1 //===- SampleProfile.cpp - Incorporate sample profiles into the IR --------===// 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 // This file implements the SampleProfileLoader transformation. This pass 10 // reads a profile file generated by a sampling profiler (e.g. Linux Perf - 11 // http://perf.wiki.kernel.org/) and generates IR metadata to reflect the 12 // profile information in the given profile. 13 // 14 // This pass generates branch weight annotations on the IR: 15 // 16 // - prof: Represents branch weights. This annotation is added to branches 17 // to indicate the weights of each edge coming out of the branch. 18 // The weight of each edge is the weight of the target block for 19 // that edge. The weight of a block B is computed as the maximum 20 // number of samples found in B. 21 // 22 //===----------------------------------------------------------------------===// 23 24 #include "llvm/Transforms/IPO/SampleProfile.h" 25 #include "llvm/ADT/ArrayRef.h" 26 #include "llvm/ADT/DenseMap.h" 27 #include "llvm/ADT/DenseSet.h" 28 #include "llvm/ADT/None.h" 29 #include "llvm/ADT/PriorityQueue.h" 30 #include "llvm/ADT/SCCIterator.h" 31 #include "llvm/ADT/SmallPtrSet.h" 32 #include "llvm/ADT/SmallSet.h" 33 #include "llvm/ADT/SmallVector.h" 34 #include "llvm/ADT/Statistic.h" 35 #include "llvm/ADT/StringMap.h" 36 #include "llvm/ADT/StringRef.h" 37 #include "llvm/ADT/Twine.h" 38 #include "llvm/Analysis/AssumptionCache.h" 39 #include "llvm/Analysis/CallGraph.h" 40 #include "llvm/Analysis/CallGraphSCCPass.h" 41 #include "llvm/Analysis/InlineAdvisor.h" 42 #include "llvm/Analysis/InlineCost.h" 43 #include "llvm/Analysis/LoopInfo.h" 44 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 45 #include "llvm/Analysis/PostDominators.h" 46 #include "llvm/Analysis/ProfileSummaryInfo.h" 47 #include "llvm/Analysis/ReplayInlineAdvisor.h" 48 #include "llvm/Analysis/TargetLibraryInfo.h" 49 #include "llvm/Analysis/TargetTransformInfo.h" 50 #include "llvm/IR/BasicBlock.h" 51 #include "llvm/IR/CFG.h" 52 #include "llvm/IR/DebugInfoMetadata.h" 53 #include "llvm/IR/DebugLoc.h" 54 #include "llvm/IR/DiagnosticInfo.h" 55 #include "llvm/IR/Dominators.h" 56 #include "llvm/IR/Function.h" 57 #include "llvm/IR/GlobalValue.h" 58 #include "llvm/IR/InstrTypes.h" 59 #include "llvm/IR/Instruction.h" 60 #include "llvm/IR/Instructions.h" 61 #include "llvm/IR/IntrinsicInst.h" 62 #include "llvm/IR/LLVMContext.h" 63 #include "llvm/IR/MDBuilder.h" 64 #include "llvm/IR/Module.h" 65 #include "llvm/IR/PassManager.h" 66 #include "llvm/IR/ValueSymbolTable.h" 67 #include "llvm/InitializePasses.h" 68 #include "llvm/Pass.h" 69 #include "llvm/ProfileData/InstrProf.h" 70 #include "llvm/ProfileData/SampleProf.h" 71 #include "llvm/ProfileData/SampleProfReader.h" 72 #include "llvm/Support/Casting.h" 73 #include "llvm/Support/CommandLine.h" 74 #include "llvm/Support/Debug.h" 75 #include "llvm/Support/ErrorHandling.h" 76 #include "llvm/Support/ErrorOr.h" 77 #include "llvm/Support/GenericDomTree.h" 78 #include "llvm/Support/raw_ostream.h" 79 #include "llvm/Transforms/IPO.h" 80 #include "llvm/Transforms/IPO/SampleContextTracker.h" 81 #include "llvm/Transforms/IPO/SampleProfileProbe.h" 82 #include "llvm/Transforms/Instrumentation.h" 83 #include "llvm/Transforms/Utils/CallPromotionUtils.h" 84 #include "llvm/Transforms/Utils/Cloning.h" 85 #include "llvm/Transforms/Utils/SampleProfileLoaderBaseImpl.h" 86 #include "llvm/Transforms/Utils/SampleProfileLoaderBaseUtil.h" 87 #include <algorithm> 88 #include <cassert> 89 #include <cstdint> 90 #include <functional> 91 #include <limits> 92 #include <map> 93 #include <memory> 94 #include <queue> 95 #include <string> 96 #include <system_error> 97 #include <utility> 98 #include <vector> 99 100 using namespace llvm; 101 using namespace sampleprof; 102 using namespace llvm::sampleprofutil; 103 using ProfileCount = Function::ProfileCount; 104 #define DEBUG_TYPE "sample-profile" 105 #define CSINLINE_DEBUG DEBUG_TYPE "-inline" 106 107 STATISTIC(NumCSInlined, 108 "Number of functions inlined with context sensitive profile"); 109 STATISTIC(NumCSNotInlined, 110 "Number of functions not inlined with context sensitive profile"); 111 STATISTIC(NumMismatchedProfile, 112 "Number of functions with CFG mismatched profile"); 113 STATISTIC(NumMatchedProfile, "Number of functions with CFG matched profile"); 114 STATISTIC(NumDuplicatedInlinesite, 115 "Number of inlined callsites with a partial distribution factor"); 116 117 STATISTIC(NumCSInlinedHitMinLimit, 118 "Number of functions with FDO inline stopped due to min size limit"); 119 STATISTIC(NumCSInlinedHitMaxLimit, 120 "Number of functions with FDO inline stopped due to max size limit"); 121 STATISTIC( 122 NumCSInlinedHitGrowthLimit, 123 "Number of functions with FDO inline stopped due to growth size limit"); 124 125 // Command line option to specify the file to read samples from. This is 126 // mainly used for debugging. 127 static cl::opt<std::string> SampleProfileFile( 128 "sample-profile-file", cl::init(""), cl::value_desc("filename"), 129 cl::desc("Profile file loaded by -sample-profile"), cl::Hidden); 130 131 // The named file contains a set of transformations that may have been applied 132 // to the symbol names between the program from which the sample data was 133 // collected and the current program's symbols. 134 static cl::opt<std::string> SampleProfileRemappingFile( 135 "sample-profile-remapping-file", cl::init(""), cl::value_desc("filename"), 136 cl::desc("Profile remapping file loaded by -sample-profile"), cl::Hidden); 137 138 static cl::opt<bool> ProfileSampleAccurate( 139 "profile-sample-accurate", cl::Hidden, cl::init(false), 140 cl::desc("If the sample profile is accurate, we will mark all un-sampled " 141 "callsite and function as having 0 samples. Otherwise, treat " 142 "un-sampled callsites and functions conservatively as unknown. ")); 143 144 static cl::opt<bool> ProfileAccurateForSymsInList( 145 "profile-accurate-for-symsinlist", cl::Hidden, cl::ZeroOrMore, 146 cl::init(true), 147 cl::desc("For symbols in profile symbol list, regard their profiles to " 148 "be accurate. It may be overriden by profile-sample-accurate. ")); 149 150 static cl::opt<bool> ProfileMergeInlinee( 151 "sample-profile-merge-inlinee", cl::Hidden, cl::init(true), 152 cl::desc("Merge past inlinee's profile to outline version if sample " 153 "profile loader decided not to inline a call site. It will " 154 "only be enabled when top-down order of profile loading is " 155 "enabled. ")); 156 157 static cl::opt<bool> ProfileTopDownLoad( 158 "sample-profile-top-down-load", cl::Hidden, cl::init(true), 159 cl::desc("Do profile annotation and inlining for functions in top-down " 160 "order of call graph during sample profile loading. It only " 161 "works for new pass manager. ")); 162 163 static cl::opt<bool> UseProfileIndirectCallEdges( 164 "use-profile-indirect-call-edges", cl::init(true), cl::Hidden, 165 cl::desc("Considering indirect call samples from profile when top-down " 166 "processing functions. Only CSSPGO is supported.")); 167 168 static cl::opt<bool> UseProfileTopDownOrder( 169 "use-profile-top-down-order", cl::init(false), cl::Hidden, 170 cl::desc("Process functions in one SCC in a top-down order " 171 "based on the input profile.")); 172 173 static cl::opt<bool> ProfileSizeInline( 174 "sample-profile-inline-size", cl::Hidden, cl::init(false), 175 cl::desc("Inline cold call sites in profile loader if it's beneficial " 176 "for code size.")); 177 178 static cl::opt<int> ProfileInlineGrowthLimit( 179 "sample-profile-inline-growth-limit", cl::Hidden, cl::init(12), 180 cl::desc("The size growth ratio limit for proirity-based sample profile " 181 "loader inlining.")); 182 183 static cl::opt<int> ProfileInlineLimitMin( 184 "sample-profile-inline-limit-min", cl::Hidden, cl::init(100), 185 cl::desc("The lower bound of size growth limit for " 186 "proirity-based sample profile loader inlining.")); 187 188 static cl::opt<int> ProfileInlineLimitMax( 189 "sample-profile-inline-limit-max", cl::Hidden, cl::init(10000), 190 cl::desc("The upper bound of size growth limit for " 191 "proirity-based sample profile loader inlining.")); 192 193 static cl::opt<int> ProfileICPThreshold( 194 "sample-profile-icp-threshold", cl::Hidden, cl::init(5), 195 cl::desc( 196 "Relative hotness threshold for indirect " 197 "call promotion in proirity-based sample profile loader inlining.")); 198 199 static cl::opt<int> SampleHotCallSiteThreshold( 200 "sample-profile-hot-inline-threshold", cl::Hidden, cl::init(3000), 201 cl::desc("Hot callsite threshold for proirity-based sample profile loader " 202 "inlining.")); 203 204 static cl::opt<bool> CallsitePrioritizedInline( 205 "sample-profile-prioritized-inline", cl::Hidden, cl::ZeroOrMore, 206 cl::init(false), 207 cl::desc("Use call site prioritized inlining for sample profile loader." 208 "Currently only CSSPGO is supported.")); 209 210 static cl::opt<int> SampleColdCallSiteThreshold( 211 "sample-profile-cold-inline-threshold", cl::Hidden, cl::init(45), 212 cl::desc("Threshold for inlining cold callsites")); 213 214 static cl::opt<std::string> ProfileInlineReplayFile( 215 "sample-profile-inline-replay", cl::init(""), cl::value_desc("filename"), 216 cl::desc( 217 "Optimization remarks file containing inline remarks to be replayed " 218 "by inlining from sample profile loader."), 219 cl::Hidden); 220 221 extern cl::opt<unsigned> MaxNumPromotions; 222 223 namespace { 224 225 using BlockWeightMap = DenseMap<const BasicBlock *, uint64_t>; 226 using EquivalenceClassMap = DenseMap<const BasicBlock *, const BasicBlock *>; 227 using Edge = std::pair<const BasicBlock *, const BasicBlock *>; 228 using EdgeWeightMap = DenseMap<Edge, uint64_t>; 229 using BlockEdgeMap = 230 DenseMap<const BasicBlock *, SmallVector<const BasicBlock *, 8>>; 231 232 class GUIDToFuncNameMapper { 233 public: 234 GUIDToFuncNameMapper(Module &M, SampleProfileReader &Reader, 235 DenseMap<uint64_t, StringRef> &GUIDToFuncNameMap) 236 : CurrentReader(Reader), CurrentModule(M), 237 CurrentGUIDToFuncNameMap(GUIDToFuncNameMap) { 238 if (!CurrentReader.useMD5()) 239 return; 240 241 for (const auto &F : CurrentModule) { 242 StringRef OrigName = F.getName(); 243 CurrentGUIDToFuncNameMap.insert( 244 {Function::getGUID(OrigName), OrigName}); 245 246 // Local to global var promotion used by optimization like thinlto 247 // will rename the var and add suffix like ".llvm.xxx" to the 248 // original local name. In sample profile, the suffixes of function 249 // names are all stripped. Since it is possible that the mapper is 250 // built in post-thin-link phase and var promotion has been done, 251 // we need to add the substring of function name without the suffix 252 // into the GUIDToFuncNameMap. 253 StringRef CanonName = FunctionSamples::getCanonicalFnName(F); 254 if (CanonName != OrigName) 255 CurrentGUIDToFuncNameMap.insert( 256 {Function::getGUID(CanonName), CanonName}); 257 } 258 259 // Update GUIDToFuncNameMap for each function including inlinees. 260 SetGUIDToFuncNameMapForAll(&CurrentGUIDToFuncNameMap); 261 } 262 263 ~GUIDToFuncNameMapper() { 264 if (!CurrentReader.useMD5()) 265 return; 266 267 CurrentGUIDToFuncNameMap.clear(); 268 269 // Reset GUIDToFuncNameMap for of each function as they're no 270 // longer valid at this point. 271 SetGUIDToFuncNameMapForAll(nullptr); 272 } 273 274 private: 275 void SetGUIDToFuncNameMapForAll(DenseMap<uint64_t, StringRef> *Map) { 276 std::queue<FunctionSamples *> FSToUpdate; 277 for (auto &IFS : CurrentReader.getProfiles()) { 278 FSToUpdate.push(&IFS.second); 279 } 280 281 while (!FSToUpdate.empty()) { 282 FunctionSamples *FS = FSToUpdate.front(); 283 FSToUpdate.pop(); 284 FS->GUIDToFuncNameMap = Map; 285 for (const auto &ICS : FS->getCallsiteSamples()) { 286 const FunctionSamplesMap &FSMap = ICS.second; 287 for (auto &IFS : FSMap) { 288 FunctionSamples &FS = const_cast<FunctionSamples &>(IFS.second); 289 FSToUpdate.push(&FS); 290 } 291 } 292 } 293 } 294 295 SampleProfileReader &CurrentReader; 296 Module &CurrentModule; 297 DenseMap<uint64_t, StringRef> &CurrentGUIDToFuncNameMap; 298 }; 299 300 // Inline candidate used by iterative callsite prioritized inliner 301 struct InlineCandidate { 302 CallBase *CallInstr; 303 const FunctionSamples *CalleeSamples; 304 // Prorated callsite count, which will be used to guide inlining. For example, 305 // if a callsite is duplicated in LTO prelink, then in LTO postlink the two 306 // copies will get their own distribution factors and their prorated counts 307 // will be used to decide if they should be inlined independently. 308 uint64_t CallsiteCount; 309 // Call site distribution factor to prorate the profile samples for a 310 // duplicated callsite. Default value is 1.0. 311 float CallsiteDistribution; 312 }; 313 314 // Inline candidate comparer using call site weight 315 struct CandidateComparer { 316 bool operator()(const InlineCandidate &LHS, const InlineCandidate &RHS) { 317 if (LHS.CallsiteCount != RHS.CallsiteCount) 318 return LHS.CallsiteCount < RHS.CallsiteCount; 319 320 // Tie breaker using GUID so we have stable/deterministic inlining order 321 assert(LHS.CalleeSamples && RHS.CalleeSamples && 322 "Expect non-null FunctionSamples"); 323 return LHS.CalleeSamples->getGUID(LHS.CalleeSamples->getName()) < 324 RHS.CalleeSamples->getGUID(RHS.CalleeSamples->getName()); 325 } 326 }; 327 328 using CandidateQueue = 329 PriorityQueue<InlineCandidate, std::vector<InlineCandidate>, 330 CandidateComparer>; 331 332 /// Sample profile pass. 333 /// 334 /// This pass reads profile data from the file specified by 335 /// -sample-profile-file and annotates every affected function with the 336 /// profile information found in that file. 337 class SampleProfileLoader final : public SampleProfileLoaderBaseImpl { 338 public: 339 SampleProfileLoader( 340 StringRef Name, StringRef RemapName, ThinOrFullLTOPhase LTOPhase, 341 std::function<AssumptionCache &(Function &)> GetAssumptionCache, 342 std::function<TargetTransformInfo &(Function &)> GetTargetTransformInfo, 343 std::function<const TargetLibraryInfo &(Function &)> GetTLI) 344 : SampleProfileLoaderBaseImpl(std::string(Name)), 345 GetAC(std::move(GetAssumptionCache)), 346 GetTTI(std::move(GetTargetTransformInfo)), GetTLI(std::move(GetTLI)), 347 RemappingFilename(std::string(RemapName)), LTOPhase(LTOPhase) {} 348 349 bool doInitialization(Module &M, FunctionAnalysisManager *FAM = nullptr); 350 bool runOnModule(Module &M, ModuleAnalysisManager *AM, 351 ProfileSummaryInfo *_PSI, CallGraph *CG); 352 353 protected: 354 bool runOnFunction(Function &F, ModuleAnalysisManager *AM); 355 bool emitAnnotations(Function &F); 356 ErrorOr<uint64_t> getInstWeight(const Instruction &I) override; 357 ErrorOr<uint64_t> getProbeWeight(const Instruction &I); 358 const FunctionSamples *findCalleeFunctionSamples(const CallBase &I) const; 359 const FunctionSamples * 360 findFunctionSamples(const Instruction &I) const override; 361 std::vector<const FunctionSamples *> 362 findIndirectCallFunctionSamples(const Instruction &I, uint64_t &Sum) const; 363 // Attempt to promote indirect call and also inline the promoted call 364 bool tryPromoteAndInlineCandidate( 365 Function &F, InlineCandidate &Candidate, uint64_t SumOrigin, 366 uint64_t &Sum, SmallVector<CallBase *, 8> *InlinedCallSites = nullptr); 367 bool inlineHotFunctions(Function &F, 368 DenseSet<GlobalValue::GUID> &InlinedGUIDs); 369 InlineCost shouldInlineCandidate(InlineCandidate &Candidate); 370 bool getInlineCandidate(InlineCandidate *NewCandidate, CallBase *CB); 371 bool 372 tryInlineCandidate(InlineCandidate &Candidate, 373 SmallVector<CallBase *, 8> *InlinedCallSites = nullptr); 374 bool 375 inlineHotFunctionsWithPriority(Function &F, 376 DenseSet<GlobalValue::GUID> &InlinedGUIDs); 377 // Inline cold/small functions in addition to hot ones 378 bool shouldInlineColdCallee(CallBase &CallInst); 379 void emitOptimizationRemarksForInlineCandidates( 380 const SmallVectorImpl<CallBase *> &Candidates, const Function &F, 381 bool Hot); 382 std::vector<Function *> buildFunctionOrder(Module &M, CallGraph *CG); 383 void addCallGraphEdges(CallGraph &CG, const FunctionSamples &Samples); 384 void replaceCallGraphEdges(CallGraph &CG, StringMap<Function *> &SymbolMap); 385 void generateMDProfMetadata(Function &F); 386 387 /// Map from function name to Function *. Used to find the function from 388 /// the function name. If the function name contains suffix, additional 389 /// entry is added to map from the stripped name to the function if there 390 /// is one-to-one mapping. 391 StringMap<Function *> SymbolMap; 392 393 std::function<AssumptionCache &(Function &)> GetAC; 394 std::function<TargetTransformInfo &(Function &)> GetTTI; 395 std::function<const TargetLibraryInfo &(Function &)> GetTLI; 396 397 /// Profile tracker for different context. 398 std::unique_ptr<SampleContextTracker> ContextTracker; 399 400 /// Name of the profile remapping file to load. 401 std::string RemappingFilename; 402 403 /// Flag indicating whether the profile input loaded successfully. 404 bool ProfileIsValid = false; 405 406 /// Flag indicating whether input profile is context-sensitive 407 bool ProfileIsCS = false; 408 409 /// Flag indicating which LTO/ThinLTO phase the pass is invoked in. 410 /// 411 /// We need to know the LTO phase because for example in ThinLTOPrelink 412 /// phase, in annotation, we should not promote indirect calls. Instead, 413 /// we will mark GUIDs that needs to be annotated to the function. 414 ThinOrFullLTOPhase LTOPhase; 415 416 /// Profle Symbol list tells whether a function name appears in the binary 417 /// used to generate the current profile. 418 std::unique_ptr<ProfileSymbolList> PSL; 419 420 /// Total number of samples collected in this profile. 421 /// 422 /// This is the sum of all the samples collected in all the functions executed 423 /// at runtime. 424 uint64_t TotalCollectedSamples = 0; 425 426 // Information recorded when we declined to inline a call site 427 // because we have determined it is too cold is accumulated for 428 // each callee function. Initially this is just the entry count. 429 struct NotInlinedProfileInfo { 430 uint64_t entryCount; 431 }; 432 DenseMap<Function *, NotInlinedProfileInfo> notInlinedCallInfo; 433 434 // GUIDToFuncNameMap saves the mapping from GUID to the symbol name, for 435 // all the function symbols defined or declared in current module. 436 DenseMap<uint64_t, StringRef> GUIDToFuncNameMap; 437 438 // All the Names used in FunctionSamples including outline function 439 // names, inline instance names and call target names. 440 StringSet<> NamesInProfile; 441 442 // For symbol in profile symbol list, whether to regard their profiles 443 // to be accurate. It is mainly decided by existance of profile symbol 444 // list and -profile-accurate-for-symsinlist flag, but it can be 445 // overriden by -profile-sample-accurate or profile-sample-accurate 446 // attribute. 447 bool ProfAccForSymsInList; 448 449 // External inline advisor used to replay inline decision from remarks. 450 std::unique_ptr<ReplayInlineAdvisor> ExternalInlineAdvisor; 451 452 // A pseudo probe helper to correlate the imported sample counts. 453 std::unique_ptr<PseudoProbeManager> ProbeManager; 454 }; 455 456 class SampleProfileLoaderLegacyPass : public ModulePass { 457 public: 458 // Class identification, replacement for typeinfo 459 static char ID; 460 461 SampleProfileLoaderLegacyPass( 462 StringRef Name = SampleProfileFile, 463 ThinOrFullLTOPhase LTOPhase = ThinOrFullLTOPhase::None) 464 : ModulePass(ID), SampleLoader( 465 Name, SampleProfileRemappingFile, LTOPhase, 466 [&](Function &F) -> AssumptionCache & { 467 return ACT->getAssumptionCache(F); 468 }, 469 [&](Function &F) -> TargetTransformInfo & { 470 return TTIWP->getTTI(F); 471 }, 472 [&](Function &F) -> TargetLibraryInfo & { 473 return TLIWP->getTLI(F); 474 }) { 475 initializeSampleProfileLoaderLegacyPassPass( 476 *PassRegistry::getPassRegistry()); 477 } 478 479 void dump() { SampleLoader.dump(); } 480 481 bool doInitialization(Module &M) override { 482 return SampleLoader.doInitialization(M); 483 } 484 485 StringRef getPassName() const override { return "Sample profile pass"; } 486 bool runOnModule(Module &M) override; 487 488 void getAnalysisUsage(AnalysisUsage &AU) const override { 489 AU.addRequired<AssumptionCacheTracker>(); 490 AU.addRequired<TargetTransformInfoWrapperPass>(); 491 AU.addRequired<TargetLibraryInfoWrapperPass>(); 492 AU.addRequired<ProfileSummaryInfoWrapperPass>(); 493 } 494 495 private: 496 SampleProfileLoader SampleLoader; 497 AssumptionCacheTracker *ACT = nullptr; 498 TargetTransformInfoWrapperPass *TTIWP = nullptr; 499 TargetLibraryInfoWrapperPass *TLIWP = nullptr; 500 }; 501 502 } // end anonymous namespace 503 504 ErrorOr<uint64_t> SampleProfileLoader::getInstWeight(const Instruction &Inst) { 505 if (FunctionSamples::ProfileIsProbeBased) 506 return getProbeWeight(Inst); 507 508 const DebugLoc &DLoc = Inst.getDebugLoc(); 509 if (!DLoc) 510 return std::error_code(); 511 512 // Ignore all intrinsics, phinodes and branch instructions. 513 // Branch and phinodes instruction usually contains debug info from sources 514 // outside of the residing basic block, thus we ignore them during annotation. 515 if (isa<BranchInst>(Inst) || isa<IntrinsicInst>(Inst) || isa<PHINode>(Inst)) 516 return std::error_code(); 517 518 // If a direct call/invoke instruction is inlined in profile 519 // (findCalleeFunctionSamples returns non-empty result), but not inlined here, 520 // it means that the inlined callsite has no sample, thus the call 521 // instruction should have 0 count. 522 if (!ProfileIsCS) 523 if (const auto *CB = dyn_cast<CallBase>(&Inst)) 524 if (!CB->isIndirectCall() && findCalleeFunctionSamples(*CB)) 525 return 0; 526 527 return getInstWeightImpl(Inst); 528 } 529 530 ErrorOr<uint64_t> SampleProfileLoader::getProbeWeight(const Instruction &Inst) { 531 assert(FunctionSamples::ProfileIsProbeBased && 532 "Profile is not pseudo probe based"); 533 Optional<PseudoProbe> Probe = extractProbe(Inst); 534 if (!Probe) 535 return std::error_code(); 536 537 const FunctionSamples *FS = findFunctionSamples(Inst); 538 if (!FS) 539 return std::error_code(); 540 541 // If a direct call/invoke instruction is inlined in profile 542 // (findCalleeFunctionSamples returns non-empty result), but not inlined here, 543 // it means that the inlined callsite has no sample, thus the call 544 // instruction should have 0 count. 545 if (const auto *CB = dyn_cast<CallBase>(&Inst)) 546 if (!CB->isIndirectCall() && findCalleeFunctionSamples(*CB)) 547 return 0; 548 549 const ErrorOr<uint64_t> &R = FS->findSamplesAt(Probe->Id, 0); 550 if (R) { 551 uint64_t Samples = R.get() * Probe->Factor; 552 bool FirstMark = CoverageTracker.markSamplesUsed(FS, Probe->Id, 0, Samples); 553 if (FirstMark) { 554 ORE->emit([&]() { 555 OptimizationRemarkAnalysis Remark(DEBUG_TYPE, "AppliedSamples", &Inst); 556 Remark << "Applied " << ore::NV("NumSamples", Samples); 557 Remark << " samples from profile (ProbeId="; 558 Remark << ore::NV("ProbeId", Probe->Id); 559 Remark << ", Factor="; 560 Remark << ore::NV("Factor", Probe->Factor); 561 Remark << ", OriginalSamples="; 562 Remark << ore::NV("OriginalSamples", R.get()); 563 Remark << ")"; 564 return Remark; 565 }); 566 } 567 LLVM_DEBUG(dbgs() << " " << Probe->Id << ":" << Inst 568 << " - weight: " << R.get() << " - factor: " 569 << format("%0.2f", Probe->Factor) << ")\n"); 570 return Samples; 571 } 572 return R; 573 } 574 575 /// Get the FunctionSamples for a call instruction. 576 /// 577 /// The FunctionSamples of a call/invoke instruction \p Inst is the inlined 578 /// instance in which that call instruction is calling to. It contains 579 /// all samples that resides in the inlined instance. We first find the 580 /// inlined instance in which the call instruction is from, then we 581 /// traverse its children to find the callsite with the matching 582 /// location. 583 /// 584 /// \param Inst Call/Invoke instruction to query. 585 /// 586 /// \returns The FunctionSamples pointer to the inlined instance. 587 const FunctionSamples * 588 SampleProfileLoader::findCalleeFunctionSamples(const CallBase &Inst) const { 589 const DILocation *DIL = Inst.getDebugLoc(); 590 if (!DIL) { 591 return nullptr; 592 } 593 594 StringRef CalleeName; 595 if (Function *Callee = Inst.getCalledFunction()) 596 CalleeName = FunctionSamples::getCanonicalFnName(*Callee); 597 598 if (ProfileIsCS) 599 return ContextTracker->getCalleeContextSamplesFor(Inst, CalleeName); 600 601 const FunctionSamples *FS = findFunctionSamples(Inst); 602 if (FS == nullptr) 603 return nullptr; 604 605 return FS->findFunctionSamplesAt(FunctionSamples::getCallSiteIdentifier(DIL), 606 CalleeName, Reader->getRemapper()); 607 } 608 609 /// Returns a vector of FunctionSamples that are the indirect call targets 610 /// of \p Inst. The vector is sorted by the total number of samples. Stores 611 /// the total call count of the indirect call in \p Sum. 612 std::vector<const FunctionSamples *> 613 SampleProfileLoader::findIndirectCallFunctionSamples( 614 const Instruction &Inst, uint64_t &Sum) const { 615 const DILocation *DIL = Inst.getDebugLoc(); 616 std::vector<const FunctionSamples *> R; 617 618 if (!DIL) { 619 return R; 620 } 621 622 auto FSCompare = [](const FunctionSamples *L, const FunctionSamples *R) { 623 assert(L && R && "Expect non-null FunctionSamples"); 624 if (L->getEntrySamples() != R->getEntrySamples()) 625 return L->getEntrySamples() > R->getEntrySamples(); 626 return FunctionSamples::getGUID(L->getName()) < 627 FunctionSamples::getGUID(R->getName()); 628 }; 629 630 if (ProfileIsCS) { 631 auto CalleeSamples = 632 ContextTracker->getIndirectCalleeContextSamplesFor(DIL); 633 if (CalleeSamples.empty()) 634 return R; 635 636 // For CSSPGO, we only use target context profile's entry count 637 // as that already includes both inlined callee and non-inlined ones.. 638 Sum = 0; 639 for (const auto *const FS : CalleeSamples) { 640 Sum += FS->getEntrySamples(); 641 R.push_back(FS); 642 } 643 llvm::sort(R, FSCompare); 644 return R; 645 } 646 647 const FunctionSamples *FS = findFunctionSamples(Inst); 648 if (FS == nullptr) 649 return R; 650 651 auto CallSite = FunctionSamples::getCallSiteIdentifier(DIL); 652 auto T = FS->findCallTargetMapAt(CallSite); 653 Sum = 0; 654 if (T) 655 for (const auto &T_C : T.get()) 656 Sum += T_C.second; 657 if (const FunctionSamplesMap *M = FS->findFunctionSamplesMapAt(CallSite)) { 658 if (M->empty()) 659 return R; 660 for (const auto &NameFS : *M) { 661 Sum += NameFS.second.getEntrySamples(); 662 R.push_back(&NameFS.second); 663 } 664 llvm::sort(R, FSCompare); 665 } 666 return R; 667 } 668 669 const FunctionSamples * 670 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const { 671 if (FunctionSamples::ProfileIsProbeBased) { 672 Optional<PseudoProbe> Probe = extractProbe(Inst); 673 if (!Probe) 674 return nullptr; 675 } 676 677 const DILocation *DIL = Inst.getDebugLoc(); 678 if (!DIL) 679 return Samples; 680 681 auto it = DILocation2SampleMap.try_emplace(DIL,nullptr); 682 if (it.second) { 683 if (ProfileIsCS) 684 it.first->second = ContextTracker->getContextSamplesFor(DIL); 685 else 686 it.first->second = 687 Samples->findFunctionSamples(DIL, Reader->getRemapper()); 688 } 689 return it.first->second; 690 } 691 692 /// If the profile count for the promotion candidate \p Candidate is 0, 693 /// it means \p Candidate has already been promoted for \p Inst. 694 static bool isPromotedBefore(const Instruction &Inst, StringRef Candidate) { 695 uint32_t NumVals = 0; 696 uint64_t TotalCount = 0; 697 std::unique_ptr<InstrProfValueData[]> ValueData = 698 std::make_unique<InstrProfValueData[]>(MaxNumPromotions); 699 bool Valid = 700 getValueProfDataFromInst(Inst, IPVK_IndirectCallTarget, MaxNumPromotions, 701 ValueData.get(), NumVals, TotalCount, true); 702 if (Valid) { 703 for (uint32_t I = 0; I < NumVals; I++) { 704 // If the promotion candidate has 0 count in the metadata, it 705 // means the candidate has been promoted for this indirect call. 706 if (ValueData[I].Value == Function::getGUID(Candidate)) 707 return ValueData[I].Count == 0; 708 } 709 } 710 return false; 711 } 712 713 /// Update indirect call target profile metadata for \p Inst. If \p Total 714 /// is given, set TotalCount of call targets counts to \p Total, otherwise 715 /// keep the original value in metadata. 716 static void 717 updateIDTMetaData(Instruction &Inst, 718 const SmallVectorImpl<InstrProfValueData> &CallTargets, 719 uint64_t Total = 0) { 720 DenseMap<uint64_t, uint64_t> ValueCountMap; 721 722 uint32_t NumVals = 0; 723 uint64_t TotalCount = 0; 724 std::unique_ptr<InstrProfValueData[]> ValueData = 725 std::make_unique<InstrProfValueData[]>(MaxNumPromotions); 726 bool Valid = 727 getValueProfDataFromInst(Inst, IPVK_IndirectCallTarget, MaxNumPromotions, 728 ValueData.get(), NumVals, TotalCount, true); 729 if (Valid) { 730 for (uint32_t I = 0; I < NumVals; I++) 731 ValueCountMap[ValueData[I].Value] = ValueData[I].Count; 732 } 733 734 for (const auto &Data : CallTargets) { 735 auto Pair = ValueCountMap.try_emplace(Data.Value, Data.Count); 736 if (Pair.second) 737 continue; 738 // Update existing profile count of the call target if it is not 0. 739 // If it is 0, the call target has been promoted so keep it as 0. 740 if (Pair.first->second != 0) 741 Pair.first->second = Data.Count; 742 else { 743 assert(Total >= Data.Count && "Total should be >= Data.Count"); 744 Total -= Data.Count; 745 } 746 } 747 748 SmallVector<InstrProfValueData, 8> NewCallTargets; 749 for (const auto &ValueCount : ValueCountMap) { 750 NewCallTargets.emplace_back( 751 InstrProfValueData{ValueCount.first, ValueCount.second}); 752 } 753 llvm::sort(NewCallTargets, 754 [](const InstrProfValueData &L, const InstrProfValueData &R) { 755 if (L.Count != R.Count) 756 return L.Count > R.Count; 757 return L.Value > R.Value; 758 }); 759 annotateValueSite(*Inst.getParent()->getParent()->getParent(), Inst, 760 NewCallTargets, Total ? Total : TotalCount, 761 IPVK_IndirectCallTarget, NewCallTargets.size()); 762 } 763 764 /// Attempt to promote indirect call and also inline the promoted call. 765 /// 766 /// \param F Caller function. 767 /// \param Candidate ICP and inline candidate. 768 /// \param Sum Sum of target counts for indirect call. 769 /// \param InlinedCallSite Output vector for new call sites exposed after 770 /// inlining. 771 bool SampleProfileLoader::tryPromoteAndInlineCandidate( 772 Function &F, InlineCandidate &Candidate, uint64_t SumOrigin, uint64_t &Sum, 773 SmallVector<CallBase *, 8> *InlinedCallSite) { 774 auto CalleeFunctionName = Candidate.CalleeSamples->getFuncName(); 775 auto R = SymbolMap.find(CalleeFunctionName); 776 if (R == SymbolMap.end() || !R->getValue()) 777 return false; 778 779 auto &CI = *Candidate.CallInstr; 780 if (isPromotedBefore(CI, R->getValue()->getName())) 781 return false; 782 783 const char *Reason = "Callee function not available"; 784 // R->getValue() != &F is to prevent promoting a recursive call. 785 // If it is a recursive call, we do not inline it as it could bloat 786 // the code exponentially. There is way to better handle this, e.g. 787 // clone the caller first, and inline the cloned caller if it is 788 // recursive. As llvm does not inline recursive calls, we will 789 // simply ignore it instead of handling it explicitly. 790 if (!R->getValue()->isDeclaration() && R->getValue()->getSubprogram() && 791 R->getValue()->hasFnAttribute("use-sample-profile") && 792 R->getValue() != &F && isLegalToPromote(CI, R->getValue(), &Reason)) { 793 // For promoted target, save 0 count in the value profile metadata so 794 // the target won't be promoted again. 795 SmallVector<InstrProfValueData, 1> SortedCallTargets = { 796 InstrProfValueData{Function::getGUID(R->getValue()->getName()), 0}}; 797 updateIDTMetaData(CI, SortedCallTargets); 798 799 auto *DI = &pgo::promoteIndirectCall( 800 CI, R->getValue(), Candidate.CallsiteCount, Sum, false, ORE); 801 if (DI) { 802 Sum -= Candidate.CallsiteCount; 803 // Prorate the indirect callsite distribution. 804 // Do not update the promoted direct callsite distribution at this 805 // point since the original distribution combined with the callee 806 // profile will be used to prorate callsites from the callee if 807 // inlined. Once not inlined, the direct callsite distribution should 808 // be prorated so that the it will reflect the real callsite counts. 809 setProbeDistributionFactor(CI, Candidate.CallsiteDistribution * Sum / 810 SumOrigin); 811 Candidate.CallInstr = DI; 812 if (isa<CallInst>(DI) || isa<InvokeInst>(DI)) { 813 bool Inlined = tryInlineCandidate(Candidate, InlinedCallSite); 814 if (!Inlined) { 815 // Prorate the direct callsite distribution so that it reflects real 816 // callsite counts. 817 setProbeDistributionFactor(*DI, Candidate.CallsiteDistribution * 818 Candidate.CallsiteCount / 819 SumOrigin); 820 } 821 return Inlined; 822 } 823 } 824 } else { 825 LLVM_DEBUG(dbgs() << "\nFailed to promote indirect call to " 826 << Candidate.CalleeSamples->getFuncName() << " because " 827 << Reason << "\n"); 828 } 829 return false; 830 } 831 832 bool SampleProfileLoader::shouldInlineColdCallee(CallBase &CallInst) { 833 if (!ProfileSizeInline) 834 return false; 835 836 Function *Callee = CallInst.getCalledFunction(); 837 if (Callee == nullptr) 838 return false; 839 840 InlineCost Cost = getInlineCost(CallInst, getInlineParams(), GetTTI(*Callee), 841 GetAC, GetTLI); 842 843 if (Cost.isNever()) 844 return false; 845 846 if (Cost.isAlways()) 847 return true; 848 849 return Cost.getCost() <= SampleColdCallSiteThreshold; 850 } 851 852 void SampleProfileLoader::emitOptimizationRemarksForInlineCandidates( 853 const SmallVectorImpl<CallBase *> &Candidates, const Function &F, 854 bool Hot) { 855 for (auto I : Candidates) { 856 Function *CalledFunction = I->getCalledFunction(); 857 if (CalledFunction) { 858 ORE->emit(OptimizationRemarkAnalysis(CSINLINE_DEBUG, "InlineAttempt", 859 I->getDebugLoc(), I->getParent()) 860 << "previous inlining reattempted for " 861 << (Hot ? "hotness: '" : "size: '") 862 << ore::NV("Callee", CalledFunction) << "' into '" 863 << ore::NV("Caller", &F) << "'"); 864 } 865 } 866 } 867 868 /// Iteratively inline hot callsites of a function. 869 /// 870 /// Iteratively traverse all callsites of the function \p F, and find if 871 /// the corresponding inlined instance exists and is hot in profile. If 872 /// it is hot enough, inline the callsites and adds new callsites of the 873 /// callee into the caller. If the call is an indirect call, first promote 874 /// it to direct call. Each indirect call is limited with a single target. 875 /// 876 /// \param F function to perform iterative inlining. 877 /// \param InlinedGUIDs a set to be updated to include all GUIDs that are 878 /// inlined in the profiled binary. 879 /// 880 /// \returns True if there is any inline happened. 881 bool SampleProfileLoader::inlineHotFunctions( 882 Function &F, DenseSet<GlobalValue::GUID> &InlinedGUIDs) { 883 // ProfAccForSymsInList is used in callsiteIsHot. The assertion makes sure 884 // Profile symbol list is ignored when profile-sample-accurate is on. 885 assert((!ProfAccForSymsInList || 886 (!ProfileSampleAccurate && 887 !F.hasFnAttribute("profile-sample-accurate"))) && 888 "ProfAccForSymsInList should be false when profile-sample-accurate " 889 "is enabled"); 890 891 DenseMap<CallBase *, const FunctionSamples *> LocalNotInlinedCallSites; 892 bool Changed = false; 893 bool LocalChanged = true; 894 while (LocalChanged) { 895 LocalChanged = false; 896 SmallVector<CallBase *, 10> CIS; 897 for (auto &BB : F) { 898 bool Hot = false; 899 SmallVector<CallBase *, 10> AllCandidates; 900 SmallVector<CallBase *, 10> ColdCandidates; 901 for (auto &I : BB.getInstList()) { 902 const FunctionSamples *FS = nullptr; 903 if (auto *CB = dyn_cast<CallBase>(&I)) { 904 if (!isa<IntrinsicInst>(I) && (FS = findCalleeFunctionSamples(*CB))) { 905 assert((!FunctionSamples::UseMD5 || FS->GUIDToFuncNameMap) && 906 "GUIDToFuncNameMap has to be populated"); 907 AllCandidates.push_back(CB); 908 if (FS->getEntrySamples() > 0 || ProfileIsCS) 909 LocalNotInlinedCallSites.try_emplace(CB, FS); 910 if (callsiteIsHot(FS, PSI, ProfAccForSymsInList)) 911 Hot = true; 912 else if (shouldInlineColdCallee(*CB)) 913 ColdCandidates.push_back(CB); 914 } 915 } 916 } 917 if (Hot || ExternalInlineAdvisor) { 918 CIS.insert(CIS.begin(), AllCandidates.begin(), AllCandidates.end()); 919 emitOptimizationRemarksForInlineCandidates(AllCandidates, F, true); 920 } else { 921 CIS.insert(CIS.begin(), ColdCandidates.begin(), ColdCandidates.end()); 922 emitOptimizationRemarksForInlineCandidates(ColdCandidates, F, false); 923 } 924 } 925 for (CallBase *I : CIS) { 926 Function *CalledFunction = I->getCalledFunction(); 927 InlineCandidate Candidate = { 928 I, 929 LocalNotInlinedCallSites.count(I) ? LocalNotInlinedCallSites[I] 930 : nullptr, 931 0 /* dummy count */, 1.0 /* dummy distribution factor */}; 932 // Do not inline recursive calls. 933 if (CalledFunction == &F) 934 continue; 935 if (I->isIndirectCall()) { 936 uint64_t Sum; 937 for (const auto *FS : findIndirectCallFunctionSamples(*I, Sum)) { 938 uint64_t SumOrigin = Sum; 939 if (LTOPhase == ThinOrFullLTOPhase::ThinLTOPreLink) { 940 FS->findInlinedFunctions(InlinedGUIDs, F.getParent(), 941 PSI->getOrCompHotCountThreshold()); 942 continue; 943 } 944 if (!callsiteIsHot(FS, PSI, ProfAccForSymsInList)) 945 continue; 946 947 Candidate = {I, FS, FS->getEntrySamples(), 1.0}; 948 if (tryPromoteAndInlineCandidate(F, Candidate, SumOrigin, Sum)) { 949 LocalNotInlinedCallSites.erase(I); 950 LocalChanged = true; 951 } 952 } 953 } else if (CalledFunction && CalledFunction->getSubprogram() && 954 !CalledFunction->isDeclaration()) { 955 if (tryInlineCandidate(Candidate)) { 956 LocalNotInlinedCallSites.erase(I); 957 LocalChanged = true; 958 } 959 } else if (LTOPhase == ThinOrFullLTOPhase::ThinLTOPreLink) { 960 findCalleeFunctionSamples(*I)->findInlinedFunctions( 961 InlinedGUIDs, F.getParent(), PSI->getOrCompHotCountThreshold()); 962 } 963 } 964 Changed |= LocalChanged; 965 } 966 967 // For CS profile, profile for not inlined context will be merged when 968 // base profile is being trieved 969 if (ProfileIsCS) 970 return Changed; 971 972 // Accumulate not inlined callsite information into notInlinedSamples 973 for (const auto &Pair : LocalNotInlinedCallSites) { 974 CallBase *I = Pair.getFirst(); 975 Function *Callee = I->getCalledFunction(); 976 if (!Callee || Callee->isDeclaration()) 977 continue; 978 979 ORE->emit(OptimizationRemarkAnalysis(CSINLINE_DEBUG, "NotInline", 980 I->getDebugLoc(), I->getParent()) 981 << "previous inlining not repeated: '" 982 << ore::NV("Callee", Callee) << "' into '" 983 << ore::NV("Caller", &F) << "'"); 984 985 ++NumCSNotInlined; 986 const FunctionSamples *FS = Pair.getSecond(); 987 if (FS->getTotalSamples() == 0 && FS->getEntrySamples() == 0) { 988 continue; 989 } 990 991 if (ProfileMergeInlinee) { 992 // A function call can be replicated by optimizations like callsite 993 // splitting or jump threading and the replicates end up sharing the 994 // sample nested callee profile instead of slicing the original inlinee's 995 // profile. We want to do merge exactly once by filtering out callee 996 // profiles with a non-zero head sample count. 997 if (FS->getHeadSamples() == 0) { 998 // Use entry samples as head samples during the merge, as inlinees 999 // don't have head samples. 1000 const_cast<FunctionSamples *>(FS)->addHeadSamples( 1001 FS->getEntrySamples()); 1002 1003 // Note that we have to do the merge right after processing function. 1004 // This allows OutlineFS's profile to be used for annotation during 1005 // top-down processing of functions' annotation. 1006 FunctionSamples *OutlineFS = Reader->getOrCreateSamplesFor(*Callee); 1007 OutlineFS->merge(*FS); 1008 } 1009 } else { 1010 auto pair = 1011 notInlinedCallInfo.try_emplace(Callee, NotInlinedProfileInfo{0}); 1012 pair.first->second.entryCount += FS->getEntrySamples(); 1013 } 1014 } 1015 return Changed; 1016 } 1017 1018 bool SampleProfileLoader::tryInlineCandidate( 1019 InlineCandidate &Candidate, SmallVector<CallBase *, 8> *InlinedCallSites) { 1020 1021 CallBase &CB = *Candidate.CallInstr; 1022 Function *CalledFunction = CB.getCalledFunction(); 1023 assert(CalledFunction && "Expect a callee with definition"); 1024 DebugLoc DLoc = CB.getDebugLoc(); 1025 BasicBlock *BB = CB.getParent(); 1026 1027 InlineCost Cost = shouldInlineCandidate(Candidate); 1028 if (Cost.isNever()) { 1029 ORE->emit(OptimizationRemarkAnalysis(CSINLINE_DEBUG, "InlineFail", DLoc, BB) 1030 << "incompatible inlining"); 1031 return false; 1032 } 1033 1034 if (!Cost) 1035 return false; 1036 1037 InlineFunctionInfo IFI(nullptr, GetAC); 1038 if (InlineFunction(CB, IFI).isSuccess()) { 1039 // The call to InlineFunction erases I, so we can't pass it here. 1040 emitInlinedInto(*ORE, DLoc, BB, *CalledFunction, *BB->getParent(), Cost, 1041 true, CSINLINE_DEBUG); 1042 1043 // Now populate the list of newly exposed call sites. 1044 if (InlinedCallSites) { 1045 InlinedCallSites->clear(); 1046 for (auto &I : IFI.InlinedCallSites) 1047 InlinedCallSites->push_back(I); 1048 } 1049 1050 if (ProfileIsCS) 1051 ContextTracker->markContextSamplesInlined(Candidate.CalleeSamples); 1052 ++NumCSInlined; 1053 1054 // Prorate inlined probes for a duplicated inlining callsite which probably 1055 // has a distribution less than 100%. Samples for an inlinee should be 1056 // distributed among the copies of the original callsite based on each 1057 // callsite's distribution factor for counts accuracy. Note that an inlined 1058 // probe may come with its own distribution factor if it has been duplicated 1059 // in the inlinee body. The two factor are multiplied to reflect the 1060 // aggregation of duplication. 1061 if (Candidate.CallsiteDistribution < 1) { 1062 for (auto &I : IFI.InlinedCallSites) { 1063 if (Optional<PseudoProbe> Probe = extractProbe(*I)) 1064 setProbeDistributionFactor(*I, Probe->Factor * 1065 Candidate.CallsiteDistribution); 1066 } 1067 NumDuplicatedInlinesite++; 1068 } 1069 1070 return true; 1071 } 1072 return false; 1073 } 1074 1075 bool SampleProfileLoader::getInlineCandidate(InlineCandidate *NewCandidate, 1076 CallBase *CB) { 1077 assert(CB && "Expect non-null call instruction"); 1078 1079 if (isa<IntrinsicInst>(CB)) 1080 return false; 1081 1082 // Find the callee's profile. For indirect call, find hottest target profile. 1083 const FunctionSamples *CalleeSamples = findCalleeFunctionSamples(*CB); 1084 if (!CalleeSamples) 1085 return false; 1086 1087 float Factor = 1.0; 1088 if (Optional<PseudoProbe> Probe = extractProbe(*CB)) 1089 Factor = Probe->Factor; 1090 1091 uint64_t CallsiteCount = 0; 1092 ErrorOr<uint64_t> Weight = getBlockWeight(CB->getParent()); 1093 if (Weight) 1094 CallsiteCount = Weight.get(); 1095 if (CalleeSamples) 1096 CallsiteCount = std::max( 1097 CallsiteCount, uint64_t(CalleeSamples->getEntrySamples() * Factor)); 1098 1099 *NewCandidate = {CB, CalleeSamples, CallsiteCount, Factor}; 1100 return true; 1101 } 1102 1103 InlineCost 1104 SampleProfileLoader::shouldInlineCandidate(InlineCandidate &Candidate) { 1105 std::unique_ptr<InlineAdvice> Advice = nullptr; 1106 if (ExternalInlineAdvisor) { 1107 Advice = ExternalInlineAdvisor->getAdvice(*Candidate.CallInstr); 1108 if (!Advice->isInliningRecommended()) { 1109 Advice->recordUnattemptedInlining(); 1110 return InlineCost::getNever("not previously inlined"); 1111 } 1112 Advice->recordInlining(); 1113 return InlineCost::getAlways("previously inlined"); 1114 } 1115 1116 // Adjust threshold based on call site hotness, only do this for callsite 1117 // prioritized inliner because otherwise cost-benefit check is done earlier. 1118 int SampleThreshold = SampleColdCallSiteThreshold; 1119 if (CallsitePrioritizedInline) { 1120 if (Candidate.CallsiteCount > PSI->getHotCountThreshold()) 1121 SampleThreshold = SampleHotCallSiteThreshold; 1122 else if (!ProfileSizeInline) 1123 return InlineCost::getNever("cold callsite"); 1124 } 1125 1126 Function *Callee = Candidate.CallInstr->getCalledFunction(); 1127 assert(Callee && "Expect a definition for inline candidate of direct call"); 1128 1129 InlineParams Params = getInlineParams(); 1130 Params.ComputeFullInlineCost = true; 1131 // Checks if there is anything in the reachable portion of the callee at 1132 // this callsite that makes this inlining potentially illegal. Need to 1133 // set ComputeFullInlineCost, otherwise getInlineCost may return early 1134 // when cost exceeds threshold without checking all IRs in the callee. 1135 // The acutal cost does not matter because we only checks isNever() to 1136 // see if it is legal to inline the callsite. 1137 InlineCost Cost = getInlineCost(*Candidate.CallInstr, Callee, Params, 1138 GetTTI(*Callee), GetAC, GetTLI); 1139 1140 // Honor always inline and never inline from call analyzer 1141 if (Cost.isNever() || Cost.isAlways()) 1142 return Cost; 1143 1144 // For old FDO inliner, we inline the call site as long as cost is not 1145 // "Never". The cost-benefit check is done earlier. 1146 if (!CallsitePrioritizedInline) { 1147 return InlineCost::get(Cost.getCost(), INT_MAX); 1148 } 1149 1150 // Otherwise only use the cost from call analyzer, but overwite threshold with 1151 // Sample PGO threshold. 1152 return InlineCost::get(Cost.getCost(), SampleThreshold); 1153 } 1154 1155 bool SampleProfileLoader::inlineHotFunctionsWithPriority( 1156 Function &F, DenseSet<GlobalValue::GUID> &InlinedGUIDs) { 1157 assert(ProfileIsCS && "Prioritiy based inliner only works with CSSPGO now"); 1158 1159 // ProfAccForSymsInList is used in callsiteIsHot. The assertion makes sure 1160 // Profile symbol list is ignored when profile-sample-accurate is on. 1161 assert((!ProfAccForSymsInList || 1162 (!ProfileSampleAccurate && 1163 !F.hasFnAttribute("profile-sample-accurate"))) && 1164 "ProfAccForSymsInList should be false when profile-sample-accurate " 1165 "is enabled"); 1166 1167 // Populating worklist with initial call sites from root inliner, along 1168 // with call site weights. 1169 CandidateQueue CQueue; 1170 InlineCandidate NewCandidate; 1171 for (auto &BB : F) { 1172 for (auto &I : BB.getInstList()) { 1173 auto *CB = dyn_cast<CallBase>(&I); 1174 if (!CB) 1175 continue; 1176 if (getInlineCandidate(&NewCandidate, CB)) 1177 CQueue.push(NewCandidate); 1178 } 1179 } 1180 1181 // Cap the size growth from profile guided inlining. This is needed even 1182 // though cost of each inline candidate already accounts for callee size, 1183 // because with top-down inlining, we can grow inliner size significantly 1184 // with large number of smaller inlinees each pass the cost check. 1185 assert(ProfileInlineLimitMax >= ProfileInlineLimitMin && 1186 "Max inline size limit should not be smaller than min inline size " 1187 "limit."); 1188 unsigned SizeLimit = F.getInstructionCount() * ProfileInlineGrowthLimit; 1189 SizeLimit = std::min(SizeLimit, (unsigned)ProfileInlineLimitMax); 1190 SizeLimit = std::max(SizeLimit, (unsigned)ProfileInlineLimitMin); 1191 if (ExternalInlineAdvisor) 1192 SizeLimit = std::numeric_limits<unsigned>::max(); 1193 1194 // Perform iterative BFS call site prioritized inlining 1195 bool Changed = false; 1196 while (!CQueue.empty() && F.getInstructionCount() < SizeLimit) { 1197 InlineCandidate Candidate = CQueue.top(); 1198 CQueue.pop(); 1199 CallBase *I = Candidate.CallInstr; 1200 Function *CalledFunction = I->getCalledFunction(); 1201 1202 if (CalledFunction == &F) 1203 continue; 1204 if (I->isIndirectCall()) { 1205 uint64_t Sum; 1206 auto CalleeSamples = findIndirectCallFunctionSamples(*I, Sum); 1207 uint64_t SumOrigin = Sum; 1208 Sum *= Candidate.CallsiteDistribution; 1209 for (const auto *FS : CalleeSamples) { 1210 // TODO: Consider disable pre-lTO ICP for MonoLTO as well 1211 if (LTOPhase == ThinOrFullLTOPhase::ThinLTOPreLink) { 1212 FS->findInlinedFunctions(InlinedGUIDs, F.getParent(), 1213 PSI->getOrCompHotCountThreshold()); 1214 continue; 1215 } 1216 uint64_t EntryCountDistributed = 1217 FS->getEntrySamples() * Candidate.CallsiteDistribution; 1218 // In addition to regular inline cost check, we also need to make sure 1219 // ICP isn't introducing excessive speculative checks even if individual 1220 // target looks beneficial to promote and inline. That means we should 1221 // only do ICP when there's a small number dominant targets. 1222 if (EntryCountDistributed < SumOrigin / ProfileICPThreshold) 1223 break; 1224 // TODO: Fix CallAnalyzer to handle all indirect calls. 1225 // For indirect call, we don't run CallAnalyzer to get InlineCost 1226 // before actual inlining. This is because we could see two different 1227 // types from the same definition, which makes CallAnalyzer choke as 1228 // it's expecting matching parameter type on both caller and callee 1229 // side. See example from PR18962 for the triggering cases (the bug was 1230 // fixed, but we generate different types). 1231 if (!PSI->isHotCount(EntryCountDistributed)) 1232 break; 1233 SmallVector<CallBase *, 8> InlinedCallSites; 1234 // Attach function profile for promoted indirect callee, and update 1235 // call site count for the promoted inline candidate too. 1236 Candidate = {I, FS, EntryCountDistributed, 1237 Candidate.CallsiteDistribution}; 1238 if (tryPromoteAndInlineCandidate(F, Candidate, SumOrigin, Sum, 1239 &InlinedCallSites)) { 1240 for (auto *CB : InlinedCallSites) { 1241 if (getInlineCandidate(&NewCandidate, CB)) 1242 CQueue.emplace(NewCandidate); 1243 } 1244 Changed = true; 1245 } 1246 } 1247 } else if (CalledFunction && CalledFunction->getSubprogram() && 1248 !CalledFunction->isDeclaration()) { 1249 SmallVector<CallBase *, 8> InlinedCallSites; 1250 if (tryInlineCandidate(Candidate, &InlinedCallSites)) { 1251 for (auto *CB : InlinedCallSites) { 1252 if (getInlineCandidate(&NewCandidate, CB)) 1253 CQueue.emplace(NewCandidate); 1254 } 1255 Changed = true; 1256 } 1257 } else if (LTOPhase == ThinOrFullLTOPhase::ThinLTOPreLink) { 1258 findCalleeFunctionSamples(*I)->findInlinedFunctions( 1259 InlinedGUIDs, F.getParent(), PSI->getOrCompHotCountThreshold()); 1260 } 1261 } 1262 1263 if (!CQueue.empty()) { 1264 if (SizeLimit == (unsigned)ProfileInlineLimitMax) 1265 ++NumCSInlinedHitMaxLimit; 1266 else if (SizeLimit == (unsigned)ProfileInlineLimitMin) 1267 ++NumCSInlinedHitMinLimit; 1268 else 1269 ++NumCSInlinedHitGrowthLimit; 1270 } 1271 1272 return Changed; 1273 } 1274 1275 /// Returns the sorted CallTargetMap \p M by count in descending order. 1276 static SmallVector<InstrProfValueData, 2> 1277 GetSortedValueDataFromCallTargets(const SampleRecord::CallTargetMap &M) { 1278 SmallVector<InstrProfValueData, 2> R; 1279 for (const auto &I : SampleRecord::SortCallTargets(M)) { 1280 R.emplace_back( 1281 InstrProfValueData{FunctionSamples::getGUID(I.first), I.second}); 1282 } 1283 return R; 1284 } 1285 1286 // Generate MD_prof metadata for every branch instruction using the 1287 // edge weights computed during propagation. 1288 void SampleProfileLoader::generateMDProfMetadata(Function &F) { 1289 // Generate MD_prof metadata for every branch instruction using the 1290 // edge weights computed during propagation. 1291 LLVM_DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n"); 1292 LLVMContext &Ctx = F.getContext(); 1293 MDBuilder MDB(Ctx); 1294 for (auto &BI : F) { 1295 BasicBlock *BB = &BI; 1296 1297 if (BlockWeights[BB]) { 1298 for (auto &I : BB->getInstList()) { 1299 if (!isa<CallInst>(I) && !isa<InvokeInst>(I)) 1300 continue; 1301 if (!cast<CallBase>(I).getCalledFunction()) { 1302 const DebugLoc &DLoc = I.getDebugLoc(); 1303 if (!DLoc) 1304 continue; 1305 const DILocation *DIL = DLoc; 1306 const FunctionSamples *FS = findFunctionSamples(I); 1307 if (!FS) 1308 continue; 1309 auto CallSite = FunctionSamples::getCallSiteIdentifier(DIL); 1310 auto T = FS->findCallTargetMapAt(CallSite); 1311 if (!T || T.get().empty()) 1312 continue; 1313 // Prorate the callsite counts to reflect what is already done to the 1314 // callsite, such as ICP or calliste cloning. 1315 if (FunctionSamples::ProfileIsProbeBased) { 1316 if (Optional<PseudoProbe> Probe = extractProbe(I)) { 1317 if (Probe->Factor < 1) 1318 T = SampleRecord::adjustCallTargets(T.get(), Probe->Factor); 1319 } 1320 } 1321 SmallVector<InstrProfValueData, 2> SortedCallTargets = 1322 GetSortedValueDataFromCallTargets(T.get()); 1323 uint64_t Sum = 0; 1324 for (const auto &C : T.get()) 1325 Sum += C.second; 1326 // With CSSPGO all indirect call targets are counted torwards the 1327 // original indirect call site in the profile, including both 1328 // inlined and non-inlined targets. 1329 if (!FunctionSamples::ProfileIsCS) { 1330 if (const FunctionSamplesMap *M = 1331 FS->findFunctionSamplesMapAt(CallSite)) { 1332 for (const auto &NameFS : *M) 1333 Sum += NameFS.second.getEntrySamples(); 1334 } 1335 } 1336 updateIDTMetaData(I, SortedCallTargets, Sum); 1337 } else if (!isa<IntrinsicInst>(&I)) { 1338 I.setMetadata(LLVMContext::MD_prof, 1339 MDB.createBranchWeights( 1340 {static_cast<uint32_t>(BlockWeights[BB])})); 1341 } 1342 } 1343 } 1344 Instruction *TI = BB->getTerminator(); 1345 if (TI->getNumSuccessors() == 1) 1346 continue; 1347 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI)) 1348 continue; 1349 1350 DebugLoc BranchLoc = TI->getDebugLoc(); 1351 LLVM_DEBUG(dbgs() << "\nGetting weights for branch at line " 1352 << ((BranchLoc) ? Twine(BranchLoc.getLine()) 1353 : Twine("<UNKNOWN LOCATION>")) 1354 << ".\n"); 1355 SmallVector<uint32_t, 4> Weights; 1356 uint32_t MaxWeight = 0; 1357 Instruction *MaxDestInst; 1358 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) { 1359 BasicBlock *Succ = TI->getSuccessor(I); 1360 Edge E = std::make_pair(BB, Succ); 1361 uint64_t Weight = EdgeWeights[E]; 1362 LLVM_DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E)); 1363 // Use uint32_t saturated arithmetic to adjust the incoming weights, 1364 // if needed. Sample counts in profiles are 64-bit unsigned values, 1365 // but internally branch weights are expressed as 32-bit values. 1366 if (Weight > std::numeric_limits<uint32_t>::max()) { 1367 LLVM_DEBUG(dbgs() << " (saturated due to uint32_t overflow)"); 1368 Weight = std::numeric_limits<uint32_t>::max(); 1369 } 1370 // Weight is added by one to avoid propagation errors introduced by 1371 // 0 weights. 1372 Weights.push_back(static_cast<uint32_t>(Weight + 1)); 1373 if (Weight != 0) { 1374 if (Weight > MaxWeight) { 1375 MaxWeight = Weight; 1376 MaxDestInst = Succ->getFirstNonPHIOrDbgOrLifetime(); 1377 } 1378 } 1379 } 1380 1381 uint64_t TempWeight; 1382 // Only set weights if there is at least one non-zero weight. 1383 // In any other case, let the analyzer set weights. 1384 // Do not set weights if the weights are present. In ThinLTO, the profile 1385 // annotation is done twice. If the first annotation already set the 1386 // weights, the second pass does not need to set it. 1387 if (MaxWeight > 0 && !TI->extractProfTotalWeight(TempWeight)) { 1388 LLVM_DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n"); 1389 TI->setMetadata(LLVMContext::MD_prof, 1390 MDB.createBranchWeights(Weights)); 1391 ORE->emit([&]() { 1392 return OptimizationRemark(DEBUG_TYPE, "PopularDest", MaxDestInst) 1393 << "most popular destination for conditional branches at " 1394 << ore::NV("CondBranchesLoc", BranchLoc); 1395 }); 1396 } else { 1397 LLVM_DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n"); 1398 } 1399 } 1400 } 1401 1402 /// Once all the branch weights are computed, we emit the MD_prof 1403 /// metadata on BB using the computed values for each of its branches. 1404 /// 1405 /// \param F The function to query. 1406 /// 1407 /// \returns true if \p F was modified. Returns false, otherwise. 1408 bool SampleProfileLoader::emitAnnotations(Function &F) { 1409 bool Changed = false; 1410 1411 if (FunctionSamples::ProfileIsProbeBased) { 1412 if (!ProbeManager->profileIsValid(F, *Samples)) { 1413 LLVM_DEBUG( 1414 dbgs() << "Profile is invalid due to CFG mismatch for Function " 1415 << F.getName()); 1416 ++NumMismatchedProfile; 1417 return false; 1418 } 1419 ++NumMatchedProfile; 1420 } else { 1421 if (getFunctionLoc(F) == 0) 1422 return false; 1423 1424 LLVM_DEBUG(dbgs() << "Line number for the first instruction in " 1425 << F.getName() << ": " << getFunctionLoc(F) << "\n"); 1426 } 1427 1428 DenseSet<GlobalValue::GUID> InlinedGUIDs; 1429 if (ProfileIsCS && CallsitePrioritizedInline) 1430 Changed |= inlineHotFunctionsWithPriority(F, InlinedGUIDs); 1431 else 1432 Changed |= inlineHotFunctions(F, InlinedGUIDs); 1433 1434 Changed |= computeAndPropagateWeights(F, InlinedGUIDs); 1435 1436 if (Changed) 1437 generateMDProfMetadata(F); 1438 1439 emitCoverageRemarks(F); 1440 return Changed; 1441 } 1442 1443 char SampleProfileLoaderLegacyPass::ID = 0; 1444 1445 INITIALIZE_PASS_BEGIN(SampleProfileLoaderLegacyPass, "sample-profile", 1446 "Sample Profile loader", false, false) 1447 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) 1448 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 1449 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass) 1450 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass) 1451 INITIALIZE_PASS_END(SampleProfileLoaderLegacyPass, "sample-profile", 1452 "Sample Profile loader", false, false) 1453 1454 // Add inlined profile call edges to the call graph. 1455 void SampleProfileLoader::addCallGraphEdges(CallGraph &CG, 1456 const FunctionSamples &Samples) { 1457 Function *Caller = SymbolMap.lookup(Samples.getFuncName()); 1458 if (!Caller || Caller->isDeclaration()) 1459 return; 1460 1461 // Skip non-inlined call edges which are not important since top down inlining 1462 // for non-CS profile is to get more precise profile matching, not to enable 1463 // more inlining. 1464 1465 for (const auto &CallsiteSamples : Samples.getCallsiteSamples()) { 1466 for (const auto &InlinedSamples : CallsiteSamples.second) { 1467 Function *Callee = SymbolMap.lookup(InlinedSamples.first); 1468 if (Callee && !Callee->isDeclaration()) 1469 CG[Caller]->addCalledFunction(nullptr, CG[Callee]); 1470 addCallGraphEdges(CG, InlinedSamples.second); 1471 } 1472 } 1473 } 1474 1475 // Replace call graph edges with dynamic call edges from the profile. 1476 void SampleProfileLoader::replaceCallGraphEdges( 1477 CallGraph &CG, StringMap<Function *> &SymbolMap) { 1478 // Remove static call edges from the call graph except for the ones from the 1479 // root which make the call graph connected. 1480 for (const auto &Node : CG) 1481 if (Node.second.get() != CG.getExternalCallingNode()) 1482 Node.second->removeAllCalledFunctions(); 1483 1484 // Add profile call edges to the call graph. 1485 if (ProfileIsCS) { 1486 ContextTracker->addCallGraphEdges(CG, SymbolMap); 1487 } else { 1488 for (const auto &Samples : Reader->getProfiles()) 1489 addCallGraphEdges(CG, Samples.second); 1490 } 1491 } 1492 1493 std::vector<Function *> 1494 SampleProfileLoader::buildFunctionOrder(Module &M, CallGraph *CG) { 1495 std::vector<Function *> FunctionOrderList; 1496 FunctionOrderList.reserve(M.size()); 1497 1498 if (!ProfileTopDownLoad || CG == nullptr) { 1499 if (ProfileMergeInlinee) { 1500 // Disable ProfileMergeInlinee if profile is not loaded in top down order, 1501 // because the profile for a function may be used for the profile 1502 // annotation of its outline copy before the profile merging of its 1503 // non-inlined inline instances, and that is not the way how 1504 // ProfileMergeInlinee is supposed to work. 1505 ProfileMergeInlinee = false; 1506 } 1507 1508 for (Function &F : M) 1509 if (!F.isDeclaration() && F.hasFnAttribute("use-sample-profile")) 1510 FunctionOrderList.push_back(&F); 1511 return FunctionOrderList; 1512 } 1513 1514 assert(&CG->getModule() == &M); 1515 1516 // Add indirect call edges from profile to augment the static call graph. 1517 // Functions will be processed in a top-down order defined by the static call 1518 // graph. Adjusting the order by considering indirect call edges from the 1519 // profile (which don't exist in the static call graph) can enable the 1520 // inlining of indirect call targets by processing the caller before them. 1521 // TODO: enable this for non-CS profile and fix the counts returning logic to 1522 // have a full support for indirect calls. 1523 if (UseProfileIndirectCallEdges && ProfileIsCS) { 1524 for (auto &Entry : *CG) { 1525 const auto *F = Entry.first; 1526 if (!F || F->isDeclaration() || !F->hasFnAttribute("use-sample-profile")) 1527 continue; 1528 auto &AllContexts = ContextTracker->getAllContextSamplesFor(F->getName()); 1529 if (AllContexts.empty()) 1530 continue; 1531 1532 for (const auto &BB : *F) { 1533 for (const auto &I : BB.getInstList()) { 1534 const auto *CB = dyn_cast<CallBase>(&I); 1535 if (!CB || !CB->isIndirectCall()) 1536 continue; 1537 const DebugLoc &DLoc = I.getDebugLoc(); 1538 if (!DLoc) 1539 continue; 1540 auto CallSite = FunctionSamples::getCallSiteIdentifier(DLoc); 1541 for (FunctionSamples *Samples : AllContexts) { 1542 if (auto CallTargets = Samples->findCallTargetMapAt(CallSite)) { 1543 for (const auto &Target : CallTargets.get()) { 1544 Function *Callee = SymbolMap.lookup(Target.first()); 1545 if (Callee && !Callee->isDeclaration()) 1546 Entry.second->addCalledFunction(nullptr, (*CG)[Callee]); 1547 } 1548 } 1549 } 1550 } 1551 } 1552 } 1553 } 1554 1555 // Compute a top-down order the profile which is used to sort functions in 1556 // one SCC later. The static processing order computed for an SCC may not 1557 // reflect the call contexts in the context-sensitive profile, thus may cause 1558 // potential inlining to be overlooked. The function order in one SCC is being 1559 // adjusted to a top-down order based on the profile to favor more inlining. 1560 DenseMap<Function *, uint64_t> ProfileOrderMap; 1561 if (UseProfileTopDownOrder || 1562 (ProfileIsCS && !UseProfileTopDownOrder.getNumOccurrences())) { 1563 // Create a static call graph. The call edges are not important since they 1564 // will be replaced by dynamic edges from the profile. 1565 CallGraph ProfileCG(M); 1566 replaceCallGraphEdges(ProfileCG, SymbolMap); 1567 scc_iterator<CallGraph *> CGI = scc_begin(&ProfileCG); 1568 uint64_t I = 0; 1569 while (!CGI.isAtEnd()) { 1570 for (CallGraphNode *Node : *CGI) { 1571 if (auto *F = Node->getFunction()) 1572 ProfileOrderMap[F] = ++I; 1573 } 1574 ++CGI; 1575 } 1576 } 1577 1578 scc_iterator<CallGraph *> CGI = scc_begin(CG); 1579 while (!CGI.isAtEnd()) { 1580 uint64_t Start = FunctionOrderList.size(); 1581 for (CallGraphNode *Node : *CGI) { 1582 auto *F = Node->getFunction(); 1583 if (F && !F->isDeclaration() && F->hasFnAttribute("use-sample-profile")) 1584 FunctionOrderList.push_back(F); 1585 } 1586 1587 // Sort nodes in SCC based on the profile top-down order. 1588 if (!ProfileOrderMap.empty()) { 1589 std::stable_sort(FunctionOrderList.begin() + Start, 1590 FunctionOrderList.end(), 1591 [&ProfileOrderMap](Function *Left, Function *Right) { 1592 return ProfileOrderMap[Left] < ProfileOrderMap[Right]; 1593 }); 1594 } 1595 1596 ++CGI; 1597 } 1598 1599 LLVM_DEBUG({ 1600 dbgs() << "Function processing order:\n"; 1601 for (auto F : reverse(FunctionOrderList)) { 1602 dbgs() << F->getName() << "\n"; 1603 } 1604 }); 1605 1606 std::reverse(FunctionOrderList.begin(), FunctionOrderList.end()); 1607 return FunctionOrderList; 1608 } 1609 1610 bool SampleProfileLoader::doInitialization(Module &M, 1611 FunctionAnalysisManager *FAM) { 1612 auto &Ctx = M.getContext(); 1613 1614 auto ReaderOrErr = 1615 SampleProfileReader::create(Filename, Ctx, RemappingFilename); 1616 if (std::error_code EC = ReaderOrErr.getError()) { 1617 std::string Msg = "Could not open profile: " + EC.message(); 1618 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg)); 1619 return false; 1620 } 1621 Reader = std::move(ReaderOrErr.get()); 1622 Reader->setSkipFlatProf(LTOPhase == ThinOrFullLTOPhase::ThinLTOPostLink); 1623 Reader->collectFuncsFrom(M); 1624 if (std::error_code EC = Reader->read()) { 1625 std::string Msg = "profile reading failed: " + EC.message(); 1626 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg)); 1627 return false; 1628 } 1629 1630 PSL = Reader->getProfileSymbolList(); 1631 1632 // While profile-sample-accurate is on, ignore symbol list. 1633 ProfAccForSymsInList = 1634 ProfileAccurateForSymsInList && PSL && !ProfileSampleAccurate; 1635 if (ProfAccForSymsInList) { 1636 NamesInProfile.clear(); 1637 if (auto NameTable = Reader->getNameTable()) 1638 NamesInProfile.insert(NameTable->begin(), NameTable->end()); 1639 CoverageTracker.setProfAccForSymsInList(true); 1640 } 1641 1642 if (FAM && !ProfileInlineReplayFile.empty()) { 1643 ExternalInlineAdvisor = std::make_unique<ReplayInlineAdvisor>( 1644 M, *FAM, Ctx, /*OriginalAdvisor=*/nullptr, ProfileInlineReplayFile, 1645 /*EmitRemarks=*/false); 1646 if (!ExternalInlineAdvisor->areReplayRemarksLoaded()) 1647 ExternalInlineAdvisor.reset(); 1648 } 1649 1650 // Apply tweaks if context-sensitive profile is available. 1651 if (Reader->profileIsCS()) { 1652 ProfileIsCS = true; 1653 FunctionSamples::ProfileIsCS = true; 1654 1655 // Enable priority-base inliner and size inline by default for CSSPGO. 1656 if (!ProfileSizeInline.getNumOccurrences()) 1657 ProfileSizeInline = true; 1658 if (!CallsitePrioritizedInline.getNumOccurrences()) 1659 CallsitePrioritizedInline = true; 1660 1661 // Tracker for profiles under different context 1662 ContextTracker = 1663 std::make_unique<SampleContextTracker>(Reader->getProfiles()); 1664 } 1665 1666 // Load pseudo probe descriptors for probe-based function samples. 1667 if (Reader->profileIsProbeBased()) { 1668 ProbeManager = std::make_unique<PseudoProbeManager>(M); 1669 if (!ProbeManager->moduleIsProbed(M)) { 1670 const char *Msg = 1671 "Pseudo-probe-based profile requires SampleProfileProbePass"; 1672 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg)); 1673 return false; 1674 } 1675 } 1676 1677 return true; 1678 } 1679 1680 ModulePass *llvm::createSampleProfileLoaderPass() { 1681 return new SampleProfileLoaderLegacyPass(); 1682 } 1683 1684 ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) { 1685 return new SampleProfileLoaderLegacyPass(Name); 1686 } 1687 1688 bool SampleProfileLoader::runOnModule(Module &M, ModuleAnalysisManager *AM, 1689 ProfileSummaryInfo *_PSI, CallGraph *CG) { 1690 GUIDToFuncNameMapper Mapper(M, *Reader, GUIDToFuncNameMap); 1691 1692 PSI = _PSI; 1693 if (M.getProfileSummary(/* IsCS */ false) == nullptr) { 1694 M.setProfileSummary(Reader->getSummary().getMD(M.getContext()), 1695 ProfileSummary::PSK_Sample); 1696 PSI->refresh(); 1697 } 1698 // Compute the total number of samples collected in this profile. 1699 for (const auto &I : Reader->getProfiles()) 1700 TotalCollectedSamples += I.second.getTotalSamples(); 1701 1702 auto Remapper = Reader->getRemapper(); 1703 // Populate the symbol map. 1704 for (const auto &N_F : M.getValueSymbolTable()) { 1705 StringRef OrigName = N_F.getKey(); 1706 Function *F = dyn_cast<Function>(N_F.getValue()); 1707 if (F == nullptr) 1708 continue; 1709 SymbolMap[OrigName] = F; 1710 auto pos = OrigName.find('.'); 1711 if (pos != StringRef::npos) { 1712 StringRef NewName = OrigName.substr(0, pos); 1713 auto r = SymbolMap.insert(std::make_pair(NewName, F)); 1714 // Failiing to insert means there is already an entry in SymbolMap, 1715 // thus there are multiple functions that are mapped to the same 1716 // stripped name. In this case of name conflicting, set the value 1717 // to nullptr to avoid confusion. 1718 if (!r.second) 1719 r.first->second = nullptr; 1720 OrigName = NewName; 1721 } 1722 // Insert the remapped names into SymbolMap. 1723 if (Remapper) { 1724 if (auto MapName = Remapper->lookUpNameInProfile(OrigName)) { 1725 if (*MapName == OrigName) 1726 continue; 1727 SymbolMap.insert(std::make_pair(*MapName, F)); 1728 } 1729 } 1730 } 1731 1732 bool retval = false; 1733 for (auto F : buildFunctionOrder(M, CG)) { 1734 assert(!F->isDeclaration()); 1735 clearFunctionData(); 1736 retval |= runOnFunction(*F, AM); 1737 } 1738 1739 // Account for cold calls not inlined.... 1740 if (!ProfileIsCS) 1741 for (const std::pair<Function *, NotInlinedProfileInfo> &pair : 1742 notInlinedCallInfo) 1743 updateProfileCallee(pair.first, pair.second.entryCount); 1744 1745 return retval; 1746 } 1747 1748 bool SampleProfileLoaderLegacyPass::runOnModule(Module &M) { 1749 ACT = &getAnalysis<AssumptionCacheTracker>(); 1750 TTIWP = &getAnalysis<TargetTransformInfoWrapperPass>(); 1751 TLIWP = &getAnalysis<TargetLibraryInfoWrapperPass>(); 1752 ProfileSummaryInfo *PSI = 1753 &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); 1754 return SampleLoader.runOnModule(M, nullptr, PSI, nullptr); 1755 } 1756 1757 bool SampleProfileLoader::runOnFunction(Function &F, ModuleAnalysisManager *AM) { 1758 LLVM_DEBUG(dbgs() << "\n\nProcessing Function " << F.getName() << "\n"); 1759 DILocation2SampleMap.clear(); 1760 // By default the entry count is initialized to -1, which will be treated 1761 // conservatively by getEntryCount as the same as unknown (None). This is 1762 // to avoid newly added code to be treated as cold. If we have samples 1763 // this will be overwritten in emitAnnotations. 1764 uint64_t initialEntryCount = -1; 1765 1766 ProfAccForSymsInList = ProfileAccurateForSymsInList && PSL; 1767 if (ProfileSampleAccurate || F.hasFnAttribute("profile-sample-accurate")) { 1768 // initialize all the function entry counts to 0. It means all the 1769 // functions without profile will be regarded as cold. 1770 initialEntryCount = 0; 1771 // profile-sample-accurate is a user assertion which has a higher precedence 1772 // than symbol list. When profile-sample-accurate is on, ignore symbol list. 1773 ProfAccForSymsInList = false; 1774 } 1775 CoverageTracker.setProfAccForSymsInList(ProfAccForSymsInList); 1776 1777 // PSL -- profile symbol list include all the symbols in sampled binary. 1778 // If ProfileAccurateForSymsInList is enabled, PSL is used to treat 1779 // old functions without samples being cold, without having to worry 1780 // about new and hot functions being mistakenly treated as cold. 1781 if (ProfAccForSymsInList) { 1782 // Initialize the entry count to 0 for functions in the list. 1783 if (PSL->contains(F.getName())) 1784 initialEntryCount = 0; 1785 1786 // Function in the symbol list but without sample will be regarded as 1787 // cold. To minimize the potential negative performance impact it could 1788 // have, we want to be a little conservative here saying if a function 1789 // shows up in the profile, no matter as outline function, inline instance 1790 // or call targets, treat the function as not being cold. This will handle 1791 // the cases such as most callsites of a function are inlined in sampled 1792 // binary but not inlined in current build (because of source code drift, 1793 // imprecise debug information, or the callsites are all cold individually 1794 // but not cold accumulatively...), so the outline function showing up as 1795 // cold in sampled binary will actually not be cold after current build. 1796 StringRef CanonName = FunctionSamples::getCanonicalFnName(F); 1797 if (NamesInProfile.count(CanonName)) 1798 initialEntryCount = -1; 1799 } 1800 1801 // Initialize entry count when the function has no existing entry 1802 // count value. 1803 if (!F.getEntryCount().hasValue()) 1804 F.setEntryCount(ProfileCount(initialEntryCount, Function::PCT_Real)); 1805 std::unique_ptr<OptimizationRemarkEmitter> OwnedORE; 1806 if (AM) { 1807 auto &FAM = 1808 AM->getResult<FunctionAnalysisManagerModuleProxy>(*F.getParent()) 1809 .getManager(); 1810 ORE = &FAM.getResult<OptimizationRemarkEmitterAnalysis>(F); 1811 } else { 1812 OwnedORE = std::make_unique<OptimizationRemarkEmitter>(&F); 1813 ORE = OwnedORE.get(); 1814 } 1815 1816 if (ProfileIsCS) 1817 Samples = ContextTracker->getBaseSamplesFor(F); 1818 else 1819 Samples = Reader->getSamplesFor(F); 1820 1821 if (Samples && !Samples->empty()) 1822 return emitAnnotations(F); 1823 return false; 1824 } 1825 1826 PreservedAnalyses SampleProfileLoaderPass::run(Module &M, 1827 ModuleAnalysisManager &AM) { 1828 FunctionAnalysisManager &FAM = 1829 AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); 1830 1831 auto GetAssumptionCache = [&](Function &F) -> AssumptionCache & { 1832 return FAM.getResult<AssumptionAnalysis>(F); 1833 }; 1834 auto GetTTI = [&](Function &F) -> TargetTransformInfo & { 1835 return FAM.getResult<TargetIRAnalysis>(F); 1836 }; 1837 auto GetTLI = [&](Function &F) -> const TargetLibraryInfo & { 1838 return FAM.getResult<TargetLibraryAnalysis>(F); 1839 }; 1840 1841 SampleProfileLoader SampleLoader( 1842 ProfileFileName.empty() ? SampleProfileFile : ProfileFileName, 1843 ProfileRemappingFileName.empty() ? SampleProfileRemappingFile 1844 : ProfileRemappingFileName, 1845 LTOPhase, GetAssumptionCache, GetTTI, GetTLI); 1846 1847 if (!SampleLoader.doInitialization(M, &FAM)) 1848 return PreservedAnalyses::all(); 1849 1850 ProfileSummaryInfo *PSI = &AM.getResult<ProfileSummaryAnalysis>(M); 1851 CallGraph &CG = AM.getResult<CallGraphAnalysis>(M); 1852 if (!SampleLoader.runOnModule(M, &AM, PSI, &CG)) 1853 return PreservedAnalyses::all(); 1854 1855 return PreservedAnalyses::none(); 1856 } 1857