1 //===- SampleProfile.cpp - Incorporate sample profiles into the IR --------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the SampleProfileLoader transformation. This pass
11 // reads a profile file generated by a sampling profiler (e.g. Linux Perf -
12 // http://perf.wiki.kernel.org/) and generates IR metadata to reflect the
13 // profile information in the given profile.
15 // This pass generates branch weight annotations on the IR:
17 // - prof: Represents branch weights. This annotation is added to branches
18 // to indicate the weights of each edge coming out of the branch.
19 // The weight of each edge is the weight of the target block for
20 // that edge. The weight of a block B is computed as the maximum
21 // number of samples found in B.
23 //===----------------------------------------------------------------------===//
25 #include "llvm/ADT/DenseMap.h"
26 #include "llvm/ADT/SmallPtrSet.h"
27 #include "llvm/ADT/SmallSet.h"
28 #include "llvm/ADT/StringRef.h"
29 #include "llvm/Analysis/LoopInfo.h"
30 #include "llvm/Analysis/PostDominators.h"
31 #include "llvm/IR/Constants.h"
32 #include "llvm/IR/DebugInfo.h"
33 #include "llvm/IR/DiagnosticInfo.h"
34 #include "llvm/IR/Dominators.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/IR/InstIterator.h"
37 #include "llvm/IR/Instructions.h"
38 #include "llvm/IR/LLVMContext.h"
39 #include "llvm/IR/MDBuilder.h"
40 #include "llvm/IR/Metadata.h"
41 #include "llvm/IR/Module.h"
42 #include "llvm/Pass.h"
43 #include "llvm/ProfileData/SampleProfReader.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/ErrorOr.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/Transforms/IPO.h"
49 #include "llvm/Transforms/Utils/Cloning.h"
53 using namespace sampleprof;
55 #define DEBUG_TYPE "sample-profile"
57 // Command line option to specify the file to read samples from. This is
58 // mainly used for debugging.
59 static cl::opt<std::string> SampleProfileFile(
60 "sample-profile-file", cl::init(""), cl::value_desc("filename"),
61 cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
62 static cl::opt<unsigned> SampleProfileMaxPropagateIterations(
63 "sample-profile-max-propagate-iterations", cl::init(100),
64 cl::desc("Maximum number of iterations to go through when propagating "
65 "sample block/edge weights through the CFG."));
66 static cl::opt<unsigned> SampleProfileRecordCoverage(
67 "sample-profile-check-record-coverage", cl::init(0), cl::value_desc("N"),
68 cl::desc("Emit a warning if less than N% of records in the input profile "
69 "are matched to the IR."));
70 static cl::opt<unsigned> SampleProfileSampleCoverage(
71 "sample-profile-check-sample-coverage", cl::init(0), cl::value_desc("N"),
72 cl::desc("Emit a warning if less than N% of samples in the input profile "
73 "are matched to the IR."));
76 typedef DenseMap<const BasicBlock *, uint64_t> BlockWeightMap;
77 typedef DenseMap<const BasicBlock *, const BasicBlock *> EquivalenceClassMap;
78 typedef std::pair<const BasicBlock *, const BasicBlock *> Edge;
79 typedef DenseMap<Edge, uint64_t> EdgeWeightMap;
80 typedef DenseMap<const BasicBlock *, SmallVector<const BasicBlock *, 8>>
83 /// \brief Sample profile pass.
85 /// This pass reads profile data from the file specified by
86 /// -sample-profile-file and annotates every affected function with the
87 /// profile information found in that file.
88 class SampleProfileLoader : public ModulePass {
90 // Class identification, replacement for typeinfo
93 SampleProfileLoader(StringRef Name = SampleProfileFile)
94 : ModulePass(ID), DT(nullptr), PDT(nullptr), LI(nullptr), Reader(),
95 Samples(nullptr), Filename(Name), ProfileIsValid(false) {
96 initializeSampleProfileLoaderPass(*PassRegistry::getPassRegistry());
99 bool doInitialization(Module &M) override;
101 void dump() { Reader->dump(); }
103 const char *getPassName() const override { return "Sample profile pass"; }
105 bool runOnModule(Module &M) override;
107 void getAnalysisUsage(AnalysisUsage &AU) const override {
108 AU.setPreservesCFG();
112 bool runOnFunction(Function &F);
113 unsigned getFunctionLoc(Function &F);
114 bool emitAnnotations(Function &F);
115 ErrorOr<uint64_t> getInstWeight(const Instruction &I) const;
116 ErrorOr<uint64_t> getBlockWeight(const BasicBlock *BB) const;
117 const FunctionSamples *findCalleeFunctionSamples(const CallInst &I) const;
118 const FunctionSamples *findFunctionSamples(const Instruction &I) const;
119 bool inlineHotFunctions(Function &F);
120 void printEdgeWeight(raw_ostream &OS, Edge E);
121 void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const;
122 void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB);
123 bool computeBlockWeights(Function &F);
124 void findEquivalenceClasses(Function &F);
125 void findEquivalencesFor(BasicBlock *BB1,
126 SmallVector<BasicBlock *, 8> Descendants,
127 DominatorTreeBase<BasicBlock> *DomTree);
128 void propagateWeights(Function &F);
129 uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
130 void buildEdges(Function &F);
131 bool propagateThroughEdges(Function &F);
132 void computeDominanceAndLoopInfo(Function &F);
133 unsigned getOffset(unsigned L, unsigned H) const;
134 void clearFunctionData();
136 /// \brief Map basic blocks to their computed weights.
138 /// The weight of a basic block is defined to be the maximum
139 /// of all the instruction weights in that block.
140 BlockWeightMap BlockWeights;
142 /// \brief Map edges to their computed weights.
144 /// Edge weights are computed by propagating basic block weights in
145 /// SampleProfile::propagateWeights.
146 EdgeWeightMap EdgeWeights;
148 /// \brief Set of visited blocks during propagation.
149 SmallPtrSet<const BasicBlock *, 128> VisitedBlocks;
151 /// \brief Set of visited edges during propagation.
152 SmallSet<Edge, 128> VisitedEdges;
154 /// \brief Equivalence classes for block weights.
156 /// Two blocks BB1 and BB2 are in the same equivalence class if they
157 /// dominate and post-dominate each other, and they are in the same loop
158 /// nest. When this happens, the two blocks are guaranteed to execute
159 /// the same number of times.
160 EquivalenceClassMap EquivalenceClass;
162 /// \brief Dominance, post-dominance and loop information.
163 std::unique_ptr<DominatorTree> DT;
164 std::unique_ptr<DominatorTreeBase<BasicBlock>> PDT;
165 std::unique_ptr<LoopInfo> LI;
167 /// \brief Predecessors for each basic block in the CFG.
168 BlockEdgeMap Predecessors;
170 /// \brief Successors for each basic block in the CFG.
171 BlockEdgeMap Successors;
173 /// \brief Profile reader object.
174 std::unique_ptr<SampleProfileReader> Reader;
176 /// \brief Samples collected for the body of this function.
177 FunctionSamples *Samples;
179 /// \brief Name of the profile file to load.
182 /// \brief Flag indicating whether the profile input loaded successfully.
186 class SampleCoverageTracker {
188 SampleCoverageTracker() : SampleCoverage(), TotalUsedSamples(0) {}
190 bool markSamplesUsed(const FunctionSamples *FS, uint32_t LineOffset,
191 uint32_t Discriminator, uint64_t Samples);
192 unsigned computeCoverage(unsigned Used, unsigned Total) const;
193 unsigned countUsedRecords(const FunctionSamples *FS) const;
194 unsigned countBodyRecords(const FunctionSamples *FS) const;
195 uint64_t getTotalUsedSamples() const { return TotalUsedSamples; }
196 uint64_t countBodySamples(const FunctionSamples *FS) const;
198 SampleCoverage.clear();
199 TotalUsedSamples = 0;
203 typedef DenseMap<LineLocation, unsigned> BodySampleCoverageMap;
204 typedef DenseMap<const FunctionSamples *, BodySampleCoverageMap>
205 FunctionSamplesCoverageMap;
207 /// Coverage map for sampling records.
209 /// This map keeps a record of sampling records that have been matched to
210 /// an IR instruction. This is used to detect some form of staleness in
211 /// profiles (see flag -sample-profile-check-coverage).
213 /// Each entry in the map corresponds to a FunctionSamples instance. This is
214 /// another map that counts how many times the sample record at the
215 /// given location has been used.
216 FunctionSamplesCoverageMap SampleCoverage;
218 /// Number of samples used from the profile.
220 /// When a sampling record is used for the first time, the samples from
221 /// that record are added to this accumulator. Coverage is later computed
222 /// based on the total number of samples available in this function and
225 /// Note that this accumulator tracks samples used from a single function
226 /// and all the inlined callsites. Strictly, we should have a map of counters
227 /// keyed by FunctionSamples pointers, but these stats are cleared after
228 /// every function, so we just need to keep a single counter.
229 uint64_t TotalUsedSamples;
232 SampleCoverageTracker CoverageTracker;
235 /// Mark as used the sample record for the given function samples at
236 /// (LineOffset, Discriminator).
238 /// \returns true if this is the first time we mark the given record.
239 bool SampleCoverageTracker::markSamplesUsed(const FunctionSamples *FS,
241 uint32_t Discriminator,
243 LineLocation Loc(LineOffset, Discriminator);
244 unsigned &Count = SampleCoverage[FS][Loc];
245 bool FirstTime = (++Count == 1);
247 TotalUsedSamples += Samples;
251 /// Return the number of sample records that were applied from this profile.
253 SampleCoverageTracker::countUsedRecords(const FunctionSamples *FS) const {
254 auto I = SampleCoverage.find(FS);
256 // The size of the coverage map for FS represents the number of records
257 // that were marked used at least once.
258 unsigned Count = (I != SampleCoverage.end()) ? I->second.size() : 0;
260 // If there are inlined callsites in this function, count the samples found
261 // in the respective bodies. However, do not bother counting callees with 0
262 // total samples, these are callees that were never invoked at runtime.
263 for (const auto &I : FS->getCallsiteSamples()) {
264 const FunctionSamples *CalleeSamples = &I.second;
265 if (CalleeSamples->getTotalSamples() > 0)
266 Count += countUsedRecords(CalleeSamples);
272 /// Return the number of sample records in the body of this profile.
274 /// The count includes all the samples in inlined callees. However, callsites
275 /// with 0 samples indicate inlined function calls that were never actually
276 /// invoked at runtime. Ignore these callsites for coverage purposes.
278 SampleCoverageTracker::countBodyRecords(const FunctionSamples *FS) const {
279 unsigned Count = FS->getBodySamples().size();
281 // Count all the callsites with non-zero samples.
282 for (const auto &I : FS->getCallsiteSamples()) {
283 const FunctionSamples *CalleeSamples = &I.second;
284 if (CalleeSamples->getTotalSamples() > 0)
285 Count += countBodyRecords(CalleeSamples);
291 /// Return the number of samples collected in the body of this profile.
293 /// The count includes all the samples in inlined callees. However, callsites
294 /// with 0 samples indicate inlined function calls that were never actually
295 /// invoked at runtime. Ignore these callsites for coverage purposes.
297 SampleCoverageTracker::countBodySamples(const FunctionSamples *FS) const {
299 for (const auto &I : FS->getBodySamples())
300 Total += I.second.getSamples();
302 // Count all the callsites with non-zero samples.
303 for (const auto &I : FS->getCallsiteSamples()) {
304 const FunctionSamples *CalleeSamples = &I.second;
305 if (CalleeSamples->getTotalSamples() > 0)
306 Total += countBodySamples(CalleeSamples);
312 /// Return the fraction of sample records used in this profile.
314 /// The returned value is an unsigned integer in the range 0-100 indicating
315 /// the percentage of sample records that were used while applying this
316 /// profile to the associated function.
317 unsigned SampleCoverageTracker::computeCoverage(unsigned Used,
318 unsigned Total) const {
319 assert(Used <= Total &&
320 "number of used records cannot exceed the total number of records");
321 return Total > 0 ? Used * 100 / Total : 100;
324 /// Clear all the per-function data used to load samples and propagate weights.
325 void SampleProfileLoader::clearFunctionData() {
326 BlockWeights.clear();
328 VisitedBlocks.clear();
329 VisitedEdges.clear();
330 EquivalenceClass.clear();
334 Predecessors.clear();
336 CoverageTracker.clear();
339 /// \brief Returns the offset of lineno \p L to head_lineno \p H
342 /// \param H Header lineno of the function
344 /// \returns offset to the header lineno. 16 bits are used to represent offset.
345 /// We assume that a single function will not exceed 65535 LOC.
346 unsigned SampleProfileLoader::getOffset(unsigned L, unsigned H) const {
347 return (L - H) & 0xffff;
350 /// \brief Print the weight of edge \p E on stream \p OS.
352 /// \param OS Stream to emit the output to.
353 /// \param E Edge to print.
354 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
355 OS << "weight[" << E.first->getName() << "->" << E.second->getName()
356 << "]: " << EdgeWeights[E] << "\n";
359 /// \brief Print the equivalence class of block \p BB on stream \p OS.
361 /// \param OS Stream to emit the output to.
362 /// \param BB Block to print.
363 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
364 const BasicBlock *BB) {
365 const BasicBlock *Equiv = EquivalenceClass[BB];
366 OS << "equivalence[" << BB->getName()
367 << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
370 /// \brief Print the weight of block \p BB on stream \p OS.
372 /// \param OS Stream to emit the output to.
373 /// \param BB Block to print.
374 void SampleProfileLoader::printBlockWeight(raw_ostream &OS,
375 const BasicBlock *BB) const {
376 const auto &I = BlockWeights.find(BB);
377 uint64_t W = (I == BlockWeights.end() ? 0 : I->second);
378 OS << "weight[" << BB->getName() << "]: " << W << "\n";
381 /// \brief Get the weight for an instruction.
383 /// The "weight" of an instruction \p Inst is the number of samples
384 /// collected on that instruction at runtime. To retrieve it, we
385 /// need to compute the line number of \p Inst relative to the start of its
386 /// function. We use HeaderLineno to compute the offset. We then
387 /// look up the samples collected for \p Inst using BodySamples.
389 /// \param Inst Instruction to query.
391 /// \returns the weight of \p Inst.
393 SampleProfileLoader::getInstWeight(const Instruction &Inst) const {
394 DebugLoc DLoc = Inst.getDebugLoc();
396 return std::error_code();
398 const FunctionSamples *FS = findFunctionSamples(Inst);
400 return std::error_code();
402 const DILocation *DIL = DLoc;
403 unsigned Lineno = DLoc.getLine();
404 unsigned HeaderLineno = DIL->getScope()->getSubprogram()->getLine();
406 uint32_t LineOffset = getOffset(Lineno, HeaderLineno);
407 uint32_t Discriminator = DIL->getDiscriminator();
408 ErrorOr<uint64_t> R = FS->findSamplesAt(LineOffset, Discriminator);
411 CoverageTracker.markSamplesUsed(FS, LineOffset, Discriminator, R.get());
413 const Function *F = Inst.getParent()->getParent();
414 LLVMContext &Ctx = F->getContext();
415 emitOptimizationRemark(
416 Ctx, DEBUG_TYPE, *F, DLoc,
417 Twine("Applied ") + Twine(*R) + " samples from profile (offset: " +
419 ((Discriminator) ? Twine(".") + Twine(Discriminator) : "") + ")");
421 DEBUG(dbgs() << " " << Lineno << "." << DIL->getDiscriminator() << ":"
422 << Inst << " (line offset: " << Lineno - HeaderLineno << "."
423 << DIL->getDiscriminator() << " - weight: " << R.get()
429 /// \brief Compute the weight of a basic block.
431 /// The weight of basic block \p BB is the maximum weight of all the
432 /// instructions in BB.
434 /// \param BB The basic block to query.
436 /// \returns the weight for \p BB.
438 SampleProfileLoader::getBlockWeight(const BasicBlock *BB) const {
441 for (auto &I : BB->getInstList()) {
442 const ErrorOr<uint64_t> &R = getInstWeight(I);
443 if (R && R.get() >= Weight) {
451 return std::error_code();
454 /// \brief Compute and store the weights of every basic block.
456 /// This populates the BlockWeights map by computing
457 /// the weights of every basic block in the CFG.
459 /// \param F The function to query.
460 bool SampleProfileLoader::computeBlockWeights(Function &F) {
461 bool Changed = false;
462 DEBUG(dbgs() << "Block weights\n");
463 for (const auto &BB : F) {
464 ErrorOr<uint64_t> Weight = getBlockWeight(&BB);
466 BlockWeights[&BB] = Weight.get();
467 VisitedBlocks.insert(&BB);
470 DEBUG(printBlockWeight(dbgs(), &BB));
476 /// \brief Get the FunctionSamples for a call instruction.
478 /// The FunctionSamples of a call instruction \p Inst is the inlined
479 /// instance in which that call instruction is calling to. It contains
480 /// all samples that resides in the inlined instance. We first find the
481 /// inlined instance in which the call instruction is from, then we
482 /// traverse its children to find the callsite with the matching
483 /// location and callee function name.
485 /// \param Inst Call instruction to query.
487 /// \returns The FunctionSamples pointer to the inlined instance.
488 const FunctionSamples *
489 SampleProfileLoader::findCalleeFunctionSamples(const CallInst &Inst) const {
490 const DILocation *DIL = Inst.getDebugLoc();
494 DISubprogram *SP = DIL->getScope()->getSubprogram();
498 Function *CalleeFunc = Inst.getCalledFunction();
503 StringRef CalleeName = CalleeFunc->getName();
504 const FunctionSamples *FS = findFunctionSamples(Inst);
508 return FS->findFunctionSamplesAt(
509 CallsiteLocation(getOffset(DIL->getLine(), SP->getLine()),
510 DIL->getDiscriminator(), CalleeName));
513 /// \brief Get the FunctionSamples for an instruction.
515 /// The FunctionSamples of an instruction \p Inst is the inlined instance
516 /// in which that instruction is coming from. We traverse the inline stack
517 /// of that instruction, and match it with the tree nodes in the profile.
519 /// \param Inst Instruction to query.
521 /// \returns the FunctionSamples pointer to the inlined instance.
522 const FunctionSamples *
523 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
524 SmallVector<CallsiteLocation, 10> S;
525 const DILocation *DIL = Inst.getDebugLoc();
529 StringRef CalleeName;
530 for (const DILocation *DIL = Inst.getDebugLoc(); DIL;
531 DIL = DIL->getInlinedAt()) {
532 DISubprogram *SP = DIL->getScope()->getSubprogram();
535 if (!CalleeName.empty()) {
536 S.push_back(CallsiteLocation(getOffset(DIL->getLine(), SP->getLine()),
537 DIL->getDiscriminator(), CalleeName));
539 CalleeName = SP->getLinkageName();
543 const FunctionSamples *FS = Samples;
544 for (int i = S.size() - 1; i >= 0 && FS != nullptr; i--) {
545 FS = FS->findFunctionSamplesAt(S[i]);
550 /// \brief Iteratively inline hot callsites of a function.
552 /// Iteratively traverse all callsites of the function \p F, and find if
553 /// the corresponding inlined instance exists and is hot in profile. If
554 /// it is hot enough, inline the callsites and adds new callsites of the
555 /// callee into the caller.
557 /// TODO: investigate the possibility of not invoking InlineFunction directly.
559 /// \param F function to perform iterative inlining.
561 /// \returns True if there is any inline happened.
562 bool SampleProfileLoader::inlineHotFunctions(Function &F) {
563 bool Changed = false;
564 LLVMContext &Ctx = F.getContext();
566 bool LocalChanged = false;
567 SmallVector<CallInst *, 10> CIS;
569 for (auto &I : BB.getInstList()) {
570 CallInst *CI = dyn_cast<CallInst>(&I);
572 const FunctionSamples *FS = findCalleeFunctionSamples(*CI);
573 if (FS && FS->getTotalSamples() > 0) {
579 for (auto CI : CIS) {
580 InlineFunctionInfo IFI;
581 Function *CalledFunction = CI->getCalledFunction();
582 DebugLoc DLoc = CI->getDebugLoc();
583 uint64_t NumSamples = findCalleeFunctionSamples(*CI)->getTotalSamples();
584 if (InlineFunction(CI, IFI)) {
586 emitOptimizationRemark(Ctx, DEBUG_TYPE, F, DLoc,
587 Twine("inlined hot callee '") +
588 CalledFunction->getName() + "' with " +
589 Twine(NumSamples) + " samples into '" +
602 /// \brief Find equivalence classes for the given block.
604 /// This finds all the blocks that are guaranteed to execute the same
605 /// number of times as \p BB1. To do this, it traverses all the
606 /// descendants of \p BB1 in the dominator or post-dominator tree.
608 /// A block BB2 will be in the same equivalence class as \p BB1 if
609 /// the following holds:
611 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
612 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
613 /// dominate BB1 in the post-dominator tree.
615 /// 2- Both BB2 and \p BB1 must be in the same loop.
617 /// For every block BB2 that meets those two requirements, we set BB2's
618 /// equivalence class to \p BB1.
620 /// \param BB1 Block to check.
621 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
622 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
623 /// with blocks from \p BB1's dominator tree, then
624 /// this is the post-dominator tree, and vice versa.
625 void SampleProfileLoader::findEquivalencesFor(
626 BasicBlock *BB1, SmallVector<BasicBlock *, 8> Descendants,
627 DominatorTreeBase<BasicBlock> *DomTree) {
628 const BasicBlock *EC = EquivalenceClass[BB1];
629 uint64_t Weight = BlockWeights[EC];
630 for (const auto *BB2 : Descendants) {
631 bool IsDomParent = DomTree->dominates(BB2, BB1);
632 bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
633 if (BB1 != BB2 && IsDomParent && IsInSameLoop) {
634 EquivalenceClass[BB2] = EC;
636 // If BB2 is heavier than BB1, make BB2 have the same weight
639 // Note that we don't worry about the opposite situation here
640 // (when BB2 is lighter than BB1). We will deal with this
641 // during the propagation phase. Right now, we just want to
642 // make sure that BB1 has the largest weight of all the
643 // members of its equivalence set.
644 Weight = std::max(Weight, BlockWeights[BB2]);
647 BlockWeights[EC] = Weight;
650 /// \brief Find equivalence classes.
652 /// Since samples may be missing from blocks, we can fill in the gaps by setting
653 /// the weights of all the blocks in the same equivalence class to the same
654 /// weight. To compute the concept of equivalence, we use dominance and loop
655 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
656 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
658 /// \param F The function to query.
659 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
660 SmallVector<BasicBlock *, 8> DominatedBBs;
661 DEBUG(dbgs() << "\nBlock equivalence classes\n");
662 // Find equivalence sets based on dominance and post-dominance information.
664 BasicBlock *BB1 = &BB;
666 // Compute BB1's equivalence class once.
667 if (EquivalenceClass.count(BB1)) {
668 DEBUG(printBlockEquivalence(dbgs(), BB1));
672 // By default, blocks are in their own equivalence class.
673 EquivalenceClass[BB1] = BB1;
675 // Traverse all the blocks dominated by BB1. We are looking for
676 // every basic block BB2 such that:
678 // 1- BB1 dominates BB2.
679 // 2- BB2 post-dominates BB1.
680 // 3- BB1 and BB2 are in the same loop nest.
682 // If all those conditions hold, it means that BB2 is executed
683 // as many times as BB1, so they are placed in the same equivalence
684 // class by making BB2's equivalence class be BB1.
685 DominatedBBs.clear();
686 DT->getDescendants(BB1, DominatedBBs);
687 findEquivalencesFor(BB1, DominatedBBs, PDT.get());
689 DEBUG(printBlockEquivalence(dbgs(), BB1));
692 // Assign weights to equivalence classes.
694 // All the basic blocks in the same equivalence class will execute
695 // the same number of times. Since we know that the head block in
696 // each equivalence class has the largest weight, assign that weight
697 // to all the blocks in that equivalence class.
698 DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n");
700 const BasicBlock *BB = &BI;
701 const BasicBlock *EquivBB = EquivalenceClass[BB];
703 BlockWeights[BB] = BlockWeights[EquivBB];
704 DEBUG(printBlockWeight(dbgs(), BB));
708 /// \brief Visit the given edge to decide if it has a valid weight.
710 /// If \p E has not been visited before, we copy to \p UnknownEdge
711 /// and increment the count of unknown edges.
713 /// \param E Edge to visit.
714 /// \param NumUnknownEdges Current number of unknown edges.
715 /// \param UnknownEdge Set if E has not been visited before.
717 /// \returns E's weight, if known. Otherwise, return 0.
718 uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
720 if (!VisitedEdges.count(E)) {
721 (*NumUnknownEdges)++;
726 return EdgeWeights[E];
729 /// \brief Propagate weights through incoming/outgoing edges.
731 /// If the weight of a basic block is known, and there is only one edge
732 /// with an unknown weight, we can calculate the weight of that edge.
734 /// Similarly, if all the edges have a known count, we can calculate the
735 /// count of the basic block, if needed.
737 /// \param F Function to process.
739 /// \returns True if new weights were assigned to edges or blocks.
740 bool SampleProfileLoader::propagateThroughEdges(Function &F) {
741 bool Changed = false;
742 DEBUG(dbgs() << "\nPropagation through edges\n");
743 for (const auto &BI : F) {
744 const BasicBlock *BB = &BI;
745 const BasicBlock *EC = EquivalenceClass[BB];
747 // Visit all the predecessor and successor edges to determine
748 // which ones have a weight assigned already. Note that it doesn't
749 // matter that we only keep track of a single unknown edge. The
750 // only case we are interested in handling is when only a single
751 // edge is unknown (see setEdgeOrBlockWeight).
752 for (unsigned i = 0; i < 2; i++) {
753 uint64_t TotalWeight = 0;
754 unsigned NumUnknownEdges = 0;
755 Edge UnknownEdge, SelfReferentialEdge;
758 // First, visit all predecessor edges.
759 for (auto *Pred : Predecessors[BB]) {
760 Edge E = std::make_pair(Pred, BB);
761 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
762 if (E.first == E.second)
763 SelfReferentialEdge = E;
766 // On the second round, visit all successor edges.
767 for (auto *Succ : Successors[BB]) {
768 Edge E = std::make_pair(BB, Succ);
769 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
773 // After visiting all the edges, there are three cases that we
774 // can handle immediately:
776 // - All the edge weights are known (i.e., NumUnknownEdges == 0).
777 // In this case, we simply check that the sum of all the edges
778 // is the same as BB's weight. If not, we change BB's weight
779 // to match. Additionally, if BB had not been visited before,
780 // we mark it visited.
782 // - Only one edge is unknown and BB has already been visited.
783 // In this case, we can compute the weight of the edge by
784 // subtracting the total block weight from all the known
785 // edge weights. If the edges weight more than BB, then the
786 // edge of the last remaining edge is set to zero.
788 // - There exists a self-referential edge and the weight of BB is
789 // known. In this case, this edge can be based on BB's weight.
790 // We add up all the other known edges and set the weight on
791 // the self-referential edge as we did in the previous case.
793 // In any other case, we must continue iterating. Eventually,
794 // all edges will get a weight, or iteration will stop when
795 // it reaches SampleProfileMaxPropagateIterations.
796 if (NumUnknownEdges <= 1) {
797 uint64_t &BBWeight = BlockWeights[EC];
798 if (NumUnknownEdges == 0) {
799 // If we already know the weight of all edges, the weight of the
800 // basic block can be computed. It should be no larger than the sum
801 // of all edge weights.
802 if (TotalWeight > BBWeight) {
803 BBWeight = TotalWeight;
805 DEBUG(dbgs() << "All edge weights for " << BB->getName()
806 << " known. Set weight for block: ";
807 printBlockWeight(dbgs(), BB););
809 if (VisitedBlocks.insert(EC).second)
811 } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) {
812 // If there is a single unknown edge and the block has been
813 // visited, then we can compute E's weight.
814 if (BBWeight >= TotalWeight)
815 EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
817 EdgeWeights[UnknownEdge] = 0;
818 VisitedEdges.insert(UnknownEdge);
820 DEBUG(dbgs() << "Set weight for edge: ";
821 printEdgeWeight(dbgs(), UnknownEdge));
823 } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) {
824 uint64_t &BBWeight = BlockWeights[BB];
825 // We have a self-referential edge and the weight of BB is known.
826 if (BBWeight >= TotalWeight)
827 EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
829 EdgeWeights[SelfReferentialEdge] = 0;
830 VisitedEdges.insert(SelfReferentialEdge);
832 DEBUG(dbgs() << "Set self-referential edge weight to: ";
833 printEdgeWeight(dbgs(), SelfReferentialEdge));
841 /// \brief Build in/out edge lists for each basic block in the CFG.
843 /// We are interested in unique edges. If a block B1 has multiple
844 /// edges to another block B2, we only add a single B1->B2 edge.
845 void SampleProfileLoader::buildEdges(Function &F) {
847 BasicBlock *B1 = &BI;
849 // Add predecessors for B1.
850 SmallPtrSet<BasicBlock *, 16> Visited;
851 if (!Predecessors[B1].empty())
852 llvm_unreachable("Found a stale predecessors list in a basic block.");
853 for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
854 BasicBlock *B2 = *PI;
855 if (Visited.insert(B2).second)
856 Predecessors[B1].push_back(B2);
859 // Add successors for B1.
861 if (!Successors[B1].empty())
862 llvm_unreachable("Found a stale successors list in a basic block.");
863 for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
864 BasicBlock *B2 = *SI;
865 if (Visited.insert(B2).second)
866 Successors[B1].push_back(B2);
871 /// \brief Propagate weights into edges
873 /// The following rules are applied to every block BB in the CFG:
875 /// - If BB has a single predecessor/successor, then the weight
876 /// of that edge is the weight of the block.
878 /// - If all incoming or outgoing edges are known except one, and the
879 /// weight of the block is already known, the weight of the unknown
880 /// edge will be the weight of the block minus the sum of all the known
881 /// edges. If the sum of all the known edges is larger than BB's weight,
882 /// we set the unknown edge weight to zero.
884 /// - If there is a self-referential edge, and the weight of the block is
885 /// known, the weight for that edge is set to the weight of the block
886 /// minus the weight of the other incoming edges to that block (if
888 void SampleProfileLoader::propagateWeights(Function &F) {
892 // Add an entry count to the function using the samples gathered
893 // at the function entry.
894 F.setEntryCount(Samples->getHeadSamples());
896 // Before propagation starts, build, for each block, a list of
897 // unique predecessors and successors. This is necessary to handle
898 // identical edges in multiway branches. Since we visit all blocks and all
899 // edges of the CFG, it is cleaner to build these lists once at the start
903 // Propagate until we converge or we go past the iteration limit.
904 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
905 Changed = propagateThroughEdges(F);
908 // Generate MD_prof metadata for every branch instruction using the
909 // edge weights computed during propagation.
910 DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
911 LLVMContext &Ctx = F.getContext();
914 BasicBlock *BB = &BI;
915 TerminatorInst *TI = BB->getTerminator();
916 if (TI->getNumSuccessors() == 1)
918 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
921 DEBUG(dbgs() << "\nGetting weights for branch at line "
922 << TI->getDebugLoc().getLine() << ".\n");
923 SmallVector<uint32_t, 4> Weights;
924 uint32_t MaxWeight = 0;
926 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
927 BasicBlock *Succ = TI->getSuccessor(I);
928 Edge E = std::make_pair(BB, Succ);
929 uint64_t Weight = EdgeWeights[E];
930 DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
931 // Use uint32_t saturated arithmetic to adjust the incoming weights,
932 // if needed. Sample counts in profiles are 64-bit unsigned values,
933 // but internally branch weights are expressed as 32-bit values.
934 if (Weight > std::numeric_limits<uint32_t>::max()) {
935 DEBUG(dbgs() << " (saturated due to uint32_t overflow)");
936 Weight = std::numeric_limits<uint32_t>::max();
938 Weights.push_back(static_cast<uint32_t>(Weight));
940 if (Weight > MaxWeight) {
942 MaxDestLoc = Succ->getFirstNonPHIOrDbgOrLifetime()->getDebugLoc();
947 // Only set weights if there is at least one non-zero weight.
948 // In any other case, let the analyzer set weights.
950 DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
951 TI->setMetadata(llvm::LLVMContext::MD_prof,
952 MDB.createBranchWeights(Weights));
953 DebugLoc BranchLoc = TI->getDebugLoc();
954 emitOptimizationRemark(
955 Ctx, DEBUG_TYPE, F, MaxDestLoc,
956 Twine("most popular destination for conditional branches at ") +
957 ((BranchLoc) ? Twine(BranchLoc->getFilename() + ":" +
958 Twine(BranchLoc.getLine()) + ":" +
959 Twine(BranchLoc.getCol()))
960 : Twine("<UNKNOWN LOCATION>")));
962 DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
967 /// \brief Get the line number for the function header.
969 /// This looks up function \p F in the current compilation unit and
970 /// retrieves the line number where the function is defined. This is
971 /// line 0 for all the samples read from the profile file. Every line
972 /// number is relative to this line.
974 /// \param F Function object to query.
976 /// \returns the line number where \p F is defined. If it returns 0,
977 /// it means that there is no debug information available for \p F.
978 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
979 if (DISubprogram *S = getDISubprogram(&F))
982 // If the start of \p F is missing, emit a diagnostic to inform the user
983 // about the missed opportunity.
984 F.getContext().diagnose(DiagnosticInfoSampleProfile(
985 "No debug information found in function " + F.getName() +
986 ": Function profile not used",
991 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
992 DT.reset(new DominatorTree);
995 PDT.reset(new DominatorTreeBase<BasicBlock>(true));
998 LI.reset(new LoopInfo);
1002 /// \brief Generate branch weight metadata for all branches in \p F.
1004 /// Branch weights are computed out of instruction samples using a
1005 /// propagation heuristic. Propagation proceeds in 3 phases:
1007 /// 1- Assignment of block weights. All the basic blocks in the function
1008 /// are initial assigned the same weight as their most frequently
1009 /// executed instruction.
1011 /// 2- Creation of equivalence classes. Since samples may be missing from
1012 /// blocks, we can fill in the gaps by setting the weights of all the
1013 /// blocks in the same equivalence class to the same weight. To compute
1014 /// the concept of equivalence, we use dominance and loop information.
1015 /// Two blocks B1 and B2 are in the same equivalence class if B1
1016 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
1018 /// 3- Propagation of block weights into edges. This uses a simple
1019 /// propagation heuristic. The following rules are applied to every
1020 /// block BB in the CFG:
1022 /// - If BB has a single predecessor/successor, then the weight
1023 /// of that edge is the weight of the block.
1025 /// - If all the edges are known except one, and the weight of the
1026 /// block is already known, the weight of the unknown edge will
1027 /// be the weight of the block minus the sum of all the known
1028 /// edges. If the sum of all the known edges is larger than BB's weight,
1029 /// we set the unknown edge weight to zero.
1031 /// - If there is a self-referential edge, and the weight of the block is
1032 /// known, the weight for that edge is set to the weight of the block
1033 /// minus the weight of the other incoming edges to that block (if
1036 /// Since this propagation is not guaranteed to finalize for every CFG, we
1037 /// only allow it to proceed for a limited number of iterations (controlled
1038 /// by -sample-profile-max-propagate-iterations).
1040 /// FIXME: Try to replace this propagation heuristic with a scheme
1041 /// that is guaranteed to finalize. A work-list approach similar to
1042 /// the standard value propagation algorithm used by SSA-CCP might
1045 /// Once all the branch weights are computed, we emit the MD_prof
1046 /// metadata on BB using the computed values for each of its branches.
1048 /// \param F The function to query.
1050 /// \returns true if \p F was modified. Returns false, otherwise.
1051 bool SampleProfileLoader::emitAnnotations(Function &F) {
1052 bool Changed = false;
1054 if (getFunctionLoc(F) == 0)
1057 DEBUG(dbgs() << "Line number for the first instruction in " << F.getName()
1058 << ": " << getFunctionLoc(F) << "\n");
1060 Changed |= inlineHotFunctions(F);
1062 // Compute basic block weights.
1063 Changed |= computeBlockWeights(F);
1066 // Compute dominance and loop info needed for propagation.
1067 computeDominanceAndLoopInfo(F);
1069 // Find equivalence classes.
1070 findEquivalenceClasses(F);
1072 // Propagate weights to all edges.
1073 propagateWeights(F);
1076 // If coverage checking was requested, compute it now.
1077 if (SampleProfileRecordCoverage) {
1078 unsigned Used = CoverageTracker.countUsedRecords(Samples);
1079 unsigned Total = CoverageTracker.countBodyRecords(Samples);
1080 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1081 if (Coverage < SampleProfileRecordCoverage) {
1082 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1083 getDISubprogram(&F)->getFilename(), getFunctionLoc(F),
1084 Twine(Used) + " of " + Twine(Total) + " available profile records (" +
1085 Twine(Coverage) + "%) were applied",
1090 if (SampleProfileSampleCoverage) {
1091 uint64_t Used = CoverageTracker.getTotalUsedSamples();
1092 uint64_t Total = CoverageTracker.countBodySamples(Samples);
1093 unsigned Coverage = CoverageTracker.computeCoverage(Used, Total);
1094 if (Coverage < SampleProfileSampleCoverage) {
1095 F.getContext().diagnose(DiagnosticInfoSampleProfile(
1096 getDISubprogram(&F)->getFilename(), getFunctionLoc(F),
1097 Twine(Used) + " of " + Twine(Total) + " available profile samples (" +
1098 Twine(Coverage) + "%) were applied",
1105 char SampleProfileLoader::ID = 0;
1106 INITIALIZE_PASS_BEGIN(SampleProfileLoader, "sample-profile",
1107 "Sample Profile loader", false, false)
1108 INITIALIZE_PASS_DEPENDENCY(AddDiscriminators)
1109 INITIALIZE_PASS_END(SampleProfileLoader, "sample-profile",
1110 "Sample Profile loader", false, false)
1112 bool SampleProfileLoader::doInitialization(Module &M) {
1113 auto &Ctx = M.getContext();
1114 auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
1115 if (std::error_code EC = ReaderOrErr.getError()) {
1116 std::string Msg = "Could not open profile: " + EC.message();
1117 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename, Msg));
1120 Reader = std::move(ReaderOrErr.get());
1121 ProfileIsValid = (Reader->read() == sampleprof_error::success);
1125 ModulePass *llvm::createSampleProfileLoaderPass() {
1126 return new SampleProfileLoader(SampleProfileFile);
1129 ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) {
1130 return new SampleProfileLoader(Name);
1133 bool SampleProfileLoader::runOnModule(Module &M) {
1134 if (!ProfileIsValid)
1137 bool retval = false;
1139 if (!F.isDeclaration()) {
1140 clearFunctionData();
1141 retval |= runOnFunction(F);
1146 bool SampleProfileLoader::runOnFunction(Function &F) {
1147 Samples = Reader->getSamplesFor(F);
1148 if (!Samples->empty())
1149 return emitAnnotations(F);