lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

   1 //===-- AMDGPUTargetMachine.cpp - TargetMachine for hw codegen targets-----===//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 //
   9 /// \file
  10 /// The AMDGPU target machine contains all of the hardware specific
  11 /// information  needed to emit code for R600 and SI GPUs.
  12 //
  13 //===----------------------------------------------------------------------===//
  14
  15 #include "AMDGPUTargetMachine.h"
  16 #include "AMDGPU.h"
  17 #include "AMDGPUAliasAnalysis.h"
  18 #include "AMDGPUCallLowering.h"
  19 #include "AMDGPUInstructionSelector.h"
  20 #include "AMDGPULegalizerInfo.h"
  21 #include "AMDGPUMacroFusion.h"
  22 #include "AMDGPUTargetObjectFile.h"
  23 #include "AMDGPUTargetTransformInfo.h"
  24 #include "GCNIterativeScheduler.h"
  25 #include "GCNSchedStrategy.h"
  26 #include "R600MachineScheduler.h"
  27 #include "SIMachineFunctionInfo.h"
  28 #include "SIMachineScheduler.h"
  29 #include "TargetInfo/AMDGPUTargetInfo.h"
  30 #include "llvm/CodeGen/GlobalISel/IRTranslator.h"
  31 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h"
  32 #include "llvm/CodeGen/GlobalISel/Legalizer.h"
  33 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h"
  34 #include "llvm/CodeGen/MIRParser/MIParser.h"
  35 #include "llvm/CodeGen/Passes.h"
  36 #include "llvm/CodeGen/TargetPassConfig.h"
  37 #include "llvm/IR/Attributes.h"
  38 #include "llvm/IR/Function.h"
  39 #include "llvm/IR/LegacyPassManager.h"
  40 #include "llvm/Pass.h"
  41 #include "llvm/Support/CommandLine.h"
  42 #include "llvm/Support/Compiler.h"
  43 #include "llvm/Support/TargetRegistry.h"
  44 #include "llvm/Target/TargetLoweringObjectFile.h"
  45 #include "llvm/Transforms/IPO.h"
  46 #include "llvm/Transforms/IPO/AlwaysInliner.h"
  47 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
  48 #include "llvm/Transforms/Scalar.h"
  49 #include "llvm/Transforms/Scalar/GVN.h"
  50 #include "llvm/Transforms/Utils.h"
  51 #include "llvm/Transforms/Vectorize.h"
  52 #include <memory>
  53
  54 using namespace llvm;
  55
  56 static cl::opt<bool> EnableR600StructurizeCFG(
  57   "r600-ir-structurize",
  58   cl::desc("Use StructurizeCFG IR pass"),
  59   cl::init(true));
  60
  61 static cl::opt<bool> EnableSROA(
  62   "amdgpu-sroa",
  63   cl::desc("Run SROA after promote alloca pass"),
  64   cl::ReallyHidden,
  65   cl::init(true));
  66
  67 static cl::opt<bool>
  68 EnableEarlyIfConversion("amdgpu-early-ifcvt", cl::Hidden,
  69                         cl::desc("Run early if-conversion"),
  70                         cl::init(false));
  71
  72 static cl::opt<bool>
  73 OptExecMaskPreRA("amdgpu-opt-exec-mask-pre-ra", cl::Hidden,
  74             cl::desc("Run pre-RA exec mask optimizations"),
  75             cl::init(true));
  76
  77 static cl::opt<bool> EnableR600IfConvert(
  78   "r600-if-convert",
  79   cl::desc("Use if conversion pass"),
  80   cl::ReallyHidden,
  81   cl::init(true));
  82
  83 // Option to disable vectorizer for tests.
  84 static cl::opt<bool> EnableLoadStoreVectorizer(
  85   "amdgpu-load-store-vectorizer",
  86   cl::desc("Enable load store vectorizer"),
  87   cl::init(true),
  88   cl::Hidden);
  89
  90 // Option to control global loads scalarization
  91 static cl::opt<bool> ScalarizeGlobal(
  92   "amdgpu-scalarize-global-loads",
  93   cl::desc("Enable global load scalarization"),
  94   cl::init(true),
  95   cl::Hidden);
  96
  97 // Option to run internalize pass.
  98 static cl::opt<bool> InternalizeSymbols(
  99   "amdgpu-internalize-symbols",
 100   cl::desc("Enable elimination of non-kernel functions and unused globals"),
 101   cl::init(false),
 102   cl::Hidden);
 103
 104 // Option to inline all early.
 105 static cl::opt<bool> EarlyInlineAll(
 106   "amdgpu-early-inline-all",
 107   cl::desc("Inline all functions early"),
 108   cl::init(false),
 109   cl::Hidden);
 110
 111 static cl::opt<bool> EnableSDWAPeephole(
 112   "amdgpu-sdwa-peephole",
 113   cl::desc("Enable SDWA peepholer"),
 114   cl::init(true));
 115
 116 static cl::opt<bool> EnableDPPCombine(
 117   "amdgpu-dpp-combine",
 118   cl::desc("Enable DPP combiner"),
 119   cl::init(true));
 120
 121 // Enable address space based alias analysis
 122 static cl::opt<bool> EnableAMDGPUAliasAnalysis("enable-amdgpu-aa", cl::Hidden,
 123   cl::desc("Enable AMDGPU Alias Analysis"),
 124   cl::init(true));
 125
 126 // Option to run late CFG structurizer
 127 static cl::opt<bool, true> LateCFGStructurize(
 128   "amdgpu-late-structurize",
 129   cl::desc("Enable late CFG structurization"),
 130   cl::location(AMDGPUTargetMachine::EnableLateStructurizeCFG),
 131   cl::Hidden);
 132
 133 static cl::opt<bool, true> EnableAMDGPUFunctionCallsOpt(
 134   "amdgpu-function-calls",
 135   cl::desc("Enable AMDGPU function call support"),
 136   cl::location(AMDGPUTargetMachine::EnableFunctionCalls),
 137   cl::init(true),
 138   cl::Hidden);
 139
 140 // Enable lib calls simplifications
 141 static cl::opt<bool> EnableLibCallSimplify(
 142   "amdgpu-simplify-libcall",
 143   cl::desc("Enable amdgpu library simplifications"),
 144   cl::init(true),
 145   cl::Hidden);
 146
 147 static cl::opt<bool> EnableLowerKernelArguments(
 148   "amdgpu-ir-lower-kernel-arguments",
 149   cl::desc("Lower kernel argument loads in IR pass"),
 150   cl::init(true),
 151   cl::Hidden);
 152
 153 static cl::opt<bool> EnableRegReassign(
 154   "amdgpu-reassign-regs",
 155   cl::desc("Enable register reassign optimizations on gfx10+"),
 156   cl::init(true),
 157   cl::Hidden);
 158
 159 // Enable atomic optimization
 160 static cl::opt<bool> EnableAtomicOptimizations(
 161   "amdgpu-atomic-optimizations",
 162   cl::desc("Enable atomic optimizations"),
 163   cl::init(false),
 164   cl::Hidden);
 165
 166 // Enable Mode register optimization
 167 static cl::opt<bool> EnableSIModeRegisterPass(
 168   "amdgpu-mode-register",
 169   cl::desc("Enable mode register pass"),
 170   cl::init(true),
 171   cl::Hidden);
 172
 173 // Option is used in lit tests to prevent deadcoding of patterns inspected.
 174 static cl::opt<bool>
 175 EnableDCEInRA("amdgpu-dce-in-ra",
 176     cl::init(true), cl::Hidden,
 177     cl::desc("Enable machine DCE inside regalloc"));
 178
 179 static cl::opt<bool> EnableScalarIRPasses(
 180   "amdgpu-scalar-ir-passes",
 181   cl::desc("Enable scalar IR passes"),
 182   cl::init(true),
 183   cl::Hidden);
 184
 185 extern "C" void LLVMInitializeAMDGPUTarget() {
 186   // Register the target
 187   RegisterTargetMachine<R600TargetMachine> X(getTheAMDGPUTarget());
 188   RegisterTargetMachine<GCNTargetMachine> Y(getTheGCNTarget());
 189
 190   PassRegistry *PR = PassRegistry::getPassRegistry();
 191   initializeR600ClauseMergePassPass(*PR);
 192   initializeR600ControlFlowFinalizerPass(*PR);
 193   initializeR600PacketizerPass(*PR);
 194   initializeR600ExpandSpecialInstrsPassPass(*PR);
 195   initializeR600VectorRegMergerPass(*PR);
 196   initializeGlobalISel(*PR);
 197   initializeAMDGPUDAGToDAGISelPass(*PR);
 198   initializeGCNDPPCombinePass(*PR);
 199   initializeSILowerI1CopiesPass(*PR);
 200   initializeSILowerSGPRSpillsPass(*PR);
 201   initializeSIFixSGPRCopiesPass(*PR);
 202   initializeSIFixVGPRCopiesPass(*PR);
 203   initializeSIFixupVectorISelPass(*PR);
 204   initializeSIFoldOperandsPass(*PR);
 205   initializeSIPeepholeSDWAPass(*PR);
 206   initializeSIShrinkInstructionsPass(*PR);
 207   initializeSIOptimizeExecMaskingPreRAPass(*PR);
 208   initializeSILoadStoreOptimizerPass(*PR);
 209   initializeAMDGPUFixFunctionBitcastsPass(*PR);
 210   initializeAMDGPUAlwaysInlinePass(*PR);
 211   initializeAMDGPUAnnotateKernelFeaturesPass(*PR);
 212   initializeAMDGPUAnnotateUniformValuesPass(*PR);
 213   initializeAMDGPUArgumentUsageInfoPass(*PR);
 214   initializeAMDGPUAtomicOptimizerPass(*PR);
 215   initializeAMDGPULowerKernelArgumentsPass(*PR);
 216   initializeAMDGPULowerKernelAttributesPass(*PR);
 217   initializeAMDGPULowerIntrinsicsPass(*PR);
 218   initializeAMDGPUOpenCLEnqueuedBlockLoweringPass(*PR);
 219   initializeAMDGPUPromoteAllocaPass(*PR);
 220   initializeAMDGPUCodeGenPreparePass(*PR);
 221   initializeAMDGPUPropagateAttributesEarlyPass(*PR);
 222   initializeAMDGPUPropagateAttributesLatePass(*PR);
 223   initializeAMDGPURewriteOutArgumentsPass(*PR);
 224   initializeAMDGPUUnifyMetadataPass(*PR);
 225   initializeSIAnnotateControlFlowPass(*PR);
 226   initializeSIInsertWaitcntsPass(*PR);
 227   initializeSIModeRegisterPass(*PR);
 228   initializeSIWholeQuadModePass(*PR);
 229   initializeSILowerControlFlowPass(*PR);
 230   initializeSIInsertSkipsPass(*PR);
 231   initializeSIMemoryLegalizerPass(*PR);
 232   initializeSIOptimizeExecMaskingPass(*PR);
 233   initializeSIPreAllocateWWMRegsPass(*PR);
 234   initializeSIFormMemoryClausesPass(*PR);
 235   initializeAMDGPUUnifyDivergentExitNodesPass(*PR);
 236   initializeAMDGPUAAWrapperPassPass(*PR);
 237   initializeAMDGPUExternalAAWrapperPass(*PR);
 238   initializeAMDGPUUseNativeCallsPass(*PR);
 239   initializeAMDGPUSimplifyLibCallsPass(*PR);
 240   initializeAMDGPUInlinerPass(*PR);
 241   initializeGCNRegBankReassignPass(*PR);
 242   initializeGCNNSAReassignPass(*PR);
 243 }
 244
 245 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
 246   return llvm::make_unique<AMDGPUTargetObjectFile>();
 247 }
 248
 249 static ScheduleDAGInstrs *createR600MachineScheduler(MachineSchedContext *C) {
 250   return new ScheduleDAGMILive(C, llvm::make_unique<R600SchedStrategy>());
 251 }
 252
 253 static ScheduleDAGInstrs *createSIMachineScheduler(MachineSchedContext *C) {
 254   return new SIScheduleDAGMI(C);
 255 }
 256
 257 static ScheduleDAGInstrs *
 258 createGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
 259   ScheduleDAGMILive *DAG =
 260     new GCNScheduleDAGMILive(C, make_unique<GCNMaxOccupancySchedStrategy>(C));
 261   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
 262   DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
 263   DAG->addMutation(createAMDGPUMacroFusionDAGMutation());
 264   return DAG;
 265 }
 266
 267 static ScheduleDAGInstrs *
 268 createIterativeGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
 269   auto DAG = new GCNIterativeScheduler(C,
 270     GCNIterativeScheduler::SCHEDULE_LEGACYMAXOCCUPANCY);
 271   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
 272   DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
 273   return DAG;
 274 }
 275
 276 static ScheduleDAGInstrs *createMinRegScheduler(MachineSchedContext *C) {
 277   return new GCNIterativeScheduler(C,
 278     GCNIterativeScheduler::SCHEDULE_MINREGFORCED);
 279 }
 280
 281 static ScheduleDAGInstrs *
 282 createIterativeILPMachineScheduler(MachineSchedContext *C) {
 283   auto DAG = new GCNIterativeScheduler(C,
 284     GCNIterativeScheduler::SCHEDULE_ILP);
 285   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
 286   DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
 287   DAG->addMutation(createAMDGPUMacroFusionDAGMutation());
 288   return DAG;
 289 }
 290
 291 static MachineSchedRegistry
 292 R600SchedRegistry("r600", "Run R600's custom scheduler",
 293                    createR600MachineScheduler);
 294
 295 static MachineSchedRegistry
 296 SISchedRegistry("si", "Run SI's custom scheduler",
 297                 createSIMachineScheduler);
 298
 299 static MachineSchedRegistry
 300 GCNMaxOccupancySchedRegistry("gcn-max-occupancy",
 301                              "Run GCN scheduler to maximize occupancy",
 302                              createGCNMaxOccupancyMachineScheduler);
 303
 304 static MachineSchedRegistry
 305 IterativeGCNMaxOccupancySchedRegistry("gcn-max-occupancy-experimental",
 306   "Run GCN scheduler to maximize occupancy (experimental)",
 307   createIterativeGCNMaxOccupancyMachineScheduler);
 308
 309 static MachineSchedRegistry
 310 GCNMinRegSchedRegistry("gcn-minreg",
 311   "Run GCN iterative scheduler for minimal register usage (experimental)",
 312   createMinRegScheduler);
 313
 314 static MachineSchedRegistry
 315 GCNILPSchedRegistry("gcn-ilp",
 316   "Run GCN iterative scheduler for ILP scheduling (experimental)",
 317   createIterativeILPMachineScheduler);
 318
 319 static StringRef computeDataLayout(const Triple &TT) {
 320   if (TT.getArch() == Triple::r600) {
 321     // 32-bit pointers.
 322       return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
 323              "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5";
 324   }
 325
 326   // 32-bit private, local, and region pointers. 64-bit global, constant and
 327   // flat, non-integral buffer fat pointers.
 328     return "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32"
 329          "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
 330          "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5"
 331          "-ni:7";
 332 }
 333
 334 LLVM_READNONE
 335 static StringRef getGPUOrDefault(const Triple &TT, StringRef GPU) {
 336   if (!GPU.empty())
 337     return GPU;
 338
 339   // Need to default to a target with flat support for HSA.
 340   if (TT.getArch() == Triple::amdgcn)
 341     return TT.getOS() == Triple::AMDHSA ? "generic-hsa" : "generic";
 342
 343   return "r600";
 344 }
 345
 346 static Reloc::Model getEffectiveRelocModel(Optional<Reloc::Model> RM) {
 347   // The AMDGPU toolchain only supports generating shared objects, so we
 348   // must always use PIC.
 349   return Reloc::PIC_;
 350 }
 351
 352 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, const Triple &TT,
 353                                          StringRef CPU, StringRef FS,
 354                                          TargetOptions Options,
 355                                          Optional<Reloc::Model> RM,
 356                                          Optional<CodeModel::Model> CM,
 357                                          CodeGenOpt::Level OptLevel)
 358     : LLVMTargetMachine(T, computeDataLayout(TT), TT, getGPUOrDefault(TT, CPU),
 359                         FS, Options, getEffectiveRelocModel(RM),
 360                         getEffectiveCodeModel(CM, CodeModel::Small), OptLevel),
 361       TLOF(createTLOF(getTargetTriple())) {
 362   initAsmInfo();
 363 }
 364
 365 bool AMDGPUTargetMachine::EnableLateStructurizeCFG = false;
 366 bool AMDGPUTargetMachine::EnableFunctionCalls = false;
 367
 368 AMDGPUTargetMachine::~AMDGPUTargetMachine() = default;
 369
 370 StringRef AMDGPUTargetMachine::getGPUName(const Function &F) const {
 371   Attribute GPUAttr = F.getFnAttribute("target-cpu");
 372   return GPUAttr.hasAttribute(Attribute::None) ?
 373     getTargetCPU() : GPUAttr.getValueAsString();
 374 }
 375
 376 StringRef AMDGPUTargetMachine::getFeatureString(const Function &F) const {
 377   Attribute FSAttr = F.getFnAttribute("target-features");
 378
 379   return FSAttr.hasAttribute(Attribute::None) ?
 380     getTargetFeatureString() :
 381     FSAttr.getValueAsString();
 382 }
 383
 384 /// Predicate for Internalize pass.
 385 static bool mustPreserveGV(const GlobalValue &GV) {
 386   if (const Function *F = dyn_cast<Function>(&GV))
 387     return F->isDeclaration() || AMDGPU::isEntryFunctionCC(F->getCallingConv());
 388
 389   return !GV.use_empty();
 390 }
 391
 392 void AMDGPUTargetMachine::adjustPassManager(PassManagerBuilder &Builder) {
 393   Builder.DivergentTarget = true;
 394
 395   bool EnableOpt = getOptLevel() > CodeGenOpt::None;
 396   bool Internalize = InternalizeSymbols;
 397   bool EarlyInline = EarlyInlineAll && EnableOpt && !EnableFunctionCalls;
 398   bool AMDGPUAA = EnableAMDGPUAliasAnalysis && EnableOpt;
 399   bool LibCallSimplify = EnableLibCallSimplify && EnableOpt;
 400
 401   if (EnableFunctionCalls) {
 402     delete Builder.Inliner;
 403     Builder.Inliner = createAMDGPUFunctionInliningPass();
 404   }
 405
 406   Builder.addExtension(
 407     PassManagerBuilder::EP_ModuleOptimizerEarly,
 408     [Internalize, EarlyInline, AMDGPUAA, this](const PassManagerBuilder &,
 409                                                legacy::PassManagerBase &PM) {
 410       if (AMDGPUAA) {
 411         PM.add(createAMDGPUAAWrapperPass());
 412         PM.add(createAMDGPUExternalAAWrapperPass());
 413       }
 414       PM.add(createAMDGPUUnifyMetadataPass());
 415       PM.add(createAMDGPUPropagateAttributesLatePass(this));
 416       if (Internalize) {
 417         PM.add(createInternalizePass(mustPreserveGV));
 418         PM.add(createGlobalDCEPass());
 419       }
 420       if (EarlyInline)
 421         PM.add(createAMDGPUAlwaysInlinePass(false));
 422   });
 423
 424   const auto &Opt = Options;
 425   Builder.addExtension(
 426     PassManagerBuilder::EP_EarlyAsPossible,
 427     [AMDGPUAA, LibCallSimplify, &Opt, this](const PassManagerBuilder &,
 428                                             legacy::PassManagerBase &PM) {
 429       if (AMDGPUAA) {
 430         PM.add(createAMDGPUAAWrapperPass());
 431         PM.add(createAMDGPUExternalAAWrapperPass());
 432       }
 433       PM.add(llvm::createAMDGPUPropagateAttributesEarlyPass(this));
 434       PM.add(llvm::createAMDGPUUseNativeCallsPass());
 435       if (LibCallSimplify)
 436         PM.add(llvm::createAMDGPUSimplifyLibCallsPass(Opt, this));
 437   });
 438
 439   Builder.addExtension(
 440     PassManagerBuilder::EP_CGSCCOptimizerLate,
 441     [](const PassManagerBuilder &, legacy::PassManagerBase &PM) {
 442       // Add infer address spaces pass to the opt pipeline after inlining
 443       // but before SROA to increase SROA opportunities.
 444       PM.add(createInferAddressSpacesPass());
 445
 446       // This should run after inlining to have any chance of doing anything,
 447       // and before other cleanup optimizations.
 448       PM.add(createAMDGPULowerKernelAttributesPass());
 449   });
 450 }
 451
 452 //===----------------------------------------------------------------------===//
 453 // R600 Target Machine (R600 -> Cayman)
 454 //===----------------------------------------------------------------------===//
 455
 456 R600TargetMachine::R600TargetMachine(const Target &T, const Triple &TT,
 457                                      StringRef CPU, StringRef FS,
 458                                      TargetOptions Options,
 459                                      Optional<Reloc::Model> RM,
 460                                      Optional<CodeModel::Model> CM,
 461                                      CodeGenOpt::Level OL, bool JIT)
 462     : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {
 463   setRequiresStructuredCFG(true);
 464
 465   // Override the default since calls aren't supported for r600.
 466   if (EnableFunctionCalls &&
 467       EnableAMDGPUFunctionCallsOpt.getNumOccurrences() == 0)
 468     EnableFunctionCalls = false;
 469 }
 470
 471 const R600Subtarget *R600TargetMachine::getSubtargetImpl(
 472   const Function &F) const {
 473   StringRef GPU = getGPUName(F);
 474   StringRef FS = getFeatureString(F);
 475
 476   SmallString<128> SubtargetKey(GPU);
 477   SubtargetKey.append(FS);
 478
 479   auto &I = SubtargetMap[SubtargetKey];
 480   if (!I) {
 481     // This needs to be done before we create a new subtarget since any
 482     // creation will depend on the TM and the code generation flags on the
 483     // function that reside in TargetOptions.
 484     resetTargetOptions(F);
 485     I = llvm::make_unique<R600Subtarget>(TargetTriple, GPU, FS, *this);
 486   }
 487
 488   return I.get();
 489 }
 490
 491 TargetTransformInfo
 492 R600TargetMachine::getTargetTransformInfo(const Function &F) {
 493   return TargetTransformInfo(R600TTIImpl(this, F));
 494 }
 495
 496 //===----------------------------------------------------------------------===//
 497 // GCN Target Machine (SI+)
 498 //===----------------------------------------------------------------------===//
 499
 500 GCNTargetMachine::GCNTargetMachine(const Target &T, const Triple &TT,
 501                                    StringRef CPU, StringRef FS,
 502                                    TargetOptions Options,
 503                                    Optional<Reloc::Model> RM,
 504                                    Optional<CodeModel::Model> CM,
 505                                    CodeGenOpt::Level OL, bool JIT)
 506     : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {}
 507
 508 const GCNSubtarget *GCNTargetMachine::getSubtargetImpl(const Function &F) const {
 509   StringRef GPU = getGPUName(F);
 510   StringRef FS = getFeatureString(F);
 511
 512   SmallString<128> SubtargetKey(GPU);
 513   SubtargetKey.append(FS);
 514
 515   auto &I = SubtargetMap[SubtargetKey];
 516   if (!I) {
 517     // This needs to be done before we create a new subtarget since any
 518     // creation will depend on the TM and the code generation flags on the
 519     // function that reside in TargetOptions.
 520     resetTargetOptions(F);
 521     I = llvm::make_unique<GCNSubtarget>(TargetTriple, GPU, FS, *this);
 522   }
 523
 524   I->setScalarizeGlobalBehavior(ScalarizeGlobal);
 525
 526   return I.get();
 527 }
 528
 529 TargetTransformInfo
 530 GCNTargetMachine::getTargetTransformInfo(const Function &F) {
 531   return TargetTransformInfo(GCNTTIImpl(this, F));
 532 }
 533
 534 //===----------------------------------------------------------------------===//
 535 // AMDGPU Pass Setup
 536 //===----------------------------------------------------------------------===//
 537
 538 namespace {
 539
 540 class AMDGPUPassConfig : public TargetPassConfig {
 541 public:
 542   AMDGPUPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM)
 543     : TargetPassConfig(TM, PM) {
 544     // Exceptions and StackMaps are not supported, so these passes will never do
 545     // anything.
 546     disablePass(&StackMapLivenessID);
 547     disablePass(&FuncletLayoutID);
 548   }
 549
 550   AMDGPUTargetMachine &getAMDGPUTargetMachine() const {
 551     return getTM<AMDGPUTargetMachine>();
 552   }
 553
 554   ScheduleDAGInstrs *
 555   createMachineScheduler(MachineSchedContext *C) const override {
 556     ScheduleDAGMILive *DAG = createGenericSchedLive(C);
 557     DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
 558     DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
 559     return DAG;
 560   }
 561
 562   void addEarlyCSEOrGVNPass();
 563   void addStraightLineScalarOptimizationPasses();
 564   void addIRPasses() override;
 565   void addCodeGenPrepare() override;
 566   bool addPreISel() override;
 567   bool addInstSelector() override;
 568   bool addGCPasses() override;
 569
 570   std::unique_ptr<CSEConfigBase> getCSEConfig() const override;
 571 };
 572
 573 std::unique_ptr<CSEConfigBase> AMDGPUPassConfig::getCSEConfig() const {
 574   return getStandardCSEConfigForOpt(TM->getOptLevel());
 575 }
 576
 577 class R600PassConfig final : public AMDGPUPassConfig {
 578 public:
 579   R600PassConfig(LLVMTargetMachine &TM, PassManagerBase &PM)
 580     : AMDGPUPassConfig(TM, PM) {}
 581
 582   ScheduleDAGInstrs *createMachineScheduler(
 583     MachineSchedContext *C) const override {
 584     return createR600MachineScheduler(C);
 585   }
 586
 587   bool addPreISel() override;
 588   bool addInstSelector() override;
 589   void addPreRegAlloc() override;
 590   void addPreSched2() override;
 591   void addPreEmitPass() override;
 592 };
 593
 594 class GCNPassConfig final : public AMDGPUPassConfig {
 595 public:
 596   GCNPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM)
 597     : AMDGPUPassConfig(TM, PM) {
 598     // It is necessary to know the register usage of the entire call graph.  We
 599     // allow calls without EnableAMDGPUFunctionCalls if they are marked
 600     // noinline, so this is always required.
 601     setRequiresCodeGenSCCOrder(true);
 602   }
 603
 604   GCNTargetMachine &getGCNTargetMachine() const {
 605     return getTM<GCNTargetMachine>();
 606   }
 607
 608   ScheduleDAGInstrs *
 609   createMachineScheduler(MachineSchedContext *C) const override;
 610
 611   bool addPreISel() override;
 612   void addMachineSSAOptimization() override;
 613   bool addILPOpts() override;
 614   bool addInstSelector() override;
 615   bool addIRTranslator() override;
 616   bool addLegalizeMachineIR() override;
 617   bool addRegBankSelect() override;
 618   bool addGlobalInstructionSelect() override;
 619   void addFastRegAlloc() override;
 620   void addOptimizedRegAlloc() override;
 621   void addPreRegAlloc() override;
 622   bool addPreRewrite() override;
 623   void addPostRegAlloc() override;
 624   void addPreSched2() override;
 625   void addPreEmitPass() override;
 626 };
 627
 628 } // end anonymous namespace
 629
 630 void AMDGPUPassConfig::addEarlyCSEOrGVNPass() {
 631   if (getOptLevel() == CodeGenOpt::Aggressive)
 632     addPass(createGVNPass());
 633   else
 634     addPass(createEarlyCSEPass());
 635 }
 636
 637 void AMDGPUPassConfig::addStraightLineScalarOptimizationPasses() {
 638   addPass(createLICMPass());
 639   addPass(createSeparateConstOffsetFromGEPPass());
 640   addPass(createSpeculativeExecutionPass());
 641   // ReassociateGEPs exposes more opportunites for SLSR. See
 642   // the example in reassociate-geps-and-slsr.ll.
 643   addPass(createStraightLineStrengthReducePass());
 644   // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or
 645   // EarlyCSE can reuse.
 646   addEarlyCSEOrGVNPass();
 647   // Run NaryReassociate after EarlyCSE/GVN to be more effective.
 648   addPass(createNaryReassociatePass());
 649   // NaryReassociate on GEPs creates redundant common expressions, so run
 650   // EarlyCSE after it.
 651   addPass(createEarlyCSEPass());
 652 }
 653
 654 void AMDGPUPassConfig::addIRPasses() {
 655   const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine();
 656
 657   // There is no reason to run these.
 658   disablePass(&StackMapLivenessID);
 659   disablePass(&FuncletLayoutID);
 660   disablePass(&PatchableFunctionID);
 661
 662   // This must occur before inlining, as the inliner will not look through
 663   // bitcast calls.
 664   addPass(createAMDGPUFixFunctionBitcastsPass());
 665
 666   // A call to propagate attributes pass in the backend in case opt was not run.
 667   addPass(createAMDGPUPropagateAttributesEarlyPass(&TM));
 668
 669   addPass(createAtomicExpandPass());
 670
 671
 672   addPass(createAMDGPULowerIntrinsicsPass());
 673
 674   // Function calls are not supported, so make sure we inline everything.
 675   addPass(createAMDGPUAlwaysInlinePass());
 676   addPass(createAlwaysInlinerLegacyPass());
 677   // We need to add the barrier noop pass, otherwise adding the function
 678   // inlining pass will cause all of the PassConfigs passes to be run
 679   // one function at a time, which means if we have a nodule with two
 680   // functions, then we will generate code for the first function
 681   // without ever running any passes on the second.
 682   addPass(createBarrierNoopPass());
 683
 684   if (TM.getTargetTriple().getArch() == Triple::amdgcn) {
 685     // TODO: May want to move later or split into an early and late one.
 686
 687     addPass(createAMDGPUCodeGenPreparePass());
 688   }
 689
 690   // Handle uses of OpenCL image2d_t, image3d_t and sampler_t arguments.
 691   if (TM.getTargetTriple().getArch() == Triple::r600)
 692     addPass(createR600OpenCLImageTypeLoweringPass());
 693
 694   // Replace OpenCL enqueued block function pointers with global variables.
 695   addPass(createAMDGPUOpenCLEnqueuedBlockLoweringPass());
 696
 697   if (TM.getOptLevel() > CodeGenOpt::None) {
 698     addPass(createInferAddressSpacesPass());
 699     addPass(createAMDGPUPromoteAlloca());
 700
 701     if (EnableSROA)
 702       addPass(createSROAPass());
 703
 704     if (EnableScalarIRPasses)
 705       addStraightLineScalarOptimizationPasses();
 706
 707     if (EnableAMDGPUAliasAnalysis) {
 708       addPass(createAMDGPUAAWrapperPass());
 709       addPass(createExternalAAWrapperPass([](Pass &P, Function &,
 710                                              AAResults &AAR) {
 711         if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>())
 712           AAR.addAAResult(WrapperPass->getResult());
 713         }));
 714     }
 715   }
 716
 717   TargetPassConfig::addIRPasses();
 718
 719   // EarlyCSE is not always strong enough to clean up what LSR produces. For
 720   // example, GVN can combine
 721   //
 722   //   %0 = add %a, %b
 723   //   %1 = add %b, %a
 724   //
 725   // and
 726   //
 727   //   %0 = shl nsw %a, 2
 728   //   %1 = shl %a, 2
 729   //
 730   // but EarlyCSE can do neither of them.
 731   if (getOptLevel() != CodeGenOpt::None && EnableScalarIRPasses)
 732     addEarlyCSEOrGVNPass();
 733 }
 734
 735 void AMDGPUPassConfig::addCodeGenPrepare() {
 736   if (TM->getTargetTriple().getArch() == Triple::amdgcn)
 737     addPass(createAMDGPUAnnotateKernelFeaturesPass());
 738
 739   if (TM->getTargetTriple().getArch() == Triple::amdgcn &&
 740       EnableLowerKernelArguments)
 741     addPass(createAMDGPULowerKernelArgumentsPass());
 742
 743   addPass(&AMDGPUPerfHintAnalysisID);
 744
 745   TargetPassConfig::addCodeGenPrepare();
 746
 747   if (EnableLoadStoreVectorizer)
 748     addPass(createLoadStoreVectorizerPass());
 749 }
 750
 751 bool AMDGPUPassConfig::addPreISel() {
 752   addPass(createLowerSwitchPass());
 753   addPass(createFlattenCFGPass());
 754   return false;
 755 }
 756
 757 bool AMDGPUPassConfig::addInstSelector() {
 758   // Defer the verifier until FinalizeISel.
 759   addPass(createAMDGPUISelDag(&getAMDGPUTargetMachine(), getOptLevel()), false);
 760   return false;
 761 }
 762
 763 bool AMDGPUPassConfig::addGCPasses() {
 764   // Do nothing. GC is not supported.
 765   return false;
 766 }
 767
 768 //===----------------------------------------------------------------------===//
 769 // R600 Pass Setup
 770 //===----------------------------------------------------------------------===//
 771
 772 bool R600PassConfig::addPreISel() {
 773   AMDGPUPassConfig::addPreISel();
 774
 775   if (EnableR600StructurizeCFG)
 776     addPass(createStructurizeCFGPass());
 777   return false;
 778 }
 779
 780 bool R600PassConfig::addInstSelector() {
 781   addPass(createR600ISelDag(&getAMDGPUTargetMachine(), getOptLevel()));
 782   return false;
 783 }
 784
 785 void R600PassConfig::addPreRegAlloc() {
 786   addPass(createR600VectorRegMerger());
 787 }
 788
 789 void R600PassConfig::addPreSched2() {
 790   addPass(createR600EmitClauseMarkers(), false);
 791   if (EnableR600IfConvert)
 792     addPass(&IfConverterID, false);
 793   addPass(createR600ClauseMergePass(), false);
 794 }
 795
 796 void R600PassConfig::addPreEmitPass() {
 797   addPass(createAMDGPUCFGStructurizerPass(), false);
 798   addPass(createR600ExpandSpecialInstrsPass(), false);
 799   addPass(&FinalizeMachineBundlesID, false);
 800   addPass(createR600Packetizer(), false);
 801   addPass(createR600ControlFlowFinalizer(), false);
 802 }
 803
 804 TargetPassConfig *R600TargetMachine::createPassConfig(PassManagerBase &PM) {
 805   return new R600PassConfig(*this, PM);
 806 }
 807
 808 //===----------------------------------------------------------------------===//
 809 // GCN Pass Setup
 810 //===----------------------------------------------------------------------===//
 811
 812 ScheduleDAGInstrs *GCNPassConfig::createMachineScheduler(
 813   MachineSchedContext *C) const {
 814   const GCNSubtarget &ST = C->MF->getSubtarget<GCNSubtarget>();
 815   if (ST.enableSIScheduler())
 816     return createSIMachineScheduler(C);
 817   return createGCNMaxOccupancyMachineScheduler(C);
 818 }
 819
 820 bool GCNPassConfig::addPreISel() {
 821   AMDGPUPassConfig::addPreISel();
 822
 823   if (EnableAtomicOptimizations) {
 824     addPass(createAMDGPUAtomicOptimizerPass());
 825   }
 826
 827   // FIXME: We need to run a pass to propagate the attributes when calls are
 828   // supported.
 829
 830   // Merge divergent exit nodes. StructurizeCFG won't recognize the multi-exit
 831   // regions formed by them.
 832   addPass(&AMDGPUUnifyDivergentExitNodesID);
 833   if (!LateCFGStructurize) {
 834     addPass(createStructurizeCFGPass(true)); // true -> SkipUniformRegions
 835   }
 836   addPass(createSinkingPass());
 837   addPass(createAMDGPUAnnotateUniformValues());
 838   if (!LateCFGStructurize) {
 839     addPass(createSIAnnotateControlFlowPass());
 840   }
 841   addPass(createLCSSAPass());
 842
 843   return false;
 844 }
 845
 846 void GCNPassConfig::addMachineSSAOptimization() {
 847   TargetPassConfig::addMachineSSAOptimization();
 848
 849   // We want to fold operands after PeepholeOptimizer has run (or as part of
 850   // it), because it will eliminate extra copies making it easier to fold the
 851   // real source operand. We want to eliminate dead instructions after, so that
 852   // we see fewer uses of the copies. We then need to clean up the dead
 853   // instructions leftover after the operands are folded as well.
 854   //
 855   // XXX - Can we get away without running DeadMachineInstructionElim again?
 856   addPass(&SIFoldOperandsID);
 857   if (EnableDPPCombine)
 858     addPass(&GCNDPPCombineID);
 859   addPass(&DeadMachineInstructionElimID);
 860   addPass(&SILoadStoreOptimizerID);
 861   if (EnableSDWAPeephole) {
 862     addPass(&SIPeepholeSDWAID);
 863     addPass(&EarlyMachineLICMID);
 864     addPass(&MachineCSEID);
 865     addPass(&SIFoldOperandsID);
 866     addPass(&DeadMachineInstructionElimID);
 867   }
 868   addPass(createSIShrinkInstructionsPass());
 869 }
 870
 871 bool GCNPassConfig::addILPOpts() {
 872   if (EnableEarlyIfConversion)
 873     addPass(&EarlyIfConverterID);
 874
 875   TargetPassConfig::addILPOpts();
 876   return false;
 877 }
 878
 879 bool GCNPassConfig::addInstSelector() {
 880   AMDGPUPassConfig::addInstSelector();
 881   addPass(&SIFixSGPRCopiesID);
 882   addPass(createSILowerI1CopiesPass());
 883   addPass(createSIFixupVectorISelPass());
 884   addPass(createSIAddIMGInitPass());
 885   // FIXME: Remove this once the phi on CF_END is cleaned up by either removing
 886   // LCSSA or other ways.
 887   addPass(&UnreachableMachineBlockElimID);
 888   return false;
 889 }
 890
 891 bool GCNPassConfig::addIRTranslator() {
 892   addPass(new IRTranslator());
 893   return false;
 894 }
 895
 896 bool GCNPassConfig::addLegalizeMachineIR() {
 897   addPass(new Legalizer());
 898   return false;
 899 }
 900
 901 bool GCNPassConfig::addRegBankSelect() {
 902   addPass(new RegBankSelect());
 903   return false;
 904 }
 905
 906 bool GCNPassConfig::addGlobalInstructionSelect() {
 907   addPass(new InstructionSelect());
 908   return false;
 909 }
 910
 911 void GCNPassConfig::addPreRegAlloc() {
 912   if (LateCFGStructurize) {
 913     addPass(createAMDGPUMachineCFGStructurizerPass());
 914   }
 915   addPass(createSIWholeQuadModePass());
 916 }
 917
 918 void GCNPassConfig::addFastRegAlloc() {
 919   // FIXME: We have to disable the verifier here because of PHIElimination +
 920   // TwoAddressInstructions disabling it.
 921
 922   // This must be run immediately after phi elimination and before
 923   // TwoAddressInstructions, otherwise the processing of the tied operand of
 924   // SI_ELSE will introduce a copy of the tied operand source after the else.
 925   insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
 926
 927   // This must be run just after RegisterCoalescing.
 928   insertPass(&RegisterCoalescerID, &SIPreAllocateWWMRegsID, false);
 929
 930   TargetPassConfig::addFastRegAlloc();
 931 }
 932
 933 void GCNPassConfig::addOptimizedRegAlloc() {
 934   if (OptExecMaskPreRA) {
 935     insertPass(&MachineSchedulerID, &SIOptimizeExecMaskingPreRAID);
 936     insertPass(&SIOptimizeExecMaskingPreRAID, &SIFormMemoryClausesID);
 937   } else {
 938     insertPass(&MachineSchedulerID, &SIFormMemoryClausesID);
 939   }
 940
 941   // This must be run immediately after phi elimination and before
 942   // TwoAddressInstructions, otherwise the processing of the tied operand of
 943   // SI_ELSE will introduce a copy of the tied operand source after the else.
 944   insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
 945
 946   // This must be run just after RegisterCoalescing.
 947   insertPass(&RegisterCoalescerID, &SIPreAllocateWWMRegsID, false);
 948
 949   if (EnableDCEInRA)
 950     insertPass(&RenameIndependentSubregsID, &DeadMachineInstructionElimID);
 951
 952   TargetPassConfig::addOptimizedRegAlloc();
 953 }
 954
 955 bool GCNPassConfig::addPreRewrite() {
 956   if (EnableRegReassign) {
 957     addPass(&GCNNSAReassignID);
 958     addPass(&GCNRegBankReassignID);
 959   }
 960   return true;
 961 }
 962
 963 void GCNPassConfig::addPostRegAlloc() {
 964   addPass(&SIFixVGPRCopiesID);
 965   if (getOptLevel() > CodeGenOpt::None)
 966     addPass(&SIOptimizeExecMaskingID);
 967   TargetPassConfig::addPostRegAlloc();
 968
 969   // Equivalent of PEI for SGPRs.
 970   addPass(&SILowerSGPRSpillsID);
 971 }
 972
 973 void GCNPassConfig::addPreSched2() {
 974 }
 975
 976 void GCNPassConfig::addPreEmitPass() {
 977   addPass(createSIMemoryLegalizerPass());
 978   addPass(createSIInsertWaitcntsPass());
 979   addPass(createSIShrinkInstructionsPass());
 980   addPass(createSIModeRegisterPass());
 981
 982   // The hazard recognizer that runs as part of the post-ra scheduler does not
 983   // guarantee to be able handle all hazards correctly. This is because if there
 984   // are multiple scheduling regions in a basic block, the regions are scheduled
 985   // bottom up, so when we begin to schedule a region we don't know what
 986   // instructions were emitted directly before it.
 987   //
 988   // Here we add a stand-alone hazard recognizer pass which can handle all
 989   // cases.
 990   //
 991   // FIXME: This stand-alone pass will emit indiv. S_NOP 0, as needed. It would
 992   // be better for it to emit S_NOP <N> when possible.
 993   addPass(&PostRAHazardRecognizerID);
 994
 995   addPass(&SIInsertSkipsPassID);
 996   addPass(&BranchRelaxationPassID);
 997 }
 998
 999 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) {
1000   return new GCNPassConfig(*this, PM);
1001 }
1002
1003 yaml::MachineFunctionInfo *GCNTargetMachine::createDefaultFuncInfoYAML() const {
1004   return new yaml::SIMachineFunctionInfo();
1005 }
1006
1007 yaml::MachineFunctionInfo *
1008 GCNTargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const {
1009   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1010   return new yaml::SIMachineFunctionInfo(*MFI,
1011                                          *MF.getSubtarget().getRegisterInfo());
1012 }
1013
1014 bool GCNTargetMachine::parseMachineFunctionInfo(
1015     const yaml::MachineFunctionInfo &MFI_, PerFunctionMIParsingState &PFS,
1016     SMDiagnostic &Error, SMRange &SourceRange) const {
1017   const yaml::SIMachineFunctionInfo &YamlMFI =
1018       reinterpret_cast<const yaml::SIMachineFunctionInfo &>(MFI_);
1019   MachineFunction &MF = PFS.MF;
1020   SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1021
1022   MFI->initializeBaseYamlFields(YamlMFI);
1023
1024   auto parseRegister = [&](const yaml::StringValue &RegName, unsigned &RegVal) {
1025     if (parseNamedRegisterReference(PFS, RegVal, RegName.Value, Error)) {
1026       SourceRange = RegName.SourceRange;
1027       return true;
1028     }
1029
1030     return false;
1031   };
1032
1033   auto diagnoseRegisterClass = [&](const yaml::StringValue &RegName) {
1034     // Create a diagnostic for a the register string literal.
1035     const MemoryBuffer &Buffer =
1036         *PFS.SM->getMemoryBuffer(PFS.SM->getMainFileID());
1037     Error = SMDiagnostic(*PFS.SM, SMLoc(), Buffer.getBufferIdentifier(), 1,
1038                          RegName.Value.size(), SourceMgr::DK_Error,
1039                          "incorrect register class for field", RegName.Value,
1040                          None, None);
1041     SourceRange = RegName.SourceRange;
1042     return true;
1043   };
1044
1045   if (parseRegister(YamlMFI.ScratchRSrcReg, MFI->ScratchRSrcReg) ||
1046       parseRegister(YamlMFI.ScratchWaveOffsetReg, MFI->ScratchWaveOffsetReg) ||
1047       parseRegister(YamlMFI.FrameOffsetReg, MFI->FrameOffsetReg) ||
1048       parseRegister(YamlMFI.StackPtrOffsetReg, MFI->StackPtrOffsetReg))
1049     return true;
1050
1051   if (MFI->ScratchRSrcReg != AMDGPU::PRIVATE_RSRC_REG &&
1052       !AMDGPU::SReg_128RegClass.contains(MFI->ScratchRSrcReg)) {
1053     return diagnoseRegisterClass(YamlMFI.ScratchRSrcReg);
1054   }
1055
1056   if (MFI->ScratchWaveOffsetReg != AMDGPU::SCRATCH_WAVE_OFFSET_REG &&
1057       !AMDGPU::SGPR_32RegClass.contains(MFI->ScratchWaveOffsetReg)) {
1058     return diagnoseRegisterClass(YamlMFI.ScratchWaveOffsetReg);
1059   }
1060
1061   if (MFI->FrameOffsetReg != AMDGPU::FP_REG &&
1062       !AMDGPU::SGPR_32RegClass.contains(MFI->FrameOffsetReg)) {
1063     return diagnoseRegisterClass(YamlMFI.FrameOffsetReg);
1064   }
1065
1066   if (MFI->StackPtrOffsetReg != AMDGPU::SP_REG &&
1067       !AMDGPU::SGPR_32RegClass.contains(MFI->StackPtrOffsetReg)) {
1068     return diagnoseRegisterClass(YamlMFI.StackPtrOffsetReg);
1069   }
1070
1071   auto parseAndCheckArgument = [&](const Optional<yaml::SIArgument> &A,
1072                                    const TargetRegisterClass &RC,
1073                                    ArgDescriptor &Arg, unsigned UserSGPRs,
1074                                    unsigned SystemSGPRs) {
1075     // Skip parsing if it's not present.
1076     if (!A)
1077       return false;
1078
1079     if (A->IsRegister) {
1080       unsigned Reg;
1081       if (parseNamedRegisterReference(PFS, Reg, A->RegisterName.Value, Error)) {
1082         SourceRange = A->RegisterName.SourceRange;
1083         return true;
1084       }
1085       if (!RC.contains(Reg))
1086         return diagnoseRegisterClass(A->RegisterName);
1087       Arg = ArgDescriptor::createRegister(Reg);
1088     } else
1089       Arg = ArgDescriptor::createStack(A->StackOffset);
1090     // Check and apply the optional mask.
1091     if (A->Mask)
1092       Arg = ArgDescriptor::createArg(Arg, A->Mask.getValue());
1093
1094     MFI->NumUserSGPRs += UserSGPRs;
1095     MFI->NumSystemSGPRs += SystemSGPRs;
1096     return false;
1097   };
1098
1099   if (YamlMFI.ArgInfo &&
1100       (parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentBuffer,
1101                              AMDGPU::SReg_128RegClass,
1102                              MFI->ArgInfo.PrivateSegmentBuffer, 4, 0) ||
1103        parseAndCheckArgument(YamlMFI.ArgInfo->DispatchPtr,
1104                              AMDGPU::SReg_64RegClass, MFI->ArgInfo.DispatchPtr,
1105                              2, 0) ||
1106        parseAndCheckArgument(YamlMFI.ArgInfo->QueuePtr, AMDGPU::SReg_64RegClass,
1107                              MFI->ArgInfo.QueuePtr, 2, 0) ||
1108        parseAndCheckArgument(YamlMFI.ArgInfo->KernargSegmentPtr,
1109                              AMDGPU::SReg_64RegClass,
1110                              MFI->ArgInfo.KernargSegmentPtr, 2, 0) ||
1111        parseAndCheckArgument(YamlMFI.ArgInfo->DispatchID,
1112                              AMDGPU::SReg_64RegClass, MFI->ArgInfo.DispatchID,
1113                              2, 0) ||
1114        parseAndCheckArgument(YamlMFI.ArgInfo->FlatScratchInit,
1115                              AMDGPU::SReg_64RegClass,
1116                              MFI->ArgInfo.FlatScratchInit, 2, 0) ||
1117        parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentSize,
1118                              AMDGPU::SGPR_32RegClass,
1119                              MFI->ArgInfo.PrivateSegmentSize, 0, 0) ||
1120        parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDX,
1121                              AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDX,
1122                              0, 1) ||
1123        parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDY,
1124                              AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDY,
1125                              0, 1) ||
1126        parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDZ,
1127                              AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDZ,
1128                              0, 1) ||
1129        parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupInfo,
1130                              AMDGPU::SGPR_32RegClass,
1131                              MFI->ArgInfo.WorkGroupInfo, 0, 1) ||
1132        parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentWaveByteOffset,
1133                              AMDGPU::SGPR_32RegClass,
1134                              MFI->ArgInfo.PrivateSegmentWaveByteOffset, 0, 1) ||
1135        parseAndCheckArgument(YamlMFI.ArgInfo->ImplicitArgPtr,
1136                              AMDGPU::SReg_64RegClass,
1137                              MFI->ArgInfo.ImplicitArgPtr, 0, 0) ||
1138        parseAndCheckArgument(YamlMFI.ArgInfo->ImplicitBufferPtr,
1139                              AMDGPU::SReg_64RegClass,
1140                              MFI->ArgInfo.ImplicitBufferPtr, 2, 0) ||
1141        parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDX,
1142                              AMDGPU::VGPR_32RegClass,
1143                              MFI->ArgInfo.WorkItemIDX, 0, 0) ||
1144        parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDY,
1145                              AMDGPU::VGPR_32RegClass,
1146                              MFI->ArgInfo.WorkItemIDY, 0, 0) ||
1147        parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDZ,
1148                              AMDGPU::VGPR_32RegClass,
1149                              MFI->ArgInfo.WorkItemIDZ, 0, 0)))
1150     return true;
1151
1152   MFI->Mode.IEEE = YamlMFI.Mode.IEEE;
1153   MFI->Mode.DX10Clamp = YamlMFI.Mode.DX10Clamp;
1154
1155   return false;
1156 }