. mov f89 = f0
  206. }
  207. { .mmf
  208. (p7) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  209. setf.d f97 = r0
  210. mov f105 = f0
  211. }
  212. ;;
  213. { .mmf
  214. (p7) LDFPD f52, f53 = [BOFFSET], 2 * SIZE
  215. setf.d f113 = r0
  216. mov f121 = f0
  217. }
  218. ;;
  219. { .mmf
  220. (p7) LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  221. setf.d f66 = r0
  222. mov f74 = f0
  223. }
  224. { .mfi
  225. setf.d f82 = r0
  226. mov f90 = f0
  227. nop __LINE__
  228. }
  229. ;;
  230. { .mmf
  231. (p7) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  232. setf.d f98 = r0
  233. mov f106 = f0
  234. }
  235. { .mfi
  236. setf.d f114 = r0
  237. mov f122 = f0
  238. adds L = 1, L
  239. }
  240. ;;
  241. { .mmf
  242. (p7) LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  243. setf.d f67 = r0
  244. mov f75 = f0
  245. }
  246. { .mfi
  247. setf.d f83 = r0
  248. mov f91 = f0
  249. cmp.eq p3, p0 = r0, r0
  250. }
  251. ;;
  252. { .mmf
  253. (p7) LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  254. setf.d f99 = r0
  255. mov f107 = f0
  256. }
  257. { .mfi
  258. setf.d f115 = r0
  259. mov f123 = f0
  260. adds PREC = CPREFETCHSIZE * SIZE, C1
  261. }
  262. ;;
  263. { .mmf
  264. CPREFETCH [PREC], LDC
  265. setf.d f68 = r0
  266. mov f76 = f0
  267. }
  268. { .mfi
  269. setf.d f84 = r0
  270. mov f92 = f0
  271. adds AOFFSET2 = 4 * SIZE, AOFFSET
  272. }
  273. ;;
  274. { .mmf
  275. CPREFETCH [PREC], LDC
  276. setf.d f100 = r0
  277. mov f108 = f0
  278. }
  279. { .mfi
  280. setf.d f116 = r0
  281. mov f124 = f0
  282. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  283. }
  284. ;;
  285. { .mmf
  286. CPREFETCH [PREC], LDC
  287. setf.d f69 = r0
  288. mov f77 = f0
  289. }
  290. { .mfi
  291. setf.d f85 = r0
  292. mov f93 = f0
  293. adds PREB = (PREFETCHSIZE - 8) * SIZE, BOFFSET
  294. }
  295. ;;
  296. { .mmf
  297. CPREFETCH [PREC], LDC
  298. setf.d f101 = r0
  299. mov f109 = f0
  300. }
  301. { .mfi
  302. setf.d f117 = r0
  303. mov f125 = f0
  304. tbit.z p12, p0 = L, 0
  305. }
  306. ;;
  307. { .mmf
  308. CPREFETCH [PREC], LDC
  309. setf.d f70 = r0
  310. mov f78 = f0
  311. }
  312. { .mfi
  313. setf.d f86 = r0
  314. mov f94 = f0
  315. shr L = L, 1
  316. }
  317. ;;
  318. { .mmf
  319. CPREFETCH [PREC], LDC
  320. setf.d f102 = r0
  321. mov f110 = f0
  322. }
  323. { .mfi
  324. setf.d f118 = r0
  325. mov f126 = f0
  326. adds L = -1, L
  327. }
  328. ;;
  329. { .mmf
  330. CPREFETCH [PREC], LDC
  331. setf.d f71 = r0
  332. mov f79 = f0
  333. }
  334. { .mfi
  335. setf.d f87 = r0
  336. mov f95 = f0
  337. mov ar.lc = L
  338. }
  339. ;;
  340. { .mmf
  341. CPREFETCH [PREC]
  342. setf.d f103 = r0
  343. mov f111 = f0
  344. }
  345. { .mfb
  346. cmp.eq p6, p0 = -1, L
  347. mov f127 = f0
  348. (p6) br.cond.dpnt .L018
  349. }
  350. ;;
  351. .align 16
  352. .L012:
  353. /* 1 */
  354. { .mfi
  355. lfetch.fault.nt1 [PREA], 16 * SIZE
  356. FMA f64 = f32, f48, f64 // A1 * B1
  357. nop __LINE__
  358. }
  359. { .mfi
  360. (p12) cmp.ne p3, p0 = 0, L
  361. FMA f72 = f32, f49, f72 // A1 * B2
  362. nop __LINE__
  363. }
  364. ;;
  365. /* 2 */
  366. { .mfb
  367. lfetch.nt1 [PREB], 16 * SIZE
  368. FMA f80 = f32, f50, f80 // A1 * B3
  369. nop __LINE__
  370. }
  371. { .mfb
  372. cmp.ne p4, p5 = 0, L
  373. FMA f88 = f32, f51, f88 // A1 * B4
  374. nop __LINE__
  375. }
  376. ;;
  377. /* 3 */
  378. { .mfb
  379. (p3) LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  380. FMA f96 = f32, f52, f96 // A1 * B5
  381. nop __LINE__
  382. }
  383. { .mfb
  384. adds C9 = 4 * SIZE, C1
  385. FMA f104 = f32, f53, f104 // A1 * B6
  386. nop __LINE__
  387. }
  388. ;;
  389. /* 4 */
  390. { .mfb
  391. (p3) LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  392. FMA f112 = f32, f54, f112 // A1 * B7
  393. nop __LINE__
  394. }
  395. { .mfb
  396. adds C10 = 4 * SIZE, C2
  397. FMA f120 = f32, f55, f120 // A1 * B8
  398. nop __LINE__
  399. }
  400. ;;
  401. /* 5 */
  402. { .mfb
  403. (p3) LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  404. FMA f65 = f33, f48, f65 // A2 * B1
  405. nop __LINE__
  406. }
  407. { .mfb
  408. adds C11 = 4 * SIZE, C3
  409. FMA f73 = f33, f49, f73 // A2 * B2
  410. nop __LINE__
  411. }
  412. ;;
  413. /* 6 */
  414. { .mfb
  415. (p3) LDFPD f60, f61 = [BOFFSET], 2 * SIZE
  416. FMA f81 = f33, f50, f81 // A2 * B3
  417. nop __LINE__
  418. }
  419. { .mfb
  420. adds C12 = 4 * SIZE, C4
  421. FMA f89 = f33, f51, f89 // A2 * B4
  422. nop __LINE__
  423. }
  424. ;;
  425. /* 7 */
  426. { .mfb
  427. (p3) LDFPD f62, f63 = [BOFFSET], 2 * SIZE
  428. FMA f97 = f33, f52, f97 // A2 * B5
  429. nop __LINE__
  430. }
  431. { .mfb
  432. adds C13 = 4 * SIZE, C5
  433. FMA f105 = f33, f53, f105 // A2 * B6
  434. nop __LINE__
  435. }
  436. ;;
  437. /* 8 */
  438. { .mfb
  439. (p3) LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  440. FMA f113 = f33, f54, f113 // A2 * B7
  441. nop __LINE__
  442. }
  443. { .mfb
  444. adds C14 = 4 * SIZE, C6
  445. FMA f121 = f33, f55, f121 // A2 * B8
  446. nop __LINE__
  447. }
  448. ;;
  449. /* 9 */
  450. { .mfb
  451. (p3) LDFPD f44, f45 = [AOFFSET], 2 * SIZE
  452. FMA f66 = f34, f48, f66 // A3 * B1
  453. nop __LINE__
  454. }
  455. { .mfb
  456. adds C15 = 4 * SIZE, C7
  457. FMA f74 = f34, f49, f74 // A3 * B2
  458. nop __LINE__
  459. }
  460. ;;
  461. /* 10 */
  462. { .mfb
  463. (p3) LDFPD f46, f47 = [AOFFSET], 2 * SIZE
  464. FMA f82 = f34, f50, f82 // A3 * B3
  465. nop __LINE__
  466. }
  467. { .mfb
  468. adds C16 = 4 * SIZE, C8
  469. FMA f90 = f34, f51, f90 // A3 * B4
  470. nop __LINE__
  471. }
  472. ;;
  473. /* 11 */
  474. { .mfb
  475. FMA f98 = f34, f52, f98 // A3 * B5
  476. nop __LINE__
  477. }
  478. { .mfb
  479. nop __LINE__
  480. FMA f106 = f34, f53, f106 // A3 * B6
  481. nop __LINE__
  482. }
  483. ;;
  484. /* 12 */
  485. { .mfb
  486. FMA f114 = f34, f54, f114 // A3 * B7
  487. nop __LINE__
  488. }
  489. { .mfb
  490. nop __LINE__
  491. FMA f122 = f34, f55, f122 // A3 * B8
  492. nop __LINE__
  493. }
  494. ;;
  495. /* 13 */
  496. { .mfb
  497. nop __LINE__
  498. FMA f67 = f35, f48, f67 // A4 * B1
  499. }
  500. { .mfb
  501. nop __LINE__
  502. FMA f75 = f35, f49, f75 // A4 * B2
  503. nop __LINE__
  504. }
  505. ;;
  506. /* 14 */
  507. { .mfb
  508. FMA f83 = f35, f50, f83 // A4 * B3
  509. nop __LINE__
  510. }
  511. { .mfb
  512. nop __LINE__
  513. FMA f91 = f35, f51, f91 // A4 * B4
  514. nop __LINE__
  515. }
  516. ;;
  517. /* 15 */
  518. { .mfb
  519. FMA f99 = f35, f52, f99 // A4 * B5
  520. nop __LINE__
  521. }
  522. { .mfb
  523. nop __LINE__
  524. FMA f107 = f35, f53, f107 // A4 * B6
  525. nop __LINE__
  526. }
  527. ;;
  528. /* 16 */
  529. { .mfb
  530. FMA f115 = f35, f54, f115 // A4 * B7
  531. nop __LINE__
  532. }
  533. { .mfb
  534. nop __LINE__
  535. FMA f123 = f35, f55, f123 // A4 * B8
  536. nop __LINE__
  537. }
  538. ;;
  539. /* 17 */
  540. { .mfb
  541. nop __LINE__
  542. FMA f68 = f36, f48, f68 // A5 * B1
  543. nop __LINE__
  544. }
  545. { .mfb
  546. nop __LINE__
  547. FMA f76 = f36, f49, f76 // A5 * B2
  548. nop __LINE__
  549. }
  550. ;;
  551. /* 18 */
  552. { .mfb
  553. nop __LINE__
  554. FMA f84 = f36, f50, f84 // A5 * B3
  555. nop __LINE__
  556. }
  557. { .mfb
  558. nop __LINE__
  559. FMA f92 = f36, f51, f92 // A5 * B4
  560. nop __LINE__
  561. }
  562. ;;
  563. /* 19 */
  564. { .mfb
  565. nop __LINE__
  566. FMA f100 = f36, f52, f100 // A5 * B5
  567. nop __LINE__
  568. }
  569. { .mfb
  570. nop __LINE__
  571. FMA f108 = f36, f53, f108 // A5 * B6
  572. nop __LINE__
  573. }
  574. ;;
  575. /* 20 */
  576. { .mfb
  577. nop __LINE__
  578. FMA f116 = f36, f54, f116 // A5 * B7
  579. nop __LINE__
  580. }
  581. { .mfb
  582. nop __LINE__
  583. FMA f124 = f36, f55, f124 // A5 * B8
  584. nop __LINE__
  585. }
  586. ;;
  587. /* 21 */
  588. { .mfb
  589. nop __LINE__
  590. FMA f69 = f37, f48, f69 // A6 * B1
  591. nop __LINE__
  592. }
  593. { .mfb
  594. nop __LINE__
  595. FMA f77 = f37, f49, f77 // A6 * B2
  596. nop __LINE__
  597. }
  598. ;;
  599. /* 22 */
  600. { .mfb
  601. nop __LINE__
  602. FMA f85 = f37, f50, f85 // A6 * B3
  603. nop __LINE__
  604. }
  605. { .mfb
  606. nop __LINE__
  607. FMA f93 = f37, f51, f93 // A6 * B4
  608. nop __LINE__
  609. }
  610. ;;
  611. /* 23 */
  612. { .mfb
  613. nop __LINE__
  614. FMA f101 = f37, f52, f101 // A6 * B5
  615. nop __LINE__
  616. }
  617. { .mfb
  618. nop __LINE__
  619. FMA f109 = f37, f53, f109 // A6 * B6
  620. nop __LINE__
  621. }
  622. ;;
  623. /* 24 */
  624. { .mfb
  625. nop __LINE__
  626. FMA f117 = f37, f54, f117 // A6 * B7
  627. nop __LINE__
  628. }
  629. { .mfb
  630. nop __LINE__
  631. FMA f125 = f37, f55, f125 // A6 * B8
  632. nop __LINE__
  633. }
  634. ;;
  635. /* 25 */
  636. { .mfb
  637. nop __LINE__
  638. FMA f70 = f38, f48, f70 // A7 * B1
  639. nop __LINE__
  640. }
  641. { .mfb
  642. nop __LINE__
  643. FMA f78 = f38, f49, f78 // A7 * B2
  644. nop __LINE__
  645. }
  646. ;;
  647. /* 26 */
  648. { .mfb
  649. nop __LINE__
  650. FMA f86 = f38, f50, f86 // A7 * B3
  651. nop __LINE__
  652. }
  653. { .mfb
  654. nop __LINE__
  655. FMA f94 = f38, f51, f94 // A7 * B4
  656. nop __LINE__
  657. }
  658. ;;
  659. /* 27 */
  660. { .mfb
  661. nop __LINE__
  662. FMA f102 = f38, f52, f102 // A7 * B5
  663. nop __LINE__
  664. }
  665. { .mfb
  666. nop __LINE__
  667. FMA f110 = f38, f53, f110 // A7 * B6
  668. nop __LINE__
  669. }
  670. ;;
  671. /* 28 */
  672. { .mfb
  673. nop __LINE__
  674. FMA f118 = f38, f54, f118 // A7 * B7
  675. nop __LINE__
  676. }
  677. { .mfb
  678. nop __LINE__
  679. FMA f126 = f38, f55, f126 // A7 * B8
  680. nop __LINE__
  681. }
  682. ;;
  683. /* 29 */
  684. { .mfb
  685. nop __LINE__
  686. FMA f71 = f39, f48, f71 // A8 * B1
  687. nop __LINE__
  688. }
  689. { .mfb
  690. nop __LINE__
  691. FMA f79 = f39, f49, f79 // A8 * B2
  692. nop __LINE__
  693. }
  694. ;;
  695. /* 30 */
  696. { .mfb
  697. (p4) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  698. FMA f87 = f39, f50, f87 // A8 * B3
  699. nop __LINE__
  700. }
  701. { .mfb
  702. nop __LINE__
  703. FMA f95 = f39, f51, f95 // A8 * B4
  704. nop __LINE__
  705. }
  706. ;;
  707. /* 31 */
  708. { .mfb
  709. (p4) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  710. FMA f103 = f39, f52, f103 // A8 * B5
  711. nop __LINE__
  712. }
  713. { .mfb
  714. nop __LINE__
  715. FMA f111 = f39, f53, f111 // A8 * B6
  716. nop __LINE__
  717. }
  718. ;;
  719. /* 32 */
  720. { .mfb
  721. nop __LINE__
  722. FMA f119 = f39, f54, f119 // A8 * B7
  723. nop __LINE__
  724. }
  725. { .mfb
  726. nop __LINE__
  727. FMA f127 = f39, f55, f127 // A8 * B8
  728. nop __LINE__
  729. }
  730. ;;
  731. /* 33 */
  732. { .mfb
  733. nop __LINE__
  734. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  735. nop __LINE__
  736. }
  737. { .mfb
  738. nop __LINE__
  739. (p3) FMA f72 = f40, f57, f72 // A1 * B2
  740. nop __LINE__
  741. }
  742. ;;
  743. /* 34 */
  744. { .mfb
  745. (p4) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  746. (p3) FMA f80 = f40, f58, f80 // A1 * B3
  747. nop __LINE__
  748. }
  749. { .mfb
  750. nop __LINE__
  751. (p3) FMA f88 = f40, f59, f88 // A1 * B4
  752. nop __LINE__
  753. }
  754. ;;
  755. /* 35 */
  756. { .mfb
  757. (p4) LDFPD f52, f53 = [BOFFSET], 2 * SIZE
  758. (p3) FMA f96 = f40, f60, f96 // A1 * B5
  759. nop __LINE__
  760. }
  761. { .mfb
  762. nop __LINE__
  763. (p3) FMA f104 = f40, f61, f104 // A1 * B6
  764. nop __LINE__
  765. }
  766. ;;
  767. /* 36 */
  768. { .mfb
  769. (p4) LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  770. (p3) FMA f112 = f40, f62, f112 // A1 * B7
  771. nop __LINE__
  772. }
  773. { .mfb
  774. nop __LINE__
  775. (p3) FMA f120 = f40, f63, f120 // A1 * B8
  776. nop __LINE__
  777. }
  778. ;;
  779. /* 37 */
  780. { .mfb
  781. (p4) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  782. (p3) FMA f65 = f41, f56, f65 // A2 * B1
  783. nop __LINE__
  784. }
  785. { .mfb
  786. nop __LINE__
  787. (p3) FMA f73 = f41, f57, f73 // A2 * B2
  788. nop __LINE__
  789. }
  790. ;;
  791. /* 38 */
  792. { .mfb
  793. (p4) LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  794. (p3) FMA f81 = f41, f58, f81 // A2 * B3
  795. nop __LINE__
  796. }
  797. { .mfb
  798. nop __LINE__
  799. (p3) FMA f89 = f41, f59, f89 // A2 * B4
  800. nop __LINE__
  801. }
  802. ;;
  803. /* 39 */
  804. { .mfb
  805. (p4) LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  806. (p3) FMA f97 = f41, f60, f97 // A2 * B5
  807. nop __LINE__
  808. }
  809. { .mfb
  810. nop __LINE__
  811. (p3) FMA f105 = f41, f61, f105 // A2 * B6
  812. nop __LINE__
  813. }
  814. ;;
  815. /* 40 */
  816. { .mfb
  817. nop __LINE__
  818. (p3) FMA f113 = f41, f62, f113 // A2 * B7
  819. nop __LINE__
  820. }
  821. { .mfb
  822. nop __LINE__
  823. (p3) FMA f121 = f41, f63, f121 // A2 * B8
  824. nop __LINE__
  825. }
  826. ;;
  827. /* 41 */
  828. { .mfb
  829. nop __LINE__
  830. (p3) FMA f66 = f42, f56, f66 // A3 * B1
  831. nop __LINE__
  832. }
  833. { .mfb
  834. nop __LINE__
  835. (p3) FMA f74 = f42, f57, f74 // A3 * B2
  836. nop __LINE__
  837. }
  838. ;;
  839. /* 42 */
  840. { .mfb
  841. nop __LINE__
  842. (p3) FMA f82 = f42, f58, f82 // A3 * B3
  843. nop __LINE__
  844. }
  845. { .mfb
  846. nop __LINE__
  847. (p3) FMA f90 = f42, f59, f90 // A3 * B4
  848. nop __LINE__
  849. }
  850. ;;
  851. /* 43 */
  852. { .mfb
  853. nop __LINE__
  854. (p3) FMA f98 = f42, f60, f98 // A3 * B5
  855. nop __LINE__
  856. }
  857. { .mfb
  858. nop __LINE__
  859. (p3) FMA f106 = f42, f61, f106 // A3 * B6
  860. nop __LINE__
  861. }
  862. ;;
  863. /* 44 */
  864. { .mfb
  865. nop __LINE__
  866. (p3) FMA f114 = f42, f62, f114 // A3 * B7
  867. nop __LINE__
  868. }
  869. { .mfb
  870. nop __LINE__
  871. (p3) FMA f122 = f42, f63, f122 // A3 * B8
  872. nop __LINE__
  873. }
  874. ;;
  875. /* 45 */
  876. { .mfb
  877. nop __LINE__
  878. (p3) FMA f67 = f43, f56, f67 // A4 * B1
  879. nop __LINE__
  880. }
  881. { .mfb
  882. nop __LINE__
  883. (p3) FMA f75 = f43, f57, f75 // A4 * B2
  884. nop __LINE__
  885. }
  886. ;;
  887. /* 46 */
  888. { .mfb
  889. nop __LINE__
  890. (p3) FMA f83 = f43, f58, f83 // A4 * B3
  891. nop __LINE__
  892. }
  893. { .mfb
  894. nop __LINE__
  895. (p3) FMA f91 = f43, f59, f91 // A4 * B4
  896. nop __LINE__
  897. }
  898. ;;
  899. /* 47 */
  900. { .mfb
  901. nop __LINE__
  902. (p3) FMA f99 = f43, f60, f99 // A4 * B5
  903. nop __LINE__
  904. }
  905. { .mfb
  906. nop __LINE__
  907. (p3) FMA f107 = f43, f61, f107 // A4 * B6
  908. nop __LINE__
  909. }
  910. ;;
  911. /* 48 */
  912. { .mfb
  913. nop __LINE__
  914. (p3) FMA f115 = f43, f62, f115 // A4 * B7
  915. nop __LINE__
  916. }
  917. { .mfb
  918. nop __LINE__
  919. (p3) FMA f123 = f43, f63, f123 // A4 * B8
  920. nop __LINE__
  921. }
  922. ;;
  923. /* 49 */
  924. { .mfb
  925. nop __LINE__
  926. (p3) FMA f68 = f44, f56, f68 // A5 * B1
  927. nop __LINE__
  928. }
  929. { .mfb
  930. nop __LINE__
  931. (p3) FMA f76 = f44, f57, f76 // A5 * B2
  932. nop __LINE__
  933. }
  934. ;;
  935. /* 50 */
  936. { .mfb
  937. nop __LINE__
  938. (p3) FMA f84 = f44, f58, f84 // A5 * B3
  939. nop __LINE__
  940. }
  941. { .mfb
  942. nop __LINE__
  943. (p3) FMA f92 = f44, f59, f92 // A5 * B4
  944. nop __LINE__
  945. }
  946. ;;
  947. /* 51 */
  948. { .mfb
  949. nop __LINE__
  950. (p3) FMA f100 = f44, f60, f100 // A5 * B5
  951. nop __LINE__
  952. }
  953. { .mfb
  954. nop __LINE__
  955. (p3) FMA f108 = f44, f61, f108 // A5 * B6
  956. nop __LINE__
  957. }
  958. ;;
  959. /* 52 */
  960. { .mfb
  961. nop __LINE__
  962. (p3) FMA f116 = f44, f62, f116 // A5 * B7
  963. nop __LINE__
  964. }
  965. { .mfb
  966. nop __LINE__
  967. (p3) FMA f124 = f44, f63, f124 // A5 * B8
  968. nop __LINE__
  969. }
  970. ;;
  971. /* 53 */
  972. { .mfb
  973. nop __LINE__
  974. (p3) FMA f69 = f45, f56, f69 // A6 * B1
  975. nop __LINE__
  976. }
  977. { .mfb
  978. nop __LINE__
  979. (p3) FMA f77 = f45, f57, f77 // A6 * B2
  980. nop __LINE__
  981. }
  982. ;;
  983. /* 54 */
  984. { .mfb
  985. nop __LINE__
  986. (p3) FMA f85 = f45, f58, f85 // A6 * B3
  987. nop __LINE__
  988. }
  989. { .mfb
  990. nop __LINE__
  991. (p3) FMA f93 = f45, f59, f93 // A6 * B4
  992. nop __LINE__
  993. }
  994. ;;
  995. /* 55 */
  996. { .mfb
  997. nop __LINE__
  998. (p3) FMA f101 = f45, f60, f101 // A6 * B5
  999. nop __LINE__
  1000. }
  1001. { .mfb
  1002. nop __LINE__
  1003. (p3) FMA f109 = f45, f61, f109 // A6 * B6
  1004. nop __LINE__
  1005. }
  1006. ;;
  1007. /* 56 */
  1008. { .mfb
  1009. nop __LINE__
  1010. (p3) FMA f117 = f45, f62, f117 // A6 * B7
  1011. nop __LINE__
  1012. }
  1013. { .mfb
  1014. nop __LINE__
  1015. (p3) FMA f125 = f45, f63, f125 // A6 * B8
  1016. nop __LINE__
  1017. }
  1018. ;;
  1019. /* 57 */
  1020. { .mfb
  1021. nop __LINE__
  1022. (p3) FMA f70 = f46, f56, f70 // A7 * B1
  1023. nop __LINE__
  1024. }
  1025. { .mfb
  1026. nop __LINE__
  1027. (p3) FMA f78 = f46, f57, f78 // A7 * B2
  1028. nop __LINE__
  1029. }
  1030. ;;
  1031. /* 58 */
  1032. { .mfb
  1033. nop __LINE__
  1034. (p3) FMA f86 = f46, f58, f86 // A7 * B3
  1035. nop __LINE__
  1036. }
  1037. { .mfb
  1038. nop __LINE__
  1039. (p3) FMA f94 = f46, f59, f94 // A7 * B4
  1040. nop __LINE__
  1041. }
  1042. ;;
  1043. /* 59 */
  1044. { .mfb
  1045. nop __LINE__
  1046. (p3) FMA f102 = f46, f60, f102 // A7 * B5
  1047. nop __LINE__
  1048. }
  1049. { .mfb
  1050. nop __LINE__
  1051. (p3) FMA f110 = f46, f61, f110 // A7 * B6
  1052. nop __LINE__
  1053. }
  1054. ;;
  1055. /* 60 */
  1056. { .mfb
  1057. nop __LINE__
  1058. (p3) FMA f118 = f46, f62, f118 // A7 * B7
  1059. nop __LINE__
  1060. }
  1061. { .mfb
  1062. nop __LINE__
  1063. (p3) FMA f126 = f46, f63, f126 // A7 * B8
  1064. nop __LINE__
  1065. }
  1066. ;;
  1067. /* 61 */
  1068. { .mfb
  1069. nop __LINE__
  1070. (p3) FMA f71 = f47, f56, f71 // A8 * B1
  1071. nop __LINE__
  1072. }
  1073. { .mfb
  1074. nop __LINE__
  1075. (p3) FMA f79 = f47, f57, f79 // A8 * B2
  1076. nop __LINE__
  1077. }
  1078. ;;
  1079. /* 62 */
  1080. { .mfb
  1081. nop __LINE__
  1082. (p3) FMA f87 = f47, f58, f87 // A8 * B3
  1083. nop __LINE__
  1084. }
  1085. { .mfb
  1086. nop __LINE__
  1087. (p3) FMA f95 = f47, f59, f95 // A8 * B4
  1088. nop __LINE__
  1089. }
  1090. ;;
  1091. /* 63 */
  1092. { .mfb
  1093. nop __LINE__
  1094. (p3) FMA f103 = f47, f60, f103 // A8 * B5
  1095. nop __LINE__
  1096. }
  1097. { .mfb
  1098. nop __LINE__
  1099. (p3) FMA f111 = f47, f61, f111 // A8 * B6
  1100. nop __LINE__
  1101. }
  1102. ;;
  1103. /* 64 */
  1104. { .mfi
  1105. nop __LINE__
  1106. (p3) FMA f119 = f47, f62, f119 // A8 * B7
  1107. adds L = -1, L
  1108. }
  1109. { .mfb
  1110. adds AOFFSET2 = 4 * SIZE, AOFFSET
  1111. (p3) FMA f127 = f47, f63, f127 // A8 * B8
  1112. br.cloop.sptk.few .L012
  1113. }
  1114. ;;
  1115. .L018:
  1116. #ifdef LT
  1117. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  1118. ;;
  1119. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  1120. ;;
  1121. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  1122. ;;
  1123. LDFPD f38, f39 = [BOFFSET], 2 * SIZE
  1124. ;;
  1125. LDFPD f40, f41 = [BOFFSET], 2 * SIZE
  1126. ;;
  1127. LDFPD f42, f43 = [BOFFSET], 2 * SIZE
  1128. ;;
  1129. LDFPD f44, f45 = [BOFFSET], 2 * SIZE
  1130. ;;
  1131. LDFPD f46, f47 = [BOFFSET], 2 * SIZE
  1132. ;;
  1133. { .mfi
  1134. LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  1135. FSUB f64 = f32, f64
  1136. nop __LINE__
  1137. }
  1138. { .mfi
  1139. nop __LINE__
  1140. FSUB f72 = f33, f72
  1141. nop __LINE__
  1142. }
  1143. ;;
  1144. { .mfi
  1145. LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  1146. FSUB f80 = f34, f80
  1147. nop __LINE__
  1148. }
  1149. { .mfi
  1150. nop __LINE__
  1151. FSUB f88 = f35, f88
  1152. nop __LINE__
  1153. }
  1154. ;;
  1155. { .mfi
  1156. LDFPD f52, f53 = [BOFFSET], 2 * SIZE
  1157. FSUB f96 = f36, f96
  1158. nop __LINE__
  1159. }
  1160. { .mfi
  1161. nop __LINE__
  1162. FSUB f104 = f37, f104
  1163. nop __LINE__
  1164. }
  1165. ;;
  1166. { .mfi
  1167. LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  1168. FSUB f112 = f38, f112
  1169. nop __LINE__
  1170. }
  1171. { .mfi
  1172. nop __LINE__
  1173. FSUB f120 = f39, f120
  1174. nop __LINE__
  1175. }
  1176. ;;
  1177. { .mfi
  1178. LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  1179. FSUB f65 = f40, f65
  1180. nop __LINE__
  1181. }
  1182. { .mfi
  1183. nop __LINE__
  1184. FSUB f73 = f41, f73
  1185. nop __LINE__
  1186. }
  1187. ;;
  1188. { .mfi
  1189. LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  1190. FSUB f81 = f42, f81
  1191. nop __LINE__
  1192. }
  1193. { .mfi
  1194. nop __LINE__
  1195. FSUB f89 = f43, f89
  1196. nop __LINE__
  1197. }
  1198. ;;
  1199. { .mfi
  1200. LDFPD f60, f61 = [BOFFSET], 2 * SIZE
  1201. FSUB f97 = f44, f97
  1202. nop __LINE__
  1203. }
  1204. { .mfi
  1205. nop __LINE__
  1206. FSUB f105 = f45, f105
  1207. nop __LINE__
  1208. }
  1209. ;;
  1210. { .mfi
  1211. LDFPD f62, f63 = [BOFFSET], 2 * SIZE
  1212. FSUB f113 = f46, f113
  1213. nop __LINE__
  1214. }
  1215. { .mfi
  1216. nop __LINE__
  1217. FSUB f121 = f47, f121
  1218. nop __LINE__
  1219. }
  1220. ;;
  1221. { .mfi
  1222. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  1223. FSUB f66 = f48, f66
  1224. nop __LINE__
  1225. }
  1226. { .mfi
  1227. nop __LINE__
  1228. FSUB f74 = f49, f74
  1229. nop __LINE__
  1230. }
  1231. ;;
  1232. { .mfi
  1233. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  1234. FSUB f82 = f50, f82
  1235. nop __LINE__
  1236. }
  1237. { .mfi
  1238. nop __LINE__
  1239. FSUB f90 = f51, f90
  1240. nop __LINE__
  1241. }
  1242. ;;
  1243. { .mfi
  1244. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  1245. FSUB f98 = f52, f98
  1246. nop __LINE__
  1247. }
  1248. { .mfi
  1249. nop __LINE__
  1250. FSUB f106 = f53, f106
  1251. nop __LINE__
  1252. }
  1253. ;;
  1254. { .mfi
  1255. LDFPD f38, f39 = [BOFFSET], 2 * SIZE
  1256. FSUB f114 = f54, f114
  1257. nop __LINE__
  1258. }
  1259. { .mfi
  1260. nop __LINE__
  1261. FSUB f122 = f55, f122
  1262. nop __LINE__
  1263. }
  1264. ;;
  1265. { .mfi
  1266. LDFPD f40, f41 = [BOFFSET], 2 * SIZE
  1267. FSUB f67 = f56, f67
  1268. nop __LINE__
  1269. }
  1270. { .mfi
  1271. nop __LINE__
  1272. FSUB f75 = f57, f75
  1273. nop __LINE__
  1274. }
  1275. ;;
  1276. { .mfi
  1277. LDFPD f42, f43 = [BOFFSET], 2 * SIZE
  1278. FSUB f83 = f58, f83
  1279. nop __LINE__
  1280. }
  1281. { .mfi
  1282. nop __LINE__
  1283. FSUB f91 = f59, f91
  1284. nop __LINE__
  1285. }
  1286. ;;
  1287. { .mfi
  1288. LDFPD f44, f45 = [BOFFSET], 2 * SIZE
  1289. FSUB f99 = f60, f99
  1290. nop __LINE__
  1291. }
  1292. { .mfi
  1293. nop __LINE__
  1294. FSUB f107 = f61, f107
  1295. nop __LINE__
  1296. }
  1297. ;;
  1298. { .mfi
  1299. LDFPD f46, f47 = [BOFFSET], 2 * SIZE
  1300. FSUB f115 = f62, f115
  1301. nop __LINE__
  1302. }
  1303. { .mfi
  1304. nop __LINE__
  1305. FSUB f123 = f63, f123
  1306. nop __LINE__
  1307. }
  1308. ;;
  1309. { .mfi
  1310. LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  1311. FSUB f68 = f32, f68
  1312. nop __LINE__
  1313. }
  1314. { .mfi
  1315. nop __LINE__
  1316. FSUB f76 = f33, f76
  1317. nop __LINE__
  1318. }
  1319. ;;
  1320. { .mfi
  1321. LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  1322. FSUB f84 = f34, f84
  1323. nop __LINE__
  1324. }
  1325. { .mfi
  1326. nop __LINE__
  1327. FSUB f92 = f35, f92
  1328. nop __LINE__
  1329. }
  1330. ;;
  1331. { .mfi
  1332. LDFPD f52, f53 = [BOFFSET], 2 * SIZE
  1333. FSUB f100 = f36, f100
  1334. nop __LINE__
  1335. }
  1336. { .mfi
  1337. nop __LINE__
  1338. FSUB f108 = f37, f108
  1339. nop __LINE__
  1340. }
  1341. ;;
  1342. { .mfi
  1343. LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  1344. FSUB f116 = f38, f116
  1345. nop __LINE__
  1346. }
  1347. { .mfi
  1348. nop __LINE__
  1349. FSUB f124 = f39, f124
  1350. nop __LINE__
  1351. }
  1352. ;;
  1353. { .mfi
  1354. LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  1355. FSUB f69 = f40, f69
  1356. nop __LINE__
  1357. }
  1358. { .mfi
  1359. nop __LINE__
  1360. FSUB f77 = f41, f77
  1361. nop __LINE__
  1362. }
  1363. ;;
  1364. { .mfi
  1365. LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  1366. FSUB f85 = f42, f85
  1367. nop __LINE__
  1368. }
  1369. { .mfi
  1370. nop __LINE__
  1371. FSUB f93 = f43, f93
  1372. nop __LINE__
  1373. }
  1374. ;;
  1375. { .mfi
  1376. LDFPD f60, f61 = [BOFFSET], 2 * SIZE
  1377. FSUB f101 = f44, f101
  1378. nop __LINE__
  1379. }
  1380. { .mfi
  1381. nop __LINE__
  1382. FSUB f109 = f45, f109
  1383. nop __LINE__
  1384. }
  1385. ;;
  1386. { .mfi
  1387. LDFPD f62, f63 = [BOFFSET]
  1388. FSUB f117 = f46, f117
  1389. adds BOFFSET = -62 * SIZE, BOFFSET
  1390. }
  1391. { .mfi
  1392. nop __LINE__
  1393. FSUB f125 = f47, f125
  1394. nop __LINE__
  1395. }
  1396. ;;
  1397. { .mfi
  1398. nop __LINE__
  1399. FSUB f70 = f48, f70
  1400. nop __LINE__
  1401. }
  1402. { .mfi
  1403. nop __LINE__
  1404. FSUB f78 = f49, f78
  1405. nop __LINE__
  1406. }
  1407. { .mfi
  1408. nop __LINE__
  1409. FSUB f86 = f50, f86
  1410. nop __LINE__
  1411. }
  1412. { .mfi
  1413. nop __LINE__
  1414. FSUB f94 = f51, f94
  1415. nop __LINE__
  1416. }
  1417. ;;
  1418. { .mfi
  1419. LDFPD f32, f33 = [AOFFSET]
  1420. FSUB f102 = f52, f102
  1421. nop __LINE__
  1422. }
  1423. { .mfi
  1424. nop __LINE__
  1425. FSUB f110 = f53, f110
  1426. nop __LINE__
  1427. }
  1428. { .mfi
  1429. nop __LINE__
  1430. FSUB f118 = f54, f118
  1431. nop __LINE__
  1432. }
  1433. { .mfi
  1434. nop __LINE__
  1435. FSUB f126 = f55, f126
  1436. adds AOFFSET = 2 * SIZE, AOFFSET
  1437. }
  1438. ;;
  1439. { .mfi
  1440. nop __LINE__
  1441. FSUB f71 = f56, f71
  1442. nop __LINE__
  1443. }
  1444. { .mfi
  1445. nop __LINE__
  1446. FSUB f79 = f57, f79
  1447. nop __LINE__
  1448. }
  1449. { .mfi
  1450. nop __LINE__
  1451. FSUB f87 = f58, f87
  1452. nop __LINE__
  1453. }
  1454. { .mfi
  1455. nop __LINE__
  1456. FSUB f95 = f59, f95
  1457. nop __LINE__
  1458. }
  1459. { .mfi
  1460. nop __LINE__
  1461. FSUB f103 = f60, f103
  1462. nop __LINE__
  1463. }
  1464. { .mfi
  1465. nop __LINE__
  1466. FSUB f111 = f61, f111
  1467. nop __LINE__
  1468. }
  1469. { .mfi
  1470. nop __LINE__
  1471. FSUB f119 = f62, f119
  1472. nop __LINE__
  1473. }
  1474. { .mfi
  1475. nop __LINE__
  1476. FSUB f127 = f63, f127
  1477. nop __LINE__
  1478. }
  1479. ;;
  1480. { .mfi
  1481. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  1482. FMPY f64 = f64, f32
  1483. nop __LINE__
  1484. }
  1485. { .mfi
  1486. nop __LINE__
  1487. FMPY f96 = f96, f32
  1488. adds BOFFSET2 = 4 * SIZE, BOFFSET
  1489. }
  1490. ;;
  1491. { .mfi
  1492. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  1493. FMPY f72 = f72, f32
  1494. nop __LINE__
  1495. }
  1496. { .mfi
  1497. nop __LINE__
  1498. FMPY f104 = f104, f32
  1499. nop __LINE__
  1500. }
  1501. ;;
  1502. { .mfi
  1503. LDFPD f38, f39 = [AOFFSET]
  1504. FMPY f80 = f80, f32
  1505. adds AOFFSET = 3 * SIZE, AOFFSET
  1506. }
  1507. { .mfi
  1508. nop __LINE__
  1509. FMPY f112 = f112, f32
  1510. nop __LINE__
  1511. }
  1512. ;;
  1513. { .mfi
  1514. LDFD f40 = [AOFFSET], 1 * SIZE
  1515. FMPY f88 = f88, f32
  1516. nop __LINE__
  1517. }
  1518. { .mfi
  1519. nop __LINE__
  1520. FMPY f120 = f120, f32
  1521. nop __LINE__
  1522. }
  1523. ;;
  1524. { .mfi
  1525. LDFPD f41, f42 = [AOFFSET], 2 * SIZE
  1526. FNMA f65 = f64, f33, f65
  1527. nop __LINE__
  1528. }
  1529. { .mfi
  1530. nop __LINE__
  1531. FNMA f97 = f96, f33, f97
  1532. nop __LINE__
  1533. }
  1534. ;;
  1535. { .mfi
  1536. LDFPD f43, f44 = [AOFFSET], 2 * SIZE
  1537. FNMA f73 = f72, f33, f73
  1538. nop __LINE__
  1539. }
  1540. { .mfi
  1541. nop __LINE__
  1542. FNMA f105 = f104, f33, f105
  1543. nop __LINE__
  1544. }
  1545. ;;
  1546. { .mfi
  1547. LDFPD f45, f46 = [AOFFSET]
  1548. FNMA f81 = f80, f33, f81
  1549. adds AOFFSET = 4 * SIZE, AOFFSET
  1550. }
  1551. { .mfi
  1552. nop __LINE__
  1553. FNMA f113 = f112, f33, f113
  1554. nop __LINE__
  1555. }
  1556. ;;
  1557. { .mfi
  1558. LDFPD f47, f48 = [AOFFSET], 2 * SIZE
  1559. FNMA f89 = f88, f33, f89
  1560. nop __LINE__
  1561. }
  1562. { .mfi
  1563. nop __LINE__
  1564. FNMA f121 = f120, f33, f121
  1565. nop __LINE__
  1566. }
  1567. ;;
  1568. { .mfi
  1569. LDFPD f49, f50 = [AOFFSET], 2 * SIZE
  1570. FNMA f66 = f64, f34, f66
  1571. nop __LINE__
  1572. }
  1573. { .mfi
  1574. nop __LINE__
  1575. FNMA f98 = f96, f34, f98
  1576. nop __LINE__
  1577. }
  1578. ;;
  1579. { .mfi
  1580. LDFPD f51, f52 = [AOFFSET]
  1581. FNMA f74 = f72, f34, f74
  1582. adds AOFFSET = 5 * SIZE, AOFFSET
  1583. }
  1584. { .mfi
  1585. nop __LINE__
  1586. FNMA f106 = f104, f34, f106
  1587. nop __LINE__
  1588. }
  1589. ;;
  1590. { .mfi
  1591. LDFD f53 = [AOFFSET], 1 * SIZE
  1592. FNMA f82 = f80, f34, f82
  1593. nop __LINE__
  1594. }
  1595. { .mfi
  1596. nop __LINE__
  1597. FNMA f114 = f112, f34, f114
  1598. nop __LINE__
  1599. }
  1600. ;;
  1601. { .mfi
  1602. LDFPD f54, f55 = [AOFFSET], 2 * SIZE
  1603. FNMA f90 = f88, f34, f90
  1604. nop __LINE__
  1605. }
  1606. { .mfi
  1607. nop __LINE__
  1608. FNMA f122 = f120, f34, f122
  1609. nop __LINE__
  1610. }
  1611. ;;
  1612. { .mfi
  1613. LDFPD f56, f57 = [AOFFSET]
  1614. FNMA f67 = f64, f35, f67
  1615. adds AOFFSET = 6 * SIZE, AOFFSET
  1616. }
  1617. { .mfi
  1618. nop __LINE__
  1619. FNMA f99 = f96, f35, f99
  1620. nop __LINE__
  1621. }
  1622. ;;
  1623. { .mfi
  1624. LDFPD f58, f59 = [AOFFSET], 2 * SIZE
  1625. FNMA f75 = f72, f35, f75
  1626. nop __LINE__
  1627. }
  1628. { .mfi
  1629. nop __LINE__
  1630. FNMA f107 = f104, f35, f107
  1631. nop __LINE__
  1632. }
  1633. ;;
  1634. { .mfi
  1635. LDFPD f60, f61 = [AOFFSET]
  1636. FNMA f83 = f80, f35, f83
  1637. adds AOFFSET = 7 * SIZE, AOFFSET
  1638. }
  1639. { .mfi
  1640. nop __LINE__
  1641. FNMA f115 = f112, f35, f115
  1642. nop __LINE__
  1643. }
  1644. ;;
  1645. { .mfi
  1646. LDFD f16 = [AOFFSET], 1 * SIZE
  1647. FNMA f91 = f88, f35, f91
  1648. nop __LINE__
  1649. }
  1650. { .mfi
  1651. nop __LINE__
  1652. FNMA f123 = f120, f35, f123
  1653. nop __LINE__
  1654. }
  1655. ;;
  1656. { .mfi
  1657. LDFPD f17, f18 = [AOFFSET]
  1658. FNMA f68 = f64, f36, f68
  1659. adds AOFFSET = 8 * SIZE, AOFFSET
  1660. }
  1661. { .mfi
  1662. nop __LINE__
  1663. FNMA f100 = f96, f36, f100
  1664. nop __LINE__
  1665. }
  1666. ;;
  1667. { .mfi
  1668. LDFPD f19, f20 = [AOFFSET]
  1669. FNMA f76 = f72, f36, f76
  1670. adds AOFFSET = 9 * SIZE, AOFFSET
  1671. }
  1672. { .mfi
  1673. nop __LINE__
  1674. FNMA f108 = f104, f36, f108
  1675. nop __LINE__
  1676. }
  1677. ;;
  1678. { .mfi
  1679. LDFD f21 = [AOFFSET]
  1680. FNMA f84 = f80, f36, f84
  1681. adds AOFFSET = -63 * SIZE, AOFFSET
  1682. }
  1683. { .mfi
  1684. nop __LINE__
  1685. FNMA f116 = f112, f36, f116
  1686. nop __LINE__
  1687. }
  1688. ;;
  1689. { .mfi
  1690. nop __LINE__
  1691. FNMA f92 = f88, f36, f92
  1692. nop __LINE__
  1693. }
  1694. { .mfi
  1695. nop __LINE__
  1696. FNMA f124 = f120, f36, f124
  1697. nop __LINE__
  1698. }
  1699. ;;
  1700. FNMA f69 = f64, f37, f69
  1701. FNMA f101 = f96, f37, f101
  1702. FNMA f77 = f72, f37, f77
  1703. FNMA f109 = f104, f37, f109
  1704. FNMA f85 = f80, f37, f85
  1705. FNMA f117 = f112, f37, f117
  1706. FNMA f93 = f88, f37, f93
  1707. FNMA f125 = f120, f37, f125
  1708. ;;
  1709. FNMA f70 = f64, f38, f70
  1710. FNMA f102 = f96, f38, f102
  1711. FNMA f78 = f72, f38, f78
  1712. FNMA f110 = f104, f38, f110
  1713. FNMA f86 = f80, f38, f86
  1714. FNMA f118 = f112, f38, f118
  1715. FNMA f94 = f88, f38, f94
  1716. FNMA f126 = f120, f38, f126
  1717. ;;
  1718. FNMA f71 = f64, f39, f71
  1719. FNMA f103 = f96, f39, f103
  1720. FNMA f79 = f72, f39, f79
  1721. FNMA f111 = f104, f39, f111
  1722. FNMA f87 = f80, f39, f87
  1723. FNMA f119 = f112, f39, f119
  1724. FNMA f95 = f88, f39, f95
  1725. FNMA f127 = f120, f39, f127
  1726. ;;
  1727. FMPY f65 = f65, f40
  1728. FMPY f97 = f97, f40
  1729. FMPY f73 = f73, f40
  1730. FMPY f105 = f105, f40
  1731. FMPY f81 = f81, f40
  1732. FMPY f113 = f113, f40
  1733. FMPY f89 = f89, f40
  1734. FMPY f121 = f121, f40
  1735. ;;
  1736. FNMA f66 = f65, f41, f66
  1737. FNMA f98 = f97, f41, f98
  1738. FNMA f74 = f73, f41, f74
  1739. FNMA f106 = f105, f41, f106
  1740. FNMA f82 = f81, f41, f82
  1741. FNMA f114 = f113, f41, f114
  1742. FNMA f90 = f89, f41, f90
  1743. FNMA f122 = f121, f41, f122
  1744. FNMA f67 = f65, f42, f67
  1745. FNMA f99 = f97, f42, f99
  1746. FNMA f75 = f73, f42, f75
  1747. FNMA f107 = f105, f42, f107
  1748. FNMA f83 = f81, f42, f83
  1749. FNMA f115 = f113, f42, f115
  1750. FNMA f91 = f89, f42, f91
  1751. FNMA f123 = f121, f42, f123
  1752. ;;
  1753. FNMA f68 = f65, f43, f68
  1754. FNMA f100 = f97, f43, f100
  1755. FNMA f76 = f73, f43, f76
  1756. FNMA f108 = f105, f43, f108
  1757. FNMA f84 = f81, f43, f84
  1758. FNMA f116 = f113, f43, f116
  1759. FNMA f92 = f89, f43, f92
  1760. FNMA f124 = f121, f43, f124
  1761. ;;
  1762. FNMA f69 = f65, f44, f69
  1763. FNMA f101 = f97, f44, f101
  1764. FNMA f77 = f73, f44, f77
  1765. FNMA f109 = f105, f44, f109
  1766. FNMA f85 = f81, f44, f85
  1767. FNMA f117 = f113, f44, f117
  1768. FNMA f93 = f89, f44, f93
  1769. FNMA f125 = f121, f44, f125
  1770. ;;
  1771. FNMA f70 = f65, f45, f70
  1772. FNMA f102 = f97, f45, f102
  1773. FNMA f78 = f73, f45, f78
  1774. FNMA f110 = f105, f45, f110
  1775. FNMA f86 = f81, f45, f86
  1776. FNMA f118 = f113, f45, f118
  1777. FNMA f94 = f89, f45, f94
  1778. FNMA f126 = f121, f45, f126
  1779. ;;
  1780. FNMA f71 = f65, f46, f71
  1781. FNMA f103 = f97, f46, f103
  1782. FNMA f79 = f73, f46, f79
  1783. FNMA f111 = f105, f46, f111
  1784. FNMA f87 = f81, f46, f87
  1785. FNMA f119 = f113, f46, f119
  1786. FNMA f95 = f89, f46, f95
  1787. FNMA f127 = f121, f46, f127
  1788. ;;
  1789. FMPY f66 = f66, f47
  1790. FMPY f98 = f98, f47
  1791. FMPY f74 = f74, f47
  1792. FMPY f106 = f106, f47
  1793. FMPY f82 = f82, f47
  1794. FMPY f114 = f114, f47
  1795. FMPY f90 = f90, f47
  1796. FMPY f122 = f122, f47
  1797. ;;
  1798. FNMA f67 = f66, f48, f67
  1799. FNMA f99 = f98, f48, f99
  1800. FNMA f75 = f74, f48, f75
  1801. FNMA f107 = f106, f48, f107
  1802. FNMA f83 = f82, f48, f83
  1803. FNMA f115 = f114, f48, f115
  1804. FNMA f91 = f90, f48, f91
  1805. FNMA f123 = f122, f48, f123
  1806. FNMA f68 = f66, f49, f68
  1807. FNMA f100 = f98, f49, f100
  1808. FNMA f76 = f74, f49, f76
  1809. FNMA f108 = f106, f49, f108
  1810. FNMA f84 = f82, f49, f84
  1811. FNMA f116 = f114, f49, f116
  1812. FNMA f92 = f90, f49, f92
  1813. FNMA f124 = f122, f49, f124
  1814. ;;
  1815. FNMA f69 = f66, f50, f69
  1816. FNMA f101 = f98, f50, f101
  1817. FNMA f77 = f74, f50, f77
  1818. FNMA f109 = f106, f50, f109
  1819. FNMA f85 = f82, f50, f85
  1820. FNMA f117 = f114, f50, f117
  1821. FNMA f93 = f90, f50, f93
  1822. FNMA f125 = f122, f50, f125
  1823. ;;
  1824. FNMA f70 = f66, f51, f70
  1825. FNMA f102 = f98, f51, f102
  1826. FNMA f78 = f74, f51, f78
  1827. FNMA f110 = f106, f51, f110
  1828. FNMA f86 = f82, f51, f86
  1829. FNMA f118 = f114, f51, f118
  1830. FNMA f94 = f90, f51, f94
  1831. FNMA f126 = f122, f51, f126
  1832. ;;
  1833. FNMA f71 = f66, f52, f71
  1834. FNMA f103 = f98, f52, f103
  1835. FNMA f79 = f74, f52, f79
  1836. FNMA f111 = f106, f52, f111
  1837. FNMA f87 = f82, f52, f87
  1838. FNMA f119 = f114, f52, f119
  1839. FNMA f95 = f90, f52, f95
  1840. FNMA f127 = f122, f52, f127
  1841. ;;
  1842. FMPY f67 = f67, f53
  1843. FMPY f99 = f99, f53
  1844. FMPY f75 = f75, f53
  1845. FMPY f107 = f107, f53
  1846. FMPY f83 = f83, f53
  1847. FMPY f115 = f115, f53
  1848. FMPY f91 = f91, f53
  1849. FMPY f123 = f123, f53
  1850. ;;
  1851. FNMA f68 = f67, f54, f68
  1852. FNMA f100 = f99, f54, f100
  1853. FNMA f76 = f75, f54, f76
  1854. FNMA f108 = f107, f54, f108
  1855. FNMA f84 = f83, f54, f84
  1856. FNMA f116 = f115, f54, f116
  1857. FNMA f92 = f91, f54, f92
  1858. FNMA f124 = f123, f54, f124
  1859. ;;
  1860. FNMA f69 = f67, f55, f69
  1861. FNMA f101 = f99, f55, f101
  1862. FNMA f77 = f75, f55, f77
  1863. FNMA f109 = f107, f55, f109
  1864. FNMA f85 = f83, f55, f85
  1865. FNMA f117 = f115, f55, f117
  1866. FNMA f93 = f91, f55, f93
  1867. FNMA f125 = f123, f55, f125
  1868. ;;
  1869. FNMA f70 = f67, f56, f70
  1870. FNMA f102 = f99, f56, f102
  1871. FNMA f78 = f75, f56, f78
  1872. FNMA f110 = f107, f56, f110
  1873. FNMA f86 = f83, f56, f86
  1874. FNMA f118 = f115, f56, f118
  1875. FNMA f94 = f91, f56, f94
  1876. FNMA f126 = f123, f56, f126
  1877. ;;
  1878. FNMA f71 = f67, f57, f71
  1879. FNMA f103 = f99, f57, f103
  1880. FNMA f79 = f75, f57, f79
  1881. FNMA f111 = f107, f57, f111
  1882. FNMA f87 = f83, f57, f87
  1883. FNMA f119 = f115, f57, f119
  1884. FNMA f95 = f91, f57, f95
  1885. FNMA f127 = f123, f57, f127
  1886. ;;
  1887. FMPY f68 = f68, f58
  1888. FMPY f100 = f100, f58
  1889. FMPY f76 = f76, f58
  1890. FMPY f108 = f108, f58
  1891. FMPY f84 = f84, f58
  1892. FMPY f116 = f116, f58
  1893. FMPY f92 = f92, f58
  1894. FMPY f124 = f124, f58
  1895. ;;
  1896. FNMA f69 = f68, f59, f69
  1897. FNMA f101 = f100, f59, f101
  1898. FNMA f77 = f76, f59, f77
  1899. FNMA f109 = f108, f59, f109
  1900. FNMA f85 = f84, f59, f85
  1901. FNMA f117 = f116, f59, f117
  1902. FNMA f93 = f92, f59, f93
  1903. FNMA f125 = f124, f59, f125
  1904. ;;
  1905. FNMA f70 = f68, f60, f70
  1906. FNMA f102 = f100, f60, f102
  1907. FNMA f78 = f76, f60, f78
  1908. FNMA f110 = f108, f60, f110
  1909. FNMA f86 = f84, f60, f86
  1910. FNMA f118 = f116, f60, f118
  1911. FNMA f94 = f92, f60, f94
  1912. FNMA f126 = f124, f60, f126
  1913. ;;
  1914. { .mfi
  1915. STFD [BOFFSET] = f64, SIZE
  1916. FNMA f71 = f68, f61, f71
  1917. }
  1918. { .mfi
  1919. STFD [BOFFSET2] = f96, SIZE
  1920. FNMA f103 = f100, f61, f103
  1921. }
  1922. ;;
  1923. { .mfi
  1924. STFD [BOFFSET] = f72, SIZE
  1925. FNMA f79 = f76, f61, f79
  1926. }
  1927. { .mfi
  1928. STFD [BOFFSET2] = f104, SIZE
  1929. FNMA f111 = f108, f61, f111
  1930. }
  1931. ;;
  1932. { .mfi
  1933. STFD [BOFFSET] = f80, SIZE
  1934. FNMA f87 = f84, f61, f87
  1935. }
  1936. { .mfi
  1937. STFD [BOFFSET2] = f112, SIZE
  1938. FNMA f119 = f116, f61, f119
  1939. }
  1940. ;;
  1941. { .mfi
  1942. STFD [BOFFSET] = f88, 5 * SIZE
  1943. FNMA f95 = f92, f61, f95
  1944. }
  1945. { .mfi
  1946. STFD [BOFFSET2] = f120, 5 * SIZE
  1947. FNMA f127 = f124, f61, f127
  1948. }
  1949. ;;
  1950. { .mfi
  1951. STFD [BOFFSET] = f65, SIZE
  1952. FMPY f69 = f69, f16
  1953. }
  1954. { .mfi
  1955. STFD [BOFFSET2] = f97, SIZE
  1956. FMPY f101 = f101, f16
  1957. }
  1958. ;;
  1959. { .mfi
  1960. STFD [BOFFSET] = f73, SIZE
  1961. FMPY f77 = f77, f16
  1962. }
  1963. { .mfi
  1964. STFD [BOFFSET2] = f105, SIZE
  1965. FMPY f109 = f109, f16
  1966. }
  1967. ;;
  1968. { .mfi
  1969. STFD [BOFFSET] = f81, SIZE
  1970. FMPY f85 = f85, f16
  1971. }
  1972. { .mfi
  1973. STFD [BOFFSET2] = f113, SIZE
  1974. FMPY f117 = f117, f16
  1975. }
  1976. ;;
  1977. { .mfi
  1978. STFD [BOFFSET] = f89, 5 * SIZE
  1979. FMPY f93 = f93, f16
  1980. }
  1981. { .mfi
  1982. STFD [BOFFSET2] = f121, 5 * SIZE
  1983. FMPY f125 = f125, f16
  1984. }
  1985. ;;
  1986. { .mfi
  1987. STFD [BOFFSET] = f66, SIZE
  1988. FNMA f70 = f69, f17, f70
  1989. }
  1990. { .mfi
  1991. STFD [BOFFSET2] = f98, SIZE
  1992. FNMA f102 = f101, f17, f102
  1993. }
  1994. ;;
  1995. { .mfi
  1996. STFD [BOFFSET] = f74, SIZE
  1997. FNMA f78 = f77, f17, f78
  1998. }
  1999. { .mfi
  2000. STFD [BOFFSET2] = f106, SIZE
  2001. FNMA f110 = f109, f17, f110
  2002. }
  2003. ;;
  2004. { .mfi
  2005. STFD [BOFFSET] = f82, SIZE
  2006. FNMA f86 = f85, f17, f86
  2007. }
  2008. { .mfi
  2009. STFD [BOFFSET2] = f114, SIZE
  2010. FNMA f118 = f117, f17, f118
  2011. }
  2012. ;;
  2013. { .mfi
  2014. STFD [BOFFSET] = f90, 5 * SIZE
  2015. FNMA f94 = f93, f17, f94
  2016. }
  2017. { .mfi
  2018. STFD [BOFFSET2] = f122, 5 * SIZE
  2019. FNMA f126 = f125, f17, f126
  2020. }
  2021. ;;
  2022. { .mfi
  2023. STFD [BOFFSET] = f67, SIZE
  2024. FNMA f71 = f69, f18, f71
  2025. }
  2026. { .mfi
  2027. STFD [BOFFSET2] = f99, SIZE
  2028. FNMA f103 = f101, f18, f103
  2029. }
  2030. ;;
  2031. { .mfi
  2032. STFD [BOFFSET] = f75, SIZE
  2033. FNMA f79 = f77, f18, f79
  2034. }
  2035. { .mfi
  2036. STFD [BOFFSET2] = f107, SIZE
  2037. FNMA f111 = f109, f18, f111
  2038. }
  2039. ;;
  2040. { .mfi
  2041. STFD [BOFFSET] = f83, SIZE
  2042. FNMA f87 = f85, f18, f87
  2043. }
  2044. { .mfi
  2045. STFD [BOFFSET2] = f115, SIZE
  2046. FNMA f119 = f117, f18, f119
  2047. }
  2048. ;;
  2049. { .mfi
  2050. STFD [BOFFSET] = f91, 5 * SIZE
  2051. FNMA f95 = f93, f18, f95
  2052. }
  2053. { .mfi
  2054. STFD [BOFFSET2] = f123, 5 * SIZE
  2055. FNMA f127 = f125, f18, f127
  2056. }
  2057. ;;
  2058. { .mfi
  2059. STFD [BOFFSET] = f68, SIZE
  2060. FMPY f70 = f70, f19
  2061. }
  2062. { .mfi
  2063. STFD [BOFFSET2] = f100, SIZE
  2064. FMPY f102 = f102, f19
  2065. }
  2066. ;;
  2067. { .mfi
  2068. STFD [BOFFSET] = f76, SIZE
  2069. FMPY f78 = f78, f19
  2070. }
  2071. { .mfi
  2072. STFD [BOFFSET2] = f108, SIZE
  2073. FMPY f110 = f110, f19
  2074. }
  2075. ;;
  2076. { .mfi
  2077. STFD [BOFFSET] = f84, SIZE
  2078. FMPY f86 = f86, f19
  2079. }
  2080. { .mfi
  2081. STFD [BOFFSET2] = f116, SIZE
  2082. FMPY f118 = f118, f19
  2083. }
  2084. ;;
  2085. { .mfi
  2086. STFD [BOFFSET] = f92, 5 * SIZE
  2087. FMPY f94 = f94, f19
  2088. }
  2089. { .mfi
  2090. STFD [BOFFSET2] = f124, 5 * SIZE
  2091. FMPY f126 = f126, f19
  2092. }
  2093. ;;
  2094. { .mfi
  2095. STFD [BOFFSET] = f69, SIZE
  2096. FNMA f71 = f70, f20, f71
  2097. }
  2098. { .mfi
  2099. STFD [BOFFSET2] = f101, SIZE
  2100. FNMA f103 = f102, f20, f103
  2101. }
  2102. ;;
  2103. { .mfi
  2104. STFD [BOFFSET] = f77, SIZE
  2105. FNMA f79 = f78, f20, f79
  2106. }
  2107. { .mfi
  2108. STFD [BOFFSET2] = f109, SIZE
  2109. FNMA f111 = f110, f20, f111
  2110. }
  2111. ;;
  2112. { .mfi
  2113. STFD [BOFFSET] = f85, SIZE
  2114. FNMA f87 = f86, f20, f87
  2115. }
  2116. { .mfi
  2117. STFD [BOFFSET2] = f117, SIZE
  2118. FNMA f119 = f118, f20, f119
  2119. }
  2120. ;;
  2121. { .mfi
  2122. STFD [BOFFSET] = f93, 5 * SIZE
  2123. FNMA f95 = f94, f20, f95
  2124. }
  2125. { .mfi
  2126. STFD [BOFFSET2] = f125, 5 * SIZE
  2127. FNMA f127 = f126, f20, f127
  2128. }
  2129. ;;
  2130. { .mfi
  2131. STFD [BOFFSET] = f70, SIZE
  2132. FMPY f71 = f71, f21
  2133. }
  2134. { .mfi
  2135. STFD [BOFFSET2] = f102, SIZE
  2136. FMPY f103 = f103, f21
  2137. }
  2138. ;;
  2139. { .mfi
  2140. STFD [BOFFSET] = f78, SIZE
  2141. FMPY f79 = f79, f21
  2142. }
  2143. { .mfi
  2144. STFD [BOFFSET2] = f110, SIZE
  2145. FMPY f111 = f111, f21
  2146. }
  2147. ;;
  2148. { .mfi
  2149. STFD [BOFFSET] = f86, SIZE
  2150. FMPY f87 = f87, f21
  2151. }
  2152. { .mfi
  2153. STFD [BOFFSET2] = f118, SIZE
  2154. FMPY f119 = f119, f21
  2155. }
  2156. ;;
  2157. { .mfi
  2158. STFD [BOFFSET] = f94, 5 * SIZE
  2159. FMPY f95 = f95, f21
  2160. }
  2161. { .mfi
  2162. STFD [BOFFSET2] = f126, 5 * SIZE
  2163. FMPY f127 = f127, f21
  2164. }
  2165. ;;
  2166. { .mmi
  2167. STFD [BOFFSET] = f71, SIZE
  2168. STFD [BOFFSET2] = f103, SIZE
  2169. }
  2170. ;;
  2171. { .mmi
  2172. STFD [BOFFSET] = f79, SIZE
  2173. STFD [BOFFSET2] = f111, SIZE
  2174. }
  2175. ;;
  2176. { .mmi
  2177. STFD [BOFFSET] = f87, SIZE
  2178. STFD [BOFFSET2] = f119, SIZE
  2179. adds C9 = 4 * SIZE, C1
  2180. }
  2181. ;;
  2182. { .mfi
  2183. STFD [BOFFSET] = f95
  2184. adds BOFFSET = - 59 * SIZE, BOFFSET
  2185. }
  2186. { .mfi
  2187. STFD [BOFFSET2] = f127
  2188. adds BOFFSET2 = - 59 * SIZE, BOFFSET2
  2189. }
  2190. ;;
  2191. #endif
  2192. #ifdef RN
  2193. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  2194. ;;
  2195. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  2196. ;;
  2197. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  2198. ;;
  2199. LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  2200. ;;
  2201. LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  2202. ;;
  2203. LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  2204. ;;
  2205. LDFPD f44, f45 = [AOFFSET], 2 * SIZE
  2206. ;;
  2207. LDFPD f46, f47 = [AOFFSET], 2 * SIZE
  2208. ;;
  2209. { .mfi
  2210. LDFPD f48, f49 = [AOFFSET], 2 * SIZE
  2211. FSUB f64 = f32, f64
  2212. }
  2213. { .mfi
  2214. FSUB f65 = f33, f65
  2215. }
  2216. ;;
  2217. { .mfi
  2218. LDFPD f50, f51 = [AOFFSET], 2 * SIZE
  2219. FSUB f66 = f34, f66
  2220. }
  2221. { .mfi
  2222. FSUB f67 = f35, f67
  2223. }
  2224. ;;
  2225. { .mfi
  2226. LDFPD f52, f53 = [AOFFSET], 2 * SIZE
  2227. FSUB f68 = f36, f68
  2228. }
  2229. { .mfi
  2230. FSUB f69 = f37, f69
  2231. }
  2232. ;;
  2233. { .mfi
  2234. LDFPD f54, f55 = [AOFFSET], 2 * SIZE
  2235. FSUB f70 = f38, f70
  2236. }
  2237. { .mfi
  2238. FSUB f71 = f39, f71
  2239. }
  2240. ;;
  2241. { .mfi
  2242. LDFPD f56, f57 = [AOFFSET], 2 * SIZE
  2243. FSUB f72 = f40, f72
  2244. }
  2245. { .mfi
  2246. FSUB f73 = f41, f73
  2247. }
  2248. ;;
  2249. { .mfi
  2250. LDFPD f58, f59 = [AOFFSET], 2 * SIZE
  2251. FSUB f74 = f42, f74
  2252. }
  2253. { .mfi
  2254. FSUB f75 = f43, f75
  2255. }
  2256. ;;
  2257. { .mfi
  2258. LDFPD f60, f61 = [AOFFSET], 2 * SIZE
  2259. FSUB f76 = f44, f76
  2260. }
  2261. { .mfi
  2262. FSUB f77 = f45, f77
  2263. }
  2264. ;;
  2265. { .mfi
  2266. LDFPD f62, f63 = [AOFFSET], 2 * SIZE
  2267. FSUB f78 = f46, f78
  2268. }
  2269. { .mfi
  2270. FSUB f79 = f47, f79
  2271. }
  2272. ;;
  2273. { .mfi
  2274. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  2275. FSUB f80 = f48, f80
  2276. nop __LINE__
  2277. }
  2278. { .mfi
  2279. nop __LINE__
  2280. FSUB f81 = f49, f81
  2281. nop __LINE__
  2282. }
  2283. ;;
  2284. { .mfi
  2285. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  2286. FSUB f82 = f50, f82
  2287. nop __LINE__
  2288. }
  2289. { .mfi
  2290. nop __LINE__
  2291. FSUB f83 = f51, f83
  2292. nop __LINE__
  2293. }
  2294. ;;
  2295. { .mfi
  2296. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  2297. FSUB f84 = f52, f84
  2298. nop __LINE__
  2299. }
  2300. { .mfi
  2301. nop __LINE__
  2302. FSUB f85 = f53, f85
  2303. nop __LINE__
  2304. }
  2305. ;;
  2306. { .mfi
  2307. LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  2308. FSUB f86 = f54, f86
  2309. nop __LINE__
  2310. }
  2311. { .mfi
  2312. nop __LINE__
  2313. FSUB f87 = f55, f87
  2314. nop __LINE__
  2315. }
  2316. ;;
  2317. { .mfi
  2318. LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  2319. FSUB f88 = f56, f88
  2320. nop __LINE__
  2321. }
  2322. { .mfi
  2323. nop __LINE__
  2324. FSUB f89 = f57, f89
  2325. nop __LINE__
  2326. }
  2327. ;;
  2328. { .mfi
  2329. LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  2330. FSUB f90 = f58, f90
  2331. nop __LINE__
  2332. }
  2333. { .mfi
  2334. nop __LINE__
  2335. FSUB f91 = f59, f91
  2336. nop __LINE__
  2337. }
  2338. ;;
  2339. { .mfi
  2340. LDFPD f44, f45 = [AOFFSET], 2 * SIZE
  2341. FSUB f92 = f60, f92
  2342. nop __LINE__
  2343. }
  2344. { .mfi
  2345. nop __LINE__
  2346. FSUB f93 = f61, f93
  2347. nop __LINE__
  2348. }
  2349. ;;
  2350. { .mfi
  2351. LDFPD f46, f47 = [AOFFSET], 2 * SIZE
  2352. FSUB f94 = f62, f94
  2353. nop __LINE__
  2354. }
  2355. { .mfi
  2356. nop __LINE__
  2357. FSUB f95 = f63, f95
  2358. nop __LINE__
  2359. }
  2360. ;;
  2361. { .mfi
  2362. LDFPD f48, f49 = [AOFFSET], 2 * SIZE
  2363. FSUB f96 = f32, f96
  2364. nop __LINE__
  2365. }
  2366. { .mfi
  2367. nop __LINE__
  2368. FSUB f97 = f33, f97
  2369. nop __LINE__
  2370. }
  2371. ;;
  2372. { .mfi
  2373. LDFPD f50, f51 = [AOFFSET], 2 * SIZE
  2374. FSUB f98 = f34, f98
  2375. nop __LINE__
  2376. }
  2377. { .mfi
  2378. nop __LINE__
  2379. FSUB f99 = f35, f99
  2380. nop __LINE__
  2381. }
  2382. ;;
  2383. { .mfi
  2384. LDFPD f52, f53 = [AOFFSET], 2 * SIZE
  2385. FSUB f100 = f36, f100
  2386. nop __LINE__
  2387. }
  2388. { .mfi
  2389. nop __LINE__
  2390. FSUB f101 = f37, f101
  2391. nop __LINE__
  2392. }
  2393. ;;
  2394. { .mfi
  2395. LDFPD f54, f55 = [AOFFSET], 2 * SIZE
  2396. FSUB f102 = f38, f102
  2397. nop __LINE__
  2398. }
  2399. { .mfi
  2400. nop __LINE__
  2401. FSUB f103 = f39, f103
  2402. nop __LINE__
  2403. }
  2404. ;;
  2405. { .mfi
  2406. LDFPD f56, f57 = [AOFFSET], 2 * SIZE
  2407. FSUB f104 = f40, f104
  2408. nop __LINE__
  2409. }
  2410. { .mfi
  2411. nop __LINE__
  2412. FSUB f105 = f41, f105
  2413. nop __LINE__
  2414. }
  2415. ;;
  2416. { .mfi
  2417. LDFPD f58, f59 = [AOFFSET], 2 * SIZE
  2418. FSUB f106 = f42, f106
  2419. nop __LINE__
  2420. }
  2421. { .mfi
  2422. nop __LINE__
  2423. FSUB f107 = f43, f107
  2424. nop __LINE__
  2425. }
  2426. ;;
  2427. { .mfi
  2428. LDFPD f60, f61 = [AOFFSET], 2 * SIZE
  2429. FSUB f108 = f44, f108
  2430. nop __LINE__
  2431. }
  2432. { .mfi
  2433. nop __LINE__
  2434. FSUB f109 = f45, f109
  2435. nop __LINE__
  2436. }
  2437. ;;
  2438. { .mfi
  2439. LDFPD f62, f63 = [AOFFSET]
  2440. FSUB f110 = f46, f110
  2441. adds AOFFSET = -62 * SIZE, AOFFSET
  2442. }
  2443. { .mfi
  2444. nop __LINE__
  2445. FSUB f111 = f47, f111
  2446. nop __LINE__
  2447. }
  2448. ;;
  2449. { .mfi
  2450. nop __LINE__
  2451. FSUB f112 = f48, f112
  2452. nop __LINE__
  2453. }
  2454. { .mfi
  2455. nop __LINE__
  2456. FSUB f113 = f49, f113
  2457. nop __LINE__
  2458. }
  2459. { .mfi
  2460. nop __LINE__
  2461. FSUB f114 = f50, f114
  2462. nop __LINE__
  2463. }
  2464. { .mfi
  2465. nop __LINE__
  2466. FSUB f115 = f51, f115
  2467. nop __LINE__
  2468. }
  2469. { .mfi
  2470. nop __LINE__
  2471. FSUB f116 = f52, f116
  2472. nop __LINE__
  2473. }
  2474. { .mfi
  2475. nop __LINE__
  2476. FSUB f117 = f53, f117
  2477. nop __LINE__
  2478. }
  2479. { .mfi
  2480. nop __LINE__
  2481. FSUB f118 = f54, f118
  2482. nop __LINE__
  2483. }
  2484. { .mfi
  2485. nop __LINE__
  2486. FSUB f119 = f55, f119
  2487. nop __LINE__
  2488. }
  2489. { .mfi
  2490. nop __LINE__
  2491. FSUB f120 = f56, f120
  2492. nop __LINE__
  2493. }
  2494. { .mfi
  2495. nop __LINE__
  2496. FSUB f121 = f57, f121
  2497. nop __LINE__
  2498. }
  2499. { .mfi
  2500. nop __LINE__
  2501. FSUB f122 = f58, f122
  2502. nop __LINE__
  2503. }
  2504. { .mfi
  2505. nop __LINE__
  2506. FSUB f123 = f59, f123
  2507. nop __LINE__
  2508. }
  2509. { .mfi
  2510. nop __LINE__
  2511. FSUB f124 = f60, f124
  2512. nop __LINE__
  2513. }
  2514. { .mfi
  2515. nop __LINE__
  2516. FSUB f125 = f61, f125
  2517. nop __LINE__
  2518. }
  2519. ;;
  2520. { .mfi
  2521. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  2522. FSUB f126 = f62, f126
  2523. nop __LINE__
  2524. }
  2525. { .mfi
  2526. nop __LINE__
  2527. FSUB f127 = f63, f127
  2528. nop __LINE__
  2529. }
  2530. ;;
  2531. { .mfi
  2532. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  2533. FMPY f64 = f64, f32
  2534. nop __LINE__
  2535. }
  2536. { .mfi
  2537. nop __LINE__
  2538. FMPY f68 = f68, f32
  2539. nop __LINE__
  2540. }
  2541. ;;
  2542. { .mfi
  2543. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  2544. FMPY f65 = f65, f32
  2545. nop __LINE__
  2546. }
  2547. { .mfi
  2548. nop __LINE__
  2549. FMPY f69 = f69, f32
  2550. nop __LINE__
  2551. }
  2552. ;;
  2553. { .mfi
  2554. LDFPD f38, f39 = [BOFFSET]
  2555. FMPY f66 = f66, f32
  2556. adds BOFFSET = 3 * SIZE, BOFFSET
  2557. }
  2558. { .mfi
  2559. nop __LINE__
  2560. FMPY f70 = f70, f32
  2561. nop __LINE__
  2562. }
  2563. ;;
  2564. { .mfi
  2565. LDFD f40 = [BOFFSET], 1 * SIZE
  2566. FMPY f67 = f67, f32
  2567. nop __LINE__
  2568. }
  2569. { .mfi
  2570. nop __LINE__
  2571. FMPY f71 = f71, f32
  2572. nop __LINE__
  2573. }
  2574. ;;
  2575. { .mfi
  2576. LDFPD f41, f42 = [BOFFSET], 2 * SIZE
  2577. FNMA f72 = f64, f33, f72
  2578. nop __LINE__
  2579. }
  2580. { .mfi
  2581. nop __LINE__
  2582. FNMA f76 = f68, f33, f76
  2583. nop __LINE__
  2584. }
  2585. ;;
  2586. { .mfi
  2587. LDFPD f43, f44 = [BOFFSET], 2 * SIZE
  2588. FNMA f73 = f65, f33, f73
  2589. nop __LINE__
  2590. }
  2591. { .mfi
  2592. nop __LINE__
  2593. FNMA f77 = f69, f33, f77
  2594. nop __LINE__
  2595. }
  2596. ;;
  2597. { .mfi
  2598. LDFPD f45, f46 = [BOFFSET]
  2599. FNMA f74 = f66, f33, f74
  2600. adds BOFFSET = 4 * SIZE, BOFFSET
  2601. }
  2602. { .mfi
  2603. nop __LINE__
  2604. FNMA f78 = f70, f33, f78
  2605. nop __LINE__
  2606. }
  2607. ;;
  2608. { .mfi
  2609. LDFPD f47, f48 = [BOFFSET], 2 * SIZE
  2610. FNMA f75 = f67, f33, f75
  2611. nop __LINE__
  2612. }
  2613. { .mfi
  2614. nop __LINE__
  2615. FNMA f79 = f71, f33, f79
  2616. nop __LINE__
  2617. }
  2618. ;;
  2619. { .mfi
  2620. LDFPD f49, f50 = [BOFFSET], 2 * SIZE
  2621. FNMA f80 = f64, f34, f80
  2622. nop __LINE__
  2623. }
  2624. { .mfi
  2625. nop __LINE__
  2626. FNMA f84 = f68, f34, f84
  2627. nop __LINE__
  2628. }
  2629. ;;
  2630. { .mfi
  2631. LDFPD f51, f52 = [BOFFSET]
  2632. FNMA f81 = f65, f34, f81
  2633. adds BOFFSET = 5 * SIZE, BOFFSET
  2634. }
  2635. { .mfi
  2636. nop __LINE__
  2637. FNMA f85 = f69, f34, f85
  2638. nop __LINE__
  2639. }
  2640. ;;
  2641. { .mfi
  2642. LDFD f53 = [BOFFSET], 1 * SIZE
  2643. FNMA f82 = f66, f34, f82
  2644. nop __LINE__
  2645. }
  2646. { .mfi
  2647. nop __LINE__
  2648. FNMA f86 = f70, f34, f86
  2649. nop __LINE__
  2650. }
  2651. ;;
  2652. { .mfi
  2653. LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  2654. FNMA f83 = f67, f34, f83
  2655. nop __LINE__
  2656. }
  2657. { .mfi
  2658. nop __LINE__
  2659. FNMA f87 = f71, f34, f87
  2660. nop __LINE__
  2661. }
  2662. ;;
  2663. { .mfi
  2664. LDFPD f56, f57 = [BOFFSET]
  2665. FNMA f88 = f64, f35, f88
  2666. adds BOFFSET = 6 * SIZE, BOFFSET
  2667. }
  2668. { .mfi
  2669. nop __LINE__
  2670. FNMA f92 = f68, f35, f92
  2671. nop __LINE__
  2672. }
  2673. ;;
  2674. { .mfi
  2675. LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  2676. FNMA f89 = f65, f35, f89
  2677. nop __LINE__
  2678. }
  2679. { .mfi
  2680. nop __LINE__
  2681. FNMA f93 = f69, f35, f93
  2682. nop __LINE__
  2683. }
  2684. ;;
  2685. { .mfi
  2686. LDFPD f60, f61 = [BOFFSET]
  2687. FNMA f90 = f66, f35, f90
  2688. adds BOFFSET = 7 * SIZE, BOFFSET
  2689. }
  2690. { .mfi
  2691. nop __LINE__
  2692. FNMA f94 = f70, f35, f94
  2693. nop __LINE__
  2694. }
  2695. ;;
  2696. { .mfi
  2697. LDFD f16 = [BOFFSET], 1 * SIZE
  2698. FNMA f91 = f67, f35, f91
  2699. nop __LINE__
  2700. }
  2701. { .mfi
  2702. nop __LINE__
  2703. FNMA f95 = f71, f35, f95
  2704. nop __LINE__
  2705. }
  2706. ;;
  2707. { .mfi
  2708. LDFPD f17, f18 = [BOFFSET]
  2709. FNMA f96 = f64, f36, f96
  2710. adds BOFFSET = 8 * SIZE, BOFFSET
  2711. }
  2712. { .mfi
  2713. nop __LINE__
  2714. FNMA f100 = f68, f36, f100
  2715. nop __LINE__
  2716. }
  2717. ;;
  2718. { .mfi
  2719. LDFPD f19, f20 = [BOFFSET]
  2720. FNMA f97 = f65, f36, f97
  2721. adds BOFFSET = 9 * SIZE, BOFFSET
  2722. }
  2723. { .mfi
  2724. nop __LINE__
  2725. FNMA f101 = f69, f36, f101
  2726. nop __LINE__
  2727. }
  2728. ;;
  2729. { .mfi
  2730. LDFD f21 = [BOFFSET]
  2731. FNMA f98 = f66, f36, f98
  2732. adds BOFFSET = -63 * SIZE, BOFFSET
  2733. }
  2734. { .mfi
  2735. nop __LINE__
  2736. FNMA f102 = f70, f36, f102
  2737. nop __LINE__
  2738. }
  2739. ;;
  2740. { .mfi
  2741. nop __LINE__
  2742. FNMA f99 = f67, f36, f99
  2743. nop __LINE__
  2744. }
  2745. { .mfi
  2746. nop __LINE__
  2747. FNMA f103 = f71, f36, f103
  2748. nop __LINE__
  2749. }
  2750. ;;
  2751. FNMA f104 = f64, f37, f104
  2752. FNMA f108 = f68, f37, f108
  2753. FNMA f105 = f65, f37, f105
  2754. FNMA f109 = f69, f37, f109
  2755. FNMA f106 = f66, f37, f106
  2756. FNMA f110 = f70, f37, f110
  2757. FNMA f107 = f67, f37, f107
  2758. FNMA f111 = f71, f37, f111
  2759. ;;
  2760. FNMA f112 = f64, f38, f112
  2761. FNMA f116 = f68, f38, f116
  2762. FNMA f113 = f65, f38, f113
  2763. FNMA f117 = f69, f38, f117
  2764. FNMA f114 = f66, f38, f114
  2765. FNMA f118 = f70, f38, f118
  2766. FNMA f115 = f67, f38, f115
  2767. FNMA f119 = f71, f38, f119
  2768. ;;
  2769. FNMA f120 = f64, f39, f120
  2770. FNMA f124 = f68, f39, f124
  2771. FNMA f121 = f65, f39, f121
  2772. FNMA f125 = f69, f39, f125
  2773. FNMA f122 = f66, f39, f122
  2774. FNMA f126 = f70, f39, f126
  2775. FNMA f123 = f67, f39, f123
  2776. FNMA f127 = f71, f39, f127
  2777. ;;
  2778. FMPY f72 = f72, f40
  2779. FMPY f76 = f76, f40
  2780. FMPY f73 = f73, f40
  2781. FMPY f77 = f77, f40
  2782. FMPY f74 = f74, f40
  2783. FMPY f78 = f78, f40
  2784. FMPY f75 = f75, f40
  2785. FMPY f79 = f79, f40
  2786. ;;
  2787. FNMA f80 = f72, f41, f80
  2788. FNMA f84 = f76, f41, f84
  2789. FNMA f81 = f73, f41, f81
  2790. FNMA f85 = f77, f41, f85
  2791. FNMA f82 = f74, f41, f82
  2792. FNMA f86 = f78, f41, f86
  2793. FNMA f83 = f75, f41, f83
  2794. FNMA f87 = f79, f41, f87
  2795. ;;
  2796. FNMA f88 = f72, f42, f88
  2797. FNMA f92 = f76, f42, f92
  2798. FNMA f89 = f73, f42, f89
  2799. FNMA f93 = f77, f42, f93
  2800. FNMA f90 = f74, f42, f90
  2801. FNMA f94 = f78, f42, f94
  2802. FNMA f91 = f75, f42, f91
  2803. FNMA f95 = f79, f42, f95
  2804. ;;
  2805. FNMA f96 = f72, f43, f96
  2806. FNMA f100 = f76, f43, f100
  2807. FNMA f97 = f73, f43, f97
  2808. FNMA f101 = f77, f43, f101
  2809. FNMA f98 = f74, f43, f98
  2810. FNMA f102 = f78, f43, f102
  2811. FNMA f99 = f75, f43, f99
  2812. FNMA f103 = f79, f43, f103
  2813. ;;
  2814. FNMA f104 = f72, f44, f104
  2815. FNMA f108 = f76, f44, f108
  2816. FNMA f105 = f73, f44, f105
  2817. FNMA f109 = f77, f44, f109
  2818. FNMA f106 = f74, f44, f106
  2819. FNMA f110 = f78, f44, f110
  2820. FNMA f107 = f75, f44, f107
  2821. FNMA f111 = f79, f44, f111
  2822. ;;
  2823. FNMA f112 = f72, f45, f112
  2824. FNMA f116 = f76, f45, f116
  2825. FNMA f113 = f73, f45, f113
  2826. FNMA f117 = f77, f45, f117
  2827. FNMA f114 = f74, f45, f114
  2828. FNMA f118 = f78, f45, f118
  2829. FNMA f115 = f75, f45, f115
  2830. FNMA f119 = f79, f45, f119
  2831. ;;
  2832. FNMA f120 = f72, f46, f120
  2833. FNMA f124 = f76, f46, f124
  2834. FNMA f121 = f73, f46, f121
  2835. FNMA f125 = f77, f46, f125
  2836. FNMA f122 = f74, f46, f122
  2837. FNMA f126 = f78, f46, f126
  2838. FNMA f123 = f75, f46, f123
  2839. FNMA f127 = f79, f46, f127
  2840. ;;
  2841. FMPY f80 = f80, f47
  2842. FMPY f84 = f84, f47
  2843. FMPY f81 = f81, f47
  2844. FMPY f85 = f85, f47
  2845. FMPY f82 = f82, f47
  2846. FMPY f86 = f86, f47
  2847. FMPY f83 = f83, f47
  2848. FMPY f87 = f87, f47
  2849. ;;
  2850. FNMA f88 = f80, f48, f88
  2851. FNMA f92 = f84, f48, f92
  2852. FNMA f89 = f81, f48, f89
  2853. FNMA f93 = f85, f48, f93
  2854. FNMA f90 = f82, f48, f90
  2855. FNMA f94 = f86, f48, f94
  2856. FNMA f91 = f83, f48, f91
  2857. FNMA f95 = f87, f48, f95
  2858. ;;
  2859. FNMA f96 = f80, f49, f96
  2860. FNMA f100 = f84, f49, f100
  2861. FNMA f97 = f81, f49, f97
  2862. FNMA f101 = f85, f49, f101
  2863. FNMA f98 = f82, f49, f98
  2864. FNMA f102 = f86, f49, f102
  2865. FNMA f99 = f83, f49, f99
  2866. FNMA f103 = f87, f49, f103
  2867. ;;
  2868. FNMA f104 = f80, f50, f104
  2869. FNMA f108 = f84, f50, f108
  2870. FNMA f105 = f81, f50, f105
  2871. FNMA f109 = f85, f50, f109
  2872. FNMA f106 = f82, f50, f106
  2873. FNMA f110 = f86, f50, f110
  2874. FNMA f107 = f83, f50, f107
  2875. FNMA f111 = f87, f50, f111
  2876. ;;
  2877. FNMA f112 = f80, f51, f112
  2878. FNMA f116 = f84, f51, f116
  2879. FNMA f113 = f81, f51, f113
  2880. FNMA f117 = f85, f51, f117
  2881. FNMA f114 = f82, f51, f114
  2882. FNMA f118 = f86, f51, f118
  2883. FNMA f115 = f83, f51, f115
  2884. FNMA f119 = f87, f51, f119
  2885. ;;
  2886. FNMA f120 = f80, f52, f120
  2887. FNMA f124 = f84, f52, f124
  2888. FNMA f121 = f81, f52, f121
  2889. FNMA f125 = f85, f52, f125
  2890. FNMA f122 = f82, f52, f122
  2891. FNMA f126 = f86, f52, f126
  2892. FNMA f123 = f83, f52, f123
  2893. FNMA f127 = f87, f52, f127
  2894. ;;
  2895. FMPY f88 = f88, f53
  2896. FMPY f92 = f92, f53
  2897. FMPY f89 = f89, f53
  2898. FMPY f93 = f93, f53
  2899. FMPY f90 = f90, f53
  2900. FMPY f94 = f94, f53
  2901. FMPY f91 = f91, f53
  2902. FMPY f95 = f95, f53
  2903. ;;
  2904. FNMA f96 = f88, f54, f96
  2905. FNMA f100 = f92, f54, f100
  2906. FNMA f97 = f89, f54, f97
  2907. FNMA f101 = f93, f54, f101
  2908. FNMA f98 = f90, f54, f98
  2909. FNMA f102 = f94, f54, f102
  2910. FNMA f99 = f91, f54, f99
  2911. FNMA f103 = f95, f54, f103
  2912. ;;
  2913. FNMA f104 = f88, f55, f104
  2914. FNMA f108 = f92, f55, f108
  2915. FNMA f105 = f89, f55, f105
  2916. FNMA f109 = f93, f55, f109
  2917. FNMA f106 = f90, f55, f106
  2918. FNMA f110 = f94, f55, f110
  2919. FNMA f107 = f91, f55, f107
  2920. FNMA f111 = f95, f55, f111
  2921. ;;
  2922. FNMA f112 = f88, f56, f112
  2923. FNMA f116 = f92, f56, f116
  2924. FNMA f113 = f89, f56, f113
  2925. FNMA f117 = f93, f56, f117
  2926. FNMA f114 = f90, f56, f114
  2927. FNMA f118 = f94, f56, f118
  2928. FNMA f115 = f91, f56, f115
  2929. FNMA f119 = f95, f56, f119
  2930. ;;
  2931. FNMA f120 = f88, f57, f120
  2932. FNMA f124 = f92, f57, f124
  2933. FNMA f121 = f89, f57, f121
  2934. FNMA f125 = f93, f57, f125
  2935. FNMA f122 = f90, f57, f122
  2936. FNMA f126 = f94, f57, f126
  2937. FNMA f123 = f91, f57, f123
  2938. FNMA f127 = f95, f57, f127
  2939. ;;
  2940. FMPY f96 = f96, f58
  2941. FMPY f100 = f100, f58
  2942. FMPY f97 = f97, f58
  2943. FMPY f101 = f101, f58
  2944. FMPY f98 = f98, f58
  2945. FMPY f102 = f102, f58
  2946. FMPY f99 = f99, f58
  2947. FMPY f103 = f103, f58
  2948. ;;
  2949. FNMA f104 = f96, f59, f104
  2950. FNMA f108 = f100, f59, f108
  2951. FNMA f105 = f97, f59, f105
  2952. FNMA f109 = f101, f59, f109
  2953. FNMA f106 = f98, f59, f106
  2954. FNMA f110 = f102, f59, f110
  2955. FNMA f107 = f99, f59, f107
  2956. FNMA f111 = f103, f59, f111
  2957. ;;
  2958. FNMA f112 = f96, f60, f112
  2959. FNMA f116 = f100, f60, f116
  2960. FNMA f113 = f97, f60, f113
  2961. FNMA f117 = f101, f60, f117
  2962. FNMA f114 = f98, f60, f114
  2963. FNMA f118 = f102, f60, f118
  2964. FNMA f115 = f99, f60, f115
  2965. FNMA f119 = f103, f60, f119
  2966. ;;
  2967. { .mfi
  2968. STFD [AOFFSET] = f64, SIZE
  2969. FNMA f120 = f96, f61, f120
  2970. }
  2971. { .mfi
  2972. STFD [AOFFSET2] = f68, SIZE
  2973. FNMA f124 = f100, f61, f124
  2974. }
  2975. ;;
  2976. { .mfi
  2977. STFD [AOFFSET] = f65, SIZE
  2978. FNMA f121 = f97, f61, f121
  2979. }
  2980. { .mfi
  2981. STFD [AOFFSET2] = f69, SIZE
  2982. FNMA f125 = f101, f61, f125
  2983. }
  2984. ;;
  2985. { .mfi
  2986. STFD [AOFFSET] = f66, SIZE
  2987. FNMA f122 = f98, f61, f122
  2988. }
  2989. { .mfi
  2990. STFD [AOFFSET2] = f70, SIZE
  2991. FNMA f126 = f102, f61, f126
  2992. }
  2993. ;;
  2994. { .mfi
  2995. STFD [AOFFSET] = f67, 5 * SIZE
  2996. FNMA f123 = f99, f61, f123
  2997. }
  2998. { .mfi
  2999. STFD [AOFFSET2] = f71, 5 * SIZE
  3000. FNMA f127 = f103, f61, f127
  3001. }
  3002. ;;
  3003. { .mfi
  3004. STFD [AOFFSET] = f72, SIZE
  3005. FMPY f104 = f104, f16
  3006. }
  3007. { .mfi
  3008. STFD [AOFFSET2] = f76, SIZE
  3009. FMPY f108 = f108, f16
  3010. }
  3011. ;;
  3012. { .mfi
  3013. STFD [AOFFSET] = f73, SIZE
  3014. FMPY f105 = f105, f16
  3015. }
  3016. { .mfi
  3017. STFD [AOFFSET2] = f77, SIZE
  3018. FMPY f109 = f109, f16
  3019. }
  3020. ;;
  3021. { .mfi
  3022. STFD [AOFFSET] = f74, SIZE
  3023. FMPY f106 = f106, f16
  3024. }
  3025. { .mfi
  3026. STFD [AOFFSET2] = f78, SIZE
  3027. FMPY f110 = f110, f16
  3028. }
  3029. ;;
  3030. { .mfi
  3031. STFD [AOFFSET] = f75, 5 * SIZE
  3032. FMPY f107 = f107, f16
  3033. }
  3034. { .mfi
  3035. STFD [AOFFSET2] = f79, 5 * SIZE
  3036. FMPY f111 = f111, f16
  3037. }
  3038. ;;
  3039. { .mfi
  3040. STFD [AOFFSET] = f80, SIZE
  3041. FNMA f112 = f104, f17, f112
  3042. }
  3043. { .mfi
  3044. STFD [AOFFSET2] = f84, SIZE
  3045. FNMA f116 = f108, f17, f116
  3046. }
  3047. ;;
  3048. { .mfi
  3049. STFD [AOFFSET] = f81, SIZE
  3050. FNMA f113 = f105, f17, f113
  3051. }
  3052. { .mfi
  3053. STFD [AOFFSET2] = f85, SIZE
  3054. FNMA f117 = f109, f17, f117
  3055. }
  3056. ;;
  3057. { .mfi
  3058. STFD [AOFFSET] = f82, SIZE
  3059. FNMA f114 = f106, f17, f114
  3060. }
  3061. { .mfi
  3062. STFD [AOFFSET2] = f86, SIZE
  3063. FNMA f118 = f110, f17, f118
  3064. }
  3065. ;;
  3066. { .mfi
  3067. STFD [AOFFSET] = f83, 5 * SIZE
  3068. FNMA f115 = f107, f17, f115
  3069. }
  3070. { .mfi
  3071. STFD [AOFFSET2] = f87, 5 * SIZE
  3072. FNMA f119 = f111, f17, f119
  3073. }
  3074. ;;
  3075. { .mfi
  3076. STFD [AOFFSET] = f88, SIZE
  3077. FNMA f120 = f104, f18, f120
  3078. }
  3079. { .mfi
  3080. STFD [AOFFSET2] = f92, SIZE
  3081. FNMA f124 = f108, f18, f124
  3082. }
  3083. ;;
  3084. { .mfi
  3085. STFD [AOFFSET] = f89, SIZE
  3086. FNMA f121 = f105, f18, f121
  3087. }
  3088. { .mfi
  3089. STFD [AOFFSET2] = f93, SIZE
  3090. FNMA f125 = f109, f18, f125
  3091. }
  3092. ;;
  3093. { .mfi
  3094. STFD [AOFFSET] = f90, SIZE
  3095. FNMA f122 = f106, f18, f122
  3096. }
  3097. { .mfi
  3098. STFD [AOFFSET2] = f94, SIZE
  3099. FNMA f126 = f110, f18, f126
  3100. }
  3101. ;;
  3102. { .mfi
  3103. STFD [AOFFSET] = f91, 5 * SIZE
  3104. FNMA f123 = f107, f18, f123
  3105. }
  3106. { .mfi
  3107. STFD [AOFFSET2] = f95, 5 * SIZE
  3108. FNMA f127 = f111, f18, f127
  3109. }
  3110. ;;
  3111. { .mfi
  3112. STFD [AOFFSET] = f96, SIZE
  3113. FMPY f112 = f112, f19
  3114. }
  3115. { .mfi
  3116. STFD [AOFFSET2] = f100, SIZE
  3117. FMPY f116 = f116, f19
  3118. }
  3119. ;;
  3120. { .mfi
  3121. STFD [AOFFSET] = f97, SIZE
  3122. FMPY f113 = f113, f19
  3123. }
  3124. { .mfi
  3125. STFD [AOFFSET2] = f101, SIZE
  3126. FMPY f117 = f117, f19
  3127. }
  3128. ;;
  3129. { .mfi
  3130. STFD [AOFFSET] = f98, SIZE
  3131. FMPY f114 = f114, f19
  3132. }
  3133. { .mfi
  3134. STFD [AOFFSET2] = f102, SIZE
  3135. FMPY f118 = f118, f19
  3136. }
  3137. ;;
  3138. { .mfi
  3139. STFD [AOFFSET] = f99, 5 * SIZE
  3140. FMPY f115 = f115, f19
  3141. }
  3142. { .mfi
  3143. STFD [AOFFSET2] = f103, 5 * SIZE
  3144. FMPY f119 = f119, f19
  3145. }
  3146. ;;
  3147. { .mfi
  3148. STFD [AOFFSET] = f104, SIZE
  3149. FNMA f120 = f112, f20, f120
  3150. }
  3151. { .mfi
  3152. STFD [AOFFSET2] = f108, SIZE
  3153. FNMA f124 = f116, f20, f124
  3154. }
  3155. ;;
  3156. { .mfi
  3157. STFD [AOFFSET] = f105, SIZE
  3158. FNMA f121 = f113, f20, f121
  3159. }
  3160. { .mfi
  3161. STFD [AOFFSET2] = f109, SIZE
  3162. FNMA f125 = f117, f20, f125
  3163. }
  3164. ;;
  3165. { .mfi
  3166. STFD [AOFFSET] = f106, SIZE
  3167. FNMA f122 = f114, f20, f122
  3168. }
  3169. { .mfi
  3170. STFD [AOFFSET2] = f110, SIZE
  3171. FNMA f126 = f118, f20, f126
  3172. }
  3173. ;;
  3174. { .mfi
  3175. STFD [AOFFSET] = f107, 5 * SIZE
  3176. FNMA f123 = f115, f20, f123
  3177. }
  3178. { .mfi
  3179. STFD [AOFFSET2] = f111, 5 * SIZE
  3180. FNMA f127 = f119, f20, f127
  3181. }
  3182. ;;
  3183. { .mfi
  3184. STFD [AOFFSET] = f112, SIZE
  3185. FMPY f120 = f120, f21
  3186. }
  3187. { .mfi
  3188. STFD [AOFFSET2] = f116, SIZE
  3189. FMPY f124 = f124, f21
  3190. }
  3191. ;;
  3192. { .mfi
  3193. STFD [AOFFSET] = f113, SIZE
  3194. FMPY f121 = f121, f21
  3195. }
  3196. { .mfi
  3197. STFD [AOFFSET2] = f117, SIZE
  3198. FMPY f125 = f125, f21
  3199. }
  3200. ;;
  3201. { .mfi
  3202. STFD [AOFFSET] = f114, SIZE
  3203. FMPY f122 = f122, f21
  3204. }
  3205. { .mfi
  3206. STFD [AOFFSET2] = f118, SIZE
  3207. FMPY f126 = f126, f21
  3208. }
  3209. ;;
  3210. { .mfi
  3211. STFD [AOFFSET] = f115, 5 * SIZE
  3212. FMPY f123 = f123, f21
  3213. }
  3214. { .mfi
  3215. STFD [AOFFSET2] = f119, 5 * SIZE
  3216. FMPY f127 = f127, f21
  3217. }
  3218. ;;
  3219. { .mmi
  3220. STFD [AOFFSET] = f120, SIZE
  3221. STFD [AOFFSET2] = f124, SIZE
  3222. }
  3223. ;;
  3224. { .mmi
  3225. STFD [AOFFSET] = f121, SIZE
  3226. STFD [AOFFSET2] = f125, SIZE
  3227. }
  3228. ;;
  3229. { .mmi
  3230. STFD [AOFFSET] = f122, SIZE
  3231. STFD [AOFFSET2] = f126, SIZE
  3232. adds C9 = 4 * SIZE, C1
  3233. }
  3234. ;;
  3235. { .mfi
  3236. STFD [AOFFSET] = f123
  3237. adds AOFFSET = - 59 * SIZE, AOFFSET
  3238. }
  3239. { .mfi
  3240. STFD [AOFFSET2] = f127
  3241. adds AOFFSET2 = - 59 * SIZE, AOFFSET2
  3242. }
  3243. ;;
  3244. #endif
  3245. { .mmf
  3246. STFD [C1 ] = f64, SIZE
  3247. STFD [C9 ] = f68, SIZE
  3248. mov f64 = f0
  3249. }
  3250. ;;
  3251. { .mmi
  3252. STFD [C1 ] = f65, SIZE
  3253. STFD [C9 ] = f69, SIZE
  3254. adds C10 = 4 * SIZE, C2
  3255. }
  3256. ;;
  3257. { .mmi
  3258. STFD [C1 ] = f66, SIZE
  3259. STFD [C9 ] = f70, SIZE
  3260. nop __LINE__
  3261. }
  3262. ;;
  3263. { .mmi
  3264. STFD [C1 ] = f67, 5 * SIZE
  3265. STFD [C9 ] = f71
  3266. adds C11 = 4 * SIZE, C3
  3267. }
  3268. ;;
  3269. { .mmf
  3270. STFD [C2 ] = f72, SIZE
  3271. STFD [C10] = f76, SIZE
  3272. mov f72 = f0
  3273. }
  3274. ;;
  3275. { .mmi
  3276. STFD [C2 ] = f73, SIZE
  3277. STFD [C10] = f77, SIZE
  3278. nop __LINE__
  3279. }
  3280. ;;
  3281. { .mmi
  3282. STFD [C2 ] = f74, SIZE
  3283. STFD [C10] = f78, SIZE
  3284. adds C12 = 4 * SIZE, C4
  3285. }
  3286. ;;
  3287. { .mmi
  3288. STFD [C2 ] = f75, 5 * SIZE
  3289. STFD [C10] = f79
  3290. nop __LINE__
  3291. }
  3292. ;;
  3293. { .mmf
  3294. STFD [C3 ] = f80, SIZE
  3295. STFD [C11] = f84, SIZE
  3296. mov f80 = f0
  3297. }
  3298. ;;
  3299. { .mmi
  3300. STFD [C3 ] = f81, SIZE
  3301. STFD [C11] = f85, SIZE
  3302. adds C13 = 4 * SIZE, C5
  3303. }
  3304. ;;
  3305. { .mmi
  3306. STFD [C3 ] = f82, SIZE
  3307. STFD [C11] = f86, SIZE
  3308. nop __LINE__
  3309. }
  3310. ;;
  3311. { .mmi
  3312. STFD [C3 ] = f83, 5 * SIZE
  3313. STFD [C11] = f87
  3314. adds C14 = 4 * SIZE, C6
  3315. }
  3316. ;;
  3317. { .mmf
  3318. STFD [C4 ] = f88, SIZE
  3319. STFD [C12] = f92, SIZE
  3320. mov f88 = f0
  3321. }
  3322. ;;
  3323. { .mmi
  3324. STFD [C4 ] = f89, SIZE
  3325. STFD [C12] = f93, SIZE
  3326. nop __LINE__
  3327. }
  3328. ;;
  3329. { .mmi
  3330. STFD [C4 ] = f90, SIZE
  3331. STFD [C12] = f94, SIZE
  3332. adds C16 = 4 * SIZE, C8
  3333. }
  3334. ;;
  3335. { .mmi
  3336. STFD [C4 ] = f91, 5 * SIZE
  3337. STFD [C12] = f95
  3338. cmp.ne p6, p0 = 1, I
  3339. }
  3340. ;;
  3341. { .mmf
  3342. STFD [C5 ] = f96, SIZE
  3343. STFD [C13] = f100, SIZE
  3344. mov f96 = f0
  3345. }
  3346. ;;
  3347. { .mmi
  3348. STFD [C5 ] = f97, SIZE
  3349. STFD [C13] = f101, SIZE
  3350. adds I = -1, I
  3351. }
  3352. ;;
  3353. { .mmi
  3354. STFD [C5 ] = f98, SIZE
  3355. STFD [C13] = f102, SIZE
  3356. nop __LINE__
  3357. }
  3358. ;;
  3359. { .mmi
  3360. STFD [C5 ] = f99, 5 * SIZE
  3361. STFD [C13] = f103
  3362. adds C15 = 4 * SIZE, C7
  3363. }
  3364. ;;
  3365. { .mmf
  3366. STFD [C6 ] = f104, SIZE
  3367. STFD [C14] = f108, SIZE
  3368. mov f104 = f0
  3369. }
  3370. ;;
  3371. { .mmi
  3372. STFD [C6 ] = f105, SIZE
  3373. STFD [C14] = f109, SIZE
  3374. nop __LINE__
  3375. }
  3376. ;;
  3377. { .mmi
  3378. STFD [C6 ] = f106, SIZE
  3379. STFD [C14] = f110, SIZE
  3380. sub L = K, KK
  3381. }
  3382. ;;
  3383. { .mmi
  3384. STFD [C6 ] = f107, 5 * SIZE
  3385. STFD [C14] = f111
  3386. nop __LINE__
  3387. }
  3388. ;;
  3389. { .mmf
  3390. STFD [C7 ] = f112, SIZE
  3391. STFD [C15] = f116, SIZE
  3392. mov f112 = f0
  3393. }
  3394. ;;
  3395. { .mmi
  3396. STFD [C7 ] = f113, SIZE
  3397. STFD [C15] = f117, SIZE
  3398. shladd L = L, BASE_SHIFT, r0
  3399. }
  3400. ;;
  3401. { .mmi
  3402. STFD [C7 ] = f114, SIZE
  3403. STFD [C15] = f118, SIZE
  3404. shladd AOFFSET = L, 3, AOFFSET
  3405. }
  3406. ;;
  3407. { .mmi
  3408. STFD [C7 ] = f115, 5 * SIZE
  3409. STFD [C15] = f119
  3410. shladd BOFFSET = L, 3, BOFFSET
  3411. }
  3412. ;;
  3413. { .mmf
  3414. STFD [C8 ] = f120, SIZE
  3415. STFD [C16] = f124, SIZE
  3416. mov f120 = f0
  3417. }
  3418. ;;
  3419. { .mmi
  3420. STFD [C8 ] = f121, SIZE
  3421. STFD [C16] = f125, SIZE
  3422. #ifdef LT
  3423. adds KK = 8, KK
  3424. #else
  3425. nop __LINE__
  3426. #endif
  3427. }
  3428. ;;
  3429. { .mmi
  3430. STFD [C8 ] = f122, SIZE
  3431. STFD [C16] = f126, SIZE
  3432. mov L = KK
  3433. }
  3434. ;;
  3435. { .mmb
  3436. STFD [C8 ] = f123, 5 * SIZE
  3437. STFD [C16] = f127
  3438. (p6) br.cond.dptk .L011
  3439. }
  3440. ;;
  3441. .L020:
  3442. { .mib
  3443. mov L = KK
  3444. tbit.z p6, p0 = M, 2
  3445. (p6) br.cond.dptk .L030
  3446. }
  3447. ;;
  3448. { .mmi
  3449. cmp.ne p7, p0 = r0, L
  3450. adds BOFFSET = 0 * SIZE, B
  3451. nop __LINE__
  3452. }
  3453. ;;
  3454. { .mmf
  3455. (p7) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  3456. setf.d f73 = r0
  3457. mov f65 = f0
  3458. }
  3459. ;;
  3460. { .mfi
  3461. setf.d f105 = r0
  3462. mov f81 = f0
  3463. adds L = 1, L
  3464. }
  3465. { .mfi
  3466. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  3467. mov f89 = f0
  3468. cmp.eq p3, p0 = r0, r0
  3469. }
  3470. ;;
  3471. { .mfi
  3472. (p7) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  3473. mov f113 = f0
  3474. tbit.z p12, p0 = L, 0
  3475. }
  3476. { .mfi
  3477. setf.d f97 = r0
  3478. mov f121 = f0
  3479. shr L = L, 1
  3480. }
  3481. ;;
  3482. { .mmf
  3483. (p7) LDFPD f52, f53 = [BOFFSET], 2 * SIZE
  3484. setf.d f66 = r0
  3485. mov f67 = f0
  3486. }
  3487. { .mfi
  3488. setf.d f74 = r0
  3489. mov f75 = f0
  3490. adds L = -1, L
  3491. }
  3492. ;;
  3493. { .mmf
  3494. (p7) LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  3495. setf.d f82 = r0
  3496. mov f83 = f0
  3497. }
  3498. { .mfi
  3499. setf.d f90 = r0
  3500. mov f91 = f0
  3501. cmp.eq p6, p0 = -1, L
  3502. }
  3503. ;;
  3504. { .mmf
  3505. (p7) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  3506. setf.d f98 = r0
  3507. mov f99 = f0
  3508. }
  3509. { .mfi
  3510. setf.d f106 = r0
  3511. mov f107 = f0
  3512. mov ar.lc = L
  3513. }
  3514. ;;
  3515. { .mmf
  3516. (p7) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  3517. setf.d f114 = r0
  3518. mov f115 = f0
  3519. }
  3520. { .mfb
  3521. setf.d f122 = r0
  3522. mov f123 = f0
  3523. (p6) br.cond.dpnt .L028
  3524. }
  3525. ;;
  3526. .L022:
  3527. { .mfi
  3528. lfetch.nt1 [PREA], 8 * SIZE
  3529. FMA f64 = f32, f48, f64 // A1 * B1
  3530. adds PREB = (PREFETCHSIZE + 0) * SIZE, BOFFSET
  3531. }
  3532. { .mfi
  3533. nop __LINE__
  3534. FMA f72 = f32, f49, f72 // A1 * B2
  3535. (p12) cmp.ne p3, p0 = 0, L
  3536. }
  3537. ;;
  3538. { .mfi
  3539. lfetch.nt1 [PREB], 16 * SIZE
  3540. FMA f80 = f32, f50, f80 // A1 * B3
  3541. cmp.ne p4, p5 = 0, L
  3542. }
  3543. { .mfb
  3544. nop __LINE__
  3545. FMA f88 = f32, f51, f88 // A1 * B4
  3546. nop __LINE__
  3547. }
  3548. ;;
  3549. { .mfi
  3550. (p3) LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  3551. FMA f96 = f32, f52, f96 // A1 * B5
  3552. (p5) adds C9 = 2 * SIZE, C1
  3553. }
  3554. { .mfi
  3555. nop __LINE__
  3556. FMA f104 = f32, f53, f104 // A1 * B6
  3557. (p5) adds C10 = 2 * SIZE, C2
  3558. }
  3559. ;;
  3560. { .mfi
  3561. (p3) LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  3562. FMA f112 = f32, f54, f112 // A1 * B7
  3563. (p5) adds C11 = 2 * SIZE, C3
  3564. }
  3565. { .mfi
  3566. nop __LINE__
  3567. FMA f120 = f32, f55, f120 // A1 * B8
  3568. (p5) adds C12 = 2 * SIZE, C4
  3569. }
  3570. ;;
  3571. { .mfi
  3572. (p3) LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  3573. FMA f65 = f33, f48, f65 // A2 * B1
  3574. (p5) adds C13 = 2 * SIZE, C5
  3575. }
  3576. { .mfi
  3577. nop __LINE__
  3578. FMA f73 = f33, f49, f73 // A2 * B2
  3579. (p5) adds C14 = 2 * SIZE, C6
  3580. }
  3581. ;;
  3582. { .mfi
  3583. (p3) LDFPD f60, f61 = [BOFFSET], 2 * SIZE
  3584. FMA f81 = f33, f50, f81 // A2 * B3
  3585. (p5) adds C15 = 2 * SIZE, C7
  3586. }
  3587. { .mfi
  3588. nop __LINE__
  3589. FMA f89 = f33, f51, f89 // A2 * B4
  3590. (p5) adds C16 = 2 * SIZE, C8
  3591. }
  3592. ;;
  3593. { .mfb
  3594. (p3) LDFPD f62, f63 = [BOFFSET], 2 * SIZE
  3595. FMA f97 = f33, f52, f97 // A2 * B5
  3596. nop __LINE__
  3597. }
  3598. { .mfb
  3599. nop __LINE__
  3600. FMA f105 = f33, f53, f105 // A2 * B6
  3601. nop __LINE__
  3602. }
  3603. ;;
  3604. { .mfb
  3605. (p3) LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  3606. FMA f113 = f33, f54, f113 // A2 * B7
  3607. nop __LINE__
  3608. }
  3609. { .mfb
  3610. nop __LINE__
  3611. FMA f121 = f33, f55, f121 // A2 * B8
  3612. nop __LINE__
  3613. }
  3614. ;;
  3615. { .mfb
  3616. nop __LINE__
  3617. FMA f66 = f34, f48, f66 // A3 * B1
  3618. nop __LINE__
  3619. }
  3620. { .mfb
  3621. nop __LINE__
  3622. FMA f74 = f34, f49, f74 // A3 * B2
  3623. nop __LINE__
  3624. }
  3625. ;;
  3626. { .mfb
  3627. nop __LINE__
  3628. FMA f82 = f34, f50, f82 // A3 * B3
  3629. nop __LINE__
  3630. }
  3631. { .mfb
  3632. nop __LINE__
  3633. FMA f90 = f34, f51, f90 // A3 * B4
  3634. nop __LINE__
  3635. }
  3636. ;;
  3637. { .mfb
  3638. nop __LINE__
  3639. FMA f98 = f34, f52, f98 // A3 * B5
  3640. nop __LINE__
  3641. }
  3642. { .mfb
  3643. nop __LINE__
  3644. FMA f106 = f34, f53, f106 // A3 * B6
  3645. nop __LINE__
  3646. }
  3647. { .mfb
  3648. nop __LINE__
  3649. FMA f114 = f34, f54, f114 // A3 * B7
  3650. nop __LINE__
  3651. }
  3652. { .mfb
  3653. nop __LINE__
  3654. FMA f122 = f34, f55, f122 // A3 * B8
  3655. nop __LINE__
  3656. }
  3657. { .mfb
  3658. nop __LINE__
  3659. FMA f67 = f35, f48, f67 // A4 * B1
  3660. nop __LINE__
  3661. }
  3662. { .mfb
  3663. nop __LINE__
  3664. FMA f75 = f35, f49, f75 // A4 * B2
  3665. nop __LINE__
  3666. }
  3667. { .mfb
  3668. nop __LINE__
  3669. FMA f83 = f35, f50, f83 // A4 * B3
  3670. nop __LINE__
  3671. }
  3672. { .mfb
  3673. nop __LINE__
  3674. FMA f91 = f35, f51, f91 // A4 * B4
  3675. nop __LINE__
  3676. }
  3677. { .mfb
  3678. (p4) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  3679. FMA f99 = f35, f52, f99 // A4 * B5
  3680. nop __LINE__
  3681. }
  3682. { .mfb
  3683. nop __LINE__
  3684. FMA f107 = f35, f53, f107 // A4 * B6
  3685. nop __LINE__
  3686. }
  3687. { .mfb
  3688. (p4) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  3689. FMA f115 = f35, f54, f115 // A4 * B7
  3690. nop __LINE__
  3691. }
  3692. { .mfb
  3693. nop __LINE__
  3694. FMA f123 = f35, f55, f123 // A4 * B8
  3695. nop __LINE__
  3696. }
  3697. ;;
  3698. { .mfb
  3699. (p4) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  3700. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  3701. nop __LINE__
  3702. }
  3703. { .mfb
  3704. nop __LINE__
  3705. (p3) FMA f72 = f40, f57, f72 // A1 * B2
  3706. nop __LINE__
  3707. }
  3708. ;;
  3709. { .mfb
  3710. (p4) LDFPD f52, f53 = [BOFFSET], 2 * SIZE
  3711. (p3) FMA f80 = f40, f58, f80 // A1 * B3
  3712. nop __LINE__
  3713. }
  3714. { .mfb
  3715. nop __LINE__
  3716. (p3) FMA f88 = f40, f59, f88 // A1 * B4
  3717. nop __LINE__
  3718. }
  3719. ;;
  3720. { .mfb
  3721. nop __LINE__
  3722. (p3) FMA f96 = f40, f60, f96 // A1 * B5
  3723. nop __LINE__
  3724. }
  3725. { .mfb
  3726. nop __LINE__
  3727. (p3) FMA f104 = f40, f61, f104 // A1 * B6
  3728. nop __LINE__
  3729. }
  3730. ;;
  3731. { .mfb
  3732. nop __LINE__
  3733. (p3) FMA f112 = f40, f62, f112 // A1 * B7
  3734. nop __LINE__
  3735. }
  3736. { .mfb
  3737. nop __LINE__
  3738. (p3) FMA f120 = f40, f63, f120 // A1 * B8
  3739. nop __LINE__
  3740. }
  3741. ;;
  3742. { .mfb
  3743. (p4) LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  3744. (p3) FMA f65 = f41, f56, f65 // A2 * B1
  3745. nop __LINE__
  3746. }
  3747. { .mfb
  3748. (p3) FMA f73 = f41, f57, f73 // A2 * B2
  3749. nop __LINE__
  3750. }
  3751. { .mfb
  3752. (p4) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  3753. (p3) FMA f81 = f41, f58, f81 // A2 * B3
  3754. nop __LINE__
  3755. }
  3756. { .mfb
  3757. (p3) FMA f89 = f41, f59, f89 // A2 * B4
  3758. nop __LINE__
  3759. }
  3760. ;;
  3761. { .mfb
  3762. nop __LINE__
  3763. (p3) FMA f97 = f41, f60, f97 // A2 * B5
  3764. nop __LINE__
  3765. }
  3766. { .mfb
  3767. nop __LINE__
  3768. (p3) FMA f105 = f41, f61, f105 // A2 * B6
  3769. nop __LINE__
  3770. }
  3771. ;;
  3772. { .mfb
  3773. nop __LINE__
  3774. (p3) FMA f113 = f41, f62, f113 // A2 * B7
  3775. nop __LINE__
  3776. }
  3777. { .mfb
  3778. nop __LINE__
  3779. (p3) FMA f121 = f41, f63, f121 // A2 * B8
  3780. nop __LINE__
  3781. }
  3782. ;;
  3783. { .mfb
  3784. nop __LINE__
  3785. (p3) FMA f66 = f42, f56, f66 // A3 * B1
  3786. nop __LINE__
  3787. }
  3788. { .mfb
  3789. nop __LINE__
  3790. (p3) FMA f74 = f42, f57, f74 // A3 * B2
  3791. nop __LINE__
  3792. }
  3793. ;;
  3794. { .mfb
  3795. nop __LINE__
  3796. (p3) FMA f82 = f42, f58, f82 // A3 * B3
  3797. nop __LINE__
  3798. }
  3799. { .mfb
  3800. nop __LINE__
  3801. (p3) FMA f90 = f42, f59, f90 // A3 * B4
  3802. nop __LINE__
  3803. }
  3804. ;;
  3805. { .mfb
  3806. nop __LINE__
  3807. (p3) FMA f98 = f42, f60, f98 // A3 * B5
  3808. nop __LINE__
  3809. }
  3810. { .mfb
  3811. nop __LINE__
  3812. (p3) FMA f106 = f42, f61, f106 // A3 * B6
  3813. nop __LINE__
  3814. }
  3815. ;;
  3816. { .mfb
  3817. nop __LINE__
  3818. (p3) FMA f114 = f42, f62, f114 // A3 * B7
  3819. nop __LINE__
  3820. }
  3821. { .mfb
  3822. nop __LINE__
  3823. (p3) FMA f122 = f42, f63, f122 // A3 * B8
  3824. nop __LINE__
  3825. }
  3826. ;;
  3827. { .mfb
  3828. nop __LINE__
  3829. (p3) FMA f67 = f43, f56, f67 // A4 * B1
  3830. nop __LINE__
  3831. }
  3832. { .mfb
  3833. nop __LINE__
  3834. (p3) FMA f75 = f43, f57, f75 // A4 * B2
  3835. nop __LINE__
  3836. }
  3837. ;;
  3838. { .mfb
  3839. nop __LINE__
  3840. (p3) FMA f83 = f43, f58, f83 // A4 * B3
  3841. nop __LINE__
  3842. }
  3843. { .mfb
  3844. nop __LINE__
  3845. (p3) FMA f91 = f43, f59, f91 // A4 * B4
  3846. nop __LINE__
  3847. }
  3848. ;;
  3849. { .mfb
  3850. nop __LINE__
  3851. (p3) FMA f99 = f43, f60, f99 // A4 * B5
  3852. nop __LINE__
  3853. }
  3854. { .mfb
  3855. nop __LINE__
  3856. (p3) FMA f107 = f43, f61, f107 // A4 * B6
  3857. nop __LINE__
  3858. }
  3859. ;;
  3860. { .mfi
  3861. nop __LINE__
  3862. (p3) FMA f115 = f43, f62, f115 // A4 * B7
  3863. adds L = -1, L
  3864. }
  3865. { .mfb
  3866. nop __LINE__
  3867. (p3) FMA f123 = f43, f63, f123 // A4 * B8
  3868. br.cloop.sptk.few .L022
  3869. }
  3870. ;;
  3871. .L028:
  3872. #ifdef LT
  3873. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  3874. ;;
  3875. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  3876. ;;
  3877. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  3878. ;;
  3879. LDFPD f38, f39 = [BOFFSET], 2 * SIZE
  3880. ;;
  3881. LDFPD f40, f41 = [BOFFSET], 2 * SIZE
  3882. ;;
  3883. LDFPD f42, f43 = [BOFFSET], 2 * SIZE
  3884. ;;
  3885. LDFPD f44, f45 = [BOFFSET], 2 * SIZE
  3886. ;;
  3887. LDFPD f46, f47 = [BOFFSET], 2 * SIZE
  3888. ;;
  3889. { .mfi
  3890. LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  3891. FSUB f64 = f32, f64
  3892. nop __LINE__
  3893. }
  3894. { .mfi
  3895. nop __LINE__
  3896. FSUB f72 = f33, f72
  3897. nop __LINE__
  3898. }
  3899. ;;
  3900. { .mfi
  3901. LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  3902. FSUB f80 = f34, f80
  3903. nop __LINE__
  3904. }
  3905. { .mfi
  3906. nop __LINE__
  3907. FSUB f88 = f35, f88
  3908. nop __LINE__
  3909. }
  3910. ;;
  3911. { .mfi
  3912. LDFPD f52, f53 = [BOFFSET], 2 * SIZE
  3913. FSUB f96 = f36, f96
  3914. nop __LINE__
  3915. }
  3916. { .mfi
  3917. nop __LINE__
  3918. FSUB f104 = f37, f104
  3919. nop __LINE__
  3920. }
  3921. ;;
  3922. { .mfi
  3923. LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  3924. FSUB f112 = f38, f112
  3925. nop __LINE__
  3926. }
  3927. { .mfi
  3928. nop __LINE__
  3929. FSUB f120 = f39, f120
  3930. nop __LINE__
  3931. }
  3932. ;;
  3933. { .mfi
  3934. LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  3935. FSUB f65 = f40, f65
  3936. nop __LINE__
  3937. }
  3938. { .mfi
  3939. nop __LINE__
  3940. FSUB f73 = f41, f73
  3941. nop __LINE__
  3942. }
  3943. ;;
  3944. { .mfi
  3945. LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  3946. FSUB f81 = f42, f81
  3947. nop __LINE__
  3948. }
  3949. { .mfi
  3950. nop __LINE__
  3951. FSUB f89 = f43, f89
  3952. nop __LINE__
  3953. }
  3954. ;;
  3955. { .mfi
  3956. LDFPD f60, f61 = [BOFFSET], 2 * SIZE
  3957. FSUB f97 = f44, f97
  3958. nop __LINE__
  3959. }
  3960. { .mfi
  3961. nop __LINE__
  3962. FSUB f105 = f45, f105
  3963. nop __LINE__
  3964. }
  3965. ;;
  3966. { .mfi
  3967. LDFPD f62, f63 = [BOFFSET]
  3968. FSUB f113 = f46, f113
  3969. adds BOFFSET = -30 * SIZE, BOFFSET
  3970. }
  3971. { .mfi
  3972. nop __LINE__
  3973. FSUB f121 = f47, f121
  3974. nop __LINE__
  3975. }
  3976. ;;
  3977. FSUB f66 = f48, f66
  3978. FSUB f74 = f49, f74
  3979. FSUB f82 = f50, f82
  3980. FSUB f90 = f51, f90
  3981. FSUB f98 = f52, f98
  3982. FSUB f106 = f53, f106
  3983. FSUB f114 = f54, f114
  3984. FSUB f122 = f55, f122
  3985. ;;
  3986. FSUB f67 = f56, f67
  3987. FSUB f75 = f57, f75
  3988. FSUB f83 = f58, f83
  3989. FSUB f91 = f59, f91
  3990. FSUB f99 = f60, f99
  3991. FSUB f107 = f61, f107
  3992. FSUB f115 = f62, f115
  3993. FSUB f123 = f63, f123
  3994. ;;
  3995. #else
  3996. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  3997. ;;
  3998. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  3999. ;;
  4000. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  4001. ;;
  4002. LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  4003. ;;
  4004. LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  4005. ;;
  4006. LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  4007. ;;
  4008. LDFPD f44, f45 = [AOFFSET], 2 * SIZE
  4009. ;;
  4010. LDFPD f46, f47 = [AOFFSET], 2 * SIZE
  4011. ;;
  4012. LDFPD f48, f49 = [AOFFSET], 2 * SIZE
  4013. ;;
  4014. LDFPD f50, f51 = [AOFFSET], 2 * SIZE
  4015. ;;
  4016. LDFPD f52, f53 = [AOFFSET], 2 * SIZE
  4017. ;;
  4018. LDFPD f54, f55 = [AOFFSET], 2 * SIZE
  4019. ;;
  4020. LDFPD f56, f57 = [AOFFSET], 2 * SIZE
  4021. ;;
  4022. LDFPD f58, f59 = [AOFFSET], 2 * SIZE
  4023. ;;
  4024. LDFPD f60, f61 = [AOFFSET], 2 * SIZE
  4025. ;;
  4026. LDFPD f62, f63 = [AOFFSET]
  4027. adds AOFFSET = -30 * SIZE, AOFFSET
  4028. ;;
  4029. FSUB f64 = f32, f64
  4030. FSUB f65 = f33, f65
  4031. FSUB f66 = f34, f66
  4032. FSUB f67 = f35, f67
  4033. FSUB f72 = f36, f72
  4034. FSUB f73 = f37, f73
  4035. FSUB f74 = f38, f74
  4036. FSUB f75 = f39, f75
  4037. FSUB f80 = f40, f80
  4038. FSUB f81 = f41, f81
  4039. FSUB f82 = f42, f82
  4040. FSUB f83 = f43, f83
  4041. FSUB f88 = f44, f88
  4042. FSUB f89 = f45, f89
  4043. FSUB f90 = f46, f90
  4044. FSUB f91 = f47, f91
  4045. ;;
  4046. FSUB f96 = f48, f96
  4047. FSUB f97 = f49, f97
  4048. FSUB f98 = f50, f98
  4049. FSUB f99 = f51, f99
  4050. ;;
  4051. FSUB f104 = f52, f104
  4052. FSUB f105 = f53, f105
  4053. FSUB f106 = f54, f106
  4054. FSUB f107 = f55, f107
  4055. ;;
  4056. FSUB f112 = f56, f112
  4057. FSUB f113 = f57, f113
  4058. FSUB f114 = f58, f114
  4059. FSUB f115 = f59, f115
  4060. ;;
  4061. FSUB f120 = f60, f120
  4062. FSUB f121 = f61, f121
  4063. FSUB f122 = f62, f122
  4064. FSUB f123 = f63, f123
  4065. ;;
  4066. #endif
  4067. #ifdef LT
  4068. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  4069. ;;
  4070. LDFPD f34, f35 = [AOFFSET]
  4071. adds AOFFSET = 3 * SIZE, AOFFSET
  4072. ;;
  4073. LDFD f36 = [AOFFSET], 1 * SIZE
  4074. ;;
  4075. LDFPD f37, f38 = [AOFFSET]
  4076. adds AOFFSET = 4 * SIZE, AOFFSET
  4077. ;;
  4078. LDFPD f39, f40 = [AOFFSET]
  4079. adds AOFFSET = 5 * SIZE, AOFFSET
  4080. ;;
  4081. LDFD f41 = [AOFFSET], -15 * SIZE
  4082. ;;
  4083. { .mfi
  4084. FMPY f64 = f64, f32
  4085. nop __LINE__
  4086. }
  4087. { .mfi
  4088. nop __LINE__
  4089. FMPY f96 = f96, f32
  4090. nop __LINE__
  4091. }
  4092. ;;
  4093. { .mfi
  4094. FMPY f72 = f72, f32
  4095. nop __LINE__
  4096. }
  4097. { .mfi
  4098. nop __LINE__
  4099. FMPY f104 = f104, f32
  4100. nop __LINE__
  4101. }
  4102. ;;
  4103. { .mfi
  4104. FMPY f80 = f80, f32
  4105. }
  4106. { .mfi
  4107. nop __LINE__
  4108. FMPY f112 = f112, f32
  4109. nop __LINE__
  4110. }
  4111. ;;
  4112. { .mfi
  4113. FMPY f88 = f88, f32
  4114. nop __LINE__
  4115. }
  4116. { .mfi
  4117. nop __LINE__
  4118. FMPY f120 = f120, f32
  4119. nop __LINE__
  4120. }
  4121. ;;
  4122. { .mfi
  4123. FNMA f65 = f64, f33, f65
  4124. nop __LINE__
  4125. }
  4126. { .mfi
  4127. nop __LINE__
  4128. FNMA f97 = f96, f33, f97
  4129. nop __LINE__
  4130. }
  4131. ;;
  4132. { .mfi
  4133. FNMA f73 = f72, f33, f73
  4134. nop __LINE__
  4135. }
  4136. { .mfi
  4137. nop __LINE__
  4138. FNMA f105 = f104, f33, f105
  4139. nop __LINE__
  4140. }
  4141. ;;
  4142. { .mfi
  4143. FNMA f81 = f80, f33, f81
  4144. }
  4145. { .mfi
  4146. nop __LINE__
  4147. FNMA f113 = f112, f33, f113
  4148. nop __LINE__
  4149. }
  4150. ;;
  4151. { .mfi
  4152. FNMA f89 = f88, f33, f89
  4153. nop __LINE__
  4154. }
  4155. { .mfi
  4156. nop __LINE__
  4157. FNMA f121 = f120, f33, f121
  4158. nop __LINE__
  4159. }
  4160. ;;
  4161. { .mfi
  4162. FNMA f66 = f64, f34, f66
  4163. nop __LINE__
  4164. }
  4165. { .mfi
  4166. nop __LINE__
  4167. FNMA f98 = f96, f34, f98
  4168. nop __LINE__
  4169. }
  4170. ;;
  4171. { .mfi
  4172. FNMA f74 = f72, f34, f74
  4173. }
  4174. { .mfi
  4175. nop __LINE__
  4176. FNMA f106 = f104, f34, f106
  4177. nop __LINE__
  4178. }
  4179. ;;
  4180. { .mfi
  4181. FNMA f82 = f80, f34, f82
  4182. nop __LINE__
  4183. }
  4184. { .mfi
  4185. nop __LINE__
  4186. FNMA f114 = f112, f34, f114
  4187. nop __LINE__
  4188. }
  4189. ;;
  4190. { .mfi
  4191. FNMA f90 = f88, f34, f90
  4192. nop __LINE__
  4193. }
  4194. { .mfi
  4195. nop __LINE__
  4196. FNMA f122 = f120, f34, f122
  4197. nop __LINE__
  4198. }
  4199. ;;
  4200. { .mfi
  4201. FNMA f67 = f64, f35, f67
  4202. }
  4203. { .mfi
  4204. nop __LINE__
  4205. FNMA f99 = f96, f35, f99
  4206. nop __LINE__
  4207. }
  4208. ;;
  4209. { .mfi
  4210. FNMA f75 = f72, f35, f75
  4211. nop __LINE__
  4212. }
  4213. { .mfi
  4214. nop __LINE__
  4215. FNMA f107 = f104, f35, f107
  4216. nop __LINE__
  4217. }
  4218. ;;
  4219. { .mfi
  4220. FNMA f83 = f80, f35, f83
  4221. }
  4222. { .mfi
  4223. nop __LINE__
  4224. FNMA f115 = f112, f35, f115
  4225. nop __LINE__
  4226. }
  4227. ;;
  4228. { .mfi
  4229. FNMA f91 = f88, f35, f91
  4230. nop __LINE__
  4231. }
  4232. { .mfi
  4233. nop __LINE__
  4234. FNMA f123 = f120, f35, f123
  4235. adds BOFFSET2 = 4 * SIZE, BOFFSET
  4236. }
  4237. ;;
  4238. FMPY f65 = f65, f36
  4239. FMPY f97 = f97, f36
  4240. FMPY f73 = f73, f36
  4241. FMPY f105 = f105, f36
  4242. FMPY f81 = f81, f36
  4243. FMPY f113 = f113, f36
  4244. FMPY f89 = f89, f36
  4245. FMPY f121 = f121, f36
  4246. ;;
  4247. FNMA f66 = f65, f37, f66
  4248. FNMA f98 = f97, f37, f98
  4249. FNMA f74 = f73, f37, f74
  4250. FNMA f106 = f105, f37, f106
  4251. FNMA f82 = f81, f37, f82
  4252. FNMA f114 = f113, f37, f114
  4253. FNMA f90 = f89, f37, f90
  4254. FNMA f122 = f121, f37, f122
  4255. ;;
  4256. FNMA f67 = f65, f38, f67
  4257. FNMA f99 = f97, f38, f99
  4258. FNMA f75 = f73, f38, f75
  4259. FNMA f107 = f105, f38, f107
  4260. FNMA f83 = f81, f38, f83
  4261. FNMA f115 = f113, f38, f115
  4262. FNMA f91 = f89, f38, f91
  4263. FNMA f123 = f121, f38, f123
  4264. ;;
  4265. { .mfi
  4266. STFD [BOFFSET] = f64, SIZE
  4267. FMPY f66 = f66, f39
  4268. }
  4269. { .mfi
  4270. STFD [BOFFSET2] = f96, SIZE
  4271. FMPY f98 = f98, f39
  4272. }
  4273. ;;
  4274. { .mfi
  4275. STFD [BOFFSET] = f72, SIZE
  4276. FMPY f74 = f74, f39
  4277. }
  4278. { .mfi
  4279. STFD [BOFFSET2] = f104, SIZE
  4280. FMPY f106 = f106, f39
  4281. }
  4282. ;;
  4283. { .mfi
  4284. STFD [BOFFSET] = f80, SIZE
  4285. FMPY f82 = f82, f39
  4286. }
  4287. { .mfi
  4288. STFD [BOFFSET2] = f112, SIZE
  4289. FMPY f114 = f114, f39
  4290. }
  4291. ;;
  4292. { .mfi
  4293. STFD [BOFFSET] = f88, 5 * SIZE
  4294. FMPY f90 = f90, f39
  4295. }
  4296. { .mfi
  4297. STFD [BOFFSET2] = f120, 5 * SIZE
  4298. FMPY f122 = f122, f39
  4299. }
  4300. ;;
  4301. { .mfi
  4302. STFD [BOFFSET] = f65, SIZE
  4303. FNMA f67 = f66, f40, f67
  4304. }
  4305. { .mfi
  4306. STFD [BOFFSET2] = f97, SIZE
  4307. FNMA f99 = f98, f40, f99
  4308. }
  4309. ;;
  4310. { .mfi
  4311. STFD [BOFFSET] = f73, SIZE
  4312. FNMA f75 = f74, f40, f75
  4313. }
  4314. { .mfi
  4315. STFD [BOFFSET2] = f105, SIZE
  4316. FNMA f107 = f106, f40, f107
  4317. }
  4318. ;;
  4319. { .mfi
  4320. STFD [BOFFSET] = f81, SIZE
  4321. FNMA f83 = f82, f40, f83
  4322. }
  4323. { .mfi
  4324. STFD [BOFFSET2] = f113, SIZE
  4325. FNMA f115 = f114, f40, f115
  4326. }
  4327. ;;
  4328. { .mfi
  4329. STFD [BOFFSET] = f89, 5 * SIZE
  4330. FNMA f91 = f90, f40, f91
  4331. }
  4332. { .mfi
  4333. STFD [BOFFSET2] = f121, 5 * SIZE
  4334. FNMA f123 = f122, f40, f123
  4335. }
  4336. ;;
  4337. { .mfi
  4338. STFD [BOFFSET] = f66, SIZE
  4339. FMPY f67 = f67, f41
  4340. }
  4341. { .mfi
  4342. STFD [BOFFSET2] = f98, SIZE
  4343. FMPY f99 = f99, f41
  4344. }
  4345. ;;
  4346. { .mfi
  4347. STFD [BOFFSET] = f74, SIZE
  4348. FMPY f75 = f75, f41
  4349. }
  4350. { .mfi
  4351. STFD [BOFFSET2] = f106, SIZE
  4352. FMPY f107 = f107, f41
  4353. }
  4354. ;;
  4355. { .mfi
  4356. STFD [BOFFSET] = f82, SIZE
  4357. FMPY f83 = f83, f41
  4358. }
  4359. { .mfi
  4360. STFD [BOFFSET2] = f114, SIZE
  4361. FMPY f115 = f115, f41
  4362. }
  4363. ;;
  4364. { .mfi
  4365. STFD [BOFFSET] = f90, 5 * SIZE
  4366. FMPY f91 = f91, f41
  4367. }
  4368. { .mfi
  4369. STFD [BOFFSET2] = f122, 5 * SIZE
  4370. FMPY f123 = f123, f41
  4371. }
  4372. ;;
  4373. { .mmf
  4374. STFD [BOFFSET] = f67, SIZE
  4375. STFD [BOFFSET2] = f99, SIZE
  4376. }
  4377. ;;
  4378. { .mmf
  4379. STFD [BOFFSET] = f75, SIZE
  4380. STFD [BOFFSET2] = f107, SIZE
  4381. }
  4382. ;;
  4383. { .mmf
  4384. STFD [BOFFSET] = f83, SIZE
  4385. STFD [BOFFSET2] = f115, SIZE
  4386. }
  4387. ;;
  4388. { .mmf
  4389. STFD [BOFFSET] = f91, -27 * SIZE
  4390. STFD [BOFFSET2] = f123, -27 * SIZE
  4391. }
  4392. ;;
  4393. #endif
  4394. #ifdef RN
  4395. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  4396. ;;
  4397. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  4398. ;;
  4399. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  4400. ;;
  4401. LDFPD f38, f39 = [BOFFSET]
  4402. adds BOFFSET = 3 * SIZE, BOFFSET
  4403. ;;
  4404. LDFD f40 = [BOFFSET], 1 * SIZE
  4405. ;;
  4406. LDFPD f41, f42 = [BOFFSET], 2 * SIZE
  4407. ;;
  4408. LDFPD f43, f44 = [BOFFSET], 2 * SIZE
  4409. ;;
  4410. LDFPD f45, f46 = [BOFFSET]
  4411. adds BOFFSET = 4 * SIZE, BOFFSET
  4412. ;;
  4413. LDFPD f47, f48 = [BOFFSET], 2 * SIZE
  4414. ;;
  4415. LDFPD f49, f50 = [BOFFSET], 2 * SIZE
  4416. ;;
  4417. LDFPD f51, f52 = [BOFFSET]
  4418. adds BOFFSET = 5 * SIZE, BOFFSET
  4419. ;;
  4420. LDFD f53 = [BOFFSET], 1 * SIZE
  4421. ;;
  4422. LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  4423. ;;
  4424. LDFPD f56, f57 = [BOFFSET]
  4425. adds BOFFSET = 6 * SIZE, BOFFSET
  4426. ;;
  4427. LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  4428. adds AOFFSET2 = 4 * SIZE, AOFFSET
  4429. ;;
  4430. LDFPD f60, f61 = [BOFFSET]
  4431. adds BOFFSET = 7 * SIZE, BOFFSET
  4432. ;;
  4433. LDFD f16 = [BOFFSET], 1 * SIZE
  4434. ;;
  4435. LDFPD f17, f18 = [BOFFSET]
  4436. adds BOFFSET = 8 * SIZE, BOFFSET
  4437. ;;
  4438. LDFPD f19, f20 = [BOFFSET]
  4439. adds BOFFSET = 9 * SIZE, BOFFSET
  4440. ;;
  4441. LDFD f21 = [BOFFSET]
  4442. adds BOFFSET = -63 * SIZE, BOFFSET
  4443. ;;
  4444. FMPY f64 = f64, f32
  4445. FMPY f65 = f65, f32
  4446. FMPY f66 = f66, f32
  4447. FMPY f67 = f67, f32
  4448. ;;
  4449. FNMA f72 = f64, f33, f72
  4450. FNMA f73 = f65, f33, f73
  4451. FNMA f74 = f66, f33, f74
  4452. FNMA f75 = f67, f33, f75
  4453. ;;
  4454. FNMA f80 = f64, f34, f80
  4455. FNMA f81 = f65, f34, f81
  4456. FNMA f82 = f66, f34, f82
  4457. FNMA f83 = f67, f34, f83
  4458. ;;
  4459. FNMA f88 = f64, f35, f88
  4460. FNMA f89 = f65, f35, f89
  4461. FNMA f90 = f66, f35, f90
  4462. FNMA f91 = f67, f35, f91
  4463. ;;
  4464. FNMA f96 = f64, f36, f96
  4465. FNMA f97 = f65, f36, f97
  4466. FNMA f98 = f66, f36, f98
  4467. FNMA f99 = f67, f36, f99
  4468. ;;
  4469. FNMA f104 = f64, f37, f104
  4470. FNMA f105 = f65, f37, f105
  4471. FNMA f106 = f66, f37, f106
  4472. FNMA f107 = f67, f37, f107
  4473. ;;
  4474. FNMA f112 = f64, f38, f112
  4475. FNMA f113 = f65, f38, f113
  4476. FNMA f114 = f66, f38, f114
  4477. FNMA f115 = f67, f38, f115
  4478. ;;
  4479. FNMA f120 = f64, f39, f120
  4480. FNMA f121 = f65, f39, f121
  4481. FNMA f122 = f66, f39, f122
  4482. FNMA f123 = f67, f39, f123
  4483. ;;
  4484. FMPY f72 = f72, f40
  4485. FMPY f73 = f73, f40
  4486. FMPY f74 = f74, f40
  4487. FMPY f75 = f75, f40
  4488. ;;
  4489. FNMA f80 = f72, f41, f80
  4490. FNMA f81 = f73, f41, f81
  4491. FNMA f82 = f74, f41, f82
  4492. FNMA f83 = f75, f41, f83
  4493. ;;
  4494. FNMA f88 = f72, f42, f88
  4495. FNMA f89 = f73, f42, f89
  4496. FNMA f90 = f74, f42, f90
  4497. FNMA f91 = f75, f42, f91
  4498. ;;
  4499. FNMA f96 = f72, f43, f96
  4500. FNMA f97 = f73, f43, f97
  4501. FNMA f98 = f74, f43, f98
  4502. FNMA f99 = f75, f43, f99
  4503. ;;
  4504. FNMA f104 = f72, f44, f104
  4505. FNMA f105 = f73, f44, f105
  4506. FNMA f106 = f74, f44, f106
  4507. FNMA f107 = f75, f44, f107
  4508. ;;
  4509. FNMA f112 = f72, f45, f112
  4510. FNMA f113 = f73, f45, f113
  4511. FNMA f114 = f74, f45, f114
  4512. FNMA f115 = f75, f45, f115
  4513. ;;
  4514. FNMA f120 = f72, f46, f120
  4515. FNMA f121 = f73, f46, f121
  4516. FNMA f122 = f74, f46, f122
  4517. FNMA f123 = f75, f46, f123
  4518. ;;
  4519. FMPY f80 = f80, f47
  4520. FMPY f81 = f81, f47
  4521. FMPY f82 = f82, f47
  4522. FMPY f83 = f83, f47
  4523. ;;
  4524. FNMA f88 = f80, f48, f88
  4525. FNMA f89 = f81, f48, f89
  4526. FNMA f90 = f82, f48, f90
  4527. FNMA f91 = f83, f48, f91
  4528. ;;
  4529. FNMA f96 = f80, f49, f96
  4530. FNMA f97 = f81, f49, f97
  4531. FNMA f98 = f82, f49, f98
  4532. FNMA f99 = f83, f49, f99
  4533. ;;
  4534. FNMA f104 = f80, f50, f104
  4535. FNMA f105 = f81, f50, f105
  4536. FNMA f106 = f82, f50, f106
  4537. FNMA f107 = f83, f50, f107
  4538. ;;
  4539. FNMA f112 = f80, f51, f112
  4540. FNMA f113 = f81, f51, f113
  4541. FNMA f114 = f82, f51, f114
  4542. FNMA f115 = f83, f51, f115
  4543. ;;
  4544. FNMA f120 = f80, f52, f120
  4545. FNMA f121 = f81, f52, f121
  4546. FNMA f122 = f82, f52, f122
  4547. FNMA f123 = f83, f52, f123
  4548. ;;
  4549. FMPY f88 = f88, f53
  4550. FMPY f89 = f89, f53
  4551. FMPY f90 = f90, f53
  4552. FMPY f91 = f91, f53
  4553. ;;
  4554. FNMA f96 = f88, f54, f96
  4555. FNMA f97 = f89, f54, f97
  4556. FNMA f98 = f90, f54, f98
  4557. FNMA f99 = f91, f54, f99
  4558. ;;
  4559. FNMA f104 = f88, f55, f104
  4560. FNMA f105 = f89, f55, f105
  4561. FNMA f106 = f90, f55, f106
  4562. FNMA f107 = f91, f55, f107
  4563. ;;
  4564. FNMA f112 = f88, f56, f112
  4565. FNMA f113 = f89, f56, f113
  4566. FNMA f114 = f90, f56, f114
  4567. FNMA f115 = f91, f56, f115
  4568. ;;
  4569. FNMA f120 = f88, f57, f120
  4570. FNMA f121 = f89, f57, f121
  4571. FNMA f122 = f90, f57, f122
  4572. FNMA f123 = f91, f57, f123
  4573. ;;
  4574. FMPY f96 = f96, f58
  4575. FMPY f97 = f97, f58
  4576. FMPY f98 = f98, f58
  4577. FMPY f99 = f99, f58
  4578. ;;
  4579. FNMA f104 = f96, f59, f104
  4580. FNMA f105 = f97, f59, f105
  4581. FNMA f106 = f98, f59, f106
  4582. FNMA f107 = f99, f59, f107
  4583. ;;
  4584. FNMA f112 = f96, f60, f112
  4585. FNMA f113 = f97, f60, f113
  4586. FNMA f114 = f98, f60, f114
  4587. FNMA f115 = f99, f60, f115
  4588. ;;
  4589. FNMA f120 = f96, f61, f120
  4590. FNMA f121 = f97, f61, f121
  4591. FNMA f122 = f98, f61, f122
  4592. FNMA f123 = f99, f61, f123
  4593. ;;
  4594. { .mfi
  4595. STFD [AOFFSET] = f64, SIZE
  4596. FMPY f104 = f104, f16
  4597. }
  4598. { .mfi
  4599. STFD [AOFFSET2] = f72, SIZE
  4600. FMPY f105 = f105, f16
  4601. }
  4602. ;;
  4603. { .mfi
  4604. STFD [AOFFSET] = f65, SIZE
  4605. FMPY f106 = f106, f16
  4606. }
  4607. { .mfi
  4608. STFD [AOFFSET2] = f73, SIZE
  4609. FMPY f107 = f107, f16
  4610. }
  4611. ;;
  4612. { .mfi
  4613. STFD [AOFFSET] = f66, SIZE
  4614. FNMA f112 = f104, f17, f112
  4615. }
  4616. { .mfi
  4617. STFD [AOFFSET2] = f74, SIZE
  4618. FNMA f113 = f105, f17, f113
  4619. }
  4620. ;;
  4621. { .mfi
  4622. STFD [AOFFSET] = f67, 5 * SIZE
  4623. FNMA f114 = f106, f17, f114
  4624. }
  4625. { .mfi
  4626. STFD [AOFFSET2] = f75, 5 * SIZE
  4627. FNMA f115 = f107, f17, f115
  4628. }
  4629. ;;
  4630. { .mfi
  4631. STFD [AOFFSET] = f80, SIZE
  4632. FNMA f120 = f104, f18, f120
  4633. }
  4634. { .mfi
  4635. STFD [AOFFSET2] = f88, SIZE
  4636. FNMA f121 = f105, f18, f121
  4637. }
  4638. ;;
  4639. { .mfi
  4640. STFD [AOFFSET] = f81, SIZE
  4641. FNMA f122 = f106, f18, f122
  4642. }
  4643. { .mfi
  4644. STFD [AOFFSET2] = f89, SIZE
  4645. FNMA f123 = f107, f18, f123
  4646. }
  4647. ;;
  4648. { .mfi
  4649. STFD [AOFFSET] = f82, SIZE
  4650. FMPY f112 = f112, f19
  4651. }
  4652. { .mfi
  4653. STFD [AOFFSET2] = f90, SIZE
  4654. FMPY f113 = f113, f19
  4655. }
  4656. ;;
  4657. { .mfi
  4658. STFD [AOFFSET] = f83, 5 * SIZE
  4659. FMPY f114 = f114, f19
  4660. }
  4661. { .mfi
  4662. STFD [AOFFSET2] = f91, 5 * SIZE
  4663. FMPY f115 = f115, f19
  4664. }
  4665. ;;
  4666. { .mfi
  4667. STFD [AOFFSET] = f96, SIZE
  4668. FNMA f120 = f112, f20, f120
  4669. }
  4670. { .mfi
  4671. STFD [AOFFSET2] = f104, SIZE
  4672. FNMA f121 = f113, f20, f121
  4673. }
  4674. ;;
  4675. { .mfi
  4676. STFD [AOFFSET] = f97, SIZE
  4677. FNMA f122 = f114, f20, f122
  4678. }
  4679. { .mfi
  4680. STFD [AOFFSET2] = f105, SIZE
  4681. FNMA f123 = f115, f20, f123
  4682. }
  4683. ;;
  4684. { .mfi
  4685. STFD [AOFFSET] = f98, SIZE
  4686. FMPY f120 = f120, f21
  4687. }
  4688. { .mfi
  4689. STFD [AOFFSET2] = f106, SIZE
  4690. FMPY f121 = f121, f21
  4691. }
  4692. ;;
  4693. { .mfi
  4694. STFD [AOFFSET] = f99, 5 * SIZE
  4695. FMPY f122 = f122, f21
  4696. }
  4697. { .mfi
  4698. STFD [AOFFSET2] = f107, 5 * SIZE
  4699. FMPY f123 = f123, f21
  4700. }
  4701. ;;
  4702. { .mmf
  4703. STFD [AOFFSET] = f112, SIZE
  4704. STFD [AOFFSET2] = f120, SIZE
  4705. }
  4706. ;;
  4707. { .mmf
  4708. STFD [AOFFSET] = f113, SIZE
  4709. STFD [AOFFSET2] = f121, SIZE
  4710. }
  4711. ;;
  4712. { .mmf
  4713. STFD [AOFFSET] = f114, SIZE
  4714. STFD [AOFFSET2] = f122, SIZE
  4715. }
  4716. ;;
  4717. { .mmf
  4718. STFD [AOFFSET] = f115, -27 * SIZE
  4719. STFD [AOFFSET2] = f123, - 27 * SIZE
  4720. }
  4721. ;;
  4722. #endif
  4723. { .mmf
  4724. STFD [C1 ] = f64, SIZE
  4725. STFD [C2 ] = f72, SIZE
  4726. mov f64 = f0
  4727. }
  4728. ;;
  4729. { .mmi
  4730. STFD [C1 ] = f65, SIZE
  4731. STFD [C2 ] = f73, SIZE
  4732. nop __LINE__
  4733. }
  4734. ;;
  4735. { .mmf
  4736. STFD [C1 ] = f66, SIZE
  4737. STFD [C2 ] = f74, SIZE
  4738. mov f72 = f0
  4739. }
  4740. ;;
  4741. { .mmi
  4742. STFD [C1 ] = f67, SIZE
  4743. STFD [C2 ] = f75, SIZE
  4744. sub L = K, KK
  4745. }
  4746. ;;
  4747. { .mmf
  4748. STFD [C3 ] = f80, SIZE
  4749. STFD [C4 ] = f88, SIZE
  4750. mov f80 = f0
  4751. }
  4752. ;;
  4753. { .mmi
  4754. STFD [C3 ] = f81, SIZE
  4755. STFD [C4 ] = f89, SIZE
  4756. shladd L = L, BASE_SHIFT, r0
  4757. }
  4758. ;;
  4759. { .mmf
  4760. STFD [C3 ] = f82, SIZE
  4761. STFD [C4 ] = f90, SIZE
  4762. mov f88 = f0
  4763. }
  4764. ;;
  4765. { .mmi
  4766. STFD [C3 ] = f83, SIZE
  4767. STFD [C4 ] = f91, SIZE
  4768. shladd AOFFSET = L, 2, AOFFSET
  4769. }
  4770. ;;
  4771. { .mmf
  4772. STFD [C5 ] = f96, SIZE
  4773. STFD [C6 ] = f104, SIZE
  4774. mov f96 = f0
  4775. }
  4776. ;;
  4777. { .mmi
  4778. STFD [C5 ] = f97, SIZE
  4779. STFD [C6 ] = f105, SIZE
  4780. shladd BOFFSET = L, 3, BOFFSET
  4781. }
  4782. ;;
  4783. { .mmf
  4784. STFD [C5 ] = f98, SIZE
  4785. STFD [C6 ] = f106, SIZE
  4786. mov f104 = f0
  4787. }
  4788. ;;
  4789. { .mmi
  4790. STFD [C5 ] = f99, SIZE
  4791. STFD [C6 ] = f107, SIZE
  4792. #ifdef LT
  4793. adds KK = 4, KK
  4794. #else
  4795. nop __LINE__
  4796. #endif
  4797. }
  4798. ;;
  4799. { .mmf
  4800. STFD [C7 ] = f112, SIZE
  4801. STFD [C8 ] = f120, SIZE
  4802. mov f112 = f0
  4803. }
  4804. ;;
  4805. { .mmi
  4806. STFD [C7 ] = f113, SIZE
  4807. STFD [C8 ] = f121, SIZE
  4808. mov L = KK
  4809. }
  4810. ;;
  4811. { .mmf
  4812. STFD [C7 ] = f114, SIZE
  4813. STFD [C8 ] = f122, SIZE
  4814. mov f120 = f0
  4815. }
  4816. ;;
  4817. { .mmi
  4818. STFD [C7 ] = f115, SIZE
  4819. STFD [C8 ] = f123, SIZE
  4820. nop __LINE__
  4821. }
  4822. ;;
  4823. .align 8
  4824. .L030:
  4825. { .mib
  4826. mov L = KK
  4827. tbit.z p6, p0 = M, 1
  4828. (p6) br.cond.dptk .L040
  4829. }
  4830. ;;
  4831. { .mmi
  4832. cmp.ne p7, p0 = r0, L
  4833. adds BOFFSET = 0 * SIZE, B
  4834. nop __LINE__
  4835. }
  4836. ;;
  4837. { .mmi
  4838. (p7) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  4839. setf.d f73 = r0
  4840. adds L = 1, L
  4841. }
  4842. ;;
  4843. { .mfi
  4844. setf.d f105 = r0
  4845. mov f81 = f0
  4846. tbit.z p12, p0 = L, 0
  4847. }
  4848. { .mfi
  4849. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  4850. mov f89 = f0
  4851. nop __LINE__
  4852. }
  4853. ;;
  4854. { .mmf
  4855. (p7) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  4856. nop __LINE__
  4857. mov f65 = f0
  4858. }
  4859. ;;
  4860. { .mfi
  4861. (p7) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  4862. mov f113 = f0
  4863. cmp.eq p3, p0 = r0, r0
  4864. }
  4865. { .mfi
  4866. setf.d f97 = r0
  4867. mov f121 = f0
  4868. shr L = L, 1
  4869. }
  4870. ;;
  4871. { .mmi
  4872. (p7) LDFPD f52, f53 = [BOFFSET], 2 * SIZE
  4873. cmp.eq p6, p0 = 0, L
  4874. adds L = -1, L
  4875. }
  4876. ;;
  4877. { .mib
  4878. (p7) LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  4879. mov ar.lc = L
  4880. (p6) br.cond.dpnt .L038
  4881. }
  4882. ;;
  4883. .L032:
  4884. { .mfb
  4885. lfetch.nt1 [PREA], 4 * SIZE
  4886. FMA f64 = f32, f48, f64 // A1 * B1
  4887. nop __LINE__
  4888. }
  4889. { .mfi
  4890. nop __LINE__
  4891. FMA f72 = f32, f49, f72 // A1 * B2
  4892. (p12) cmp.ne p3, p0 = 0, L
  4893. }
  4894. ;;
  4895. { .mfi
  4896. lfetch.nt1 [PREB], 16 * SIZE
  4897. FMA f80 = f32, f50, f80 // A1 * B3
  4898. cmp.ne p4, p5 = 0, L
  4899. }
  4900. { .mfb
  4901. nop __LINE__
  4902. FMA f88 = f32, f51, f88 // A1 * B4
  4903. nop __LINE__
  4904. }
  4905. ;;
  4906. { .mfb
  4907. (p3) LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  4908. FMA f96 = f32, f52, f96 // A1 * B5
  4909. nop __LINE__
  4910. }
  4911. { .mfb
  4912. nop __LINE__
  4913. FMA f104 = f32, f53, f104 // A1 * B6
  4914. nop __LINE__
  4915. }
  4916. ;;
  4917. { .mfb
  4918. (p3) LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  4919. FMA f112 = f32, f54, f112 // A1 * B7
  4920. nop __LINE__
  4921. }
  4922. { .mfb
  4923. nop __LINE__
  4924. FMA f120 = f32, f55, f120 // A1 * B8
  4925. nop __LINE__
  4926. }
  4927. ;;
  4928. { .mfb
  4929. (p3) LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  4930. FMA f65 = f33, f48, f65 // A2 * B1
  4931. nop __LINE__
  4932. }
  4933. { .mfb
  4934. nop __LINE__
  4935. FMA f73 = f33, f49, f73 // A2 * B2
  4936. nop __LINE__
  4937. }
  4938. ;;
  4939. { .mfb
  4940. (p3) LDFPD f60, f61 = [BOFFSET], 2 * SIZE
  4941. FMA f81 = f33, f50, f81 // A2 * B3
  4942. nop __LINE__
  4943. }
  4944. { .mfb
  4945. nop __LINE__
  4946. FMA f89 = f33, f51, f89 // A2 * B4
  4947. nop __LINE__
  4948. }
  4949. ;;
  4950. { .mfb
  4951. (p3) LDFPD f62, f63 = [BOFFSET], 2 * SIZE
  4952. FMA f97 = f33, f52, f97 // A2 * B5
  4953. nop __LINE__
  4954. }
  4955. { .mfb
  4956. nop __LINE__
  4957. FMA f105 = f33, f53, f105 // A2 * B6
  4958. nop __LINE__
  4959. }
  4960. ;;
  4961. { .mfb
  4962. nop __LINE__
  4963. FMA f113 = f33, f54, f113 // A2 * B7
  4964. nop __LINE__
  4965. }
  4966. { .mfb
  4967. nop __LINE__
  4968. FMA f121 = f33, f55, f121 // A2 * B8
  4969. nop __LINE__
  4970. }
  4971. ;;
  4972. { .mfb
  4973. (p4) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  4974. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  4975. nop __LINE__
  4976. }
  4977. { .mfb
  4978. (p4) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  4979. (p3) FMA f72 = f40, f57, f72 // A1 * B2
  4980. nop __LINE__
  4981. }
  4982. ;;
  4983. { .mfb
  4984. (p4) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  4985. (p3) FMA f80 = f40, f58, f80 // A1 * B3
  4986. nop __LINE__
  4987. }
  4988. { .mfb
  4989. nop __LINE__
  4990. (p3) FMA f88 = f40, f59, f88 // A1 * B4
  4991. nop __LINE__
  4992. }
  4993. ;;
  4994. { .mfb
  4995. (p4) LDFPD f52, f53 = [BOFFSET], 2 * SIZE
  4996. (p3) FMA f96 = f40, f60, f96 // A1 * B5
  4997. nop __LINE__
  4998. }
  4999. { .mfb
  5000. nop __LINE__
  5001. (p3) FMA f104 = f40, f61, f104 // A1 * B6
  5002. nop __LINE__
  5003. }
  5004. ;;
  5005. { .mfb
  5006. nop __LINE__
  5007. (p3) FMA f112 = f40, f62, f112 // A1 * B7
  5008. nop __LINE__
  5009. }
  5010. { .mfb
  5011. nop __LINE__
  5012. (p3) FMA f120 = f40, f63, f120 // A1 * B8
  5013. nop __LINE__
  5014. }
  5015. ;;
  5016. { .mfb
  5017. (p4) LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  5018. (p3) FMA f65 = f41, f56, f65 // A2 * B1
  5019. nop __LINE__
  5020. }
  5021. { .mfb
  5022. (p3) FMA f73 = f41, f57, f73 // A2 * B2
  5023. nop __LINE__
  5024. }
  5025. { .mfb
  5026. nop __LINE__
  5027. (p3) FMA f81 = f41, f58, f81 // A2 * B3
  5028. nop __LINE__
  5029. }
  5030. { .mfb
  5031. nop __LINE__
  5032. (p3) FMA f89 = f41, f59, f89 // A2 * B4
  5033. nop __LINE__
  5034. }
  5035. ;;
  5036. { .mfb
  5037. nop __LINE__
  5038. (p3) FMA f97 = f41, f60, f97 // A2 * B5
  5039. nop __LINE__
  5040. }
  5041. { .mfb
  5042. nop __LINE__
  5043. (p3) FMA f105 = f41, f61, f105 // A2 * B6
  5044. nop __LINE__
  5045. }
  5046. ;;
  5047. { .mfi
  5048. nop __LINE__
  5049. (p3) FMA f113 = f41, f62, f113 // A2 * B7
  5050. adds L = -1, L
  5051. }
  5052. { .mfb
  5053. nop __LINE__
  5054. (p3) FMA f121 = f41, f63, f121 // A2 * B8
  5055. br.cloop.sptk.few .L032
  5056. }
  5057. ;;
  5058. .L038:
  5059. #ifdef LT
  5060. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  5061. ;;
  5062. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  5063. ;;
  5064. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  5065. ;;
  5066. LDFPD f38, f39 = [BOFFSET], 2 * SIZE
  5067. ;;
  5068. LDFPD f40, f41 = [BOFFSET], 2 * SIZE
  5069. ;;
  5070. LDFPD f42, f43 = [BOFFSET], 2 * SIZE
  5071. ;;
  5072. LDFPD f44, f45 = [BOFFSET], 2 * SIZE
  5073. ;;
  5074. LDFPD f46, f47 = [BOFFSET]
  5075. adds BOFFSET = -14 * SIZE, BOFFSET
  5076. ;;
  5077. { .mfi
  5078. FSUB f64 = f32, f64
  5079. nop __LINE__
  5080. }
  5081. { .mfi
  5082. nop __LINE__
  5083. FSUB f72 = f33, f72
  5084. nop __LINE__
  5085. }
  5086. ;;
  5087. { .mfi
  5088. FSUB f80 = f34, f80
  5089. nop __LINE__
  5090. }
  5091. { .mfi
  5092. nop __LINE__
  5093. FSUB f88 = f35, f88
  5094. nop __LINE__
  5095. }
  5096. ;;
  5097. { .mfi
  5098. FSUB f96 = f36, f96
  5099. nop __LINE__
  5100. }
  5101. { .mfi
  5102. nop __LINE__
  5103. FSUB f104 = f37, f104
  5104. nop __LINE__
  5105. }
  5106. ;;
  5107. { .mfi
  5108. FSUB f112 = f38, f112
  5109. nop __LINE__
  5110. }
  5111. { .mfi
  5112. nop __LINE__
  5113. FSUB f120 = f39, f120
  5114. nop __LINE__
  5115. }
  5116. ;;
  5117. { .mfi
  5118. FSUB f65 = f40, f65
  5119. nop __LINE__
  5120. }
  5121. { .mfi
  5122. nop __LINE__
  5123. FSUB f73 = f41, f73
  5124. nop __LINE__
  5125. }
  5126. ;;
  5127. { .mfi
  5128. FSUB f81 = f42, f81
  5129. nop __LINE__
  5130. }
  5131. { .mfi
  5132. nop __LINE__
  5133. FSUB f89 = f43, f89
  5134. nop __LINE__
  5135. }
  5136. ;;
  5137. { .mfi
  5138. FSUB f97 = f44, f97
  5139. nop __LINE__
  5140. }
  5141. { .mfi
  5142. nop __LINE__
  5143. FSUB f105 = f45, f105
  5144. nop __LINE__
  5145. }
  5146. ;;
  5147. { .mfi
  5148. FSUB f113 = f46, f113
  5149. }
  5150. { .mfi
  5151. nop __LINE__
  5152. FSUB f121 = f47, f121
  5153. nop __LINE__
  5154. }
  5155. ;;
  5156. { .mmi
  5157. LDFPD f32, f33 = [AOFFSET]
  5158. nop __LINE__
  5159. adds AOFFSET = 3 * SIZE, AOFFSET
  5160. }
  5161. ;;
  5162. { .mfi
  5163. LDFD f34 = [AOFFSET], - 3 * SIZE
  5164. FMPY f64 = f64, f32
  5165. nop __LINE__
  5166. }
  5167. { .mfi
  5168. nop __LINE__
  5169. FMPY f96 = f96, f32
  5170. nop __LINE__
  5171. }
  5172. ;;
  5173. { .mfi
  5174. FMPY f72 = f72, f32
  5175. nop __LINE__
  5176. }
  5177. { .mfi
  5178. nop __LINE__
  5179. FMPY f104 = f104, f32
  5180. nop __LINE__
  5181. }
  5182. ;;
  5183. { .mfi
  5184. FMPY f80 = f80, f32
  5185. }
  5186. { .mfi
  5187. nop __LINE__
  5188. FMPY f112 = f112, f32
  5189. nop __LINE__
  5190. }
  5191. ;;
  5192. { .mfi
  5193. FMPY f88 = f88, f32
  5194. nop __LINE__
  5195. }
  5196. { .mfi
  5197. nop __LINE__
  5198. FMPY f120 = f120, f32
  5199. nop __LINE__
  5200. }
  5201. ;;
  5202. { .mfi
  5203. FNMA f65 = f64, f33, f65
  5204. nop __LINE__
  5205. }
  5206. { .mfi
  5207. nop __LINE__
  5208. FNMA f97 = f96, f33, f97
  5209. nop __LINE__
  5210. }
  5211. ;;
  5212. { .mfi
  5213. FNMA f73 = f72, f33, f73
  5214. nop __LINE__
  5215. }
  5216. { .mfi
  5217. nop __LINE__
  5218. FNMA f105 = f104, f33, f105
  5219. nop __LINE__
  5220. }
  5221. ;;
  5222. { .mfi
  5223. FNMA f81 = f80, f33, f81
  5224. }
  5225. { .mfi
  5226. nop __LINE__
  5227. FNMA f113 = f112, f33, f113
  5228. nop __LINE__
  5229. }
  5230. ;;
  5231. { .mfi
  5232. FNMA f89 = f88, f33, f89
  5233. nop __LINE__
  5234. }
  5235. { .mfi
  5236. nop __LINE__
  5237. FNMA f121 = f120, f33, f121
  5238. adds BOFFSET2 = 4 * SIZE, BOFFSET
  5239. }
  5240. ;;
  5241. { .mfi
  5242. STFD [BOFFSET] = f64, SIZE
  5243. FMPY f65 = f65, f34
  5244. }
  5245. { .mfi
  5246. STFD [BOFFSET2] = f96, SIZE
  5247. FMPY f97 = f97, f34
  5248. }
  5249. ;;
  5250. { .mfi
  5251. STFD [BOFFSET] = f72, SIZE
  5252. FMPY f73 = f73, f34
  5253. }
  5254. { .mfi
  5255. STFD [BOFFSET2] = f104, SIZE
  5256. FMPY f105 = f105, f34
  5257. }
  5258. ;;
  5259. { .mfi
  5260. STFD [BOFFSET] = f80, SIZE
  5261. FMPY f81 = f81, f34
  5262. sub L = K, KK
  5263. }
  5264. { .mfi
  5265. STFD [BOFFSET2] = f112, SIZE
  5266. FMPY f113 = f113, f34
  5267. }
  5268. ;;
  5269. { .mfi
  5270. STFD [BOFFSET] = f88, 5 * SIZE
  5271. FMPY f89 = f89, f34
  5272. shladd L = L, BASE_SHIFT, r0
  5273. }
  5274. { .mfi
  5275. STFD [BOFFSET2] = f120, 5 * SIZE
  5276. FMPY f121 = f121, f34
  5277. }
  5278. ;;
  5279. { .mmi
  5280. STFD [BOFFSET] = f65, SIZE
  5281. STFD [BOFFSET2] = f97, SIZE
  5282. }
  5283. ;;
  5284. { .mmi
  5285. STFD [BOFFSET] = f73, SIZE
  5286. STFD [BOFFSET2] = f105, SIZE
  5287. }
  5288. ;;
  5289. { .mmi
  5290. STFD [BOFFSET] = f81, SIZE
  5291. STFD [BOFFSET2] = f113, SIZE
  5292. }
  5293. ;;
  5294. { .mmi
  5295. STFD [BOFFSET] = f89, -11 * SIZE
  5296. STFD [BOFFSET2] = f121, -11 * SIZE
  5297. }
  5298. #endif
  5299. #ifdef RN
  5300. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  5301. ;;
  5302. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  5303. ;;
  5304. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  5305. ;;
  5306. LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  5307. ;;
  5308. LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  5309. ;;
  5310. LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  5311. ;;
  5312. LDFPD f44, f45 = [AOFFSET], 2 * SIZE
  5313. ;;
  5314. LDFPD f46, f47 = [AOFFSET]
  5315. adds AOFFSET = -14 * SIZE, AOFFSET
  5316. ;;
  5317. FSUB f64 = f32, f64
  5318. FSUB f65 = f33, f65
  5319. FSUB f72 = f34, f72
  5320. FSUB f73 = f35, f73
  5321. FSUB f80 = f36, f80
  5322. FSUB f81 = f37, f81
  5323. FSUB f88 = f38, f88
  5324. FSUB f89 = f39, f89
  5325. FSUB f96 = f40, f96
  5326. FSUB f97 = f41, f97
  5327. FSUB f104 = f42, f104
  5328. FSUB f105 = f43, f105
  5329. FSUB f112 = f44, f112
  5330. FSUB f113 = f45, f113
  5331. FSUB f120 = f46, f120
  5332. FSUB f121 = f47, f121
  5333. ;;
  5334. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  5335. ;;
  5336. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  5337. ;;
  5338. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  5339. ;;
  5340. LDFPD f38, f39 = [BOFFSET]
  5341. adds BOFFSET = 3 * SIZE, BOFFSET
  5342. ;;
  5343. LDFD f40 = [BOFFSET], 1 * SIZE
  5344. ;;
  5345. LDFPD f41, f42 = [BOFFSET], 2 * SIZE
  5346. ;;
  5347. LDFPD f43, f44 = [BOFFSET], 2 * SIZE
  5348. ;;
  5349. LDFPD f45, f46 = [BOFFSET]
  5350. adds BOFFSET = 4 * SIZE, BOFFSET
  5351. ;;
  5352. LDFPD f47, f48 = [BOFFSET], 2 * SIZE
  5353. ;;
  5354. LDFPD f49, f50 = [BOFFSET], 2 * SIZE
  5355. ;;
  5356. LDFPD f51, f52 = [BOFFSET]
  5357. adds BOFFSET = 5 * SIZE, BOFFSET
  5358. ;;
  5359. LDFD f53 = [BOFFSET], 1 * SIZE
  5360. ;;
  5361. LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  5362. ;;
  5363. LDFPD f56, f57 = [BOFFSET]
  5364. adds BOFFSET = 6 * SIZE, BOFFSET
  5365. ;;
  5366. LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  5367. ;;
  5368. LDFPD f60, f61 = [BOFFSET]
  5369. adds BOFFSET = 7 * SIZE, BOFFSET
  5370. ;;
  5371. LDFD f16 = [BOFFSET], 1 * SIZE
  5372. ;;
  5373. LDFPD f17, f18 = [BOFFSET]
  5374. adds BOFFSET = 8 * SIZE, BOFFSET
  5375. ;;
  5376. LDFPD f19, f20 = [BOFFSET]
  5377. adds BOFFSET = 9 * SIZE, BOFFSET
  5378. ;;
  5379. LDFD f21 = [BOFFSET]
  5380. adds BOFFSET = -63 * SIZE, BOFFSET
  5381. adds AOFFSET2 = 4 * SIZE, AOFFSET
  5382. ;;
  5383. FMPY f64 = f64, f32
  5384. FMPY f65 = f65, f32
  5385. ;;
  5386. FNMA f72 = f64, f33, f72
  5387. FNMA f73 = f65, f33, f73
  5388. FNMA f80 = f64, f34, f80
  5389. FNMA f81 = f65, f34, f81
  5390. ;;
  5391. FNMA f88 = f64, f35, f88
  5392. FNMA f89 = f65, f35, f89
  5393. FNMA f96 = f64, f36, f96
  5394. FNMA f97 = f65, f36, f97
  5395. FMPY f72 = f72, f40
  5396. FMPY f73 = f73, f40
  5397. FNMA f104 = f64, f37, f104
  5398. FNMA f105 = f65, f37, f105
  5399. FNMA f112 = f64, f38, f112
  5400. FNMA f113 = f65, f38, f113
  5401. FNMA f120 = f64, f39, f120
  5402. FNMA f121 = f65, f39, f121
  5403. ;;
  5404. FNMA f80 = f72, f41, f80
  5405. FNMA f81 = f73, f41, f81
  5406. FNMA f88 = f72, f42, f88
  5407. FNMA f89 = f73, f42, f89
  5408. ;;
  5409. FNMA f96 = f72, f43, f96
  5410. FNMA f97 = f73, f43, f97
  5411. FNMA f104 = f72, f44, f104
  5412. FNMA f105 = f73, f44, f105
  5413. FMPY f80 = f80, f47
  5414. FMPY f81 = f81, f47
  5415. FNMA f112 = f72, f45, f112
  5416. FNMA f113 = f73, f45, f113
  5417. FNMA f120 = f72, f46, f120
  5418. FNMA f121 = f73, f46, f121
  5419. ;;
  5420. FNMA f88 = f80, f48, f88
  5421. FNMA f89 = f81, f48, f89
  5422. FNMA f96 = f80, f49, f96
  5423. FNMA f97 = f81, f49, f97
  5424. FNMA f104 = f80, f50, f104
  5425. FNMA f105 = f81, f50, f105
  5426. FNMA f112 = f80, f51, f112
  5427. FNMA f113 = f81, f51, f113
  5428. ;;
  5429. FMPY f88 = f88, f53
  5430. FMPY f89 = f89, f53
  5431. FNMA f120 = f80, f52, f120
  5432. FNMA f121 = f81, f52, f121
  5433. ;;
  5434. FNMA f96 = f88, f54, f96
  5435. FNMA f97 = f89, f54, f97
  5436. FNMA f104 = f88, f55, f104
  5437. FNMA f105 = f89, f55, f105
  5438. FNMA f112 = f88, f56, f112
  5439. FNMA f113 = f89, f56, f113
  5440. FNMA f120 = f88, f57, f120
  5441. FNMA f121 = f89, f57, f121
  5442. ;;
  5443. FMPY f96 = f96, f58
  5444. FMPY f97 = f97, f58
  5445. ;;
  5446. FNMA f104 = f96, f59, f104
  5447. FNMA f105 = f97, f59, f105
  5448. FNMA f112 = f96, f60, f112
  5449. FNMA f113 = f97, f60, f113
  5450. FNMA f120 = f96, f61, f120
  5451. FNMA f121 = f97, f61, f121
  5452. ;;
  5453. FMPY f104 = f104, f16
  5454. FMPY f105 = f105, f16
  5455. ;;
  5456. FNMA f112 = f104, f17, f112
  5457. FNMA f113 = f105, f17, f113
  5458. ;;
  5459. { .mfi
  5460. STFD [AOFFSET] = f64, SIZE
  5461. FNMA f120 = f104, f18, f120
  5462. }
  5463. { .mfi
  5464. STFD [AOFFSET2] = f80, SIZE
  5465. FNMA f121 = f105, f18, f121
  5466. }
  5467. ;;
  5468. { .mfi
  5469. STFD [AOFFSET] = f65, SIZE
  5470. FMPY f112 = f112, f19
  5471. }
  5472. { .mfi
  5473. STFD [AOFFSET2] = f81, SIZE
  5474. FMPY f113 = f113, f19
  5475. }
  5476. ;;
  5477. { .mfi
  5478. STFD [AOFFSET] = f72, SIZE
  5479. FNMA f120 = f112, f20, f120
  5480. sub L = K, KK
  5481. }
  5482. { .mfi
  5483. STFD [AOFFSET2] = f88, SIZE
  5484. FNMA f121 = f113, f20, f121
  5485. }
  5486. ;;
  5487. { .mfi
  5488. STFD [AOFFSET] = f73, 5 * SIZE
  5489. FMPY f120 = f120, f21
  5490. shladd L = L, BASE_SHIFT, r0
  5491. }
  5492. { .mfi
  5493. STFD [AOFFSET2] = f89, 5 * SIZE
  5494. FMPY f121 = f121, f21
  5495. }
  5496. ;;
  5497. { .mmi
  5498. STFD [AOFFSET] = f96, SIZE
  5499. STFD [AOFFSET2] = f112, SIZE
  5500. nop __LINE__
  5501. }
  5502. ;;
  5503. { .mmi
  5504. STFD [AOFFSET] = f97, SIZE
  5505. STFD [AOFFSET2] = f113, SIZE
  5506. nop __LINE__
  5507. }
  5508. ;;
  5509. { .mmi
  5510. STFD [AOFFSET] = f104, SIZE
  5511. STFD [AOFFSET2] = f120, SIZE
  5512. nop __LINE__
  5513. }
  5514. ;;
  5515. { .mmi
  5516. STFD [AOFFSET] = f105, -11 * SIZE
  5517. STFD [AOFFSET2] = f121, - 11 * SIZE
  5518. nop __LINE__
  5519. }
  5520. ;;
  5521. #endif
  5522. { .mmf
  5523. STFD [C1 ] = f64, SIZE
  5524. STFD [C2 ] = f72, SIZE
  5525. mov f72 = f0
  5526. }
  5527. ;;
  5528. { .mmf
  5529. STFD [C1 ] = f65, SIZE
  5530. STFD [C2 ] = f73, SIZE
  5531. mov f64 = f0
  5532. }
  5533. ;;
  5534. { .mmf
  5535. STFD [C3 ] = f80, SIZE
  5536. STFD [C4 ] = f88, SIZE
  5537. mov f88 = f0
  5538. }
  5539. ;;
  5540. { .mmf
  5541. STFD [C3 ] = f81, SIZE
  5542. STFD [C4 ] = f89, SIZE
  5543. mov f80 = f0
  5544. }
  5545. ;;
  5546. { .mmf
  5547. STFD [C5 ] = f96, SIZE
  5548. STFD [C6 ] = f104, SIZE
  5549. mov f96 = f0
  5550. }
  5551. ;;
  5552. { .mmf
  5553. STFD [C5 ] = f97, SIZE
  5554. STFD [C6 ] = f105, SIZE
  5555. mov f104 = f0
  5556. }
  5557. ;;
  5558. { .mmf
  5559. STFD [C7 ] = f112, SIZE
  5560. STFD [C8 ] = f120, SIZE
  5561. mov f112 = f0
  5562. }
  5563. ;;
  5564. { .mmf
  5565. STFD [C7 ] = f113, SIZE
  5566. STFD [C8 ] = f121, SIZE
  5567. mov f120 = f0
  5568. }
  5569. { .mmi
  5570. shladd AOFFSET = L, 1, AOFFSET
  5571. shladd BOFFSET = L, 3, BOFFSET
  5572. #ifdef LT
  5573. adds KK = 2, KK
  5574. #else
  5575. nop __LINE__
  5576. #endif
  5577. }
  5578. ;;
  5579. .align 8
  5580. .L040:
  5581. { .mib
  5582. mov L = KK
  5583. tbit.z p6, p0 = M, 0
  5584. (p6) br.cond.dptk .L049
  5585. }
  5586. ;;
  5587. { .mmi
  5588. cmp.ne p7, p0 = r0, L
  5589. adds BOFFSET = 0 * SIZE, B
  5590. }
  5591. ;;
  5592. { .mmf
  5593. (p7) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  5594. }
  5595. ;;
  5596. { .mmi
  5597. adds L = 1, L
  5598. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  5599. cmp.eq p3, p0 = r0, r0
  5600. }
  5601. ;;
  5602. { .mii
  5603. (p7) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  5604. tbit.z p12, p0 = L, 0
  5605. shr L = L, 1
  5606. }
  5607. ;;
  5608. { .mmi
  5609. (p7) LDFPD f52, f53 = [BOFFSET], 2 * SIZE
  5610. adds L = -1, L
  5611. }
  5612. ;;
  5613. { .mmi
  5614. (p7) LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  5615. cmp.eq p6, p0 = -1, L
  5616. }
  5617. ;;
  5618. { .mib
  5619. (p7) LDFD f32 = [AOFFSET], 1 * SIZE
  5620. mov ar.lc = L
  5621. (p6) br.cond.dpnt .L048
  5622. }
  5623. ;;
  5624. .L042:
  5625. { .mfb
  5626. lfetch.nt1 [PREB], 16 * SIZE
  5627. FMA f64 = f32, f48, f64 // A1 * B1
  5628. nop __LINE__
  5629. }
  5630. { .mfb
  5631. (p12) cmp.ne p3, p0 = 0, L
  5632. FMA f72 = f32, f49, f72 // A1 * B2
  5633. nop __LINE__
  5634. }
  5635. ;;
  5636. { .mfi
  5637. (p3) LDFD f40 = [AOFFSET], 1 * SIZE
  5638. FMA f80 = f32, f50, f80 // A1 * B3
  5639. cmp.ne p4, p5 = 0, L
  5640. }
  5641. { .mfb
  5642. (p3) LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  5643. FMA f88 = f32, f51, f88 // A1 * B4
  5644. nop __LINE__
  5645. }
  5646. ;;
  5647. { .mfb
  5648. (p3) LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  5649. FMA f96 = f32, f52, f96 // A1 * B5
  5650. nop __LINE__
  5651. }
  5652. { .mfb
  5653. nop __LINE__
  5654. FMA f104 = f32, f53, f104 // A1 * B6
  5655. nop __LINE__
  5656. }
  5657. ;;
  5658. { .mfb
  5659. (p3) LDFPD f60, f61 = [BOFFSET], 2 * SIZE
  5660. FMA f112 = f32, f54, f112 // A1 * B7
  5661. nop __LINE__
  5662. }
  5663. { .mfb
  5664. nop __LINE__
  5665. FMA f120 = f32, f55, f120 // A1 * B8
  5666. nop __LINE__
  5667. }
  5668. ;;
  5669. { .mfb
  5670. (p4) LDFD f32 = [AOFFSET], 1 * SIZE
  5671. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  5672. nop __LINE__
  5673. }
  5674. { .mfb
  5675. (p3) LDFPD f62, f63 = [BOFFSET], 2 * SIZE
  5676. (p3) FMA f72 = f40, f57, f72 // A1 * B2
  5677. nop __LINE__
  5678. }
  5679. ;;
  5680. { .mfb
  5681. (p4) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  5682. (p3) FMA f80 = f40, f58, f80 // A1 * B3
  5683. nop __LINE__
  5684. }
  5685. { .mfb
  5686. nop __LINE__
  5687. (p3) FMA f88 = f40, f59, f88 // A1 * B4
  5688. nop __LINE__
  5689. }
  5690. ;;
  5691. { .mfb
  5692. (p4) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  5693. (p3) FMA f96 = f40, f60, f96 // A1 * B5
  5694. nop __LINE__
  5695. }
  5696. { .mfb
  5697. nop __LINE__
  5698. (p3) FMA f104 = f40, f61, f104 // A1 * B6
  5699. nop __LINE__
  5700. }
  5701. ;;
  5702. { .mfi
  5703. (p4) LDFPD f52, f53 = [BOFFSET], 2 * SIZE
  5704. (p3) FMA f112 = f40, f62, f112 // A1 * B7
  5705. adds L = -1, L
  5706. }
  5707. { .mmb
  5708. nop __LINE__
  5709. nop __LINE__
  5710. nop __LINE__
  5711. }
  5712. ;;
  5713. { .mfb
  5714. (p4) LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  5715. (p3) FMA f120 = f40, f63, f120 // A1 * B8
  5716. nop __LINE__
  5717. }
  5718. { .mmb
  5719. nop __LINE__
  5720. nop __LINE__
  5721. br.cloop.sptk.few .L042
  5722. }
  5723. ;;
  5724. .L048:
  5725. adds AOFFSET2 = 4 * SIZE, AOFFSET
  5726. adds BOFFSET2 = 4 * SIZE, BOFFSET
  5727. ;;
  5728. #ifdef LT
  5729. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  5730. ;;
  5731. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  5732. ;;
  5733. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  5734. ;;
  5735. LDFPD f38, f39 = [BOFFSET]
  5736. adds BOFFSET = -6 * SIZE, BOFFSET
  5737. ;;
  5738. { .mfi
  5739. FSUB f64 = f32, f64
  5740. nop __LINE__
  5741. }
  5742. { .mfi
  5743. nop __LINE__
  5744. FSUB f72 = f33, f72
  5745. nop __LINE__
  5746. }
  5747. ;;
  5748. { .mfi
  5749. FSUB f80 = f34, f80
  5750. nop __LINE__
  5751. }
  5752. { .mfi
  5753. nop __LINE__
  5754. FSUB f88 = f35, f88
  5755. nop __LINE__
  5756. }
  5757. ;;
  5758. { .mfi
  5759. FSUB f96 = f36, f96
  5760. nop __LINE__
  5761. }
  5762. { .mfi
  5763. nop __LINE__
  5764. FSUB f104 = f37, f104
  5765. nop __LINE__
  5766. }
  5767. ;;
  5768. { .mfi
  5769. FSUB f112 = f38, f112
  5770. nop __LINE__
  5771. }
  5772. { .mfi
  5773. nop __LINE__
  5774. FSUB f120 = f39, f120
  5775. nop __LINE__
  5776. }
  5777. ;;
  5778. #else
  5779. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  5780. ;;
  5781. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  5782. ;;
  5783. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  5784. ;;
  5785. LDFPD f38, f39 = [AOFFSET]
  5786. adds AOFFSET = -6 * SIZE, AOFFSET
  5787. ;;
  5788. FSUB f64 = f32, f64
  5789. FSUB f72 = f33, f72
  5790. FSUB f80 = f34, f80
  5791. FSUB f88 = f35, f88
  5792. FSUB f96 = f36, f96
  5793. FSUB f104 = f37, f104
  5794. FSUB f112 = f38, f112
  5795. FSUB f120 = f39, f120
  5796. ;;
  5797. #endif
  5798. #ifdef LT
  5799. LDFD f32 = [AOFFSET]
  5800. ;;
  5801. { .mfi
  5802. FMPY f64 = f64, f32
  5803. nop __LINE__
  5804. }
  5805. { .mfi
  5806. nop __LINE__
  5807. FMPY f96 = f96, f32
  5808. nop __LINE__
  5809. }
  5810. ;;
  5811. { .mfi
  5812. FMPY f72 = f72, f32
  5813. nop __LINE__
  5814. }
  5815. { .mfi
  5816. nop __LINE__
  5817. FMPY f104 = f104, f32
  5818. nop __LINE__
  5819. }
  5820. ;;
  5821. { .mfi
  5822. FMPY f80 = f80, f32
  5823. }
  5824. { .mfi
  5825. nop __LINE__
  5826. FMPY f112 = f112, f32
  5827. nop __LINE__
  5828. }
  5829. ;;
  5830. { .mfi
  5831. FMPY f88 = f88, f32
  5832. nop __LINE__
  5833. }
  5834. { .mfi
  5835. nop __LINE__
  5836. FMPY f120 = f120, f32
  5837. nop __LINE__
  5838. }
  5839. ;;
  5840. { .mfi
  5841. STFD [BOFFSET] = f64, SIZE
  5842. }
  5843. { .mfi
  5844. STFD [BOFFSET2] = f96, SIZE
  5845. }
  5846. ;;
  5847. { .mfi
  5848. STFD [BOFFSET] = f72, SIZE
  5849. }
  5850. { .mfi
  5851. STFD [BOFFSET2] = f104, SIZE
  5852. }
  5853. ;;
  5854. { .mfi
  5855. STFD [BOFFSET] = f80, SIZE
  5856. }
  5857. { .mfi
  5858. STFD [BOFFSET2] = f112, SIZE
  5859. }
  5860. ;;
  5861. { .mfi
  5862. STFD [BOFFSET] = f88, -3 * SIZE
  5863. }
  5864. { .mfi
  5865. STFD [BOFFSET2] = f120, -3 * SIZE
  5866. }
  5867. ;;
  5868. #endif
  5869. #ifdef RN
  5870. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  5871. ;;
  5872. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  5873. ;;
  5874. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  5875. ;;
  5876. LDFPD f38, f39 = [BOFFSET]
  5877. adds BOFFSET = 3 * SIZE, BOFFSET
  5878. ;;
  5879. LDFD f40 = [BOFFSET], 1 * SIZE
  5880. ;;
  5881. LDFPD f41, f42 = [BOFFSET], 2 * SIZE
  5882. ;;
  5883. LDFPD f43, f44 = [BOFFSET], 2 * SIZE
  5884. ;;
  5885. LDFPD f45, f46 = [BOFFSET]
  5886. adds BOFFSET = 4 * SIZE, BOFFSET
  5887. ;;
  5888. LDFPD f47, f48 = [BOFFSET], 2 * SIZE
  5889. ;;
  5890. LDFPD f49, f50 = [BOFFSET], 2 * SIZE
  5891. ;;
  5892. LDFPD f51, f52 = [BOFFSET]
  5893. adds BOFFSET = 5 * SIZE, BOFFSET
  5894. ;;
  5895. LDFD f53 = [BOFFSET], 1 * SIZE
  5896. ;;
  5897. LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  5898. ;;
  5899. LDFPD f56, f57 = [BOFFSET]
  5900. adds BOFFSET = 6 * SIZE, BOFFSET
  5901. ;;
  5902. LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  5903. ;;
  5904. LDFPD f60, f61 = [BOFFSET]
  5905. adds BOFFSET = 7 * SIZE, BOFFSET
  5906. ;;
  5907. LDFD f16 = [BOFFSET], 1 * SIZE
  5908. ;;
  5909. LDFPD f17, f18 = [BOFFSET]
  5910. adds BOFFSET = 8 * SIZE, BOFFSET
  5911. ;;
  5912. LDFPD f19, f20 = [BOFFSET]
  5913. adds BOFFSET = 9 * SIZE, BOFFSET
  5914. ;;
  5915. LDFD f21 = [BOFFSET]
  5916. adds BOFFSET = -63 * SIZE, BOFFSET
  5917. ;;
  5918. FMPY f64 = f64, f32
  5919. ;;
  5920. FNMA f72 = f64, f33, f72
  5921. ;;
  5922. FNMA f80 = f64, f34, f80
  5923. ;;
  5924. FNMA f88 = f64, f35, f88
  5925. ;;
  5926. FNMA f96 = f64, f36, f96
  5927. ;;
  5928. FNMA f104 = f64, f37, f104
  5929. ;;
  5930. FNMA f112 = f64, f38, f112
  5931. ;;
  5932. FNMA f120 = f64, f39, f120
  5933. ;;
  5934. FMPY f72 = f72, f40
  5935. ;;
  5936. FNMA f80 = f72, f41, f80
  5937. ;;
  5938. FNMA f88 = f72, f42, f88
  5939. ;;
  5940. FNMA f96 = f72, f43, f96
  5941. ;;
  5942. FNMA f104 = f72, f44, f104
  5943. ;;
  5944. FNMA f112 = f72, f45, f112
  5945. ;;
  5946. FNMA f120 = f72, f46, f120
  5947. ;;
  5948. FMPY f80 = f80, f47
  5949. ;;
  5950. FNMA f88 = f80, f48, f88
  5951. ;;
  5952. FNMA f96 = f80, f49, f96
  5953. ;;
  5954. FNMA f104 = f80, f50, f104
  5955. ;;
  5956. FNMA f112 = f80, f51, f112
  5957. ;;
  5958. FNMA f120 = f80, f52, f120
  5959. ;;
  5960. FMPY f88 = f88, f53
  5961. ;;
  5962. FNMA f96 = f88, f54, f96
  5963. ;;
  5964. FNMA f104 = f88, f55, f104
  5965. ;;
  5966. FNMA f112 = f88, f56, f112
  5967. ;;
  5968. FNMA f120 = f88, f57, f120
  5969. ;;
  5970. FMPY f96 = f96, f58
  5971. ;;
  5972. FNMA f104 = f96, f59, f104
  5973. ;;
  5974. FNMA f112 = f96, f60, f112
  5975. ;;
  5976. FNMA f120 = f96, f61, f120
  5977. ;;
  5978. FMPY f104 = f104, f16
  5979. ;;
  5980. FNMA f112 = f104, f17, f112
  5981. ;;
  5982. FNMA f120 = f104, f18, f120
  5983. ;;
  5984. FMPY f112 = f112, f19
  5985. ;;
  5986. FNMA f120 = f112, f20, f120
  5987. ;;
  5988. FMPY f120 = f120, f21
  5989. ;;
  5990. STFD [AOFFSET] = f64, SIZE
  5991. STFD [AOFFSET2] = f96, SIZE
  5992. ;;
  5993. STFD [AOFFSET] = f72, SIZE
  5994. STFD [AOFFSET2] = f104, SIZE
  5995. ;;
  5996. STFD [AOFFSET] = f80, SIZE
  5997. STFD [AOFFSET2] = f112, SIZE
  5998. ;;
  5999. STFD [AOFFSET] = f88, -3 * SIZE
  6000. STFD [AOFFSET2] = f120, - 3 * SIZE
  6001. ;;
  6002. #endif
  6003. STFD [C1 ] = f64, SIZE
  6004. STFD [C2 ] = f72, SIZE
  6005. STFD [C3 ] = f80, SIZE
  6006. STFD [C4 ] = f88, SIZE
  6007. STFD [C5 ] = f96, SIZE
  6008. STFD [C6 ] = f104, SIZE
  6009. STFD [C7 ] = f112, SIZE
  6010. STFD [C8 ] = f120, SIZE
  6011. ;;
  6012. mov f64 = f0
  6013. mov f72 = f0
  6014. mov f80 = f0
  6015. mov f88 = f0
  6016. mov f96 = f0
  6017. mov f104 = f0
  6018. mov f112 = f0
  6019. mov f120 = f0
  6020. ;;
  6021. sub L = K, KK
  6022. ;;
  6023. shladd L = L, BASE_SHIFT, r0
  6024. ;;
  6025. add AOFFSET = L, AOFFSET
  6026. ;;
  6027. shladd BOFFSET = L, 3, BOFFSET
  6028. ;;
  6029. #ifdef LT
  6030. adds KK = 1, KK
  6031. #else
  6032. nop __LINE__
  6033. #endif
  6034. ;;
  6035. mov L = KK
  6036. ;;
  6037. .align 8
  6038. .L049:
  6039. mov B = BOFFSET
  6040. #ifdef RN
  6041. adds KK = 8, KK
  6042. #endif
  6043. ;;
  6044. { .mmi
  6045. mov AOFFSET = A
  6046. }
  6047. ;;
  6048. { .mmb
  6049. nop __LINE__
  6050. cmp.lt p6, p0 = 0, J
  6051. (p6) br.cond.dptk .L010
  6052. }
  6053. ;;
  6054. .align 8
  6055. .L050:
  6056. { .mib
  6057. setf.d f64 = r0
  6058. tbit.z p6, p0 = N, 2
  6059. (p6) br.cond.dpnt .L090
  6060. }
  6061. ;;
  6062. { .mfi
  6063. setf.d f72 = r0
  6064. mov f80 = f0
  6065. shr I = M, 3
  6066. }
  6067. { .mfi
  6068. mov C1 = C // coffset1 = c + 0 * ldc
  6069. mov f88 = f0
  6070. #ifdef LT
  6071. mov KK = OFFSET
  6072. #else
  6073. nop __LINE__
  6074. #endif
  6075. }
  6076. ;;
  6077. { .mmf
  6078. cmp.eq p6, p7 = 0, I
  6079. mov AORIG = A
  6080. mov f65 = f0
  6081. }
  6082. { .mmf
  6083. add C2 = LDC, C // coffset2 = c + 1 * ldc
  6084. shladd C3 = LDC, 1, C // coffset3 = c + 2 * ldc
  6085. mov f73 = f0
  6086. }
  6087. ;;
  6088. { .mfi
  6089. shladd C = LDC, 2, C // coffset += 8 * ldc
  6090. mov f81 = f0
  6091. mov L = KK
  6092. }{ .mfb
  6093. shladd C4 = LDC, 1, C2
  6094. mov f89 = f0
  6095. (p6) br.cond.dpnt .L060
  6096. }
  6097. ;;
  6098. .align 16
  6099. .L052:
  6100. { .mmi
  6101. cmp.ne p7, p0 = r0, L
  6102. adds BOFFSET = 0 * SIZE, B
  6103. }
  6104. ;;
  6105. { .mmi
  6106. (p7) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  6107. nop __LINE__
  6108. nop __LINE__
  6109. }
  6110. ;;
  6111. { .mfi
  6112. (p7) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  6113. mov f66 = f0
  6114. nop __LINE__
  6115. }
  6116. { .mfi
  6117. (p7) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  6118. mov f74 = f0
  6119. nop __LINE__
  6120. }
  6121. ;;
  6122. { .mmf
  6123. (p7) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  6124. setf.d f82 = r0
  6125. mov f90 = f0
  6126. }
  6127. ;;
  6128. { .mmf
  6129. (p7) LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  6130. setf.d f67 = r0
  6131. mov f75 = f0
  6132. }
  6133. { .mfi
  6134. setf.d f83 = r0
  6135. mov f91 = f0
  6136. cmp.eq p3, p0 = r0, r0
  6137. }
  6138. ;;
  6139. { .mmf
  6140. (p7) LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  6141. }
  6142. { .mfi
  6143. adds PREC = CPREFETCHSIZE * SIZE, C1
  6144. }
  6145. ;;
  6146. { .mmf
  6147. CPREFETCH [PREC], LDC
  6148. setf.d f68 = r0
  6149. mov f76 = f0
  6150. }
  6151. { .mfi
  6152. setf.d f84 = r0
  6153. mov f92 = f0
  6154. adds L = 1, L
  6155. }
  6156. ;;
  6157. { .mmf
  6158. CPREFETCH [PREC], LDC
  6159. }
  6160. { .mfi
  6161. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  6162. }
  6163. ;;
  6164. { .mmf
  6165. CPREFETCH [PREC], LDC
  6166. setf.d f69 = r0
  6167. mov f77 = f0
  6168. }
  6169. { .mfi
  6170. setf.d f85 = r0
  6171. mov f93 = f0
  6172. adds PREB = (PREFETCHSIZE - 8) * SIZE, BOFFSET
  6173. }
  6174. ;;
  6175. { .mmf
  6176. CPREFETCH [PREC]
  6177. }
  6178. ;;
  6179. { .mfi
  6180. setf.d f70 = r0
  6181. mov f78 = f0
  6182. tbit.z p12, p0 = L, 0
  6183. }
  6184. { .mfi
  6185. setf.d f86 = r0
  6186. mov f94 = f0
  6187. shr L = L, 1
  6188. }
  6189. ;;
  6190. { .mfi
  6191. setf.d f71 = r0
  6192. adds L = -1, L
  6193. }
  6194. ;;
  6195. { .mfi
  6196. setf.d f87 = r0
  6197. mov f79 = f0
  6198. mov ar.lc = L
  6199. }
  6200. { .mfb
  6201. cmp.eq p6, p0 = -1, L
  6202. mov f95 = f0
  6203. (p6) br.cond.dpnt .L058
  6204. }
  6205. ;;
  6206. .align 8
  6207. .L053:
  6208. { .mfb
  6209. lfetch.nt1 [PREA], 16 * SIZE
  6210. FMA f64 = f32, f48, f64 // A1 * B1
  6211. nop __LINE__
  6212. }
  6213. { .mfi
  6214. nop __LINE__
  6215. FMA f72 = f32, f49, f72 // A1 * B2
  6216. (p12) cmp.ne p3, p0 = 0, L
  6217. }
  6218. ;;
  6219. { .mfi
  6220. lfetch.nt1 [PREB], 8 * SIZE
  6221. FMA f80 = f32, f50, f80 // A1 * B3
  6222. cmp.ne p4, p5 = 0, L
  6223. }
  6224. { .mfi
  6225. nop __LINE__
  6226. FMA f88 = f32, f51, f88 // A1 * B4
  6227. adds C9 = 4 * SIZE, C1
  6228. }
  6229. ;;
  6230. { .mfi
  6231. (p3) LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  6232. FMA f65 = f33, f48, f65 // A2 * B1
  6233. adds C10 = 4 * SIZE, C2
  6234. }
  6235. { .mfi
  6236. nop __LINE__
  6237. FMA f73 = f33, f49, f73 // A2 * B2
  6238. adds C11 = 4 * SIZE, C3
  6239. }
  6240. ;;
  6241. { .mfi
  6242. (p3) LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  6243. FMA f81 = f33, f50, f81 // A2 * B3
  6244. adds C12 = 4 * SIZE, C4
  6245. }
  6246. { .mfb
  6247. nop __LINE__
  6248. FMA f89 = f33, f51, f89 // A2 * B4
  6249. nop __LINE__
  6250. }
  6251. ;;
  6252. { .mfb
  6253. (p3) LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  6254. FMA f66 = f34, f48, f66 // A3 * B1
  6255. nop __LINE__
  6256. }
  6257. { .mfb
  6258. nop __LINE__
  6259. FMA f74 = f34, f49, f74 // A3 * B2
  6260. nop __LINE__
  6261. }
  6262. ;;
  6263. { .mfb
  6264. (p3) LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  6265. FMA f82 = f34, f50, f82 // A3 * B3
  6266. nop __LINE__
  6267. }
  6268. { .mfb
  6269. nop __LINE__
  6270. FMA f90 = f34, f51, f90 // A3 * B4
  6271. nop __LINE__
  6272. }
  6273. ;;
  6274. { .mfb
  6275. (p3) LDFPD f44, f45 = [AOFFSET], 2 * SIZE
  6276. FMA f67 = f35, f48, f67 // A4 * B1
  6277. nop __LINE__
  6278. }
  6279. { .mfb
  6280. nop __LINE__
  6281. FMA f75 = f35, f49, f75 // A4 * B2
  6282. nop __LINE__
  6283. }
  6284. ;;
  6285. { .mfb
  6286. (p3) LDFPD f46, f47 = [AOFFSET], 2 * SIZE
  6287. FMA f83 = f35, f50, f83 // A4 * B3
  6288. nop __LINE__
  6289. }
  6290. { .mfb
  6291. nop __LINE__
  6292. FMA f91 = f35, f51, f91 // A4 * B4
  6293. nop __LINE__
  6294. }
  6295. ;;
  6296. { .mfb
  6297. nop __LINE__
  6298. FMA f68 = f36, f48, f68 // A5 * B1
  6299. nop __LINE__
  6300. }
  6301. { .mfb
  6302. nop __LINE__
  6303. FMA f76 = f36, f49, f76 // A5 * B2
  6304. nop __LINE__
  6305. }
  6306. ;;
  6307. { .mfb
  6308. nop __LINE__
  6309. FMA f84 = f36, f50, f84 // A5 * B3
  6310. nop __LINE__
  6311. }
  6312. { .mfb
  6313. nop __LINE__
  6314. FMA f92 = f36, f51, f92 // A5 * B4
  6315. nop __LINE__
  6316. }
  6317. ;;
  6318. { .mfb
  6319. nop __LINE__
  6320. FMA f69 = f37, f48, f69 // A6 * B1
  6321. nop __LINE__
  6322. }
  6323. { .mfb
  6324. nop __LINE__
  6325. FMA f77 = f37, f49, f77 // A6 * B2
  6326. nop __LINE__
  6327. }
  6328. ;;
  6329. { .mfb
  6330. nop __LINE__
  6331. FMA f85 = f37, f50, f85 // A6 * B3
  6332. nop __LINE__
  6333. }
  6334. { .mfb
  6335. nop __LINE__
  6336. FMA f93 = f37, f51, f93 // A6 * B4
  6337. nop __LINE__
  6338. }
  6339. ;;
  6340. { .mfb
  6341. nop __LINE__
  6342. FMA f70 = f38, f48, f70 // A7 * B1
  6343. nop __LINE__
  6344. }
  6345. { .mfb
  6346. nop __LINE__
  6347. FMA f78 = f38, f49, f78 // A7 * B2
  6348. nop __LINE__
  6349. }
  6350. ;;
  6351. { .mfb
  6352. nop __LINE__
  6353. FMA f86 = f38, f50, f86 // A7 * B3
  6354. nop __LINE__
  6355. }
  6356. { .mfb
  6357. nop __LINE__
  6358. FMA f94 = f38, f51, f94 // A7 * B4
  6359. nop __LINE__
  6360. }
  6361. ;;
  6362. { .mfb
  6363. (p4) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  6364. FMA f71 = f39, f48, f71 // A8 * B1
  6365. nop __LINE__
  6366. }
  6367. { .mfb
  6368. nop __LINE__
  6369. FMA f79 = f39, f49, f79 // A8 * B2
  6370. nop __LINE__
  6371. }
  6372. ;;
  6373. { .mfb
  6374. (p4) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  6375. FMA f87 = f39, f50, f87 // A8 * B3
  6376. nop __LINE__
  6377. }
  6378. { .mfb
  6379. nop __LINE__
  6380. FMA f95 = f39, f51, f95 // A8 * B4
  6381. nop __LINE__
  6382. }
  6383. ;;
  6384. { .mfb
  6385. (p4) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  6386. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  6387. nop __LINE__
  6388. }
  6389. { .mfb
  6390. nop __LINE__
  6391. (p3) FMA f72 = f40, f57, f72 // A1 * B2
  6392. nop __LINE__
  6393. }
  6394. ;;
  6395. { .mfb
  6396. (p4) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  6397. (p3) FMA f80 = f40, f58, f80 // A1 * B3
  6398. nop __LINE__
  6399. }
  6400. { .mfb
  6401. nop __LINE__
  6402. (p3) FMA f88 = f40, f59, f88 // A1 * B4
  6403. nop __LINE__
  6404. }
  6405. ;;
  6406. { .mfb
  6407. (p4) LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  6408. (p3) FMA f65 = f41, f56, f65 // A2 * B1
  6409. nop __LINE__
  6410. }
  6411. { .mfb
  6412. nop __LINE__
  6413. (p3) FMA f73 = f41, f57, f73 // A2 * B2
  6414. nop __LINE__
  6415. }
  6416. ;;
  6417. { .mfb
  6418. (p4) LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  6419. (p3) FMA f81 = f41, f58, f81 // A2 * B3
  6420. nop __LINE__
  6421. }
  6422. { .mfb
  6423. nop __LINE__
  6424. (p3) FMA f89 = f41, f59, f89 // A2 * B4
  6425. nop __LINE__
  6426. }
  6427. ;;
  6428. { .mfb
  6429. nop __LINE__
  6430. (p3) FMA f66 = f42, f56, f66 // A3 * B1
  6431. nop __LINE__
  6432. }
  6433. { .mfb
  6434. nop __LINE__
  6435. (p3) FMA f74 = f42, f57, f74 // A3 * B2
  6436. nop __LINE__
  6437. }
  6438. ;;
  6439. { .mfb
  6440. nop __LINE__
  6441. (p3) FMA f82 = f42, f58, f82 // A3 * B3
  6442. nop __LINE__
  6443. }
  6444. { .mfb
  6445. nop __LINE__
  6446. (p3) FMA f90 = f42, f59, f90 // A3 * B4
  6447. nop __LINE__
  6448. }
  6449. ;;
  6450. { .mfb
  6451. nop __LINE__
  6452. (p3) FMA f67 = f43, f56, f67 // A4 * B1
  6453. nop __LINE__
  6454. }
  6455. { .mfb
  6456. nop __LINE__
  6457. (p3) FMA f75 = f43, f57, f75 // A4 * B2
  6458. nop __LINE__
  6459. }
  6460. ;;
  6461. { .mfb
  6462. nop __LINE__
  6463. (p3) FMA f83 = f43, f58, f83 // A4 * B3
  6464. nop __LINE__
  6465. }
  6466. { .mfb
  6467. nop __LINE__
  6468. (p3) FMA f91 = f43, f59, f91 // A4 * B4
  6469. nop __LINE__
  6470. }
  6471. ;;
  6472. { .mfb
  6473. nop __LINE__
  6474. (p3) FMA f68 = f44, f56, f68 // A5 * B1
  6475. nop __LINE__
  6476. }
  6477. { .mfb
  6478. nop __LINE__
  6479. (p3) FMA f76 = f44, f57, f76 // A5 * B2
  6480. nop __LINE__
  6481. }
  6482. ;;
  6483. { .mfb
  6484. nop __LINE__
  6485. (p3) FMA f84 = f44, f58, f84 // A5 * B3
  6486. nop __LINE__
  6487. }
  6488. { .mfb
  6489. nop __LINE__
  6490. (p3) FMA f92 = f44, f59, f92 // A5 * B4
  6491. nop __LINE__
  6492. }
  6493. ;;
  6494. { .mfb
  6495. nop __LINE__
  6496. (p3) FMA f69 = f45, f56, f69 // A6 * B1
  6497. nop __LINE__
  6498. }
  6499. { .mfb
  6500. nop __LINE__
  6501. (p3) FMA f77 = f45, f57, f77 // A6 * B2
  6502. nop __LINE__
  6503. }
  6504. ;;
  6505. { .mfb
  6506. nop __LINE__
  6507. (p3) FMA f85 = f45, f58, f85 // A6 * B3
  6508. nop __LINE__
  6509. }
  6510. { .mfb
  6511. nop __LINE__
  6512. (p3) FMA f93 = f45, f59, f93 // A6 * B4
  6513. nop __LINE__
  6514. }
  6515. ;;
  6516. { .mfb
  6517. nop __LINE__
  6518. (p3) FMA f70 = f46, f56, f70 // A7 * B1
  6519. nop __LINE__
  6520. }
  6521. { .mfb
  6522. nop __LINE__
  6523. (p3) FMA f78 = f46, f57, f78 // A7 * B2
  6524. nop __LINE__
  6525. }
  6526. ;;
  6527. { .mfb
  6528. nop __LINE__
  6529. (p3) FMA f86 = f46, f58, f86 // A7 * B3
  6530. nop __LINE__
  6531. }
  6532. { .mfb
  6533. nop __LINE__
  6534. (p3) FMA f94 = f46, f59, f94 // A7 * B4
  6535. nop __LINE__
  6536. }
  6537. ;;
  6538. { .mfb
  6539. nop __LINE__
  6540. (p3) FMA f71 = f47, f56, f71 // A8 * B1
  6541. nop __LINE__
  6542. }
  6543. { .mfb
  6544. nop __LINE__
  6545. (p3) FMA f79 = f47, f57, f79 // A8 * B2
  6546. nop __LINE__
  6547. }
  6548. ;;
  6549. { .mfi
  6550. nop __LINE__
  6551. (p3) FMA f87 = f47, f58, f87 // A8 * B3
  6552. adds L = -1, L
  6553. }
  6554. { .mfb
  6555. nop __LINE__
  6556. (p3) FMA f95 = f47, f59, f95 // A8 * B4
  6557. br.cloop.sptk.few .L053
  6558. }
  6559. ;;
  6560. .align 8
  6561. .L058:
  6562. adds AOFFSET2 = 4 * SIZE, AOFFSET
  6563. adds BOFFSET2 = 4 * SIZE, BOFFSET
  6564. ;;
  6565. #ifdef LT
  6566. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  6567. ;;
  6568. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  6569. ;;
  6570. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  6571. ;;
  6572. LDFPD f38, f39 = [BOFFSET], 2 * SIZE
  6573. ;;
  6574. LDFPD f40, f41 = [BOFFSET], 2 * SIZE
  6575. ;;
  6576. LDFPD f42, f43 = [BOFFSET], 2 * SIZE
  6577. ;;
  6578. LDFPD f44, f45 = [BOFFSET], 2 * SIZE
  6579. ;;
  6580. LDFPD f46, f47 = [BOFFSET], 2 * SIZE
  6581. ;;
  6582. LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  6583. ;;
  6584. LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  6585. ;;
  6586. LDFPD f52, f53 = [BOFFSET], 2 * SIZE
  6587. ;;
  6588. LDFPD f54, f55 = [BOFFSET], 2 * SIZE
  6589. ;;
  6590. LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  6591. ;;
  6592. LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  6593. ;;
  6594. LDFPD f60, f61 = [BOFFSET], 2 * SIZE
  6595. ;;
  6596. LDFPD f62, f63 = [BOFFSET]
  6597. adds BOFFSET = -30 * SIZE, BOFFSET
  6598. ;;
  6599. FSUB f64 = f32, f64
  6600. FSUB f72 = f33, f72
  6601. FSUB f80 = f34, f80
  6602. FSUB f88 = f35, f88
  6603. FSUB f65 = f36, f65
  6604. FSUB f73 = f37, f73
  6605. FSUB f81 = f38, f81
  6606. FSUB f89 = f39, f89
  6607. FSUB f66 = f40, f66
  6608. FSUB f74 = f41, f74
  6609. FSUB f82 = f42, f82
  6610. FSUB f90 = f43, f90
  6611. FSUB f67 = f44, f67
  6612. FSUB f75 = f45, f75
  6613. FSUB f83 = f46, f83
  6614. FSUB f91 = f47, f91
  6615. FSUB f68 = f48, f68
  6616. FSUB f76 = f49, f76
  6617. FSUB f84 = f50, f84
  6618. FSUB f92 = f51, f92
  6619. FSUB f69 = f52, f69
  6620. FSUB f77 = f53, f77
  6621. FSUB f85 = f54, f85
  6622. FSUB f93 = f55, f93
  6623. FSUB f70 = f56, f70
  6624. FSUB f78 = f57, f78
  6625. FSUB f86 = f58, f86
  6626. FSUB f94 = f59, f94
  6627. FSUB f71 = f60, f71
  6628. FSUB f79 = f61, f79
  6629. FSUB f87 = f62, f87
  6630. FSUB f95 = f63, f95
  6631. ;;
  6632. #else
  6633. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  6634. ;;
  6635. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  6636. ;;
  6637. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  6638. ;;
  6639. LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  6640. ;;
  6641. LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  6642. ;;
  6643. LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  6644. ;;
  6645. LDFPD f44, f45 = [AOFFSET], 2 * SIZE
  6646. ;;
  6647. LDFPD f46, f47 = [AOFFSET], 2 * SIZE
  6648. ;;
  6649. LDFPD f48, f49 = [AOFFSET], 2 * SIZE
  6650. ;;
  6651. LDFPD f50, f51 = [AOFFSET], 2 * SIZE
  6652. ;;
  6653. LDFPD f52, f53 = [AOFFSET], 2 * SIZE
  6654. ;;
  6655. LDFPD f54, f55 = [AOFFSET], 2 * SIZE
  6656. ;;
  6657. LDFPD f56, f57 = [AOFFSET], 2 * SIZE
  6658. ;;
  6659. LDFPD f58, f59 = [AOFFSET], 2 * SIZE
  6660. ;;
  6661. LDFPD f60, f61 = [AOFFSET], 2 * SIZE
  6662. ;;
  6663. LDFPD f62, f63 = [AOFFSET]
  6664. adds AOFFSET = -30 * SIZE, AOFFSET
  6665. ;;
  6666. FSUB f64 = f32, f64
  6667. FSUB f65 = f33, f65
  6668. FSUB f66 = f34, f66
  6669. FSUB f67 = f35, f67
  6670. FSUB f68 = f36, f68
  6671. FSUB f69 = f37, f69
  6672. FSUB f70 = f38, f70
  6673. FSUB f71 = f39, f71
  6674. ;;
  6675. FSUB f72 = f40, f72
  6676. FSUB f73 = f41, f73
  6677. FSUB f74 = f42, f74
  6678. FSUB f75 = f43, f75
  6679. FSUB f76 = f44, f76
  6680. FSUB f77 = f45, f77
  6681. FSUB f78 = f46, f78
  6682. FSUB f79 = f47, f79
  6683. ;;
  6684. FSUB f80 = f48, f80
  6685. FSUB f81 = f49, f81
  6686. FSUB f82 = f50, f82
  6687. FSUB f83 = f51, f83
  6688. FSUB f84 = f52, f84
  6689. FSUB f85 = f53, f85
  6690. FSUB f86 = f54, f86
  6691. FSUB f87 = f55, f87
  6692. FSUB f88 = f56, f88
  6693. FSUB f89 = f57, f89
  6694. FSUB f90 = f58, f90
  6695. FSUB f91 = f59, f91
  6696. FSUB f92 = f60, f92
  6697. FSUB f93 = f61, f93
  6698. FSUB f94 = f62, f94
  6699. FSUB f95 = f63, f95
  6700. ;;
  6701. #endif
  6702. #ifdef LT
  6703. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  6704. ;;
  6705. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  6706. ;;
  6707. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  6708. ;;
  6709. LDFPD f38, f39 = [AOFFSET]
  6710. adds AOFFSET = 3 * SIZE, AOFFSET
  6711. ;;
  6712. LDFD f40 = [AOFFSET], 1 * SIZE
  6713. ;;
  6714. LDFPD f41, f42 = [AOFFSET], 2 * SIZE
  6715. ;;
  6716. LDFPD f43, f44 = [AOFFSET], 2 * SIZE
  6717. ;;
  6718. LDFPD f45, f46 = [AOFFSET]
  6719. adds AOFFSET = 4 * SIZE, AOFFSET
  6720. ;;
  6721. LDFPD f47, f48 = [AOFFSET], 2 * SIZE
  6722. ;;
  6723. LDFPD f49, f50 = [AOFFSET], 2 * SIZE
  6724. ;;
  6725. LDFPD f51, f52 = [AOFFSET]
  6726. adds AOFFSET = 5 * SIZE, AOFFSET
  6727. ;;
  6728. LDFD f53 = [AOFFSET], 1 * SIZE
  6729. ;;
  6730. LDFPD f54, f55 = [AOFFSET], 2 * SIZE
  6731. ;;
  6732. LDFPD f56, f57 = [AOFFSET]
  6733. adds AOFFSET = 6 * SIZE, AOFFSET
  6734. ;;
  6735. LDFPD f58, f59 = [AOFFSET], 2 * SIZE
  6736. ;;
  6737. LDFPD f60, f61 = [AOFFSET]
  6738. adds AOFFSET = 7 * SIZE, AOFFSET
  6739. ;;
  6740. LDFD f16 = [AOFFSET], 1 * SIZE
  6741. ;;
  6742. LDFPD f17, f18 = [AOFFSET]
  6743. adds AOFFSET = 8 * SIZE, AOFFSET
  6744. ;;
  6745. LDFPD f19, f20 = [AOFFSET]
  6746. adds AOFFSET = 9 * SIZE, AOFFSET
  6747. ;;
  6748. LDFD f21 = [AOFFSET]
  6749. adds AOFFSET = -63 * SIZE, AOFFSET
  6750. ;;
  6751. FMPY f64 = f64, f32
  6752. FMPY f72 = f72, f32
  6753. FMPY f80 = f80, f32
  6754. FMPY f88 = f88, f32
  6755. ;;
  6756. FNMA f65 = f64, f33, f65
  6757. FNMA f73 = f72, f33, f73
  6758. FNMA f81 = f80, f33, f81
  6759. FNMA f89 = f88, f33, f89
  6760. ;;
  6761. FNMA f66 = f64, f34, f66
  6762. FNMA f74 = f72, f34, f74
  6763. FNMA f82 = f80, f34, f82
  6764. FNMA f90 = f88, f34, f90
  6765. ;;
  6766. FNMA f67 = f64, f35, f67
  6767. FNMA f75 = f72, f35, f75
  6768. FNMA f83 = f80, f35, f83
  6769. FNMA f91 = f88, f35, f91
  6770. ;;
  6771. FNMA f68 = f64, f36, f68
  6772. FNMA f76 = f72, f36, f76
  6773. FNMA f84 = f80, f36, f84
  6774. FNMA f92 = f88, f36, f92
  6775. ;;
  6776. FNMA f69 = f64, f37, f69
  6777. FNMA f77 = f72, f37, f77
  6778. FNMA f85 = f80, f37, f85
  6779. FNMA f93 = f88, f37, f93
  6780. ;;
  6781. FNMA f70 = f64, f38, f70
  6782. FNMA f78 = f72, f38, f78
  6783. FNMA f86 = f80, f38, f86
  6784. FNMA f94 = f88, f38, f94
  6785. ;;
  6786. FNMA f71 = f64, f39, f71
  6787. FNMA f79 = f72, f39, f79
  6788. FNMA f87 = f80, f39, f87
  6789. FNMA f95 = f88, f39, f95
  6790. ;;
  6791. FMPY f65 = f65, f40
  6792. FMPY f73 = f73, f40
  6793. FMPY f81 = f81, f40
  6794. FMPY f89 = f89, f40
  6795. ;;
  6796. FNMA f66 = f65, f41, f66
  6797. FNMA f74 = f73, f41, f74
  6798. FNMA f82 = f81, f41, f82
  6799. FNMA f90 = f89, f41, f90
  6800. ;;
  6801. FNMA f67 = f65, f42, f67
  6802. FNMA f75 = f73, f42, f75
  6803. FNMA f83 = f81, f42, f83
  6804. FNMA f91 = f89, f42, f91
  6805. ;;
  6806. FNMA f68 = f65, f43, f68
  6807. FNMA f76 = f73, f43, f76
  6808. FNMA f84 = f81, f43, f84
  6809. FNMA f92 = f89, f43, f92
  6810. ;;
  6811. FNMA f69 = f65, f44, f69
  6812. FNMA f77 = f73, f44, f77
  6813. FNMA f85 = f81, f44, f85
  6814. FNMA f93 = f89, f44, f93
  6815. ;;
  6816. FNMA f70 = f65, f45, f70
  6817. FNMA f78 = f73, f45, f78
  6818. FNMA f86 = f81, f45, f86
  6819. FNMA f94 = f89, f45, f94
  6820. ;;
  6821. FNMA f71 = f65, f46, f71
  6822. FNMA f79 = f73, f46, f79
  6823. FNMA f87 = f81, f46, f87
  6824. FNMA f95 = f89, f46, f95
  6825. ;;
  6826. FMPY f66 = f66, f47
  6827. FMPY f74 = f74, f47
  6828. FMPY f82 = f82, f47
  6829. FMPY f90 = f90, f47
  6830. ;;
  6831. FNMA f67 = f66, f48, f67
  6832. FNMA f75 = f74, f48, f75
  6833. FNMA f83 = f82, f48, f83
  6834. FNMA f91 = f90, f48, f91
  6835. ;;
  6836. FNMA f68 = f66, f49, f68
  6837. FNMA f76 = f74, f49, f76
  6838. FNMA f84 = f82, f49, f84
  6839. FNMA f92 = f90, f49, f92
  6840. ;;
  6841. FNMA f69 = f66, f50, f69
  6842. FNMA f77 = f74, f50, f77
  6843. FNMA f85 = f82, f50, f85
  6844. FNMA f93 = f90, f50, f93
  6845. ;;
  6846. FNMA f70 = f66, f51, f70
  6847. FNMA f78 = f74, f51, f78
  6848. FNMA f86 = f82, f51, f86
  6849. FNMA f94 = f90, f51, f94
  6850. ;;
  6851. FNMA f71 = f66, f52, f71
  6852. FNMA f79 = f74, f52, f79
  6853. FNMA f87 = f82, f52, f87
  6854. FNMA f95 = f90, f52, f95
  6855. ;;
  6856. FMPY f67 = f67, f53
  6857. FMPY f75 = f75, f53
  6858. FMPY f83 = f83, f53
  6859. FMPY f91 = f91, f53
  6860. ;;
  6861. FNMA f68 = f67, f54, f68
  6862. FNMA f76 = f75, f54, f76
  6863. FNMA f84 = f83, f54, f84
  6864. FNMA f92 = f91, f54, f92
  6865. ;;
  6866. FNMA f69 = f67, f55, f69
  6867. FNMA f77 = f75, f55, f77
  6868. FNMA f85 = f83, f55, f85
  6869. FNMA f93 = f91, f55, f93
  6870. ;;
  6871. FNMA f70 = f67, f56, f70
  6872. FNMA f78 = f75, f56, f78
  6873. FNMA f86 = f83, f56, f86
  6874. FNMA f94 = f91, f56, f94
  6875. ;;
  6876. FNMA f71 = f67, f57, f71
  6877. FNMA f79 = f75, f57, f79
  6878. FNMA f87 = f83, f57, f87
  6879. FNMA f95 = f91, f57, f95
  6880. ;;
  6881. FMPY f68 = f68, f58
  6882. FMPY f76 = f76, f58
  6883. FMPY f84 = f84, f58
  6884. FMPY f92 = f92, f58
  6885. ;;
  6886. FNMA f69 = f68, f59, f69
  6887. FNMA f77 = f76, f59, f77
  6888. FNMA f85 = f84, f59, f85
  6889. FNMA f93 = f92, f59, f93
  6890. ;;
  6891. FNMA f70 = f68, f60, f70
  6892. FNMA f78 = f76, f60, f78
  6893. FNMA f86 = f84, f60, f86
  6894. FNMA f94 = f92, f60, f94
  6895. ;;
  6896. FNMA f71 = f68, f61, f71
  6897. FNMA f79 = f76, f61, f79
  6898. FNMA f87 = f84, f61, f87
  6899. FNMA f95 = f92, f61, f95
  6900. ;;
  6901. FMPY f69 = f69, f16
  6902. FMPY f77 = f77, f16
  6903. FMPY f85 = f85, f16
  6904. FMPY f93 = f93, f16
  6905. ;;
  6906. FNMA f70 = f69, f17, f70
  6907. FNMA f78 = f77, f17, f78
  6908. FNMA f86 = f85, f17, f86
  6909. FNMA f94 = f93, f17, f94
  6910. ;;
  6911. FNMA f71 = f69, f18, f71
  6912. FNMA f79 = f77, f18, f79
  6913. FNMA f87 = f85, f18, f87
  6914. FNMA f95 = f93, f18, f95
  6915. ;;
  6916. FMPY f70 = f70, f19
  6917. FMPY f78 = f78, f19
  6918. FMPY f86 = f86, f19
  6919. FMPY f94 = f94, f19
  6920. ;;
  6921. FNMA f71 = f70, f20, f71
  6922. FNMA f79 = f78, f20, f79
  6923. FNMA f87 = f86, f20, f87
  6924. FNMA f95 = f94, f20, f95
  6925. ;;
  6926. FMPY f71 = f71, f21
  6927. FMPY f79 = f79, f21
  6928. FMPY f87 = f87, f21
  6929. FMPY f95 = f95, f21
  6930. ;;
  6931. STFD [BOFFSET] = f64, SIZE
  6932. STFD [BOFFSET2] = f65, SIZE
  6933. ;;
  6934. STFD [BOFFSET] = f72, SIZE
  6935. STFD [BOFFSET2] = f73, SIZE
  6936. ;;
  6937. STFD [BOFFSET] = f80, SIZE
  6938. STFD [BOFFSET2] = f81, SIZE
  6939. ;;
  6940. STFD [BOFFSET] = f88, 5 * SIZE
  6941. STFD [BOFFSET2] = f89, 5 * SIZE
  6942. ;;
  6943. STFD [BOFFSET] = f66, SIZE
  6944. STFD [BOFFSET2] = f67, SIZE
  6945. ;;
  6946. STFD [BOFFSET] = f74, SIZE
  6947. STFD [BOFFSET2] = f75, SIZE
  6948. ;;
  6949. STFD [BOFFSET] = f82, SIZE
  6950. STFD [BOFFSET2] = f83, SIZE
  6951. ;;
  6952. STFD [BOFFSET] = f90, 5 * SIZE
  6953. STFD [BOFFSET2] = f91, 5 * SIZE
  6954. ;;
  6955. STFD [BOFFSET] = f68, SIZE
  6956. STFD [BOFFSET2] = f69, SIZE
  6957. ;;
  6958. STFD [BOFFSET] = f76, SIZE
  6959. STFD [BOFFSET2] = f77, SIZE
  6960. ;;
  6961. STFD [BOFFSET] = f84, SIZE
  6962. STFD [BOFFSET2] = f85, SIZE
  6963. ;;
  6964. STFD [BOFFSET] = f92, 5 * SIZE
  6965. STFD [BOFFSET2] = f93, 5 * SIZE
  6966. ;;
  6967. STFD [BOFFSET] = f70, SIZE
  6968. STFD [BOFFSET2] = f71, SIZE
  6969. ;;
  6970. STFD [BOFFSET] = f78, SIZE
  6971. STFD [BOFFSET2] = f79, SIZE
  6972. ;;
  6973. STFD [BOFFSET] = f86, SIZE
  6974. STFD [BOFFSET2] = f87, SIZE
  6975. ;;
  6976. STFD [BOFFSET] = f94
  6977. STFD [BOFFSET2] = f95
  6978. adds C9 = 4 * SIZE, C1
  6979. adds BOFFSET = - 27 * SIZE, BOFFSET
  6980. adds BOFFSET2 = - 27 * SIZE, BOFFSET2
  6981. ;;
  6982. #endif
  6983. #ifdef RN
  6984. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  6985. ;;
  6986. LDFPD f34, f35 = [BOFFSET]
  6987. adds BOFFSET = 3 * SIZE, BOFFSET
  6988. ;;
  6989. LDFD f36 = [BOFFSET], 1 * SIZE
  6990. ;;
  6991. LDFPD f37, f38 = [BOFFSET]
  6992. adds BOFFSET = 4 * SIZE, BOFFSET
  6993. ;;
  6994. LDFPD f39, f40 = [BOFFSET]
  6995. adds BOFFSET = 5 * SIZE, BOFFSET
  6996. ;;
  6997. LDFD f41 = [BOFFSET], -15 * SIZE
  6998. ;;
  6999. FMPY f64 = f64, f32
  7000. FMPY f68 = f68, f32
  7001. FMPY f65 = f65, f32
  7002. FMPY f69 = f69, f32
  7003. FMPY f66 = f66, f32
  7004. FMPY f70 = f70, f32
  7005. FMPY f67 = f67, f32
  7006. FMPY f71 = f71, f32
  7007. ;;
  7008. FNMA f72 = f64, f33, f72
  7009. FNMA f76 = f68, f33, f76
  7010. FNMA f73 = f65, f33, f73
  7011. FNMA f77 = f69, f33, f77
  7012. FNMA f74 = f66, f33, f74
  7013. FNMA f78 = f70, f33, f78
  7014. FNMA f75 = f67, f33, f75
  7015. FNMA f79 = f71, f33, f79
  7016. ;;
  7017. FNMA f80 = f64, f34, f80
  7018. FNMA f84 = f68, f34, f84
  7019. FNMA f81 = f65, f34, f81
  7020. FNMA f85 = f69, f34, f85
  7021. FNMA f82 = f66, f34, f82
  7022. FNMA f86 = f70, f34, f86
  7023. FNMA f83 = f67, f34, f83
  7024. FNMA f87 = f71, f34, f87
  7025. ;;
  7026. FNMA f88 = f64, f35, f88
  7027. FNMA f92 = f68, f35, f92
  7028. FNMA f89 = f65, f35, f89
  7029. FNMA f93 = f69, f35, f93
  7030. FNMA f90 = f66, f35, f90
  7031. FNMA f94 = f70, f35, f94
  7032. FNMA f91 = f67, f35, f91
  7033. FNMA f95 = f71, f35, f95
  7034. ;;
  7035. FMPY f72 = f72, f36
  7036. FMPY f76 = f76, f36
  7037. FMPY f73 = f73, f36
  7038. FMPY f77 = f77, f36
  7039. FMPY f74 = f74, f36
  7040. FMPY f78 = f78, f36
  7041. FMPY f75 = f75, f36
  7042. FMPY f79 = f79, f36
  7043. ;;
  7044. FNMA f80 = f72, f37, f80
  7045. FNMA f84 = f76, f37, f84
  7046. FNMA f81 = f73, f37, f81
  7047. FNMA f85 = f77, f37, f85
  7048. FNMA f82 = f74, f37, f82
  7049. FNMA f86 = f78, f37, f86
  7050. FNMA f83 = f75, f37, f83
  7051. FNMA f87 = f79, f37, f87
  7052. ;;
  7053. FNMA f88 = f72, f38, f88
  7054. FNMA f92 = f76, f38, f92
  7055. FNMA f89 = f73, f38, f89
  7056. FNMA f93 = f77, f38, f93
  7057. FNMA f90 = f74, f38, f90
  7058. FNMA f94 = f78, f38, f94
  7059. FNMA f91 = f75, f38, f91
  7060. FNMA f95 = f79, f38, f95
  7061. ;;
  7062. FMPY f80 = f80, f39
  7063. FMPY f84 = f84, f39
  7064. FMPY f81 = f81, f39
  7065. FMPY f85 = f85, f39
  7066. FMPY f82 = f82, f39
  7067. FMPY f86 = f86, f39
  7068. FMPY f83 = f83, f39
  7069. FMPY f87 = f87, f39
  7070. ;;
  7071. FNMA f88 = f80, f40, f88
  7072. FNMA f92 = f84, f40, f92
  7073. FNMA f89 = f81, f40, f89
  7074. FNMA f93 = f85, f40, f93
  7075. FNMA f90 = f82, f40, f90
  7076. FNMA f94 = f86, f40, f94
  7077. FNMA f91 = f83, f40, f91
  7078. FNMA f95 = f87, f40, f95
  7079. ;;
  7080. FMPY f88 = f88, f41
  7081. FMPY f92 = f92, f41
  7082. FMPY f89 = f89, f41
  7083. FMPY f93 = f93, f41
  7084. FMPY f90 = f90, f41
  7085. FMPY f94 = f94, f41
  7086. FMPY f91 = f91, f41
  7087. FMPY f95 = f95, f41
  7088. ;;
  7089. STFD [AOFFSET] = f64, SIZE
  7090. STFD [AOFFSET2] = f68, SIZE
  7091. ;;
  7092. STFD [AOFFSET] = f65, SIZE
  7093. STFD [AOFFSET2] = f69, SIZE
  7094. ;;
  7095. STFD [AOFFSET] = f66, SIZE
  7096. STFD [AOFFSET2] = f70, SIZE
  7097. ;;
  7098. STFD [AOFFSET] = f67, 5 * SIZE
  7099. STFD [AOFFSET2] = f71, 5 * SIZE
  7100. ;;
  7101. STFD [AOFFSET] = f72, SIZE
  7102. STFD [AOFFSET2] = f76, SIZE
  7103. ;;
  7104. STFD [AOFFSET] = f73, SIZE
  7105. STFD [AOFFSET2] = f77, SIZE
  7106. ;;
  7107. STFD [AOFFSET] = f74, SIZE
  7108. STFD [AOFFSET2] = f78, SIZE
  7109. ;;
  7110. STFD [AOFFSET] = f75, 5 * SIZE
  7111. STFD [AOFFSET2] = f79, 5 * SIZE
  7112. ;;
  7113. STFD [AOFFSET] = f80, SIZE
  7114. STFD [AOFFSET2] = f84, SIZE
  7115. ;;
  7116. STFD [AOFFSET] = f81, SIZE
  7117. STFD [AOFFSET2] = f85, SIZE
  7118. ;;
  7119. STFD [AOFFSET] = f82, SIZE
  7120. STFD [AOFFSET2] = f86, SIZE
  7121. ;;
  7122. STFD [AOFFSET] = f83, 5 * SIZE
  7123. STFD [AOFFSET2] = f87, 5 * SIZE
  7124. ;;
  7125. STFD [AOFFSET] = f88, SIZE
  7126. STFD [AOFFSET2] = f92, SIZE
  7127. ;;
  7128. STFD [AOFFSET] = f89, SIZE
  7129. STFD [AOFFSET2] = f93, SIZE
  7130. ;;
  7131. STFD [AOFFSET] = f90, SIZE
  7132. STFD [AOFFSET2] = f94, SIZE
  7133. ;;
  7134. STFD [AOFFSET] = f91, -27 * SIZE
  7135. STFD [AOFFSET2] = f95, -27 * SIZE
  7136. ;;
  7137. #endif
  7138. adds C9 = 4 * SIZE, C1
  7139. ;;
  7140. { .mmf
  7141. STFD [C1 ] = f64, SIZE
  7142. STFD [C9 ] = f68, SIZE
  7143. mov f64 = f0
  7144. }
  7145. ;;
  7146. { .mmi
  7147. STFD [C1 ] = f65, SIZE
  7148. STFD [C9 ] = f69, SIZE
  7149. adds C10 = 4 * SIZE, C2
  7150. }
  7151. ;;
  7152. { .mmi
  7153. STFD [C1 ] = f66, SIZE
  7154. STFD [C9 ] = f70, SIZE
  7155. }
  7156. ;;
  7157. { .mmi
  7158. STFD [C1 ] = f67, 5 * SIZE
  7159. STFD [C9 ] = f71
  7160. adds C11 = 4 * SIZE, C3
  7161. }
  7162. ;;
  7163. { .mmf
  7164. STFD [C2 ] = f72, SIZE
  7165. STFD [C10] = f76, SIZE
  7166. mov f72 = f0
  7167. }
  7168. ;;
  7169. { .mmi
  7170. STFD [C2 ] = f73, SIZE
  7171. STFD [C10] = f77, SIZE
  7172. }
  7173. ;;
  7174. { .mmi
  7175. STFD [C2 ] = f74, SIZE
  7176. STFD [C10] = f78, SIZE
  7177. adds C12 = 4 * SIZE, C4
  7178. }
  7179. ;;
  7180. { .mmi
  7181. STFD [C2 ] = f75, 5 * SIZE
  7182. STFD [C10] = f79
  7183. }
  7184. ;;
  7185. { .mmf
  7186. STFD [C3 ] = f80, SIZE
  7187. STFD [C11] = f84, SIZE
  7188. }
  7189. ;;
  7190. { .mmi
  7191. STFD [C3 ] = f81, SIZE
  7192. STFD [C11] = f85, SIZE
  7193. }
  7194. ;;
  7195. { .mmi
  7196. STFD [C3 ] = f82, SIZE
  7197. STFD [C11] = f86, SIZE
  7198. }
  7199. ;;
  7200. { .mmi
  7201. STFD [C3 ] = f83, 5 * SIZE
  7202. STFD [C11] = f87
  7203. }
  7204. ;;
  7205. { .mmf
  7206. STFD [C4 ] = f88, SIZE
  7207. STFD [C12] = f92, SIZE
  7208. }
  7209. ;;
  7210. { .mmi
  7211. STFD [C4 ] = f89, SIZE
  7212. STFD [C12] = f93, SIZE
  7213. }
  7214. ;;
  7215. { .mmi
  7216. STFD [C4 ] = f90, SIZE
  7217. STFD [C12] = f94, SIZE
  7218. }
  7219. ;;
  7220. { .mmi
  7221. STFD [C4 ] = f91, 5 * SIZE
  7222. STFD [C12] = f95
  7223. cmp.ne p6, p0 = 1, I
  7224. }
  7225. ;;
  7226. adds I = -1, I
  7227. ;;
  7228. { .mmi
  7229. sub L = K, KK
  7230. }
  7231. ;;
  7232. { .mmi
  7233. shladd L = L, BASE_SHIFT, r0
  7234. }
  7235. ;;
  7236. ;;
  7237. { .mmi
  7238. shladd AOFFSET = L, 3, AOFFSET
  7239. }
  7240. ;;
  7241. { .mmi
  7242. shladd BOFFSET = L, 2, BOFFSET
  7243. }
  7244. ;;
  7245. { .mmi
  7246. #ifdef LT
  7247. adds KK = 8, KK
  7248. #else
  7249. nop __LINE__
  7250. #endif
  7251. }
  7252. ;;
  7253. { .mmi
  7254. mov L = KK
  7255. }
  7256. ;;
  7257. mov f64 = f0
  7258. mov f72 = f0
  7259. mov f80 = f0
  7260. mov f88 = f0
  7261. mov f65 = f0
  7262. mov f73 = f0
  7263. mov f81 = f0
  7264. mov f89 = f0
  7265. { .mmb
  7266. (p6) br.cond.dptk .L052
  7267. }
  7268. ;;
  7269. .align 8
  7270. .L060:
  7271. tbit.z p6, p7 = M, 2
  7272. (p6) br.cond.dptk .L070
  7273. ;;
  7274. { .mib
  7275. mov L = KK
  7276. }
  7277. ;;
  7278. { .mmi
  7279. cmp.ne p7, p0 = r0, L
  7280. adds BOFFSET = 0 * SIZE, B
  7281. }
  7282. ;;
  7283. { .mmf
  7284. (p7) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  7285. mov f65 = f0
  7286. }
  7287. ;;
  7288. { .mfi
  7289. adds L = 1, L
  7290. }
  7291. { .mfi
  7292. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  7293. cmp.eq p3, p0 = r0, r0
  7294. }
  7295. ;;
  7296. { .mfi
  7297. (p7) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  7298. tbit.z p12, p0 = L, 0
  7299. }
  7300. { .mfi
  7301. shr L = L, 1
  7302. }
  7303. ;;
  7304. { .mfi
  7305. adds L = -1, L
  7306. }
  7307. ;;
  7308. { .mfi
  7309. cmp.eq p6, p0 = -1, L
  7310. }
  7311. ;;
  7312. { .mmf
  7313. (p7) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  7314. }
  7315. { .mfi
  7316. mov ar.lc = L
  7317. }
  7318. ;;
  7319. mov f66 = f0
  7320. mov f67 = f0
  7321. mov f74 = f0
  7322. mov f75 = f0
  7323. mov f82 = f0
  7324. mov f83 = f0
  7325. mov f90 = f0
  7326. mov f91 = f0
  7327. ;;
  7328. { .mmf
  7329. (p7) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  7330. }
  7331. { .mfb
  7332. (p6) br.cond.dpnt .L068
  7333. }
  7334. ;;
  7335. .align 8
  7336. .L062:
  7337. { .mfi
  7338. lfetch.nt1 [PREA], 8 * SIZE
  7339. FMA f64 = f32, f48, f64 // A1 * B1
  7340. cmp.ne p4, p5 = 0, L
  7341. }
  7342. { .mfi
  7343. nop __LINE__
  7344. FMA f72 = f32, f49, f72 // A1 * B2
  7345. (p12) cmp.ne p3, p0 = 0, L
  7346. }
  7347. ;;
  7348. { .mfi
  7349. lfetch.nt1 [PREB], 8 * SIZE
  7350. FMA f80 = f32, f50, f80 // A1 * B3
  7351. (p5) adds C9 = 2 * SIZE, C1
  7352. }
  7353. { .mfi
  7354. nop __LINE__
  7355. FMA f88 = f32, f51, f88 // A1 * B4
  7356. (p5) adds C10 = 2 * SIZE, C2
  7357. }
  7358. ;;
  7359. { .mfi
  7360. (p3) LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  7361. FMA f65 = f33, f48, f65 // A2 * B1
  7362. (p5) adds C11 = 2 * SIZE, C3
  7363. }
  7364. { .mfi
  7365. nop __LINE__
  7366. FMA f73 = f33, f49, f73 // A2 * B2
  7367. (p5) adds C12 = 2 * SIZE, C4
  7368. }
  7369. ;;
  7370. { .mfb
  7371. (p3) LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  7372. FMA f81 = f33, f50, f81 // A2 * B3
  7373. nop __LINE__
  7374. }
  7375. { .mfb
  7376. nop __LINE__
  7377. FMA f89 = f33, f51, f89 // A2 * B4
  7378. nop __LINE__
  7379. }
  7380. ;;
  7381. { .mfb
  7382. (p3) LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  7383. FMA f66 = f34, f48, f66 // A3 * B1
  7384. nop __LINE__
  7385. }
  7386. { .mfb
  7387. nop __LINE__
  7388. FMA f74 = f34, f49, f74 // A3 * B2
  7389. nop __LINE__
  7390. }
  7391. ;;
  7392. { .mfb
  7393. (p3) LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  7394. FMA f82 = f34, f50, f82 // A3 * B3
  7395. nop __LINE__
  7396. }
  7397. { .mfb
  7398. nop __LINE__
  7399. FMA f90 = f34, f51, f90 // A3 * B4
  7400. nop __LINE__
  7401. }
  7402. ;;
  7403. { .mfb
  7404. (p4) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  7405. FMA f67 = f35, f48, f67 // A4 * B1
  7406. }
  7407. { .mfb
  7408. nop __LINE__
  7409. FMA f75 = f35, f49, f75 // A4 * B2
  7410. nop __LINE__
  7411. }
  7412. { .mfb
  7413. (p4) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  7414. FMA f83 = f35, f50, f83 // A4 * B3
  7415. nop __LINE__
  7416. }
  7417. { .mfb
  7418. nop __LINE__
  7419. FMA f91 = f35, f51, f91 // A4 * B4
  7420. nop __LINE__
  7421. }
  7422. ;;
  7423. { .mfb
  7424. (p4) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  7425. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  7426. nop __LINE__
  7427. }
  7428. { .mfb
  7429. nop __LINE__
  7430. (p3) FMA f72 = f40, f57, f72 // A1 * B2
  7431. nop __LINE__
  7432. }
  7433. ;;
  7434. { .mfb
  7435. (p4) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  7436. (p3) FMA f80 = f40, f58, f80 // A1 * B3
  7437. nop __LINE__
  7438. }
  7439. { .mfb
  7440. nop __LINE__
  7441. (p3) FMA f88 = f40, f59, f88 // A1 * B4
  7442. nop __LINE__
  7443. }
  7444. ;;
  7445. { .mfb
  7446. nop __LINE__
  7447. (p3) FMA f65 = f41, f56, f65 // A2 * B1
  7448. nop __LINE__
  7449. }
  7450. { .mfb
  7451. nop __LINE__
  7452. (p3) FMA f73 = f41, f57, f73 // A2 * B2
  7453. nop __LINE__
  7454. }
  7455. ;;
  7456. { .mfb
  7457. nop __LINE__
  7458. (p3) FMA f81 = f41, f58, f81 // A2 * B3
  7459. nop __LINE__
  7460. }
  7461. { .mfb
  7462. nop __LINE__
  7463. (p3) FMA f89 = f41, f59, f89 // A2 * B4
  7464. nop __LINE__
  7465. }
  7466. ;;
  7467. { .mfb
  7468. nop __LINE__
  7469. (p3) FMA f66 = f42, f56, f66 // A3 * B1
  7470. nop __LINE__
  7471. }
  7472. { .mfb
  7473. nop __LINE__
  7474. (p3) FMA f74 = f42, f57, f74 // A3 * B2
  7475. nop __LINE__
  7476. }
  7477. ;;
  7478. { .mfb
  7479. nop __LINE__
  7480. (p3) FMA f82 = f42, f58, f82 // A3 * B3
  7481. nop __LINE__
  7482. }
  7483. { .mfb
  7484. nop __LINE__
  7485. (p3) FMA f90 = f42, f59, f90 // A3 * B4
  7486. nop __LINE__
  7487. }
  7488. ;;
  7489. { .mfb
  7490. nop __LINE__
  7491. (p3) FMA f67 = f43, f56, f67 // A4 * B1
  7492. nop __LINE__
  7493. }
  7494. { .mfb
  7495. nop __LINE__
  7496. (p3) FMA f75 = f43, f57, f75 // A4 * B2
  7497. nop __LINE__
  7498. }
  7499. ;;
  7500. { .mfi
  7501. nop __LINE__
  7502. (p3) FMA f83 = f43, f58, f83 // A4 * B3
  7503. adds L = -1, L
  7504. }
  7505. { .mfb
  7506. nop __LINE__
  7507. (p3) FMA f91 = f43, f59, f91 // A4 * B4
  7508. br.cloop.sptk.few .L062
  7509. }
  7510. ;;
  7511. .align 8
  7512. .L068:
  7513. adds AOFFSET2 = 4 * SIZE, AOFFSET
  7514. adds BOFFSET2 = 4 * SIZE, BOFFSET
  7515. ;;
  7516. #ifdef LT
  7517. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  7518. ;;
  7519. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  7520. ;;
  7521. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  7522. ;;
  7523. LDFPD f38, f39 = [BOFFSET], 2 * SIZE
  7524. ;;
  7525. LDFPD f40, f41 = [BOFFSET], 2 * SIZE
  7526. ;;
  7527. LDFPD f42, f43 = [BOFFSET], 2 * SIZE
  7528. ;;
  7529. LDFPD f44, f45 = [BOFFSET], 2 * SIZE
  7530. ;;
  7531. LDFPD f46, f47 = [BOFFSET]
  7532. adds BOFFSET = -14 * SIZE, BOFFSET
  7533. ;;
  7534. FSUB f64 = f32, f64
  7535. FSUB f72 = f33, f72
  7536. FSUB f80 = f34, f80
  7537. FSUB f88 = f35, f88
  7538. ;;
  7539. FSUB f65 = f36, f65
  7540. FSUB f73 = f37, f73
  7541. FSUB f81 = f38, f81
  7542. FSUB f89 = f39, f89
  7543. ;;
  7544. FSUB f66 = f40, f66
  7545. FSUB f74 = f41, f74
  7546. FSUB f82 = f42, f82
  7547. FSUB f90 = f43, f90
  7548. ;;
  7549. FSUB f67 = f44, f67
  7550. FSUB f75 = f45, f75
  7551. FSUB f83 = f46, f83
  7552. FSUB f91 = f47, f91
  7553. ;;
  7554. #else
  7555. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  7556. ;;
  7557. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  7558. ;;
  7559. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  7560. ;;
  7561. LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  7562. ;;
  7563. LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  7564. ;;
  7565. LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  7566. ;;
  7567. LDFPD f44, f45 = [AOFFSET], 2 * SIZE
  7568. ;;
  7569. LDFPD f46, f47 = [AOFFSET]
  7570. adds AOFFSET = -14 * SIZE, AOFFSET
  7571. ;;
  7572. FSUB f64 = f32, f64
  7573. FSUB f65 = f33, f65
  7574. FSUB f66 = f34, f66
  7575. FSUB f67 = f35, f67
  7576. FSUB f72 = f36, f72
  7577. FSUB f73 = f37, f73
  7578. FSUB f74 = f38, f74
  7579. FSUB f75 = f39, f75
  7580. FSUB f80 = f40, f80
  7581. FSUB f81 = f41, f81
  7582. FSUB f82 = f42, f82
  7583. FSUB f83 = f43, f83
  7584. FSUB f88 = f44, f88
  7585. FSUB f89 = f45, f89
  7586. FSUB f90 = f46, f90
  7587. FSUB f91 = f47, f91
  7588. ;;
  7589. #endif
  7590. #ifdef LT
  7591. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  7592. ;;
  7593. LDFPD f34, f35 = [AOFFSET]
  7594. adds AOFFSET = 3 * SIZE, AOFFSET
  7595. ;;
  7596. LDFD f36 = [AOFFSET], 1 * SIZE
  7597. ;;
  7598. LDFPD f37, f38 = [AOFFSET]
  7599. adds AOFFSET = 4 * SIZE, AOFFSET
  7600. ;;
  7601. LDFPD f39, f40 = [AOFFSET]
  7602. adds AOFFSET = 5 * SIZE, AOFFSET
  7603. ;;
  7604. LDFD f41 = [AOFFSET], -15 * SIZE
  7605. ;;
  7606. FMPY f64 = f64, f32
  7607. FMPY f72 = f72, f32
  7608. FMPY f80 = f80, f32
  7609. FMPY f88 = f88, f32
  7610. ;;
  7611. FNMA f65 = f64, f33, f65
  7612. FNMA f73 = f72, f33, f73
  7613. FNMA f81 = f80, f33, f81
  7614. FNMA f89 = f88, f33, f89
  7615. ;;
  7616. FNMA f66 = f64, f34, f66
  7617. FNMA f74 = f72, f34, f74
  7618. FNMA f82 = f80, f34, f82
  7619. FNMA f90 = f88, f34, f90
  7620. ;;
  7621. FNMA f67 = f64, f35, f67
  7622. FNMA f75 = f72, f35, f75
  7623. FNMA f83 = f80, f35, f83
  7624. FNMA f91 = f88, f35, f91
  7625. ;;
  7626. FMPY f65 = f65, f36
  7627. FMPY f73 = f73, f36
  7628. FMPY f81 = f81, f36
  7629. FMPY f89 = f89, f36
  7630. ;;
  7631. FNMA f66 = f65, f37, f66
  7632. FNMA f74 = f73, f37, f74
  7633. FNMA f82 = f81, f37, f82
  7634. FNMA f90 = f89, f37, f90
  7635. ;;
  7636. FNMA f67 = f65, f38, f67
  7637. FNMA f75 = f73, f38, f75
  7638. FNMA f83 = f81, f38, f83
  7639. FNMA f91 = f89, f38, f91
  7640. ;;
  7641. FMPY f66 = f66, f39
  7642. FMPY f74 = f74, f39
  7643. FMPY f82 = f82, f39
  7644. FMPY f90 = f90, f39
  7645. ;;
  7646. FNMA f67 = f66, f40, f67
  7647. FNMA f75 = f74, f40, f75
  7648. FNMA f83 = f82, f40, f83
  7649. FNMA f91 = f90, f40, f91
  7650. ;;
  7651. FMPY f67 = f67, f41
  7652. FMPY f75 = f75, f41
  7653. FMPY f83 = f83, f41
  7654. FMPY f91 = f91, f41
  7655. ;;
  7656. STFD [BOFFSET] = f64, SIZE
  7657. STFD [BOFFSET2] = f65, SIZE
  7658. ;;
  7659. STFD [BOFFSET] = f72, SIZE
  7660. STFD [BOFFSET2] = f73, SIZE
  7661. ;;
  7662. STFD [BOFFSET] = f80, SIZE
  7663. STFD [BOFFSET2] = f81, SIZE
  7664. ;;
  7665. STFD [BOFFSET] = f88, 5 * SIZE
  7666. STFD [BOFFSET2] = f89, 5 * SIZE
  7667. ;;
  7668. STFD [BOFFSET] = f66, SIZE
  7669. STFD [BOFFSET2] = f67, SIZE
  7670. ;;
  7671. STFD [BOFFSET] = f74, SIZE
  7672. STFD [BOFFSET2] = f75, SIZE
  7673. ;;
  7674. STFD [BOFFSET] = f82, SIZE
  7675. STFD [BOFFSET2] = f83, SIZE
  7676. ;;
  7677. STFD [BOFFSET] = f90, -11 * SIZE
  7678. STFD [BOFFSET2] = f91, -11 * SIZE
  7679. ;;
  7680. #endif
  7681. #ifdef RN
  7682. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  7683. ;;
  7684. LDFPD f34, f35 = [BOFFSET]
  7685. adds BOFFSET = 3 * SIZE, BOFFSET
  7686. ;;
  7687. LDFD f36 = [BOFFSET], 1 * SIZE
  7688. ;;
  7689. LDFPD f37, f38 = [BOFFSET]
  7690. adds BOFFSET = 4 * SIZE, BOFFSET
  7691. ;;
  7692. LDFPD f39, f40 = [BOFFSET]
  7693. adds BOFFSET = 5 * SIZE, BOFFSET
  7694. ;;
  7695. LDFD f41 = [BOFFSET], -15 * SIZE
  7696. ;;
  7697. FMPY f64 = f64, f32
  7698. FMPY f65 = f65, f32
  7699. FMPY f66 = f66, f32
  7700. FMPY f67 = f67, f32
  7701. ;;
  7702. FNMA f72 = f64, f33, f72
  7703. FNMA f73 = f65, f33, f73
  7704. FNMA f74 = f66, f33, f74
  7705. FNMA f75 = f67, f33, f75
  7706. ;;
  7707. FNMA f80 = f64, f34, f80
  7708. FNMA f81 = f65, f34, f81
  7709. FNMA f82 = f66, f34, f82
  7710. FNMA f83 = f67, f34, f83
  7711. ;;
  7712. FNMA f88 = f64, f35, f88
  7713. FNMA f89 = f65, f35, f89
  7714. FNMA f90 = f66, f35, f90
  7715. FNMA f91 = f67, f35, f91
  7716. ;;
  7717. FMPY f72 = f72, f36
  7718. FMPY f73 = f73, f36
  7719. FMPY f74 = f74, f36
  7720. FMPY f75 = f75, f36
  7721. ;;
  7722. FNMA f80 = f72, f37, f80
  7723. FNMA f81 = f73, f37, f81
  7724. FNMA f82 = f74, f37, f82
  7725. FNMA f83 = f75, f37, f83
  7726. ;;
  7727. FNMA f88 = f72, f38, f88
  7728. FNMA f89 = f73, f38, f89
  7729. FNMA f90 = f74, f38, f90
  7730. FNMA f91 = f75, f38, f91
  7731. ;;
  7732. FMPY f80 = f80, f39
  7733. FMPY f81 = f81, f39
  7734. FMPY f82 = f82, f39
  7735. FMPY f83 = f83, f39
  7736. ;;
  7737. FNMA f88 = f80, f40, f88
  7738. FNMA f89 = f81, f40, f89
  7739. FNMA f90 = f82, f40, f90
  7740. FNMA f91 = f83, f40, f91
  7741. ;;
  7742. FMPY f88 = f88, f41
  7743. FMPY f89 = f89, f41
  7744. FMPY f90 = f90, f41
  7745. FMPY f91 = f91, f41
  7746. ;;
  7747. STFD [AOFFSET] = f64, SIZE
  7748. STFD [AOFFSET2] = f72, SIZE
  7749. ;;
  7750. STFD [AOFFSET] = f65, SIZE
  7751. STFD [AOFFSET2] = f73, SIZE
  7752. ;;
  7753. STFD [AOFFSET] = f66, SIZE
  7754. STFD [AOFFSET2] = f74, SIZE
  7755. ;;
  7756. STFD [AOFFSET] = f67, 5 * SIZE
  7757. STFD [AOFFSET2] = f75, 5 * SIZE
  7758. ;;
  7759. STFD [AOFFSET] = f80, SIZE
  7760. STFD [AOFFSET2] = f88, SIZE
  7761. ;;
  7762. STFD [AOFFSET] = f81, SIZE
  7763. STFD [AOFFSET2] = f89, SIZE
  7764. ;;
  7765. STFD [AOFFSET] = f82, SIZE
  7766. STFD [AOFFSET2] = f90, SIZE
  7767. ;;
  7768. STFD [AOFFSET] = f83, -11 * SIZE
  7769. STFD [AOFFSET2] = f91, -11 * SIZE
  7770. ;;
  7771. #endif
  7772. { .mmf
  7773. STFD [C1 ] = f64, SIZE
  7774. mov f64 = f0
  7775. }
  7776. ;;
  7777. { .mmi
  7778. STFD [C1 ] = f65, SIZE
  7779. }
  7780. ;;
  7781. { .mmi
  7782. STFD [C1 ] = f66, SIZE
  7783. }
  7784. ;;
  7785. { .mmi
  7786. STFD [C1 ] = f67, SIZE
  7787. }
  7788. ;;
  7789. { .mmf
  7790. STFD [C2 ] = f72, SIZE
  7791. mov f72 = f0
  7792. }
  7793. ;;
  7794. { .mmi
  7795. STFD [C2 ] = f73, SIZE
  7796. }
  7797. ;;
  7798. { .mmi
  7799. STFD [C2 ] = f74, SIZE
  7800. }
  7801. ;;
  7802. { .mmi
  7803. STFD [C2 ] = f75, SIZE
  7804. }
  7805. ;;
  7806. { .mmf
  7807. STFD [C3 ] = f80, SIZE
  7808. mov f80 = f0
  7809. }
  7810. ;;
  7811. { .mmi
  7812. STFD [C3 ] = f81, SIZE
  7813. }
  7814. ;;
  7815. { .mmi
  7816. STFD [C3 ] = f82, SIZE
  7817. }
  7818. ;;
  7819. { .mmi
  7820. STFD [C3 ] = f83, SIZE
  7821. }
  7822. ;;
  7823. { .mmf
  7824. STFD [C4 ] = f88, SIZE
  7825. mov f88 = f0
  7826. }
  7827. ;;
  7828. { .mmi
  7829. STFD [C4 ] = f89, SIZE
  7830. }
  7831. ;;
  7832. { .mmi
  7833. STFD [C4 ] = f90, SIZE
  7834. }
  7835. ;;
  7836. { .mmi
  7837. STFD [C4 ] = f91, SIZE
  7838. nop __LINE__
  7839. }
  7840. ;;
  7841. mov f65 = f0
  7842. ;;
  7843. mov f73 = f0
  7844. ;;
  7845. { .mmi
  7846. sub L = K, KK
  7847. }
  7848. ;;
  7849. { .mmf
  7850. mov f81 = f0
  7851. }
  7852. ;;
  7853. { .mmi
  7854. shladd L = L, BASE_SHIFT, r0
  7855. }
  7856. ;;
  7857. { .mmi
  7858. shladd AOFFSET = L, 2, AOFFSET
  7859. }
  7860. ;;
  7861. { .mmi
  7862. shladd BOFFSET = L, 2, BOFFSET
  7863. }
  7864. ;;
  7865. { .mmf
  7866. mov f89 = f0
  7867. }
  7868. ;;
  7869. { .mmi
  7870. #ifdef LT
  7871. adds KK = 4, KK
  7872. #else
  7873. nop __LINE__
  7874. #endif
  7875. }
  7876. ;;
  7877. .align 8
  7878. .L070:
  7879. tbit.z p6,p7 = M, 1
  7880. (p6) br.cond.dptk .L080
  7881. ;;
  7882. { .mib
  7883. mov L = KK
  7884. }
  7885. ;;
  7886. { .mmi
  7887. cmp.ne p7, p0 = r0, L
  7888. adds BOFFSET = 0 * SIZE, B
  7889. }
  7890. ;;
  7891. { .mmf
  7892. (p7) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  7893. setf.d f73 = r0
  7894. mov f65 = f0
  7895. }
  7896. ;;
  7897. { .mfi
  7898. mov f81 = f0
  7899. adds L = 1, L
  7900. }
  7901. { .mfi
  7902. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  7903. mov f89 = f0
  7904. cmp.eq p3, p0 = r0, r0
  7905. }
  7906. ;;
  7907. { .mfi
  7908. (p7) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  7909. tbit.z p12, p0 = L, 0
  7910. }
  7911. { .mfi
  7912. shr L = L, 1
  7913. }
  7914. ;;
  7915. { .mmf
  7916. adds L = -1, L
  7917. }
  7918. ;;
  7919. { .mmf
  7920. cmp.eq p6, p0 = -1, L
  7921. }
  7922. ;;
  7923. { .mib
  7924. (p7) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  7925. mov ar.lc = L
  7926. (p6) br.cond.dpnt .L078
  7927. }
  7928. ;;
  7929. .align 8
  7930. .L072:
  7931. { .mfb
  7932. lfetch.nt1 [PREA], 4 * SIZE
  7933. FMA f64 = f32, f48, f64 // A1 * B1
  7934. nop __LINE__
  7935. }
  7936. { .mfi
  7937. nop __LINE__
  7938. FMA f72 = f32, f49, f72 // A1 * B2
  7939. (p12) cmp.ne p3, p0 = 0, L
  7940. }
  7941. ;;
  7942. { .mfi
  7943. lfetch.nt1 [PREB], 8 * SIZE
  7944. FMA f80 = f32, f50, f80 // A1 * B3
  7945. cmp.ne p4, p5 = 0, L
  7946. }
  7947. { .mfb
  7948. nop __LINE__
  7949. FMA f88 = f32, f51, f88 // A1 * B4
  7950. nop __LINE__
  7951. }
  7952. ;;
  7953. { .mfi
  7954. (p3) LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  7955. FMA f65 = f33, f48, f65 // A2 * B1
  7956. }
  7957. { .mfi
  7958. nop __LINE__
  7959. FMA f73 = f33, f49, f73 // A2 * B2
  7960. }
  7961. ;;
  7962. { .mfi
  7963. (p3) LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  7964. FMA f81 = f33, f50, f81 // A2 * B3
  7965. }
  7966. { .mmf
  7967. nop __LINE__
  7968. nop __LINE__
  7969. FMA f89 = f33, f51, f89 // A2 * B4
  7970. }
  7971. ;;
  7972. { .mfb
  7973. (p3) LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  7974. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  7975. nop __LINE__
  7976. }
  7977. { .mmf
  7978. nop __LINE__
  7979. nop __LINE__
  7980. (p3) FMA f72 = f40, f57, f72 // A1 * B2
  7981. }
  7982. ;;
  7983. { .mfb
  7984. (p4) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  7985. (p3) FMA f80 = f40, f58, f80 // A1 * B3
  7986. nop __LINE__
  7987. }
  7988. { .mmf
  7989. nop __LINE__
  7990. nop __LINE__
  7991. (p3) FMA f88 = f40, f59, f88 // A1 * B4
  7992. }
  7993. ;;
  7994. { .mfb
  7995. (p4) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  7996. (p3) FMA f65 = f41, f56, f65 // A2 * B1
  7997. nop __LINE__
  7998. }
  7999. { .mfb
  8000. nop __LINE__
  8001. (p3) FMA f73 = f41, f57, f73 // A2 * B2
  8002. nop __LINE__
  8003. }
  8004. ;;
  8005. { .mfi
  8006. (p4) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  8007. (p3) FMA f81 = f41, f58, f81 // A2 * B3
  8008. adds L = -1, L
  8009. }
  8010. { .mfb
  8011. nop __LINE__
  8012. (p3) FMA f89 = f41, f59, f89 // A2 * B4
  8013. br.cloop.sptk.few .L072
  8014. }
  8015. ;;
  8016. .L078:
  8017. adds AOFFSET2 = 4 * SIZE, AOFFSET
  8018. adds BOFFSET2 = 4 * SIZE, BOFFSET
  8019. ;;
  8020. #ifdef LT
  8021. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  8022. ;;
  8023. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  8024. ;;
  8025. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  8026. ;;
  8027. LDFPD f38, f39 = [BOFFSET]
  8028. adds BOFFSET = -6 * SIZE, BOFFSET
  8029. ;;
  8030. FSUB f64 = f32, f64
  8031. FSUB f72 = f33, f72
  8032. FSUB f80 = f34, f80
  8033. FSUB f88 = f35, f88
  8034. FSUB f65 = f36, f65
  8035. FSUB f73 = f37, f73
  8036. FSUB f81 = f38, f81
  8037. FSUB f89 = f39, f89
  8038. ;;
  8039. #else
  8040. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  8041. ;;
  8042. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  8043. ;;
  8044. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  8045. ;;
  8046. LDFPD f38, f39 = [AOFFSET]
  8047. adds AOFFSET = -6 * SIZE, AOFFSET
  8048. ;;
  8049. FSUB f64 = f32, f64
  8050. FSUB f65 = f33, f65
  8051. FSUB f72 = f34, f72
  8052. FSUB f73 = f35, f73
  8053. FSUB f80 = f36, f80
  8054. FSUB f81 = f37, f81
  8055. FSUB f88 = f38, f88
  8056. FSUB f89 = f39, f89
  8057. ;;
  8058. #endif
  8059. #ifdef LT
  8060. LDFPD f32, f33 = [AOFFSET]
  8061. adds AOFFSET = 3 * SIZE, AOFFSET
  8062. ;;
  8063. LDFD f34 = [AOFFSET], - 3 * SIZE
  8064. ;;
  8065. FMPY f64 = f64, f32
  8066. FMPY f72 = f72, f32
  8067. FMPY f80 = f80, f32
  8068. FMPY f88 = f88, f32
  8069. ;;
  8070. FNMA f65 = f64, f33, f65
  8071. FNMA f73 = f72, f33, f73
  8072. FNMA f81 = f80, f33, f81
  8073. FNMA f89 = f88, f33, f89
  8074. ;;
  8075. FMPY f65 = f65, f34
  8076. FMPY f73 = f73, f34
  8077. FMPY f81 = f81, f34
  8078. FMPY f89 = f89, f34
  8079. ;;
  8080. STFD [BOFFSET] = f64, SIZE
  8081. STFD [BOFFSET2] = f65, SIZE
  8082. ;;
  8083. STFD [BOFFSET] = f72, SIZE
  8084. STFD [BOFFSET2] = f73, SIZE
  8085. ;;
  8086. STFD [BOFFSET] = f80, SIZE
  8087. STFD [BOFFSET2] = f81, SIZE
  8088. ;;
  8089. STFD [BOFFSET] = f88, -3 * SIZE
  8090. STFD [BOFFSET2] = f89, -3 * SIZE
  8091. ;;
  8092. #endif
  8093. #ifdef RN
  8094. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  8095. ;;
  8096. LDFPD f34, f35 = [BOFFSET]
  8097. adds BOFFSET = 3 * SIZE, BOFFSET
  8098. ;;
  8099. LDFD f36 = [BOFFSET], 1 * SIZE
  8100. ;;
  8101. LDFPD f37, f38 = [BOFFSET]
  8102. adds BOFFSET = 4 * SIZE, BOFFSET
  8103. ;;
  8104. LDFPD f39, f40 = [BOFFSET]
  8105. adds BOFFSET = 5 * SIZE, BOFFSET
  8106. ;;
  8107. LDFD f41 = [BOFFSET], -15 * SIZE
  8108. ;;
  8109. FMPY f64 = f64, f32
  8110. FMPY f65 = f65, f32
  8111. ;;
  8112. FNMA f72 = f64, f33, f72
  8113. FNMA f73 = f65, f33, f73
  8114. ;;
  8115. FNMA f80 = f64, f34, f80
  8116. FNMA f81 = f65, f34, f81
  8117. ;;
  8118. FNMA f88 = f64, f35, f88
  8119. FNMA f89 = f65, f35, f89
  8120. ;;
  8121. FMPY f72 = f72, f36
  8122. FMPY f73 = f73, f36
  8123. ;;
  8124. FNMA f80 = f72, f37, f80
  8125. FNMA f81 = f73, f37, f81
  8126. ;;
  8127. FNMA f88 = f72, f38, f88
  8128. FNMA f89 = f73, f38, f89
  8129. ;;
  8130. FMPY f80 = f80, f39
  8131. FMPY f81 = f81, f39
  8132. ;;
  8133. FNMA f88 = f80, f40, f88
  8134. FNMA f89 = f81, f40, f89
  8135. ;;
  8136. FMPY f88 = f88, f41
  8137. FMPY f89 = f89, f41
  8138. ;;
  8139. STFD [AOFFSET] = f64, SIZE
  8140. STFD [AOFFSET2] = f80, SIZE
  8141. ;;
  8142. STFD [AOFFSET] = f65, SIZE
  8143. STFD [AOFFSET2] = f81, SIZE
  8144. ;;
  8145. STFD [AOFFSET] = f72, SIZE
  8146. STFD [AOFFSET2] = f88, SIZE
  8147. ;;
  8148. STFD [AOFFSET] = f73, -3 * SIZE
  8149. STFD [AOFFSET2] = f89, -3 * SIZE
  8150. ;;
  8151. #endif
  8152. STFD [C1 ] = f64, SIZE
  8153. mov f64 = f0
  8154. ;;
  8155. STFD [C1 ] = f65, SIZE
  8156. ;;
  8157. STFD [C2 ] = f72, SIZE
  8158. mov f72 = f0
  8159. ;;
  8160. STFD [C2 ] = f73, SIZE
  8161. ;;
  8162. STFD [C3 ] = f80, SIZE
  8163. mov f80 = f0
  8164. ;;
  8165. STFD [C3 ] = f81, SIZE
  8166. ;;
  8167. STFD [C4 ] = f88, SIZE
  8168. mov f88 = f0
  8169. ;;
  8170. STFD [C4 ] = f89, SIZE
  8171. ;;
  8172. mov f96 = f0
  8173. ;;
  8174. mov f104 = f0
  8175. ;;
  8176. sub L = K, KK
  8177. ;;
  8178. mov f112 = f0
  8179. ;;
  8180. { .mmi
  8181. shladd L = L, BASE_SHIFT, r0
  8182. }
  8183. ;;
  8184. { .mmi
  8185. shladd AOFFSET = L, 1, AOFFSET
  8186. }
  8187. ;;
  8188. { .mmi
  8189. shladd BOFFSET = L, 2, BOFFSET
  8190. }
  8191. ;;
  8192. { .mmf
  8193. mov f120 = f0
  8194. }
  8195. ;;
  8196. { .mmi
  8197. #ifdef LT
  8198. adds KK = 2, KK
  8199. #else
  8200. nop __LINE__
  8201. #endif
  8202. }
  8203. ;;
  8204. .align 8
  8205. .L080:
  8206. tbit.z p6,p7 = M, 0
  8207. (p6) br.cond.dptk .L089
  8208. { .mib
  8209. mov L = KK
  8210. }
  8211. ;;
  8212. { .mmi
  8213. cmp.ne p7, p0 = r0, L
  8214. adds BOFFSET = 0 * SIZE, B
  8215. }
  8216. ;;
  8217. { .mmf
  8218. (p7) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  8219. }
  8220. ;;
  8221. { .mmi
  8222. adds L = 1, L
  8223. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  8224. cmp.eq p3, p0 = r0, r0
  8225. }
  8226. ;;
  8227. { .mii
  8228. (p7) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  8229. tbit.z p12, p0 = L, 0
  8230. shr L = L, 1
  8231. }
  8232. ;;
  8233. { .mmi
  8234. adds L = -1, L
  8235. }
  8236. ;;
  8237. { .mmi
  8238. cmp.eq p6, p0 = -1, L
  8239. }
  8240. ;;
  8241. { .mib
  8242. (p7) LDFD f32 = [AOFFSET], 1 * SIZE
  8243. mov ar.lc = L
  8244. (p6) br.cond.dpnt .L088
  8245. }
  8246. ;;
  8247. .L082:
  8248. { .mfb
  8249. cmp.ne p4, p5 = 0, L
  8250. FMA f64 = f32, f48, f64 // A1 * B1
  8251. nop __LINE__
  8252. }
  8253. { .mfi
  8254. (p12) cmp.ne p3, p0 = 0, L
  8255. FMA f72 = f32, f49, f72 // A1 * B2
  8256. nop __LINE__
  8257. }
  8258. ;;
  8259. { .mfb
  8260. (p3) LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  8261. FMA f80 = f32, f50, f80 // A1 * B3
  8262. nop __LINE__
  8263. }
  8264. { .mfb
  8265. (p3) LDFD f40 = [AOFFSET], 1 * SIZE
  8266. FMA f88 = f32, f51, f88 // A1 * B4
  8267. nop __LINE__
  8268. }
  8269. ;;
  8270. { .mfb
  8271. (p3) LDFPD f58, f59 = [BOFFSET], 2 * SIZE
  8272. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  8273. nop __LINE__
  8274. }
  8275. { .mfb
  8276. nop __LINE__
  8277. (p3) FMA f72 = f40, f57, f72 // A1 * B2
  8278. nop __LINE__
  8279. }
  8280. ;;
  8281. { .mmf
  8282. (p4) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  8283. (p4) LDFD f32 = [AOFFSET], 1 * SIZE
  8284. (p3) FMA f80 = f40, f58, f80 // A1 * B3
  8285. }
  8286. { .mmf
  8287. nop __LINE__
  8288. nop __LINE__
  8289. (p3) FMA f88 = f40, f59, f88 // A1 * B4
  8290. }
  8291. ;;
  8292. { .mib
  8293. (p4) LDFPD f50, f51 = [BOFFSET], 2 * SIZE
  8294. nop __LINE__
  8295. nop __LINE__
  8296. }
  8297. { .mmb
  8298. nop __LINE__
  8299. adds L = -1, L
  8300. br.cloop.sptk.few .L082
  8301. }
  8302. ;;
  8303. .L088:
  8304. adds AOFFSET2 = 4 * SIZE, AOFFSET
  8305. adds BOFFSET2 = 4 * SIZE, BOFFSET
  8306. ;;
  8307. #ifdef LT
  8308. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  8309. ;;
  8310. LDFPD f34, f35 = [BOFFSET]
  8311. adds BOFFSET = -2 * SIZE, BOFFSET
  8312. ;;
  8313. FSUB f64 = f32, f64
  8314. FSUB f72 = f33, f72
  8315. FSUB f80 = f34, f80
  8316. FSUB f88 = f35, f88
  8317. ;;
  8318. #else
  8319. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  8320. ;;
  8321. LDFPD f34, f35 = [AOFFSET]
  8322. adds AOFFSET = -2 * SIZE, AOFFSET
  8323. ;;
  8324. FSUB f64 = f32, f64
  8325. FSUB f72 = f33, f72
  8326. FSUB f80 = f34, f80
  8327. FSUB f88 = f35, f88
  8328. ;;
  8329. #endif
  8330. #ifdef LT
  8331. LDFD f32 = [AOFFSET]
  8332. ;;
  8333. FMPY f64 = f64, f32
  8334. FMPY f72 = f72, f32
  8335. FMPY f80 = f80, f32
  8336. FMPY f88 = f88, f32
  8337. ;;
  8338. STFD [BOFFSET] = f64, SIZE
  8339. ;;
  8340. STFD [BOFFSET] = f72, SIZE
  8341. ;;
  8342. STFD [BOFFSET] = f80, SIZE
  8343. ;;
  8344. STFD [BOFFSET] = f88, -3 * SIZE
  8345. ;;
  8346. #endif
  8347. #ifdef RN
  8348. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  8349. ;;
  8350. LDFPD f34, f35 = [BOFFSET]
  8351. adds BOFFSET = 3 * SIZE, BOFFSET
  8352. ;;
  8353. LDFD f36 = [BOFFSET], 1 * SIZE
  8354. ;;
  8355. LDFPD f37, f38 = [BOFFSET]
  8356. adds BOFFSET = 4 * SIZE, BOFFSET
  8357. ;;
  8358. LDFPD f39, f40 = [BOFFSET]
  8359. adds BOFFSET = 5 * SIZE, BOFFSET
  8360. ;;
  8361. LDFD f41 = [BOFFSET], -15 * SIZE
  8362. FMPY f64 = f64, f32
  8363. ;;
  8364. FNMA f72 = f64, f33, f72
  8365. ;;
  8366. FNMA f80 = f64, f34, f80
  8367. ;;
  8368. FNMA f88 = f64, f35, f88
  8369. ;;
  8370. FMPY f72 = f72, f36
  8371. ;;
  8372. FNMA f80 = f72, f37, f80
  8373. ;;
  8374. FNMA f88 = f72, f38, f88
  8375. ;;
  8376. FMPY f80 = f80, f39
  8377. ;;
  8378. FNMA f88 = f80, f40, f88
  8379. ;;
  8380. FMPY f88 = f88, f41
  8381. ;;
  8382. STFD [AOFFSET] = f64, SIZE
  8383. ;;
  8384. STFD [AOFFSET] = f72, SIZE
  8385. ;;
  8386. STFD [AOFFSET] = f80, SIZE
  8387. ;;
  8388. STFD [AOFFSET] = f88, -3 * SIZE
  8389. ;;
  8390. #endif
  8391. STFD [C1 ] = f64, SIZE
  8392. STFD [C2 ] = f72, SIZE
  8393. STFD [C3 ] = f80, SIZE
  8394. STFD [C4 ] = f88, SIZE
  8395. ;;
  8396. mov f64 = f0
  8397. mov f72 = f0
  8398. mov f80 = f0
  8399. mov f88 = f0
  8400. ;;
  8401. sub L = K, KK
  8402. ;;
  8403. shladd L = L, BASE_SHIFT, r0
  8404. ;;
  8405. add AOFFSET = L, AOFFSET
  8406. ;;
  8407. shladd BOFFSET = L, 2, BOFFSET
  8408. ;;
  8409. #ifdef LT
  8410. adds KK = 1, KK
  8411. #else
  8412. nop __LINE__
  8413. #endif
  8414. ;;
  8415. mov L = KK
  8416. ;;
  8417. .align 8
  8418. .L089:
  8419. mov B = BOFFSET
  8420. #ifdef RN
  8421. adds KK = 4, KK
  8422. #endif
  8423. ;;
  8424. mov AOFFSET = A
  8425. ;;
  8426. .align 16
  8427. .L090:
  8428. tbit.z p6, p0 = N, 1
  8429. (p6) br.cond.dpnt .L130
  8430. ;;
  8431. mov f64 = f0
  8432. mov f65 = f0
  8433. mov f66 = f0
  8434. mov f67 = f0
  8435. mov f72 = f0
  8436. mov f73 = f0
  8437. mov f74 = f0
  8438. mov f75 = f0
  8439. ;;
  8440. { .mfi
  8441. shr I = M, 3
  8442. }
  8443. { .mfi
  8444. mov C1 = C // coffset1 = c + 0 * ldc
  8445. #ifdef LT
  8446. mov KK = OFFSET
  8447. #else
  8448. nop __LINE__
  8449. #endif
  8450. }
  8451. ;;
  8452. { .mmf
  8453. cmp.eq p6, p7 = 0, I
  8454. mov AORIG = A
  8455. }
  8456. { .mmf
  8457. add C2 = LDC, C // coffset2 = c + 1 * ldc
  8458. }
  8459. ;;
  8460. { .mfi
  8461. shladd C = LDC, 1, C // coffset += 8 * ldc
  8462. mov f81 = f0
  8463. mov L = KK
  8464. }{ .mfb
  8465. (p6) br.cond.dpnt .L100
  8466. }
  8467. ;;
  8468. .align 16
  8469. .L092:
  8470. { .mmi
  8471. cmp.ne p7, p0 = r0, L
  8472. adds BOFFSET = 0 * SIZE, B
  8473. }
  8474. ;;
  8475. { .mmi
  8476. (p7) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  8477. nop __LINE__
  8478. nop __LINE__
  8479. }
  8480. ;;
  8481. (p7) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  8482. ;;
  8483. { .mmf
  8484. (p7) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  8485. }
  8486. ;;
  8487. { .mmf
  8488. (p7) LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  8489. }
  8490. { .mfi
  8491. cmp.eq p3, p0 = r0, r0
  8492. }
  8493. ;;
  8494. { .mmf
  8495. (p7) LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  8496. }
  8497. { .mfi
  8498. adds PREC = CPREFETCHSIZE * SIZE, C1
  8499. }
  8500. ;;
  8501. { .mmf
  8502. CPREFETCH [PREC], LDC
  8503. }
  8504. { .mfi
  8505. adds L = 1, L
  8506. }
  8507. ;;
  8508. { .mmf
  8509. CPREFETCH [PREC]
  8510. }
  8511. { .mfi
  8512. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  8513. }
  8514. ;;
  8515. { .mfi
  8516. adds PREB = (PREFETCHSIZE - 8) * SIZE, BOFFSET
  8517. }
  8518. ;;
  8519. { .mfi
  8520. tbit.z p12, p0 = L, 0
  8521. }
  8522. { .mfi
  8523. shr L = L, 1
  8524. }
  8525. ;;
  8526. { .mfi
  8527. adds L = -1, L
  8528. }
  8529. ;;
  8530. { .mfi
  8531. mov ar.lc = L
  8532. }
  8533. ;;
  8534. mov f68 = f0
  8535. mov f69 = f0
  8536. mov f70 = f0
  8537. mov f71 = f0
  8538. mov f76 = f0
  8539. mov f77 = f0
  8540. mov f78 = f0
  8541. mov f79 = f0
  8542. ;;
  8543. { .mfb
  8544. cmp.eq p6, p0 = -1, L
  8545. (p6) br.cond.dpnt .L098
  8546. }
  8547. ;;
  8548. .align 8
  8549. .L093:
  8550. /* 1 */
  8551. { .mfi
  8552. lfetch.nt1 [PREA], 16 * SIZE
  8553. FMA f64 = f32, f48, f64 // A1 * B1
  8554. cmp.ne p4, p5 = 0, L
  8555. }
  8556. { .mfi
  8557. nop __LINE__
  8558. FMA f72 = f32, f49, f72 // A1 * B2
  8559. (p12) cmp.ne p3, p0 = 0, L
  8560. }
  8561. ;;
  8562. { .mfi
  8563. lfetch.nt1 [PREB], 4 * SIZE
  8564. FMA f65 = f33, f48, f65 // A2 * B1
  8565. adds C9 = 4 * SIZE, C1
  8566. }
  8567. { .mfi
  8568. nop __LINE__
  8569. FMA f73 = f33, f49, f73 // A2 * B2
  8570. adds C10 = 4 * SIZE, C2
  8571. }
  8572. ;;
  8573. { .mfi
  8574. (p3) LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  8575. FMA f66 = f34, f48, f66 // A3 * B1
  8576. adds C11 = 4 * SIZE, C3
  8577. }
  8578. { .mfi
  8579. nop __LINE__
  8580. FMA f74 = f34, f49, f74 // A3 * B2
  8581. adds C12 = 4 * SIZE, C4
  8582. }
  8583. ;;
  8584. { .mfb
  8585. (p3) LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  8586. FMA f67 = f35, f48, f67 // A4 * B1
  8587. nop __LINE__
  8588. }
  8589. { .mfb
  8590. nop __LINE__
  8591. FMA f75 = f35, f49, f75 // A4 * B2
  8592. nop __LINE__
  8593. }
  8594. ;;
  8595. { .mfb
  8596. (p3) LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  8597. FMA f68 = f36, f48, f68 // A5 * B1
  8598. nop __LINE__
  8599. }
  8600. { .mfb
  8601. nop __LINE__
  8602. FMA f76 = f36, f49, f76 // A5 * B2
  8603. nop __LINE__
  8604. }
  8605. ;;
  8606. { .mfb
  8607. (p3) LDFPD f44, f45 = [AOFFSET], 2 * SIZE
  8608. FMA f69 = f37, f48, f69 // A6 * B1
  8609. nop __LINE__
  8610. }
  8611. { .mfb
  8612. nop __LINE__
  8613. FMA f77 = f37, f49, f77 // A6 * B2
  8614. nop __LINE__
  8615. }
  8616. ;;
  8617. { .mfb
  8618. (p3) LDFPD f46, f47 = [AOFFSET], 2 * SIZE
  8619. FMA f70 = f38, f48, f70 // A7 * B1
  8620. nop __LINE__
  8621. }
  8622. { .mfb
  8623. nop __LINE__
  8624. FMA f78 = f38, f49, f78 // A7 * B2
  8625. nop __LINE__
  8626. }
  8627. ;;
  8628. { .mfb
  8629. (p4) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  8630. FMA f71 = f39, f48, f71 // A8 * B1
  8631. nop __LINE__
  8632. }
  8633. { .mfb
  8634. nop __LINE__
  8635. FMA f79 = f39, f49, f79 // A8 * B2
  8636. nop __LINE__
  8637. }
  8638. ;;
  8639. { .mfb
  8640. (p4) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  8641. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  8642. nop __LINE__
  8643. }
  8644. { .mfb
  8645. nop __LINE__
  8646. (p3) FMA f72 = f40, f57, f72 // A1 * B2
  8647. nop __LINE__
  8648. }
  8649. ;;
  8650. { .mfb
  8651. (p4) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  8652. (p3) FMA f65 = f41, f56, f65 // A2 * B1
  8653. nop __LINE__
  8654. }
  8655. { .mfb
  8656. nop __LINE__
  8657. (p3) FMA f73 = f41, f57, f73 // A2 * B2
  8658. nop __LINE__
  8659. }
  8660. ;;
  8661. { .mfb
  8662. (p4) LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  8663. (p3) FMA f66 = f42, f56, f66 // A3 * B1
  8664. nop __LINE__
  8665. }
  8666. { .mfb
  8667. nop __LINE__
  8668. (p3) FMA f74 = f42, f57, f74 // A3 * B2
  8669. nop __LINE__
  8670. }
  8671. ;;
  8672. { .mfb
  8673. (p4) LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  8674. (p3) FMA f67 = f43, f56, f67 // A4 * B1
  8675. nop __LINE__
  8676. }
  8677. { .mfb
  8678. nop __LINE__
  8679. (p3) FMA f75 = f43, f57, f75 // A4 * B2
  8680. nop __LINE__
  8681. }
  8682. ;;
  8683. { .mfb
  8684. nop __LINE__
  8685. (p3) FMA f68 = f44, f56, f68 // A5 * B1
  8686. nop __LINE__
  8687. }
  8688. { .mfb
  8689. nop __LINE__
  8690. (p3) FMA f76 = f44, f57, f76 // A5 * B2
  8691. nop __LINE__
  8692. }
  8693. ;;
  8694. { .mfb
  8695. nop __LINE__
  8696. (p3) FMA f69 = f45, f56, f69 // A6 * B1
  8697. nop __LINE__
  8698. }
  8699. { .mfb
  8700. nop __LINE__
  8701. (p3) FMA f77 = f45, f57, f77 // A6 * B2
  8702. nop __LINE__
  8703. }
  8704. ;;
  8705. { .mfb
  8706. nop __LINE__
  8707. (p3) FMA f70 = f46, f56, f70 // A7 * B1
  8708. nop __LINE__
  8709. }
  8710. { .mfb
  8711. nop __LINE__
  8712. (p3) FMA f78 = f46, f57, f78 // A7 * B2
  8713. nop __LINE__
  8714. }
  8715. ;;
  8716. { .mfi
  8717. nop __LINE__
  8718. (p3) FMA f71 = f47, f56, f71 // A8 * B1
  8719. adds L = -1, L
  8720. }
  8721. { .mfb
  8722. nop __LINE__
  8723. (p3) FMA f79 = f47, f57, f79 // A8 * B2
  8724. br.cloop.sptk.few .L093
  8725. }
  8726. ;;
  8727. .align 8
  8728. .L098:
  8729. adds AOFFSET2 = 4 * SIZE, AOFFSET
  8730. adds BOFFSET2 = 4 * SIZE, BOFFSET
  8731. ;;
  8732. #ifdef LT
  8733. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  8734. ;;
  8735. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  8736. ;;
  8737. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  8738. ;;
  8739. LDFPD f38, f39 = [BOFFSET], 2 * SIZE
  8740. ;;
  8741. LDFPD f40, f41 = [BOFFSET], 2 * SIZE
  8742. ;;
  8743. LDFPD f42, f43 = [BOFFSET], 2 * SIZE
  8744. ;;
  8745. LDFPD f44, f45 = [BOFFSET], 2 * SIZE
  8746. ;;
  8747. LDFPD f46, f47 = [BOFFSET]
  8748. adds BOFFSET = -14 * SIZE, BOFFSET
  8749. ;;
  8750. FSUB f64 = f32, f64
  8751. FSUB f72 = f33, f72
  8752. FSUB f65 = f34, f65
  8753. FSUB f73 = f35, f73
  8754. FSUB f66 = f36, f66
  8755. FSUB f74 = f37, f74
  8756. FSUB f67 = f38, f67
  8757. FSUB f75 = f39, f75
  8758. FSUB f68 = f40, f68
  8759. FSUB f76 = f41, f76
  8760. FSUB f69 = f42, f69
  8761. FSUB f77 = f43, f77
  8762. FSUB f70 = f44, f70
  8763. FSUB f78 = f45, f78
  8764. FSUB f71 = f46, f71
  8765. FSUB f79 = f47, f79
  8766. ;;
  8767. #else
  8768. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  8769. ;;
  8770. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  8771. ;;
  8772. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  8773. ;;
  8774. LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  8775. ;;
  8776. LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  8777. ;;
  8778. LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  8779. ;;
  8780. LDFPD f44, f45 = [AOFFSET], 2 * SIZE
  8781. ;;
  8782. LDFPD f46, f47 = [AOFFSET]
  8783. adds AOFFSET = -14 * SIZE, AOFFSET
  8784. ;;
  8785. FSUB f64 = f32, f64
  8786. FSUB f65 = f33, f65
  8787. FSUB f66 = f34, f66
  8788. FSUB f67 = f35, f67
  8789. FSUB f68 = f36, f68
  8790. FSUB f69 = f37, f69
  8791. FSUB f70 = f38, f70
  8792. FSUB f71 = f39, f71
  8793. ;;
  8794. FSUB f72 = f40, f72
  8795. FSUB f73 = f41, f73
  8796. FSUB f74 = f42, f74
  8797. FSUB f75 = f43, f75
  8798. FSUB f76 = f44, f76
  8799. FSUB f77 = f45, f77
  8800. FSUB f78 = f46, f78
  8801. FSUB f79 = f47, f79
  8802. ;;
  8803. #endif
  8804. #ifdef LT
  8805. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  8806. ;;
  8807. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  8808. ;;
  8809. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  8810. ;;
  8811. LDFPD f38, f39 = [AOFFSET]
  8812. adds AOFFSET = 3 * SIZE, AOFFSET
  8813. ;;
  8814. LDFD f40 = [AOFFSET], 1 * SIZE
  8815. ;;
  8816. LDFPD f41, f42 = [AOFFSET], 2 * SIZE
  8817. ;;
  8818. LDFPD f43, f44 = [AOFFSET], 2 * SIZE
  8819. ;;
  8820. LDFPD f45, f46 = [AOFFSET]
  8821. adds AOFFSET = 4 * SIZE, AOFFSET
  8822. ;;
  8823. LDFPD f47, f48 = [AOFFSET], 2 * SIZE
  8824. ;;
  8825. LDFPD f49, f50 = [AOFFSET], 2 * SIZE
  8826. ;;
  8827. LDFPD f51, f52 = [AOFFSET]
  8828. adds AOFFSET = 5 * SIZE, AOFFSET
  8829. ;;
  8830. LDFD f53 = [AOFFSET], 1 * SIZE
  8831. ;;
  8832. LDFPD f54, f55 = [AOFFSET], 2 * SIZE
  8833. ;;
  8834. LDFPD f56, f57 = [AOFFSET]
  8835. adds AOFFSET = 6 * SIZE, AOFFSET
  8836. ;;
  8837. LDFPD f58, f59 = [AOFFSET], 2 * SIZE
  8838. ;;
  8839. LDFPD f60, f61 = [AOFFSET]
  8840. adds AOFFSET = 7 * SIZE, AOFFSET
  8841. ;;
  8842. LDFD f16 = [AOFFSET], 1 * SIZE
  8843. ;;
  8844. LDFPD f17, f18 = [AOFFSET]
  8845. adds AOFFSET = 8 * SIZE, AOFFSET
  8846. ;;
  8847. LDFPD f19, f20 = [AOFFSET]
  8848. adds AOFFSET = 9 * SIZE, AOFFSET
  8849. ;;
  8850. LDFD f21 = [AOFFSET]
  8851. adds AOFFSET = -63 * SIZE, AOFFSET
  8852. ;;
  8853. FMPY f64 = f64, f32
  8854. FMPY f72 = f72, f32
  8855. ;;
  8856. FNMA f65 = f64, f33, f65
  8857. FNMA f73 = f72, f33, f73
  8858. ;;
  8859. FNMA f66 = f64, f34, f66
  8860. FNMA f74 = f72, f34, f74
  8861. ;;
  8862. FNMA f67 = f64, f35, f67
  8863. FNMA f75 = f72, f35, f75
  8864. ;;
  8865. FNMA f68 = f64, f36, f68
  8866. FNMA f76 = f72, f36, f76
  8867. ;;
  8868. FNMA f69 = f64, f37, f69
  8869. FNMA f77 = f72, f37, f77
  8870. ;;
  8871. FNMA f70 = f64, f38, f70
  8872. FNMA f78 = f72, f38, f78
  8873. ;;
  8874. FNMA f71 = f64, f39, f71
  8875. FNMA f79 = f72, f39, f79
  8876. ;;
  8877. FMPY f65 = f65, f40
  8878. FMPY f73 = f73, f40
  8879. ;;
  8880. FNMA f66 = f65, f41, f66
  8881. FNMA f74 = f73, f41, f74
  8882. ;;
  8883. FNMA f67 = f65, f42, f67
  8884. FNMA f75 = f73, f42, f75
  8885. ;;
  8886. FNMA f68 = f65, f43, f68
  8887. FNMA f76 = f73, f43, f76
  8888. ;;
  8889. FNMA f69 = f65, f44, f69
  8890. FNMA f77 = f73, f44, f77
  8891. ;;
  8892. FNMA f70 = f65, f45, f70
  8893. FNMA f78 = f73, f45, f78
  8894. ;;
  8895. FNMA f71 = f65, f46, f71
  8896. FNMA f79 = f73, f46, f79
  8897. ;;
  8898. FMPY f66 = f66, f47
  8899. FMPY f74 = f74, f47
  8900. ;;
  8901. FNMA f67 = f66, f48, f67
  8902. FNMA f75 = f74, f48, f75
  8903. ;;
  8904. FNMA f68 = f66, f49, f68
  8905. FNMA f76 = f74, f49, f76
  8906. ;;
  8907. FNMA f69 = f66, f50, f69
  8908. FNMA f77 = f74, f50, f77
  8909. ;;
  8910. FNMA f70 = f66, f51, f70
  8911. FNMA f78 = f74, f51, f78
  8912. ;;
  8913. FNMA f71 = f66, f52, f71
  8914. FNMA f79 = f74, f52, f79
  8915. ;;
  8916. FMPY f67 = f67, f53
  8917. FMPY f75 = f75, f53
  8918. ;;
  8919. FNMA f68 = f67, f54, f68
  8920. FNMA f76 = f75, f54, f76
  8921. ;;
  8922. FNMA f69 = f67, f55, f69
  8923. FNMA f77 = f75, f55, f77
  8924. ;;
  8925. FNMA f70 = f67, f56, f70
  8926. FNMA f78 = f75, f56, f78
  8927. ;;
  8928. FNMA f71 = f67, f57, f71
  8929. FNMA f79 = f75, f57, f79
  8930. ;;
  8931. FMPY f68 = f68, f58
  8932. FMPY f76 = f76, f58
  8933. ;;
  8934. FNMA f69 = f68, f59, f69
  8935. FNMA f77 = f76, f59, f77
  8936. ;;
  8937. FNMA f70 = f68, f60, f70
  8938. FNMA f78 = f76, f60, f78
  8939. ;;
  8940. FNMA f71 = f68, f61, f71
  8941. FNMA f79 = f76, f61, f79
  8942. ;;
  8943. FMPY f69 = f69, f16
  8944. FMPY f77 = f77, f16
  8945. ;;
  8946. FNMA f70 = f69, f17, f70
  8947. FNMA f78 = f77, f17, f78
  8948. ;;
  8949. FNMA f71 = f69, f18, f71
  8950. FNMA f79 = f77, f18, f79
  8951. ;;
  8952. FMPY f70 = f70, f19
  8953. FMPY f78 = f78, f19
  8954. ;;
  8955. FNMA f71 = f70, f20, f71
  8956. FNMA f79 = f78, f20, f79
  8957. ;;
  8958. FMPY f71 = f71, f21
  8959. FMPY f79 = f79, f21
  8960. ;;
  8961. STFD [BOFFSET] = f64, SIZE
  8962. STFD [BOFFSET2] = f66, SIZE
  8963. ;;
  8964. STFD [BOFFSET] = f72, SIZE
  8965. STFD [BOFFSET2] = f74, SIZE
  8966. ;;
  8967. STFD [BOFFSET] = f65, SIZE
  8968. STFD [BOFFSET2] = f67, SIZE
  8969. ;;
  8970. STFD [BOFFSET] = f73, 5 * SIZE
  8971. STFD [BOFFSET2] = f75, 5 * SIZE
  8972. ;;
  8973. STFD [BOFFSET] = f68, SIZE
  8974. STFD [BOFFSET2] = f70, SIZE
  8975. ;;
  8976. STFD [BOFFSET] = f76, SIZE
  8977. STFD [BOFFSET2] = f78, SIZE
  8978. ;;
  8979. STFD [BOFFSET] = f69, SIZE
  8980. STFD [BOFFSET2] = f71, SIZE
  8981. ;;
  8982. STFD [BOFFSET] = f77, -11 * SIZE
  8983. STFD [BOFFSET2] = f79, -11 * SIZE
  8984. ;;
  8985. adds C9 = 4 * SIZE, C1
  8986. ;;
  8987. #endif
  8988. #ifdef RN
  8989. LDFPD f32, f33 = [BOFFSET]
  8990. adds BOFFSET = 3 * SIZE, BOFFSET
  8991. ;;
  8992. LDFD f34 = [BOFFSET], -3 * SIZE
  8993. ;;
  8994. FMPY f64 = f64, f32
  8995. FMPY f68 = f68, f32
  8996. FMPY f65 = f65, f32
  8997. FMPY f69 = f69, f32
  8998. FMPY f66 = f66, f32
  8999. FMPY f70 = f70, f32
  9000. FMPY f67 = f67, f32
  9001. FMPY f71 = f71, f32
  9002. ;;
  9003. FNMA f72 = f64, f33, f72
  9004. FNMA f76 = f68, f33, f76
  9005. FNMA f73 = f65, f33, f73
  9006. FNMA f77 = f69, f33, f77
  9007. FNMA f74 = f66, f33, f74
  9008. FNMA f78 = f70, f33, f78
  9009. FNMA f75 = f67, f33, f75
  9010. FNMA f79 = f71, f33, f79
  9011. ;;
  9012. FMPY f72 = f72, f34
  9013. FMPY f76 = f76, f34
  9014. FMPY f73 = f73, f34
  9015. FMPY f77 = f77, f34
  9016. FMPY f74 = f74, f34
  9017. FMPY f78 = f78, f34
  9018. FMPY f75 = f75, f34
  9019. FMPY f79 = f79, f34
  9020. ;;
  9021. STFD [AOFFSET] = f64, SIZE
  9022. STFD [AOFFSET2] = f68, SIZE
  9023. ;;
  9024. STFD [AOFFSET] = f65, SIZE
  9025. STFD [AOFFSET2] = f69, SIZE
  9026. ;;
  9027. STFD [AOFFSET] = f66, SIZE
  9028. STFD [AOFFSET2] = f70, SIZE
  9029. ;;
  9030. STFD [AOFFSET] = f67, 5 * SIZE
  9031. STFD [AOFFSET2] = f71, 5 * SIZE
  9032. ;;
  9033. STFD [AOFFSET] = f72, SIZE
  9034. STFD [AOFFSET2] = f76, SIZE
  9035. ;;
  9036. STFD [AOFFSET] = f73, SIZE
  9037. STFD [AOFFSET2] = f77, SIZE
  9038. ;;
  9039. STFD [AOFFSET] = f74, SIZE
  9040. STFD [AOFFSET2] = f78, SIZE
  9041. ;;
  9042. STFD [AOFFSET] = f75, -11 * SIZE
  9043. STFD [AOFFSET2] = f79, -11 * SIZE
  9044. ;;
  9045. #endif
  9046. adds C9 = 4 * SIZE, C1
  9047. ;;
  9048. { .mmf
  9049. STFD [C1 ] = f64, SIZE
  9050. STFD [C9 ] = f68, SIZE
  9051. mov f64 = f0
  9052. }
  9053. ;;
  9054. { .mmi
  9055. STFD [C1 ] = f65, SIZE
  9056. STFD [C9 ] = f69, SIZE
  9057. adds C10 = 4 * SIZE, C2
  9058. }
  9059. ;;
  9060. { .mmi
  9061. STFD [C1 ] = f66, SIZE
  9062. STFD [C9 ] = f70, SIZE
  9063. }
  9064. ;;
  9065. { .mmi
  9066. STFD [C1 ] = f67, 5 * SIZE
  9067. STFD [C9 ] = f71
  9068. adds C11 = 4 * SIZE, C3
  9069. }
  9070. ;;
  9071. { .mmf
  9072. STFD [C2 ] = f72, SIZE
  9073. STFD [C10] = f76, SIZE
  9074. mov f72 = f0
  9075. }
  9076. ;;
  9077. { .mmi
  9078. STFD [C2 ] = f73, SIZE
  9079. STFD [C10] = f77, SIZE
  9080. }
  9081. ;;
  9082. { .mmi
  9083. STFD [C2 ] = f74, SIZE
  9084. STFD [C10] = f78, SIZE
  9085. adds C12 = 4 * SIZE, C4
  9086. }
  9087. ;;
  9088. { .mmi
  9089. STFD [C2 ] = f75, 5 * SIZE
  9090. STFD [C10] = f79
  9091. }
  9092. ;;
  9093. { .mmf
  9094. cmp.ne p6, p0 = 1, I
  9095. }
  9096. ;;
  9097. adds I = -1, I
  9098. ;;
  9099. { .mmi
  9100. sub L = K, KK
  9101. }
  9102. ;;
  9103. { .mmi
  9104. shladd L = L, BASE_SHIFT, r0
  9105. }
  9106. ;;
  9107. ;;
  9108. shladd AOFFSET = L, 3, AOFFSET
  9109. shladd BOFFSET = L, 1, BOFFSET
  9110. ;;
  9111. { .mmi
  9112. #ifdef LT
  9113. adds KK = 8, KK
  9114. #else
  9115. nop __LINE__
  9116. #endif
  9117. }
  9118. ;;
  9119. mov L = KK
  9120. mov f64 = f0
  9121. mov f65 = f0
  9122. mov f66 = f0
  9123. mov f67 = f0
  9124. mov f72 = f0
  9125. mov f73 = f0
  9126. mov f74 = f0
  9127. mov f75 = f0
  9128. (p6) br.cond.dptk .L092
  9129. ;;
  9130. .align 8
  9131. .L100:
  9132. { .mib
  9133. mov L = KK
  9134. tbit.z p6, p7 = M, 2
  9135. (p6) br.cond.dptk .L110
  9136. }
  9137. ;;
  9138. cmp.ne p7, p0 = r0, L
  9139. adds BOFFSET = 0 * SIZE, B
  9140. ;;
  9141. { .mmf
  9142. (p7) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  9143. mov f65 = f0
  9144. }
  9145. ;;
  9146. adds L = 1, L
  9147. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  9148. cmp.eq p3, p0 = r0, r0
  9149. ;;
  9150. { .mfi
  9151. tbit.z p12, p0 = L, 0
  9152. }
  9153. { .mfi
  9154. shr L = L, 1
  9155. }
  9156. ;;
  9157. { .mfi
  9158. adds L = -1, L
  9159. }
  9160. ;;
  9161. { .mfi
  9162. cmp.eq p6, p0 = -1, L
  9163. }
  9164. ;;
  9165. { .mmf
  9166. (p7) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  9167. }
  9168. { .mfi
  9169. mov ar.lc = L
  9170. }
  9171. ;;
  9172. { .mmf
  9173. (p7) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  9174. }
  9175. { .mfb
  9176. (p6) br.cond.dpnt .L108
  9177. }
  9178. ;;
  9179. .L102:
  9180. { .mfi
  9181. lfetch.nt1 [PREA], 8 * SIZE
  9182. FMA f64 = f32, f48, f64 // A1 * B1
  9183. cmp.ne p4, p5 = 0, L
  9184. }
  9185. { .mfi
  9186. adds PREB = (PREFETCHSIZE + 0) * SIZE, BOFFSET
  9187. FMA f72 = f32, f49, f72 // A1 * B2
  9188. (p12) cmp.ne p3, p0 = 0, L
  9189. }
  9190. ;;
  9191. { .mfi
  9192. lfetch.nt1 [PREB], 4 * SIZE
  9193. FMA f65 = f33, f48, f65 // A2 * B1
  9194. adds C9 = 2 * SIZE, C1
  9195. }
  9196. { .mfi
  9197. nop __LINE__
  9198. FMA f73 = f33, f49, f73 // A2 * B2
  9199. adds C10 = 2 * SIZE, C2
  9200. }
  9201. ;;
  9202. { .mfb
  9203. (p3) LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  9204. FMA f66 = f34, f48, f66 // A3 * B1
  9205. nop __LINE__
  9206. }
  9207. { .mfb
  9208. nop __LINE__
  9209. FMA f74 = f34, f49, f74 // A3 * B2
  9210. nop __LINE__
  9211. }
  9212. ;;
  9213. { .mfb
  9214. (p3) LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  9215. FMA f67 = f35, f48, f67 // A4 * B1
  9216. nop __LINE__
  9217. }
  9218. { .mfb
  9219. nop __LINE__
  9220. FMA f75 = f35, f49, f75 // A4 * B2
  9221. nop __LINE__
  9222. }
  9223. ;;
  9224. { .mfb
  9225. (p3) LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  9226. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  9227. nop __LINE__
  9228. }
  9229. { .mfb
  9230. nop __LINE__
  9231. (p3) FMA f72 = f40, f57, f72 // A1 * B2
  9232. nop __LINE__
  9233. }
  9234. ;;
  9235. { .mfb
  9236. (p4) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  9237. (p3) FMA f65 = f41, f56, f65 // A2 * B1
  9238. nop __LINE__
  9239. }
  9240. { .mfb
  9241. nop __LINE__
  9242. (p3) FMA f73 = f41, f57, f73 // A2 * B2
  9243. nop __LINE__
  9244. }
  9245. ;;
  9246. { .mfb
  9247. (p4) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  9248. (p3) FMA f66 = f42, f56, f66 // A3 * B1
  9249. nop __LINE__
  9250. }
  9251. { .mfb
  9252. nop __LINE__
  9253. (p3) FMA f74 = f42, f57, f74 // A3 * B2
  9254. nop __LINE__
  9255. }
  9256. ;;
  9257. { .mfi
  9258. (p4) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  9259. (p3) FMA f67 = f43, f56, f67 // A4 * B1
  9260. adds L = -1, L
  9261. }
  9262. { .mfb
  9263. nop __LINE__
  9264. (p3) FMA f75 = f43, f57, f75 // A4 * B2
  9265. br.cloop.sptk.few .L102
  9266. }
  9267. ;;
  9268. .align 8
  9269. .L108:
  9270. adds AOFFSET2 = 4 * SIZE, AOFFSET
  9271. adds BOFFSET2 = 4 * SIZE, BOFFSET
  9272. ;;
  9273. #ifdef LT
  9274. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  9275. ;;
  9276. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  9277. ;;
  9278. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  9279. ;;
  9280. LDFPD f38, f39 = [BOFFSET]
  9281. adds BOFFSET = -6 * SIZE, BOFFSET
  9282. ;;
  9283. FSUB f64 = f32, f64
  9284. FSUB f72 = f33, f72
  9285. ;;
  9286. FSUB f65 = f34, f65
  9287. FSUB f73 = f35, f73
  9288. ;;
  9289. FSUB f66 = f36, f66
  9290. FSUB f74 = f37, f74
  9291. ;;
  9292. FSUB f67 = f38, f67
  9293. FSUB f75 = f39, f75
  9294. ;;
  9295. #else
  9296. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  9297. ;;
  9298. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  9299. ;;
  9300. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  9301. ;;
  9302. LDFPD f38, f39 = [AOFFSET]
  9303. adds AOFFSET = -6 * SIZE, AOFFSET
  9304. ;;
  9305. FSUB f64 = f32, f64
  9306. FSUB f65 = f33, f65
  9307. FSUB f66 = f34, f66
  9308. FSUB f67 = f35, f67
  9309. FSUB f72 = f36, f72
  9310. FSUB f73 = f37, f73
  9311. FSUB f74 = f38, f74
  9312. FSUB f75 = f39, f75
  9313. ;;
  9314. #endif
  9315. #ifdef LT
  9316. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  9317. ;;
  9318. LDFPD f34, f35 = [AOFFSET]
  9319. adds AOFFSET = 3 * SIZE, AOFFSET
  9320. ;;
  9321. LDFD f36 = [AOFFSET], 1 * SIZE
  9322. ;;
  9323. LDFPD f37, f38 = [AOFFSET]
  9324. adds AOFFSET = 4 * SIZE, AOFFSET
  9325. ;;
  9326. LDFPD f39, f40 = [AOFFSET]
  9327. adds AOFFSET = 5 * SIZE, AOFFSET
  9328. ;;
  9329. LDFD f41 = [AOFFSET], -15 * SIZE
  9330. ;;
  9331. FMPY f64 = f64, f32
  9332. FMPY f72 = f72, f32
  9333. ;;
  9334. FNMA f65 = f64, f33, f65
  9335. FNMA f73 = f72, f33, f73
  9336. ;;
  9337. FNMA f66 = f64, f34, f66
  9338. FNMA f74 = f72, f34, f74
  9339. ;;
  9340. FNMA f67 = f64, f35, f67
  9341. FNMA f75 = f72, f35, f75
  9342. ;;
  9343. FMPY f65 = f65, f36
  9344. FMPY f73 = f73, f36
  9345. ;;
  9346. FNMA f66 = f65, f37, f66
  9347. FNMA f74 = f73, f37, f74
  9348. ;;
  9349. FNMA f67 = f65, f38, f67
  9350. FNMA f75 = f73, f38, f75
  9351. ;;
  9352. FMPY f66 = f66, f39
  9353. FMPY f74 = f74, f39
  9354. ;;
  9355. FNMA f67 = f66, f40, f67
  9356. FNMA f75 = f74, f40, f75
  9357. ;;
  9358. FMPY f67 = f67, f41
  9359. FMPY f75 = f75, f41
  9360. ;;
  9361. STFD [BOFFSET] = f64, SIZE
  9362. STFD [BOFFSET2] = f66, SIZE
  9363. ;;
  9364. STFD [BOFFSET] = f72, SIZE
  9365. STFD [BOFFSET2] = f74, SIZE
  9366. ;;
  9367. STFD [BOFFSET] = f65, SIZE
  9368. STFD [BOFFSET2] = f67, SIZE
  9369. ;;
  9370. STFD [BOFFSET] = f73, -3 * SIZE
  9371. STFD [BOFFSET2] = f75, -3 * SIZE
  9372. ;;
  9373. #endif
  9374. #ifdef RN
  9375. LDFPD f32, f33 = [BOFFSET]
  9376. adds BOFFSET = 3 * SIZE, BOFFSET
  9377. ;;
  9378. LDFD f34 = [BOFFSET], -3 * SIZE
  9379. ;;
  9380. FMPY f64 = f64, f32
  9381. FMPY f65 = f65, f32
  9382. FMPY f66 = f66, f32
  9383. FMPY f67 = f67, f32
  9384. ;;
  9385. FNMA f72 = f64, f33, f72
  9386. FNMA f73 = f65, f33, f73
  9387. FNMA f74 = f66, f33, f74
  9388. FNMA f75 = f67, f33, f75
  9389. ;;
  9390. FMPY f72 = f72, f34
  9391. FMPY f73 = f73, f34
  9392. FMPY f74 = f74, f34
  9393. FMPY f75 = f75, f34
  9394. ;;
  9395. STFD [AOFFSET] = f64, SIZE
  9396. STFD [AOFFSET2] = f72, SIZE
  9397. ;;
  9398. STFD [AOFFSET] = f65, SIZE
  9399. STFD [AOFFSET2] = f73, SIZE
  9400. ;;
  9401. STFD [AOFFSET] = f66, SIZE
  9402. STFD [AOFFSET2] = f74, SIZE
  9403. ;;
  9404. STFD [AOFFSET] = f67, -3 * SIZE
  9405. STFD [AOFFSET2] = f75, -3 * SIZE
  9406. ;;
  9407. #endif
  9408. { .mmf
  9409. STFD [C1 ] = f64, SIZE
  9410. mov f64 = f0
  9411. }
  9412. ;;
  9413. { .mmi
  9414. STFD [C1 ] = f65, SIZE
  9415. }
  9416. ;;
  9417. { .mmi
  9418. STFD [C1 ] = f66, SIZE
  9419. }
  9420. ;;
  9421. { .mmi
  9422. STFD [C1 ] = f67, SIZE
  9423. }
  9424. ;;
  9425. { .mmf
  9426. STFD [C2 ] = f72, SIZE
  9427. mov f72 = f0
  9428. }
  9429. ;;
  9430. { .mmi
  9431. STFD [C2 ] = f73, SIZE
  9432. }
  9433. ;;
  9434. { .mmi
  9435. STFD [C2 ] = f74, SIZE
  9436. }
  9437. ;;
  9438. { .mmi
  9439. STFD [C2 ] = f75, SIZE
  9440. }
  9441. ;;
  9442. mov f65 = f0
  9443. mov f73 = f0
  9444. mov f66 = f0
  9445. mov f74 = f0
  9446. mov f67 = f0
  9447. mov f75 = f0
  9448. ;;
  9449. { .mmi
  9450. sub L = K, KK
  9451. }
  9452. ;;
  9453. { .mmi
  9454. shladd L = L, BASE_SHIFT, r0
  9455. }
  9456. ;;
  9457. shladd AOFFSET = L, 2, AOFFSET
  9458. ;;
  9459. shladd BOFFSET = L, 1, BOFFSET
  9460. ;;
  9461. #ifdef LT
  9462. adds KK = 4, KK
  9463. nop __LINE__
  9464. #endif
  9465. ;;
  9466. .align 8
  9467. .L110:
  9468. { .mib
  9469. tbit.z p6, p7 = M, 1
  9470. (p6) br.cond.dptk .L120
  9471. }
  9472. ;;
  9473. { .mib
  9474. mov L = KK
  9475. }
  9476. ;;
  9477. { .mmi
  9478. cmp.ne p7, p0 = r0, L
  9479. adds BOFFSET = 0 * SIZE, B
  9480. }
  9481. ;;
  9482. { .mmf
  9483. (p7) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  9484. }
  9485. ;;
  9486. adds L = 1, L
  9487. ;;
  9488. { .mfi
  9489. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  9490. cmp.eq p3, p0 = r0, r0
  9491. }
  9492. ;;
  9493. { .mfi
  9494. tbit.z p12, p0 = L, 0
  9495. }
  9496. { .mfi
  9497. shr L = L, 1
  9498. }
  9499. ;;
  9500. { .mmf
  9501. adds L = -1, L
  9502. }
  9503. ;;
  9504. { .mmf
  9505. cmp.eq p6, p0 = -1, L
  9506. }
  9507. ;;
  9508. { .mib
  9509. (p7) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  9510. mov ar.lc = L
  9511. (p6) br.cond.dpnt .L118
  9512. }
  9513. ;;
  9514. .L112:
  9515. { .mfi
  9516. lfetch.nt1 [PREA], 4 * SIZE
  9517. FMA f64 = f32, f48, f64 // A1 * B1
  9518. cmp.ne p4, p5 = 0, L
  9519. }
  9520. { .mfi
  9521. lfetch.nt1 [PREB], 4 * SIZE
  9522. FMA f72 = f32, f49, f72 // A1 * B2
  9523. (p12) cmp.ne p3, p0 = 0, L
  9524. }
  9525. ;;
  9526. { .mmf
  9527. (p3) LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  9528. (p3) LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  9529. FMA f65 = f33, f48, f65 // A2 * B1
  9530. }
  9531. { .mmf
  9532. nop __LINE__
  9533. nop __LINE__
  9534. FMA f73 = f33, f49, f73 // A2 * B2
  9535. }
  9536. ;;
  9537. { .mfb
  9538. (p4) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  9539. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  9540. nop __LINE__
  9541. }
  9542. { .mfb
  9543. nop __LINE__
  9544. (p3) FMA f72 = f40, f57, f72 // A1 * B2
  9545. nop __LINE__
  9546. }
  9547. ;;
  9548. { .mfi
  9549. (p4) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  9550. (p3) FMA f65 = f41, f56, f65 // A2 * B1
  9551. adds L = -1, L
  9552. }
  9553. { .mfb
  9554. nop __LINE__
  9555. (p3) FMA f73 = f41, f57, f73 // A2 * B2
  9556. br.cloop.sptk.few .L112
  9557. }
  9558. ;;
  9559. .align 8
  9560. .L118:
  9561. adds AOFFSET2 = 4 * SIZE, AOFFSET
  9562. adds BOFFSET2 = 4 * SIZE, BOFFSET
  9563. ;;
  9564. #ifdef LT
  9565. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  9566. ;;
  9567. LDFPD f34, f35 = [BOFFSET]
  9568. adds BOFFSET = -2 * SIZE, BOFFSET
  9569. ;;
  9570. FSUB f64 = f32, f64
  9571. FSUB f72 = f33, f72
  9572. FSUB f65 = f34, f65
  9573. FSUB f73 = f35, f73
  9574. ;;
  9575. #else
  9576. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  9577. ;;
  9578. LDFPD f34, f35 = [AOFFSET]
  9579. adds AOFFSET = -2 * SIZE, AOFFSET
  9580. ;;
  9581. FSUB f64 = f32, f64
  9582. FSUB f65 = f33, f65
  9583. FSUB f72 = f34, f72
  9584. FSUB f73 = f35, f73
  9585. ;;
  9586. #endif
  9587. #ifdef LT
  9588. LDFPD f32, f33 = [AOFFSET]
  9589. adds AOFFSET = 3 * SIZE, AOFFSET
  9590. ;;
  9591. LDFD f34 = [AOFFSET], - 3 * SIZE
  9592. ;;
  9593. FMPY f64 = f64, f32
  9594. FMPY f72 = f72, f32
  9595. ;;
  9596. FNMA f65 = f64, f33, f65
  9597. FNMA f73 = f72, f33, f73
  9598. ;;
  9599. FMPY f65 = f65, f34
  9600. FMPY f73 = f73, f34
  9601. ;;
  9602. STFD [BOFFSET] = f64, SIZE
  9603. ;;
  9604. STFD [BOFFSET] = f72, SIZE
  9605. ;;
  9606. STFD [BOFFSET] = f65, SIZE
  9607. ;;
  9608. STFD [BOFFSET] = f73, -3 * SIZE
  9609. ;;
  9610. #endif
  9611. #ifdef RN
  9612. LDFPD f32, f33 = [BOFFSET]
  9613. adds BOFFSET = 3 * SIZE, BOFFSET
  9614. ;;
  9615. LDFD f34 = [BOFFSET], -3 * SIZE
  9616. ;;
  9617. FMPY f64 = f64, f32
  9618. FMPY f65 = f65, f32
  9619. ;;
  9620. FNMA f72 = f64, f33, f72
  9621. FNMA f73 = f65, f33, f73
  9622. ;;
  9623. FMPY f72 = f72, f34
  9624. FMPY f73 = f73, f34
  9625. ;;
  9626. STFD [AOFFSET] = f64, SIZE
  9627. ;;
  9628. STFD [AOFFSET] = f65, SIZE
  9629. ;;
  9630. STFD [AOFFSET] = f72, SIZE
  9631. ;;
  9632. STFD [AOFFSET] = f73, -3 * SIZE
  9633. ;;
  9634. #endif
  9635. STFD [C1 ] = f64, SIZE
  9636. mov f64 = f0
  9637. ;;
  9638. STFD [C1 ] = f65, SIZE
  9639. ;;
  9640. STFD [C2 ] = f72, SIZE
  9641. mov f72 = f0
  9642. ;;
  9643. STFD [C2 ] = f73, SIZE
  9644. ;;
  9645. mov f65 = f0
  9646. mov f73 = f0
  9647. ;;
  9648. sub L = K, KK
  9649. ;;
  9650. { .mmi
  9651. shladd L = L, BASE_SHIFT, r0
  9652. }
  9653. ;;
  9654. { .mmi
  9655. shladd AOFFSET = L, 1, AOFFSET
  9656. }
  9657. ;;
  9658. { .mmi
  9659. shladd BOFFSET = L, 1, BOFFSET
  9660. }
  9661. ;;
  9662. { .mmi
  9663. #ifdef LT
  9664. adds KK = 2, KK
  9665. #else
  9666. nop __LINE__
  9667. #endif
  9668. }
  9669. ;;
  9670. .align 8
  9671. .L120:
  9672. tbit.z p6, p7 = M, 0
  9673. (p6) br.cond.dptk .L129
  9674. ;;
  9675. { .mib
  9676. mov L = KK
  9677. }
  9678. ;;
  9679. { .mmi
  9680. cmp.ne p7, p0 = r0, L
  9681. adds BOFFSET = 0 * SIZE, B
  9682. }
  9683. ;;
  9684. { .mmf
  9685. (p7) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  9686. }
  9687. ;;
  9688. { .mmi
  9689. adds L = 1, L
  9690. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  9691. cmp.eq p3, p0 = r0, r0
  9692. }
  9693. ;;
  9694. { .mii
  9695. tbit.z p12, p0 = L, 0
  9696. shr L = L, 1
  9697. }
  9698. ;;
  9699. { .mmi
  9700. adds L = -1, L
  9701. }
  9702. ;;
  9703. { .mmi
  9704. cmp.eq p6, p0 = -1, L
  9705. }
  9706. ;;
  9707. { .mib
  9708. (p7) LDFD f32 = [AOFFSET], 1 * SIZE
  9709. mov ar.lc = L
  9710. (p6) br.cond.dpnt .L128
  9711. }
  9712. ;;
  9713. .align 8
  9714. .L122:
  9715. { .mfi
  9716. FMA f64 = f32, f48, f64 // A1 * B1
  9717. cmp.ne p4, p5 = 0, L
  9718. }
  9719. { .mfi
  9720. nop __LINE__
  9721. FMA f72 = f32, f49, f72 // A1 * B2
  9722. (p12) cmp.ne p3, p0 = 0, L
  9723. }
  9724. ;;
  9725. { .mmi
  9726. (p3) LDFPD f56, f57 = [BOFFSET], 2 * SIZE
  9727. (p3) LDFD f40 = [AOFFSET], 1 * SIZE
  9728. nop __LINE__
  9729. }
  9730. { .mmi
  9731. nop __LINE__
  9732. nop __LINE__
  9733. nop __LINE__
  9734. }
  9735. ;;
  9736. { .mfi
  9737. (p4) LDFPD f48, f49 = [BOFFSET], 2 * SIZE
  9738. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  9739. adds L = -1, L
  9740. }
  9741. { .mfb
  9742. (p4) LDFD f32 = [AOFFSET], 1 * SIZE
  9743. (p3) FMA f72 = f40, f57, f72 // A1 * B2
  9744. br.cloop.sptk.few .L122
  9745. }
  9746. ;;
  9747. .L128:
  9748. adds AOFFSET2 = 4 * SIZE, AOFFSET
  9749. adds BOFFSET2 = 4 * SIZE, BOFFSET
  9750. ;;
  9751. #ifdef LT
  9752. LDFPD f32, f33 = [BOFFSET]
  9753. ;;
  9754. FSUB f64 = f32, f64
  9755. FSUB f72 = f33, f72
  9756. ;;
  9757. #else
  9758. LDFPD f32, f33 = [AOFFSET]
  9759. ;;
  9760. FSUB f64 = f32, f64
  9761. FSUB f72 = f33, f72
  9762. ;;
  9763. #endif
  9764. #ifdef LT
  9765. LDFD f32 = [AOFFSET]
  9766. ;;
  9767. FMPY f64 = f64, f32
  9768. FMPY f72 = f72, f32
  9769. ;;
  9770. STFD [BOFFSET] = f64, SIZE
  9771. ;;
  9772. STFD [BOFFSET] = f72, -SIZE
  9773. ;;
  9774. #endif
  9775. #ifdef RN
  9776. LDFPD f32, f33 = [BOFFSET]
  9777. adds BOFFSET = 3 * SIZE, BOFFSET
  9778. ;;
  9779. LDFD f34 = [BOFFSET], -3 * SIZE
  9780. ;;
  9781. FMPY f64 = f64, f32
  9782. ;;
  9783. FNMA f72 = f64, f33, f72
  9784. ;;
  9785. FMPY f72 = f72, f34
  9786. ;;
  9787. STFD [AOFFSET] = f64, SIZE
  9788. ;;
  9789. STFD [AOFFSET] = f72, -SIZE
  9790. ;;
  9791. #endif
  9792. STFD [C1 ] = f64, SIZE
  9793. STFD [C2 ] = f72, SIZE
  9794. mov f64 = f0
  9795. mov f72 = f0
  9796. ;;
  9797. sub L = K, KK
  9798. ;;
  9799. shladd L = L, BASE_SHIFT, r0
  9800. ;;
  9801. add AOFFSET = L, AOFFSET
  9802. ;;
  9803. shladd BOFFSET = L, 1, BOFFSET
  9804. ;;
  9805. #ifdef LT
  9806. adds KK = 1, KK
  9807. #else
  9808. nop __LINE__
  9809. #endif
  9810. ;;
  9811. mov L = KK
  9812. ;;
  9813. .align 8
  9814. .L129:
  9815. mov B = BOFFSET
  9816. #ifdef RN
  9817. adds KK = 2, KK
  9818. #endif
  9819. ;;
  9820. mov AOFFSET = A
  9821. ;;
  9822. .align 16
  9823. .L130:
  9824. tbit.z p6, p0 = N, 0
  9825. (p6) br.cond.dpnt .L999
  9826. ;;
  9827. mov f64 = f0
  9828. mov f65 = f0
  9829. mov f66 = f0
  9830. mov f67 = f0
  9831. mov f68 = f0
  9832. mov f69 = f0
  9833. mov f70 = f0
  9834. mov f71 = f0
  9835. ;;
  9836. { .mfi
  9837. shr I = M, 3
  9838. }
  9839. { .mfi
  9840. mov C1 = C // coffset1 = c + 0 * ldc
  9841. #ifdef LT
  9842. mov KK = OFFSET
  9843. #else
  9844. nop __LINE__
  9845. #endif
  9846. }
  9847. ;;
  9848. { .mmf
  9849. cmp.eq p6, p7 = 0, I
  9850. mov AORIG = A
  9851. }
  9852. ;;
  9853. { .mfi
  9854. add C = C, LDC // coffset += 8 * ldc
  9855. mov L = KK
  9856. }{ .mfb
  9857. (p6) br.cond.dpnt .L140
  9858. }
  9859. ;;
  9860. .align 16
  9861. .L132:
  9862. { .mmi
  9863. cmp.ne p7, p0 = r0, L
  9864. adds BOFFSET = 0 * SIZE, B
  9865. }
  9866. ;;
  9867. { .mmi
  9868. (p7) LDFD f48 = [BOFFSET], 1 * SIZE
  9869. nop __LINE__
  9870. nop __LINE__
  9871. }
  9872. ;;
  9873. (p7) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  9874. ;;
  9875. { .mmf
  9876. (p7) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  9877. }
  9878. ;;
  9879. { .mmf
  9880. (p7) LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  9881. }
  9882. { .mfi
  9883. cmp.eq p3, p0 = r0, r0
  9884. }
  9885. ;;
  9886. { .mmf
  9887. (p7) LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  9888. }
  9889. { .mfi
  9890. adds PREC = CPREFETCHSIZE * SIZE, C1
  9891. }
  9892. ;;
  9893. { .mmf
  9894. CPREFETCH [PREC]
  9895. }
  9896. { .mfi
  9897. adds L = 1, L
  9898. }
  9899. ;;
  9900. { .mfi
  9901. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  9902. }
  9903. ;;
  9904. { .mfi
  9905. adds PREB = (PREFETCHSIZE - 8) * SIZE, BOFFSET
  9906. }
  9907. ;;
  9908. { .mfi
  9909. tbit.z p12, p0 = L, 0
  9910. }
  9911. { .mfi
  9912. shr L = L, 1
  9913. }
  9914. ;;
  9915. { .mfi
  9916. adds L = -1, L
  9917. }
  9918. ;;
  9919. { .mfi
  9920. mov ar.lc = L
  9921. }
  9922. ;;
  9923. { .mfb
  9924. cmp.eq p6, p0 = -1, L
  9925. (p6) br.cond.dpnt .L138
  9926. }
  9927. ;;
  9928. .align 16
  9929. .L133:
  9930. { .mfi
  9931. lfetch.nt1 [PREA], 16 * SIZE
  9932. FMA f64 = f32, f48, f64 // A1 * B1
  9933. cmp.ne p4, p5 = 0, L
  9934. }
  9935. { .mfi
  9936. adds PREB = (PREFETCHSIZE + 0) * SIZE, BOFFSET
  9937. FMA f65 = f33, f48, f65 // A2 * B1
  9938. (p12) cmp.ne p3, p0 = 0, L
  9939. }
  9940. ;;
  9941. { .mfi
  9942. (p3) LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  9943. FMA f66 = f34, f48, f66 // A3 * B1
  9944. adds C9 = 4 * SIZE, C1
  9945. }
  9946. { .mmf
  9947. (p3) LDFD f56 = [BOFFSET], 1 * SIZE
  9948. nop __LINE__
  9949. FMA f67 = f35, f48, f67 // A4 * B1
  9950. }
  9951. ;;
  9952. { .mfb
  9953. (p3) LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  9954. FMA f68 = f36, f48, f68 // A5 * B1
  9955. nop __LINE__
  9956. }
  9957. { .mfb
  9958. nop __LINE__
  9959. FMA f69 = f37, f48, f69 // A6 * B1
  9960. nop __LINE__
  9961. }
  9962. ;;
  9963. { .mfb
  9964. (p3) LDFPD f44, f45 = [AOFFSET], 2 * SIZE
  9965. FMA f70 = f38, f48, f70 // A7 * B1
  9966. nop __LINE__
  9967. }
  9968. { .mfb
  9969. nop __LINE__
  9970. FMA f71 = f39, f48, f71 // A8 * B1
  9971. nop __LINE__
  9972. }
  9973. ;;
  9974. { .mfb
  9975. (p3) LDFPD f46, f47 = [AOFFSET], 2 * SIZE
  9976. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  9977. nop __LINE__
  9978. }
  9979. { .mfb
  9980. nop __LINE__
  9981. (p3) FMA f65 = f41, f56, f65 // A2 * B1
  9982. nop __LINE__
  9983. }
  9984. ;;
  9985. { .mfb
  9986. (p4) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  9987. (p3) FMA f66 = f42, f56, f66 // A3 * B1
  9988. nop __LINE__
  9989. }
  9990. { .mmf
  9991. (p4) LDFD f48 = [BOFFSET], 1 * SIZE
  9992. nop __LINE__
  9993. (p3) FMA f67 = f43, f56, f67 // A4 * B1
  9994. }
  9995. ;;
  9996. { .mfb
  9997. (p4) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  9998. (p3) FMA f68 = f44, f56, f68 // A5 * B1
  9999. nop __LINE__
  10000. }
  10001. { .mfb
  10002. nop __LINE__
  10003. (p3) FMA f69 = f45, f56, f69 // A6 * B1
  10004. nop __LINE__
  10005. }
  10006. ;;
  10007. { .mfi
  10008. (p4) LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  10009. (p3) FMA f70 = f46, f56, f70 // A7 * B1
  10010. adds L = -1, L
  10011. }
  10012. { .mfb
  10013. nop __LINE__
  10014. (p3) FMA f71 = f47, f56, f71 // A8 * B1
  10015. nop __LINE__
  10016. }
  10017. ;;
  10018. { .mfb
  10019. (p4) LDFPD f38, f39 = [AOFFSET], 2 * SIZE
  10020. nop __LINE__
  10021. br.cloop.sptk.few .L133
  10022. }
  10023. ;;
  10024. .L138:
  10025. adds AOFFSET2 = 4 * SIZE, AOFFSET
  10026. adds BOFFSET2 = 4 * SIZE, BOFFSET
  10027. ;;
  10028. #ifdef LT
  10029. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  10030. ;;
  10031. LDFPD f34, f35 = [BOFFSET], 2 * SIZE
  10032. ;;
  10033. LDFPD f36, f37 = [BOFFSET], 2 * SIZE
  10034. ;;
  10035. LDFPD f38, f39 = [BOFFSET]
  10036. adds BOFFSET = -6 * SIZE, BOFFSET
  10037. ;;
  10038. FSUB f64 = f32, f64
  10039. FSUB f65 = f33, f65
  10040. FSUB f66 = f34, f66
  10041. FSUB f67 = f35, f67
  10042. FSUB f68 = f36, f68
  10043. FSUB f69 = f37, f69
  10044. FSUB f70 = f38, f70
  10045. FSUB f71 = f39, f71
  10046. ;;
  10047. #else
  10048. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  10049. ;;
  10050. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  10051. ;;
  10052. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  10053. ;;
  10054. LDFPD f38, f39 = [AOFFSET]
  10055. adds AOFFSET = -6 * SIZE, AOFFSET
  10056. ;;
  10057. FSUB f64 = f32, f64
  10058. FSUB f65 = f33, f65
  10059. FSUB f66 = f34, f66
  10060. FSUB f67 = f35, f67
  10061. FSUB f68 = f36, f68
  10062. FSUB f69 = f37, f69
  10063. FSUB f70 = f38, f70
  10064. FSUB f71 = f39, f71
  10065. ;;
  10066. #endif
  10067. #ifdef LT
  10068. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  10069. ;;
  10070. LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  10071. ;;
  10072. LDFPD f36, f37 = [AOFFSET], 2 * SIZE
  10073. ;;
  10074. LDFPD f38, f39 = [AOFFSET]
  10075. adds AOFFSET = 3 * SIZE, AOFFSET
  10076. ;;
  10077. LDFD f40 = [AOFFSET], 1 * SIZE
  10078. ;;
  10079. LDFPD f41, f42 = [AOFFSET], 2 * SIZE
  10080. ;;
  10081. LDFPD f43, f44 = [AOFFSET], 2 * SIZE
  10082. ;;
  10083. LDFPD f45, f46 = [AOFFSET]
  10084. adds AOFFSET = 4 * SIZE, AOFFSET
  10085. ;;
  10086. LDFPD f47, f48 = [AOFFSET], 2 * SIZE
  10087. ;;
  10088. LDFPD f49, f50 = [AOFFSET], 2 * SIZE
  10089. ;;
  10090. LDFPD f51, f52 = [AOFFSET]
  10091. adds AOFFSET = 5 * SIZE, AOFFSET
  10092. ;;
  10093. LDFD f53 = [AOFFSET], 1 * SIZE
  10094. ;;
  10095. LDFPD f54, f55 = [AOFFSET], 2 * SIZE
  10096. ;;
  10097. LDFPD f56, f57 = [AOFFSET]
  10098. adds AOFFSET = 6 * SIZE, AOFFSET
  10099. ;;
  10100. LDFPD f58, f59 = [AOFFSET], 2 * SIZE
  10101. ;;
  10102. LDFPD f60, f61 = [AOFFSET]
  10103. adds AOFFSET = 7 * SIZE, AOFFSET
  10104. ;;
  10105. LDFD f16 = [AOFFSET], 1 * SIZE
  10106. ;;
  10107. LDFPD f17, f18 = [AOFFSET]
  10108. adds AOFFSET = 8 * SIZE, AOFFSET
  10109. ;;
  10110. LDFPD f19, f20 = [AOFFSET]
  10111. adds AOFFSET = 9 * SIZE, AOFFSET
  10112. ;;
  10113. LDFD f21 = [AOFFSET]
  10114. adds AOFFSET = -63 * SIZE, AOFFSET
  10115. ;;
  10116. FMPY f64 = f64, f32
  10117. ;;
  10118. FNMA f65 = f64, f33, f65
  10119. ;;
  10120. FNMA f66 = f64, f34, f66
  10121. ;;
  10122. FNMA f67 = f64, f35, f67
  10123. ;;
  10124. FNMA f68 = f64, f36, f68
  10125. ;;
  10126. FNMA f69 = f64, f37, f69
  10127. ;;
  10128. FNMA f70 = f64, f38, f70
  10129. ;;
  10130. FNMA f71 = f64, f39, f71
  10131. ;;
  10132. FMPY f65 = f65, f40
  10133. ;;
  10134. FNMA f66 = f65, f41, f66
  10135. ;;
  10136. FNMA f67 = f65, f42, f67
  10137. ;;
  10138. FNMA f68 = f65, f43, f68
  10139. ;;
  10140. FNMA f69 = f65, f44, f69
  10141. ;;
  10142. FNMA f70 = f65, f45, f70
  10143. ;;
  10144. FNMA f71 = f65, f46, f71
  10145. ;;
  10146. FMPY f66 = f66, f47
  10147. ;;
  10148. FNMA f67 = f66, f48, f67
  10149. ;;
  10150. FNMA f68 = f66, f49, f68
  10151. ;;
  10152. FNMA f69 = f66, f50, f69
  10153. ;;
  10154. FNMA f70 = f66, f51, f70
  10155. ;;
  10156. FNMA f71 = f66, f52, f71
  10157. ;;
  10158. FMPY f67 = f67, f53
  10159. ;;
  10160. FNMA f68 = f67, f54, f68
  10161. ;;
  10162. FNMA f69 = f67, f55, f69
  10163. ;;
  10164. FNMA f70 = f67, f56, f70
  10165. ;;
  10166. FNMA f71 = f67, f57, f71
  10167. ;;
  10168. FMPY f68 = f68, f58
  10169. ;;
  10170. FNMA f69 = f68, f59, f69
  10171. ;;
  10172. FNMA f70 = f68, f60, f70
  10173. ;;
  10174. FNMA f71 = f68, f61, f71
  10175. ;;
  10176. FMPY f69 = f69, f16
  10177. ;;
  10178. FNMA f70 = f69, f17, f70
  10179. ;;
  10180. FNMA f71 = f69, f18, f71
  10181. ;;
  10182. FMPY f70 = f70, f19
  10183. ;;
  10184. FNMA f71 = f70, f20, f71
  10185. ;;
  10186. FMPY f71 = f71, f21
  10187. ;;
  10188. STFD [BOFFSET] = f64, SIZE
  10189. STFD [BOFFSET2] = f68, SIZE
  10190. ;;
  10191. STFD [BOFFSET] = f65, SIZE
  10192. STFD [BOFFSET2] = f69, SIZE
  10193. ;;
  10194. STFD [BOFFSET] = f66, SIZE
  10195. STFD [BOFFSET2] = f70, SIZE
  10196. ;;
  10197. STFD [BOFFSET] = f67, -3 * SIZE
  10198. STFD [BOFFSET2] = f71, -3 * SIZE
  10199. ;;
  10200. adds C9 = 4 * SIZE, C1
  10201. ;;
  10202. #endif
  10203. #ifdef RN
  10204. LDFD f32 = [BOFFSET]
  10205. ;;
  10206. FMPY f64 = f64, f32
  10207. FMPY f68 = f68, f32
  10208. FMPY f65 = f65, f32
  10209. FMPY f69 = f69, f32
  10210. FMPY f66 = f66, f32
  10211. FMPY f70 = f70, f32
  10212. FMPY f67 = f67, f32
  10213. FMPY f71 = f71, f32
  10214. ;;
  10215. STFD [AOFFSET] = f64, SIZE
  10216. STFD [AOFFSET2] = f68, SIZE
  10217. ;;
  10218. STFD [AOFFSET] = f65, SIZE
  10219. STFD [AOFFSET2] = f69, SIZE
  10220. ;;
  10221. STFD [AOFFSET] = f66, SIZE
  10222. STFD [AOFFSET2] = f70, SIZE
  10223. ;;
  10224. STFD [AOFFSET] = f67, -3 * SIZE
  10225. STFD [AOFFSET2] = f71, -3 * SIZE
  10226. ;;
  10227. #endif
  10228. adds C9 = 4 * SIZE, C1
  10229. ;;
  10230. { .mmf
  10231. STFD [C1 ] = f64, SIZE
  10232. STFD [C9 ] = f68, SIZE
  10233. mov f64 = f0
  10234. }
  10235. ;;
  10236. { .mmi
  10237. STFD [C1 ] = f65, SIZE
  10238. STFD [C9 ] = f69, SIZE
  10239. }
  10240. ;;
  10241. { .mmi
  10242. STFD [C1 ] = f66, SIZE
  10243. STFD [C9 ] = f70, SIZE
  10244. }
  10245. ;;
  10246. { .mmi
  10247. STFD [C1 ] = f67, 5 * SIZE
  10248. STFD [C9 ] = f71
  10249. }
  10250. ;;
  10251. { .mmf
  10252. cmp.ne p6, p0 = 1, I
  10253. }
  10254. ;;
  10255. adds I = -1, I
  10256. ;;
  10257. { .mmi
  10258. sub L = K, KK
  10259. }
  10260. ;;
  10261. { .mmi
  10262. shladd L = L, BASE_SHIFT, r0
  10263. }
  10264. ;;
  10265. ;;
  10266. { .mmi
  10267. shladd AOFFSET = L, 3, AOFFSET
  10268. }
  10269. ;;
  10270. { .mmi
  10271. add BOFFSET = L, BOFFSET
  10272. }
  10273. ;;
  10274. { .mmi
  10275. #ifdef LT
  10276. adds KK = 8, KK
  10277. #else
  10278. nop __LINE__
  10279. #endif
  10280. }
  10281. ;;
  10282. { .mmi
  10283. mov L = KK
  10284. }
  10285. ;;
  10286. mov f64 = f0
  10287. mov f65 = f0
  10288. mov f66 = f0
  10289. mov f67 = f0
  10290. mov f68 = f0
  10291. mov f69 = f0
  10292. mov f70 = f0
  10293. mov f71 = f0
  10294. (p6) br.cond.dptk .L132
  10295. .align 8
  10296. .L140:
  10297. tbit.z p6, p7 = M, 2
  10298. (p6) br.cond.dptk .L150
  10299. ;;
  10300. { .mib
  10301. mov L = KK
  10302. }
  10303. ;;
  10304. { .mmi
  10305. cmp.ne p7, p0 = r0, L
  10306. adds BOFFSET = 0 * SIZE, B
  10307. }
  10308. ;;
  10309. { .mmf
  10310. (p7) LDFD f48 = [BOFFSET], 1 * SIZE
  10311. mov f65 = f0
  10312. }
  10313. ;;
  10314. { .mfi
  10315. adds L = 1, L
  10316. }
  10317. { .mfi
  10318. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  10319. cmp.eq p3, p0 = r0, r0
  10320. }
  10321. ;;
  10322. { .mfi
  10323. tbit.z p12, p0 = L, 0
  10324. }
  10325. { .mfi
  10326. shr L = L, 1
  10327. }
  10328. ;;
  10329. { .mfi
  10330. adds L = -1, L
  10331. }
  10332. ;;
  10333. { .mfi
  10334. cmp.eq p6, p0 = -1, L
  10335. }
  10336. ;;
  10337. { .mmf
  10338. (p7) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  10339. }
  10340. { .mfi
  10341. mov ar.lc = L
  10342. }
  10343. ;;
  10344. { .mmf
  10345. (p7) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  10346. }
  10347. { .mfb
  10348. (p6) br.cond.dpnt .L148
  10349. }
  10350. ;;
  10351. .L142:
  10352. { .mfi
  10353. lfetch.nt1 [PREA], 8 * SIZE
  10354. FMA f64 = f32, f48, f64 // A1 * B1
  10355. cmp.ne p4, p5 = 0, L
  10356. }
  10357. { .mfi
  10358. nop __LINE__
  10359. FMA f65 = f33, f48, f65 // A2 * B1
  10360. (p12) cmp.ne p3, p0 = 0, L
  10361. }
  10362. ;;
  10363. { .mfi
  10364. (p3) LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  10365. FMA f66 = f34, f48, f66 // A3 * B1
  10366. (p5) adds C9 = 2 * SIZE, C1
  10367. }
  10368. { .mmf
  10369. nop __LINE__
  10370. (p3) LDFD f56 = [BOFFSET], 1 * SIZE
  10371. FMA f67 = f35, f48, f67 // A4 * B1
  10372. }
  10373. ;;
  10374. { .mfi
  10375. (p3) LDFPD f42, f43 = [AOFFSET], 2 * SIZE
  10376. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  10377. (p5) adds C10 = 2 * SIZE, C2
  10378. }
  10379. { .mfb
  10380. nop __LINE__
  10381. (p3) FMA f65 = f41, f56, f65 // A2 * B1
  10382. nop __LINE__
  10383. }
  10384. ;;
  10385. { .mfb
  10386. (p4) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  10387. (p3) FMA f66 = f42, f56, f66 // A3 * B1
  10388. nop __LINE__
  10389. }
  10390. { .mmf
  10391. (p4) LDFD f48 = [BOFFSET], 1 * SIZE
  10392. nop __LINE__
  10393. (p3) FMA f67 = f43, f56, f67 // A4 * B1
  10394. }
  10395. ;;
  10396. { .mfi
  10397. (p4) LDFPD f34, f35 = [AOFFSET], 2 * SIZE
  10398. nop __LINE__
  10399. adds L = -1, L
  10400. }
  10401. { .mfb
  10402. nop __LINE__
  10403. nop.f 0
  10404. br.cloop.sptk.few .L142
  10405. }
  10406. ;;
  10407. .L148:
  10408. adds AOFFSET2 = 4 * SIZE, AOFFSET
  10409. adds BOFFSET2 = 4 * SIZE, BOFFSET
  10410. ;;
  10411. #ifdef LT
  10412. LDFPD f32, f33 = [BOFFSET], 2 * SIZE
  10413. ;;
  10414. LDFPD f34, f35 = [BOFFSET]
  10415. adds BOFFSET = -2 * SIZE, BOFFSET
  10416. ;;
  10417. FSUB f64 = f32, f64
  10418. FSUB f65 = f33, f65
  10419. FSUB f66 = f34, f66
  10420. FSUB f67 = f35, f67
  10421. ;;
  10422. #else
  10423. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  10424. ;;
  10425. LDFPD f34, f35 = [AOFFSET]
  10426. adds AOFFSET = -2 * SIZE, AOFFSET
  10427. ;;
  10428. FSUB f64 = f32, f64
  10429. FSUB f65 = f33, f65
  10430. FSUB f66 = f34, f66
  10431. FSUB f67 = f35, f67
  10432. ;;
  10433. #endif
  10434. #ifdef LT
  10435. LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  10436. ;;
  10437. LDFPD f34, f35 = [AOFFSET]
  10438. adds AOFFSET = 3 * SIZE, AOFFSET
  10439. ;;
  10440. LDFD f36 = [AOFFSET], 1 * SIZE
  10441. ;;
  10442. LDFPD f37, f38 = [AOFFSET]
  10443. adds AOFFSET = 4 * SIZE, AOFFSET
  10444. ;;
  10445. LDFPD f39, f40 = [AOFFSET]
  10446. adds AOFFSET = 5 * SIZE, AOFFSET
  10447. ;;
  10448. LDFD f41 = [AOFFSET], -15 * SIZE
  10449. ;;
  10450. FMPY f64 = f64, f32
  10451. ;;
  10452. FNMA f65 = f64, f33, f65
  10453. ;;
  10454. FNMA f66 = f64, f34, f66
  10455. ;;
  10456. FNMA f67 = f64, f35, f67
  10457. ;;
  10458. FMPY f65 = f65, f36
  10459. ;;
  10460. FNMA f66 = f65, f37, f66
  10461. ;;
  10462. FNMA f67 = f65, f38, f67
  10463. ;;
  10464. FMPY f66 = f66, f39
  10465. ;;
  10466. FNMA f67 = f66, f40, f67
  10467. ;;
  10468. FMPY f67 = f67, f41
  10469. ;;
  10470. STFD [BOFFSET] = f64, SIZE
  10471. ;;
  10472. STFD [BOFFSET] = f65, SIZE
  10473. ;;
  10474. STFD [BOFFSET] = f66, SIZE
  10475. ;;
  10476. STFD [BOFFSET] = f67, -3 * SIZE
  10477. ;;
  10478. #endif
  10479. #ifdef RN
  10480. LDFD f32 = [BOFFSET]
  10481. ;;
  10482. FMPY f64 = f64, f32
  10483. FMPY f65 = f65, f32
  10484. FMPY f66 = f66, f32
  10485. FMPY f67 = f67, f32
  10486. ;;
  10487. STFD [AOFFSET] = f64, SIZE
  10488. ;;
  10489. STFD [AOFFSET] = f65, SIZE
  10490. ;;
  10491. STFD [AOFFSET] = f66, SIZE
  10492. ;;
  10493. STFD [AOFFSET] = f67, -3 * SIZE
  10494. ;;
  10495. #endif
  10496. { .mmf
  10497. STFD [C1 ] = f64, SIZE
  10498. mov f64 = f0
  10499. }
  10500. ;;
  10501. { .mmi
  10502. STFD [C1 ] = f65, SIZE
  10503. }
  10504. ;;
  10505. { .mmi
  10506. STFD [C1 ] = f66, SIZE
  10507. }
  10508. ;;
  10509. { .mmi
  10510. STFD [C1 ] = f67, SIZE
  10511. }
  10512. ;;
  10513. { .mmf
  10514. mov f72 = f0
  10515. }
  10516. ;;
  10517. mov f65 = f0
  10518. mov f73 = f0
  10519. mov f66 = f0
  10520. mov f74 = f0
  10521. mov f67 = f0
  10522. mov f75 = f0
  10523. ;;
  10524. { .mmi
  10525. sub L = K, KK
  10526. }
  10527. ;;
  10528. { .mmi
  10529. shladd L = L, BASE_SHIFT, r0
  10530. }
  10531. ;;
  10532. { .mmi
  10533. shladd AOFFSET = L, 2, AOFFSET
  10534. }
  10535. ;;
  10536. { .mmi
  10537. add BOFFSET = L, BOFFSET
  10538. }
  10539. ;;
  10540. { .mmi
  10541. #ifdef LT
  10542. adds KK = 4, KK
  10543. #else
  10544. nop __LINE__
  10545. #endif
  10546. }
  10547. ;;
  10548. .align 8
  10549. .L150:
  10550. tbit.z p6, p7 = M, 1
  10551. (p6) br.cond.dptk .L160
  10552. ;;
  10553. { .mib
  10554. mov L = KK
  10555. }
  10556. ;;
  10557. { .mmi
  10558. cmp.ne p7, p0 = r0, L
  10559. adds BOFFSET = 0 * SIZE, B
  10560. }
  10561. ;;
  10562. { .mmf
  10563. (p7) LDFD f48 = [BOFFSET], 1 * SIZE
  10564. }
  10565. ;;
  10566. { .mfi
  10567. adds L = 1, L
  10568. }
  10569. { .mfi
  10570. adds PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
  10571. cmp.eq p3, p0 = r0, r0
  10572. }
  10573. ;;
  10574. { .mfi
  10575. tbit.z p12, p0 = L, 0
  10576. }
  10577. { .mfi
  10578. shr L = L, 1
  10579. }
  10580. ;;
  10581. { .mmf
  10582. adds L = -1, L
  10583. }
  10584. ;;
  10585. { .mmf
  10586. cmp.eq p6, p0 = -1, L
  10587. }
  10588. ;;
  10589. { .mib
  10590. (p7) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  10591. mov ar.lc = L
  10592. (p6) br.cond.dpnt .L158
  10593. }
  10594. ;;
  10595. .L152:
  10596. { .mfi
  10597. cmp.ne p4, p5 = 0, L
  10598. FMA f64 = f32, f48, f64 // A1 * B1
  10599. (p12) cmp.ne p3, p0 = 0, L
  10600. }
  10601. ;;
  10602. { .mmf
  10603. (p3) LDFD f56 = [BOFFSET], 1 * SIZE
  10604. (p3) LDFPD f40, f41 = [AOFFSET], 2 * SIZE
  10605. FMA f65 = f33, f48, f65 // A2 * B1
  10606. }
  10607. ;;
  10608. { .mfi
  10609. (p4) LDFPD f32, f33 = [AOFFSET], 2 * SIZE
  10610. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  10611. adds L = -1, L
  10612. }
  10613. ;;
  10614. { .mfb
  10615. (p4) LDFD f48 = [BOFFSET], 1 * SIZE
  10616. (p3) FMA f65 = f41, f56, f65 // A2 * B1
  10617. br.cloop.sptk.few .L152
  10618. }
  10619. ;;
  10620. .L158:
  10621. adds AOFFSET2 = 4 * SIZE, AOFFSET
  10622. adds BOFFSET2 = 4 * SIZE, BOFFSET
  10623. ;;
  10624. #ifdef LT
  10625. LDFPD f32, f33 = [BOFFSET]
  10626. ;;
  10627. FSUB f64 = f32, f64
  10628. FSUB f65 = f33, f65
  10629. ;;
  10630. #else
  10631. LDFPD f32, f33 = [AOFFSET]
  10632. ;;
  10633. FSUB f64 = f32, f64
  10634. FSUB f65 = f33, f65
  10635. ;;
  10636. #endif
  10637. #ifdef LT
  10638. LDFPD f32, f33 = [AOFFSET]
  10639. adds AOFFSET = 3 * SIZE, AOFFSET
  10640. ;;
  10641. LDFD f34 = [AOFFSET], - 3 * SIZE
  10642. ;;
  10643. FMPY f64 = f64, f32
  10644. ;;
  10645. FNMA f65 = f64, f33, f65
  10646. ;;
  10647. FMPY f65 = f65, f34
  10648. ;;
  10649. STFD [BOFFSET] = f64, SIZE
  10650. ;;
  10651. STFD [BOFFSET] = f65, -SIZE
  10652. ;;
  10653. #endif
  10654. #ifdef RN
  10655. LDFD f32 = [BOFFSET]
  10656. ;;
  10657. FMPY f64 = f64, f32
  10658. FMPY f65 = f65, f32
  10659. ;;
  10660. STFD [AOFFSET] = f64, SIZE
  10661. ;;
  10662. STFD [AOFFSET] = f65, - SIZE
  10663. ;;
  10664. #endif
  10665. STFD [C1 ] = f64, SIZE
  10666. ;;
  10667. STFD [C1 ] = f65, SIZE
  10668. ;;
  10669. mov f64 = f0
  10670. mov f65 = f0
  10671. ;;
  10672. sub L = K, KK
  10673. ;;
  10674. { .mmi
  10675. shladd L = L, BASE_SHIFT, r0
  10676. }
  10677. ;;
  10678. { .mmi
  10679. shladd AOFFSET = L, 1, AOFFSET
  10680. }
  10681. ;;
  10682. { .mmi
  10683. add BOFFSET = L, BOFFSET
  10684. }
  10685. ;;
  10686. { .mmi
  10687. #ifdef LT
  10688. adds KK = 2, KK
  10689. #else
  10690. nop __LINE__
  10691. #endif
  10692. }
  10693. ;;
  10694. .align 8
  10695. .L160:
  10696. { .mib
  10697. mov L = KK
  10698. tbit.z p6, p7 = M, 0
  10699. (p6) br.cond.dptk .L169
  10700. }
  10701. ;;
  10702. { .mmi
  10703. cmp.ne p7, p0 = r0, L
  10704. adds BOFFSET = 0 * SIZE, B
  10705. }
  10706. ;;
  10707. { .mmi
  10708. (p7) LDFD f48 = [BOFFSET], 1 * SIZE
  10709. nop __LINE__
  10710. adds L = 1, L
  10711. }
  10712. ;;
  10713. { .mii
  10714. tbit.z p12, p0 = L, 0
  10715. shr L = L, 1
  10716. }
  10717. ;;
  10718. { .mmi
  10719. cmp.eq p6, p0 = 0, L
  10720. adds L = -1, L
  10721. cmp.eq p3, p0 = r0, r0
  10722. }
  10723. ;;
  10724. { .mib
  10725. (p7) LDFD f32 = [AOFFSET], 1 * SIZE
  10726. mov ar.lc = L
  10727. (p6) br.cond.dpnt .L168
  10728. }
  10729. ;;
  10730. .align 8
  10731. .L162:
  10732. { .mmf
  10733. cmp.ne p4, p5 = 0, L
  10734. (p12) cmp.ne p3, p0 = 0, L
  10735. FMA f64 = f32, f48, f64 // A1 * B1
  10736. }
  10737. ;;
  10738. { .mmi
  10739. (p3) LDFD f56 = [BOFFSET], 1 * SIZE
  10740. (p3) LDFD f40 = [AOFFSET], 1 * SIZE
  10741. nop __LINE__
  10742. }
  10743. ;;
  10744. { .mmi
  10745. (p4) LDFD f32 = [AOFFSET], 1 * SIZE
  10746. nop __LINE__
  10747. adds L = -1, L
  10748. }
  10749. { .mfb
  10750. (p4) LDFD f48 = [BOFFSET], 1 * SIZE
  10751. (p3) FMA f64 = f40, f56, f64 // A1 * B1
  10752. br.cloop.sptk.few .L162
  10753. }
  10754. ;;
  10755. .align 8
  10756. .L168:
  10757. #ifdef LT
  10758. { .mmi
  10759. LDFD f32 = [BOFFSET]
  10760. LDFD f33 = [AOFFSET]
  10761. nop __LINE__
  10762. }
  10763. ;;
  10764. #else
  10765. { .mmi
  10766. LDFD f32 = [AOFFSET]
  10767. LDFD f33 = [BOFFSET]
  10768. nop __LINE__
  10769. }
  10770. ;;
  10771. #endif
  10772. { .mmf
  10773. sub L = K, KK
  10774. nop __LINE__
  10775. FSUB f64 = f32, f64
  10776. }
  10777. ;;
  10778. #ifdef LT
  10779. adds KK = 1, KK
  10780. #else
  10781. nop __LINE__
  10782. #endif
  10783. ;;
  10784. mov L = KK
  10785. ;;
  10786. FMPY f64 = f64, f33
  10787. ;;
  10788. #ifdef LT
  10789. { .mmf
  10790. STFD [BOFFSET] = f64
  10791. STFD [C1 ] = f64, SIZE
  10792. mov f64 = f0
  10793. }
  10794. ;;
  10795. #else
  10796. { .mmf
  10797. STFD [AOFFSET] = f64
  10798. STFD [C1 ] = f64, SIZE
  10799. mov f64 = f0
  10800. }
  10801. ;;
  10802. #endif
  10803. shladd AOFFSET = L, BASE_SHIFT, AOFFSET
  10804. shladd BOFFSET = L, BASE_SHIFT, BOFFSET
  10805. ;;
  10806. .align 8
  10807. .L169:
  10808. { .mii
  10809. mov B = BOFFSET
  10810. #ifdef RN
  10811. adds KK = 1, KK
  10812. #else
  10813. nop __LINE__
  10814. #endif
  10815. mov AOFFSET = A
  10816. }
  10817. ;;
  10818. .align 16
  10819. .L999:
  10820. mov r8 = r0
  10821. adds r9 = 1 * 16, SP
  10822. ;;
  10823. ldf.fill f16 = [SP], 32
  10824. ldf.fill f17 = [r9], 32
  10825. ;;
  10826. ldf.fill f18 = [SP], 32
  10827. ldf.fill f19 = [r9], 32
  10828. ;;
  10829. ldf.fill f20 = [SP], 32
  10830. ldf.fill f21 = [r9], 32
  10831. ;;
  10832. mov ar.lc = ARLC
  10833. ;;
  10834. mov pr = PR, -1
  10835. ;;
  10836. mov ar.pfs = ARPFS
  10837. ;;
  10838. br.ret.sptk.many b0
  10839. EPILOGUE