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 weights 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));
204 Weights.push_back(Weight->getLimitedValue(WeightLimit));
206 assert(Weights.size() == TI->getNumSuccessors() && "Checked above");
207 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
208 setEdgeWeight(BB, i, Weights[i]);
213 /// \brief Calculate edge weights for edges leading to cold blocks.
215 /// A cold block is one post-dominated by a block with a call to a
216 /// cold function. Those edges are unlikely to be taken, so we give
217 /// them relatively low weight.
219 /// Return true if we could compute the weights for cold edges.
220 /// Return false, otherwise.
221 bool BranchProbabilityInfo::calcColdCallHeuristics(BasicBlock *BB) {
222 TerminatorInst *TI = BB->getTerminator();
223 if (TI->getNumSuccessors() == 0)
226 // Determine which successors are post-dominated by a cold block.
227 SmallVector<unsigned, 4> ColdEdges;
228 SmallVector<unsigned, 4> NormalEdges;
229 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I)
230 if (PostDominatedByColdCall.count(*I))
231 ColdEdges.push_back(I.getSuccessorIndex());
233 NormalEdges.push_back(I.getSuccessorIndex());
235 // If all successors are in the set of blocks post-dominated by cold calls,
236 // this block is in the set post-dominated by cold calls.
237 if (ColdEdges.size() == TI->getNumSuccessors())
238 PostDominatedByColdCall.insert(BB);
240 // Otherwise, if the block itself contains a cold function, add it to the
241 // set of blocks postdominated by a cold call.
242 assert(!PostDominatedByColdCall.count(BB));
243 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
244 if (CallInst *CI = dyn_cast<CallInst>(I))
245 if (CI->hasFnAttr(Attribute::Cold)) {
246 PostDominatedByColdCall.insert(BB);
251 // Skip probabilities if this block has a single successor.
252 if (TI->getNumSuccessors() == 1 || ColdEdges.empty())
255 uint32_t ColdWeight =
256 std::max(CC_TAKEN_WEIGHT / (unsigned) ColdEdges.size(), MIN_WEIGHT);
257 for (SmallVectorImpl<unsigned>::iterator I = ColdEdges.begin(),
260 setEdgeWeight(BB, *I, ColdWeight);
262 if (NormalEdges.empty())
264 uint32_t NormalWeight = std::max(
265 CC_NONTAKEN_WEIGHT / (unsigned) NormalEdges.size(), NORMAL_WEIGHT);
266 for (SmallVectorImpl<unsigned>::iterator I = NormalEdges.begin(),
267 E = NormalEdges.end();
269 setEdgeWeight(BB, *I, NormalWeight);
274 // Calculate Edge Weights using "Pointer Heuristics". Predict a comparsion
275 // between two pointer or pointer and NULL will fail.
276 bool BranchProbabilityInfo::calcPointerHeuristics(BasicBlock *BB) {
277 BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
278 if (!BI || !BI->isConditional())
281 Value *Cond = BI->getCondition();
282 ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
283 if (!CI || !CI->isEquality())
286 Value *LHS = CI->getOperand(0);
288 if (!LHS->getType()->isPointerTy())
291 assert(CI->getOperand(1)->getType()->isPointerTy());
293 // p != 0 -> isProb = true
294 // p == 0 -> isProb = false
295 // p != q -> isProb = true
296 // p == q -> isProb = false;
297 unsigned TakenIdx = 0, NonTakenIdx = 1;
298 bool isProb = CI->getPredicate() == ICmpInst::ICMP_NE;
300 std::swap(TakenIdx, NonTakenIdx);
302 setEdgeWeight(BB, TakenIdx, PH_TAKEN_WEIGHT);
303 setEdgeWeight(BB, NonTakenIdx, PH_NONTAKEN_WEIGHT);
307 // Calculate Edge Weights using "Loop Branch Heuristics". Predict backedges
308 // as taken, exiting edges as not-taken.
309 bool BranchProbabilityInfo::calcLoopBranchHeuristics(BasicBlock *BB) {
310 Loop *L = LI->getLoopFor(BB);
314 SmallVector<unsigned, 8> BackEdges;
315 SmallVector<unsigned, 8> ExitingEdges;
316 SmallVector<unsigned, 8> InEdges; // Edges from header to the loop.
318 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
319 if (!L->contains(*I))
320 ExitingEdges.push_back(I.getSuccessorIndex());
321 else if (L->getHeader() == *I)
322 BackEdges.push_back(I.getSuccessorIndex());
324 InEdges.push_back(I.getSuccessorIndex());
327 if (BackEdges.empty() && ExitingEdges.empty())
330 if (uint32_t numBackEdges = BackEdges.size()) {
331 uint32_t backWeight = LBH_TAKEN_WEIGHT / numBackEdges;
332 if (backWeight < NORMAL_WEIGHT)
333 backWeight = NORMAL_WEIGHT;
335 for (SmallVectorImpl<unsigned>::iterator EI = BackEdges.begin(),
336 EE = BackEdges.end(); EI != EE; ++EI) {
337 setEdgeWeight(BB, *EI, backWeight);
341 if (uint32_t numInEdges = InEdges.size()) {
342 uint32_t inWeight = LBH_TAKEN_WEIGHT / numInEdges;
343 if (inWeight < NORMAL_WEIGHT)
344 inWeight = NORMAL_WEIGHT;
346 for (SmallVectorImpl<unsigned>::iterator EI = InEdges.begin(),
347 EE = InEdges.end(); EI != EE; ++EI) {
348 setEdgeWeight(BB, *EI, inWeight);
352 if (uint32_t numExitingEdges = ExitingEdges.size()) {
353 uint32_t exitWeight = LBH_NONTAKEN_WEIGHT / numExitingEdges;
354 if (exitWeight < MIN_WEIGHT)
355 exitWeight = MIN_WEIGHT;
357 for (SmallVectorImpl<unsigned>::iterator EI = ExitingEdges.begin(),
358 EE = ExitingEdges.end(); EI != EE; ++EI) {
359 setEdgeWeight(BB, *EI, exitWeight);
366 bool BranchProbabilityInfo::calcZeroHeuristics(BasicBlock *BB) {
367 BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
368 if (!BI || !BI->isConditional())
371 Value *Cond = BI->getCondition();
372 ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
376 Value *RHS = CI->getOperand(1);
377 ConstantInt *CV = dyn_cast<ConstantInt>(RHS);
381 // If the LHS is the result of AND'ing a value with a single bit bitmask,
382 // we don't have information about probabilities.
383 if (Instruction *LHS = dyn_cast<Instruction>(CI->getOperand(0)))
384 if (LHS->getOpcode() == Instruction::And)
385 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(LHS->getOperand(1)))
386 if (AndRHS->getUniqueInteger().isPowerOf2())
391 switch (CI->getPredicate()) {
392 case CmpInst::ICMP_EQ:
393 // X == 0 -> Unlikely
396 case CmpInst::ICMP_NE:
400 case CmpInst::ICMP_SLT:
404 case CmpInst::ICMP_SGT:
411 } else if (CV->isOne() && CI->getPredicate() == CmpInst::ICMP_SLT) {
412 // InstCombine canonicalizes X <= 0 into X < 1.
413 // X <= 0 -> Unlikely
415 } else if (CV->isAllOnesValue()) {
416 switch (CI->getPredicate()) {
417 case CmpInst::ICMP_EQ:
418 // X == -1 -> Unlikely
421 case CmpInst::ICMP_NE:
425 case CmpInst::ICMP_SGT:
426 // InstCombine canonicalizes X >= 0 into X > -1.
437 unsigned TakenIdx = 0, NonTakenIdx = 1;
440 std::swap(TakenIdx, NonTakenIdx);
442 setEdgeWeight(BB, TakenIdx, ZH_TAKEN_WEIGHT);
443 setEdgeWeight(BB, NonTakenIdx, ZH_NONTAKEN_WEIGHT);
448 bool BranchProbabilityInfo::calcFloatingPointHeuristics(BasicBlock *BB) {
449 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
450 if (!BI || !BI->isConditional())
453 Value *Cond = BI->getCondition();
454 FCmpInst *FCmp = dyn_cast<FCmpInst>(Cond);
459 if (FCmp->isEquality()) {
460 // f1 == f2 -> Unlikely
461 // f1 != f2 -> Likely
462 isProb = !FCmp->isTrueWhenEqual();
463 } else if (FCmp->getPredicate() == FCmpInst::FCMP_ORD) {
466 } else if (FCmp->getPredicate() == FCmpInst::FCMP_UNO) {
473 unsigned TakenIdx = 0, NonTakenIdx = 1;
476 std::swap(TakenIdx, NonTakenIdx);
478 setEdgeWeight(BB, TakenIdx, FPH_TAKEN_WEIGHT);
479 setEdgeWeight(BB, NonTakenIdx, FPH_NONTAKEN_WEIGHT);
484 bool BranchProbabilityInfo::calcInvokeHeuristics(BasicBlock *BB) {
485 InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator());
489 setEdgeWeight(BB, 0/*Index for Normal*/, IH_TAKEN_WEIGHT);
490 setEdgeWeight(BB, 1/*Index for Unwind*/, IH_NONTAKEN_WEIGHT);
494 void BranchProbabilityInfo::getAnalysisUsage(AnalysisUsage &AU) const {
495 AU.addRequired<LoopInfoWrapperPass>();
496 AU.setPreservesAll();
499 bool BranchProbabilityInfo::runOnFunction(Function &F) {
500 DEBUG(dbgs() << "---- Branch Probability Info : " << F.getName()
502 LastF = &F; // Store the last function we ran on for printing.
503 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
504 assert(PostDominatedByUnreachable.empty());
505 assert(PostDominatedByColdCall.empty());
507 // Walk the basic blocks in post-order so that we can build up state about
508 // the successors of a block iteratively.
509 for (auto BB : post_order(&F.getEntryBlock())) {
510 DEBUG(dbgs() << "Computing probabilities for " << BB->getName() << "\n");
511 if (calcUnreachableHeuristics(BB))
513 if (calcMetadataWeights(BB))
515 if (calcColdCallHeuristics(BB))
517 if (calcLoopBranchHeuristics(BB))
519 if (calcPointerHeuristics(BB))
521 if (calcZeroHeuristics(BB))
523 if (calcFloatingPointHeuristics(BB))
525 calcInvokeHeuristics(BB);
528 PostDominatedByUnreachable.clear();
529 PostDominatedByColdCall.clear();
533 void BranchProbabilityInfo::print(raw_ostream &OS, const Module *) const {
534 OS << "---- Branch Probabilities ----\n";
535 // We print the probabilities from the last function the analysis ran over,
536 // or the function it is currently running over.
537 assert(LastF && "Cannot print prior to running over a function");
538 for (Function::const_iterator BI = LastF->begin(), BE = LastF->end();
540 for (succ_const_iterator SI = succ_begin(BI), SE = succ_end(BI);
542 printEdgeProbability(OS << " ", BI, *SI);
547 uint32_t BranchProbabilityInfo::getSumForBlock(const BasicBlock *BB) const {
550 for (succ_const_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
551 uint32_t Weight = getEdgeWeight(BB, I.getSuccessorIndex());
552 uint32_t PrevSum = Sum;
555 assert(Sum >= PrevSum); (void) PrevSum;
561 bool BranchProbabilityInfo::
562 isEdgeHot(const BasicBlock *Src, const BasicBlock *Dst) const {
563 // Hot probability is at least 4/5 = 80%
564 // FIXME: Compare against a static "hot" BranchProbability.
565 return getEdgeProbability(Src, Dst) > BranchProbability(4, 5);
568 BasicBlock *BranchProbabilityInfo::getHotSucc(BasicBlock *BB) const {
570 uint32_t MaxWeight = 0;
571 BasicBlock *MaxSucc = nullptr;
573 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
574 BasicBlock *Succ = *I;
575 uint32_t Weight = getEdgeWeight(BB, Succ);
576 uint32_t PrevSum = Sum;
579 assert(Sum > PrevSum); (void) PrevSum;
581 if (Weight > MaxWeight) {
587 // Hot probability is at least 4/5 = 80%
588 if (BranchProbability(MaxWeight, Sum) > BranchProbability(4, 5))
594 /// Get the raw edge weight for the edge. If can't find it, return
595 /// DEFAULT_WEIGHT value. Here an edge is specified using PredBlock and an index
596 /// to the successors.
597 uint32_t BranchProbabilityInfo::
598 getEdgeWeight(const BasicBlock *Src, unsigned IndexInSuccessors) const {
599 DenseMap<Edge, uint32_t>::const_iterator I =
600 Weights.find(std::make_pair(Src, IndexInSuccessors));
602 if (I != Weights.end())
605 return DEFAULT_WEIGHT;
608 uint32_t BranchProbabilityInfo::getEdgeWeight(const BasicBlock *Src,
609 succ_const_iterator Dst) const {
610 return getEdgeWeight(Src, Dst.getSuccessorIndex());
613 /// Get the raw edge weight calculated for the block pair. This returns the sum
614 /// of all raw edge weights from Src to Dst.
615 uint32_t BranchProbabilityInfo::
616 getEdgeWeight(const BasicBlock *Src, const BasicBlock *Dst) const {
618 bool FoundWeight = false;
619 DenseMap<Edge, uint32_t>::const_iterator MapI;
620 for (succ_const_iterator I = succ_begin(Src), E = succ_end(Src); I != E; ++I)
622 MapI = Weights.find(std::make_pair(Src, I.getSuccessorIndex()));
623 if (MapI != Weights.end()) {
625 Weight += MapI->second;
628 return (!FoundWeight) ? DEFAULT_WEIGHT : Weight;
631 /// Set the edge weight for a given edge specified by PredBlock and an index
632 /// to the successors.
633 void BranchProbabilityInfo::
634 setEdgeWeight(const BasicBlock *Src, unsigned IndexInSuccessors,
636 Weights[std::make_pair(Src, IndexInSuccessors)] = Weight;
637 DEBUG(dbgs() << "set edge " << Src->getName() << " -> "
638 << IndexInSuccessors << " successor weight to "
642 /// Get an edge's probability, relative to other out-edges from Src.
643 BranchProbability BranchProbabilityInfo::
644 getEdgeProbability(const BasicBlock *Src, unsigned IndexInSuccessors) const {
645 uint32_t N = getEdgeWeight(Src, IndexInSuccessors);
646 uint32_t D = getSumForBlock(Src);
648 return BranchProbability(N, D);
651 /// Get the probability of going from Src to Dst. It returns the sum of all
652 /// probabilities for edges from Src to Dst.
653 BranchProbability BranchProbabilityInfo::
654 getEdgeProbability(const BasicBlock *Src, const BasicBlock *Dst) const {
656 uint32_t N = getEdgeWeight(Src, Dst);
657 uint32_t D = getSumForBlock(Src);
659 return BranchProbability(N, D);
663 BranchProbabilityInfo::printEdgeProbability(raw_ostream &OS,
664 const BasicBlock *Src,
665 const BasicBlock *Dst) const {
667 const BranchProbability Prob = getEdgeProbability(Src, Dst);
668 OS << "edge " << Src->getName() << " -> " << Dst->getName()
669 << " probability is " << Prob
670 << (isEdgeHot(Src, Dst) ? " [HOT edge]\n" : "\n");