1 //===-- AMDGPUTargetMachine.cpp - TargetMachine for hw codegen targets-----===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file
11 /// \brief The AMDGPU target machine contains all of the hardware specific
12 /// information  needed to emit code for R600 and SI GPUs.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "AMDGPUTargetMachine.h"
17 #include "AMDGPU.h"
18 #include "AMDGPUAliasAnalysis.h"
19 #include "AMDGPUCallLowering.h"
20 #include "AMDGPUInstructionSelector.h"
21 #include "AMDGPULegalizerInfo.h"
22 #include "AMDGPUMacroFusion.h"
23 #include "AMDGPUTargetObjectFile.h"
24 #include "AMDGPUTargetTransformInfo.h"
25 #include "GCNIterativeScheduler.h"
26 #include "GCNSchedStrategy.h"
27 #include "R600MachineScheduler.h"
28 #include "SIMachineScheduler.h"
29 #include "llvm/CodeGen/GlobalISel/IRTranslator.h"
30 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h"
31 #include "llvm/CodeGen/GlobalISel/Legalizer.h"
32 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h"
33 #include "llvm/CodeGen/Passes.h"
34 #include "llvm/CodeGen/TargetLoweringObjectFile.h"
35 #include "llvm/CodeGen/TargetPassConfig.h"
36 #include "llvm/IR/Attributes.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/LegacyPassManager.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Compiler.h"
42 #include "llvm/Support/TargetRegistry.h"
43 #include "llvm/Transforms/IPO.h"
44 #include "llvm/Transforms/IPO/AlwaysInliner.h"
45 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
46 #include "llvm/Transforms/Scalar.h"
47 #include "llvm/Transforms/Scalar/GVN.h"
48 #include "llvm/Transforms/Vectorize.h"
49 #include <memory>
50 
51 using namespace llvm;
52 
53 static cl::opt<bool> EnableR600StructurizeCFG(
54   "r600-ir-structurize",
55   cl::desc("Use StructurizeCFG IR pass"),
56   cl::init(true));
57 
58 static cl::opt<bool> EnableSROA(
59   "amdgpu-sroa",
60   cl::desc("Run SROA after promote alloca pass"),
61   cl::ReallyHidden,
62   cl::init(true));
63 
64 static cl::opt<bool>
65 EnableEarlyIfConversion("amdgpu-early-ifcvt", cl::Hidden,
66                         cl::desc("Run early if-conversion"),
67                         cl::init(false));
68 
69 static cl::opt<bool> EnableR600IfConvert(
70   "r600-if-convert",
71   cl::desc("Use if conversion pass"),
72   cl::ReallyHidden,
73   cl::init(true));
74 
75 // Option to disable vectorizer for tests.
76 static cl::opt<bool> EnableLoadStoreVectorizer(
77   "amdgpu-load-store-vectorizer",
78   cl::desc("Enable load store vectorizer"),
79   cl::init(true),
80   cl::Hidden);
81 
82 // Option to control global loads scalarization
83 static cl::opt<bool> ScalarizeGlobal(
84   "amdgpu-scalarize-global-loads",
85   cl::desc("Enable global load scalarization"),
86   cl::init(true),
87   cl::Hidden);
88 
89 // Option to run internalize pass.
90 static cl::opt<bool> InternalizeSymbols(
91   "amdgpu-internalize-symbols",
92   cl::desc("Enable elimination of non-kernel functions and unused globals"),
93   cl::init(false),
94   cl::Hidden);
95 
96 // Option to inline all early.
97 static cl::opt<bool> EarlyInlineAll(
98   "amdgpu-early-inline-all",
99   cl::desc("Inline all functions early"),
100   cl::init(false),
101   cl::Hidden);
102 
103 static cl::opt<bool> EnableSDWAPeephole(
104   "amdgpu-sdwa-peephole",
105   cl::desc("Enable SDWA peepholer"),
106   cl::init(true));
107 
108 // Enable address space based alias analysis
109 static cl::opt<bool> EnableAMDGPUAliasAnalysis("enable-amdgpu-aa", cl::Hidden,
110   cl::desc("Enable AMDGPU Alias Analysis"),
111   cl::init(true));
112 
113 // Option to enable new waitcnt insertion pass.
114 static cl::opt<bool> EnableSIInsertWaitcntsPass(
115   "enable-si-insert-waitcnts",
116   cl::desc("Use new waitcnt insertion pass"),
117   cl::init(true));
118 
119 // Option to run late CFG structurizer
120 static cl::opt<bool, true> LateCFGStructurize(
121   "amdgpu-late-structurize",
122   cl::desc("Enable late CFG structurization"),
123   cl::location(AMDGPUTargetMachine::EnableLateStructurizeCFG),
124   cl::Hidden);
125 
126 static cl::opt<bool> EnableAMDGPUFunctionCalls(
127   "amdgpu-function-calls",
128   cl::Hidden,
129   cl::desc("Enable AMDGPU function call support"),
130   cl::init(false));
131 
132 // Enable lib calls simplifications
133 static cl::opt<bool> EnableLibCallSimplify(
134   "amdgpu-simplify-libcall",
135   cl::desc("Enable mdgpu library simplifications"),
136   cl::init(true),
137   cl::Hidden);
138 
139 extern "C" void LLVMInitializeAMDGPUTarget() {
140   // Register the target
141   RegisterTargetMachine<R600TargetMachine> X(getTheAMDGPUTarget());
142   RegisterTargetMachine<GCNTargetMachine> Y(getTheGCNTarget());
143 
144   PassRegistry *PR = PassRegistry::getPassRegistry();
145   initializeR600ClauseMergePassPass(*PR);
146   initializeR600ControlFlowFinalizerPass(*PR);
147   initializeR600PacketizerPass(*PR);
148   initializeR600ExpandSpecialInstrsPassPass(*PR);
149   initializeR600VectorRegMergerPass(*PR);
150   initializeAMDGPUDAGToDAGISelPass(*PR);
151   initializeSILowerI1CopiesPass(*PR);
152   initializeSIFixSGPRCopiesPass(*PR);
153   initializeSIFixVGPRCopiesPass(*PR);
154   initializeSIFoldOperandsPass(*PR);
155   initializeSIPeepholeSDWAPass(*PR);
156   initializeSIShrinkInstructionsPass(*PR);
157   initializeSIOptimizeExecMaskingPreRAPass(*PR);
158   initializeSILoadStoreOptimizerPass(*PR);
159   initializeAMDGPUAlwaysInlinePass(*PR);
160   initializeAMDGPUAnnotateKernelFeaturesPass(*PR);
161   initializeAMDGPUAnnotateUniformValuesPass(*PR);
162   initializeAMDGPUArgumentUsageInfoPass(*PR);
163   initializeAMDGPULowerIntrinsicsPass(*PR);
164   initializeAMDGPUOpenCLEnqueuedBlockLoweringPass(*PR);
165   initializeAMDGPUPromoteAllocaPass(*PR);
166   initializeAMDGPUCodeGenPreparePass(*PR);
167   initializeAMDGPURewriteOutArgumentsPass(*PR);
168   initializeAMDGPUUnifyMetadataPass(*PR);
169   initializeSIAnnotateControlFlowPass(*PR);
170   initializeSIInsertWaitsPass(*PR);
171   initializeSIInsertWaitcntsPass(*PR);
172   initializeSIWholeQuadModePass(*PR);
173   initializeSILowerControlFlowPass(*PR);
174   initializeSIInsertSkipsPass(*PR);
175   initializeSIMemoryLegalizerPass(*PR);
176   initializeSIDebuggerInsertNopsPass(*PR);
177   initializeSIOptimizeExecMaskingPass(*PR);
178   initializeSIFixWWMLivenessPass(*PR);
179   initializeAMDGPUUnifyDivergentExitNodesPass(*PR);
180   initializeAMDGPUAAWrapperPassPass(*PR);
181   initializeAMDGPUUseNativeCallsPass(*PR);
182   initializeAMDGPUSimplifyLibCallsPass(*PR);
183   initializeAMDGPUInlinerPass(*PR);
184 }
185 
186 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
187   return llvm::make_unique<AMDGPUTargetObjectFile>();
188 }
189 
190 static ScheduleDAGInstrs *createR600MachineScheduler(MachineSchedContext *C) {
191   return new ScheduleDAGMILive(C, llvm::make_unique<R600SchedStrategy>());
192 }
193 
194 static ScheduleDAGInstrs *createSIMachineScheduler(MachineSchedContext *C) {
195   return new SIScheduleDAGMI(C);
196 }
197 
198 static ScheduleDAGInstrs *
199 createGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
200   ScheduleDAGMILive *DAG =
201     new GCNScheduleDAGMILive(C, make_unique<GCNMaxOccupancySchedStrategy>(C));
202   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
203   DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
204   DAG->addMutation(createAMDGPUMacroFusionDAGMutation());
205   return DAG;
206 }
207 
208 static ScheduleDAGInstrs *
209 createIterativeGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
210   auto DAG = new GCNIterativeScheduler(C,
211     GCNIterativeScheduler::SCHEDULE_LEGACYMAXOCCUPANCY);
212   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
213   DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
214   return DAG;
215 }
216 
217 static ScheduleDAGInstrs *createMinRegScheduler(MachineSchedContext *C) {
218   return new GCNIterativeScheduler(C,
219     GCNIterativeScheduler::SCHEDULE_MINREGFORCED);
220 }
221 
222 static ScheduleDAGInstrs *
223 createIterativeILPMachineScheduler(MachineSchedContext *C) {
224   auto DAG = new GCNIterativeScheduler(C,
225     GCNIterativeScheduler::SCHEDULE_ILP);
226   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
227   DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
228   DAG->addMutation(createAMDGPUMacroFusionDAGMutation());
229   return DAG;
230 }
231 
232 static MachineSchedRegistry
233 R600SchedRegistry("r600", "Run R600's custom scheduler",
234                    createR600MachineScheduler);
235 
236 static MachineSchedRegistry
237 SISchedRegistry("si", "Run SI's custom scheduler",
238                 createSIMachineScheduler);
239 
240 static MachineSchedRegistry
241 GCNMaxOccupancySchedRegistry("gcn-max-occupancy",
242                              "Run GCN scheduler to maximize occupancy",
243                              createGCNMaxOccupancyMachineScheduler);
244 
245 static MachineSchedRegistry
246 IterativeGCNMaxOccupancySchedRegistry("gcn-max-occupancy-experimental",
247   "Run GCN scheduler to maximize occupancy (experimental)",
248   createIterativeGCNMaxOccupancyMachineScheduler);
249 
250 static MachineSchedRegistry
251 GCNMinRegSchedRegistry("gcn-minreg",
252   "Run GCN iterative scheduler for minimal register usage (experimental)",
253   createMinRegScheduler);
254 
255 static MachineSchedRegistry
256 GCNILPSchedRegistry("gcn-ilp",
257   "Run GCN iterative scheduler for ILP scheduling (experimental)",
258   createIterativeILPMachineScheduler);
259 
260 static StringRef computeDataLayout(const Triple &TT) {
261   if (TT.getArch() == Triple::r600) {
262     // 32-bit pointers.
263     if (TT.getEnvironmentName() == "amdgiz" ||
264         TT.getEnvironmentName() == "amdgizcl")
265       return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
266              "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-A5";
267     return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
268             "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64";
269   }
270 
271   // 32-bit private, local, and region pointers. 64-bit global, constant and
272   // flat.
273   if (TT.getEnvironmentName() == "amdgiz" ||
274       TT.getEnvironmentName() == "amdgizcl")
275     return "e-p:64:64-p1:64:64-p2:64:64-p3:32:32-p4:32:32-p5:32:32"
276          "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
277          "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-A5";
278   return "e-p:32:32-p1:64:64-p2:64:64-p3:32:32-p4:64:64-p5:32:32"
279       "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
280       "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64";
281 }
282 
283 LLVM_READNONE
284 static StringRef getGPUOrDefault(const Triple &TT, StringRef GPU) {
285   if (!GPU.empty())
286     return GPU;
287 
288   if (TT.getArch() == Triple::amdgcn)
289     return "generic";
290 
291   return "r600";
292 }
293 
294 static Reloc::Model getEffectiveRelocModel(Optional<Reloc::Model> RM) {
295   // The AMDGPU toolchain only supports generating shared objects, so we
296   // must always use PIC.
297   return Reloc::PIC_;
298 }
299 
300 static CodeModel::Model getEffectiveCodeModel(Optional<CodeModel::Model> CM) {
301   if (CM)
302     return *CM;
303   return CodeModel::Small;
304 }
305 
306 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, const Triple &TT,
307                                          StringRef CPU, StringRef FS,
308                                          TargetOptions Options,
309                                          Optional<Reloc::Model> RM,
310                                          Optional<CodeModel::Model> CM,
311                                          CodeGenOpt::Level OptLevel)
312     : LLVMTargetMachine(T, computeDataLayout(TT), TT, getGPUOrDefault(TT, CPU),
313                         FS, Options, getEffectiveRelocModel(RM),
314                         getEffectiveCodeModel(CM), OptLevel),
315       TLOF(createTLOF(getTargetTriple())) {
316   AS = AMDGPU::getAMDGPUAS(TT);
317   initAsmInfo();
318 }
319 
320 AMDGPUTargetMachine::~AMDGPUTargetMachine() = default;
321 
322 bool AMDGPUTargetMachine::EnableLateStructurizeCFG = false;
323 
324 StringRef AMDGPUTargetMachine::getGPUName(const Function &F) const {
325   Attribute GPUAttr = F.getFnAttribute("target-cpu");
326   return GPUAttr.hasAttribute(Attribute::None) ?
327     getTargetCPU() : GPUAttr.getValueAsString();
328 }
329 
330 StringRef AMDGPUTargetMachine::getFeatureString(const Function &F) const {
331   Attribute FSAttr = F.getFnAttribute("target-features");
332 
333   return FSAttr.hasAttribute(Attribute::None) ?
334     getTargetFeatureString() :
335     FSAttr.getValueAsString();
336 }
337 
338 static ImmutablePass *createAMDGPUExternalAAWrapperPass() {
339   return createExternalAAWrapperPass([](Pass &P, Function &, AAResults &AAR) {
340       if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>())
341         AAR.addAAResult(WrapperPass->getResult());
342       });
343 }
344 
345 /// Predicate for Internalize pass.
346 static bool mustPreserveGV(const GlobalValue &GV) {
347   if (const Function *F = dyn_cast<Function>(&GV))
348     return F->isDeclaration() || AMDGPU::isEntryFunctionCC(F->getCallingConv());
349 
350   return !GV.use_empty();
351 }
352 
353 void AMDGPUTargetMachine::adjustPassManager(PassManagerBuilder &Builder) {
354   Builder.DivergentTarget = true;
355 
356   bool EnableOpt = getOptLevel() > CodeGenOpt::None;
357   bool Internalize = InternalizeSymbols;
358   bool EarlyInline = EarlyInlineAll && EnableOpt && !EnableAMDGPUFunctionCalls;
359   bool AMDGPUAA = EnableAMDGPUAliasAnalysis && EnableOpt;
360   bool LibCallSimplify = EnableLibCallSimplify && EnableOpt;
361 
362   if (EnableAMDGPUFunctionCalls) {
363     delete Builder.Inliner;
364     Builder.Inliner = createAMDGPUFunctionInliningPass();
365   }
366 
367   if (Internalize) {
368     // If we're generating code, we always have the whole program available. The
369     // relocations expected for externally visible functions aren't supported,
370     // so make sure every non-entry function is hidden.
371     Builder.addExtension(
372       PassManagerBuilder::EP_EnabledOnOptLevel0,
373       [](const PassManagerBuilder &, legacy::PassManagerBase &PM) {
374         PM.add(createInternalizePass(mustPreserveGV));
375       });
376   }
377 
378   Builder.addExtension(
379     PassManagerBuilder::EP_ModuleOptimizerEarly,
380     [Internalize, EarlyInline, AMDGPUAA](const PassManagerBuilder &,
381                                          legacy::PassManagerBase &PM) {
382       if (AMDGPUAA) {
383         PM.add(createAMDGPUAAWrapperPass());
384         PM.add(createAMDGPUExternalAAWrapperPass());
385       }
386       PM.add(createAMDGPUUnifyMetadataPass());
387       if (Internalize) {
388         PM.add(createInternalizePass(mustPreserveGV));
389         PM.add(createGlobalDCEPass());
390       }
391       if (EarlyInline)
392         PM.add(createAMDGPUAlwaysInlinePass(false));
393   });
394 
395   const auto &Opt = Options;
396   Builder.addExtension(
397     PassManagerBuilder::EP_EarlyAsPossible,
398     [AMDGPUAA, LibCallSimplify, &Opt](const PassManagerBuilder &,
399                                       legacy::PassManagerBase &PM) {
400       if (AMDGPUAA) {
401         PM.add(createAMDGPUAAWrapperPass());
402         PM.add(createAMDGPUExternalAAWrapperPass());
403       }
404       PM.add(llvm::createAMDGPUUseNativeCallsPass());
405       if (LibCallSimplify)
406         PM.add(llvm::createAMDGPUSimplifyLibCallsPass(Opt));
407   });
408 
409   Builder.addExtension(
410     PassManagerBuilder::EP_CGSCCOptimizerLate,
411     [](const PassManagerBuilder &, legacy::PassManagerBase &PM) {
412       // Add infer address spaces pass to the opt pipeline after inlining
413       // but before SROA to increase SROA opportunities.
414       PM.add(createInferAddressSpacesPass());
415   });
416 }
417 
418 //===----------------------------------------------------------------------===//
419 // R600 Target Machine (R600 -> Cayman)
420 //===----------------------------------------------------------------------===//
421 
422 R600TargetMachine::R600TargetMachine(const Target &T, const Triple &TT,
423                                      StringRef CPU, StringRef FS,
424                                      TargetOptions Options,
425                                      Optional<Reloc::Model> RM,
426                                      Optional<CodeModel::Model> CM,
427                                      CodeGenOpt::Level OL, bool JIT)
428     : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {
429   setRequiresStructuredCFG(true);
430 }
431 
432 const R600Subtarget *R600TargetMachine::getSubtargetImpl(
433   const Function &F) const {
434   StringRef GPU = getGPUName(F);
435   StringRef FS = getFeatureString(F);
436 
437   SmallString<128> SubtargetKey(GPU);
438   SubtargetKey.append(FS);
439 
440   auto &I = SubtargetMap[SubtargetKey];
441   if (!I) {
442     // This needs to be done before we create a new subtarget since any
443     // creation will depend on the TM and the code generation flags on the
444     // function that reside in TargetOptions.
445     resetTargetOptions(F);
446     I = llvm::make_unique<R600Subtarget>(TargetTriple, GPU, FS, *this);
447   }
448 
449   return I.get();
450 }
451 
452 //===----------------------------------------------------------------------===//
453 // GCN Target Machine (SI+)
454 //===----------------------------------------------------------------------===//
455 
456 GCNTargetMachine::GCNTargetMachine(const Target &T, const Triple &TT,
457                                    StringRef CPU, StringRef FS,
458                                    TargetOptions Options,
459                                    Optional<Reloc::Model> RM,
460                                    Optional<CodeModel::Model> CM,
461                                    CodeGenOpt::Level OL, bool JIT)
462     : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {}
463 
464 const SISubtarget *GCNTargetMachine::getSubtargetImpl(const Function &F) const {
465   StringRef GPU = getGPUName(F);
466   StringRef FS = getFeatureString(F);
467 
468   SmallString<128> SubtargetKey(GPU);
469   SubtargetKey.append(FS);
470 
471   auto &I = SubtargetMap[SubtargetKey];
472   if (!I) {
473     // This needs to be done before we create a new subtarget since any
474     // creation will depend on the TM and the code generation flags on the
475     // function that reside in TargetOptions.
476     resetTargetOptions(F);
477     I = llvm::make_unique<SISubtarget>(TargetTriple, GPU, FS, *this);
478   }
479 
480   I->setScalarizeGlobalBehavior(ScalarizeGlobal);
481 
482   return I.get();
483 }
484 
485 //===----------------------------------------------------------------------===//
486 // AMDGPU Pass Setup
487 //===----------------------------------------------------------------------===//
488 
489 namespace {
490 
491 class AMDGPUPassConfig : public TargetPassConfig {
492 public:
493   AMDGPUPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM)
494     : TargetPassConfig(TM, PM) {
495     // Exceptions and StackMaps are not supported, so these passes will never do
496     // anything.
497     disablePass(&StackMapLivenessID);
498     disablePass(&FuncletLayoutID);
499   }
500 
501   AMDGPUTargetMachine &getAMDGPUTargetMachine() const {
502     return getTM<AMDGPUTargetMachine>();
503   }
504 
505   ScheduleDAGInstrs *
506   createMachineScheduler(MachineSchedContext *C) const override {
507     ScheduleDAGMILive *DAG = createGenericSchedLive(C);
508     DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
509     DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
510     return DAG;
511   }
512 
513   void addEarlyCSEOrGVNPass();
514   void addStraightLineScalarOptimizationPasses();
515   void addIRPasses() override;
516   void addCodeGenPrepare() override;
517   bool addPreISel() override;
518   bool addInstSelector() override;
519   bool addGCPasses() override;
520 };
521 
522 class R600PassConfig final : public AMDGPUPassConfig {
523 public:
524   R600PassConfig(LLVMTargetMachine &TM, PassManagerBase &PM)
525     : AMDGPUPassConfig(TM, PM) {}
526 
527   ScheduleDAGInstrs *createMachineScheduler(
528     MachineSchedContext *C) const override {
529     return createR600MachineScheduler(C);
530   }
531 
532   bool addPreISel() override;
533   bool addInstSelector() override;
534   void addPreRegAlloc() override;
535   void addPreSched2() override;
536   void addPreEmitPass() override;
537 };
538 
539 class GCNPassConfig final : public AMDGPUPassConfig {
540 public:
541   GCNPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM)
542     : AMDGPUPassConfig(TM, PM) {
543     // It is necessary to know the register usage of the entire call graph.  We
544     // allow calls without EnableAMDGPUFunctionCalls if they are marked
545     // noinline, so this is always required.
546     setRequiresCodeGenSCCOrder(true);
547   }
548 
549   GCNTargetMachine &getGCNTargetMachine() const {
550     return getTM<GCNTargetMachine>();
551   }
552 
553   ScheduleDAGInstrs *
554   createMachineScheduler(MachineSchedContext *C) const override;
555 
556   bool addPreISel() override;
557   void addMachineSSAOptimization() override;
558   bool addILPOpts() override;
559   bool addInstSelector() override;
560   bool addIRTranslator() override;
561   bool addLegalizeMachineIR() override;
562   bool addRegBankSelect() override;
563   bool addGlobalInstructionSelect() override;
564   void addFastRegAlloc(FunctionPass *RegAllocPass) override;
565   void addOptimizedRegAlloc(FunctionPass *RegAllocPass) override;
566   void addPreRegAlloc() override;
567   void addPostRegAlloc() override;
568   void addPreSched2() override;
569   void addPreEmitPass() override;
570 };
571 
572 } // end anonymous namespace
573 
574 TargetTransformInfo
575 AMDGPUTargetMachine::getTargetTransformInfo(const Function &F) {
576   return TargetTransformInfo(AMDGPUTTIImpl(this, F));
577 }
578 
579 void AMDGPUPassConfig::addEarlyCSEOrGVNPass() {
580   if (getOptLevel() == CodeGenOpt::Aggressive)
581     addPass(createGVNPass());
582   else
583     addPass(createEarlyCSEPass());
584 }
585 
586 void AMDGPUPassConfig::addStraightLineScalarOptimizationPasses() {
587   addPass(createSeparateConstOffsetFromGEPPass());
588   addPass(createSpeculativeExecutionPass());
589   // ReassociateGEPs exposes more opportunites for SLSR. See
590   // the example in reassociate-geps-and-slsr.ll.
591   addPass(createStraightLineStrengthReducePass());
592   // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or
593   // EarlyCSE can reuse.
594   addEarlyCSEOrGVNPass();
595   // Run NaryReassociate after EarlyCSE/GVN to be more effective.
596   addPass(createNaryReassociatePass());
597   // NaryReassociate on GEPs creates redundant common expressions, so run
598   // EarlyCSE after it.
599   addPass(createEarlyCSEPass());
600 }
601 
602 void AMDGPUPassConfig::addIRPasses() {
603   const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine();
604 
605   // There is no reason to run these.
606   disablePass(&StackMapLivenessID);
607   disablePass(&FuncletLayoutID);
608   disablePass(&PatchableFunctionID);
609 
610   addPass(createAMDGPULowerIntrinsicsPass());
611 
612   if (TM.getTargetTriple().getArch() == Triple::r600 ||
613       !EnableAMDGPUFunctionCalls) {
614     // Function calls are not supported, so make sure we inline everything.
615     addPass(createAMDGPUAlwaysInlinePass());
616     addPass(createAlwaysInlinerLegacyPass());
617     // We need to add the barrier noop pass, otherwise adding the function
618     // inlining pass will cause all of the PassConfigs passes to be run
619     // one function at a time, which means if we have a nodule with two
620     // functions, then we will generate code for the first function
621     // without ever running any passes on the second.
622     addPass(createBarrierNoopPass());
623   }
624 
625   if (TM.getTargetTriple().getArch() == Triple::amdgcn) {
626     // TODO: May want to move later or split into an early and late one.
627 
628     addPass(createAMDGPUCodeGenPreparePass());
629   }
630 
631   // Handle uses of OpenCL image2d_t, image3d_t and sampler_t arguments.
632   addPass(createAMDGPUOpenCLImageTypeLoweringPass());
633 
634   // Replace OpenCL enqueued block function pointers with global variables.
635   addPass(createAMDGPUOpenCLEnqueuedBlockLoweringPass());
636 
637   if (TM.getOptLevel() > CodeGenOpt::None) {
638     addPass(createInferAddressSpacesPass());
639     addPass(createAMDGPUPromoteAlloca());
640 
641     if (EnableSROA)
642       addPass(createSROAPass());
643 
644     addStraightLineScalarOptimizationPasses();
645 
646     if (EnableAMDGPUAliasAnalysis) {
647       addPass(createAMDGPUAAWrapperPass());
648       addPass(createExternalAAWrapperPass([](Pass &P, Function &,
649                                              AAResults &AAR) {
650         if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>())
651           AAR.addAAResult(WrapperPass->getResult());
652         }));
653     }
654   }
655 
656   TargetPassConfig::addIRPasses();
657 
658   // EarlyCSE is not always strong enough to clean up what LSR produces. For
659   // example, GVN can combine
660   //
661   //   %0 = add %a, %b
662   //   %1 = add %b, %a
663   //
664   // and
665   //
666   //   %0 = shl nsw %a, 2
667   //   %1 = shl %a, 2
668   //
669   // but EarlyCSE can do neither of them.
670   if (getOptLevel() != CodeGenOpt::None)
671     addEarlyCSEOrGVNPass();
672 }
673 
674 void AMDGPUPassConfig::addCodeGenPrepare() {
675   TargetPassConfig::addCodeGenPrepare();
676 
677   if (EnableLoadStoreVectorizer)
678     addPass(createLoadStoreVectorizerPass());
679 }
680 
681 bool AMDGPUPassConfig::addPreISel() {
682   addPass(createFlattenCFGPass());
683   return false;
684 }
685 
686 bool AMDGPUPassConfig::addInstSelector() {
687   addPass(createAMDGPUISelDag(&getAMDGPUTargetMachine(), getOptLevel()));
688   return false;
689 }
690 
691 bool AMDGPUPassConfig::addGCPasses() {
692   // Do nothing. GC is not supported.
693   return false;
694 }
695 
696 //===----------------------------------------------------------------------===//
697 // R600 Pass Setup
698 //===----------------------------------------------------------------------===//
699 
700 bool R600PassConfig::addPreISel() {
701   AMDGPUPassConfig::addPreISel();
702 
703   if (EnableR600StructurizeCFG)
704     addPass(createStructurizeCFGPass());
705   return false;
706 }
707 
708 bool R600PassConfig::addInstSelector() {
709   addPass(createR600ISelDag(&getAMDGPUTargetMachine(), getOptLevel()));
710   return false;
711 }
712 
713 void R600PassConfig::addPreRegAlloc() {
714   addPass(createR600VectorRegMerger());
715 }
716 
717 void R600PassConfig::addPreSched2() {
718   addPass(createR600EmitClauseMarkers(), false);
719   if (EnableR600IfConvert)
720     addPass(&IfConverterID, false);
721   addPass(createR600ClauseMergePass(), false);
722 }
723 
724 void R600PassConfig::addPreEmitPass() {
725   addPass(createAMDGPUCFGStructurizerPass(), false);
726   addPass(createR600ExpandSpecialInstrsPass(), false);
727   addPass(&FinalizeMachineBundlesID, false);
728   addPass(createR600Packetizer(), false);
729   addPass(createR600ControlFlowFinalizer(), false);
730 }
731 
732 TargetPassConfig *R600TargetMachine::createPassConfig(PassManagerBase &PM) {
733   return new R600PassConfig(*this, PM);
734 }
735 
736 //===----------------------------------------------------------------------===//
737 // GCN Pass Setup
738 //===----------------------------------------------------------------------===//
739 
740 ScheduleDAGInstrs *GCNPassConfig::createMachineScheduler(
741   MachineSchedContext *C) const {
742   const SISubtarget &ST = C->MF->getSubtarget<SISubtarget>();
743   if (ST.enableSIScheduler())
744     return createSIMachineScheduler(C);
745   return createGCNMaxOccupancyMachineScheduler(C);
746 }
747 
748 bool GCNPassConfig::addPreISel() {
749   AMDGPUPassConfig::addPreISel();
750 
751   // FIXME: We need to run a pass to propagate the attributes when calls are
752   // supported.
753   addPass(createAMDGPUAnnotateKernelFeaturesPass());
754 
755   // Merge divergent exit nodes. StructurizeCFG won't recognize the multi-exit
756   // regions formed by them.
757   addPass(&AMDGPUUnifyDivergentExitNodesID);
758   if (!LateCFGStructurize) {
759     addPass(createStructurizeCFGPass(true)); // true -> SkipUniformRegions
760   }
761   addPass(createSinkingPass());
762   addPass(createAMDGPUAnnotateUniformValues());
763   if (!LateCFGStructurize) {
764     addPass(createSIAnnotateControlFlowPass());
765   }
766 
767   return false;
768 }
769 
770 void GCNPassConfig::addMachineSSAOptimization() {
771   TargetPassConfig::addMachineSSAOptimization();
772 
773   // We want to fold operands after PeepholeOptimizer has run (or as part of
774   // it), because it will eliminate extra copies making it easier to fold the
775   // real source operand. We want to eliminate dead instructions after, so that
776   // we see fewer uses of the copies. We then need to clean up the dead
777   // instructions leftover after the operands are folded as well.
778   //
779   // XXX - Can we get away without running DeadMachineInstructionElim again?
780   addPass(&SIFoldOperandsID);
781   addPass(&DeadMachineInstructionElimID);
782   addPass(&SILoadStoreOptimizerID);
783   if (EnableSDWAPeephole) {
784     addPass(&SIPeepholeSDWAID);
785     addPass(&EarlyMachineLICMID);
786     addPass(&MachineCSEID);
787     addPass(&SIFoldOperandsID);
788     addPass(&DeadMachineInstructionElimID);
789   }
790   addPass(createSIShrinkInstructionsPass());
791 }
792 
793 bool GCNPassConfig::addILPOpts() {
794   if (EnableEarlyIfConversion)
795     addPass(&EarlyIfConverterID);
796 
797   TargetPassConfig::addILPOpts();
798   return false;
799 }
800 
801 bool GCNPassConfig::addInstSelector() {
802   AMDGPUPassConfig::addInstSelector();
803   addPass(createSILowerI1CopiesPass());
804   addPass(&SIFixSGPRCopiesID);
805   return false;
806 }
807 
808 bool GCNPassConfig::addIRTranslator() {
809   addPass(new IRTranslator());
810   return false;
811 }
812 
813 bool GCNPassConfig::addLegalizeMachineIR() {
814   addPass(new Legalizer());
815   return false;
816 }
817 
818 bool GCNPassConfig::addRegBankSelect() {
819   addPass(new RegBankSelect());
820   return false;
821 }
822 
823 bool GCNPassConfig::addGlobalInstructionSelect() {
824   addPass(new InstructionSelect());
825   return false;
826 }
827 
828 void GCNPassConfig::addPreRegAlloc() {
829   if (LateCFGStructurize) {
830     addPass(createAMDGPUMachineCFGStructurizerPass());
831   }
832   addPass(createSIWholeQuadModePass());
833 }
834 
835 void GCNPassConfig::addFastRegAlloc(FunctionPass *RegAllocPass) {
836   // FIXME: We have to disable the verifier here because of PHIElimination +
837   // TwoAddressInstructions disabling it.
838 
839   // This must be run immediately after phi elimination and before
840   // TwoAddressInstructions, otherwise the processing of the tied operand of
841   // SI_ELSE will introduce a copy of the tied operand source after the else.
842   insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
843 
844   // This must be run after SILowerControlFlow, since it needs to use the
845   // machine-level CFG, but before register allocation.
846   insertPass(&SILowerControlFlowID, &SIFixWWMLivenessID, false);
847 
848   TargetPassConfig::addFastRegAlloc(RegAllocPass);
849 }
850 
851 void GCNPassConfig::addOptimizedRegAlloc(FunctionPass *RegAllocPass) {
852   insertPass(&MachineSchedulerID, &SIOptimizeExecMaskingPreRAID);
853 
854   // This must be run immediately after phi elimination and before
855   // TwoAddressInstructions, otherwise the processing of the tied operand of
856   // SI_ELSE will introduce a copy of the tied operand source after the else.
857   insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
858 
859   // This must be run after SILowerControlFlow, since it needs to use the
860   // machine-level CFG, but before register allocation.
861   insertPass(&SILowerControlFlowID, &SIFixWWMLivenessID, false);
862 
863   TargetPassConfig::addOptimizedRegAlloc(RegAllocPass);
864 }
865 
866 void GCNPassConfig::addPostRegAlloc() {
867   addPass(&SIFixVGPRCopiesID);
868   addPass(&SIOptimizeExecMaskingID);
869   TargetPassConfig::addPostRegAlloc();
870 }
871 
872 void GCNPassConfig::addPreSched2() {
873 }
874 
875 void GCNPassConfig::addPreEmitPass() {
876   // The hazard recognizer that runs as part of the post-ra scheduler does not
877   // guarantee to be able handle all hazards correctly. This is because if there
878   // are multiple scheduling regions in a basic block, the regions are scheduled
879   // bottom up, so when we begin to schedule a region we don't know what
880   // instructions were emitted directly before it.
881   //
882   // Here we add a stand-alone hazard recognizer pass which can handle all
883   // cases.
884   addPass(&PostRAHazardRecognizerID);
885 
886   if (EnableSIInsertWaitcntsPass)
887     addPass(createSIInsertWaitcntsPass());
888   else
889     addPass(createSIInsertWaitsPass());
890   addPass(createSIShrinkInstructionsPass());
891   addPass(&SIInsertSkipsPassID);
892   addPass(createSIMemoryLegalizerPass());
893   addPass(createSIDebuggerInsertNopsPass());
894   addPass(&BranchRelaxationPassID);
895 }
896 
897 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) {
898   return new GCNPassConfig(*this, PM);
899 }
900