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/raw_ostream.h"
47 #include "llvm/Transforms/IPO.h"
51 using namespace sampleprof;
53 #define DEBUG_TYPE "sample-profile"
55 // Command line option to specify the file to read samples from. This is
56 // mainly used for debugging.
57 static cl::opt<std::string> SampleProfileFile(
58 "sample-profile-file", cl::init(""), cl::value_desc("filename"),
59 cl::desc("Profile file loaded by -sample-profile"), cl::Hidden);
60 static cl::opt<unsigned> SampleProfileMaxPropagateIterations(
61 "sample-profile-max-propagate-iterations", cl::init(100),
62 cl::desc("Maximum number of iterations to go through when propagating "
63 "sample block/edge weights through the CFG."));
66 typedef DenseMap<BasicBlock *, unsigned> BlockWeightMap;
67 typedef DenseMap<BasicBlock *, BasicBlock *> EquivalenceClassMap;
68 typedef std::pair<BasicBlock *, BasicBlock *> Edge;
69 typedef DenseMap<Edge, unsigned> EdgeWeightMap;
70 typedef DenseMap<BasicBlock *, SmallVector<BasicBlock *, 8>> BlockEdgeMap;
72 /// \brief Sample profile pass.
74 /// This pass reads profile data from the file specified by
75 /// -sample-profile-file and annotates every affected function with the
76 /// profile information found in that file.
77 class SampleProfileLoader : public ModulePass {
79 // Class identification, replacement for typeinfo
82 SampleProfileLoader(StringRef Name = SampleProfileFile)
83 : ModulePass(ID), DT(nullptr), PDT(nullptr), LI(nullptr),
84 Ctx(nullptr), Reader(), Samples(nullptr), Filename(Name),
85 ProfileIsValid(false) {
86 initializeSampleProfileLoaderPass(*PassRegistry::getPassRegistry());
89 bool doInitialization(Module &M) override;
91 void dump() { Reader->dump(); }
93 const char *getPassName() const override { return "Sample profile pass"; }
95 bool runOnModule(Module &M) override;
97 void getAnalysisUsage(AnalysisUsage &AU) const override {
100 AU.addRequired<LoopInfoWrapperPass>();
101 AU.addPreserved<LoopInfoWrapperPass>();
103 AU.addRequired<DominatorTreeWrapperPass>();
104 AU.addPreserved<DominatorTreeWrapperPass>();
106 AU.addRequired<PostDominatorTree>();
107 AU.addPreserved<PostDominatorTree>();
111 bool runOnFunction(Function &F);
112 unsigned getFunctionLoc(Function &F);
113 bool emitAnnotations(Function &F);
114 unsigned getInstWeight(Instruction &I);
115 unsigned getBlockWeight(BasicBlock *BB);
116 void printEdgeWeight(raw_ostream &OS, Edge E);
117 void printBlockWeight(raw_ostream &OS, BasicBlock *BB);
118 void printBlockEquivalence(raw_ostream &OS, BasicBlock *BB);
119 bool computeBlockWeights(Function &F);
120 void findEquivalenceClasses(Function &F);
121 void findEquivalencesFor(BasicBlock *BB1,
122 SmallVector<BasicBlock *, 8> Descendants,
123 DominatorTreeBase<BasicBlock> *DomTree);
124 void propagateWeights(Function &F);
125 unsigned visitEdge(Edge E, unsigned *NumUnknownEdges, Edge *UnknownEdge);
126 void buildEdges(Function &F);
127 bool propagateThroughEdges(Function &F);
129 /// \brief Line number for the function header. Used to compute absolute
130 /// line numbers from the relative line numbers found in the profile.
131 unsigned HeaderLineno;
133 /// \brief Map basic blocks to their computed weights.
135 /// The weight of a basic block is defined to be the maximum
136 /// of all the instruction weights in that block.
137 BlockWeightMap BlockWeights;
139 /// \brief Map edges to their computed weights.
141 /// Edge weights are computed by propagating basic block weights in
142 /// SampleProfile::propagateWeights.
143 EdgeWeightMap EdgeWeights;
145 /// \brief Set of visited blocks during propagation.
146 SmallPtrSet<BasicBlock *, 128> VisitedBlocks;
148 /// \brief Set of visited edges during propagation.
149 SmallSet<Edge, 128> VisitedEdges;
151 /// \brief Equivalence classes for block weights.
153 /// Two blocks BB1 and BB2 are in the same equivalence class if they
154 /// dominate and post-dominate each other, and they are in the same loop
155 /// nest. When this happens, the two blocks are guaranteed to execute
156 /// the same number of times.
157 EquivalenceClassMap EquivalenceClass;
159 /// \brief Dominance, post-dominance and loop information.
161 PostDominatorTree *PDT;
164 /// \brief Predecessors for each basic block in the CFG.
165 BlockEdgeMap Predecessors;
167 /// \brief Successors for each basic block in the CFG.
168 BlockEdgeMap Successors;
170 /// \brief LLVM context holding the debug data we need.
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.
187 /// \brief Print the weight of edge \p E on stream \p OS.
189 /// \param OS Stream to emit the output to.
190 /// \param E Edge to print.
191 void SampleProfileLoader::printEdgeWeight(raw_ostream &OS, Edge E) {
192 OS << "weight[" << E.first->getName() << "->" << E.second->getName()
193 << "]: " << EdgeWeights[E] << "\n";
196 /// \brief Print the equivalence class of block \p BB on stream \p OS.
198 /// \param OS Stream to emit the output to.
199 /// \param BB Block to print.
200 void SampleProfileLoader::printBlockEquivalence(raw_ostream &OS,
202 BasicBlock *Equiv = EquivalenceClass[BB];
203 OS << "equivalence[" << BB->getName()
204 << "]: " << ((Equiv) ? EquivalenceClass[BB]->getName() : "NONE") << "\n";
207 /// \brief Print the weight of block \p BB on stream \p OS.
209 /// \param OS Stream to emit the output to.
210 /// \param BB Block to print.
211 void SampleProfileLoader::printBlockWeight(raw_ostream &OS, BasicBlock *BB) {
212 OS << "weight[" << BB->getName() << "]: " << BlockWeights[BB] << "\n";
215 /// \brief Get the weight for an instruction.
217 /// The "weight" of an instruction \p Inst is the number of samples
218 /// collected on that instruction at runtime. To retrieve it, we
219 /// need to compute the line number of \p Inst relative to the start of its
220 /// function. We use HeaderLineno to compute the offset. We then
221 /// look up the samples collected for \p Inst using BodySamples.
223 /// \param Inst Instruction to query.
225 /// \returns The profiled weight of I.
226 unsigned SampleProfileLoader::getInstWeight(Instruction &Inst) {
227 DebugLoc DLoc = Inst.getDebugLoc();
231 unsigned Lineno = DLoc.getLine();
232 if (Lineno < HeaderLineno)
235 const DILocation *DIL = DLoc;
236 int LOffset = Lineno - HeaderLineno;
237 unsigned Discriminator = DIL->getDiscriminator();
238 unsigned Weight = Samples->samplesAt(LOffset, Discriminator);
239 DEBUG(dbgs() << " " << Lineno << "." << Discriminator << ":" << Inst
240 << " (line offset: " << LOffset << "." << Discriminator
241 << " - weight: " << Weight << ")\n");
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. The weight of \p BB is computed and cached in
249 /// the BlockWeights map.
251 /// \param BB The basic block to query.
253 /// \returns The computed weight of BB.
254 unsigned SampleProfileLoader::getBlockWeight(BasicBlock *BB) {
255 // If we've computed BB's weight before, return it.
256 std::pair<BlockWeightMap::iterator, bool> Entry =
257 BlockWeights.insert(std::make_pair(BB, 0));
259 return Entry.first->second;
261 // Otherwise, compute and cache BB's weight.
263 for (auto &I : BB->getInstList()) {
264 unsigned InstWeight = getInstWeight(I);
265 if (InstWeight > Weight)
268 Entry.first->second = Weight;
272 /// \brief Compute and store the weights of every basic block.
274 /// This populates the BlockWeights map by computing
275 /// the weights of every basic block in the CFG.
277 /// \param F The function to query.
278 bool SampleProfileLoader::computeBlockWeights(Function &F) {
279 bool Changed = false;
280 DEBUG(dbgs() << "Block weights\n");
282 unsigned Weight = getBlockWeight(&BB);
283 Changed |= (Weight > 0);
284 DEBUG(printBlockWeight(dbgs(), &BB));
290 /// \brief Find equivalence classes for the given block.
292 /// This finds all the blocks that are guaranteed to execute the same
293 /// number of times as \p BB1. To do this, it traverses all the
294 /// descendants of \p BB1 in the dominator or post-dominator tree.
296 /// A block BB2 will be in the same equivalence class as \p BB1 if
297 /// the following holds:
299 /// 1- \p BB1 is a descendant of BB2 in the opposite tree. So, if BB2
300 /// is a descendant of \p BB1 in the dominator tree, then BB2 should
301 /// dominate BB1 in the post-dominator tree.
303 /// 2- Both BB2 and \p BB1 must be in the same loop.
305 /// For every block BB2 that meets those two requirements, we set BB2's
306 /// equivalence class to \p BB1.
308 /// \param BB1 Block to check.
309 /// \param Descendants Descendants of \p BB1 in either the dom or pdom tree.
310 /// \param DomTree Opposite dominator tree. If \p Descendants is filled
311 /// with blocks from \p BB1's dominator tree, then
312 /// this is the post-dominator tree, and vice versa.
313 void SampleProfileLoader::findEquivalencesFor(
314 BasicBlock *BB1, SmallVector<BasicBlock *, 8> Descendants,
315 DominatorTreeBase<BasicBlock> *DomTree) {
316 for (auto *BB2 : Descendants) {
317 bool IsDomParent = DomTree->dominates(BB2, BB1);
318 bool IsInSameLoop = LI->getLoopFor(BB1) == LI->getLoopFor(BB2);
319 if (BB1 != BB2 && VisitedBlocks.insert(BB2).second && IsDomParent &&
321 EquivalenceClass[BB2] = BB1;
323 // If BB2 is heavier than BB1, make BB2 have the same weight
326 // Note that we don't worry about the opposite situation here
327 // (when BB2 is lighter than BB1). We will deal with this
328 // during the propagation phase. Right now, we just want to
329 // make sure that BB1 has the largest weight of all the
330 // members of its equivalence set.
331 unsigned &BB1Weight = BlockWeights[BB1];
332 unsigned &BB2Weight = BlockWeights[BB2];
333 BB1Weight = std::max(BB1Weight, BB2Weight);
338 /// \brief Find equivalence classes.
340 /// Since samples may be missing from blocks, we can fill in the gaps by setting
341 /// the weights of all the blocks in the same equivalence class to the same
342 /// weight. To compute the concept of equivalence, we use dominance and loop
343 /// information. Two blocks B1 and B2 are in the same equivalence class if B1
344 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
346 /// \param F The function to query.
347 void SampleProfileLoader::findEquivalenceClasses(Function &F) {
348 SmallVector<BasicBlock *, 8> DominatedBBs;
349 DEBUG(dbgs() << "\nBlock equivalence classes\n");
350 // Find equivalence sets based on dominance and post-dominance information.
352 BasicBlock *BB1 = &BB;
354 // Compute BB1's equivalence class once.
355 if (EquivalenceClass.count(BB1)) {
356 DEBUG(printBlockEquivalence(dbgs(), BB1));
360 // By default, blocks are in their own equivalence class.
361 EquivalenceClass[BB1] = BB1;
363 // Traverse all the blocks dominated by BB1. We are looking for
364 // every basic block BB2 such that:
366 // 1- BB1 dominates BB2.
367 // 2- BB2 post-dominates BB1.
368 // 3- BB1 and BB2 are in the same loop nest.
370 // If all those conditions hold, it means that BB2 is executed
371 // as many times as BB1, so they are placed in the same equivalence
372 // class by making BB2's equivalence class be BB1.
373 DominatedBBs.clear();
374 DT->getDescendants(BB1, DominatedBBs);
375 findEquivalencesFor(BB1, DominatedBBs, PDT->DT);
377 // Repeat the same logic for all the blocks post-dominated by BB1.
378 // We are looking for every basic block BB2 such that:
380 // 1- BB1 post-dominates BB2.
381 // 2- BB2 dominates BB1.
382 // 3- BB1 and BB2 are in the same loop nest.
384 // If all those conditions hold, BB2's equivalence class is BB1.
385 DominatedBBs.clear();
386 PDT->getDescendants(BB1, DominatedBBs);
387 findEquivalencesFor(BB1, DominatedBBs, DT);
389 DEBUG(printBlockEquivalence(dbgs(), BB1));
392 // Assign weights to equivalence classes.
394 // All the basic blocks in the same equivalence class will execute
395 // the same number of times. Since we know that the head block in
396 // each equivalence class has the largest weight, assign that weight
397 // to all the blocks in that equivalence class.
398 DEBUG(dbgs() << "\nAssign the same weight to all blocks in the same class\n");
400 BasicBlock *BB = &BI;
401 BasicBlock *EquivBB = EquivalenceClass[BB];
403 BlockWeights[BB] = BlockWeights[EquivBB];
404 DEBUG(printBlockWeight(dbgs(), BB));
408 /// \brief Visit the given edge to decide if it has a valid weight.
410 /// If \p E has not been visited before, we copy to \p UnknownEdge
411 /// and increment the count of unknown edges.
413 /// \param E Edge to visit.
414 /// \param NumUnknownEdges Current number of unknown edges.
415 /// \param UnknownEdge Set if E has not been visited before.
417 /// \returns E's weight, if known. Otherwise, return 0.
418 unsigned SampleProfileLoader::visitEdge(Edge E, unsigned *NumUnknownEdges,
420 if (!VisitedEdges.count(E)) {
421 (*NumUnknownEdges)++;
426 return EdgeWeights[E];
429 /// \brief Propagate weights through incoming/outgoing edges.
431 /// If the weight of a basic block is known, and there is only one edge
432 /// with an unknown weight, we can calculate the weight of that edge.
434 /// Similarly, if all the edges have a known count, we can calculate the
435 /// count of the basic block, if needed.
437 /// \param F Function to process.
439 /// \returns True if new weights were assigned to edges or blocks.
440 bool SampleProfileLoader::propagateThroughEdges(Function &F) {
441 bool Changed = false;
442 DEBUG(dbgs() << "\nPropagation through edges\n");
444 BasicBlock *BB = &BI;
446 // Visit all the predecessor and successor edges to determine
447 // which ones have a weight assigned already. Note that it doesn't
448 // matter that we only keep track of a single unknown edge. The
449 // only case we are interested in handling is when only a single
450 // edge is unknown (see setEdgeOrBlockWeight).
451 for (unsigned i = 0; i < 2; i++) {
452 unsigned TotalWeight = 0;
453 unsigned NumUnknownEdges = 0;
454 Edge UnknownEdge, SelfReferentialEdge;
457 // First, visit all predecessor edges.
458 for (auto *Pred : Predecessors[BB]) {
459 Edge E = std::make_pair(Pred, BB);
460 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
461 if (E.first == E.second)
462 SelfReferentialEdge = E;
465 // On the second round, visit all successor edges.
466 for (auto *Succ : Successors[BB]) {
467 Edge E = std::make_pair(BB, Succ);
468 TotalWeight += visitEdge(E, &NumUnknownEdges, &UnknownEdge);
472 // After visiting all the edges, there are three cases that we
473 // can handle immediately:
475 // - All the edge weights are known (i.e., NumUnknownEdges == 0).
476 // In this case, we simply check that the sum of all the edges
477 // is the same as BB's weight. If not, we change BB's weight
478 // to match. Additionally, if BB had not been visited before,
479 // we mark it visited.
481 // - Only one edge is unknown and BB has already been visited.
482 // In this case, we can compute the weight of the edge by
483 // subtracting the total block weight from all the known
484 // edge weights. If the edges weight more than BB, then the
485 // edge of the last remaining edge is set to zero.
487 // - There exists a self-referential edge and the weight of BB is
488 // known. In this case, this edge can be based on BB's weight.
489 // We add up all the other known edges and set the weight on
490 // the self-referential edge as we did in the previous case.
492 // In any other case, we must continue iterating. Eventually,
493 // all edges will get a weight, or iteration will stop when
494 // it reaches SampleProfileMaxPropagateIterations.
495 if (NumUnknownEdges <= 1) {
496 unsigned &BBWeight = BlockWeights[BB];
497 if (NumUnknownEdges == 0) {
498 // If we already know the weight of all edges, the weight of the
499 // basic block can be computed. It should be no larger than the sum
500 // of all edge weights.
501 if (TotalWeight > BBWeight) {
502 BBWeight = TotalWeight;
504 DEBUG(dbgs() << "All edge weights for " << BB->getName()
505 << " known. Set weight for block: ";
506 printBlockWeight(dbgs(), BB););
508 if (VisitedBlocks.insert(BB).second)
510 } else if (NumUnknownEdges == 1 && VisitedBlocks.count(BB)) {
511 // If there is a single unknown edge and the block has been
512 // visited, then we can compute E's weight.
513 if (BBWeight >= TotalWeight)
514 EdgeWeights[UnknownEdge] = BBWeight - TotalWeight;
516 EdgeWeights[UnknownEdge] = 0;
517 VisitedEdges.insert(UnknownEdge);
519 DEBUG(dbgs() << "Set weight for edge: ";
520 printEdgeWeight(dbgs(), UnknownEdge));
522 } else if (SelfReferentialEdge.first && VisitedBlocks.count(BB)) {
523 unsigned &BBWeight = BlockWeights[BB];
524 // We have a self-referential edge and the weight of BB is known.
525 if (BBWeight >= TotalWeight)
526 EdgeWeights[SelfReferentialEdge] = BBWeight - TotalWeight;
528 EdgeWeights[SelfReferentialEdge] = 0;
529 VisitedEdges.insert(SelfReferentialEdge);
531 DEBUG(dbgs() << "Set self-referential edge weight to: ";
532 printEdgeWeight(dbgs(), SelfReferentialEdge));
540 /// \brief Build in/out edge lists for each basic block in the CFG.
542 /// We are interested in unique edges. If a block B1 has multiple
543 /// edges to another block B2, we only add a single B1->B2 edge.
544 void SampleProfileLoader::buildEdges(Function &F) {
546 BasicBlock *B1 = &BI;
548 // Add predecessors for B1.
549 SmallPtrSet<BasicBlock *, 16> Visited;
550 if (!Predecessors[B1].empty())
551 llvm_unreachable("Found a stale predecessors list in a basic block.");
552 for (pred_iterator PI = pred_begin(B1), PE = pred_end(B1); PI != PE; ++PI) {
553 BasicBlock *B2 = *PI;
554 if (Visited.insert(B2).second)
555 Predecessors[B1].push_back(B2);
558 // Add successors for B1.
560 if (!Successors[B1].empty())
561 llvm_unreachable("Found a stale successors list in a basic block.");
562 for (succ_iterator SI = succ_begin(B1), SE = succ_end(B1); SI != SE; ++SI) {
563 BasicBlock *B2 = *SI;
564 if (Visited.insert(B2).second)
565 Successors[B1].push_back(B2);
570 /// \brief Propagate weights into edges
572 /// The following rules are applied to every block BB in the CFG:
574 /// - If BB has a single predecessor/successor, then the weight
575 /// of that edge is the weight of the block.
577 /// - If all incoming or outgoing edges are known except one, and the
578 /// weight of the block is already known, the weight of the unknown
579 /// edge will be the weight of the block minus the sum of all the known
580 /// edges. If the sum of all the known edges is larger than BB's weight,
581 /// we set the unknown edge weight to zero.
583 /// - If there is a self-referential edge, and the weight of the block is
584 /// known, the weight for that edge is set to the weight of the block
585 /// minus the weight of the other incoming edges to that block (if
587 void SampleProfileLoader::propagateWeights(Function &F) {
591 // Add an entry count to the function using the samples gathered
592 // at the function entry.
593 F.setEntryCount(Samples->getHeadSamples());
595 // Before propagation starts, build, for each block, a list of
596 // unique predecessors and successors. This is necessary to handle
597 // identical edges in multiway branches. Since we visit all blocks and all
598 // edges of the CFG, it is cleaner to build these lists once at the start
602 // Propagate until we converge or we go past the iteration limit.
603 while (Changed && i++ < SampleProfileMaxPropagateIterations) {
604 Changed = propagateThroughEdges(F);
607 // Generate MD_prof metadata for every branch instruction using the
608 // edge weights computed during propagation.
609 DEBUG(dbgs() << "\nPropagation complete. Setting branch weights\n");
610 MDBuilder MDB(F.getContext());
612 BasicBlock *BB = &BI;
613 TerminatorInst *TI = BB->getTerminator();
614 if (TI->getNumSuccessors() == 1)
616 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
619 DEBUG(dbgs() << "\nGetting weights for branch at line "
620 << TI->getDebugLoc().getLine() << ".\n");
621 SmallVector<unsigned, 4> Weights;
622 bool AllWeightsZero = true;
623 for (unsigned I = 0; I < TI->getNumSuccessors(); ++I) {
624 BasicBlock *Succ = TI->getSuccessor(I);
625 Edge E = std::make_pair(BB, Succ);
626 unsigned Weight = EdgeWeights[E];
627 DEBUG(dbgs() << "\t"; printEdgeWeight(dbgs(), E));
628 Weights.push_back(Weight);
630 AllWeightsZero = false;
633 // Only set weights if there is at least one non-zero weight.
634 // In any other case, let the analyzer set weights.
635 if (!AllWeightsZero) {
636 DEBUG(dbgs() << "SUCCESS. Found non-zero weights.\n");
637 TI->setMetadata(llvm::LLVMContext::MD_prof,
638 MDB.createBranchWeights(Weights));
640 DEBUG(dbgs() << "SKIPPED. All branch weights are zero.\n");
645 /// \brief Get the line number for the function header.
647 /// This looks up function \p F in the current compilation unit and
648 /// retrieves the line number where the function is defined. This is
649 /// line 0 for all the samples read from the profile file. Every line
650 /// number is relative to this line.
652 /// \param F Function object to query.
654 /// \returns the line number where \p F is defined. If it returns 0,
655 /// it means that there is no debug information available for \p F.
656 unsigned SampleProfileLoader::getFunctionLoc(Function &F) {
657 if (DISubprogram *S = getDISubprogram(&F))
660 // If could not find the start of \p F, emit a diagnostic to inform the user
661 // about the missed opportunity.
662 F.getContext().diagnose(DiagnosticInfoSampleProfile(
663 "No debug information found in function " + F.getName() +
664 ": Function profile not used",
669 /// \brief Generate branch weight metadata for all branches in \p F.
671 /// Branch weights are computed out of instruction samples using a
672 /// propagation heuristic. Propagation proceeds in 3 phases:
674 /// 1- Assignment of block weights. All the basic blocks in the function
675 /// are initial assigned the same weight as their most frequently
676 /// executed instruction.
678 /// 2- Creation of equivalence classes. Since samples may be missing from
679 /// blocks, we can fill in the gaps by setting the weights of all the
680 /// blocks in the same equivalence class to the same weight. To compute
681 /// the concept of equivalence, we use dominance and loop information.
682 /// Two blocks B1 and B2 are in the same equivalence class if B1
683 /// dominates B2, B2 post-dominates B1 and both are in the same loop.
685 /// 3- Propagation of block weights into edges. This uses a simple
686 /// propagation heuristic. The following rules are applied to every
687 /// block BB in the CFG:
689 /// - If BB has a single predecessor/successor, then the weight
690 /// of that edge is the weight of the block.
692 /// - If all the edges are known except one, and the weight of the
693 /// block is already known, the weight of the unknown edge will
694 /// be the weight of the block minus the sum of all the known
695 /// edges. If the sum of all the known edges is larger than BB's weight,
696 /// we set the unknown edge weight to zero.
698 /// - If there is a self-referential edge, and the weight of the block is
699 /// known, the weight for that edge is set to the weight of the block
700 /// minus the weight of the other incoming edges to that block (if
703 /// Since this propagation is not guaranteed to finalize for every CFG, we
704 /// only allow it to proceed for a limited number of iterations (controlled
705 /// by -sample-profile-max-propagate-iterations).
707 /// FIXME: Try to replace this propagation heuristic with a scheme
708 /// that is guaranteed to finalize. A work-list approach similar to
709 /// the standard value propagation algorithm used by SSA-CCP might
712 /// Once all the branch weights are computed, we emit the MD_prof
713 /// metadata on BB using the computed values for each of its branches.
715 /// \param F The function to query.
717 /// \returns true if \p F was modified. Returns false, otherwise.
718 bool SampleProfileLoader::emitAnnotations(Function &F) {
719 bool Changed = false;
721 // Initialize invariants used during computation and propagation.
722 HeaderLineno = getFunctionLoc(F);
723 if (HeaderLineno == 0)
726 DEBUG(dbgs() << "Line number for the first instruction in " << F.getName()
727 << ": " << HeaderLineno << "\n");
729 // Compute basic block weights.
730 Changed |= computeBlockWeights(F);
733 // Find equivalence classes.
734 findEquivalenceClasses(F);
736 // Propagate weights to all edges.
743 char SampleProfileLoader::ID = 0;
744 INITIALIZE_PASS_BEGIN(SampleProfileLoader, "sample-profile",
745 "Sample Profile loader", false, false)
746 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
747 INITIALIZE_PASS_DEPENDENCY(PostDominatorTree)
748 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
749 INITIALIZE_PASS_DEPENDENCY(AddDiscriminators)
750 INITIALIZE_PASS_END(SampleProfileLoader, "sample-profile",
751 "Sample Profile loader", false, false)
753 bool SampleProfileLoader::doInitialization(Module &M) {
754 auto& Ctx = M.getContext();
755 auto ReaderOrErr = SampleProfileReader::create(Filename, Ctx);
756 if (std::error_code EC = ReaderOrErr.getError()) {
757 std::string Msg = "Could not open profile: " + EC.message();
758 Ctx.diagnose(DiagnosticInfoSampleProfile(Filename.data(), Msg));
761 Reader = std::move(ReaderOrErr.get());
762 ProfileIsValid = (Reader->read() == sampleprof_error::success);
766 ModulePass *llvm::createSampleProfileLoaderPass() {
767 return new SampleProfileLoader(SampleProfileFile);
770 ModulePass *llvm::createSampleProfileLoaderPass(StringRef Name) {
771 return new SampleProfileLoader(Name);
774 bool SampleProfileLoader::runOnModule(Module &M) {
777 if (!F.isDeclaration())
778 retval |= runOnFunction(F);
782 bool SampleProfileLoader::runOnFunction(Function &F) {
786 DT = &getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
787 PDT = &getAnalysis<PostDominatorTree>(F);
788 LI = &getAnalysis<LoopInfoWrapperPass>(F).getLoopInfo();
789 Ctx = &F.getParent()->getContext();
790 Samples = Reader->getSamplesFor(F);
791 if (!Samples->empty())
792 return emitAnnotations(F);