1 //===- SampleProfile.cpp - Incorporate sample profiles into the IR --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the SampleProfileLoader transformation. This pass
11 // reads a profile file generated by a sampling profiler (e.g. Linux Perf -
12 // http://perf.wiki.kernel.org/) and generates IR metadata to reflect the
13 // profile information in the given profile.
15 // This pass generates branch weight annotations on the IR:
17 // - prof: Represents branch weights. This annotation is added to branches
18 // to indicate the weights of each edge coming out of the branch.
19 // The weight of each edge is the weight of the target block for
20 // that edge. The weight of a block B is computed as the maximum
21 // number of samples found in B.
23 //===----------------------------------------------------------------------===//
25 #include "llvm/ADT/DenseMap.h"
26 #include "llvm/ADT/SmallPtrSet.h"
27 #include "llvm/ADT/SmallSet.h"
28 #include "llvm/ADT/StringRef.h"
29 #include "llvm/Analysis/LoopInfo.h"
30 #include "llvm/Analysis/PostDominators.h"
31 #include "llvm/IR/Constants.h"
32 #include "llvm/IR/DebugInfo.h"
33 #include "llvm/IR/DiagnosticInfo.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/InstIterator.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/LLVMContext.h"
39 #include "llvm/IR/MDBuilder.h"
40 #include "llvm/IR/Metadata.h"
41 #include "llvm/IR/Module.h"
42 #include "llvm/Pass.h"
43 #include "llvm/ProfileData/SampleProfReader.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/ErrorOr.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/Transforms/IPO.h"
49 #include "llvm/Transforms/Utils/Cloning.h"
53 using namespace sampleprof;
55 #define DEBUG_TYPE "sample-profile"
57 // Command line option to specify the file to read samples from. This is
58 // mainly used for debugging.
59 static cl::opt<std::string> SampleProfileFile(
60 "sample-profile-file", cl::init(""), cl::value_desc("filename"),
61 cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
62 static cl::opt<unsigned> SampleProfileMaxPropagateIterations(
63 "sample-profile-max-propagate-iterations", cl::init(100),
64 cl::desc("Maximum number of iterations to go through when propagating "
65 "sample block/edge weights through the CFG."));
66 static cl::opt<unsigned> SampleProfileCoverage(
67 "sample-profile-check-coverage", cl::init(0), cl::value_desc("N"),
68 cl::desc("Emit a warning if less than N% of samples in the input profile "
69 "are matched to the IR."));
72 typedef DenseMap<const BasicBlock *, uint64_t> BlockWeightMap;
73 typedef DenseMap<const BasicBlock *, const BasicBlock *> EquivalenceClassMap;
74 typedef std::pair<const BasicBlock *, const BasicBlock *> Edge;
75 typedef DenseMap<Edge, uint64_t> EdgeWeightMap;
76 typedef DenseMap<const BasicBlock *, SmallVector<const BasicBlock *, 8>>
79 /// \brief Sample profile pass.
81 /// This pass reads profile data from the file specified by
82 /// -sample-profile-file and annotates every affected function with the
83 /// profile information found in that file.
84 class SampleProfileLoader : public ModulePass {
86 // Class identification, replacement for typeinfo
89 SampleProfileLoader(StringRef Name = SampleProfileFile)
90 : ModulePass(ID), DT(nullptr), PDT(nullptr), LI(nullptr), Reader(),
91 Samples(nullptr), Filename(Name), ProfileIsValid(false) {
92 initializeSampleProfileLoaderPass(*PassRegistry::getPassRegistry());
95 bool doInitialization(Module &M) override;
97 void dump() { Reader->dump(); }
99 const char *getPassName() const override { return "Sample profile pass"; }
101 bool runOnModule(Module &M) override;
103 void getAnalysisUsage(AnalysisUsage &AU) const override {
104 AU.setPreservesCFG();
108 bool runOnFunction(Function &F);
109 unsigned getFunctionLoc(Function &F);
110 bool emitAnnotations(Function &F);
111 ErrorOr<uint64_t> getInstWeight(const Instruction &I) const;
112 ErrorOr<uint64_t> getBlockWeight(const BasicBlock *BB) const;
113 const FunctionSamples *findCalleeFunctionSamples(const CallInst &I) const;
114 const FunctionSamples *findFunctionSamples(const Instruction &I) const;
115 bool inlineHotFunctions(Function &F);
116 void printEdgeWeight(raw_ostream &OS, Edge E);
117 void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const;
118 void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB);
119 bool computeBlockWeights(Function &F);
120 void findEquivalenceClasses(Function &F);
121 void findEquivalencesFor(BasicBlock *BB1,
122 SmallVector<BasicBlock *, 8> Descendants,
123 DominatorTreeBase<BasicBlock> *DomTree);
124 void propagateWeights(Function &F);
125 uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
126 void buildEdges(Function &F);
127 bool propagateThroughEdges(Function &F);
128 void computeDominanceAndLoopInfo(Function &F);
129 unsigned getOffset(unsigned L, unsigned H) const;
130 void clearFunctionData();
132 /// \brief Map basic blocks to their computed weights.
134 /// The weight of a basic block is defined to be the maximum
135 /// of all the instruction weights in that block.
136 BlockWeightMap BlockWeights;
138 /// \brief Map edges to their computed weights.
140 /// Edge weights are computed by propagating basic block weights in
141 /// SampleProfile::propagateWeights.
142 EdgeWeightMap EdgeWeights;
144 /// \brief Set of visited blocks during propagation.
145 SmallPtrSet<const BasicBlock *, 128> VisitedBlocks;
147 /// \brief Set of visited edges during propagation.
148 SmallSet<Edge, 128> VisitedEdges;
150 /// \brief Equivalence classes for block weights.
152 /// Two blocks BB1 and BB2 are in the same equivalence class if they
153 /// dominate and post-dominate each other, and they are in the same loop
154 /// nest. When this happens, the two blocks are guaranteed to execute
155 /// the same number of times.
156 EquivalenceClassMap EquivalenceClass;
158 /// \brief Dominance, post-dominance and loop information.
159 std::unique_ptr<DominatorTree> DT;
160 std::unique_ptr<DominatorTreeBase<BasicBlock>> PDT;
161 std::unique_ptr<LoopInfo> LI;
163 /// \brief Predecessors for each basic block in the CFG.
164 BlockEdgeMap Predecessors;
166 /// \brief Successors for each basic block in the CFG.
167 BlockEdgeMap Successors;
169 /// \brief Profile reader object.
170 std::unique_ptr<SampleProfileReader> Reader;
172 /// \brief Samples collected for the body of this function.
173 FunctionSamples *Samples;
175 /// \brief Name of the profile file to load.
178 /// \brief Flag indicating whether the profile input loaded successfully.
182 class SampleCoverageTracker {
184 SampleCoverageTracker() : SampleCoverage() {}
186 bool markSamplesUsed(const FunctionSamples *Samples, uint32_t LineOffset,
187 uint32_t Discriminator);
188 unsigned computeCoverage(unsigned Used, unsigned Total) const;
189 unsigned countUsedSamples(const FunctionSamples *Samples) const;
190 unsigned countBodySamples(const FunctionSamples *Samples) const;
193 typedef DenseMap<LineLocation, unsigned> BodySampleCoverageMap;
194 typedef DenseMap<const FunctionSamples *, BodySampleCoverageMap>
195 FunctionSamplesCoverageMap;
197 /// Coverage map for sampling records.
199 /// This map keeps a record of sampling records that have been matched to
200 /// an IR instruction. This is used to detect some form of staleness in
201 /// profiles (see flag -sample-profile-check-coverage).
203 /// Each entry in the map corresponds to a FunctionSamples instance. This is
204 /// another map that counts how many times the sample record at the
205 /// given location has been used.
206 FunctionSamplesCoverageMap SampleCoverage;
209 SampleCoverageTracker CoverageTracker;
212 /// Mark as used the sample record for the given function samples at
213 /// (LineOffset, Discriminator).
215 /// \returns true if this is the first time we mark the given record.
216 bool SampleCoverageTracker::markSamplesUsed(const FunctionSamples *Samples,
218 uint32_t Discriminator) {
219 LineLocation Loc(LineOffset, Discriminator);
220 unsigned &Count = SampleCoverage[Samples][Loc];
224 /// Return the number of sample records that were applied from this profile.
226 SampleCoverageTracker::countUsedSamples(const FunctionSamples *Samples) const {
227 auto I = SampleCoverage.find(Samples);
228 unsigned Count = (I != SampleCoverage.end()) ? I->second.size() : 0;
229 for (const auto &I : Samples->getCallsiteSamples())
230 Count += countUsedSamples(&I.second);
234 /// Return the number of sample records in the body of this profile.
236 /// The count includes all the samples in inlined callees.
238 SampleCoverageTracker::countBodySamples(const FunctionSamples *Samples) const {
239 unsigned Count = Samples->getBodySamples().size();
240 for (const auto &I : Samples->getCallsiteSamples())
241 Count += countBodySamples(&I.second);
245 /// Return the fraction of sample records used in this profile.
247 /// The returned value is an unsigned integer in the range 0-100 indicating
248 /// the percentage of sample records that were used while applying this
249 /// profile to the associated function.
250 unsigned SampleCoverageTracker::computeCoverage(unsigned Used,
251 unsigned Total) const {
252 assert(Used <= Total &&
253 "number of used records cannot exceed the total number of records");
254 return Total > 0 ? Used * 100 / Total : 100;
257 /// Clear all the per-function data used to load samples and propagate weights.
258 void SampleProfileLoader::clearFunctionData() {
259 BlockWeights.clear();
261 VisitedBlocks.clear();
262 VisitedEdges.clear();
263 EquivalenceClass.clear();
267 Predecessors.clear();
271 /// \brief Returns the offset of lineno \p L to head_lineno \p H
274 /// \param H Header lineno of the function
276 /// \returns offset to the header lineno. 16 bits are used to represent offset.
277 /// We assume that a single function will not exceed 65535 LOC.
278 unsigned SampleProfileLoader::getOffset(unsigned L, unsigned H) const {
279 return (L - H) & 0xffff;
282 /// \brief Print the weight of edge \p E on stream \p OS.
284 /// \param OS Stream to emit the output to.
285 /// \param E Edge to print.
286 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
287 OS << "weight[" << E.first->getName() << "->" << E.second->getName()
288 << "]: " << EdgeWeights[E] << "\n";
291 /// \brief Print the equivalence class of block \p BB on stream \p OS.
293 /// \param OS Stream to emit the output to.
294 /// \param BB Block to print.
295 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
296 const BasicBlock *BB) {
297 const BasicBlock *Equiv = EquivalenceClass[BB];
298 OS << "equivalence[" << BB->getName()
299 << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
302 /// \brief Print the weight of block \p BB on stream \p OS.
304 /// \param OS Stream to emit the output to.
305 /// \param BB Block to print.
306 void SampleProfileLoader::printBlockWeight(raw_ostream &OS,
307 const BasicBlock *BB) const {
308 const auto &I = BlockWeights.find(BB);
309 uint64_t W = (I == BlockWeights.end() ? 0 : I->second);
310 OS << "weight[" << BB->getName() << "]: " << W << "\n";
313 /// \brief Get the weight for an instruction.
315 /// The "weight" of an instruction \p Inst is the number of samples
316 /// collected on that instruction at runtime. To retrieve it, we
317 /// need to compute the line number of \p Inst relative to the start of its
318 /// function. We use HeaderLineno to compute the offset. We then
319 /// look up the samples collected for \p Inst using BodySamples.
321 /// \param Inst Instruction to query.
323 /// \returns the weight of \p Inst.
325 SampleProfileLoader::getInstWeight(const Instruction &Inst) const {
326 DebugLoc DLoc = Inst.getDebugLoc();
328 return std::error_code();
330 const FunctionSamples *FS = findFunctionSamples(Inst);
332 return std::error_code();
334 const DILocation *DIL = DLoc;
335 unsigned Lineno = DLoc.getLine();
336 unsigned HeaderLineno = DIL->getScope()->getSubprogram()->getLine();
338 uint32_t LineOffset = getOffset(Lineno, HeaderLineno);
339 uint32_t Discriminator = DIL->getDiscriminator();
340 ErrorOr<uint64_t> R = FS->findSamplesAt(LineOffset, Discriminator);
343 CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator);
345 const Function *F = Inst.getParent()->getParent();
346 LLVMContext &Ctx = F->getContext();
347 emitOptimizationRemark(Ctx, DEBUG_TYPE, *F, DLoc,
348 Twine("Applied ") + Twine(*R) +
349 " samples from profile");
351 DEBUG(dbgs() << " " << Lineno << "." << DIL->getDiscriminator() << ":"
352 << Inst << " (line offset: " << Lineno - HeaderLineno << "."
353 << DIL->getDiscriminator() << " - weight: " << R.get()
359 /// \brief Compute the weight of a basic block.
361 /// The weight of basic block \p BB is the maximum weight of all the
362 /// instructions in BB.
364 /// \param BB The basic block to query.
366 /// \returns the weight for \p BB.
368 SampleProfileLoader::getBlockWeight(const BasicBlock *BB) const {
371 for (auto &I : BB->getInstList()) {
372 const ErrorOr<uint64_t> &R = getInstWeight(I);
373 if (R && R.get() >= Weight) {
381 return std::error_code();
384 /// \brief Compute and store the weights of every basic block.
386 /// This populates the BlockWeights map by computing
387 /// the weights of every basic block in the CFG.
389 /// \param F The function to query.
390 bool SampleProfileLoader::computeBlockWeights(Function &F) {
391 bool Changed = false;
392 DEBUG(dbgs() << "Block weights\n");
393 for (const auto &BB : F) {
394 ErrorOr<uint64_t> Weight = getBlockWeight(&BB);
396 BlockWeights[&BB] = Weight.get();
397 VisitedBlocks.insert(&BB);
400 DEBUG(printBlockWeight(dbgs(), &BB));
406 /// \brief Get the FunctionSamples for a call instruction.
408 /// The FunctionSamples of a call instruction \p Inst is the inlined
409 /// instance in which that call instruction is calling to. It contains
410 /// all samples that resides in the inlined instance. We first find the
411 /// inlined instance in which the call instruction is from, then we
412 /// traverse its children to find the callsite with the matching
413 /// location and callee function name.
415 /// \param Inst Call instruction to query.
417 /// \returns The FunctionSamples pointer to the inlined instance.
418 const FunctionSamples *
419 SampleProfileLoader::findCalleeFunctionSamples(const CallInst &Inst) const {
420 const DILocation *DIL = Inst.getDebugLoc();
424 DISubprogram *SP = DIL->getScope()->getSubprogram();
428 Function *CalleeFunc = Inst.getCalledFunction();
433 StringRef CalleeName = CalleeFunc->getName();
434 const FunctionSamples *FS = findFunctionSamples(Inst);
438 return FS->findFunctionSamplesAt(
439 CallsiteLocation(getOffset(DIL->getLine(), SP->getLine()),
440 DIL->getDiscriminator(), CalleeName));
443 /// \brief Get the FunctionSamples for an instruction.
445 /// The FunctionSamples of an instruction \p Inst is the inlined instance
446 /// in which that instruction is coming from. We traverse the inline stack
447 /// of that instruction, and match it with the tree nodes in the profile.
449 /// \param Inst Instruction to query.
451 /// \returns the FunctionSamples pointer to the inlined instance.
452 const FunctionSamples *
453 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
454 SmallVector<CallsiteLocation, 10> S;
455 const DILocation *DIL = Inst.getDebugLoc();
459 StringRef CalleeName;
460 for (const DILocation *DIL = Inst.getDebugLoc(); DIL;
461 DIL = DIL->getInlinedAt()) {
462 DISubprogram *SP = DIL->getScope()->getSubprogram();
465 if (!CalleeName.empty()) {
466 S.push_back(CallsiteLocation(getOffset(DIL->getLine(), SP->getLine()),
467 DIL->getDiscriminator(), CalleeName));
469 CalleeName = SP->getLinkageName();
473 const FunctionSamples *FS = Samples;
474 for (int i = S.size() - 1; i >= 0 && FS != nullptr; i--) {
475 FS = FS->findFunctionSamplesAt(S[i]);
480 /// \brief Iteratively inline hot callsites of a function.
482 /// Iteratively traverse all callsites of the function \p F, and find if
483 /// the corresponding inlined instance exists and is hot in profile. If
484 /// it is hot enough, inline the callsites and adds new callsites of the
485 /// callee into the caller.
487 /// TODO: investigate the possibility of not invoking InlineFunction directly.
489 /// \param F function to perform iterative inlining.
491 /// \returns True if there is any inline happened.
492 bool SampleProfileLoader::inlineHotFunctions(Function &F) {
493 bool Changed = false;
494 LLVMContext &Ctx = F.getContext();
496 bool LocalChanged = false;
497 SmallVector<CallInst *, 10> CIS;
499 for (auto &I : BB.getInstList()) {
500 CallInst *CI = dyn_cast<CallInst>(&I);
502 const FunctionSamples *FS = findCalleeFunctionSamples(*CI);
503 if (FS && FS->getTotalSamples() > 0) {
509 for (auto CI : CIS) {
510 InlineFunctionInfo IFI;
511 Function *CalledFunction = CI->getCalledFunction();
512 DebugLoc DLoc = CI->getDebugLoc();
513 uint64_t NumSamples = findCalleeFunctionSamples(*CI)->getTotalSamples();
514 if (InlineFunction(CI, IFI)) {
516 emitOptimizationRemark(Ctx, DEBUG_TYPE, F, DLoc,
517 Twine("inlined hot callee '") +
518 CalledFunction->getName() + "' with " +
519 Twine(NumSamples) + " samples into '" +
532 /// \brief Find equivalence classes for the given block.
534 /// This finds all the blocks that are guaranteed to execute the same
535 /// number of times as \p BB1. To do this, it traverses all the
536 /// descendants of \p BB1 in the dominator or post-dominator tree.
538 /// A block BB2 will be in the same equivalence class as \p BB1 if
539 /// the following holds:
541 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
542 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
543 /// dominate BB1 in the post-dominator tree.
545 /// 2- Both BB2 and \p BB1 must be in the same loop.
547 /// For every block BB2 that meets those two requirements, we set BB2's
548 /// equivalence class to \p BB1.
550 /// \param BB1 Block to check.
551 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
552 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
553 /// with blocks from \p BB1's dominator tree, then
554 /// this is the post-dominator tree, and vice versa.
555 void SampleProfileLoader::findEquivalencesFor(
556 BasicBlock *BB1, SmallVector<BasicBlock *, 8> Descendants,
557 DominatorTreeBase<BasicBlock> *DomTree) {
558 const BasicBlock *EC = EquivalenceClass[BB1];
559 uint64_t Weight = BlockWeights[EC];
560 for (const auto *BB2 : Descendants) {
561 bool IsDomParent = DomTree->dominates(BB2, BB1);
562 bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
563 if (BB1 != BB2 && IsDomParent && IsInSameLoop) {
564 EquivalenceClass[BB2] = EC;
566 // If BB2 is heavier than BB1, make BB2 have the same weight
569 // Note that we don't worry about the opposite situation here
570 // (when BB2 is lighter than BB1). We will deal with this
571 // during the propagation phase. Right now, we just want to
572 // make sure that BB1 has the largest weight of all the
573 // members of its equivalence set.
574 Weight = std::max(Weight, BlockWeights[BB2]);
577 BlockWeights[EC] = Weight;
580 /// \brief Find equivalence classes.
582 /// Since samples may be missing from blocks, we can fill in the gaps by setting
583 /// the weights of all the blocks in the same equivalence class to the same
584 /// weight. To compute the concept of equivalence, we use dominance and loop
585 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
586 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
588 /// \param F The function to query.
589 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
590 SmallVector<BasicBlock *, 8> DominatedBBs;
591 DEBUG(dbgs() << "\nBlock equivalence classes\n");
592 // Find equivalence sets based on dominance and post-dominance information.
594 BasicBlock *BB1 = &BB;
596 // Compute BB1's equivalence class once.
597 if (EquivalenceClass.count(BB1)) {
598 DEBUG(printBlockEquivalence(dbgs(), BB1));
602 // By default, blocks are in their own equivalence class.
603 EquivalenceClass[BB1] = BB1;
605 // Traverse all the blocks dominated by BB1. We are looking for
606 // every basic block BB2 such that:
608 // 1- BB1 dominates BB2.
609 // 2- BB2 post-dominates BB1.
610 // 3- BB1 and BB2 are in the same loop nest.
612 // If all those conditions hold, it means that BB2 is executed
613 // as many times as BB1, so they are placed in the same equivalence
614 // class by making BB2's equivalence class be BB1.
615 DominatedBBs.clear();
616 DT->getDescendants(BB1, DominatedBBs);
617 findEquivalencesFor(BB1, DominatedBBs, PDT.get());
619 DEBUG(printBlockEquivalence(dbgs(), BB1));
622 // Assign weights to equivalence classes.
624 // All the basic blocks in the same equivalence class will execute
625 // the same number of times. Since we know that the head block in
626 // each equivalence class has the largest weight, assign that weight
627 // to all the blocks in that equivalence class.
628 DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n");
630 const BasicBlock *BB = &BI;
631 const BasicBlock *EquivBB = EquivalenceClass[BB];
633 BlockWeights[BB] = BlockWeights[EquivBB];
634 DEBUG(printBlockWeight(dbgs(), BB));
638 /// \brief Visit the given edge to decide if it has a valid weight.
640 /// If \p E has not been visited before, we copy to \p UnknownEdge
641 /// and increment the count of unknown edges.
643 /// \param E Edge to visit.
644 /// \param NumUnknownEdges Current number of unknown edges.
645 /// \param UnknownEdge Set if E has not been visited before.
647 /// \returns E's weight, if known. Otherwise, return 0.
648 uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
650 if (!VisitedEdges.count(E)) {
651 (*NumUnknownEdges)++;
656 return EdgeWeights[E];
659 /// \brief Propagate weights through incoming/outgoing edges.
661 /// If the weight of a basic block is known, and there is only one edge
662 /// with an unknown weight, we can calculate the weight of that edge.
664 /// Similarly, if all the edges have a known count, we can calculate the
665 /// count of the basic block, if needed.
667 /// \param F Function to process.
669 /// \returns True if new weights were assigned to edges or blocks.
670 bool SampleProfileLoader::propagateThroughEdges(Function &F) {
671 bool Changed = false;
672 DEBUG(dbgs() << "\nPropagation through edges\n");
673 for (const auto &BI : F) {
674 const BasicBlock *BB = &BI;
675 const BasicBlock *EC = EquivalenceClass[BB];
677 // Visit all the predecessor and successor edges to determine
678 // which ones have a weight assigned already. Note that it doesn't
679 // matter that we only keep track of a single unknown edge. The
680 // only case we are interested in handling is when only a single
681 // edge is unknown (see setEdgeOrBlockWeight).
682 for (unsigned i = 0; i < 2; i++) {
683 uint64_t TotalWeight = 0;
684 unsigned NumUnknownEdges = 0;
685 Edge UnknownEdge, SelfReferentialEdge;
688 // First, visit all predecessor edges.
689 for (auto *Pred : Predecessors[BB]) {
690 Edge E = std::make_pair(Pred, BB);
691 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
692 if (E.first == E.second)
693 SelfReferentialEdge = E;
696 // On the second round, visit all successor edges.
697 for (auto *Succ : Successors[BB]) {
698 Edge E = std::make_pair(BB, Succ);
699 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
703 // After visiting all the edges, there are three cases that we
704 // can handle immediately:
706 // - All the edge weights are known (i.e., NumUnknownEdges == 0).
707 // In this case, we simply check that the sum of all the edges
708 // is the same as BB's weight. If not, we change BB's weight
709 // to match. Additionally, if BB had not been visited before,
710 // we mark it visited.
712 // - Only one edge is unknown and BB has already been visited.
713 // In this case, we can compute the weight of the edge by
714 // subtracting the total block weight from all the known
715 // edge weights. If the edges weight more than BB, then the
716 // edge of the last remaining edge is set to zero.
718 // - There exists a self-referential edge and the weight of BB is
719 // known. In this case, this edge can be based on BB's weight.
720 // We add up all the other known edges and set the weight on
721 // the self-referential edge as we did in the previous case.
723 // In any other case, we must continue iterating. Eventually,
724 // all edges will get a weight, or iteration will stop when
725 // it reaches SampleProfileMaxPropagateIterations.
726 if (NumUnknownEdges <= 1) {
727 uint64_t &BBWeight = BlockWeights[EC];
728 if (NumUnknownEdges == 0) {
729 // If we already know the weight of all edges, the weight of the
730 // basic block can be computed. It should be no larger than the sum
731 // of all edge weights.
732 if (TotalWeight > BBWeight) {
733 BBWeight = TotalWeight;
735 DEBUG(dbgs() << "All edge weights for " << BB->getName()
736 << " known. Set weight for block: ";
737 printBlockWeight(dbgs(), BB););
739 if (VisitedBlocks.insert(EC).second)
741 } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) {
742 // If there is a single unknown edge and the block has been
743 // visited, then we can compute E's weight.
744 if (BBWeight >= TotalWeight)
745 EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
747 EdgeWeights[UnknownEdge] = 0;
748 VisitedEdges.insert(UnknownEdge);
750 DEBUG(dbgs() << "Set weight for edge: ";
751 printEdgeWeight(dbgs(), UnknownEdge));
753 } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) {
754 uint64_t &BBWeight = BlockWeights[BB];
755 // We have a self-referential edge and the weight of BB is known.
756 if (BBWeight >= TotalWeight)
757 EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
759 EdgeWeights[SelfReferentialEdge] = 0;
760 VisitedEdges.insert(SelfReferentialEdge);
762 DEBUG(dbgs() << "Set self-referential edge weight to: ";
763 printEdgeWeight(dbgs(), SelfReferentialEdge));
771 /// \brief Build in/out edge lists for each basic block in the CFG.
773 /// We are interested in unique edges. If a block B1 has multiple
774 /// edges to another block B2, we only add a single B1->B2 edge.
775 void SampleProfileLoader::buildEdges(Function &F) {
777 BasicBlock *B1 = &BI;
779 // Add predecessors for B1.
780 SmallPtrSet<BasicBlock *, 16> Visited;
781 if (!Predecessors[B1].empty())
782 llvm_unreachable("Found a stale predecessors list in a basic block.");
783 for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
784 BasicBlock *B2 = *PI;
785 if (Visited.insert(B2).second)
786 Predecessors[B1].push_back(B2);
789 // Add successors for B1.
791 if (!Successors[B1].empty())
792 llvm_unreachable("Found a stale successors list in a basic block.");
793 for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
794 BasicBlock *B2 = *SI;
795 if (Visited.insert(B2).second)
796 Successors[B1].push_back(B2);
801 /// \brief Propagate weights into edges
803 /// The following rules are applied to every block BB in the CFG:
805 /// - If BB has a single predecessor/successor, then the weight
806 /// of that edge is the weight of the block.
808 /// - If all incoming or outgoing edges are known except one, and the
809 /// weight of the block is already known, the weight of the unknown
810 /// edge will be the weight of the block minus the sum of all the known
811 /// edges. If the sum of all the known edges is larger than BB's weight,
812 /// we set the unknown edge weight to zero.
814 /// - If there is a self-referential edge, and the weight of the block is
815 /// known, the weight for that edge is set to the weight of the block
816 /// minus the weight of the other incoming edges to that block (if
818 void SampleProfileLoader::propagateWeights(Function &F) {
822 // Add an entry count to the function using the samples gathered
823 // at the function entry.
824 F.setEntryCount(Samples->getHeadSamples());
826 // Before propagation starts, build, for each block, a list of
827 // unique predecessors and successors. This is necessary to handle
828 // identical edges in multiway branches. Since we visit all blocks and all
829 // edges of the CFG, it is cleaner to build these lists once at the start
833 // Propagate until we converge or we go past the iteration limit.
834 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
835 Changed = propagateThroughEdges(F);
838 // Generate MD_prof metadata for every branch instruction using the
839 // edge weights computed during propagation.
840 DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
841 LLVMContext &Ctx = F.getContext();
844 BasicBlock *BB = &BI;
845 TerminatorInst *TI = BB->getTerminator();
846 if (TI->getNumSuccessors() == 1)
848 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
851 DEBUG(dbgs() << "\nGetting weights for branch at line "
852 << TI->getDebugLoc().getLine() << ".\n");
853 SmallVector<uint32_t, 4> Weights;
854 uint32_t MaxWeight = 0;
856 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
857 BasicBlock *Succ = TI->getSuccessor(I);
858 Edge E = std::make_pair(BB, Succ);
859 uint64_t Weight = EdgeWeights[E];
860 DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
861 // Use uint32_t saturated arithmetic to adjust the incoming weights,
862 // if needed. Sample counts in profiles are 64-bit unsigned values,
863 // but internally branch weights are expressed as 32-bit values.
864 if (Weight > std::numeric_limits<uint32_t>::max()) {
865 DEBUG(dbgs() << " (saturated due to uint32_t overflow)");
866 Weight = std::numeric_limits<uint32_t>::max();
868 Weights.push_back(static_cast<uint32_t>(Weight));
870 if (Weight > MaxWeight) {
872 MaxDestLoc = Succ->getFirstNonPHIOrDbgOrLifetime()->getDebugLoc();
877 // Only set weights if there is at least one non-zero weight.
878 // In any other case, let the analyzer set weights.
880 DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
881 TI->setMetadata(llvm::LLVMContext::MD_prof,
882 MDB.createBranchWeights(Weights));
883 DebugLoc BranchLoc = TI->getDebugLoc();
884 emitOptimizationRemark(
885 Ctx, DEBUG_TYPE, F, MaxDestLoc,
886 Twine("most popular destination for conditional branches at ") +
887 ((BranchLoc) ? Twine(BranchLoc->getFilename() + ":" +
888 Twine(BranchLoc.getLine()) + ":" +
889 Twine(BranchLoc.getCol()))
890 : Twine("<UNKNOWN LOCATION>")));
892 DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
897 /// \brief Get the line number for the function header.
899 /// This looks up function \p F in the current compilation unit and
900 /// retrieves the line number where the function is defined. This is
901 /// line 0 for all the samples read from the profile file. Every line
902 /// number is relative to this line.
904 /// \param F Function object to query.
906 /// \returns the line number where \p F is defined. If it returns 0,
907 /// it means that there is no debug information available for \p F.
908 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
909 if (DISubprogram *S = getDISubprogram(&F))
912 // If the start of \p F is missing, emit a diagnostic to inform the user
913 // about the missed opportunity.
914 F.getContext().diagnose(DiagnosticInfoSampleProfile(
915 "No debug information found in function " + F.getName() +
916 ": Function profile not used",
921 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
922 DT.reset(new DominatorTree);
925 PDT.reset(new DominatorTreeBase<BasicBlock>(true));
928 LI.reset(new LoopInfo);
932 /// \brief Generate branch weight metadata for all branches in \p F.
934 /// Branch weights are computed out of instruction samples using a
935 /// propagation heuristic. Propagation proceeds in 3 phases:
937 /// 1- Assignment of block weights. All the basic blocks in the function
938 /// are initial assigned the same weight as their most frequently
939 /// executed instruction.
941 /// 2- Creation of equivalence classes. Since samples may be missing from
942 /// blocks, we can fill in the gaps by setting the weights of all the
943 /// blocks in the same equivalence class to the same weight. To compute
944 /// the concept of equivalence, we use dominance and loop information.
945 /// Two blocks B1 and B2 are in the same equivalence class if B1
946 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
948 /// 3- Propagation of block weights into edges. This uses a simple
949 /// propagation heuristic. The following rules are applied to every
950 /// block BB in the CFG:
952 /// - If BB has a single predecessor/successor, then the weight
953 /// of that edge is the weight of the block.
955 /// - If all the edges are known except one, and the weight of the
956 /// block is already known, the weight of the unknown edge will
957 /// be the weight of the block minus the sum of all the known
958 /// edges. If the sum of all the known edges is larger than BB's weight,
959 /// we set the unknown edge weight to zero.
961 /// - If there is a self-referential edge, and the weight of the block is
962 /// known, the weight for that edge is set to the weight of the block
963 /// minus the weight of the other incoming edges to that block (if
966 /// Since this propagation is not guaranteed to finalize for every CFG, we
967 /// only allow it to proceed for a limited number of iterations (controlled
968 /// by -sample-profile-max-propagate-iterations).
970 /// FIXME: Try to replace this propagation heuristic with a scheme
971 /// that is guaranteed to finalize. A work-list approach similar to
972 /// the standard value propagation algorithm used by SSA-CCP might
975 /// Once all the branch weights are computed, we emit the MD_prof
976 /// metadata on BB using the computed values for each of its branches.
978 /// \param F The function to query.
980 /// \returns true if \p F was modified. Returns false, otherwise.
981 bool SampleProfileLoader::emitAnnotations(Function &F) {
982 bool Changed = false;
984 if (getFunctionLoc(F) == 0)
987 DEBUG(dbgs() << "Line number for the first instruction in " << F.getName()
988 << ": " << getFunctionLoc(F) << "\n");
990 Changed |= inlineHotFunctions(F);
992 // Compute basic block weights.
993 Changed |= computeBlockWeights(F);
996 // Compute dominance and loop info needed for propagation.
997 computeDominanceAndLoopInfo(F);
999 // Find equivalence classes.
1000 findEquivalenceClasses(F);
1002 // Propagate weights to all edges.
1003 propagateWeights(F);
1006 // If coverage checking was requested, compute it now.
1007 if (SampleProfileCoverage) {
1008 unsigned Used = CoverageTracker.countUsedSamples(Samples);
1009 unsigned Total = CoverageTracker.countBodySamples(Samples);
1010 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1011 if (Coverage < SampleProfileCoverage) {
1012 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1013 getDISubprogram(&F)->getFilename(), getFunctionLoc(F),
1014 Twine(Used) + " of " + Twine(Total) + " available profile records (" +
1015 Twine(Coverage) + "%) were applied",
1023 char SampleProfileLoader::ID = 0;
1024 INITIALIZE_PASS_BEGIN(SampleProfileLoader, "sample-profile",
1025 "Sample Profile loader", false, false)
1026 INITIALIZE_PASS_DEPENDENCY(AddDiscriminators)
1027 INITIALIZE_PASS_END(SampleProfileLoader, "sample-profile",
1028 "Sample Profile loader", false, false)
1030 bool SampleProfileLoader::doInitialization(Module &M) {
1031 auto &Ctx = M.getContext();
1032 auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
1033 if (std::error_code EC = ReaderOrErr.getError()) {
1034 std::string Msg = "Could not open profile: " + EC.message();
1035 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg));
1038 Reader = std::move(ReaderOrErr.get());
1039 ProfileIsValid = (Reader->read() == sampleprof_error::success);
1043 ModulePass *llvm::createSampleProfileLoaderPass() {
1044 return new SampleProfileLoader(SampleProfileFile);
1047 ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) {
1048 return new SampleProfileLoader(Name);
1051 bool SampleProfileLoader::runOnModule(Module &M) {
1052 if (!ProfileIsValid)
1055 bool retval = false;
1057 if (!F.isDeclaration()) {
1058 clearFunctionData();
1059 retval |= runOnFunction(F);
1064 bool SampleProfileLoader::runOnFunction(Function &F) {
1065 Samples = Reader->getSamplesFor(F);
1066 if (!Samples->empty())
1067 return emitAnnotations(F);