1 //===-- AMDGPUTargetMachine.cpp - TargetMachine for hw codegen targets-----===//
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
7 //===----------------------------------------------------------------------===//
10 /// The AMDGPU target machine contains all of the hardware specific
11 /// information needed to emit code for SI+ GPUs.
13 //===----------------------------------------------------------------------===//
15 #include "AMDGPUTargetMachine.h"
17 #include "AMDGPUAliasAnalysis.h"
18 #include "AMDGPUExportClustering.h"
19 #include "AMDGPUMacroFusion.h"
20 #include "AMDGPUTargetObjectFile.h"
21 #include "AMDGPUTargetTransformInfo.h"
22 #include "GCNIterativeScheduler.h"
23 #include "GCNSchedStrategy.h"
25 #include "R600TargetMachine.h"
26 #include "SIMachineFunctionInfo.h"
27 #include "SIMachineScheduler.h"
28 #include "TargetInfo/AMDGPUTargetInfo.h"
29 #include "llvm/Analysis/CGSCCPassManager.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/Localizer.h"
34 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h"
35 #include "llvm/CodeGen/MIRParser/MIParser.h"
36 #include "llvm/CodeGen/Passes.h"
37 #include "llvm/CodeGen/RegAllocRegistry.h"
38 #include "llvm/CodeGen/TargetPassConfig.h"
39 #include "llvm/IR/IntrinsicsAMDGPU.h"
40 #include "llvm/IR/LegacyPassManager.h"
41 #include "llvm/IR/PassManager.h"
42 #include "llvm/IR/PatternMatch.h"
43 #include "llvm/InitializePasses.h"
44 #include "llvm/MC/TargetRegistry.h"
45 #include "llvm/Passes/PassBuilder.h"
46 #include "llvm/Transforms/IPO.h"
47 #include "llvm/Transforms/IPO/AlwaysInliner.h"
48 #include "llvm/Transforms/IPO/GlobalDCE.h"
49 #include "llvm/Transforms/IPO/Internalize.h"
50 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
51 #include "llvm/Transforms/Scalar.h"
52 #include "llvm/Transforms/Scalar/GVN.h"
53 #include "llvm/Transforms/Scalar/InferAddressSpaces.h"
54 #include "llvm/Transforms/Utils.h"
55 #include "llvm/Transforms/Utils/SimplifyLibCalls.h"
56 #include "llvm/Transforms/Vectorize.h"
61 class SGPRRegisterRegAlloc : public RegisterRegAllocBase<SGPRRegisterRegAlloc> {
63 SGPRRegisterRegAlloc(const char *N, const char *D, FunctionPassCtor C)
64 : RegisterRegAllocBase(N, D, C) {}
67 class VGPRRegisterRegAlloc : public RegisterRegAllocBase<VGPRRegisterRegAlloc> {
69 VGPRRegisterRegAlloc(const char *N, const char *D, FunctionPassCtor C)
70 : RegisterRegAllocBase(N, D, C) {}
73 static bool onlyAllocateSGPRs(const TargetRegisterInfo &TRI,
74 const TargetRegisterClass &RC) {
75 return static_cast<const SIRegisterInfo &>(TRI).isSGPRClass(&RC);
78 static bool onlyAllocateVGPRs(const TargetRegisterInfo &TRI,
79 const TargetRegisterClass &RC) {
80 return !static_cast<const SIRegisterInfo &>(TRI).isSGPRClass(&RC);
84 /// -{sgpr|vgpr}-regalloc=... command line option.
85 static FunctionPass *useDefaultRegisterAllocator() { return nullptr; }
87 /// A dummy default pass factory indicates whether the register allocator is
88 /// overridden on the command line.
89 static llvm::once_flag InitializeDefaultSGPRRegisterAllocatorFlag;
90 static llvm::once_flag InitializeDefaultVGPRRegisterAllocatorFlag;
92 static SGPRRegisterRegAlloc
93 defaultSGPRRegAlloc("default",
94 "pick SGPR register allocator based on -O option",
95 useDefaultRegisterAllocator);
97 static cl::opt<SGPRRegisterRegAlloc::FunctionPassCtor, false,
98 RegisterPassParser<SGPRRegisterRegAlloc>>
99 SGPRRegAlloc("sgpr-regalloc", cl::Hidden, cl::init(&useDefaultRegisterAllocator),
100 cl::desc("Register allocator to use for SGPRs"));
102 static cl::opt<VGPRRegisterRegAlloc::FunctionPassCtor, false,
103 RegisterPassParser<VGPRRegisterRegAlloc>>
104 VGPRRegAlloc("vgpr-regalloc", cl::Hidden, cl::init(&useDefaultRegisterAllocator),
105 cl::desc("Register allocator to use for VGPRs"));
108 static void initializeDefaultSGPRRegisterAllocatorOnce() {
109 RegisterRegAlloc::FunctionPassCtor Ctor = SGPRRegisterRegAlloc::getDefault();
113 SGPRRegisterRegAlloc::setDefault(SGPRRegAlloc);
117 static void initializeDefaultVGPRRegisterAllocatorOnce() {
118 RegisterRegAlloc::FunctionPassCtor Ctor = VGPRRegisterRegAlloc::getDefault();
122 VGPRRegisterRegAlloc::setDefault(VGPRRegAlloc);
126 static FunctionPass *createBasicSGPRRegisterAllocator() {
127 return createBasicRegisterAllocator(onlyAllocateSGPRs);
130 static FunctionPass *createGreedySGPRRegisterAllocator() {
131 return createGreedyRegisterAllocator(onlyAllocateSGPRs);
134 static FunctionPass *createFastSGPRRegisterAllocator() {
135 return createFastRegisterAllocator(onlyAllocateSGPRs, false);
138 static FunctionPass *createBasicVGPRRegisterAllocator() {
139 return createBasicRegisterAllocator(onlyAllocateVGPRs);
142 static FunctionPass *createGreedyVGPRRegisterAllocator() {
143 return createGreedyRegisterAllocator(onlyAllocateVGPRs);
146 static FunctionPass *createFastVGPRRegisterAllocator() {
147 return createFastRegisterAllocator(onlyAllocateVGPRs, true);
150 static SGPRRegisterRegAlloc basicRegAllocSGPR(
151 "basic", "basic register allocator", createBasicSGPRRegisterAllocator);
152 static SGPRRegisterRegAlloc greedyRegAllocSGPR(
153 "greedy", "greedy register allocator", createGreedySGPRRegisterAllocator);
155 static SGPRRegisterRegAlloc fastRegAllocSGPR(
156 "fast", "fast register allocator", createFastSGPRRegisterAllocator);
159 static VGPRRegisterRegAlloc basicRegAllocVGPR(
160 "basic", "basic register allocator", createBasicVGPRRegisterAllocator);
161 static VGPRRegisterRegAlloc greedyRegAllocVGPR(
162 "greedy", "greedy register allocator", createGreedyVGPRRegisterAllocator);
164 static VGPRRegisterRegAlloc fastRegAllocVGPR(
165 "fast", "fast register allocator", createFastVGPRRegisterAllocator);
168 static cl::opt<bool> EnableSROA(
170 cl::desc("Run SROA after promote alloca pass"),
175 EnableEarlyIfConversion("amdgpu-early-ifcvt", cl::Hidden,
176 cl::desc("Run early if-conversion"),
180 OptExecMaskPreRA("amdgpu-opt-exec-mask-pre-ra", cl::Hidden,
181 cl::desc("Run pre-RA exec mask optimizations"),
184 // Option to disable vectorizer for tests.
185 static cl::opt<bool> EnableLoadStoreVectorizer(
186 "amdgpu-load-store-vectorizer",
187 cl::desc("Enable load store vectorizer"),
191 // Option to control global loads scalarization
192 static cl::opt<bool> ScalarizeGlobal(
193 "amdgpu-scalarize-global-loads",
194 cl::desc("Enable global load scalarization"),
198 // Option to run internalize pass.
199 static cl::opt<bool> InternalizeSymbols(
200 "amdgpu-internalize-symbols",
201 cl::desc("Enable elimination of non-kernel functions and unused globals"),
205 // Option to inline all early.
206 static cl::opt<bool> EarlyInlineAll(
207 "amdgpu-early-inline-all",
208 cl::desc("Inline all functions early"),
212 static cl::opt<bool> EnableSDWAPeephole(
213 "amdgpu-sdwa-peephole",
214 cl::desc("Enable SDWA peepholer"),
217 static cl::opt<bool> EnableDPPCombine(
218 "amdgpu-dpp-combine",
219 cl::desc("Enable DPP combiner"),
222 // Enable address space based alias analysis
223 static cl::opt<bool> EnableAMDGPUAliasAnalysis("enable-amdgpu-aa", cl::Hidden,
224 cl::desc("Enable AMDGPU Alias Analysis"),
227 // Option to run late CFG structurizer
228 static cl::opt<bool, true> LateCFGStructurize(
229 "amdgpu-late-structurize",
230 cl::desc("Enable late CFG structurization"),
231 cl::location(AMDGPUTargetMachine::EnableLateStructurizeCFG),
234 // Enable lib calls simplifications
235 static cl::opt<bool> EnableLibCallSimplify(
236 "amdgpu-simplify-libcall",
237 cl::desc("Enable amdgpu library simplifications"),
241 static cl::opt<bool> EnableLowerKernelArguments(
242 "amdgpu-ir-lower-kernel-arguments",
243 cl::desc("Lower kernel argument loads in IR pass"),
247 static cl::opt<bool> EnableRegReassign(
248 "amdgpu-reassign-regs",
249 cl::desc("Enable register reassign optimizations on gfx10+"),
253 static cl::opt<bool> OptVGPRLiveRange(
254 "amdgpu-opt-vgpr-liverange",
255 cl::desc("Enable VGPR liverange optimizations for if-else structure"),
256 cl::init(true), cl::Hidden);
258 // Enable atomic optimization
259 static cl::opt<bool> EnableAtomicOptimizations(
260 "amdgpu-atomic-optimizations",
261 cl::desc("Enable atomic optimizations"),
265 // Enable Mode register optimization
266 static cl::opt<bool> EnableSIModeRegisterPass(
267 "amdgpu-mode-register",
268 cl::desc("Enable mode register pass"),
272 // Option is used in lit tests to prevent deadcoding of patterns inspected.
274 EnableDCEInRA("amdgpu-dce-in-ra",
275 cl::init(true), cl::Hidden,
276 cl::desc("Enable machine DCE inside regalloc"));
278 static cl::opt<bool> EnableScalarIRPasses(
279 "amdgpu-scalar-ir-passes",
280 cl::desc("Enable scalar IR passes"),
284 static cl::opt<bool> EnableStructurizerWorkarounds(
285 "amdgpu-enable-structurizer-workarounds",
286 cl::desc("Enable workarounds for the StructurizeCFG pass"), cl::init(true),
289 static cl::opt<bool> EnableLDSReplaceWithPointer(
290 "amdgpu-enable-lds-replace-with-pointer",
291 cl::desc("Enable LDS replace with pointer pass"), cl::init(false),
294 static cl::opt<bool, true> EnableLowerModuleLDS(
295 "amdgpu-enable-lower-module-lds", cl::desc("Enable lower module lds pass"),
296 cl::location(AMDGPUTargetMachine::EnableLowerModuleLDS), cl::init(true),
299 static cl::opt<bool> EnablePreRAOptimizations(
300 "amdgpu-enable-pre-ra-optimizations",
301 cl::desc("Enable Pre-RA optimizations pass"), cl::init(true),
304 static cl::opt<bool> EnablePromoteKernelArguments(
305 "amdgpu-enable-promote-kernel-arguments",
306 cl::desc("Enable promotion of flat kernel pointer arguments to global"),
307 cl::Hidden, cl::init(true));
309 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUTarget() {
310 // Register the target
311 RegisterTargetMachine<R600TargetMachine> X(getTheAMDGPUTarget());
312 RegisterTargetMachine<GCNTargetMachine> Y(getTheGCNTarget());
314 PassRegistry *PR = PassRegistry::getPassRegistry();
315 initializeR600ClauseMergePassPass(*PR);
316 initializeR600ControlFlowFinalizerPass(*PR);
317 initializeR600PacketizerPass(*PR);
318 initializeR600ExpandSpecialInstrsPassPass(*PR);
319 initializeR600VectorRegMergerPass(*PR);
320 initializeGlobalISel(*PR);
321 initializeAMDGPUDAGToDAGISelPass(*PR);
322 initializeGCNDPPCombinePass(*PR);
323 initializeSILowerI1CopiesPass(*PR);
324 initializeSILowerSGPRSpillsPass(*PR);
325 initializeSIFixSGPRCopiesPass(*PR);
326 initializeSIFixVGPRCopiesPass(*PR);
327 initializeSIFoldOperandsPass(*PR);
328 initializeSIPeepholeSDWAPass(*PR);
329 initializeSIShrinkInstructionsPass(*PR);
330 initializeSIOptimizeExecMaskingPreRAPass(*PR);
331 initializeSIOptimizeVGPRLiveRangePass(*PR);
332 initializeSILoadStoreOptimizerPass(*PR);
333 initializeAMDGPUFixFunctionBitcastsPass(*PR);
334 initializeAMDGPUCtorDtorLoweringPass(*PR);
335 initializeAMDGPUAlwaysInlinePass(*PR);
336 initializeAMDGPUAttributorPass(*PR);
337 initializeAMDGPUAnnotateKernelFeaturesPass(*PR);
338 initializeAMDGPUAnnotateUniformValuesPass(*PR);
339 initializeAMDGPUArgumentUsageInfoPass(*PR);
340 initializeAMDGPUAtomicOptimizerPass(*PR);
341 initializeAMDGPULowerKernelArgumentsPass(*PR);
342 initializeAMDGPUPromoteKernelArgumentsPass(*PR);
343 initializeAMDGPULowerKernelAttributesPass(*PR);
344 initializeAMDGPULowerIntrinsicsPass(*PR);
345 initializeAMDGPUOpenCLEnqueuedBlockLoweringPass(*PR);
346 initializeAMDGPUPostLegalizerCombinerPass(*PR);
347 initializeAMDGPUPreLegalizerCombinerPass(*PR);
348 initializeAMDGPURegBankCombinerPass(*PR);
349 initializeAMDGPUPromoteAllocaPass(*PR);
350 initializeAMDGPUPromoteAllocaToVectorPass(*PR);
351 initializeAMDGPUCodeGenPreparePass(*PR);
352 initializeAMDGPULateCodeGenPreparePass(*PR);
353 initializeAMDGPUPropagateAttributesEarlyPass(*PR);
354 initializeAMDGPUPropagateAttributesLatePass(*PR);
355 initializeAMDGPUReplaceLDSUseWithPointerPass(*PR);
356 initializeAMDGPULowerModuleLDSPass(*PR);
357 initializeAMDGPURewriteOutArgumentsPass(*PR);
358 initializeAMDGPUUnifyMetadataPass(*PR);
359 initializeSIAnnotateControlFlowPass(*PR);
360 initializeSIInsertHardClausesPass(*PR);
361 initializeSIInsertWaitcntsPass(*PR);
362 initializeSIModeRegisterPass(*PR);
363 initializeSIWholeQuadModePass(*PR);
364 initializeSILowerControlFlowPass(*PR);
365 initializeSIPreEmitPeepholePass(*PR);
366 initializeSILateBranchLoweringPass(*PR);
367 initializeSIMemoryLegalizerPass(*PR);
368 initializeSIOptimizeExecMaskingPass(*PR);
369 initializeSIPreAllocateWWMRegsPass(*PR);
370 initializeSIFormMemoryClausesPass(*PR);
371 initializeSIPostRABundlerPass(*PR);
372 initializeAMDGPUUnifyDivergentExitNodesPass(*PR);
373 initializeAMDGPUAAWrapperPassPass(*PR);
374 initializeAMDGPUExternalAAWrapperPass(*PR);
375 initializeAMDGPUUseNativeCallsPass(*PR);
376 initializeAMDGPUSimplifyLibCallsPass(*PR);
377 initializeAMDGPUPrintfRuntimeBindingPass(*PR);
378 initializeAMDGPUResourceUsageAnalysisPass(*PR);
379 initializeGCNNSAReassignPass(*PR);
380 initializeGCNPreRAOptimizationsPass(*PR);
383 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
384 return std::make_unique<AMDGPUTargetObjectFile>();
387 static ScheduleDAGInstrs *createSIMachineScheduler(MachineSchedContext *C) {
388 return new SIScheduleDAGMI(C);
391 static ScheduleDAGInstrs *
392 createGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
393 ScheduleDAGMILive *DAG =
394 new GCNScheduleDAGMILive(C, std::make_unique<GCNMaxOccupancySchedStrategy>(C));
395 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
396 DAG->addMutation(createAMDGPUMacroFusionDAGMutation());
397 DAG->addMutation(createAMDGPUExportClusteringDAGMutation());
401 static ScheduleDAGInstrs *
402 createIterativeGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
403 auto DAG = new GCNIterativeScheduler(C,
404 GCNIterativeScheduler::SCHEDULE_LEGACYMAXOCCUPANCY);
405 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
409 static ScheduleDAGInstrs *createMinRegScheduler(MachineSchedContext *C) {
410 return new GCNIterativeScheduler(C,
411 GCNIterativeScheduler::SCHEDULE_MINREGFORCED);
414 static ScheduleDAGInstrs *
415 createIterativeILPMachineScheduler(MachineSchedContext *C) {
416 auto DAG = new GCNIterativeScheduler(C,
417 GCNIterativeScheduler::SCHEDULE_ILP);
418 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
419 DAG->addMutation(createAMDGPUMacroFusionDAGMutation());
423 static MachineSchedRegistry
424 SISchedRegistry("si", "Run SI's custom scheduler",
425 createSIMachineScheduler);
427 static MachineSchedRegistry
428 GCNMaxOccupancySchedRegistry("gcn-max-occupancy",
429 "Run GCN scheduler to maximize occupancy",
430 createGCNMaxOccupancyMachineScheduler);
432 static MachineSchedRegistry
433 IterativeGCNMaxOccupancySchedRegistry("gcn-max-occupancy-experimental",
434 "Run GCN scheduler to maximize occupancy (experimental)",
435 createIterativeGCNMaxOccupancyMachineScheduler);
437 static MachineSchedRegistry
438 GCNMinRegSchedRegistry("gcn-minreg",
439 "Run GCN iterative scheduler for minimal register usage (experimental)",
440 createMinRegScheduler);
442 static MachineSchedRegistry
443 GCNILPSchedRegistry("gcn-ilp",
444 "Run GCN iterative scheduler for ILP scheduling (experimental)",
445 createIterativeILPMachineScheduler);
447 static StringRef computeDataLayout(const Triple &TT) {
448 if (TT.getArch() == Triple::r600) {
450 return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
451 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1";
454 // 32-bit private, local, and region pointers. 64-bit global, constant and
455 // flat, non-integral buffer fat pointers.
456 return "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32"
457 "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
458 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1"
463 static StringRef getGPUOrDefault(const Triple &TT, StringRef GPU) {
467 // Need to default to a target with flat support for HSA.
468 if (TT.getArch() == Triple::amdgcn)
469 return TT.getOS() == Triple::AMDHSA ? "generic-hsa" : "generic";
474 static Reloc::Model getEffectiveRelocModel(Optional<Reloc::Model> RM) {
475 // The AMDGPU toolchain only supports generating shared objects, so we
476 // must always use PIC.
480 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, const Triple &TT,
481 StringRef CPU, StringRef FS,
482 TargetOptions Options,
483 Optional<Reloc::Model> RM,
484 Optional<CodeModel::Model> CM,
485 CodeGenOpt::Level OptLevel)
486 : LLVMTargetMachine(T, computeDataLayout(TT), TT, getGPUOrDefault(TT, CPU),
487 FS, Options, getEffectiveRelocModel(RM),
488 getEffectiveCodeModel(CM, CodeModel::Small), OptLevel),
489 TLOF(createTLOF(getTargetTriple())) {
491 if (TT.getArch() == Triple::amdgcn) {
492 if (getMCSubtargetInfo()->checkFeatures("+wavefrontsize64"))
493 MRI.reset(llvm::createGCNMCRegisterInfo(AMDGPUDwarfFlavour::Wave64));
494 else if (getMCSubtargetInfo()->checkFeatures("+wavefrontsize32"))
495 MRI.reset(llvm::createGCNMCRegisterInfo(AMDGPUDwarfFlavour::Wave32));
499 bool AMDGPUTargetMachine::EnableLateStructurizeCFG = false;
500 bool AMDGPUTargetMachine::EnableFunctionCalls = false;
501 bool AMDGPUTargetMachine::EnableLowerModuleLDS = true;
503 AMDGPUTargetMachine::~AMDGPUTargetMachine() = default;
505 StringRef AMDGPUTargetMachine::getGPUName(const Function &F) const {
506 Attribute GPUAttr = F.getFnAttribute("target-cpu");
507 return GPUAttr.isValid() ? GPUAttr.getValueAsString() : getTargetCPU();
510 StringRef AMDGPUTargetMachine::getFeatureString(const Function &F) const {
511 Attribute FSAttr = F.getFnAttribute("target-features");
513 return FSAttr.isValid() ? FSAttr.getValueAsString()
514 : getTargetFeatureString();
517 /// Predicate for Internalize pass.
518 static bool mustPreserveGV(const GlobalValue &GV) {
519 if (const Function *F = dyn_cast<Function>(&GV))
520 return F->isDeclaration() || F->getName().startswith("__asan_") ||
521 F->getName().startswith("__sanitizer_") ||
522 AMDGPU::isEntryFunctionCC(F->getCallingConv());
524 GV.removeDeadConstantUsers();
525 return !GV.use_empty();
528 void AMDGPUTargetMachine::adjustPassManager(PassManagerBuilder &Builder) {
529 Builder.DivergentTarget = true;
531 bool EnableOpt = getOptLevel() > CodeGenOpt::None;
532 bool Internalize = InternalizeSymbols;
533 bool EarlyInline = EarlyInlineAll && EnableOpt && !EnableFunctionCalls;
534 bool AMDGPUAA = EnableAMDGPUAliasAnalysis && EnableOpt;
535 bool LibCallSimplify = EnableLibCallSimplify && EnableOpt;
536 bool PromoteKernelArguments =
537 EnablePromoteKernelArguments && getOptLevel() > CodeGenOpt::Less;
539 if (EnableFunctionCalls) {
540 delete Builder.Inliner;
541 Builder.Inliner = createFunctionInliningPass();
544 Builder.addExtension(
545 PassManagerBuilder::EP_ModuleOptimizerEarly,
546 [Internalize, EarlyInline, AMDGPUAA, this](const PassManagerBuilder &,
547 legacy::PassManagerBase &PM) {
549 PM.add(createAMDGPUAAWrapperPass());
550 PM.add(createAMDGPUExternalAAWrapperPass());
552 PM.add(createAMDGPUUnifyMetadataPass());
553 PM.add(createAMDGPUPrintfRuntimeBinding());
555 PM.add(createInternalizePass(mustPreserveGV));
556 PM.add(createAMDGPUPropagateAttributesLatePass(this));
558 PM.add(createGlobalDCEPass());
560 PM.add(createAMDGPUAlwaysInlinePass(false));
563 Builder.addExtension(
564 PassManagerBuilder::EP_EarlyAsPossible,
565 [AMDGPUAA, LibCallSimplify, this](const PassManagerBuilder &,
566 legacy::PassManagerBase &PM) {
568 PM.add(createAMDGPUAAWrapperPass());
569 PM.add(createAMDGPUExternalAAWrapperPass());
571 PM.add(llvm::createAMDGPUPropagateAttributesEarlyPass(this));
572 PM.add(llvm::createAMDGPUUseNativeCallsPass());
574 PM.add(llvm::createAMDGPUSimplifyLibCallsPass(this));
577 Builder.addExtension(
578 PassManagerBuilder::EP_CGSCCOptimizerLate,
579 [EnableOpt, PromoteKernelArguments](const PassManagerBuilder &,
580 legacy::PassManagerBase &PM) {
581 // Add promote kernel arguments pass to the opt pipeline right before
582 // infer address spaces which is needed to do actual address space
584 if (PromoteKernelArguments)
585 PM.add(createAMDGPUPromoteKernelArgumentsPass());
587 // Add infer address spaces pass to the opt pipeline after inlining
588 // but before SROA to increase SROA opportunities.
589 PM.add(createInferAddressSpacesPass());
591 // This should run after inlining to have any chance of doing anything,
592 // and before other cleanup optimizations.
593 PM.add(createAMDGPULowerKernelAttributesPass());
595 // Promote alloca to vector before SROA and loop unroll. If we manage
596 // to eliminate allocas before unroll we may choose to unroll less.
598 PM.add(createAMDGPUPromoteAllocaToVector());
602 void AMDGPUTargetMachine::registerDefaultAliasAnalyses(AAManager &AAM) {
603 AAM.registerFunctionAnalysis<AMDGPUAA>();
606 void AMDGPUTargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) {
607 PB.registerPipelineParsingCallback(
608 [this](StringRef PassName, ModulePassManager &PM,
609 ArrayRef<PassBuilder::PipelineElement>) {
610 if (PassName == "amdgpu-propagate-attributes-late") {
611 PM.addPass(AMDGPUPropagateAttributesLatePass(*this));
614 if (PassName == "amdgpu-unify-metadata") {
615 PM.addPass(AMDGPUUnifyMetadataPass());
618 if (PassName == "amdgpu-printf-runtime-binding") {
619 PM.addPass(AMDGPUPrintfRuntimeBindingPass());
622 if (PassName == "amdgpu-always-inline") {
623 PM.addPass(AMDGPUAlwaysInlinePass());
626 if (PassName == "amdgpu-replace-lds-use-with-pointer") {
627 PM.addPass(AMDGPUReplaceLDSUseWithPointerPass());
630 if (PassName == "amdgpu-lower-module-lds") {
631 PM.addPass(AMDGPULowerModuleLDSPass());
636 PB.registerPipelineParsingCallback(
637 [this](StringRef PassName, FunctionPassManager &PM,
638 ArrayRef<PassBuilder::PipelineElement>) {
639 if (PassName == "amdgpu-simplifylib") {
640 PM.addPass(AMDGPUSimplifyLibCallsPass(*this));
643 if (PassName == "amdgpu-usenative") {
644 PM.addPass(AMDGPUUseNativeCallsPass());
647 if (PassName == "amdgpu-promote-alloca") {
648 PM.addPass(AMDGPUPromoteAllocaPass(*this));
651 if (PassName == "amdgpu-promote-alloca-to-vector") {
652 PM.addPass(AMDGPUPromoteAllocaToVectorPass(*this));
655 if (PassName == "amdgpu-lower-kernel-attributes") {
656 PM.addPass(AMDGPULowerKernelAttributesPass());
659 if (PassName == "amdgpu-propagate-attributes-early") {
660 PM.addPass(AMDGPUPropagateAttributesEarlyPass(*this));
663 if (PassName == "amdgpu-promote-kernel-arguments") {
664 PM.addPass(AMDGPUPromoteKernelArgumentsPass());
670 PB.registerAnalysisRegistrationCallback([](FunctionAnalysisManager &FAM) {
671 FAM.registerPass([&] { return AMDGPUAA(); });
674 PB.registerParseAACallback([](StringRef AAName, AAManager &AAM) {
675 if (AAName == "amdgpu-aa") {
676 AAM.registerFunctionAnalysis<AMDGPUAA>();
682 PB.registerPipelineStartEPCallback(
683 [this](ModulePassManager &PM, OptimizationLevel Level) {
684 FunctionPassManager FPM;
685 FPM.addPass(AMDGPUPropagateAttributesEarlyPass(*this));
686 FPM.addPass(AMDGPUUseNativeCallsPass());
687 if (EnableLibCallSimplify && Level != OptimizationLevel::O0)
688 FPM.addPass(AMDGPUSimplifyLibCallsPass(*this));
689 PM.addPass(createModuleToFunctionPassAdaptor(std::move(FPM)));
692 PB.registerPipelineEarlySimplificationEPCallback(
693 [this](ModulePassManager &PM, OptimizationLevel Level) {
694 if (Level == OptimizationLevel::O0)
697 PM.addPass(AMDGPUUnifyMetadataPass());
698 PM.addPass(AMDGPUPrintfRuntimeBindingPass());
700 if (InternalizeSymbols) {
701 PM.addPass(InternalizePass(mustPreserveGV));
703 PM.addPass(AMDGPUPropagateAttributesLatePass(*this));
704 if (InternalizeSymbols) {
705 PM.addPass(GlobalDCEPass());
707 if (EarlyInlineAll && !EnableFunctionCalls)
708 PM.addPass(AMDGPUAlwaysInlinePass());
711 PB.registerCGSCCOptimizerLateEPCallback(
712 [this](CGSCCPassManager &PM, OptimizationLevel Level) {
713 if (Level == OptimizationLevel::O0)
716 FunctionPassManager FPM;
718 // Add promote kernel arguments pass to the opt pipeline right before
719 // infer address spaces which is needed to do actual address space
721 if (Level.getSpeedupLevel() > OptimizationLevel::O1.getSpeedupLevel() &&
722 EnablePromoteKernelArguments)
723 FPM.addPass(AMDGPUPromoteKernelArgumentsPass());
725 // Add infer address spaces pass to the opt pipeline after inlining
726 // but before SROA to increase SROA opportunities.
727 FPM.addPass(InferAddressSpacesPass());
729 // This should run after inlining to have any chance of doing
730 // anything, and before other cleanup optimizations.
731 FPM.addPass(AMDGPULowerKernelAttributesPass());
733 if (Level != OptimizationLevel::O0) {
734 // Promote alloca to vector before SROA and loop unroll. If we
735 // manage to eliminate allocas before unroll we may choose to unroll
737 FPM.addPass(AMDGPUPromoteAllocaToVectorPass(*this));
740 PM.addPass(createCGSCCToFunctionPassAdaptor(std::move(FPM)));
744 int64_t AMDGPUTargetMachine::getNullPointerValue(unsigned AddrSpace) {
745 return (AddrSpace == AMDGPUAS::LOCAL_ADDRESS ||
746 AddrSpace == AMDGPUAS::PRIVATE_ADDRESS ||
747 AddrSpace == AMDGPUAS::REGION_ADDRESS)
752 bool AMDGPUTargetMachine::isNoopAddrSpaceCast(unsigned SrcAS,
753 unsigned DestAS) const {
754 return AMDGPU::isFlatGlobalAddrSpace(SrcAS) &&
755 AMDGPU::isFlatGlobalAddrSpace(DestAS);
758 unsigned AMDGPUTargetMachine::getAssumedAddrSpace(const Value *V) const {
759 const auto *LD = dyn_cast<LoadInst>(V);
761 return AMDGPUAS::UNKNOWN_ADDRESS_SPACE;
763 // It must be a generic pointer loaded.
764 assert(V->getType()->isPointerTy() &&
765 V->getType()->getPointerAddressSpace() == AMDGPUAS::FLAT_ADDRESS);
767 const auto *Ptr = LD->getPointerOperand();
768 if (Ptr->getType()->getPointerAddressSpace() != AMDGPUAS::CONSTANT_ADDRESS)
769 return AMDGPUAS::UNKNOWN_ADDRESS_SPACE;
770 // For a generic pointer loaded from the constant memory, it could be assumed
771 // as a global pointer since the constant memory is only populated on the
772 // host side. As implied by the offload programming model, only global
773 // pointers could be referenced on the host side.
774 return AMDGPUAS::GLOBAL_ADDRESS;
777 std::pair<const Value *, unsigned>
778 AMDGPUTargetMachine::getPredicatedAddrSpace(const Value *V) const {
779 if (auto *II = dyn_cast<IntrinsicInst>(V)) {
780 switch (II->getIntrinsicID()) {
781 case Intrinsic::amdgcn_is_shared:
782 return std::make_pair(II->getArgOperand(0), AMDGPUAS::LOCAL_ADDRESS);
783 case Intrinsic::amdgcn_is_private:
784 return std::make_pair(II->getArgOperand(0), AMDGPUAS::PRIVATE_ADDRESS);
788 return std::make_pair(nullptr, -1);
790 // Check the global pointer predication based on
791 // (!is_share(p) && !is_private(p)). Note that logic 'and' is commutative and
792 // the order of 'is_shared' and 'is_private' is not significant.
795 const_cast<Value *>(V),
796 m_c_And(m_Not(m_Intrinsic<Intrinsic::amdgcn_is_shared>(m_Value(Ptr))),
797 m_Not(m_Intrinsic<Intrinsic::amdgcn_is_private>(
799 return std::make_pair(Ptr, AMDGPUAS::GLOBAL_ADDRESS);
801 return std::make_pair(nullptr, -1);
804 //===----------------------------------------------------------------------===//
805 // GCN Target Machine (SI+)
806 //===----------------------------------------------------------------------===//
808 GCNTargetMachine::GCNTargetMachine(const Target &T, const Triple &TT,
809 StringRef CPU, StringRef FS,
810 TargetOptions Options,
811 Optional<Reloc::Model> RM,
812 Optional<CodeModel::Model> CM,
813 CodeGenOpt::Level OL, bool JIT)
814 : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {}
816 const TargetSubtargetInfo *
817 GCNTargetMachine::getSubtargetImpl(const Function &F) const {
818 StringRef GPU = getGPUName(F);
819 StringRef FS = getFeatureString(F);
821 SmallString<128> SubtargetKey(GPU);
822 SubtargetKey.append(FS);
824 auto &I = SubtargetMap[SubtargetKey];
826 // This needs to be done before we create a new subtarget since any
827 // creation will depend on the TM and the code generation flags on the
828 // function that reside in TargetOptions.
829 resetTargetOptions(F);
830 I = std::make_unique<GCNSubtarget>(TargetTriple, GPU, FS, *this);
833 I->setScalarizeGlobalBehavior(ScalarizeGlobal);
839 GCNTargetMachine::getTargetTransformInfo(const Function &F) {
840 return TargetTransformInfo(GCNTTIImpl(this, F));
843 //===----------------------------------------------------------------------===//
845 //===----------------------------------------------------------------------===//
847 std::unique_ptr<CSEConfigBase> llvm::AMDGPUPassConfig::getCSEConfig() const {
848 return getStandardCSEConfigForOpt(TM->getOptLevel());
853 class GCNPassConfig final : public AMDGPUPassConfig {
855 GCNPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM)
856 : AMDGPUPassConfig(TM, PM) {
857 // It is necessary to know the register usage of the entire call graph. We
858 // allow calls without EnableAMDGPUFunctionCalls if they are marked
859 // noinline, so this is always required.
860 setRequiresCodeGenSCCOrder(true);
861 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID);
864 GCNTargetMachine &getGCNTargetMachine() const {
865 return getTM<GCNTargetMachine>();
869 createMachineScheduler(MachineSchedContext *C) const override;
872 createPostMachineScheduler(MachineSchedContext *C) const override {
873 ScheduleDAGMI *DAG = createGenericSchedPostRA(C);
874 const GCNSubtarget &ST = C->MF->getSubtarget<GCNSubtarget>();
875 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
876 DAG->addMutation(ST.createFillMFMAShadowMutation(DAG->TII));
880 bool addPreISel() override;
881 void addMachineSSAOptimization() override;
882 bool addILPOpts() override;
883 bool addInstSelector() override;
884 bool addIRTranslator() override;
885 void addPreLegalizeMachineIR() override;
886 bool addLegalizeMachineIR() override;
887 void addPreRegBankSelect() override;
888 bool addRegBankSelect() override;
889 void addPreGlobalInstructionSelect() override;
890 bool addGlobalInstructionSelect() override;
891 void addFastRegAlloc() override;
892 void addOptimizedRegAlloc() override;
894 FunctionPass *createSGPRAllocPass(bool Optimized);
895 FunctionPass *createVGPRAllocPass(bool Optimized);
896 FunctionPass *createRegAllocPass(bool Optimized) override;
898 bool addRegAssignAndRewriteFast() override;
899 bool addRegAssignAndRewriteOptimized() override;
901 void addPreRegAlloc() override;
902 bool addPreRewrite() override;
903 void addPostRegAlloc() override;
904 void addPreSched2() override;
905 void addPreEmitPass() override;
908 } // end anonymous namespace
910 AMDGPUPassConfig::AMDGPUPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM)
911 : TargetPassConfig(TM, PM) {
912 // Exceptions and StackMaps are not supported, so these passes will never do
914 disablePass(&StackMapLivenessID);
915 disablePass(&FuncletLayoutID);
916 // Garbage collection is not supported.
917 disablePass(&GCLoweringID);
918 disablePass(&ShadowStackGCLoweringID);
921 void AMDGPUPassConfig::addEarlyCSEOrGVNPass() {
922 if (getOptLevel() == CodeGenOpt::Aggressive)
923 addPass(createGVNPass());
925 addPass(createEarlyCSEPass());
928 void AMDGPUPassConfig::addStraightLineScalarOptimizationPasses() {
929 addPass(createLICMPass());
930 addPass(createSeparateConstOffsetFromGEPPass());
931 addPass(createSpeculativeExecutionPass());
932 // ReassociateGEPs exposes more opportunities for SLSR. See
933 // the example in reassociate-geps-and-slsr.ll.
934 addPass(createStraightLineStrengthReducePass());
935 // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or
936 // EarlyCSE can reuse.
937 addEarlyCSEOrGVNPass();
938 // Run NaryReassociate after EarlyCSE/GVN to be more effective.
939 addPass(createNaryReassociatePass());
940 // NaryReassociate on GEPs creates redundant common expressions, so run
941 // EarlyCSE after it.
942 addPass(createEarlyCSEPass());
945 void AMDGPUPassConfig::addIRPasses() {
946 const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine();
948 // There is no reason to run these.
949 disablePass(&StackMapLivenessID);
950 disablePass(&FuncletLayoutID);
951 disablePass(&PatchableFunctionID);
953 addPass(createAMDGPUPrintfRuntimeBinding());
954 addPass(createAMDGPUCtorDtorLoweringPass());
956 // This must occur before inlining, as the inliner will not look through
958 addPass(createAMDGPUFixFunctionBitcastsPass());
960 // A call to propagate attributes pass in the backend in case opt was not run.
961 addPass(createAMDGPUPropagateAttributesEarlyPass(&TM));
963 addPass(createAMDGPULowerIntrinsicsPass());
965 // Function calls are not supported, so make sure we inline everything.
966 addPass(createAMDGPUAlwaysInlinePass());
967 addPass(createAlwaysInlinerLegacyPass());
968 // We need to add the barrier noop pass, otherwise adding the function
969 // inlining pass will cause all of the PassConfigs passes to be run
970 // one function at a time, which means if we have a nodule with two
971 // functions, then we will generate code for the first function
972 // without ever running any passes on the second.
973 addPass(createBarrierNoopPass());
975 // Handle uses of OpenCL image2d_t, image3d_t and sampler_t arguments.
976 if (TM.getTargetTriple().getArch() == Triple::r600)
977 addPass(createR600OpenCLImageTypeLoweringPass());
979 // Replace OpenCL enqueued block function pointers with global variables.
980 addPass(createAMDGPUOpenCLEnqueuedBlockLoweringPass());
982 // Can increase LDS used by kernel so runs before PromoteAlloca
983 if (EnableLowerModuleLDS) {
984 // The pass "amdgpu-replace-lds-use-with-pointer" need to be run before the
985 // pass "amdgpu-lower-module-lds", and also it required to be run only if
986 // "amdgpu-lower-module-lds" pass is enabled.
987 if (EnableLDSReplaceWithPointer)
988 addPass(createAMDGPUReplaceLDSUseWithPointerPass());
990 addPass(createAMDGPULowerModuleLDSPass());
993 if (TM.getOptLevel() > CodeGenOpt::None)
994 addPass(createInferAddressSpacesPass());
996 addPass(createAtomicExpandPass());
998 if (TM.getOptLevel() > CodeGenOpt::None) {
999 addPass(createAMDGPUPromoteAlloca());
1002 addPass(createSROAPass());
1003 if (isPassEnabled(EnableScalarIRPasses))
1004 addStraightLineScalarOptimizationPasses();
1006 if (EnableAMDGPUAliasAnalysis) {
1007 addPass(createAMDGPUAAWrapperPass());
1008 addPass(createExternalAAWrapperPass([](Pass &P, Function &,
1010 if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>())
1011 AAR.addAAResult(WrapperPass->getResult());
1015 if (TM.getTargetTriple().getArch() == Triple::amdgcn) {
1016 // TODO: May want to move later or split into an early and late one.
1017 addPass(createAMDGPUCodeGenPreparePass());
1021 TargetPassConfig::addIRPasses();
1023 // EarlyCSE is not always strong enough to clean up what LSR produces. For
1024 // example, GVN can combine
1031 // %0 = shl nsw %a, 2
1034 // but EarlyCSE can do neither of them.
1035 if (isPassEnabled(EnableScalarIRPasses))
1036 addEarlyCSEOrGVNPass();
1039 void AMDGPUPassConfig::addCodeGenPrepare() {
1040 if (TM->getTargetTriple().getArch() == Triple::amdgcn) {
1041 addPass(createAMDGPUAttributorPass());
1043 // FIXME: This pass adds 2 hacky attributes that can be replaced with an
1044 // analysis, and should be removed.
1045 addPass(createAMDGPUAnnotateKernelFeaturesPass());
1048 if (TM->getTargetTriple().getArch() == Triple::amdgcn &&
1049 EnableLowerKernelArguments)
1050 addPass(createAMDGPULowerKernelArgumentsPass());
1052 TargetPassConfig::addCodeGenPrepare();
1054 if (isPassEnabled(EnableLoadStoreVectorizer))
1055 addPass(createLoadStoreVectorizerPass());
1057 // LowerSwitch pass may introduce unreachable blocks that can
1058 // cause unexpected behavior for subsequent passes. Placing it
1059 // here seems better that these blocks would get cleaned up by
1060 // UnreachableBlockElim inserted next in the pass flow.
1061 addPass(createLowerSwitchPass());
1064 bool AMDGPUPassConfig::addPreISel() {
1065 if (TM->getOptLevel() > CodeGenOpt::None)
1066 addPass(createFlattenCFGPass());
1070 bool AMDGPUPassConfig::addInstSelector() {
1071 addPass(createAMDGPUISelDag(&getAMDGPUTargetMachine(), getOptLevel()));
1075 bool AMDGPUPassConfig::addGCPasses() {
1076 // Do nothing. GC is not supported.
1080 llvm::ScheduleDAGInstrs *
1081 AMDGPUPassConfig::createMachineScheduler(MachineSchedContext *C) const {
1082 ScheduleDAGMILive *DAG = createGenericSchedLive(C);
1083 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
1087 //===----------------------------------------------------------------------===//
1089 //===----------------------------------------------------------------------===//
1091 ScheduleDAGInstrs *GCNPassConfig::createMachineScheduler(
1092 MachineSchedContext *C) const {
1093 const GCNSubtarget &ST = C->MF->getSubtarget<GCNSubtarget>();
1094 if (ST.enableSIScheduler())
1095 return createSIMachineScheduler(C);
1096 return createGCNMaxOccupancyMachineScheduler(C);
1099 bool GCNPassConfig::addPreISel() {
1100 AMDGPUPassConfig::addPreISel();
1102 if (TM->getOptLevel() > CodeGenOpt::None)
1103 addPass(createAMDGPULateCodeGenPreparePass());
1105 if (isPassEnabled(EnableAtomicOptimizations, CodeGenOpt::Less)) {
1106 addPass(createAMDGPUAtomicOptimizerPass());
1109 if (TM->getOptLevel() > CodeGenOpt::None)
1110 addPass(createSinkingPass());
1112 // Merge divergent exit nodes. StructurizeCFG won't recognize the multi-exit
1113 // regions formed by them.
1114 addPass(&AMDGPUUnifyDivergentExitNodesID);
1115 if (!LateCFGStructurize) {
1116 if (EnableStructurizerWorkarounds) {
1117 addPass(createFixIrreduciblePass());
1118 addPass(createUnifyLoopExitsPass());
1120 addPass(createStructurizeCFGPass(false)); // true -> SkipUniformRegions
1122 addPass(createAMDGPUAnnotateUniformValues());
1123 if (!LateCFGStructurize) {
1124 addPass(createSIAnnotateControlFlowPass());
1126 addPass(createLCSSAPass());
1128 if (TM->getOptLevel() > CodeGenOpt::Less)
1129 addPass(&AMDGPUPerfHintAnalysisID);
1134 void GCNPassConfig::addMachineSSAOptimization() {
1135 TargetPassConfig::addMachineSSAOptimization();
1137 // We want to fold operands after PeepholeOptimizer has run (or as part of
1138 // it), because it will eliminate extra copies making it easier to fold the
1139 // real source operand. We want to eliminate dead instructions after, so that
1140 // we see fewer uses of the copies. We then need to clean up the dead
1141 // instructions leftover after the operands are folded as well.
1143 // XXX - Can we get away without running DeadMachineInstructionElim again?
1144 addPass(&SIFoldOperandsID);
1145 if (EnableDPPCombine)
1146 addPass(&GCNDPPCombineID);
1147 addPass(&SILoadStoreOptimizerID);
1148 if (isPassEnabled(EnableSDWAPeephole)) {
1149 addPass(&SIPeepholeSDWAID);
1150 addPass(&EarlyMachineLICMID);
1151 addPass(&MachineCSEID);
1152 addPass(&SIFoldOperandsID);
1154 addPass(&DeadMachineInstructionElimID);
1155 addPass(createSIShrinkInstructionsPass());
1158 bool GCNPassConfig::addILPOpts() {
1159 if (EnableEarlyIfConversion)
1160 addPass(&EarlyIfConverterID);
1162 TargetPassConfig::addILPOpts();
1166 bool GCNPassConfig::addInstSelector() {
1167 AMDGPUPassConfig::addInstSelector();
1168 addPass(&SIFixSGPRCopiesID);
1169 addPass(createSILowerI1CopiesPass());
1173 bool GCNPassConfig::addIRTranslator() {
1174 addPass(new IRTranslator(getOptLevel()));
1178 void GCNPassConfig::addPreLegalizeMachineIR() {
1179 bool IsOptNone = getOptLevel() == CodeGenOpt::None;
1180 addPass(createAMDGPUPreLegalizeCombiner(IsOptNone));
1181 addPass(new Localizer());
1184 bool GCNPassConfig::addLegalizeMachineIR() {
1185 addPass(new Legalizer());
1189 void GCNPassConfig::addPreRegBankSelect() {
1190 bool IsOptNone = getOptLevel() == CodeGenOpt::None;
1191 addPass(createAMDGPUPostLegalizeCombiner(IsOptNone));
1194 bool GCNPassConfig::addRegBankSelect() {
1195 addPass(new RegBankSelect());
1199 void GCNPassConfig::addPreGlobalInstructionSelect() {
1200 bool IsOptNone = getOptLevel() == CodeGenOpt::None;
1201 addPass(createAMDGPURegBankCombiner(IsOptNone));
1204 bool GCNPassConfig::addGlobalInstructionSelect() {
1205 addPass(new InstructionSelect(getOptLevel()));
1209 void GCNPassConfig::addPreRegAlloc() {
1210 if (LateCFGStructurize) {
1211 addPass(createAMDGPUMachineCFGStructurizerPass());
1215 void GCNPassConfig::addFastRegAlloc() {
1216 // FIXME: We have to disable the verifier here because of PHIElimination +
1217 // TwoAddressInstructions disabling it.
1219 // This must be run immediately after phi elimination and before
1220 // TwoAddressInstructions, otherwise the processing of the tied operand of
1221 // SI_ELSE will introduce a copy of the tied operand source after the else.
1222 insertPass(&PHIEliminationID, &SILowerControlFlowID);
1224 insertPass(&TwoAddressInstructionPassID, &SIWholeQuadModeID);
1225 insertPass(&TwoAddressInstructionPassID, &SIPreAllocateWWMRegsID);
1227 TargetPassConfig::addFastRegAlloc();
1230 void GCNPassConfig::addOptimizedRegAlloc() {
1231 // Allow the scheduler to run before SIWholeQuadMode inserts exec manipulation
1232 // instructions that cause scheduling barriers.
1233 insertPass(&MachineSchedulerID, &SIWholeQuadModeID);
1234 insertPass(&MachineSchedulerID, &SIPreAllocateWWMRegsID);
1236 if (OptExecMaskPreRA)
1237 insertPass(&MachineSchedulerID, &SIOptimizeExecMaskingPreRAID);
1239 if (isPassEnabled(EnablePreRAOptimizations))
1240 insertPass(&RenameIndependentSubregsID, &GCNPreRAOptimizationsID);
1242 // This is not an essential optimization and it has a noticeable impact on
1243 // compilation time, so we only enable it from O2.
1244 if (TM->getOptLevel() > CodeGenOpt::Less)
1245 insertPass(&MachineSchedulerID, &SIFormMemoryClausesID);
1247 // FIXME: when an instruction has a Killed operand, and the instruction is
1248 // inside a bundle, seems only the BUNDLE instruction appears as the Kills of
1249 // the register in LiveVariables, this would trigger a failure in verifier,
1250 // we should fix it and enable the verifier.
1251 if (OptVGPRLiveRange)
1252 insertPass(&LiveVariablesID, &SIOptimizeVGPRLiveRangeID);
1253 // This must be run immediately after phi elimination and before
1254 // TwoAddressInstructions, otherwise the processing of the tied operand of
1255 // SI_ELSE will introduce a copy of the tied operand source after the else.
1256 insertPass(&PHIEliminationID, &SILowerControlFlowID);
1259 insertPass(&DetectDeadLanesID, &DeadMachineInstructionElimID);
1261 TargetPassConfig::addOptimizedRegAlloc();
1264 bool GCNPassConfig::addPreRewrite() {
1265 if (EnableRegReassign)
1266 addPass(&GCNNSAReassignID);
1270 FunctionPass *GCNPassConfig::createSGPRAllocPass(bool Optimized) {
1271 // Initialize the global default.
1272 llvm::call_once(InitializeDefaultSGPRRegisterAllocatorFlag,
1273 initializeDefaultSGPRRegisterAllocatorOnce);
1275 RegisterRegAlloc::FunctionPassCtor Ctor = SGPRRegisterRegAlloc::getDefault();
1276 if (Ctor != useDefaultRegisterAllocator)
1280 return createGreedyRegisterAllocator(onlyAllocateSGPRs);
1282 return createFastRegisterAllocator(onlyAllocateSGPRs, false);
1285 FunctionPass *GCNPassConfig::createVGPRAllocPass(bool Optimized) {
1286 // Initialize the global default.
1287 llvm::call_once(InitializeDefaultVGPRRegisterAllocatorFlag,
1288 initializeDefaultVGPRRegisterAllocatorOnce);
1290 RegisterRegAlloc::FunctionPassCtor Ctor = VGPRRegisterRegAlloc::getDefault();
1291 if (Ctor != useDefaultRegisterAllocator)
1295 return createGreedyVGPRRegisterAllocator();
1297 return createFastVGPRRegisterAllocator();
1300 FunctionPass *GCNPassConfig::createRegAllocPass(bool Optimized) {
1301 llvm_unreachable("should not be used");
1304 static const char RegAllocOptNotSupportedMessage[] =
1305 "-regalloc not supported with amdgcn. Use -sgpr-regalloc and -vgpr-regalloc";
1307 bool GCNPassConfig::addRegAssignAndRewriteFast() {
1308 if (!usingDefaultRegAlloc())
1309 report_fatal_error(RegAllocOptNotSupportedMessage);
1311 addPass(createSGPRAllocPass(false));
1313 // Equivalent of PEI for SGPRs.
1314 addPass(&SILowerSGPRSpillsID);
1316 addPass(createVGPRAllocPass(false));
1320 bool GCNPassConfig::addRegAssignAndRewriteOptimized() {
1321 if (!usingDefaultRegAlloc())
1322 report_fatal_error(RegAllocOptNotSupportedMessage);
1324 addPass(createSGPRAllocPass(true));
1326 // Commit allocated register changes. This is mostly necessary because too
1327 // many things rely on the use lists of the physical registers, such as the
1328 // verifier. This is only necessary with allocators which use LiveIntervals,
1329 // since FastRegAlloc does the replacements itself.
1330 addPass(createVirtRegRewriter(false));
1332 // Equivalent of PEI for SGPRs.
1333 addPass(&SILowerSGPRSpillsID);
1335 addPass(createVGPRAllocPass(true));
1338 addPass(&VirtRegRewriterID);
1343 void GCNPassConfig::addPostRegAlloc() {
1344 addPass(&SIFixVGPRCopiesID);
1345 if (getOptLevel() > CodeGenOpt::None)
1346 addPass(&SIOptimizeExecMaskingID);
1347 TargetPassConfig::addPostRegAlloc();
1350 void GCNPassConfig::addPreSched2() {
1351 if (TM->getOptLevel() > CodeGenOpt::None)
1352 addPass(createSIShrinkInstructionsPass());
1353 addPass(&SIPostRABundlerID);
1356 void GCNPassConfig::addPreEmitPass() {
1357 addPass(createSIMemoryLegalizerPass());
1358 addPass(createSIInsertWaitcntsPass());
1360 addPass(createSIModeRegisterPass());
1362 if (getOptLevel() > CodeGenOpt::None)
1363 addPass(&SIInsertHardClausesID);
1365 addPass(&SILateBranchLoweringPassID);
1366 if (getOptLevel() > CodeGenOpt::None)
1367 addPass(&SIPreEmitPeepholeID);
1368 // The hazard recognizer that runs as part of the post-ra scheduler does not
1369 // guarantee to be able handle all hazards correctly. This is because if there
1370 // are multiple scheduling regions in a basic block, the regions are scheduled
1371 // bottom up, so when we begin to schedule a region we don't know what
1372 // instructions were emitted directly before it.
1374 // Here we add a stand-alone hazard recognizer pass which can handle all
1376 addPass(&PostRAHazardRecognizerID);
1377 addPass(&BranchRelaxationPassID);
1380 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) {
1381 return new GCNPassConfig(*this, PM);
1384 yaml::MachineFunctionInfo *GCNTargetMachine::createDefaultFuncInfoYAML() const {
1385 return new yaml::SIMachineFunctionInfo();
1388 yaml::MachineFunctionInfo *
1389 GCNTargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const {
1390 const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1391 return new yaml::SIMachineFunctionInfo(
1392 *MFI, *MF.getSubtarget().getRegisterInfo(), MF);
1395 bool GCNTargetMachine::parseMachineFunctionInfo(
1396 const yaml::MachineFunctionInfo &MFI_, PerFunctionMIParsingState &PFS,
1397 SMDiagnostic &Error, SMRange &SourceRange) const {
1398 const yaml::SIMachineFunctionInfo &YamlMFI =
1399 reinterpret_cast<const yaml::SIMachineFunctionInfo &>(MFI_);
1400 MachineFunction &MF = PFS.MF;
1401 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1403 if (MFI->initializeBaseYamlFields(YamlMFI, MF, PFS, Error, SourceRange))
1406 if (MFI->Occupancy == 0) {
1407 // Fixup the subtarget dependent default value.
1408 const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
1409 MFI->Occupancy = ST.computeOccupancy(MF.getFunction(), MFI->getLDSSize());
1412 auto parseRegister = [&](const yaml::StringValue &RegName, Register &RegVal) {
1414 if (parseNamedRegisterReference(PFS, TempReg, RegName.Value, Error)) {
1415 SourceRange = RegName.SourceRange;
1423 auto diagnoseRegisterClass = [&](const yaml::StringValue &RegName) {
1424 // Create a diagnostic for a the register string literal.
1425 const MemoryBuffer &Buffer =
1426 *PFS.SM->getMemoryBuffer(PFS.SM->getMainFileID());
1427 Error = SMDiagnostic(*PFS.SM, SMLoc(), Buffer.getBufferIdentifier(), 1,
1428 RegName.Value.size(), SourceMgr::DK_Error,
1429 "incorrect register class for field", RegName.Value,
1431 SourceRange = RegName.SourceRange;
1435 if (parseRegister(YamlMFI.ScratchRSrcReg, MFI->ScratchRSrcReg) ||
1436 parseRegister(YamlMFI.FrameOffsetReg, MFI->FrameOffsetReg) ||
1437 parseRegister(YamlMFI.StackPtrOffsetReg, MFI->StackPtrOffsetReg))
1440 if (MFI->ScratchRSrcReg != AMDGPU::PRIVATE_RSRC_REG &&
1441 !AMDGPU::SGPR_128RegClass.contains(MFI->ScratchRSrcReg)) {
1442 return diagnoseRegisterClass(YamlMFI.ScratchRSrcReg);
1445 if (MFI->FrameOffsetReg != AMDGPU::FP_REG &&
1446 !AMDGPU::SGPR_32RegClass.contains(MFI->FrameOffsetReg)) {
1447 return diagnoseRegisterClass(YamlMFI.FrameOffsetReg);
1450 if (MFI->StackPtrOffsetReg != AMDGPU::SP_REG &&
1451 !AMDGPU::SGPR_32RegClass.contains(MFI->StackPtrOffsetReg)) {
1452 return diagnoseRegisterClass(YamlMFI.StackPtrOffsetReg);
1455 auto parseAndCheckArgument = [&](const Optional<yaml::SIArgument> &A,
1456 const TargetRegisterClass &RC,
1457 ArgDescriptor &Arg, unsigned UserSGPRs,
1458 unsigned SystemSGPRs) {
1459 // Skip parsing if it's not present.
1463 if (A->IsRegister) {
1465 if (parseNamedRegisterReference(PFS, Reg, A->RegisterName.Value, Error)) {
1466 SourceRange = A->RegisterName.SourceRange;
1469 if (!RC.contains(Reg))
1470 return diagnoseRegisterClass(A->RegisterName);
1471 Arg = ArgDescriptor::createRegister(Reg);
1473 Arg = ArgDescriptor::createStack(A->StackOffset);
1474 // Check and apply the optional mask.
1476 Arg = ArgDescriptor::createArg(Arg, A->Mask.getValue());
1478 MFI->NumUserSGPRs += UserSGPRs;
1479 MFI->NumSystemSGPRs += SystemSGPRs;
1483 if (YamlMFI.ArgInfo &&
1484 (parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentBuffer,
1485 AMDGPU::SGPR_128RegClass,
1486 MFI->ArgInfo.PrivateSegmentBuffer, 4, 0) ||
1487 parseAndCheckArgument(YamlMFI.ArgInfo->DispatchPtr,
1488 AMDGPU::SReg_64RegClass, MFI->ArgInfo.DispatchPtr,
1490 parseAndCheckArgument(YamlMFI.ArgInfo->QueuePtr, AMDGPU::SReg_64RegClass,
1491 MFI->ArgInfo.QueuePtr, 2, 0) ||
1492 parseAndCheckArgument(YamlMFI.ArgInfo->KernargSegmentPtr,
1493 AMDGPU::SReg_64RegClass,
1494 MFI->ArgInfo.KernargSegmentPtr, 2, 0) ||
1495 parseAndCheckArgument(YamlMFI.ArgInfo->DispatchID,
1496 AMDGPU::SReg_64RegClass, MFI->ArgInfo.DispatchID,
1498 parseAndCheckArgument(YamlMFI.ArgInfo->FlatScratchInit,
1499 AMDGPU::SReg_64RegClass,
1500 MFI->ArgInfo.FlatScratchInit, 2, 0) ||
1501 parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentSize,
1502 AMDGPU::SGPR_32RegClass,
1503 MFI->ArgInfo.PrivateSegmentSize, 0, 0) ||
1504 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDX,
1505 AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDX,
1507 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDY,
1508 AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDY,
1510 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDZ,
1511 AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDZ,
1513 parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupInfo,
1514 AMDGPU::SGPR_32RegClass,
1515 MFI->ArgInfo.WorkGroupInfo, 0, 1) ||
1516 parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentWaveByteOffset,
1517 AMDGPU::SGPR_32RegClass,
1518 MFI->ArgInfo.PrivateSegmentWaveByteOffset, 0, 1) ||
1519 parseAndCheckArgument(YamlMFI.ArgInfo->ImplicitArgPtr,
1520 AMDGPU::SReg_64RegClass,
1521 MFI->ArgInfo.ImplicitArgPtr, 0, 0) ||
1522 parseAndCheckArgument(YamlMFI.ArgInfo->ImplicitBufferPtr,
1523 AMDGPU::SReg_64RegClass,
1524 MFI->ArgInfo.ImplicitBufferPtr, 2, 0) ||
1525 parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDX,
1526 AMDGPU::VGPR_32RegClass,
1527 MFI->ArgInfo.WorkItemIDX, 0, 0) ||
1528 parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDY,
1529 AMDGPU::VGPR_32RegClass,
1530 MFI->ArgInfo.WorkItemIDY, 0, 0) ||
1531 parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDZ,
1532 AMDGPU::VGPR_32RegClass,
1533 MFI->ArgInfo.WorkItemIDZ, 0, 0)))
1536 MFI->Mode.IEEE = YamlMFI.Mode.IEEE;
1537 MFI->Mode.DX10Clamp = YamlMFI.Mode.DX10Clamp;
1538 MFI->Mode.FP32InputDenormals = YamlMFI.Mode.FP32InputDenormals;
1539 MFI->Mode.FP32OutputDenormals = YamlMFI.Mode.FP32OutputDenormals;
1540 MFI->Mode.FP64FP16InputDenormals = YamlMFI.Mode.FP64FP16InputDenormals;
1541 MFI->Mode.FP64FP16OutputDenormals = YamlMFI.Mode.FP64FP16OutputDenormals;