a64.cpp 92 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157
  1. /* This file is part of the dynarmic project.
  2. * Copyright (c) 2018 MerryMage
  3. * SPDX-License-Identifier: 0BSD
  4. */
  5. #include <catch2/catch_test_macros.hpp>
  6. #include <oaknut/oaknut.hpp>
  7. #include "./testenv.h"
  8. #include "dynarmic/common/fp/fpsr.h"
  9. #include "dynarmic/interface/exclusive_monitor.h"
  10. using namespace Dynarmic;
  11. using namespace oaknut::util;
  12. TEST_CASE("A64: ADD", "[a64]") {
  13. A64TestEnv env;
  14. A64::Jit jit{A64::UserConfig{&env}};
  15. env.code_mem.emplace_back(0x8b020020); // ADD X0, X1, X2
  16. env.code_mem.emplace_back(0x14000000); // B .
  17. jit.SetRegister(0, 0);
  18. jit.SetRegister(1, 1);
  19. jit.SetRegister(2, 2);
  20. jit.SetPC(0);
  21. env.ticks_left = 2;
  22. jit.Run();
  23. REQUIRE(jit.GetRegister(0) == 3);
  24. REQUIRE(jit.GetRegister(1) == 1);
  25. REQUIRE(jit.GetRegister(2) == 2);
  26. REQUIRE(jit.GetPC() == 4);
  27. }
  28. TEST_CASE("A64: ADD{V,P}", "[a64]") {
  29. A64TestEnv env;
  30. A64::Jit jit{A64::UserConfig{&env}};
  31. env.code_mem.emplace_back(0x0E31B801); // ADDV b1, v0.8b
  32. env.code_mem.emplace_back(0x4E31B802); // ADDV b2, v0.16b
  33. env.code_mem.emplace_back(0x0E71B803); // ADDV h3, v0.4h
  34. env.code_mem.emplace_back(0x4E71B804); // ADDV h4, v0.8h
  35. env.code_mem.emplace_back(0x0EA0BC05); // ADDP v5.2s, v0.2s, v0.2s
  36. env.code_mem.emplace_back(0x4EB1B806); // ADDV s6, v0.4s
  37. env.code_mem.emplace_back(0x14000000); // B .
  38. jit.SetVector(0, {0x0101010101010101, 0x0101010101010101});
  39. jit.SetPC(0);
  40. env.ticks_left = 7;
  41. jit.Run();
  42. REQUIRE(jit.GetVector(1) == Vector{0x0000000000000008, 0x0000000000000000});
  43. REQUIRE(jit.GetVector(2) == Vector{0x0000000000000010, 0x0000000000000000});
  44. REQUIRE(jit.GetVector(3) == Vector{0x0000000000000404, 0x0000000000000000});
  45. REQUIRE(jit.GetVector(4) == Vector{0x0000000000000808, 0x0000000000000000});
  46. REQUIRE(jit.GetVector(5) == Vector{0x0202020202020202, 0x0000000000000000});
  47. REQUIRE(jit.GetVector(6) == Vector{0x0000000004040404, 0x0000000000000000});
  48. }
  49. TEST_CASE("A64: CLZ", "[a64]") {
  50. A64TestEnv env;
  51. A64::Jit jit{A64::UserConfig{&env}};
  52. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  53. code.CLZ(V3.B16(), V0.B16());
  54. code.CLZ(V4.H8(), V1.H8());
  55. code.CLZ(V5.S4(), V2.S4());
  56. jit.SetPC(0);
  57. jit.SetVector(0, {0xeff0fafbfcfdfeff, 0xff7f3f1f0f070301});
  58. jit.SetVector(1, {0xfffcfffdfffeffff, 0x000F000700030001});
  59. jit.SetVector(2, {0xfffffffdfffffffe, 0x0000000300000001});
  60. env.ticks_left = env.code_mem.size();
  61. jit.Run();
  62. REQUIRE(jit.GetVector(3) == Vector{0x0, 0x0001020304050607});
  63. REQUIRE(jit.GetVector(4) == Vector{0x0, 0x000c000d000e000f});
  64. REQUIRE(jit.GetVector(5) == Vector{0x0, 0x0000001e0000001f});
  65. }
  66. TEST_CASE("A64: UADDL{V,P}", "[a64]") {
  67. A64TestEnv env;
  68. A64::Jit jit{A64::UserConfig{&env}};
  69. env.code_mem.emplace_back(0x2E303801); // UADDLV h1, v0.8b
  70. env.code_mem.emplace_back(0x6E303802); // UADDLV h2, v0.16b
  71. env.code_mem.emplace_back(0x2E703803); // UADDLV s3, v0.4h
  72. env.code_mem.emplace_back(0x6E703804); // UADDLV s4, v0.8h
  73. env.code_mem.emplace_back(0x2EA02805); // UADDLP v5.1d, v0.2s
  74. env.code_mem.emplace_back(0x6EB03806); // UADDLV d6, v0.4s
  75. env.code_mem.emplace_back(0x14000000); // B .
  76. jit.SetVector(0, {0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF});
  77. jit.SetPC(0);
  78. env.ticks_left = 7;
  79. jit.Run();
  80. REQUIRE(jit.GetVector(1) == Vector{0x00000000000007f8, 0x0000000000000000});
  81. REQUIRE(jit.GetVector(2) == Vector{0x0000000000000ff0, 0x0000000000000000});
  82. REQUIRE(jit.GetVector(3) == Vector{0x000000000003fffc, 0x0000000000000000});
  83. REQUIRE(jit.GetVector(4) == Vector{0x000000000007fff8, 0x0000000000000000});
  84. REQUIRE(jit.GetVector(5) == Vector{0x00000001fffffffe, 0x0000000000000000});
  85. REQUIRE(jit.GetVector(6) == Vector{0x00000003fffffffc, 0x0000000000000000});
  86. }
  87. TEST_CASE("A64: SADDL{V,P}", "[a64]") {
  88. A64TestEnv env;
  89. A64::Jit jit{A64::UserConfig{&env}};
  90. env.code_mem.emplace_back(0x0E303801); // SADDLV h1, v0.8b
  91. env.code_mem.emplace_back(0x4E303802); // SADDLV h2, v0.16b
  92. env.code_mem.emplace_back(0x0E703803); // SADDLV s3, v0.4h
  93. env.code_mem.emplace_back(0x4E703804); // SADDLV s4, v0.8h
  94. env.code_mem.emplace_back(0x0EA02805); // SADDLP v5.1d, v0.2s
  95. env.code_mem.emplace_back(0x4EB03806); // SADDLV d6, v0.4s
  96. env.code_mem.emplace_back(0x14000000); // B .
  97. jit.SetVector(0, {0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF});
  98. jit.SetPC(0);
  99. env.ticks_left = 7;
  100. jit.Run();
  101. REQUIRE(jit.GetVector(1) == Vector{0x000000000000fff8, 0x0000000000000000});
  102. REQUIRE(jit.GetVector(2) == Vector{0x000000000000fff0, 0x0000000000000000});
  103. REQUIRE(jit.GetVector(3) == Vector{0x00000000fffffffc, 0x0000000000000000});
  104. REQUIRE(jit.GetVector(4) == Vector{0x00000000fffffff8, 0x0000000000000000});
  105. REQUIRE(jit.GetVector(5) == Vector{0xfffffffffffffffe, 0x0000000000000000});
  106. REQUIRE(jit.GetVector(6) == Vector{0xfffffffffffffffc, 0x0000000000000000});
  107. }
  108. TEST_CASE("A64: VQADD", "[a64]") {
  109. A64TestEnv env;
  110. A64::Jit jit{A64::UserConfig{&env}};
  111. env.code_mem.emplace_back(0x6e210c02); // UQADD v2.16b, v0.16b, v1.16b
  112. env.code_mem.emplace_back(0x4e210c03); // SQADD v3.16b, v0.16b, v1.16b
  113. env.code_mem.emplace_back(0x6e610c04); // UQADD v4.8h, v0.8h, v1.8h
  114. env.code_mem.emplace_back(0x4e610c05); // SQADD v5.8h, v0.8h, v1.8h
  115. env.code_mem.emplace_back(0x6ea10c06); // UQADD v6.4s, v0.4s, v1.4s
  116. env.code_mem.emplace_back(0x4ea10c07); // SQADD v7.4s, v0.4s, v1.4s
  117. env.code_mem.emplace_back(0x6ee10c08); // UQADD v8.2d, v0.2d, v1.2d
  118. env.code_mem.emplace_back(0x4ee10c09); // SQADD v9.2d, v0.2d, v1.2d
  119. env.code_mem.emplace_back(0x14000000); // B .
  120. jit.SetVector(0, {0x7F7F7F7F7F7F7F7F, 0x7FFFFFFF7FFF7FFF});
  121. jit.SetVector(1, {0x8010FF00807F0000, 0x8000000080008000});
  122. jit.SetPC(0);
  123. env.ticks_left = 9;
  124. jit.Run();
  125. REQUIRE(jit.GetVector(2) == Vector{0xff8fff7ffffe7f7f, 0xffffffffffffffff});
  126. REQUIRE(jit.GetVector(3) == Vector{0xff7f7e7fff7f7f7f, 0xffffffffffffffff});
  127. REQUIRE(jit.GetVector(4) == Vector{0xff8ffffffffe7f7f, 0xffffffffffffffff});
  128. REQUIRE(jit.GetVector(5) == Vector{0xff8f7e7ffffe7f7f, 0xffffffffffffffff});
  129. REQUIRE(jit.GetVector(6) == Vector{0xff907e7ffffe7f7f, 0xffffffffffffffff});
  130. REQUIRE(jit.GetVector(7) == Vector{0xff907e7ffffe7f7f, 0xffffffffffffffff});
  131. REQUIRE(jit.GetVector(8) == Vector{0xff907e7ffffe7f7f, 0xffffffffffffffff});
  132. REQUIRE(jit.GetVector(9) == Vector{0xff907e7ffffe7f7f, 0xffffffffffffffff});
  133. }
  134. TEST_CASE("A64: VQSUB", "[a64]") {
  135. A64TestEnv env;
  136. A64::Jit jit{A64::UserConfig{&env}};
  137. env.code_mem.emplace_back(0x6e212c02); // UQSUB v2.16b, v0.16b, v1.16b
  138. env.code_mem.emplace_back(0x4e212c03); // SQSUB v3.16b, v0.16b, v1.16b
  139. env.code_mem.emplace_back(0x6e612c04); // UQSUB v4.8h, v0.8h, v1.8h
  140. env.code_mem.emplace_back(0x4e612c05); // SQSUB v5.8h, v0.8h, v1.8h
  141. env.code_mem.emplace_back(0x6ea12c06); // UQSUB v6.4s, v0.4s, v1.4s
  142. env.code_mem.emplace_back(0x4ea12c07); // SQSUB v7.4s, v0.4s, v1.4s
  143. env.code_mem.emplace_back(0x6ee12c08); // UQSUB v8.2d, v0.2d, v1.2d
  144. env.code_mem.emplace_back(0x4ee12c09); // SQSUB v9.2d, v0.2d, v1.2d
  145. env.code_mem.emplace_back(0x14000000); // B .
  146. jit.SetVector(0, {0x8010FF00807F0000, 0x8000000080008000});
  147. jit.SetVector(1, {0x7F7F7F7F7F7F7F7F, 0x7FFFFFFF7FFF7FFF});
  148. jit.SetPC(0);
  149. env.ticks_left = 9;
  150. jit.Run();
  151. REQUIRE(jit.GetVector(2) == Vector{0x0100800001000000, 0x0100000001000100});
  152. REQUIRE(jit.GetVector(3) == Vector{0x8091808180008181, 0x8001010180018001});
  153. REQUIRE(jit.GetVector(4) == Vector{0x00917f8101000000, 0x0001000000010001});
  154. REQUIRE(jit.GetVector(5) == Vector{0x8000800080008081, 0x8000000180008000});
  155. REQUIRE(jit.GetVector(6) == Vector{0x00917f8100ff8081, 0x0000000100010001});
  156. REQUIRE(jit.GetVector(7) == Vector{0x8000000080000000, 0x8000000080000000});
  157. REQUIRE(jit.GetVector(8) == Vector{0x00917f8100ff8081, 0x0000000100010001});
  158. REQUIRE(jit.GetVector(9) == Vector{0x8000000000000000, 0x8000000000000000});
  159. }
  160. TEST_CASE("A64: REV", "[a64]") {
  161. A64TestEnv env;
  162. A64::Jit jit{A64::UserConfig{&env}};
  163. env.code_mem.emplace_back(0xdac00c00); // REV X0, X0
  164. env.code_mem.emplace_back(0x5ac00821); // REV W1, W1
  165. env.code_mem.emplace_back(0x14000000); // B .
  166. jit.SetRegister(0, 0xaabbccddeeff1100);
  167. jit.SetRegister(1, 0xaabbccdd);
  168. jit.SetPC(0);
  169. env.ticks_left = 3;
  170. jit.Run();
  171. REQUIRE(jit.GetRegister(0) == 0x11ffeeddccbbaa);
  172. REQUIRE(jit.GetRegister(1) == 0xddccbbaa);
  173. REQUIRE(jit.GetPC() == 8);
  174. }
  175. TEST_CASE("A64: REV32", "[a64]") {
  176. A64TestEnv env;
  177. A64::Jit jit{A64::UserConfig{&env}};
  178. env.code_mem.emplace_back(0xdac00800); // REV32 X0, X0
  179. env.code_mem.emplace_back(0x14000000); // B .
  180. jit.SetRegister(0, 0xaabbccddeeff1100);
  181. jit.SetPC(0);
  182. env.ticks_left = 2;
  183. jit.Run();
  184. REQUIRE(jit.GetRegister(0) == 0xddccbbaa0011ffee);
  185. REQUIRE(jit.GetPC() == 4);
  186. }
  187. TEST_CASE("A64: REV16", "[a64]") {
  188. A64TestEnv env;
  189. A64::Jit jit{A64::UserConfig{&env}};
  190. env.code_mem.emplace_back(0xdac00400); // REV16 X0, X0
  191. env.code_mem.emplace_back(0x5ac00421); // REV16 W1, W1
  192. env.code_mem.emplace_back(0x14000000); // B .
  193. jit.SetRegister(0, 0xaabbccddeeff1100);
  194. jit.SetRegister(1, 0xaabbccdd);
  195. jit.SetPC(0);
  196. env.ticks_left = 3;
  197. jit.Run();
  198. REQUIRE(jit.GetRegister(0) == 0xbbaaddccffee0011);
  199. REQUIRE(jit.GetRegister(1) == 0xbbaaddcc);
  200. REQUIRE(jit.GetPC() == 8);
  201. }
  202. TEST_CASE("A64: SSHL", "[a64]") {
  203. A64TestEnv env;
  204. A64::Jit jit{A64::UserConfig{&env}};
  205. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  206. code.SSHL(V4.B16(), V4.B16(), V0.B16());
  207. code.SSHL(V5.H8(), V5.H8(), V1.H8());
  208. code.SSHL(V6.S4(), V6.S4(), V2.S4());
  209. code.SSHL(V7.D2(), V7.D2(), V3.D2());
  210. code.SSHL(V17.D2(), V17.D2(), V13.D2());
  211. jit.SetPC(0);
  212. jit.SetVector(0, {0xEFF0FAFBFCFDFEFF, 0x0807050403020100});
  213. jit.SetVector(1, {0x00FCFFFDFFFEFFFF, 0xFF04000300020001});
  214. jit.SetVector(2, {0x000000FDFFFFFFFE, 0xFFFFFF0200000001});
  215. jit.SetVector(3, {0xFFFFFFFFFFFFFFFF, 0x0000000000000001});
  216. jit.SetVector(13, {0x00000000000000FF, 0xFFFFFFFFFFFFFF01});
  217. jit.SetVector(4, {0x8080808080808080, 0xFFFFFFFFFFFFFFFF});
  218. jit.SetVector(5, {0x8000800080008000, 0xFFFFFFFFFFFFFFFF});
  219. jit.SetVector(6, {0x8000000080000000, 0xFFFFFFFFFFFFFFFF});
  220. jit.SetVector(7, {0x8000000000000000, 0xFFFFFFFFFFFFFFFF});
  221. jit.SetVector(17, {0x8000000000000000, 0xFFFFFFFFFFFFFFFF});
  222. env.ticks_left = env.code_mem.size();
  223. jit.Run();
  224. CHECK(jit.GetVector(4) == Vector{0xfffffefcf8f0e0c0, 0x0080e0f0f8fcfeff});
  225. CHECK(jit.GetVector(5) == Vector{0xf800f000e000c000, 0xfff0fff8fffcfffe});
  226. CHECK(jit.GetVector(6) == Vector{0xf0000000e0000000, 0xfffffffcfffffffe});
  227. CHECK(jit.GetVector(7) == Vector{0xc000000000000000, 0xfffffffffffffffe});
  228. CHECK(jit.GetVector(17) == Vector{0xc000000000000000, 0xfffffffffffffffe});
  229. }
  230. TEST_CASE("A64: USHL", "[a64]") {
  231. A64TestEnv env;
  232. A64::Jit jit{A64::UserConfig{&env}};
  233. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  234. code.USHL(V4.B16(), V4.B16(), V0.B16());
  235. code.USHL(V14.B8(), V14.B8(), V10.B8());
  236. code.USHL(V5.H8(), V5.H8(), V1.H8());
  237. code.USHL(V15.H4(), V15.H4(), V11.H4());
  238. code.USHL(V6.S4(), V6.S4(), V2.S4());
  239. code.USHL(V16.S2(), V16.S2(), V12.S2());
  240. code.USHL(V7.D2(), V7.D2(), V3.D2());
  241. code.USHL(V17.D2(), V17.D2(), V13.D2());
  242. jit.SetPC(0);
  243. jit.SetVector(0, {0x10FE0E0D0C0B0A09, 0x0807050403020100});
  244. jit.SetVector(10, {0xF6F7F8F9FAFBFCFD});
  245. jit.SetVector(1, {0xFFFE000700060005, 0x0004000300020001});
  246. jit.SetVector(11, {0x00F1FF0F00F08010});
  247. jit.SetVector(2, {0xFFFFFFFE00000003, 0x0000000200000001});
  248. jit.SetVector(12, {0x000000E18000001F});
  249. jit.SetVector(3, {0xFFFFFFFFFFFFFFFE, 0x0000000000000001});
  250. jit.SetVector(13, {0x00000000000000C1, 0xFF0000000000003F});
  251. jit.SetVector(4, {0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF});
  252. jit.SetVector(14, {0x8080808080808080});
  253. jit.SetVector(5, {0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF});
  254. jit.SetVector(15, {0x80000001FFFFFFFF});
  255. jit.SetVector(6, {0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF});
  256. jit.SetVector(16, {0x8000000000000001});
  257. jit.SetVector(7, {0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF});
  258. jit.SetVector(17, {0x8000000000000000, 0x0000000000000001});
  259. env.ticks_left = env.code_mem.size();
  260. jit.Run();
  261. CHECK(jit.GetVector(4) == Vector{0x003f000000000000, 0x0080e0f0f8fcfeff});
  262. CHECK(jit.GetVector(14) == Vector{0x0000000102040810});
  263. CHECK(jit.GetVector(5) == Vector{0x3fffff80ffc0ffe0, 0xfff0fff8fffcfffe});
  264. CHECK(jit.GetVector(15) == Vector{0x0001800000000000});
  265. CHECK(jit.GetVector(6) == Vector{0x3ffffffffffffff8, 0xfffffffcfffffffe});
  266. CHECK(jit.GetVector(16) == Vector{0x0000000180000000});
  267. CHECK(jit.GetVector(7) == Vector{0x3fffffffffffffff, 0xfffffffffffffffe});
  268. CHECK(jit.GetVector(17) == Vector{0x0000000000000001, 0x8000000000000000});
  269. }
  270. TEST_CASE("A64: URSHL", "[a64]") {
  271. A64TestEnv env;
  272. A64::Jit jit{A64::UserConfig{&env}};
  273. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  274. code.URSHL(V0.S4(), V1.S4(), V2.S4());
  275. code.URSHL(V3.S4(), V4.S4(), V5.S4());
  276. code.URSHL(V6.D2(), V7.D2(), V8.D2());
  277. code.URSHL(V9.D2(), V10.D2(), V11.D2());
  278. jit.SetVector(1, Vector{0xffffffff'18ba6a6a, 0x7fffffff'943b954f});
  279. jit.SetVector(2, Vector{0xffffffe0'80000013, 0x00aabbe1'abcdef21});
  280. jit.SetVector(4, Vector{0x0000000b'0000000f, 0xffffffff'ffffffff});
  281. jit.SetVector(5, Vector{0x000000fd'fedcbafd, 0x000000ff'00000001});
  282. jit.SetVector(7, Vector{0x824817df73adca9f, 0x35e511704656e7a4});
  283. jit.SetVector(8, Vector{0x000000000000002a, 0xffffffffffffffc9});
  284. jit.SetVector(10, Vector{0xffffffffffffffff, 0x96dc5c140705cd04});
  285. jit.SetVector(11, Vector{0xffffffffffffffc1, 0x00555555555555f5});
  286. env.ticks_left = env.code_mem.size();
  287. jit.Run();
  288. CHECK(jit.GetVector(0) == Vector{0x00000001'53500000, 0x00000001'00000000});
  289. CHECK(jit.GetVector(3) == Vector{0x00000001'00000002, 0x80000000'fffffffe});
  290. CHECK(jit.GetVector(6) == Vector{0xb72a7c0000000000, 0x0000000000006c});
  291. CHECK(jit.GetVector(9) == Vector{0x0000000000000002, 0x12db8b8280e0ba});
  292. }
  293. TEST_CASE("A64: XTN", "[a64]") {
  294. A64TestEnv env;
  295. A64::Jit jit{A64::UserConfig{&env}};
  296. env.code_mem.emplace_back(0x0e212803); // XTN v3.8b, v0.8h
  297. env.code_mem.emplace_back(0x0e612824); // XTN v4.4h, v1.4s
  298. env.code_mem.emplace_back(0x0ea12845); // XTN v5.2s, v2.2d
  299. env.code_mem.emplace_back(0x14000000); // B .
  300. jit.SetPC(0);
  301. jit.SetVector(0, {0x3333222211110000, 0x7777666655554444});
  302. jit.SetVector(1, {0x1111111100000000, 0x3333333322222222});
  303. jit.SetVector(2, {0x0000000000000000, 0x1111111111111111});
  304. env.ticks_left = 4;
  305. jit.Run();
  306. REQUIRE(jit.GetVector(3) == Vector{0x7766554433221100, 0x0000000000000000});
  307. REQUIRE(jit.GetVector(4) == Vector{0x3333222211110000, 0x0000000000000000});
  308. REQUIRE(jit.GetVector(5) == Vector{0x1111111100000000, 0x0000000000000000});
  309. }
  310. TEST_CASE("A64: TBL", "[a64]") {
  311. A64TestEnv env;
  312. A64::Jit jit{A64::UserConfig{&env}};
  313. env.code_mem.emplace_back(0x0e000100); // TBL v0.8b, { v8.16b }, v0.8b
  314. env.code_mem.emplace_back(0x4e010101); // TBL v1.16b, { v8.16b }, v1.16b
  315. env.code_mem.emplace_back(0x0e022102); // TBL v2.8b, { v8.16b, v9.16b }, v2.8b
  316. env.code_mem.emplace_back(0x4e032103); // TBL v3.16b, { v8.16b, v9.16b }, v3.16b
  317. env.code_mem.emplace_back(0x0e044104); // TBL v4.8b, { v8.16b, v9.16b, v10.16b }, v4.8b
  318. env.code_mem.emplace_back(0x4e054105); // TBL v5.16b, { v8.16b, v9.16b, v10.16b }, v5.16b
  319. env.code_mem.emplace_back(0x0e066106); // TBL v6.8b, { v8.16b, v9.16b, v10.16b, v11.16b }, v6.8b
  320. env.code_mem.emplace_back(0x4e076107); // TBL v7.16b, { v8.16b, v9.16b, v10.16b, v11.16b }, v7.16b
  321. env.code_mem.emplace_back(0x14000000); // B .
  322. // Indices
  323. // 'FF' intended to test out-of-index
  324. jit.SetVector(0, {0x000102030405'FF'07, 0x08090a0b0c0d0e0f});
  325. jit.SetVector(1, {0x000102030405'FF'07, 0x08090a0b0c0d0e0f});
  326. jit.SetVector(2, {0x100011011202'FF'03, 0x1404150516061707});
  327. jit.SetVector(3, {0x100011011202'FF'03, 0x1404150516061707});
  328. jit.SetVector(4, {0x201000211101'FF'12, 0x0233231303241404});
  329. jit.SetVector(5, {0x201000211101'FF'12, 0x0233231303241404});
  330. jit.SetVector(6, {0x403010004131'FF'01, 0x4232120243332303});
  331. jit.SetVector(7, {0x403010004131'FF'01, 0x4232120243332303});
  332. // Table
  333. jit.SetVector(8, {0x7766554433221100, 0xffeeddccbbaa9988});
  334. jit.SetVector(9, {0xffffffffffffffff, 0xffffffffffffffff});
  335. jit.SetVector(10, {0xeeeeeeeeeeeeeeee, 0xeeeeeeeeeeeeeeee});
  336. jit.SetVector(11, {0xdddddddddddddddd, 0xdddddddddddddddd});
  337. jit.SetPC(0);
  338. env.ticks_left = 9;
  339. jit.Run();
  340. REQUIRE(jit.GetVector(0) == Vector{0x001122334455'00'77, 0x0000000000000000});
  341. REQUIRE(jit.GetVector(1) == Vector{0x001122334455'00'77, 0x8899aabbccddeeff});
  342. REQUIRE(jit.GetVector(2) == Vector{0xff00ff11ff22'00'33, 0x0000000000000000});
  343. REQUIRE(jit.GetVector(3) == Vector{0xff00ff11ff22'00'33, 0xff44ff55ff66ff77});
  344. REQUIRE(jit.GetVector(4) == Vector{0xeeff00eeff11'00'ff, 0x0000000000000000});
  345. REQUIRE(jit.GetVector(5) == Vector{0xeeff00eeff11'00'ff, 0x2200eeff33eeff44});
  346. REQUIRE(jit.GetVector(6) == Vector{0x00ddff0000dd'00'11, 0x0000000000000000});
  347. REQUIRE(jit.GetVector(7) == Vector{0x00ddff0000dd'00'11, 0x00ddff2200ddee33});
  348. }
  349. TEST_CASE("A64: TBX", "[a64]") {
  350. A64TestEnv env;
  351. A64::Jit jit{A64::UserConfig{&env}};
  352. env.code_mem.emplace_back(0x0e001100); // TBX v0.8b, { v8.16b }, v0.8b
  353. env.code_mem.emplace_back(0x4e011101); // TBX v1.16b, { v8.16b }, v1.16b
  354. env.code_mem.emplace_back(0x0e023102); // TBX v2.8b, { v8.16b, v9.16b }, v2.8b
  355. env.code_mem.emplace_back(0x4e033103); // TBX v3.16b, { v8.16b, v9.16b }, v3.16b
  356. env.code_mem.emplace_back(0x0e045104); // TBX v4.8b, { v8.16b, v9.16b, v10.16b }, v4.8b
  357. env.code_mem.emplace_back(0x4e055105); // TBX v5.16b, { v8.16b, v9.16b, v10.16b }, v5.16b
  358. env.code_mem.emplace_back(0x0e067106); // TBX v6.8b, { v8.16b, v9.16b, v10.16b, v11.16b }, v6.8b
  359. env.code_mem.emplace_back(0x4e077107); // TBX v7.16b, { v8.16b, v9.16b, v10.16b, v11.16b }, v7.16b
  360. env.code_mem.emplace_back(0x14000000); // B .
  361. // Indices
  362. // 'FF' intended to test out-of-index
  363. jit.SetVector(0, {0x000102030405'FF'07, 0x08090a0b0c0d0e0f});
  364. jit.SetVector(1, {0x000102030405'FF'07, 0x08090a0b0c0d0e0f});
  365. jit.SetVector(2, {0x100011011202'FF'03, 0x1404150516061707});
  366. jit.SetVector(3, {0x100011011202'FF'03, 0x1404150516061707});
  367. jit.SetVector(4, {0x201000211101'FF'12, 0x0233231303241404});
  368. jit.SetVector(5, {0x201000211101'FF'12, 0x0233231303241404});
  369. jit.SetVector(6, {0x403010004131'FF'01, 0x4232120243332303});
  370. jit.SetVector(7, {0x403010004131'FF'01, 0x4232120243332303});
  371. // Table
  372. jit.SetVector(8, {0x7766554433221100, 0xffeeddccbbaa9988});
  373. jit.SetVector(9, {0xffffffffffffffff, 0xffffffffffffffff});
  374. jit.SetVector(10, {0xeeeeeeeeeeeeeeee, 0xeeeeeeeeeeeeeeee});
  375. jit.SetVector(11, {0xdddddddddddddddd, 0xdddddddddddddddd});
  376. jit.SetPC(0);
  377. env.ticks_left = 9;
  378. jit.Run();
  379. REQUIRE(jit.GetVector(0) == Vector{0x001122334455'FF'77, 0x0000000000000000});
  380. REQUIRE(jit.GetVector(1) == Vector{0x001122334455'FF'77, 0x8899aabbccddeeff});
  381. REQUIRE(jit.GetVector(2) == Vector{0xff00ff11ff22'FF'33, 0x0000000000000000});
  382. REQUIRE(jit.GetVector(3) == Vector{0xff00ff11ff22'FF'33, 0xff44ff55ff66ff77});
  383. REQUIRE(jit.GetVector(4) == Vector{0xeeff00eeff11'FF'ff, 0x0000000000000000});
  384. REQUIRE(jit.GetVector(5) == Vector{0xeeff00eeff11'FF'ff, 0x2233eeff33eeff44});
  385. REQUIRE(jit.GetVector(6) == Vector{0x40ddff0041dd'FF'11, 0x0000000000000000});
  386. REQUIRE(jit.GetVector(7) == Vector{0x40ddff0041dd'FF'11, 0x42ddff2243ddee33});
  387. }
  388. TEST_CASE("A64: AND", "[a64]") {
  389. A64TestEnv env;
  390. A64::Jit jit{A64::UserConfig{&env}};
  391. env.code_mem.emplace_back(0x8a020020); // AND X0, X1, X2
  392. env.code_mem.emplace_back(0x14000000); // B .
  393. jit.SetRegister(0, 0);
  394. jit.SetRegister(1, 1);
  395. jit.SetRegister(2, 3);
  396. jit.SetPC(0);
  397. env.ticks_left = 2;
  398. jit.Run();
  399. REQUIRE(jit.GetRegister(0) == 1);
  400. REQUIRE(jit.GetRegister(1) == 1);
  401. REQUIRE(jit.GetRegister(2) == 3);
  402. REQUIRE(jit.GetPC() == 4);
  403. }
  404. TEST_CASE("A64: Bitmasks", "[a64]") {
  405. A64TestEnv env;
  406. A64::Jit jit{A64::UserConfig{&env}};
  407. env.code_mem.emplace_back(0x3200c3e0); // ORR W0, WZR, #0x01010101
  408. env.code_mem.emplace_back(0x320c8fe1); // ORR W1, WZR, #0x00F000F0
  409. env.code_mem.emplace_back(0x320003e2); // ORR W2, WZR, #1
  410. env.code_mem.emplace_back(0x14000000); // B .
  411. jit.SetPC(0);
  412. env.ticks_left = 4;
  413. jit.Run();
  414. REQUIRE(jit.GetRegister(0) == 0x01010101);
  415. REQUIRE(jit.GetRegister(1) == 0x00F000F0);
  416. REQUIRE(jit.GetRegister(2) == 1);
  417. REQUIRE(jit.GetPC() == 12);
  418. }
  419. TEST_CASE("A64: ANDS NZCV", "[a64]") {
  420. A64TestEnv env;
  421. A64::Jit jit{A64::UserConfig{&env}};
  422. env.code_mem.emplace_back(0x6a020020); // ANDS W0, W1, W2
  423. env.code_mem.emplace_back(0x14000000); // B .
  424. SECTION("N=1, Z=0") {
  425. jit.SetRegister(0, 0);
  426. jit.SetRegister(1, 0xFFFFFFFF);
  427. jit.SetRegister(2, 0xFFFFFFFF);
  428. jit.SetPC(0);
  429. env.ticks_left = 2;
  430. jit.Run();
  431. REQUIRE(jit.GetRegister(0) == 0xFFFFFFFF);
  432. REQUIRE(jit.GetRegister(1) == 0xFFFFFFFF);
  433. REQUIRE(jit.GetRegister(2) == 0xFFFFFFFF);
  434. REQUIRE(jit.GetPC() == 4);
  435. REQUIRE((jit.GetPstate() & 0xF0000000) == 0x80000000);
  436. }
  437. SECTION("N=0, Z=1") {
  438. jit.SetRegister(0, 0);
  439. jit.SetRegister(1, 0xFFFFFFFF);
  440. jit.SetRegister(2, 0x00000000);
  441. jit.SetPC(0);
  442. env.ticks_left = 2;
  443. jit.Run();
  444. REQUIRE(jit.GetRegister(0) == 0x00000000);
  445. REQUIRE(jit.GetRegister(1) == 0xFFFFFFFF);
  446. REQUIRE(jit.GetRegister(2) == 0x00000000);
  447. REQUIRE(jit.GetPC() == 4);
  448. REQUIRE((jit.GetPstate() & 0xF0000000) == 0x40000000);
  449. }
  450. SECTION("N=0, Z=0") {
  451. jit.SetRegister(0, 0);
  452. jit.SetRegister(1, 0x12345678);
  453. jit.SetRegister(2, 0x7324a993);
  454. jit.SetPC(0);
  455. env.ticks_left = 2;
  456. jit.Run();
  457. REQUIRE(jit.GetRegister(0) == 0x12240010);
  458. REQUIRE(jit.GetRegister(1) == 0x12345678);
  459. REQUIRE(jit.GetRegister(2) == 0x7324a993);
  460. REQUIRE(jit.GetPC() == 4);
  461. REQUIRE((jit.GetPstate() & 0xF0000000) == 0x00000000);
  462. }
  463. }
  464. TEST_CASE("A64: CBZ", "[a64]") {
  465. A64TestEnv env;
  466. A64::Jit jit{A64::UserConfig{&env}};
  467. env.code_mem.emplace_back(0x34000060); // 0x00 : CBZ X0, label
  468. env.code_mem.emplace_back(0x320003e2); // 0x04 : MOV X2, 1
  469. env.code_mem.emplace_back(0x14000000); // 0x08 : B.
  470. env.code_mem.emplace_back(0x321f03e2); // 0x0C : label: MOV X2, 2
  471. env.code_mem.emplace_back(0x14000000); // 0x10 : B .
  472. SECTION("no branch") {
  473. jit.SetPC(0);
  474. jit.SetRegister(0, 1);
  475. env.ticks_left = 4;
  476. jit.Run();
  477. REQUIRE(jit.GetRegister(2) == 1);
  478. REQUIRE(jit.GetPC() == 8);
  479. }
  480. SECTION("branch") {
  481. jit.SetPC(0);
  482. jit.SetRegister(0, 0);
  483. env.ticks_left = 4;
  484. jit.Run();
  485. REQUIRE(jit.GetRegister(2) == 2);
  486. REQUIRE(jit.GetPC() == 16);
  487. }
  488. }
  489. TEST_CASE("A64: TBZ", "[a64]") {
  490. A64TestEnv env;
  491. A64::Jit jit{A64::UserConfig{&env}};
  492. env.code_mem.emplace_back(0x36180060); // 0x00 : TBZ X0, 3, label
  493. env.code_mem.emplace_back(0x320003e2); // 0x04 : MOV X2, 1
  494. env.code_mem.emplace_back(0x14000000); // 0x08 : B .
  495. env.code_mem.emplace_back(0x321f03e2); // 0x0C : label: MOV X2, 2
  496. env.code_mem.emplace_back(0x14000000); // 0x10 : B .
  497. SECTION("no branch") {
  498. jit.SetPC(0);
  499. jit.SetRegister(0, 0xFF);
  500. env.ticks_left = 4;
  501. jit.Run();
  502. REQUIRE(jit.GetRegister(2) == 1);
  503. REQUIRE(jit.GetPC() == 8);
  504. }
  505. SECTION("branch with zero") {
  506. jit.SetPC(0);
  507. jit.SetRegister(0, 0);
  508. env.ticks_left = 4;
  509. jit.Run();
  510. REQUIRE(jit.GetRegister(2) == 2);
  511. REQUIRE(jit.GetPC() == 16);
  512. }
  513. SECTION("branch with non-zero") {
  514. jit.SetPC(0);
  515. jit.SetRegister(0, 1);
  516. env.ticks_left = 4;
  517. jit.Run();
  518. REQUIRE(jit.GetRegister(2) == 2);
  519. REQUIRE(jit.GetPC() == 16);
  520. }
  521. }
  522. TEST_CASE("A64: FABD", "[a64]") {
  523. A64TestEnv env;
  524. A64::Jit jit{A64::UserConfig{&env}};
  525. env.code_mem.emplace_back(0x6eb5d556); // FABD.4S V22, V10, V21
  526. env.code_mem.emplace_back(0x14000000); // B .
  527. jit.SetPC(0);
  528. jit.SetVector(10, {0xb4858ac77ff39a87, 0x9fce5e14c4873176});
  529. jit.SetVector(21, {0x56d3f085ff890e2b, 0x6e4b0a41801a2d00});
  530. env.ticks_left = 2;
  531. jit.Run();
  532. REQUIRE(jit.GetVector(22) == Vector{0x56d3f0857fc90e2b, 0x6e4b0a4144873176});
  533. }
  534. TEST_CASE("A64: FABS", "[a64]") {
  535. A64TestEnv env;
  536. A64::Jit jit{A64::UserConfig{&env}};
  537. env.code_mem.emplace_back(0x4ef8f804); // FABS v4.8h, v0.8h
  538. env.code_mem.emplace_back(0x4ea0f825); // FABS v5.4s, v1.4s
  539. env.code_mem.emplace_back(0x4ee0f846); // FABS v6.2d, v2.2d
  540. env.code_mem.emplace_back(0x14000000); // B .
  541. jit.SetPC(0);
  542. jit.SetVector(0, {0xffffffffffffffff, 0xffffffffffff8000});
  543. jit.SetVector(1, {0xffbfffffffc00000, 0xff80000080000000});
  544. jit.SetVector(2, {0xffffffffffffffff, 0x8000000000000000});
  545. env.ticks_left = 4;
  546. jit.Run();
  547. REQUIRE(jit.GetVector(4) == Vector{0x7fff7fff7fff7fff, 0x7fff7fff7fff0000});
  548. REQUIRE(jit.GetVector(5) == Vector{0x7fbfffff7fc00000, 0x7f80000000000000});
  549. REQUIRE(jit.GetVector(6) == Vector{0x7fffffffffffffff, 0x0000000000000000});
  550. }
  551. TEST_CASE("A64: FMIN (example)", "[a64]") {
  552. A64TestEnv env;
  553. A64::Jit jit{A64::UserConfig{&env}};
  554. env.code_mem.emplace_back(0x4ea1f400); // FMIN.4S V0, V0, V1
  555. env.code_mem.emplace_back(0x4ee3f442); // FMIN.2D V2, V2, V3
  556. env.code_mem.emplace_back(0x14000000); // B .
  557. jit.SetPC(0);
  558. jit.SetVector(0, {0x7fc00000'09503366, 0x00000000'7f984a37});
  559. jit.SetVector(1, {0xc1200000'00000001, 0x6e4b0a41'ffffffff});
  560. jit.SetVector(2, {0x7fc0000009503366, 0x3ff0000000000000});
  561. jit.SetVector(3, {0xbff0000000000000, 0x6e4b0a41ffffffff});
  562. env.ticks_left = 2;
  563. jit.Run();
  564. REQUIRE(jit.GetVector(0) == Vector{0x7fc00000'00000001, 0x00000000'7fd84a37});
  565. REQUIRE(jit.GetVector(2) == Vector{0xbff0000000000000, 0x3ff0000000000000});
  566. }
  567. TEST_CASE("A64: FMAX (example)", "[a64]") {
  568. A64TestEnv env;
  569. A64::Jit jit{A64::UserConfig{&env}};
  570. env.code_mem.emplace_back(0x4e21f400); // FMAX.4S V0, V0, V1
  571. env.code_mem.emplace_back(0x4e63f442); // FMAX.2D V2, V2, V3
  572. env.code_mem.emplace_back(0x14000000); // B .
  573. jit.SetPC(0);
  574. jit.SetVector(0, {0x7fc00000'09503366, 0x00000000'7f984a37});
  575. jit.SetVector(1, {0xc1200000'00000001, 0x6e4b0a41'ffffffff});
  576. jit.SetVector(2, {0x7fc0000009503366, 0x3ff0000000000000});
  577. jit.SetVector(3, {0xbff0000000000000, 0x6e4b0a41ffffffff});
  578. env.ticks_left = 2;
  579. jit.Run();
  580. REQUIRE(jit.GetVector(0) == Vector{0x7fc00000'09503366, 0x6e4b0a41'7fd84a37});
  581. REQUIRE(jit.GetVector(2) == Vector{0x7fc0000009503366, 0x6e4b0a41ffffffff});
  582. }
  583. TEST_CASE("A64: FMINNM (example)", "[a64]") {
  584. A64TestEnv env;
  585. A64::Jit jit{A64::UserConfig{&env}};
  586. env.code_mem.emplace_back(0x4ea1c400); // FMINNM.4S V0, V0, V1
  587. env.code_mem.emplace_back(0x4ee3c442); // FMINNM.2D V2, V2, V3
  588. env.code_mem.emplace_back(0x14000000); // B .
  589. jit.SetPC(0);
  590. jit.SetVector(0, {0x7fc00000'09503366, 0x00000000'7f984a37});
  591. jit.SetVector(1, {0xc1200000'00000001, 0x6e4b0a41'ffffffff});
  592. jit.SetVector(2, {0x7fc0000009503366, 0x3ff0000000000000});
  593. jit.SetVector(3, {0xfff0000000000000, 0xffffffffffffffff});
  594. env.ticks_left = 2;
  595. jit.Run();
  596. REQUIRE(jit.GetVector(0) == Vector{0xc1200000'00000001, 0x00000000'7fd84a37});
  597. REQUIRE(jit.GetVector(2) == Vector{0xfff0000000000000, 0x3ff0000000000000});
  598. }
  599. TEST_CASE("A64: FMAXNM (example)", "[a64]") {
  600. A64TestEnv env;
  601. A64::Jit jit{A64::UserConfig{&env}};
  602. env.code_mem.emplace_back(0x4e21c400); // FMAXNM.4S V0, V0, V1
  603. env.code_mem.emplace_back(0x4e63c442); // FMAXNM.2D V2, V2, V3
  604. env.code_mem.emplace_back(0x14000000); // B .
  605. jit.SetPC(0);
  606. jit.SetVector(0, {0x7fc00000'09503366, 0x00000000'7f984a37});
  607. jit.SetVector(1, {0xc1200000'00000001, 0x6e4b0a41'ffffffff});
  608. jit.SetVector(2, {0x7fc0000009503366, 0x3ff0000000000000});
  609. jit.SetVector(3, {0xfff0000000000000, 0xffffffffffffffff});
  610. env.ticks_left = 2;
  611. jit.Run();
  612. REQUIRE(jit.GetVector(0) == Vector{0xc1200000'09503366, 0x6e4b0a41'7fd84a37});
  613. REQUIRE(jit.GetVector(2) == Vector{0x7fc0000009503366, 0x3ff0000000000000});
  614. }
  615. TEST_CASE("A64: FMAXNM (example 2)", "[a64]") {
  616. A64TestEnv env;
  617. A64::Jit jit{A64::UserConfig{&env}};
  618. env.code_mem.emplace_back(0x4e3bc6fd); // FMAXNM.4S V29, V23, V27
  619. env.code_mem.emplace_back(0x14000000); // B .
  620. jit.SetPC(0);
  621. jit.SetFpcr(0x01400000);
  622. jit.SetVector(23, {0xb485877c'42280000, 0x317285d3'b5c8e5d3});
  623. jit.SetVector(27, {0xbc48d091'c79b271e, 0xff800001'3304c3ef});
  624. env.ticks_left = 2;
  625. jit.Run();
  626. REQUIRE(jit.GetVector(29) == Vector{0xb485877c'42280000, 0xffc00001'3304c3ef});
  627. }
  628. TEST_CASE("A64: 128-bit exclusive read/write", "[a64]") {
  629. A64TestEnv env;
  630. ExclusiveMonitor monitor{1};
  631. A64::UserConfig conf;
  632. conf.callbacks = &env;
  633. conf.processor_id = 0;
  634. SECTION("Global Monitor") {
  635. conf.global_monitor = &monitor;
  636. }
  637. A64::Jit jit{conf};
  638. env.code_mem.emplace_back(0xc87f0861); // LDXP X1, X2, [X3]
  639. env.code_mem.emplace_back(0xc8241865); // STXP W4, X5, X6, [X3]
  640. env.code_mem.emplace_back(0x14000000); // B .
  641. jit.SetPC(0);
  642. jit.SetRegister(3, 0x1234567812345678);
  643. jit.SetRegister(4, 0xbaadbaadbaadbaad);
  644. jit.SetRegister(5, 0xaf00d1e5badcafe0);
  645. jit.SetRegister(6, 0xd0d0cacad0d0caca);
  646. env.ticks_left = 3;
  647. jit.Run();
  648. REQUIRE(jit.GetRegister(1) == 0x7f7e7d7c7b7a7978);
  649. REQUIRE(jit.GetRegister(2) == 0x8786858483828180);
  650. REQUIRE(jit.GetRegister(4) == 0);
  651. REQUIRE(env.MemoryRead64(0x1234567812345678) == 0xaf00d1e5badcafe0);
  652. REQUIRE(env.MemoryRead64(0x1234567812345680) == 0xd0d0cacad0d0caca);
  653. }
  654. TEST_CASE("A64: CNTPCT_EL0", "[a64]") {
  655. A64TestEnv env;
  656. A64::Jit jit{A64::UserConfig{&env}};
  657. env.code_mem.emplace_back(0xd53be021); // MRS X1, CNTPCT_EL0
  658. env.code_mem.emplace_back(0xd503201f); // NOP
  659. env.code_mem.emplace_back(0xd503201f); // NOP
  660. env.code_mem.emplace_back(0xd503201f); // NOP
  661. env.code_mem.emplace_back(0xd503201f); // NOP
  662. env.code_mem.emplace_back(0xd503201f); // NOP
  663. env.code_mem.emplace_back(0xd503201f); // NOP
  664. env.code_mem.emplace_back(0xd53be022); // MRS X2, CNTPCT_EL0
  665. env.code_mem.emplace_back(0xcb010043); // SUB X3, X2, X1
  666. env.code_mem.emplace_back(0x14000000); // B .
  667. env.ticks_left = 10;
  668. jit.Run();
  669. REQUIRE(jit.GetRegister(3) == 7);
  670. }
  671. TEST_CASE("A64: FNMSUB 1", "[a64]") {
  672. A64TestEnv env;
  673. A64::Jit jit{A64::UserConfig{&env}};
  674. env.code_mem.emplace_back(0x1f618a9c); // FNMSUB D28, D20, D1, D2
  675. env.code_mem.emplace_back(0x14000000); // B .
  676. jit.SetPC(0);
  677. jit.SetVector(20, {0xe73a51346164bd6c, 0x8080000000002b94});
  678. jit.SetVector(1, {0xbf8000007fffffff, 0xffffffff00002b94});
  679. jit.SetVector(2, {0x0000000000000000, 0xc79b271e3f000000});
  680. env.ticks_left = 2;
  681. jit.Run();
  682. REQUIRE(jit.GetVector(28) == Vector{0x66ca513533ee6076, 0x0000000000000000});
  683. }
  684. TEST_CASE("A64: FNMSUB 2", "[a64]") {
  685. A64TestEnv env;
  686. A64::Jit jit{A64::UserConfig{&env}};
  687. env.code_mem.emplace_back(0x1f2ab88e); // FNMSUB S14, S4, S10, S14
  688. env.code_mem.emplace_back(0x14000000); // B .
  689. jit.SetPC(0);
  690. jit.SetVector(4, {0x3c9623b101398437, 0x7ff0abcd0ba98d27});
  691. jit.SetVector(10, {0xffbfffff3eaaaaab, 0x3f0000003f8147ae});
  692. jit.SetVector(14, {0x80000000007fffff, 0xe73a513400000000});
  693. jit.SetFpcr(0x00400000);
  694. env.ticks_left = 2;
  695. jit.Run();
  696. REQUIRE(jit.GetVector(14) == Vector{0x0000000080045284, 0x0000000000000000});
  697. }
  698. TEST_CASE("A64: FMADD", "[a64]") {
  699. A64TestEnv env;
  700. A64::Jit jit{A64::UserConfig{&env}};
  701. env.code_mem.emplace_back(0x1f5e0e4a); // FMADD D10, D18, D30, D3
  702. env.code_mem.emplace_back(0x14000000); // B .
  703. jit.SetPC(0);
  704. jit.SetVector(18, {0x8000007600800000, 0x7ff812347f800000});
  705. jit.SetVector(30, {0xff984a3700000000, 0xe73a513480800000});
  706. jit.SetVector(3, {0x3f000000ff7fffff, 0x8139843780000000});
  707. jit.SetFpcr(0x00400000);
  708. env.ticks_left = 2;
  709. jit.Run();
  710. REQUIRE(jit.GetVector(10) == Vector{0x3f059921bf0dbfff, 0x0000000000000000});
  711. }
  712. TEST_CASE("A64: FMLA.4S(lane)", "[a64]") {
  713. A64TestEnv env;
  714. A64::Jit jit{A64::UserConfig{&env}};
  715. env.code_mem.emplace_back(0x4f8f11c0); // FMLA.4S V0, V14, V15[0]
  716. env.code_mem.emplace_back(0x4faf11c1); // FMLA.4S V1, V14, V15[1]
  717. env.code_mem.emplace_back(0x4f8f19c2); // FMLA.4S V2, V14, V15[2]
  718. env.code_mem.emplace_back(0x4faf19c3); // FMLA.4S V3, V14, V15[3]
  719. env.code_mem.emplace_back(0x14000000); // B .
  720. jit.SetPC(0);
  721. jit.SetVector(0, {0x3ff00000'3ff00000, 0x00000000'00000000});
  722. jit.SetVector(1, {0x3ff00000'3ff00000, 0x00000000'00000000});
  723. jit.SetVector(2, {0x3ff00000'3ff00000, 0x00000000'00000000});
  724. jit.SetVector(3, {0x3ff00000'3ff00000, 0x00000000'00000000});
  725. jit.SetVector(14, {0x3ff00000'3ff00000, 0x3ff00000'3ff00000});
  726. jit.SetVector(15, {0x3ff00000'40000000, 0x40400000'40800000});
  727. env.ticks_left = 5;
  728. jit.Run();
  729. REQUIRE(jit.GetVector(0) == Vector{0x40b4000040b40000, 0x4070000040700000});
  730. REQUIRE(jit.GetVector(1) == Vector{0x40ac800040ac8000, 0x4061000040610000});
  731. REQUIRE(jit.GetVector(2) == Vector{0x4116000041160000, 0x40f0000040f00000});
  732. REQUIRE(jit.GetVector(3) == Vector{0x40f0000040f00000, 0x40b4000040b40000});
  733. }
  734. TEST_CASE("A64: FMUL.4S(lane)", "[a64]") {
  735. A64TestEnv env;
  736. A64::Jit jit{A64::UserConfig{&env}};
  737. env.code_mem.emplace_back(0x4f8f91c0); // FMUL.4S V0, V14, V15[0]
  738. env.code_mem.emplace_back(0x4faf91c1); // FMUL.4S V1, V14, V15[1]
  739. env.code_mem.emplace_back(0x4f8f99c2); // FMUL.4S V2, V14, V15[2]
  740. env.code_mem.emplace_back(0x4faf99c3); // FMUL.4S V3, V14, V15[3]
  741. env.code_mem.emplace_back(0x14000000); // B .
  742. jit.SetPC(0);
  743. jit.SetVector(14, {0x3ff00000'3ff00000, 0x3ff00000'3ff00000});
  744. jit.SetVector(15, {0x3ff00000'40000000, 0x40400000'40800000});
  745. env.ticks_left = 5;
  746. jit.Run();
  747. REQUIRE(jit.GetVector(0) == Vector{0x4070000040700000, 0x4070000040700000});
  748. REQUIRE(jit.GetVector(1) == Vector{0x4061000040610000, 0x4061000040610000});
  749. REQUIRE(jit.GetVector(2) == Vector{0x40f0000040f00000, 0x40f0000040f00000});
  750. REQUIRE(jit.GetVector(3) == Vector{0x40b4000040b40000, 0x40b4000040b40000});
  751. }
  752. TEST_CASE("A64: FMLA.4S (denormal)", "[a64]") {
  753. A64TestEnv env;
  754. A64::Jit jit{A64::UserConfig{&env}};
  755. env.code_mem.emplace_back(0x4e2fcccc); // FMLA.4S V12, V6, V15
  756. env.code_mem.emplace_back(0x14000000); // B .
  757. jit.SetPC(0);
  758. jit.SetVector(12, {0x3c9623b17ff80000, 0xbff0000080000076});
  759. jit.SetVector(6, {0x7ff80000ff800000, 0x09503366c1200000});
  760. jit.SetVector(15, {0x3ff0000080636d24, 0xbf800000e73a5134});
  761. jit.SetFpcr(0x01000000);
  762. env.ticks_left = 2;
  763. jit.Run();
  764. REQUIRE(jit.GetVector(12) == Vector{0x7ff800007fc00000, 0xbff0000068e8e581});
  765. }
  766. TEST_CASE("A64: FMLA.4S (0x80800000)", "[a64]") {
  767. A64TestEnv env;
  768. A64::Jit jit{A64::UserConfig{&env}};
  769. env.code_mem.emplace_back(0x4e38cc2b); // FMLA.4S V11, V1, V24
  770. env.code_mem.emplace_back(0x14000000); // B .
  771. jit.SetPC(0);
  772. jit.SetVector(11, {0xc79b271efff05678, 0xffc0000080800000});
  773. jit.SetVector(1, {0x00636d2400800000, 0x0966320bb26bddee});
  774. jit.SetVector(24, {0x460e8c84fff00000, 0x8ba98d2780800002});
  775. jit.SetFpcr(0x03000000);
  776. env.ticks_left = 2;
  777. jit.Run();
  778. REQUIRE(jit.GetVector(11) == Vector{0xc79b271e7fc00000, 0x7fc0000080000000});
  779. }
  780. // x64 has different rounding behaviour to AArch64.
  781. // AArch64 performs rounding after flushing-to-zero.
  782. // x64 performs rounding before flushing-to-zero.
  783. TEST_CASE("A64: FMADD (0x80800000)", "[a64]") {
  784. A64TestEnv env;
  785. A64::Jit jit{A64::UserConfig{&env}};
  786. env.code_mem.emplace_back(0x1f0f7319); // FMADD S25, S24, S15, S28
  787. env.code_mem.emplace_back(0x14000000); // B .
  788. jit.SetPC(0);
  789. jit.SetVector(24, {0x00800000, 0});
  790. jit.SetVector(15, {0x0ba98d27, 0});
  791. jit.SetVector(28, {0x80800000, 0});
  792. jit.SetFpcr(0x01000000);
  793. env.ticks_left = 2;
  794. jit.Run();
  795. REQUIRE(jit.GetVector(25) == Vector{0x80000000, 0});
  796. }
  797. TEST_CASE("A64: FNEG failed to zero upper", "[a64]") {
  798. A64TestEnv env;
  799. A64::Jit jit{A64::UserConfig{&env}};
  800. env.code_mem.emplace_back(0x2ea0fb50); // FNEG.2S V16, V26
  801. env.code_mem.emplace_back(0x2e207a1c); // SQNEG.8B V28, V16
  802. env.code_mem.emplace_back(0x14000000); // B .
  803. jit.SetPC(0);
  804. jit.SetVector(26, {0x071286fde8f34a90, 0x837cffa8be382f60});
  805. jit.SetFpcr(0x01000000);
  806. env.ticks_left = 6;
  807. jit.Run();
  808. REQUIRE(jit.GetVector(28) == Vector{0x79ee7a03980db670, 0});
  809. REQUIRE(FP::FPSR{jit.GetFpsr()}.QC() == false);
  810. }
  811. TEST_CASE("A64: FRSQRTS", "[a64]") {
  812. A64TestEnv env;
  813. A64::Jit jit{A64::UserConfig{&env}};
  814. env.code_mem.emplace_back(0x5eb8fcad); // FRSQRTS S13, S5, S24
  815. env.code_mem.emplace_back(0x14000000); // B .
  816. // These particular values result in an intermediate value during
  817. // the calculation that is close to infinity. We want to verify
  818. // that this special case is handled appropriately.
  819. jit.SetPC(0);
  820. jit.SetVector(5, {0xfc6a0206, 0});
  821. jit.SetVector(24, {0xfc6a0206, 0});
  822. jit.SetFpcr(0x00400000);
  823. env.ticks_left = 2;
  824. jit.Run();
  825. REQUIRE(jit.GetVector(13) == Vector{0xff7fffff, 0});
  826. }
  827. TEST_CASE("A64: SQDMULH.8H (saturate)", "[a64]") {
  828. A64TestEnv env;
  829. A64::Jit jit{A64::UserConfig{&env}};
  830. env.code_mem.emplace_back(0x4e62b420); // SQDMULH.8H V0, V1, V2
  831. env.code_mem.emplace_back(0x14000000); // B .
  832. // Make sure that saturating values are tested
  833. jit.SetPC(0);
  834. jit.SetVector(1, {0x7fff80007ffe8001, 0x7fff80007ffe8001});
  835. jit.SetVector(2, {0x7fff80007ffe8001, 0x80007fff80017ffe});
  836. jit.SetFpsr(0);
  837. env.ticks_left = 2;
  838. jit.Run();
  839. REQUIRE(jit.GetVector(0) == Vector{0x7ffe7fff7ffc7ffe, 0x8001800180028002});
  840. REQUIRE(FP::FPSR{jit.GetFpsr()}.QC() == true);
  841. }
  842. TEST_CASE("A64: SQDMULH.4S (saturate)", "[a64]") {
  843. A64TestEnv env;
  844. A64::Jit jit{A64::UserConfig{&env}};
  845. env.code_mem.emplace_back(0x4ea2b420); // SQDMULH.4S V0, V1, V2
  846. env.code_mem.emplace_back(0x14000000); // B .
  847. // Make sure that saturating values are tested
  848. jit.SetPC(0);
  849. jit.SetVector(1, {0x7fffffff80000000, 0x7fffffff80000000});
  850. jit.SetVector(2, {0x7fffffff80000000, 0x800000007fffffff});
  851. jit.SetFpsr(0);
  852. env.ticks_left = 2;
  853. jit.Run();
  854. REQUIRE(jit.GetVector(0) == Vector{0x7ffffffe7fffffff, 0x8000000180000001});
  855. REQUIRE(FP::FPSR{jit.GetFpsr()}.QC() == true);
  856. }
  857. TEST_CASE("A64: This is an infinite loop if fast dispatch is enabled", "[a64]") {
  858. A64TestEnv env;
  859. A64::UserConfig conf{&env};
  860. conf.optimizations &= ~OptimizationFlag::FastDispatch;
  861. A64::Jit jit{conf};
  862. env.code_mem.emplace_back(0x2ef998fa);
  863. env.code_mem.emplace_back(0x2ef41c11);
  864. env.code_mem.emplace_back(0x0f07fdd8);
  865. env.code_mem.emplace_back(0x9ac90d09);
  866. env.code_mem.emplace_back(0xd63f0120); // BLR X9
  867. env.code_mem.emplace_back(0x14000000); // B .
  868. env.ticks_left = 6;
  869. jit.Run();
  870. }
  871. TEST_CASE("A64: EXTR", "[a64]") {
  872. A64TestEnv env;
  873. A64::Jit jit{A64::UserConfig{&env}};
  874. env.code_mem.emplace_back(0x93d8fef7); // EXTR X23, X23, X24, #63
  875. env.code_mem.emplace_back(0x14000000); // B .
  876. jit.SetPC(0);
  877. jit.SetRegister(23, 0);
  878. jit.SetRegister(24, 1);
  879. env.ticks_left = 2;
  880. jit.Run();
  881. REQUIRE(jit.GetRegister(23) == 0);
  882. }
  883. TEST_CASE("A64: Isolated GetNZCVFromOp", "[a64]") {
  884. A64TestEnv env;
  885. A64::Jit jit{A64::UserConfig{&env}};
  886. env.code_mem.emplace_back(0xaa1f03f5); // MOV X21, XZR
  887. env.code_mem.emplace_back(0x912a02da); // ADD X26, X22, #0xa80
  888. env.code_mem.emplace_back(0x913662dc); // ADD X28, X22, #0xd98
  889. env.code_mem.emplace_back(0x320003e8); // MOV W8, #1
  890. env.code_mem.emplace_back(0xa9006bfc); // STP X28, X26, [SP]
  891. env.code_mem.emplace_back(0x7200011f); // TST W8, #1
  892. env.code_mem.emplace_back(0xf94007e8); // LDR X8, [SP, #8]
  893. env.code_mem.emplace_back(0x321e03e3); // MOV W3, #4
  894. env.code_mem.emplace_back(0xaa1303e2); // MOV X2, X19
  895. env.code_mem.emplace_back(0x9a881357); // CSEL X23, X26, X8, NE
  896. env.code_mem.emplace_back(0xf94003e8); // LDR X8, [SP]
  897. env.code_mem.emplace_back(0xaa1703e0); // MOV X0, X23
  898. env.code_mem.emplace_back(0x9a881396); // CSEL X22, X28, X8, NE
  899. env.code_mem.emplace_back(0x92407ea8); // AND X8, X21, #0xffffffff
  900. env.code_mem.emplace_back(0x1ac8269b); // LSR W27, W20, W8
  901. env.code_mem.emplace_back(0x0b1b0768); // ADD W8, W27, W27, LSL #1
  902. env.code_mem.emplace_back(0x937f7d01); // SBFIZ X1, X8, #1, #32
  903. env.code_mem.emplace_back(0x2a1f03e4); // MOV W4, WZR
  904. env.code_mem.emplace_back(0x531e7779); // LSL W25, W27, #2
  905. env.code_mem.emplace_back(0x14000000); // B .
  906. jit.SetPC(0);
  907. env.ticks_left = 20;
  908. jit.Run();
  909. }
  910. TEST_CASE("A64: Optimization failure when folding ADD", "[a64]") {
  911. A64TestEnv env;
  912. A64::Jit jit{A64::UserConfig{&env}};
  913. env.code_mem.emplace_back(0xbc4f84be); // LDR S30, [X5], #248
  914. env.code_mem.emplace_back(0x9a0c00ea); // ADC X10, X7, X12
  915. env.code_mem.emplace_back(0x5a1a0079); // SBC W25, W3, W26
  916. env.code_mem.emplace_back(0x9b0e2be9); // MADD X9, XZR, X14, X10
  917. env.code_mem.emplace_back(0xfa5fe8a9); // CCMP X5, #31, #9, AL
  918. env.code_mem.emplace_back(0x14000000); // B .
  919. jit.SetPC(0);
  920. jit.SetRegister(0, 0x46e15845dba57924);
  921. jit.SetRegister(1, 0x6f60d04350581fea);
  922. jit.SetRegister(2, 0x85cface50edcfc03);
  923. jit.SetRegister(3, 0x47e1e8906e10ec5a);
  924. jit.SetRegister(4, 0x70717c9450b6b707);
  925. jit.SetRegister(5, 0x300d83205baeaff4);
  926. jit.SetRegister(6, 0xb7890de7c6fee082);
  927. jit.SetRegister(7, 0xa89fb6d6f1b42f4a);
  928. jit.SetRegister(8, 0x04e36b8aada91d4f);
  929. jit.SetRegister(9, 0xa03bf6bde71c6ac5);
  930. jit.SetRegister(10, 0x319374d14baa83b0);
  931. jit.SetRegister(11, 0x5a78fc0fffca7c5f);
  932. jit.SetRegister(12, 0xc012b5063f43b8ad);
  933. jit.SetRegister(13, 0x821ade159d39fea1);
  934. jit.SetRegister(14, 0x41f97b2f5525c25e);
  935. jit.SetRegister(15, 0xab0cd3653cb93738);
  936. jit.SetRegister(16, 0x50dfcb55a4ebd554);
  937. jit.SetRegister(17, 0x30dd7d18ae52df03);
  938. jit.SetRegister(18, 0x4e53b20d252bf085);
  939. jit.SetRegister(19, 0x013582d71f5fd42a);
  940. jit.SetRegister(20, 0x97a151539dad44e7);
  941. jit.SetRegister(21, 0xa6fcc6bb220a2ad3);
  942. jit.SetRegister(22, 0x4c84d3c84a6c5c5c);
  943. jit.SetRegister(23, 0x1a7596a5ef930dff);
  944. jit.SetRegister(24, 0x06248d96a02ff210);
  945. jit.SetRegister(25, 0xfcb8772aec4b1dfd);
  946. jit.SetRegister(26, 0x63619787b6a17665);
  947. jit.SetRegister(27, 0xbd50c3352d001e40);
  948. jit.SetRegister(28, 0x4e186aae63c81553);
  949. jit.SetRegister(29, 0x57462b7163bd6508);
  950. jit.SetRegister(30, 0xa977c850d16d562c);
  951. jit.SetSP(0x000000da9b761d8c);
  952. jit.SetFpsr(0x03480000);
  953. jit.SetPstate(0x30000000);
  954. env.ticks_left = 6;
  955. jit.Run();
  956. REQUIRE(jit.GetRegister(0) == 0x46e15845dba57924);
  957. REQUIRE(jit.GetRegister(1) == 0x6f60d04350581fea);
  958. REQUIRE(jit.GetRegister(2) == 0x85cface50edcfc03);
  959. REQUIRE(jit.GetRegister(3) == 0x47e1e8906e10ec5a);
  960. REQUIRE(jit.GetRegister(4) == 0x70717c9450b6b707);
  961. REQUIRE(jit.GetRegister(5) == 0x300d83205baeb0ec);
  962. REQUIRE(jit.GetRegister(6) == 0xb7890de7c6fee082);
  963. REQUIRE(jit.GetRegister(7) == 0xa89fb6d6f1b42f4a);
  964. REQUIRE(jit.GetRegister(8) == 0x04e36b8aada91d4f);
  965. REQUIRE(jit.GetRegister(9) == 0x68b26bdd30f7e7f8);
  966. REQUIRE(jit.GetRegister(10) == 0x68b26bdd30f7e7f8);
  967. REQUIRE(jit.GetRegister(11) == 0x5a78fc0fffca7c5f);
  968. REQUIRE(jit.GetRegister(12) == 0xc012b5063f43b8ad);
  969. REQUIRE(jit.GetRegister(13) == 0x821ade159d39fea1);
  970. REQUIRE(jit.GetRegister(14) == 0x41f97b2f5525c25e);
  971. REQUIRE(jit.GetRegister(15) == 0xab0cd3653cb93738);
  972. REQUIRE(jit.GetRegister(16) == 0x50dfcb55a4ebd554);
  973. REQUIRE(jit.GetRegister(17) == 0x30dd7d18ae52df03);
  974. REQUIRE(jit.GetRegister(18) == 0x4e53b20d252bf085);
  975. REQUIRE(jit.GetRegister(19) == 0x013582d71f5fd42a);
  976. REQUIRE(jit.GetRegister(20) == 0x97a151539dad44e7);
  977. REQUIRE(jit.GetRegister(21) == 0xa6fcc6bb220a2ad3);
  978. REQUIRE(jit.GetRegister(22) == 0x4c84d3c84a6c5c5c);
  979. REQUIRE(jit.GetRegister(23) == 0x1a7596a5ef930dff);
  980. REQUIRE(jit.GetRegister(24) == 0x06248d96a02ff210);
  981. REQUIRE(jit.GetRegister(25) == 0x00000000b76f75f5);
  982. REQUIRE(jit.GetRegister(26) == 0x63619787b6a17665);
  983. REQUIRE(jit.GetRegister(27) == 0xbd50c3352d001e40);
  984. REQUIRE(jit.GetRegister(28) == 0x4e186aae63c81553);
  985. REQUIRE(jit.GetRegister(29) == 0x57462b7163bd6508);
  986. REQUIRE(jit.GetRegister(30) == 0xa977c850d16d562c);
  987. REQUIRE(jit.GetPstate() == 0x20000000);
  988. REQUIRE(jit.GetVector(30) == Vector{0xf7f6f5f4, 0});
  989. }
  990. TEST_CASE("A64: Cache Maintenance Instructions", "[a64]") {
  991. class CacheMaintenanceTestEnv final : public A64TestEnv {
  992. void InstructionCacheOperationRaised(A64::InstructionCacheOperation op, VAddr value) override {
  993. REQUIRE(op == A64::InstructionCacheOperation::InvalidateByVAToPoU);
  994. REQUIRE(value == 0xcafed00d);
  995. }
  996. void DataCacheOperationRaised(A64::DataCacheOperation op, VAddr value) override {
  997. REQUIRE(op == A64::DataCacheOperation::InvalidateByVAToPoC);
  998. REQUIRE(value == 0xcafebabe);
  999. }
  1000. };
  1001. CacheMaintenanceTestEnv env;
  1002. A64::UserConfig conf{&env};
  1003. conf.hook_data_cache_operations = true;
  1004. A64::Jit jit{conf};
  1005. jit.SetRegister(0, 0xcafed00d);
  1006. jit.SetRegister(1, 0xcafebabe);
  1007. env.code_mem.emplace_back(0xd50b7520); // ic ivau, x0
  1008. env.code_mem.emplace_back(0xd5087621); // dc ivac, x1
  1009. env.code_mem.emplace_back(0x14000000); // B .
  1010. env.ticks_left = 3;
  1011. jit.Run();
  1012. }
  1013. TEST_CASE("A64: Memory access (fastmem)", "[a64]") {
  1014. constexpr size_t address_width = 12;
  1015. constexpr size_t memory_size = 1ull << address_width; // 4K
  1016. constexpr size_t page_size = 4 * 1024;
  1017. constexpr size_t buffer_size = 2 * page_size;
  1018. char buffer[buffer_size];
  1019. void* buffer_ptr = reinterpret_cast<void*>(buffer);
  1020. size_t buffer_size_nconst = buffer_size;
  1021. char* backing_memory = reinterpret_cast<char*>(std::align(page_size, memory_size, buffer_ptr, buffer_size_nconst));
  1022. A64FastmemTestEnv env{backing_memory};
  1023. Dynarmic::A64::UserConfig config{&env};
  1024. config.fastmem_pointer = reinterpret_cast<uintptr_t>(backing_memory);
  1025. config.fastmem_address_space_bits = address_width;
  1026. config.recompile_on_fastmem_failure = false;
  1027. config.silently_mirror_fastmem = true;
  1028. config.processor_id = 0;
  1029. Dynarmic::A64::Jit jit{config};
  1030. memset(backing_memory, 0, memory_size);
  1031. memcpy(backing_memory + 0x100, "Lorem ipsum dolor sit amet, consectetur adipiscing elit.", 57);
  1032. env.MemoryWrite32(0, 0xA9401404); // LDP X4, X5, [X0]
  1033. env.MemoryWrite32(4, 0xF9400046); // LDR X6, [X2]
  1034. env.MemoryWrite32(8, 0xA9001424); // STP X4, X5, [X1]
  1035. env.MemoryWrite32(12, 0xF9000066); // STR X6, [X3]
  1036. env.MemoryWrite32(16, 0x14000000); // B .
  1037. jit.SetRegister(0, 0x100);
  1038. jit.SetRegister(1, 0x1F0);
  1039. jit.SetRegister(2, 0x10F);
  1040. jit.SetRegister(3, 0x1FF);
  1041. jit.SetPC(0);
  1042. jit.SetSP(memory_size - 1);
  1043. jit.SetFpsr(0x03480000);
  1044. jit.SetPstate(0x30000000);
  1045. env.ticks_left = 5;
  1046. jit.Run();
  1047. REQUIRE(strncmp(backing_memory + 0x100, backing_memory + 0x1F0, 23) == 0);
  1048. }
  1049. TEST_CASE("A64: SQRDMULH QC flag when output invalidated", "[a64]") {
  1050. A64TestEnv env;
  1051. A64::Jit jit{A64::UserConfig{&env}};
  1052. env.code_mem.emplace_back(0x0fbcd38b); // SQRDMULH.2S V11, V28, V28[1]
  1053. env.code_mem.emplace_back(0x7ef0f8eb); // FMINP.2D D11, V7
  1054. env.code_mem.emplace_back(0x14000000); // B .
  1055. jit.SetPC(0);
  1056. jit.SetVector(7, {0xb1b5'd0b1'4e54'e281, 0xb4cb'4fec'8563'1032});
  1057. jit.SetVector(28, {0x8000'0000'0000'0000, 0x0000'0000'0000'0000});
  1058. jit.SetFpcr(0x05400000);
  1059. env.ticks_left = 3;
  1060. jit.Run();
  1061. REQUIRE(jit.GetFpsr() == 0x08000000);
  1062. REQUIRE(jit.GetVector(11) == Vector{0xb4cb'4fec'8563'1032, 0x0000'0000'0000'0000});
  1063. }
  1064. TEST_CASE("A64: SDIV maximally", "[a64]") {
  1065. A64TestEnv env;
  1066. A64::Jit jit{A64::UserConfig{&env}};
  1067. env.code_mem.emplace_back(0x9ac00c22); // SDIV X2, X1, X0
  1068. env.code_mem.emplace_back(0x14000000); // B .
  1069. jit.SetRegister(0, 0xffffffffffffffff);
  1070. jit.SetRegister(1, 0x8000000000000000);
  1071. jit.SetRegister(2, 0xffffffffffffffff);
  1072. jit.SetPC(0);
  1073. env.ticks_left = 2;
  1074. jit.Run();
  1075. REQUIRE(jit.GetRegister(0) == 0xffffffffffffffff);
  1076. REQUIRE(jit.GetRegister(1) == 0x8000000000000000);
  1077. REQUIRE(jit.GetRegister(2) == 0x8000000000000000);
  1078. REQUIRE(jit.GetPC() == 4);
  1079. }
  1080. // Restricted register set required to trigger:
  1081. // const HostLocList any_gpr = { HostLoc::RAX, HostLoc::RBX, HostLoc::RCX, HostLoc::R13, HostLoc::R14 };
  1082. // const HostLocList any_xmm = { HostLoc::XMM1, HostLoc::XMM2, HostLoc::XMM3, HostLoc::XMM4, HostLoc::XMM5, HostLoc::XMM6 };
  1083. TEST_CASE("A64: rand1", "[a64]") {
  1084. A64TestEnv env;
  1085. A64::Jit jit{A64::UserConfig{&env}};
  1086. env.code_mem = {0x2ea2e69a, 0x6f7168e7, 0x7eb0f816, 0x6ebd369d, 0x1e65c302, 0x1e63011c, 0x1e67c349, 0x0f861bd6, 0x9e59cbbc, 0x5e61cb8b, 0x6e218b01, 0x4eb2409f, 0x7f7c2452, 0x7e207a8d, 0xd503369f};
  1087. env.code_mem.emplace_back(0x14000000); // B .
  1088. jit.SetRegister(0, 0x67e1d59cc30a788c);
  1089. jit.SetRegister(1, 0x0e771a2a79dfb060);
  1090. jit.SetRegister(2, 0x35cc7e7831247f7c);
  1091. jit.SetRegister(3, 0x63a22cce1f9cde66);
  1092. jit.SetRegister(4, 0xb6a022d8406543a3);
  1093. jit.SetRegister(5, 0x6712e272c4ad27a0);
  1094. jit.SetRegister(6, 0x9d2a01c3bc374837);
  1095. jit.SetRegister(7, 0x83bc2f62feb76043);
  1096. jit.SetRegister(8, 0x9ba9e8c3d543f1bf);
  1097. jit.SetRegister(9, 0xe4aee4636277b787);
  1098. jit.SetRegister(10, 0x9cd9e201dacc233b);
  1099. jit.SetRegister(11, 0x39e0a5c3bb44efc9);
  1100. jit.SetRegister(12, 0xca229296c29f8742);
  1101. jit.SetRegister(13, 0x4cdf038f1323ff2d);
  1102. jit.SetRegister(14, 0x377ad499a81b1f5a);
  1103. jit.SetRegister(15, 0x8217307060f11c6d);
  1104. jit.SetRegister(16, 0xd1af2e75ea62dba7);
  1105. jit.SetRegister(17, 0x77661148c760e9d6);
  1106. jit.SetRegister(18, 0xf05a251f9cf60f9e);
  1107. jit.SetRegister(19, 0xf54301927e8fa020);
  1108. jit.SetRegister(20, 0x534c76f6f6d6805c);
  1109. jit.SetRegister(21, 0x60240c3e727aae2d);
  1110. jit.SetRegister(22, 0x52b82c212af254d6);
  1111. jit.SetRegister(23, 0xb0ad501210d12c07);
  1112. jit.SetRegister(24, 0x596a9119514f3460);
  1113. jit.SetRegister(25, 0xa933e19b69b2c6f7);
  1114. jit.SetRegister(26, 0x6f3693ec0f5e7708);
  1115. jit.SetRegister(27, 0xc6a3908a03fb9737);
  1116. jit.SetRegister(28, 0x113ba38d50953b60);
  1117. jit.SetRegister(29, 0xbe5395907134511e);
  1118. jit.SetRegister(30, 0x9a5d96aa066e5c39);
  1119. jit.SetPC(0);
  1120. jit.SetSP(0x000000c6bec5a48c);
  1121. jit.SetVector(0, {0x0faa90e6561b1ffb, 0xb8c1c925ee613293});
  1122. jit.SetVector(1, {0x3fa365cf7a4f3eaa, 0xbd0fabf98eb5c061});
  1123. jit.SetVector(2, {0x3d7722d0e4444b00, 0xf30ba88476b79615});
  1124. jit.SetVector(3, {0xf794f4953fb4a413, 0xedd6426638cf0242});
  1125. jit.SetVector(4, {0x1ddfdd8985c58693, 0xc344d565e68ab18b});
  1126. jit.SetVector(5, {0x600fcef72b18ae5f, 0x3af9964747ff06b9});
  1127. jit.SetVector(6, {0x276b755d4452ec74, 0xf5579ddb0f2146b4});
  1128. jit.SetVector(7, {0xd1823739c80439e5, 0xd8c4bc8cf08fce6e});
  1129. jit.SetVector(8, {0x0e4c8796dca46ad0, 0x53293d124cd38d6e});
  1130. jit.SetVector(9, {0x860e30c54fcbe0b8, 0x09c57c6b723e45f5});
  1131. jit.SetVector(10, {0xe3652801c3d11ddb, 0x4ef5f76fa85d28b9});
  1132. jit.SetVector(11, {0xa6c22b4e20d5a3a2, 0x5b98938307afb538});
  1133. jit.SetVector(12, {0x915960a26d2d8c02, 0x0ecdf8bc35c8a184});
  1134. jit.SetVector(13, {0xa79a1f506ed066b4, 0x23de2152171ce4c6});
  1135. jit.SetVector(14, {0xd4b85ed863708645, 0x3cf7b2693ac76d3f});
  1136. jit.SetVector(15, {0x8900b9888729557b, 0x2eeeef32083bf9b9});
  1137. jit.SetVector(16, {0x0b40331c7fc30b54, 0xcb5fb7d6ca96ccca});
  1138. jit.SetVector(17, {0x0040b87ea24910c7, 0x97f925750c5da4c5});
  1139. jit.SetVector(18, {0xf19de744c8c88b3d, 0xa1406fae21f53d8c});
  1140. jit.SetVector(19, {0x02b6e985e99a6a3d, 0xe470d5328c9b2af5});
  1141. jit.SetVector(20, {0x6bfb919ed9752198, 0xcaab56c2adc2c486});
  1142. jit.SetVector(21, {0x4c1dd31e9fb91bae, 0xe1d4a4b936d1dfab});
  1143. jit.SetVector(22, {0x5d8c08ee0dbe758a, 0xb1b25da077a0ba26});
  1144. jit.SetVector(23, {0xf1f3377346a6e4db, 0x4995274fe7e17908});
  1145. jit.SetVector(24, {0xa1c4d7cca6fe8a95, 0xb267a94646819606});
  1146. jit.SetVector(25, {0x8bbe1a250a008e73, 0xc729df1ac7eeb7d3});
  1147. jit.SetVector(26, {0x48c23bc8ce6857d5, 0x35bb31ef278268d7});
  1148. jit.SetVector(27, {0x0473d63f3f0c5075, 0xf4bb5d79938901f4});
  1149. jit.SetVector(28, {0x01e2930f7313493e, 0xdc6ef4adadcc8e37});
  1150. jit.SetVector(29, {0x2c500da43b460d13, 0x7bb4520d5580a648});
  1151. jit.SetVector(30, {0xdf4e3d139b825da0, 0x19fea0310522fda2});
  1152. jit.SetVector(31, {0xf8b440b8d5e25111, 0x73758151a32b6b13});
  1153. jit.SetPstate(0x60000000);
  1154. jit.SetFpcr(0x01080000);
  1155. env.ticks_left = 16;
  1156. jit.Run();
  1157. REQUIRE(jit.GetRegister(0) == 0x67e1d59cc30a788c);
  1158. REQUIRE(jit.GetRegister(1) == 0x0e771a2a79dfb060);
  1159. REQUIRE(jit.GetRegister(2) == 0x35cc7e7831247f7c);
  1160. REQUIRE(jit.GetRegister(3) == 0x63a22cce1f9cde66);
  1161. REQUIRE(jit.GetRegister(4) == 0xb6a022d8406543a3);
  1162. REQUIRE(jit.GetRegister(5) == 0x6712e272c4ad27a0);
  1163. REQUIRE(jit.GetRegister(6) == 0x9d2a01c3bc374837);
  1164. REQUIRE(jit.GetRegister(7) == 0x83bc2f62feb76043);
  1165. REQUIRE(jit.GetRegister(8) == 0x9ba9e8c3d543f1bf);
  1166. REQUIRE(jit.GetRegister(9) == 0xe4aee4636277b787);
  1167. REQUIRE(jit.GetRegister(10) == 0x9cd9e201dacc233b);
  1168. REQUIRE(jit.GetRegister(11) == 0x39e0a5c3bb44efc9);
  1169. REQUIRE(jit.GetRegister(12) == 0xca229296c29f8742);
  1170. REQUIRE(jit.GetRegister(13) == 0x4cdf038f1323ff2d);
  1171. REQUIRE(jit.GetRegister(14) == 0x377ad499a81b1f5a);
  1172. REQUIRE(jit.GetRegister(15) == 0x8217307060f11c6d);
  1173. REQUIRE(jit.GetRegister(16) == 0xd1af2e75ea62dba7);
  1174. REQUIRE(jit.GetRegister(17) == 0x77661148c760e9d6);
  1175. REQUIRE(jit.GetRegister(18) == 0xf05a251f9cf60f9e);
  1176. REQUIRE(jit.GetRegister(19) == 0xf54301927e8fa020);
  1177. REQUIRE(jit.GetRegister(20) == 0x534c76f6f6d6805c);
  1178. REQUIRE(jit.GetRegister(21) == 0x60240c3e727aae2d);
  1179. REQUIRE(jit.GetRegister(22) == 0x52b82c212af254d6);
  1180. REQUIRE(jit.GetRegister(23) == 0xb0ad501210d12c07);
  1181. REQUIRE(jit.GetRegister(24) == 0x596a9119514f3460);
  1182. REQUIRE(jit.GetRegister(25) == 0xa933e19b69b2c6f7);
  1183. REQUIRE(jit.GetRegister(26) == 0x6f3693ec0f5e7708);
  1184. REQUIRE(jit.GetRegister(27) == 0xc6a3908a03fb9737);
  1185. REQUIRE(jit.GetRegister(28) == 0x0000000000000000);
  1186. REQUIRE(jit.GetRegister(29) == 0xbe5395907134511e);
  1187. REQUIRE(jit.GetRegister(30) == 0x9a5d96aa066e5c39);
  1188. }
  1189. TEST_CASE("A64: rand2", "[a64][.]") {
  1190. A64TestEnv env;
  1191. A64::Jit jit{A64::UserConfig{.callbacks = &env, .fastmem_pointer = 0xffffffff00000000}};
  1192. env.code_mem = {0xea80f352, 0x6e65e59d, 0x1e20c343, 0x2e3a7192, 0x2e267249, 0xd500405f, 0x6f01f461, 0x6eb684fc, 0x58028edd, 0x0ea5f5b6, 0x0ea069fb, 0x2e769517, 0x5e066063, 0x1e65c3f5, 0x4f00ff52, 0x93401cf6, 0x1e274248, 0x6f67aaf5, 0x5e0c0782, 0x5ef43f3c, 0x2e6595b7, 0x4e20590f, 0xb35aa451, 0x6ee2c5ed, 0x4e32bf46, 0x2ea1ba8f, 0x2f68a85e, 0x9237d90a, 0x5e23dd10, 0x0e762e32, 0x4e31a8cf, 0xce1f3360, 0x781a4ac0, 0x13834066, 0x5fa8101c, 0x6f7c5594, 0x0e71bb68, 0xbc0b3e8f, 0x785dbbda, 0x6f51e794, 0xce50af75, 0x1ad728ec, 0x6ee0da4c, 0xb84efa14, 0x2eb3f613, 0x4e287ade, 0x4eb8c734, 0x2e83f4e8, 0x0e397c80, 0xd08f93f8, 0xce718e48, 0x0f672a0d, 0x2e9edd40, 0x0e14128b, 0x6f5942e6, 0x8b3a0f03, 0x3c5d16b9, 0x7f7e3743, 0x4f4c54e4, 0x0ea0a9e9, 0x9e59dbe6, 0x6e7ddcd3, 0xcec08377, 0x9ba759f8, 0x2ea5046e, 0x0e24c569, 0xb8979780, 0x4e31b98c, 0x4efe4f46, 0x4ea7c762, 0x7e61c9c6, 0x6e30c880, 0x1ada0c25, 0x4e603a2f, 0xda9d7218, 0x0d40c5d9, 0x5e214b05, 0x9ba9efc5, 0x5e61b81e, 0x6e7bc31c, 0x0e61a163, 0x9e5832d2, 0x4e772248, 0x4e3d17c8, 0x92624f60, 0x7a1a02dc, 0x79891f65, 0x6eb45036, 0x0e321ee8, 0x4e2566f0, 0x4ea02b9b, 0x0f9dcb3d, 0x2e21b9f9, 0x0e21a8c3, 0xda1700bd, 0x6ea0fb38, 0x7e607a0b, 0x72845817, 0x7f61068e, 0x0d60e529, 0x4ea0ca5c, 0x1a94b20f, 0x8b87419d, 0x7ea9ed71, 0x2ea1a86e, 0x4d40c4da, 0x5ea0eada, 0x784ba96e, 0x7eb6ee02, 0x3db1c710, 0x0e217836, 0x7ee0bb96, 0x4e786c08, 0x4e976a08, 0x489ffe86, 0x4e79fc9b, 0x0e21cbce, 0x5ef7fc65, 0x4ea1286d, 0xd29c771e, 0x6f5c2839, 0x0ea00a9d, 0x6ee44c06, 0x5ee1d858, 0x5ef2fda6, 0x7eb0c9fe, 0x7f762791, 0x2e212ae6, 0x4e61c9db, 0x13003c57, 0x5ee1b8f8, 0x0f2396d2, 0x6ea0db1e, 0x0e71ba82, 0xab29c807, 0x6ef8f8b3, 0x1f18d4a1, 0x0e261d15, 0x1e290081, 0x1b0c7d12, 0x4e7771c3, 0xf845f1e4, 0x4d40c9e8, 0xce778452, 0x6eb9879d, 0x6e21c93d, 0xcec0829f, 0x52a0969f, 0x1e772b4f, 0x7ee1da88, 0x5f52fe0a, 0x7f3387b1, 0x5e214850, 0x1e65c025, 0x0e2ca294, 0x2e614829, 0x1e640077, 0x9e240048, 0x4ebe9537, 0x9bb7925e, 0x38b669c5, 0x2840d089, 0x6f43e648, 0x2e662d28, 0x4eabaff3, 0x6e734cc7, 0x0e31baee, 0x7ee0d93c, 0x5e282bde, 0x7e21bba4, 0x4e6c75fa, 0x5ac01217, 0x7f4304af, 0x1e7878ed, 0x1ada2196, 0x7ee1aba3, 0x93407f3c, 0x4f6c34eb, 0x6e3447a9, 0x7e7ae545, 0x5e0802bb, 0x6eeae63a, 0x7ee1da62, 0x5e280bb3, 0xf81d4009, 0x1e603b21, 0x5e281a14, 0x6eb0a99b, 0x1e266a25, 0x0d60cafe, 0x0e0b6194, 0x7a4ed2c5, 0x92b762ec, 0x4e6b5749, 0x3c16a6e5, 0x4ea0a92b, 0x0fa58b6a, 0x5f76148c, 0x6e30c95f, 0x1e6540fd, 0x5e28e40f, 0x0d403fd4, 0x7e30da36, 0x7fda9b51, 0x2ea04bde, 0x1e25c3d2, 0x1ee0434c, 0x5e21d8e7, 0x5ee1ba51, 0x5e61aba9, 0x4e2849fb, 0x5ee098ea, 0x4e60f63d, 0x0f280443, 0x5ee0da27, 0x2e78a6ce, 0x78054afc, 0x4e14286b, 0x4e218bd8, 0x2a3d2551, 0x3a04017a, 0x5f4317cd, 0x0e604a37, 0x9a834614, 0x0e2edf4d, 0x7a51a0a0, 0x5f8e9043, 0x6ea06bb2, 0xaa2857dd, 0x7a1903fc, 0x301ba9ba, 0x9ac929cd, 0x4e061ff0, 0x2e38fcfc, 0x0e2f614a, 0x7ee0d8e4, 0x6e73afda, 0x7f4156f7, 0x0e6078bf, 0x4ee1d9ed, 0x93403fbe, 0xce6f8640, 0x4e3855e3, 0x6f76fe23, 0x112466e8, 0x1e358a90, 0x7f45272c, 0x6ea19a9d, 0x8a696350, 0x1e3900f6, 0x5e61c866, 0x0e3fbfd0, 0x5ee09ad0, 0x0e651d27, 0x4dffc35e, 0x2e20c6ce, 0x0fbe118d, 0x1e656a15, 0xd1357365, 0x0e20a847, 0xce4a835c, 0x4e203905, 0x2e60090d, 0x7f4a27bb, 0x1e64c316, 0xce7d86a4, 0x7ebded2d, 0x6e70a97e, 0x4eb9a42b, 0x0e209bef, 0x6f151730, 0x0e7e30f7, 0x4e724509, 0xd503375f, 0xce58b6ae, 0x5e21a9b8, 0xcb2ca538, 0x5ac01131, 0x6ea19a24, 0xeb40c8b3, 0xc8df7d65, 0x78108341, 0x3218ab9b, 0x0f3da7dd, 0x2e003089, 0x4e21cab5, 0x8aa5c924, 0x1a94950c, 0x123e506f, 0x13117e37, 0x1ee6005b, 0x5ac00647, 0x5eec8cd5, 0x7ef0fb3d, 0x9223272a, 0x5ee0cb02, 0x6e66071d, 0x6ea1dbbf, 0x5e61c903, 0x5ac015ea, 0x93db6206, 0x7e62b5e3, 0x6ea0c87b, 0xdac0090e, 0x48df7d90, 0x6e206ba5, 0x9e2503c2, 0x6e25fc89, 0x4d60e2db, 0x1e3e22a0, 0x2eb81c19, 0x7856ea00, 0x5fbfb22d, 0x1e630244, 0x4e202a83, 0x1f50a722, 0x7f7b55d2, 0x0fae89b9, 0x4e781d73, 0xce738c3a, 0x4f15a591, 0x6e21c7e1, 0x586ff77e, 0x8a5d3592, 0x93401c67, 0x5e61cb86, 0xce6bc2c1, 0x6e393f10, 0x9bb70ec3, 0xdac0098c, 0x4da84b95, 0x7f494476, 0x9ace5c11, 0x7e61ca14, 0x4f7a60ef, 0x1ad32b39, 0x0ea3777f, 0x5e61da7f, 0x4f1404e2, 0x4e3244e2, 0x6e1b1ceb, 0x0dee5aac, 0x4e2f9dc4, 0x5ea1b8c3, 0x1e59f863, 0xd500403f, 0x4e3ae7d0, 0x4ef5c6ea, 0x08dffe3b, 0x6e36f4f6, 0x2e764f29, 0x0e726f23, 0x5f42375b, 0x7f71fc40, 0x6e618aad, 0x93403e5b, 0x0e205976, 0x0e7250c4, 0x6eb0abc9, 0x2e2049f0, 0x5f14754d, 0x7f6ce468, 0x6f950bbe, 0x6e31aa47, 0x4eb83396, 0x0dccc952, 0x2ea1ca90, 0xce69c701, 0xb0bed69e, 0x7c5dec39, 0x4e2868a2, 0x0e591b08, 0x5f34e6dd, 0x3a449184, 0x5e3ce6de, 0x4ea149b7, 0x4e7ad29b, 0xba198503, 0x1f683e8f, 0xfa52f2a7, 0x6e30dffc, 0x4e6c3d17, 0x2eae3248, 0xd503349f, 0x1e60002c, 0x0f180680, 0x9e240049, 0x6f75774e, 0xa90d8678, 0x9ad924c4, 0x7eb0f85b, 0x0e205aaf, 0x7ee08899, 0x5f4bffd8, 0x1b0ff5f3, 0x4ee11dcd, 0x2e218948, 0x0dcb2733, 0x4eac107c, 0x4ea04a53, 0x4e287b44, 0x0e60b82a, 0x5ee0ebbc, 0xce454ff1, 0x5e1761e7, 0x5e09202f, 0x0e0c0754, 0x1e72e6b9, 0x7e21da70, 0x0fbdb20c, 0x5efb8c84, 0xd500401f, 0x3a47526e, 0x1e680acf, 0x7f7375fc, 0xf80522da, 0x4ee60c02, 0x4d40c2e7, 0x6f89096b, 0x7ee1bb6e, 0x5e280b4a, 0x1e3120c8, 0x7eb2ef96, 0x4fd012dd, 0x0f3027ef, 0x4e2078a8, 0xd503201f, 0x2e2312d9, 0x6ebf1c6e, 0x5ee1f8df, 0x4e607a46, 0x6e30c877, 0x6c09d2d1, 0x4e61abd8, 0x0e35267e, 0x6ac17728, 0x0e861aa0, 0x6f63fe26, 0x6f157628, 0x6f30a5f9, 0x4d60cc0c, 0x4e21cb59, 0x2e68a3fb, 0x7efae601, 0x6ea0f82c, 0x9b25ec12, 0x1a1a0305, 0x0e043fe1, 0x6e73c0ed, 0x6ea1b8c0, 0x7e20380b, 0x0f0534e8, 0x1f56bc7d, 0xba0c0128, 0x1e672160, 0x6e7b259b, 0x7ee07b5d, 0x9a820443, 0x4e040581, 0x2f1d87e8, 0x1acd2f5b, 0x6e20794f, 0x2e6a3c93, 0xc8dffe13, 0xce5ab1c6, 0x6eea55f6, 0x4ea039b3, 0x0d602fec, 0x2e246e2f, 0x7857be39, 0xb80608fb, 0x1e67c017, 0x9bcf7f63, 0x0f92d857, 0x5e0812f7, 0x1e210172, 0x7e6128e9, 0x7ea94d41, 0x981179e1, 0x1effb018, 0x2e600828, 0x0eb9c6b2, 0x6ee1baae, 0x4ea0db28, 0x2ea1487b, 0x4ea6c7f0, 0x2e2374c7, 0x7e30d8dd, 0xb9991fa7, 0x4e791e3e, 0x889f7c4b, 0x0e6c753c, 0x1e740ad1, 0x1e244324, 0x1ef33010, 0x5ac01102, 0x9bd97fba, 0x6e290143, 0x1e2220d8, 0x4d8d5aee, 0x6f28570b, 0xfa4ab0c1, 0xdac00b14, 0x7ea1a90e, 0x2e3027d8, 0x6f25a733, 0x4e61a96e, 0x4e1a2fcb, 0x0e22fe0a, 0xc8df7cd0, 0x5e280a55, 0x4e012b20, 0x7e70dbf4, 0x520c5a4e, 0x6ea6c57f, 0x0e861af8, 0xd503233f, 0x889ffe3c, 0x5e274ea9, 0x4e21a89a, 0x0e170c02, 0x6efd4c0b, 0xd5033ebf, 0x6e61a92c, 0x2e205b72, 0x789fb828, 0x0e626e94, 0x2ea6724c, 0x9a10028b, 0x2c6c51fc, 0x5a9de6b9, 0x6e6881f3, 0x5ee0ea6b, 0x0faec36e, 0x0e955bca, 0x1acf206d, 0x7f6f571b, 0x4e286930, 0x12b41ceb, 0x1e770b7a, 0x0ea18ac2, 0x5e282aaf, 0xf2b7fa1e, 0x1ac34311, 0x13167d11, 0x4ea63412, 0x6e758038, 0x2f1d85d6, 0x0f275480, 0x0ead6c71, 0x6e204b69, 0x1e6303f4, 0x5e0031ef, 0x13001e40, 0x7a16006f, 0x6e6ae4c0, 0x0f0f242f, 0x6e674f50, 0x4e606b7a, 0x7e6ee684, 0x1e6b5957, 0x7ea1bbab, 0x7ea0b6cb, 0xce4da241, 0x0ea1b953, 0x0eb2af4b, 0x9ac309d0, 0x6e61d8bd, 0x5ea0d890, 0x5f47d1e7, 0xfa5981ca, 0x1e7f7959, 0x6ef24dd8, 0x0e0a41d1, 0x5ee0e898, 0x4e6038e2, 0x13097d65, 0x6f839088, 0x9e290265, 0x0e208824, 0x2e65af79, 0x6f36a561, 0x9ad3204b, 0x0e21482e, 0x1e24431d, 0xd50330bf, 0x0df641aa, 0x6e602a83, 0xce30505f, 0x5e025238, 0xd503201f, 0x4e608880, 0x4de9c38d, 0x5e0f5348, 0x6eb48ca9, 0x50fda31b, 0x2e251eec, 0x7842ba50, 0xd8a1cd86, 0x2ea09862, 0x0ea09983, 0x2ea333b0, 0x0ea6032c, 0x4f94801b, 0x7e3ee57d, 0x38135e4f, 0xd8fdd9dd, 0x5ee0fcde, 0x9e64033d, 0x6e37f547, 0x6e3dd7ef, 0x13003f3d, 0x0e602f9f, 0x4e7ad014, 0x9b3b6857, 0x5ea0cb67, 0x0eb31c9f, 0x4e7c5372, 0x5e61b8c0, 0x0ea19b23, 0x0ee6e1df, 0x6e63a626, 0x2f139405, 0x7eb0f96d, 0x9e588c63, 0x2e714c3a, 0x6e8c941e, 0x0f61b331, 0x6f01f625, 0x4e78d4ea, 0x6f403709, 0x1a0300da, 0xda0102c8, 0x7e61d9fd, 0xb89469bb, 0x0c838780, 0x2e60a590, 0x4dfd29e1, 0x4e150f2e, 0xce2810bc, 0x5f541591, 0x9ee60259, 0x2eb40e56, 0x5e014027, 0x2ef71faf, 0x4e2d452f, 0x5ee0a813, 0x4eb03301, 0x38443acf, 0x6eabd502, 0x0e2ee71e, 0x5a960364, 0xce7ec596, 0x7efbed09, 0x4ef42ea2, 0x0eb30ea5, 0x5ee0d9f8, 0x6f513552, 0xf89eb3fa, 0x7ea2eca6, 0x9b00cc19, 0xf897409e, 0x1e73485f, 0x381afa77, 0x0f169f3b, 0x5ee1aa70, 0x5e1803ee, 0x0dbf5a4c, 0xce78c7a6, 0x9b0b260c, 0x2ef8fa19, 0x6e70aa4b, 0xce45b805, 0x2ea08e86, 0x4ee0bafd, 0x2ea09a1f, 0x4e218900, 0x6e744f13, 0xce518653, 0xf81b7a68, 0xce45ac5e, 0x7e62e416, 0x1a1b02b6, 0x7e21db48, 0x381daaaf, 0x6b2c0987, 0x0e2ec651, 0x4eae8502, 0x9bde7ca0, 0x6f47201f, 0x7e61a8a3, 0x6e60d5db, 0x4e2879de, 0xf81d194e, 0x4f1b8d05, 0x4d0048b2, 0x6e203be9, 0x4e3e7eb1, 0x0e260ef8, 0x2e688518, 0x7e3fec46, 0xdac00843, 0xf85c8917, 0x2e212a0f, 0x0e8196da, 0xd503359f, 0xce4c81f2, 0x6ee19992, 0x6e21ca79, 0x4d40c1d2, 0x4f5816ef, 0x4e34c3ea, 0x4df7c283, 0x7ef7eeb6, 0x18e276ce, 0xab0d21c0, 0xd5032f7f, 0x4ea00dbf, 0x5ac01251, 0xd0121955, 0x7f1495e4, 0x7ef0fa11, 0x5e24dd9c, 0x9add25b5, 0x0eb2bdef, 0x9e1977c7, 0x6f4b26bd, 0x0e200a9c, 0x9b4f7c00, 0x0ea0392e, 0x7e212a2c, 0x0b248b90, 0x1acc27a1, 0x2e701c90, 0x5ee1b870, 0x5e280aba, 0x5ea0780e, 0x1e264246, 0x4e052d04, 0x0e731dc4, 0xce461997, 0x9a9e9413, 0x3d462048, 0x5ea1fac5, 0x2ea0c8c4, 0x9a030280, 0x2ebda4b8, 0x5eef8614, 0x6eadc4e0, 0xbd035a8f, 0x4e606b84, 0x4eb1aba1, 0x4e286928, 0x4e2858cc, 0x9add0ce9, 0x4e070d65, 0x5fd399d5, 0x0f03fde7, 0x6ee90c74, 0x4ef8e31e, 0x381d986a, 0x5ea0ebf4, 0x5ea0d87e, 0x2e76ac9e, 0x6eb36cd4, 0x2e6e1c4c, 0x2e2feebc, 0x1ace4b03, 0x5ee0db12, 0x5ea0e9b1, 0x2e1c32d5, 0x5fa49a09, 0x0e258737, 0x7e21ca8e, 0xce4f9988, 0x5f7f56a6, 0x0e739766, 0x4e28586c, 0x6e619908, 0xd500401f, 0xf88b9252, 0x6e251c8e, 0x9e20015b, 0x7f1486b9, 0x717c339b, 0x1f31ff70, 0x4ea0eb62, 0x9acb0926, 0x489f7d85, 0x4e209b54, 0x2e84cf03, 0x2e65946c, 0x0e7d80cd, 0xc8dffecc, 0xce668bd8, 0x6e2188af, 0xeb4ada34, 0x2b25ec33, 0x0d40e6e7, 0x4eb2c757, 0x4ec82ad0, 0x7e21cb0a, 0x0e21a847, 0x4e0b1ec0, 0x381e6ac0, 0x6e61c8f5, 0x0f10071c, 0x2ee21daa, 0x5e61ab31, 0x6e218892, 0x2e7e7cb5, 0x6f2826aa, 0x7f6b54df, 0x4eaa2620, 0xdac00034, 0x4f6477be, 0x7e6148ea, 0x4eef1f57, 0x78459aeb, 0x2ebc3f10, 0x2e35f4eb, 0x4fbf19ce, 0xd8d0e58e, 0x2e21bbc7, 0x6ee0cab6, 0x9bc57e3f, 0x2f854037, 0x4e92181c, 0x6e6d1f89, 0x0f305545, 0x4ee19a57, 0x0e887bdf, 0x5e1a4185, 0x7ef0c821, 0x2eb6607c, 0x2ea0d9b8, 0x9e0380f4, 0x2ebf1c83, 0x1e62597d, 0x7f6e2548, 0x5ac00205, 0x4e616adb, 0xce638b8c, 0x5e1653cf, 0x2e6069be, 0x0e2ac641, 0x1e33c76f, 0xce44956d, 0x9bb90d31, 0x1e24c20a, 0x7ee038c1, 0x93407e5e, 0x4e280127, 0xc8df7f7d, 0xba42f263, 0x1e6f199c, 0x6e212889, 0x6e92f60e, 0x6ebdc499, 0x8b9acbf8, 0x4d40c581, 0x3a020250, 0x6e6a6716, 0x9248403b, 0x9081ffea, 0x4e603856, 0x9ad1242b, 0x6f270579, 0x1a070349, 0xcec08133, 0xd503305f, 0x5a1a00ca, 0x2e60b8a2, 0x0e5f28fd, 0x0e31a3da, 0x7e61cbc1, 0xd503399f, 0x5f5e54aa, 0x0eb8bdea, 0x4eba8f10, 0x4e2a2e60, 0x2f3da7d6, 0x1e58e297, 0x6e71aa3e, 0x6b86701a, 0xce4fa5e6, 0x4ee7c463, 0x8a79307f, 0x0ebea541, 0x2e218af4, 0x4e774f8a, 0xb9b95dc5, 0x6e61abd5, 0x4dd1e814, 0x4da72098, 0x98307582, 0x3a512101, 0x7ef95497, 0x1ace5535, 0x5a0c0349, 0x4e28581b, 0x6ebf1c02, 0x5ea1da23, 0x1e274314, 0x5e25dd29, 0x6e75f594, 0x6eaf6ed5, 0x4e214abe, 0x4e064172, 0x2e21c8f4, 0xf84c5b08, 0x1e244312, 0x14000000};
  1193. env.code_mem.emplace_back(0x14000000); // B .
  1194. jit.SetRegister(0, 0x866524401a1d4e47);
  1195. jit.SetRegister(1, 0x02ca8cec51301b60);
  1196. jit.SetRegister(2, 0x0d2e0921242a853d);
  1197. jit.SetRegister(3, 0x5ce3dda7d19ec198);
  1198. jit.SetRegister(4, 0x8a608e22fb3f50d9);
  1199. jit.SetRegister(5, 0x97eab1c959f550bb);
  1200. jit.SetRegister(6, 0xdb6d004e7503e72a);
  1201. jit.SetRegister(7, 0xbc585cf4f01fee85);
  1202. jit.SetRegister(8, 0xd7873927978802ca);
  1203. jit.SetRegister(9, 0xf64d146839cc0275);
  1204. jit.SetRegister(10, 0xada655f0c8013f78);
  1205. jit.SetRegister(11, 0x9c06b18d34ad718a);
  1206. jit.SetRegister(12, 0xaa46ab9693a7549f);
  1207. jit.SetRegister(13, 0xdc0392ca7ded1f12);
  1208. jit.SetRegister(14, 0xb86b5a280b452d1e);
  1209. jit.SetRegister(15, 0x4cafeaf58ccf472e);
  1210. jit.SetRegister(16, 0x21fcba85c1ed26ba);
  1211. jit.SetRegister(17, 0xca8075f2eb56e277);
  1212. jit.SetRegister(18, 0x3f06bc758608d762);
  1213. jit.SetRegister(19, 0xbbc5a0aecff698e5);
  1214. jit.SetRegister(20, 0x02170439baa29e14);
  1215. jit.SetRegister(21, 0x0e7a29e1ab81b89b);
  1216. jit.SetRegister(22, 0xe8af1b958d645884);
  1217. jit.SetRegister(23, 0x86691d7e0500e2e9);
  1218. jit.SetRegister(24, 0x4983e6e57f0602c1);
  1219. jit.SetRegister(25, 0x4077d562a05048c5);
  1220. jit.SetRegister(26, 0x7019154cfcba3e12);
  1221. jit.SetRegister(27, 0xfb17997ce5f6a4ce);
  1222. jit.SetRegister(28, 0x6eb7a6b778e3dbca);
  1223. jit.SetRegister(29, 0x2ca051e70a4743be);
  1224. jit.SetRegister(30, 0x91fcc5fdd8a78378);
  1225. jit.SetPC(100);
  1226. jit.SetSP(0x000000cdfadeaff0);
  1227. env.code_mem_start_address = 100;
  1228. jit.SetVector(0, {0x4d5a180ac0ffdac8, 0xfc6eb113cd5ff2a8});
  1229. jit.SetVector(1, {0x39f8cecc9de9cefd, 0x3a6b35d333d89a6b});
  1230. jit.SetVector(2, {0x791fd8290bbdd2f4, 0xdc0e5e7aee311411});
  1231. jit.SetVector(3, {0xd97db4cbd67fe7de, 0x50042a5e0b94f71c});
  1232. jit.SetVector(4, {0xe2b93543509f65a7, 0xaa1b6433c337c5b9});
  1233. jit.SetVector(5, {0xd93ee9fc22c5edf7, 0xe9042e8f2a2279d3});
  1234. jit.SetVector(6, {0x988cf27e5c9928ad, 0xc1a39aa7429018af});
  1235. jit.SetVector(7, {0x8f24fd7c96752d5e, 0x211ed066df4bf60d});
  1236. jit.SetVector(8, {0xec12260921aa0e5d, 0xcb98d7c3aa39bb54});
  1237. jit.SetVector(9, {0x8ae0d63bef16836b, 0x54b582f6c7c563d5});
  1238. jit.SetVector(10, {0xd36cb5833320a802, 0x94afbd35a90c0d01});
  1239. jit.SetVector(11, {0xf80d24f3de920bb5, 0x8505fd820fdca5ac});
  1240. jit.SetVector(12, {0xc4d5ee040479c10a, 0xb9a65305f855b401});
  1241. jit.SetVector(13, {0xe258117dea0e2e1d, 0x50b6e47f2cbbf98f});
  1242. jit.SetVector(14, {0x8c46631befe40367, 0x76ef634acc1d252e});
  1243. jit.SetVector(15, {0x31ba2e4997445a39, 0xeea2b7e296ed9a10});
  1244. jit.SetVector(16, {0xb1b6ad7f6888ad82, 0x22d61f3a89e351f2});
  1245. jit.SetVector(17, {0x38556d902cb1e166, 0xd94cd8ece8871a9b});
  1246. jit.SetVector(18, {0x8022388e51111894, 0x8319843c0f97c296});
  1247. jit.SetVector(19, {0x80950f4f1988738e, 0x2b51d501a2ac843e});
  1248. jit.SetVector(20, {0xd959d91895a0e304, 0xd86a18f9fbca97cd});
  1249. jit.SetVector(21, {0x9b06de585c91b8f6, 0x6a27b488c3137c9c});
  1250. jit.SetVector(22, {0x95970398b8941fde, 0x85f81fbbf5989d74});
  1251. jit.SetVector(23, {0x33926666f9db44d7, 0xf36ed3933d067e0f});
  1252. jit.SetVector(24, {0x1aefb2ab9a149525, 0xbff5abf69badf81f});
  1253. jit.SetVector(25, {0x88492c5b044f4d83, 0x3fc4029fe302c62c});
  1254. jit.SetVector(26, {0x0cfcc374a4866662, 0xcec449f82b95bc0f});
  1255. jit.SetVector(27, {0x54506ca290052cf6, 0x22f41aa29a475adb});
  1256. jit.SetVector(28, {0x7baf46a55161f432, 0xe7426c082b417919});
  1257. jit.SetVector(29, {0x03a801b9d543654e, 0xb78f7f602ad245ee});
  1258. jit.SetVector(30, {0x656014c093d5ef4a, 0x180caaef9d32e7ab});
  1259. jit.SetVector(31, {0xb6f6e9d497f143b9, 0x1c52381350356431});
  1260. jit.SetPstate(0xb0000000);
  1261. jit.SetFpcr(0x01000000);
  1262. env.ticks_left = 110;
  1263. jit.Run();
  1264. REQUIRE(jit.GetVector(0) == Vector{0x0101010211914707, 0x090000007fd9991a});
  1265. REQUIRE(jit.GetVector(1) == Vector{0x00000000fffffffe, 0x0000000000000000});
  1266. REQUIRE(jit.GetVector(2) == Vector{0x05004503877a2f45, 0x0000000000000000});
  1267. REQUIRE(jit.GetVector(3) == Vector{0x000000007f800000, 0x0000000000000000});
  1268. REQUIRE(jit.GetVector(4) == Vector{0xffffffff00000000, 0x0000000000000000});
  1269. REQUIRE(jit.GetVector(5) == Vector{0xda00894d7886d0bb, 0x5cc5a3b2ca6afb26});
  1270. REQUIRE(jit.GetVector(6) == Vector{0x0000000000000000, 0xfffffffd00000000});
  1271. REQUIRE(jit.GetVector(7) == Vector{0x0000000000000000, 0x0000000000000000});
  1272. REQUIRE(jit.GetVector(8) == Vector{0x0000000000000000, 0x0000000000000000});
  1273. REQUIRE(jit.GetVector(9) == Vector{0x00000000ff800000, 0x0000000000000000});
  1274. REQUIRE(jit.GetVector(10) == Vector{0xc000000000000000, 0x0000000000000000});
  1275. REQUIRE(jit.GetVector(11) == Vector{0xffff000000000000, 0x0000000000000000});
  1276. REQUIRE(jit.GetVector(12) == Vector{0x0c0bd08451d5d9b3, 0x0000000000000000});
  1277. REQUIRE(jit.GetVector(13) == Vector{0x0000000000000000, 0xdc1e34ac00000000});
  1278. REQUIRE(jit.GetVector(14) == Vector{0x00000000ffffffff, 0x0000000000000000});
  1279. REQUIRE(jit.GetVector(15) == Vector{0xfbdfff7cf38fba7d, 0xfffffffffffffffe});
  1280. REQUIRE(jit.GetVector(16) == Vector{0x0000000000000000, 0x0000000000000000});
  1281. REQUIRE(jit.GetVector(17) == Vector{0x0000000000000000, 0x0000000000000000});
  1282. REQUIRE(jit.GetVector(18) == Vector{0x00000000ffffffff, 0x0000000000000000});
  1283. REQUIRE(jit.GetVector(19) == Vector{0x0000000000000000, 0x090000007fd9991a});
  1284. REQUIRE(jit.GetVector(20) == Vector{0x0000000000000000, 0x0000000000000000});
  1285. REQUIRE(jit.GetVector(21) == Vector{0xdbdad9d8dbdad9d8, 0xdbdad9d8dbdad9d8});
  1286. REQUIRE(jit.GetVector(22) == Vector{0xdbdad9d8dbdad9d8, 0xdbdad9d8dbdad9d8});
  1287. REQUIRE(jit.GetVector(23) == Vector{0xffffffff00000000, 0x0000000000000000});
  1288. REQUIRE(jit.GetVector(24) == Vector{0xffffffffffffffff, 0x0000000000000000});
  1289. REQUIRE(jit.GetVector(25) == Vector{0x0000007f00000000, 0x0000000000000000});
  1290. REQUIRE(jit.GetVector(26) == Vector{0x0000000000000000, 0x0000000000000000});
  1291. REQUIRE(jit.GetVector(27) == Vector{0x3a7d96116b237d60, 0x0c6bd37dd698d82a});
  1292. REQUIRE(jit.GetVector(28) == Vector{0x8000000000000000, 0x0000000000000000});
  1293. REQUIRE(jit.GetVector(29) == Vector{0xb3b2000000000000, 0x0000000000000000});
  1294. REQUIRE(jit.GetVector(30) == Vector{0x0000000000000000, 0x8080808080808080});
  1295. REQUIRE(jit.GetVector(31) == Vector{0xb3b2b3b200000000, 0x0000000000000000});
  1296. }
  1297. TEST_CASE("A64: SABD", "[a64]") {
  1298. A64TestEnv env;
  1299. A64::Jit jit{A64::UserConfig{&env}};
  1300. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  1301. code.SABD(V0.B16(), V3.B16(), V4.B16());
  1302. code.SABD(V1.H8(), V5.H8(), V6.H8());
  1303. code.SABD(V2.S4(), V7.S4(), V8.S4());
  1304. constexpr std::array<Vector, 9> vectors = {
  1305. // expected output vectors (int8, int16, int32)
  1306. Vector{0xa8'4a'cd'0f'7b'2b'78'49, 0x00'ff'88'01'29'34'10'1d},
  1307. Vector{0x1b8c'83cc'4640'37e5, 0x1696'ab90'3d96'2155},
  1308. Vector{0x1c656335'733d91c4, 0x1a488da4'b025dc65},
  1309. // int8 input vectors [3-4]
  1310. Vector{0x81'60'7e'60'c4'd6'20'34, 0x12'7f'f7'00'3f'db'0b'a0},
  1311. Vector{0x29'16'b1'6f'3f'ab'a8'7d, 0x12'80'7f'ff'16'0f'fb'83},
  1312. // int16 input vectors [5-6]
  1313. Vector{0x8bbd'c450'2dd9'7179, 0xf171'966c'33f2'423b},
  1314. Vector{0xa749'481c'e799'3994, 0xdadb'41fc'f65c'20e6},
  1315. // int32 input vectors [7-8]
  1316. Vector{0x57816e27'df8b9293, 0xe1808186'495e497a},
  1317. Vector{0x73e6d15c'52c92457, 0xfbc90f2a'99386d15},
  1318. };
  1319. jit.SetPC(0);
  1320. jit.SetVector(3, vectors[3]);
  1321. jit.SetVector(4, vectors[4]);
  1322. jit.SetVector(5, vectors[5]);
  1323. jit.SetVector(6, vectors[6]);
  1324. jit.SetVector(7, vectors[7]);
  1325. jit.SetVector(8, vectors[8]);
  1326. env.ticks_left = env.code_mem.size();
  1327. jit.Run();
  1328. CHECK(jit.GetVector(0) == vectors[0]);
  1329. CHECK(jit.GetVector(1) == vectors[1]);
  1330. CHECK(jit.GetVector(2) == vectors[2]);
  1331. // ensure the correct results are not being produced randomly
  1332. jit.SetPC(0);
  1333. jit.SetVectors(std::array<Vector, 32>{});
  1334. jit.SetVector(3, vectors[4]);
  1335. jit.SetVector(4, vectors[3]);
  1336. jit.SetVector(5, vectors[6]);
  1337. jit.SetVector(6, vectors[5]);
  1338. jit.SetVector(7, vectors[8]);
  1339. jit.SetVector(8, vectors[7]);
  1340. env.ticks_left = 4;
  1341. jit.Run();
  1342. CHECK(jit.GetVector(0) == vectors[0]);
  1343. CHECK(jit.GetVector(1) == vectors[1]);
  1344. CHECK(jit.GetVector(2) == vectors[2]);
  1345. }
  1346. TEST_CASE("A64: UZP{1,2}.2D", "[a64]") {
  1347. A64TestEnv env;
  1348. A64::Jit jit{A64::UserConfig{&env}};
  1349. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  1350. code.UZP1(V2.D2(), V0.D2(), V1.D2());
  1351. code.UZP2(V3.D2(), V0.D2(), V1.D2());
  1352. jit.SetPC(0);
  1353. jit.SetVector(0, {0xF0F1F2F3F4F5F6F7, 0xE0E1E2E3E4E5E6E7});
  1354. jit.SetVector(1, {0xA0A1A2A3A4A5A6A7, 0xB0B1B2B3B4B5B6B7});
  1355. env.ticks_left = env.code_mem.size();
  1356. jit.Run();
  1357. REQUIRE(jit.GetVector(2) == Vector{0xF0F1F2F3F4F5F6F7, 0xA0A1A2A3A4A5A6A7});
  1358. REQUIRE(jit.GetVector(3) == Vector{0xE0E1E2E3E4E5E6E7, 0xB0B1B2B3B4B5B6B7});
  1359. }
  1360. TEST_CASE("A64: UZP{1,2}.S", "[a64]") {
  1361. A64TestEnv env;
  1362. A64::Jit jit{A64::UserConfig{&env}};
  1363. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  1364. code.UZP1(V2.S2(), V0.S2(), V1.S2());
  1365. code.UZP2(V3.S2(), V0.S2(), V1.S2());
  1366. code.UZP1(V4.S4(), V0.S4(), V1.S4());
  1367. code.UZP2(V5.S4(), V0.S4(), V1.S4());
  1368. jit.SetPC(0);
  1369. jit.SetVector(0, {0xF4F5F6F7'F0F1F2F3, 0xE4E5E6E7'E0E1E2E3});
  1370. jit.SetVector(1, {0xA4A5A6A7'A0A1A2A3, 0xB4B5B6B7'B0B1B2B3});
  1371. env.ticks_left = env.code_mem.size();
  1372. jit.Run();
  1373. REQUIRE(jit.GetVector(2) == Vector{0xA0A1A2A3'F0F1F2F3, 0});
  1374. REQUIRE(jit.GetVector(3) == Vector{0xA4A5A6A7'F4F5F6F7, 0});
  1375. REQUIRE(jit.GetVector(4) == Vector{0xE0E1E2E3'F0F1F2F3, 0xB0B1B2B3'A0A1A2A3});
  1376. REQUIRE(jit.GetVector(5) == Vector{0xE4E5E6E7'F4F5F6F7, 0xB4B5B6B7'A4A5A6A7});
  1377. }
  1378. TEST_CASE("A64: UZP{1,2}.H", "[a64]") {
  1379. A64TestEnv env;
  1380. A64::Jit jit{A64::UserConfig{&env}};
  1381. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  1382. code.UZP1(V2.H4(), V0.H4(), V1.H4());
  1383. code.UZP2(V3.H4(), V0.H4(), V1.H4());
  1384. code.UZP1(V4.H8(), V0.H8(), V1.H8());
  1385. code.UZP2(V5.H8(), V0.H8(), V1.H8());
  1386. jit.SetPC(0);
  1387. jit.SetVector(0, {0xF6F7'F4F5'F2F3'F0F1, 0xE6E7'E4E5'E2E3'E0E1});
  1388. jit.SetVector(1, {0xA6A7'A4A5'A2A3'A0A1, 0xB6B7'B4B5'B2B3'B0B1});
  1389. env.ticks_left = env.code_mem.size();
  1390. jit.Run();
  1391. REQUIRE(jit.GetVector(2) == Vector{0xA4A5'A0A1'F4F5'F0F1, 0});
  1392. REQUIRE(jit.GetVector(3) == Vector{0xA6A7'A2A3'F6F7'F2F3, 0});
  1393. REQUIRE(jit.GetVector(4) == Vector{0xE4E5'E0E1'F4F5'F0F1, 0xB4B5'B0B1'A4A5'A0A1});
  1394. REQUIRE(jit.GetVector(5) == Vector{0xE6E7'E2E3'F6F7'F2F3, 0xB6B7'B2B3'A6A7'A2A3});
  1395. }
  1396. TEST_CASE("A64: UZP{1,2}.B", "[a64]") {
  1397. A64TestEnv env;
  1398. A64::Jit jit{A64::UserConfig{&env}};
  1399. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  1400. code.UZP1(V2.B8(), V0.B8(), V1.B8());
  1401. code.UZP2(V3.B8(), V0.B8(), V1.B8());
  1402. code.UZP1(V4.B16(), V0.B16(), V1.B16());
  1403. code.UZP2(V5.B16(), V0.B16(), V1.B16());
  1404. jit.SetPC(0);
  1405. jit.SetVector(0, {0xF7'F6'F5'F4'F3'F2'F1'F0, 0xE7'E6'E5'E4'E3'E2'E1'E0});
  1406. jit.SetVector(1, {0xA7'A6'A5'A4'A3'A2'A1'A0, 0xB7'B6'B5'B4'B3'B2'B1'B0});
  1407. env.ticks_left = env.code_mem.size();
  1408. jit.Run();
  1409. REQUIRE(jit.GetVector(2) == Vector{0xA6'A4'A2'A0'F6'F4'F2'F0, 0});
  1410. REQUIRE(jit.GetVector(3) == Vector{0xA7'A5'A3'A1'F7'F5'F3'F1, 0});
  1411. REQUIRE(jit.GetVector(4) == Vector{0xE6'E4'E2'E0'F6'F4'F2'F0, 0xB6'B4'B2'B0'A6'A4'A2'A0});
  1412. REQUIRE(jit.GetVector(5) == Vector{0xE7'E5'E3'E1'F7'F5'F3'F1, 0xB7'B5'B3'B1'A7'A5'A3'A1});
  1413. }
  1414. TEST_CASE("A64: {S,U}MIN.S, {S,U}MAX.S", "[a64]") {
  1415. A64TestEnv env;
  1416. A64::Jit jit{A64::UserConfig{&env}};
  1417. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  1418. code.SMIN(V2.S4(), V0.S4(), V1.S4());
  1419. code.UMIN(V3.S4(), V0.S4(), V1.S4());
  1420. code.SMAX(V4.S4(), V0.S4(), V1.S4());
  1421. code.UMAX(V5.S4(), V0.S4(), V1.S4());
  1422. code.SMIN(V12.S4(), V1.S4(), V0.S4());
  1423. code.UMIN(V13.S4(), V1.S4(), V0.S4());
  1424. code.SMAX(V14.S4(), V1.S4(), V0.S4());
  1425. code.UMAX(V15.S4(), V1.S4(), V0.S4());
  1426. constexpr std::array<Vector, 6> vectors = {
  1427. // initial input vectors [0-1]
  1428. Vector{0x7FFFFFFF'00000002, 0x76543209'01234567},
  1429. Vector{0x80000000'00000003, 0x76543210'F1234567},
  1430. // expected output vectors [2-5]
  1431. Vector{0x80000000'00000002, 0x76543209'F1234567},
  1432. Vector{0x7FFFFFFF'00000002, 0x76543209'01234567},
  1433. Vector{0x7FFFFFFF'00000003, 0x76543210'01234567},
  1434. Vector{0x80000000'00000003, 0x76543210'F1234567},
  1435. };
  1436. jit.SetPC(0);
  1437. jit.SetVector(0, vectors[0]);
  1438. jit.SetVector(1, vectors[1]);
  1439. env.ticks_left = env.code_mem.size();
  1440. jit.Run();
  1441. CHECK(jit.GetVector(2) == vectors[2]);
  1442. CHECK(jit.GetVector(3) == vectors[3]);
  1443. CHECK(jit.GetVector(4) == vectors[4]);
  1444. CHECK(jit.GetVector(5) == vectors[5]);
  1445. CHECK(jit.GetVector(12) == vectors[2]);
  1446. CHECK(jit.GetVector(13) == vectors[3]);
  1447. CHECK(jit.GetVector(14) == vectors[4]);
  1448. CHECK(jit.GetVector(15) == vectors[5]);
  1449. }
  1450. TEST_CASE("A64: {S,U}MIN.H, {S,U}MAX.H", "[a64]") {
  1451. A64TestEnv env;
  1452. A64::Jit jit{A64::UserConfig{&env}};
  1453. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  1454. code.SMIN(V2.H8(), V0.H8(), V1.H8());
  1455. code.UMIN(V3.H8(), V0.H8(), V1.H8());
  1456. code.SMAX(V4.H8(), V0.H8(), V1.H8());
  1457. code.UMAX(V5.H8(), V0.H8(), V1.H8());
  1458. code.SMIN(V12.H8(), V1.H8(), V0.H8());
  1459. code.UMIN(V13.H8(), V1.H8(), V0.H8());
  1460. code.SMAX(V14.H8(), V1.H8(), V0.H8());
  1461. code.UMAX(V15.H8(), V1.H8(), V0.H8());
  1462. constexpr std::array<Vector, 6> vectors = {
  1463. // initial input vectors [0-1]
  1464. Vector{0x0123'0000'0002'7FFE, 0x8764'0123'7FFF'FFFE},
  1465. Vector{0xF123'FFFF'0003'7FFF, 0x8765'0124'8000'FFFF},
  1466. // expected output vectors [2-5]
  1467. Vector{0xF123'FFFF'0002'7FFE, 0x8764'0123'8000'FFFE},
  1468. Vector{0x0123'0000'0002'7FFE, 0x8764'0123'7FFF'FFFE},
  1469. Vector{0x0123'0000'0003'7FFF, 0x8765'0124'7FFF'FFFF},
  1470. Vector{0xF123'FFFF'0003'7FFF, 0x8765'0124'8000'FFFF},
  1471. };
  1472. jit.SetPC(0);
  1473. jit.SetVector(0, vectors[0]);
  1474. jit.SetVector(1, vectors[1]);
  1475. env.ticks_left = env.code_mem.size();
  1476. jit.Run();
  1477. CHECK(jit.GetVector(2) == vectors[2]);
  1478. CHECK(jit.GetVector(3) == vectors[3]);
  1479. CHECK(jit.GetVector(4) == vectors[4]);
  1480. CHECK(jit.GetVector(5) == vectors[5]);
  1481. CHECK(jit.GetVector(12) == vectors[2]);
  1482. CHECK(jit.GetVector(13) == vectors[3]);
  1483. CHECK(jit.GetVector(14) == vectors[4]);
  1484. CHECK(jit.GetVector(15) == vectors[5]);
  1485. }
  1486. TEST_CASE("A64: {S,U}MIN.B, {S,U}MAX.B", "[a64]") {
  1487. A64TestEnv env;
  1488. A64::Jit jit{A64::UserConfig{&env}};
  1489. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  1490. code.SMIN(V2.B16(), V0.B16(), V1.B16());
  1491. code.UMIN(V3.B16(), V0.B16(), V1.B16());
  1492. code.SMAX(V4.B16(), V0.B16(), V1.B16());
  1493. code.UMAX(V5.B16(), V0.B16(), V1.B16());
  1494. code.SMIN(V12.B16(), V1.B16(), V0.B16());
  1495. code.UMIN(V13.B16(), V1.B16(), V0.B16());
  1496. code.SMAX(V14.B16(), V1.B16(), V0.B16());
  1497. code.UMAX(V15.B16(), V1.B16(), V0.B16());
  1498. constexpr std::array<Vector, 6> vectors = {
  1499. // initial input vectors [0-1]
  1500. Vector{0x40'70'F0'A0'02'7E'7F'FE, 0xC2'B0'7E'7F'00'18'9A'12},
  1501. Vector{0x41'71'F1'A1'03'7F'80'FF, 0xC3'B1'82'81'FF'81'99'34},
  1502. // expected output vectors [2-5]
  1503. Vector{0x40'70'F0'A0'02'7E'80'FE, 0xC2'B0'82'81'FF'81'99'12},
  1504. Vector{0x40'70'F0'A0'02'7E'7F'FE, 0xC2'B0'7E'7F'00'18'99'12},
  1505. Vector{0x41'71'F1'A1'03'7F'7F'FF, 0xC3'B1'7E'7F'00'18'9A'34},
  1506. Vector{0x41'71'F1'A1'03'7F'80'FF, 0xC3'B1'82'81'FF'81'9A'34},
  1507. };
  1508. jit.SetPC(0);
  1509. jit.SetVector(0, vectors[0]);
  1510. jit.SetVector(1, vectors[1]);
  1511. env.ticks_left = env.code_mem.size();
  1512. jit.Run();
  1513. CHECK(jit.GetVector(2) == vectors[2]);
  1514. CHECK(jit.GetVector(3) == vectors[3]);
  1515. CHECK(jit.GetVector(4) == vectors[4]);
  1516. CHECK(jit.GetVector(5) == vectors[5]);
  1517. CHECK(jit.GetVector(12) == vectors[2]);
  1518. CHECK(jit.GetVector(13) == vectors[3]);
  1519. CHECK(jit.GetVector(14) == vectors[4]);
  1520. CHECK(jit.GetVector(15) == vectors[5]);
  1521. }
  1522. TEST_CASE("A64: {S,U}MINP.S, {S,U}MAXP.S", "[a64]") {
  1523. A64TestEnv env;
  1524. A64::Jit jit{A64::UserConfig{&env}};
  1525. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  1526. code.SMINP(V2.S2(), V0.S2(), V1.S2());
  1527. code.UMINP(V3.S2(), V0.S2(), V1.S2());
  1528. code.SMINP(V4.S4(), V0.S4(), V1.S4());
  1529. code.UMINP(V5.S4(), V0.S4(), V1.S4());
  1530. code.SMAXP(V6.S2(), V0.S2(), V1.S2());
  1531. code.UMAXP(V7.S2(), V0.S2(), V1.S2());
  1532. code.SMAXP(V8.S4(), V0.S4(), V1.S4());
  1533. code.UMAXP(V9.S4(), V0.S4(), V1.S4());
  1534. constexpr std::array<Vector, 12> vectors = {
  1535. // initial input vectors [0-1]
  1536. Vector{0x00000003'00000002, 0xF1234567'01234567},
  1537. Vector{0x80000000'7FFFFFFF, 0x76543210'76543209},
  1538. // expected output vectors [2-9]
  1539. Vector{0x80000000'00000002, 0},
  1540. Vector{0x7FFFFFFF'00000002, 0},
  1541. Vector{0xF1234567'00000002, 0x76543209'80000000},
  1542. Vector{0x01234567'00000002, 0x76543209'7FFFFFFF},
  1543. Vector{0x7FFFFFFF'00000003, 0},
  1544. Vector{0x80000000'00000003, 0},
  1545. Vector{0x01234567'00000003, 0x76543210'7FFFFFFF},
  1546. Vector{0xF1234567'00000003, 0x76543210'80000000},
  1547. // input vectors with elements swapped pairwise [10-11]
  1548. Vector{0x00000002'00000003, 0x01234567'F1234567},
  1549. Vector{0x7FFFFFFF'80000000, 0x76543209'76543210},
  1550. };
  1551. jit.SetPC(0);
  1552. jit.SetVector(0, vectors[0]);
  1553. jit.SetVector(1, vectors[1]);
  1554. env.ticks_left = env.code_mem.size();
  1555. jit.Run();
  1556. CHECK(jit.GetVector(2) == vectors[2]);
  1557. CHECK(jit.GetVector(3) == vectors[3]);
  1558. CHECK(jit.GetVector(4) == vectors[4]);
  1559. CHECK(jit.GetVector(5) == vectors[5]);
  1560. CHECK(jit.GetVector(6) == vectors[6]);
  1561. CHECK(jit.GetVector(7) == vectors[7]);
  1562. CHECK(jit.GetVector(8) == vectors[8]);
  1563. CHECK(jit.GetVector(9) == vectors[9]);
  1564. // run the same tests again but with the input vectors swapped pairwise,
  1565. // to ensure we aren't randomly producing the correct values
  1566. jit.SetPC(0);
  1567. jit.SetVectors(std::array<Vector, 32>{});
  1568. jit.SetVector(0, vectors[10]);
  1569. jit.SetVector(1, vectors[11]);
  1570. env.ticks_left = env.code_mem.size();
  1571. jit.Run();
  1572. CHECK(jit.GetVector(2) == vectors[2]);
  1573. CHECK(jit.GetVector(3) == vectors[3]);
  1574. CHECK(jit.GetVector(4) == vectors[4]);
  1575. CHECK(jit.GetVector(5) == vectors[5]);
  1576. CHECK(jit.GetVector(6) == vectors[6]);
  1577. CHECK(jit.GetVector(7) == vectors[7]);
  1578. CHECK(jit.GetVector(8) == vectors[8]);
  1579. CHECK(jit.GetVector(9) == vectors[9]);
  1580. }
  1581. TEST_CASE("A64: {S,U}MINP.H, {S,U}MAXP.H", "[a64]") {
  1582. A64TestEnv env;
  1583. A64::Jit jit{A64::UserConfig{&env}};
  1584. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  1585. code.SMINP(V2.H4(), V0.H4(), V1.H4());
  1586. code.UMINP(V3.H4(), V0.H4(), V1.H4());
  1587. code.SMINP(V4.H8(), V0.H8(), V1.H8());
  1588. code.UMINP(V5.H8(), V0.H8(), V1.H8());
  1589. code.SMAXP(V6.H4(), V0.H4(), V1.H4());
  1590. code.UMAXP(V7.H4(), V0.H4(), V1.H4());
  1591. code.SMAXP(V8.H8(), V0.H8(), V1.H8());
  1592. code.UMAXP(V9.H8(), V0.H8(), V1.H8());
  1593. constexpr std::array<Vector, 12> vectors = {
  1594. // initial input vectors [0-1]
  1595. Vector{0x0003'0002'7FFF'7FFE, 0xF123'0123'FFFF'0000},
  1596. Vector{0x8000'7FFF'FFFF'FFFE, 0x8765'8764'0123'0124},
  1597. // expected output vectors [2-9]
  1598. Vector{0x8000'FFFE'0002'7FFE, 0},
  1599. Vector{0x7FFF'FFFE'0002'7FFE, 0},
  1600. Vector{0xF123'FFFF'0002'7FFE, 0x8764'0123'8000'FFFE},
  1601. Vector{0x0123'0000'0002'7FFE, 0x8764'0123'7FFF'FFFE},
  1602. Vector{0x7FFF'FFFF'0003'7FFF, 0},
  1603. Vector{0x8000'FFFF'0003'7FFF, 0},
  1604. Vector{0x0123'0000'0003'7FFF, 0x8765'0124'7FFF'FFFF},
  1605. Vector{0xF123'FFFF'0003'7FFF, 0x8765'0124'8000'FFFF},
  1606. // input vectors with elements swapped pairwise [10-11]
  1607. Vector{0x0002'0003'7FFE'7FFF, 0x0123'F123'0000'FFFF},
  1608. Vector{0x7FFF'8000'FFFE'FFFF, 0x8764'8765'0124'0123},
  1609. };
  1610. jit.SetPC(0);
  1611. jit.SetVector(0, vectors[0]);
  1612. jit.SetVector(1, vectors[1]);
  1613. env.ticks_left = env.code_mem.size();
  1614. jit.Run();
  1615. CHECK(jit.GetVector(2) == vectors[2]);
  1616. CHECK(jit.GetVector(3) == vectors[3]);
  1617. CHECK(jit.GetVector(4) == vectors[4]);
  1618. CHECK(jit.GetVector(5) == vectors[5]);
  1619. CHECK(jit.GetVector(6) == vectors[6]);
  1620. CHECK(jit.GetVector(7) == vectors[7]);
  1621. CHECK(jit.GetVector(8) == vectors[8]);
  1622. CHECK(jit.GetVector(9) == vectors[9]);
  1623. // run the same tests again but with the input vectors swapped pairwise,
  1624. // to ensure we aren't randomly producing the correct values
  1625. jit.SetPC(0);
  1626. jit.SetVectors(std::array<Vector, 32>{});
  1627. jit.SetVector(0, vectors[10]);
  1628. jit.SetVector(1, vectors[11]);
  1629. env.ticks_left = env.code_mem.size();
  1630. jit.Run();
  1631. CHECK(jit.GetVector(2) == vectors[2]);
  1632. CHECK(jit.GetVector(3) == vectors[3]);
  1633. CHECK(jit.GetVector(4) == vectors[4]);
  1634. CHECK(jit.GetVector(5) == vectors[5]);
  1635. CHECK(jit.GetVector(6) == vectors[6]);
  1636. CHECK(jit.GetVector(7) == vectors[7]);
  1637. CHECK(jit.GetVector(8) == vectors[8]);
  1638. CHECK(jit.GetVector(9) == vectors[9]);
  1639. }
  1640. TEST_CASE("A64: {S,U}MINP.B, {S,U}MAXP.B", "[a64]") {
  1641. A64TestEnv env;
  1642. A64::Jit jit{A64::UserConfig{&env}};
  1643. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  1644. code.SMINP(V2.B8(), V0.B8(), V1.B8());
  1645. code.UMINP(V3.B8(), V0.B8(), V1.B8());
  1646. code.SMINP(V4.B16(), V0.B16(), V1.B16());
  1647. code.UMINP(V5.B16(), V0.B16(), V1.B16());
  1648. code.SMAXP(V6.B8(), V0.B8(), V1.B8());
  1649. code.UMAXP(V7.B8(), V0.B8(), V1.B8());
  1650. code.SMAXP(V8.B16(), V0.B16(), V1.B16());
  1651. code.UMAXP(V9.B16(), V0.B16(), V1.B16());
  1652. constexpr std::array<Vector, 12> vectors = {
  1653. // initial input vectors [0-1]
  1654. Vector{0x02'03'7F'7E'80'7F'FF'FE, 0x40'41'70'71'F0'F1'A0'A1},
  1655. Vector{0xFF'00'81'18'99'9A'12'34, 0xC3'C2'B1'B0'82'7E'81'7F},
  1656. // expected output vectors [2-9]
  1657. Vector{0xFF'81'99'12'02'7E'80'FE, 0},
  1658. Vector{0x00'18'99'12'02'7E'7F'FE, 0},
  1659. Vector{0x40'70'F0'A0'02'7E'80'FE, 0xC2'B0'82'81'FF'81'99'12},
  1660. Vector{0x40'70'F0'A0'02'7E'7F'FE, 0xC2'B0'7E'7F'00'18'99'12},
  1661. Vector{0x00'18'9A'34'03'7F'7F'FF, 0},
  1662. Vector{0xFF'81'9A'34'03'7F'80'FF, 0},
  1663. Vector{0x41'71'F1'A1'03'7F'7F'FF, 0xC3'B1'7E'7F'00'18'9A'34},
  1664. Vector{0x41'71'F1'A1'03'7F'80'FF, 0xC3'B1'82'81'FF'81'9A'34},
  1665. // input vectors with elements swapped pairwise [10-11]
  1666. Vector{0x03'02'7E'7F'7F'80'FE'FF, 0x41'40'71'70'F1'F0'A1'A0},
  1667. Vector{0x00'FF'18'81'9A'99'34'12, 0xC2'C3'B0'B1'7E'82'7F'81},
  1668. };
  1669. jit.SetPC(0);
  1670. jit.SetVector(0, vectors[0]);
  1671. jit.SetVector(1, vectors[1]);
  1672. env.ticks_left = env.code_mem.size();
  1673. jit.Run();
  1674. CHECK(jit.GetVector(2) == vectors[2]);
  1675. CHECK(jit.GetVector(3) == vectors[3]);
  1676. CHECK(jit.GetVector(4) == vectors[4]);
  1677. CHECK(jit.GetVector(5) == vectors[5]);
  1678. CHECK(jit.GetVector(6) == vectors[6]);
  1679. CHECK(jit.GetVector(7) == vectors[7]);
  1680. CHECK(jit.GetVector(8) == vectors[8]);
  1681. CHECK(jit.GetVector(9) == vectors[9]);
  1682. // run the same tests again but with the input vectors swapped pairwise,
  1683. // to ensure we aren't randomly producing the correct values
  1684. jit.SetPC(0);
  1685. jit.SetVectors(std::array<Vector, 32>{});
  1686. jit.SetVector(0, vectors[10]);
  1687. jit.SetVector(1, vectors[11]);
  1688. env.ticks_left = env.code_mem.size();
  1689. jit.Run();
  1690. CHECK(jit.GetVector(2) == vectors[2]);
  1691. CHECK(jit.GetVector(3) == vectors[3]);
  1692. CHECK(jit.GetVector(4) == vectors[4]);
  1693. CHECK(jit.GetVector(5) == vectors[5]);
  1694. CHECK(jit.GetVector(6) == vectors[6]);
  1695. CHECK(jit.GetVector(7) == vectors[7]);
  1696. CHECK(jit.GetVector(8) == vectors[8]);
  1697. CHECK(jit.GetVector(9) == vectors[9]);
  1698. }
  1699. TEST_CASE("A64: SQABS", "[a64]") {
  1700. A64TestEnv env;
  1701. A64::Jit jit{A64::UserConfig{&env}};
  1702. oaknut::VectorCodeGenerator code{env.code_mem, nullptr};
  1703. // should set QC flag
  1704. code.SQABS(V0.B16(), V0.B16());
  1705. code.MRS(X0, oaknut::SystemReg::FPSR);
  1706. code.MSR(oaknut::SystemReg::FPSR, XZR);
  1707. code.SQABS(V1.H8(), V1.H8());
  1708. code.MRS(X1, oaknut::SystemReg::FPSR);
  1709. code.MSR(oaknut::SystemReg::FPSR, XZR);
  1710. code.SQABS(V2.S4(), V2.S4());
  1711. code.MRS(X2, oaknut::SystemReg::FPSR);
  1712. code.MSR(oaknut::SystemReg::FPSR, XZR);
  1713. code.SQABS(V3.D2(), V3.D2());
  1714. code.MRS(X3, oaknut::SystemReg::FPSR);
  1715. code.MSR(oaknut::SystemReg::FPSR, XZR);
  1716. // should not set QC flag
  1717. code.SQABS(V10.B16(), V10.B16());
  1718. code.MRS(X10, oaknut::SystemReg::FPSR);
  1719. code.MSR(oaknut::SystemReg::FPSR, XZR);
  1720. code.SQABS(V11.H8(), V11.H8());
  1721. code.MRS(X11, oaknut::SystemReg::FPSR);
  1722. code.MSR(oaknut::SystemReg::FPSR, XZR);
  1723. code.SQABS(V12.S4(), V12.S4());
  1724. code.MRS(X12, oaknut::SystemReg::FPSR);
  1725. code.MSR(oaknut::SystemReg::FPSR, XZR);
  1726. code.SQABS(V13.D2(), V13.D2());
  1727. code.MRS(X13, oaknut::SystemReg::FPSR);
  1728. jit.SetPC(0);
  1729. jit.SetFpsr(0);
  1730. // contains one value that will be saturated
  1731. jit.SetVector(0, Vector{0x2B'7F'EC'D6'77'CE'80'10, 0x9D'EA'82'45'81'CD'42'FC});
  1732. jit.SetVector(1, Vector{0x3D74'9114'8000'B0BE, 0x3F0F'E281'CE50'0616});
  1733. jit.SetVector(2, Vector{0x218630B5'BEC18D71, 0x9042167E'80000000});
  1734. jit.SetVector(3, Vector{0x89C1B48FBC43F53B, 0x8000000000000000});
  1735. // contains no values that will be saturated
  1736. jit.SetVector(10, Vector{0x2B'7F'EC'D6'77'CE'00'10, 0x9D'EA'82'45'81'CD'42'FC});
  1737. jit.SetVector(11, Vector{0x3D74'9114'0000'B0BE, 0x3F0F'E281'CE50'0616});
  1738. jit.SetVector(12, Vector{0x218630B5'BEC18D71, 0x9042167E'00000000});
  1739. jit.SetVector(13, Vector{0x89C1B48FBC43F53B, 0x5FDD5D671D399E2});
  1740. env.ticks_left = env.code_mem.size();
  1741. jit.Run();
  1742. CHECK(jit.GetVector(0) == Vector{0x2B'7F'14'2A'77'32'7F'10, 0x63'16'7E'45'7F'33'42'04});
  1743. CHECK(FP::FPSR{(uint32_t)jit.GetRegister(0)}.QC() == 1);
  1744. CHECK(jit.GetVector(1) == Vector{0x3D74'6EEC'7FFF'4F42, 0x3F0F'1D7F'31B0'0616});
  1745. CHECK(FP::FPSR{(uint32_t)jit.GetRegister(1)}.QC() == 1);
  1746. CHECK(jit.GetVector(2) == Vector{0x218630B5'413E728F, 0x6FBDE982'7FFFFFFF});
  1747. CHECK(FP::FPSR{(uint32_t)jit.GetRegister(2)}.QC() == 1);
  1748. CHECK(jit.GetVector(3) == Vector{0x763E4B7043BC0AC5, 0x7FFFFFFFFFFFFFFF});
  1749. CHECK(FP::FPSR{(uint32_t)jit.GetRegister(3)}.QC() == 1);
  1750. CHECK(jit.GetVector(10) == Vector{0x2B'7F'14'2A'77'32'00'10, 0x63'16'7E'45'7F'33'42'04});
  1751. CHECK(FP::FPSR{(uint32_t)jit.GetRegister(10)}.QC() == 0);
  1752. CHECK(jit.GetVector(11) == Vector{0x3D74'6EEC'0000'4F42, 0x3F0F'1D7F'31B0'0616});
  1753. CHECK(FP::FPSR{(uint32_t)jit.GetRegister(11)}.QC() == 0);
  1754. CHECK(jit.GetVector(12) == Vector{0x218630B5'413E728F, 0x6FBDE982'00000000});
  1755. CHECK(FP::FPSR{(uint32_t)jit.GetRegister(12)}.QC() == 0);
  1756. CHECK(jit.GetVector(13) == Vector{0x763E4B7043BC0AC5, 0x5FDD5D671D399E2});
  1757. CHECK(FP::FPSR{(uint32_t)jit.GetRegister(13)}.QC() == 0);
  1758. }