Index: include/llvm/IR/DiagnosticInfo.h =================================================================== --- include/llvm/IR/DiagnosticInfo.h +++ include/llvm/IR/DiagnosticInfo.h @@ -60,6 +60,7 @@ DK_OptimizationRemarkAnalysisAliasing, DK_OptimizationFailure, DK_MIRParser, + DK_PGOProfile, DK_FirstPluginKind }; @@ -250,6 +251,31 @@ const Twine &Msg; }; +/// Diagnostic information for the PGO profiler. +class DiagnosticInfoPGOProfile : public DiagnosticInfo { +public: + DiagnosticInfoPGOProfile(const char *FileName, const Twine &Msg, + DiagnosticSeverity Severity = DS_Error) + : DiagnosticInfo(DK_PGOProfile, Severity), FileName(FileName), Msg(Msg) {} + + /// \see DiagnosticInfo::print. + void print(DiagnosticPrinter &DP) const override; + + static bool classof(const DiagnosticInfo *DI) { + return DI->getKind() == DK_PGOProfile; + } + + const char *getFileName() const { return FileName; } + const Twine &getMsg() const { return Msg; } + +private: + /// Name of the input file associated with this diagnostic. + const char *FileName; + + /// Message to report. + const Twine &Msg; +}; + /// Common features for diagnostics dealing with optimization remarks. class DiagnosticInfoOptimizationBase : public DiagnosticInfo { public: Index: include/llvm/InitializePasses.h =================================================================== --- include/llvm/InitializePasses.h +++ include/llvm/InitializePasses.h @@ -117,6 +117,8 @@ void initializeExpandPostRAPass(PassRegistry&); void initializeAAResultsWrapperPassPass(PassRegistry &); void initializeGCOVProfilerPass(PassRegistry&); +void initializePGOInstrumentationGenPass(PassRegistry&); +void initializePGOInstrumentationUsePass(PassRegistry&); void initializeInstrProfilingPass(PassRegistry&); void initializeAddressSanitizerPass(PassRegistry&); void initializeAddressSanitizerModulePass(PassRegistry&); Index: include/llvm/LinkAllPasses.h =================================================================== --- include/llvm/LinkAllPasses.h +++ include/llvm/LinkAllPasses.h @@ -85,6 +85,8 @@ (void) llvm::createDomOnlyViewerPass(); (void) llvm::createDomViewerPass(); (void) llvm::createGCOVProfilerPass(); + (void) llvm::createPGOInstrumentationGenPass(); + (void) llvm::createPGOInstrumentationUsePass(); (void) llvm::createInstrProfilingPass(); (void) llvm::createFunctionInliningPass(); (void) llvm::createAlwaysInlinerPass(); Index: include/llvm/Transforms/Instrumentation.h =================================================================== --- include/llvm/Transforms/Instrumentation.h +++ include/llvm/Transforms/Instrumentation.h @@ -79,6 +79,11 @@ ModulePass *createGCOVProfilerPass(const GCOVOptions &Options = GCOVOptions::getDefault()); +// PGO Instrumention +ModulePass *createPGOInstrumentationGenPass(); +ModulePass * +createPGOInstrumentationUsePass(StringRef Filename = StringRef("")); + /// Options for the frontend instrumentation based profiling pass. struct InstrProfOptions { InstrProfOptions() : NoRedZone(false) {} @@ -149,6 +154,29 @@ /// protect against stack-based overflow vulnerabilities. FunctionPass *createSafeStackPass(const TargetMachine *TM = nullptr); +/// \brief Calculate what to divide by to scale counts. +/// +/// Given the maximum count, calculate a divisor that will scale all the +/// weights to strictly less than UINT32_MAX. +static inline uint64_t calculateCountScale(uint64_t MaxCount) { + return MaxCount < UINT32_MAX ? 1 : MaxCount / UINT32_MAX + 1; +} + +/// \brief Scale an individual branch count (and add 1). +/// +/// Scale a 64-bit weight down to 32-bits using \c Scale. +/// +/// According to Laplace's Rule of Succession, it is better to compute the +/// count based on the count plus 1, so universally add 1 to the value. +/// +/// \pre \c Scale was calculated by \a calculateCountScale() with a count no +/// greater than \c Count. +static inline uint32_t scaleBranchCount(uint64_t Count, uint64_t Scale) { + uint64_t Scaled = Count / Scale + 1; + assert(Scaled <= UINT32_MAX && "overflow 32-bits"); + return Scaled; +} + } // End llvm namespace #endif Index: lib/IR/DiagnosticInfo.cpp =================================================================== --- lib/IR/DiagnosticInfo.cpp +++ lib/IR/DiagnosticInfo.cpp @@ -132,6 +132,12 @@ DP << getMsg(); } +void DiagnosticInfoPGOProfile::print(DiagnosticPrinter &DP) const { + if (getFileName()) + DP << getFileName() << ": "; + DP << getMsg(); +} + bool DiagnosticInfoOptimizationBase::isLocationAvailable() const { return getDebugLoc(); } Index: lib/Transforms/IPO/LLVMBuild.txt =================================================================== --- lib/Transforms/IPO/LLVMBuild.txt +++ lib/Transforms/IPO/LLVMBuild.txt @@ -20,4 +20,4 @@ name = IPO parent = Transforms library_name = ipo -required_libraries = Analysis Core InstCombine ProfileData Scalar Support TransformUtils Vectorize +required_libraries = Analysis Core InstCombine ProfileData Scalar Support TransformUtils Vectorize Instrumentation Index: lib/Transforms/Instrumentation/CFGMST.h =================================================================== --- /dev/null +++ lib/Transforms/Instrumentation/CFGMST.h @@ -0,0 +1,210 @@ +//===-- CFGMST.h - Minimum Spanning Tree for CFG -------*- C++ -*-===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This file implements Union-find algorithm to compute Minimum Spanning Tree +// for a given CFG. +// +//===----------------------------------------------------------------------===// + +#include "llvm/ADT/DenseMap.h" +#include "llvm/Support/Debug.h" +#include "llvm/Support/raw_ostream.h" +#include "llvm/Support/BranchProbability.h" +#include "llvm/Transforms/Utils/BasicBlockUtils.h" +#include "llvm/Analysis/CFG.h" +#include "llvm/Analysis/BranchProbabilityInfo.h" +#include "llvm/Analysis/BlockFrequencyInfo.h" +#include +#include +#include + +namespace llvm { + +#define DEBUG_TYPE "cfgmst" + +template class CFGMST { +public: + Function &F; + + // Store all the edges in CFG. It may contain some stale edges + // when Removed is set. + std::vector> AllEdges; + + // This map records the auxiliary information for each BB. + DenseMap> BBInfos; + + // Find the root group of the G and compress the path from G to the root. + BBInfo *findAndCompressGroup(BBInfo *G) { + if (G->Group != G) + G->Group = findAndCompressGroup(static_cast(G->Group)); + return static_cast(G->Group); + } + + // Union BB1 and BB2 into the same group and return true. + // Returns false if BB1 and BB2 are already in the same group. + bool unionGroups(const BasicBlock *BB1, const BasicBlock *BB2) { + BBInfo *BB1G = findAndCompressGroup(&getBBInfo(BB1)); + BBInfo *BB2G = findAndCompressGroup(&getBBInfo(BB2)); + + if (BB1G == BB2G) + return false; + + // Make the smaller rank tree a direct child or the root of high rank tree. + if (BB1G->Rank < BB2G->Rank) + BB1G->Group = BB2G; + else { + BB2G->Group = BB1G; + // If the ranks are the same, increment root of one tree by one. + if (BB1G->Rank == BB2G->Rank) + BB1G->Rank++; + } + return true; + } + + // Give BB, return the auxiliary information. + BBInfo &getBBInfo(const BasicBlock *BB) const { + auto It = BBInfos.find(BB); + assert(It->second.get() != nullptr); + return *It->second.get(); + } + + // Traverse the CFG using a stack. Find all the edges and assign the weight. + // Edges with large weight will be put into MST first so they are less likely + // to be instrumented. + void buildEdges() { + DEBUG(dbgs() << "Build Edge on " << F.getName() << "\n"); + + const BasicBlock *BB = &(F.getEntryBlock()); + // Add a fake edge to the entry. + addEdge(nullptr, BB, BFI->getEntryFreq()); + + // Special handling for single BB functions. + if (succ_empty(BB)) { + addEdge(BB, nullptr, BFI->getEntryFreq()); + return; + } + +#define CRITICAL_EDGE_MULTIPLIER 1000 + + for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { + TerminatorInst *TI = BB->getTerminator(); + uint64_t BBWeight = BFI->getBlockFreq(&*BB).getFrequency(); + uint64_t Weight; + if (int successors = TI->getNumSuccessors()) { + for (uint32_t i = 0; i != successors; ++i) { + BasicBlock *TargetBB = TI->getSuccessor(i); + bool Critical = isCriticalEdge(TI, i); + uint64_t scaleFactor = BBWeight; + if (Critical) { + if (scaleFactor < UINT64_MAX / CRITICAL_EDGE_MULTIPLIER) + scaleFactor *= CRITICAL_EDGE_MULTIPLIER; + else + scaleFactor = UINT64_MAX; + } + Weight = BPI->getEdgeProbability(&*BB, TargetBB).scale(scaleFactor); + addEdge(&*BB, TargetBB, Weight).IsCritical = Critical; + DEBUG(dbgs() << " Edge: from " << BB->getName() << " to " + << TargetBB->getName() << " w=" << Weight << "\n"); + } + } else { + addEdge(&*BB, nullptr, BBWeight); + DEBUG(dbgs() << " Edge: from " << BB->getName() << " to exit" + << " w = " << BBWeight << "\n"); + } + } + } + + // Sort CFG edges based on its weight. + void sortEdgesByWeight() { + std::sort( + AllEdges.begin(), AllEdges.end(), + [](const std::unique_ptr &lhs, const std::unique_ptr &rhs) { + return lhs->Weight > rhs->Weight; + }); + } + + // Traverse all the edges and compute the Minimum Weight Spanning Tree + // using union-find algorithm. + void computeMinimumSpanningTree() { + // First, put all the critical edge with landing-pad as the Dest to MST. + // This works around the insufficient support of critical edges split + // when destination BB is a landing pad. + for (auto &Ei : AllEdges) { + if (Ei->Removed) + continue; + if (Ei->IsCritical) { + if (Ei->DestBB && Ei->DestBB->isLandingPad()) { + if (unionGroups(Ei->SrcBB, Ei->DestBB)) + Ei->InMST = true; + } + } + } + + for (auto &Ei : AllEdges) { + if (Ei->Removed) + continue; + if (unionGroups(Ei->SrcBB, Ei->DestBB)) + Ei->InMST = true; + } + } + + // Dump the Debug information about the instrumentation. + void dumpEdges(raw_ostream &OS, const StringRef Message = StringRef()) const { + if (!Message.empty()) + OS << Message << "\n"; + OS << " Number of Basic Blocks: " << BBInfos.size() << "\n"; + for (auto &BI : BBInfos) { + const BasicBlock *BB = BI.first; + OS << " BB: " << (BB == nullptr ? "FakeNode" : BB->getName()) << " " + << BI.second->infoString() << "\n"; + } + + OS << " Number of Edges: " << AllEdges.size() + << " (*: Instrument, C: CriticalEdge, -: Removed)\n"; + uint32_t Count = 0; + for (auto &EI : AllEdges) { + OS << " Edge " << Count++ << ": " << getBBInfo(EI->SrcBB).Index << "-->" + << getBBInfo(EI->DestBB).Index << EI->infoString() << "\n"; + }; + } + + // Add an edge to AllEdges with weight W. + Edge &addEdge(const BasicBlock *Src, const BasicBlock *Dest, uint64_t W) { + uint32_t Index = BBInfos.size(); + auto Iter = BBInfos.end(); + bool Inserted; + std::tie(Iter, Inserted) = BBInfos.insert(std::make_pair(Src, nullptr)); + if (Inserted) { + // Newly inserted, update the real info. + Iter->second = std::move(std::unique_ptr(new BBInfo(Index))); + Index++; + } + std::tie(Iter, Inserted) = BBInfos.insert(std::make_pair(Dest, nullptr)); + if (Inserted) + // Newly inserted, update the real info. + Iter->second = std::move(std::unique_ptr(new BBInfo(Index))); + AllEdges.emplace_back(new Edge(Src, Dest, W)); + return *(AllEdges.back().get()); + } + + BranchProbabilityInfo *BPI; + BlockFrequencyInfo *BFI; + +public: + CFGMST(Function &Func, BranchProbabilityInfo *BPI_ = nullptr, + BlockFrequencyInfo *BFI_ = nullptr) + : F(Func), BPI(BPI_), BFI(BFI_) { + buildEdges(); + sortEdgesByWeight(); + computeMinimumSpanningTree(); + } +}; + +#undef DEBUG_TYPE // "cfgmst" +} // end namespace llvm Index: lib/Transforms/Instrumentation/CMakeLists.txt =================================================================== --- lib/Transforms/Instrumentation/CMakeLists.txt +++ lib/Transforms/Instrumentation/CMakeLists.txt @@ -6,6 +6,7 @@ MemorySanitizer.cpp Instrumentation.cpp InstrProfiling.cpp + PGOInstrumentation.cpp SafeStack.cpp SanitizerCoverage.cpp ThreadSanitizer.cpp Index: lib/Transforms/Instrumentation/Instrumentation.cpp =================================================================== --- lib/Transforms/Instrumentation/Instrumentation.cpp +++ lib/Transforms/Instrumentation/Instrumentation.cpp @@ -60,6 +60,8 @@ initializeAddressSanitizerModulePass(Registry); initializeBoundsCheckingPass(Registry); initializeGCOVProfilerPass(Registry); + initializePGOInstrumentationGenPass(Registry); + initializePGOInstrumentationUsePass(Registry); initializeInstrProfilingPass(Registry); initializeMemorySanitizerPass(Registry); initializeThreadSanitizerPass(Registry); Index: lib/Transforms/Instrumentation/LLVMBuild.txt =================================================================== --- lib/Transforms/Instrumentation/LLVMBuild.txt +++ lib/Transforms/Instrumentation/LLVMBuild.txt @@ -19,4 +19,4 @@ type = Library name = Instrumentation parent = Transforms -required_libraries = Analysis Core MC Support TransformUtils +required_libraries = Analysis Core MC Support TransformUtils ProfileData Index: lib/Transforms/Instrumentation/PGOInstrumentation.cpp =================================================================== --- /dev/null +++ lib/Transforms/Instrumentation/PGOInstrumentation.cpp @@ -0,0 +1,705 @@ +//===- PGOInstru.cpp - PGO Instrumentation --------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This file implements PGO instrumentation. It contains two passes: +// (1) Pass PGOInstrumentationGen which instruments the IR to generate edge +// count profile, and +// (2) Pass PGOInstrumentationUse which reads the edge count profile and +// annotates the branch weight. +// These two passes are mutually exclusive, and they are called at the same +// compilation point (so they see the same IR). For PGOInstrumentationGen, +// the real work is done in class PGOGenFunc (per function) For +// PGOInstrumentationUse, the real work in done in class PGOUseFunc and the +// profile is opened in module level and passed to each PGOUseFunc instance. +// Both PGOGenFunc and PGOUseFunc class has member instance of class +// FuncPGOInstrumentation, which does the shared work for the instrumentation. +// +// Class PGOEdge represents a CFG edge and some auxiliary information. Class +// BBInfo contains auxiliary information for a BB. These two classes are used +// in PGOGenFunc. Class PGOUseEdge and UseBBInfo are the derived class of +// PGOEdge and BBInfo, respectively. They contains extra data structure used +// in populating profile counters. +// +// The main algorithm for the instrumentation is: first compute the minimal +// spanning tree (MST) of the CFG. All edges that not in the MST will be +// instrumented. The MST implementation is in Class CFGMST. +// +//===----------------------------------------------------------------------===// + +#include "llvm/Transforms/Instrumentation.h" +#include "llvm/ADT/Statistic.h" +#include "llvm/ADT/DenseMap.h" +#include "llvm/ADT/STLExtras.h" +#include "llvm/IR/InstIterator.h" +#include "llvm/IR/Instructions.h" +#include "llvm/IR/IntrinsicInst.h" +#include "llvm/IR/IRBuilder.h" +#include "llvm/IR/MDBuilder.h" +#include "llvm/IR/DiagnosticInfo.h" +#include "llvm/Pass.h" +#include "llvm/IR/Module.h" +#include "llvm/Support/Debug.h" +#include "llvm/Support/BranchProbability.h" +#include "llvm/Support/JamCRC.h" +#include "llvm/Transforms/Utils/BasicBlockUtils.h" +#include "llvm/ProfileData/InstrProfReader.h" +#include "llvm/Analysis/CFG.h" +#include "llvm/Analysis/BranchProbabilityInfo.h" +#include "llvm/Analysis/BlockFrequencyInfo.h" +#include +#include +#include +#include "CFGMST.h" + +using namespace llvm; + +#define DEBUG_TYPE "pgo-instr" + +STATISTIC(NumOfPGOInstrument, "Number of edges instrumented."); +STATISTIC(NumOfPGOEdge, "Number of edges."); +STATISTIC(NumOfPGOBB, "Number of basic-blocks."); +STATISTIC(NumOfPGOSplit, "Number of critical edge splits."); +STATISTIC(NumOfPGOFunc, "Number of functions having valid profile counts."); +STATISTIC(NumOfPGOMismatch, "Number of functions having mismatch profile."); +STATISTIC(NumOfPGOMissing, "Number of functions without profile."); + +namespace { +class PGOInstrumentationGen : public ModulePass { +public: + static char ID; + + PGOInstrumentationGen() : ModulePass(ID) { + initializePGOInstrumentationGenPass(*PassRegistry::getPassRegistry()); + } + + const char *getPassName() const override { + return "PGOInstrumentationGenPass"; + } + +private: + bool runOnModule(Module &M) override; + + void getAnalysisUsage(AnalysisUsage &AU) const override { + AU.addRequired(); + AU.addRequired(); + } +}; + +class PGOInstrumentationUse : public ModulePass { +public: + static char ID; + + // Provide the profile filename as the parameter. + PGOInstrumentationUse(StringRef Filename = StringRef("")) + : ModulePass(ID), ProfileFileName(Filename) { + initializePGOInstrumentationUsePass(*PassRegistry::getPassRegistry()); + } + + const char *getPassName() const override { + return "PGOInstrumentationUsePass"; + } + +private: + StringRef ProfileFileName; + std::unique_ptr PGOReader; + bool runOnModule(Module &M) override; + + void getAnalysisUsage(AnalysisUsage &AU) const override { + AU.addRequired(); + AU.addRequired(); + } +}; +} // end anonymous namespace + +char PGOInstrumentationGen::ID = 0; +INITIALIZE_PASS_BEGIN(PGOInstrumentationGen, "pgo-instr-gen", + "PGO instrumentation.", false, false) +INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass) +INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass) +INITIALIZE_PASS_END(PGOInstrumentationGen, "pgo-instr-gen", + "PGO instrumentation.", false, false) + +ModulePass *llvm::createPGOInstrumentationGenPass() { + return new PGOInstrumentationGen(); +} + +char PGOInstrumentationUse::ID = 0; +INITIALIZE_PASS_BEGIN(PGOInstrumentationUse, "pgo-instr-use", + "Read PGO instrumentation profile.", false, false) +INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass) +INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass) +INITIALIZE_PASS_END(PGOInstrumentationUse, "pgo-instr-use", + "Read PGO instrumentation profile.", false, false) + +ModulePass *llvm::createPGOInstrumentationUsePass(StringRef Filename) { + return new PGOInstrumentationUse(Filename); +} + +namespace { +/// \brief An MST based instrumentation for PGO +/// +/// Implements a Minimum Spanning Tree (MST) based instrumentation for PGO +/// in the function level. +// +// This class implements the CFG edges. Note the CFG can be a multi-graph. +struct PGOEdge { + const BasicBlock *SrcBB; + const BasicBlock *DestBB; + uint64_t Weight; + bool InMST; + bool Removed; + bool IsCritical; + PGOEdge(const BasicBlock *Src, const BasicBlock *Dest, unsigned W = 1) + : SrcBB(Src), DestBB(Dest), Weight(W), InMST(false), Removed(false), + IsCritical(false) {} + // Return the information string of an edge. + const std::string infoString() const { + std::string Str = (Removed ? "-" : " "); + Str += (InMST ? " " : "*"); + Str += (IsCritical ? "c" : " "); + Str += " W=" + std::to_string(Weight); + return Str; + } +}; + +// This class stores the auxiliary information for each BB. +struct BBInfo { + BBInfo *Group; + uint32_t Index; + uint32_t Rank; + + BBInfo(unsigned IX) : Group(this), Index(IX), Rank(0) {} + + // Return the information string of this object. + const std::string infoString() const { + return "Index=" + std::to_string(Index); + } +}; + +// This class implements the CFG edges. Note the CFG can be a multi-graph. +template class FuncPGOInstrumentation { +private: + Function &F; + void computeCFGHash(); + +public: + std::string FuncName; + GlobalVariable *FuncNameVar; + // CFG hash value for this function. + uint64_t FunctionHash; + + // The Minimum Spanning Tree of function CFG. + CFGMST MST; + + // Give an edge, find the BB that will be instrumented. + // Return nullptr if there is no BB to be instrumented. + BasicBlock *getInstrBB(Edge *E); + + // Return the auxiliary BB information. + BBInfo &getBBInfo(const BasicBlock *BB) const { return MST.getBBInfo(BB); } + + // Dump edges and BB information. + void dumpInfo(std::string Str = "") const { + std::string Message = "Dump Function " + FuncName + " Hash: " + + std::to_string(FunctionHash) + "\t" + Str; + MST.dumpEdges(dbgs(), Message); + } + + FuncPGOInstrumentation(Function &Func, bool CreateGlobalVar = false, + BranchProbabilityInfo *BPI_ = nullptr, + BlockFrequencyInfo *BFI_ = nullptr) + : F(Func), FunctionHash(0), MST(F, BPI_, BFI_) { + FuncName = getPGOFuncName(F); + computeCFGHash(); + DEBUG(dumpInfo("after CFGMST")); + + NumOfPGOBB += MST.BBInfos.size(); + for (auto &Ei : MST.AllEdges) { + if (Ei->Removed) + continue; + NumOfPGOEdge++; + if (!Ei->InMST) + NumOfPGOInstrument++; + } + + if (CreateGlobalVar) + FuncNameVar = createPGOFuncNameVar(F, FuncName); + }; +}; + +// Compute Hash value for the CFG: the lower 32 bits are CRC32 of the index +// value of each BB in the CFG. The higher 32 bits record the number of edges. +template +void FuncPGOInstrumentation::computeCFGHash() { + std::vector Indexes; + JamCRC JC; + for (auto &BB : F) { + const TerminatorInst *TI = BB.getTerminator(); + for (unsigned s = 0, e = TI->getNumSuccessors(); s != e; ++s) { + BasicBlock *Succ = TI->getSuccessor(s); + uint32_t Index = getBBInfo(Succ).Index; + for (int i = 0; i < sizeof(uint32_t) / sizeof(char); i++) + Indexes.push_back((char)(Index >> (i * sizeof(char)))); + } + } + JC.update(Indexes); + FunctionHash = MST.AllEdges.size() << 32 | JC.getCRC(); +} + +template +BasicBlock *FuncPGOInstrumentation::getInstrBB(Edge *E) { + if (E->InMST || E->Removed) + return nullptr; + + BasicBlock *SrcBB = const_cast(E->SrcBB); + BasicBlock *DestBB = const_cast(E->DestBB); + // For a fake edge, instrument the real BB. + if (SrcBB == nullptr) + return DestBB; + if (DestBB == nullptr) + return SrcBB; + + // Instrument the SrcBB if it has a single successor, + // otherwise, the DestBB if this is not a critical edge. + TerminatorInst *TI = SrcBB->getTerminator(); + if (TI->getNumSuccessors() <= 1) + return SrcBB; + if (!E->IsCritical) + return DestBB; + + // For a critical edge, we have to split. Instrument the newly + // created BB. + NumOfPGOSplit++; + DEBUG(dbgs() << "Split critical edge: " << getBBInfo(SrcBB).Index << " --> " + << getBBInfo(DestBB).Index << "\n"); + unsigned SuccNum = GetSuccessorNumber(SrcBB, DestBB); + BasicBlock *InstrBB = SplitCriticalEdge(TI, SuccNum); + assert(InstrBB && "Critical edge is not split"); + + E->Removed = true; + return InstrBB; +} + +class PGOGenFunc { +private: + Module *M; + // This member stores the shared information with class PGOUseFunc. + FuncPGOInstrumentation FuncInfo; + +public: + PGOGenFunc(Function &Func, Module *Modu, BranchProbabilityInfo *BPI_, + BlockFrequencyInfo *BFI_) + : M(Modu), FuncInfo(Func, true, BPI_, BFI_) {} + + // Add the instrumentation intrinsic calls. + void instrumentCFG(); +}; + +// Visit all edge and instrument the edges not in MST. +// Critical edges will be split. +void PGOGenFunc::instrumentCFG() { + unsigned NumCounters = 0; + for (auto &Ei : FuncInfo.MST.AllEdges) { + if (!Ei->InMST && !Ei->Removed) + NumCounters++; + } + + for (unsigned i = 0, j = 0, E = FuncInfo.MST.AllEdges.size(); i < E; i++) { + PGOEdge *Ei = FuncInfo.MST.AllEdges[i].get(); + BasicBlock *InstrBB = FuncInfo.getInstrBB(Ei); + if (!InstrBB) + continue; + + IRBuilder<> Builder(InstrBB, InstrBB->getFirstInsertionPt()); + assert(Builder.GetInsertPoint() != InstrBB->end() && + "Cannot get the Instrumentation point"); + auto *I8PtrTy = Type::getInt8PtrTy(M->getContext()); + Builder.CreateCall( + Intrinsic::getDeclaration(M, Intrinsic::instrprof_increment), + {llvm::ConstantExpr::getBitCast(FuncInfo.FuncNameVar, I8PtrTy), + Builder.getInt64(FuncInfo.FunctionHash), Builder.getInt32(NumCounters), + Builder.getInt32(j++)}); + } +} + +struct PGOUseEdge : public PGOEdge { + bool CountValid; + uint64_t CountValue; + PGOUseEdge(const BasicBlock *Src, const BasicBlock *Dest, unsigned W = 1) + : PGOEdge(Src, Dest, W), CountValid(false), CountValue(0) {} + + // Set edge count value + void setEdgeCount(uint64_t Value) { + CountValue = Value; + CountValid = true; + } + + // Return the information string for this object. + const std::string infoString() const { + if (!CountValid) + return PGOEdge::infoString(); + return PGOEdge::infoString() + " Count=" + std::to_string(CountValue); + } +}; + +typedef SmallVector DirectEdges; + +// This class stores the auxiliary information for each BB. +struct UseBBInfo : public BBInfo { + uint64_t CountValue; + bool CountValid; + int32_t UnknownCountInEdge; + int32_t UnknownCountOutEdge; + DirectEdges InEdges; + DirectEdges OutEdges; + UseBBInfo(unsigned IX) + : BBInfo(IX), CountValue(0), CountValid(false), UnknownCountInEdge(0), + UnknownCountOutEdge(0) {} + UseBBInfo(unsigned IX, uint64_t C) + : BBInfo(IX), CountValue(C), CountValid(true), UnknownCountInEdge(0), + UnknownCountOutEdge(0) {} + + // Set the profile count value for this BB. + void setBBInfoCount(uint64_t Value) { + CountValue = Value; + CountValid = true; + } + + // Return the information string of this object. + const std::string infoString() const { + if (!CountValid) + return BBInfo::infoString(); + return BBInfo::infoString() + " Count=" + std::to_string(CountValue); + } +}; + +// Sum up the count values for all the edges. +static uint64_t sumEdgeCount(const ArrayRef Edges) { + uint64_t Total = 0; + for (auto &Ei : Edges) { + if (Ei->Removed) + continue; + Total += Ei->CountValue; + } + return Total; +} + +class PGOUseFunc { +private: + Function &F; + Module *M; + // This member stores the shared information with class PGOGenFunc. + FuncPGOInstrumentation FuncInfo; + + // Return the auxiliary BB information. + UseBBInfo &getBBInfo(const BasicBlock *BB) const { + return FuncInfo.getBBInfo(BB); + } + + // The maximum count value in the profile. This is only used in PGO use + // compilation. + uint64_t ProgramMaxCount; + + // Find the Instrumented BB and set the value. + void setInstrumentedCounts(const std::vector *CountFromProfile); + + // Set the edge counter value for the unknown edge -- there should be only + // one unknown edge. + void setEdgeCount(DirectEdges &Edges, uint64_t Value); + + // Return FuncName string; + const std::string getFuncName() const { return FuncInfo.FuncName; } + + // Set the hot/cold inline hints based on the count values. + void applyFunctionAttributes(uint64_t EntryCount, uint64_t MaxCount) { + if (ProgramMaxCount == 0) + return; + // Threshold of the hot functions. + const BranchProbability HotFunctionThreshold(1, 100); + // Threshold of the cold functions. + const BranchProbability ColdFunctionThreshold(2, 10000); + if (EntryCount >= HotFunctionThreshold.scale(ProgramMaxCount)) + F.addFnAttr(llvm::Attribute::InlineHint); + else if (MaxCount <= ColdFunctionThreshold.scale(ProgramMaxCount)) + F.addFnAttr(llvm::Attribute::Cold); + } + +public: + PGOUseFunc(Function &Func, Module *Modu, + BranchProbabilityInfo *BPI_ = nullptr, + BlockFrequencyInfo *BFI_ = nullptr) + : F(Func), M(Modu), FuncInfo(Func, false, BPI_, BFI_) {} + + // Read counts for the instrumented BB from profile. + bool readCounters(IndexedInstrProfReader *PGOReader); + + // Populate the counts for all BBs. + void populateCounters(); + + // Set the branch weights based on the count values. + void setBranchWeights(); +}; + +// Visit all the edges and assign the count value for the instrumented +// edges and the BB. +void PGOUseFunc::setInstrumentedCounts( + const std::vector *CountFromProfile) { + for (unsigned i = 0, j = 0, E = FuncInfo.MST.AllEdges.size(); i < E; i++) { + PGOUseEdge *Ei = FuncInfo.MST.AllEdges[i].get(); + BasicBlock *InstrBB = FuncInfo.getInstrBB(Ei); + if (!InstrBB) + continue; + uint64_t CountValue = (*CountFromProfile)[j++]; + + if (Ei->Removed) { + BasicBlock *SrcBB = const_cast(Ei->SrcBB); + BasicBlock *DestBB = const_cast(Ei->DestBB); + + // Add new edge of SrcBB->InstrBB. + PGOUseEdge &NewEdge = FuncInfo.MST.addEdge(SrcBB, InstrBB, 0); + NewEdge.setEdgeCount(CountValue); + + // Add new edge of InstrBB->DestBB. + PGOUseEdge &NewEdge1 = FuncInfo.MST.addEdge(InstrBB, DestBB, 0); + NewEdge1.setEdgeCount(CountValue); + NewEdge1.InMST = true; + } else + Ei->setEdgeCount(CountValue); + + getBBInfo(InstrBB).setBBInfoCount(CountValue); + } +} + +// Set the count value for the unknown edges. There should be one and only one +// unknown edge in Edges vector. +void PGOUseFunc::setEdgeCount(DirectEdges &Edges, uint64_t Value) { + for (auto &Ei : Edges) { + if (Ei->CountValid) + continue; + Ei->setEdgeCount(Value); + + getBBInfo(Ei->SrcBB).UnknownCountOutEdge--; + getBBInfo(Ei->DestBB).UnknownCountInEdge--; + return; + } + llvm_unreachable("Cannot find the unknown count edge"); +} + +// Read the profile from ProfileFileName and assign the value to the +// instrumented BB and the edges. This function also updates ProgramMaxCount. +// Return true if the profile are successfully read, and false on errors. +bool PGOUseFunc::readCounters(IndexedInstrProfReader *PGOReader) { + auto &Ctx = M->getContext(); + ErrorOr Result = + PGOReader->getInstrProfRecord(FuncInfo.FuncName, FuncInfo.FunctionHash); + if (std::error_code EC = Result.getError()) { + if (EC == instrprof_error::unknown_function) + NumOfPGOMissing++; + else if (EC == instrprof_error::hash_mismatch || + EC == llvm::instrprof_error::malformed) + NumOfPGOMismatch++; + + std::string Msg = EC.message() + std::string(" ") + F.getName().str(); + Ctx.diagnose( + DiagnosticInfoPGOProfile(M->getName().data(), Msg, DS_Warning)); + return false; + } + std::vector CountFromProfile = Result.get().Counts; + + NumOfPGOFunc++; + DEBUG(dbgs() << CountFromProfile.size() << " counts\n"); + uint64_t ValueSum = 0; + for (unsigned i = 0, E = CountFromProfile.size(); i < E; i++) { + DEBUG(dbgs() << " " << i << ": " << CountFromProfile[i] << "\n"); + ValueSum += CountFromProfile[i]; + } + + DEBUG(dbgs() << "SUM = " << ValueSum << "\n"); + + getBBInfo(nullptr).UnknownCountOutEdge = 2; + getBBInfo(nullptr).UnknownCountInEdge = 2; + + setInstrumentedCounts(&CountFromProfile); + ProgramMaxCount = PGOReader->getMaximumFunctionCount(); + return true; +} + +// Populate the counters from instrumented BBs to all BBs. +// In the end of this operation, all BBs should have a valid count value. +void PGOUseFunc::populateCounters() { + // First set up Count variable for all BBs. + for (auto &Ei : FuncInfo.MST.AllEdges) { + if (Ei->Removed) + continue; + + BasicBlock *SrcBB = const_cast(Ei->SrcBB); + BasicBlock *DestBB = const_cast(Ei->DestBB); + UseBBInfo &SrcInfo = getBBInfo(SrcBB); + UseBBInfo &DestInfo = getBBInfo(DestBB); + SrcInfo.OutEdges.push_back(Ei.get()); + DestInfo.InEdges.push_back(Ei.get()); + SrcInfo.UnknownCountOutEdge++; + DestInfo.UnknownCountInEdge++; + + if (!Ei->CountValid) + continue; + DestInfo.UnknownCountInEdge--; + SrcInfo.UnknownCountOutEdge--; + } + + bool Changes = true; + unsigned NumPasses = 0; + while (Changes) { + NumPasses++; + Changes = false; + + // For efficient traversal, it's better to start from the end as most + // of the instrumented edges are at the end. + for (auto &BBB : reverse(F)) { + BasicBlock *BB = &BBB; + UseBBInfo &Count = getBBInfo(BB); + if (!Count.CountValid) { + if (Count.UnknownCountOutEdge == 0) { + Count.CountValue = sumEdgeCount(Count.OutEdges); + Count.CountValid = true; + Changes = true; + } else if (Count.UnknownCountInEdge == 0) { + Count.CountValue = sumEdgeCount(Count.InEdges); + Count.CountValid = true; + Changes = true; + } + } + if (Count.CountValid) { + if (Count.UnknownCountOutEdge == 1) { + uint64_t Total = Count.CountValue - sumEdgeCount(Count.OutEdges); + setEdgeCount(Count.OutEdges, Total); + Changes = true; + } + if (Count.UnknownCountInEdge == 1) { + uint64_t Total = Count.CountValue - sumEdgeCount(Count.InEdges); + setEdgeCount(Count.InEdges, Total); + Changes = true; + } + } + } + } + + DEBUG(dbgs() << "Populate counts in " << NumPasses << " passes.\n"); + // Assert every BB has a valid counter. + uint64_t FuncEntryCount = getBBInfo(&*F.begin()).CountValue; + uint64_t FuncMaxCount = FuncEntryCount; + for (auto &BB : F) { + assert(getBBInfo(&BB).CountValid && "BB count is not valid"); + uint64_t Count = getBBInfo(&BB).CountValue; + if (Count > FuncMaxCount) + FuncMaxCount = Count; + } + applyFunctionAttributes(FuncEntryCount, FuncMaxCount); + + DEBUG(FuncInfo.dumpInfo("after reading profile.")); +} + +// Assign the scaled count values to the BB with multiple out edges. +void PGOUseFunc::setBranchWeights() { + // Generate MD_prof metadata for every branch instruction. + DEBUG(dbgs() << "\nSetting branch weights.\n"); + MDBuilder MDB(M->getContext()); + for (auto &BB : F) { + TerminatorInst *TI = BB.getTerminator(); + if (TI->getNumSuccessors() < 2) + continue; + if (!isa(TI) && !isa(TI)) + continue; + if (getBBInfo(&BB).CountValue == 0) + continue; + + // We have a non-zero Branch BB. + const UseBBInfo &BBCountInfo = getBBInfo(&BB); + unsigned Size = BBCountInfo.OutEdges.size(); + SmallVector EdgeCounts(Size, 0); + uint64_t MaxCount = 0; + for (unsigned s = 0; s < Size; s++) { + const PGOUseEdge *E = BBCountInfo.OutEdges[s]; + const BasicBlock *SrcBB = E->SrcBB; + const BasicBlock *DestBB = E->DestBB; + if (DestBB == 0) + continue; + unsigned SuccNum = GetSuccessorNumber(const_cast(SrcBB), + const_cast(DestBB)); + uint64_t EdgeCount = E->CountValue; + if (EdgeCount > MaxCount) + MaxCount = EdgeCount; + EdgeCounts[SuccNum] = EdgeCount; + } + assert(MaxCount > 0 && "Bad max count"); + uint64_t Scale = calculateCountScale(MaxCount); + SmallVector Weights; + for (const auto &ECI : EdgeCounts) + Weights.push_back(scaleBranchCount(ECI, Scale)); + + TI->setMetadata(llvm::LLVMContext::MD_prof, + MDB.createBranchWeights(Weights)); + } +} +} // end anonymous namespace + +void static instrumentOnFunc(PGOGenFunc &Func) { Func.instrumentCFG(); } + +bool PGOInstrumentationGen::runOnModule(Module &M) { + for (auto &F : M) { + if (F.isDeclaration()) + continue; + BranchProbabilityInfo *BPI = + &(getAnalysis(F).getBPI()); + BlockFrequencyInfo *BFI = + &(getAnalysis(F).getBFI()); + PGOGenFunc Func(F, &M, BPI, BFI); + instrumentOnFunc(Func); + } + return true; +} + +void static setPGOCountOnFunc(PGOUseFunc &Func, + IndexedInstrProfReader *PGOReader) { + if (Func.readCounters(PGOReader)) { + Func.populateCounters(); + Func.setBranchWeights(); + } +} + +bool PGOInstrumentationUse::runOnModule(Module &M) { + DEBUG(dbgs() << "Read in profile counters: "); + auto &Ctx = M.getContext(); + // Read the counter array from file. + auto ReaderOrErr = IndexedInstrProfReader::create(ProfileFileName); + if (std::error_code EC = ReaderOrErr.getError()) { + Ctx.diagnose( + DiagnosticInfoPGOProfile(ProfileFileName.data(), EC.message())); + return false; + } + + PGOReader = std::move(ReaderOrErr.get()); + if (!PGOReader) { + Ctx.diagnose(DiagnosticInfoPGOProfile(ProfileFileName.data(), + "Cannot get PGOReader")); + return false; + } + + for (auto &F : M) { + if (F.isDeclaration()) + continue; + BranchProbabilityInfo *BPI = + &(getAnalysis(F).getBPI()); + BlockFrequencyInfo *BFI = + &(getAnalysis(F).getBFI()); + PGOUseFunc Func(F, &M, BPI, BFI); + setPGOCountOnFunc(Func, PGOReader.get()); + } + return true; +}