optimizer.cpp 33 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881
  1. // Copyright (c) 2016 Google Inc.
  2. //
  3. // Licensed under the Apache License, Version 2.0 (the "License");
  4. // you may not use this file except in compliance with the License.
  5. // You may obtain a copy of the License at
  6. //
  7. // http://www.apache.org/licenses/LICENSE-2.0
  8. //
  9. // Unless required by applicable law or agreed to in writing, software
  10. // distributed under the License is distributed on an "AS IS" BASIS,
  11. // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  12. // See the License for the specific language governing permissions and
  13. // limitations under the License.
  14. #include "spirv-tools/optimizer.hpp"
  15. #include <memory>
  16. #include <string>
  17. #include <unordered_map>
  18. #include <utility>
  19. #include <vector>
  20. #include <source/spirv_optimizer_options.h>
  21. #include "source/opt/build_module.h"
  22. #include "source/opt/log.h"
  23. #include "source/opt/pass_manager.h"
  24. #include "source/opt/passes.h"
  25. #include "source/util/make_unique.h"
  26. #include "source/util/string_utils.h"
  27. namespace spvtools {
  28. struct Optimizer::PassToken::Impl {
  29. Impl(std::unique_ptr<opt::Pass> p) : pass(std::move(p)) {}
  30. std::unique_ptr<opt::Pass> pass; // Internal implementation pass.
  31. };
  32. Optimizer::PassToken::PassToken(
  33. std::unique_ptr<Optimizer::PassToken::Impl> impl)
  34. : impl_(std::move(impl)) {}
  35. Optimizer::PassToken::PassToken(std::unique_ptr<opt::Pass>&& pass)
  36. : impl_(MakeUnique<Optimizer::PassToken::Impl>(std::move(pass))) {}
  37. Optimizer::PassToken::PassToken(PassToken&& that)
  38. : impl_(std::move(that.impl_)) {}
  39. Optimizer::PassToken& Optimizer::PassToken::operator=(PassToken&& that) {
  40. impl_ = std::move(that.impl_);
  41. return *this;
  42. }
  43. Optimizer::PassToken::~PassToken() {}
  44. struct Optimizer::Impl {
  45. explicit Impl(spv_target_env env) : target_env(env), pass_manager() {}
  46. spv_target_env target_env; // Target environment.
  47. opt::PassManager pass_manager; // Internal implementation pass manager.
  48. };
  49. Optimizer::Optimizer(spv_target_env env) : impl_(new Impl(env)) {}
  50. Optimizer::~Optimizer() {}
  51. void Optimizer::SetMessageConsumer(MessageConsumer c) {
  52. // All passes' message consumer needs to be updated.
  53. for (uint32_t i = 0; i < impl_->pass_manager.NumPasses(); ++i) {
  54. impl_->pass_manager.GetPass(i)->SetMessageConsumer(c);
  55. }
  56. impl_->pass_manager.SetMessageConsumer(std::move(c));
  57. }
  58. const MessageConsumer& Optimizer::consumer() const {
  59. return impl_->pass_manager.consumer();
  60. }
  61. Optimizer& Optimizer::RegisterPass(PassToken&& p) {
  62. // Change to use the pass manager's consumer.
  63. p.impl_->pass->SetMessageConsumer(consumer());
  64. impl_->pass_manager.AddPass(std::move(p.impl_->pass));
  65. return *this;
  66. }
  67. // The legalization passes take a spir-v shader generated by an HLSL front-end
  68. // and turn it into a valid vulkan spir-v shader. There are two ways in which
  69. // the code will be invalid at the start:
  70. //
  71. // 1) There will be opaque objects, like images, which will be passed around
  72. // in intermediate objects. Valid spir-v will have to replace the use of
  73. // the opaque object with an intermediate object that is the result of the
  74. // load of the global opaque object.
  75. //
  76. // 2) There will be variables that contain pointers to structured or uniform
  77. // buffers. It be legal, the variables must be eliminated, and the
  78. // references to the structured buffers must use the result of OpVariable
  79. // in the Uniform storage class.
  80. //
  81. // Optimization in this list must accept shaders with these relaxation of the
  82. // rules. There is not guarantee that this list of optimizations is able to
  83. // legalize all inputs, but it is on a best effort basis.
  84. //
  85. // The legalization problem is essentially a very general copy propagation
  86. // problem. The optimization we use are all used to either do copy propagation
  87. // or enable more copy propagation.
  88. Optimizer& Optimizer::RegisterLegalizationPasses() {
  89. return
  90. // Remove unreachable block so that merge return works.
  91. RegisterPass(CreateDeadBranchElimPass())
  92. // Merge the returns so we can inline.
  93. .RegisterPass(CreateMergeReturnPass())
  94. // Make sure uses and definitions are in the same function.
  95. .RegisterPass(CreateInlineExhaustivePass())
  96. // Make private variable function scope
  97. .RegisterPass(CreateEliminateDeadFunctionsPass())
  98. .RegisterPass(CreatePrivateToLocalPass())
  99. // Fix up the storage classes that DXC may have purposely generated
  100. // incorrectly. All functions are inlined, and a lot of dead code has
  101. // been removed.
  102. .RegisterPass(CreateFixStorageClassPass())
  103. // Propagate the value stored to the loads in very simple cases.
  104. .RegisterPass(CreateLocalSingleBlockLoadStoreElimPass())
  105. .RegisterPass(CreateLocalSingleStoreElimPass())
  106. .RegisterPass(CreateAggressiveDCEPass())
  107. // Split up aggregates so they are easier to deal with.
  108. .RegisterPass(CreateScalarReplacementPass(0))
  109. // Remove loads and stores so everything is in intermediate values.
  110. // Takes care of copy propagation of non-members.
  111. .RegisterPass(CreateLocalSingleBlockLoadStoreElimPass())
  112. .RegisterPass(CreateLocalSingleStoreElimPass())
  113. .RegisterPass(CreateAggressiveDCEPass())
  114. .RegisterPass(CreateLocalMultiStoreElimPass())
  115. .RegisterPass(CreateAggressiveDCEPass())
  116. // Propagate constants to get as many constant conditions on branches
  117. // as possible.
  118. .RegisterPass(CreateCCPPass())
  119. .RegisterPass(CreateLoopUnrollPass(true))
  120. .RegisterPass(CreateDeadBranchElimPass())
  121. // Copy propagate members. Cleans up code sequences generated by
  122. // scalar replacement. Also important for removing OpPhi nodes.
  123. .RegisterPass(CreateSimplificationPass())
  124. .RegisterPass(CreateAggressiveDCEPass())
  125. .RegisterPass(CreateCopyPropagateArraysPass())
  126. // May need loop unrolling here see
  127. // https://github.com/Microsoft/DirectXShaderCompiler/pull/930
  128. // Get rid of unused code that contain traces of illegal code
  129. // or unused references to unbound external objects
  130. .RegisterPass(CreateVectorDCEPass())
  131. .RegisterPass(CreateDeadInsertElimPass())
  132. .RegisterPass(CreateReduceLoadSizePass())
  133. .RegisterPass(CreateAggressiveDCEPass());
  134. }
  135. Optimizer& Optimizer::RegisterPerformancePasses() {
  136. return RegisterPass(CreateDeadBranchElimPass())
  137. .RegisterPass(CreateMergeReturnPass())
  138. .RegisterPass(CreateInlineExhaustivePass())
  139. .RegisterPass(CreateAggressiveDCEPass())
  140. .RegisterPass(CreatePrivateToLocalPass())
  141. .RegisterPass(CreateLocalSingleBlockLoadStoreElimPass())
  142. .RegisterPass(CreateLocalSingleStoreElimPass())
  143. .RegisterPass(CreateAggressiveDCEPass())
  144. .RegisterPass(CreateScalarReplacementPass())
  145. .RegisterPass(CreateLocalAccessChainConvertPass())
  146. .RegisterPass(CreateLocalSingleBlockLoadStoreElimPass())
  147. .RegisterPass(CreateLocalSingleStoreElimPass())
  148. .RegisterPass(CreateAggressiveDCEPass())
  149. .RegisterPass(CreateLocalMultiStoreElimPass())
  150. .RegisterPass(CreateAggressiveDCEPass())
  151. .RegisterPass(CreateCCPPass())
  152. .RegisterPass(CreateAggressiveDCEPass())
  153. .RegisterPass(CreateRedundancyEliminationPass())
  154. .RegisterPass(CreateCombineAccessChainsPass())
  155. .RegisterPass(CreateSimplificationPass())
  156. .RegisterPass(CreateVectorDCEPass())
  157. .RegisterPass(CreateDeadInsertElimPass())
  158. .RegisterPass(CreateDeadBranchElimPass())
  159. .RegisterPass(CreateSimplificationPass())
  160. .RegisterPass(CreateIfConversionPass())
  161. .RegisterPass(CreateCopyPropagateArraysPass())
  162. .RegisterPass(CreateReduceLoadSizePass())
  163. .RegisterPass(CreateAggressiveDCEPass())
  164. .RegisterPass(CreateBlockMergePass())
  165. .RegisterPass(CreateRedundancyEliminationPass())
  166. .RegisterPass(CreateDeadBranchElimPass())
  167. .RegisterPass(CreateBlockMergePass())
  168. .RegisterPass(CreateSimplificationPass());
  169. }
  170. Optimizer& Optimizer::RegisterSizePasses() {
  171. return RegisterPass(CreateDeadBranchElimPass())
  172. .RegisterPass(CreateMergeReturnPass())
  173. .RegisterPass(CreateInlineExhaustivePass())
  174. .RegisterPass(CreateAggressiveDCEPass())
  175. .RegisterPass(CreatePrivateToLocalPass())
  176. .RegisterPass(CreateScalarReplacementPass())
  177. .RegisterPass(CreateLocalAccessChainConvertPass())
  178. .RegisterPass(CreateLocalSingleBlockLoadStoreElimPass())
  179. .RegisterPass(CreateLocalSingleStoreElimPass())
  180. .RegisterPass(CreateAggressiveDCEPass())
  181. .RegisterPass(CreateSimplificationPass())
  182. .RegisterPass(CreateDeadInsertElimPass())
  183. .RegisterPass(CreateLocalMultiStoreElimPass())
  184. .RegisterPass(CreateAggressiveDCEPass())
  185. .RegisterPass(CreateCCPPass())
  186. .RegisterPass(CreateAggressiveDCEPass())
  187. .RegisterPass(CreateDeadBranchElimPass())
  188. .RegisterPass(CreateIfConversionPass())
  189. .RegisterPass(CreateAggressiveDCEPass())
  190. .RegisterPass(CreateBlockMergePass())
  191. .RegisterPass(CreateSimplificationPass())
  192. .RegisterPass(CreateDeadInsertElimPass())
  193. .RegisterPass(CreateRedundancyEliminationPass())
  194. .RegisterPass(CreateCFGCleanupPass())
  195. .RegisterPass(CreateAggressiveDCEPass());
  196. }
  197. Optimizer& Optimizer::RegisterVulkanToWebGPUPasses() {
  198. return RegisterPass(CreateStripDebugInfoPass())
  199. .RegisterPass(CreateStripAtomicCounterMemoryPass())
  200. .RegisterPass(CreateGenerateWebGPUInitializersPass())
  201. .RegisterPass(CreateLegalizeVectorShufflePass())
  202. .RegisterPass(CreateSplitInvalidUnreachablePass())
  203. .RegisterPass(CreateEliminateDeadConstantPass())
  204. .RegisterPass(CreateFlattenDecorationPass())
  205. .RegisterPass(CreateAggressiveDCEPass())
  206. .RegisterPass(CreateDeadBranchElimPass())
  207. .RegisterPass(CreateCompactIdsPass());
  208. }
  209. Optimizer& Optimizer::RegisterWebGPUToVulkanPasses() {
  210. return RegisterPass(CreateDecomposeInitializedVariablesPass())
  211. .RegisterPass(CreateCompactIdsPass());
  212. }
  213. bool Optimizer::RegisterPassesFromFlags(const std::vector<std::string>& flags) {
  214. for (const auto& flag : flags) {
  215. if (!RegisterPassFromFlag(flag)) {
  216. return false;
  217. }
  218. }
  219. return true;
  220. }
  221. bool Optimizer::FlagHasValidForm(const std::string& flag) const {
  222. if (flag == "-O" || flag == "-Os") {
  223. return true;
  224. } else if (flag.size() > 2 && flag.substr(0, 2) == "--") {
  225. return true;
  226. }
  227. Errorf(consumer(), nullptr, {},
  228. "%s is not a valid flag. Flag passes should have the form "
  229. "'--pass_name[=pass_args]'. Special flag names also accepted: -O "
  230. "and -Os.",
  231. flag.c_str());
  232. return false;
  233. }
  234. bool Optimizer::RegisterPassFromFlag(const std::string& flag) {
  235. if (!FlagHasValidForm(flag)) {
  236. return false;
  237. }
  238. // Split flags of the form --pass_name=pass_args.
  239. auto p = utils::SplitFlagArgs(flag);
  240. std::string pass_name = p.first;
  241. std::string pass_args = p.second;
  242. // FIXME(dnovillo): This should be re-factored so that pass names can be
  243. // automatically checked against Pass::name() and PassToken instances created
  244. // via a template function. Additionally, class Pass should have a desc()
  245. // method that describes the pass (so it can be used in --help).
  246. //
  247. // Both Pass::name() and Pass::desc() should be static class members so they
  248. // can be invoked without creating a pass instance.
  249. if (pass_name == "strip-atomic-counter-memory") {
  250. RegisterPass(CreateStripAtomicCounterMemoryPass());
  251. } else if (pass_name == "strip-debug") {
  252. RegisterPass(CreateStripDebugInfoPass());
  253. } else if (pass_name == "strip-reflect") {
  254. RegisterPass(CreateStripReflectInfoPass());
  255. } else if (pass_name == "set-spec-const-default-value") {
  256. if (pass_args.size() > 0) {
  257. auto spec_ids_vals =
  258. opt::SetSpecConstantDefaultValuePass::ParseDefaultValuesString(
  259. pass_args.c_str());
  260. if (!spec_ids_vals) {
  261. Errorf(consumer(), nullptr, {},
  262. "Invalid argument for --set-spec-const-default-value: %s",
  263. pass_args.c_str());
  264. return false;
  265. }
  266. RegisterPass(
  267. CreateSetSpecConstantDefaultValuePass(std::move(*spec_ids_vals)));
  268. } else {
  269. Errorf(consumer(), nullptr, {},
  270. "Invalid spec constant value string '%s'. Expected a string of "
  271. "<spec id>:<default value> pairs.",
  272. pass_args.c_str());
  273. return false;
  274. }
  275. } else if (pass_name == "if-conversion") {
  276. RegisterPass(CreateIfConversionPass());
  277. } else if (pass_name == "freeze-spec-const") {
  278. RegisterPass(CreateFreezeSpecConstantValuePass());
  279. } else if (pass_name == "inline-entry-points-exhaustive") {
  280. RegisterPass(CreateInlineExhaustivePass());
  281. } else if (pass_name == "inline-entry-points-opaque") {
  282. RegisterPass(CreateInlineOpaquePass());
  283. } else if (pass_name == "combine-access-chains") {
  284. RegisterPass(CreateCombineAccessChainsPass());
  285. } else if (pass_name == "convert-local-access-chains") {
  286. RegisterPass(CreateLocalAccessChainConvertPass());
  287. } else if (pass_name == "eliminate-dead-code-aggressive") {
  288. RegisterPass(CreateAggressiveDCEPass());
  289. } else if (pass_name == "propagate-line-info") {
  290. RegisterPass(CreatePropagateLineInfoPass());
  291. } else if (pass_name == "eliminate-redundant-line-info") {
  292. RegisterPass(CreateRedundantLineInfoElimPass());
  293. } else if (pass_name == "eliminate-insert-extract") {
  294. RegisterPass(CreateInsertExtractElimPass());
  295. } else if (pass_name == "eliminate-local-single-block") {
  296. RegisterPass(CreateLocalSingleBlockLoadStoreElimPass());
  297. } else if (pass_name == "eliminate-local-single-store") {
  298. RegisterPass(CreateLocalSingleStoreElimPass());
  299. } else if (pass_name == "merge-blocks") {
  300. RegisterPass(CreateBlockMergePass());
  301. } else if (pass_name == "merge-return") {
  302. RegisterPass(CreateMergeReturnPass());
  303. } else if (pass_name == "eliminate-dead-branches") {
  304. RegisterPass(CreateDeadBranchElimPass());
  305. } else if (pass_name == "eliminate-dead-functions") {
  306. RegisterPass(CreateEliminateDeadFunctionsPass());
  307. } else if (pass_name == "eliminate-local-multi-store") {
  308. RegisterPass(CreateLocalMultiStoreElimPass());
  309. } else if (pass_name == "eliminate-dead-const") {
  310. RegisterPass(CreateEliminateDeadConstantPass());
  311. } else if (pass_name == "eliminate-dead-inserts") {
  312. RegisterPass(CreateDeadInsertElimPass());
  313. } else if (pass_name == "eliminate-dead-variables") {
  314. RegisterPass(CreateDeadVariableEliminationPass());
  315. } else if (pass_name == "eliminate-dead-members") {
  316. RegisterPass(CreateEliminateDeadMembersPass());
  317. } else if (pass_name == "fold-spec-const-op-composite") {
  318. RegisterPass(CreateFoldSpecConstantOpAndCompositePass());
  319. } else if (pass_name == "loop-unswitch") {
  320. RegisterPass(CreateLoopUnswitchPass());
  321. } else if (pass_name == "scalar-replacement") {
  322. if (pass_args.size() == 0) {
  323. RegisterPass(CreateScalarReplacementPass());
  324. } else {
  325. int limit = -1;
  326. if (pass_args.find_first_not_of("0123456789") == std::string::npos) {
  327. limit = atoi(pass_args.c_str());
  328. }
  329. if (limit >= 0) {
  330. RegisterPass(CreateScalarReplacementPass(limit));
  331. } else {
  332. Error(consumer(), nullptr, {},
  333. "--scalar-replacement must have no arguments or a non-negative "
  334. "integer argument");
  335. return false;
  336. }
  337. }
  338. } else if (pass_name == "strength-reduction") {
  339. RegisterPass(CreateStrengthReductionPass());
  340. } else if (pass_name == "unify-const") {
  341. RegisterPass(CreateUnifyConstantPass());
  342. } else if (pass_name == "flatten-decorations") {
  343. RegisterPass(CreateFlattenDecorationPass());
  344. } else if (pass_name == "compact-ids") {
  345. RegisterPass(CreateCompactIdsPass());
  346. } else if (pass_name == "cfg-cleanup") {
  347. RegisterPass(CreateCFGCleanupPass());
  348. } else if (pass_name == "local-redundancy-elimination") {
  349. RegisterPass(CreateLocalRedundancyEliminationPass());
  350. } else if (pass_name == "loop-invariant-code-motion") {
  351. RegisterPass(CreateLoopInvariantCodeMotionPass());
  352. } else if (pass_name == "reduce-load-size") {
  353. RegisterPass(CreateReduceLoadSizePass());
  354. } else if (pass_name == "redundancy-elimination") {
  355. RegisterPass(CreateRedundancyEliminationPass());
  356. } else if (pass_name == "private-to-local") {
  357. RegisterPass(CreatePrivateToLocalPass());
  358. } else if (pass_name == "remove-duplicates") {
  359. RegisterPass(CreateRemoveDuplicatesPass());
  360. } else if (pass_name == "workaround-1209") {
  361. RegisterPass(CreateWorkaround1209Pass());
  362. } else if (pass_name == "replace-invalid-opcode") {
  363. RegisterPass(CreateReplaceInvalidOpcodePass());
  364. } else if (pass_name == "inst-bindless-check") {
  365. RegisterPass(CreateInstBindlessCheckPass(7, 23, true, true, 1));
  366. RegisterPass(CreateSimplificationPass());
  367. RegisterPass(CreateDeadBranchElimPass());
  368. RegisterPass(CreateBlockMergePass());
  369. RegisterPass(CreateAggressiveDCEPass());
  370. } else if (pass_name == "simplify-instructions") {
  371. RegisterPass(CreateSimplificationPass());
  372. } else if (pass_name == "ssa-rewrite") {
  373. RegisterPass(CreateSSARewritePass());
  374. } else if (pass_name == "copy-propagate-arrays") {
  375. RegisterPass(CreateCopyPropagateArraysPass());
  376. } else if (pass_name == "loop-fission") {
  377. int register_threshold_to_split =
  378. (pass_args.size() > 0) ? atoi(pass_args.c_str()) : -1;
  379. if (register_threshold_to_split > 0) {
  380. RegisterPass(CreateLoopFissionPass(
  381. static_cast<size_t>(register_threshold_to_split)));
  382. } else {
  383. Error(consumer(), nullptr, {},
  384. "--loop-fission must have a positive integer argument");
  385. return false;
  386. }
  387. } else if (pass_name == "loop-fusion") {
  388. int max_registers_per_loop =
  389. (pass_args.size() > 0) ? atoi(pass_args.c_str()) : -1;
  390. if (max_registers_per_loop > 0) {
  391. RegisterPass(
  392. CreateLoopFusionPass(static_cast<size_t>(max_registers_per_loop)));
  393. } else {
  394. Error(consumer(), nullptr, {},
  395. "--loop-fusion must have a positive integer argument");
  396. return false;
  397. }
  398. } else if (pass_name == "loop-unroll") {
  399. RegisterPass(CreateLoopUnrollPass(true));
  400. } else if (pass_name == "upgrade-memory-model") {
  401. RegisterPass(CreateUpgradeMemoryModelPass());
  402. } else if (pass_name == "vector-dce") {
  403. RegisterPass(CreateVectorDCEPass());
  404. } else if (pass_name == "loop-unroll-partial") {
  405. int factor = (pass_args.size() > 0) ? atoi(pass_args.c_str()) : 0;
  406. if (factor > 0) {
  407. RegisterPass(CreateLoopUnrollPass(false, factor));
  408. } else {
  409. Error(consumer(), nullptr, {},
  410. "--loop-unroll-partial must have a positive integer argument");
  411. return false;
  412. }
  413. } else if (pass_name == "loop-peeling") {
  414. RegisterPass(CreateLoopPeelingPass());
  415. } else if (pass_name == "loop-peeling-threshold") {
  416. int factor = (pass_args.size() > 0) ? atoi(pass_args.c_str()) : 0;
  417. if (factor > 0) {
  418. opt::LoopPeelingPass::SetLoopPeelingThreshold(factor);
  419. } else {
  420. Error(consumer(), nullptr, {},
  421. "--loop-peeling-threshold must have a positive integer argument");
  422. return false;
  423. }
  424. } else if (pass_name == "ccp") {
  425. RegisterPass(CreateCCPPass());
  426. } else if (pass_name == "code-sink") {
  427. RegisterPass(CreateCodeSinkingPass());
  428. } else if (pass_name == "fix-storage-class") {
  429. RegisterPass(CreateFixStorageClassPass());
  430. } else if (pass_name == "O") {
  431. RegisterPerformancePasses();
  432. } else if (pass_name == "Os") {
  433. RegisterSizePasses();
  434. } else if (pass_name == "legalize-hlsl") {
  435. RegisterLegalizationPasses();
  436. } else if (pass_name == "generate-webgpu-initializers") {
  437. RegisterPass(CreateGenerateWebGPUInitializersPass());
  438. } else if (pass_name == "legalize-vector-shuffle") {
  439. RegisterPass(CreateLegalizeVectorShufflePass());
  440. } else if (pass_name == "split-invalid-unreachable") {
  441. RegisterPass(CreateLegalizeVectorShufflePass());
  442. } else if (pass_name == "decompose-initialized-variables") {
  443. RegisterPass(CreateDecomposeInitializedVariablesPass());
  444. } else {
  445. Errorf(consumer(), nullptr, {},
  446. "Unknown flag '--%s'. Use --help for a list of valid flags",
  447. pass_name.c_str());
  448. return false;
  449. }
  450. return true;
  451. }
  452. void Optimizer::SetTargetEnv(const spv_target_env env) {
  453. impl_->target_env = env;
  454. }
  455. bool Optimizer::Run(const uint32_t* original_binary,
  456. const size_t original_binary_size,
  457. std::vector<uint32_t>* optimized_binary) const {
  458. return Run(original_binary, original_binary_size, optimized_binary,
  459. OptimizerOptions());
  460. }
  461. bool Optimizer::Run(const uint32_t* original_binary,
  462. const size_t original_binary_size,
  463. std::vector<uint32_t>* optimized_binary,
  464. const ValidatorOptions& validator_options,
  465. bool skip_validation) const {
  466. OptimizerOptions opt_options;
  467. opt_options.set_run_validator(!skip_validation);
  468. opt_options.set_validator_options(validator_options);
  469. return Run(original_binary, original_binary_size, optimized_binary,
  470. opt_options);
  471. }
  472. bool Optimizer::Run(const uint32_t* original_binary,
  473. const size_t original_binary_size,
  474. std::vector<uint32_t>* optimized_binary,
  475. const spv_optimizer_options opt_options) const {
  476. spvtools::SpirvTools tools(impl_->target_env);
  477. tools.SetMessageConsumer(impl_->pass_manager.consumer());
  478. if (opt_options->run_validator_ &&
  479. !tools.Validate(original_binary, original_binary_size,
  480. &opt_options->val_options_)) {
  481. return false;
  482. }
  483. std::unique_ptr<opt::IRContext> context = BuildModule(
  484. impl_->target_env, consumer(), original_binary, original_binary_size);
  485. if (context == nullptr) return false;
  486. context->set_max_id_bound(opt_options->max_id_bound_);
  487. context->set_preserve_bindings(opt_options->preserve_bindings_);
  488. context->set_preserve_spec_constants(opt_options->preserve_spec_constants_);
  489. impl_->pass_manager.SetValidatorOptions(&opt_options->val_options_);
  490. impl_->pass_manager.SetTargetEnv(impl_->target_env);
  491. auto status = impl_->pass_manager.Run(context.get());
  492. bool binary_changed = false;
  493. if (status == opt::Pass::Status::SuccessWithChange) {
  494. binary_changed = true;
  495. } else if (status == opt::Pass::Status::SuccessWithoutChange) {
  496. if (optimized_binary->size() != original_binary_size ||
  497. (memcmp(optimized_binary->data(), original_binary,
  498. original_binary_size) != 0)) {
  499. binary_changed = true;
  500. Log(consumer(), SPV_MSG_WARNING, nullptr, {},
  501. "Binary unexpectedly changed despite optimizer saying there was no "
  502. "change");
  503. }
  504. }
  505. if (binary_changed) {
  506. optimized_binary->clear();
  507. context->module()->ToBinary(optimized_binary, /* skip_nop = */ true);
  508. }
  509. return status != opt::Pass::Status::Failure;
  510. }
  511. Optimizer& Optimizer::SetPrintAll(std::ostream* out) {
  512. impl_->pass_manager.SetPrintAll(out);
  513. return *this;
  514. }
  515. Optimizer& Optimizer::SetTimeReport(std::ostream* out) {
  516. impl_->pass_manager.SetTimeReport(out);
  517. return *this;
  518. }
  519. Optimizer& Optimizer::SetValidateAfterAll(bool validate) {
  520. impl_->pass_manager.SetValidateAfterAll(validate);
  521. return *this;
  522. }
  523. Optimizer::PassToken CreateNullPass() {
  524. return MakeUnique<Optimizer::PassToken::Impl>(MakeUnique<opt::NullPass>());
  525. }
  526. Optimizer::PassToken CreateStripAtomicCounterMemoryPass() {
  527. return MakeUnique<Optimizer::PassToken::Impl>(
  528. MakeUnique<opt::StripAtomicCounterMemoryPass>());
  529. }
  530. Optimizer::PassToken CreateStripDebugInfoPass() {
  531. return MakeUnique<Optimizer::PassToken::Impl>(
  532. MakeUnique<opt::StripDebugInfoPass>());
  533. }
  534. Optimizer::PassToken CreateStripReflectInfoPass() {
  535. return MakeUnique<Optimizer::PassToken::Impl>(
  536. MakeUnique<opt::StripReflectInfoPass>());
  537. }
  538. Optimizer::PassToken CreateEliminateDeadFunctionsPass() {
  539. return MakeUnique<Optimizer::PassToken::Impl>(
  540. MakeUnique<opt::EliminateDeadFunctionsPass>());
  541. }
  542. Optimizer::PassToken CreateEliminateDeadMembersPass() {
  543. return MakeUnique<Optimizer::PassToken::Impl>(
  544. MakeUnique<opt::EliminateDeadMembersPass>());
  545. }
  546. Optimizer::PassToken CreateSetSpecConstantDefaultValuePass(
  547. const std::unordered_map<uint32_t, std::string>& id_value_map) {
  548. return MakeUnique<Optimizer::PassToken::Impl>(
  549. MakeUnique<opt::SetSpecConstantDefaultValuePass>(id_value_map));
  550. }
  551. Optimizer::PassToken CreateSetSpecConstantDefaultValuePass(
  552. const std::unordered_map<uint32_t, std::vector<uint32_t>>& id_value_map) {
  553. return MakeUnique<Optimizer::PassToken::Impl>(
  554. MakeUnique<opt::SetSpecConstantDefaultValuePass>(id_value_map));
  555. }
  556. Optimizer::PassToken CreateFlattenDecorationPass() {
  557. return MakeUnique<Optimizer::PassToken::Impl>(
  558. MakeUnique<opt::FlattenDecorationPass>());
  559. }
  560. Optimizer::PassToken CreateFreezeSpecConstantValuePass() {
  561. return MakeUnique<Optimizer::PassToken::Impl>(
  562. MakeUnique<opt::FreezeSpecConstantValuePass>());
  563. }
  564. Optimizer::PassToken CreateFoldSpecConstantOpAndCompositePass() {
  565. return MakeUnique<Optimizer::PassToken::Impl>(
  566. MakeUnique<opt::FoldSpecConstantOpAndCompositePass>());
  567. }
  568. Optimizer::PassToken CreateUnifyConstantPass() {
  569. return MakeUnique<Optimizer::PassToken::Impl>(
  570. MakeUnique<opt::UnifyConstantPass>());
  571. }
  572. Optimizer::PassToken CreateEliminateDeadConstantPass() {
  573. return MakeUnique<Optimizer::PassToken::Impl>(
  574. MakeUnique<opt::EliminateDeadConstantPass>());
  575. }
  576. Optimizer::PassToken CreateDeadVariableEliminationPass() {
  577. return MakeUnique<Optimizer::PassToken::Impl>(
  578. MakeUnique<opt::DeadVariableElimination>());
  579. }
  580. Optimizer::PassToken CreateStrengthReductionPass() {
  581. return MakeUnique<Optimizer::PassToken::Impl>(
  582. MakeUnique<opt::StrengthReductionPass>());
  583. }
  584. Optimizer::PassToken CreateBlockMergePass() {
  585. return MakeUnique<Optimizer::PassToken::Impl>(
  586. MakeUnique<opt::BlockMergePass>());
  587. }
  588. Optimizer::PassToken CreateInlineExhaustivePass() {
  589. return MakeUnique<Optimizer::PassToken::Impl>(
  590. MakeUnique<opt::InlineExhaustivePass>());
  591. }
  592. Optimizer::PassToken CreateInlineOpaquePass() {
  593. return MakeUnique<Optimizer::PassToken::Impl>(
  594. MakeUnique<opt::InlineOpaquePass>());
  595. }
  596. Optimizer::PassToken CreateLocalAccessChainConvertPass() {
  597. return MakeUnique<Optimizer::PassToken::Impl>(
  598. MakeUnique<opt::LocalAccessChainConvertPass>());
  599. }
  600. Optimizer::PassToken CreateLocalSingleBlockLoadStoreElimPass() {
  601. return MakeUnique<Optimizer::PassToken::Impl>(
  602. MakeUnique<opt::LocalSingleBlockLoadStoreElimPass>());
  603. }
  604. Optimizer::PassToken CreateLocalSingleStoreElimPass() {
  605. return MakeUnique<Optimizer::PassToken::Impl>(
  606. MakeUnique<opt::LocalSingleStoreElimPass>());
  607. }
  608. Optimizer::PassToken CreateInsertExtractElimPass() {
  609. return MakeUnique<Optimizer::PassToken::Impl>(
  610. MakeUnique<opt::SimplificationPass>());
  611. }
  612. Optimizer::PassToken CreateDeadInsertElimPass() {
  613. return MakeUnique<Optimizer::PassToken::Impl>(
  614. MakeUnique<opt::DeadInsertElimPass>());
  615. }
  616. Optimizer::PassToken CreateDeadBranchElimPass() {
  617. return MakeUnique<Optimizer::PassToken::Impl>(
  618. MakeUnique<opt::DeadBranchElimPass>());
  619. }
  620. Optimizer::PassToken CreateLocalMultiStoreElimPass() {
  621. return MakeUnique<Optimizer::PassToken::Impl>(
  622. MakeUnique<opt::LocalMultiStoreElimPass>());
  623. }
  624. Optimizer::PassToken CreateAggressiveDCEPass() {
  625. return MakeUnique<Optimizer::PassToken::Impl>(
  626. MakeUnique<opt::AggressiveDCEPass>());
  627. }
  628. Optimizer::PassToken CreatePropagateLineInfoPass() {
  629. return MakeUnique<Optimizer::PassToken::Impl>(
  630. MakeUnique<opt::ProcessLinesPass>(opt::kLinesPropagateLines));
  631. }
  632. Optimizer::PassToken CreateRedundantLineInfoElimPass() {
  633. return MakeUnique<Optimizer::PassToken::Impl>(
  634. MakeUnique<opt::ProcessLinesPass>(opt::kLinesEliminateDeadLines));
  635. }
  636. Optimizer::PassToken CreateCompactIdsPass() {
  637. return MakeUnique<Optimizer::PassToken::Impl>(
  638. MakeUnique<opt::CompactIdsPass>());
  639. }
  640. Optimizer::PassToken CreateMergeReturnPass() {
  641. return MakeUnique<Optimizer::PassToken::Impl>(
  642. MakeUnique<opt::MergeReturnPass>());
  643. }
  644. std::vector<const char*> Optimizer::GetPassNames() const {
  645. std::vector<const char*> v;
  646. for (uint32_t i = 0; i < impl_->pass_manager.NumPasses(); i++) {
  647. v.push_back(impl_->pass_manager.GetPass(i)->name());
  648. }
  649. return v;
  650. }
  651. Optimizer::PassToken CreateCFGCleanupPass() {
  652. return MakeUnique<Optimizer::PassToken::Impl>(
  653. MakeUnique<opt::CFGCleanupPass>());
  654. }
  655. Optimizer::PassToken CreateLocalRedundancyEliminationPass() {
  656. return MakeUnique<Optimizer::PassToken::Impl>(
  657. MakeUnique<opt::LocalRedundancyEliminationPass>());
  658. }
  659. Optimizer::PassToken CreateLoopFissionPass(size_t threshold) {
  660. return MakeUnique<Optimizer::PassToken::Impl>(
  661. MakeUnique<opt::LoopFissionPass>(threshold));
  662. }
  663. Optimizer::PassToken CreateLoopFusionPass(size_t max_registers_per_loop) {
  664. return MakeUnique<Optimizer::PassToken::Impl>(
  665. MakeUnique<opt::LoopFusionPass>(max_registers_per_loop));
  666. }
  667. Optimizer::PassToken CreateLoopInvariantCodeMotionPass() {
  668. return MakeUnique<Optimizer::PassToken::Impl>(MakeUnique<opt::LICMPass>());
  669. }
  670. Optimizer::PassToken CreateLoopPeelingPass() {
  671. return MakeUnique<Optimizer::PassToken::Impl>(
  672. MakeUnique<opt::LoopPeelingPass>());
  673. }
  674. Optimizer::PassToken CreateLoopUnswitchPass() {
  675. return MakeUnique<Optimizer::PassToken::Impl>(
  676. MakeUnique<opt::LoopUnswitchPass>());
  677. }
  678. Optimizer::PassToken CreateRedundancyEliminationPass() {
  679. return MakeUnique<Optimizer::PassToken::Impl>(
  680. MakeUnique<opt::RedundancyEliminationPass>());
  681. }
  682. Optimizer::PassToken CreateRemoveDuplicatesPass() {
  683. return MakeUnique<Optimizer::PassToken::Impl>(
  684. MakeUnique<opt::RemoveDuplicatesPass>());
  685. }
  686. Optimizer::PassToken CreateScalarReplacementPass(uint32_t size_limit) {
  687. return MakeUnique<Optimizer::PassToken::Impl>(
  688. MakeUnique<opt::ScalarReplacementPass>(size_limit));
  689. }
  690. Optimizer::PassToken CreatePrivateToLocalPass() {
  691. return MakeUnique<Optimizer::PassToken::Impl>(
  692. MakeUnique<opt::PrivateToLocalPass>());
  693. }
  694. Optimizer::PassToken CreateCCPPass() {
  695. return MakeUnique<Optimizer::PassToken::Impl>(MakeUnique<opt::CCPPass>());
  696. }
  697. Optimizer::PassToken CreateWorkaround1209Pass() {
  698. return MakeUnique<Optimizer::PassToken::Impl>(
  699. MakeUnique<opt::Workaround1209>());
  700. }
  701. Optimizer::PassToken CreateIfConversionPass() {
  702. return MakeUnique<Optimizer::PassToken::Impl>(
  703. MakeUnique<opt::IfConversion>());
  704. }
  705. Optimizer::PassToken CreateReplaceInvalidOpcodePass() {
  706. return MakeUnique<Optimizer::PassToken::Impl>(
  707. MakeUnique<opt::ReplaceInvalidOpcodePass>());
  708. }
  709. Optimizer::PassToken CreateSimplificationPass() {
  710. return MakeUnique<Optimizer::PassToken::Impl>(
  711. MakeUnique<opt::SimplificationPass>());
  712. }
  713. Optimizer::PassToken CreateLoopUnrollPass(bool fully_unroll, int factor) {
  714. return MakeUnique<Optimizer::PassToken::Impl>(
  715. MakeUnique<opt::LoopUnroller>(fully_unroll, factor));
  716. }
  717. Optimizer::PassToken CreateSSARewritePass() {
  718. return MakeUnique<Optimizer::PassToken::Impl>(
  719. MakeUnique<opt::SSARewritePass>());
  720. }
  721. Optimizer::PassToken CreateCopyPropagateArraysPass() {
  722. return MakeUnique<Optimizer::PassToken::Impl>(
  723. MakeUnique<opt::CopyPropagateArrays>());
  724. }
  725. Optimizer::PassToken CreateVectorDCEPass() {
  726. return MakeUnique<Optimizer::PassToken::Impl>(MakeUnique<opt::VectorDCE>());
  727. }
  728. Optimizer::PassToken CreateReduceLoadSizePass() {
  729. return MakeUnique<Optimizer::PassToken::Impl>(
  730. MakeUnique<opt::ReduceLoadSize>());
  731. }
  732. Optimizer::PassToken CreateCombineAccessChainsPass() {
  733. return MakeUnique<Optimizer::PassToken::Impl>(
  734. MakeUnique<opt::CombineAccessChains>());
  735. }
  736. Optimizer::PassToken CreateUpgradeMemoryModelPass() {
  737. return MakeUnique<Optimizer::PassToken::Impl>(
  738. MakeUnique<opt::UpgradeMemoryModel>());
  739. }
  740. Optimizer::PassToken CreateInstBindlessCheckPass(uint32_t desc_set,
  741. uint32_t shader_id,
  742. bool input_length_enable,
  743. bool input_init_enable,
  744. uint32_t version) {
  745. return MakeUnique<Optimizer::PassToken::Impl>(
  746. MakeUnique<opt::InstBindlessCheckPass>(desc_set, shader_id,
  747. input_length_enable,
  748. input_init_enable, version));
  749. }
  750. Optimizer::PassToken CreateCodeSinkingPass() {
  751. return MakeUnique<Optimizer::PassToken::Impl>(
  752. MakeUnique<opt::CodeSinkingPass>());
  753. }
  754. Optimizer::PassToken CreateGenerateWebGPUInitializersPass() {
  755. return MakeUnique<Optimizer::PassToken::Impl>(
  756. MakeUnique<opt::GenerateWebGPUInitializersPass>());
  757. }
  758. Optimizer::PassToken CreateFixStorageClassPass() {
  759. return MakeUnique<Optimizer::PassToken::Impl>(
  760. MakeUnique<opt::FixStorageClass>());
  761. }
  762. Optimizer::PassToken CreateLegalizeVectorShufflePass() {
  763. return MakeUnique<Optimizer::PassToken::Impl>(
  764. MakeUnique<opt::LegalizeVectorShufflePass>());
  765. }
  766. Optimizer::PassToken CreateDecomposeInitializedVariablesPass() {
  767. return MakeUnique<Optimizer::PassToken::Impl>(
  768. MakeUnique<opt::DecomposeInitializedVariablesPass>());
  769. }
  770. Optimizer::PassToken CreateSplitInvalidUnreachablePass() {
  771. return MakeUnique<Optimizer::PassToken::Impl>(
  772. MakeUnique<opt::SplitInvalidUnreachablePass>());
  773. }
  774. } // namespace spvtools