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 "AMDGPUExportClustering.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 "MCTargetDesc/AMDGPUMCTargetDesc.h"
28 #include "R600MachineScheduler.h"
29 #include "SIMachineFunctionInfo.h"
30 #include "SIMachineScheduler.h"
31 #include "TargetInfo/AMDGPUTargetInfo.h"
32 #include "llvm/CodeGen/GlobalISel/IRTranslator.h"
33 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h"
34 #include "llvm/CodeGen/GlobalISel/Legalizer.h"
35 #include "llvm/CodeGen/GlobalISel/Localizer.h"
36 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h"
37 #include "llvm/CodeGen/MIRParser/MIParser.h"
38 #include "llvm/CodeGen/Passes.h"
39 #include "llvm/CodeGen/TargetPassConfig.h"
40 #include "llvm/IR/Attributes.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/LegacyPassManager.h"
43 #include "llvm/InitializePasses.h"
44 #include "llvm/Pass.h"
45 #include "llvm/Support/CommandLine.h"
46 #include "llvm/Support/Compiler.h"
47 #include "llvm/Support/TargetRegistry.h"
48 #include "llvm/Target/TargetLoweringObjectFile.h"
49 #include "llvm/Transforms/IPO.h"
50 #include "llvm/Transforms/IPO/AlwaysInliner.h"
51 #include "llvm/Transforms/IPO/PassManagerBuilder.h"
52 #include "llvm/Transforms/Scalar.h"
53 #include "llvm/Transforms/Scalar/GVN.h"
54 #include "llvm/Transforms/Utils.h"
55 #include "llvm/Transforms/Vectorize.h"
56 #include <memory>
57 
58 using namespace llvm;
59 
60 static cl::opt<bool> EnableR600StructurizeCFG(
61   "r600-ir-structurize",
62   cl::desc("Use StructurizeCFG IR pass"),
63   cl::init(true));
64 
65 static cl::opt<bool> EnableSROA(
66   "amdgpu-sroa",
67   cl::desc("Run SROA after promote alloca pass"),
68   cl::ReallyHidden,
69   cl::init(true));
70 
71 static cl::opt<bool>
72 EnableEarlyIfConversion("amdgpu-early-ifcvt", cl::Hidden,
73                         cl::desc("Run early if-conversion"),
74                         cl::init(false));
75 
76 static cl::opt<bool>
77 OptExecMaskPreRA("amdgpu-opt-exec-mask-pre-ra", cl::Hidden,
78             cl::desc("Run pre-RA exec mask optimizations"),
79             cl::init(true));
80 
81 static cl::opt<bool> EnableR600IfConvert(
82   "r600-if-convert",
83   cl::desc("Use if conversion pass"),
84   cl::ReallyHidden,
85   cl::init(true));
86 
87 // Option to disable vectorizer for tests.
88 static cl::opt<bool> EnableLoadStoreVectorizer(
89   "amdgpu-load-store-vectorizer",
90   cl::desc("Enable load store vectorizer"),
91   cl::init(true),
92   cl::Hidden);
93 
94 // Option to control global loads scalarization
95 static cl::opt<bool> ScalarizeGlobal(
96   "amdgpu-scalarize-global-loads",
97   cl::desc("Enable global load scalarization"),
98   cl::init(true),
99   cl::Hidden);
100 
101 // Option to run internalize pass.
102 static cl::opt<bool> InternalizeSymbols(
103   "amdgpu-internalize-symbols",
104   cl::desc("Enable elimination of non-kernel functions and unused globals"),
105   cl::init(false),
106   cl::Hidden);
107 
108 // Option to inline all early.
109 static cl::opt<bool> EarlyInlineAll(
110   "amdgpu-early-inline-all",
111   cl::desc("Inline all functions early"),
112   cl::init(false),
113   cl::Hidden);
114 
115 static cl::opt<bool> EnableSDWAPeephole(
116   "amdgpu-sdwa-peephole",
117   cl::desc("Enable SDWA peepholer"),
118   cl::init(true));
119 
120 static cl::opt<bool> EnableDPPCombine(
121   "amdgpu-dpp-combine",
122   cl::desc("Enable DPP combiner"),
123   cl::init(true));
124 
125 // Enable address space based alias analysis
126 static cl::opt<bool> EnableAMDGPUAliasAnalysis("enable-amdgpu-aa", cl::Hidden,
127   cl::desc("Enable AMDGPU Alias Analysis"),
128   cl::init(true));
129 
130 // Option to run late CFG structurizer
131 static cl::opt<bool, true> LateCFGStructurize(
132   "amdgpu-late-structurize",
133   cl::desc("Enable late CFG structurization"),
134   cl::location(AMDGPUTargetMachine::EnableLateStructurizeCFG),
135   cl::Hidden);
136 
137 static cl::opt<bool, true> EnableAMDGPUFunctionCallsOpt(
138   "amdgpu-function-calls",
139   cl::desc("Enable AMDGPU function call support"),
140   cl::location(AMDGPUTargetMachine::EnableFunctionCalls),
141   cl::init(true),
142   cl::Hidden);
143 
144 static cl::opt<bool, true> EnableAMDGPUFixedFunctionABIOpt(
145   "amdgpu-fixed-function-abi",
146   cl::desc("Enable all implicit function arguments"),
147   cl::location(AMDGPUTargetMachine::EnableFixedFunctionABI),
148   cl::init(false),
149   cl::Hidden);
150 
151 // Enable lib calls simplifications
152 static cl::opt<bool> EnableLibCallSimplify(
153   "amdgpu-simplify-libcall",
154   cl::desc("Enable amdgpu library simplifications"),
155   cl::init(true),
156   cl::Hidden);
157 
158 static cl::opt<bool> EnableLowerKernelArguments(
159   "amdgpu-ir-lower-kernel-arguments",
160   cl::desc("Lower kernel argument loads in IR pass"),
161   cl::init(true),
162   cl::Hidden);
163 
164 static cl::opt<bool> EnableRegReassign(
165   "amdgpu-reassign-regs",
166   cl::desc("Enable register reassign optimizations on gfx10+"),
167   cl::init(true),
168   cl::Hidden);
169 
170 // Enable atomic optimization
171 static cl::opt<bool> EnableAtomicOptimizations(
172   "amdgpu-atomic-optimizations",
173   cl::desc("Enable atomic optimizations"),
174   cl::init(false),
175   cl::Hidden);
176 
177 // Enable Mode register optimization
178 static cl::opt<bool> EnableSIModeRegisterPass(
179   "amdgpu-mode-register",
180   cl::desc("Enable mode register pass"),
181   cl::init(true),
182   cl::Hidden);
183 
184 // Option is used in lit tests to prevent deadcoding of patterns inspected.
185 static cl::opt<bool>
186 EnableDCEInRA("amdgpu-dce-in-ra",
187     cl::init(true), cl::Hidden,
188     cl::desc("Enable machine DCE inside regalloc"));
189 
190 static cl::opt<bool> EnableScalarIRPasses(
191   "amdgpu-scalar-ir-passes",
192   cl::desc("Enable scalar IR passes"),
193   cl::init(true),
194   cl::Hidden);
195 
196 static cl::opt<bool> EnableStructurizerWorkarounds(
197     "amdgpu-enable-structurizer-workarounds",
198     cl::desc("Enable workarounds for the StructurizeCFG pass"), cl::init(true),
199     cl::Hidden);
200 
201 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAMDGPUTarget() {
202   // Register the target
203   RegisterTargetMachine<R600TargetMachine> X(getTheAMDGPUTarget());
204   RegisterTargetMachine<GCNTargetMachine> Y(getTheGCNTarget());
205 
206   PassRegistry *PR = PassRegistry::getPassRegistry();
207   initializeR600ClauseMergePassPass(*PR);
208   initializeR600ControlFlowFinalizerPass(*PR);
209   initializeR600PacketizerPass(*PR);
210   initializeR600ExpandSpecialInstrsPassPass(*PR);
211   initializeR600VectorRegMergerPass(*PR);
212   initializeGlobalISel(*PR);
213   initializeAMDGPUDAGToDAGISelPass(*PR);
214   initializeGCNDPPCombinePass(*PR);
215   initializeSILowerI1CopiesPass(*PR);
216   initializeSILowerSGPRSpillsPass(*PR);
217   initializeSIFixSGPRCopiesPass(*PR);
218   initializeSIFixVGPRCopiesPass(*PR);
219   initializeSIFoldOperandsPass(*PR);
220   initializeSIPeepholeSDWAPass(*PR);
221   initializeSIShrinkInstructionsPass(*PR);
222   initializeSIOptimizeExecMaskingPreRAPass(*PR);
223   initializeSILoadStoreOptimizerPass(*PR);
224   initializeAMDGPUFixFunctionBitcastsPass(*PR);
225   initializeAMDGPUAlwaysInlinePass(*PR);
226   initializeAMDGPUAnnotateKernelFeaturesPass(*PR);
227   initializeAMDGPUAnnotateUniformValuesPass(*PR);
228   initializeAMDGPUArgumentUsageInfoPass(*PR);
229   initializeAMDGPUAtomicOptimizerPass(*PR);
230   initializeAMDGPULowerKernelArgumentsPass(*PR);
231   initializeAMDGPULowerKernelAttributesPass(*PR);
232   initializeAMDGPULowerIntrinsicsPass(*PR);
233   initializeAMDGPUOpenCLEnqueuedBlockLoweringPass(*PR);
234   initializeAMDGPUPostLegalizerCombinerPass(*PR);
235   initializeAMDGPUPreLegalizerCombinerPass(*PR);
236   initializeAMDGPUPromoteAllocaPass(*PR);
237   initializeAMDGPUPromoteAllocaToVectorPass(*PR);
238   initializeAMDGPUCodeGenPreparePass(*PR);
239   initializeAMDGPUPropagateAttributesEarlyPass(*PR);
240   initializeAMDGPUPropagateAttributesLatePass(*PR);
241   initializeAMDGPURewriteOutArgumentsPass(*PR);
242   initializeAMDGPUUnifyMetadataPass(*PR);
243   initializeSIAnnotateControlFlowPass(*PR);
244   initializeSIInsertHardClausesPass(*PR);
245   initializeSIInsertWaitcntsPass(*PR);
246   initializeSIModeRegisterPass(*PR);
247   initializeSIWholeQuadModePass(*PR);
248   initializeSILowerControlFlowPass(*PR);
249   initializeSIRemoveShortExecBranchesPass(*PR);
250   initializeSIPreEmitPeepholePass(*PR);
251   initializeSIInsertSkipsPass(*PR);
252   initializeSIMemoryLegalizerPass(*PR);
253   initializeSIOptimizeExecMaskingPass(*PR);
254   initializeSIPreAllocateWWMRegsPass(*PR);
255   initializeSIFormMemoryClausesPass(*PR);
256   initializeSIPostRABundlerPass(*PR);
257   initializeAMDGPUUnifyDivergentExitNodesPass(*PR);
258   initializeAMDGPUAAWrapperPassPass(*PR);
259   initializeAMDGPUExternalAAWrapperPass(*PR);
260   initializeAMDGPUUseNativeCallsPass(*PR);
261   initializeAMDGPUSimplifyLibCallsPass(*PR);
262   initializeAMDGPUInlinerPass(*PR);
263   initializeAMDGPUPrintfRuntimeBindingPass(*PR);
264   initializeGCNRegBankReassignPass(*PR);
265   initializeGCNNSAReassignPass(*PR);
266   initializeSIAddIMGInitPass(*PR);
267 }
268 
269 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
270   return std::make_unique<AMDGPUTargetObjectFile>();
271 }
272 
273 static ScheduleDAGInstrs *createR600MachineScheduler(MachineSchedContext *C) {
274   return new ScheduleDAGMILive(C, std::make_unique<R600SchedStrategy>());
275 }
276 
277 static ScheduleDAGInstrs *createSIMachineScheduler(MachineSchedContext *C) {
278   return new SIScheduleDAGMI(C);
279 }
280 
281 static ScheduleDAGInstrs *
282 createGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
283   ScheduleDAGMILive *DAG =
284     new GCNScheduleDAGMILive(C, std::make_unique<GCNMaxOccupancySchedStrategy>(C));
285   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
286   DAG->addMutation(createAMDGPUMacroFusionDAGMutation());
287   DAG->addMutation(createAMDGPUExportClusteringDAGMutation());
288   return DAG;
289 }
290 
291 static ScheduleDAGInstrs *
292 createIterativeGCNMaxOccupancyMachineScheduler(MachineSchedContext *C) {
293   auto DAG = new GCNIterativeScheduler(C,
294     GCNIterativeScheduler::SCHEDULE_LEGACYMAXOCCUPANCY);
295   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
296   return DAG;
297 }
298 
299 static ScheduleDAGInstrs *createMinRegScheduler(MachineSchedContext *C) {
300   return new GCNIterativeScheduler(C,
301     GCNIterativeScheduler::SCHEDULE_MINREGFORCED);
302 }
303 
304 static ScheduleDAGInstrs *
305 createIterativeILPMachineScheduler(MachineSchedContext *C) {
306   auto DAG = new GCNIterativeScheduler(C,
307     GCNIterativeScheduler::SCHEDULE_ILP);
308   DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
309   DAG->addMutation(createAMDGPUMacroFusionDAGMutation());
310   return DAG;
311 }
312 
313 static MachineSchedRegistry
314 R600SchedRegistry("r600", "Run R600's custom scheduler",
315                    createR600MachineScheduler);
316 
317 static MachineSchedRegistry
318 SISchedRegistry("si", "Run SI's custom scheduler",
319                 createSIMachineScheduler);
320 
321 static MachineSchedRegistry
322 GCNMaxOccupancySchedRegistry("gcn-max-occupancy",
323                              "Run GCN scheduler to maximize occupancy",
324                              createGCNMaxOccupancyMachineScheduler);
325 
326 static MachineSchedRegistry
327 IterativeGCNMaxOccupancySchedRegistry("gcn-max-occupancy-experimental",
328   "Run GCN scheduler to maximize occupancy (experimental)",
329   createIterativeGCNMaxOccupancyMachineScheduler);
330 
331 static MachineSchedRegistry
332 GCNMinRegSchedRegistry("gcn-minreg",
333   "Run GCN iterative scheduler for minimal register usage (experimental)",
334   createMinRegScheduler);
335 
336 static MachineSchedRegistry
337 GCNILPSchedRegistry("gcn-ilp",
338   "Run GCN iterative scheduler for ILP scheduling (experimental)",
339   createIterativeILPMachineScheduler);
340 
341 static StringRef computeDataLayout(const Triple &TT) {
342   if (TT.getArch() == Triple::r600) {
343     // 32-bit pointers.
344       return "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
345              "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5";
346   }
347 
348   // 32-bit private, local, and region pointers. 64-bit global, constant and
349   // flat, non-integral buffer fat pointers.
350     return "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32"
351          "-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128"
352          "-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5"
353          "-ni:7";
354 }
355 
356 LLVM_READNONE
357 static StringRef getGPUOrDefault(const Triple &TT, StringRef GPU) {
358   if (!GPU.empty())
359     return GPU;
360 
361   // Need to default to a target with flat support for HSA.
362   if (TT.getArch() == Triple::amdgcn)
363     return TT.getOS() == Triple::AMDHSA ? "generic-hsa" : "generic";
364 
365   return "r600";
366 }
367 
368 static Reloc::Model getEffectiveRelocModel(Optional<Reloc::Model> RM) {
369   // The AMDGPU toolchain only supports generating shared objects, so we
370   // must always use PIC.
371   return Reloc::PIC_;
372 }
373 
374 AMDGPUTargetMachine::AMDGPUTargetMachine(const Target &T, const Triple &TT,
375                                          StringRef CPU, StringRef FS,
376                                          TargetOptions Options,
377                                          Optional<Reloc::Model> RM,
378                                          Optional<CodeModel::Model> CM,
379                                          CodeGenOpt::Level OptLevel)
380     : LLVMTargetMachine(T, computeDataLayout(TT), TT, getGPUOrDefault(TT, CPU),
381                         FS, Options, getEffectiveRelocModel(RM),
382                         getEffectiveCodeModel(CM, CodeModel::Small), OptLevel),
383       TLOF(createTLOF(getTargetTriple())) {
384   initAsmInfo();
385   if (TT.getArch() == Triple::amdgcn) {
386     if (getMCSubtargetInfo()->checkFeatures("+wavefrontsize64"))
387       MRI.reset(llvm::createGCNMCRegisterInfo(AMDGPUDwarfFlavour::Wave64));
388     else if (getMCSubtargetInfo()->checkFeatures("+wavefrontsize32"))
389       MRI.reset(llvm::createGCNMCRegisterInfo(AMDGPUDwarfFlavour::Wave32));
390   }
391 }
392 
393 bool AMDGPUTargetMachine::EnableLateStructurizeCFG = false;
394 bool AMDGPUTargetMachine::EnableFunctionCalls = false;
395 bool AMDGPUTargetMachine::EnableFixedFunctionABI = false;
396 
397 AMDGPUTargetMachine::~AMDGPUTargetMachine() = default;
398 
399 StringRef AMDGPUTargetMachine::getGPUName(const Function &F) const {
400   Attribute GPUAttr = F.getFnAttribute("target-cpu");
401   return GPUAttr.isValid() ? GPUAttr.getValueAsString() : getTargetCPU();
402 }
403 
404 StringRef AMDGPUTargetMachine::getFeatureString(const Function &F) const {
405   Attribute FSAttr = F.getFnAttribute("target-features");
406 
407   return FSAttr.isValid() ? FSAttr.getValueAsString()
408                           : getTargetFeatureString();
409 }
410 
411 /// Predicate for Internalize pass.
412 static bool mustPreserveGV(const GlobalValue &GV) {
413   if (const Function *F = dyn_cast<Function>(&GV))
414     return F->isDeclaration() || AMDGPU::isEntryFunctionCC(F->getCallingConv());
415 
416   return !GV.use_empty();
417 }
418 
419 void AMDGPUTargetMachine::adjustPassManager(PassManagerBuilder &Builder) {
420   Builder.DivergentTarget = true;
421 
422   bool EnableOpt = getOptLevel() > CodeGenOpt::None;
423   bool Internalize = InternalizeSymbols;
424   bool EarlyInline = EarlyInlineAll && EnableOpt && !EnableFunctionCalls;
425   bool AMDGPUAA = EnableAMDGPUAliasAnalysis && EnableOpt;
426   bool LibCallSimplify = EnableLibCallSimplify && EnableOpt;
427 
428   if (EnableFunctionCalls) {
429     delete Builder.Inliner;
430     Builder.Inliner = createAMDGPUFunctionInliningPass();
431   }
432 
433   Builder.addExtension(
434     PassManagerBuilder::EP_ModuleOptimizerEarly,
435     [Internalize, EarlyInline, AMDGPUAA, this](const PassManagerBuilder &,
436                                                legacy::PassManagerBase &PM) {
437       if (AMDGPUAA) {
438         PM.add(createAMDGPUAAWrapperPass());
439         PM.add(createAMDGPUExternalAAWrapperPass());
440       }
441       PM.add(createAMDGPUUnifyMetadataPass());
442       PM.add(createAMDGPUPrintfRuntimeBinding());
443       if (Internalize)
444         PM.add(createInternalizePass(mustPreserveGV));
445       PM.add(createAMDGPUPropagateAttributesLatePass(this));
446       if (Internalize)
447         PM.add(createGlobalDCEPass());
448       if (EarlyInline)
449         PM.add(createAMDGPUAlwaysInlinePass(false));
450   });
451 
452   Builder.addExtension(
453     PassManagerBuilder::EP_EarlyAsPossible,
454     [AMDGPUAA, LibCallSimplify, this](const PassManagerBuilder &,
455                                       legacy::PassManagerBase &PM) {
456       if (AMDGPUAA) {
457         PM.add(createAMDGPUAAWrapperPass());
458         PM.add(createAMDGPUExternalAAWrapperPass());
459       }
460       PM.add(llvm::createAMDGPUPropagateAttributesEarlyPass(this));
461       PM.add(llvm::createAMDGPUUseNativeCallsPass());
462       if (LibCallSimplify)
463         PM.add(llvm::createAMDGPUSimplifyLibCallsPass(this));
464   });
465 
466   Builder.addExtension(
467     PassManagerBuilder::EP_CGSCCOptimizerLate,
468     [EnableOpt](const PassManagerBuilder &, legacy::PassManagerBase &PM) {
469       // Add infer address spaces pass to the opt pipeline after inlining
470       // but before SROA to increase SROA opportunities.
471       PM.add(createInferAddressSpacesPass());
472 
473       // This should run after inlining to have any chance of doing anything,
474       // and before other cleanup optimizations.
475       PM.add(createAMDGPULowerKernelAttributesPass());
476 
477       // Promote alloca to vector before SROA and loop unroll. If we manage
478       // to eliminate allocas before unroll we may choose to unroll less.
479       if (EnableOpt)
480         PM.add(createAMDGPUPromoteAllocaToVector());
481   });
482 
483   Builder.addExtension(
484       PassManagerBuilder::EP_LoopOptimizerEnd,
485       [](const PassManagerBuilder &, legacy::PassManagerBase &PM) {
486         // Add SROA after loop unrolling as more promotable patterns are
487         // exposed after small loops are fully unrolled.
488         PM.add(createSROAPass());
489       });
490 }
491 
492 //===----------------------------------------------------------------------===//
493 // R600 Target Machine (R600 -> Cayman)
494 //===----------------------------------------------------------------------===//
495 
496 R600TargetMachine::R600TargetMachine(const Target &T, const Triple &TT,
497                                      StringRef CPU, StringRef FS,
498                                      TargetOptions Options,
499                                      Optional<Reloc::Model> RM,
500                                      Optional<CodeModel::Model> CM,
501                                      CodeGenOpt::Level OL, bool JIT)
502     : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {
503   setRequiresStructuredCFG(true);
504 
505   // Override the default since calls aren't supported for r600.
506   if (EnableFunctionCalls &&
507       EnableAMDGPUFunctionCallsOpt.getNumOccurrences() == 0)
508     EnableFunctionCalls = false;
509 }
510 
511 const R600Subtarget *R600TargetMachine::getSubtargetImpl(
512   const Function &F) const {
513   StringRef GPU = getGPUName(F);
514   StringRef FS = getFeatureString(F);
515 
516   SmallString<128> SubtargetKey(GPU);
517   SubtargetKey.append(FS);
518 
519   auto &I = SubtargetMap[SubtargetKey];
520   if (!I) {
521     // This needs to be done before we create a new subtarget since any
522     // creation will depend on the TM and the code generation flags on the
523     // function that reside in TargetOptions.
524     resetTargetOptions(F);
525     I = std::make_unique<R600Subtarget>(TargetTriple, GPU, FS, *this);
526   }
527 
528   return I.get();
529 }
530 
531 bool AMDGPUTargetMachine::isNoopAddrSpaceCast(unsigned SrcAS,
532                                               unsigned DestAS) const {
533   return AMDGPU::isFlatGlobalAddrSpace(SrcAS) &&
534          AMDGPU::isFlatGlobalAddrSpace(DestAS);
535 }
536 
537 TargetTransformInfo
538 R600TargetMachine::getTargetTransformInfo(const Function &F) {
539   return TargetTransformInfo(R600TTIImpl(this, F));
540 }
541 
542 //===----------------------------------------------------------------------===//
543 // GCN Target Machine (SI+)
544 //===----------------------------------------------------------------------===//
545 
546 GCNTargetMachine::GCNTargetMachine(const Target &T, const Triple &TT,
547                                    StringRef CPU, StringRef FS,
548                                    TargetOptions Options,
549                                    Optional<Reloc::Model> RM,
550                                    Optional<CodeModel::Model> CM,
551                                    CodeGenOpt::Level OL, bool JIT)
552     : AMDGPUTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL) {}
553 
554 const GCNSubtarget *GCNTargetMachine::getSubtargetImpl(const Function &F) const {
555   StringRef GPU = getGPUName(F);
556   StringRef FS = getFeatureString(F);
557 
558   SmallString<128> SubtargetKey(GPU);
559   SubtargetKey.append(FS);
560 
561   auto &I = SubtargetMap[SubtargetKey];
562   if (!I) {
563     // This needs to be done before we create a new subtarget since any
564     // creation will depend on the TM and the code generation flags on the
565     // function that reside in TargetOptions.
566     resetTargetOptions(F);
567     I = std::make_unique<GCNSubtarget>(TargetTriple, GPU, FS, *this);
568   }
569 
570   I->setScalarizeGlobalBehavior(ScalarizeGlobal);
571 
572   return I.get();
573 }
574 
575 TargetTransformInfo
576 GCNTargetMachine::getTargetTransformInfo(const Function &F) {
577   return TargetTransformInfo(GCNTTIImpl(this, F));
578 }
579 
580 //===----------------------------------------------------------------------===//
581 // AMDGPU Pass Setup
582 //===----------------------------------------------------------------------===//
583 
584 namespace {
585 
586 class AMDGPUPassConfig : public TargetPassConfig {
587 public:
588   AMDGPUPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM)
589     : TargetPassConfig(TM, PM) {
590     // Exceptions and StackMaps are not supported, so these passes will never do
591     // anything.
592     disablePass(&StackMapLivenessID);
593     disablePass(&FuncletLayoutID);
594   }
595 
596   AMDGPUTargetMachine &getAMDGPUTargetMachine() const {
597     return getTM<AMDGPUTargetMachine>();
598   }
599 
600   ScheduleDAGInstrs *
601   createMachineScheduler(MachineSchedContext *C) const override {
602     ScheduleDAGMILive *DAG = createGenericSchedLive(C);
603     DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
604     return DAG;
605   }
606 
607   void addEarlyCSEOrGVNPass();
608   void addStraightLineScalarOptimizationPasses();
609   void addIRPasses() override;
610   void addCodeGenPrepare() override;
611   bool addPreISel() override;
612   bool addInstSelector() override;
613   bool addGCPasses() override;
614 
615   std::unique_ptr<CSEConfigBase> getCSEConfig() const override;
616 };
617 
618 std::unique_ptr<CSEConfigBase> AMDGPUPassConfig::getCSEConfig() const {
619   return getStandardCSEConfigForOpt(TM->getOptLevel());
620 }
621 
622 class R600PassConfig final : public AMDGPUPassConfig {
623 public:
624   R600PassConfig(LLVMTargetMachine &TM, PassManagerBase &PM)
625     : AMDGPUPassConfig(TM, PM) {}
626 
627   ScheduleDAGInstrs *createMachineScheduler(
628     MachineSchedContext *C) const override {
629     return createR600MachineScheduler(C);
630   }
631 
632   bool addPreISel() override;
633   bool addInstSelector() override;
634   void addPreRegAlloc() override;
635   void addPreSched2() override;
636   void addPreEmitPass() override;
637 };
638 
639 class GCNPassConfig final : public AMDGPUPassConfig {
640 public:
641   GCNPassConfig(LLVMTargetMachine &TM, PassManagerBase &PM)
642     : AMDGPUPassConfig(TM, PM) {
643     // It is necessary to know the register usage of the entire call graph.  We
644     // allow calls without EnableAMDGPUFunctionCalls if they are marked
645     // noinline, so this is always required.
646     setRequiresCodeGenSCCOrder(true);
647   }
648 
649   GCNTargetMachine &getGCNTargetMachine() const {
650     return getTM<GCNTargetMachine>();
651   }
652 
653   ScheduleDAGInstrs *
654   createMachineScheduler(MachineSchedContext *C) const override;
655 
656   bool addPreISel() override;
657   void addMachineSSAOptimization() override;
658   bool addILPOpts() override;
659   bool addInstSelector() override;
660   bool addIRTranslator() override;
661   void addPreLegalizeMachineIR() override;
662   bool addLegalizeMachineIR() override;
663   void addPreRegBankSelect() override;
664   bool addRegBankSelect() override;
665   bool addGlobalInstructionSelect() override;
666   void addFastRegAlloc() override;
667   void addOptimizedRegAlloc() override;
668   void addPreRegAlloc() override;
669   bool addPreRewrite() override;
670   void addPostRegAlloc() override;
671   void addPreSched2() override;
672   void addPreEmitPass() override;
673 };
674 
675 } // end anonymous namespace
676 
677 void AMDGPUPassConfig::addEarlyCSEOrGVNPass() {
678   if (getOptLevel() == CodeGenOpt::Aggressive)
679     addPass(createGVNPass());
680   else
681     addPass(createEarlyCSEPass());
682 }
683 
684 void AMDGPUPassConfig::addStraightLineScalarOptimizationPasses() {
685   addPass(createLICMPass());
686   addPass(createSeparateConstOffsetFromGEPPass());
687   addPass(createSpeculativeExecutionPass());
688   // ReassociateGEPs exposes more opportunites for SLSR. See
689   // the example in reassociate-geps-and-slsr.ll.
690   addPass(createStraightLineStrengthReducePass());
691   // SeparateConstOffsetFromGEP and SLSR creates common expressions which GVN or
692   // EarlyCSE can reuse.
693   addEarlyCSEOrGVNPass();
694   // Run NaryReassociate after EarlyCSE/GVN to be more effective.
695   addPass(createNaryReassociatePass());
696   // NaryReassociate on GEPs creates redundant common expressions, so run
697   // EarlyCSE after it.
698   addPass(createEarlyCSEPass());
699 }
700 
701 void AMDGPUPassConfig::addIRPasses() {
702   const AMDGPUTargetMachine &TM = getAMDGPUTargetMachine();
703 
704   // There is no reason to run these.
705   disablePass(&StackMapLivenessID);
706   disablePass(&FuncletLayoutID);
707   disablePass(&PatchableFunctionID);
708 
709   addPass(createAMDGPUPrintfRuntimeBinding());
710 
711   // This must occur before inlining, as the inliner will not look through
712   // bitcast calls.
713   addPass(createAMDGPUFixFunctionBitcastsPass());
714 
715   // A call to propagate attributes pass in the backend in case opt was not run.
716   addPass(createAMDGPUPropagateAttributesEarlyPass(&TM));
717 
718   addPass(createAtomicExpandPass());
719 
720 
721   addPass(createAMDGPULowerIntrinsicsPass());
722 
723   // Function calls are not supported, so make sure we inline everything.
724   addPass(createAMDGPUAlwaysInlinePass());
725   addPass(createAlwaysInlinerLegacyPass());
726   // We need to add the barrier noop pass, otherwise adding the function
727   // inlining pass will cause all of the PassConfigs passes to be run
728   // one function at a time, which means if we have a nodule with two
729   // functions, then we will generate code for the first function
730   // without ever running any passes on the second.
731   addPass(createBarrierNoopPass());
732 
733   // Handle uses of OpenCL image2d_t, image3d_t and sampler_t arguments.
734   if (TM.getTargetTriple().getArch() == Triple::r600)
735     addPass(createR600OpenCLImageTypeLoweringPass());
736 
737   // Replace OpenCL enqueued block function pointers with global variables.
738   addPass(createAMDGPUOpenCLEnqueuedBlockLoweringPass());
739 
740   if (TM.getOptLevel() > CodeGenOpt::None) {
741     addPass(createInferAddressSpacesPass());
742     addPass(createAMDGPUPromoteAlloca());
743 
744     if (EnableSROA)
745       addPass(createSROAPass());
746 
747     if (EnableScalarIRPasses)
748       addStraightLineScalarOptimizationPasses();
749 
750     if (EnableAMDGPUAliasAnalysis) {
751       addPass(createAMDGPUAAWrapperPass());
752       addPass(createExternalAAWrapperPass([](Pass &P, Function &,
753                                              AAResults &AAR) {
754         if (auto *WrapperPass = P.getAnalysisIfAvailable<AMDGPUAAWrapperPass>())
755           AAR.addAAResult(WrapperPass->getResult());
756         }));
757     }
758   }
759 
760   if (TM.getTargetTriple().getArch() == Triple::amdgcn) {
761     // TODO: May want to move later or split into an early and late one.
762     addPass(createAMDGPUCodeGenPreparePass());
763   }
764 
765   TargetPassConfig::addIRPasses();
766 
767   // EarlyCSE is not always strong enough to clean up what LSR produces. For
768   // example, GVN can combine
769   //
770   //   %0 = add %a, %b
771   //   %1 = add %b, %a
772   //
773   // and
774   //
775   //   %0 = shl nsw %a, 2
776   //   %1 = shl %a, 2
777   //
778   // but EarlyCSE can do neither of them.
779   if (getOptLevel() != CodeGenOpt::None && EnableScalarIRPasses)
780     addEarlyCSEOrGVNPass();
781 }
782 
783 void AMDGPUPassConfig::addCodeGenPrepare() {
784   if (TM->getTargetTriple().getArch() == Triple::amdgcn)
785     addPass(createAMDGPUAnnotateKernelFeaturesPass());
786 
787   if (TM->getTargetTriple().getArch() == Triple::amdgcn &&
788       EnableLowerKernelArguments)
789     addPass(createAMDGPULowerKernelArgumentsPass());
790 
791   addPass(&AMDGPUPerfHintAnalysisID);
792 
793   TargetPassConfig::addCodeGenPrepare();
794 
795   if (EnableLoadStoreVectorizer)
796     addPass(createLoadStoreVectorizerPass());
797 
798   // LowerSwitch pass may introduce unreachable blocks that can
799   // cause unexpected behavior for subsequent passes. Placing it
800   // here seems better that these blocks would get cleaned up by
801   // UnreachableBlockElim inserted next in the pass flow.
802   addPass(createLowerSwitchPass());
803 }
804 
805 bool AMDGPUPassConfig::addPreISel() {
806   addPass(createFlattenCFGPass());
807   return false;
808 }
809 
810 bool AMDGPUPassConfig::addInstSelector() {
811   // Defer the verifier until FinalizeISel.
812   addPass(createAMDGPUISelDag(&getAMDGPUTargetMachine(), getOptLevel()), false);
813   return false;
814 }
815 
816 bool AMDGPUPassConfig::addGCPasses() {
817   // Do nothing. GC is not supported.
818   return false;
819 }
820 
821 //===----------------------------------------------------------------------===//
822 // R600 Pass Setup
823 //===----------------------------------------------------------------------===//
824 
825 bool R600PassConfig::addPreISel() {
826   AMDGPUPassConfig::addPreISel();
827 
828   if (EnableR600StructurizeCFG)
829     addPass(createStructurizeCFGPass());
830   return false;
831 }
832 
833 bool R600PassConfig::addInstSelector() {
834   addPass(createR600ISelDag(&getAMDGPUTargetMachine(), getOptLevel()));
835   return false;
836 }
837 
838 void R600PassConfig::addPreRegAlloc() {
839   addPass(createR600VectorRegMerger());
840 }
841 
842 void R600PassConfig::addPreSched2() {
843   addPass(createR600EmitClauseMarkers(), false);
844   if (EnableR600IfConvert)
845     addPass(&IfConverterID, false);
846   addPass(createR600ClauseMergePass(), false);
847 }
848 
849 void R600PassConfig::addPreEmitPass() {
850   addPass(createAMDGPUCFGStructurizerPass(), false);
851   addPass(createR600ExpandSpecialInstrsPass(), false);
852   addPass(&FinalizeMachineBundlesID, false);
853   addPass(createR600Packetizer(), false);
854   addPass(createR600ControlFlowFinalizer(), false);
855 }
856 
857 TargetPassConfig *R600TargetMachine::createPassConfig(PassManagerBase &PM) {
858   return new R600PassConfig(*this, PM);
859 }
860 
861 //===----------------------------------------------------------------------===//
862 // GCN Pass Setup
863 //===----------------------------------------------------------------------===//
864 
865 ScheduleDAGInstrs *GCNPassConfig::createMachineScheduler(
866   MachineSchedContext *C) const {
867   const GCNSubtarget &ST = C->MF->getSubtarget<GCNSubtarget>();
868   if (ST.enableSIScheduler())
869     return createSIMachineScheduler(C);
870   return createGCNMaxOccupancyMachineScheduler(C);
871 }
872 
873 bool GCNPassConfig::addPreISel() {
874   AMDGPUPassConfig::addPreISel();
875 
876   if (EnableAtomicOptimizations) {
877     addPass(createAMDGPUAtomicOptimizerPass());
878   }
879 
880   // FIXME: We need to run a pass to propagate the attributes when calls are
881   // supported.
882 
883   // Merge divergent exit nodes. StructurizeCFG won't recognize the multi-exit
884   // regions formed by them.
885   addPass(&AMDGPUUnifyDivergentExitNodesID);
886   if (!LateCFGStructurize) {
887     if (EnableStructurizerWorkarounds) {
888       addPass(createFixIrreduciblePass());
889       addPass(createUnifyLoopExitsPass());
890     }
891     addPass(createStructurizeCFGPass(false)); // true -> SkipUniformRegions
892   }
893   addPass(createSinkingPass());
894   addPass(createAMDGPUAnnotateUniformValues());
895   if (!LateCFGStructurize) {
896     addPass(createSIAnnotateControlFlowPass());
897   }
898   addPass(createLCSSAPass());
899 
900   return false;
901 }
902 
903 void GCNPassConfig::addMachineSSAOptimization() {
904   TargetPassConfig::addMachineSSAOptimization();
905 
906   // We want to fold operands after PeepholeOptimizer has run (or as part of
907   // it), because it will eliminate extra copies making it easier to fold the
908   // real source operand. We want to eliminate dead instructions after, so that
909   // we see fewer uses of the copies. We then need to clean up the dead
910   // instructions leftover after the operands are folded as well.
911   //
912   // XXX - Can we get away without running DeadMachineInstructionElim again?
913   addPass(&SIFoldOperandsID);
914   if (EnableDPPCombine)
915     addPass(&GCNDPPCombineID);
916   addPass(&DeadMachineInstructionElimID);
917   addPass(&SILoadStoreOptimizerID);
918   if (EnableSDWAPeephole) {
919     addPass(&SIPeepholeSDWAID);
920     addPass(&EarlyMachineLICMID);
921     addPass(&MachineCSEID);
922     addPass(&SIFoldOperandsID);
923     addPass(&DeadMachineInstructionElimID);
924   }
925   addPass(createSIShrinkInstructionsPass());
926 }
927 
928 bool GCNPassConfig::addILPOpts() {
929   if (EnableEarlyIfConversion)
930     addPass(&EarlyIfConverterID);
931 
932   TargetPassConfig::addILPOpts();
933   return false;
934 }
935 
936 bool GCNPassConfig::addInstSelector() {
937   AMDGPUPassConfig::addInstSelector();
938   addPass(&SIFixSGPRCopiesID);
939   addPass(createSILowerI1CopiesPass());
940   addPass(createSIAddIMGInitPass());
941   return false;
942 }
943 
944 bool GCNPassConfig::addIRTranslator() {
945   addPass(new IRTranslator(getOptLevel()));
946   return false;
947 }
948 
949 void GCNPassConfig::addPreLegalizeMachineIR() {
950   bool IsOptNone = getOptLevel() == CodeGenOpt::None;
951   addPass(createAMDGPUPreLegalizeCombiner(IsOptNone));
952   addPass(new Localizer());
953 }
954 
955 bool GCNPassConfig::addLegalizeMachineIR() {
956   addPass(new Legalizer());
957   return false;
958 }
959 
960 void GCNPassConfig::addPreRegBankSelect() {
961   bool IsOptNone = getOptLevel() == CodeGenOpt::None;
962   addPass(createAMDGPUPostLegalizeCombiner(IsOptNone));
963 }
964 
965 bool GCNPassConfig::addRegBankSelect() {
966   addPass(new RegBankSelect());
967   return false;
968 }
969 
970 bool GCNPassConfig::addGlobalInstructionSelect() {
971   addPass(new InstructionSelect());
972   return false;
973 }
974 
975 void GCNPassConfig::addPreRegAlloc() {
976   if (LateCFGStructurize) {
977     addPass(createAMDGPUMachineCFGStructurizerPass());
978   }
979   addPass(createSIWholeQuadModePass());
980 }
981 
982 void GCNPassConfig::addFastRegAlloc() {
983   // FIXME: We have to disable the verifier here because of PHIElimination +
984   // TwoAddressInstructions disabling it.
985 
986   // This must be run immediately after phi elimination and before
987   // TwoAddressInstructions, otherwise the processing of the tied operand of
988   // SI_ELSE will introduce a copy of the tied operand source after the else.
989   insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
990 
991   // This must be run just after RegisterCoalescing.
992   insertPass(&RegisterCoalescerID, &SIPreAllocateWWMRegsID, false);
993 
994   TargetPassConfig::addFastRegAlloc();
995 }
996 
997 void GCNPassConfig::addOptimizedRegAlloc() {
998   if (OptExecMaskPreRA)
999     insertPass(&MachineSchedulerID, &SIOptimizeExecMaskingPreRAID);
1000   insertPass(&MachineSchedulerID, &SIFormMemoryClausesID);
1001 
1002   // This must be run immediately after phi elimination and before
1003   // TwoAddressInstructions, otherwise the processing of the tied operand of
1004   // SI_ELSE will introduce a copy of the tied operand source after the else.
1005   insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
1006 
1007   // This must be run just after RegisterCoalescing.
1008   insertPass(&RegisterCoalescerID, &SIPreAllocateWWMRegsID, false);
1009 
1010   if (EnableDCEInRA)
1011     insertPass(&DetectDeadLanesID, &DeadMachineInstructionElimID);
1012 
1013   TargetPassConfig::addOptimizedRegAlloc();
1014 }
1015 
1016 bool GCNPassConfig::addPreRewrite() {
1017   if (EnableRegReassign) {
1018     addPass(&GCNNSAReassignID);
1019     addPass(&GCNRegBankReassignID);
1020   }
1021   return true;
1022 }
1023 
1024 void GCNPassConfig::addPostRegAlloc() {
1025   addPass(&SIFixVGPRCopiesID);
1026   if (getOptLevel() > CodeGenOpt::None)
1027     addPass(&SIOptimizeExecMaskingID);
1028   TargetPassConfig::addPostRegAlloc();
1029 
1030   // Equivalent of PEI for SGPRs.
1031   addPass(&SILowerSGPRSpillsID);
1032 }
1033 
1034 void GCNPassConfig::addPreSched2() {
1035   addPass(&SIPostRABundlerID);
1036 }
1037 
1038 void GCNPassConfig::addPreEmitPass() {
1039   addPass(createSIMemoryLegalizerPass());
1040   addPass(createSIInsertWaitcntsPass());
1041   addPass(createSIShrinkInstructionsPass());
1042   addPass(createSIModeRegisterPass());
1043 
1044   // The hazard recognizer that runs as part of the post-ra scheduler does not
1045   // guarantee to be able handle all hazards correctly. This is because if there
1046   // are multiple scheduling regions in a basic block, the regions are scheduled
1047   // bottom up, so when we begin to schedule a region we don't know what
1048   // instructions were emitted directly before it.
1049   //
1050   // Here we add a stand-alone hazard recognizer pass which can handle all
1051   // cases.
1052   //
1053   // FIXME: This stand-alone pass will emit indiv. S_NOP 0, as needed. It would
1054   // be better for it to emit S_NOP <N> when possible.
1055   addPass(&PostRAHazardRecognizerID);
1056   if (getOptLevel() > CodeGenOpt::None)
1057     addPass(&SIInsertHardClausesID);
1058 
1059   addPass(&SIRemoveShortExecBranchesID);
1060   addPass(&SIInsertSkipsPassID);
1061   addPass(&SIPreEmitPeepholeID);
1062   addPass(&BranchRelaxationPassID);
1063 }
1064 
1065 TargetPassConfig *GCNTargetMachine::createPassConfig(PassManagerBase &PM) {
1066   return new GCNPassConfig(*this, PM);
1067 }
1068 
1069 yaml::MachineFunctionInfo *GCNTargetMachine::createDefaultFuncInfoYAML() const {
1070   return new yaml::SIMachineFunctionInfo();
1071 }
1072 
1073 yaml::MachineFunctionInfo *
1074 GCNTargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const {
1075   const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1076   return new yaml::SIMachineFunctionInfo(*MFI,
1077                                          *MF.getSubtarget().getRegisterInfo());
1078 }
1079 
1080 bool GCNTargetMachine::parseMachineFunctionInfo(
1081     const yaml::MachineFunctionInfo &MFI_, PerFunctionMIParsingState &PFS,
1082     SMDiagnostic &Error, SMRange &SourceRange) const {
1083   const yaml::SIMachineFunctionInfo &YamlMFI =
1084       reinterpret_cast<const yaml::SIMachineFunctionInfo &>(MFI_);
1085   MachineFunction &MF = PFS.MF;
1086   SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
1087 
1088   MFI->initializeBaseYamlFields(YamlMFI);
1089 
1090   auto parseRegister = [&](const yaml::StringValue &RegName, Register &RegVal) {
1091     Register TempReg;
1092     if (parseNamedRegisterReference(PFS, TempReg, RegName.Value, Error)) {
1093       SourceRange = RegName.SourceRange;
1094       return true;
1095     }
1096     RegVal = TempReg;
1097 
1098     return false;
1099   };
1100 
1101   auto diagnoseRegisterClass = [&](const yaml::StringValue &RegName) {
1102     // Create a diagnostic for a the register string literal.
1103     const MemoryBuffer &Buffer =
1104         *PFS.SM->getMemoryBuffer(PFS.SM->getMainFileID());
1105     Error = SMDiagnostic(*PFS.SM, SMLoc(), Buffer.getBufferIdentifier(), 1,
1106                          RegName.Value.size(), SourceMgr::DK_Error,
1107                          "incorrect register class for field", RegName.Value,
1108                          None, None);
1109     SourceRange = RegName.SourceRange;
1110     return true;
1111   };
1112 
1113   if (parseRegister(YamlMFI.ScratchRSrcReg, MFI->ScratchRSrcReg) ||
1114       parseRegister(YamlMFI.FrameOffsetReg, MFI->FrameOffsetReg) ||
1115       parseRegister(YamlMFI.StackPtrOffsetReg, MFI->StackPtrOffsetReg))
1116     return true;
1117 
1118   if (MFI->ScratchRSrcReg != AMDGPU::PRIVATE_RSRC_REG &&
1119       !AMDGPU::SGPR_128RegClass.contains(MFI->ScratchRSrcReg)) {
1120     return diagnoseRegisterClass(YamlMFI.ScratchRSrcReg);
1121   }
1122 
1123   if (MFI->FrameOffsetReg != AMDGPU::FP_REG &&
1124       !AMDGPU::SGPR_32RegClass.contains(MFI->FrameOffsetReg)) {
1125     return diagnoseRegisterClass(YamlMFI.FrameOffsetReg);
1126   }
1127 
1128   if (MFI->StackPtrOffsetReg != AMDGPU::SP_REG &&
1129       !AMDGPU::SGPR_32RegClass.contains(MFI->StackPtrOffsetReg)) {
1130     return diagnoseRegisterClass(YamlMFI.StackPtrOffsetReg);
1131   }
1132 
1133   auto parseAndCheckArgument = [&](const Optional<yaml::SIArgument> &A,
1134                                    const TargetRegisterClass &RC,
1135                                    ArgDescriptor &Arg, unsigned UserSGPRs,
1136                                    unsigned SystemSGPRs) {
1137     // Skip parsing if it's not present.
1138     if (!A)
1139       return false;
1140 
1141     if (A->IsRegister) {
1142       Register Reg;
1143       if (parseNamedRegisterReference(PFS, Reg, A->RegisterName.Value, Error)) {
1144         SourceRange = A->RegisterName.SourceRange;
1145         return true;
1146       }
1147       if (!RC.contains(Reg))
1148         return diagnoseRegisterClass(A->RegisterName);
1149       Arg = ArgDescriptor::createRegister(Reg);
1150     } else
1151       Arg = ArgDescriptor::createStack(A->StackOffset);
1152     // Check and apply the optional mask.
1153     if (A->Mask)
1154       Arg = ArgDescriptor::createArg(Arg, A->Mask.getValue());
1155 
1156     MFI->NumUserSGPRs += UserSGPRs;
1157     MFI->NumSystemSGPRs += SystemSGPRs;
1158     return false;
1159   };
1160 
1161   if (YamlMFI.ArgInfo &&
1162       (parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentBuffer,
1163                              AMDGPU::SGPR_128RegClass,
1164                              MFI->ArgInfo.PrivateSegmentBuffer, 4, 0) ||
1165        parseAndCheckArgument(YamlMFI.ArgInfo->DispatchPtr,
1166                              AMDGPU::SReg_64RegClass, MFI->ArgInfo.DispatchPtr,
1167                              2, 0) ||
1168        parseAndCheckArgument(YamlMFI.ArgInfo->QueuePtr, AMDGPU::SReg_64RegClass,
1169                              MFI->ArgInfo.QueuePtr, 2, 0) ||
1170        parseAndCheckArgument(YamlMFI.ArgInfo->KernargSegmentPtr,
1171                              AMDGPU::SReg_64RegClass,
1172                              MFI->ArgInfo.KernargSegmentPtr, 2, 0) ||
1173        parseAndCheckArgument(YamlMFI.ArgInfo->DispatchID,
1174                              AMDGPU::SReg_64RegClass, MFI->ArgInfo.DispatchID,
1175                              2, 0) ||
1176        parseAndCheckArgument(YamlMFI.ArgInfo->FlatScratchInit,
1177                              AMDGPU::SReg_64RegClass,
1178                              MFI->ArgInfo.FlatScratchInit, 2, 0) ||
1179        parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentSize,
1180                              AMDGPU::SGPR_32RegClass,
1181                              MFI->ArgInfo.PrivateSegmentSize, 0, 0) ||
1182        parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDX,
1183                              AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDX,
1184                              0, 1) ||
1185        parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDY,
1186                              AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDY,
1187                              0, 1) ||
1188        parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupIDZ,
1189                              AMDGPU::SGPR_32RegClass, MFI->ArgInfo.WorkGroupIDZ,
1190                              0, 1) ||
1191        parseAndCheckArgument(YamlMFI.ArgInfo->WorkGroupInfo,
1192                              AMDGPU::SGPR_32RegClass,
1193                              MFI->ArgInfo.WorkGroupInfo, 0, 1) ||
1194        parseAndCheckArgument(YamlMFI.ArgInfo->PrivateSegmentWaveByteOffset,
1195                              AMDGPU::SGPR_32RegClass,
1196                              MFI->ArgInfo.PrivateSegmentWaveByteOffset, 0, 1) ||
1197        parseAndCheckArgument(YamlMFI.ArgInfo->ImplicitArgPtr,
1198                              AMDGPU::SReg_64RegClass,
1199                              MFI->ArgInfo.ImplicitArgPtr, 0, 0) ||
1200        parseAndCheckArgument(YamlMFI.ArgInfo->ImplicitBufferPtr,
1201                              AMDGPU::SReg_64RegClass,
1202                              MFI->ArgInfo.ImplicitBufferPtr, 2, 0) ||
1203        parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDX,
1204                              AMDGPU::VGPR_32RegClass,
1205                              MFI->ArgInfo.WorkItemIDX, 0, 0) ||
1206        parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDY,
1207                              AMDGPU::VGPR_32RegClass,
1208                              MFI->ArgInfo.WorkItemIDY, 0, 0) ||
1209        parseAndCheckArgument(YamlMFI.ArgInfo->WorkItemIDZ,
1210                              AMDGPU::VGPR_32RegClass,
1211                              MFI->ArgInfo.WorkItemIDZ, 0, 0)))
1212     return true;
1213 
1214   MFI->Mode.IEEE = YamlMFI.Mode.IEEE;
1215   MFI->Mode.DX10Clamp = YamlMFI.Mode.DX10Clamp;
1216   MFI->Mode.FP32InputDenormals = YamlMFI.Mode.FP32InputDenormals;
1217   MFI->Mode.FP32OutputDenormals = YamlMFI.Mode.FP32OutputDenormals;
1218   MFI->Mode.FP64FP16InputDenormals = YamlMFI.Mode.FP64FP16InputDenormals;
1219   MFI->Mode.FP64FP16OutputDenormals = YamlMFI.Mode.FP64FP16OutputDenormals;
1220 
1221   return false;
1222 }
1223