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/Transforms/SampleProfile.h"
26 #include "llvm/ADT/DenseMap.h"
27 #include "llvm/ADT/SmallPtrSet.h"
28 #include "llvm/ADT/SmallSet.h"
29 #include "llvm/ADT/StringRef.h"
30 #include "llvm/Analysis/AssumptionCache.h"
31 #include "llvm/Analysis/LoopInfo.h"
32 #include "llvm/Analysis/PostDominators.h"
33 #include "llvm/IR/Constants.h"
34 #include "llvm/IR/DebugInfo.h"
35 #include "llvm/IR/DiagnosticInfo.h"
36 #include "llvm/IR/Dominators.h"
37 #include "llvm/IR/Function.h"
38 #include "llvm/IR/InstIterator.h"
39 #include "llvm/IR/Instructions.h"
40 #include "llvm/IR/IntrinsicInst.h"
41 #include "llvm/IR/LLVMContext.h"
42 #include "llvm/IR/MDBuilder.h"
43 #include "llvm/IR/Metadata.h"
44 #include "llvm/IR/Module.h"
45 #include "llvm/Pass.h"
46 #include "llvm/ProfileData/SampleProfReader.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/ErrorOr.h"
50 #include "llvm/Support/Format.h"
51 #include "llvm/Support/raw_ostream.h"
52 #include "llvm/Transforms/IPO.h"
53 #include "llvm/Transforms/Utils/Cloning.h"
57 using namespace sampleprof;
59 #define DEBUG_TYPE "sample-profile"
61 // Command line option to specify the file to read samples from. This is
62 // mainly used for debugging.
63 static cl::opt<std::string> SampleProfileFile(
64 "sample-profile-file", cl::init(""), cl::value_desc("filename"),
65 cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
66 static cl::opt<unsigned> SampleProfileMaxPropagateIterations(
67 "sample-profile-max-propagate-iterations", cl::init(100),
68 cl::desc("Maximum number of iterations to go through when propagating "
69 "sample block/edge weights through the CFG."));
70 static cl::opt<unsigned> SampleProfileRecordCoverage(
71 "sample-profile-check-record-coverage", cl::init(0), cl::value_desc("N"),
72 cl::desc("Emit a warning if less than N% of records in the input profile "
73 "are matched to the IR."));
74 static cl::opt<unsigned> SampleProfileSampleCoverage(
75 "sample-profile-check-sample-coverage", cl::init(0), cl::value_desc("N"),
76 cl::desc("Emit a warning if less than N% of samples in the input profile "
77 "are matched to the IR."));
78 static cl::opt<double> SampleProfileHotThreshold(
79 "sample-profile-inline-hot-threshold", cl::init(0.1), cl::value_desc("N"),
80 cl::desc("Inlined functions that account for more than N% of all samples "
81 "collected in the parent function, will be inlined again."));
84 typedef DenseMap<const BasicBlock *, uint64_t> BlockWeightMap;
85 typedef DenseMap<const BasicBlock *, const BasicBlock *> EquivalenceClassMap;
86 typedef std::pair<const BasicBlock *, const BasicBlock *> Edge;
87 typedef DenseMap<Edge, uint64_t> EdgeWeightMap;
88 typedef DenseMap<const BasicBlock *, SmallVector<const BasicBlock *, 8>>
91 class SampleCoverageTracker {
93 SampleCoverageTracker() : SampleCoverage(), TotalUsedSamples(0) {}
95 bool markSamplesUsed(const FunctionSamples *FS, uint32_t LineOffset,
96 uint32_t Discriminator, uint64_t Samples);
97 unsigned computeCoverage(unsigned Used, unsigned Total) const;
98 unsigned countUsedRecords(const FunctionSamples *FS) const;
99 unsigned countBodyRecords(const FunctionSamples *FS) const;
100 uint64_t getTotalUsedSamples() const { return TotalUsedSamples; }
101 uint64_t countBodySamples(const FunctionSamples *FS) const;
103 SampleCoverage.clear();
104 TotalUsedSamples = 0;
108 typedef std::map<LineLocation, unsigned> BodySampleCoverageMap;
109 typedef DenseMap<const FunctionSamples *, BodySampleCoverageMap>
110 FunctionSamplesCoverageMap;
112 /// Coverage map for sampling records.
114 /// This map keeps a record of sampling records that have been matched to
115 /// an IR instruction. This is used to detect some form of staleness in
116 /// profiles (see flag -sample-profile-check-coverage).
118 /// Each entry in the map corresponds to a FunctionSamples instance. This is
119 /// another map that counts how many times the sample record at the
120 /// given location has been used.
121 FunctionSamplesCoverageMap SampleCoverage;
123 /// Number of samples used from the profile.
125 /// When a sampling record is used for the first time, the samples from
126 /// that record are added to this accumulator. Coverage is later computed
127 /// based on the total number of samples available in this function and
130 /// Note that this accumulator tracks samples used from a single function
131 /// and all the inlined callsites. Strictly, we should have a map of counters
132 /// keyed by FunctionSamples pointers, but these stats are cleared after
133 /// every function, so we just need to keep a single counter.
134 uint64_t TotalUsedSamples;
137 /// \brief Sample profile pass.
139 /// This pass reads profile data from the file specified by
140 /// -sample-profile-file and annotates every affected function with the
141 /// profile information found in that file.
142 class SampleProfileLoader {
144 SampleProfileLoader(StringRef Name = SampleProfileFile)
145 : DT(nullptr), PDT(nullptr), LI(nullptr), ACT(nullptr), Reader(),
146 Samples(nullptr), Filename(Name), ProfileIsValid(false),
147 TotalCollectedSamples(0) {}
149 bool doInitialization(Module &M);
150 bool runOnModule(Module &M);
151 void setACT(AssumptionCacheTracker *A) { ACT = A; }
153 void dump() { Reader->dump(); }
156 bool runOnFunction(Function &F);
157 unsigned getFunctionLoc(Function &F);
158 bool emitAnnotations(Function &F);
159 ErrorOr<uint64_t> getInstWeight(const Instruction &I);
160 ErrorOr<uint64_t> getBlockWeight(const BasicBlock *BB);
161 const FunctionSamples *findCalleeFunctionSamples(const Instruction &I) const;
162 const FunctionSamples *findFunctionSamples(const Instruction &I) const;
163 bool inlineHotFunctions(Function &F);
164 void printEdgeWeight(raw_ostream &OS, Edge E);
165 void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const;
166 void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB);
167 bool computeBlockWeights(Function &F);
168 void findEquivalenceClasses(Function &F);
169 void findEquivalencesFor(BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
170 DominatorTreeBase<BasicBlock> *DomTree);
171 void propagateWeights(Function &F);
172 uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
173 void buildEdges(Function &F);
174 bool propagateThroughEdges(Function &F, bool UpdateBlockCount);
175 void computeDominanceAndLoopInfo(Function &F);
176 unsigned getOffset(unsigned L, unsigned H) const;
177 void clearFunctionData();
179 /// \brief Map basic blocks to their computed weights.
181 /// The weight of a basic block is defined to be the maximum
182 /// of all the instruction weights in that block.
183 BlockWeightMap BlockWeights;
185 /// \brief Map edges to their computed weights.
187 /// Edge weights are computed by propagating basic block weights in
188 /// SampleProfile::propagateWeights.
189 EdgeWeightMap EdgeWeights;
191 /// \brief Set of visited blocks during propagation.
192 SmallPtrSet<const BasicBlock *, 32> VisitedBlocks;
194 /// \brief Set of visited edges during propagation.
195 SmallSet<Edge, 32> VisitedEdges;
197 /// \brief Equivalence classes for block weights.
199 /// Two blocks BB1 and BB2 are in the same equivalence class if they
200 /// dominate and post-dominate each other, and they are in the same loop
201 /// nest. When this happens, the two blocks are guaranteed to execute
202 /// the same number of times.
203 EquivalenceClassMap EquivalenceClass;
205 /// \brief Dominance, post-dominance and loop information.
206 std::unique_ptr<DominatorTree> DT;
207 std::unique_ptr<DominatorTreeBase<BasicBlock>> PDT;
208 std::unique_ptr<LoopInfo> LI;
210 AssumptionCacheTracker *ACT;
212 /// \brief Predecessors for each basic block in the CFG.
213 BlockEdgeMap Predecessors;
215 /// \brief Successors for each basic block in the CFG.
216 BlockEdgeMap Successors;
218 SampleCoverageTracker CoverageTracker;
220 /// \brief Profile reader object.
221 std::unique_ptr<SampleProfileReader> Reader;
223 /// \brief Samples collected for the body of this function.
224 FunctionSamples *Samples;
226 /// \brief Name of the profile file to load.
227 std::string Filename;
229 /// \brief Flag indicating whether the profile input loaded successfully.
232 /// \brief Total number of samples collected in this profile.
234 /// This is the sum of all the samples collected in all the functions executed
236 uint64_t TotalCollectedSamples;
239 class SampleProfileLoaderLegacyPass : public ModulePass {
241 // Class identification, replacement for typeinfo
244 SampleProfileLoaderLegacyPass(StringRef Name = SampleProfileFile)
245 : ModulePass(ID), SampleLoader(Name) {
246 initializeSampleProfileLoaderLegacyPassPass(
247 *PassRegistry::getPassRegistry());
250 void dump() { SampleLoader.dump(); }
252 bool doInitialization(Module &M) override {
253 return SampleLoader.doInitialization(M);
255 StringRef getPassName() const override { return "Sample profile pass"; }
256 bool runOnModule(Module &M) override;
258 void getAnalysisUsage(AnalysisUsage &AU) const override {
259 AU.addRequired<AssumptionCacheTracker>();
263 SampleProfileLoader SampleLoader;
266 /// Return true if the given callsite is hot wrt to its caller.
268 /// Functions that were inlined in the original binary will be represented
269 /// in the inline stack in the sample profile. If the profile shows that
270 /// the original inline decision was "good" (i.e., the callsite is executed
271 /// frequently), then we will recreate the inline decision and apply the
272 /// profile from the inlined callsite.
274 /// To decide whether an inlined callsite is hot, we compute the fraction
275 /// of samples used by the callsite with respect to the total number of samples
276 /// collected in the caller.
278 /// If that fraction is larger than the default given by
279 /// SampleProfileHotThreshold, the callsite will be inlined again.
280 bool callsiteIsHot(const FunctionSamples *CallerFS,
281 const FunctionSamples *CallsiteFS) {
283 return false; // The callsite was not inlined in the original binary.
285 uint64_t ParentTotalSamples = CallerFS->getTotalSamples();
286 if (ParentTotalSamples == 0)
287 return false; // Avoid division by zero.
289 uint64_t CallsiteTotalSamples = CallsiteFS->getTotalSamples();
290 if (CallsiteTotalSamples == 0)
291 return false; // Callsite is trivially cold.
293 double PercentSamples =
294 (double)CallsiteTotalSamples / (double)ParentTotalSamples * 100.0;
295 return PercentSamples >= SampleProfileHotThreshold;
299 /// Mark as used the sample record for the given function samples at
300 /// (LineOffset, Discriminator).
302 /// \returns true if this is the first time we mark the given record.
303 bool SampleCoverageTracker::markSamplesUsed(const FunctionSamples *FS,
305 uint32_t Discriminator,
307 LineLocation Loc(LineOffset, Discriminator);
308 unsigned &Count = SampleCoverage[FS][Loc];
309 bool FirstTime = (++Count == 1);
311 TotalUsedSamples += Samples;
315 /// Return the number of sample records that were applied from this profile.
317 /// This count does not include records from cold inlined callsites.
319 SampleCoverageTracker::countUsedRecords(const FunctionSamples *FS) const {
320 auto I = SampleCoverage.find(FS);
322 // The size of the coverage map for FS represents the number of records
323 // that were marked used at least once.
324 unsigned Count = (I != SampleCoverage.end()) ? I->second.size() : 0;
326 // If there are inlined callsites in this function, count the samples found
327 // in the respective bodies. However, do not bother counting callees with 0
328 // total samples, these are callees that were never invoked at runtime.
329 for (const auto &I : FS->getCallsiteSamples()) {
330 const FunctionSamples *CalleeSamples = &I.second;
331 if (callsiteIsHot(FS, CalleeSamples))
332 Count += countUsedRecords(CalleeSamples);
338 /// Return the number of sample records in the body of this profile.
340 /// This count does not include records from cold inlined callsites.
342 SampleCoverageTracker::countBodyRecords(const FunctionSamples *FS) const {
343 unsigned Count = FS->getBodySamples().size();
345 // Only count records in hot callsites.
346 for (const auto &I : FS->getCallsiteSamples()) {
347 const FunctionSamples *CalleeSamples = &I.second;
348 if (callsiteIsHot(FS, CalleeSamples))
349 Count += countBodyRecords(CalleeSamples);
355 /// Return the number of samples collected in the body of this profile.
357 /// This count does not include samples from cold inlined callsites.
359 SampleCoverageTracker::countBodySamples(const FunctionSamples *FS) const {
361 for (const auto &I : FS->getBodySamples())
362 Total += I.second.getSamples();
364 // Only count samples in hot callsites.
365 for (const auto &I : FS->getCallsiteSamples()) {
366 const FunctionSamples *CalleeSamples = &I.second;
367 if (callsiteIsHot(FS, CalleeSamples))
368 Total += countBodySamples(CalleeSamples);
374 /// Return the fraction of sample records used in this profile.
376 /// The returned value is an unsigned integer in the range 0-100 indicating
377 /// the percentage of sample records that were used while applying this
378 /// profile to the associated function.
379 unsigned SampleCoverageTracker::computeCoverage(unsigned Used,
380 unsigned Total) const {
381 assert(Used <= Total &&
382 "number of used records cannot exceed the total number of records");
383 return Total > 0 ? Used * 100 / Total : 100;
386 /// Clear all the per-function data used to load samples and propagate weights.
387 void SampleProfileLoader::clearFunctionData() {
388 BlockWeights.clear();
390 VisitedBlocks.clear();
391 VisitedEdges.clear();
392 EquivalenceClass.clear();
396 Predecessors.clear();
398 CoverageTracker.clear();
401 /// \brief Returns the offset of lineno \p L to head_lineno \p H
404 /// \param H Header lineno of the function
406 /// \returns offset to the header lineno. 16 bits are used to represent offset.
407 /// We assume that a single function will not exceed 65535 LOC.
408 unsigned SampleProfileLoader::getOffset(unsigned L, unsigned H) const {
409 return (L - H) & 0xffff;
412 /// \brief Print the weight of edge \p E on stream \p OS.
414 /// \param OS Stream to emit the output to.
415 /// \param E Edge to print.
416 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
417 OS << "weight[" << E.first->getName() << "->" << E.second->getName()
418 << "]: " << EdgeWeights[E] << "\n";
421 /// \brief Print the equivalence class of block \p BB on stream \p OS.
423 /// \param OS Stream to emit the output to.
424 /// \param BB Block to print.
425 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
426 const BasicBlock *BB) {
427 const BasicBlock *Equiv = EquivalenceClass[BB];
428 OS << "equivalence[" << BB->getName()
429 << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
432 /// \brief Print the weight of block \p BB on stream \p OS.
434 /// \param OS Stream to emit the output to.
435 /// \param BB Block to print.
436 void SampleProfileLoader::printBlockWeight(raw_ostream &OS,
437 const BasicBlock *BB) const {
438 const auto &I = BlockWeights.find(BB);
439 uint64_t W = (I == BlockWeights.end() ? 0 : I->second);
440 OS << "weight[" << BB->getName() << "]: " << W << "\n";
443 /// \brief Get the weight for an instruction.
445 /// The "weight" of an instruction \p Inst is the number of samples
446 /// collected on that instruction at runtime. To retrieve it, we
447 /// need to compute the line number of \p Inst relative to the start of its
448 /// function. We use HeaderLineno to compute the offset. We then
449 /// look up the samples collected for \p Inst using BodySamples.
451 /// \param Inst Instruction to query.
453 /// \returns the weight of \p Inst.
455 SampleProfileLoader::getInstWeight(const Instruction &Inst) {
456 const DebugLoc &DLoc = Inst.getDebugLoc();
458 return std::error_code();
460 const FunctionSamples *FS = findFunctionSamples(Inst);
462 return std::error_code();
464 // Ignore all intrinsics and branch instructions.
465 // Branch instruction usually contains debug info from sources outside of
466 // the residing basic block, thus we ignore them during annotation.
467 if (isa<BranchInst>(Inst) || isa<IntrinsicInst>(Inst))
468 return std::error_code();
470 // If a call/invoke instruction is inlined in profile, but not inlined here,
471 // it means that the inlined callsite has no sample, thus the call
472 // instruction should have 0 count.
473 bool IsCall = isa<CallInst>(Inst) || isa<InvokeInst>(Inst);
474 if (IsCall && findCalleeFunctionSamples(Inst))
477 const DILocation *DIL = DLoc;
478 unsigned Lineno = DLoc.getLine();
479 unsigned HeaderLineno = DIL->getScope()->getSubprogram()->getLine();
481 uint32_t LineOffset = getOffset(Lineno, HeaderLineno);
482 uint32_t Discriminator = DIL->getDiscriminator();
483 ErrorOr<uint64_t> R = IsCall
484 ? FS->findCallSamplesAt(LineOffset, Discriminator)
485 : FS->findSamplesAt(LineOffset, Discriminator);
488 CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get());
490 const Function *F = Inst.getParent()->getParent();
491 LLVMContext &Ctx = F->getContext();
492 emitOptimizationRemark(
493 Ctx, DEBUG_TYPE, *F, DLoc,
494 Twine("Applied ") + Twine(*R) + " samples from profile (offset: " +
496 ((Discriminator) ? Twine(".") + Twine(Discriminator) : "") + ")");
498 DEBUG(dbgs() << " " << Lineno << "." << DIL->getDiscriminator() << ":"
499 << Inst << " (line offset: " << Lineno - HeaderLineno << "."
500 << DIL->getDiscriminator() << " - weight: " << R.get()
506 /// \brief Compute the weight of a basic block.
508 /// The weight of basic block \p BB is the maximum weight of all the
509 /// instructions in BB.
511 /// \param BB The basic block to query.
513 /// \returns the weight for \p BB.
515 SampleProfileLoader::getBlockWeight(const BasicBlock *BB) {
517 bool HasWeight = false;
518 for (auto &I : BB->getInstList()) {
519 const ErrorOr<uint64_t> &R = getInstWeight(I);
521 Max = std::max(Max, R.get());
525 return HasWeight ? ErrorOr<uint64_t>(Max) : std::error_code();
528 /// \brief Compute and store the weights of every basic block.
530 /// This populates the BlockWeights map by computing
531 /// the weights of every basic block in the CFG.
533 /// \param F The function to query.
534 bool SampleProfileLoader::computeBlockWeights(Function &F) {
535 bool Changed = false;
536 DEBUG(dbgs() << "Block weights\n");
537 for (const auto &BB : F) {
538 ErrorOr<uint64_t> Weight = getBlockWeight(&BB);
540 BlockWeights[&BB] = Weight.get();
541 VisitedBlocks.insert(&BB);
544 DEBUG(printBlockWeight(dbgs(), &BB));
550 /// \brief Get the FunctionSamples for a call instruction.
552 /// The FunctionSamples of a call/invoke instruction \p Inst is the inlined
553 /// instance in which that call instruction is calling to. It contains
554 /// all samples that resides in the inlined instance. We first find the
555 /// inlined instance in which the call instruction is from, then we
556 /// traverse its children to find the callsite with the matching
559 /// \param Inst Call/Invoke instruction to query.
561 /// \returns The FunctionSamples pointer to the inlined instance.
562 const FunctionSamples *
563 SampleProfileLoader::findCalleeFunctionSamples(const Instruction &Inst) const {
564 const DILocation *DIL = Inst.getDebugLoc();
568 DISubprogram *SP = DIL->getScope()->getSubprogram();
572 const FunctionSamples *FS = findFunctionSamples(Inst);
576 return FS->findFunctionSamplesAt(LineLocation(
577 getOffset(DIL->getLine(), SP->getLine()), DIL->getDiscriminator()));
580 /// \brief Get the FunctionSamples for an instruction.
582 /// The FunctionSamples of an instruction \p Inst is the inlined instance
583 /// in which that instruction is coming from. We traverse the inline stack
584 /// of that instruction, and match it with the tree nodes in the profile.
586 /// \param Inst Instruction to query.
588 /// \returns the FunctionSamples pointer to the inlined instance.
589 const FunctionSamples *
590 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
591 SmallVector<LineLocation, 10> S;
592 const DILocation *DIL = Inst.getDebugLoc();
596 for (DIL = DIL->getInlinedAt(); DIL; DIL = DIL->getInlinedAt()) {
597 DISubprogram *SP = DIL->getScope()->getSubprogram();
600 S.push_back(LineLocation(getOffset(DIL->getLine(), SP->getLine()),
601 DIL->getDiscriminator()));
605 const FunctionSamples *FS = Samples;
606 for (int i = S.size() - 1; i >= 0 && FS != nullptr; i--) {
607 FS = FS->findFunctionSamplesAt(S[i]);
612 /// \brief Iteratively inline hot callsites of a function.
614 /// Iteratively traverse all callsites of the function \p F, and find if
615 /// the corresponding inlined instance exists and is hot in profile. If
616 /// it is hot enough, inline the callsites and adds new callsites of the
617 /// callee into the caller.
619 /// TODO: investigate the possibility of not invoking InlineFunction directly.
621 /// \param F function to perform iterative inlining.
623 /// \returns True if there is any inline happened.
624 bool SampleProfileLoader::inlineHotFunctions(Function &F) {
625 bool Changed = false;
626 LLVMContext &Ctx = F.getContext();
627 std::function<AssumptionCache &(Function &)> GetAssumptionCache = [&](
628 Function &F) -> AssumptionCache & { return ACT->getAssumptionCache(F); };
630 bool LocalChanged = false;
631 SmallVector<Instruction *, 10> CIS;
634 SmallVector<Instruction *, 10> Candidates;
635 for (auto &I : BB.getInstList()) {
636 const FunctionSamples *FS = nullptr;
637 if ((isa<CallInst>(I) || isa<InvokeInst>(I)) &&
638 (FS = findCalleeFunctionSamples(I))) {
639 Candidates.push_back(&I);
640 if (callsiteIsHot(Samples, FS))
645 CIS.insert(CIS.begin(), Candidates.begin(), Candidates.end());
649 InlineFunctionInfo IFI(nullptr, ACT ? &GetAssumptionCache : nullptr);
651 Function *CalledFunction = CS.getCalledFunction();
652 if (!CalledFunction || !CalledFunction->getSubprogram())
654 DebugLoc DLoc = I->getDebugLoc();
655 uint64_t NumSamples = findCalleeFunctionSamples(*I)->getTotalSamples();
656 if (InlineFunction(CS, IFI)) {
658 emitOptimizationRemark(Ctx, DEBUG_TYPE, F, DLoc,
659 Twine("inlined hot callee '") +
660 CalledFunction->getName() + "' with " +
661 Twine(NumSamples) + " samples into '" +
674 /// \brief Find equivalence classes for the given block.
676 /// This finds all the blocks that are guaranteed to execute the same
677 /// number of times as \p BB1. To do this, it traverses all the
678 /// descendants of \p BB1 in the dominator or post-dominator tree.
680 /// A block BB2 will be in the same equivalence class as \p BB1 if
681 /// the following holds:
683 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
684 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
685 /// dominate BB1 in the post-dominator tree.
687 /// 2- Both BB2 and \p BB1 must be in the same loop.
689 /// For every block BB2 that meets those two requirements, we set BB2's
690 /// equivalence class to \p BB1.
692 /// \param BB1 Block to check.
693 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
694 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
695 /// with blocks from \p BB1's dominator tree, then
696 /// this is the post-dominator tree, and vice versa.
697 void SampleProfileLoader::findEquivalencesFor(
698 BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
699 DominatorTreeBase<BasicBlock> *DomTree) {
700 const BasicBlock *EC = EquivalenceClass[BB1];
701 uint64_t Weight = BlockWeights[EC];
702 for (const auto *BB2 : Descendants) {
703 bool IsDomParent = DomTree->dominates(BB2, BB1);
704 bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
705 if (BB1 != BB2 && IsDomParent && IsInSameLoop) {
706 EquivalenceClass[BB2] = EC;
707 // If BB2 is visited, then the entire EC should be marked as visited.
708 if (VisitedBlocks.count(BB2)) {
709 VisitedBlocks.insert(EC);
712 // If BB2 is heavier than BB1, make BB2 have the same weight
715 // Note that we don't worry about the opposite situation here
716 // (when BB2 is lighter than BB1). We will deal with this
717 // during the propagation phase. Right now, we just want to
718 // make sure that BB1 has the largest weight of all the
719 // members of its equivalence set.
720 Weight = std::max(Weight, BlockWeights[BB2]);
723 if (EC == &EC->getParent()->getEntryBlock()) {
724 BlockWeights[EC] = Samples->getHeadSamples() + 1;
726 BlockWeights[EC] = Weight;
730 /// \brief Find equivalence classes.
732 /// Since samples may be missing from blocks, we can fill in the gaps by setting
733 /// the weights of all the blocks in the same equivalence class to the same
734 /// weight. To compute the concept of equivalence, we use dominance and loop
735 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
736 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
738 /// \param F The function to query.
739 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
740 SmallVector<BasicBlock *, 8> DominatedBBs;
741 DEBUG(dbgs() << "\nBlock equivalence classes\n");
742 // Find equivalence sets based on dominance and post-dominance information.
744 BasicBlock *BB1 = &BB;
746 // Compute BB1's equivalence class once.
747 if (EquivalenceClass.count(BB1)) {
748 DEBUG(printBlockEquivalence(dbgs(), BB1));
752 // By default, blocks are in their own equivalence class.
753 EquivalenceClass[BB1] = BB1;
755 // Traverse all the blocks dominated by BB1. We are looking for
756 // every basic block BB2 such that:
758 // 1- BB1 dominates BB2.
759 // 2- BB2 post-dominates BB1.
760 // 3- BB1 and BB2 are in the same loop nest.
762 // If all those conditions hold, it means that BB2 is executed
763 // as many times as BB1, so they are placed in the same equivalence
764 // class by making BB2's equivalence class be BB1.
765 DominatedBBs.clear();
766 DT->getDescendants(BB1, DominatedBBs);
767 findEquivalencesFor(BB1, DominatedBBs, PDT.get());
769 DEBUG(printBlockEquivalence(dbgs(), BB1));
772 // Assign weights to equivalence classes.
774 // All the basic blocks in the same equivalence class will execute
775 // the same number of times. Since we know that the head block in
776 // each equivalence class has the largest weight, assign that weight
777 // to all the blocks in that equivalence class.
778 DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n");
780 const BasicBlock *BB = &BI;
781 const BasicBlock *EquivBB = EquivalenceClass[BB];
783 BlockWeights[BB] = BlockWeights[EquivBB];
784 DEBUG(printBlockWeight(dbgs(), BB));
788 /// \brief Visit the given edge to decide if it has a valid weight.
790 /// If \p E has not been visited before, we copy to \p UnknownEdge
791 /// and increment the count of unknown edges.
793 /// \param E Edge to visit.
794 /// \param NumUnknownEdges Current number of unknown edges.
795 /// \param UnknownEdge Set if E has not been visited before.
797 /// \returns E's weight, if known. Otherwise, return 0.
798 uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
800 if (!VisitedEdges.count(E)) {
801 (*NumUnknownEdges)++;
806 return EdgeWeights[E];
809 /// \brief Propagate weights through incoming/outgoing edges.
811 /// If the weight of a basic block is known, and there is only one edge
812 /// with an unknown weight, we can calculate the weight of that edge.
814 /// Similarly, if all the edges have a known count, we can calculate the
815 /// count of the basic block, if needed.
817 /// \param F Function to process.
818 /// \param UpdateBlockCount Whether we should update basic block counts that
819 /// has already been annotated.
821 /// \returns True if new weights were assigned to edges or blocks.
822 bool SampleProfileLoader::propagateThroughEdges(Function &F,
823 bool UpdateBlockCount) {
824 bool Changed = false;
825 DEBUG(dbgs() << "\nPropagation through edges\n");
826 for (const auto &BI : F) {
827 const BasicBlock *BB = &BI;
828 const BasicBlock *EC = EquivalenceClass[BB];
830 // Visit all the predecessor and successor edges to determine
831 // which ones have a weight assigned already. Note that it doesn't
832 // matter that we only keep track of a single unknown edge. The
833 // only case we are interested in handling is when only a single
834 // edge is unknown (see setEdgeOrBlockWeight).
835 for (unsigned i = 0; i < 2; i++) {
836 uint64_t TotalWeight = 0;
837 unsigned NumUnknownEdges = 0, NumTotalEdges = 0;
838 Edge UnknownEdge, SelfReferentialEdge, SingleEdge;
841 // First, visit all predecessor edges.
842 NumTotalEdges = Predecessors[BB].size();
843 for (auto *Pred : Predecessors[BB]) {
844 Edge E = std::make_pair(Pred, BB);
845 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
846 if (E.first == E.second)
847 SelfReferentialEdge = E;
849 if (NumTotalEdges == 1) {
850 SingleEdge = std::make_pair(Predecessors[BB][0], BB);
853 // On the second round, visit all successor edges.
854 NumTotalEdges = Successors[BB].size();
855 for (auto *Succ : Successors[BB]) {
856 Edge E = std::make_pair(BB, Succ);
857 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
859 if (NumTotalEdges == 1) {
860 SingleEdge = std::make_pair(BB, Successors[BB][0]);
864 // After visiting all the edges, there are three cases that we
865 // can handle immediately:
867 // - All the edge weights are known (i.e., NumUnknownEdges == 0).
868 // In this case, we simply check that the sum of all the edges
869 // is the same as BB's weight. If not, we change BB's weight
870 // to match. Additionally, if BB had not been visited before,
871 // we mark it visited.
873 // - Only one edge is unknown and BB has already been visited.
874 // In this case, we can compute the weight of the edge by
875 // subtracting the total block weight from all the known
876 // edge weights. If the edges weight more than BB, then the
877 // edge of the last remaining edge is set to zero.
879 // - There exists a self-referential edge and the weight of BB is
880 // known. In this case, this edge can be based on BB's weight.
881 // We add up all the other known edges and set the weight on
882 // the self-referential edge as we did in the previous case.
884 // In any other case, we must continue iterating. Eventually,
885 // all edges will get a weight, or iteration will stop when
886 // it reaches SampleProfileMaxPropagateIterations.
887 if (NumUnknownEdges <= 1) {
888 uint64_t &BBWeight = BlockWeights[EC];
889 if (NumUnknownEdges == 0) {
890 if (!VisitedBlocks.count(EC)) {
891 // If we already know the weight of all edges, the weight of the
892 // basic block can be computed. It should be no larger than the sum
893 // of all edge weights.
894 if (TotalWeight > BBWeight) {
895 BBWeight = TotalWeight;
897 DEBUG(dbgs() << "All edge weights for " << BB->getName()
898 << " known. Set weight for block: ";
899 printBlockWeight(dbgs(), BB););
901 } else if (NumTotalEdges == 1 &&
902 EdgeWeights[SingleEdge] < BlockWeights[EC]) {
903 // If there is only one edge for the visited basic block, use the
904 // block weight to adjust edge weight if edge weight is smaller.
905 EdgeWeights[SingleEdge] = BlockWeights[EC];
908 } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) {
909 // If there is a single unknown edge and the block has been
910 // visited, then we can compute E's weight.
911 if (BBWeight >= TotalWeight)
912 EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
914 EdgeWeights[UnknownEdge] = 0;
915 const BasicBlock *OtherEC;
917 OtherEC = EquivalenceClass[UnknownEdge.first];
919 OtherEC = EquivalenceClass[UnknownEdge.second];
920 // Edge weights should never exceed the BB weights it connects.
921 if (VisitedBlocks.count(OtherEC) &&
922 EdgeWeights[UnknownEdge] > BlockWeights[OtherEC])
923 EdgeWeights[UnknownEdge] = BlockWeights[OtherEC];
924 VisitedEdges.insert(UnknownEdge);
926 DEBUG(dbgs() << "Set weight for edge: ";
927 printEdgeWeight(dbgs(), UnknownEdge));
929 } else if (VisitedBlocks.count(EC) && BlockWeights[EC] == 0) {
930 // If a block Weights 0, all its in/out edges should weight 0.
932 for (auto *Pred : Predecessors[BB]) {
933 Edge E = std::make_pair(Pred, BB);
935 VisitedEdges.insert(E);
938 for (auto *Succ : Successors[BB]) {
939 Edge E = std::make_pair(BB, Succ);
941 VisitedEdges.insert(E);
944 } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) {
945 uint64_t &BBWeight = BlockWeights[BB];
946 // We have a self-referential edge and the weight of BB is known.
947 if (BBWeight >= TotalWeight)
948 EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
950 EdgeWeights[SelfReferentialEdge] = 0;
951 VisitedEdges.insert(SelfReferentialEdge);
953 DEBUG(dbgs() << "Set self-referential edge weight to: ";
954 printEdgeWeight(dbgs(), SelfReferentialEdge));
956 if (UpdateBlockCount && !VisitedBlocks.count(EC) && TotalWeight > 0) {
957 BlockWeights[EC] = TotalWeight;
958 VisitedBlocks.insert(EC);
967 /// \brief Build in/out edge lists for each basic block in the CFG.
969 /// We are interested in unique edges. If a block B1 has multiple
970 /// edges to another block B2, we only add a single B1->B2 edge.
971 void SampleProfileLoader::buildEdges(Function &F) {
973 BasicBlock *B1 = &BI;
975 // Add predecessors for B1.
976 SmallPtrSet<BasicBlock *, 16> Visited;
977 if (!Predecessors[B1].empty())
978 llvm_unreachable("Found a stale predecessors list in a basic block.");
979 for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
980 BasicBlock *B2 = *PI;
981 if (Visited.insert(B2).second)
982 Predecessors[B1].push_back(B2);
985 // Add successors for B1.
987 if (!Successors[B1].empty())
988 llvm_unreachable("Found a stale successors list in a basic block.");
989 for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
990 BasicBlock *B2 = *SI;
991 if (Visited.insert(B2).second)
992 Successors[B1].push_back(B2);
997 /// \brief Propagate weights into edges
999 /// The following rules are applied to every block BB in the CFG:
1001 /// - If BB has a single predecessor/successor, then the weight
1002 /// of that edge is the weight of the block.
1004 /// - If all incoming or outgoing edges are known except one, and the
1005 /// weight of the block is already known, the weight of the unknown
1006 /// edge will be the weight of the block minus the sum of all the known
1007 /// edges. If the sum of all the known edges is larger than BB's weight,
1008 /// we set the unknown edge weight to zero.
1010 /// - If there is a self-referential edge, and the weight of the block is
1011 /// known, the weight for that edge is set to the weight of the block
1012 /// minus the weight of the other incoming edges to that block (if
1014 void SampleProfileLoader::propagateWeights(Function &F) {
1015 bool Changed = true;
1018 // Add an entry count to the function using the samples gathered
1019 // at the function entry.
1020 F.setEntryCount(Samples->getHeadSamples() + 1);
1022 // If BB weight is larger than its corresponding loop's header BB weight,
1023 // use the BB weight to replace the loop header BB weight.
1024 for (auto &BI : F) {
1025 BasicBlock *BB = &BI;
1026 Loop *L = LI->getLoopFor(BB);
1030 BasicBlock *Header = L->getHeader();
1031 if (Header && BlockWeights[BB] > BlockWeights[Header]) {
1032 BlockWeights[Header] = BlockWeights[BB];
1036 // Before propagation starts, build, for each block, a list of
1037 // unique predecessors and successors. This is necessary to handle
1038 // identical edges in multiway branches. Since we visit all blocks and all
1039 // edges of the CFG, it is cleaner to build these lists once at the start
1043 // Propagate until we converge or we go past the iteration limit.
1044 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1045 Changed = propagateThroughEdges(F, false);
1048 // The first propagation propagates BB counts from annotated BBs to unknown
1049 // BBs. The 2nd propagation pass resets edges weights, and use all BB weights
1050 // to propagate edge weights.
1051 VisitedEdges.clear();
1053 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1054 Changed = propagateThroughEdges(F, false);
1057 // The 3rd propagation pass allows adjust annotated BB weights that are
1060 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1061 Changed = propagateThroughEdges(F, true);
1064 // Generate MD_prof metadata for every branch instruction using the
1065 // edge weights computed during propagation.
1066 DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
1067 LLVMContext &Ctx = F.getContext();
1069 for (auto &BI : F) {
1070 BasicBlock *BB = &BI;
1072 if (BlockWeights[BB]) {
1073 for (auto &I : BB->getInstList()) {
1074 if (CallInst *CI = dyn_cast<CallInst>(&I)) {
1075 if (!dyn_cast<IntrinsicInst>(&I)) {
1076 SmallVector<uint32_t, 1> Weights;
1077 Weights.push_back(BlockWeights[BB]);
1078 CI->setMetadata(LLVMContext::MD_prof,
1079 MDB.createBranchWeights(Weights));
1084 TerminatorInst *TI = BB->getTerminator();
1085 if (TI->getNumSuccessors() == 1)
1087 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
1090 DEBUG(dbgs() << "\nGetting weights for branch at line "
1091 << TI->getDebugLoc().getLine() << ".\n");
1092 SmallVector<uint32_t, 4> Weights;
1093 uint32_t MaxWeight = 0;
1094 DebugLoc MaxDestLoc;
1095 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
1096 BasicBlock *Succ = TI->getSuccessor(I);
1097 Edge E = std::make_pair(BB, Succ);
1098 uint64_t Weight = EdgeWeights[E];
1099 DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
1100 // Use uint32_t saturated arithmetic to adjust the incoming weights,
1101 // if needed. Sample counts in profiles are 64-bit unsigned values,
1102 // but internally branch weights are expressed as 32-bit values.
1103 if (Weight > std::numeric_limits<uint32_t>::max()) {
1104 DEBUG(dbgs() << " (saturated due to uint32_t overflow)");
1105 Weight = std::numeric_limits<uint32_t>::max();
1107 // Weight is added by one to avoid propagation errors introduced by
1109 Weights.push_back(static_cast<uint32_t>(Weight + 1));
1111 if (Weight > MaxWeight) {
1113 MaxDestLoc = Succ->getFirstNonPHIOrDbgOrLifetime()->getDebugLoc();
1118 // Only set weights if there is at least one non-zero weight.
1119 // In any other case, let the analyzer set weights.
1120 if (MaxWeight > 0) {
1121 DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
1122 TI->setMetadata(llvm::LLVMContext::MD_prof,
1123 MDB.createBranchWeights(Weights));
1124 DebugLoc BranchLoc = TI->getDebugLoc();
1125 emitOptimizationRemark(
1126 Ctx, DEBUG_TYPE, F, MaxDestLoc,
1127 Twine("most popular destination for conditional branches at ") +
1128 ((BranchLoc) ? Twine(BranchLoc->getFilename() + ":" +
1129 Twine(BranchLoc.getLine()) + ":" +
1130 Twine(BranchLoc.getCol()))
1131 : Twine("<UNKNOWN LOCATION>")));
1133 DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
1138 /// \brief Get the line number for the function header.
1140 /// This looks up function \p F in the current compilation unit and
1141 /// retrieves the line number where the function is defined. This is
1142 /// line 0 for all the samples read from the profile file. Every line
1143 /// number is relative to this line.
1145 /// \param F Function object to query.
1147 /// \returns the line number where \p F is defined. If it returns 0,
1148 /// it means that there is no debug information available for \p F.
1149 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
1150 if (DISubprogram *S = F.getSubprogram())
1151 return S->getLine();
1153 // If the start of \p F is missing, emit a diagnostic to inform the user
1154 // about the missed opportunity.
1155 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1156 "No debug information found in function " + F.getName() +
1157 ": Function profile not used",
1162 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
1163 DT.reset(new DominatorTree);
1166 PDT.reset(new DominatorTreeBase<BasicBlock>(true));
1167 PDT->recalculate(F);
1169 LI.reset(new LoopInfo);
1173 /// \brief Generate branch weight metadata for all branches in \p F.
1175 /// Branch weights are computed out of instruction samples using a
1176 /// propagation heuristic. Propagation proceeds in 3 phases:
1178 /// 1- Assignment of block weights. All the basic blocks in the function
1179 /// are initial assigned the same weight as their most frequently
1180 /// executed instruction.
1182 /// 2- Creation of equivalence classes. Since samples may be missing from
1183 /// blocks, we can fill in the gaps by setting the weights of all the
1184 /// blocks in the same equivalence class to the same weight. To compute
1185 /// the concept of equivalence, we use dominance and loop information.
1186 /// Two blocks B1 and B2 are in the same equivalence class if B1
1187 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
1189 /// 3- Propagation of block weights into edges. This uses a simple
1190 /// propagation heuristic. The following rules are applied to every
1191 /// block BB in the CFG:
1193 /// - If BB has a single predecessor/successor, then the weight
1194 /// of that edge is the weight of the block.
1196 /// - If all the edges are known except one, and the weight of the
1197 /// block is already known, the weight of the unknown edge will
1198 /// be the weight of the block minus the sum of all the known
1199 /// edges. If the sum of all the known edges is larger than BB's weight,
1200 /// we set the unknown edge weight to zero.
1202 /// - If there is a self-referential edge, and the weight of the block is
1203 /// known, the weight for that edge is set to the weight of the block
1204 /// minus the weight of the other incoming edges to that block (if
1207 /// Since this propagation is not guaranteed to finalize for every CFG, we
1208 /// only allow it to proceed for a limited number of iterations (controlled
1209 /// by -sample-profile-max-propagate-iterations).
1211 /// FIXME: Try to replace this propagation heuristic with a scheme
1212 /// that is guaranteed to finalize. A work-list approach similar to
1213 /// the standard value propagation algorithm used by SSA-CCP might
1216 /// Once all the branch weights are computed, we emit the MD_prof
1217 /// metadata on BB using the computed values for each of its branches.
1219 /// \param F The function to query.
1221 /// \returns true if \p F was modified. Returns false, otherwise.
1222 bool SampleProfileLoader::emitAnnotations(Function &F) {
1223 bool Changed = false;
1225 if (getFunctionLoc(F) == 0)
1228 DEBUG(dbgs() << "Line number for the first instruction in " << F.getName()
1229 << ": " << getFunctionLoc(F) << "\n");
1231 Changed |= inlineHotFunctions(F);
1233 // Compute basic block weights.
1234 Changed |= computeBlockWeights(F);
1237 // Compute dominance and loop info needed for propagation.
1238 computeDominanceAndLoopInfo(F);
1240 // Find equivalence classes.
1241 findEquivalenceClasses(F);
1243 // Propagate weights to all edges.
1244 propagateWeights(F);
1247 // If coverage checking was requested, compute it now.
1248 if (SampleProfileRecordCoverage) {
1249 unsigned Used = CoverageTracker.countUsedRecords(Samples);
1250 unsigned Total = CoverageTracker.countBodyRecords(Samples);
1251 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1252 if (Coverage < SampleProfileRecordCoverage) {
1253 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1254 F.getSubprogram()->getFilename(), getFunctionLoc(F),
1255 Twine(Used) + " of " + Twine(Total) + " available profile records (" +
1256 Twine(Coverage) + "%) were applied",
1261 if (SampleProfileSampleCoverage) {
1262 uint64_t Used = CoverageTracker.getTotalUsedSamples();
1263 uint64_t Total = CoverageTracker.countBodySamples(Samples);
1264 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1265 if (Coverage < SampleProfileSampleCoverage) {
1266 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1267 F.getSubprogram()->getFilename(), getFunctionLoc(F),
1268 Twine(Used) + " of " + Twine(Total) + " available profile samples (" +
1269 Twine(Coverage) + "%) were applied",
1276 char SampleProfileLoaderLegacyPass::ID = 0;
1277 INITIALIZE_PASS_BEGIN(SampleProfileLoaderLegacyPass, "sample-profile",
1278 "Sample Profile loader", false, false)
1279 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1280 INITIALIZE_PASS_END(SampleProfileLoaderLegacyPass, "sample-profile",
1281 "Sample Profile loader", false, false)
1283 bool SampleProfileLoader::doInitialization(Module &M) {
1284 auto &Ctx = M.getContext();
1285 auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
1286 if (std::error_code EC = ReaderOrErr.getError()) {
1287 std::string Msg = "Could not open profile: " + EC.message();
1288 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg));
1291 Reader = std::move(ReaderOrErr.get());
1292 ProfileIsValid = (Reader->read() == sampleprof_error::success);
1296 ModulePass *llvm::createSampleProfileLoaderPass() {
1297 return new SampleProfileLoaderLegacyPass(SampleProfileFile);
1300 ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) {
1301 return new SampleProfileLoaderLegacyPass(Name);
1304 bool SampleProfileLoader::runOnModule(Module &M) {
1305 if (!ProfileIsValid)
1308 // Compute the total number of samples collected in this profile.
1309 for (const auto &I : Reader->getProfiles())
1310 TotalCollectedSamples += I.second.getTotalSamples();
1312 bool retval = false;
1314 if (!F.isDeclaration()) {
1315 clearFunctionData();
1316 retval |= runOnFunction(F);
1318 if (M.getProfileSummary() == nullptr)
1319 M.setProfileSummary(Reader->getSummary().getMD(M.getContext()));
1323 bool SampleProfileLoaderLegacyPass::runOnModule(Module &M) {
1324 // FIXME: pass in AssumptionCache correctly for the new pass manager.
1325 SampleLoader.setACT(&getAnalysis<AssumptionCacheTracker>());
1326 return SampleLoader.runOnModule(M);
1329 bool SampleProfileLoader::runOnFunction(Function &F) {
1331 Samples = Reader->getSamplesFor(F);
1332 if (!Samples->empty())
1333 return emitAnnotations(F);
1337 PreservedAnalyses SampleProfileLoaderPass::run(Module &M,
1338 ModuleAnalysisManager &AM) {
1340 SampleProfileLoader SampleLoader(SampleProfileFile);
1342 SampleLoader.doInitialization(M);
1344 if (!SampleLoader.runOnModule(M))
1345 return PreservedAnalyses::all();
1347 return PreservedAnalyses::none();