1 //===-- AMDGPUTargetMachine.cpp - TargetMachine for hw codegen targets-----===//
2 //
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
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 /// \file
10 /// The AMDGPU target machine contains all of the hardware specific
11 /// information  needed to emit code for R600 and SI GPUs.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "AMDGPUTargetMachine.h"
16 #include "AMDGPU.h"
17 #include "AMDGPUAliasAnalysis.h"
18 #include "AMDGPUCallLowering.h"
19 #include "AMDGPUInstructionSelector.h"
20 #include "AMDGPULegalizerInfo.h"
21 #include "AMDGPUMacroFusion.h"
22 #include "AMDGPUTargetObjectFile.h"
23 #include "AMDGPUTargetTransformInfo.h"
24 #include "GCNIterativeScheduler.h"
25 #include "GCNSchedStrategy.h"
26 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
27 #include "R600MachineScheduler.h"
28 #include "SIMachineFunctionInfo.h"
29 #include "SIMachineScheduler.h"
30 #include "TargetInfo/AMDGPUTargetInfo.h"
31 #include "llvm/CodeGen/GlobalISel/IRTranslator.h"
32 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h"
33 #include "llvm/CodeGen/GlobalISel/Legalizer.h"
34 #include "llvm/CodeGen/GlobalISel/Localizer.h"
35 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h"
36 #include "llvm/CodeGen/MIRParser/MIParser.h"
37 #include "llvm/CodeGen/Passes.h"
38 #include "llvm/CodeGen/TargetPassConfig.h"
39 #include "llvm/IR/Attributes.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/LegacyPassManager.h"
42 #include "llvm/InitializePasses.h"
43 #include "llvm/Pass.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Compiler.h"
46 #include "llvm/Support/TargetRegistry.h"
47 #include "llvm/Target/TargetLoweringObjectFile.h"
48 #include "llvm/Transforms/IPO.h"
49 #include "llvm/Transforms/IPO/AlwaysInliner.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/Utils.h"
54 #include "llvm/Transforms/Vectorize.h"
55 #include <memory>
56 
57 using namespace llvm;
58 
59 static cl::opt<bool> EnableR600StructurizeCFG(
60   "r600-ir-structurize",
61   cl::desc("Use StructurizeCFG IR pass"),
62   cl::init(true));
63 
64 static cl::opt<bool> EnableSROA(
65   "amdgpu-sroa",
66   cl::desc("Run SROA after promote alloca pass"),
67   cl::ReallyHidden,
68   cl::init(true));
69 
70 static cl::opt<bool>
71 EnableEarlyIfConversion("amdgpu-early-ifcvt", cl::Hidden,
72                         cl::desc("Run early if-conversion"),
73                         cl::init(false));
74 
75 static cl::opt<bool>
76 OptExecMaskPreRA("amdgpu-opt-exec-mask-pre-ra", cl::Hidden,
77             cl::desc("Run pre-RA exec mask optimizations"),
78             cl::init(true));
79 
80 static cl::opt<bool> EnableR600IfConvert(
81   "r600-if-convert",
82   cl::desc("Use if conversion pass"),
83   cl::ReallyHidden,
84   cl::init(true));
85 
86 // Option to disable vectorizer for tests.
87 static cl::opt<bool> EnableLoadStoreVectorizer(
88   "amdgpu-load-store-vectorizer",
89   cl::desc("Enable load store vectorizer"),
90   cl::init(true),
91   cl::Hidden);
92 
93 // Option to control global loads scalarization
94 static cl::opt<bool> ScalarizeGlobal(
95   "amdgpu-scalarize-global-loads",
96   cl::desc("Enable global load scalarization"),
97   cl::init(true),
98   cl::Hidden);
99 
100 // Option to run internalize pass.
101 static cl::opt<bool> InternalizeSymbols(
102   "amdgpu-internalize-symbols",
103   cl::desc("Enable elimination of non-kernel functions and unused globals"),
104   cl::init(false),
105   cl::Hidden);
106 
107 // Option to inline all early.
108 static cl::opt<bool> EarlyInlineAll(
109   "amdgpu-early-inline-all",
110   cl::desc("Inline all functions early"),
111   cl::init(false),
112   cl::Hidden);
113 
114 static cl::opt<bool> EnableSDWAPeephole(
115   "amdgpu-sdwa-peephole",
116   cl::desc("Enable SDWA peepholer"),
117   cl::init(true));
118 
119 static cl::opt<bool> EnableDPPCombine(
120   "amdgpu-dpp-combine",
121   cl::desc("Enable DPP combiner"),
122   cl::init(true));
123 
124 // Enable address space based alias analysis
125 static cl::opt<bool> EnableAMDGPUAliasAnalysis("enable-amdgpu-aa", cl::Hidden,
126   cl::desc("Enable AMDGPU Alias Analysis"),
127   cl::init(true));
128 
129 // Option to run late CFG structurizer
130 static cl::opt<bool, true> LateCFGStructurize(
131   "amdgpu-late-structurize",
132   cl::desc("Enable late CFG structurization"),
133   cl::location(AMDGPUTargetMachine::EnableLateStructurizeCFG),
134   cl::Hidden);
135 
136 static cl::opt<bool, true> EnableAMDGPUFunctionCallsOpt(
137   "amdgpu-function-calls",
138   cl::desc("Enable AMDGPU function call support"),
139   cl::location(AMDGPUTargetMachine::EnableFunctionCalls),
140   cl::init(true),
141   cl::Hidden);
142 
143 static cl::opt<bool, true> EnableAMDGPUFixedFunctionABIOpt(
144   "amdgpu-fixed-function-abi",
145   cl::desc("Enable all implicit function arguments"),
146   cl::location(AMDGPUTargetMachine::EnableFixedFunctionABI),
147   cl::init(false),
148   cl::Hidden);
149 
150 // Enable lib calls simplifications
151 static cl::opt<bool> EnableLibCallSimplify(
152   "amdgpu-simplify-libcall",
153   cl::desc("Enable amdgpu library simplifications"),
154   cl::init(true),
155   cl::Hidden);
156 
157 static cl::opt<bool> EnableLowerKernelArguments(
158   "amdgpu-ir-lower-kernel-arguments",
159   cl::desc("Lower kernel argument loads in IR pass"),
160   cl::init(true),
161   cl::Hidden);
162 
163 static cl::opt<bool> EnableRegReassign(
164   "amdgpu-reassign-regs",
165   cl::desc("Enable register reassign optimizations on gfx10+"),
166   cl::init(true),
167   cl::Hidden);
168 
169 // Enable atomic optimization
170 static cl::opt<bool> EnableAtomicOptimizations(
171   "amdgpu-atomic-optimizations",
172   cl::desc("Enable atomic optimizations"),
173   cl::init(false),
174   cl::Hidden);
175 
176 // Enable Mode register optimization
177 static cl::opt<bool> EnableSIModeRegisterPass(
178   "amdgpu-mode-register",
179   cl::desc("Enable mode register pass"),
180   cl::init(true),
181   cl::Hidden);
182 
183 // Option is used in lit tests to prevent deadcoding of patterns inspected.
184 static cl::opt<bool>
185 EnableDCEInRA("amdgpu-dce-in-ra",
186     cl::init(true), cl::Hidden,
187     cl::desc("Enable machine DCE inside regalloc"));
188 
189 static cl::opt<bool> EnableScalarIRPasses(
190   "amdgpu-scalar-ir-passes",
191   cl::desc("Enable scalar IR passes"),
192   cl::init(true),
193   cl::Hidden);
194 
195 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUTarget() {
196   // Register the target
197   RegisterTargetMachine<R600TargetMachine> X(getTheAMDGPUTarget());
198   RegisterTargetMachine<GCNTargetMachine> Y(getTheGCNTarget());
199 
200   PassRegistry *PR = PassRegistry::getPassRegistry();
201   initializeR600ClauseMergePassPass(*PR);
202   initializeR600ControlFlowFinalizerPass(*PR);
203   initializeR600PacketizerPass(*PR);
204   initializeR600ExpandSpecialInstrsPassPass(*PR);
205   initializeR600VectorRegMergerPass(*PR);
206   initializeGlobalISel(*PR);
207   initializeAMDGPUDAGToDAGISelPass(*PR);
208   initializeGCNDPPCombinePass(*PR);
209   initializeSILowerI1CopiesPass(*PR);
210   initializeSILowerSGPRSpillsPass(*PR);
211   initializeSIFixSGPRCopiesPass(*PR);
212   initializeSIFixVGPRCopiesPass(*PR);
213   initializeSIFixupVectorISelPass(*PR);
214   initializeSIFoldOperandsPass(*PR);
215   initializeSIPeepholeSDWAPass(*PR);
216   initializeSIShrinkInstructionsPass(*PR);
217   initializeSIOptimizeExecMaskingPreRAPass(*PR);
218   initializeSILoadStoreOptimizerPass(*PR);
219   initializeAMDGPUFixFunctionBitcastsPass(*PR);
220   initializeAMDGPUAlwaysInlinePass(*PR);
221   initializeAMDGPUAnnotateKernelFeaturesPass(*PR);
222   initializeAMDGPUAnnotateUniformValuesPass(*PR);
223   initializeAMDGPUArgumentUsageInfoPass(*PR);
224   initializeAMDGPUAtomicOptimizerPass(*PR);
225   initializeAMDGPULowerKernelArgumentsPass(*PR);
226   initializeAMDGPULowerKernelAttributesPass(*PR);
227   initializeAMDGPULowerIntrinsicsPass(*PR);
228   initializeAMDGPUOpenCLEnqueuedBlockLoweringPass(*PR);
229   initializeAMDGPUPostLegalizerCombinerPass(*PR);
230   initializeAMDGPUPreLegalizerCombinerPass(*PR);
231   initializeAMDGPUPromoteAllocaPass(*PR);
232   initializeAMDGPUCodeGenPreparePass(*PR);
233   initializeAMDGPUPropagateAttributesEarlyPass(*PR);
234   initializeAMDGPUPropagateAttributesLatePass(*PR);
235   initializeAMDGPURewriteOutArgumentsPass(*PR);
236   initializeAMDGPUUnifyMetadataPass(*PR);
237   initializeSIAnnotateControlFlowPass(*PR);
238   initializeSIInsertWaitcntsPass(*PR);
239   initializeSIModeRegisterPass(*PR);
240   initializeSIWholeQuadModePass(*PR);
241   initializeSILowerControlFlowPass(*PR);
242   initializeSIRemoveShortExecBranchesPass(*PR);
243   initializeSIInsertSkipsPass(*PR);
244   initializeSIMemoryLegalizerPass(*PR);
245   initializeSIOptimizeExecMaskingPass(*PR);
246   initializeSIPreAllocateWWMRegsPass(*PR);
247   initializeSIFormMemoryClausesPass(*PR);
248   initializeSIPostRABundlerPass(*PR);
249   initializeAMDGPUUnifyDivergentExitNodesPass(*PR);
250   initializeAMDGPUAAWrapperPassPass(*PR);
251   initializeAMDGPUExternalAAWrapperPass(*PR);
252   initializeAMDGPUUseNativeCallsPass(*PR);
253   initializeAMDGPUSimplifyLibCallsPass(*PR);
254   initializeAMDGPUInlinerPass(*PR);
255   initializeAMDGPUPrintfRuntimeBindingPass(*PR);
256   initializeGCNRegBankReassignPass(*PR);
257   initializeGCNNSAReassignPass(*PR);
258   initializeSIAddIMGInitPass(*PR);
259 }
260 
261 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
262   return std::make_unique<AMDGPUTargetObjectFile>();
263 }
264 
265 static ScheduleDAGInstrs *createR600MachineScheduler(MachineSchedContext *C) {
266   return new ScheduleDAGMILive(C, std::make_unique<R600SchedStrategy>());
267 }
268 
269 static ScheduleDAGInstrs *createSIMachineScheduler(MachineSchedContext *C) {
270   return new SIScheduleDAGMI(C);
271 }
272 
273 static ScheduleDAGInstrs *
274 createGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
275   ScheduleDAGMILive *DAG =
276     new GCNScheduleDAGMILive(C, std::make_unique<GCNMaxOccupancySchedStrategy>(C));
277   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
278   DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
279   DAG->addMutation(createAMDGPUMacroFusionDAGMutation());
280   return DAG;
281 }
282 
283 static ScheduleDAGInstrs *
284 createIterativeGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
285   auto DAG = new GCNIterativeScheduler(C,
286     GCNIterativeScheduler::SCHEDULE_LEGACYMAXOCCUPANCY);
287   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
288   DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
289   return DAG;
290 }
291 
292 static ScheduleDAGInstrs *createMinRegScheduler(MachineSchedContext *C) {
293   return new GCNIterativeScheduler(C,
294     GCNIterativeScheduler::SCHEDULE_MINREGFORCED);
295 }
296 
297 static ScheduleDAGInstrs *
298 createIterativeILPMachineScheduler(MachineSchedContext *C) {
299   auto DAG = new GCNIterativeScheduler(C,
300     GCNIterativeScheduler::SCHEDULE_ILP);
301   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
302   DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
303   DAG->addMutation(createAMDGPUMacroFusionDAGMutation());
304   return DAG;
305 }
306 
307 static MachineSchedRegistry
308 R600SchedRegistry("r600", "Run R600's custom scheduler",
309                    createR600MachineScheduler);
310 
311 static MachineSchedRegistry
312 SISchedRegistry("si", "Run SI's custom scheduler",
313                 createSIMachineScheduler);
314 
315 static MachineSchedRegistry
316 GCNMaxOccupancySchedRegistry("gcn-max-occupancy",
317                              "Run GCN scheduler to maximize occupancy",
318                              createGCNMaxOccupancyMachineScheduler);
319 
320 static MachineSchedRegistry
321 IterativeGCNMaxOccupancySchedRegistry("gcn-max-occupancy-experimental",
322   "Run GCN scheduler to maximize occupancy (experimental)",
323   createIterativeGCNMaxOccupancyMachineScheduler);
324 
325 static MachineSchedRegistry
326 GCNMinRegSchedRegistry("gcn-minreg",
327   "Run GCN iterative scheduler for minimal register usage (experimental)",
328   createMinRegScheduler);
329 
330 static MachineSchedRegistry
331 GCNILPSchedRegistry("gcn-ilp",
332   "Run GCN iterative scheduler for ILP scheduling (experimental)",
333   createIterativeILPMachineScheduler);
334 
335 static StringRef computeDataLayout(const Triple &TT) {
336   if (TT.getArch() == Triple::r600) {
337     // 32-bit pointers.
338       return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
339              "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5";
340   }
341 
342   // 32-bit private, local, and region pointers. 64-bit global, constant and
343   // flat, non-integral buffer fat pointers.
344     return "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32"
345          "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
346          "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5"
347          "-ni:7";
348 }
349 
350 LLVM_READNONE
351 static StringRef getGPUOrDefault(const Triple &TT, StringRef GPU) {
352   if (!GPU.empty())
353     return GPU;
354 
355   // Need to default to a target with flat support for HSA.
356   if (TT.getArch() == Triple::amdgcn)
357     return TT.getOS() == Triple::AMDHSA ? "generic-hsa" : "generic";
358 
359   return "r600";
360 }
361 
362 static Reloc::Model getEffectiveRelocModel(Optional<Reloc::Model> RM) {
363   // The AMDGPU toolchain only supports generating shared objects, so we
364   // must always use PIC.
365   return Reloc::PIC_;
366 }
367 
368 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, const Triple &TT,
369                                          StringRef CPU, StringRef FS,
370                                          TargetOptions Options,
371                                          Optional<Reloc::Model> RM,
372                                          Optional<CodeModel::Model> CM,
373                                          CodeGenOpt::Level OptLevel)
374     : LLVMTargetMachine(T, computeDataLayout(TT), TT, getGPUOrDefault(TT, CPU),
375                         FS, Options, getEffectiveRelocModel(RM),
376                         getEffectiveCodeModel(CM, CodeModel::Small), OptLevel),
377       TLOF(createTLOF(getTargetTriple())) {
378   initAsmInfo();
379   if (TT.getArch() == Triple::amdgcn) {
380     if (getMCSubtargetInfo()->checkFeatures("+wavefrontsize64"))
381       MRI.reset(llvm::createGCNMCRegisterInfo(AMDGPUDwarfFlavour::Wave64));
382     else if (getMCSubtargetInfo()->checkFeatures("+wavefrontsize32"))
383       MRI.reset(llvm::createGCNMCRegisterInfo(AMDGPUDwarfFlavour::Wave32));
384   }
385 }
386 
387 bool AMDGPUTargetMachine::EnableLateStructurizeCFG = false;
388 bool AMDGPUTargetMachine::EnableFunctionCalls = false;
389 bool AMDGPUTargetMachine::EnableFixedFunctionABI = false;
390 
391 AMDGPUTargetMachine::~AMDGPUTargetMachine() = default;
392 
393 StringRef AMDGPUTargetMachine::getGPUName(const Function &F) const {
394   Attribute GPUAttr = F.getFnAttribute("target-cpu");
395   return GPUAttr.hasAttribute(Attribute::None) ?
396     getTargetCPU() : GPUAttr.getValueAsString();
397 }
398 
399 StringRef AMDGPUTargetMachine::getFeatureString(const Function &F) const {
400   Attribute FSAttr = F.getFnAttribute("target-features");
401 
402   return FSAttr.hasAttribute(Attribute::None) ?
403     getTargetFeatureString() :
404     FSAttr.getValueAsString();
405 }
406 
407 /// Predicate for Internalize pass.
408 static bool mustPreserveGV(const GlobalValue &GV) {
409   if (const Function *F = dyn_cast<Function>(&GV))
410     return F->isDeclaration() || AMDGPU::isEntryFunctionCC(F->getCallingConv());
411 
412   return !GV.use_empty();
413 }
414 
415 void AMDGPUTargetMachine::adjustPassManager(PassManagerBuilder &Builder) {
416   Builder.DivergentTarget = true;
417 
418   bool EnableOpt = getOptLevel() > CodeGenOpt::None;
419   bool Internalize = InternalizeSymbols;
420   bool EarlyInline = EarlyInlineAll && EnableOpt && !EnableFunctionCalls;
421   bool AMDGPUAA = EnableAMDGPUAliasAnalysis && EnableOpt;
422   bool LibCallSimplify = EnableLibCallSimplify && EnableOpt;
423 
424   if (EnableFunctionCalls) {
425     delete Builder.Inliner;
426     Builder.Inliner = createAMDGPUFunctionInliningPass();
427   }
428 
429   Builder.addExtension(
430     PassManagerBuilder::EP_ModuleOptimizerEarly,
431     [Internalize, EarlyInline, AMDGPUAA, this](const PassManagerBuilder &,
432                                                legacy::PassManagerBase &PM) {
433       if (AMDGPUAA) {
434         PM.add(createAMDGPUAAWrapperPass());
435         PM.add(createAMDGPUExternalAAWrapperPass());
436       }
437       PM.add(createAMDGPUUnifyMetadataPass());
438       PM.add(createAMDGPUPrintfRuntimeBinding());
439       PM.add(createAMDGPUPropagateAttributesLatePass(this));
440       if (Internalize) {
441         PM.add(createInternalizePass(mustPreserveGV));
442         PM.add(createGlobalDCEPass());
443       }
444       if (EarlyInline)
445         PM.add(createAMDGPUAlwaysInlinePass(false));
446   });
447 
448   const auto &Opt = Options;
449   Builder.addExtension(
450     PassManagerBuilder::EP_EarlyAsPossible,
451     [AMDGPUAA, LibCallSimplify, &Opt, this](const PassManagerBuilder &,
452                                             legacy::PassManagerBase &PM) {
453       if (AMDGPUAA) {
454         PM.add(createAMDGPUAAWrapperPass());
455         PM.add(createAMDGPUExternalAAWrapperPass());
456       }
457       PM.add(llvm::createAMDGPUPropagateAttributesEarlyPass(this));
458       PM.add(llvm::createAMDGPUUseNativeCallsPass());
459       if (LibCallSimplify)
460         PM.add(llvm::createAMDGPUSimplifyLibCallsPass(Opt, this));
461   });
462 
463   Builder.addExtension(
464     PassManagerBuilder::EP_CGSCCOptimizerLate,
465     [](const PassManagerBuilder &, legacy::PassManagerBase &PM) {
466       // Add infer address spaces pass to the opt pipeline after inlining
467       // but before SROA to increase SROA opportunities.
468       PM.add(createInferAddressSpacesPass());
469 
470       // This should run after inlining to have any chance of doing anything,
471       // and before other cleanup optimizations.
472       PM.add(createAMDGPULowerKernelAttributesPass());
473   });
474 }
475 
476 //===----------------------------------------------------------------------===//
477 // R600 Target Machine (R600 -> Cayman)
478 //===----------------------------------------------------------------------===//
479 
480 R600TargetMachine::R600TargetMachine(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 OL, bool JIT)
486     : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {
487   setRequiresStructuredCFG(true);
488 
489   // Override the default since calls aren't supported for r600.
490   if (EnableFunctionCalls &&
491       EnableAMDGPUFunctionCallsOpt.getNumOccurrences() == 0)
492     EnableFunctionCalls = false;
493 }
494 
495 const R600Subtarget *R600TargetMachine::getSubtargetImpl(
496   const Function &F) const {
497   StringRef GPU = getGPUName(F);
498   StringRef FS = getFeatureString(F);
499 
500   SmallString<128> SubtargetKey(GPU);
501   SubtargetKey.append(FS);
502 
503   auto &I = SubtargetMap[SubtargetKey];
504   if (!I) {
505     // This needs to be done before we create a new subtarget since any
506     // creation will depend on the TM and the code generation flags on the
507     // function that reside in TargetOptions.
508     resetTargetOptions(F);
509     I = std::make_unique<R600Subtarget>(TargetTriple, GPU, FS, *this);
510   }
511 
512   return I.get();
513 }
514 
515 TargetTransformInfo
516 R600TargetMachine::getTargetTransformInfo(const Function &F) {
517   return TargetTransformInfo(R600TTIImpl(this, F));
518 }
519 
520 //===----------------------------------------------------------------------===//
521 // GCN Target Machine (SI+)
522 //===----------------------------------------------------------------------===//
523 
524 GCNTargetMachine::GCNTargetMachine(const Target &T, const Triple &TT,
525                                    StringRef CPU, StringRef FS,
526                                    TargetOptions Options,
527                                    Optional<Reloc::Model> RM,
528                                    Optional<CodeModel::Model> CM,
529                                    CodeGenOpt::Level OL, bool JIT)
530     : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {}
531 
532 const GCNSubtarget *GCNTargetMachine::getSubtargetImpl(const Function &F) const {
533   StringRef GPU = getGPUName(F);
534   StringRef FS = getFeatureString(F);
535 
536   SmallString<128> SubtargetKey(GPU);
537   SubtargetKey.append(FS);
538 
539   auto &I = SubtargetMap[SubtargetKey];
540   if (!I) {
541     // This needs to be done before we create a new subtarget since any
542     // creation will depend on the TM and the code generation flags on the
543     // function that reside in TargetOptions.
544     resetTargetOptions(F);
545     I = std::make_unique<GCNSubtarget>(TargetTriple, GPU, FS, *this);
546   }
547 
548   I->setScalarizeGlobalBehavior(ScalarizeGlobal);
549 
550   return I.get();
551 }
552 
553 TargetTransformInfo
554 GCNTargetMachine::getTargetTransformInfo(const Function &F) {
555   return TargetTransformInfo(GCNTTIImpl(this, F));
556 }
557 
558 //===----------------------------------------------------------------------===//
559 // AMDGPU Pass Setup
560 //===----------------------------------------------------------------------===//
561 
562 namespace {
563 
564 class AMDGPUPassConfig : public TargetPassConfig {
565 public:
566   AMDGPUPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM)
567     : TargetPassConfig(TM, PM) {
568     // Exceptions and StackMaps are not supported, so these passes will never do
569     // anything.
570     disablePass(&StackMapLivenessID);
571     disablePass(&FuncletLayoutID);
572   }
573 
574   AMDGPUTargetMachine &getAMDGPUTargetMachine() const {
575     return getTM<AMDGPUTargetMachine>();
576   }
577 
578   ScheduleDAGInstrs *
579   createMachineScheduler(MachineSchedContext *C) const override {
580     ScheduleDAGMILive *DAG = createGenericSchedLive(C);
581     DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
582     DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
583     return DAG;
584   }
585 
586   void addEarlyCSEOrGVNPass();
587   void addStraightLineScalarOptimizationPasses();
588   void addIRPasses() override;
589   void addCodeGenPrepare() override;
590   bool addPreISel() override;
591   bool addInstSelector() override;
592   bool addGCPasses() override;
593 
594   std::unique_ptr<CSEConfigBase> getCSEConfig() const override;
595 };
596 
597 std::unique_ptr<CSEConfigBase> AMDGPUPassConfig::getCSEConfig() const {
598   return getStandardCSEConfigForOpt(TM->getOptLevel());
599 }
600 
601 class R600PassConfig final : public AMDGPUPassConfig {
602 public:
603   R600PassConfig(LLVMTargetMachine &TM, PassManagerBase &PM)
604     : AMDGPUPassConfig(TM, PM) {}
605 
606   ScheduleDAGInstrs *createMachineScheduler(
607     MachineSchedContext *C) const override {
608     return createR600MachineScheduler(C);
609   }
610 
611   bool addPreISel() override;
612   bool addInstSelector() override;
613   void addPreRegAlloc() override;
614   void addPreSched2() override;
615   void addPreEmitPass() override;
616 };
617 
618 class GCNPassConfig final : public AMDGPUPassConfig {
619 public:
620   GCNPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM)
621     : AMDGPUPassConfig(TM, PM) {
622     // It is necessary to know the register usage of the entire call graph.  We
623     // allow calls without EnableAMDGPUFunctionCalls if they are marked
624     // noinline, so this is always required.
625     setRequiresCodeGenSCCOrder(true);
626   }
627 
628   GCNTargetMachine &getGCNTargetMachine() const {
629     return getTM<GCNTargetMachine>();
630   }
631 
632   ScheduleDAGInstrs *
633   createMachineScheduler(MachineSchedContext *C) const override;
634 
635   bool addPreISel() override;
636   void addMachineSSAOptimization() override;
637   bool addILPOpts() override;
638   bool addInstSelector() override;
639   bool addIRTranslator() override;
640   void addPreLegalizeMachineIR() override;
641   bool addLegalizeMachineIR() override;
642   void addPreRegBankSelect() override;
643   bool addRegBankSelect() override;
644   bool addGlobalInstructionSelect() override;
645   void addFastRegAlloc() override;
646   void addOptimizedRegAlloc() override;
647   void addPreRegAlloc() override;
648   bool addPreRewrite() override;
649   void addPostRegAlloc() override;
650   void addPreSched2() override;
651   void addPreEmitPass() override;
652 };
653 
654 } // end anonymous namespace
655 
656 void AMDGPUPassConfig::addEarlyCSEOrGVNPass() {
657   if (getOptLevel() == CodeGenOpt::Aggressive)
658     addPass(createGVNPass());
659   else
660     addPass(createEarlyCSEPass());
661 }
662 
663 void AMDGPUPassConfig::addStraightLineScalarOptimizationPasses() {
664   addPass(createLICMPass());
665   addPass(createSeparateConstOffsetFromGEPPass());
666   addPass(createSpeculativeExecutionPass());
667   // ReassociateGEPs exposes more opportunites for SLSR. See
668   // the example in reassociate-geps-and-slsr.ll.
669   addPass(createStraightLineStrengthReducePass());
670   // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or
671   // EarlyCSE can reuse.
672   addEarlyCSEOrGVNPass();
673   // Run NaryReassociate after EarlyCSE/GVN to be more effective.
674   addPass(createNaryReassociatePass());
675   // NaryReassociate on GEPs creates redundant common expressions, so run
676   // EarlyCSE after it.
677   addPass(createEarlyCSEPass());
678 }
679 
680 void AMDGPUPassConfig::addIRPasses() {
681   const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine();
682 
683   // There is no reason to run these.
684   disablePass(&StackMapLivenessID);
685   disablePass(&FuncletLayoutID);
686   disablePass(&PatchableFunctionID);
687 
688   addPass(createAMDGPUPrintfRuntimeBinding());
689 
690   // This must occur before inlining, as the inliner will not look through
691   // bitcast calls.
692   addPass(createAMDGPUFixFunctionBitcastsPass());
693 
694   // A call to propagate attributes pass in the backend in case opt was not run.
695   addPass(createAMDGPUPropagateAttributesEarlyPass(&TM));
696 
697   addPass(createAtomicExpandPass());
698 
699 
700   addPass(createAMDGPULowerIntrinsicsPass());
701 
702   // Function calls are not supported, so make sure we inline everything.
703   addPass(createAMDGPUAlwaysInlinePass());
704   addPass(createAlwaysInlinerLegacyPass());
705   // We need to add the barrier noop pass, otherwise adding the function
706   // inlining pass will cause all of the PassConfigs passes to be run
707   // one function at a time, which means if we have a nodule with two
708   // functions, then we will generate code for the first function
709   // without ever running any passes on the second.
710   addPass(createBarrierNoopPass());
711 
712   // Handle uses of OpenCL image2d_t, image3d_t and sampler_t arguments.
713   if (TM.getTargetTriple().getArch() == Triple::r600)
714     addPass(createR600OpenCLImageTypeLoweringPass());
715 
716   // Replace OpenCL enqueued block function pointers with global variables.
717   addPass(createAMDGPUOpenCLEnqueuedBlockLoweringPass());
718 
719   if (TM.getOptLevel() > CodeGenOpt::None) {
720     addPass(createInferAddressSpacesPass());
721     addPass(createAMDGPUPromoteAlloca());
722 
723     if (EnableSROA)
724       addPass(createSROAPass());
725 
726     if (EnableScalarIRPasses)
727       addStraightLineScalarOptimizationPasses();
728 
729     if (EnableAMDGPUAliasAnalysis) {
730       addPass(createAMDGPUAAWrapperPass());
731       addPass(createExternalAAWrapperPass([](Pass &P, Function &,
732                                              AAResults &AAR) {
733         if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>())
734           AAR.addAAResult(WrapperPass->getResult());
735         }));
736     }
737   }
738 
739   if (TM.getTargetTriple().getArch() == Triple::amdgcn) {
740     // TODO: May want to move later or split into an early and late one.
741     addPass(createAMDGPUCodeGenPreparePass());
742   }
743 
744   TargetPassConfig::addIRPasses();
745 
746   // EarlyCSE is not always strong enough to clean up what LSR produces. For
747   // example, GVN can combine
748   //
749   //   %0 = add %a, %b
750   //   %1 = add %b, %a
751   //
752   // and
753   //
754   //   %0 = shl nsw %a, 2
755   //   %1 = shl %a, 2
756   //
757   // but EarlyCSE can do neither of them.
758   if (getOptLevel() != CodeGenOpt::None && EnableScalarIRPasses)
759     addEarlyCSEOrGVNPass();
760 }
761 
762 void AMDGPUPassConfig::addCodeGenPrepare() {
763   if (TM->getTargetTriple().getArch() == Triple::amdgcn)
764     addPass(createAMDGPUAnnotateKernelFeaturesPass());
765 
766   if (TM->getTargetTriple().getArch() == Triple::amdgcn &&
767       EnableLowerKernelArguments)
768     addPass(createAMDGPULowerKernelArgumentsPass());
769 
770   addPass(&AMDGPUPerfHintAnalysisID);
771 
772   TargetPassConfig::addCodeGenPrepare();
773 
774   if (EnableLoadStoreVectorizer)
775     addPass(createLoadStoreVectorizerPass());
776 }
777 
778 bool AMDGPUPassConfig::addPreISel() {
779   addPass(createLowerSwitchPass());
780   addPass(createFlattenCFGPass());
781   return false;
782 }
783 
784 bool AMDGPUPassConfig::addInstSelector() {
785   // Defer the verifier until FinalizeISel.
786   addPass(createAMDGPUISelDag(&getAMDGPUTargetMachine(), getOptLevel()), false);
787   return false;
788 }
789 
790 bool AMDGPUPassConfig::addGCPasses() {
791   // Do nothing. GC is not supported.
792   return false;
793 }
794 
795 //===----------------------------------------------------------------------===//
796 // R600 Pass Setup
797 //===----------------------------------------------------------------------===//
798 
799 bool R600PassConfig::addPreISel() {
800   AMDGPUPassConfig::addPreISel();
801 
802   if (EnableR600StructurizeCFG)
803     addPass(createStructurizeCFGPass());
804   return false;
805 }
806 
807 bool R600PassConfig::addInstSelector() {
808   addPass(createR600ISelDag(&getAMDGPUTargetMachine(), getOptLevel()));
809   return false;
810 }
811 
812 void R600PassConfig::addPreRegAlloc() {
813   addPass(createR600VectorRegMerger());
814 }
815 
816 void R600PassConfig::addPreSched2() {
817   addPass(createR600EmitClauseMarkers(), false);
818   if (EnableR600IfConvert)
819     addPass(&IfConverterID, false);
820   addPass(createR600ClauseMergePass(), false);
821 }
822 
823 void R600PassConfig::addPreEmitPass() {
824   addPass(createAMDGPUCFGStructurizerPass(), false);
825   addPass(createR600ExpandSpecialInstrsPass(), false);
826   addPass(&FinalizeMachineBundlesID, false);
827   addPass(createR600Packetizer(), false);
828   addPass(createR600ControlFlowFinalizer(), false);
829 }
830 
831 TargetPassConfig *R600TargetMachine::createPassConfig(PassManagerBase &PM) {
832   return new R600PassConfig(*this, PM);
833 }
834 
835 //===----------------------------------------------------------------------===//
836 // GCN Pass Setup
837 //===----------------------------------------------------------------------===//
838 
839 ScheduleDAGInstrs *GCNPassConfig::createMachineScheduler(
840   MachineSchedContext *C) const {
841   const GCNSubtarget &ST = C->MF->getSubtarget<GCNSubtarget>();
842   if (ST.enableSIScheduler())
843     return createSIMachineScheduler(C);
844   return createGCNMaxOccupancyMachineScheduler(C);
845 }
846 
847 bool GCNPassConfig::addPreISel() {
848   AMDGPUPassConfig::addPreISel();
849 
850   if (EnableAtomicOptimizations) {
851     addPass(createAMDGPUAtomicOptimizerPass());
852   }
853 
854   // FIXME: We need to run a pass to propagate the attributes when calls are
855   // supported.
856 
857   // Merge divergent exit nodes. StructurizeCFG won't recognize the multi-exit
858   // regions formed by them.
859   addPass(&AMDGPUUnifyDivergentExitNodesID);
860   if (!LateCFGStructurize) {
861     addPass(createStructurizeCFGPass(true)); // true -> SkipUniformRegions
862   }
863   addPass(createSinkingPass());
864   addPass(createAMDGPUAnnotateUniformValues());
865   if (!LateCFGStructurize) {
866     addPass(createSIAnnotateControlFlowPass());
867   }
868   addPass(createLCSSAPass());
869 
870   return false;
871 }
872 
873 void GCNPassConfig::addMachineSSAOptimization() {
874   TargetPassConfig::addMachineSSAOptimization();
875 
876   // We want to fold operands after PeepholeOptimizer has run (or as part of
877   // it), because it will eliminate extra copies making it easier to fold the
878   // real source operand. We want to eliminate dead instructions after, so that
879   // we see fewer uses of the copies. We then need to clean up the dead
880   // instructions leftover after the operands are folded as well.
881   //
882   // XXX - Can we get away without running DeadMachineInstructionElim again?
883   addPass(&SIFoldOperandsID);
884   if (EnableDPPCombine)
885     addPass(&GCNDPPCombineID);
886   addPass(&DeadMachineInstructionElimID);
887   addPass(&SILoadStoreOptimizerID);
888   if (EnableSDWAPeephole) {
889     addPass(&SIPeepholeSDWAID);
890     addPass(&EarlyMachineLICMID);
891     addPass(&MachineCSEID);
892     addPass(&SIFoldOperandsID);
893     addPass(&DeadMachineInstructionElimID);
894   }
895   addPass(createSIShrinkInstructionsPass());
896 }
897 
898 bool GCNPassConfig::addILPOpts() {
899   if (EnableEarlyIfConversion)
900     addPass(&EarlyIfConverterID);
901 
902   TargetPassConfig::addILPOpts();
903   return false;
904 }
905 
906 bool GCNPassConfig::addInstSelector() {
907   AMDGPUPassConfig::addInstSelector();
908   addPass(&SIFixSGPRCopiesID);
909   addPass(createSILowerI1CopiesPass());
910   addPass(createSIFixupVectorISelPass());
911   addPass(createSIAddIMGInitPass());
912   return false;
913 }
914 
915 bool GCNPassConfig::addIRTranslator() {
916   addPass(new IRTranslator());
917   return false;
918 }
919 
920 void GCNPassConfig::addPreLegalizeMachineIR() {
921   bool IsOptNone = getOptLevel() == CodeGenOpt::None;
922   addPass(createAMDGPUPreLegalizeCombiner(IsOptNone));
923   addPass(new Localizer());
924 }
925 
926 bool GCNPassConfig::addLegalizeMachineIR() {
927   addPass(new Legalizer());
928   return false;
929 }
930 
931 void GCNPassConfig::addPreRegBankSelect() {
932   bool IsOptNone = getOptLevel() == CodeGenOpt::None;
933   addPass(createAMDGPUPostLegalizeCombiner(IsOptNone));
934 }
935 
936 bool GCNPassConfig::addRegBankSelect() {
937   addPass(new RegBankSelect());
938   return false;
939 }
940 
941 bool GCNPassConfig::addGlobalInstructionSelect() {
942   addPass(new InstructionSelect());
943   return false;
944 }
945 
946 void GCNPassConfig::addPreRegAlloc() {
947   if (LateCFGStructurize) {
948     addPass(createAMDGPUMachineCFGStructurizerPass());
949   }
950   addPass(createSIWholeQuadModePass());
951 }
952 
953 void GCNPassConfig::addFastRegAlloc() {
954   // FIXME: We have to disable the verifier here because of PHIElimination +
955   // TwoAddressInstructions disabling it.
956 
957   // This must be run immediately after phi elimination and before
958   // TwoAddressInstructions, otherwise the processing of the tied operand of
959   // SI_ELSE will introduce a copy of the tied operand source after the else.
960   insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
961 
962   // This must be run just after RegisterCoalescing.
963   insertPass(&RegisterCoalescerID, &SIPreAllocateWWMRegsID, false);
964 
965   TargetPassConfig::addFastRegAlloc();
966 }
967 
968 void GCNPassConfig::addOptimizedRegAlloc() {
969   if (OptExecMaskPreRA) {
970     insertPass(&MachineSchedulerID, &SIOptimizeExecMaskingPreRAID);
971     insertPass(&SIOptimizeExecMaskingPreRAID, &SIFormMemoryClausesID);
972   } else {
973     insertPass(&MachineSchedulerID, &SIFormMemoryClausesID);
974   }
975 
976   // This must be run immediately after phi elimination and before
977   // TwoAddressInstructions, otherwise the processing of the tied operand of
978   // SI_ELSE will introduce a copy of the tied operand source after the else.
979   insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
980 
981   // This must be run just after RegisterCoalescing.
982   insertPass(&RegisterCoalescerID, &SIPreAllocateWWMRegsID, false);
983 
984   if (EnableDCEInRA)
985     insertPass(&DetectDeadLanesID, &DeadMachineInstructionElimID);
986 
987   TargetPassConfig::addOptimizedRegAlloc();
988 }
989 
990 bool GCNPassConfig::addPreRewrite() {
991   if (EnableRegReassign) {
992     addPass(&GCNNSAReassignID);
993     addPass(&GCNRegBankReassignID);
994   }
995   return true;
996 }
997 
998 void GCNPassConfig::addPostRegAlloc() {
999   addPass(&SIFixVGPRCopiesID);
1000   if (getOptLevel() > CodeGenOpt::None)
1001     addPass(&SIOptimizeExecMaskingID);
1002   TargetPassConfig::addPostRegAlloc();
1003 
1004   // Equivalent of PEI for SGPRs.
1005   addPass(&SILowerSGPRSpillsID);
1006 }
1007 
1008 void GCNPassConfig::addPreSched2() {
1009   addPass(&SIPostRABundlerID);
1010 }
1011 
1012 void GCNPassConfig::addPreEmitPass() {
1013   addPass(createSIMemoryLegalizerPass());
1014   addPass(createSIInsertWaitcntsPass());
1015   addPass(createSIShrinkInstructionsPass());
1016   addPass(createSIModeRegisterPass());
1017 
1018   // The hazard recognizer that runs as part of the post-ra scheduler does not
1019   // guarantee to be able handle all hazards correctly. This is because if there
1020   // are multiple scheduling regions in a basic block, the regions are scheduled
1021   // bottom up, so when we begin to schedule a region we don't know what
1022   // instructions were emitted directly before it.
1023   //
1024   // Here we add a stand-alone hazard recognizer pass which can handle all
1025   // cases.
1026   //
1027   // FIXME: This stand-alone pass will emit indiv. S_NOP 0, as needed. It would
1028   // be better for it to emit S_NOP <N> when possible.
1029   addPass(&PostRAHazardRecognizerID);
1030 
1031   addPass(&SIRemoveShortExecBranchesID);
1032   addPass(&SIInsertSkipsPassID);
1033   addPass(&BranchRelaxationPassID);
1034 }
1035 
1036 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) {
1037   return new GCNPassConfig(*this, PM);
1038 }
1039 
1040 yaml::MachineFunctionInfo *GCNTargetMachine::createDefaultFuncInfoYAML() const {
1041   return new yaml::SIMachineFunctionInfo();
1042 }
1043 
1044 yaml::MachineFunctionInfo *
1045 GCNTargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const {
1046   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1047   return new yaml::SIMachineFunctionInfo(*MFI,
1048                                          *MF.getSubtarget().getRegisterInfo());
1049 }
1050 
1051 bool GCNTargetMachine::parseMachineFunctionInfo(
1052     const yaml::MachineFunctionInfo &MFI_, PerFunctionMIParsingState &PFS,
1053     SMDiagnostic &Error, SMRange &SourceRange) const {
1054   const yaml::SIMachineFunctionInfo &YamlMFI =
1055       reinterpret_cast<const yaml::SIMachineFunctionInfo &>(MFI_);
1056   MachineFunction &MF = PFS.MF;
1057   SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1058 
1059   MFI->initializeBaseYamlFields(YamlMFI);
1060 
1061   auto parseRegister = [&](const yaml::StringValue &RegName, Register &RegVal) {
1062     // FIXME: Update parseNamedRegsiterReference to take a Register.
1063     unsigned TempReg;
1064     if (parseNamedRegisterReference(PFS, TempReg, RegName.Value, Error)) {
1065       SourceRange = RegName.SourceRange;
1066       return true;
1067     }
1068     RegVal = TempReg;
1069 
1070     return false;
1071   };
1072 
1073   auto diagnoseRegisterClass = [&](const yaml::StringValue &RegName) {
1074     // Create a diagnostic for a the register string literal.
1075     const MemoryBuffer &Buffer =
1076         *PFS.SM->getMemoryBuffer(PFS.SM->getMainFileID());
1077     Error = SMDiagnostic(*PFS.SM, SMLoc(), Buffer.getBufferIdentifier(), 1,
1078                          RegName.Value.size(), SourceMgr::DK_Error,
1079                          "incorrect register class for field", RegName.Value,
1080                          None, None);
1081     SourceRange = RegName.SourceRange;
1082     return true;
1083   };
1084 
1085   if (parseRegister(YamlMFI.ScratchRSrcReg, MFI->ScratchRSrcReg) ||
1086       parseRegister(YamlMFI.FrameOffsetReg, MFI->FrameOffsetReg) ||
1087       parseRegister(YamlMFI.StackPtrOffsetReg, MFI->StackPtrOffsetReg))
1088     return true;
1089 
1090   if (MFI->ScratchRSrcReg != AMDGPU::PRIVATE_RSRC_REG &&
1091       !AMDGPU::SGPR_128RegClass.contains(MFI->ScratchRSrcReg)) {
1092     return diagnoseRegisterClass(YamlMFI.ScratchRSrcReg);
1093   }
1094 
1095   if (MFI->FrameOffsetReg != AMDGPU::FP_REG &&
1096       !AMDGPU::SGPR_32RegClass.contains(MFI->FrameOffsetReg)) {
1097     return diagnoseRegisterClass(YamlMFI.FrameOffsetReg);
1098   }
1099 
1100   if (MFI->StackPtrOffsetReg != AMDGPU::SP_REG &&
1101       !AMDGPU::SGPR_32RegClass.contains(MFI->StackPtrOffsetReg)) {
1102     return diagnoseRegisterClass(YamlMFI.StackPtrOffsetReg);
1103   }
1104 
1105   auto parseAndCheckArgument = [&](const Optional<yaml::SIArgument> &A,
1106                                    const TargetRegisterClass &RC,
1107                                    ArgDescriptor &Arg, unsigned UserSGPRs,
1108                                    unsigned SystemSGPRs) {
1109     // Skip parsing if it's not present.
1110     if (!A)
1111       return false;
1112 
1113     if (A->IsRegister) {
1114       unsigned Reg;
1115       if (parseNamedRegisterReference(PFS, Reg, A->RegisterName.Value, Error)) {
1116         SourceRange = A->RegisterName.SourceRange;
1117         return true;
1118       }
1119       if (!RC.contains(Reg))
1120         return diagnoseRegisterClass(A->RegisterName);
1121       Arg = ArgDescriptor::createRegister(Reg);
1122     } else
1123       Arg = ArgDescriptor::createStack(A->StackOffset);
1124     // Check and apply the optional mask.
1125     if (A->Mask)
1126       Arg = ArgDescriptor::createArg(Arg, A->Mask.getValue());
1127 
1128     MFI->NumUserSGPRs += UserSGPRs;
1129     MFI->NumSystemSGPRs += SystemSGPRs;
1130     return false;
1131   };
1132 
1133   if (YamlMFI.ArgInfo &&
1134       (parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentBuffer,
1135                              AMDGPU::SGPR_128RegClass,
1136                              MFI->ArgInfo.PrivateSegmentBuffer, 4, 0) ||
1137        parseAndCheckArgument(YamlMFI.ArgInfo->DispatchPtr,
1138                              AMDGPU::SReg_64RegClass, MFI->ArgInfo.DispatchPtr,
1139                              2, 0) ||
1140        parseAndCheckArgument(YamlMFI.ArgInfo->QueuePtr, AMDGPU::SReg_64RegClass,
1141                              MFI->ArgInfo.QueuePtr, 2, 0) ||
1142        parseAndCheckArgument(YamlMFI.ArgInfo->KernargSegmentPtr,
1143                              AMDGPU::SReg_64RegClass,
1144                              MFI->ArgInfo.KernargSegmentPtr, 2, 0) ||
1145        parseAndCheckArgument(YamlMFI.ArgInfo->DispatchID,
1146                              AMDGPU::SReg_64RegClass, MFI->ArgInfo.DispatchID,
1147                              2, 0) ||
1148        parseAndCheckArgument(YamlMFI.ArgInfo->FlatScratchInit,
1149                              AMDGPU::SReg_64RegClass,
1150                              MFI->ArgInfo.FlatScratchInit, 2, 0) ||
1151        parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentSize,
1152                              AMDGPU::SGPR_32RegClass,
1153                              MFI->ArgInfo.PrivateSegmentSize, 0, 0) ||
1154        parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDX,
1155                              AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDX,
1156                              0, 1) ||
1157        parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDY,
1158                              AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDY,
1159                              0, 1) ||
1160        parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDZ,
1161                              AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDZ,
1162                              0, 1) ||
1163        parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupInfo,
1164                              AMDGPU::SGPR_32RegClass,
1165                              MFI->ArgInfo.WorkGroupInfo, 0, 1) ||
1166        parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentWaveByteOffset,
1167                              AMDGPU::SGPR_32RegClass,
1168                              MFI->ArgInfo.PrivateSegmentWaveByteOffset, 0, 1) ||
1169        parseAndCheckArgument(YamlMFI.ArgInfo->ImplicitArgPtr,
1170                              AMDGPU::SReg_64RegClass,
1171                              MFI->ArgInfo.ImplicitArgPtr, 0, 0) ||
1172        parseAndCheckArgument(YamlMFI.ArgInfo->ImplicitBufferPtr,
1173                              AMDGPU::SReg_64RegClass,
1174                              MFI->ArgInfo.ImplicitBufferPtr, 2, 0) ||
1175        parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDX,
1176                              AMDGPU::VGPR_32RegClass,
1177                              MFI->ArgInfo.WorkItemIDX, 0, 0) ||
1178        parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDY,
1179                              AMDGPU::VGPR_32RegClass,
1180                              MFI->ArgInfo.WorkItemIDY, 0, 0) ||
1181        parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDZ,
1182                              AMDGPU::VGPR_32RegClass,
1183                              MFI->ArgInfo.WorkItemIDZ, 0, 0)))
1184     return true;
1185 
1186   MFI->Mode.IEEE = YamlMFI.Mode.IEEE;
1187   MFI->Mode.DX10Clamp = YamlMFI.Mode.DX10Clamp;
1188   MFI->Mode.FP32InputDenormals = YamlMFI.Mode.FP32InputDenormals;
1189   MFI->Mode.FP32OutputDenormals = YamlMFI.Mode.FP32OutputDenormals;
1190   MFI->Mode.FP64FP16InputDenormals = YamlMFI.Mode.FP64FP16InputDenormals;
1191   MFI->Mode.FP64FP16OutputDenormals = YamlMFI.Mode.FP64FP16OutputDenormals;
1192 
1193   return false;
1194 }
1195