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 extern "C" void LLVMInitializeAMDGPUTarget() {
122 // Register the target
123 RegisterTargetMachine<R600TargetMachine> X(getTheAMDGPUTarget());
124 RegisterTargetMachine<GCNTargetMachine> Y(getTheGCNTarget());
126 PassRegistry *PR = PassRegistry::getPassRegistry();
127 initializeSILowerI1CopiesPass(*PR);
128 initializeSIFixSGPRCopiesPass(*PR);
129 initializeSIFixVGPRCopiesPass(*PR);
130 initializeSIFoldOperandsPass(*PR);
131 initializeSIPeepholeSDWAPass(*PR);
132 initializeSIShrinkInstructionsPass(*PR);
133 initializeSIFixControlFlowLiveIntervalsPass(*PR);
134 initializeSILoadStoreOptimizerPass(*PR);
135 initializeAMDGPUAnnotateKernelFeaturesPass(*PR);
136 initializeAMDGPUAnnotateUniformValuesPass(*PR);
137 initializeAMDGPULowerIntrinsicsPass(*PR);
138 initializeAMDGPUPromoteAllocaPass(*PR);
139 initializeAMDGPUCodeGenPreparePass(*PR);
140 initializeAMDGPUUnifyMetadataPass(*PR);
141 initializeSIAnnotateControlFlowPass(*PR);
142 initializeSIInsertWaitsPass(*PR);
143 initializeSIInsertWaitcntsPass(*PR);
144 initializeSIWholeQuadModePass(*PR);
145 initializeSILowerControlFlowPass(*PR);
146 initializeSIInsertSkipsPass(*PR);
147 initializeSIDebuggerInsertNopsPass(*PR);
148 initializeSIOptimizeExecMaskingPass(*PR);
149 initializeAMDGPUUnifyDivergentExitNodesPass(*PR);
150 initializeAMDGPUAAWrapperPassPass(*PR);
153 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
154 return llvm::make_unique<AMDGPUTargetObjectFile>();
157 static ScheduleDAGInstrs *createR600MachineScheduler(MachineSchedContext *C) {
158 return new ScheduleDAGMILive(C, llvm::make_unique<R600SchedStrategy>());
161 static ScheduleDAGInstrs *createSIMachineScheduler(MachineSchedContext *C) {
162 return new SIScheduleDAGMI(C);
165 static ScheduleDAGInstrs *
166 createGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
167 ScheduleDAGMILive *DAG =
168 new GCNScheduleDAGMILive(C, make_unique<GCNMaxOccupancySchedStrategy>(C));
169 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
170 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
174 static ScheduleDAGInstrs *
175 createIterativeGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
176 auto DAG = new GCNIterativeScheduler(C,
177 GCNIterativeScheduler::SCHEDULE_LEGACYMAXOCCUPANCY);
178 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
179 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
183 static ScheduleDAGInstrs *createMinRegScheduler(MachineSchedContext *C) {
184 return new GCNIterativeScheduler(C,
185 GCNIterativeScheduler::SCHEDULE_MINREGFORCED);
188 static MachineSchedRegistry
189 R600SchedRegistry("r600", "Run R600's custom scheduler",
190 createR600MachineScheduler);
192 static MachineSchedRegistry
193 SISchedRegistry("si", "Run SI's custom scheduler",
194 createSIMachineScheduler);
196 static MachineSchedRegistry
197 GCNMaxOccupancySchedRegistry("gcn-max-occupancy",
198 "Run GCN scheduler to maximize occupancy",
199 createGCNMaxOccupancyMachineScheduler);
201 static MachineSchedRegistry
202 IterativeGCNMaxOccupancySchedRegistry("gcn-max-occupancy-experimental",
203 "Run GCN scheduler to maximize occupancy (experimental)",
204 createIterativeGCNMaxOccupancyMachineScheduler);
206 static MachineSchedRegistry
207 GCNMinRegSchedRegistry("gcn-minreg",
208 "Run GCN iterative scheduler for minimal register usage (experimental)",
209 createMinRegScheduler);
211 static StringRef computeDataLayout(const Triple &TT) {
212 if (TT.getArch() == Triple::r600) {
214 return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
215 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64";
218 // 32-bit private, local, and region pointers. 64-bit global, constant and
220 if (TT.getEnvironmentName() == "amdgiz" ||
221 TT.getEnvironmentName() == "amdgizcl")
222 return "e-p:64:64-p1:64:64-p2:64:64-p3:32:32-p4:32:32-p5:32:32"
223 "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
224 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-A5";
225 return "e-p:32:32-p1:64:64-p2:64:64-p3:32:32-p4:64:64-p5:32:32"
226 "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
227 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64";
231 static StringRef getGPUOrDefault(const Triple &TT, StringRef GPU) {
235 // HSA only supports CI+, so change the default GPU to a CI for HSA.
236 if (TT.getArch() == Triple::amdgcn)
237 return (TT.getOS() == Triple::AMDHSA) ? "kaveri" : "tahiti";
242 static Reloc::Model getEffectiveRelocModel(Optional<Reloc::Model> RM) {
243 // The AMDGPU toolchain only supports generating shared objects, so we
244 // must always use PIC.
248 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, const Triple &TT,
249 StringRef CPU, StringRef FS,
250 TargetOptions Options,
251 Optional<Reloc::Model> RM,
253 CodeGenOpt::Level OptLevel)
254 : LLVMTargetMachine(T, computeDataLayout(TT), TT, getGPUOrDefault(TT, CPU),
255 FS, Options, getEffectiveRelocModel(RM), CM, OptLevel),
256 TLOF(createTLOF(getTargetTriple())) {
257 AS = AMDGPU::getAMDGPUAS(TT);
261 AMDGPUTargetMachine::~AMDGPUTargetMachine() = default;
263 StringRef AMDGPUTargetMachine::getGPUName(const Function &F) const {
264 Attribute GPUAttr = F.getFnAttribute("target-cpu");
265 return GPUAttr.hasAttribute(Attribute::None) ?
266 getTargetCPU() : GPUAttr.getValueAsString();
269 StringRef AMDGPUTargetMachine::getFeatureString(const Function &F) const {
270 Attribute FSAttr = F.getFnAttribute("target-features");
272 return FSAttr.hasAttribute(Attribute::None) ?
273 getTargetFeatureString() :
274 FSAttr.getValueAsString();
277 static ImmutablePass *createAMDGPUExternalAAWrapperPass() {
278 return createExternalAAWrapperPass([](Pass &P, Function &, AAResults &AAR) {
279 if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>())
280 AAR.addAAResult(WrapperPass->getResult());
284 void AMDGPUTargetMachine::adjustPassManager(PassManagerBuilder &Builder) {
285 Builder.DivergentTarget = true;
287 bool Internalize = InternalizeSymbols &&
288 (getOptLevel() > CodeGenOpt::None) &&
289 (getTargetTriple().getArch() == Triple::amdgcn);
290 bool EarlyInline = EarlyInlineAll &&
291 (getOptLevel() > CodeGenOpt::None);
292 bool AMDGPUAA = EnableAMDGPUAliasAnalysis && getOptLevel() > CodeGenOpt::None;
294 Builder.addExtension(
295 PassManagerBuilder::EP_ModuleOptimizerEarly,
296 [Internalize, EarlyInline, AMDGPUAA](const PassManagerBuilder &,
297 legacy::PassManagerBase &PM) {
299 PM.add(createAMDGPUAAWrapperPass());
300 PM.add(createAMDGPUExternalAAWrapperPass());
302 PM.add(createAMDGPUUnifyMetadataPass());
304 PM.add(createInternalizePass([=](const GlobalValue &GV) -> bool {
305 if (const Function *F = dyn_cast<Function>(&GV)) {
306 if (F->isDeclaration())
308 switch (F->getCallingConv()) {
311 case CallingConv::AMDGPU_VS:
312 case CallingConv::AMDGPU_GS:
313 case CallingConv::AMDGPU_PS:
314 case CallingConv::AMDGPU_CS:
315 case CallingConv::AMDGPU_KERNEL:
316 case CallingConv::SPIR_KERNEL:
320 return !GV.use_empty();
322 PM.add(createGlobalDCEPass());
325 PM.add(createAMDGPUAlwaysInlinePass(false));
328 Builder.addExtension(
329 PassManagerBuilder::EP_EarlyAsPossible,
330 [AMDGPUAA](const PassManagerBuilder &, legacy::PassManagerBase &PM) {
332 PM.add(createAMDGPUAAWrapperPass());
333 PM.add(createAMDGPUExternalAAWrapperPass());
338 //===----------------------------------------------------------------------===//
339 // R600 Target Machine (R600 -> Cayman)
340 //===----------------------------------------------------------------------===//
342 R600TargetMachine::R600TargetMachine(const Target &T, const Triple &TT,
343 StringRef CPU, StringRef FS,
344 TargetOptions Options,
345 Optional<Reloc::Model> RM,
346 CodeModel::Model CM, CodeGenOpt::Level OL)
347 : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {
348 setRequiresStructuredCFG(true);
351 const R600Subtarget *R600TargetMachine::getSubtargetImpl(
352 const Function &F) const {
353 StringRef GPU = getGPUName(F);
354 StringRef FS = getFeatureString(F);
356 SmallString<128> SubtargetKey(GPU);
357 SubtargetKey.append(FS);
359 auto &I = SubtargetMap[SubtargetKey];
361 // This needs to be done before we create a new subtarget since any
362 // creation will depend on the TM and the code generation flags on the
363 // function that reside in TargetOptions.
364 resetTargetOptions(F);
365 I = llvm::make_unique<R600Subtarget>(TargetTriple, GPU, FS, *this);
371 //===----------------------------------------------------------------------===//
372 // GCN Target Machine (SI+)
373 //===----------------------------------------------------------------------===//
375 #ifdef LLVM_BUILD_GLOBAL_ISEL
378 struct SIGISelActualAccessor : public GISelAccessor {
379 std::unique_ptr<AMDGPUCallLowering> CallLoweringInfo;
380 std::unique_ptr<InstructionSelector> InstSelector;
381 std::unique_ptr<LegalizerInfo> Legalizer;
382 std::unique_ptr<RegisterBankInfo> RegBankInfo;
383 const AMDGPUCallLowering *getCallLowering() const override {
384 return CallLoweringInfo.get();
386 const InstructionSelector *getInstructionSelector() const override {
387 return InstSelector.get();
389 const LegalizerInfo *getLegalizerInfo() const override {
390 return Legalizer.get();
392 const RegisterBankInfo *getRegBankInfo() const override {
393 return RegBankInfo.get();
397 } // end anonymous namespace
400 GCNTargetMachine::GCNTargetMachine(const Target &T, const Triple &TT,
401 StringRef CPU, StringRef FS,
402 TargetOptions Options,
403 Optional<Reloc::Model> RM,
404 CodeModel::Model CM, CodeGenOpt::Level OL)
405 : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {}
407 const SISubtarget *GCNTargetMachine::getSubtargetImpl(const Function &F) const {
408 StringRef GPU = getGPUName(F);
409 StringRef FS = getFeatureString(F);
411 SmallString<128> SubtargetKey(GPU);
412 SubtargetKey.append(FS);
414 auto &I = SubtargetMap[SubtargetKey];
416 // This needs to be done before we create a new subtarget since any
417 // creation will depend on the TM and the code generation flags on the
418 // function that reside in TargetOptions.
419 resetTargetOptions(F);
420 I = llvm::make_unique<SISubtarget>(TargetTriple, GPU, FS, *this);
422 #ifndef LLVM_BUILD_GLOBAL_ISEL
423 GISelAccessor *GISel = new GISelAccessor();
425 SIGISelActualAccessor *GISel = new SIGISelActualAccessor();
426 GISel->CallLoweringInfo.reset(
427 new AMDGPUCallLowering(*I->getTargetLowering()));
428 GISel->Legalizer.reset(new AMDGPULegalizerInfo());
430 GISel->RegBankInfo.reset(new AMDGPURegisterBankInfo(*I->getRegisterInfo()));
431 GISel->InstSelector.reset(new AMDGPUInstructionSelector(*I,
432 *static_cast<AMDGPURegisterBankInfo*>(GISel->RegBankInfo.get())));
435 I->setGISelAccessor(*GISel);
438 I->setScalarizeGlobalBehavior(ScalarizeGlobal);
443 //===----------------------------------------------------------------------===//
445 //===----------------------------------------------------------------------===//
449 class AMDGPUPassConfig : public TargetPassConfig {
451 AMDGPUPassConfig(TargetMachine *TM, PassManagerBase &PM)
452 : TargetPassConfig(TM, PM) {
453 // Exceptions and StackMaps are not supported, so these passes will never do
455 disablePass(&StackMapLivenessID);
456 disablePass(&FuncletLayoutID);
459 AMDGPUTargetMachine &getAMDGPUTargetMachine() const {
460 return getTM<AMDGPUTargetMachine>();
464 createMachineScheduler(MachineSchedContext *C) const override {
465 ScheduleDAGMILive *DAG = createGenericSchedLive(C);
466 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
467 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
471 void addEarlyCSEOrGVNPass();
472 void addStraightLineScalarOptimizationPasses();
473 void addIRPasses() override;
474 void addCodeGenPrepare() override;
475 bool addPreISel() override;
476 bool addInstSelector() override;
477 bool addGCPasses() override;
480 class R600PassConfig final : public AMDGPUPassConfig {
482 R600PassConfig(TargetMachine *TM, PassManagerBase &PM)
483 : AMDGPUPassConfig(TM, PM) {}
485 ScheduleDAGInstrs *createMachineScheduler(
486 MachineSchedContext *C) const override {
487 return createR600MachineScheduler(C);
490 bool addPreISel() override;
491 void addPreRegAlloc() override;
492 void addPreSched2() override;
493 void addPreEmitPass() override;
496 class GCNPassConfig final : public AMDGPUPassConfig {
498 GCNPassConfig(TargetMachine *TM, PassManagerBase &PM)
499 : AMDGPUPassConfig(TM, PM) {}
501 GCNTargetMachine &getGCNTargetMachine() const {
502 return getTM<GCNTargetMachine>();
506 createMachineScheduler(MachineSchedContext *C) const override;
508 bool addPreISel() override;
509 void addMachineSSAOptimization() override;
510 bool addILPOpts() override;
511 bool addInstSelector() override;
512 #ifdef LLVM_BUILD_GLOBAL_ISEL
513 bool addIRTranslator() override;
514 bool addLegalizeMachineIR() override;
515 bool addRegBankSelect() override;
516 bool addGlobalInstructionSelect() override;
518 void addFastRegAlloc(FunctionPass *RegAllocPass) override;
519 void addOptimizedRegAlloc(FunctionPass *RegAllocPass) override;
520 void addPreRegAlloc() override;
521 void addPostRegAlloc() override;
522 void addPreSched2() override;
523 void addPreEmitPass() override;
526 } // end anonymous namespace
528 TargetIRAnalysis AMDGPUTargetMachine::getTargetIRAnalysis() {
529 return TargetIRAnalysis([this](const Function &F) {
530 return TargetTransformInfo(AMDGPUTTIImpl(this, F));
534 void AMDGPUPassConfig::addEarlyCSEOrGVNPass() {
535 if (getOptLevel() == CodeGenOpt::Aggressive)
536 addPass(createGVNPass());
538 addPass(createEarlyCSEPass());
541 void AMDGPUPassConfig::addStraightLineScalarOptimizationPasses() {
542 addPass(createSeparateConstOffsetFromGEPPass());
543 addPass(createSpeculativeExecutionPass());
544 // ReassociateGEPs exposes more opportunites for SLSR. See
545 // the example in reassociate-geps-and-slsr.ll.
546 addPass(createStraightLineStrengthReducePass());
547 // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or
548 // EarlyCSE can reuse.
549 addEarlyCSEOrGVNPass();
550 // Run NaryReassociate after EarlyCSE/GVN to be more effective.
551 addPass(createNaryReassociatePass());
552 // NaryReassociate on GEPs creates redundant common expressions, so run
553 // EarlyCSE after it.
554 addPass(createEarlyCSEPass());
557 void AMDGPUPassConfig::addIRPasses() {
558 const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine();
560 // There is no reason to run these.
561 disablePass(&StackMapLivenessID);
562 disablePass(&FuncletLayoutID);
563 disablePass(&PatchableFunctionID);
565 addPass(createAMDGPULowerIntrinsicsPass(&TM));
567 // Function calls are not supported, so make sure we inline everything.
568 addPass(createAMDGPUAlwaysInlinePass());
569 addPass(createAlwaysInlinerLegacyPass());
570 // We need to add the barrier noop pass, otherwise adding the function
571 // inlining pass will cause all of the PassConfigs passes to be run
572 // one function at a time, which means if we have a nodule with two
573 // functions, then we will generate code for the first function
574 // without ever running any passes on the second.
575 addPass(createBarrierNoopPass());
577 if (TM.getTargetTriple().getArch() == Triple::amdgcn) {
578 // TODO: May want to move later or split into an early and late one.
580 addPass(createAMDGPUCodeGenPreparePass(
581 static_cast<const GCNTargetMachine *>(&TM)));
584 // Handle uses of OpenCL image2d_t, image3d_t and sampler_t arguments.
585 addPass(createAMDGPUOpenCLImageTypeLoweringPass());
587 if (TM.getOptLevel() > CodeGenOpt::None) {
588 addPass(createInferAddressSpacesPass());
589 addPass(createAMDGPUPromoteAlloca(&TM));
592 addPass(createSROAPass());
594 addStraightLineScalarOptimizationPasses();
596 if (EnableAMDGPUAliasAnalysis) {
597 addPass(createAMDGPUAAWrapperPass());
598 addPass(createExternalAAWrapperPass([](Pass &P, Function &,
600 if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>())
601 AAR.addAAResult(WrapperPass->getResult());
606 TargetPassConfig::addIRPasses();
608 // EarlyCSE is not always strong enough to clean up what LSR produces. For
609 // example, GVN can combine
616 // %0 = shl nsw %a, 2
619 // but EarlyCSE can do neither of them.
620 if (getOptLevel() != CodeGenOpt::None)
621 addEarlyCSEOrGVNPass();
624 void AMDGPUPassConfig::addCodeGenPrepare() {
625 TargetPassConfig::addCodeGenPrepare();
627 if (EnableLoadStoreVectorizer)
628 addPass(createLoadStoreVectorizerPass());
631 bool AMDGPUPassConfig::addPreISel() {
632 addPass(createFlattenCFGPass());
636 bool AMDGPUPassConfig::addInstSelector() {
637 addPass(createAMDGPUISelDag(getAMDGPUTargetMachine(), getOptLevel()));
641 bool AMDGPUPassConfig::addGCPasses() {
642 // Do nothing. GC is not supported.
646 //===----------------------------------------------------------------------===//
648 //===----------------------------------------------------------------------===//
650 bool R600PassConfig::addPreISel() {
651 AMDGPUPassConfig::addPreISel();
653 if (EnableR600StructurizeCFG)
654 addPass(createStructurizeCFGPass());
658 void R600PassConfig::addPreRegAlloc() {
659 addPass(createR600VectorRegMerger(*TM));
662 void R600PassConfig::addPreSched2() {
663 addPass(createR600EmitClauseMarkers(), false);
664 if (EnableR600IfConvert)
665 addPass(&IfConverterID, false);
666 addPass(createR600ClauseMergePass(*TM), false);
669 void R600PassConfig::addPreEmitPass() {
670 addPass(createAMDGPUCFGStructurizerPass(), false);
671 addPass(createR600ExpandSpecialInstrsPass(*TM), false);
672 addPass(&FinalizeMachineBundlesID, false);
673 addPass(createR600Packetizer(*TM), false);
674 addPass(createR600ControlFlowFinalizer(*TM), false);
677 TargetPassConfig *R600TargetMachine::createPassConfig(PassManagerBase &PM) {
678 return new R600PassConfig(this, PM);
681 //===----------------------------------------------------------------------===//
683 //===----------------------------------------------------------------------===//
685 ScheduleDAGInstrs *GCNPassConfig::createMachineScheduler(
686 MachineSchedContext *C) const {
687 const SISubtarget &ST = C->MF->getSubtarget<SISubtarget>();
688 if (ST.enableSIScheduler())
689 return createSIMachineScheduler(C);
690 return createGCNMaxOccupancyMachineScheduler(C);
693 bool GCNPassConfig::addPreISel() {
694 AMDGPUPassConfig::addPreISel();
696 // FIXME: We need to run a pass to propagate the attributes when calls are
698 const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine();
699 addPass(createAMDGPUAnnotateKernelFeaturesPass(&TM));
701 // Merge divergent exit nodes. StructurizeCFG won't recognize the multi-exit
702 // regions formed by them.
703 addPass(&AMDGPUUnifyDivergentExitNodesID);
704 addPass(createStructurizeCFGPass(true)); // true -> SkipUniformRegions
705 addPass(createSinkingPass());
706 addPass(createSITypeRewriter());
707 addPass(createAMDGPUAnnotateUniformValues());
708 addPass(createSIAnnotateControlFlowPass());
713 void GCNPassConfig::addMachineSSAOptimization() {
714 TargetPassConfig::addMachineSSAOptimization();
716 // We want to fold operands after PeepholeOptimizer has run (or as part of
717 // it), because it will eliminate extra copies making it easier to fold the
718 // real source operand. We want to eliminate dead instructions after, so that
719 // we see fewer uses of the copies. We then need to clean up the dead
720 // instructions leftover after the operands are folded as well.
722 // XXX - Can we get away without running DeadMachineInstructionElim again?
723 addPass(&SIFoldOperandsID);
724 addPass(&DeadMachineInstructionElimID);
725 addPass(&SILoadStoreOptimizerID);
726 addPass(createSIShrinkInstructionsPass());
727 if (EnableSDWAPeephole) {
728 addPass(&SIPeepholeSDWAID);
729 addPass(&DeadMachineInstructionElimID);
733 bool GCNPassConfig::addILPOpts() {
734 if (EnableEarlyIfConversion)
735 addPass(&EarlyIfConverterID);
737 TargetPassConfig::addILPOpts();
741 bool GCNPassConfig::addInstSelector() {
742 AMDGPUPassConfig::addInstSelector();
743 addPass(createSILowerI1CopiesPass());
744 addPass(&SIFixSGPRCopiesID);
748 #ifdef LLVM_BUILD_GLOBAL_ISEL
749 bool GCNPassConfig::addIRTranslator() {
750 addPass(new IRTranslator());
754 bool GCNPassConfig::addLegalizeMachineIR() {
755 addPass(new Legalizer());
759 bool GCNPassConfig::addRegBankSelect() {
760 addPass(new RegBankSelect());
764 bool GCNPassConfig::addGlobalInstructionSelect() {
765 addPass(new InstructionSelect());
771 void GCNPassConfig::addPreRegAlloc() {
772 addPass(createSIWholeQuadModePass());
775 void GCNPassConfig::addFastRegAlloc(FunctionPass *RegAllocPass) {
776 // FIXME: We have to disable the verifier here because of PHIElimination +
777 // TwoAddressInstructions disabling it.
779 // This must be run immediately after phi elimination and before
780 // TwoAddressInstructions, otherwise the processing of the tied operand of
781 // SI_ELSE will introduce a copy of the tied operand source after the else.
782 insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
784 TargetPassConfig::addFastRegAlloc(RegAllocPass);
787 void GCNPassConfig::addOptimizedRegAlloc(FunctionPass *RegAllocPass) {
788 // This needs to be run directly before register allocation because earlier
789 // passes might recompute live intervals.
790 insertPass(&MachineSchedulerID, &SIFixControlFlowLiveIntervalsID);
792 // This must be run immediately after phi elimination and before
793 // TwoAddressInstructions, otherwise the processing of the tied operand of
794 // SI_ELSE will introduce a copy of the tied operand source after the else.
795 insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
797 TargetPassConfig::addOptimizedRegAlloc(RegAllocPass);
800 void GCNPassConfig::addPostRegAlloc() {
801 addPass(&SIFixVGPRCopiesID);
802 addPass(&SIOptimizeExecMaskingID);
803 TargetPassConfig::addPostRegAlloc();
806 void GCNPassConfig::addPreSched2() {
809 void GCNPassConfig::addPreEmitPass() {
810 // The hazard recognizer that runs as part of the post-ra scheduler does not
811 // guarantee to be able handle all hazards correctly. This is because if there
812 // are multiple scheduling regions in a basic block, the regions are scheduled
813 // bottom up, so when we begin to schedule a region we don't know what
814 // instructions were emitted directly before it.
816 // Here we add a stand-alone hazard recognizer pass which can handle all
818 addPass(&PostRAHazardRecognizerID);
820 if (EnableSIInsertWaitcntsPass)
821 addPass(createSIInsertWaitcntsPass());
823 addPass(createSIInsertWaitsPass());
824 addPass(createSIShrinkInstructionsPass());
825 addPass(&SIInsertSkipsPassID);
826 addPass(createSIDebuggerInsertNopsPass());
827 addPass(&BranchRelaxationPassID);
830 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) {
831 return new GCNPassConfig(this, PM);