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 "AMDGPUCallLowering.h"
19 #include "AMDGPUTargetObjectFile.h"
20 #include "AMDGPUTargetTransformInfo.h"
21 #include "R600ISelLowering.h"
22 #include "R600InstrInfo.h"
23 #include "R600MachineScheduler.h"
24 #include "SIISelLowering.h"
25 #include "SIInstrInfo.h"
27 #include "llvm/Analysis/Passes.h"
28 #include "llvm/CodeGen/GlobalISel/IRTranslator.h"
29 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
30 #include "llvm/CodeGen/MachineModuleInfo.h"
31 #include "llvm/CodeGen/Passes.h"
32 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
33 #include "llvm/CodeGen/TargetPassConfig.h"
34 #include "llvm/IR/Verifier.h"
35 #include "llvm/MC/MCAsmInfo.h"
36 #include "llvm/IR/LegacyPassManager.h"
37 #include "llvm/Support/TargetRegistry.h"
38 #include "llvm/Support/raw_os_ostream.h"
39 #include "llvm/Transforms/IPO.h"
40 #include "llvm/Transforms/Scalar.h"
41 #include "llvm/Transforms/Scalar/GVN.h"
42 #include "llvm/Transforms/Vectorize.h"
46 static cl::opt<bool> EnableR600StructurizeCFG(
47 "r600-ir-structurize",
48 cl::desc("Use StructurizeCFG IR pass"),
51 static cl::opt<bool> EnableSROA(
53 cl::desc("Run SROA after promote alloca pass"),
57 static cl::opt<bool> EnableR600IfConvert(
59 cl::desc("Use if conversion pass"),
63 // Option to disable vectorizer for tests.
64 static cl::opt<bool> EnableLoadStoreVectorizer(
65 "amdgpu-load-store-vectorizer",
66 cl::desc("Enable load store vectorizer"),
70 extern "C" void LLVMInitializeAMDGPUTarget() {
71 // Register the target
72 RegisterTargetMachine<R600TargetMachine> X(TheAMDGPUTarget);
73 RegisterTargetMachine<GCNTargetMachine> Y(TheGCNTarget);
75 PassRegistry *PR = PassRegistry::getPassRegistry();
76 initializeSILowerI1CopiesPass(*PR);
77 initializeSIFixSGPRCopiesPass(*PR);
78 initializeSIFoldOperandsPass(*PR);
79 initializeSIShrinkInstructionsPass(*PR);
80 initializeSIFixControlFlowLiveIntervalsPass(*PR);
81 initializeSILoadStoreOptimizerPass(*PR);
82 initializeAMDGPUAnnotateKernelFeaturesPass(*PR);
83 initializeAMDGPUAnnotateUniformValuesPass(*PR);
84 initializeAMDGPUPromoteAllocaPass(*PR);
85 initializeAMDGPUCodeGenPreparePass(*PR);
86 initializeSIAnnotateControlFlowPass(*PR);
87 initializeSIDebuggerInsertNopsPass(*PR);
88 initializeSIInsertWaitsPass(*PR);
89 initializeSIWholeQuadModePass(*PR);
90 initializeSILowerControlFlowPass(*PR);
91 initializeSIDebuggerInsertNopsPass(*PR);
94 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
95 return make_unique<AMDGPUTargetObjectFile>();
98 static ScheduleDAGInstrs *createR600MachineScheduler(MachineSchedContext *C) {
99 return new ScheduleDAGMILive(C, make_unique<R600SchedStrategy>());
102 static MachineSchedRegistry
103 R600SchedRegistry("r600", "Run R600's custom scheduler",
104 createR600MachineScheduler);
106 static MachineSchedRegistry
107 SISchedRegistry("si", "Run SI's custom scheduler",
108 createSIMachineScheduler);
110 static StringRef computeDataLayout(const Triple &TT) {
111 if (TT.getArch() == Triple::r600) {
113 return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
114 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64";
117 // 32-bit private, local, and region pointers. 64-bit global, constant and
119 return "e-p:32:32-p1:64:64-p2:64:64-p3:32:32-p4:64:64-p5:32:32"
120 "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
121 "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64";
125 static StringRef getGPUOrDefault(const Triple &TT, StringRef GPU) {
129 // HSA only supports CI+, so change the default GPU to a CI for HSA.
130 if (TT.getArch() == Triple::amdgcn)
131 return (TT.getOS() == Triple::AMDHSA) ? "kaveri" : "tahiti";
136 static Reloc::Model getEffectiveRelocModel(Optional<Reloc::Model> RM) {
137 // The AMDGPU toolchain only supports generating shared objects, so we
138 // must always use PIC.
142 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, const Triple &TT,
143 StringRef CPU, StringRef FS,
144 TargetOptions Options,
145 Optional<Reloc::Model> RM,
147 CodeGenOpt::Level OptLevel)
148 : LLVMTargetMachine(T, computeDataLayout(TT), TT, getGPUOrDefault(TT, CPU),
149 FS, Options, getEffectiveRelocModel(RM), CM, OptLevel),
150 TLOF(createTLOF(getTargetTriple())),
152 setRequiresStructuredCFG(true);
156 AMDGPUTargetMachine::~AMDGPUTargetMachine() { }
158 StringRef AMDGPUTargetMachine::getGPUName(const Function &F) const {
159 Attribute GPUAttr = F.getFnAttribute("target-cpu");
160 return GPUAttr.hasAttribute(Attribute::None) ?
161 getTargetCPU() : GPUAttr.getValueAsString();
164 StringRef AMDGPUTargetMachine::getFeatureString(const Function &F) const {
165 Attribute FSAttr = F.getFnAttribute("target-features");
167 return FSAttr.hasAttribute(Attribute::None) ?
168 getTargetFeatureString() :
169 FSAttr.getValueAsString();
172 //===----------------------------------------------------------------------===//
173 // R600 Target Machine (R600 -> Cayman)
174 //===----------------------------------------------------------------------===//
176 R600TargetMachine::R600TargetMachine(const Target &T, const Triple &TT,
177 StringRef CPU, StringRef FS,
178 TargetOptions Options,
179 Optional<Reloc::Model> RM,
180 CodeModel::Model CM, CodeGenOpt::Level OL)
181 : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {}
183 const R600Subtarget *R600TargetMachine::getSubtargetImpl(
184 const Function &F) const {
185 StringRef GPU = getGPUName(F);
186 StringRef FS = getFeatureString(F);
188 SmallString<128> SubtargetKey(GPU);
189 SubtargetKey.append(FS);
191 auto &I = SubtargetMap[SubtargetKey];
193 // This needs to be done before we create a new subtarget since any
194 // creation will depend on the TM and the code generation flags on the
195 // function that reside in TargetOptions.
196 resetTargetOptions(F);
197 I = llvm::make_unique<R600Subtarget>(TargetTriple, GPU, FS, *this);
203 //===----------------------------------------------------------------------===//
204 // GCN Target Machine (SI+)
205 //===----------------------------------------------------------------------===//
207 #ifdef LLVM_BUILD_GLOBAL_ISEL
209 struct SIGISelActualAccessor : public GISelAccessor {
210 std::unique_ptr<AMDGPUCallLowering> CallLoweringInfo;
211 const AMDGPUCallLowering *getCallLowering() const override {
212 return CallLoweringInfo.get();
215 } // End anonymous namespace.
218 GCNTargetMachine::GCNTargetMachine(const Target &T, const Triple &TT,
219 StringRef CPU, StringRef FS,
220 TargetOptions Options,
221 Optional<Reloc::Model> RM,
222 CodeModel::Model CM, CodeGenOpt::Level OL)
223 : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {}
225 const SISubtarget *GCNTargetMachine::getSubtargetImpl(const Function &F) const {
226 StringRef GPU = getGPUName(F);
227 StringRef FS = getFeatureString(F);
229 SmallString<128> SubtargetKey(GPU);
230 SubtargetKey.append(FS);
232 auto &I = SubtargetMap[SubtargetKey];
234 // This needs to be done before we create a new subtarget since any
235 // creation will depend on the TM and the code generation flags on the
236 // function that reside in TargetOptions.
237 resetTargetOptions(F);
238 I = llvm::make_unique<SISubtarget>(TargetTriple, GPU, FS, *this);
240 #ifndef LLVM_BUILD_GLOBAL_ISEL
241 GISelAccessor *GISel = new GISelAccessor();
243 SIGISelActualAccessor *GISel = new SIGISelActualAccessor();
244 GISel->CallLoweringInfo.reset(
245 new AMDGPUCallLowering(*I->getTargetLowering()));
248 I->setGISelAccessor(*GISel);
254 //===----------------------------------------------------------------------===//
256 //===----------------------------------------------------------------------===//
260 class AMDGPUPassConfig : public TargetPassConfig {
262 AMDGPUPassConfig(TargetMachine *TM, PassManagerBase &PM)
263 : TargetPassConfig(TM, PM) {
265 // Exceptions and StackMaps are not supported, so these passes will never do
267 disablePass(&StackMapLivenessID);
268 disablePass(&FuncletLayoutID);
271 AMDGPUTargetMachine &getAMDGPUTargetMachine() const {
272 return getTM<AMDGPUTargetMachine>();
275 void addEarlyCSEOrGVNPass();
276 void addStraightLineScalarOptimizationPasses();
277 void addIRPasses() override;
278 void addCodeGenPrepare() override;
279 bool addPreISel() override;
280 bool addInstSelector() override;
281 bool addGCPasses() override;
284 class R600PassConfig final : public AMDGPUPassConfig {
286 R600PassConfig(TargetMachine *TM, PassManagerBase &PM)
287 : AMDGPUPassConfig(TM, PM) { }
289 ScheduleDAGInstrs *createMachineScheduler(
290 MachineSchedContext *C) const override {
291 return createR600MachineScheduler(C);
294 bool addPreISel() override;
295 void addPreRegAlloc() override;
296 void addPreSched2() override;
297 void addPreEmitPass() override;
300 class GCNPassConfig final : public AMDGPUPassConfig {
302 GCNPassConfig(TargetMachine *TM, PassManagerBase &PM)
303 : AMDGPUPassConfig(TM, PM) { }
305 GCNTargetMachine &getGCNTargetMachine() const {
306 return getTM<GCNTargetMachine>();
310 createMachineScheduler(MachineSchedContext *C) const override;
312 void addIRPasses() override;
313 bool addPreISel() override;
314 void addMachineSSAOptimization() override;
315 bool addInstSelector() override;
316 #ifdef LLVM_BUILD_GLOBAL_ISEL
317 bool addIRTranslator() override;
318 bool addRegBankSelect() override;
320 void addFastRegAlloc(FunctionPass *RegAllocPass) override;
321 void addOptimizedRegAlloc(FunctionPass *RegAllocPass) override;
322 void addPreRegAlloc() override;
323 void addPreSched2() override;
324 void addPreEmitPass() override;
327 } // End of anonymous namespace
329 TargetIRAnalysis AMDGPUTargetMachine::getTargetIRAnalysis() {
330 return TargetIRAnalysis([this](const Function &F) {
331 return TargetTransformInfo(AMDGPUTTIImpl(this, F));
335 void AMDGPUPassConfig::addEarlyCSEOrGVNPass() {
336 if (getOptLevel() == CodeGenOpt::Aggressive)
337 addPass(createGVNPass());
339 addPass(createEarlyCSEPass());
342 void AMDGPUPassConfig::addStraightLineScalarOptimizationPasses() {
343 addPass(createSeparateConstOffsetFromGEPPass());
344 addPass(createSpeculativeExecutionPass());
345 // ReassociateGEPs exposes more opportunites for SLSR. See
346 // the example in reassociate-geps-and-slsr.ll.
347 addPass(createStraightLineStrengthReducePass());
348 // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or
349 // EarlyCSE can reuse.
350 addEarlyCSEOrGVNPass();
351 // Run NaryReassociate after EarlyCSE/GVN to be more effective.
352 addPass(createNaryReassociatePass());
353 // NaryReassociate on GEPs creates redundant common expressions, so run
354 // EarlyCSE after it.
355 addPass(createEarlyCSEPass());
358 void AMDGPUPassConfig::addIRPasses() {
359 // There is no reason to run these.
360 disablePass(&StackMapLivenessID);
361 disablePass(&FuncletLayoutID);
362 disablePass(&PatchableFunctionID);
364 // Function calls are not supported, so make sure we inline everything.
365 addPass(createAMDGPUAlwaysInlinePass());
366 addPass(createAlwaysInlinerPass());
367 // We need to add the barrier noop pass, otherwise adding the function
368 // inlining pass will cause all of the PassConfigs passes to be run
369 // one function at a time, which means if we have a nodule with two
370 // functions, then we will generate code for the first function
371 // without ever running any passes on the second.
372 addPass(createBarrierNoopPass());
374 // Handle uses of OpenCL image2d_t, image3d_t and sampler_t arguments.
375 addPass(createAMDGPUOpenCLImageTypeLoweringPass());
377 const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine();
378 if (TM.getOptLevel() > CodeGenOpt::None) {
379 addPass(createAMDGPUPromoteAlloca(&TM));
382 addPass(createSROAPass());
385 addStraightLineScalarOptimizationPasses();
387 TargetPassConfig::addIRPasses();
389 // EarlyCSE is not always strong enough to clean up what LSR produces. For
390 // example, GVN can combine
397 // %0 = shl nsw %a, 2
400 // but EarlyCSE can do neither of them.
401 if (getOptLevel() != CodeGenOpt::None)
402 addEarlyCSEOrGVNPass();
405 void AMDGPUPassConfig::addCodeGenPrepare() {
406 TargetPassConfig::addCodeGenPrepare();
408 if (EnableLoadStoreVectorizer)
409 addPass(createLoadStoreVectorizerPass());
412 bool AMDGPUPassConfig::addPreISel() {
413 addPass(createFlattenCFGPass());
417 bool AMDGPUPassConfig::addInstSelector() {
418 addPass(createAMDGPUISelDag(getAMDGPUTargetMachine()));
422 bool AMDGPUPassConfig::addGCPasses() {
423 // Do nothing. GC is not supported.
427 //===----------------------------------------------------------------------===//
429 //===----------------------------------------------------------------------===//
431 bool R600PassConfig::addPreISel() {
432 AMDGPUPassConfig::addPreISel();
434 if (EnableR600StructurizeCFG)
435 addPass(createStructurizeCFGPass());
439 void R600PassConfig::addPreRegAlloc() {
440 addPass(createR600VectorRegMerger(*TM));
443 void R600PassConfig::addPreSched2() {
444 addPass(createR600EmitClauseMarkers(), false);
445 if (EnableR600IfConvert)
446 addPass(&IfConverterID, false);
447 addPass(createR600ClauseMergePass(*TM), false);
450 void R600PassConfig::addPreEmitPass() {
451 addPass(createAMDGPUCFGStructurizerPass(), false);
452 addPass(createR600ExpandSpecialInstrsPass(*TM), false);
453 addPass(&FinalizeMachineBundlesID, false);
454 addPass(createR600Packetizer(*TM), false);
455 addPass(createR600ControlFlowFinalizer(*TM), false);
458 TargetPassConfig *R600TargetMachine::createPassConfig(PassManagerBase &PM) {
459 return new R600PassConfig(this, PM);
462 //===----------------------------------------------------------------------===//
464 //===----------------------------------------------------------------------===//
466 ScheduleDAGInstrs *GCNPassConfig::createMachineScheduler(
467 MachineSchedContext *C) const {
468 const SISubtarget &ST = C->MF->getSubtarget<SISubtarget>();
469 if (ST.enableSIScheduler())
470 return createSIMachineScheduler(C);
474 bool GCNPassConfig::addPreISel() {
475 AMDGPUPassConfig::addPreISel();
477 // FIXME: We need to run a pass to propagate the attributes when calls are
479 addPass(&AMDGPUAnnotateKernelFeaturesID);
480 addPass(createStructurizeCFGPass(true)); // true -> SkipUniformRegions
481 addPass(createSinkingPass());
482 addPass(createSITypeRewriter());
483 addPass(createAMDGPUAnnotateUniformValues());
484 addPass(createSIAnnotateControlFlowPass());
489 void GCNPassConfig::addMachineSSAOptimization() {
490 TargetPassConfig::addMachineSSAOptimization();
492 // We want to fold operands after PeepholeOptimizer has run (or as part of
493 // it), because it will eliminate extra copies making it easier to fold the
494 // real source operand. We want to eliminate dead instructions after, so that
495 // we see fewer uses of the copies. We then need to clean up the dead
496 // instructions leftover after the operands are folded as well.
498 // XXX - Can we get away without running DeadMachineInstructionElim again?
499 addPass(&SIFoldOperandsID);
500 addPass(&DeadMachineInstructionElimID);
503 void GCNPassConfig::addIRPasses() {
504 // TODO: May want to move later or split into an early and late one.
505 addPass(createAMDGPUCodeGenPreparePass(&getGCNTargetMachine()));
507 AMDGPUPassConfig::addIRPasses();
510 bool GCNPassConfig::addInstSelector() {
511 AMDGPUPassConfig::addInstSelector();
512 addPass(createSILowerI1CopiesPass());
513 addPass(&SIFixSGPRCopiesID);
517 #ifdef LLVM_BUILD_GLOBAL_ISEL
518 bool GCNPassConfig::addIRTranslator() {
519 addPass(new IRTranslator());
523 bool GCNPassConfig::addRegBankSelect() {
528 void GCNPassConfig::addPreRegAlloc() {
529 // This needs to be run directly before register allocation because
530 // earlier passes might recompute live intervals.
531 // TODO: handle CodeGenOpt::None; fast RA ignores spill weights set by the pass
532 if (getOptLevel() > CodeGenOpt::None) {
533 insertPass(&MachineSchedulerID, &SIFixControlFlowLiveIntervalsID);
536 if (getOptLevel() > CodeGenOpt::None) {
537 // Don't do this with no optimizations since it throws away debug info by
538 // merging nonadjacent loads.
540 // This should be run after scheduling, but before register allocation. It
541 // also need extra copies to the address operand to be eliminated.
543 // FIXME: Move pre-RA and remove extra reg coalescer run.
544 insertPass(&MachineSchedulerID, &SILoadStoreOptimizerID);
545 insertPass(&MachineSchedulerID, &RegisterCoalescerID);
548 addPass(createSIShrinkInstructionsPass());
549 addPass(createSIWholeQuadModePass());
552 void GCNPassConfig::addFastRegAlloc(FunctionPass *RegAllocPass) {
553 TargetPassConfig::addFastRegAlloc(RegAllocPass);
556 void GCNPassConfig::addOptimizedRegAlloc(FunctionPass *RegAllocPass) {
557 TargetPassConfig::addOptimizedRegAlloc(RegAllocPass);
560 void GCNPassConfig::addPreSched2() {
563 void GCNPassConfig::addPreEmitPass() {
564 // The hazard recognizer that runs as part of the post-ra scheduler does not
565 // guarantee to be able handle all hazards correctly. This is because if there
566 // are multiple scheduling regions in a basic block, the regions are scheduled
567 // bottom up, so when we begin to schedule a region we don't know what
568 // instructions were emitted directly before it.
570 // Here we add a stand-alone hazard recognizer pass which can handle all
572 addPass(&PostRAHazardRecognizerID);
574 addPass(createSIInsertWaitsPass());
575 addPass(createSIShrinkInstructionsPass());
576 addPass(createSILowerControlFlowPass());
577 addPass(createSIDebuggerInsertNopsPass());
580 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) {
581 return new GCNPassConfig(this, PM);