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 /// \brief Calculate edge weights for successors lead to unreachable.
120 /// Predict that a successor which leads necessarily to an
121 /// unreachable-terminated block as extremely unlikely.
122 bool BranchProbabilityInfo::calcUnreachableHeuristics(BasicBlock *BB) {
123 TerminatorInst *TI = BB->getTerminator();
124 if (TI->getNumSuccessors() == 0) {
125 if (isa<UnreachableInst>(TI))
126 PostDominatedByUnreachable.insert(BB);
130 SmallVector<unsigned, 4> UnreachableEdges;
131 SmallVector<unsigned, 4> ReachableEdges;
133 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
134 if (PostDominatedByUnreachable.count(*I))
135 UnreachableEdges.push_back(I.getSuccessorIndex());
137 ReachableEdges.push_back(I.getSuccessorIndex());
140 // If all successors are in the set of blocks post-dominated by unreachable,
141 // this block is too.
142 if (UnreachableEdges.size() == TI->getNumSuccessors())
143 PostDominatedByUnreachable.insert(BB);
145 // Skip probabilities if this block has a single successor or if all were
147 if (TI->getNumSuccessors() == 1 || UnreachableEdges.empty())
150 uint32_t UnreachableWeight =
151 std::max(UR_TAKEN_WEIGHT / (unsigned)UnreachableEdges.size(), MIN_WEIGHT);
152 for (SmallVectorImpl<unsigned>::iterator I = UnreachableEdges.begin(),
153 E = UnreachableEdges.end();
155 setEdgeWeight(BB, *I, UnreachableWeight);
157 if (ReachableEdges.empty())
159 uint32_t ReachableWeight =
160 std::max(UR_NONTAKEN_WEIGHT / (unsigned)ReachableEdges.size(),
162 for (SmallVectorImpl<unsigned>::iterator I = ReachableEdges.begin(),
163 E = ReachableEdges.end();
165 setEdgeWeight(BB, *I, ReachableWeight);
170 // Propagate existing explicit probabilities from either profile data or
171 // 'expect' intrinsic processing.
172 bool BranchProbabilityInfo::calcMetadataWeights(BasicBlock *BB) {
173 TerminatorInst *TI = BB->getTerminator();
174 if (TI->getNumSuccessors() == 1)
176 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
179 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
183 // Check that the number of successors is manageable.
184 assert(TI->getNumSuccessors() < UINT32_MAX && "Too many successors");
186 // Ensure there are weights for all of the successors. Note that the first
187 // operand to the metadata node is a name, not a weight.
188 if (WeightsNode->getNumOperands() != TI->getNumSuccessors() + 1)
191 // Build up the final weights that will be used in a temporary buffer.
192 // Compute the sum of all weights to later decide whether they need to
193 // be scaled to fit in 32 bits.
194 uint64_t WeightSum = 0;
195 SmallVector<uint32_t, 2> Weights;
196 Weights.reserve(TI->getNumSuccessors());
197 for (unsigned i = 1, e = WeightsNode->getNumOperands(); i != e; ++i) {
198 ConstantInt *Weight =
199 mdconst::dyn_extract<ConstantInt>(WeightsNode->getOperand(i));
202 assert(Weight->getValue().getActiveBits() <= 32 &&
203 "Too many bits for uint32_t");
204 Weights.push_back(Weight->getZExtValue());
205 WeightSum += Weights.back();
207 assert(Weights.size() == TI->getNumSuccessors() && "Checked above");
209 // If the sum of weights does not fit in 32 bits, scale every weight down
211 uint64_t ScalingFactor =
212 (WeightSum > UINT32_MAX) ? WeightSum / UINT32_MAX + 1 : 1;
215 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
216 uint32_t W = Weights[i] / ScalingFactor;
218 setEdgeWeight(BB, i, W);
220 assert(WeightSum <= UINT32_MAX &&
221 "Expected weights to scale down to 32 bits");
226 /// \brief Calculate edge weights for edges leading to cold blocks.
228 /// A cold block is one post-dominated by a block with a call to a
229 /// cold function. Those edges are unlikely to be taken, so we give
230 /// them relatively low weight.
232 /// Return true if we could compute the weights for cold edges.
233 /// Return false, otherwise.
234 bool BranchProbabilityInfo::calcColdCallHeuristics(BasicBlock *BB) {
235 TerminatorInst *TI = BB->getTerminator();
236 if (TI->getNumSuccessors() == 0)
239 // Determine which successors are post-dominated by a cold block.
240 SmallVector<unsigned, 4> ColdEdges;
241 SmallVector<unsigned, 4> NormalEdges;
242 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I)
243 if (PostDominatedByColdCall.count(*I))
244 ColdEdges.push_back(I.getSuccessorIndex());
246 NormalEdges.push_back(I.getSuccessorIndex());
248 // If all successors are in the set of blocks post-dominated by cold calls,
249 // this block is in the set post-dominated by cold calls.
250 if (ColdEdges.size() == TI->getNumSuccessors())
251 PostDominatedByColdCall.insert(BB);
253 // Otherwise, if the block itself contains a cold function, add it to the
254 // set of blocks postdominated by a cold call.
255 assert(!PostDominatedByColdCall.count(BB));
256 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
257 if (CallInst *CI = dyn_cast<CallInst>(I))
258 if (CI->hasFnAttr(Attribute::Cold)) {
259 PostDominatedByColdCall.insert(BB);
264 // Skip probabilities if this block has a single successor.
265 if (TI->getNumSuccessors() == 1 || ColdEdges.empty())
268 uint32_t ColdWeight =
269 std::max(CC_TAKEN_WEIGHT / (unsigned) ColdEdges.size(), MIN_WEIGHT);
270 for (SmallVectorImpl<unsigned>::iterator I = ColdEdges.begin(),
273 setEdgeWeight(BB, *I, ColdWeight);
275 if (NormalEdges.empty())
277 uint32_t NormalWeight = std::max(
278 CC_NONTAKEN_WEIGHT / (unsigned) NormalEdges.size(), NORMAL_WEIGHT);
279 for (SmallVectorImpl<unsigned>::iterator I = NormalEdges.begin(),
280 E = NormalEdges.end();
282 setEdgeWeight(BB, *I, NormalWeight);
287 // Calculate Edge Weights using "Pointer Heuristics". Predict a comparsion
288 // between two pointer or pointer and NULL will fail.
289 bool BranchProbabilityInfo::calcPointerHeuristics(BasicBlock *BB) {
290 BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
291 if (!BI || !BI->isConditional())
294 Value *Cond = BI->getCondition();
295 ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
296 if (!CI || !CI->isEquality())
299 Value *LHS = CI->getOperand(0);
301 if (!LHS->getType()->isPointerTy())
304 assert(CI->getOperand(1)->getType()->isPointerTy());
306 // p != 0 -> isProb = true
307 // p == 0 -> isProb = false
308 // p != q -> isProb = true
309 // p == q -> isProb = false;
310 unsigned TakenIdx = 0, NonTakenIdx = 1;
311 bool isProb = CI->getPredicate() == ICmpInst::ICMP_NE;
313 std::swap(TakenIdx, NonTakenIdx);
315 setEdgeWeight(BB, TakenIdx, PH_TAKEN_WEIGHT);
316 setEdgeWeight(BB, NonTakenIdx, PH_NONTAKEN_WEIGHT);
320 // Calculate Edge Weights using "Loop Branch Heuristics". Predict backedges
321 // as taken, exiting edges as not-taken.
322 bool BranchProbabilityInfo::calcLoopBranchHeuristics(BasicBlock *BB) {
323 Loop *L = LI->getLoopFor(BB);
327 SmallVector<unsigned, 8> BackEdges;
328 SmallVector<unsigned, 8> ExitingEdges;
329 SmallVector<unsigned, 8> InEdges; // Edges from header to the loop.
331 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
332 if (!L->contains(*I))
333 ExitingEdges.push_back(I.getSuccessorIndex());
334 else if (L->getHeader() == *I)
335 BackEdges.push_back(I.getSuccessorIndex());
337 InEdges.push_back(I.getSuccessorIndex());
340 if (BackEdges.empty() && ExitingEdges.empty())
343 if (uint32_t numBackEdges = BackEdges.size()) {
344 uint32_t backWeight = LBH_TAKEN_WEIGHT / numBackEdges;
345 if (backWeight < NORMAL_WEIGHT)
346 backWeight = NORMAL_WEIGHT;
348 for (SmallVectorImpl<unsigned>::iterator EI = BackEdges.begin(),
349 EE = BackEdges.end(); EI != EE; ++EI) {
350 setEdgeWeight(BB, *EI, backWeight);
354 if (uint32_t numInEdges = InEdges.size()) {
355 uint32_t inWeight = LBH_TAKEN_WEIGHT / numInEdges;
356 if (inWeight < NORMAL_WEIGHT)
357 inWeight = NORMAL_WEIGHT;
359 for (SmallVectorImpl<unsigned>::iterator EI = InEdges.begin(),
360 EE = InEdges.end(); EI != EE; ++EI) {
361 setEdgeWeight(BB, *EI, inWeight);
365 if (uint32_t numExitingEdges = ExitingEdges.size()) {
366 uint32_t exitWeight = LBH_NONTAKEN_WEIGHT / numExitingEdges;
367 if (exitWeight < MIN_WEIGHT)
368 exitWeight = MIN_WEIGHT;
370 for (SmallVectorImpl<unsigned>::iterator EI = ExitingEdges.begin(),
371 EE = ExitingEdges.end(); EI != EE; ++EI) {
372 setEdgeWeight(BB, *EI, exitWeight);
379 bool BranchProbabilityInfo::calcZeroHeuristics(BasicBlock *BB) {
380 BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
381 if (!BI || !BI->isConditional())
384 Value *Cond = BI->getCondition();
385 ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
389 Value *RHS = CI->getOperand(1);
390 ConstantInt *CV = dyn_cast<ConstantInt>(RHS);
394 // If the LHS is the result of AND'ing a value with a single bit bitmask,
395 // we don't have information about probabilities.
396 if (Instruction *LHS = dyn_cast<Instruction>(CI->getOperand(0)))
397 if (LHS->getOpcode() == Instruction::And)
398 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(LHS->getOperand(1)))
399 if (AndRHS->getUniqueInteger().isPowerOf2())
404 switch (CI->getPredicate()) {
405 case CmpInst::ICMP_EQ:
406 // X == 0 -> Unlikely
409 case CmpInst::ICMP_NE:
413 case CmpInst::ICMP_SLT:
417 case CmpInst::ICMP_SGT:
424 } else if (CV->isOne() && CI->getPredicate() == CmpInst::ICMP_SLT) {
425 // InstCombine canonicalizes X <= 0 into X < 1.
426 // X <= 0 -> Unlikely
428 } else if (CV->isAllOnesValue()) {
429 switch (CI->getPredicate()) {
430 case CmpInst::ICMP_EQ:
431 // X == -1 -> Unlikely
434 case CmpInst::ICMP_NE:
438 case CmpInst::ICMP_SGT:
439 // InstCombine canonicalizes X >= 0 into X > -1.
450 unsigned TakenIdx = 0, NonTakenIdx = 1;
453 std::swap(TakenIdx, NonTakenIdx);
455 setEdgeWeight(BB, TakenIdx, ZH_TAKEN_WEIGHT);
456 setEdgeWeight(BB, NonTakenIdx, ZH_NONTAKEN_WEIGHT);
461 bool BranchProbabilityInfo::calcFloatingPointHeuristics(BasicBlock *BB) {
462 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
463 if (!BI || !BI->isConditional())
466 Value *Cond = BI->getCondition();
467 FCmpInst *FCmp = dyn_cast<FCmpInst>(Cond);
472 if (FCmp->isEquality()) {
473 // f1 == f2 -> Unlikely
474 // f1 != f2 -> Likely
475 isProb = !FCmp->isTrueWhenEqual();
476 } else if (FCmp->getPredicate() == FCmpInst::FCMP_ORD) {
479 } else if (FCmp->getPredicate() == FCmpInst::FCMP_UNO) {
486 unsigned TakenIdx = 0, NonTakenIdx = 1;
489 std::swap(TakenIdx, NonTakenIdx);
491 setEdgeWeight(BB, TakenIdx, FPH_TAKEN_WEIGHT);
492 setEdgeWeight(BB, NonTakenIdx, FPH_NONTAKEN_WEIGHT);
497 bool BranchProbabilityInfo::calcInvokeHeuristics(BasicBlock *BB) {
498 InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator());
502 setEdgeWeight(BB, 0/*Index for Normal*/, IH_TAKEN_WEIGHT);
503 setEdgeWeight(BB, 1/*Index for Unwind*/, IH_NONTAKEN_WEIGHT);
507 void BranchProbabilityInfo::getAnalysisUsage(AnalysisUsage &AU) const {
508 AU.addRequired<LoopInfoWrapperPass>();
509 AU.setPreservesAll();
512 bool BranchProbabilityInfo::runOnFunction(Function &F) {
513 DEBUG(dbgs() << "---- Branch Probability Info : " << F.getName()
515 LastF = &F; // Store the last function we ran on for printing.
516 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
517 assert(PostDominatedByUnreachable.empty());
518 assert(PostDominatedByColdCall.empty());
520 // Walk the basic blocks in post-order so that we can build up state about
521 // the successors of a block iteratively.
522 for (auto BB : post_order(&F.getEntryBlock())) {
523 DEBUG(dbgs() << "Computing probabilities for " << BB->getName() << "\n");
524 if (calcUnreachableHeuristics(BB))
526 if (calcMetadataWeights(BB))
528 if (calcColdCallHeuristics(BB))
530 if (calcLoopBranchHeuristics(BB))
532 if (calcPointerHeuristics(BB))
534 if (calcZeroHeuristics(BB))
536 if (calcFloatingPointHeuristics(BB))
538 calcInvokeHeuristics(BB);
541 PostDominatedByUnreachable.clear();
542 PostDominatedByColdCall.clear();
546 void BranchProbabilityInfo::releaseMemory() {
550 void BranchProbabilityInfo::print(raw_ostream &OS, const Module *) const {
551 OS << "---- Branch Probabilities ----\n";
552 // We print the probabilities from the last function the analysis ran over,
553 // or the function it is currently running over.
554 assert(LastF && "Cannot print prior to running over a function");
555 for (Function::const_iterator BI = LastF->begin(), BE = LastF->end();
557 for (succ_const_iterator SI = succ_begin(BI), SE = succ_end(BI);
559 printEdgeProbability(OS << " ", BI, *SI);
564 uint32_t BranchProbabilityInfo::getSumForBlock(const BasicBlock *BB) const {
567 for (succ_const_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
568 uint32_t Weight = getEdgeWeight(BB, I.getSuccessorIndex());
569 uint32_t PrevSum = Sum;
572 assert(Sum >= PrevSum); (void) PrevSum;
578 bool BranchProbabilityInfo::
579 isEdgeHot(const BasicBlock *Src, const BasicBlock *Dst) const {
580 // Hot probability is at least 4/5 = 80%
581 // FIXME: Compare against a static "hot" BranchProbability.
582 return getEdgeProbability(Src, Dst) > BranchProbability(4, 5);
585 BasicBlock *BranchProbabilityInfo::getHotSucc(BasicBlock *BB) const {
587 uint32_t MaxWeight = 0;
588 BasicBlock *MaxSucc = nullptr;
590 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
591 BasicBlock *Succ = *I;
592 uint32_t Weight = getEdgeWeight(BB, Succ);
593 uint32_t PrevSum = Sum;
596 assert(Sum > PrevSum); (void) PrevSum;
598 if (Weight > MaxWeight) {
604 // Hot probability is at least 4/5 = 80%
605 if (BranchProbability(MaxWeight, Sum) > BranchProbability(4, 5))
611 /// Get the raw edge weight for the edge. If can't find it, return
612 /// DEFAULT_WEIGHT value. Here an edge is specified using PredBlock and an index
613 /// to the successors.
614 uint32_t BranchProbabilityInfo::
615 getEdgeWeight(const BasicBlock *Src, unsigned IndexInSuccessors) const {
616 DenseMap<Edge, uint32_t>::const_iterator I =
617 Weights.find(std::make_pair(Src, IndexInSuccessors));
619 if (I != Weights.end())
622 return DEFAULT_WEIGHT;
625 uint32_t BranchProbabilityInfo::getEdgeWeight(const BasicBlock *Src,
626 succ_const_iterator Dst) const {
627 return getEdgeWeight(Src, Dst.getSuccessorIndex());
630 /// Get the raw edge weight calculated for the block pair. This returns the sum
631 /// of all raw edge weights from Src to Dst.
632 uint32_t BranchProbabilityInfo::
633 getEdgeWeight(const BasicBlock *Src, const BasicBlock *Dst) const {
635 bool FoundWeight = false;
636 DenseMap<Edge, uint32_t>::const_iterator MapI;
637 for (succ_const_iterator I = succ_begin(Src), E = succ_end(Src); I != E; ++I)
639 MapI = Weights.find(std::make_pair(Src, I.getSuccessorIndex()));
640 if (MapI != Weights.end()) {
642 Weight += MapI->second;
645 return (!FoundWeight) ? DEFAULT_WEIGHT : Weight;
648 /// Set the edge weight for a given edge specified by PredBlock and an index
649 /// to the successors.
650 void BranchProbabilityInfo::
651 setEdgeWeight(const BasicBlock *Src, unsigned IndexInSuccessors,
653 Weights[std::make_pair(Src, IndexInSuccessors)] = Weight;
654 DEBUG(dbgs() << "set edge " << Src->getName() << " -> "
655 << IndexInSuccessors << " successor weight to "
659 /// Get an edge's probability, relative to other out-edges from Src.
660 BranchProbability BranchProbabilityInfo::
661 getEdgeProbability(const BasicBlock *Src, unsigned IndexInSuccessors) const {
662 uint32_t N = getEdgeWeight(Src, IndexInSuccessors);
663 uint32_t D = getSumForBlock(Src);
665 return BranchProbability(N, D);
668 /// Get the probability of going from Src to Dst. It returns the sum of all
669 /// probabilities for edges from Src to Dst.
670 BranchProbability BranchProbabilityInfo::
671 getEdgeProbability(const BasicBlock *Src, const BasicBlock *Dst) const {
673 uint32_t N = getEdgeWeight(Src, Dst);
674 uint32_t D = getSumForBlock(Src);
676 return BranchProbability(N, D);
680 BranchProbabilityInfo::printEdgeProbability(raw_ostream &OS,
681 const BasicBlock *Src,
682 const BasicBlock *Dst) const {
684 const BranchProbability Prob = getEdgeProbability(Src, Dst);
685 OS << "edge " << Src->getName() << " -> " << Dst->getName()
686 << " probability is " << Prob
687 << (isEdgeHot(Src, Dst) ? " [HOT edge]\n" : "\n");