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/CFG.h"
18 #include "llvm/IR/Constants.h"
19 #include "llvm/IR/Function.h"
20 #include "llvm/IR/Instructions.h"
21 #include "llvm/IR/LLVMContext.h"
22 #include "llvm/IR/Metadata.h"
23 #include "llvm/Support/Debug.h"
24 #include "llvm/Support/raw_ostream.h"
28 #define DEBUG_TYPE "branch-prob"
30 INITIALIZE_PASS_BEGIN(BranchProbabilityInfo, "branch-prob",
31 "Branch Probability Analysis", false, true)
32 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
33 INITIALIZE_PASS_END(BranchProbabilityInfo, "branch-prob",
34 "Branch Probability Analysis", false, true)
36 char BranchProbabilityInfo::ID = 0;
38 // Weights are for internal use only. They are used by heuristics to help to
39 // estimate edges' probability. Example:
41 // Using "Loop Branch Heuristics" we predict weights of edges for the
56 // Probability of the edge BB2->BB1 = 124 / (124 + 4) = 0.96875
57 // Probability of the edge BB2->BB3 = 4 / (124 + 4) = 0.03125
58 static const uint32_t LBH_TAKEN_WEIGHT = 124;
59 static const uint32_t LBH_NONTAKEN_WEIGHT = 4;
61 /// \brief Unreachable-terminating branch taken weight.
63 /// This is the weight for a branch being taken to a block that terminates
64 /// (eventually) in unreachable. These are predicted as unlikely as possible.
65 static const uint32_t UR_TAKEN_WEIGHT = 1;
67 /// \brief Unreachable-terminating branch not-taken weight.
69 /// This is the weight for a branch not being taken toward a block that
70 /// terminates (eventually) in unreachable. Such a branch is essentially never
71 /// taken. Set the weight to an absurdly high value so that nested loops don't
72 /// easily subsume it.
73 static const uint32_t UR_NONTAKEN_WEIGHT = 1024*1024 - 1;
75 /// \brief Weight for a branch taken going into a cold block.
77 /// This is the weight for a branch taken toward a block marked
78 /// cold. A block is marked cold if it's postdominated by a
79 /// block containing a call to a cold function. Cold functions
80 /// are those marked with attribute 'cold'.
81 static const uint32_t CC_TAKEN_WEIGHT = 4;
83 /// \brief Weight for a branch not-taken into a cold block.
85 /// This is the weight for a branch not taken toward a block marked
87 static const uint32_t CC_NONTAKEN_WEIGHT = 64;
89 static const uint32_t PH_TAKEN_WEIGHT = 20;
90 static const uint32_t PH_NONTAKEN_WEIGHT = 12;
92 static const uint32_t ZH_TAKEN_WEIGHT = 20;
93 static const uint32_t ZH_NONTAKEN_WEIGHT = 12;
95 static const uint32_t FPH_TAKEN_WEIGHT = 20;
96 static const uint32_t FPH_NONTAKEN_WEIGHT = 12;
98 /// \brief Invoke-terminating normal branch taken weight
100 /// This is the weight for branching to the normal destination of an invoke
101 /// instruction. We expect this to happen most of the time. Set the weight to an
102 /// absurdly high value so that nested loops subsume it.
103 static const uint32_t IH_TAKEN_WEIGHT = 1024 * 1024 - 1;
105 /// \brief Invoke-terminating normal branch not-taken weight.
107 /// This is the weight for branching to the unwind destination of an invoke
108 /// instruction. This is essentially never taken.
109 static const uint32_t IH_NONTAKEN_WEIGHT = 1;
111 // Standard weight value. Used when none of the heuristics set weight for
113 static const uint32_t NORMAL_WEIGHT = 16;
115 // Minimum weight of an edge. Please note, that weight is NEVER 0.
116 static const uint32_t MIN_WEIGHT = 1;
118 static uint32_t getMaxWeightFor(BasicBlock *BB) {
119 return UINT32_MAX / BB->getTerminator()->getNumSuccessors();
123 /// \brief Calculate edge weights for successors lead to unreachable.
125 /// Predict that a successor which leads necessarily to an
126 /// unreachable-terminated block as extremely unlikely.
127 bool BranchProbabilityInfo::calcUnreachableHeuristics(BasicBlock *BB) {
128 TerminatorInst *TI = BB->getTerminator();
129 if (TI->getNumSuccessors() == 0) {
130 if (isa<UnreachableInst>(TI))
131 PostDominatedByUnreachable.insert(BB);
135 SmallVector<unsigned, 4> UnreachableEdges;
136 SmallVector<unsigned, 4> ReachableEdges;
138 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
139 if (PostDominatedByUnreachable.count(*I))
140 UnreachableEdges.push_back(I.getSuccessorIndex());
142 ReachableEdges.push_back(I.getSuccessorIndex());
145 // If all successors are in the set of blocks post-dominated by unreachable,
146 // this block is too.
147 if (UnreachableEdges.size() == TI->getNumSuccessors())
148 PostDominatedByUnreachable.insert(BB);
150 // Skip probabilities if this block has a single successor or if all were
152 if (TI->getNumSuccessors() == 1 || UnreachableEdges.empty())
155 uint32_t UnreachableWeight =
156 std::max(UR_TAKEN_WEIGHT / (unsigned)UnreachableEdges.size(), MIN_WEIGHT);
157 for (SmallVectorImpl<unsigned>::iterator I = UnreachableEdges.begin(),
158 E = UnreachableEdges.end();
160 setEdgeWeight(BB, *I, UnreachableWeight);
162 if (ReachableEdges.empty())
164 uint32_t ReachableWeight =
165 std::max(UR_NONTAKEN_WEIGHT / (unsigned)ReachableEdges.size(),
167 for (SmallVectorImpl<unsigned>::iterator I = ReachableEdges.begin(),
168 E = ReachableEdges.end();
170 setEdgeWeight(BB, *I, ReachableWeight);
175 // Propagate existing explicit probabilities from either profile data or
176 // 'expect' intrinsic processing.
177 bool BranchProbabilityInfo::calcMetadataWeights(BasicBlock *BB) {
178 TerminatorInst *TI = BB->getTerminator();
179 if (TI->getNumSuccessors() == 1)
181 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
184 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
188 // Ensure there are weights for all of the successors. Note that the first
189 // operand to the metadata node is a name, not a weight.
190 if (WeightsNode->getNumOperands() != TI->getNumSuccessors() + 1)
193 // Build up the final weights that will be used in a temporary buffer, but
194 // don't add them until all weihts are present. Each weight value is clamped
195 // to [1, getMaxWeightFor(BB)].
196 uint32_t WeightLimit = getMaxWeightFor(BB);
197 SmallVector<uint32_t, 2> Weights;
198 Weights.reserve(TI->getNumSuccessors());
199 for (unsigned i = 1, e = WeightsNode->getNumOperands(); i != e; ++i) {
200 ConstantInt *Weight =
201 mdconst::dyn_extract<ConstantInt>(WeightsNode->getOperand(i));
205 std::max<uint32_t>(1, Weight->getLimitedValue(WeightLimit)));
207 assert(Weights.size() == TI->getNumSuccessors() && "Checked above");
208 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
209 setEdgeWeight(BB, i, Weights[i]);
214 /// \brief Calculate edge weights for edges leading to cold blocks.
216 /// A cold block is one post-dominated by a block with a call to a
217 /// cold function. Those edges are unlikely to be taken, so we give
218 /// them relatively low weight.
220 /// Return true if we could compute the weights for cold edges.
221 /// Return false, otherwise.
222 bool BranchProbabilityInfo::calcColdCallHeuristics(BasicBlock *BB) {
223 TerminatorInst *TI = BB->getTerminator();
224 if (TI->getNumSuccessors() == 0)
227 // Determine which successors are post-dominated by a cold block.
228 SmallVector<unsigned, 4> ColdEdges;
229 SmallVector<unsigned, 4> NormalEdges;
230 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I)
231 if (PostDominatedByColdCall.count(*I))
232 ColdEdges.push_back(I.getSuccessorIndex());
234 NormalEdges.push_back(I.getSuccessorIndex());
236 // If all successors are in the set of blocks post-dominated by cold calls,
237 // this block is in the set post-dominated by cold calls.
238 if (ColdEdges.size() == TI->getNumSuccessors())
239 PostDominatedByColdCall.insert(BB);
241 // Otherwise, if the block itself contains a cold function, add it to the
242 // set of blocks postdominated by a cold call.
243 assert(!PostDominatedByColdCall.count(BB));
244 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
245 if (CallInst *CI = dyn_cast<CallInst>(I))
246 if (CI->hasFnAttr(Attribute::Cold)) {
247 PostDominatedByColdCall.insert(BB);
252 // Skip probabilities if this block has a single successor.
253 if (TI->getNumSuccessors() == 1 || ColdEdges.empty())
256 uint32_t ColdWeight =
257 std::max(CC_TAKEN_WEIGHT / (unsigned) ColdEdges.size(), MIN_WEIGHT);
258 for (SmallVectorImpl<unsigned>::iterator I = ColdEdges.begin(),
261 setEdgeWeight(BB, *I, ColdWeight);
263 if (NormalEdges.empty())
265 uint32_t NormalWeight = std::max(
266 CC_NONTAKEN_WEIGHT / (unsigned) NormalEdges.size(), NORMAL_WEIGHT);
267 for (SmallVectorImpl<unsigned>::iterator I = NormalEdges.begin(),
268 E = NormalEdges.end();
270 setEdgeWeight(BB, *I, NormalWeight);
275 // Calculate Edge Weights using "Pointer Heuristics". Predict a comparsion
276 // between two pointer or pointer and NULL will fail.
277 bool BranchProbabilityInfo::calcPointerHeuristics(BasicBlock *BB) {
278 BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
279 if (!BI || !BI->isConditional())
282 Value *Cond = BI->getCondition();
283 ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
284 if (!CI || !CI->isEquality())
287 Value *LHS = CI->getOperand(0);
289 if (!LHS->getType()->isPointerTy())
292 assert(CI->getOperand(1)->getType()->isPointerTy());
294 // p != 0 -> isProb = true
295 // p == 0 -> isProb = false
296 // p != q -> isProb = true
297 // p == q -> isProb = false;
298 unsigned TakenIdx = 0, NonTakenIdx = 1;
299 bool isProb = CI->getPredicate() == ICmpInst::ICMP_NE;
301 std::swap(TakenIdx, NonTakenIdx);
303 setEdgeWeight(BB, TakenIdx, PH_TAKEN_WEIGHT);
304 setEdgeWeight(BB, NonTakenIdx, PH_NONTAKEN_WEIGHT);
308 // Calculate Edge Weights using "Loop Branch Heuristics". Predict backedges
309 // as taken, exiting edges as not-taken.
310 bool BranchProbabilityInfo::calcLoopBranchHeuristics(BasicBlock *BB) {
311 Loop *L = LI->getLoopFor(BB);
315 SmallVector<unsigned, 8> BackEdges;
316 SmallVector<unsigned, 8> ExitingEdges;
317 SmallVector<unsigned, 8> InEdges; // Edges from header to the loop.
319 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
320 if (!L->contains(*I))
321 ExitingEdges.push_back(I.getSuccessorIndex());
322 else if (L->getHeader() == *I)
323 BackEdges.push_back(I.getSuccessorIndex());
325 InEdges.push_back(I.getSuccessorIndex());
328 if (BackEdges.empty() && ExitingEdges.empty())
331 if (uint32_t numBackEdges = BackEdges.size()) {
332 uint32_t backWeight = LBH_TAKEN_WEIGHT / numBackEdges;
333 if (backWeight < NORMAL_WEIGHT)
334 backWeight = NORMAL_WEIGHT;
336 for (SmallVectorImpl<unsigned>::iterator EI = BackEdges.begin(),
337 EE = BackEdges.end(); EI != EE; ++EI) {
338 setEdgeWeight(BB, *EI, backWeight);
342 if (uint32_t numInEdges = InEdges.size()) {
343 uint32_t inWeight = LBH_TAKEN_WEIGHT / numInEdges;
344 if (inWeight < NORMAL_WEIGHT)
345 inWeight = NORMAL_WEIGHT;
347 for (SmallVectorImpl<unsigned>::iterator EI = InEdges.begin(),
348 EE = InEdges.end(); EI != EE; ++EI) {
349 setEdgeWeight(BB, *EI, inWeight);
353 if (uint32_t numExitingEdges = ExitingEdges.size()) {
354 uint32_t exitWeight = LBH_NONTAKEN_WEIGHT / numExitingEdges;
355 if (exitWeight < MIN_WEIGHT)
356 exitWeight = MIN_WEIGHT;
358 for (SmallVectorImpl<unsigned>::iterator EI = ExitingEdges.begin(),
359 EE = ExitingEdges.end(); EI != EE; ++EI) {
360 setEdgeWeight(BB, *EI, exitWeight);
367 bool BranchProbabilityInfo::calcZeroHeuristics(BasicBlock *BB) {
368 BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
369 if (!BI || !BI->isConditional())
372 Value *Cond = BI->getCondition();
373 ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
377 Value *RHS = CI->getOperand(1);
378 ConstantInt *CV = dyn_cast<ConstantInt>(RHS);
384 switch (CI->getPredicate()) {
385 case CmpInst::ICMP_EQ:
386 // X == 0 -> Unlikely
389 case CmpInst::ICMP_NE:
393 case CmpInst::ICMP_SLT:
397 case CmpInst::ICMP_SGT:
404 } else if (CV->isOne() && CI->getPredicate() == CmpInst::ICMP_SLT) {
405 // InstCombine canonicalizes X <= 0 into X < 1.
406 // X <= 0 -> Unlikely
408 } else if (CV->isAllOnesValue()) {
409 switch (CI->getPredicate()) {
410 case CmpInst::ICMP_EQ:
411 // X == -1 -> Unlikely
414 case CmpInst::ICMP_NE:
418 case CmpInst::ICMP_SGT:
419 // InstCombine canonicalizes X >= 0 into X > -1.
430 unsigned TakenIdx = 0, NonTakenIdx = 1;
433 std::swap(TakenIdx, NonTakenIdx);
435 setEdgeWeight(BB, TakenIdx, ZH_TAKEN_WEIGHT);
436 setEdgeWeight(BB, NonTakenIdx, ZH_NONTAKEN_WEIGHT);
441 bool BranchProbabilityInfo::calcFloatingPointHeuristics(BasicBlock *BB) {
442 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
443 if (!BI || !BI->isConditional())
446 Value *Cond = BI->getCondition();
447 FCmpInst *FCmp = dyn_cast<FCmpInst>(Cond);
452 if (FCmp->isEquality()) {
453 // f1 == f2 -> Unlikely
454 // f1 != f2 -> Likely
455 isProb = !FCmp->isTrueWhenEqual();
456 } else if (FCmp->getPredicate() == FCmpInst::FCMP_ORD) {
459 } else if (FCmp->getPredicate() == FCmpInst::FCMP_UNO) {
466 unsigned TakenIdx = 0, NonTakenIdx = 1;
469 std::swap(TakenIdx, NonTakenIdx);
471 setEdgeWeight(BB, TakenIdx, FPH_TAKEN_WEIGHT);
472 setEdgeWeight(BB, NonTakenIdx, FPH_NONTAKEN_WEIGHT);
477 bool BranchProbabilityInfo::calcInvokeHeuristics(BasicBlock *BB) {
478 InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator());
482 setEdgeWeight(BB, 0/*Index for Normal*/, IH_TAKEN_WEIGHT);
483 setEdgeWeight(BB, 1/*Index for Unwind*/, IH_NONTAKEN_WEIGHT);
487 void BranchProbabilityInfo::getAnalysisUsage(AnalysisUsage &AU) const {
488 AU.addRequired<LoopInfoWrapperPass>();
489 AU.setPreservesAll();
492 bool BranchProbabilityInfo::runOnFunction(Function &F) {
493 DEBUG(dbgs() << "---- Branch Probability Info : " << F.getName()
495 LastF = &F; // Store the last function we ran on for printing.
496 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
497 assert(PostDominatedByUnreachable.empty());
498 assert(PostDominatedByColdCall.empty());
500 // Walk the basic blocks in post-order so that we can build up state about
501 // the successors of a block iteratively.
502 for (po_iterator<BasicBlock *> I = po_begin(&F.getEntryBlock()),
503 E = po_end(&F.getEntryBlock());
505 DEBUG(dbgs() << "Computing probabilities for " << I->getName() << "\n");
506 if (calcUnreachableHeuristics(*I))
508 if (calcMetadataWeights(*I))
510 if (calcColdCallHeuristics(*I))
512 if (calcLoopBranchHeuristics(*I))
514 if (calcPointerHeuristics(*I))
516 if (calcZeroHeuristics(*I))
518 if (calcFloatingPointHeuristics(*I))
520 calcInvokeHeuristics(*I);
523 PostDominatedByUnreachable.clear();
524 PostDominatedByColdCall.clear();
528 void BranchProbabilityInfo::print(raw_ostream &OS, const Module *) const {
529 OS << "---- Branch Probabilities ----\n";
530 // We print the probabilities from the last function the analysis ran over,
531 // or the function it is currently running over.
532 assert(LastF && "Cannot print prior to running over a function");
533 for (Function::const_iterator BI = LastF->begin(), BE = LastF->end();
535 for (succ_const_iterator SI = succ_begin(BI), SE = succ_end(BI);
537 printEdgeProbability(OS << " ", BI, *SI);
542 uint32_t BranchProbabilityInfo::getSumForBlock(const BasicBlock *BB) const {
545 for (succ_const_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
546 uint32_t Weight = getEdgeWeight(BB, I.getSuccessorIndex());
547 uint32_t PrevSum = Sum;
550 assert(Sum > PrevSum); (void) PrevSum;
556 bool BranchProbabilityInfo::
557 isEdgeHot(const BasicBlock *Src, const BasicBlock *Dst) const {
558 // Hot probability is at least 4/5 = 80%
559 // FIXME: Compare against a static "hot" BranchProbability.
560 return getEdgeProbability(Src, Dst) > BranchProbability(4, 5);
563 BasicBlock *BranchProbabilityInfo::getHotSucc(BasicBlock *BB) const {
565 uint32_t MaxWeight = 0;
566 BasicBlock *MaxSucc = nullptr;
568 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
569 BasicBlock *Succ = *I;
570 uint32_t Weight = getEdgeWeight(BB, Succ);
571 uint32_t PrevSum = Sum;
574 assert(Sum > PrevSum); (void) PrevSum;
576 if (Weight > MaxWeight) {
582 // Hot probability is at least 4/5 = 80%
583 if (BranchProbability(MaxWeight, Sum) > BranchProbability(4, 5))
589 /// Get the raw edge weight for the edge. If can't find it, return
590 /// DEFAULT_WEIGHT value. Here an edge is specified using PredBlock and an index
591 /// to the successors.
592 uint32_t BranchProbabilityInfo::
593 getEdgeWeight(const BasicBlock *Src, unsigned IndexInSuccessors) const {
594 DenseMap<Edge, uint32_t>::const_iterator I =
595 Weights.find(std::make_pair(Src, IndexInSuccessors));
597 if (I != Weights.end())
600 return DEFAULT_WEIGHT;
603 uint32_t BranchProbabilityInfo::getEdgeWeight(const BasicBlock *Src,
604 succ_const_iterator Dst) const {
605 return getEdgeWeight(Src, Dst.getSuccessorIndex());
608 /// Get the raw edge weight calculated for the block pair. This returns the sum
609 /// of all raw edge weights from Src to Dst.
610 uint32_t BranchProbabilityInfo::
611 getEdgeWeight(const BasicBlock *Src, const BasicBlock *Dst) const {
613 DenseMap<Edge, uint32_t>::const_iterator MapI;
614 for (succ_const_iterator I = succ_begin(Src), E = succ_end(Src); I != E; ++I)
616 MapI = Weights.find(std::make_pair(Src, I.getSuccessorIndex()));
617 if (MapI != Weights.end())
618 Weight += MapI->second;
620 return (Weight == 0) ? DEFAULT_WEIGHT : Weight;
623 /// Set the edge weight for a given edge specified by PredBlock and an index
624 /// to the successors.
625 void BranchProbabilityInfo::
626 setEdgeWeight(const BasicBlock *Src, unsigned IndexInSuccessors,
628 Weights[std::make_pair(Src, IndexInSuccessors)] = Weight;
629 DEBUG(dbgs() << "set edge " << Src->getName() << " -> "
630 << IndexInSuccessors << " successor weight to "
634 /// Get an edge's probability, relative to other out-edges from Src.
635 BranchProbability BranchProbabilityInfo::
636 getEdgeProbability(const BasicBlock *Src, unsigned IndexInSuccessors) const {
637 uint32_t N = getEdgeWeight(Src, IndexInSuccessors);
638 uint32_t D = getSumForBlock(Src);
640 return BranchProbability(N, D);
643 /// Get the probability of going from Src to Dst. It returns the sum of all
644 /// probabilities for edges from Src to Dst.
645 BranchProbability BranchProbabilityInfo::
646 getEdgeProbability(const BasicBlock *Src, const BasicBlock *Dst) const {
648 uint32_t N = getEdgeWeight(Src, Dst);
649 uint32_t D = getSumForBlock(Src);
651 return BranchProbability(N, D);
655 BranchProbabilityInfo::printEdgeProbability(raw_ostream &OS,
656 const BasicBlock *Src,
657 const BasicBlock *Dst) const {
659 const BranchProbability Prob = getEdgeProbability(Src, Dst);
660 OS << "edge " << Src->getName() << " -> " << Dst->getName()
661 << " probability is " << Prob
662 << (isEdgeHot(Src, Dst) ? " [HOT edge]\n" : "\n");