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 *, unsigned> BlockWeightMap;
69 typedef DenseMap<const BasicBlock *, const BasicBlock *> EquivalenceClassMap;
70 typedef std::pair<const BasicBlock *, const BasicBlock *> Edge;
71 typedef DenseMap<Edge, unsigned> 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<unsigned> getInstWeight(const Instruction &I) const;
108 ErrorOr<unsigned> 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 unsigned visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
122 void buildEdges(Function &F);
123 bool propagateThroughEdges(Function &F);
124 void computeDominanceAndLoopInfo(Function &F);
126 /// \brief Map basic blocks to their computed weights.
128 /// The weight of a basic block is defined to be the maximum
129 /// of all the instruction weights in that block.
130 BlockWeightMap BlockWeights;
132 /// \brief Map edges to their computed weights.
134 /// Edge weights are computed by propagating basic block weights in
135 /// SampleProfile::propagateWeights.
136 EdgeWeightMap EdgeWeights;
138 /// \brief Set of visited blocks during propagation.
139 SmallPtrSet<const BasicBlock *, 128> VisitedBlocks;
141 /// \brief Set of visited edges during propagation.
142 SmallSet<Edge, 128> VisitedEdges;
144 /// \brief Equivalence classes for block weights.
146 /// Two blocks BB1 and BB2 are in the same equivalence class if they
147 /// dominate and post-dominate each other, and they are in the same loop
148 /// nest. When this happens, the two blocks are guaranteed to execute
149 /// the same number of times.
150 EquivalenceClassMap EquivalenceClass;
152 /// \brief Dominance, post-dominance and loop information.
153 std::unique_ptr<DominatorTree> DT;
154 std::unique_ptr<DominatorTreeBase<BasicBlock>> PDT;
155 std::unique_ptr<LoopInfo> LI;
157 /// \brief Predecessors for each basic block in the CFG.
158 BlockEdgeMap Predecessors;
160 /// \brief Successors for each basic block in the CFG.
161 BlockEdgeMap Successors;
163 /// \brief Profile reader object.
164 std::unique_ptr<SampleProfileReader> Reader;
166 /// \brief Samples collected for the body of this function.
167 FunctionSamples *Samples;
169 /// \brief Name of the profile file to load.
172 /// \brief Flag indicating whether the profile input loaded successfully.
177 /// \brief Print the weight of edge \p E on stream \p OS.
179 /// \param OS Stream to emit the output to.
180 /// \param E Edge to print.
181 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
182 OS << "weight[" << E.first->getName() << "->" << E.second->getName()
183 << "]: " << EdgeWeights[E] << "\n";
186 /// \brief Print the equivalence class of block \p BB on stream \p OS.
188 /// \param OS Stream to emit the output to.
189 /// \param BB Block to print.
190 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
191 const BasicBlock *BB) {
192 const BasicBlock *Equiv = EquivalenceClass[BB];
193 OS << "equivalence[" << BB->getName()
194 << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
197 /// \brief Print the weight of block \p BB on stream \p OS.
199 /// \param OS Stream to emit the output to.
200 /// \param BB Block to print.
201 void SampleProfileLoader::printBlockWeight(raw_ostream &OS,
202 const BasicBlock *BB) const {
203 const auto &I = BlockWeights.find(BB);
204 unsigned W = (I == BlockWeights.end() ? 0 : I->second);
205 OS << "weight[" << BB->getName() << "]: " << W << "\n";
208 /// \brief Get the weight for an instruction.
210 /// The "weight" of an instruction \p Inst is the number of samples
211 /// collected on that instruction at runtime. To retrieve it, we
212 /// need to compute the line number of \p Inst relative to the start of its
213 /// function. We use HeaderLineno to compute the offset. We then
214 /// look up the samples collected for \p Inst using BodySamples.
216 /// \param Inst Instruction to query.
218 /// \returns the weight of \p Inst.
220 SampleProfileLoader::getInstWeight(const Instruction &Inst) const {
221 DebugLoc DLoc = Inst.getDebugLoc();
223 return std::error_code();
225 const FunctionSamples *FS = findFunctionSamples(Inst);
227 return std::error_code();
229 const DILocation *DIL = DLoc;
230 unsigned Lineno = DLoc.getLine();
231 unsigned HeaderLineno = DIL->getScope()->getSubprogram()->getLine();
232 if (Lineno < HeaderLineno)
233 return std::error_code();
235 ErrorOr<unsigned> R =
236 FS->findSamplesAt(Lineno - HeaderLineno, DIL->getDiscriminator());
238 DEBUG(dbgs() << " " << Lineno << "." << DIL->getDiscriminator() << ":"
239 << Inst << " (line offset: " << Lineno - HeaderLineno << "."
240 << DIL->getDiscriminator() << " - weight: " << R.get()
245 /// \brief Compute the weight of a basic block.
247 /// The weight of basic block \p BB is the maximum weight of all the
248 /// instructions in BB.
250 /// \param BB The basic block to query.
252 /// \returns the weight for \p BB.
254 SampleProfileLoader::getBlockWeight(const BasicBlock *BB) const {
257 for (auto &I : BB->getInstList()) {
258 const ErrorOr<unsigned> &R = getInstWeight(I);
259 if (R && R.get() >= Weight) {
267 return std::error_code();
270 /// \brief Compute and store the weights of every basic block.
272 /// This populates the BlockWeights map by computing
273 /// the weights of every basic block in the CFG.
275 /// \param F The function to query.
276 bool SampleProfileLoader::computeBlockWeights(Function &F) {
277 bool Changed = false;
278 DEBUG(dbgs() << "Block weights\n");
279 for (const auto &BB : F) {
280 ErrorOr<unsigned> Weight = getBlockWeight(&BB);
282 BlockWeights[&BB] = Weight.get();
283 VisitedBlocks.insert(&BB);
286 DEBUG(printBlockWeight(dbgs(), &BB));
292 /// \brief Get the FunctionSamples for a call instruction.
294 /// The FunctionSamples of a call instruction \p Inst is the inlined
295 /// instance in which that call instruction is calling to. It contains
296 /// all samples that resides in the inlined instance. We first find the
297 /// inlined instance in which the call instruction is from, then we
298 /// traverse its children to find the callsite with the matching
299 /// location and callee function name.
301 /// \param Inst Call instruction to query.
303 /// \returns The FunctionSamples pointer to the inlined instance.
304 const FunctionSamples *
305 SampleProfileLoader::findCalleeFunctionSamples(const CallInst &Inst) const {
306 const DILocation *DIL = Inst.getDebugLoc();
310 DISubprogram *SP = DIL->getScope()->getSubprogram();
311 if (!SP || DIL->getLine() < SP->getLine())
314 Function *CalleeFunc = Inst.getCalledFunction();
319 StringRef CalleeName = CalleeFunc->getName();
320 const FunctionSamples *FS = findFunctionSamples(Inst);
324 return FS->findFunctionSamplesAt(CallsiteLocation(
325 DIL->getLine() - SP->getLine(), DIL->getDiscriminator(), CalleeName));
328 /// \brief Get the FunctionSamples for an instruction.
330 /// The FunctionSamples of an instruction \p Inst is the inlined instance
331 /// in which that instruction is coming from. We traverse the inline stack
332 /// of that instruction, and match it with the tree nodes in the profile.
334 /// \param Inst Instruction to query.
336 /// \returns the FunctionSamples pointer to the inlined instance.
337 const FunctionSamples *
338 SampleProfileLoader::findFunctionSamples(const Instruction &Inst) const {
339 SmallVector<CallsiteLocation, 10> S;
340 const DILocation *DIL = Inst.getDebugLoc();
344 StringRef CalleeName;
345 for (const DILocation *DIL = Inst.getDebugLoc(); DIL;
346 DIL = DIL->getInlinedAt()) {
347 DISubprogram *SP = DIL->getScope()->getSubprogram();
348 if (!SP || DIL->getLine() < SP->getLine())
350 if (!CalleeName.empty()) {
351 S.push_back(CallsiteLocation(DIL->getLine() - SP->getLine(),
352 DIL->getDiscriminator(), CalleeName));
354 CalleeName = SP->getLinkageName();
358 const FunctionSamples *FS = Samples;
359 for (int i = S.size() - 1; i >= 0 && FS != nullptr; i--) {
360 FS = FS->findFunctionSamplesAt(S[i]);
365 /// \brief Iteratively inline hot callsites of a function.
367 /// Iteratively traverse all callsites of the function \p F, and find if
368 /// the corresponding inlined instance exists and is hot in profile. If
369 /// it is hot enough, inline the callsites and adds new callsites of the
370 /// callee into the caller.
372 /// TODO: investigate the possibility of not invoking InlineFunction directly.
374 /// \param F function to perform iterative inlining.
376 /// \returns True if there is any inline happened.
377 bool SampleProfileLoader::inlineHotFunctions(Function &F) {
378 bool Changed = false;
380 bool LocalChanged = false;
381 SmallVector<CallInst *, 10> CIS;
383 for (auto &I : BB.getInstList()) {
384 CallInst *CI = dyn_cast<CallInst>(&I);
386 const FunctionSamples *FS = findCalleeFunctionSamples(*CI);
387 if (FS && FS->getTotalSamples() > 0) {
393 for (auto CI : CIS) {
394 InlineFunctionInfo IFI;
395 if (InlineFunction(CI, IFI))
407 /// \brief Find equivalence classes for the given block.
409 /// This finds all the blocks that are guaranteed to execute the same
410 /// number of times as \p BB1. To do this, it traverses all the
411 /// descendants of \p BB1 in the dominator or post-dominator tree.
413 /// A block BB2 will be in the same equivalence class as \p BB1 if
414 /// the following holds:
416 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
417 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
418 /// dominate BB1 in the post-dominator tree.
420 /// 2- Both BB2 and \p BB1 must be in the same loop.
422 /// For every block BB2 that meets those two requirements, we set BB2's
423 /// equivalence class to \p BB1.
425 /// \param BB1 Block to check.
426 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
427 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
428 /// with blocks from \p BB1's dominator tree, then
429 /// this is the post-dominator tree, and vice versa.
430 void SampleProfileLoader::findEquivalencesFor(
431 BasicBlock *BB1, SmallVector<BasicBlock *, 8> Descendants,
432 DominatorTreeBase<BasicBlock> *DomTree) {
433 const BasicBlock *EC = EquivalenceClass[BB1];
434 unsigned Weight = BlockWeights[EC];
435 for (const auto *BB2 : Descendants) {
436 bool IsDomParent = DomTree->dominates(BB2, BB1);
437 bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
438 if (BB1 != BB2 && IsDomParent && IsInSameLoop) {
439 EquivalenceClass[BB2] = EC;
441 // If BB2 is heavier than BB1, make BB2 have the same weight
444 // Note that we don't worry about the opposite situation here
445 // (when BB2 is lighter than BB1). We will deal with this
446 // during the propagation phase. Right now, we just want to
447 // make sure that BB1 has the largest weight of all the
448 // members of its equivalence set.
449 Weight = std::max(Weight, BlockWeights[BB2]);
452 BlockWeights[EC] = Weight;
455 /// \brief Find equivalence classes.
457 /// Since samples may be missing from blocks, we can fill in the gaps by setting
458 /// the weights of all the blocks in the same equivalence class to the same
459 /// weight. To compute the concept of equivalence, we use dominance and loop
460 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
461 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
463 /// \param F The function to query.
464 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
465 SmallVector<BasicBlock *, 8> DominatedBBs;
466 DEBUG(dbgs() << "\nBlock equivalence classes\n");
467 // Find equivalence sets based on dominance and post-dominance information.
469 BasicBlock *BB1 = &BB;
471 // Compute BB1's equivalence class once.
472 if (EquivalenceClass.count(BB1)) {
473 DEBUG(printBlockEquivalence(dbgs(), BB1));
477 // By default, blocks are in their own equivalence class.
478 EquivalenceClass[BB1] = BB1;
480 // Traverse all the blocks dominated by BB1. We are looking for
481 // every basic block BB2 such that:
483 // 1- BB1 dominates BB2.
484 // 2- BB2 post-dominates BB1.
485 // 3- BB1 and BB2 are in the same loop nest.
487 // If all those conditions hold, it means that BB2 is executed
488 // as many times as BB1, so they are placed in the same equivalence
489 // class by making BB2's equivalence class be BB1.
490 DominatedBBs.clear();
491 DT->getDescendants(BB1, DominatedBBs);
492 findEquivalencesFor(BB1, DominatedBBs, PDT.get());
494 DEBUG(printBlockEquivalence(dbgs(), BB1));
497 // Assign weights to equivalence classes.
499 // All the basic blocks in the same equivalence class will execute
500 // the same number of times. Since we know that the head block in
501 // each equivalence class has the largest weight, assign that weight
502 // to all the blocks in that equivalence class.
503 DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n");
505 const BasicBlock *BB = &BI;
506 const BasicBlock *EquivBB = EquivalenceClass[BB];
508 BlockWeights[BB] = BlockWeights[EquivBB];
509 DEBUG(printBlockWeight(dbgs(), BB));
513 /// \brief Visit the given edge to decide if it has a valid weight.
515 /// If \p E has not been visited before, we copy to \p UnknownEdge
516 /// and increment the count of unknown edges.
518 /// \param E Edge to visit.
519 /// \param NumUnknownEdges Current number of unknown edges.
520 /// \param UnknownEdge Set if E has not been visited before.
522 /// \returns E's weight, if known. Otherwise, return 0.
523 unsigned SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
525 if (!VisitedEdges.count(E)) {
526 (*NumUnknownEdges)++;
531 return EdgeWeights[E];
534 /// \brief Propagate weights through incoming/outgoing edges.
536 /// If the weight of a basic block is known, and there is only one edge
537 /// with an unknown weight, we can calculate the weight of that edge.
539 /// Similarly, if all the edges have a known count, we can calculate the
540 /// count of the basic block, if needed.
542 /// \param F Function to process.
544 /// \returns True if new weights were assigned to edges or blocks.
545 bool SampleProfileLoader::propagateThroughEdges(Function &F) {
546 bool Changed = false;
547 DEBUG(dbgs() << "\nPropagation through edges\n");
548 for (const auto &BI : F) {
549 const BasicBlock *BB = &BI;
550 const BasicBlock *EC = EquivalenceClass[BB];
552 // Visit all the predecessor and successor edges to determine
553 // which ones have a weight assigned already. Note that it doesn't
554 // matter that we only keep track of a single unknown edge. The
555 // only case we are interested in handling is when only a single
556 // edge is unknown (see setEdgeOrBlockWeight).
557 for (unsigned i = 0; i < 2; i++) {
558 unsigned TotalWeight = 0;
559 unsigned NumUnknownEdges = 0;
560 Edge UnknownEdge, SelfReferentialEdge;
563 // First, visit all predecessor edges.
564 for (auto *Pred : Predecessors[BB]) {
565 Edge E = std::make_pair(Pred, BB);
566 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
567 if (E.first == E.second)
568 SelfReferentialEdge = E;
571 // On the second round, visit all successor edges.
572 for (auto *Succ : Successors[BB]) {
573 Edge E = std::make_pair(BB, Succ);
574 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
578 // After visiting all the edges, there are three cases that we
579 // can handle immediately:
581 // - All the edge weights are known (i.e., NumUnknownEdges == 0).
582 // In this case, we simply check that the sum of all the edges
583 // is the same as BB's weight. If not, we change BB's weight
584 // to match. Additionally, if BB had not been visited before,
585 // we mark it visited.
587 // - Only one edge is unknown and BB has already been visited.
588 // In this case, we can compute the weight of the edge by
589 // subtracting the total block weight from all the known
590 // edge weights. If the edges weight more than BB, then the
591 // edge of the last remaining edge is set to zero.
593 // - There exists a self-referential edge and the weight of BB is
594 // known. In this case, this edge can be based on BB's weight.
595 // We add up all the other known edges and set the weight on
596 // the self-referential edge as we did in the previous case.
598 // In any other case, we must continue iterating. Eventually,
599 // all edges will get a weight, or iteration will stop when
600 // it reaches SampleProfileMaxPropagateIterations.
601 if (NumUnknownEdges <= 1) {
602 unsigned &BBWeight = BlockWeights[EC];
603 if (NumUnknownEdges == 0) {
604 // If we already know the weight of all edges, the weight of the
605 // basic block can be computed. It should be no larger than the sum
606 // of all edge weights.
607 if (TotalWeight > BBWeight) {
608 BBWeight = TotalWeight;
610 DEBUG(dbgs() << "All edge weights for " << BB->getName()
611 << " known. Set weight for block: ";
612 printBlockWeight(dbgs(), BB););
614 if (VisitedBlocks.insert(EC).second)
616 } else if (NumUnknownEdges == 1 && VisitedBlocks.count(EC)) {
617 // If there is a single unknown edge and the block has been
618 // visited, then we can compute E's weight.
619 if (BBWeight >= TotalWeight)
620 EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
622 EdgeWeights[UnknownEdge] = 0;
623 VisitedEdges.insert(UnknownEdge);
625 DEBUG(dbgs() << "Set weight for edge: ";
626 printEdgeWeight(dbgs(), UnknownEdge));
628 } else if (SelfReferentialEdge.first && VisitedBlocks.count(EC)) {
629 unsigned &BBWeight = BlockWeights[BB];
630 // We have a self-referential edge and the weight of BB is known.
631 if (BBWeight >= TotalWeight)
632 EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
634 EdgeWeights[SelfReferentialEdge] = 0;
635 VisitedEdges.insert(SelfReferentialEdge);
637 DEBUG(dbgs() << "Set self-referential edge weight to: ";
638 printEdgeWeight(dbgs(), SelfReferentialEdge));
646 /// \brief Build in/out edge lists for each basic block in the CFG.
648 /// We are interested in unique edges. If a block B1 has multiple
649 /// edges to another block B2, we only add a single B1->B2 edge.
650 void SampleProfileLoader::buildEdges(Function &F) {
652 BasicBlock *B1 = &BI;
654 // Add predecessors for B1.
655 SmallPtrSet<BasicBlock *, 16> Visited;
656 if (!Predecessors[B1].empty())
657 llvm_unreachable("Found a stale predecessors list in a basic block.");
658 for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
659 BasicBlock *B2 = *PI;
660 if (Visited.insert(B2).second)
661 Predecessors[B1].push_back(B2);
664 // Add successors for B1.
666 if (!Successors[B1].empty())
667 llvm_unreachable("Found a stale successors list in a basic block.");
668 for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
669 BasicBlock *B2 = *SI;
670 if (Visited.insert(B2).second)
671 Successors[B1].push_back(B2);
676 /// \brief Propagate weights into edges
678 /// The following rules are applied to every block BB in the CFG:
680 /// - If BB has a single predecessor/successor, then the weight
681 /// of that edge is the weight of the block.
683 /// - If all incoming or outgoing edges are known except one, and the
684 /// weight of the block is already known, the weight of the unknown
685 /// edge will be the weight of the block minus the sum of all the known
686 /// edges. If the sum of all the known edges is larger than BB's weight,
687 /// we set the unknown edge weight to zero.
689 /// - If there is a self-referential edge, and the weight of the block is
690 /// known, the weight for that edge is set to the weight of the block
691 /// minus the weight of the other incoming edges to that block (if
693 void SampleProfileLoader::propagateWeights(Function &F) {
697 // Add an entry count to the function using the samples gathered
698 // at the function entry.
699 F.setEntryCount(Samples->getHeadSamples());
701 // Before propagation starts, build, for each block, a list of
702 // unique predecessors and successors. This is necessary to handle
703 // identical edges in multiway branches. Since we visit all blocks and all
704 // edges of the CFG, it is cleaner to build these lists once at the start
708 // Propagate until we converge or we go past the iteration limit.
709 while (Changed && i++ < SampleProfileMaxPropagateIterations) {
710 Changed = propagateThroughEdges(F);
713 // Generate MD_prof metadata for every branch instruction using the
714 // edge weights computed during propagation.
715 DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
716 MDBuilder MDB(F.getContext());
718 BasicBlock *BB = &BI;
719 TerminatorInst *TI = BB->getTerminator();
720 if (TI->getNumSuccessors() == 1)
722 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
725 DEBUG(dbgs() << "\nGetting weights for branch at line "
726 << TI->getDebugLoc().getLine() << ".\n");
727 SmallVector<unsigned, 4> Weights;
728 bool AllWeightsZero = true;
729 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
730 BasicBlock *Succ = TI->getSuccessor(I);
731 Edge E = std::make_pair(BB, Succ);
732 unsigned Weight = EdgeWeights[E];
733 DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
734 Weights.push_back(Weight);
736 AllWeightsZero = false;
739 // Only set weights if there is at least one non-zero weight.
740 // In any other case, let the analyzer set weights.
741 if (!AllWeightsZero) {
742 DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
743 TI->setMetadata(llvm::LLVMContext::MD_prof,
744 MDB.createBranchWeights(Weights));
746 DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
751 /// \brief Get the line number for the function header.
753 /// This looks up function \p F in the current compilation unit and
754 /// retrieves the line number where the function is defined. This is
755 /// line 0 for all the samples read from the profile file. Every line
756 /// number is relative to this line.
758 /// \param F Function object to query.
760 /// \returns the line number where \p F is defined. If it returns 0,
761 /// it means that there is no debug information available for \p F.
762 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
763 if (DISubprogram *S = getDISubprogram(&F))
766 // If could not find the start of \p F, emit a diagnostic to inform the user
767 // about the missed opportunity.
768 F.getContext().diagnose(DiagnosticInfoSampleProfile(
769 "No debug information found in function " + F.getName() +
770 ": Function profile not used",
775 void SampleProfileLoader::computeDominanceAndLoopInfo(Function &F) {
776 DT.reset(new DominatorTree);
779 PDT.reset(new DominatorTreeBase<BasicBlock>(true));
782 LI.reset(new LoopInfo);
786 /// \brief Generate branch weight metadata for all branches in \p F.
788 /// Branch weights are computed out of instruction samples using a
789 /// propagation heuristic. Propagation proceeds in 3 phases:
791 /// 1- Assignment of block weights. All the basic blocks in the function
792 /// are initial assigned the same weight as their most frequently
793 /// executed instruction.
795 /// 2- Creation of equivalence classes. Since samples may be missing from
796 /// blocks, we can fill in the gaps by setting the weights of all the
797 /// blocks in the same equivalence class to the same weight. To compute
798 /// the concept of equivalence, we use dominance and loop information.
799 /// Two blocks B1 and B2 are in the same equivalence class if B1
800 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
802 /// 3- Propagation of block weights into edges. This uses a simple
803 /// propagation heuristic. The following rules are applied to every
804 /// block BB in the CFG:
806 /// - If BB has a single predecessor/successor, then the weight
807 /// of that edge is the weight of the block.
809 /// - If all the edges are known except one, and the weight of the
810 /// block is already known, the weight of the unknown edge will
811 /// be the weight of the block minus the sum of all the known
812 /// edges. If the sum of all the known edges is larger than BB's weight,
813 /// we set the unknown edge weight to zero.
815 /// - If there is a self-referential edge, and the weight of the block is
816 /// known, the weight for that edge is set to the weight of the block
817 /// minus the weight of the other incoming edges to that block (if
820 /// Since this propagation is not guaranteed to finalize for every CFG, we
821 /// only allow it to proceed for a limited number of iterations (controlled
822 /// by -sample-profile-max-propagate-iterations).
824 /// FIXME: Try to replace this propagation heuristic with a scheme
825 /// that is guaranteed to finalize. A work-list approach similar to
826 /// the standard value propagation algorithm used by SSA-CCP might
829 /// Once all the branch weights are computed, we emit the MD_prof
830 /// metadata on BB using the computed values for each of its branches.
832 /// \param F The function to query.
834 /// \returns true if \p F was modified. Returns false, otherwise.
835 bool SampleProfileLoader::emitAnnotations(Function &F) {
836 bool Changed = false;
838 if (getFunctionLoc(F) == 0)
841 DEBUG(dbgs() << "Line number for the first instruction in " << F.getName()
842 << ": " << getFunctionLoc(F) << "\n");
844 Changed |= inlineHotFunctions(F);
846 // Compute basic block weights.
847 Changed |= computeBlockWeights(F);
850 // Compute dominance and loop info needed for propagation.
851 computeDominanceAndLoopInfo(F);
853 // Find equivalence classes.
854 findEquivalenceClasses(F);
856 // Propagate weights to all edges.
863 char SampleProfileLoader::ID = 0;
864 INITIALIZE_PASS_BEGIN(SampleProfileLoader, "sample-profile",
865 "Sample Profile loader", false, false)
866 INITIALIZE_PASS_DEPENDENCY(AddDiscriminators)
867 INITIALIZE_PASS_END(SampleProfileLoader, "sample-profile",
868 "Sample Profile loader", false, false)
870 bool SampleProfileLoader::doInitialization(Module &M) {
871 auto &Ctx = M.getContext();
872 auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
873 if (std::error_code EC = ReaderOrErr.getError()) {
874 std::string Msg = "Could not open profile: " + EC.message();
875 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename.data(), Msg));
878 Reader = std::move(ReaderOrErr.get());
879 ProfileIsValid = (Reader->read() == sampleprof_error::success);
883 ModulePass *llvm::createSampleProfileLoaderPass() {
884 return new SampleProfileLoader(SampleProfileFile);
887 ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) {
888 return new SampleProfileLoader(Name);
891 bool SampleProfileLoader::runOnModule(Module &M) {
894 if (!F.isDeclaration())
895 retval |= runOnFunction(F);
899 bool SampleProfileLoader::runOnFunction(Function &F) {
903 Samples = Reader->getSamplesFor(F);
904 if (!Samples->empty())
905 return emitAnnotations(F);