1 //===-- AMDGPUTargetMachine.cpp - TargetMachine for hw codegen targets-----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief The AMDGPU target machine contains all of the hardware specific
12 /// information needed to emit code for R600 and SI GPUs.
14 //===----------------------------------------------------------------------===//
16 #include "AMDGPUTargetMachine.h"
18 #include "AMDGPUAliasAnalysis.h"
19 #include "AMDGPUCallLowering.h"
20 #include "AMDGPUInstructionSelector.h"
21 #include "AMDGPULegalizerInfo.h"
22 #ifdef LLVM_BUILD_GLOBAL_ISEL
23 #include "AMDGPURegisterBankInfo.h"
25 #include "AMDGPUTargetObjectFile.h"
26 #include "AMDGPUTargetTransformInfo.h"
27 #include "GCNIterativeScheduler.h"
28 #include "GCNSchedStrategy.h"
29 #include "R600MachineScheduler.h"
30 #include "SIMachineScheduler.h"
31 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h"
32 #include "llvm/CodeGen/GlobalISel/IRTranslator.h"
33 #include "llvm/CodeGen/GlobalISel/Legalizer.h"
34 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h"
35 #include "llvm/CodeGen/Passes.h"
36 #include "llvm/CodeGen/TargetPassConfig.h"
37 #include "llvm/Support/TargetRegistry.h"
38 #include "llvm/Transforms/IPO.h"
39 #include "llvm/Transforms/IPO/AlwaysInliner.h"
40 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
41 #include "llvm/Transforms/Scalar.h"
42 #include "llvm/Transforms/Scalar/GVN.h"
43 #include "llvm/Transforms/Vectorize.h"
44 #include "llvm/IR/Attributes.h"
45 #include "llvm/IR/Function.h"
46 #include "llvm/IR/LegacyPassManager.h"
47 #include "llvm/Pass.h"
48 #include "llvm/Support/CommandLine.h"
49 #include "llvm/Support/Compiler.h"
50 #include "llvm/Target/TargetLoweringObjectFile.h"
55 static cl::opt<bool> EnableR600StructurizeCFG(
56 "r600-ir-structurize",
57 cl::desc("Use StructurizeCFG IR pass"),
60 static cl::opt<bool> EnableSROA(
62 cl::desc("Run SROA after promote alloca pass"),
67 EnableEarlyIfConversion("amdgpu-early-ifcvt", cl::Hidden,
68 cl::desc("Run early if-conversion"),
71 static cl::opt<bool> EnableR600IfConvert(
73 cl::desc("Use if conversion pass"),
77 // Option to disable vectorizer for tests.
78 static cl::opt<bool> EnableLoadStoreVectorizer(
79 "amdgpu-load-store-vectorizer",
80 cl::desc("Enable load store vectorizer"),
84 // Option to to control global loads scalarization
85 static cl::opt<bool> ScalarizeGlobal(
86 "amdgpu-scalarize-global-loads",
87 cl::desc("Enable global load scalarization"),
91 // Option to run internalize pass.
92 static cl::opt<bool> InternalizeSymbols(
93 "amdgpu-internalize-symbols",
94 cl::desc("Enable elimination of non-kernel functions and unused globals"),
98 // Option to inline all early.
99 static cl::opt<bool> EarlyInlineAll(
100 "amdgpu-early-inline-all",
101 cl::desc("Inline all functions early"),
105 static cl::opt<bool> EnableSDWAPeephole(
106 "amdgpu-sdwa-peephole",
107 cl::desc("Enable SDWA peepholer"),
110 // Enable address space based alias analysis
111 static cl::opt<bool> EnableAMDGPUAliasAnalysis("enable-amdgpu-aa", cl::Hidden,
112 cl::desc("Enable AMDGPU Alias Analysis"),
115 // Option to enable new waitcnt insertion pass.
116 static cl::opt<bool> EnableSIInsertWaitcntsPass(
117 "enable-si-insert-waitcnts",
118 cl::desc("Use new waitcnt insertion pass"),
121 // Option to run late CFG structurizer
122 static cl::opt<bool> LateCFGStructurize(
123 "amdgpu-late-structurize",
124 cl::desc("Enable late CFG structurization"),
128 extern "C" void LLVMInitializeAMDGPUTarget() {
129 // Register the target
130 RegisterTargetMachine<R600TargetMachine> X(getTheAMDGPUTarget());
131 RegisterTargetMachine<GCNTargetMachine> Y(getTheGCNTarget());
133 PassRegistry *PR = PassRegistry::getPassRegistry();
134 initializeSILowerI1CopiesPass(*PR);
135 initializeSIFixSGPRCopiesPass(*PR);
136 initializeSIFixVGPRCopiesPass(*PR);
137 initializeSIFoldOperandsPass(*PR);
138 initializeSIPeepholeSDWAPass(*PR);
139 initializeSIShrinkInstructionsPass(*PR);
140 initializeSIFixControlFlowLiveIntervalsPass(*PR);
141 initializeSILoadStoreOptimizerPass(*PR);
142 initializeAMDGPUAnnotateKernelFeaturesPass(*PR);
143 initializeAMDGPUAnnotateUniformValuesPass(*PR);
144 initializeAMDGPULowerIntrinsicsPass(*PR);
145 initializeAMDGPUPromoteAllocaPass(*PR);
146 initializeAMDGPUCodeGenPreparePass(*PR);
147 initializeAMDGPUUnifyMetadataPass(*PR);
148 initializeSIAnnotateControlFlowPass(*PR);
149 initializeSIInsertWaitsPass(*PR);
150 initializeSIInsertWaitcntsPass(*PR);
151 initializeSIWholeQuadModePass(*PR);
152 initializeSILowerControlFlowPass(*PR);
153 initializeSIInsertSkipsPass(*PR);
154 initializeSIDebuggerInsertNopsPass(*PR);
155 initializeSIOptimizeExecMaskingPass(*PR);
156 initializeAMDGPUUnifyDivergentExitNodesPass(*PR);
157 initializeAMDGPUAAWrapperPassPass(*PR);
160 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
161 return llvm::make_unique<AMDGPUTargetObjectFile>();
164 static ScheduleDAGInstrs *createR600MachineScheduler(MachineSchedContext *C) {
165 return new ScheduleDAGMILive(C, llvm::make_unique<R600SchedStrategy>());
168 static ScheduleDAGInstrs *createSIMachineScheduler(MachineSchedContext *C) {
169 return new SIScheduleDAGMI(C);
172 static ScheduleDAGInstrs *
173 createGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
174 ScheduleDAGMILive *DAG =
175 new GCNScheduleDAGMILive(C, make_unique<GCNMaxOccupancySchedStrategy>(C));
176 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
177 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
181 static ScheduleDAGInstrs *
182 createIterativeGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
183 auto DAG = new GCNIterativeScheduler(C,
184 GCNIterativeScheduler::SCHEDULE_LEGACYMAXOCCUPANCY);
185 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
186 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
190 static ScheduleDAGInstrs *createMinRegScheduler(MachineSchedContext *C) {
191 return new GCNIterativeScheduler(C,
192 GCNIterativeScheduler::SCHEDULE_MINREGFORCED);
195 static MachineSchedRegistry
196 R600SchedRegistry("r600", "Run R600's custom scheduler",
197 createR600MachineScheduler);
199 static MachineSchedRegistry
200 SISchedRegistry("si", "Run SI's custom scheduler",
201 createSIMachineScheduler);
203 static MachineSchedRegistry
204 GCNMaxOccupancySchedRegistry("gcn-max-occupancy",
205 "Run GCN scheduler to maximize occupancy",
206 createGCNMaxOccupancyMachineScheduler);
208 static MachineSchedRegistry
209 IterativeGCNMaxOccupancySchedRegistry("gcn-max-occupancy-experimental",
210 "Run GCN scheduler to maximize occupancy (experimental)",
211 createIterativeGCNMaxOccupancyMachineScheduler);
213 static MachineSchedRegistry
214 GCNMinRegSchedRegistry("gcn-minreg",
215 "Run GCN iterative scheduler for minimal register usage (experimental)",
216 createMinRegScheduler);
218 static StringRef computeDataLayout(const Triple &TT) {
219 if (TT.getArch() == Triple::r600) {
221 return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
222 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64";
225 // 32-bit private, local, and region pointers. 64-bit global, constant and
227 if (TT.getEnvironmentName() == "amdgiz" ||
228 TT.getEnvironmentName() == "amdgizcl")
229 return "e-p:64:64-p1:64:64-p2:64:64-p3:32:32-p4:32:32-p5:32:32"
230 "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
231 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-A5";
232 return "e-p:32:32-p1:64:64-p2:64:64-p3:32:32-p4:64:64-p5:32:32"
233 "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
234 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64";
238 static StringRef getGPUOrDefault(const Triple &TT, StringRef GPU) {
242 // HSA only supports CI+, so change the default GPU to a CI for HSA.
243 if (TT.getArch() == Triple::amdgcn)
244 return (TT.getOS() == Triple::AMDHSA) ? "kaveri" : "tahiti";
249 static Reloc::Model getEffectiveRelocModel(Optional<Reloc::Model> RM) {
250 // The AMDGPU toolchain only supports generating shared objects, so we
251 // must always use PIC.
255 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, const Triple &TT,
256 StringRef CPU, StringRef FS,
257 TargetOptions Options,
258 Optional<Reloc::Model> RM,
260 CodeGenOpt::Level OptLevel)
261 : LLVMTargetMachine(T, computeDataLayout(TT), TT, getGPUOrDefault(TT, CPU),
262 FS, Options, getEffectiveRelocModel(RM), CM, OptLevel),
263 TLOF(createTLOF(getTargetTriple())) {
264 AS = AMDGPU::getAMDGPUAS(TT);
268 AMDGPUTargetMachine::~AMDGPUTargetMachine() = default;
270 StringRef AMDGPUTargetMachine::getGPUName(const Function &F) const {
271 Attribute GPUAttr = F.getFnAttribute("target-cpu");
272 return GPUAttr.hasAttribute(Attribute::None) ?
273 getTargetCPU() : GPUAttr.getValueAsString();
276 StringRef AMDGPUTargetMachine::getFeatureString(const Function &F) const {
277 Attribute FSAttr = F.getFnAttribute("target-features");
279 return FSAttr.hasAttribute(Attribute::None) ?
280 getTargetFeatureString() :
281 FSAttr.getValueAsString();
284 static ImmutablePass *createAMDGPUExternalAAWrapperPass() {
285 return createExternalAAWrapperPass([](Pass &P, Function &, AAResults &AAR) {
286 if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>())
287 AAR.addAAResult(WrapperPass->getResult());
291 void AMDGPUTargetMachine::adjustPassManager(PassManagerBuilder &Builder) {
292 Builder.DivergentTarget = true;
294 bool Internalize = InternalizeSymbols &&
295 (getOptLevel() > CodeGenOpt::None) &&
296 (getTargetTriple().getArch() == Triple::amdgcn);
297 bool EarlyInline = EarlyInlineAll &&
298 (getOptLevel() > CodeGenOpt::None);
299 bool AMDGPUAA = EnableAMDGPUAliasAnalysis && getOptLevel() > CodeGenOpt::None;
301 Builder.addExtension(
302 PassManagerBuilder::EP_ModuleOptimizerEarly,
303 [Internalize, EarlyInline, AMDGPUAA](const PassManagerBuilder &,
304 legacy::PassManagerBase &PM) {
306 PM.add(createAMDGPUAAWrapperPass());
307 PM.add(createAMDGPUExternalAAWrapperPass());
309 PM.add(createAMDGPUUnifyMetadataPass());
311 PM.add(createInternalizePass([=](const GlobalValue &GV) -> bool {
312 if (const Function *F = dyn_cast<Function>(&GV)) {
313 if (F->isDeclaration())
315 switch (F->getCallingConv()) {
318 case CallingConv::AMDGPU_VS:
319 case CallingConv::AMDGPU_HS:
320 case CallingConv::AMDGPU_GS:
321 case CallingConv::AMDGPU_PS:
322 case CallingConv::AMDGPU_CS:
323 case CallingConv::AMDGPU_KERNEL:
324 case CallingConv::SPIR_KERNEL:
328 return !GV.use_empty();
330 PM.add(createGlobalDCEPass());
333 PM.add(createAMDGPUAlwaysInlinePass(false));
336 Builder.addExtension(
337 PassManagerBuilder::EP_EarlyAsPossible,
338 [AMDGPUAA](const PassManagerBuilder &, legacy::PassManagerBase &PM) {
340 PM.add(createAMDGPUAAWrapperPass());
341 PM.add(createAMDGPUExternalAAWrapperPass());
346 //===----------------------------------------------------------------------===//
347 // R600 Target Machine (R600 -> Cayman)
348 //===----------------------------------------------------------------------===//
350 R600TargetMachine::R600TargetMachine(const Target &T, const Triple &TT,
351 StringRef CPU, StringRef FS,
352 TargetOptions Options,
353 Optional<Reloc::Model> RM,
354 CodeModel::Model CM, CodeGenOpt::Level OL)
355 : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {
356 setRequiresStructuredCFG(true);
359 const R600Subtarget *R600TargetMachine::getSubtargetImpl(
360 const Function &F) const {
361 StringRef GPU = getGPUName(F);
362 StringRef FS = getFeatureString(F);
364 SmallString<128> SubtargetKey(GPU);
365 SubtargetKey.append(FS);
367 auto &I = SubtargetMap[SubtargetKey];
369 // This needs to be done before we create a new subtarget since any
370 // creation will depend on the TM and the code generation flags on the
371 // function that reside in TargetOptions.
372 resetTargetOptions(F);
373 I = llvm::make_unique<R600Subtarget>(TargetTriple, GPU, FS, *this);
379 //===----------------------------------------------------------------------===//
380 // GCN Target Machine (SI+)
381 //===----------------------------------------------------------------------===//
383 #ifdef LLVM_BUILD_GLOBAL_ISEL
386 struct SIGISelActualAccessor : public GISelAccessor {
387 std::unique_ptr<AMDGPUCallLowering> CallLoweringInfo;
388 std::unique_ptr<InstructionSelector> InstSelector;
389 std::unique_ptr<LegalizerInfo> Legalizer;
390 std::unique_ptr<RegisterBankInfo> RegBankInfo;
391 const AMDGPUCallLowering *getCallLowering() const override {
392 return CallLoweringInfo.get();
394 const InstructionSelector *getInstructionSelector() const override {
395 return InstSelector.get();
397 const LegalizerInfo *getLegalizerInfo() const override {
398 return Legalizer.get();
400 const RegisterBankInfo *getRegBankInfo() const override {
401 return RegBankInfo.get();
405 } // end anonymous namespace
408 GCNTargetMachine::GCNTargetMachine(const Target &T, const Triple &TT,
409 StringRef CPU, StringRef FS,
410 TargetOptions Options,
411 Optional<Reloc::Model> RM,
412 CodeModel::Model CM, CodeGenOpt::Level OL)
413 : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {}
415 const SISubtarget *GCNTargetMachine::getSubtargetImpl(const Function &F) const {
416 StringRef GPU = getGPUName(F);
417 StringRef FS = getFeatureString(F);
419 SmallString<128> SubtargetKey(GPU);
420 SubtargetKey.append(FS);
422 auto &I = SubtargetMap[SubtargetKey];
424 // This needs to be done before we create a new subtarget since any
425 // creation will depend on the TM and the code generation flags on the
426 // function that reside in TargetOptions.
427 resetTargetOptions(F);
428 I = llvm::make_unique<SISubtarget>(TargetTriple, GPU, FS, *this);
430 #ifndef LLVM_BUILD_GLOBAL_ISEL
431 GISelAccessor *GISel = new GISelAccessor();
433 SIGISelActualAccessor *GISel = new SIGISelActualAccessor();
434 GISel->CallLoweringInfo.reset(
435 new AMDGPUCallLowering(*I->getTargetLowering()));
436 GISel->Legalizer.reset(new AMDGPULegalizerInfo());
438 GISel->RegBankInfo.reset(new AMDGPURegisterBankInfo(*I->getRegisterInfo()));
439 GISel->InstSelector.reset(new AMDGPUInstructionSelector(*I,
440 *static_cast<AMDGPURegisterBankInfo*>(GISel->RegBankInfo.get())));
443 I->setGISelAccessor(*GISel);
446 I->setScalarizeGlobalBehavior(ScalarizeGlobal);
451 //===----------------------------------------------------------------------===//
453 //===----------------------------------------------------------------------===//
457 class AMDGPUPassConfig : public TargetPassConfig {
459 AMDGPUPassConfig(TargetMachine *TM, PassManagerBase &PM)
460 : TargetPassConfig(TM, PM) {
461 // Exceptions and StackMaps are not supported, so these passes will never do
463 disablePass(&StackMapLivenessID);
464 disablePass(&FuncletLayoutID);
467 AMDGPUTargetMachine &getAMDGPUTargetMachine() const {
468 return getTM<AMDGPUTargetMachine>();
472 createMachineScheduler(MachineSchedContext *C) const override {
473 ScheduleDAGMILive *DAG = createGenericSchedLive(C);
474 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
475 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
479 void addEarlyCSEOrGVNPass();
480 void addStraightLineScalarOptimizationPasses();
481 void addIRPasses() override;
482 void addCodeGenPrepare() override;
483 bool addPreISel() override;
484 bool addInstSelector() override;
485 bool addGCPasses() override;
488 class R600PassConfig final : public AMDGPUPassConfig {
490 R600PassConfig(TargetMachine *TM, PassManagerBase &PM)
491 : AMDGPUPassConfig(TM, PM) {}
493 ScheduleDAGInstrs *createMachineScheduler(
494 MachineSchedContext *C) const override {
495 return createR600MachineScheduler(C);
498 bool addPreISel() override;
499 void addPreRegAlloc() override;
500 void addPreSched2() override;
501 void addPreEmitPass() override;
504 class GCNPassConfig final : public AMDGPUPassConfig {
506 GCNPassConfig(TargetMachine *TM, PassManagerBase &PM)
507 : AMDGPUPassConfig(TM, PM) {}
509 GCNTargetMachine &getGCNTargetMachine() const {
510 return getTM<GCNTargetMachine>();
514 createMachineScheduler(MachineSchedContext *C) const override;
516 bool addPreISel() override;
517 void addMachineSSAOptimization() override;
518 bool addILPOpts() override;
519 bool addInstSelector() override;
520 #ifdef LLVM_BUILD_GLOBAL_ISEL
521 bool addIRTranslator() override;
522 bool addLegalizeMachineIR() override;
523 bool addRegBankSelect() override;
524 bool addGlobalInstructionSelect() override;
526 void addFastRegAlloc(FunctionPass *RegAllocPass) override;
527 void addOptimizedRegAlloc(FunctionPass *RegAllocPass) override;
528 void addPreRegAlloc() override;
529 void addPostRegAlloc() override;
530 void addPreSched2() override;
531 void addPreEmitPass() override;
534 } // end anonymous namespace
536 TargetIRAnalysis AMDGPUTargetMachine::getTargetIRAnalysis() {
537 return TargetIRAnalysis([this](const Function &F) {
538 return TargetTransformInfo(AMDGPUTTIImpl(this, F));
542 void AMDGPUPassConfig::addEarlyCSEOrGVNPass() {
543 if (getOptLevel() == CodeGenOpt::Aggressive)
544 addPass(createGVNPass());
546 addPass(createEarlyCSEPass());
549 void AMDGPUPassConfig::addStraightLineScalarOptimizationPasses() {
550 addPass(createSeparateConstOffsetFromGEPPass());
551 addPass(createSpeculativeExecutionPass());
552 // ReassociateGEPs exposes more opportunites for SLSR. See
553 // the example in reassociate-geps-and-slsr.ll.
554 addPass(createStraightLineStrengthReducePass());
555 // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or
556 // EarlyCSE can reuse.
557 addEarlyCSEOrGVNPass();
558 // Run NaryReassociate after EarlyCSE/GVN to be more effective.
559 addPass(createNaryReassociatePass());
560 // NaryReassociate on GEPs creates redundant common expressions, so run
561 // EarlyCSE after it.
562 addPass(createEarlyCSEPass());
565 void AMDGPUPassConfig::addIRPasses() {
566 const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine();
568 // There is no reason to run these.
569 disablePass(&StackMapLivenessID);
570 disablePass(&FuncletLayoutID);
571 disablePass(&PatchableFunctionID);
573 addPass(createAMDGPULowerIntrinsicsPass(&TM));
575 // Function calls are not supported, so make sure we inline everything.
576 addPass(createAMDGPUAlwaysInlinePass());
577 addPass(createAlwaysInlinerLegacyPass());
578 // We need to add the barrier noop pass, otherwise adding the function
579 // inlining pass will cause all of the PassConfigs passes to be run
580 // one function at a time, which means if we have a nodule with two
581 // functions, then we will generate code for the first function
582 // without ever running any passes on the second.
583 addPass(createBarrierNoopPass());
585 if (TM.getTargetTriple().getArch() == Triple::amdgcn) {
586 // TODO: May want to move later or split into an early and late one.
588 addPass(createAMDGPUCodeGenPreparePass(
589 static_cast<const GCNTargetMachine *>(&TM)));
592 // Handle uses of OpenCL image2d_t, image3d_t and sampler_t arguments.
593 addPass(createAMDGPUOpenCLImageTypeLoweringPass());
595 if (TM.getOptLevel() > CodeGenOpt::None) {
596 addPass(createInferAddressSpacesPass());
597 addPass(createAMDGPUPromoteAlloca(&TM));
600 addPass(createSROAPass());
602 addStraightLineScalarOptimizationPasses();
604 if (EnableAMDGPUAliasAnalysis) {
605 addPass(createAMDGPUAAWrapperPass());
606 addPass(createExternalAAWrapperPass([](Pass &P, Function &,
608 if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>())
609 AAR.addAAResult(WrapperPass->getResult());
614 TargetPassConfig::addIRPasses();
616 // EarlyCSE is not always strong enough to clean up what LSR produces. For
617 // example, GVN can combine
624 // %0 = shl nsw %a, 2
627 // but EarlyCSE can do neither of them.
628 if (getOptLevel() != CodeGenOpt::None)
629 addEarlyCSEOrGVNPass();
632 void AMDGPUPassConfig::addCodeGenPrepare() {
633 TargetPassConfig::addCodeGenPrepare();
635 if (EnableLoadStoreVectorizer)
636 addPass(createLoadStoreVectorizerPass());
639 bool AMDGPUPassConfig::addPreISel() {
640 addPass(createFlattenCFGPass());
644 bool AMDGPUPassConfig::addInstSelector() {
645 addPass(createAMDGPUISelDag(getAMDGPUTargetMachine(), getOptLevel()));
649 bool AMDGPUPassConfig::addGCPasses() {
650 // Do nothing. GC is not supported.
654 //===----------------------------------------------------------------------===//
656 //===----------------------------------------------------------------------===//
658 bool R600PassConfig::addPreISel() {
659 AMDGPUPassConfig::addPreISel();
661 if (EnableR600StructurizeCFG)
662 addPass(createStructurizeCFGPass());
666 void R600PassConfig::addPreRegAlloc() {
667 addPass(createR600VectorRegMerger(*TM));
670 void R600PassConfig::addPreSched2() {
671 addPass(createR600EmitClauseMarkers(), false);
672 if (EnableR600IfConvert)
673 addPass(&IfConverterID, false);
674 addPass(createR600ClauseMergePass(*TM), false);
677 void R600PassConfig::addPreEmitPass() {
678 addPass(createAMDGPUCFGStructurizerPass(), false);
679 addPass(createR600ExpandSpecialInstrsPass(*TM), false);
680 addPass(&FinalizeMachineBundlesID, false);
681 addPass(createR600Packetizer(*TM), false);
682 addPass(createR600ControlFlowFinalizer(*TM), false);
685 TargetPassConfig *R600TargetMachine::createPassConfig(PassManagerBase &PM) {
686 return new R600PassConfig(this, PM);
689 //===----------------------------------------------------------------------===//
691 //===----------------------------------------------------------------------===//
693 ScheduleDAGInstrs *GCNPassConfig::createMachineScheduler(
694 MachineSchedContext *C) const {
695 const SISubtarget &ST = C->MF->getSubtarget<SISubtarget>();
696 if (ST.enableSIScheduler())
697 return createSIMachineScheduler(C);
698 return createGCNMaxOccupancyMachineScheduler(C);
701 bool GCNPassConfig::addPreISel() {
702 AMDGPUPassConfig::addPreISel();
704 // FIXME: We need to run a pass to propagate the attributes when calls are
706 const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine();
707 addPass(createAMDGPUAnnotateKernelFeaturesPass(&TM));
709 // Merge divergent exit nodes. StructurizeCFG won't recognize the multi-exit
710 // regions formed by them.
711 addPass(&AMDGPUUnifyDivergentExitNodesID);
712 if (!LateCFGStructurize) {
713 addPass(createStructurizeCFGPass(true)); // true -> SkipUniformRegions
715 addPass(createSinkingPass());
716 addPass(createSITypeRewriter());
717 addPass(createAMDGPUAnnotateUniformValues());
718 if (!LateCFGStructurize) {
719 addPass(createSIAnnotateControlFlowPass());
725 void GCNPassConfig::addMachineSSAOptimization() {
726 TargetPassConfig::addMachineSSAOptimization();
728 // We want to fold operands after PeepholeOptimizer has run (or as part of
729 // it), because it will eliminate extra copies making it easier to fold the
730 // real source operand. We want to eliminate dead instructions after, so that
731 // we see fewer uses of the copies. We then need to clean up the dead
732 // instructions leftover after the operands are folded as well.
734 // XXX - Can we get away without running DeadMachineInstructionElim again?
735 addPass(&SIFoldOperandsID);
736 addPass(&DeadMachineInstructionElimID);
737 addPass(&SILoadStoreOptimizerID);
738 addPass(createSIShrinkInstructionsPass());
739 if (EnableSDWAPeephole) {
740 addPass(&SIPeepholeSDWAID);
741 addPass(&DeadMachineInstructionElimID);
745 bool GCNPassConfig::addILPOpts() {
746 if (EnableEarlyIfConversion)
747 addPass(&EarlyIfConverterID);
749 TargetPassConfig::addILPOpts();
753 bool GCNPassConfig::addInstSelector() {
754 AMDGPUPassConfig::addInstSelector();
755 addPass(createSILowerI1CopiesPass());
756 addPass(&SIFixSGPRCopiesID);
760 #ifdef LLVM_BUILD_GLOBAL_ISEL
761 bool GCNPassConfig::addIRTranslator() {
762 addPass(new IRTranslator());
766 bool GCNPassConfig::addLegalizeMachineIR() {
767 addPass(new Legalizer());
771 bool GCNPassConfig::addRegBankSelect() {
772 addPass(new RegBankSelect());
776 bool GCNPassConfig::addGlobalInstructionSelect() {
777 addPass(new InstructionSelect());
783 void GCNPassConfig::addPreRegAlloc() {
784 if (LateCFGStructurize) {
785 addPass(createAMDGPUMachineCFGStructurizerPass());
787 addPass(createSIWholeQuadModePass());
790 void GCNPassConfig::addFastRegAlloc(FunctionPass *RegAllocPass) {
791 // FIXME: We have to disable the verifier here because of PHIElimination +
792 // TwoAddressInstructions disabling it.
794 // This must be run immediately after phi elimination and before
795 // TwoAddressInstructions, otherwise the processing of the tied operand of
796 // SI_ELSE will introduce a copy of the tied operand source after the else.
797 insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
799 TargetPassConfig::addFastRegAlloc(RegAllocPass);
802 void GCNPassConfig::addOptimizedRegAlloc(FunctionPass *RegAllocPass) {
803 // This needs to be run directly before register allocation because earlier
804 // passes might recompute live intervals.
805 insertPass(&MachineSchedulerID, &SIFixControlFlowLiveIntervalsID);
807 // This must be run immediately after phi elimination and before
808 // TwoAddressInstructions, otherwise the processing of the tied operand of
809 // SI_ELSE will introduce a copy of the tied operand source after the else.
810 insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
812 TargetPassConfig::addOptimizedRegAlloc(RegAllocPass);
815 void GCNPassConfig::addPostRegAlloc() {
816 addPass(&SIFixVGPRCopiesID);
817 addPass(&SIOptimizeExecMaskingID);
818 TargetPassConfig::addPostRegAlloc();
821 void GCNPassConfig::addPreSched2() {
824 void GCNPassConfig::addPreEmitPass() {
825 // The hazard recognizer that runs as part of the post-ra scheduler does not
826 // guarantee to be able handle all hazards correctly. This is because if there
827 // are multiple scheduling regions in a basic block, the regions are scheduled
828 // bottom up, so when we begin to schedule a region we don't know what
829 // instructions were emitted directly before it.
831 // Here we add a stand-alone hazard recognizer pass which can handle all
833 addPass(&PostRAHazardRecognizerID);
835 if (EnableSIInsertWaitcntsPass)
836 addPass(createSIInsertWaitcntsPass());
838 addPass(createSIInsertWaitsPass());
839 addPass(createSIShrinkInstructionsPass());
840 addPass(&SIInsertSkipsPassID);
841 addPass(createSIDebuggerInsertNopsPass());
842 addPass(&BranchRelaxationPassID);
845 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) {
846 return new GCNPassConfig(this, PM);