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 const FunctionSamples *findFunctionSamples(const Instruction &I) const;
166 bool inlineHotFunctions(Function &F);
167 void printEdgeWeight(raw_ostream &OS, Edge E);
168 void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const;
169 void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB);
170 bool computeBlockWeights(Function &F);
171 void findEquivalenceClasses(Function &F);
172 void findEquivalencesFor(BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
173 DominatorTreeBase<BasicBlock> *DomTree);
174 void propagateWeights(Function &F);
175 uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
176 void buildEdges(Function &F);
177 bool propagateThroughEdges(Function &F, bool UpdateBlockCount);
178 void computeDominanceAndLoopInfo(Function &F);
179 unsigned getOffset(const DILocation *DIL) const;
180 void clearFunctionData();
182 /// \brief Map basic blocks to their computed weights.
184 /// The weight of a basic block is defined to be the maximum
185 /// of all the instruction weights in that block.
186 BlockWeightMap BlockWeights;
188 /// \brief Map edges to their computed weights.
190 /// Edge weights are computed by propagating basic block weights in
191 /// SampleProfile::propagateWeights.
192 EdgeWeightMap EdgeWeights;
194 /// \brief Set of visited blocks during propagation.
195 SmallPtrSet<const BasicBlock *, 32> VisitedBlocks;
197 /// \brief Set of visited edges during propagation.
198 SmallSet<Edge, 32> VisitedEdges;
200 /// \brief Equivalence classes for block weights.
202 /// Two blocks BB1 and BB2 are in the same equivalence class if they
203 /// dominate and post-dominate each other, and they are in the same loop
204 /// nest. When this happens, the two blocks are guaranteed to execute
205 /// the same number of times.
206 EquivalenceClassMap EquivalenceClass;
208 /// \brief Dominance, post-dominance and loop information.
209 std::unique_ptr<DominatorTree> DT;
210 std::unique_ptr<DominatorTreeBase<BasicBlock>> PDT;
211 std::unique_ptr<LoopInfo> LI;
213 AssumptionCacheTracker *ACT;
215 /// \brief Predecessors for each basic block in the CFG.
216 BlockEdgeMap Predecessors;
218 /// \brief Successors for each basic block in the CFG.
219 BlockEdgeMap Successors;
221 SampleCoverageTracker CoverageTracker;
223 /// \brief Profile reader object.
224 std::unique_ptr<SampleProfileReader> Reader;
226 /// \brief Samples collected for the body of this function.
227 FunctionSamples *Samples;
229 /// \brief Name of the profile file to load.
230 std::string Filename;
232 /// \brief Flag indicating whether the profile input loaded successfully.
235 /// \brief Total number of samples collected in this profile.
237 /// This is the sum of all the samples collected in all the functions executed
239 uint64_t TotalCollectedSamples;
242 class SampleProfileLoaderLegacyPass : public ModulePass {
244 // Class identification, replacement for typeinfo
247 SampleProfileLoaderLegacyPass(StringRef Name = SampleProfileFile)
248 : ModulePass(ID), SampleLoader(Name) {
249 initializeSampleProfileLoaderLegacyPassPass(
250 *PassRegistry::getPassRegistry());
253 void dump() { SampleLoader.dump(); }
255 bool doInitialization(Module &M) override {
256 return SampleLoader.doInitialization(M);
258 StringRef getPassName() const override { return "Sample profile pass"; }
259 bool runOnModule(Module &M) override;
261 void getAnalysisUsage(AnalysisUsage &AU) const override {
262 AU.addRequired<AssumptionCacheTracker>();
266 SampleProfileLoader SampleLoader;
269 /// Return true if the given callsite is hot wrt to its caller.
271 /// Functions that were inlined in the original binary will be represented
272 /// in the inline stack in the sample profile. If the profile shows that
273 /// the original inline decision was "good" (i.e., the callsite is executed
274 /// frequently), then we will recreate the inline decision and apply the
275 /// profile from the inlined callsite.
277 /// To decide whether an inlined callsite is hot, we compute the fraction
278 /// of samples used by the callsite with respect to the total number of samples
279 /// collected in the caller.
281 /// If that fraction is larger than the default given by
282 /// SampleProfileHotThreshold, the callsite will be inlined again.
283 bool callsiteIsHot(const FunctionSamples *CallerFS,
284 const FunctionSamples *CallsiteFS) {
286 return false; // The callsite was not inlined in the original binary.
288 uint64_t ParentTotalSamples = CallerFS->getTotalSamples();
289 if (ParentTotalSamples == 0)
290 return false; // Avoid division by zero.
292 uint64_t CallsiteTotalSamples = CallsiteFS->getTotalSamples();
293 if (CallsiteTotalSamples == 0)
294 return false; // Callsite is trivially cold.
296 double PercentSamples =
297 (double)CallsiteTotalSamples / (double)ParentTotalSamples * 100.0;
298 return PercentSamples >= SampleProfileHotThreshold;
302 /// Mark as used the sample record for the given function samples at
303 /// (LineOffset, Discriminator).
305 /// \returns true if this is the first time we mark the given record.
306 bool SampleCoverageTracker::markSamplesUsed(const FunctionSamples *FS,
308 uint32_t Discriminator,
310 LineLocation Loc(LineOffset, Discriminator);
311 unsigned &Count = SampleCoverage[FS][Loc];
312 bool FirstTime = (++Count == 1);
314 TotalUsedSamples += Samples;
318 /// Return the number of sample records that were applied from this profile.
320 /// This count does not include records from cold inlined callsites.
322 SampleCoverageTracker::countUsedRecords(const FunctionSamples *FS) const {
323 auto I = SampleCoverage.find(FS);
325 // The size of the coverage map for FS represents the number of records
326 // that were marked used at least once.
327 unsigned Count = (I != SampleCoverage.end()) ? I->second.size() : 0;
329 // If there are inlined callsites in this function, count the samples found
330 // in the respective bodies. However, do not bother counting callees with 0
331 // total samples, these are callees that were never invoked at runtime.
332 for (const auto &I : FS->getCallsiteSamples()) {
333 const FunctionSamples *CalleeSamples = &I.second;
334 if (callsiteIsHot(FS, CalleeSamples))
335 Count += countUsedRecords(CalleeSamples);
341 /// Return the number of sample records in the body of this profile.
343 /// This count does not include records from cold inlined callsites.
345 SampleCoverageTracker::countBodyRecords(const FunctionSamples *FS) const {
346 unsigned Count = FS->getBodySamples().size();
348 // Only count records in hot callsites.
349 for (const auto &I : FS->getCallsiteSamples()) {
350 const FunctionSamples *CalleeSamples = &I.second;
351 if (callsiteIsHot(FS, CalleeSamples))
352 Count += countBodyRecords(CalleeSamples);
358 /// Return the number of samples collected in the body of this profile.
360 /// This count does not include samples from cold inlined callsites.
362 SampleCoverageTracker::countBodySamples(const FunctionSamples *FS) const {
364 for (const auto &I : FS->getBodySamples())
365 Total += I.second.getSamples();
367 // Only count samples in hot callsites.
368 for (const auto &I : FS->getCallsiteSamples()) {
369 const FunctionSamples *CalleeSamples = &I.second;
370 if (callsiteIsHot(FS, CalleeSamples))
371 Total += countBodySamples(CalleeSamples);
377 /// Return the fraction of sample records used in this profile.
379 /// The returned value is an unsigned integer in the range 0-100 indicating
380 /// the percentage of sample records that were used while applying this
381 /// profile to the associated function.
382 unsigned SampleCoverageTracker::computeCoverage(unsigned Used,
383 unsigned Total) const {
384 assert(Used <= Total &&
385 "number of used records cannot exceed the total number of records");
386 return Total > 0 ? Used * 100 / Total : 100;
389 /// Clear all the per-function data used to load samples and propagate weights.
390 void SampleProfileLoader::clearFunctionData() {
391 BlockWeights.clear();
393 VisitedBlocks.clear();
394 VisitedEdges.clear();
395 EquivalenceClass.clear();
399 Predecessors.clear();
401 CoverageTracker.clear();
404 /// Returns the line offset to the start line of the subprogram.
405 /// We assume that a single function will not exceed 65535 LOC.
406 unsigned SampleProfileLoader::getOffset(const DILocation *DIL) const {
407 return (DIL->getLine() - DIL->getScope()->getSubprogram()->getLine()) &
411 /// \brief Print the weight of edge \p E on stream \p OS.
413 /// \param OS Stream to emit the output to.
414 /// \param E Edge to print.
415 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
416 OS << "weight[" << E.first->getName() << "->" << E.second->getName()
417 << "]: " << EdgeWeights[E] << "\n";
420 /// \brief Print the equivalence class of block \p BB on stream \p OS.
422 /// \param OS Stream to emit the output to.
423 /// \param BB Block to print.
424 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
425 const BasicBlock *BB) {
426 const BasicBlock *Equiv = EquivalenceClass[BB];
427 OS << "equivalence[" << BB->getName()
428 << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
431 /// \brief Print the weight of block \p BB on stream \p OS.
433 /// \param OS Stream to emit the output to.
434 /// \param BB Block to print.
435 void SampleProfileLoader::printBlockWeight(raw_ostream &OS,
436 const BasicBlock *BB) const {
437 const auto &I = BlockWeights.find(BB);
438 uint64_t W = (I == BlockWeights.end() ? 0 : I->second);
439 OS << "weight[" << BB->getName() << "]: " << W << "\n";
442 /// \brief Get the weight for an instruction.
444 /// The "weight" of an instruction \p Inst is the number of samples
445 /// collected on that instruction at runtime. To retrieve it, we
446 /// need to compute the line number of \p Inst relative to the start of its
447 /// function. We use HeaderLineno to compute the offset. We then
448 /// look up the samples collected for \p Inst using BodySamples.
450 /// \param Inst Instruction to query.
452 /// \returns the weight of \p Inst.
453 ErrorOr<uint64_t> SampleProfileLoader::getInstWeight(const Instruction &Inst) {
454 const DebugLoc &DLoc = Inst.getDebugLoc();
456 return std::error_code();
458 const FunctionSamples *FS = findFunctionSamples(Inst);
460 return std::error_code();
462 // Ignore all intrinsics and branch instructions.
463 // Branch instruction usually contains debug info from sources outside of
464 // the residing basic block, thus we ignore them during annotation.
465 if (isa<BranchInst>(Inst) || isa<IntrinsicInst>(Inst))
466 return std::error_code();
468 // If a call/invoke instruction is inlined in profile, but not inlined here,
469 // it means that the inlined callsite has no sample, thus the call
470 // instruction should have 0 count.
471 bool IsCall = isa<CallInst>(Inst) || isa<InvokeInst>(Inst);
472 if (IsCall && findCalleeFunctionSamples(Inst))
475 const DILocation *DIL = DLoc;
476 uint32_t LineOffset = getOffset(DIL);
477 uint32_t Discriminator = DIL->getBaseDiscriminator();
478 ErrorOr<uint64_t> R = IsCall
479 ? FS->findCallSamplesAt(LineOffset, Discriminator)
480 : FS->findSamplesAt(LineOffset, Discriminator);
483 CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get());
485 const Function *F = Inst.getParent()->getParent();
486 LLVMContext &Ctx = F->getContext();
487 emitOptimizationRemark(
488 Ctx, DEBUG_TYPE, *F, DLoc,
489 Twine("Applied ") + Twine(*R) +
490 " samples from profile (offset: " + Twine(LineOffset) +
491 ((Discriminator) ? Twine(".") + Twine(Discriminator) : "") + ")");
493 DEBUG(dbgs() << " " << DLoc.getLine() << "."
494 << DIL->getBaseDiscriminator() << ":" << Inst
495 << " (line offset: " << LineOffset << "."
496 << DIL->getBaseDiscriminator() << " - weight: " << R.get()
502 /// \brief Compute the weight of a basic block.
504 /// The weight of basic block \p BB is the maximum weight of all the
505 /// instructions in BB.
507 /// \param BB The basic block to query.
509 /// \returns the weight for \p BB.
510 ErrorOr<uint64_t> SampleProfileLoader::getBlockWeight(const BasicBlock *BB) {
512 bool HasWeight = false;
513 for (auto &I : BB->getInstList()) {
514 const ErrorOr<uint64_t> &R = getInstWeight(I);
516 Max = std::max(Max, R.get());
520 return HasWeight ? ErrorOr<uint64_t>(Max) : std::error_code();
523 /// \brief Compute and store the weights of every basic block.
525 /// This populates the BlockWeights map by computing
526 /// the weights of every basic block in the CFG.
528 /// \param F The function to query.
529 bool SampleProfileLoader::computeBlockWeights(Function &F) {
530 bool Changed = false;
531 DEBUG(dbgs() << "Block weights\n");
532 for (const auto &BB : F) {
533 ErrorOr<uint64_t> Weight = getBlockWeight(&BB);
535 BlockWeights[&BB] = Weight.get();
536 VisitedBlocks.insert(&BB);
539 DEBUG(printBlockWeight(dbgs(), &BB));
545 /// \brief Get the FunctionSamples for a call instruction.
547 /// The FunctionSamples of a call/invoke instruction \p Inst is the inlined
548 /// instance in which that call instruction is calling to. It contains
549 /// all samples that resides in the inlined instance. We first find the
550 /// inlined instance in which the call instruction is from, then we
551 /// traverse its children to find the callsite with the matching
554 /// \param Inst Call/Invoke instruction to query.
556 /// \returns The FunctionSamples pointer to the inlined instance.
557 const FunctionSamples *
558 SampleProfileLoader::findCalleeFunctionSamples(const Instruction &Inst) const {
559 const DILocation *DIL = Inst.getDebugLoc();
563 const FunctionSamples *FS = findFunctionSamples(Inst);
567 return FS->findFunctionSamplesAt(
568 LineLocation(getOffset(DIL), DIL->getBaseDiscriminator()));
571 /// \brief Get the FunctionSamples for an instruction.
573 /// The FunctionSamples of an instruction \p Inst is the inlined instance
574 /// in which that instruction is coming from. We traverse the inline stack
575 /// of that instruction, and match it with the tree nodes in the profile.
577 /// \param Inst Instruction to query.
579 /// \returns the FunctionSamples pointer to the inlined instance.
580 const FunctionSamples *
581 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
582 SmallVector<LineLocation, 10> S;
583 const DILocation *DIL = Inst.getDebugLoc();
587 for (DIL = DIL->getInlinedAt(); DIL; DIL = DIL->getInlinedAt())
588 S.push_back(LineLocation(getOffset(DIL), DIL->getBaseDiscriminator()));
591 const FunctionSamples *FS = Samples;
592 for (int i = S.size() - 1; i >= 0 && FS != nullptr; i--) {
593 FS = FS->findFunctionSamplesAt(S[i]);
598 /// \brief Iteratively inline hot callsites of a function.
600 /// Iteratively traverse all callsites of the function \p F, and find if
601 /// the corresponding inlined instance exists and is hot in profile. If
602 /// it is hot enough, inline the callsites and adds new callsites of the
603 /// callee into the caller. If the call is an indirect call, first promote
604 /// it to direct call. Each indirect call is limited with a single target.
606 /// \param F function to perform iterative inlining.
608 /// \returns True if there is any inline happened.
609 bool SampleProfileLoader::inlineHotFunctions(Function &F) {
610 DenseSet<Instruction *> PromotedInsns;
611 bool Changed = false;
612 LLVMContext &Ctx = F.getContext();
613 std::function<AssumptionCache &(Function &)> GetAssumptionCache = [&](
614 Function &F) -> AssumptionCache & { return ACT->getAssumptionCache(F); };
616 bool LocalChanged = false;
617 SmallVector<Instruction *, 10> CIS;
620 SmallVector<Instruction *, 10> Candidates;
621 for (auto &I : BB.getInstList()) {
622 const FunctionSamples *FS = nullptr;
623 if ((isa<CallInst>(I) || isa<InvokeInst>(I)) &&
624 (FS = findCalleeFunctionSamples(I))) {
625 Candidates.push_back(&I);
626 if (callsiteIsHot(Samples, FS))
631 CIS.insert(CIS.begin(), Candidates.begin(), Candidates.end());
635 InlineFunctionInfo IFI(nullptr, ACT ? &GetAssumptionCache : nullptr);
636 Function *CalledFunction = CallSite(I).getCalledFunction();
638 if (!CalledFunction && !PromotedInsns.count(I) &&
639 CallSite(I).isIndirectCall()) {
640 auto CalleeFunctionName = findCalleeFunctionSamples(*I)->getName();
641 const char *Reason = "Callee function not available";
642 CalledFunction = F.getParent()->getFunction(CalleeFunctionName);
643 if (CalledFunction && isLegalToPromote(I, CalledFunction, &Reason)) {
644 // The indirect target was promoted and inlined in the profile, as a
645 // result, we do not have profile info for the branch probability.
646 // We set the probability to 80% taken to indicate that the static
647 // call is likely taken.
648 DI = dyn_cast<Instruction>(
649 promoteIndirectCall(I, CalledFunction, 80, 100, false)
650 ->stripPointerCasts());
651 PromotedInsns.insert(I);
653 DEBUG(dbgs() << "\nFailed to promote indirect call to "
654 << CalleeFunctionName << " because " << Reason << "\n");
658 if (!CalledFunction || !CalledFunction->getSubprogram())
660 DebugLoc DLoc = I->getDebugLoc();
661 uint64_t NumSamples = findCalleeFunctionSamples(*I)->getTotalSamples();
662 if (InlineFunction(CallSite(DI), IFI)) {
664 emitOptimizationRemark(Ctx, DEBUG_TYPE, F, DLoc,
665 Twine("inlined hot callee '") +
666 CalledFunction->getName() + "' with " +
667 Twine(NumSamples) + " samples into '" +
680 /// \brief Find equivalence classes for the given block.
682 /// This finds all the blocks that are guaranteed to execute the same
683 /// number of times as \p BB1. To do this, it traverses all the
684 /// descendants of \p BB1 in the dominator or post-dominator tree.
686 /// A block BB2 will be in the same equivalence class as \p BB1 if
687 /// the following holds:
689 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
690 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
691 /// dominate BB1 in the post-dominator tree.
693 /// 2- Both BB2 and \p BB1 must be in the same loop.
695 /// For every block BB2 that meets those two requirements, we set BB2's
696 /// equivalence class to \p BB1.
698 /// \param BB1 Block to check.
699 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
700 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
701 /// with blocks from \p BB1's dominator tree, then
702 /// this is the post-dominator tree, and vice versa.
703 void SampleProfileLoader::findEquivalencesFor(
704 BasicBlock *BB1, ArrayRef<BasicBlock *> Descendants,
705 DominatorTreeBase<BasicBlock> *DomTree) {
706 const BasicBlock *EC = EquivalenceClass[BB1];
707 uint64_t Weight = BlockWeights[EC];
708 for (const auto *BB2 : Descendants) {
709 bool IsDomParent = DomTree->dominates(BB2, BB1);
710 bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
711 if (BB1 != BB2 && IsDomParent && IsInSameLoop) {
712 EquivalenceClass[BB2] = EC;
713 // If BB2 is visited, then the entire EC should be marked as visited.
714 if (VisitedBlocks.count(BB2)) {
715 VisitedBlocks.insert(EC);
718 // If BB2 is heavier than BB1, make BB2 have the same weight
721 // Note that we don't worry about the opposite situation here
722 // (when BB2 is lighter than BB1). We will deal with this
723 // during the propagation phase. Right now, we just want to
724 // make sure that BB1 has the largest weight of all the
725 // members of its equivalence set.
726 Weight = std::max(Weight, BlockWeights[BB2]);
729 if (EC == &EC->getParent()->getEntryBlock()) {
730 BlockWeights[EC] = Samples->getHeadSamples() + 1;
732 BlockWeights[EC] = Weight;
736 /// \brief Find equivalence classes.
738 /// Since samples may be missing from blocks, we can fill in the gaps by setting
739 /// the weights of all the blocks in the same equivalence class to the same
740 /// weight. To compute the concept of equivalence, we use dominance and loop
741 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
742 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
744 /// \param F The function to query.
745 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
746 SmallVector<BasicBlock *, 8> DominatedBBs;
747 DEBUG(dbgs() << "\nBlock equivalence classes\n");
748 // Find equivalence sets based on dominance and post-dominance information.
750 BasicBlock *BB1 = &BB;
752 // Compute BB1's equivalence class once.
753 if (EquivalenceClass.count(BB1)) {
754 DEBUG(printBlockEquivalence(dbgs(), BB1));
758 // By default, blocks are in their own equivalence class.
759 EquivalenceClass[BB1] = BB1;
761 // Traverse all the blocks dominated by BB1. We are looking for
762 // every basic block BB2 such that:
764 // 1- BB1 dominates BB2.
765 // 2- BB2 post-dominates BB1.
766 // 3- BB1 and BB2 are in the same loop nest.
768 // If all those conditions hold, it means that BB2 is executed
769 // as many times as BB1, so they are placed in the same equivalence
770 // class by making BB2's equivalence class be BB1.
771 DominatedBBs.clear();
772 DT->getDescendants(BB1, DominatedBBs);
773 findEquivalencesFor(BB1, DominatedBBs, PDT.get());
775 DEBUG(printBlockEquivalence(dbgs(), BB1));
778 // Assign weights to equivalence classes.
780 // All the basic blocks in the same equivalence class will execute
781 // the same number of times. Since we know that the head block in
782 // each equivalence class has the largest weight, assign that weight
783 // to all the blocks in that equivalence class.
784 DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n");
786 const BasicBlock *BB = &BI;
787 const BasicBlock *EquivBB = EquivalenceClass[BB];
789 BlockWeights[BB] = BlockWeights[EquivBB];
790 DEBUG(printBlockWeight(dbgs(), BB));
794 /// \brief Visit the given edge to decide if it has a valid weight.
796 /// If \p E has not been visited before, we copy to \p UnknownEdge
797 /// and increment the count of unknown edges.
799 /// \param E Edge to visit.
800 /// \param NumUnknownEdges Current number of unknown edges.
801 /// \param UnknownEdge Set if E has not been visited before.
803 /// \returns E's weight, if known. Otherwise, return 0.
804 uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
806 if (!VisitedEdges.count(E)) {
807 (*NumUnknownEdges)++;
812 return EdgeWeights[E];
815 /// \brief Propagate weights through incoming/outgoing edges.
817 /// If the weight of a basic block is known, and there is only one edge
818 /// with an unknown weight, we can calculate the weight of that edge.
820 /// Similarly, if all the edges have a known count, we can calculate the
821 /// count of the basic block, if needed.
823 /// \param F Function to process.
824 /// \param UpdateBlockCount Whether we should update basic block counts that
825 /// has already been annotated.
827 /// \returns True if new weights were assigned to edges or blocks.
828 bool SampleProfileLoader::propagateThroughEdges(Function &F,
829 bool UpdateBlockCount) {
830 bool Changed = false;
831 DEBUG(dbgs() << "\nPropagation through edges\n");
832 for (const auto &BI : F) {
833 const BasicBlock *BB = &BI;
834 const BasicBlock *EC = EquivalenceClass[BB];
836 // Visit all the predecessor and successor edges to determine
837 // which ones have a weight assigned already. Note that it doesn't
838 // matter that we only keep track of a single unknown edge. The
839 // only case we are interested in handling is when only a single
840 // edge is unknown (see setEdgeOrBlockWeight).
841 for (unsigned i = 0; i < 2; i++) {
842 uint64_t TotalWeight = 0;
843 unsigned NumUnknownEdges = 0, NumTotalEdges = 0;
844 Edge UnknownEdge, SelfReferentialEdge, SingleEdge;
847 // First, visit all predecessor edges.
848 NumTotalEdges = Predecessors[BB].size();
849 for (auto *Pred : Predecessors[BB]) {
850 Edge E = std::make_pair(Pred, BB);
851 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
852 if (E.first == E.second)
853 SelfReferentialEdge = E;
855 if (NumTotalEdges == 1) {
856 SingleEdge = std::make_pair(Predecessors[BB][0], BB);
859 // On the second round, visit all successor edges.
860 NumTotalEdges = Successors[BB].size();
861 for (auto *Succ : Successors[BB]) {
862 Edge E = std::make_pair(BB, Succ);
863 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
865 if (NumTotalEdges == 1) {
866 SingleEdge = std::make_pair(BB, Successors[BB][0]);
870 // After visiting all the edges, there are three cases that we
871 // can handle immediately:
873 // - All the edge weights are known (i.e., NumUnknownEdges == 0).
874 // In this case, we simply check that the sum of all the edges
875 // is the same as BB's weight. If not, we change BB's weight
876 // to match. Additionally, if BB had not been visited before,
877 // we mark it visited.
879 // - Only one edge is unknown and BB has already been visited.
880 // In this case, we can compute the weight of the edge by
881 // subtracting the total block weight from all the known
882 // edge weights. If the edges weight more than BB, then the
883 // edge of the last remaining edge is set to zero.
885 // - There exists a self-referential edge and the weight of BB is
886 // known. In this case, this edge can be based on BB's weight.
887 // We add up all the other known edges and set the weight on
888 // the self-referential edge as we did in the previous case.
890 // In any other case, we must continue iterating. Eventually,
891 // all edges will get a weight, or iteration will stop when
892 // it reaches SampleProfileMaxPropagateIterations.
893 if (NumUnknownEdges <= 1) {
894 uint64_t &BBWeight = BlockWeights[EC];
895 if (NumUnknownEdges == 0) {
896 if (!VisitedBlocks.count(EC)) {
897 // If we already know the weight of all edges, the weight of the
898 // basic block can be computed. It should be no larger than the sum
899 // of all edge weights.
900 if (TotalWeight > BBWeight) {
901 BBWeight = TotalWeight;
903 DEBUG(dbgs() << "All edge weights for " << BB->getName()
904 << " known. Set weight for block: ";
905 printBlockWeight(dbgs(), BB););
907 } else if (NumTotalEdges == 1 &&
908 EdgeWeights[SingleEdge] < BlockWeights[EC]) {
909 // If there is only one edge for the visited basic block, use the
910 // block weight to adjust edge weight if edge weight is smaller.
911 EdgeWeights[SingleEdge] = BlockWeights[EC];
914 } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) {
915 // If there is a single unknown edge and the block has been
916 // visited, then we can compute E's weight.
917 if (BBWeight >= TotalWeight)
918 EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
920 EdgeWeights[UnknownEdge] = 0;
921 const BasicBlock *OtherEC;
923 OtherEC = EquivalenceClass[UnknownEdge.first];
925 OtherEC = EquivalenceClass[UnknownEdge.second];
926 // Edge weights should never exceed the BB weights it connects.
927 if (VisitedBlocks.count(OtherEC) &&
928 EdgeWeights[UnknownEdge] > BlockWeights[OtherEC])
929 EdgeWeights[UnknownEdge] = BlockWeights[OtherEC];
930 VisitedEdges.insert(UnknownEdge);
932 DEBUG(dbgs() << "Set weight for edge: ";
933 printEdgeWeight(dbgs(), UnknownEdge));
935 } else if (VisitedBlocks.count(EC) && BlockWeights[EC] == 0) {
936 // If a block Weights 0, all its in/out edges should weight 0.
938 for (auto *Pred : Predecessors[BB]) {
939 Edge E = std::make_pair(Pred, BB);
941 VisitedEdges.insert(E);
944 for (auto *Succ : Successors[BB]) {
945 Edge E = std::make_pair(BB, Succ);
947 VisitedEdges.insert(E);
950 } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) {
951 uint64_t &BBWeight = BlockWeights[BB];
952 // We have a self-referential edge and the weight of BB is known.
953 if (BBWeight >= TotalWeight)
954 EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
956 EdgeWeights[SelfReferentialEdge] = 0;
957 VisitedEdges.insert(SelfReferentialEdge);
959 DEBUG(dbgs() << "Set self-referential edge weight to: ";
960 printEdgeWeight(dbgs(), SelfReferentialEdge));
962 if (UpdateBlockCount && !VisitedBlocks.count(EC) && TotalWeight > 0) {
963 BlockWeights[EC] = TotalWeight;
964 VisitedBlocks.insert(EC);
973 /// \brief Build in/out edge lists for each basic block in the CFG.
975 /// We are interested in unique edges. If a block B1 has multiple
976 /// edges to another block B2, we only add a single B1->B2 edge.
977 void SampleProfileLoader::buildEdges(Function &F) {
979 BasicBlock *B1 = &BI;
981 // Add predecessors for B1.
982 SmallPtrSet<BasicBlock *, 16> Visited;
983 if (!Predecessors[B1].empty())
984 llvm_unreachable("Found a stale predecessors list in a basic block.");
985 for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
986 BasicBlock *B2 = *PI;
987 if (Visited.insert(B2).second)
988 Predecessors[B1].push_back(B2);
991 // Add successors for B1.
993 if (!Successors[B1].empty())
994 llvm_unreachable("Found a stale successors list in a basic block.");
995 for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
996 BasicBlock *B2 = *SI;
997 if (Visited.insert(B2).second)
998 Successors[B1].push_back(B2);
1003 /// Sorts the CallTargetMap \p M by count in descending order and stores the
1004 /// sorted result in \p Sorted. Returns the total counts.
1005 static uint64_t SortCallTargets(SmallVector<InstrProfValueData, 2> &Sorted,
1006 const SampleRecord::CallTargetMap &M) {
1009 for (auto I = M.begin(); I != M.end(); ++I) {
1010 Sum += I->getValue();
1011 Sorted.push_back({Function::getGUID(I->getKey()), I->getValue()});
1013 std::sort(Sorted.begin(), Sorted.end(),
1014 [](const InstrProfValueData &L, const InstrProfValueData &R) {
1015 if (L.Count == R.Count)
1016 return L.Value > R.Value;
1018 return L.Count > R.Count;
1023 /// \brief Propagate weights into edges
1025 /// The following rules are applied to every block BB in the CFG:
1027 /// - If BB has a single predecessor/successor, then the weight
1028 /// of that edge is the weight of the block.
1030 /// - If all incoming or outgoing edges are known except one, and the
1031 /// weight of the block is already known, the weight of the unknown
1032 /// edge will be the weight of the block minus the sum of all the known
1033 /// edges. If the sum of all the known edges is larger than BB's weight,
1034 /// we set the unknown edge weight to zero.
1036 /// - If there is a self-referential edge, and the weight of the block is
1037 /// known, the weight for that edge is set to the weight of the block
1038 /// minus the weight of the other incoming edges to that block (if
1040 void SampleProfileLoader::propagateWeights(Function &F) {
1041 bool Changed = true;
1044 // Add an entry count to the function using the samples gathered
1045 // at the function entry.
1046 F.setEntryCount(Samples->getHeadSamples() + 1);
1048 // If BB weight is larger than its corresponding loop's header BB weight,
1049 // use the BB weight to replace the loop header BB weight.
1050 for (auto &BI : F) {
1051 BasicBlock *BB = &BI;
1052 Loop *L = LI->getLoopFor(BB);
1056 BasicBlock *Header = L->getHeader();
1057 if (Header && BlockWeights[BB] > BlockWeights[Header]) {
1058 BlockWeights[Header] = BlockWeights[BB];
1062 // Before propagation starts, build, for each block, a list of
1063 // unique predecessors and successors. This is necessary to handle
1064 // identical edges in multiway branches. Since we visit all blocks and all
1065 // edges of the CFG, it is cleaner to build these lists once at the start
1069 // Propagate until we converge or we go past the iteration limit.
1070 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1071 Changed = propagateThroughEdges(F, false);
1074 // The first propagation propagates BB counts from annotated BBs to unknown
1075 // BBs. The 2nd propagation pass resets edges weights, and use all BB weights
1076 // to propagate edge weights.
1077 VisitedEdges.clear();
1079 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1080 Changed = propagateThroughEdges(F, false);
1083 // The 3rd propagation pass allows adjust annotated BB weights that are
1086 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
1087 Changed = propagateThroughEdges(F, true);
1090 // Generate MD_prof metadata for every branch instruction using the
1091 // edge weights computed during propagation.
1092 DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
1093 LLVMContext &Ctx = F.getContext();
1095 for (auto &BI : F) {
1096 BasicBlock *BB = &BI;
1098 if (BlockWeights[BB]) {
1099 for (auto &I : BB->getInstList()) {
1100 if (!isa<CallInst>(I) && !isa<InvokeInst>(I))
1103 if (!CS.getCalledFunction()) {
1104 const DebugLoc &DLoc = I.getDebugLoc();
1107 const DILocation *DIL = DLoc;
1108 uint32_t LineOffset = getOffset(DIL);
1109 uint32_t Discriminator = DIL->getBaseDiscriminator();
1111 const FunctionSamples *FS = findFunctionSamples(I);
1114 auto T = FS->findCallTargetMapAt(LineOffset, Discriminator);
1115 if (!T || T.get().size() == 0)
1117 SmallVector<InstrProfValueData, 2> SortedCallTargets;
1118 uint64_t Sum = SortCallTargets(SortedCallTargets, T.get());
1119 annotateValueSite(*I.getParent()->getParent()->getParent(), I,
1120 SortedCallTargets, Sum, IPVK_IndirectCallTarget,
1121 SortedCallTargets.size());
1122 } else if (!dyn_cast<IntrinsicInst>(&I)) {
1123 SmallVector<uint32_t, 1> Weights;
1124 Weights.push_back(BlockWeights[BB]);
1125 I.setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights));
1129 TerminatorInst *TI = BB->getTerminator();
1130 if (TI->getNumSuccessors() == 1)
1132 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
1135 DEBUG(dbgs() << "\nGetting weights for branch at line "
1136 << TI->getDebugLoc().getLine() << ".\n");
1137 SmallVector<uint32_t, 4> Weights;
1138 uint32_t MaxWeight = 0;
1139 DebugLoc MaxDestLoc;
1140 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
1141 BasicBlock *Succ = TI->getSuccessor(I);
1142 Edge E = std::make_pair(BB, Succ);
1143 uint64_t Weight = EdgeWeights[E];
1144 DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
1145 // Use uint32_t saturated arithmetic to adjust the incoming weights,
1146 // if needed. Sample counts in profiles are 64-bit unsigned values,
1147 // but internally branch weights are expressed as 32-bit values.
1148 if (Weight > std::numeric_limits<uint32_t>::max()) {
1149 DEBUG(dbgs() << " (saturated due to uint32_t overflow)");
1150 Weight = std::numeric_limits<uint32_t>::max();
1152 // Weight is added by one to avoid propagation errors introduced by
1154 Weights.push_back(static_cast<uint32_t>(Weight + 1));
1156 if (Weight > MaxWeight) {
1158 MaxDestLoc = Succ->getFirstNonPHIOrDbgOrLifetime()->getDebugLoc();
1163 // Only set weights if there is at least one non-zero weight.
1164 // In any other case, let the analyzer set weights.
1165 if (MaxWeight > 0) {
1166 DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
1167 TI->setMetadata(llvm::LLVMContext::MD_prof,
1168 MDB.createBranchWeights(Weights));
1169 DebugLoc BranchLoc = TI->getDebugLoc();
1170 emitOptimizationRemark(
1171 Ctx, DEBUG_TYPE, F, MaxDestLoc,
1172 Twine("most popular destination for conditional branches at ") +
1173 ((BranchLoc) ? Twine(BranchLoc->getFilename() + ":" +
1174 Twine(BranchLoc.getLine()) + ":" +
1175 Twine(BranchLoc.getCol()))
1176 : Twine("<UNKNOWN LOCATION>")));
1178 DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
1183 /// \brief Get the line number for the function header.
1185 /// This looks up function \p F in the current compilation unit and
1186 /// retrieves the line number where the function is defined. This is
1187 /// line 0 for all the samples read from the profile file. Every line
1188 /// number is relative to this line.
1190 /// \param F Function object to query.
1192 /// \returns the line number where \p F is defined. If it returns 0,
1193 /// it means that there is no debug information available for \p F.
1194 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
1195 if (DISubprogram *S = F.getSubprogram())
1196 return S->getLine();
1198 // If the start of \p F is missing, emit a diagnostic to inform the user
1199 // about the missed opportunity.
1200 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1201 "No debug information found in function " + F.getName() +
1202 ": Function profile not used",
1207 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
1208 DT.reset(new DominatorTree);
1211 PDT.reset(new DominatorTreeBase<BasicBlock>(true));
1212 PDT->recalculate(F);
1214 LI.reset(new LoopInfo);
1218 /// \brief Generate branch weight metadata for all branches in \p F.
1220 /// Branch weights are computed out of instruction samples using a
1221 /// propagation heuristic. Propagation proceeds in 3 phases:
1223 /// 1- Assignment of block weights. All the basic blocks in the function
1224 /// are initial assigned the same weight as their most frequently
1225 /// executed instruction.
1227 /// 2- Creation of equivalence classes. Since samples may be missing from
1228 /// blocks, we can fill in the gaps by setting the weights of all the
1229 /// blocks in the same equivalence class to the same weight. To compute
1230 /// the concept of equivalence, we use dominance and loop information.
1231 /// Two blocks B1 and B2 are in the same equivalence class if B1
1232 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
1234 /// 3- Propagation of block weights into edges. This uses a simple
1235 /// propagation heuristic. The following rules are applied to every
1236 /// block BB in the CFG:
1238 /// - If BB has a single predecessor/successor, then the weight
1239 /// of that edge is the weight of the block.
1241 /// - If all the edges are known except one, and the weight of the
1242 /// block is already known, the weight of the unknown edge will
1243 /// be the weight of the block minus the sum of all the known
1244 /// edges. If the sum of all the known edges is larger than BB's weight,
1245 /// we set the unknown edge weight to zero.
1247 /// - If there is a self-referential edge, and the weight of the block is
1248 /// known, the weight for that edge is set to the weight of the block
1249 /// minus the weight of the other incoming edges to that block (if
1252 /// Since this propagation is not guaranteed to finalize for every CFG, we
1253 /// only allow it to proceed for a limited number of iterations (controlled
1254 /// by -sample-profile-max-propagate-iterations).
1256 /// FIXME: Try to replace this propagation heuristic with a scheme
1257 /// that is guaranteed to finalize. A work-list approach similar to
1258 /// the standard value propagation algorithm used by SSA-CCP might
1261 /// Once all the branch weights are computed, we emit the MD_prof
1262 /// metadata on BB using the computed values for each of its branches.
1264 /// \param F The function to query.
1266 /// \returns true if \p F was modified. Returns false, otherwise.
1267 bool SampleProfileLoader::emitAnnotations(Function &F) {
1268 bool Changed = false;
1270 if (getFunctionLoc(F) == 0)
1273 DEBUG(dbgs() << "Line number for the first instruction in " << F.getName()
1274 << ": " << getFunctionLoc(F) << "\n");
1276 Changed |= inlineHotFunctions(F);
1278 // Compute basic block weights.
1279 Changed |= computeBlockWeights(F);
1282 // Compute dominance and loop info needed for propagation.
1283 computeDominanceAndLoopInfo(F);
1285 // Find equivalence classes.
1286 findEquivalenceClasses(F);
1288 // Propagate weights to all edges.
1289 propagateWeights(F);
1292 // If coverage checking was requested, compute it now.
1293 if (SampleProfileRecordCoverage) {
1294 unsigned Used = CoverageTracker.countUsedRecords(Samples);
1295 unsigned Total = CoverageTracker.countBodyRecords(Samples);
1296 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1297 if (Coverage < SampleProfileRecordCoverage) {
1298 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1299 F.getSubprogram()->getFilename(), getFunctionLoc(F),
1300 Twine(Used) + " of " + Twine(Total) + " available profile records (" +
1301 Twine(Coverage) + "%) were applied",
1306 if (SampleProfileSampleCoverage) {
1307 uint64_t Used = CoverageTracker.getTotalUsedSamples();
1308 uint64_t Total = CoverageTracker.countBodySamples(Samples);
1309 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1310 if (Coverage < SampleProfileSampleCoverage) {
1311 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1312 F.getSubprogram()->getFilename(), getFunctionLoc(F),
1313 Twine(Used) + " of " + Twine(Total) + " available profile samples (" +
1314 Twine(Coverage) + "%) were applied",
1321 char SampleProfileLoaderLegacyPass::ID = 0;
1322 INITIALIZE_PASS_BEGIN(SampleProfileLoaderLegacyPass, "sample-profile",
1323 "Sample Profile loader", false, false)
1324 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1325 INITIALIZE_PASS_END(SampleProfileLoaderLegacyPass, "sample-profile",
1326 "Sample Profile loader", false, false)
1328 bool SampleProfileLoader::doInitialization(Module &M) {
1329 auto &Ctx = M.getContext();
1330 auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
1331 if (std::error_code EC = ReaderOrErr.getError()) {
1332 std::string Msg = "Could not open profile: " + EC.message();
1333 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg));
1336 Reader = std::move(ReaderOrErr.get());
1337 ProfileIsValid = (Reader->read() == sampleprof_error::success);
1341 ModulePass *llvm::createSampleProfileLoaderPass() {
1342 return new SampleProfileLoaderLegacyPass(SampleProfileFile);
1345 ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) {
1346 return new SampleProfileLoaderLegacyPass(Name);
1349 bool SampleProfileLoader::runOnModule(Module &M) {
1350 if (!ProfileIsValid)
1353 // Compute the total number of samples collected in this profile.
1354 for (const auto &I : Reader->getProfiles())
1355 TotalCollectedSamples += I.second.getTotalSamples();
1357 bool retval = false;
1359 if (!F.isDeclaration()) {
1360 clearFunctionData();
1361 retval |= runOnFunction(F);
1363 if (M.getProfileSummary() == nullptr)
1364 M.setProfileSummary(Reader->getSummary().getMD(M.getContext()));
1368 bool SampleProfileLoaderLegacyPass::runOnModule(Module &M) {
1369 // FIXME: pass in AssumptionCache correctly for the new pass manager.
1370 SampleLoader.setACT(&getAnalysis<AssumptionCacheTracker>());
1371 return SampleLoader.runOnModule(M);
1374 bool SampleProfileLoader::runOnFunction(Function &F) {
1376 Samples = Reader->getSamplesFor(F);
1377 if (!Samples->empty())
1378 return emitAnnotations(F);
1382 PreservedAnalyses SampleProfileLoaderPass::run(Module &M,
1383 ModuleAnalysisManager &AM) {
1385 SampleProfileLoader SampleLoader(SampleProfileFile);
1387 SampleLoader.doInitialization(M);
1389 if (!SampleLoader.runOnModule(M))
1390 return PreservedAnalyses::all();
1392 return PreservedAnalyses::none();