1 //===-- BranchProbabilityInfo.cpp - Branch Probability Analysis -----------===//
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 // Loops should be simplified before this analysis.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Analysis/BranchProbabilityInfo.h"
15 #include "llvm/ADT/PostOrderIterator.h"
16 #include "llvm/Analysis/LoopInfo.h"
17 #include "llvm/IR/Constants.h"
18 #include "llvm/IR/Function.h"
19 #include "llvm/IR/Instructions.h"
20 #include "llvm/IR/LLVMContext.h"
21 #include "llvm/IR/Metadata.h"
22 #include "llvm/Support/CFG.h"
23 #include "llvm/Support/Debug.h"
27 INITIALIZE_PASS_BEGIN(BranchProbabilityInfo, "branch-prob",
28 "Branch Probability Analysis", false, true)
29 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
30 INITIALIZE_PASS_END(BranchProbabilityInfo, "branch-prob",
31 "Branch Probability Analysis", false, true)
33 char BranchProbabilityInfo::ID = 0;
35 // Weights are for internal use only. They are used by heuristics to help to
36 // estimate edges' probability. Example:
38 // Using "Loop Branch Heuristics" we predict weights of edges for the
53 // Probability of the edge BB2->BB1 = 124 / (124 + 4) = 0.96875
54 // Probability of the edge BB2->BB3 = 4 / (124 + 4) = 0.03125
55 static const uint32_t LBH_TAKEN_WEIGHT = 124;
56 static const uint32_t LBH_NONTAKEN_WEIGHT = 4;
58 /// \brief Unreachable-terminating branch taken weight.
60 /// This is the weight for a branch being taken to a block that terminates
61 /// (eventually) in unreachable. These are predicted as unlikely as possible.
62 static const uint32_t UR_TAKEN_WEIGHT = 1;
64 /// \brief Unreachable-terminating branch not-taken weight.
66 /// This is the weight for a branch not being taken toward a block that
67 /// terminates (eventually) in unreachable. Such a branch is essentially never
68 /// taken. Set the weight to an absurdly high value so that nested loops don't
69 /// easily subsume it.
70 static const uint32_t UR_NONTAKEN_WEIGHT = 1024*1024 - 1;
72 /// \brief Weight for a branch taken going into a cold block.
74 /// This is the weight for a branch taken toward a block marked
75 /// cold. A block is marked cold if it's postdominated by a
76 /// block containing a call to a cold function. Cold functions
77 /// are those marked with attribute 'cold'.
78 static const uint32_t CC_TAKEN_WEIGHT = 4;
80 /// \brief Weight for a branch not-taken into a cold block.
82 /// This is the weight for a branch not taken toward a block marked
84 static const uint32_t CC_NONTAKEN_WEIGHT = 64;
86 static const uint32_t PH_TAKEN_WEIGHT = 20;
87 static const uint32_t PH_NONTAKEN_WEIGHT = 12;
89 static const uint32_t ZH_TAKEN_WEIGHT = 20;
90 static const uint32_t ZH_NONTAKEN_WEIGHT = 12;
92 static const uint32_t FPH_TAKEN_WEIGHT = 20;
93 static const uint32_t FPH_NONTAKEN_WEIGHT = 12;
95 /// \brief Invoke-terminating normal branch taken weight
97 /// This is the weight for branching to the normal destination of an invoke
98 /// instruction. We expect this to happen most of the time. Set the weight to an
99 /// absurdly high value so that nested loops subsume it.
100 static const uint32_t IH_TAKEN_WEIGHT = 1024 * 1024 - 1;
102 /// \brief Invoke-terminating normal branch not-taken weight.
104 /// This is the weight for branching to the unwind destination of an invoke
105 /// instruction. This is essentially never taken.
106 static const uint32_t IH_NONTAKEN_WEIGHT = 1;
108 // Standard weight value. Used when none of the heuristics set weight for
110 static const uint32_t NORMAL_WEIGHT = 16;
112 // Minimum weight of an edge. Please note, that weight is NEVER 0.
113 static const uint32_t MIN_WEIGHT = 1;
115 static uint32_t getMaxWeightFor(BasicBlock *BB) {
116 return UINT32_MAX / BB->getTerminator()->getNumSuccessors();
120 /// \brief Calculate edge weights for successors lead to unreachable.
122 /// Predict that a successor which leads necessarily to an
123 /// unreachable-terminated block as extremely unlikely.
124 bool BranchProbabilityInfo::calcUnreachableHeuristics(BasicBlock *BB) {
125 TerminatorInst *TI = BB->getTerminator();
126 if (TI->getNumSuccessors() == 0) {
127 if (isa<UnreachableInst>(TI))
128 PostDominatedByUnreachable.insert(BB);
132 SmallVector<unsigned, 4> UnreachableEdges;
133 SmallVector<unsigned, 4> ReachableEdges;
135 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
136 if (PostDominatedByUnreachable.count(*I))
137 UnreachableEdges.push_back(I.getSuccessorIndex());
139 ReachableEdges.push_back(I.getSuccessorIndex());
142 // If all successors are in the set of blocks post-dominated by unreachable,
143 // this block is too.
144 if (UnreachableEdges.size() == TI->getNumSuccessors())
145 PostDominatedByUnreachable.insert(BB);
147 // Skip probabilities if this block has a single successor or if all were
149 if (TI->getNumSuccessors() == 1 || UnreachableEdges.empty())
152 uint32_t UnreachableWeight =
153 std::max(UR_TAKEN_WEIGHT / (unsigned)UnreachableEdges.size(), MIN_WEIGHT);
154 for (SmallVectorImpl<unsigned>::iterator I = UnreachableEdges.begin(),
155 E = UnreachableEdges.end();
157 setEdgeWeight(BB, *I, UnreachableWeight);
159 if (ReachableEdges.empty())
161 uint32_t ReachableWeight =
162 std::max(UR_NONTAKEN_WEIGHT / (unsigned)ReachableEdges.size(),
164 for (SmallVectorImpl<unsigned>::iterator I = ReachableEdges.begin(),
165 E = ReachableEdges.end();
167 setEdgeWeight(BB, *I, ReachableWeight);
172 // Propagate existing explicit probabilities from either profile data or
173 // 'expect' intrinsic processing.
174 bool BranchProbabilityInfo::calcMetadataWeights(BasicBlock *BB) {
175 TerminatorInst *TI = BB->getTerminator();
176 if (TI->getNumSuccessors() == 1)
178 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
181 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
185 // Ensure there are weights for all of the successors. Note that the first
186 // operand to the metadata node is a name, not a weight.
187 if (WeightsNode->getNumOperands() != TI->getNumSuccessors() + 1)
190 // Build up the final weights that will be used in a temporary buffer, but
191 // don't add them until all weihts are present. Each weight value is clamped
192 // to [1, getMaxWeightFor(BB)].
193 uint32_t WeightLimit = getMaxWeightFor(BB);
194 SmallVector<uint32_t, 2> Weights;
195 Weights.reserve(TI->getNumSuccessors());
196 for (unsigned i = 1, e = WeightsNode->getNumOperands(); i != e; ++i) {
197 ConstantInt *Weight = dyn_cast<ConstantInt>(WeightsNode->getOperand(i));
201 std::max<uint32_t>(1, Weight->getLimitedValue(WeightLimit)));
203 assert(Weights.size() == TI->getNumSuccessors() && "Checked above");
204 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
205 setEdgeWeight(BB, i, Weights[i]);
210 /// \brief Calculate edge weights for edges leading to cold blocks.
212 /// A cold block is one post-dominated by a block with a call to a
213 /// cold function. Those edges are unlikely to be taken, so we give
214 /// them relatively low weight.
216 /// Return true if we could compute the weights for cold edges.
217 /// Return false, otherwise.
218 bool BranchProbabilityInfo::calcColdCallHeuristics(BasicBlock *BB) {
219 TerminatorInst *TI = BB->getTerminator();
220 if (TI->getNumSuccessors() == 0)
223 // Determine which successors are post-dominated by a cold block.
224 SmallVector<unsigned, 4> ColdEdges;
225 SmallVector<unsigned, 4> NormalEdges;
226 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I)
227 if (PostDominatedByColdCall.count(*I))
228 ColdEdges.push_back(I.getSuccessorIndex());
230 NormalEdges.push_back(I.getSuccessorIndex());
232 // If all successors are in the set of blocks post-dominated by cold calls,
233 // this block is in the set post-dominated by cold calls.
234 if (ColdEdges.size() == TI->getNumSuccessors())
235 PostDominatedByColdCall.insert(BB);
237 // Otherwise, if the block itself contains a cold function, add it to the
238 // set of blocks postdominated by a cold call.
239 assert(!PostDominatedByColdCall.count(BB));
240 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
241 if (CallInst *CI = dyn_cast<CallInst>(I))
242 if (CI->hasFnAttr(Attribute::Cold)) {
243 PostDominatedByColdCall.insert(BB);
248 // Skip probabilities if this block has a single successor.
249 if (TI->getNumSuccessors() == 1 || ColdEdges.empty())
252 uint32_t ColdWeight =
253 std::max(CC_TAKEN_WEIGHT / (unsigned) ColdEdges.size(), MIN_WEIGHT);
254 for (SmallVectorImpl<unsigned>::iterator I = ColdEdges.begin(),
257 setEdgeWeight(BB, *I, ColdWeight);
259 if (NormalEdges.empty())
261 uint32_t NormalWeight = std::max(
262 CC_NONTAKEN_WEIGHT / (unsigned) NormalEdges.size(), NORMAL_WEIGHT);
263 for (SmallVectorImpl<unsigned>::iterator I = NormalEdges.begin(),
264 E = NormalEdges.end();
266 setEdgeWeight(BB, *I, NormalWeight);
271 // Calculate Edge Weights using "Pointer Heuristics". Predict a comparsion
272 // between two pointer or pointer and NULL will fail.
273 bool BranchProbabilityInfo::calcPointerHeuristics(BasicBlock *BB) {
274 BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
275 if (!BI || !BI->isConditional())
278 Value *Cond = BI->getCondition();
279 ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
280 if (!CI || !CI->isEquality())
283 Value *LHS = CI->getOperand(0);
285 if (!LHS->getType()->isPointerTy())
288 assert(CI->getOperand(1)->getType()->isPointerTy());
290 // p != 0 -> isProb = true
291 // p == 0 -> isProb = false
292 // p != q -> isProb = true
293 // p == q -> isProb = false;
294 unsigned TakenIdx = 0, NonTakenIdx = 1;
295 bool isProb = CI->getPredicate() == ICmpInst::ICMP_NE;
297 std::swap(TakenIdx, NonTakenIdx);
299 setEdgeWeight(BB, TakenIdx, PH_TAKEN_WEIGHT);
300 setEdgeWeight(BB, NonTakenIdx, PH_NONTAKEN_WEIGHT);
304 // Calculate Edge Weights using "Loop Branch Heuristics". Predict backedges
305 // as taken, exiting edges as not-taken.
306 bool BranchProbabilityInfo::calcLoopBranchHeuristics(BasicBlock *BB) {
307 Loop *L = LI->getLoopFor(BB);
311 SmallVector<unsigned, 8> BackEdges;
312 SmallVector<unsigned, 8> ExitingEdges;
313 SmallVector<unsigned, 8> InEdges; // Edges from header to the loop.
315 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
316 if (!L->contains(*I))
317 ExitingEdges.push_back(I.getSuccessorIndex());
318 else if (L->getHeader() == *I)
319 BackEdges.push_back(I.getSuccessorIndex());
321 InEdges.push_back(I.getSuccessorIndex());
324 if (uint32_t numBackEdges = BackEdges.size()) {
325 uint32_t backWeight = LBH_TAKEN_WEIGHT / numBackEdges;
326 if (backWeight < NORMAL_WEIGHT)
327 backWeight = NORMAL_WEIGHT;
329 for (SmallVectorImpl<unsigned>::iterator EI = BackEdges.begin(),
330 EE = BackEdges.end(); EI != EE; ++EI) {
331 setEdgeWeight(BB, *EI, backWeight);
335 if (uint32_t numInEdges = InEdges.size()) {
336 uint32_t inWeight = LBH_TAKEN_WEIGHT / numInEdges;
337 if (inWeight < NORMAL_WEIGHT)
338 inWeight = NORMAL_WEIGHT;
340 for (SmallVectorImpl<unsigned>::iterator EI = InEdges.begin(),
341 EE = InEdges.end(); EI != EE; ++EI) {
342 setEdgeWeight(BB, *EI, inWeight);
346 if (uint32_t numExitingEdges = ExitingEdges.size()) {
347 uint32_t exitWeight = LBH_NONTAKEN_WEIGHT / numExitingEdges;
348 if (exitWeight < MIN_WEIGHT)
349 exitWeight = MIN_WEIGHT;
351 for (SmallVectorImpl<unsigned>::iterator EI = ExitingEdges.begin(),
352 EE = ExitingEdges.end(); EI != EE; ++EI) {
353 setEdgeWeight(BB, *EI, exitWeight);
360 bool BranchProbabilityInfo::calcZeroHeuristics(BasicBlock *BB) {
361 BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
362 if (!BI || !BI->isConditional())
365 Value *Cond = BI->getCondition();
366 ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
370 Value *RHS = CI->getOperand(1);
371 ConstantInt *CV = dyn_cast<ConstantInt>(RHS);
377 switch (CI->getPredicate()) {
378 case CmpInst::ICMP_EQ:
379 // X == 0 -> Unlikely
382 case CmpInst::ICMP_NE:
386 case CmpInst::ICMP_SLT:
390 case CmpInst::ICMP_SGT:
397 } else if (CV->isOne() && CI->getPredicate() == CmpInst::ICMP_SLT) {
398 // InstCombine canonicalizes X <= 0 into X < 1.
399 // X <= 0 -> Unlikely
401 } else if (CV->isAllOnesValue()) {
402 switch (CI->getPredicate()) {
403 case CmpInst::ICMP_EQ:
404 // X == -1 -> Unlikely
407 case CmpInst::ICMP_NE:
411 case CmpInst::ICMP_SGT:
412 // InstCombine canonicalizes X >= 0 into X > -1.
423 unsigned TakenIdx = 0, NonTakenIdx = 1;
426 std::swap(TakenIdx, NonTakenIdx);
428 setEdgeWeight(BB, TakenIdx, ZH_TAKEN_WEIGHT);
429 setEdgeWeight(BB, NonTakenIdx, ZH_NONTAKEN_WEIGHT);
434 bool BranchProbabilityInfo::calcFloatingPointHeuristics(BasicBlock *BB) {
435 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
436 if (!BI || !BI->isConditional())
439 Value *Cond = BI->getCondition();
440 FCmpInst *FCmp = dyn_cast<FCmpInst>(Cond);
445 if (FCmp->isEquality()) {
446 // f1 == f2 -> Unlikely
447 // f1 != f2 -> Likely
448 isProb = !FCmp->isTrueWhenEqual();
449 } else if (FCmp->getPredicate() == FCmpInst::FCMP_ORD) {
452 } else if (FCmp->getPredicate() == FCmpInst::FCMP_UNO) {
459 unsigned TakenIdx = 0, NonTakenIdx = 1;
462 std::swap(TakenIdx, NonTakenIdx);
464 setEdgeWeight(BB, TakenIdx, FPH_TAKEN_WEIGHT);
465 setEdgeWeight(BB, NonTakenIdx, FPH_NONTAKEN_WEIGHT);
470 bool BranchProbabilityInfo::calcInvokeHeuristics(BasicBlock *BB) {
471 InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator());
475 setEdgeWeight(BB, 0/*Index for Normal*/, IH_TAKEN_WEIGHT);
476 setEdgeWeight(BB, 1/*Index for Unwind*/, IH_NONTAKEN_WEIGHT);
480 void BranchProbabilityInfo::getAnalysisUsage(AnalysisUsage &AU) const {
481 AU.addRequired<LoopInfo>();
482 AU.setPreservesAll();
485 bool BranchProbabilityInfo::runOnFunction(Function &F) {
486 LastF = &F; // Store the last function we ran on for printing.
487 LI = &getAnalysis<LoopInfo>();
488 assert(PostDominatedByUnreachable.empty());
489 assert(PostDominatedByColdCall.empty());
491 // Walk the basic blocks in post-order so that we can build up state about
492 // the successors of a block iteratively.
493 for (po_iterator<BasicBlock *> I = po_begin(&F.getEntryBlock()),
494 E = po_end(&F.getEntryBlock());
496 DEBUG(dbgs() << "Computing probabilities for " << I->getName() << "\n");
497 if (calcUnreachableHeuristics(*I))
499 if (calcMetadataWeights(*I))
501 if (calcColdCallHeuristics(*I))
503 if (calcLoopBranchHeuristics(*I))
505 if (calcPointerHeuristics(*I))
507 if (calcZeroHeuristics(*I))
509 if (calcFloatingPointHeuristics(*I))
511 calcInvokeHeuristics(*I);
514 PostDominatedByUnreachable.clear();
515 PostDominatedByColdCall.clear();
519 void BranchProbabilityInfo::print(raw_ostream &OS, const Module *) const {
520 OS << "---- Branch Probabilities ----\n";
521 // We print the probabilities from the last function the analysis ran over,
522 // or the function it is currently running over.
523 assert(LastF && "Cannot print prior to running over a function");
524 for (Function::const_iterator BI = LastF->begin(), BE = LastF->end();
526 for (succ_const_iterator SI = succ_begin(BI), SE = succ_end(BI);
528 printEdgeProbability(OS << " ", BI, *SI);
533 uint32_t BranchProbabilityInfo::getSumForBlock(const BasicBlock *BB) const {
536 for (succ_const_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
537 uint32_t Weight = getEdgeWeight(BB, I.getSuccessorIndex());
538 uint32_t PrevSum = Sum;
541 assert(Sum > PrevSum); (void) PrevSum;
547 bool BranchProbabilityInfo::
548 isEdgeHot(const BasicBlock *Src, const BasicBlock *Dst) const {
549 // Hot probability is at least 4/5 = 80%
550 // FIXME: Compare against a static "hot" BranchProbability.
551 return getEdgeProbability(Src, Dst) > BranchProbability(4, 5);
554 BasicBlock *BranchProbabilityInfo::getHotSucc(BasicBlock *BB) const {
556 uint32_t MaxWeight = 0;
557 BasicBlock *MaxSucc = 0;
559 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
560 BasicBlock *Succ = *I;
561 uint32_t Weight = getEdgeWeight(BB, Succ);
562 uint32_t PrevSum = Sum;
565 assert(Sum > PrevSum); (void) PrevSum;
567 if (Weight > MaxWeight) {
573 // Hot probability is at least 4/5 = 80%
574 if (BranchProbability(MaxWeight, Sum) > BranchProbability(4, 5))
580 /// Get the raw edge weight for the edge. If can't find it, return
581 /// DEFAULT_WEIGHT value. Here an edge is specified using PredBlock and an index
582 /// to the successors.
583 uint32_t BranchProbabilityInfo::
584 getEdgeWeight(const BasicBlock *Src, unsigned IndexInSuccessors) const {
585 DenseMap<Edge, uint32_t>::const_iterator I =
586 Weights.find(std::make_pair(Src, IndexInSuccessors));
588 if (I != Weights.end())
591 return DEFAULT_WEIGHT;
594 /// Get the raw edge weight calculated for the block pair. This returns the sum
595 /// of all raw edge weights from Src to Dst.
596 uint32_t BranchProbabilityInfo::
597 getEdgeWeight(const BasicBlock *Src, const BasicBlock *Dst) const {
599 DenseMap<Edge, uint32_t>::const_iterator MapI;
600 for (succ_const_iterator I = succ_begin(Src), E = succ_end(Src); I != E; ++I)
602 MapI = Weights.find(std::make_pair(Src, I.getSuccessorIndex()));
603 if (MapI != Weights.end())
604 Weight += MapI->second;
606 return (Weight == 0) ? DEFAULT_WEIGHT : Weight;
609 /// Set the edge weight for a given edge specified by PredBlock and an index
610 /// to the successors.
611 void BranchProbabilityInfo::
612 setEdgeWeight(const BasicBlock *Src, unsigned IndexInSuccessors,
614 Weights[std::make_pair(Src, IndexInSuccessors)] = Weight;
615 DEBUG(dbgs() << "set edge " << Src->getName() << " -> "
616 << IndexInSuccessors << " successor weight to "
620 /// Get an edge's probability, relative to other out-edges from Src.
621 BranchProbability BranchProbabilityInfo::
622 getEdgeProbability(const BasicBlock *Src, unsigned IndexInSuccessors) const {
623 uint32_t N = getEdgeWeight(Src, IndexInSuccessors);
624 uint32_t D = getSumForBlock(Src);
626 return BranchProbability(N, D);
629 /// Get the probability of going from Src to Dst. It returns the sum of all
630 /// probabilities for edges from Src to Dst.
631 BranchProbability BranchProbabilityInfo::
632 getEdgeProbability(const BasicBlock *Src, const BasicBlock *Dst) const {
634 uint32_t N = getEdgeWeight(Src, Dst);
635 uint32_t D = getSumForBlock(Src);
637 return BranchProbability(N, D);
641 BranchProbabilityInfo::printEdgeProbability(raw_ostream &OS,
642 const BasicBlock *Src,
643 const BasicBlock *Dst) const {
645 const BranchProbability Prob = getEdgeProbability(Src, Dst);
646 OS << "edge " << Src->getName() << " -> " << Dst->getName()
647 << " probability is " << Prob
648 << (isEdgeHot(Src, Dst) ? " [HOT edge]\n" : "\n");