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/IR/ValueSymbolTable.h"
47 #include "llvm/Pass.h"
48 #include "llvm/ProfileData/InstrProf.h"
49 #include "llvm/ProfileData/SampleProfReader.h"
50 #include "llvm/Support/CommandLine.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorOr.h"
53 #include "llvm/Support/Format.h"
54 #include "llvm/Support/raw_ostream.h"
55 #include "llvm/Transforms/IPO.h"
56 #include "llvm/Transforms/Instrumentation.h"
57 #include "llvm/Transforms/Utils/Cloning.h"
61 using namespace sampleprof;
63 #define DEBUG_TYPE "sample-profile"
65 // Command line option to specify the file to read samples from. This is
66 // mainly used for debugging.
67 static cl::opt<std::string> SampleProfileFile(
68 "sample-profile-file", cl::init(""), cl::value_desc("filename"),
69 cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
70 static cl::opt<unsigned> SampleProfileMaxPropagateIterations(
71 "sample-profile-max-propagate-iterations", cl::init(100),
72 cl::desc("Maximum number of iterations to go through when propagating "
73 "sample block/edge weights through the CFG."));
74 static cl::opt<unsigned> SampleProfileRecordCoverage(
75 "sample-profile-check-record-coverage", cl::init(0), cl::value_desc("N"),
76 cl::desc("Emit a warning if less than N% of records in the input profile "
77 "are matched to the IR."));
78 static cl::opt<unsigned> SampleProfileSampleCoverage(
79 "sample-profile-check-sample-coverage", cl::init(0), cl::value_desc("N"),
80 cl::desc("Emit a warning if less than N% of samples in the input profile "
81 "are matched to the IR."));
82 static cl::opt<double> SampleProfileHotThreshold(
83 "sample-profile-inline-hot-threshold", cl::init(0.1), cl::value_desc("N"),
84 cl::desc("Inlined functions that account for more than N% of all samples "
85 "collected in the parent function, will be inlined again."));
88 typedef DenseMap<const BasicBlock *, uint64_t> BlockWeightMap;
89 typedef DenseMap<const BasicBlock *, const BasicBlock *> EquivalenceClassMap;
90 typedef std::pair<const BasicBlock *, const BasicBlock *> Edge;
91 typedef DenseMap<Edge, uint64_t> EdgeWeightMap;
92 typedef DenseMap<const BasicBlock *, SmallVector<const BasicBlock *, 8>>
95 class SampleCoverageTracker {
97 SampleCoverageTracker() : SampleCoverage(), TotalUsedSamples(0) {}
99 bool markSamplesUsed(const FunctionSamples *FS, uint32_t LineOffset,
100 uint32_t Discriminator, uint64_t Samples);
101 unsigned computeCoverage(unsigned Used, unsigned Total) const;
102 unsigned countUsedRecords(const FunctionSamples *FS) const;
103 unsigned countBodyRecords(const FunctionSamples *FS) const;
104 uint64_t getTotalUsedSamples() const { return TotalUsedSamples; }
105 uint64_t countBodySamples(const FunctionSamples *FS) const;
107 SampleCoverage.clear();
108 TotalUsedSamples = 0;
112 typedef std::map<LineLocation, unsigned> BodySampleCoverageMap;
113 typedef DenseMap<const FunctionSamples *, BodySampleCoverageMap>
114 FunctionSamplesCoverageMap;
116 /// Coverage map for sampling records.
118 /// This map keeps a record of sampling records that have been matched to
119 /// an IR instruction. This is used to detect some form of staleness in
120 /// profiles (see flag -sample-profile-check-coverage).
122 /// Each entry in the map corresponds to a FunctionSamples instance. This is
123 /// another map that counts how many times the sample record at the
124 /// given location has been used.
125 FunctionSamplesCoverageMap SampleCoverage;
127 /// Number of samples used from the profile.
129 /// When a sampling record is used for the first time, the samples from
130 /// that record are added to this accumulator. Coverage is later computed
131 /// based on the total number of samples available in this function and
134 /// Note that this accumulator tracks samples used from a single function
135 /// and all the inlined callsites. Strictly, we should have a map of counters
136 /// keyed by FunctionSamples pointers, but these stats are cleared after
137 /// every function, so we just need to keep a single counter.
138 uint64_t TotalUsedSamples;
141 /// \brief Sample profile pass.
143 /// This pass reads profile data from the file specified by
144 /// -sample-profile-file and annotates every affected function with the
145 /// profile information found in that file.
146 class SampleProfileLoader {
148 SampleProfileLoader(StringRef Name = SampleProfileFile)
149 : DT(nullptr), PDT(nullptr), LI(nullptr), ACT(nullptr), Reader(),
150 Samples(nullptr), Filename(Name), ProfileIsValid(false),
151 TotalCollectedSamples(0) {}
153 bool doInitialization(Module &M);
154 bool runOnModule(Module &M);
155 void setACT(AssumptionCacheTracker *A) { ACT = A; }
157 void dump() { Reader->dump(); }
160 bool runOnFunction(Function &F);
161 unsigned getFunctionLoc(Function &F);
162 bool emitAnnotations(Function &F);
163 ErrorOr<uint64_t> getInstWeight(const Instruction &I);
164 ErrorOr<uint64_t> getBlockWeight(const BasicBlock *BB);
165 const FunctionSamples *findCalleeFunctionSamples(const Instruction &I) const;
166 std::vector<const FunctionSamples *>
167 findIndirectCallFunctionSamples(const Instruction &I) const;
168 const FunctionSamples *findFunctionSamples(const Instruction &I) const;
169 bool inlineHotFunctions(Function &F,
170 DenseSet<GlobalValue::GUID> &ImportGUIDs);
171 void printEdgeWeight(raw_ostream &OS, Edge E);
172 void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const;
173 void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB);
174 bool computeBlockWeights(Function &F);
175 void findEquivalenceClasses(Function &F);
176 void findEquivalencesFor(BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
177 DominatorTreeBase<BasicBlock> *DomTree);
178 void propagateWeights(Function &F);
179 uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
180 void buildEdges(Function &F);
181 bool propagateThroughEdges(Function &F, bool UpdateBlockCount);
182 void computeDominanceAndLoopInfo(Function &F);
183 unsigned getOffset(const DILocation *DIL) const;
184 void clearFunctionData();
186 /// \brief Map basic blocks to their computed weights.
188 /// The weight of a basic block is defined to be the maximum
189 /// of all the instruction weights in that block.
190 BlockWeightMap BlockWeights;
192 /// \brief Map edges to their computed weights.
194 /// Edge weights are computed by propagating basic block weights in
195 /// SampleProfile::propagateWeights.
196 EdgeWeightMap EdgeWeights;
198 /// \brief Set of visited blocks during propagation.
199 SmallPtrSet<const BasicBlock *, 32> VisitedBlocks;
201 /// \brief Set of visited edges during propagation.
202 SmallSet<Edge, 32> VisitedEdges;
204 /// \brief Equivalence classes for block weights.
206 /// Two blocks BB1 and BB2 are in the same equivalence class if they
207 /// dominate and post-dominate each other, and they are in the same loop
208 /// nest. When this happens, the two blocks are guaranteed to execute
209 /// the same number of times.
210 EquivalenceClassMap EquivalenceClass;
212 /// Map from function name to Function *. Used to find the function from
213 /// the function name. If the function name contains suffix, additional
214 /// entry is added to map from the stripped name to the function if there
215 /// is one-to-one mapping.
216 StringMap<Function *> SymbolMap;
218 /// \brief Dominance, post-dominance and loop information.
219 std::unique_ptr<DominatorTree> DT;
220 std::unique_ptr<DominatorTreeBase<BasicBlock>> PDT;
221 std::unique_ptr<LoopInfo> LI;
223 AssumptionCacheTracker *ACT;
225 /// \brief Predecessors for each basic block in the CFG.
226 BlockEdgeMap Predecessors;
228 /// \brief Successors for each basic block in the CFG.
229 BlockEdgeMap Successors;
231 SampleCoverageTracker CoverageTracker;
233 /// \brief Profile reader object.
234 std::unique_ptr<SampleProfileReader> Reader;
236 /// \brief Samples collected for the body of this function.
237 FunctionSamples *Samples;
239 /// \brief Name of the profile file to load.
240 std::string Filename;
242 /// \brief Flag indicating whether the profile input loaded successfully.
245 /// \brief Total number of samples collected in this profile.
247 /// This is the sum of all the samples collected in all the functions executed
249 uint64_t TotalCollectedSamples;
252 class SampleProfileLoaderLegacyPass : public ModulePass {
254 // Class identification, replacement for typeinfo
257 SampleProfileLoaderLegacyPass(StringRef Name = SampleProfileFile)
258 : ModulePass(ID), SampleLoader(Name) {
259 initializeSampleProfileLoaderLegacyPassPass(
260 *PassRegistry::getPassRegistry());
263 void dump() { SampleLoader.dump(); }
265 bool doInitialization(Module &M) override {
266 return SampleLoader.doInitialization(M);
268 StringRef getPassName() const override { return "Sample profile pass"; }
269 bool runOnModule(Module &M) override;
271 void getAnalysisUsage(AnalysisUsage &AU) const override {
272 AU.addRequired<AssumptionCacheTracker>();
276 SampleProfileLoader SampleLoader;
279 /// Return true if the given callsite is hot wrt to its caller.
281 /// Functions that were inlined in the original binary will be represented
282 /// in the inline stack in the sample profile. If the profile shows that
283 /// the original inline decision was "good" (i.e., the callsite is executed
284 /// frequently), then we will recreate the inline decision and apply the
285 /// profile from the inlined callsite.
287 /// To decide whether an inlined callsite is hot, we compute the fraction
288 /// of samples used by the callsite with respect to the total number of samples
289 /// collected in the caller.
291 /// If that fraction is larger than the default given by
292 /// SampleProfileHotThreshold, the callsite will be inlined again.
293 bool callsiteIsHot(const FunctionSamples *CallerFS,
294 const FunctionSamples *CallsiteFS) {
296 return false; // The callsite was not inlined in the original binary.
298 uint64_t ParentTotalSamples = CallerFS->getTotalSamples();
299 if (ParentTotalSamples == 0)
300 return false; // Avoid division by zero.
302 uint64_t CallsiteTotalSamples = CallsiteFS->getTotalSamples();
303 if (CallsiteTotalSamples == 0)
304 return false; // Callsite is trivially cold.
306 double PercentSamples =
307 (double)CallsiteTotalSamples / (double)ParentTotalSamples * 100.0;
308 return PercentSamples >= SampleProfileHotThreshold;
312 /// Mark as used the sample record for the given function samples at
313 /// (LineOffset, Discriminator).
315 /// \returns true if this is the first time we mark the given record.
316 bool SampleCoverageTracker::markSamplesUsed(const FunctionSamples *FS,
318 uint32_t Discriminator,
320 LineLocation Loc(LineOffset, Discriminator);
321 unsigned &Count = SampleCoverage[FS][Loc];
322 bool FirstTime = (++Count == 1);
324 TotalUsedSamples += Samples;
328 /// Return the number of sample records that were applied from this profile.
330 /// This count does not include records from cold inlined callsites.
332 SampleCoverageTracker::countUsedRecords(const FunctionSamples *FS) const {
333 auto I = SampleCoverage.find(FS);
335 // The size of the coverage map for FS represents the number of records
336 // that were marked used at least once.
337 unsigned Count = (I != SampleCoverage.end()) ? I->second.size() : 0;
339 // If there are inlined callsites in this function, count the samples found
340 // in the respective bodies. However, do not bother counting callees with 0
341 // total samples, these are callees that were never invoked at runtime.
342 for (const auto &I : FS->getCallsiteSamples())
343 for (const auto &J : I.second) {
344 const FunctionSamples *CalleeSamples = &J.second;
345 if (callsiteIsHot(FS, CalleeSamples))
346 Count += countUsedRecords(CalleeSamples);
352 /// Return the number of sample records in the body of this profile.
354 /// This count does not include records from cold inlined callsites.
356 SampleCoverageTracker::countBodyRecords(const FunctionSamples *FS) const {
357 unsigned Count = FS->getBodySamples().size();
359 // Only count records in hot callsites.
360 for (const auto &I : FS->getCallsiteSamples())
361 for (const auto &J : I.second) {
362 const FunctionSamples *CalleeSamples = &J.second;
363 if (callsiteIsHot(FS, CalleeSamples))
364 Count += countBodyRecords(CalleeSamples);
370 /// Return the number of samples collected in the body of this profile.
372 /// This count does not include samples from cold inlined callsites.
374 SampleCoverageTracker::countBodySamples(const FunctionSamples *FS) const {
376 for (const auto &I : FS->getBodySamples())
377 Total += I.second.getSamples();
379 // Only count samples in hot callsites.
380 for (const auto &I : FS->getCallsiteSamples())
381 for (const auto &J : I.second) {
382 const FunctionSamples *CalleeSamples = &J.second;
383 if (callsiteIsHot(FS, CalleeSamples))
384 Total += countBodySamples(CalleeSamples);
390 /// Return the fraction of sample records used in this profile.
392 /// The returned value is an unsigned integer in the range 0-100 indicating
393 /// the percentage of sample records that were used while applying this
394 /// profile to the associated function.
395 unsigned SampleCoverageTracker::computeCoverage(unsigned Used,
396 unsigned Total) const {
397 assert(Used <= Total &&
398 "number of used records cannot exceed the total number of records");
399 return Total > 0 ? Used * 100 / Total : 100;
402 /// Clear all the per-function data used to load samples and propagate weights.
403 void SampleProfileLoader::clearFunctionData() {
404 BlockWeights.clear();
406 VisitedBlocks.clear();
407 VisitedEdges.clear();
408 EquivalenceClass.clear();
412 Predecessors.clear();
414 CoverageTracker.clear();
417 /// Returns the line offset to the start line of the subprogram.
418 /// We assume that a single function will not exceed 65535 LOC.
419 unsigned SampleProfileLoader::getOffset(const DILocation *DIL) const {
420 return (DIL->getLine() - DIL->getScope()->getSubprogram()->getLine()) &
424 /// \brief Print the weight of edge \p E on stream \p OS.
426 /// \param OS Stream to emit the output to.
427 /// \param E Edge to print.
428 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
429 OS << "weight[" << E.first->getName() << "->" << E.second->getName()
430 << "]: " << EdgeWeights[E] << "\n";
433 /// \brief Print the equivalence class of block \p BB on stream \p OS.
435 /// \param OS Stream to emit the output to.
436 /// \param BB Block to print.
437 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
438 const BasicBlock *BB) {
439 const BasicBlock *Equiv = EquivalenceClass[BB];
440 OS << "equivalence[" << BB->getName()
441 << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
444 /// \brief Print the weight of block \p BB on stream \p OS.
446 /// \param OS Stream to emit the output to.
447 /// \param BB Block to print.
448 void SampleProfileLoader::printBlockWeight(raw_ostream &OS,
449 const BasicBlock *BB) const {
450 const auto &I = BlockWeights.find(BB);
451 uint64_t W = (I == BlockWeights.end() ? 0 : I->second);
452 OS << "weight[" << BB->getName() << "]: " << W << "\n";
455 /// \brief Get the weight for an instruction.
457 /// The "weight" of an instruction \p Inst is the number of samples
458 /// collected on that instruction at runtime. To retrieve it, we
459 /// need to compute the line number of \p Inst relative to the start of its
460 /// function. We use HeaderLineno to compute the offset. We then
461 /// look up the samples collected for \p Inst using BodySamples.
463 /// \param Inst Instruction to query.
465 /// \returns the weight of \p Inst.
466 ErrorOr<uint64_t> SampleProfileLoader::getInstWeight(const Instruction &Inst) {
467 const DebugLoc &DLoc = Inst.getDebugLoc();
469 return std::error_code();
471 const FunctionSamples *FS = findFunctionSamples(Inst);
473 return std::error_code();
475 // Ignore all intrinsics and branch instructions.
476 // Branch instruction usually contains debug info from sources outside of
477 // the residing basic block, thus we ignore them during annotation.
478 if (isa<BranchInst>(Inst) || isa<IntrinsicInst>(Inst))
479 return std::error_code();
481 // If a call/invoke instruction is inlined in profile, but not inlined here,
482 // it means that the inlined callsite has no sample, thus the call
483 // instruction should have 0 count.
484 if ((isa<CallInst>(Inst) || isa<InvokeInst>(Inst)) &&
485 findCalleeFunctionSamples(Inst))
488 const DILocation *DIL = DLoc;
489 uint32_t LineOffset = getOffset(DIL);
490 uint32_t Discriminator = DIL->getBaseDiscriminator();
491 ErrorOr<uint64_t> R = FS->findSamplesAt(LineOffset, Discriminator);
494 CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get());
496 const Function *F = Inst.getParent()->getParent();
497 LLVMContext &Ctx = F->getContext();
498 emitOptimizationRemark(
499 Ctx, DEBUG_TYPE, *F, DLoc,
500 Twine("Applied ") + Twine(*R) +
501 " samples from profile (offset: " + Twine(LineOffset) +
502 ((Discriminator) ? Twine(".") + Twine(Discriminator) : "") + ")");
504 DEBUG(dbgs() << " " << DLoc.getLine() << "."
505 << DIL->getBaseDiscriminator() << ":" << Inst
506 << " (line offset: " << LineOffset << "."
507 << DIL->getBaseDiscriminator() << " - weight: " << R.get()
513 /// \brief Compute the weight of a basic block.
515 /// The weight of basic block \p BB is the maximum weight of all the
516 /// instructions in BB.
518 /// \param BB The basic block to query.
520 /// \returns the weight for \p BB.
521 ErrorOr<uint64_t> SampleProfileLoader::getBlockWeight(const BasicBlock *BB) {
523 bool HasWeight = false;
524 for (auto &I : BB->getInstList()) {
525 const ErrorOr<uint64_t> &R = getInstWeight(I);
527 Max = std::max(Max, R.get());
531 return HasWeight ? ErrorOr<uint64_t>(Max) : std::error_code();
534 /// \brief Compute and store the weights of every basic block.
536 /// This populates the BlockWeights map by computing
537 /// the weights of every basic block in the CFG.
539 /// \param F The function to query.
540 bool SampleProfileLoader::computeBlockWeights(Function &F) {
541 bool Changed = false;
542 DEBUG(dbgs() << "Block weights\n");
543 for (const auto &BB : F) {
544 ErrorOr<uint64_t> Weight = getBlockWeight(&BB);
546 BlockWeights[&BB] = Weight.get();
547 VisitedBlocks.insert(&BB);
550 DEBUG(printBlockWeight(dbgs(), &BB));
556 /// \brief Get the FunctionSamples for a call instruction.
558 /// The FunctionSamples of a call/invoke instruction \p Inst is the inlined
559 /// instance in which that call instruction is calling to. It contains
560 /// all samples that resides in the inlined instance. We first find the
561 /// inlined instance in which the call instruction is from, then we
562 /// traverse its children to find the callsite with the matching
565 /// \param Inst Call/Invoke instruction to query.
567 /// \returns The FunctionSamples pointer to the inlined instance.
568 const FunctionSamples *
569 SampleProfileLoader::findCalleeFunctionSamples(const Instruction &Inst) const {
570 const DILocation *DIL = Inst.getDebugLoc();
575 StringRef CalleeName;
576 if (const CallInst *CI = dyn_cast<CallInst>(&Inst))
577 if (Function *Callee = CI->getCalledFunction())
578 CalleeName = Callee->getName();
580 const FunctionSamples *FS = findFunctionSamples(Inst);
584 return FS->findFunctionSamplesAt(
585 LineLocation(getOffset(DIL), DIL->getBaseDiscriminator()), CalleeName);
588 /// Returns a vector of FunctionSamples that are the indirect call targets
589 /// of \p Inst. The vector is sorted by the total number of samples.
590 std::vector<const FunctionSamples *>
591 SampleProfileLoader::findIndirectCallFunctionSamples(
592 const Instruction &Inst) const {
593 const DILocation *DIL = Inst.getDebugLoc();
594 std::vector<const FunctionSamples *> R;
600 const FunctionSamples *FS = findFunctionSamples(Inst);
604 if (const FunctionSamplesMap *M = FS->findFunctionSamplesMapAt(
605 LineLocation(getOffset(DIL), DIL->getBaseDiscriminator()))) {
608 for (const auto &NameFS : *M) {
609 R.push_back(&NameFS.second);
611 std::sort(R.begin(), R.end(),
612 [](const FunctionSamples *L, const FunctionSamples *R) {
613 return L->getTotalSamples() > R->getTotalSamples();
619 /// \brief Get the FunctionSamples for an instruction.
621 /// The FunctionSamples of an instruction \p Inst is the inlined instance
622 /// in which that instruction is coming from. We traverse the inline stack
623 /// of that instruction, and match it with the tree nodes in the profile.
625 /// \param Inst Instruction to query.
627 /// \returns the FunctionSamples pointer to the inlined instance.
628 const FunctionSamples *
629 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
630 SmallVector<std::pair<LineLocation, StringRef>, 10> S;
631 const DILocation *DIL = Inst.getDebugLoc();
635 const DILocation *PrevDIL = DIL;
636 for (DIL = DIL->getInlinedAt(); DIL; DIL = DIL->getInlinedAt()) {
637 S.push_back(std::make_pair(
638 LineLocation(getOffset(DIL), DIL->getBaseDiscriminator()),
639 PrevDIL->getScope()->getSubprogram()->getLinkageName()));
644 const FunctionSamples *FS = Samples;
645 for (int i = S.size() - 1; i >= 0 && FS != nullptr; i--) {
646 FS = FS->findFunctionSamplesAt(S[i].first, S[i].second);
651 /// \brief Iteratively inline hot callsites of a function.
653 /// Iteratively traverse all callsites of the function \p F, and find if
654 /// the corresponding inlined instance exists and is hot in profile. If
655 /// it is hot enough, inline the callsites and adds new callsites of the
656 /// callee into the caller. If the call is an indirect call, first promote
657 /// it to direct call. Each indirect call is limited with a single target.
659 /// \param F function to perform iterative inlining.
660 /// \param ImportGUIDs a set to be updated to include all GUIDs that come
661 /// from a different module but inlined in the profiled binary.
663 /// \returns True if there is any inline happened.
664 bool SampleProfileLoader::inlineHotFunctions(
665 Function &F, DenseSet<GlobalValue::GUID> &ImportGUIDs) {
666 DenseSet<Instruction *> PromotedInsns;
667 bool Changed = false;
668 LLVMContext &Ctx = F.getContext();
669 std::function<AssumptionCache &(Function &)> GetAssumptionCache = [&](
670 Function &F) -> AssumptionCache & { return ACT->getAssumptionCache(F); };
672 bool LocalChanged = false;
673 SmallVector<Instruction *, 10> CIS;
676 SmallVector<Instruction *, 10> Candidates;
677 for (auto &I : BB.getInstList()) {
678 const FunctionSamples *FS = nullptr;
679 if ((isa<CallInst>(I) || isa<InvokeInst>(I)) &&
680 !isa<IntrinsicInst>(I) && (FS = findCalleeFunctionSamples(I))) {
681 Candidates.push_back(&I);
682 if (callsiteIsHot(Samples, FS))
687 CIS.insert(CIS.begin(), Candidates.begin(), Candidates.end());
691 InlineFunctionInfo IFI(nullptr, ACT ? &GetAssumptionCache : nullptr);
692 Function *CalledFunction = CallSite(I).getCalledFunction();
694 if (!CalledFunction && !PromotedInsns.count(I) &&
695 CallSite(I).isIndirectCall())
696 for (const auto *FS : findIndirectCallFunctionSamples(*I)) {
697 auto CalleeFunctionName = FS->getName();
698 const char *Reason = "Callee function not available";
699 auto R = SymbolMap.find(CalleeFunctionName);
700 if (R == SymbolMap.end())
702 CalledFunction = R->getValue();
703 if (CalledFunction && isLegalToPromote(I, CalledFunction, &Reason)) {
704 // The indirect target was promoted and inlined in the profile, as a
705 // result, we do not have profile info for the branch probability.
706 // We set the probability to 80% taken to indicate that the static
707 // call is likely taken.
708 DI = dyn_cast<Instruction>(
709 promoteIndirectCall(I, CalledFunction, 80, 100, false)
710 ->stripPointerCasts());
711 PromotedInsns.insert(I);
713 DEBUG(dbgs() << "\nFailed to promote indirect call to "
714 << CalleeFunctionName << " because " << Reason
719 if (!CalledFunction || !CalledFunction->getSubprogram()) {
720 findCalleeFunctionSamples(*I)->findImportedFunctions(
721 ImportGUIDs, F.getParent(),
722 Samples->getTotalSamples() * SampleProfileHotThreshold / 100);
725 DebugLoc DLoc = I->getDebugLoc();
726 if (InlineFunction(CallSite(DI), IFI)) {
728 emitOptimizationRemark(Ctx, DEBUG_TYPE, F, DLoc,
729 Twine("inlined hot callee '") +
730 CalledFunction->getName() + "' into '" +
743 /// \brief Find equivalence classes for the given block.
745 /// This finds all the blocks that are guaranteed to execute the same
746 /// number of times as \p BB1. To do this, it traverses all the
747 /// descendants of \p BB1 in the dominator or post-dominator tree.
749 /// A block BB2 will be in the same equivalence class as \p BB1 if
750 /// the following holds:
752 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
753 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
754 /// dominate BB1 in the post-dominator tree.
756 /// 2- Both BB2 and \p BB1 must be in the same loop.
758 /// For every block BB2 that meets those two requirements, we set BB2's
759 /// equivalence class to \p BB1.
761 /// \param BB1 Block to check.
762 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
763 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
764 /// with blocks from \p BB1's dominator tree, then
765 /// this is the post-dominator tree, and vice versa.
766 void SampleProfileLoader::findEquivalencesFor(
767 BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
768 DominatorTreeBase<BasicBlock> *DomTree) {
769 const BasicBlock *EC = EquivalenceClass[BB1];
770 uint64_t Weight = BlockWeights[EC];
771 for (const auto *BB2 : Descendants) {
772 bool IsDomParent = DomTree->dominates(BB2, BB1);
773 bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
774 if (BB1 != BB2 && IsDomParent && IsInSameLoop) {
775 EquivalenceClass[BB2] = EC;
776 // If BB2 is visited, then the entire EC should be marked as visited.
777 if (VisitedBlocks.count(BB2)) {
778 VisitedBlocks.insert(EC);
781 // If BB2 is heavier than BB1, make BB2 have the same weight
784 // Note that we don't worry about the opposite situation here
785 // (when BB2 is lighter than BB1). We will deal with this
786 // during the propagation phase. Right now, we just want to
787 // make sure that BB1 has the largest weight of all the
788 // members of its equivalence set.
789 Weight = std::max(Weight, BlockWeights[BB2]);
792 if (EC == &EC->getParent()->getEntryBlock()) {
793 BlockWeights[EC] = Samples->getHeadSamples() + 1;
795 BlockWeights[EC] = Weight;
799 /// \brief Find equivalence classes.
801 /// Since samples may be missing from blocks, we can fill in the gaps by setting
802 /// the weights of all the blocks in the same equivalence class to the same
803 /// weight. To compute the concept of equivalence, we use dominance and loop
804 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
805 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
807 /// \param F The function to query.
808 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
809 SmallVector<BasicBlock *, 8> DominatedBBs;
810 DEBUG(dbgs() << "\nBlock equivalence classes\n");
811 // Find equivalence sets based on dominance and post-dominance information.
813 BasicBlock *BB1 = &BB;
815 // Compute BB1's equivalence class once.
816 if (EquivalenceClass.count(BB1)) {
817 DEBUG(printBlockEquivalence(dbgs(), BB1));
821 // By default, blocks are in their own equivalence class.
822 EquivalenceClass[BB1] = BB1;
824 // Traverse all the blocks dominated by BB1. We are looking for
825 // every basic block BB2 such that:
827 // 1- BB1 dominates BB2.
828 // 2- BB2 post-dominates BB1.
829 // 3- BB1 and BB2 are in the same loop nest.
831 // If all those conditions hold, it means that BB2 is executed
832 // as many times as BB1, so they are placed in the same equivalence
833 // class by making BB2's equivalence class be BB1.
834 DominatedBBs.clear();
835 DT->getDescendants(BB1, DominatedBBs);
836 findEquivalencesFor(BB1, DominatedBBs, PDT.get());
838 DEBUG(printBlockEquivalence(dbgs(), BB1));
841 // Assign weights to equivalence classes.
843 // All the basic blocks in the same equivalence class will execute
844 // the same number of times. Since we know that the head block in
845 // each equivalence class has the largest weight, assign that weight
846 // to all the blocks in that equivalence class.
847 DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n");
849 const BasicBlock *BB = &BI;
850 const BasicBlock *EquivBB = EquivalenceClass[BB];
852 BlockWeights[BB] = BlockWeights[EquivBB];
853 DEBUG(printBlockWeight(dbgs(), BB));
857 /// \brief Visit the given edge to decide if it has a valid weight.
859 /// If \p E has not been visited before, we copy to \p UnknownEdge
860 /// and increment the count of unknown edges.
862 /// \param E Edge to visit.
863 /// \param NumUnknownEdges Current number of unknown edges.
864 /// \param UnknownEdge Set if E has not been visited before.
866 /// \returns E's weight, if known. Otherwise, return 0.
867 uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
869 if (!VisitedEdges.count(E)) {
870 (*NumUnknownEdges)++;
875 return EdgeWeights[E];
878 /// \brief Propagate weights through incoming/outgoing edges.
880 /// If the weight of a basic block is known, and there is only one edge
881 /// with an unknown weight, we can calculate the weight of that edge.
883 /// Similarly, if all the edges have a known count, we can calculate the
884 /// count of the basic block, if needed.
886 /// \param F Function to process.
887 /// \param UpdateBlockCount Whether we should update basic block counts that
888 /// has already been annotated.
890 /// \returns True if new weights were assigned to edges or blocks.
891 bool SampleProfileLoader::propagateThroughEdges(Function &F,
892 bool UpdateBlockCount) {
893 bool Changed = false;
894 DEBUG(dbgs() << "\nPropagation through edges\n");
895 for (const auto &BI : F) {
896 const BasicBlock *BB = &BI;
897 const BasicBlock *EC = EquivalenceClass[BB];
899 // Visit all the predecessor and successor edges to determine
900 // which ones have a weight assigned already. Note that it doesn't
901 // matter that we only keep track of a single unknown edge. The
902 // only case we are interested in handling is when only a single
903 // edge is unknown (see setEdgeOrBlockWeight).
904 for (unsigned i = 0; i < 2; i++) {
905 uint64_t TotalWeight = 0;
906 unsigned NumUnknownEdges = 0, NumTotalEdges = 0;
907 Edge UnknownEdge, SelfReferentialEdge, SingleEdge;
910 // First, visit all predecessor edges.
911 NumTotalEdges = Predecessors[BB].size();
912 for (auto *Pred : Predecessors[BB]) {
913 Edge E = std::make_pair(Pred, BB);
914 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
915 if (E.first == E.second)
916 SelfReferentialEdge = E;
918 if (NumTotalEdges == 1) {
919 SingleEdge = std::make_pair(Predecessors[BB][0], BB);
922 // On the second round, visit all successor edges.
923 NumTotalEdges = Successors[BB].size();
924 for (auto *Succ : Successors[BB]) {
925 Edge E = std::make_pair(BB, Succ);
926 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
928 if (NumTotalEdges == 1) {
929 SingleEdge = std::make_pair(BB, Successors[BB][0]);
933 // After visiting all the edges, there are three cases that we
934 // can handle immediately:
936 // - All the edge weights are known (i.e., NumUnknownEdges == 0).
937 // In this case, we simply check that the sum of all the edges
938 // is the same as BB's weight. If not, we change BB's weight
939 // to match. Additionally, if BB had not been visited before,
940 // we mark it visited.
942 // - Only one edge is unknown and BB has already been visited.
943 // In this case, we can compute the weight of the edge by
944 // subtracting the total block weight from all the known
945 // edge weights. If the edges weight more than BB, then the
946 // edge of the last remaining edge is set to zero.
948 // - There exists a self-referential edge and the weight of BB is
949 // known. In this case, this edge can be based on BB's weight.
950 // We add up all the other known edges and set the weight on
951 // the self-referential edge as we did in the previous case.
953 // In any other case, we must continue iterating. Eventually,
954 // all edges will get a weight, or iteration will stop when
955 // it reaches SampleProfileMaxPropagateIterations.
956 if (NumUnknownEdges <= 1) {
957 uint64_t &BBWeight = BlockWeights[EC];
958 if (NumUnknownEdges == 0) {
959 if (!VisitedBlocks.count(EC)) {
960 // If we already know the weight of all edges, the weight of the
961 // basic block can be computed. It should be no larger than the sum
962 // of all edge weights.
963 if (TotalWeight > BBWeight) {
964 BBWeight = TotalWeight;
966 DEBUG(dbgs() << "All edge weights for " << BB->getName()
967 << " known. Set weight for block: ";
968 printBlockWeight(dbgs(), BB););
970 } else if (NumTotalEdges == 1 &&
971 EdgeWeights[SingleEdge] < BlockWeights[EC]) {
972 // If there is only one edge for the visited basic block, use the
973 // block weight to adjust edge weight if edge weight is smaller.
974 EdgeWeights[SingleEdge] = BlockWeights[EC];
977 } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) {
978 // If there is a single unknown edge and the block has been
979 // visited, then we can compute E's weight.
980 if (BBWeight >= TotalWeight)
981 EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
983 EdgeWeights[UnknownEdge] = 0;
984 const BasicBlock *OtherEC;
986 OtherEC = EquivalenceClass[UnknownEdge.first];
988 OtherEC = EquivalenceClass[UnknownEdge.second];
989 // Edge weights should never exceed the BB weights it connects.
990 if (VisitedBlocks.count(OtherEC) &&
991 EdgeWeights[UnknownEdge] > BlockWeights[OtherEC])
992 EdgeWeights[UnknownEdge] = BlockWeights[OtherEC];
993 VisitedEdges.insert(UnknownEdge);
995 DEBUG(dbgs() << "Set weight for edge: ";
996 printEdgeWeight(dbgs(), UnknownEdge));
998 } else if (VisitedBlocks.count(EC) && BlockWeights[EC] == 0) {
999 // If a block Weights 0, all its in/out edges should weight 0.
1001 for (auto *Pred : Predecessors[BB]) {
1002 Edge E = std::make_pair(Pred, BB);
1004 VisitedEdges.insert(E);
1007 for (auto *Succ : Successors[BB]) {
1008 Edge E = std::make_pair(BB, Succ);
1010 VisitedEdges.insert(E);
1013 } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) {
1014 uint64_t &BBWeight = BlockWeights[BB];
1015 // We have a self-referential edge and the weight of BB is known.
1016 if (BBWeight >= TotalWeight)
1017 EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
1019 EdgeWeights[SelfReferentialEdge] = 0;
1020 VisitedEdges.insert(SelfReferentialEdge);
1022 DEBUG(dbgs() << "Set self-referential edge weight to: ";
1023 printEdgeWeight(dbgs(), SelfReferentialEdge));
1025 if (UpdateBlockCount && !VisitedBlocks.count(EC) && TotalWeight > 0) {
1026 BlockWeights[EC] = TotalWeight;
1027 VisitedBlocks.insert(EC);
1036 /// \brief Build in/out edge lists for each basic block in the CFG.
1038 /// We are interested in unique edges. If a block B1 has multiple
1039 /// edges to another block B2, we only add a single B1->B2 edge.
1040 void SampleProfileLoader::buildEdges(Function &F) {
1041 for (auto &BI : F) {
1042 BasicBlock *B1 = &BI;
1044 // Add predecessors for B1.
1045 SmallPtrSet<BasicBlock *, 16> Visited;
1046 if (!Predecessors[B1].empty())
1047 llvm_unreachable("Found a stale predecessors list in a basic block.");
1048 for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
1049 BasicBlock *B2 = *PI;
1050 if (Visited.insert(B2).second)
1051 Predecessors[B1].push_back(B2);
1054 // Add successors for B1.
1056 if (!Successors[B1].empty())
1057 llvm_unreachable("Found a stale successors list in a basic block.");
1058 for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
1059 BasicBlock *B2 = *SI;
1060 if (Visited.insert(B2).second)
1061 Successors[B1].push_back(B2);
1066 /// Sorts the CallTargetMap \p M by count in descending order and stores the
1067 /// sorted result in \p Sorted. Returns the total counts.
1068 static uint64_t SortCallTargets(SmallVector<InstrProfValueData, 2> &Sorted,
1069 const SampleRecord::CallTargetMap &M) {
1072 for (auto I = M.begin(); I != M.end(); ++I) {
1073 Sum += I->getValue();
1074 Sorted.push_back({Function::getGUID(I->getKey()), I->getValue()});
1076 std::sort(Sorted.begin(), Sorted.end(),
1077 [](const InstrProfValueData &L, const InstrProfValueData &R) {
1078 if (L.Count == R.Count)
1079 return L.Value > R.Value;
1081 return L.Count > R.Count;
1086 /// \brief Propagate weights into edges
1088 /// The following rules are applied to every block BB in the CFG:
1090 /// - If BB has a single predecessor/successor, then the weight
1091 /// of that edge is the weight of the block.
1093 /// - If all incoming or outgoing edges are known except one, and the
1094 /// weight of the block is already known, the weight of the unknown
1095 /// edge will be the weight of the block minus the sum of all the known
1096 /// edges. If the sum of all the known edges is larger than BB's weight,
1097 /// we set the unknown edge weight to zero.
1099 /// - If there is a self-referential edge, and the weight of the block is
1100 /// known, the weight for that edge is set to the weight of the block
1101 /// minus the weight of the other incoming edges to that block (if
1103 void SampleProfileLoader::propagateWeights(Function &F) {
1104 bool Changed = true;
1107 // If BB weight is larger than its corresponding loop's header BB weight,
1108 // use the BB weight to replace the loop header BB weight.
1109 for (auto &BI : F) {
1110 BasicBlock *BB = &BI;
1111 Loop *L = LI->getLoopFor(BB);
1115 BasicBlock *Header = L->getHeader();
1116 if (Header && BlockWeights[BB] > BlockWeights[Header]) {
1117 BlockWeights[Header] = BlockWeights[BB];
1121 // Before propagation starts, build, for each block, a list of
1122 // unique predecessors and successors. This is necessary to handle
1123 // identical edges in multiway branches. Since we visit all blocks and all
1124 // edges of the CFG, it is cleaner to build these lists once at the start
1128 // Propagate until we converge or we go past the iteration limit.
1129 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1130 Changed = propagateThroughEdges(F, false);
1133 // The first propagation propagates BB counts from annotated BBs to unknown
1134 // BBs. The 2nd propagation pass resets edges weights, and use all BB weights
1135 // to propagate edge weights.
1136 VisitedEdges.clear();
1138 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1139 Changed = propagateThroughEdges(F, false);
1142 // The 3rd propagation pass allows adjust annotated BB weights that are
1145 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1146 Changed = propagateThroughEdges(F, true);
1149 // Generate MD_prof metadata for every branch instruction using the
1150 // edge weights computed during propagation.
1151 DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
1152 LLVMContext &Ctx = F.getContext();
1154 for (auto &BI : F) {
1155 BasicBlock *BB = &BI;
1157 if (BlockWeights[BB]) {
1158 for (auto &I : BB->getInstList()) {
1159 if (!isa<CallInst>(I) && !isa<InvokeInst>(I))
1162 if (!CS.getCalledFunction()) {
1163 const DebugLoc &DLoc = I.getDebugLoc();
1166 const DILocation *DIL = DLoc;
1167 uint32_t LineOffset = getOffset(DIL);
1168 uint32_t Discriminator = DIL->getBaseDiscriminator();
1170 const FunctionSamples *FS = findFunctionSamples(I);
1173 auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
1174 if (!T || T.get().size() == 0)
1176 SmallVector<InstrProfValueData, 2> SortedCallTargets;
1177 uint64_t Sum = SortCallTargets(SortedCallTargets, T.get());
1178 annotateValueSite(*I.getParent()->getParent()->getParent(), I,
1179 SortedCallTargets, Sum, IPVK_IndirectCallTarget,
1180 SortedCallTargets.size());
1181 } else if (!dyn_cast<IntrinsicInst>(&I)) {
1182 SmallVector<uint32_t, 1> Weights;
1183 Weights.push_back(BlockWeights[BB]);
1184 I.setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights));
1188 TerminatorInst *TI = BB->getTerminator();
1189 if (TI->getNumSuccessors() == 1)
1191 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
1194 DebugLoc BranchLoc = TI->getDebugLoc();
1195 DEBUG(dbgs() << "\nGetting weights for branch at line "
1196 << ((BranchLoc) ? Twine(BranchLoc.getLine())
1197 : Twine("<UNKNOWN LOCATION>"))
1199 SmallVector<uint32_t, 4> Weights;
1200 uint32_t MaxWeight = 0;
1201 DebugLoc MaxDestLoc;
1202 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
1203 BasicBlock *Succ = TI->getSuccessor(I);
1204 Edge E = std::make_pair(BB, Succ);
1205 uint64_t Weight = EdgeWeights[E];
1206 DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
1207 // Use uint32_t saturated arithmetic to adjust the incoming weights,
1208 // if needed. Sample counts in profiles are 64-bit unsigned values,
1209 // but internally branch weights are expressed as 32-bit values.
1210 if (Weight > std::numeric_limits<uint32_t>::max()) {
1211 DEBUG(dbgs() << " (saturated due to uint32_t overflow)");
1212 Weight = std::numeric_limits<uint32_t>::max();
1214 // Weight is added by one to avoid propagation errors introduced by
1216 Weights.push_back(static_cast<uint32_t>(Weight + 1));
1218 if (Weight > MaxWeight) {
1220 MaxDestLoc = Succ->getFirstNonPHIOrDbgOrLifetime()->getDebugLoc();
1225 uint64_t TempWeight;
1226 // Only set weights if there is at least one non-zero weight.
1227 // In any other case, let the analyzer set weights.
1228 // Do not set weights if the weights are present. In ThinLTO, the profile
1229 // annotation is done twice. If the first annotation already set the
1230 // weights, the second pass does not need to set it.
1231 if (MaxWeight > 0 && !TI->extractProfTotalWeight(TempWeight)) {
1232 DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
1233 TI->setMetadata(llvm::LLVMContext::MD_prof,
1234 MDB.createBranchWeights(Weights));
1235 emitOptimizationRemark(
1236 Ctx, DEBUG_TYPE, F, MaxDestLoc,
1237 Twine("most popular destination for conditional branches at ") +
1238 ((BranchLoc) ? Twine(BranchLoc->getFilename() + ":" +
1239 Twine(BranchLoc.getLine()) + ":" +
1240 Twine(BranchLoc.getCol()))
1241 : Twine("<UNKNOWN LOCATION>")));
1243 DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
1248 /// \brief Get the line number for the function header.
1250 /// This looks up function \p F in the current compilation unit and
1251 /// retrieves the line number where the function is defined. This is
1252 /// line 0 for all the samples read from the profile file. Every line
1253 /// number is relative to this line.
1255 /// \param F Function object to query.
1257 /// \returns the line number where \p F is defined. If it returns 0,
1258 /// it means that there is no debug information available for \p F.
1259 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
1260 if (DISubprogram *S = F.getSubprogram())
1261 return S->getLine();
1263 // If the start of \p F is missing, emit a diagnostic to inform the user
1264 // about the missed opportunity.
1265 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1266 "No debug information found in function " + F.getName() +
1267 ": Function profile not used",
1272 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
1273 DT.reset(new DominatorTree);
1276 PDT.reset(new DominatorTreeBase<BasicBlock>(true));
1277 PDT->recalculate(F);
1279 LI.reset(new LoopInfo);
1283 /// \brief Generate branch weight metadata for all branches in \p F.
1285 /// Branch weights are computed out of instruction samples using a
1286 /// propagation heuristic. Propagation proceeds in 3 phases:
1288 /// 1- Assignment of block weights. All the basic blocks in the function
1289 /// are initial assigned the same weight as their most frequently
1290 /// executed instruction.
1292 /// 2- Creation of equivalence classes. Since samples may be missing from
1293 /// blocks, we can fill in the gaps by setting the weights of all the
1294 /// blocks in the same equivalence class to the same weight. To compute
1295 /// the concept of equivalence, we use dominance and loop information.
1296 /// Two blocks B1 and B2 are in the same equivalence class if B1
1297 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
1299 /// 3- Propagation of block weights into edges. This uses a simple
1300 /// propagation heuristic. The following rules are applied to every
1301 /// block BB in the CFG:
1303 /// - If BB has a single predecessor/successor, then the weight
1304 /// of that edge is the weight of the block.
1306 /// - If all the edges are known except one, and the weight of the
1307 /// block is already known, the weight of the unknown edge will
1308 /// be the weight of the block minus the sum of all the known
1309 /// edges. If the sum of all the known edges is larger than BB's weight,
1310 /// we set the unknown edge weight to zero.
1312 /// - If there is a self-referential edge, and the weight of the block is
1313 /// known, the weight for that edge is set to the weight of the block
1314 /// minus the weight of the other incoming edges to that block (if
1317 /// Since this propagation is not guaranteed to finalize for every CFG, we
1318 /// only allow it to proceed for a limited number of iterations (controlled
1319 /// by -sample-profile-max-propagate-iterations).
1321 /// FIXME: Try to replace this propagation heuristic with a scheme
1322 /// that is guaranteed to finalize. A work-list approach similar to
1323 /// the standard value propagation algorithm used by SSA-CCP might
1326 /// Once all the branch weights are computed, we emit the MD_prof
1327 /// metadata on BB using the computed values for each of its branches.
1329 /// \param F The function to query.
1331 /// \returns true if \p F was modified. Returns false, otherwise.
1332 bool SampleProfileLoader::emitAnnotations(Function &F) {
1333 bool Changed = false;
1335 if (getFunctionLoc(F) == 0)
1338 DEBUG(dbgs() << "Line number for the first instruction in " << F.getName()
1339 << ": " << getFunctionLoc(F) << "\n");
1341 DenseSet<GlobalValue::GUID> ImportGUIDs;
1342 Changed |= inlineHotFunctions(F, ImportGUIDs);
1344 // Compute basic block weights.
1345 Changed |= computeBlockWeights(F);
1348 // Add an entry count to the function using the samples gathered at the
1349 // function entry. Also sets the GUIDs that comes from a different
1350 // module but inlined in the profiled binary. This is aiming at making
1351 // the IR match the profiled binary before annotation.
1352 F.setEntryCount(Samples->getHeadSamples() + 1, &ImportGUIDs);
1354 // Compute dominance and loop info needed for propagation.
1355 computeDominanceAndLoopInfo(F);
1357 // Find equivalence classes.
1358 findEquivalenceClasses(F);
1360 // Propagate weights to all edges.
1361 propagateWeights(F);
1364 // If coverage checking was requested, compute it now.
1365 if (SampleProfileRecordCoverage) {
1366 unsigned Used = CoverageTracker.countUsedRecords(Samples);
1367 unsigned Total = CoverageTracker.countBodyRecords(Samples);
1368 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1369 if (Coverage < SampleProfileRecordCoverage) {
1370 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1371 F.getSubprogram()->getFilename(), getFunctionLoc(F),
1372 Twine(Used) + " of " + Twine(Total) + " available profile records (" +
1373 Twine(Coverage) + "%) were applied",
1378 if (SampleProfileSampleCoverage) {
1379 uint64_t Used = CoverageTracker.getTotalUsedSamples();
1380 uint64_t Total = CoverageTracker.countBodySamples(Samples);
1381 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1382 if (Coverage < SampleProfileSampleCoverage) {
1383 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1384 F.getSubprogram()->getFilename(), getFunctionLoc(F),
1385 Twine(Used) + " of " + Twine(Total) + " available profile samples (" +
1386 Twine(Coverage) + "%) were applied",
1393 char SampleProfileLoaderLegacyPass::ID = 0;
1394 INITIALIZE_PASS_BEGIN(SampleProfileLoaderLegacyPass, "sample-profile",
1395 "Sample Profile loader", false, false)
1396 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1397 INITIALIZE_PASS_END(SampleProfileLoaderLegacyPass, "sample-profile",
1398 "Sample Profile loader", false, false)
1400 bool SampleProfileLoader::doInitialization(Module &M) {
1401 auto &Ctx = M.getContext();
1402 auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
1403 if (std::error_code EC = ReaderOrErr.getError()) {
1404 std::string Msg = "Could not open profile: " + EC.message();
1405 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg));
1408 Reader = std::move(ReaderOrErr.get());
1409 ProfileIsValid = (Reader->read() == sampleprof_error::success);
1413 ModulePass *llvm::createSampleProfileLoaderPass() {
1414 return new SampleProfileLoaderLegacyPass(SampleProfileFile);
1417 ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) {
1418 return new SampleProfileLoaderLegacyPass(Name);
1421 bool SampleProfileLoader::runOnModule(Module &M) {
1422 if (!ProfileIsValid)
1425 // Compute the total number of samples collected in this profile.
1426 for (const auto &I : Reader->getProfiles())
1427 TotalCollectedSamples += I.second.getTotalSamples();
1429 // Populate the symbol map.
1430 for (const auto &N_F : M.getValueSymbolTable()) {
1431 std::string OrigName = N_F.getKey();
1432 Function *F = dyn_cast<Function>(N_F.getValue());
1435 SymbolMap[OrigName] = F;
1436 auto pos = OrigName.find('.');
1437 if (pos != std::string::npos) {
1438 std::string NewName = OrigName.substr(0, pos);
1439 auto r = SymbolMap.insert(std::make_pair(NewName, F));
1440 // Failiing to insert means there is already an entry in SymbolMap,
1441 // thus there are multiple functions that are mapped to the same
1442 // stripped name. In this case of name conflicting, set the value
1443 // to nullptr to avoid confusion.
1445 r.first->second = nullptr;
1449 bool retval = false;
1451 if (!F.isDeclaration()) {
1452 clearFunctionData();
1453 retval |= runOnFunction(F);
1455 if (M.getProfileSummary() == nullptr)
1456 M.setProfileSummary(Reader->getSummary().getMD(M.getContext()));
1460 bool SampleProfileLoaderLegacyPass::runOnModule(Module &M) {
1461 // FIXME: pass in AssumptionCache correctly for the new pass manager.
1462 SampleLoader.setACT(&getAnalysis<AssumptionCacheTracker>());
1463 return SampleLoader.runOnModule(M);
1466 bool SampleProfileLoader::runOnFunction(Function &F) {
1468 Samples = Reader->getSamplesFor(F);
1469 if (Samples && !Samples->empty())
1470 return emitAnnotations(F);
1474 PreservedAnalyses SampleProfileLoaderPass::run(Module &M,
1475 ModuleAnalysisManager &AM) {
1477 SampleProfileLoader SampleLoader(SampleProfileFile);
1479 SampleLoader.doInitialization(M);
1481 if (!SampleLoader.runOnModule(M))
1482 return PreservedAnalyses::all();
1484 return PreservedAnalyses::none();