From 51c40d78daadef14ce2ace0a1ff89354c9282328 Mon Sep 17 00:00:00 2001 From: Mehdi Amini Date: Tue, 16 Feb 2021 22:10:51 +0000 Subject: [PATCH] Revert "[SampleFDO][NFC] Refactor SampleProfile.cpp" This reverts commit 310b35304cdf5a230c042904655583c5532d3e91. The build is broken with -DBUILD_SHARED_LIBS=ON : lib/ProfileData/CMakeFiles/LLVMProfileData.dir/SampleProfileLoaderBaseUtil.cpp.o: In function `llvm::sampleprofutil::callsiteIsHot(llvm::sampleprof::FunctionSamples const*, llvm::ProfileSummaryInfo*, bool)': SampleProfileLoaderBaseUtil.cpp:(.text._ZN4llvm14sampleprofutil13callsiteIsHotEPKNS_10sampleprof15FunctionSamplesEPNS_18ProfileSummaryInfoEb+0x1a): undefined reference to `llvm::ProfileSummaryInfo::isColdCount(unsigned long) const' SampleProfileLoaderBaseUtil.cpp:(.text._ZN4llvm14sampleprofutil13callsiteIsHotEPKNS_10sampleprof15FunctionSamplesEPNS_18ProfileSummaryInfoEb+0x28): undefined reference to `llvm::ProfileSummaryInfo::isHotCount(unsigned long) const' ... --- .../ProfileData/SampleProfileLoaderBaseImpl.h | 863 -------------- .../ProfileData/SampleProfileLoaderBaseUtil.h | 97 -- lib/ProfileData/CMakeLists.txt | 1 - .../SampleProfileLoaderBaseUtil.cpp | 192 ---- lib/Transforms/IPO/SampleProfile.cpp | 1001 ++++++++++++++++- 5 files changed, 992 insertions(+), 1162 deletions(-) delete mode 100644 include/llvm/ProfileData/SampleProfileLoaderBaseImpl.h delete mode 100644 include/llvm/ProfileData/SampleProfileLoaderBaseUtil.h delete mode 100644 lib/ProfileData/SampleProfileLoaderBaseUtil.cpp diff --git a/include/llvm/ProfileData/SampleProfileLoaderBaseImpl.h b/include/llvm/ProfileData/SampleProfileLoaderBaseImpl.h deleted file mode 100644 index 48d7ae41e9d..00000000000 --- a/include/llvm/ProfileData/SampleProfileLoaderBaseImpl.h +++ /dev/null @@ -1,863 +0,0 @@ -////===- SampleProfileLoadBaseImpl.h - Profile loader base impl --*- C++-*-===// -// -// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. -// See https://llvm.org/LICENSE.txt for license information. -// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception -// -//===----------------------------------------------------------------------===// -// -/// \file -/// This file provides the interface for the sampled PGO profile loader base -/// implementation. -// -//===----------------------------------------------------------------------===// - -#ifndef LLVM_TRANSFORMS_IPO_SAMPLEPROFILELOADERIMPL_H -#define LLVM_TRANSFORMS_IPO_SAMPLEPROFILELOADERIMPL_H - -#include "llvm/ADT/ArrayRef.h" -#include "llvm/ADT/DenseMap.h" -#include "llvm/ADT/DenseSet.h" -#include "llvm/ADT/SmallPtrSet.h" -#include "llvm/ADT/SmallSet.h" -#include "llvm/ADT/SmallVector.h" -#include "llvm/Analysis/LoopInfo.h" -#include "llvm/Analysis/OptimizationRemarkEmitter.h" -#include "llvm/Analysis/PostDominators.h" -#include "llvm/Analysis/ProfileSummaryInfo.h" -#include "llvm/IR/BasicBlock.h" -#include "llvm/IR/CFG.h" -#include "llvm/IR/DebugInfoMetadata.h" -#include "llvm/IR/DebugLoc.h" -#include "llvm/IR/Dominators.h" -#include "llvm/IR/Function.h" -#include "llvm/IR/Instruction.h" -#include "llvm/IR/Instructions.h" -#include "llvm/IR/Module.h" -#include "llvm/ProfileData/SampleProf.h" -#include "llvm/ProfileData/SampleProfReader.h" -#include "llvm/ProfileData/SampleProfileLoaderBaseUtil.h" -#include "llvm/Support/CommandLine.h" -#include "llvm/Support/GenericDomTree.h" -#include "llvm/Support/raw_ostream.h" - -namespace llvm { -using namespace llvm; -using namespace sampleprof; -using ProfileCount = Function::ProfileCount; -namespace sampleprofutil { -bool callsiteIsHot(const SampleCoverageTracker *CT, - const FunctionSamples *CallsiteFS, ProfileSummaryInfo *PSI, - bool ProfAccForSymsInList); -} // namespace sampleprofutil -using namespace sampleprofutil; - -#define DEBUG_TYPE "sample-profile-impl" - -using BlockWeightMap = DenseMap; -using EquivalenceClassMap = DenseMap; -using Edge = std::pair; -using EdgeWeightMap = DenseMap; -using BlockEdgeMap = - DenseMap>; - -extern cl::opt SampleProfileMaxPropagateIterations; -extern cl::opt SampleProfileRecordCoverage; -extern cl::opt SampleProfileSampleCoverage; -extern cl::opt NoWarnSampleUnused; - -class SampleProfileLoaderBaseImpl { -public: - SampleProfileLoaderBaseImpl(std::string Name) : Filename(Name) {} - void dump() { Reader->dump(); } - -protected: - ~SampleProfileLoaderBaseImpl() = default; - friend class SampleCoverageTracker; - - unsigned getFunctionLoc(Function &F); - virtual ErrorOr getInstWeight(const Instruction &Inst); - ErrorOr getInstWeightImpl(const Instruction &Inst); - ErrorOr getBlockWeight(const BasicBlock *BB); - mutable DenseMap - DILocation2SampleMap; - virtual const FunctionSamples * - findFunctionSamples(const Instruction &I) const; - 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); - template - 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 clearFunctionData(); - void computeDominanceAndLoopInfo(Function &F); - bool - computeAndPropagateWeights(Function &F, - const DenseSet &InlinedGUIDs); - void emitCoverageRemarks(Function &F); - - /// 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; - - /// Map edges to their computed weights. - /// - /// Edge weights are computed by propagating basic block weights in - /// SampleProfile::propagateWeights. - EdgeWeightMap EdgeWeights; - - /// Set of visited blocks during propagation. - SmallPtrSet VisitedBlocks; - - /// Set of visited edges during propagation. - SmallSet VisitedEdges; - - /// 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; - - /// Dominance, post-dominance and loop information. - std::unique_ptr DT; - std::unique_ptr PDT; - std::unique_ptr LI; - - /// Predecessors for each basic block in the CFG. - BlockEdgeMap Predecessors; - - /// Successors for each basic block in the CFG. - BlockEdgeMap Successors; - - /// Profile coverage tracker. - SampleCoverageTracker CoverageTracker; - - /// Profile reader object. - std::unique_ptr Reader; - - /// Samples collected for the body of this function. - FunctionSamples *Samples = nullptr; - - /// Name of the profile file to load. - std::string Filename; - - /// Profile Summary Info computed from sample profile. - ProfileSummaryInfo *PSI = nullptr; - - /// Optimization Remark Emitter used to emit diagnostic remarks. - OptimizationRemarkEmitter *ORE = nullptr; -}; - -/// Clear all the per-function data used to load samples and propagate weights. -void SampleProfileLoaderBaseImpl::clearFunctionData() { - BlockWeights.clear(); - EdgeWeights.clear(); - VisitedBlocks.clear(); - VisitedEdges.clear(); - EquivalenceClass.clear(); - DT = nullptr; - PDT = nullptr; - LI = nullptr; - Predecessors.clear(); - Successors.clear(); - CoverageTracker.clear(); -} - -#ifndef NDEBUG -/// 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 SampleProfileLoaderBaseImpl::printEdgeWeight(raw_ostream &OS, Edge E) { - OS << "weight[" << E.first->getName() << "->" << E.second->getName() - << "]: " << EdgeWeights[E] << "\n"; -} - -/// 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 SampleProfileLoaderBaseImpl::printBlockEquivalence(raw_ostream &OS, - const BasicBlock *BB) { - const BasicBlock *Equiv = EquivalenceClass[BB]; - OS << "equivalence[" << BB->getName() - << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n"; -} - -/// 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 SampleProfileLoaderBaseImpl::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"; -} -#endif - -/// 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 -SampleProfileLoaderBaseImpl::getInstWeight(const Instruction &Inst) { - return getInstWeightImpl(Inst); -} - -ErrorOr -SampleProfileLoaderBaseImpl::getInstWeightImpl(const Instruction &Inst) { - const FunctionSamples *FS = findFunctionSamples(Inst); - if (!FS) - return std::error_code(); - - const DebugLoc &DLoc = Inst.getDebugLoc(); - if (!DLoc) - return std::error_code(); - - const DILocation *DIL = DLoc; - uint32_t LineOffset = FunctionSamples::getOffset(DIL); - uint32_t Discriminator = DIL->getBaseDiscriminator(); - ErrorOr R = FS->findSamplesAt(LineOffset, Discriminator); - if (R) { - bool FirstMark = - CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get()); - if (FirstMark) { - ORE->emit([&]() { - OptimizationRemarkAnalysis Remark(DEBUG_TYPE, "AppliedSamples", &Inst); - Remark << "Applied " << ore::NV("NumSamples", *R); - Remark << " samples from profile (offset: "; - Remark << ore::NV("LineOffset", LineOffset); - if (Discriminator) { - Remark << "."; - Remark << ore::NV("Discriminator", Discriminator); - } - Remark << ")"; - return Remark; - }); - } - LLVM_DEBUG(dbgs() << " " << DLoc.getLine() << "." - << DIL->getBaseDiscriminator() << ":" << Inst - << " (line offset: " << LineOffset << "." - << DIL->getBaseDiscriminator() << " - weight: " << R.get() - << ")\n"); - } - return R; -} - -/// 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 -SampleProfileLoaderBaseImpl::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(); -} - -/// 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 SampleProfileLoaderBaseImpl::computeBlockWeights(Function &F) { - bool Changed = false; - LLVM_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; - } - LLVM_DEBUG(printBlockWeight(dbgs(), &BB)); - } - - return Changed; -} - -/// 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 *SampleProfileLoaderBaseImpl::findFunctionSamples( - const Instruction &Inst) const { - const DILocation *DIL = Inst.getDebugLoc(); - if (!DIL) - return Samples; - - auto it = DILocation2SampleMap.try_emplace(DIL, nullptr); - if (it.second) { - it.first->second = Samples->findFunctionSamples(DIL, Reader->getRemapper()); - } - return it.first->second; -} - -/// 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. -template -void SampleProfileLoaderBaseImpl::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; - } -} - -/// 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 SampleProfileLoaderBaseImpl::findEquivalenceClasses(Function &F) { - SmallVector DominatedBBs; - LLVM_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)) { - LLVM_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()); - - LLVM_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. - LLVM_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]; - LLVM_DEBUG(printBlockWeight(dbgs(), BB)); - } -} - -/// 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 SampleProfileLoaderBaseImpl::visitEdge(Edge E, - unsigned *NumUnknownEdges, - Edge *UnknownEdge) { - if (!VisitedEdges.count(E)) { - (*NumUnknownEdges)++; - *UnknownEdge = E; - return 0; - } - - return EdgeWeights[E]; -} - -/// 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 SampleProfileLoaderBaseImpl::propagateThroughEdges(Function &F, - bool UpdateBlockCount) { - bool Changed = false; - LLVM_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; - LLVM_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; - LLVM_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; - LLVM_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; -} - -/// 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 SampleProfileLoaderBaseImpl::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 (BasicBlock *B2 : predecessors(B1)) - 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 (BasicBlock *B2 : successors(B1)) - if (Visited.insert(B2).second) - Successors[B1].push_back(B2); - } -} - -/// 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 SampleProfileLoaderBaseImpl::propagateWeights(Function &F) { - bool Changed = true; - unsigned I = 0; - - // 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 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. -/// -/// \param F The function to query. -/// -/// \returns true if \p F was modified. Returns false, otherwise. -bool SampleProfileLoaderBaseImpl::computeAndPropagateWeights( - Function &F, const DenseSet &InlinedGUIDs) { - bool Changed = (InlinedGUIDs.size() != 0); - - // Compute basic block weights. - Changed |= computeBlockWeights(F); - - if (Changed) { - // Add an entry count to the function using the samples gathered at the - // function entry. - // Sets the GUIDs that are inlined in the profiled binary. This is used - // for ThinLink to make correct liveness analysis, and also make the IR - // match the profiled binary before annotation. - F.setEntryCount( - ProfileCount(Samples->getHeadSamples() + 1, Function::PCT_Real), - &InlinedGUIDs); - - // Compute dominance and loop info needed for propagation. - computeDominanceAndLoopInfo(F); - - // Find equivalence classes. - findEquivalenceClasses(F); - - // Propagate weights to all edges. - propagateWeights(F); - } - - return Changed; -} - -void SampleProfileLoaderBaseImpl::emitCoverageRemarks(Function &F) { - // If coverage checking was requested, compute it now. - if (SampleProfileRecordCoverage) { - unsigned Used = CoverageTracker.countUsedRecords(Samples, PSI); - unsigned Total = CoverageTracker.countBodyRecords(Samples, PSI); - 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, PSI); - 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)); - } - } -} - -/// 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 SampleProfileLoaderBaseImpl::getFunctionLoc(Function &F) { - if (DISubprogram *S = F.getSubprogram()) - return S->getLine(); - - if (NoWarnSampleUnused) - return 0; - - // 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 SampleProfileLoaderBaseImpl::computeDominanceAndLoopInfo(Function &F) { - DT.reset(new DominatorTree); - DT->recalculate(F); - - PDT.reset(new PostDominatorTree(F)); - - LI.reset(new LoopInfo); - LI->analyze(*DT); -} - -#undef DEBUG_TYPE - -} // namespace llvm -#endif // LLVM_TRANSFORMS_IPO_SAMPLEPROFILELOADERIMPL_H diff --git a/include/llvm/ProfileData/SampleProfileLoaderBaseUtil.h b/include/llvm/ProfileData/SampleProfileLoaderBaseUtil.h deleted file mode 100644 index 37dc8d8187d..00000000000 --- a/include/llvm/ProfileData/SampleProfileLoaderBaseUtil.h +++ /dev/null @@ -1,97 +0,0 @@ -////===- SampleProfileLoadBaseUtil.h - Profile loader util func --*- C++-*-===// -// -// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. -// See https://llvm.org/LICENSE.txt for license information. -// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception -// -//===----------------------------------------------------------------------===// -// -/// \file -/// This file provides the utility functions for the sampled PGO loader base -/// implementation. -// -//===----------------------------------------------------------------------===// - -#ifndef LLVM_TRANSFORMS_IPO_SAMPLEPROFILELOADERUTIL_H -#define LLVM_TRANSFORMS_IPO_SAMPLEPROFILELOADERUTIL_H - -#include "llvm/ADT/DenseMap.h" -#include "llvm/Analysis/ProfileSummaryInfo.h" -#include "llvm/IR/BasicBlock.h" -#include "llvm/IR/CFG.h" -#include "llvm/IR/DebugLoc.h" -#include "llvm/IR/Function.h" -#include "llvm/ProfileData/SampleProf.h" -#include "llvm/Support/CommandLine.h" - -namespace llvm { -using namespace sampleprof; - -extern cl::opt SampleProfileMaxPropagateIterations; -extern cl::opt SampleProfileRecordCoverage; -extern cl::opt SampleProfileSampleCoverage; -extern cl::opt NoWarnSampleUnused; - -namespace sampleprofutil { - -class SampleCoverageTracker { -public: - 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, - ProfileSummaryInfo *PSI) const; - unsigned countBodyRecords(const FunctionSamples *FS, - ProfileSummaryInfo *PSI) const; - uint64_t getTotalUsedSamples() const { return TotalUsedSamples; } - uint64_t countBodySamples(const FunctionSamples *FS, - ProfileSummaryInfo *PSI) const; - - void clear() { - SampleCoverage.clear(); - TotalUsedSamples = 0; - } - void setProfAccForSymsInList(bool V) { ProfAccForSymsInList = V; } - -private: - using BodySampleCoverageMap = std::map; - using FunctionSamplesCoverageMap = - DenseMap; - - /// 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 = 0; - - // For symbol in profile symbol list, whether to regard their profiles - // to be accurate. This is passed from the SampleLoader instance. - bool ProfAccForSymsInList = false; -}; - -/// Return true if the given callsite is hot wrt to hot cutoff threshold. -bool callsiteIsHot(const FunctionSamples *CallsiteFS, ProfileSummaryInfo *PSI, - bool ProfAccForSymsInList); - -} // end of namespace sampleprofutil -} // end of namespace llvm - -#endif // LLVM_TRANSFORMS_IPO_SAMPLEPROFILELOADERUTIL_H diff --git a/lib/ProfileData/CMakeLists.txt b/lib/ProfileData/CMakeLists.txt index 4125fac918a..2a377e4d74d 100644 --- a/lib/ProfileData/CMakeLists.txt +++ b/lib/ProfileData/CMakeLists.txt @@ -5,7 +5,6 @@ add_llvm_component_library(LLVMProfileData InstrProfWriter.cpp ProfileSummaryBuilder.cpp SampleProf.cpp - SampleProfileLoaderBaseUtil.cpp SampleProfReader.cpp SampleProfWriter.cpp diff --git a/lib/ProfileData/SampleProfileLoaderBaseUtil.cpp b/lib/ProfileData/SampleProfileLoaderBaseUtil.cpp deleted file mode 100644 index 87ef4573704..00000000000 --- a/lib/ProfileData/SampleProfileLoaderBaseUtil.cpp +++ /dev/null @@ -1,192 +0,0 @@ -//===- SampleProfileLoaderBaseUtil.cpp - Profile loader Util func ---------===// -// -// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. -// See https://llvm.org/LICENSE.txt for license information. -// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception -// -//===----------------------------------------------------------------------===// -// -// This file implements the SampleProfileLoader base utility functions. -// -//===----------------------------------------------------------------------===// - -#include "llvm/ProfileData/SampleProfileLoaderBaseUtil.h" - -namespace llvm { - -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.")); - -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.")); - -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.")); - -cl::opt NoWarnSampleUnused( - "no-warn-sample-unused", cl::init(false), cl::Hidden, - cl::desc("Use this option to turn off/on warnings about function with " - "samples but without debug information to use those samples. ")); - -namespace sampleprofutil { - -/// Return true if the given callsite is hot wrt to hot cutoff threshold. -/// -/// 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 compare the callsite -/// sample count with the hot cutoff computed by ProfileSummaryInfo, it is -/// regarded as hot if the count is above the cutoff value. -/// -/// When ProfileAccurateForSymsInList is enabled and profile symbol list -/// is present, functions in the profile symbol list but without profile will -/// be regarded as cold and much less inlining will happen in CGSCC inlining -/// pass, so we tend to lower the hot criteria here to allow more early -/// inlining to happen for warm callsites and it is helpful for performance. -bool callsiteIsHot(const FunctionSamples *CallsiteFS, ProfileSummaryInfo *PSI, - bool ProfAccForSymsInList) { - if (!CallsiteFS) - return false; // The callsite was not inlined in the original binary. - - assert(PSI && "PSI is expected to be non null"); - uint64_t CallsiteTotalSamples = CallsiteFS->getTotalSamples(); - if (ProfAccForSymsInList) - return !PSI->isColdCount(CallsiteTotalSamples); - else - return PSI->isHotCount(CallsiteTotalSamples); -} - -#if 0 -void clearCoverageTracker(SampleCoverageTracker *CT) { CT->clear(); } - -bool markSamplesUsed(SampleCoverageTracker *CT, const FunctionSamples *FS, - uint32_t LineOffset, uint32_t Discriminator, - uint64_t Samples) { - return CT->markSamplesUsed(FS, LineOffset, Discriminator, Samples); -} -unsigned computeCoverage(const SampleCoverageTracker *CT, unsigned Used, - unsigned Total) { - return CT->computeCoverage(Used, Total); -} -unsigned countUsedRecords(const SampleCoverageTracker *CT, - const FunctionSamples *FS, ProfileSummaryInfo *PSI) { - return CT->countUsedRecords(FS, PSI); -} -unsigned countBodyRecords(const SampleCoverageTracker *CT, - const FunctionSamples *FS, ProfileSummaryInfo *PSI) { - return CT->countBodyRecords(FS, PSI); -} -uint64_t getTotalUsedSamples(const SampleCoverageTracker *CT) { - return CT->getTotalUsedSamples(); -} -uint64_t countBodySamples(const SampleCoverageTracker *CT, - const FunctionSamples *FS, ProfileSummaryInfo *PSI) { - return CT->countBodySamples(FS, PSI); -} -#endif - -/// 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, - ProfileSummaryInfo *PSI) 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()) - for (const auto &J : I.second) { - const FunctionSamples *CalleeSamples = &J.second; - if (callsiteIsHot(CalleeSamples, PSI, ProfAccForSymsInList)) - Count += countUsedRecords(CalleeSamples, PSI); - } - - 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, - ProfileSummaryInfo *PSI) const { - unsigned Count = FS->getBodySamples().size(); - - // Only count records in hot callsites. - for (const auto &I : FS->getCallsiteSamples()) - for (const auto &J : I.second) { - const FunctionSamples *CalleeSamples = &J.second; - if (callsiteIsHot(CalleeSamples, PSI, ProfAccForSymsInList)) - Count += countBodyRecords(CalleeSamples, PSI); - } - - 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, - ProfileSummaryInfo *PSI) 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()) - for (const auto &J : I.second) { - const FunctionSamples *CalleeSamples = &J.second; - if (callsiteIsHot(CalleeSamples, PSI, ProfAccForSymsInList)) - Total += countBodySamples(CalleeSamples, PSI); - } - - 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; -} - -} // end of namespace sampleprofutil -} // end of namespace llvm diff --git a/lib/Transforms/IPO/SampleProfile.cpp b/lib/Transforms/IPO/SampleProfile.cpp index 4387a264036..812700ac25d 100644 --- a/lib/Transforms/IPO/SampleProfile.cpp +++ b/lib/Transforms/IPO/SampleProfile.cpp @@ -69,8 +69,6 @@ #include "llvm/ProfileData/InstrProf.h" #include "llvm/ProfileData/SampleProf.h" #include "llvm/ProfileData/SampleProfReader.h" -#include "llvm/ProfileData/SampleProfileLoaderBaseImpl.h" -#include "llvm/ProfileData/SampleProfileLoaderBaseUtil.h" #include "llvm/Support/Casting.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" @@ -99,7 +97,6 @@ using namespace llvm; using namespace sampleprof; -using namespace llvm::sampleprofutil; using ProfileCount = Function::ProfileCount; #define DEBUG_TYPE "sample-profile" #define CSINLINE_DEBUG DEBUG_TYPE "-inline" @@ -135,6 +132,26 @@ static cl::opt SampleProfileRemappingFile( "sample-profile-remapping-file", cl::init(""), cl::value_desc("filename"), cl::desc("Profile remapping 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 NoWarnSampleUnused( + "no-warn-sample-unused", cl::init(false), cl::Hidden, + cl::desc("Use this option to turn off/on warnings about function with " + "samples but without debug information to use those samples. ")); + static cl::opt ProfileSampleAccurate( "profile-sample-accurate", cl::Hidden, cl::init(false), cl::desc("If the sample profile is accurate, we will mark all un-sampled " @@ -227,12 +244,65 @@ using EdgeWeightMap = DenseMap; using BlockEdgeMap = DenseMap>; +class SampleCoverageTracker { +public: + 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, + ProfileSummaryInfo *PSI) const; + unsigned countBodyRecords(const FunctionSamples *FS, + ProfileSummaryInfo *PSI) const; + uint64_t getTotalUsedSamples() const { return TotalUsedSamples; } + uint64_t countBodySamples(const FunctionSamples *FS, + ProfileSummaryInfo *PSI) const; + + void clear() { + SampleCoverage.clear(); + TotalUsedSamples = 0; + } + inline void setProfAccForSymsInList(bool V) { ProfAccForSymsInList = V; } + +private: + using BodySampleCoverageMap = std::map; + using FunctionSamplesCoverageMap = + DenseMap; + + /// 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 = 0; + + // For symbol in profile symbol list, whether to regard their profiles + // to be accurate. This is passed from the SampleLoader instance. + bool ProfAccForSymsInList = false; +}; + class GUIDToFuncNameMapper { public: GUIDToFuncNameMapper(Module &M, SampleProfileReader &Reader, - DenseMap &GUIDToFuncNameMap) + DenseMap &GUIDToFuncNameMap) : CurrentReader(Reader), CurrentModule(M), - CurrentGUIDToFuncNameMap(GUIDToFuncNameMap) { + CurrentGUIDToFuncNameMap(GUIDToFuncNameMap) { if (!CurrentReader.useMD5()) return; @@ -327,6 +397,99 @@ using CandidateQueue = PriorityQueue, CandidateComparer>; +class SampleProfileLoaderBaseImpl { +public: + SampleProfileLoaderBaseImpl(std::string Name) : Filename(Name) {} + void dump() { Reader->dump(); } + +protected: + friend class SampleCoverageTracker; + + ~SampleProfileLoaderBaseImpl() = default; + + unsigned getFunctionLoc(Function &F); + virtual ErrorOr getInstWeight(const Instruction &Inst); + ErrorOr getInstWeightImpl(const Instruction &Inst); + ErrorOr getBlockWeight(const BasicBlock *BB); + mutable DenseMap + DILocation2SampleMap; + virtual const FunctionSamples * + findFunctionSamples(const Instruction &I) const; + 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); + template + 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 clearFunctionData(); + void computeDominanceAndLoopInfo(Function &F); + bool + computeAndPropagateWeights(Function &F, + const DenseSet &InlinedGUIDs); + void emitCoverageRemarks(Function &F); + + /// 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; + + /// Map edges to their computed weights. + /// + /// Edge weights are computed by propagating basic block weights in + /// SampleProfile::propagateWeights. + EdgeWeightMap EdgeWeights; + + /// Set of visited blocks during propagation. + SmallPtrSet VisitedBlocks; + + /// Set of visited edges during propagation. + SmallSet VisitedEdges; + + /// 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; + + /// Dominance, post-dominance and loop information. + std::unique_ptr DT; + std::unique_ptr PDT; + std::unique_ptr LI; + + /// Predecessors for each basic block in the CFG. + BlockEdgeMap Predecessors; + + /// Successors for each basic block in the CFG. + BlockEdgeMap Successors; + + SampleCoverageTracker CoverageTracker; + + /// Profile reader object. + std::unique_ptr Reader; + + /// Samples collected for the body of this function. + FunctionSamples *Samples = nullptr; + + /// Name of the profile file to load. + std::string Filename; + + /// Profile Summary Info computed from sample profile. + ProfileSummaryInfo *PSI = nullptr; + + /// Optimization Remark Emitter used to emit diagnostic remarks. + OptimizationRemarkEmitter *ORE = nullptr; +}; + /// Sample profile pass. /// /// This pass reads profile data from the file specified by @@ -500,6 +663,193 @@ private: } // end anonymous namespace +/// Return true if the given callsite is hot wrt to hot cutoff threshold. +/// +/// 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 compare the callsite +/// sample count with the hot cutoff computed by ProfileSummaryInfo, it is +/// regarded as hot if the count is above the cutoff value. +/// +/// When ProfileAccurateForSymsInList is enabled and profile symbol list +/// is present, functions in the profile symbol list but without profile will +/// be regarded as cold and much less inlining will happen in CGSCC inlining +/// pass, so we tend to lower the hot criteria here to allow more early +/// inlining to happen for warm callsites and it is helpful for performance. +static bool callsiteIsHot(const FunctionSamples *CallsiteFS, + ProfileSummaryInfo *PSI, bool ProfAccForSymsInList) { + if (!CallsiteFS) + return false; // The callsite was not inlined in the original binary. + + assert(PSI && "PSI is expected to be non null"); + uint64_t CallsiteTotalSamples = CallsiteFS->getTotalSamples(); + if (ProfAccForSymsInList) + return !PSI->isColdCount(CallsiteTotalSamples); + else + return PSI->isHotCount(CallsiteTotalSamples); +} + +/// 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, + ProfileSummaryInfo *PSI) 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()) + for (const auto &J : I.second) { + const FunctionSamples *CalleeSamples = &J.second; + if (callsiteIsHot(CalleeSamples, PSI, ProfAccForSymsInList)) + Count += countUsedRecords(CalleeSamples, PSI); + } + + 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, + ProfileSummaryInfo *PSI) const { + unsigned Count = FS->getBodySamples().size(); + + // Only count records in hot callsites. + for (const auto &I : FS->getCallsiteSamples()) + for (const auto &J : I.second) { + const FunctionSamples *CalleeSamples = &J.second; + if (callsiteIsHot(CalleeSamples, PSI, ProfAccForSymsInList)) + Count += countBodyRecords(CalleeSamples, PSI); + } + + 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, + ProfileSummaryInfo *PSI) 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()) + for (const auto &J : I.second) { + const FunctionSamples *CalleeSamples = &J.second; + if (callsiteIsHot(CalleeSamples, PSI, ProfAccForSymsInList)) + Total += countBodySamples(CalleeSamples, PSI); + } + + 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 SampleProfileLoaderBaseImpl::clearFunctionData() { + BlockWeights.clear(); + EdgeWeights.clear(); + VisitedBlocks.clear(); + VisitedEdges.clear(); + EquivalenceClass.clear(); + DT = nullptr; + PDT = nullptr; + LI = nullptr; + Predecessors.clear(); + Successors.clear(); + CoverageTracker.clear(); +} + +#ifndef NDEBUG +/// 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 SampleProfileLoaderBaseImpl::printEdgeWeight(raw_ostream &OS, Edge E) { + OS << "weight[" << E.first->getName() << "->" << E.second->getName() + << "]: " << EdgeWeights[E] << "\n"; +} + +/// 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 SampleProfileLoaderBaseImpl::printBlockEquivalence(raw_ostream &OS, + const BasicBlock *BB) { + const BasicBlock *Equiv = EquivalenceClass[BB]; + OS << "equivalence[" << BB->getName() + << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n"; +} + +/// 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 SampleProfileLoaderBaseImpl::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"; +} +#endif + +/// 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 +SampleProfileLoaderBaseImpl::getInstWeight(const Instruction &Inst) { + return getInstWeightImpl(Inst); +} + ErrorOr SampleProfileLoader::getInstWeight(const Instruction &Inst) { if (FunctionSamples::ProfileIsProbeBased) return getProbeWeight(Inst); @@ -526,6 +876,46 @@ ErrorOr SampleProfileLoader::getInstWeight(const Instruction &Inst) { return getInstWeightImpl(Inst); } +ErrorOr +SampleProfileLoaderBaseImpl::getInstWeightImpl(const Instruction &Inst) { + const FunctionSamples *FS = findFunctionSamples(Inst); + if (!FS) + return std::error_code(); + + const DebugLoc &DLoc = Inst.getDebugLoc(); + if (!DLoc) + return std::error_code(); + + const DILocation *DIL = DLoc; + uint32_t LineOffset = FunctionSamples::getOffset(DIL); + uint32_t Discriminator = DIL->getBaseDiscriminator(); + ErrorOr R = FS->findSamplesAt(LineOffset, Discriminator); + if (R) { + bool FirstMark = + CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get()); + if (FirstMark) { + ORE->emit([&]() { + OptimizationRemarkAnalysis Remark(DEBUG_TYPE, "AppliedSamples", &Inst); + Remark << "Applied " << ore::NV("NumSamples", *R); + Remark << " samples from profile (offset: "; + Remark << ore::NV("LineOffset", LineOffset); + if (Discriminator) { + Remark << "."; + Remark << ore::NV("Discriminator", Discriminator); + } + Remark << ")"; + return Remark; + }); + } + LLVM_DEBUG(dbgs() << " " << DLoc.getLine() << "." + << DIL->getBaseDiscriminator() << ":" << Inst + << " (line offset: " << LineOffset << "." + << DIL->getBaseDiscriminator() << " - weight: " << R.get() + << ")\n"); + } + return R; +} + ErrorOr SampleProfileLoader::getProbeWeight(const Instruction &Inst) { assert(FunctionSamples::ProfileIsProbeBased && "Profile is not pseudo probe based"); @@ -571,6 +961,50 @@ ErrorOr SampleProfileLoader::getProbeWeight(const Instruction &Inst) { return R; } +/// 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 +SampleProfileLoaderBaseImpl::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(); +} + +/// 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 SampleProfileLoaderBaseImpl::computeBlockWeights(Function &F) { + bool Changed = false; + LLVM_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; + } + LLVM_DEBUG(printBlockWeight(dbgs(), &BB)); + } + + return Changed; +} + /// Get the FunctionSamples for a call instruction. /// /// The FunctionSamples of a call/invoke instruction \p Inst is the inlined @@ -665,6 +1099,28 @@ SampleProfileLoader::findIndirectCallFunctionSamples( return R; } +/// 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 *SampleProfileLoaderBaseImpl::findFunctionSamples( + const Instruction &Inst) const { + const DILocation *DIL = Inst.getDebugLoc(); + if (!DIL) + return Samples; + + auto it = DILocation2SampleMap.try_emplace(DIL, nullptr); + if (it.second) { + it.first->second = Samples->findFunctionSamples(DIL, Reader->getRemapper()); + } + return it.first->second; +} + const FunctionSamples * SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const { if (FunctionSamples::ProfileIsProbeBased) { @@ -1200,17 +1656,507 @@ bool SampleProfileLoader::inlineHotFunctionsWithPriority( return Changed; } +/// 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. +template +void SampleProfileLoaderBaseImpl::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; + } +} + +/// 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 SampleProfileLoaderBaseImpl::findEquivalenceClasses(Function &F) { + SmallVector DominatedBBs; + LLVM_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)) { + LLVM_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()); + + LLVM_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. + LLVM_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]; + LLVM_DEBUG(printBlockWeight(dbgs(), BB)); + } +} + +/// 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 SampleProfileLoaderBaseImpl::visitEdge(Edge E, + unsigned *NumUnknownEdges, + Edge *UnknownEdge) { + if (!VisitedEdges.count(E)) { + (*NumUnknownEdges)++; + *UnknownEdge = E; + return 0; + } + + return EdgeWeights[E]; +} + +/// 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 SampleProfileLoaderBaseImpl::propagateThroughEdges(Function &F, + bool UpdateBlockCount) { + bool Changed = false; + LLVM_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; + LLVM_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; + LLVM_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; + LLVM_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; +} + +/// 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 SampleProfileLoaderBaseImpl::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 (BasicBlock *B2 : predecessors(B1)) + 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 (BasicBlock *B2 : successors(B1)) + if (Visited.insert(B2).second) + Successors[B1].push_back(B2); + } +} + /// Returns the sorted CallTargetMap \p M by count in descending order. -static SmallVector -GetSortedValueDataFromCallTargets(const SampleRecord::CallTargetMap &M) { +static SmallVector GetSortedValueDataFromCallTargets( + const SampleRecord::CallTargetMap & M) { SmallVector R; for (const auto &I : SampleRecord::SortCallTargets(M)) { - R.emplace_back( - InstrProfValueData{FunctionSamples::getGUID(I.first), I.second}); + R.emplace_back(InstrProfValueData{FunctionSamples::getGUID(I.first), I.second}); } return R; } +/// 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 SampleProfileLoaderBaseImpl::propagateWeights(Function &F) { + bool Changed = true; + unsigned I = 0; + + // 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 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. +/// +/// \param F The function to query. +/// +/// \returns true if \p F was modified. Returns false, otherwise. +bool SampleProfileLoaderBaseImpl::computeAndPropagateWeights( + Function &F, const DenseSet &InlinedGUIDs) { + bool Changed = (InlinedGUIDs.size() != 0); + + // Compute basic block weights. + Changed |= computeBlockWeights(F); + + if (Changed) { + // Add an entry count to the function using the samples gathered at the + // function entry. + // Sets the GUIDs that are inlined in the profiled binary. This is used + // for ThinLink to make correct liveness analysis, and also make the IR + // match the profiled binary before annotation. + F.setEntryCount( + ProfileCount(Samples->getHeadSamples() + 1, Function::PCT_Real), + &InlinedGUIDs); + + // Compute dominance and loop info needed for propagation. + computeDominanceAndLoopInfo(F); + + // Find equivalence classes. + findEquivalenceClasses(F); + + // Propagate weights to all edges. + propagateWeights(F); + } + + return Changed; +} + +void SampleProfileLoaderBaseImpl::emitCoverageRemarks(Function &F) { + // If coverage checking was requested, compute it now. + if (SampleProfileRecordCoverage) { + unsigned Used = CoverageTracker.countUsedRecords(Samples, PSI); + unsigned Total = CoverageTracker.countBodyRecords(Samples, PSI); + 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, PSI); + 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)); + } + } +} + // Generate MD_prof metadata for every branch instruction using the // edge weights computed during propagation. void SampleProfileLoader::generateMDProfMetadata(Function &F) { @@ -1318,6 +2264,43 @@ void SampleProfileLoader::generateMDProfMetadata(Function &F) { } } +/// 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 SampleProfileLoaderBaseImpl::getFunctionLoc(Function &F) { + if (DISubprogram *S = F.getSubprogram()) + return S->getLine(); + + if (NoWarnSampleUnused) + return 0; + + // 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 SampleProfileLoaderBaseImpl::computeDominanceAndLoopInfo(Function &F) { + DT.reset(new DominatorTree); + DT->recalculate(F); + + PDT.reset(new PostDominatorTree(F)); + + LI.reset(new LoopInfo); + LI->analyze(*DT); +} + /// Once all the branch weights are computed, we emit the MD_prof /// metadata on BB using the computed values for each of its branches. ///