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."));
68 typedef DenseMap<const BasicBlock *, uint64_t> BlockWeightMap;
69 typedef DenseMap<const BasicBlock *, const BasicBlock *> EquivalenceClassMap;
70 typedef std::pair<const BasicBlock *, const BasicBlock *> Edge;
71 typedef DenseMap<Edge, uint64_t> EdgeWeightMap;
72 typedef DenseMap<const BasicBlock *, SmallVector<const BasicBlock *, 8>>
75 /// \brief Sample profile pass.
77 /// This pass reads profile data from the file specified by
78 /// -sample-profile-file and annotates every affected function with the
79 /// profile information found in that file.
80 class SampleProfileLoader : public ModulePass {
82 // Class identification, replacement for typeinfo
85 SampleProfileLoader(StringRef Name = SampleProfileFile)
86 : ModulePass(ID), DT(nullptr), PDT(nullptr), LI(nullptr), Reader(),
87 Samples(nullptr), Filename(Name), ProfileIsValid(false) {
88 initializeSampleProfileLoaderPass(*PassRegistry::getPassRegistry());
91 bool doInitialization(Module &M) override;
93 void dump() { Reader->dump(); }
95 const char *getPassName() const override { return "Sample profile pass"; }
97 bool runOnModule(Module &M) override;
99 void getAnalysisUsage(AnalysisUsage &AU) const override {
100 AU.setPreservesCFG();
104 bool runOnFunction(Function &F);
105 unsigned getFunctionLoc(Function &F);
106 bool emitAnnotations(Function &F);
107 ErrorOr<uint64_t> getInstWeight(const Instruction &I) const;
108 ErrorOr<uint64_t> getBlockWeight(const BasicBlock *BB) const;
109 const FunctionSamples *findCalleeFunctionSamples(const CallInst &I) const;
110 const FunctionSamples *findFunctionSamples(const Instruction &I) const;
111 bool inlineHotFunctions(Function &F);
112 void printEdgeWeight(raw_ostream &OS, Edge E);
113 void printBlockWeight(raw_ostream &OS, const BasicBlock *BB) const;
114 void printBlockEquivalence(raw_ostream &OS, const BasicBlock *BB);
115 bool computeBlockWeights(Function &F);
116 void findEquivalenceClasses(Function &F);
117 void findEquivalencesFor(BasicBlock *BB1,
118 SmallVector<BasicBlock *, 8> Descendants,
119 DominatorTreeBase<BasicBlock> *DomTree);
120 void propagateWeights(Function &F);
121 uint64_t visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
122 void buildEdges(Function &F);
123 bool propagateThroughEdges(Function &F);
124 void computeDominanceAndLoopInfo(Function &F);
125 unsigned getOffset(unsigned L, unsigned H) const;
127 /// \brief Map basic blocks to their computed weights.
129 /// The weight of a basic block is defined to be the maximum
130 /// of all the instruction weights in that block.
131 BlockWeightMap BlockWeights;
133 /// \brief Map edges to their computed weights.
135 /// Edge weights are computed by propagating basic block weights in
136 /// SampleProfile::propagateWeights.
137 EdgeWeightMap EdgeWeights;
139 /// \brief Set of visited blocks during propagation.
140 SmallPtrSet<const BasicBlock *, 128> VisitedBlocks;
142 /// \brief Set of visited edges during propagation.
143 SmallSet<Edge, 128> VisitedEdges;
145 /// \brief Equivalence classes for block weights.
147 /// Two blocks BB1 and BB2 are in the same equivalence class if they
148 /// dominate and post-dominate each other, and they are in the same loop
149 /// nest. When this happens, the two blocks are guaranteed to execute
150 /// the same number of times.
151 EquivalenceClassMap EquivalenceClass;
153 /// \brief Dominance, post-dominance and loop information.
154 std::unique_ptr<DominatorTree> DT;
155 std::unique_ptr<DominatorTreeBase<BasicBlock>> PDT;
156 std::unique_ptr<LoopInfo> LI;
158 /// \brief Predecessors for each basic block in the CFG.
159 BlockEdgeMap Predecessors;
161 /// \brief Successors for each basic block in the CFG.
162 BlockEdgeMap Successors;
164 /// \brief Profile reader object.
165 std::unique_ptr<SampleProfileReader> Reader;
167 /// \brief Samples collected for the body of this function.
168 FunctionSamples *Samples;
170 /// \brief Name of the profile file to load.
173 /// \brief Flag indicating whether the profile input loaded successfully.
178 /// \brief Returns the offset of lineno \p L to head_lineno \p H
181 /// \param H Header lineno of the function
183 /// \returns offset to the header lineno. 16 bits are used to represent offset.
184 /// We assume that a single function will not exceed 65535 LOC.
185 unsigned SampleProfileLoader::getOffset(unsigned L, unsigned H) const {
186 return (L - H) & 0xffff;
189 /// \brief Print the weight of edge \p E on stream \p OS.
191 /// \param OS Stream to emit the output to.
192 /// \param E Edge to print.
193 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
194 OS << "weight[" << E.first->getName() << "->" << E.second->getName()
195 << "]: " << EdgeWeights[E] << "\n";
198 /// \brief Print the equivalence class of block \p BB on stream \p OS.
200 /// \param OS Stream to emit the output to.
201 /// \param BB Block to print.
202 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
203 const BasicBlock *BB) {
204 const BasicBlock *Equiv = EquivalenceClass[BB];
205 OS << "equivalence[" << BB->getName()
206 << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
209 /// \brief Print the weight of block \p BB on stream \p OS.
211 /// \param OS Stream to emit the output to.
212 /// \param BB Block to print.
213 void SampleProfileLoader::printBlockWeight(raw_ostream &OS,
214 const BasicBlock *BB) const {
215 const auto &I = BlockWeights.find(BB);
216 uint64_t W = (I == BlockWeights.end() ? 0 : I->second);
217 OS << "weight[" << BB->getName() << "]: " << W << "\n";
220 /// \brief Get the weight for an instruction.
222 /// The "weight" of an instruction \p Inst is the number of samples
223 /// collected on that instruction at runtime. To retrieve it, we
224 /// need to compute the line number of \p Inst relative to the start of its
225 /// function. We use HeaderLineno to compute the offset. We then
226 /// look up the samples collected for \p Inst using BodySamples.
228 /// \param Inst Instruction to query.
230 /// \returns the weight of \p Inst.
232 SampleProfileLoader::getInstWeight(const Instruction &Inst) const {
233 DebugLoc DLoc = Inst.getDebugLoc();
235 return std::error_code();
237 const FunctionSamples *FS = findFunctionSamples(Inst);
239 return std::error_code();
241 const DILocation *DIL = DLoc;
242 unsigned Lineno = DLoc.getLine();
243 unsigned HeaderLineno = DIL->getScope()->getSubprogram()->getLine();
245 ErrorOr<uint64_t> R = FS->findSamplesAt(getOffset(Lineno, HeaderLineno),
246 DIL->getDiscriminator());
248 DEBUG(dbgs() << " " << Lineno << "." << DIL->getDiscriminator() << ":"
249 << Inst << " (line offset: " << Lineno - HeaderLineno << "."
250 << DIL->getDiscriminator() << " - weight: " << R.get()
255 /// \brief Compute the weight of a basic block.
257 /// The weight of basic block \p BB is the maximum weight of all the
258 /// instructions in BB.
260 /// \param BB The basic block to query.
262 /// \returns the weight for \p BB.
264 SampleProfileLoader::getBlockWeight(const BasicBlock *BB) const {
267 for (auto &I : BB->getInstList()) {
268 const ErrorOr<uint64_t> &R = getInstWeight(I);
269 if (R && R.get() >= Weight) {
277 return std::error_code();
280 /// \brief Compute and store the weights of every basic block.
282 /// This populates the BlockWeights map by computing
283 /// the weights of every basic block in the CFG.
285 /// \param F The function to query.
286 bool SampleProfileLoader::computeBlockWeights(Function &F) {
287 bool Changed = false;
288 DEBUG(dbgs() << "Block weights\n");
289 for (const auto &BB : F) {
290 ErrorOr<uint64_t> Weight = getBlockWeight(&BB);
292 BlockWeights[&BB] = Weight.get();
293 VisitedBlocks.insert(&BB);
296 DEBUG(printBlockWeight(dbgs(), &BB));
302 /// \brief Get the FunctionSamples for a call instruction.
304 /// The FunctionSamples of a call instruction \p Inst is the inlined
305 /// instance in which that call instruction is calling to. It contains
306 /// all samples that resides in the inlined instance. We first find the
307 /// inlined instance in which the call instruction is from, then we
308 /// traverse its children to find the callsite with the matching
309 /// location and callee function name.
311 /// \param Inst Call instruction to query.
313 /// \returns The FunctionSamples pointer to the inlined instance.
314 const FunctionSamples *
315 SampleProfileLoader::findCalleeFunctionSamples(const CallInst &Inst) const {
316 const DILocation *DIL = Inst.getDebugLoc();
320 DISubprogram *SP = DIL->getScope()->getSubprogram();
324 Function *CalleeFunc = Inst.getCalledFunction();
329 StringRef CalleeName = CalleeFunc->getName();
330 const FunctionSamples *FS = findFunctionSamples(Inst);
334 return FS->findFunctionSamplesAt(
335 CallsiteLocation(getOffset(DIL->getLine(), SP->getLine()),
336 DIL->getDiscriminator(), CalleeName));
339 /// \brief Get the FunctionSamples for an instruction.
341 /// The FunctionSamples of an instruction \p Inst is the inlined instance
342 /// in which that instruction is coming from. We traverse the inline stack
343 /// of that instruction, and match it with the tree nodes in the profile.
345 /// \param Inst Instruction to query.
347 /// \returns the FunctionSamples pointer to the inlined instance.
348 const FunctionSamples *
349 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
350 SmallVector<CallsiteLocation, 10> S;
351 const DILocation *DIL = Inst.getDebugLoc();
355 StringRef CalleeName;
356 for (const DILocation *DIL = Inst.getDebugLoc(); DIL;
357 DIL = DIL->getInlinedAt()) {
358 DISubprogram *SP = DIL->getScope()->getSubprogram();
361 if (!CalleeName.empty()) {
362 S.push_back(CallsiteLocation(getOffset(DIL->getLine(), SP->getLine()),
363 DIL->getDiscriminator(), CalleeName));
365 CalleeName = SP->getLinkageName();
369 const FunctionSamples *FS = Samples;
370 for (int i = S.size() - 1; i >= 0 && FS != nullptr; i--) {
371 FS = FS->findFunctionSamplesAt(S[i]);
376 /// \brief Iteratively inline hot callsites of a function.
378 /// Iteratively traverse all callsites of the function \p F, and find if
379 /// the corresponding inlined instance exists and is hot in profile. If
380 /// it is hot enough, inline the callsites and adds new callsites of the
381 /// callee into the caller.
383 /// TODO: investigate the possibility of not invoking InlineFunction directly.
385 /// \param F function to perform iterative inlining.
387 /// \returns True if there is any inline happened.
388 bool SampleProfileLoader::inlineHotFunctions(Function &F) {
389 bool Changed = false;
390 LLVMContext &Ctx = F.getContext();
392 bool LocalChanged = false;
393 SmallVector<CallInst *, 10> CIS;
395 for (auto &I : BB.getInstList()) {
396 CallInst *CI = dyn_cast<CallInst>(&I);
398 const FunctionSamples *FS = findCalleeFunctionSamples(*CI);
399 if (FS && FS->getTotalSamples() > 0) {
405 for (auto CI : CIS) {
406 InlineFunctionInfo IFI;
407 Function *CalledFunction = CI->getCalledFunction();
408 DebugLoc DLoc = CI->getDebugLoc();
409 uint64_t NumSamples = findCalleeFunctionSamples(*CI)->getTotalSamples();
410 if (InlineFunction(CI, IFI)) {
412 emitOptimizationRemark(Ctx, DEBUG_TYPE, F, DLoc,
413 Twine("inlined hot callee '") +
414 CalledFunction->getName() + "' with " +
415 Twine(NumSamples) + " samples into '" +
428 /// \brief Find equivalence classes for the given block.
430 /// This finds all the blocks that are guaranteed to execute the same
431 /// number of times as \p BB1. To do this, it traverses all the
432 /// descendants of \p BB1 in the dominator or post-dominator tree.
434 /// A block BB2 will be in the same equivalence class as \p BB1 if
435 /// the following holds:
437 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
438 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
439 /// dominate BB1 in the post-dominator tree.
441 /// 2- Both BB2 and \p BB1 must be in the same loop.
443 /// For every block BB2 that meets those two requirements, we set BB2's
444 /// equivalence class to \p BB1.
446 /// \param BB1 Block to check.
447 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
448 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
449 /// with blocks from \p BB1's dominator tree, then
450 /// this is the post-dominator tree, and vice versa.
451 void SampleProfileLoader::findEquivalencesFor(
452 BasicBlock *BB1, SmallVector<BasicBlock *, 8> Descendants,
453 DominatorTreeBase<BasicBlock> *DomTree) {
454 const BasicBlock *EC = EquivalenceClass[BB1];
455 uint64_t Weight = BlockWeights[EC];
456 for (const auto *BB2 : Descendants) {
457 bool IsDomParent = DomTree->dominates(BB2, BB1);
458 bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
459 if (BB1 != BB2 && IsDomParent && IsInSameLoop) {
460 EquivalenceClass[BB2] = EC;
462 // If BB2 is heavier than BB1, make BB2 have the same weight
465 // Note that we don't worry about the opposite situation here
466 // (when BB2 is lighter than BB1). We will deal with this
467 // during the propagation phase. Right now, we just want to
468 // make sure that BB1 has the largest weight of all the
469 // members of its equivalence set.
470 Weight = std::max(Weight, BlockWeights[BB2]);
473 BlockWeights[EC] = Weight;
476 /// \brief Find equivalence classes.
478 /// Since samples may be missing from blocks, we can fill in the gaps by setting
479 /// the weights of all the blocks in the same equivalence class to the same
480 /// weight. To compute the concept of equivalence, we use dominance and loop
481 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
482 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
484 /// \param F The function to query.
485 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
486 SmallVector<BasicBlock *, 8> DominatedBBs;
487 DEBUG(dbgs() << "\nBlock equivalence classes\n");
488 // Find equivalence sets based on dominance and post-dominance information.
490 BasicBlock *BB1 = &BB;
492 // Compute BB1's equivalence class once.
493 if (EquivalenceClass.count(BB1)) {
494 DEBUG(printBlockEquivalence(dbgs(), BB1));
498 // By default, blocks are in their own equivalence class.
499 EquivalenceClass[BB1] = BB1;
501 // Traverse all the blocks dominated by BB1. We are looking for
502 // every basic block BB2 such that:
504 // 1- BB1 dominates BB2.
505 // 2- BB2 post-dominates BB1.
506 // 3- BB1 and BB2 are in the same loop nest.
508 // If all those conditions hold, it means that BB2 is executed
509 // as many times as BB1, so they are placed in the same equivalence
510 // class by making BB2's equivalence class be BB1.
511 DominatedBBs.clear();
512 DT->getDescendants(BB1, DominatedBBs);
513 findEquivalencesFor(BB1, DominatedBBs, PDT.get());
515 DEBUG(printBlockEquivalence(dbgs(), BB1));
518 // Assign weights to equivalence classes.
520 // All the basic blocks in the same equivalence class will execute
521 // the same number of times. Since we know that the head block in
522 // each equivalence class has the largest weight, assign that weight
523 // to all the blocks in that equivalence class.
524 DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n");
526 const BasicBlock *BB = &BI;
527 const BasicBlock *EquivBB = EquivalenceClass[BB];
529 BlockWeights[BB] = BlockWeights[EquivBB];
530 DEBUG(printBlockWeight(dbgs(), BB));
534 /// \brief Visit the given edge to decide if it has a valid weight.
536 /// If \p E has not been visited before, we copy to \p UnknownEdge
537 /// and increment the count of unknown edges.
539 /// \param E Edge to visit.
540 /// \param NumUnknownEdges Current number of unknown edges.
541 /// \param UnknownEdge Set if E has not been visited before.
543 /// \returns E's weight, if known. Otherwise, return 0.
544 uint64_t SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
546 if (!VisitedEdges.count(E)) {
547 (*NumUnknownEdges)++;
552 return EdgeWeights[E];
555 /// \brief Propagate weights through incoming/outgoing edges.
557 /// If the weight of a basic block is known, and there is only one edge
558 /// with an unknown weight, we can calculate the weight of that edge.
560 /// Similarly, if all the edges have a known count, we can calculate the
561 /// count of the basic block, if needed.
563 /// \param F Function to process.
565 /// \returns True if new weights were assigned to edges or blocks.
566 bool SampleProfileLoader::propagateThroughEdges(Function &F) {
567 bool Changed = false;
568 DEBUG(dbgs() << "\nPropagation through edges\n");
569 for (const auto &BI : F) {
570 const BasicBlock *BB = &BI;
571 const BasicBlock *EC = EquivalenceClass[BB];
573 // Visit all the predecessor and successor edges to determine
574 // which ones have a weight assigned already. Note that it doesn't
575 // matter that we only keep track of a single unknown edge. The
576 // only case we are interested in handling is when only a single
577 // edge is unknown (see setEdgeOrBlockWeight).
578 for (unsigned i = 0; i < 2; i++) {
579 uint64_t TotalWeight = 0;
580 unsigned NumUnknownEdges = 0;
581 Edge UnknownEdge, SelfReferentialEdge;
584 // First, visit all predecessor edges.
585 for (auto *Pred : Predecessors[BB]) {
586 Edge E = std::make_pair(Pred, BB);
587 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
588 if (E.first == E.second)
589 SelfReferentialEdge = E;
592 // On the second round, visit all successor edges.
593 for (auto *Succ : Successors[BB]) {
594 Edge E = std::make_pair(BB, Succ);
595 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
599 // After visiting all the edges, there are three cases that we
600 // can handle immediately:
602 // - All the edge weights are known (i.e., NumUnknownEdges == 0).
603 // In this case, we simply check that the sum of all the edges
604 // is the same as BB's weight. If not, we change BB's weight
605 // to match. Additionally, if BB had not been visited before,
606 // we mark it visited.
608 // - Only one edge is unknown and BB has already been visited.
609 // In this case, we can compute the weight of the edge by
610 // subtracting the total block weight from all the known
611 // edge weights. If the edges weight more than BB, then the
612 // edge of the last remaining edge is set to zero.
614 // - There exists a self-referential edge and the weight of BB is
615 // known. In this case, this edge can be based on BB's weight.
616 // We add up all the other known edges and set the weight on
617 // the self-referential edge as we did in the previous case.
619 // In any other case, we must continue iterating. Eventually,
620 // all edges will get a weight, or iteration will stop when
621 // it reaches SampleProfileMaxPropagateIterations.
622 if (NumUnknownEdges <= 1) {
623 uint64_t &BBWeight = BlockWeights[EC];
624 if (NumUnknownEdges == 0) {
625 // If we already know the weight of all edges, the weight of the
626 // basic block can be computed. It should be no larger than the sum
627 // of all edge weights.
628 if (TotalWeight > BBWeight) {
629 BBWeight = TotalWeight;
631 DEBUG(dbgs() << "All edge weights for " << BB->getName()
632 << " known. Set weight for block: ";
633 printBlockWeight(dbgs(), BB););
635 if (VisitedBlocks.insert(EC).second)
637 } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) {
638 // If there is a single unknown edge and the block has been
639 // visited, then we can compute E's weight.
640 if (BBWeight >= TotalWeight)
641 EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
643 EdgeWeights[UnknownEdge] = 0;
644 VisitedEdges.insert(UnknownEdge);
646 DEBUG(dbgs() << "Set weight for edge: ";
647 printEdgeWeight(dbgs(), UnknownEdge));
649 } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) {
650 uint64_t &BBWeight = BlockWeights[BB];
651 // We have a self-referential edge and the weight of BB is known.
652 if (BBWeight >= TotalWeight)
653 EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
655 EdgeWeights[SelfReferentialEdge] = 0;
656 VisitedEdges.insert(SelfReferentialEdge);
658 DEBUG(dbgs() << "Set self-referential edge weight to: ";
659 printEdgeWeight(dbgs(), SelfReferentialEdge));
667 /// \brief Build in/out edge lists for each basic block in the CFG.
669 /// We are interested in unique edges. If a block B1 has multiple
670 /// edges to another block B2, we only add a single B1->B2 edge.
671 void SampleProfileLoader::buildEdges(Function &F) {
673 BasicBlock *B1 = &BI;
675 // Add predecessors for B1.
676 SmallPtrSet<BasicBlock *, 16> Visited;
677 if (!Predecessors[B1].empty())
678 llvm_unreachable("Found a stale predecessors list in a basic block.");
679 for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
680 BasicBlock *B2 = *PI;
681 if (Visited.insert(B2).second)
682 Predecessors[B1].push_back(B2);
685 // Add successors for B1.
687 if (!Successors[B1].empty())
688 llvm_unreachable("Found a stale successors list in a basic block.");
689 for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
690 BasicBlock *B2 = *SI;
691 if (Visited.insert(B2).second)
692 Successors[B1].push_back(B2);
697 /// \brief Propagate weights into edges
699 /// The following rules are applied to every block BB in the CFG:
701 /// - If BB has a single predecessor/successor, then the weight
702 /// of that edge is the weight of the block.
704 /// - If all incoming or outgoing edges are known except one, and the
705 /// weight of the block is already known, the weight of the unknown
706 /// edge will be the weight of the block minus the sum of all the known
707 /// edges. If the sum of all the known edges is larger than BB's weight,
708 /// we set the unknown edge weight to zero.
710 /// - If there is a self-referential edge, and the weight of the block is
711 /// known, the weight for that edge is set to the weight of the block
712 /// minus the weight of the other incoming edges to that block (if
714 void SampleProfileLoader::propagateWeights(Function &F) {
718 // Add an entry count to the function using the samples gathered
719 // at the function entry.
720 F.setEntryCount(Samples->getHeadSamples());
722 // Before propagation starts, build, for each block, a list of
723 // unique predecessors and successors. This is necessary to handle
724 // identical edges in multiway branches. Since we visit all blocks and all
725 // edges of the CFG, it is cleaner to build these lists once at the start
729 // Propagate until we converge or we go past the iteration limit.
730 while (Changed && I++ < SampleProfileMaxPropagateIterations) {
731 Changed = propagateThroughEdges(F);
734 // Generate MD_prof metadata for every branch instruction using the
735 // edge weights computed during propagation.
736 DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
737 LLVMContext &Ctx = F.getContext();
740 BasicBlock *BB = &BI;
741 TerminatorInst *TI = BB->getTerminator();
742 if (TI->getNumSuccessors() == 1)
744 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
747 DEBUG(dbgs() << "\nGetting weights for branch at line "
748 << TI->getDebugLoc().getLine() << ".\n");
749 SmallVector<uint32_t, 4> Weights;
750 uint32_t MaxWeight = 0;
752 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
753 BasicBlock *Succ = TI->getSuccessor(I);
754 Edge E = std::make_pair(BB, Succ);
755 uint64_t Weight = EdgeWeights[E];
756 DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
757 // Use uint32_t saturated arithmetic to adjust the incoming weights,
758 // if needed. Sample counts in profiles are 64-bit unsigned values,
759 // but internally branch weights are expressed as 32-bit values.
760 if (Weight > std::numeric_limits<uint32_t>::max()) {
761 DEBUG(dbgs() << " (saturated due to uint32_t overflow)");
762 Weight = std::numeric_limits<uint32_t>::max();
764 Weights.push_back(static_cast<uint32_t>(Weight));
766 if (Weight > MaxWeight) {
768 MaxDestLoc = Succ->getFirstNonPHIOrDbgOrLifetime()->getDebugLoc();
773 // Only set weights if there is at least one non-zero weight.
774 // In any other case, let the analyzer set weights.
776 DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
777 TI->setMetadata(llvm::LLVMContext::MD_prof,
778 MDB.createBranchWeights(Weights));
779 DebugLoc BranchLoc = TI->getDebugLoc();
780 emitOptimizationRemark(
781 Ctx, DEBUG_TYPE, F, MaxDestLoc,
782 Twine("most popular destination for conditional branches at ") +
783 ((BranchLoc) ? Twine(BranchLoc->getFilename() + ":" +
784 Twine(BranchLoc.getLine()) + ":" +
785 Twine(BranchLoc.getCol()))
786 : Twine("<UNKNOWN LOCATION>")));
788 DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
793 /// \brief Get the line number for the function header.
795 /// This looks up function \p F in the current compilation unit and
796 /// retrieves the line number where the function is defined. This is
797 /// line 0 for all the samples read from the profile file. Every line
798 /// number is relative to this line.
800 /// \param F Function object to query.
802 /// \returns the line number where \p F is defined. If it returns 0,
803 /// it means that there is no debug information available for \p F.
804 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
805 if (DISubprogram *S = getDISubprogram(&F))
808 // If the start of \p F is missing, emit a diagnostic to inform the user
809 // about the missed opportunity.
810 F.getContext().diagnose(DiagnosticInfoSampleProfile(
811 "No debug information found in function " + F.getName() +
812 ": Function profile not used",
817 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
818 DT.reset(new DominatorTree);
821 PDT.reset(new DominatorTreeBase<BasicBlock>(true));
824 LI.reset(new LoopInfo);
828 /// \brief Generate branch weight metadata for all branches in \p F.
830 /// Branch weights are computed out of instruction samples using a
831 /// propagation heuristic. Propagation proceeds in 3 phases:
833 /// 1- Assignment of block weights. All the basic blocks in the function
834 /// are initial assigned the same weight as their most frequently
835 /// executed instruction.
837 /// 2- Creation of equivalence classes. Since samples may be missing from
838 /// blocks, we can fill in the gaps by setting the weights of all the
839 /// blocks in the same equivalence class to the same weight. To compute
840 /// the concept of equivalence, we use dominance and loop information.
841 /// Two blocks B1 and B2 are in the same equivalence class if B1
842 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
844 /// 3- Propagation of block weights into edges. This uses a simple
845 /// propagation heuristic. The following rules are applied to every
846 /// block BB in the CFG:
848 /// - If BB has a single predecessor/successor, then the weight
849 /// of that edge is the weight of the block.
851 /// - If all the edges are known except one, and the weight of the
852 /// block is already known, the weight of the unknown edge will
853 /// be the weight of the block minus the sum of all the known
854 /// edges. If the sum of all the known edges is larger than BB's weight,
855 /// we set the unknown edge weight to zero.
857 /// - If there is a self-referential edge, and the weight of the block is
858 /// known, the weight for that edge is set to the weight of the block
859 /// minus the weight of the other incoming edges to that block (if
862 /// Since this propagation is not guaranteed to finalize for every CFG, we
863 /// only allow it to proceed for a limited number of iterations (controlled
864 /// by -sample-profile-max-propagate-iterations).
866 /// FIXME: Try to replace this propagation heuristic with a scheme
867 /// that is guaranteed to finalize. A work-list approach similar to
868 /// the standard value propagation algorithm used by SSA-CCP might
871 /// Once all the branch weights are computed, we emit the MD_prof
872 /// metadata on BB using the computed values for each of its branches.
874 /// \param F The function to query.
876 /// \returns true if \p F was modified. Returns false, otherwise.
877 bool SampleProfileLoader::emitAnnotations(Function &F) {
878 bool Changed = false;
880 if (getFunctionLoc(F) == 0)
883 DEBUG(dbgs() << "Line number for the first instruction in " << F.getName()
884 << ": " << getFunctionLoc(F) << "\n");
886 Changed |= inlineHotFunctions(F);
888 // Compute basic block weights.
889 Changed |= computeBlockWeights(F);
892 // Compute dominance and loop info needed for propagation.
893 computeDominanceAndLoopInfo(F);
895 // Find equivalence classes.
896 findEquivalenceClasses(F);
898 // Propagate weights to all edges.
905 char SampleProfileLoader::ID = 0;
906 INITIALIZE_PASS_BEGIN(SampleProfileLoader, "sample-profile",
907 "Sample Profile loader", false, false)
908 INITIALIZE_PASS_DEPENDENCY(AddDiscriminators)
909 INITIALIZE_PASS_END(SampleProfileLoader, "sample-profile",
910 "Sample Profile loader", false, false)
912 bool SampleProfileLoader::doInitialization(Module &M) {
913 auto &Ctx = M.getContext();
914 auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
915 if (std::error_code EC = ReaderOrErr.getError()) {
916 std::string Msg = "Could not open profile: " + EC.message();
917 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename.data(), Msg));
920 Reader = std::move(ReaderOrErr.get());
921 ProfileIsValid = (Reader->read() == sampleprof_error::success);
925 ModulePass *llvm::createSampleProfileLoaderPass() {
926 return new SampleProfileLoader(SampleProfileFile);
929 ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) {
930 return new SampleProfileLoader(Name);
933 bool SampleProfileLoader::runOnModule(Module &M) {
936 if (!F.isDeclaration())
937 retval |= runOnFunction(F);
941 bool SampleProfileLoader::runOnFunction(Function &F) {
945 Samples = Reader->getSamplesFor(F);
946 if (!Samples->empty())
947 return emitAnnotations(F);