//===- SampleProfile.cpp - Incorporate sample profiles into the IR --------===// // // The LLVM Compiler Infrastructure // // This file is distributed under the University of Illinois Open Source // License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // // This file implements the SampleProfileLoader transformation. This pass // reads a profile file generated by a sampling profiler (e.g. Linux Perf - // http://perf.wiki.kernel.org/) and generates IR metadata to reflect the // profile information in the given profile. // // This pass generates branch weight annotations on the IR: // // - prof: Represents branch weights. This annotation is added to branches // to indicate the weights of each edge coming out of the branch. // The weight of each edge is the weight of the target block for // that edge. The weight of a block B is computed as the maximum // number of samples found in B. // //===----------------------------------------------------------------------===// #include "llvm/Transforms/SampleProfile.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/SmallSet.h" #include "llvm/ADT/StringRef.h" #include "llvm/Analysis/AssumptionCache.h" #include "llvm/Analysis/LoopInfo.h" #include "llvm/Analysis/PostDominators.h" #include "llvm/IR/Constants.h" #include "llvm/IR/DebugInfo.h" #include "llvm/IR/DiagnosticInfo.h" #include "llvm/IR/Dominators.h" #include "llvm/IR/Function.h" #include "llvm/IR/InstIterator.h" #include "llvm/IR/Instructions.h" #include "llvm/IR/IntrinsicInst.h" #include "llvm/IR/LLVMContext.h" #include "llvm/IR/MDBuilder.h" #include "llvm/IR/Metadata.h" #include "llvm/IR/Module.h" #include "llvm/Pass.h" #include "llvm/ProfileData/SampleProfReader.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorOr.h" #include "llvm/Support/Format.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Transforms/IPO.h" #include "llvm/Transforms/Utils/Cloning.h" #include using namespace llvm; using namespace sampleprof; #define DEBUG_TYPE "sample-profile" // Command line option to specify the file to read samples from. This is // mainly used for debugging. static cl::opt SampleProfileFile( "sample-profile-file", cl::init(""), cl::value_desc("filename"), cl::desc("Profile file loaded by -sample-profile"), cl::Hidden); static cl::opt SampleProfileMaxPropagateIterations( "sample-profile-max-propagate-iterations", cl::init(100), cl::desc("Maximum number of iterations to go through when propagating " "sample block/edge weights through the CFG.")); static cl::opt SampleProfileRecordCoverage( "sample-profile-check-record-coverage", cl::init(0), cl::value_desc("N"), cl::desc("Emit a warning if less than N% of records in the input profile " "are matched to the IR.")); static cl::opt SampleProfileSampleCoverage( "sample-profile-check-sample-coverage", cl::init(0), cl::value_desc("N"), cl::desc("Emit a warning if less than N% of samples in the input profile " "are matched to the IR.")); static cl::opt SampleProfileHotThreshold( "sample-profile-inline-hot-threshold", cl::init(0.1), cl::value_desc("N"), cl::desc("Inlined functions that account for more than N% of all samples " "collected in the parent function, will be inlined again.")); namespace { typedef DenseMap BlockWeightMap; typedef DenseMap EquivalenceClassMap; typedef std::pair Edge; typedef DenseMap EdgeWeightMap; typedef DenseMap> BlockEdgeMap; class SampleCoverageTracker { public: SampleCoverageTracker() : SampleCoverage(), TotalUsedSamples(0) {} bool markSamplesUsed(const FunctionSamples *FS, uint32_t LineOffset, uint32_t Discriminator, uint64_t Samples); unsigned computeCoverage(unsigned Used, unsigned Total) const; unsigned countUsedRecords(const FunctionSamples *FS) const; unsigned countBodyRecords(const FunctionSamples *FS) const; uint64_t getTotalUsedSamples() const { return TotalUsedSamples; } uint64_t countBodySamples(const FunctionSamples *FS) const; void clear() { SampleCoverage.clear(); TotalUsedSamples = 0; } private: typedef std::map BodySampleCoverageMap; typedef DenseMap FunctionSamplesCoverageMap; /// Coverage map for sampling records. /// /// This map keeps a record of sampling records that have been matched to /// an IR instruction. This is used to detect some form of staleness in /// profiles (see flag -sample-profile-check-coverage). /// /// Each entry in the map corresponds to a FunctionSamples instance. This is /// another map that counts how many times the sample record at the /// given location has been used. FunctionSamplesCoverageMap SampleCoverage; /// Number of samples used from the profile. /// /// When a sampling record is used for the first time, the samples from /// that record are added to this accumulator. Coverage is later computed /// based on the total number of samples available in this function and /// its callsites. /// /// Note that this accumulator tracks samples used from a single function /// and all the inlined callsites. Strictly, we should have a map of counters /// keyed by FunctionSamples pointers, but these stats are cleared after /// every function, so we just need to keep a single counter. uint64_t TotalUsedSamples; }; /// \brief Sample profile pass. /// /// This pass reads profile data from the file specified by /// -sample-profile-file and annotates every affected function with the /// profile information found in that file. class SampleProfileLoader { public: SampleProfileLoader(StringRef Name = SampleProfileFile) : DT(nullptr), PDT(nullptr), LI(nullptr), ACT(nullptr), Reader(), Samples(nullptr), Filename(Name), ProfileIsValid(false), TotalCollectedSamples(0) {} bool doInitialization(Module &M); bool runOnModule(Module &M); void setACT(AssumptionCacheTracker *A) { ACT = A; } void dump() { Reader->dump(); } protected: bool runOnFunction(Function &F); unsigned getFunctionLoc(Function &F); bool emitAnnotations(Function &F); ErrorOr getInstWeight(const Instruction &I); ErrorOr getBlockWeight(const BasicBlock *BB); const FunctionSamples *findCalleeFunctionSamples(const Instruction &I) const; const FunctionSamples *findFunctionSamples(const Instruction &I) const; bool inlineHotFunctions(Function &F); void printEdgeWeight(raw_ostream &OS, Edge E); void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const; void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB); bool computeBlockWeights(Function &F); void findEquivalenceClasses(Function &F); void findEquivalencesFor(BasicBlock *BB1, ArrayRef Descendants, DominatorTreeBase *DomTree); void propagateWeights(Function &F); uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge); void buildEdges(Function &F); bool propagateThroughEdges(Function &F, bool UpdateBlockCount); void computeDominanceAndLoopInfo(Function &F); unsigned getOffset(unsigned L, unsigned H) const; void clearFunctionData(); /// \brief Map basic blocks to their computed weights. /// /// The weight of a basic block is defined to be the maximum /// of all the instruction weights in that block. BlockWeightMap BlockWeights; /// \brief Map edges to their computed weights. /// /// Edge weights are computed by propagating basic block weights in /// SampleProfile::propagateWeights. EdgeWeightMap EdgeWeights; /// \brief Set of visited blocks during propagation. SmallPtrSet VisitedBlocks; /// \brief Set of visited edges during propagation. SmallSet VisitedEdges; /// \brief Equivalence classes for block weights. /// /// Two blocks BB1 and BB2 are in the same equivalence class if they /// dominate and post-dominate each other, and they are in the same loop /// nest. When this happens, the two blocks are guaranteed to execute /// the same number of times. EquivalenceClassMap EquivalenceClass; /// \brief Dominance, post-dominance and loop information. std::unique_ptr DT; std::unique_ptr> PDT; std::unique_ptr LI; AssumptionCacheTracker *ACT; /// \brief Predecessors for each basic block in the CFG. BlockEdgeMap Predecessors; /// \brief Successors for each basic block in the CFG. BlockEdgeMap Successors; SampleCoverageTracker CoverageTracker; /// \brief Profile reader object. std::unique_ptr Reader; /// \brief Samples collected for the body of this function. FunctionSamples *Samples; /// \brief Name of the profile file to load. std::string Filename; /// \brief Flag indicating whether the profile input loaded successfully. bool ProfileIsValid; /// \brief Total number of samples collected in this profile. /// /// This is the sum of all the samples collected in all the functions executed /// at runtime. uint64_t TotalCollectedSamples; }; class SampleProfileLoaderLegacyPass : public ModulePass { public: // Class identification, replacement for typeinfo static char ID; SampleProfileLoaderLegacyPass(StringRef Name = SampleProfileFile) : ModulePass(ID), SampleLoader(Name) { initializeSampleProfileLoaderLegacyPassPass( *PassRegistry::getPassRegistry()); } void dump() { SampleLoader.dump(); } bool doInitialization(Module &M) override { return SampleLoader.doInitialization(M); } StringRef getPassName() const override { return "Sample profile pass"; } bool runOnModule(Module &M) override; void getAnalysisUsage(AnalysisUsage &AU) const override { AU.addRequired(); } private: SampleProfileLoader SampleLoader; }; /// Return true if the given callsite is hot wrt to its caller. /// /// Functions that were inlined in the original binary will be represented /// in the inline stack in the sample profile. If the profile shows that /// the original inline decision was "good" (i.e., the callsite is executed /// frequently), then we will recreate the inline decision and apply the /// profile from the inlined callsite. /// /// To decide whether an inlined callsite is hot, we compute the fraction /// of samples used by the callsite with respect to the total number of samples /// collected in the caller. /// /// If that fraction is larger than the default given by /// SampleProfileHotThreshold, the callsite will be inlined again. bool callsiteIsHot(const FunctionSamples *CallerFS, const FunctionSamples *CallsiteFS) { if (!CallsiteFS) return false; // The callsite was not inlined in the original binary. uint64_t ParentTotalSamples = CallerFS->getTotalSamples(); if (ParentTotalSamples == 0) return false; // Avoid division by zero. uint64_t CallsiteTotalSamples = CallsiteFS->getTotalSamples(); if (CallsiteTotalSamples == 0) return false; // Callsite is trivially cold. double PercentSamples = (double)CallsiteTotalSamples / (double)ParentTotalSamples * 100.0; return PercentSamples >= SampleProfileHotThreshold; } } /// Mark as used the sample record for the given function samples at /// (LineOffset, Discriminator). /// /// \returns true if this is the first time we mark the given record. bool SampleCoverageTracker::markSamplesUsed(const FunctionSamples *FS, uint32_t LineOffset, uint32_t Discriminator, uint64_t Samples) { LineLocation Loc(LineOffset, Discriminator); unsigned &Count = SampleCoverage[FS][Loc]; bool FirstTime = (++Count == 1); if (FirstTime) TotalUsedSamples += Samples; return FirstTime; } /// Return the number of sample records that were applied from this profile. /// /// This count does not include records from cold inlined callsites. unsigned SampleCoverageTracker::countUsedRecords(const FunctionSamples *FS) const { auto I = SampleCoverage.find(FS); // The size of the coverage map for FS represents the number of records // that were marked used at least once. unsigned Count = (I != SampleCoverage.end()) ? I->second.size() : 0; // If there are inlined callsites in this function, count the samples found // in the respective bodies. However, do not bother counting callees with 0 // total samples, these are callees that were never invoked at runtime. for (const auto &I : FS->getCallsiteSamples()) { const FunctionSamples *CalleeSamples = &I.second; if (callsiteIsHot(FS, CalleeSamples)) Count += countUsedRecords(CalleeSamples); } return Count; } /// Return the number of sample records in the body of this profile. /// /// This count does not include records from cold inlined callsites. unsigned SampleCoverageTracker::countBodyRecords(const FunctionSamples *FS) const { unsigned Count = FS->getBodySamples().size(); // Only count records in hot callsites. for (const auto &I : FS->getCallsiteSamples()) { const FunctionSamples *CalleeSamples = &I.second; if (callsiteIsHot(FS, CalleeSamples)) Count += countBodyRecords(CalleeSamples); } return Count; } /// Return the number of samples collected in the body of this profile. /// /// This count does not include samples from cold inlined callsites. uint64_t SampleCoverageTracker::countBodySamples(const FunctionSamples *FS) const { uint64_t Total = 0; for (const auto &I : FS->getBodySamples()) Total += I.second.getSamples(); // Only count samples in hot callsites. for (const auto &I : FS->getCallsiteSamples()) { const FunctionSamples *CalleeSamples = &I.second; if (callsiteIsHot(FS, CalleeSamples)) Total += countBodySamples(CalleeSamples); } return Total; } /// Return the fraction of sample records used in this profile. /// /// The returned value is an unsigned integer in the range 0-100 indicating /// the percentage of sample records that were used while applying this /// profile to the associated function. unsigned SampleCoverageTracker::computeCoverage(unsigned Used, unsigned Total) const { assert(Used <= Total && "number of used records cannot exceed the total number of records"); return Total > 0 ? Used * 100 / Total : 100; } /// Clear all the per-function data used to load samples and propagate weights. void SampleProfileLoader::clearFunctionData() { BlockWeights.clear(); EdgeWeights.clear(); VisitedBlocks.clear(); VisitedEdges.clear(); EquivalenceClass.clear(); DT = nullptr; PDT = nullptr; LI = nullptr; Predecessors.clear(); Successors.clear(); CoverageTracker.clear(); } /// \brief Returns the offset of lineno \p L to head_lineno \p H /// /// \param L Lineno /// \param H Header lineno of the function /// /// \returns offset to the header lineno. 16 bits are used to represent offset. /// We assume that a single function will not exceed 65535 LOC. unsigned SampleProfileLoader::getOffset(unsigned L, unsigned H) const { return (L - H) & 0xffff; } /// \brief Print the weight of edge \p E on stream \p OS. /// /// \param OS Stream to emit the output to. /// \param E Edge to print. void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) { OS << "weight[" << E.first->getName() << "->" << E.second->getName() << "]: " << EdgeWeights[E] << "\n"; } /// \brief Print the equivalence class of block \p BB on stream \p OS. /// /// \param OS Stream to emit the output to. /// \param BB Block to print. void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB) { const BasicBlock *Equiv = EquivalenceClass[BB]; OS << "equivalence[" << BB->getName() << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n"; } /// \brief Print the weight of block \p BB on stream \p OS. /// /// \param OS Stream to emit the output to. /// \param BB Block to print. void SampleProfileLoader::printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const { const auto &I = BlockWeights.find(BB); uint64_t W = (I == BlockWeights.end() ? 0 : I->second); OS << "weight[" << BB->getName() << "]: " << W << "\n"; } /// \brief Get the weight for an instruction. /// /// The "weight" of an instruction \p Inst is the number of samples /// collected on that instruction at runtime. To retrieve it, we /// need to compute the line number of \p Inst relative to the start of its /// function. We use HeaderLineno to compute the offset. We then /// look up the samples collected for \p Inst using BodySamples. /// /// \param Inst Instruction to query. /// /// \returns the weight of \p Inst. ErrorOr SampleProfileLoader::getInstWeight(const Instruction &Inst) { const DebugLoc &DLoc = Inst.getDebugLoc(); if (!DLoc) return std::error_code(); const FunctionSamples *FS = findFunctionSamples(Inst); if (!FS) return std::error_code(); // Ignore all intrinsics and branch instructions. // Branch instruction usually contains debug info from sources outside of // the residing basic block, thus we ignore them during annotation. if (isa(Inst) || isa(Inst)) return std::error_code(); // If a call/invoke instruction is inlined in profile, but not inlined here, // it means that the inlined callsite has no sample, thus the call // instruction should have 0 count. bool IsCall = isa(Inst) || isa(Inst); if (IsCall && findCalleeFunctionSamples(Inst)) return 0; const DILocation *DIL = DLoc; unsigned Lineno = DLoc.getLine(); unsigned HeaderLineno = DIL->getScope()->getSubprogram()->getLine(); uint32_t LineOffset = getOffset(Lineno, HeaderLineno); uint32_t Discriminator = DIL->getDiscriminator(); ErrorOr R = IsCall ? FS->findCallSamplesAt(LineOffset, Discriminator) : FS->findSamplesAt(LineOffset, Discriminator); if (R) { bool FirstMark = CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get()); if (FirstMark) { const Function *F = Inst.getParent()->getParent(); LLVMContext &Ctx = F->getContext(); emitOptimizationRemark( Ctx, DEBUG_TYPE, *F, DLoc, Twine("Applied ") + Twine(*R) + " samples from profile (offset: " + Twine(LineOffset) + ((Discriminator) ? Twine(".") + Twine(Discriminator) : "") + ")"); } DEBUG(dbgs() << " " << Lineno << "." << DIL->getDiscriminator() << ":" << Inst << " (line offset: " << Lineno - HeaderLineno << "." << DIL->getDiscriminator() << " - weight: " << R.get() << ")\n"); } return R; } /// \brief Compute the weight of a basic block. /// /// The weight of basic block \p BB is the maximum weight of all the /// instructions in BB. /// /// \param BB The basic block to query. /// /// \returns the weight for \p BB. ErrorOr SampleProfileLoader::getBlockWeight(const BasicBlock *BB) { uint64_t Max = 0; bool HasWeight = false; for (auto &I : BB->getInstList()) { const ErrorOr &R = getInstWeight(I); if (R) { Max = std::max(Max, R.get()); HasWeight = true; } } return HasWeight ? ErrorOr(Max) : std::error_code(); } /// \brief Compute and store the weights of every basic block. /// /// This populates the BlockWeights map by computing /// the weights of every basic block in the CFG. /// /// \param F The function to query. bool SampleProfileLoader::computeBlockWeights(Function &F) { bool Changed = false; DEBUG(dbgs() << "Block weights\n"); for (const auto &BB : F) { ErrorOr Weight = getBlockWeight(&BB); if (Weight) { BlockWeights[&BB] = Weight.get(); VisitedBlocks.insert(&BB); Changed = true; } DEBUG(printBlockWeight(dbgs(), &BB)); } return Changed; } /// \brief Get the FunctionSamples for a call instruction. /// /// The FunctionSamples of a call/invoke instruction \p Inst is the inlined /// instance in which that call instruction is calling to. It contains /// all samples that resides in the inlined instance. We first find the /// inlined instance in which the call instruction is from, then we /// traverse its children to find the callsite with the matching /// location. /// /// \param Inst Call/Invoke instruction to query. /// /// \returns The FunctionSamples pointer to the inlined instance. const FunctionSamples * SampleProfileLoader::findCalleeFunctionSamples(const Instruction &Inst) const { const DILocation *DIL = Inst.getDebugLoc(); if (!DIL) { return nullptr; } DISubprogram *SP = DIL->getScope()->getSubprogram(); if (!SP) return nullptr; const FunctionSamples *FS = findFunctionSamples(Inst); if (FS == nullptr) return nullptr; return FS->findFunctionSamplesAt(LineLocation( getOffset(DIL->getLine(), SP->getLine()), DIL->getDiscriminator())); } /// \brief Get the FunctionSamples for an instruction. /// /// The FunctionSamples of an instruction \p Inst is the inlined instance /// in which that instruction is coming from. We traverse the inline stack /// of that instruction, and match it with the tree nodes in the profile. /// /// \param Inst Instruction to query. /// /// \returns the FunctionSamples pointer to the inlined instance. const FunctionSamples * SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const { SmallVector S; const DILocation *DIL = Inst.getDebugLoc(); if (!DIL) { return Samples; } for (DIL = DIL->getInlinedAt(); DIL; DIL = DIL->getInlinedAt()) { DISubprogram *SP = DIL->getScope()->getSubprogram(); if (!SP) return nullptr; S.push_back(LineLocation(getOffset(DIL->getLine(), SP->getLine()), DIL->getDiscriminator())); } if (S.size() == 0) return Samples; const FunctionSamples *FS = Samples; for (int i = S.size() - 1; i >= 0 && FS != nullptr; i--) { FS = FS->findFunctionSamplesAt(S[i]); } return FS; } /// \brief Iteratively inline hot callsites of a function. /// /// Iteratively traverse all callsites of the function \p F, and find if /// the corresponding inlined instance exists and is hot in profile. If /// it is hot enough, inline the callsites and adds new callsites of the /// callee into the caller. /// /// TODO: investigate the possibility of not invoking InlineFunction directly. /// /// \param F function to perform iterative inlining. /// /// \returns True if there is any inline happened. bool SampleProfileLoader::inlineHotFunctions(Function &F) { bool Changed = false; LLVMContext &Ctx = F.getContext(); std::function GetAssumptionCache = [&]( Function &F) -> AssumptionCache & { return ACT->getAssumptionCache(F); }; while (true) { bool LocalChanged = false; SmallVector CIS; for (auto &BB : F) { bool Hot = false; SmallVector Candidates; for (auto &I : BB.getInstList()) { const FunctionSamples *FS = nullptr; if ((isa(I) || isa(I)) && (FS = findCalleeFunctionSamples(I))) { Candidates.push_back(&I); if (callsiteIsHot(Samples, FS)) Hot = true; } } if (Hot) { CIS.insert(CIS.begin(), Candidates.begin(), Candidates.end()); } } for (auto I : CIS) { InlineFunctionInfo IFI(nullptr, ACT ? &GetAssumptionCache : nullptr); CallSite CS(I); Function *CalledFunction = CS.getCalledFunction(); if (!CalledFunction || !CalledFunction->getSubprogram()) continue; DebugLoc DLoc = I->getDebugLoc(); uint64_t NumSamples = findCalleeFunctionSamples(*I)->getTotalSamples(); if (InlineFunction(CS, IFI)) { LocalChanged = true; emitOptimizationRemark(Ctx, DEBUG_TYPE, F, DLoc, Twine("inlined hot callee '") + CalledFunction->getName() + "' with " + Twine(NumSamples) + " samples into '" + F.getName() + "'"); } } if (LocalChanged) { Changed = true; } else { break; } } return Changed; } /// \brief Find equivalence classes for the given block. /// /// This finds all the blocks that are guaranteed to execute the same /// number of times as \p BB1. To do this, it traverses all the /// descendants of \p BB1 in the dominator or post-dominator tree. /// /// A block BB2 will be in the same equivalence class as \p BB1 if /// the following holds: /// /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2 /// is a descendant of \p BB1 in the dominator tree, then BB2 should /// dominate BB1 in the post-dominator tree. /// /// 2- Both BB2 and \p BB1 must be in the same loop. /// /// For every block BB2 that meets those two requirements, we set BB2's /// equivalence class to \p BB1. /// /// \param BB1 Block to check. /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree. /// \param DomTree Opposite dominator tree. If \p Descendants is filled /// with blocks from \p BB1's dominator tree, then /// this is the post-dominator tree, and vice versa. void SampleProfileLoader::findEquivalencesFor( BasicBlock *BB1, ArrayRef Descendants, DominatorTreeBase *DomTree) { const BasicBlock *EC = EquivalenceClass[BB1]; uint64_t Weight = BlockWeights[EC]; for (const auto *BB2 : Descendants) { bool IsDomParent = DomTree->dominates(BB2, BB1); bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2); if (BB1 != BB2 && IsDomParent && IsInSameLoop) { EquivalenceClass[BB2] = EC; // If BB2 is visited, then the entire EC should be marked as visited. if (VisitedBlocks.count(BB2)) { VisitedBlocks.insert(EC); } // If BB2 is heavier than BB1, make BB2 have the same weight // as BB1. // // Note that we don't worry about the opposite situation here // (when BB2 is lighter than BB1). We will deal with this // during the propagation phase. Right now, we just want to // make sure that BB1 has the largest weight of all the // members of its equivalence set. Weight = std::max(Weight, BlockWeights[BB2]); } } if (EC == &EC->getParent()->getEntryBlock()) { BlockWeights[EC] = Samples->getHeadSamples() + 1; } else { BlockWeights[EC] = Weight; } } /// \brief Find equivalence classes. /// /// Since samples may be missing from blocks, we can fill in the gaps by setting /// the weights of all the blocks in the same equivalence class to the same /// weight. To compute the concept of equivalence, we use dominance and loop /// information. Two blocks B1 and B2 are in the same equivalence class if B1 /// dominates B2, B2 post-dominates B1 and both are in the same loop. /// /// \param F The function to query. void SampleProfileLoader::findEquivalenceClasses(Function &F) { SmallVector DominatedBBs; DEBUG(dbgs() << "\nBlock equivalence classes\n"); // Find equivalence sets based on dominance and post-dominance information. for (auto &BB : F) { BasicBlock *BB1 = &BB; // Compute BB1's equivalence class once. if (EquivalenceClass.count(BB1)) { DEBUG(printBlockEquivalence(dbgs(), BB1)); continue; } // By default, blocks are in their own equivalence class. EquivalenceClass[BB1] = BB1; // Traverse all the blocks dominated by BB1. We are looking for // every basic block BB2 such that: // // 1- BB1 dominates BB2. // 2- BB2 post-dominates BB1. // 3- BB1 and BB2 are in the same loop nest. // // If all those conditions hold, it means that BB2 is executed // as many times as BB1, so they are placed in the same equivalence // class by making BB2's equivalence class be BB1. DominatedBBs.clear(); DT->getDescendants(BB1, DominatedBBs); findEquivalencesFor(BB1, DominatedBBs, PDT.get()); DEBUG(printBlockEquivalence(dbgs(), BB1)); } // Assign weights to equivalence classes. // // All the basic blocks in the same equivalence class will execute // the same number of times. Since we know that the head block in // each equivalence class has the largest weight, assign that weight // to all the blocks in that equivalence class. DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n"); for (auto &BI : F) { const BasicBlock *BB = &BI; const BasicBlock *EquivBB = EquivalenceClass[BB]; if (BB != EquivBB) BlockWeights[BB] = BlockWeights[EquivBB]; DEBUG(printBlockWeight(dbgs(), BB)); } } /// \brief Visit the given edge to decide if it has a valid weight. /// /// If \p E has not been visited before, we copy to \p UnknownEdge /// and increment the count of unknown edges. /// /// \param E Edge to visit. /// \param NumUnknownEdges Current number of unknown edges. /// \param UnknownEdge Set if E has not been visited before. /// /// \returns E's weight, if known. Otherwise, return 0. uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge) { if (!VisitedEdges.count(E)) { (*NumUnknownEdges)++; *UnknownEdge = E; return 0; } return EdgeWeights[E]; } /// \brief Propagate weights through incoming/outgoing edges. /// /// If the weight of a basic block is known, and there is only one edge /// with an unknown weight, we can calculate the weight of that edge. /// /// Similarly, if all the edges have a known count, we can calculate the /// count of the basic block, if needed. /// /// \param F Function to process. /// \param UpdateBlockCount Whether we should update basic block counts that /// has already been annotated. /// /// \returns True if new weights were assigned to edges or blocks. bool SampleProfileLoader::propagateThroughEdges(Function &F, bool UpdateBlockCount) { bool Changed = false; DEBUG(dbgs() << "\nPropagation through edges\n"); for (const auto &BI : F) { const BasicBlock *BB = &BI; const BasicBlock *EC = EquivalenceClass[BB]; // Visit all the predecessor and successor edges to determine // which ones have a weight assigned already. Note that it doesn't // matter that we only keep track of a single unknown edge. The // only case we are interested in handling is when only a single // edge is unknown (see setEdgeOrBlockWeight). for (unsigned i = 0; i < 2; i++) { uint64_t TotalWeight = 0; unsigned NumUnknownEdges = 0, NumTotalEdges = 0; Edge UnknownEdge, SelfReferentialEdge, SingleEdge; if (i == 0) { // First, visit all predecessor edges. NumTotalEdges = Predecessors[BB].size(); for (auto *Pred : Predecessors[BB]) { Edge E = std::make_pair(Pred, BB); TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge); if (E.first == E.second) SelfReferentialEdge = E; } if (NumTotalEdges == 1) { SingleEdge = std::make_pair(Predecessors[BB][0], BB); } } else { // On the second round, visit all successor edges. NumTotalEdges = Successors[BB].size(); for (auto *Succ : Successors[BB]) { Edge E = std::make_pair(BB, Succ); TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge); } if (NumTotalEdges == 1) { SingleEdge = std::make_pair(BB, Successors[BB][0]); } } // After visiting all the edges, there are three cases that we // can handle immediately: // // - All the edge weights are known (i.e., NumUnknownEdges == 0). // In this case, we simply check that the sum of all the edges // is the same as BB's weight. If not, we change BB's weight // to match. Additionally, if BB had not been visited before, // we mark it visited. // // - Only one edge is unknown and BB has already been visited. // In this case, we can compute the weight of the edge by // subtracting the total block weight from all the known // edge weights. If the edges weight more than BB, then the // edge of the last remaining edge is set to zero. // // - There exists a self-referential edge and the weight of BB is // known. In this case, this edge can be based on BB's weight. // We add up all the other known edges and set the weight on // the self-referential edge as we did in the previous case. // // In any other case, we must continue iterating. Eventually, // all edges will get a weight, or iteration will stop when // it reaches SampleProfileMaxPropagateIterations. if (NumUnknownEdges <= 1) { uint64_t &BBWeight = BlockWeights[EC]; if (NumUnknownEdges == 0) { if (!VisitedBlocks.count(EC)) { // If we already know the weight of all edges, the weight of the // basic block can be computed. It should be no larger than the sum // of all edge weights. if (TotalWeight > BBWeight) { BBWeight = TotalWeight; Changed = true; DEBUG(dbgs() << "All edge weights for " << BB->getName() << " known. Set weight for block: "; printBlockWeight(dbgs(), BB);); } } else if (NumTotalEdges == 1 && EdgeWeights[SingleEdge] < BlockWeights[EC]) { // If there is only one edge for the visited basic block, use the // block weight to adjust edge weight if edge weight is smaller. EdgeWeights[SingleEdge] = BlockWeights[EC]; Changed = true; } } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) { // If there is a single unknown edge and the block has been // visited, then we can compute E's weight. if (BBWeight >= TotalWeight) EdgeWeights[UnknownEdge] = BBWeight - TotalWeight; else EdgeWeights[UnknownEdge] = 0; const BasicBlock *OtherEC; if (i == 0) OtherEC = EquivalenceClass[UnknownEdge.first]; else OtherEC = EquivalenceClass[UnknownEdge.second]; // Edge weights should never exceed the BB weights it connects. if (VisitedBlocks.count(OtherEC) && EdgeWeights[UnknownEdge] > BlockWeights[OtherEC]) EdgeWeights[UnknownEdge] = BlockWeights[OtherEC]; VisitedEdges.insert(UnknownEdge); Changed = true; DEBUG(dbgs() << "Set weight for edge: "; printEdgeWeight(dbgs(), UnknownEdge)); } } else if (VisitedBlocks.count(EC) && BlockWeights[EC] == 0) { // If a block Weights 0, all its in/out edges should weight 0. if (i == 0) { for (auto *Pred : Predecessors[BB]) { Edge E = std::make_pair(Pred, BB); EdgeWeights[E] = 0; VisitedEdges.insert(E); } } else { for (auto *Succ : Successors[BB]) { Edge E = std::make_pair(BB, Succ); EdgeWeights[E] = 0; VisitedEdges.insert(E); } } } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) { uint64_t &BBWeight = BlockWeights[BB]; // We have a self-referential edge and the weight of BB is known. if (BBWeight >= TotalWeight) EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight; else EdgeWeights[SelfReferentialEdge] = 0; VisitedEdges.insert(SelfReferentialEdge); Changed = true; DEBUG(dbgs() << "Set self-referential edge weight to: "; printEdgeWeight(dbgs(), SelfReferentialEdge)); } if (UpdateBlockCount && !VisitedBlocks.count(EC) && TotalWeight > 0) { BlockWeights[EC] = TotalWeight; VisitedBlocks.insert(EC); Changed = true; } } } return Changed; } /// \brief Build in/out edge lists for each basic block in the CFG. /// /// We are interested in unique edges. If a block B1 has multiple /// edges to another block B2, we only add a single B1->B2 edge. void SampleProfileLoader::buildEdges(Function &F) { for (auto &BI : F) { BasicBlock *B1 = &BI; // Add predecessors for B1. SmallPtrSet Visited; if (!Predecessors[B1].empty()) llvm_unreachable("Found a stale predecessors list in a basic block."); for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) { BasicBlock *B2 = *PI; if (Visited.insert(B2).second) Predecessors[B1].push_back(B2); } // Add successors for B1. Visited.clear(); if (!Successors[B1].empty()) llvm_unreachable("Found a stale successors list in a basic block."); for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) { BasicBlock *B2 = *SI; if (Visited.insert(B2).second) Successors[B1].push_back(B2); } } } /// \brief Propagate weights into edges /// /// The following rules are applied to every block BB in the CFG: /// /// - If BB has a single predecessor/successor, then the weight /// of that edge is the weight of the block. /// /// - If all incoming or outgoing edges are known except one, and the /// weight of the block is already known, the weight of the unknown /// edge will be the weight of the block minus the sum of all the known /// edges. If the sum of all the known edges is larger than BB's weight, /// we set the unknown edge weight to zero. /// /// - If there is a self-referential edge, and the weight of the block is /// known, the weight for that edge is set to the weight of the block /// minus the weight of the other incoming edges to that block (if /// known). void SampleProfileLoader::propagateWeights(Function &F) { bool Changed = true; unsigned I = 0; // Add an entry count to the function using the samples gathered // at the function entry. F.setEntryCount(Samples->getHeadSamples() + 1); // If BB weight is larger than its corresponding loop's header BB weight, // use the BB weight to replace the loop header BB weight. for (auto &BI : F) { BasicBlock *BB = &BI; Loop *L = LI->getLoopFor(BB); if (!L) { continue; } BasicBlock *Header = L->getHeader(); if (Header && BlockWeights[BB] > BlockWeights[Header]) { BlockWeights[Header] = BlockWeights[BB]; } } // Before propagation starts, build, for each block, a list of // unique predecessors and successors. This is necessary to handle // identical edges in multiway branches. Since we visit all blocks and all // edges of the CFG, it is cleaner to build these lists once at the start // of the pass. buildEdges(F); // Propagate until we converge or we go past the iteration limit. while (Changed && I++ < SampleProfileMaxPropagateIterations) { Changed = propagateThroughEdges(F, false); } // The first propagation propagates BB counts from annotated BBs to unknown // BBs. The 2nd propagation pass resets edges weights, and use all BB weights // to propagate edge weights. VisitedEdges.clear(); Changed = true; while (Changed && I++ < SampleProfileMaxPropagateIterations) { Changed = propagateThroughEdges(F, false); } // The 3rd propagation pass allows adjust annotated BB weights that are // obviously wrong. Changed = true; while (Changed && I++ < SampleProfileMaxPropagateIterations) { Changed = propagateThroughEdges(F, true); } // Generate MD_prof metadata for every branch instruction using the // edge weights computed during propagation. DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n"); LLVMContext &Ctx = F.getContext(); MDBuilder MDB(Ctx); for (auto &BI : F) { BasicBlock *BB = &BI; if (BlockWeights[BB]) { for (auto &I : BB->getInstList()) { if (CallInst *CI = dyn_cast(&I)) { if (!dyn_cast(&I)) { SmallVector Weights; Weights.push_back(BlockWeights[BB]); CI->setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights)); } } } } TerminatorInst *TI = BB->getTerminator(); if (TI->getNumSuccessors() == 1) continue; if (!isa(TI) && !isa(TI)) continue; DEBUG(dbgs() << "\nGetting weights for branch at line " << TI->getDebugLoc().getLine() << ".\n"); SmallVector Weights; uint32_t MaxWeight = 0; DebugLoc MaxDestLoc; for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) { BasicBlock *Succ = TI->getSuccessor(I); Edge E = std::make_pair(BB, Succ); uint64_t Weight = EdgeWeights[E]; DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E)); // Use uint32_t saturated arithmetic to adjust the incoming weights, // if needed. Sample counts in profiles are 64-bit unsigned values, // but internally branch weights are expressed as 32-bit values. if (Weight > std::numeric_limits::max()) { DEBUG(dbgs() << " (saturated due to uint32_t overflow)"); Weight = std::numeric_limits::max(); } // Weight is added by one to avoid propagation errors introduced by // 0 weights. Weights.push_back(static_cast(Weight + 1)); if (Weight != 0) { if (Weight > MaxWeight) { MaxWeight = Weight; MaxDestLoc = Succ->getFirstNonPHIOrDbgOrLifetime()->getDebugLoc(); } } } // Only set weights if there is at least one non-zero weight. // In any other case, let the analyzer set weights. if (MaxWeight > 0) { DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n"); TI->setMetadata(llvm::LLVMContext::MD_prof, MDB.createBranchWeights(Weights)); DebugLoc BranchLoc = TI->getDebugLoc(); emitOptimizationRemark( Ctx, DEBUG_TYPE, F, MaxDestLoc, Twine("most popular destination for conditional branches at ") + ((BranchLoc) ? Twine(BranchLoc->getFilename() + ":" + Twine(BranchLoc.getLine()) + ":" + Twine(BranchLoc.getCol())) : Twine(""))); } else { DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n"); } } } /// \brief Get the line number for the function header. /// /// This looks up function \p F in the current compilation unit and /// retrieves the line number where the function is defined. This is /// line 0 for all the samples read from the profile file. Every line /// number is relative to this line. /// /// \param F Function object to query. /// /// \returns the line number where \p F is defined. If it returns 0, /// it means that there is no debug information available for \p F. unsigned SampleProfileLoader::getFunctionLoc(Function &F) { if (DISubprogram *S = F.getSubprogram()) return S->getLine(); // If the start of \p F is missing, emit a diagnostic to inform the user // about the missed opportunity. F.getContext().diagnose(DiagnosticInfoSampleProfile( "No debug information found in function " + F.getName() + ": Function profile not used", DS_Warning)); return 0; } void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) { DT.reset(new DominatorTree); DT->recalculate(F); PDT.reset(new DominatorTreeBase(true)); PDT->recalculate(F); LI.reset(new LoopInfo); LI->analyze(*DT); } /// \brief Generate branch weight metadata for all branches in \p F. /// /// Branch weights are computed out of instruction samples using a /// propagation heuristic. Propagation proceeds in 3 phases: /// /// 1- Assignment of block weights. All the basic blocks in the function /// are initial assigned the same weight as their most frequently /// executed instruction. /// /// 2- Creation of equivalence classes. Since samples may be missing from /// blocks, we can fill in the gaps by setting the weights of all the /// blocks in the same equivalence class to the same weight. To compute /// the concept of equivalence, we use dominance and loop information. /// Two blocks B1 and B2 are in the same equivalence class if B1 /// dominates B2, B2 post-dominates B1 and both are in the same loop. /// /// 3- Propagation of block weights into edges. This uses a simple /// propagation heuristic. The following rules are applied to every /// block BB in the CFG: /// /// - If BB has a single predecessor/successor, then the weight /// of that edge is the weight of the block. /// /// - If all the edges are known except one, and the weight of the /// block is already known, the weight of the unknown edge will /// be the weight of the block minus the sum of all the known /// edges. If the sum of all the known edges is larger than BB's weight, /// we set the unknown edge weight to zero. /// /// - If there is a self-referential edge, and the weight of the block is /// known, the weight for that edge is set to the weight of the block /// minus the weight of the other incoming edges to that block (if /// known). /// /// Since this propagation is not guaranteed to finalize for every CFG, we /// only allow it to proceed for a limited number of iterations (controlled /// by -sample-profile-max-propagate-iterations). /// /// FIXME: Try to replace this propagation heuristic with a scheme /// that is guaranteed to finalize. A work-list approach similar to /// the standard value propagation algorithm used by SSA-CCP might /// work here. /// /// Once all the branch weights are computed, we emit the MD_prof /// metadata on BB using the computed values for each of its branches. /// /// \param F The function to query. /// /// \returns true if \p F was modified. Returns false, otherwise. bool SampleProfileLoader::emitAnnotations(Function &F) { bool Changed = false; if (getFunctionLoc(F) == 0) return false; DEBUG(dbgs() << "Line number for the first instruction in " << F.getName() << ": " << getFunctionLoc(F) << "\n"); Changed |= inlineHotFunctions(F); // Compute basic block weights. Changed |= computeBlockWeights(F); if (Changed) { // Compute dominance and loop info needed for propagation. computeDominanceAndLoopInfo(F); // Find equivalence classes. findEquivalenceClasses(F); // Propagate weights to all edges. propagateWeights(F); } // If coverage checking was requested, compute it now. if (SampleProfileRecordCoverage) { unsigned Used = CoverageTracker.countUsedRecords(Samples); unsigned Total = CoverageTracker.countBodyRecords(Samples); unsigned Coverage = CoverageTracker.computeCoverage(Used, Total); if (Coverage < SampleProfileRecordCoverage) { F.getContext().diagnose(DiagnosticInfoSampleProfile( F.getSubprogram()->getFilename(), getFunctionLoc(F), Twine(Used) + " of " + Twine(Total) + " available profile records (" + Twine(Coverage) + "%) were applied", DS_Warning)); } } if (SampleProfileSampleCoverage) { uint64_t Used = CoverageTracker.getTotalUsedSamples(); uint64_t Total = CoverageTracker.countBodySamples(Samples); unsigned Coverage = CoverageTracker.computeCoverage(Used, Total); if (Coverage < SampleProfileSampleCoverage) { F.getContext().diagnose(DiagnosticInfoSampleProfile( F.getSubprogram()->getFilename(), getFunctionLoc(F), Twine(Used) + " of " + Twine(Total) + " available profile samples (" + Twine(Coverage) + "%) were applied", DS_Warning)); } } return Changed; } char SampleProfileLoaderLegacyPass::ID = 0; INITIALIZE_PASS_BEGIN(SampleProfileLoaderLegacyPass, "sample-profile", "Sample Profile loader", false, false) INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) INITIALIZE_PASS_END(SampleProfileLoaderLegacyPass, "sample-profile", "Sample Profile loader", false, false) bool SampleProfileLoader::doInitialization(Module &M) { auto &Ctx = M.getContext(); auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx); if (std::error_code EC = ReaderOrErr.getError()) { std::string Msg = "Could not open profile: " + EC.message(); Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg)); return false; } Reader = std::move(ReaderOrErr.get()); ProfileIsValid = (Reader->read() == sampleprof_error::success); return true; } ModulePass *llvm::createSampleProfileLoaderPass() { return new SampleProfileLoaderLegacyPass(SampleProfileFile); } ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) { return new SampleProfileLoaderLegacyPass(Name); } bool SampleProfileLoader::runOnModule(Module &M) { if (!ProfileIsValid) return false; // Compute the total number of samples collected in this profile. for (const auto &I : Reader->getProfiles()) TotalCollectedSamples += I.second.getTotalSamples(); bool retval = false; for (auto &F : M) if (!F.isDeclaration()) { clearFunctionData(); retval |= runOnFunction(F); } if (M.getProfileSummary() == nullptr) M.setProfileSummary(Reader->getSummary().getMD(M.getContext())); return retval; } bool SampleProfileLoaderLegacyPass::runOnModule(Module &M) { // FIXME: pass in AssumptionCache correctly for the new pass manager. SampleLoader.setACT(&getAnalysis()); return SampleLoader.runOnModule(M); } bool SampleProfileLoader::runOnFunction(Function &F) { F.setEntryCount(0); Samples = Reader->getSamplesFor(F); if (!Samples->empty()) return emitAnnotations(F); return false; } PreservedAnalyses SampleProfileLoaderPass::run(Module &M, ModuleAnalysisManager &AM) { SampleProfileLoader SampleLoader(SampleProfileFile); SampleLoader.doInitialization(M); if (!SampleLoader.runOnModule(M)) return PreservedAnalyses::all(); return PreservedAnalyses::none(); }