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/GlobalValue.h"
39 #include "llvm/IR/InstIterator.h"
40 #include "llvm/IR/Instructions.h"
41 #include "llvm/IR/IntrinsicInst.h"
42 #include "llvm/IR/LLVMContext.h"
43 #include "llvm/IR/MDBuilder.h"
44 #include "llvm/IR/Metadata.h"
45 #include "llvm/IR/Module.h"
46 #include "llvm/Pass.h"
47 #include "llvm/ProfileData/InstrProf.h"
48 #include "llvm/ProfileData/SampleProfReader.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorOr.h"
52 #include "llvm/Support/Format.h"
53 #include "llvm/Support/raw_ostream.h"
54 #include "llvm/Transforms/IPO.h"
55 #include "llvm/Transforms/Instrumentation.h"
56 #include "llvm/Transforms/Utils/Cloning.h"
60 using namespace sampleprof;
62 #define DEBUG_TYPE "sample-profile"
64 // Command line option to specify the file to read samples from. This is
65 // mainly used for debugging.
66 static cl::opt<std::string> SampleProfileFile(
67 "sample-profile-file", cl::init(""), cl::value_desc("filename"),
68 cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
69 static cl::opt<unsigned> SampleProfileMaxPropagateIterations(
70 "sample-profile-max-propagate-iterations", cl::init(100),
71 cl::desc("Maximum number of iterations to go through when propagating "
72 "sample block/edge weights through the CFG."));
73 static cl::opt<unsigned> SampleProfileRecordCoverage(
74 "sample-profile-check-record-coverage", cl::init(0), cl::value_desc("N"),
75 cl::desc("Emit a warning if less than N% of records in the input profile "
76 "are matched to the IR."));
77 static cl::opt<unsigned> SampleProfileSampleCoverage(
78 "sample-profile-check-sample-coverage", cl::init(0), cl::value_desc("N"),
79 cl::desc("Emit a warning if less than N% of samples in the input profile "
80 "are matched to the IR."));
81 static cl::opt<double> SampleProfileHotThreshold(
82 "sample-profile-inline-hot-threshold", cl::init(0.1), cl::value_desc("N"),
83 cl::desc("Inlined functions that account for more than N% of all samples "
84 "collected in the parent function, will be inlined again."));
87 typedef DenseMap<const BasicBlock *, uint64_t> BlockWeightMap;
88 typedef DenseMap<const BasicBlock *, const BasicBlock *> EquivalenceClassMap;
89 typedef std::pair<const BasicBlock *, const BasicBlock *> Edge;
90 typedef DenseMap<Edge, uint64_t> EdgeWeightMap;
91 typedef DenseMap<const BasicBlock *, SmallVector<const BasicBlock *, 8>>
94 class SampleCoverageTracker {
96 SampleCoverageTracker() : SampleCoverage(), TotalUsedSamples(0) {}
98 bool markSamplesUsed(const FunctionSamples *FS, uint32_t LineOffset,
99 uint32_t Discriminator, uint64_t Samples);
100 unsigned computeCoverage(unsigned Used, unsigned Total) const;
101 unsigned countUsedRecords(const FunctionSamples *FS) const;
102 unsigned countBodyRecords(const FunctionSamples *FS) const;
103 uint64_t getTotalUsedSamples() const { return TotalUsedSamples; }
104 uint64_t countBodySamples(const FunctionSamples *FS) const;
106 SampleCoverage.clear();
107 TotalUsedSamples = 0;
111 typedef std::map<LineLocation, unsigned> BodySampleCoverageMap;
112 typedef DenseMap<const FunctionSamples *, BodySampleCoverageMap>
113 FunctionSamplesCoverageMap;
115 /// Coverage map for sampling records.
117 /// This map keeps a record of sampling records that have been matched to
118 /// an IR instruction. This is used to detect some form of staleness in
119 /// profiles (see flag -sample-profile-check-coverage).
121 /// Each entry in the map corresponds to a FunctionSamples instance. This is
122 /// another map that counts how many times the sample record at the
123 /// given location has been used.
124 FunctionSamplesCoverageMap SampleCoverage;
126 /// Number of samples used from the profile.
128 /// When a sampling record is used for the first time, the samples from
129 /// that record are added to this accumulator. Coverage is later computed
130 /// based on the total number of samples available in this function and
133 /// Note that this accumulator tracks samples used from a single function
134 /// and all the inlined callsites. Strictly, we should have a map of counters
135 /// keyed by FunctionSamples pointers, but these stats are cleared after
136 /// every function, so we just need to keep a single counter.
137 uint64_t TotalUsedSamples;
140 /// \brief Sample profile pass.
142 /// This pass reads profile data from the file specified by
143 /// -sample-profile-file and annotates every affected function with the
144 /// profile information found in that file.
145 class SampleProfileLoader {
147 SampleProfileLoader(StringRef Name = SampleProfileFile)
148 : DT(nullptr), PDT(nullptr), LI(nullptr), ACT(nullptr), Reader(),
149 Samples(nullptr), Filename(Name), ProfileIsValid(false),
150 TotalCollectedSamples(0) {}
152 bool doInitialization(Module &M);
153 bool runOnModule(Module &M);
154 void setACT(AssumptionCacheTracker *A) { ACT = A; }
156 void dump() { Reader->dump(); }
159 bool runOnFunction(Function &F);
160 unsigned getFunctionLoc(Function &F);
161 bool emitAnnotations(Function &F);
162 ErrorOr<uint64_t> getInstWeight(const Instruction &I);
163 ErrorOr<uint64_t> getBlockWeight(const BasicBlock *BB);
164 const FunctionSamples *findCalleeFunctionSamples(const Instruction &I) const;
165 std::vector<const FunctionSamples *>
166 findIndirectCallFunctionSamples(const Instruction &I) const;
167 const FunctionSamples *findFunctionSamples(const Instruction &I) const;
168 bool inlineHotFunctions(Function &F,
169 DenseSet<GlobalValue::GUID> &ImportGUIDs);
170 void printEdgeWeight(raw_ostream &OS, Edge E);
171 void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const;
172 void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB);
173 bool computeBlockWeights(Function &F);
174 void findEquivalenceClasses(Function &F);
175 void findEquivalencesFor(BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
176 DominatorTreeBase<BasicBlock> *DomTree);
177 void propagateWeights(Function &F);
178 uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
179 void buildEdges(Function &F);
180 bool propagateThroughEdges(Function &F, bool UpdateBlockCount);
181 void computeDominanceAndLoopInfo(Function &F);
182 unsigned getOffset(const DILocation *DIL) const;
183 void clearFunctionData();
185 /// \brief Map basic blocks to their computed weights.
187 /// The weight of a basic block is defined to be the maximum
188 /// of all the instruction weights in that block.
189 BlockWeightMap BlockWeights;
191 /// \brief Map edges to their computed weights.
193 /// Edge weights are computed by propagating basic block weights in
194 /// SampleProfile::propagateWeights.
195 EdgeWeightMap EdgeWeights;
197 /// \brief Set of visited blocks during propagation.
198 SmallPtrSet<const BasicBlock *, 32> VisitedBlocks;
200 /// \brief Set of visited edges during propagation.
201 SmallSet<Edge, 32> VisitedEdges;
203 /// \brief Equivalence classes for block weights.
205 /// Two blocks BB1 and BB2 are in the same equivalence class if they
206 /// dominate and post-dominate each other, and they are in the same loop
207 /// nest. When this happens, the two blocks are guaranteed to execute
208 /// the same number of times.
209 EquivalenceClassMap EquivalenceClass;
211 /// \brief Dominance, post-dominance and loop information.
212 std::unique_ptr<DominatorTree> DT;
213 std::unique_ptr<DominatorTreeBase<BasicBlock>> PDT;
214 std::unique_ptr<LoopInfo> LI;
216 AssumptionCacheTracker *ACT;
218 /// \brief Predecessors for each basic block in the CFG.
219 BlockEdgeMap Predecessors;
221 /// \brief Successors for each basic block in the CFG.
222 BlockEdgeMap Successors;
224 SampleCoverageTracker CoverageTracker;
226 /// \brief Profile reader object.
227 std::unique_ptr<SampleProfileReader> Reader;
229 /// \brief Samples collected for the body of this function.
230 FunctionSamples *Samples;
232 /// \brief Name of the profile file to load.
233 std::string Filename;
235 /// \brief Flag indicating whether the profile input loaded successfully.
238 /// \brief Total number of samples collected in this profile.
240 /// This is the sum of all the samples collected in all the functions executed
242 uint64_t TotalCollectedSamples;
245 class SampleProfileLoaderLegacyPass : public ModulePass {
247 // Class identification, replacement for typeinfo
250 SampleProfileLoaderLegacyPass(StringRef Name = SampleProfileFile)
251 : ModulePass(ID), SampleLoader(Name) {
252 initializeSampleProfileLoaderLegacyPassPass(
253 *PassRegistry::getPassRegistry());
256 void dump() { SampleLoader.dump(); }
258 bool doInitialization(Module &M) override {
259 return SampleLoader.doInitialization(M);
261 StringRef getPassName() const override { return "Sample profile pass"; }
262 bool runOnModule(Module &M) override;
264 void getAnalysisUsage(AnalysisUsage &AU) const override {
265 AU.addRequired<AssumptionCacheTracker>();
269 SampleProfileLoader SampleLoader;
272 /// Return true if the given callsite is hot wrt to its caller.
274 /// Functions that were inlined in the original binary will be represented
275 /// in the inline stack in the sample profile. If the profile shows that
276 /// the original inline decision was "good" (i.e., the callsite is executed
277 /// frequently), then we will recreate the inline decision and apply the
278 /// profile from the inlined callsite.
280 /// To decide whether an inlined callsite is hot, we compute the fraction
281 /// of samples used by the callsite with respect to the total number of samples
282 /// collected in the caller.
284 /// If that fraction is larger than the default given by
285 /// SampleProfileHotThreshold, the callsite will be inlined again.
286 bool callsiteIsHot(const FunctionSamples *CallerFS,
287 const FunctionSamples *CallsiteFS) {
289 return false; // The callsite was not inlined in the original binary.
291 uint64_t ParentTotalSamples = CallerFS->getTotalSamples();
292 if (ParentTotalSamples == 0)
293 return false; // Avoid division by zero.
295 uint64_t CallsiteTotalSamples = CallsiteFS->getTotalSamples();
296 if (CallsiteTotalSamples == 0)
297 return false; // Callsite is trivially cold.
299 double PercentSamples =
300 (double)CallsiteTotalSamples / (double)ParentTotalSamples * 100.0;
301 return PercentSamples >= SampleProfileHotThreshold;
305 /// Mark as used the sample record for the given function samples at
306 /// (LineOffset, Discriminator).
308 /// \returns true if this is the first time we mark the given record.
309 bool SampleCoverageTracker::markSamplesUsed(const FunctionSamples *FS,
311 uint32_t Discriminator,
313 LineLocation Loc(LineOffset, Discriminator);
314 unsigned &Count = SampleCoverage[FS][Loc];
315 bool FirstTime = (++Count == 1);
317 TotalUsedSamples += Samples;
321 /// Return the number of sample records that were applied from this profile.
323 /// This count does not include records from cold inlined callsites.
325 SampleCoverageTracker::countUsedRecords(const FunctionSamples *FS) const {
326 auto I = SampleCoverage.find(FS);
328 // The size of the coverage map for FS represents the number of records
329 // that were marked used at least once.
330 unsigned Count = (I != SampleCoverage.end()) ? I->second.size() : 0;
332 // If there are inlined callsites in this function, count the samples found
333 // in the respective bodies. However, do not bother counting callees with 0
334 // total samples, these are callees that were never invoked at runtime.
335 for (const auto &I : FS->getCallsiteSamples())
336 for (const auto &J : I.second) {
337 const FunctionSamples *CalleeSamples = &J.second;
338 if (callsiteIsHot(FS, CalleeSamples))
339 Count += countUsedRecords(CalleeSamples);
345 /// Return the number of sample records in the body of this profile.
347 /// This count does not include records from cold inlined callsites.
349 SampleCoverageTracker::countBodyRecords(const FunctionSamples *FS) const {
350 unsigned Count = FS->getBodySamples().size();
352 // Only count records in hot callsites.
353 for (const auto &I : FS->getCallsiteSamples())
354 for (const auto &J : I.second) {
355 const FunctionSamples *CalleeSamples = &J.second;
356 if (callsiteIsHot(FS, CalleeSamples))
357 Count += countBodyRecords(CalleeSamples);
363 /// Return the number of samples collected in the body of this profile.
365 /// This count does not include samples from cold inlined callsites.
367 SampleCoverageTracker::countBodySamples(const FunctionSamples *FS) const {
369 for (const auto &I : FS->getBodySamples())
370 Total += I.second.getSamples();
372 // Only count samples in hot callsites.
373 for (const auto &I : FS->getCallsiteSamples())
374 for (const auto &J : I.second) {
375 const FunctionSamples *CalleeSamples = &J.second;
376 if (callsiteIsHot(FS, CalleeSamples))
377 Total += countBodySamples(CalleeSamples);
383 /// Return the fraction of sample records used in this profile.
385 /// The returned value is an unsigned integer in the range 0-100 indicating
386 /// the percentage of sample records that were used while applying this
387 /// profile to the associated function.
388 unsigned SampleCoverageTracker::computeCoverage(unsigned Used,
389 unsigned Total) const {
390 assert(Used <= Total &&
391 "number of used records cannot exceed the total number of records");
392 return Total > 0 ? Used * 100 / Total : 100;
395 /// Clear all the per-function data used to load samples and propagate weights.
396 void SampleProfileLoader::clearFunctionData() {
397 BlockWeights.clear();
399 VisitedBlocks.clear();
400 VisitedEdges.clear();
401 EquivalenceClass.clear();
405 Predecessors.clear();
407 CoverageTracker.clear();
410 /// Returns the line offset to the start line of the subprogram.
411 /// We assume that a single function will not exceed 65535 LOC.
412 unsigned SampleProfileLoader::getOffset(const DILocation *DIL) const {
413 return (DIL->getLine() - DIL->getScope()->getSubprogram()->getLine()) &
417 /// \brief Print the weight of edge \p E on stream \p OS.
419 /// \param OS Stream to emit the output to.
420 /// \param E Edge to print.
421 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
422 OS << "weight[" << E.first->getName() << "->" << E.second->getName()
423 << "]: " << EdgeWeights[E] << "\n";
426 /// \brief Print the equivalence class of block \p BB on stream \p OS.
428 /// \param OS Stream to emit the output to.
429 /// \param BB Block to print.
430 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
431 const BasicBlock *BB) {
432 const BasicBlock *Equiv = EquivalenceClass[BB];
433 OS << "equivalence[" << BB->getName()
434 << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
437 /// \brief Print the weight of block \p BB on stream \p OS.
439 /// \param OS Stream to emit the output to.
440 /// \param BB Block to print.
441 void SampleProfileLoader::printBlockWeight(raw_ostream &OS,
442 const BasicBlock *BB) const {
443 const auto &I = BlockWeights.find(BB);
444 uint64_t W = (I == BlockWeights.end() ? 0 : I->second);
445 OS << "weight[" << BB->getName() << "]: " << W << "\n";
448 /// \brief Get the weight for an instruction.
450 /// The "weight" of an instruction \p Inst is the number of samples
451 /// collected on that instruction at runtime. To retrieve it, we
452 /// need to compute the line number of \p Inst relative to the start of its
453 /// function. We use HeaderLineno to compute the offset. We then
454 /// look up the samples collected for \p Inst using BodySamples.
456 /// \param Inst Instruction to query.
458 /// \returns the weight of \p Inst.
459 ErrorOr<uint64_t> SampleProfileLoader::getInstWeight(const Instruction &Inst) {
460 const DebugLoc &DLoc = Inst.getDebugLoc();
462 return std::error_code();
464 const FunctionSamples *FS = findFunctionSamples(Inst);
466 return std::error_code();
468 // Ignore all intrinsics and branch instructions.
469 // Branch instruction usually contains debug info from sources outside of
470 // the residing basic block, thus we ignore them during annotation.
471 if (isa<BranchInst>(Inst) || isa<IntrinsicInst>(Inst))
472 return std::error_code();
474 // If a call/invoke instruction is inlined in profile, but not inlined here,
475 // it means that the inlined callsite has no sample, thus the call
476 // instruction should have 0 count.
477 if ((isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) &&
478 findCalleeFunctionSamples(Inst))
481 const DILocation *DIL = DLoc;
482 uint32_t LineOffset = getOffset(DIL);
483 uint32_t Discriminator = DIL->getBaseDiscriminator();
484 ErrorOr<uint64_t> R = FS->findSamplesAt(LineOffset, Discriminator);
487 CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get());
489 const Function *F = Inst.getParent()->getParent();
490 LLVMContext &Ctx = F->getContext();
491 emitOptimizationRemark(
492 Ctx, DEBUG_TYPE, *F, DLoc,
493 Twine("Applied ") + Twine(*R) +
494 " samples from profile (offset: " + Twine(LineOffset) +
495 ((Discriminator) ? Twine(".") + Twine(Discriminator) : "") + ")");
497 DEBUG(dbgs() << " " << DLoc.getLine() << "."
498 << DIL->getBaseDiscriminator() << ":" << Inst
499 << " (line offset: " << LineOffset << "."
500 << DIL->getBaseDiscriminator() << " - 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.
514 ErrorOr<uint64_t> SampleProfileLoader::getBlockWeight(const BasicBlock *BB) {
516 bool HasWeight = false;
517 for (auto &I : BB->getInstList()) {
518 const ErrorOr<uint64_t> &R = getInstWeight(I);
520 Max = std::max(Max, R.get());
524 return HasWeight ? ErrorOr<uint64_t>(Max) : std::error_code();
527 /// \brief Compute and store the weights of every basic block.
529 /// This populates the BlockWeights map by computing
530 /// the weights of every basic block in the CFG.
532 /// \param F The function to query.
533 bool SampleProfileLoader::computeBlockWeights(Function &F) {
534 bool Changed = false;
535 DEBUG(dbgs() << "Block weights\n");
536 for (const auto &BB : F) {
537 ErrorOr<uint64_t> Weight = getBlockWeight(&BB);
539 BlockWeights[&BB] = Weight.get();
540 VisitedBlocks.insert(&BB);
543 DEBUG(printBlockWeight(dbgs(), &BB));
549 /// \brief Get the FunctionSamples for a call instruction.
551 /// The FunctionSamples of a call/invoke instruction \p Inst is the inlined
552 /// instance in which that call instruction is calling to. It contains
553 /// all samples that resides in the inlined instance. We first find the
554 /// inlined instance in which the call instruction is from, then we
555 /// traverse its children to find the callsite with the matching
558 /// \param Inst Call/Invoke instruction to query.
560 /// \returns The FunctionSamples pointer to the inlined instance.
561 const FunctionSamples *
562 SampleProfileLoader::findCalleeFunctionSamples(const Instruction &Inst) const {
563 const DILocation *DIL = Inst.getDebugLoc();
568 StringRef CalleeName;
569 if (const CallInst *CI = dyn_cast<CallInst>(&Inst))
570 if (Function *Callee = CI->getCalledFunction())
571 CalleeName = Callee->getName();
573 const FunctionSamples *FS = findFunctionSamples(Inst);
577 return FS->findFunctionSamplesAt(
578 LineLocation(getOffset(DIL), DIL->getBaseDiscriminator()), CalleeName);
581 /// Returns a vector of FunctionSamples that are the indirect call targets
582 /// of \p Inst. The vector is sorted by the total number of samples.
583 std::vector<const FunctionSamples *>
584 SampleProfileLoader::findIndirectCallFunctionSamples(
585 const Instruction &Inst) const {
586 const DILocation *DIL = Inst.getDebugLoc();
587 std::vector<const FunctionSamples *> R;
593 const FunctionSamples *FS = findFunctionSamples(Inst);
597 if (const FunctionSamplesMap *M = FS->findFunctionSamplesMapAt(
598 LineLocation(getOffset(DIL), DIL->getBaseDiscriminator()))) {
601 for (const auto &NameFS : *M) {
602 R.push_back(&NameFS.second);
604 std::sort(R.begin(), R.end(),
605 [](const FunctionSamples *L, const FunctionSamples *R) {
606 return L->getTotalSamples() > R->getTotalSamples();
612 /// \brief Get the FunctionSamples for an instruction.
614 /// The FunctionSamples of an instruction \p Inst is the inlined instance
615 /// in which that instruction is coming from. We traverse the inline stack
616 /// of that instruction, and match it with the tree nodes in the profile.
618 /// \param Inst Instruction to query.
620 /// \returns the FunctionSamples pointer to the inlined instance.
621 const FunctionSamples *
622 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
623 SmallVector<std::pair<LineLocation, StringRef>, 10> S;
624 const DILocation *DIL = Inst.getDebugLoc();
628 const DILocation *PrevDIL = DIL;
629 for (DIL = DIL->getInlinedAt(); DIL; DIL = DIL->getInlinedAt()) {
630 S.push_back(std::make_pair(
631 LineLocation(getOffset(DIL), DIL->getBaseDiscriminator()),
632 PrevDIL->getScope()->getSubprogram()->getLinkageName()));
637 const FunctionSamples *FS = Samples;
638 for (int i = S.size() - 1; i >= 0 && FS != nullptr; i--) {
639 FS = FS->findFunctionSamplesAt(S[i].first, S[i].second);
644 /// \brief Iteratively inline hot callsites of a function.
646 /// Iteratively traverse all callsites of the function \p F, and find if
647 /// the corresponding inlined instance exists and is hot in profile. If
648 /// it is hot enough, inline the callsites and adds new callsites of the
649 /// callee into the caller. If the call is an indirect call, first promote
650 /// it to direct call. Each indirect call is limited with a single target.
652 /// \param F function to perform iterative inlining.
653 /// \param ImportGUIDs a set to be updated to include all GUIDs that come
654 /// from a different module but inlined in the profiled binary.
656 /// \returns True if there is any inline happened.
657 bool SampleProfileLoader::inlineHotFunctions(
658 Function &F, DenseSet<GlobalValue::GUID> &ImportGUIDs) {
659 DenseSet<Instruction *> PromotedInsns;
660 bool Changed = false;
661 LLVMContext &Ctx = F.getContext();
662 std::function<AssumptionCache &(Function &)> GetAssumptionCache = [&](
663 Function &F) -> AssumptionCache & { return ACT->getAssumptionCache(F); };
665 bool LocalChanged = false;
666 SmallVector<Instruction *, 10> CIS;
669 SmallVector<Instruction *, 10> Candidates;
670 for (auto &I : BB.getInstList()) {
671 const FunctionSamples *FS = nullptr;
672 if ((isa<CallInst>(I) || isa<InvokeInst>(I)) &&
673 (FS = findCalleeFunctionSamples(I))) {
674 Candidates.push_back(&I);
675 if (callsiteIsHot(Samples, FS))
680 CIS.insert(CIS.begin(), Candidates.begin(), Candidates.end());
684 InlineFunctionInfo IFI(nullptr, ACT ? &GetAssumptionCache : nullptr);
685 Function *CalledFunction = CallSite(I).getCalledFunction();
687 if (!CalledFunction && !PromotedInsns.count(I) &&
688 CallSite(I).isIndirectCall())
689 for (const auto *FS : findIndirectCallFunctionSamples(*I)) {
690 auto CalleeFunctionName = FS->getName();
691 const char *Reason = "Callee function not available";
692 CalledFunction = F.getParent()->getFunction(CalleeFunctionName);
693 if (CalledFunction && isLegalToPromote(I, CalledFunction, &Reason)) {
694 // The indirect target was promoted and inlined in the profile, as a
695 // result, we do not have profile info for the branch probability.
696 // We set the probability to 80% taken to indicate that the static
697 // call is likely taken.
698 DI = dyn_cast<Instruction>(
699 promoteIndirectCall(I, CalledFunction, 80, 100, false)
700 ->stripPointerCasts());
701 PromotedInsns.insert(I);
703 DEBUG(dbgs() << "\nFailed to promote indirect call to "
704 << CalleeFunctionName << " because " << Reason
709 if (!CalledFunction || !CalledFunction->getSubprogram()) {
710 findCalleeFunctionSamples(*I)->findImportedFunctions(
711 ImportGUIDs, F.getParent(),
712 Samples->getTotalSamples() * SampleProfileHotThreshold / 100);
715 DebugLoc DLoc = I->getDebugLoc();
716 if (InlineFunction(CallSite(DI), IFI)) {
718 emitOptimizationRemark(Ctx, DEBUG_TYPE, F, DLoc,
719 Twine("inlined hot callee '") +
720 CalledFunction->getName() + "' into '" +
733 /// \brief Find equivalence classes for the given block.
735 /// This finds all the blocks that are guaranteed to execute the same
736 /// number of times as \p BB1. To do this, it traverses all the
737 /// descendants of \p BB1 in the dominator or post-dominator tree.
739 /// A block BB2 will be in the same equivalence class as \p BB1 if
740 /// the following holds:
742 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
743 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
744 /// dominate BB1 in the post-dominator tree.
746 /// 2- Both BB2 and \p BB1 must be in the same loop.
748 /// For every block BB2 that meets those two requirements, we set BB2's
749 /// equivalence class to \p BB1.
751 /// \param BB1 Block to check.
752 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
753 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
754 /// with blocks from \p BB1's dominator tree, then
755 /// this is the post-dominator tree, and vice versa.
756 void SampleProfileLoader::findEquivalencesFor(
757 BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
758 DominatorTreeBase<BasicBlock> *DomTree) {
759 const BasicBlock *EC = EquivalenceClass[BB1];
760 uint64_t Weight = BlockWeights[EC];
761 for (const auto *BB2 : Descendants) {
762 bool IsDomParent = DomTree->dominates(BB2, BB1);
763 bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
764 if (BB1 != BB2 && IsDomParent && IsInSameLoop) {
765 EquivalenceClass[BB2] = EC;
766 // If BB2 is visited, then the entire EC should be marked as visited.
767 if (VisitedBlocks.count(BB2)) {
768 VisitedBlocks.insert(EC);
771 // If BB2 is heavier than BB1, make BB2 have the same weight
774 // Note that we don't worry about the opposite situation here
775 // (when BB2 is lighter than BB1). We will deal with this
776 // during the propagation phase. Right now, we just want to
777 // make sure that BB1 has the largest weight of all the
778 // members of its equivalence set.
779 Weight = std::max(Weight, BlockWeights[BB2]);
782 if (EC == &EC->getParent()->getEntryBlock()) {
783 BlockWeights[EC] = Samples->getHeadSamples() + 1;
785 BlockWeights[EC] = Weight;
789 /// \brief Find equivalence classes.
791 /// Since samples may be missing from blocks, we can fill in the gaps by setting
792 /// the weights of all the blocks in the same equivalence class to the same
793 /// weight. To compute the concept of equivalence, we use dominance and loop
794 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
795 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
797 /// \param F The function to query.
798 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
799 SmallVector<BasicBlock *, 8> DominatedBBs;
800 DEBUG(dbgs() << "\nBlock equivalence classes\n");
801 // Find equivalence sets based on dominance and post-dominance information.
803 BasicBlock *BB1 = &BB;
805 // Compute BB1's equivalence class once.
806 if (EquivalenceClass.count(BB1)) {
807 DEBUG(printBlockEquivalence(dbgs(), BB1));
811 // By default, blocks are in their own equivalence class.
812 EquivalenceClass[BB1] = BB1;
814 // Traverse all the blocks dominated by BB1. We are looking for
815 // every basic block BB2 such that:
817 // 1- BB1 dominates BB2.
818 // 2- BB2 post-dominates BB1.
819 // 3- BB1 and BB2 are in the same loop nest.
821 // If all those conditions hold, it means that BB2 is executed
822 // as many times as BB1, so they are placed in the same equivalence
823 // class by making BB2's equivalence class be BB1.
824 DominatedBBs.clear();
825 DT->getDescendants(BB1, DominatedBBs);
826 findEquivalencesFor(BB1, DominatedBBs, PDT.get());
828 DEBUG(printBlockEquivalence(dbgs(), BB1));
831 // Assign weights to equivalence classes.
833 // All the basic blocks in the same equivalence class will execute
834 // the same number of times. Since we know that the head block in
835 // each equivalence class has the largest weight, assign that weight
836 // to all the blocks in that equivalence class.
837 DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n");
839 const BasicBlock *BB = &BI;
840 const BasicBlock *EquivBB = EquivalenceClass[BB];
842 BlockWeights[BB] = BlockWeights[EquivBB];
843 DEBUG(printBlockWeight(dbgs(), BB));
847 /// \brief Visit the given edge to decide if it has a valid weight.
849 /// If \p E has not been visited before, we copy to \p UnknownEdge
850 /// and increment the count of unknown edges.
852 /// \param E Edge to visit.
853 /// \param NumUnknownEdges Current number of unknown edges.
854 /// \param UnknownEdge Set if E has not been visited before.
856 /// \returns E's weight, if known. Otherwise, return 0.
857 uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
859 if (!VisitedEdges.count(E)) {
860 (*NumUnknownEdges)++;
865 return EdgeWeights[E];
868 /// \brief Propagate weights through incoming/outgoing edges.
870 /// If the weight of a basic block is known, and there is only one edge
871 /// with an unknown weight, we can calculate the weight of that edge.
873 /// Similarly, if all the edges have a known count, we can calculate the
874 /// count of the basic block, if needed.
876 /// \param F Function to process.
877 /// \param UpdateBlockCount Whether we should update basic block counts that
878 /// has already been annotated.
880 /// \returns True if new weights were assigned to edges or blocks.
881 bool SampleProfileLoader::propagateThroughEdges(Function &F,
882 bool UpdateBlockCount) {
883 bool Changed = false;
884 DEBUG(dbgs() << "\nPropagation through edges\n");
885 for (const auto &BI : F) {
886 const BasicBlock *BB = &BI;
887 const BasicBlock *EC = EquivalenceClass[BB];
889 // Visit all the predecessor and successor edges to determine
890 // which ones have a weight assigned already. Note that it doesn't
891 // matter that we only keep track of a single unknown edge. The
892 // only case we are interested in handling is when only a single
893 // edge is unknown (see setEdgeOrBlockWeight).
894 for (unsigned i = 0; i < 2; i++) {
895 uint64_t TotalWeight = 0;
896 unsigned NumUnknownEdges = 0, NumTotalEdges = 0;
897 Edge UnknownEdge, SelfReferentialEdge, SingleEdge;
900 // First, visit all predecessor edges.
901 NumTotalEdges = Predecessors[BB].size();
902 for (auto *Pred : Predecessors[BB]) {
903 Edge E = std::make_pair(Pred, BB);
904 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
905 if (E.first == E.second)
906 SelfReferentialEdge = E;
908 if (NumTotalEdges == 1) {
909 SingleEdge = std::make_pair(Predecessors[BB][0], BB);
912 // On the second round, visit all successor edges.
913 NumTotalEdges = Successors[BB].size();
914 for (auto *Succ : Successors[BB]) {
915 Edge E = std::make_pair(BB, Succ);
916 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
918 if (NumTotalEdges == 1) {
919 SingleEdge = std::make_pair(BB, Successors[BB][0]);
923 // After visiting all the edges, there are three cases that we
924 // can handle immediately:
926 // - All the edge weights are known (i.e., NumUnknownEdges == 0).
927 // In this case, we simply check that the sum of all the edges
928 // is the same as BB's weight. If not, we change BB's weight
929 // to match. Additionally, if BB had not been visited before,
930 // we mark it visited.
932 // - Only one edge is unknown and BB has already been visited.
933 // In this case, we can compute the weight of the edge by
934 // subtracting the total block weight from all the known
935 // edge weights. If the edges weight more than BB, then the
936 // edge of the last remaining edge is set to zero.
938 // - There exists a self-referential edge and the weight of BB is
939 // known. In this case, this edge can be based on BB's weight.
940 // We add up all the other known edges and set the weight on
941 // the self-referential edge as we did in the previous case.
943 // In any other case, we must continue iterating. Eventually,
944 // all edges will get a weight, or iteration will stop when
945 // it reaches SampleProfileMaxPropagateIterations.
946 if (NumUnknownEdges <= 1) {
947 uint64_t &BBWeight = BlockWeights[EC];
948 if (NumUnknownEdges == 0) {
949 if (!VisitedBlocks.count(EC)) {
950 // If we already know the weight of all edges, the weight of the
951 // basic block can be computed. It should be no larger than the sum
952 // of all edge weights.
953 if (TotalWeight > BBWeight) {
954 BBWeight = TotalWeight;
956 DEBUG(dbgs() << "All edge weights for " << BB->getName()
957 << " known. Set weight for block: ";
958 printBlockWeight(dbgs(), BB););
960 } else if (NumTotalEdges == 1 &&
961 EdgeWeights[SingleEdge] < BlockWeights[EC]) {
962 // If there is only one edge for the visited basic block, use the
963 // block weight to adjust edge weight if edge weight is smaller.
964 EdgeWeights[SingleEdge] = BlockWeights[EC];
967 } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) {
968 // If there is a single unknown edge and the block has been
969 // visited, then we can compute E's weight.
970 if (BBWeight >= TotalWeight)
971 EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
973 EdgeWeights[UnknownEdge] = 0;
974 const BasicBlock *OtherEC;
976 OtherEC = EquivalenceClass[UnknownEdge.first];
978 OtherEC = EquivalenceClass[UnknownEdge.second];
979 // Edge weights should never exceed the BB weights it connects.
980 if (VisitedBlocks.count(OtherEC) &&
981 EdgeWeights[UnknownEdge] > BlockWeights[OtherEC])
982 EdgeWeights[UnknownEdge] = BlockWeights[OtherEC];
983 VisitedEdges.insert(UnknownEdge);
985 DEBUG(dbgs() << "Set weight for edge: ";
986 printEdgeWeight(dbgs(), UnknownEdge));
988 } else if (VisitedBlocks.count(EC) && BlockWeights[EC] == 0) {
989 // If a block Weights 0, all its in/out edges should weight 0.
991 for (auto *Pred : Predecessors[BB]) {
992 Edge E = std::make_pair(Pred, BB);
994 VisitedEdges.insert(E);
997 for (auto *Succ : Successors[BB]) {
998 Edge E = std::make_pair(BB, Succ);
1000 VisitedEdges.insert(E);
1003 } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) {
1004 uint64_t &BBWeight = BlockWeights[BB];
1005 // We have a self-referential edge and the weight of BB is known.
1006 if (BBWeight >= TotalWeight)
1007 EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
1009 EdgeWeights[SelfReferentialEdge] = 0;
1010 VisitedEdges.insert(SelfReferentialEdge);
1012 DEBUG(dbgs() << "Set self-referential edge weight to: ";
1013 printEdgeWeight(dbgs(), SelfReferentialEdge));
1015 if (UpdateBlockCount && !VisitedBlocks.count(EC) && TotalWeight > 0) {
1016 BlockWeights[EC] = TotalWeight;
1017 VisitedBlocks.insert(EC);
1026 /// \brief Build in/out edge lists for each basic block in the CFG.
1028 /// We are interested in unique edges. If a block B1 has multiple
1029 /// edges to another block B2, we only add a single B1->B2 edge.
1030 void SampleProfileLoader::buildEdges(Function &F) {
1031 for (auto &BI : F) {
1032 BasicBlock *B1 = &BI;
1034 // Add predecessors for B1.
1035 SmallPtrSet<BasicBlock *, 16> Visited;
1036 if (!Predecessors[B1].empty())
1037 llvm_unreachable("Found a stale predecessors list in a basic block.");
1038 for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
1039 BasicBlock *B2 = *PI;
1040 if (Visited.insert(B2).second)
1041 Predecessors[B1].push_back(B2);
1044 // Add successors for B1.
1046 if (!Successors[B1].empty())
1047 llvm_unreachable("Found a stale successors list in a basic block.");
1048 for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
1049 BasicBlock *B2 = *SI;
1050 if (Visited.insert(B2).second)
1051 Successors[B1].push_back(B2);
1056 /// Sorts the CallTargetMap \p M by count in descending order and stores the
1057 /// sorted result in \p Sorted. Returns the total counts.
1058 static uint64_t SortCallTargets(SmallVector<InstrProfValueData, 2> &Sorted,
1059 const SampleRecord::CallTargetMap &M) {
1062 for (auto I = M.begin(); I != M.end(); ++I) {
1063 Sum += I->getValue();
1064 Sorted.push_back({Function::getGUID(I->getKey()), I->getValue()});
1066 std::sort(Sorted.begin(), Sorted.end(),
1067 [](const InstrProfValueData &L, const InstrProfValueData &R) {
1068 if (L.Count == R.Count)
1069 return L.Value > R.Value;
1071 return L.Count > R.Count;
1076 /// \brief Propagate weights into edges
1078 /// The following rules are applied to every block BB in the CFG:
1080 /// - If BB has a single predecessor/successor, then the weight
1081 /// of that edge is the weight of the block.
1083 /// - If all incoming or outgoing edges are known except one, and the
1084 /// weight of the block is already known, the weight of the unknown
1085 /// edge will be the weight of the block minus the sum of all the known
1086 /// edges. If the sum of all the known edges is larger than BB's weight,
1087 /// we set the unknown edge weight to zero.
1089 /// - If there is a self-referential edge, and the weight of the block is
1090 /// known, the weight for that edge is set to the weight of the block
1091 /// minus the weight of the other incoming edges to that block (if
1093 void SampleProfileLoader::propagateWeights(Function &F) {
1094 bool Changed = true;
1097 // If BB weight is larger than its corresponding loop's header BB weight,
1098 // use the BB weight to replace the loop header BB weight.
1099 for (auto &BI : F) {
1100 BasicBlock *BB = &BI;
1101 Loop *L = LI->getLoopFor(BB);
1105 BasicBlock *Header = L->getHeader();
1106 if (Header && BlockWeights[BB] > BlockWeights[Header]) {
1107 BlockWeights[Header] = BlockWeights[BB];
1111 // Before propagation starts, build, for each block, a list of
1112 // unique predecessors and successors. This is necessary to handle
1113 // identical edges in multiway branches. Since we visit all blocks and all
1114 // edges of the CFG, it is cleaner to build these lists once at the start
1118 // Propagate until we converge or we go past the iteration limit.
1119 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1120 Changed = propagateThroughEdges(F, false);
1123 // The first propagation propagates BB counts from annotated BBs to unknown
1124 // BBs. The 2nd propagation pass resets edges weights, and use all BB weights
1125 // to propagate edge weights.
1126 VisitedEdges.clear();
1128 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1129 Changed = propagateThroughEdges(F, false);
1132 // The 3rd propagation pass allows adjust annotated BB weights that are
1135 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1136 Changed = propagateThroughEdges(F, true);
1139 // Generate MD_prof metadata for every branch instruction using the
1140 // edge weights computed during propagation.
1141 DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
1142 LLVMContext &Ctx = F.getContext();
1144 for (auto &BI : F) {
1145 BasicBlock *BB = &BI;
1147 if (BlockWeights[BB]) {
1148 for (auto &I : BB->getInstList()) {
1149 if (!isa<CallInst>(I) && !isa<InvokeInst>(I))
1152 if (!CS.getCalledFunction()) {
1153 const DebugLoc &DLoc = I.getDebugLoc();
1156 const DILocation *DIL = DLoc;
1157 uint32_t LineOffset = getOffset(DIL);
1158 uint32_t Discriminator = DIL->getBaseDiscriminator();
1160 const FunctionSamples *FS = findFunctionSamples(I);
1163 auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
1164 if (!T || T.get().size() == 0)
1166 SmallVector<InstrProfValueData, 2> SortedCallTargets;
1167 uint64_t Sum = SortCallTargets(SortedCallTargets, T.get());
1168 annotateValueSite(*I.getParent()->getParent()->getParent(), I,
1169 SortedCallTargets, Sum, IPVK_IndirectCallTarget,
1170 SortedCallTargets.size());
1171 } else if (!dyn_cast<IntrinsicInst>(&I)) {
1172 SmallVector<uint32_t, 1> Weights;
1173 Weights.push_back(BlockWeights[BB]);
1174 I.setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights));
1178 TerminatorInst *TI = BB->getTerminator();
1179 if (TI->getNumSuccessors() == 1)
1181 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
1184 DEBUG(dbgs() << "\nGetting weights for branch at line "
1185 << TI->getDebugLoc().getLine() << ".\n");
1186 SmallVector<uint32_t, 4> Weights;
1187 uint32_t MaxWeight = 0;
1188 DebugLoc MaxDestLoc;
1189 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
1190 BasicBlock *Succ = TI->getSuccessor(I);
1191 Edge E = std::make_pair(BB, Succ);
1192 uint64_t Weight = EdgeWeights[E];
1193 DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
1194 // Use uint32_t saturated arithmetic to adjust the incoming weights,
1195 // if needed. Sample counts in profiles are 64-bit unsigned values,
1196 // but internally branch weights are expressed as 32-bit values.
1197 if (Weight > std::numeric_limits<uint32_t>::max()) {
1198 DEBUG(dbgs() << " (saturated due to uint32_t overflow)");
1199 Weight = std::numeric_limits<uint32_t>::max();
1201 // Weight is added by one to avoid propagation errors introduced by
1203 Weights.push_back(static_cast<uint32_t>(Weight + 1));
1205 if (Weight > MaxWeight) {
1207 MaxDestLoc = Succ->getFirstNonPHIOrDbgOrLifetime()->getDebugLoc();
1212 uint64_t TempWeight;
1213 // Only set weights if there is at least one non-zero weight.
1214 // In any other case, let the analyzer set weights.
1215 // Do not set weights if the weights are present. In ThinLTO, the profile
1216 // annotation is done twice. If the first annotation already set the
1217 // weights, the second pass does not need to set it.
1218 if (MaxWeight > 0 && !TI->extractProfTotalWeight(TempWeight)) {
1219 DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
1220 TI->setMetadata(llvm::LLVMContext::MD_prof,
1221 MDB.createBranchWeights(Weights));
1222 DebugLoc BranchLoc = TI->getDebugLoc();
1223 emitOptimizationRemark(
1224 Ctx, DEBUG_TYPE, F, MaxDestLoc,
1225 Twine("most popular destination for conditional branches at ") +
1226 ((BranchLoc) ? Twine(BranchLoc->getFilename() + ":" +
1227 Twine(BranchLoc.getLine()) + ":" +
1228 Twine(BranchLoc.getCol()))
1229 : Twine("<UNKNOWN LOCATION>")));
1231 DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
1236 /// \brief Get the line number for the function header.
1238 /// This looks up function \p F in the current compilation unit and
1239 /// retrieves the line number where the function is defined. This is
1240 /// line 0 for all the samples read from the profile file. Every line
1241 /// number is relative to this line.
1243 /// \param F Function object to query.
1245 /// \returns the line number where \p F is defined. If it returns 0,
1246 /// it means that there is no debug information available for \p F.
1247 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
1248 if (DISubprogram *S = F.getSubprogram())
1249 return S->getLine();
1251 // If the start of \p F is missing, emit a diagnostic to inform the user
1252 // about the missed opportunity.
1253 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1254 "No debug information found in function " + F.getName() +
1255 ": Function profile not used",
1260 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
1261 DT.reset(new DominatorTree);
1264 PDT.reset(new DominatorTreeBase<BasicBlock>(true));
1265 PDT->recalculate(F);
1267 LI.reset(new LoopInfo);
1271 /// \brief Generate branch weight metadata for all branches in \p F.
1273 /// Branch weights are computed out of instruction samples using a
1274 /// propagation heuristic. Propagation proceeds in 3 phases:
1276 /// 1- Assignment of block weights. All the basic blocks in the function
1277 /// are initial assigned the same weight as their most frequently
1278 /// executed instruction.
1280 /// 2- Creation of equivalence classes. Since samples may be missing from
1281 /// blocks, we can fill in the gaps by setting the weights of all the
1282 /// blocks in the same equivalence class to the same weight. To compute
1283 /// the concept of equivalence, we use dominance and loop information.
1284 /// Two blocks B1 and B2 are in the same equivalence class if B1
1285 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
1287 /// 3- Propagation of block weights into edges. This uses a simple
1288 /// propagation heuristic. The following rules are applied to every
1289 /// block BB in the CFG:
1291 /// - If BB has a single predecessor/successor, then the weight
1292 /// of that edge is the weight of the block.
1294 /// - If all the edges are known except one, and the weight of the
1295 /// block is already known, the weight of the unknown edge will
1296 /// be the weight of the block minus the sum of all the known
1297 /// edges. If the sum of all the known edges is larger than BB's weight,
1298 /// we set the unknown edge weight to zero.
1300 /// - If there is a self-referential edge, and the weight of the block is
1301 /// known, the weight for that edge is set to the weight of the block
1302 /// minus the weight of the other incoming edges to that block (if
1305 /// Since this propagation is not guaranteed to finalize for every CFG, we
1306 /// only allow it to proceed for a limited number of iterations (controlled
1307 /// by -sample-profile-max-propagate-iterations).
1309 /// FIXME: Try to replace this propagation heuristic with a scheme
1310 /// that is guaranteed to finalize. A work-list approach similar to
1311 /// the standard value propagation algorithm used by SSA-CCP might
1314 /// Once all the branch weights are computed, we emit the MD_prof
1315 /// metadata on BB using the computed values for each of its branches.
1317 /// \param F The function to query.
1319 /// \returns true if \p F was modified. Returns false, otherwise.
1320 bool SampleProfileLoader::emitAnnotations(Function &F) {
1321 bool Changed = false;
1323 if (getFunctionLoc(F) == 0)
1326 DEBUG(dbgs() << "Line number for the first instruction in " << F.getName()
1327 << ": " << getFunctionLoc(F) << "\n");
1329 DenseSet<GlobalValue::GUID> ImportGUIDs;
1330 Changed |= inlineHotFunctions(F, ImportGUIDs);
1332 // Compute basic block weights.
1333 Changed |= computeBlockWeights(F);
1336 // Add an entry count to the function using the samples gathered at the
1337 // function entry. Also sets the GUIDs that comes from a different
1338 // module but inlined in the profiled binary. This is aiming at making
1339 // the IR match the profiled binary before annotation.
1340 F.setEntryCount(Samples->getHeadSamples() + 1, &ImportGUIDs);
1342 // Compute dominance and loop info needed for propagation.
1343 computeDominanceAndLoopInfo(F);
1345 // Find equivalence classes.
1346 findEquivalenceClasses(F);
1348 // Propagate weights to all edges.
1349 propagateWeights(F);
1352 // If coverage checking was requested, compute it now.
1353 if (SampleProfileRecordCoverage) {
1354 unsigned Used = CoverageTracker.countUsedRecords(Samples);
1355 unsigned Total = CoverageTracker.countBodyRecords(Samples);
1356 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1357 if (Coverage < SampleProfileRecordCoverage) {
1358 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1359 F.getSubprogram()->getFilename(), getFunctionLoc(F),
1360 Twine(Used) + " of " + Twine(Total) + " available profile records (" +
1361 Twine(Coverage) + "%) were applied",
1366 if (SampleProfileSampleCoverage) {
1367 uint64_t Used = CoverageTracker.getTotalUsedSamples();
1368 uint64_t Total = CoverageTracker.countBodySamples(Samples);
1369 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1370 if (Coverage < SampleProfileSampleCoverage) {
1371 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1372 F.getSubprogram()->getFilename(), getFunctionLoc(F),
1373 Twine(Used) + " of " + Twine(Total) + " available profile samples (" +
1374 Twine(Coverage) + "%) were applied",
1381 char SampleProfileLoaderLegacyPass::ID = 0;
1382 INITIALIZE_PASS_BEGIN(SampleProfileLoaderLegacyPass, "sample-profile",
1383 "Sample Profile loader", false, false)
1384 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1385 INITIALIZE_PASS_END(SampleProfileLoaderLegacyPass, "sample-profile",
1386 "Sample Profile loader", false, false)
1388 bool SampleProfileLoader::doInitialization(Module &M) {
1389 auto &Ctx = M.getContext();
1390 auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
1391 if (std::error_code EC = ReaderOrErr.getError()) {
1392 std::string Msg = "Could not open profile: " + EC.message();
1393 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg));
1396 Reader = std::move(ReaderOrErr.get());
1397 ProfileIsValid = (Reader->read() == sampleprof_error::success);
1401 ModulePass *llvm::createSampleProfileLoaderPass() {
1402 return new SampleProfileLoaderLegacyPass(SampleProfileFile);
1405 ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) {
1406 return new SampleProfileLoaderLegacyPass(Name);
1409 bool SampleProfileLoader::runOnModule(Module &M) {
1410 if (!ProfileIsValid)
1413 // Compute the total number of samples collected in this profile.
1414 for (const auto &I : Reader->getProfiles())
1415 TotalCollectedSamples += I.second.getTotalSamples();
1417 bool retval = false;
1419 if (!F.isDeclaration()) {
1420 clearFunctionData();
1421 retval |= runOnFunction(F);
1423 if (M.getProfileSummary() == nullptr)
1424 M.setProfileSummary(Reader->getSummary().getMD(M.getContext()));
1428 bool SampleProfileLoaderLegacyPass::runOnModule(Module &M) {
1429 // FIXME: pass in AssumptionCache correctly for the new pass manager.
1430 SampleLoader.setACT(&getAnalysis<AssumptionCacheTracker>());
1431 return SampleLoader.runOnModule(M);
1434 bool SampleProfileLoader::runOnFunction(Function &F) {
1436 Samples = Reader->getSamplesFor(F);
1437 if (Samples && !Samples->empty())
1438 return emitAnnotations(F);
1442 PreservedAnalyses SampleProfileLoaderPass::run(Module &M,
1443 ModuleAnalysisManager &AM) {
1445 SampleProfileLoader SampleLoader(SampleProfileFile);
1447 SampleLoader.doInitialization(M);
1449 if (!SampleLoader.runOnModule(M))
1450 return PreservedAnalyses::all();
1452 return PreservedAnalyses::none();