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"
27 #define DEBUG_TYPE "branch-prob"
29 INITIALIZE_PASS_BEGIN(BranchProbabilityInfo, "branch-prob",
30 "Branch Probability Analysis", false, true)
31 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
32 INITIALIZE_PASS_END(BranchProbabilityInfo, "branch-prob",
33 "Branch Probability Analysis", false, true)
35 char BranchProbabilityInfo::ID = 0;
37 // Weights are for internal use only. They are used by heuristics to help to
38 // estimate edges' probability. Example:
40 // Using "Loop Branch Heuristics" we predict weights of edges for the
55 // Probability of the edge BB2->BB1 = 124 / (124 + 4) = 0.96875
56 // Probability of the edge BB2->BB3 = 4 / (124 + 4) = 0.03125
57 static const uint32_t LBH_TAKEN_WEIGHT = 124;
58 static const uint32_t LBH_NONTAKEN_WEIGHT = 4;
60 /// \brief Unreachable-terminating branch taken weight.
62 /// This is the weight for a branch being taken to a block that terminates
63 /// (eventually) in unreachable. These are predicted as unlikely as possible.
64 static const uint32_t UR_TAKEN_WEIGHT = 1;
66 /// \brief Unreachable-terminating branch not-taken weight.
68 /// This is the weight for a branch not being taken toward a block that
69 /// terminates (eventually) in unreachable. Such a branch is essentially never
70 /// taken. Set the weight to an absurdly high value so that nested loops don't
71 /// easily subsume it.
72 static const uint32_t UR_NONTAKEN_WEIGHT = 1024*1024 - 1;
74 /// \brief Weight for a branch taken going into a cold block.
76 /// This is the weight for a branch taken toward a block marked
77 /// cold. A block is marked cold if it's postdominated by a
78 /// block containing a call to a cold function. Cold functions
79 /// are those marked with attribute 'cold'.
80 static const uint32_t CC_TAKEN_WEIGHT = 4;
82 /// \brief Weight for a branch not-taken into a cold block.
84 /// This is the weight for a branch not taken toward a block marked
86 static const uint32_t CC_NONTAKEN_WEIGHT = 64;
88 static const uint32_t PH_TAKEN_WEIGHT = 20;
89 static const uint32_t PH_NONTAKEN_WEIGHT = 12;
91 static const uint32_t ZH_TAKEN_WEIGHT = 20;
92 static const uint32_t ZH_NONTAKEN_WEIGHT = 12;
94 static const uint32_t FPH_TAKEN_WEIGHT = 20;
95 static const uint32_t FPH_NONTAKEN_WEIGHT = 12;
97 /// \brief Invoke-terminating normal branch taken weight
99 /// This is the weight for branching to the normal destination of an invoke
100 /// instruction. We expect this to happen most of the time. Set the weight to an
101 /// absurdly high value so that nested loops subsume it.
102 static const uint32_t IH_TAKEN_WEIGHT = 1024 * 1024 - 1;
104 /// \brief Invoke-terminating normal branch not-taken weight.
106 /// This is the weight for branching to the unwind destination of an invoke
107 /// instruction. This is essentially never taken.
108 static const uint32_t IH_NONTAKEN_WEIGHT = 1;
110 // Standard weight value. Used when none of the heuristics set weight for
112 static const uint32_t NORMAL_WEIGHT = 16;
114 // Minimum weight of an edge. Please note, that weight is NEVER 0.
115 static const uint32_t MIN_WEIGHT = 1;
117 static uint32_t getMaxWeightFor(BasicBlock *BB) {
118 return UINT32_MAX / BB->getTerminator()->getNumSuccessors();
122 /// \brief Calculate edge weights for successors lead to unreachable.
124 /// Predict that a successor which leads necessarily to an
125 /// unreachable-terminated block as extremely unlikely.
126 bool BranchProbabilityInfo::calcUnreachableHeuristics(BasicBlock *BB) {
127 TerminatorInst *TI = BB->getTerminator();
128 if (TI->getNumSuccessors() == 0) {
129 if (isa<UnreachableInst>(TI))
130 PostDominatedByUnreachable.insert(BB);
134 SmallVector<unsigned, 4> UnreachableEdges;
135 SmallVector<unsigned, 4> ReachableEdges;
137 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
138 if (PostDominatedByUnreachable.count(*I))
139 UnreachableEdges.push_back(I.getSuccessorIndex());
141 ReachableEdges.push_back(I.getSuccessorIndex());
144 // If all successors are in the set of blocks post-dominated by unreachable,
145 // this block is too.
146 if (UnreachableEdges.size() == TI->getNumSuccessors())
147 PostDominatedByUnreachable.insert(BB);
149 // Skip probabilities if this block has a single successor or if all were
151 if (TI->getNumSuccessors() == 1 || UnreachableEdges.empty())
154 uint32_t UnreachableWeight =
155 std::max(UR_TAKEN_WEIGHT / (unsigned)UnreachableEdges.size(), MIN_WEIGHT);
156 for (SmallVectorImpl<unsigned>::iterator I = UnreachableEdges.begin(),
157 E = UnreachableEdges.end();
159 setEdgeWeight(BB, *I, UnreachableWeight);
161 if (ReachableEdges.empty())
163 uint32_t ReachableWeight =
164 std::max(UR_NONTAKEN_WEIGHT / (unsigned)ReachableEdges.size(),
166 for (SmallVectorImpl<unsigned>::iterator I = ReachableEdges.begin(),
167 E = ReachableEdges.end();
169 setEdgeWeight(BB, *I, ReachableWeight);
174 // Propagate existing explicit probabilities from either profile data or
175 // 'expect' intrinsic processing.
176 bool BranchProbabilityInfo::calcMetadataWeights(BasicBlock *BB) {
177 TerminatorInst *TI = BB->getTerminator();
178 if (TI->getNumSuccessors() == 1)
180 if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
183 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
187 // Ensure there are weights for all of the successors. Note that the first
188 // operand to the metadata node is a name, not a weight.
189 if (WeightsNode->getNumOperands() != TI->getNumSuccessors() + 1)
192 // Build up the final weights that will be used in a temporary buffer, but
193 // don't add them until all weihts are present. Each weight value is clamped
194 // to [1, getMaxWeightFor(BB)].
195 uint32_t WeightLimit = getMaxWeightFor(BB);
196 SmallVector<uint32_t, 2> Weights;
197 Weights.reserve(TI->getNumSuccessors());
198 for (unsigned i = 1, e = WeightsNode->getNumOperands(); i != e; ++i) {
199 ConstantInt *Weight =
200 mdconst::dyn_extract<ConstantInt>(WeightsNode->getOperand(i));
204 std::max<uint32_t>(1, 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);
383 switch (CI->getPredicate()) {
384 case CmpInst::ICMP_EQ:
385 // X == 0 -> Unlikely
388 case CmpInst::ICMP_NE:
392 case CmpInst::ICMP_SLT:
396 case CmpInst::ICMP_SGT:
403 } else if (CV->isOne() && CI->getPredicate() == CmpInst::ICMP_SLT) {
404 // InstCombine canonicalizes X <= 0 into X < 1.
405 // X <= 0 -> Unlikely
407 } else if (CV->isAllOnesValue()) {
408 switch (CI->getPredicate()) {
409 case CmpInst::ICMP_EQ:
410 // X == -1 -> Unlikely
413 case CmpInst::ICMP_NE:
417 case CmpInst::ICMP_SGT:
418 // InstCombine canonicalizes X >= 0 into X > -1.
429 unsigned TakenIdx = 0, NonTakenIdx = 1;
432 std::swap(TakenIdx, NonTakenIdx);
434 setEdgeWeight(BB, TakenIdx, ZH_TAKEN_WEIGHT);
435 setEdgeWeight(BB, NonTakenIdx, ZH_NONTAKEN_WEIGHT);
440 bool BranchProbabilityInfo::calcFloatingPointHeuristics(BasicBlock *BB) {
441 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
442 if (!BI || !BI->isConditional())
445 Value *Cond = BI->getCondition();
446 FCmpInst *FCmp = dyn_cast<FCmpInst>(Cond);
451 if (FCmp->isEquality()) {
452 // f1 == f2 -> Unlikely
453 // f1 != f2 -> Likely
454 isProb = !FCmp->isTrueWhenEqual();
455 } else if (FCmp->getPredicate() == FCmpInst::FCMP_ORD) {
458 } else if (FCmp->getPredicate() == FCmpInst::FCMP_UNO) {
465 unsigned TakenIdx = 0, NonTakenIdx = 1;
468 std::swap(TakenIdx, NonTakenIdx);
470 setEdgeWeight(BB, TakenIdx, FPH_TAKEN_WEIGHT);
471 setEdgeWeight(BB, NonTakenIdx, FPH_NONTAKEN_WEIGHT);
476 bool BranchProbabilityInfo::calcInvokeHeuristics(BasicBlock *BB) {
477 InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator());
481 setEdgeWeight(BB, 0/*Index for Normal*/, IH_TAKEN_WEIGHT);
482 setEdgeWeight(BB, 1/*Index for Unwind*/, IH_NONTAKEN_WEIGHT);
486 void BranchProbabilityInfo::getAnalysisUsage(AnalysisUsage &AU) const {
487 AU.addRequired<LoopInfoWrapperPass>();
488 AU.setPreservesAll();
491 bool BranchProbabilityInfo::runOnFunction(Function &F) {
492 DEBUG(dbgs() << "---- Branch Probability Info : " << F.getName()
494 LastF = &F; // Store the last function we ran on for printing.
495 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
496 assert(PostDominatedByUnreachable.empty());
497 assert(PostDominatedByColdCall.empty());
499 // Walk the basic blocks in post-order so that we can build up state about
500 // the successors of a block iteratively.
501 for (po_iterator<BasicBlock *> I = po_begin(&F.getEntryBlock()),
502 E = po_end(&F.getEntryBlock());
504 DEBUG(dbgs() << "Computing probabilities for " << I->getName() << "\n");
505 if (calcUnreachableHeuristics(*I))
507 if (calcMetadataWeights(*I))
509 if (calcColdCallHeuristics(*I))
511 if (calcLoopBranchHeuristics(*I))
513 if (calcPointerHeuristics(*I))
515 if (calcZeroHeuristics(*I))
517 if (calcFloatingPointHeuristics(*I))
519 calcInvokeHeuristics(*I);
522 PostDominatedByUnreachable.clear();
523 PostDominatedByColdCall.clear();
527 void BranchProbabilityInfo::print(raw_ostream &OS, const Module *) const {
528 OS << "---- Branch Probabilities ----\n";
529 // We print the probabilities from the last function the analysis ran over,
530 // or the function it is currently running over.
531 assert(LastF && "Cannot print prior to running over a function");
532 for (Function::const_iterator BI = LastF->begin(), BE = LastF->end();
534 for (succ_const_iterator SI = succ_begin(BI), SE = succ_end(BI);
536 printEdgeProbability(OS << " ", BI, *SI);
541 uint32_t BranchProbabilityInfo::getSumForBlock(const BasicBlock *BB) const {
544 for (succ_const_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
545 uint32_t Weight = getEdgeWeight(BB, I.getSuccessorIndex());
546 uint32_t PrevSum = Sum;
549 assert(Sum > PrevSum); (void) PrevSum;
555 bool BranchProbabilityInfo::
556 isEdgeHot(const BasicBlock *Src, const BasicBlock *Dst) const {
557 // Hot probability is at least 4/5 = 80%
558 // FIXME: Compare against a static "hot" BranchProbability.
559 return getEdgeProbability(Src, Dst) > BranchProbability(4, 5);
562 BasicBlock *BranchProbabilityInfo::getHotSucc(BasicBlock *BB) const {
564 uint32_t MaxWeight = 0;
565 BasicBlock *MaxSucc = nullptr;
567 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
568 BasicBlock *Succ = *I;
569 uint32_t Weight = getEdgeWeight(BB, Succ);
570 uint32_t PrevSum = Sum;
573 assert(Sum > PrevSum); (void) PrevSum;
575 if (Weight > MaxWeight) {
581 // Hot probability is at least 4/5 = 80%
582 if (BranchProbability(MaxWeight, Sum) > BranchProbability(4, 5))
588 /// Get the raw edge weight for the edge. If can't find it, return
589 /// DEFAULT_WEIGHT value. Here an edge is specified using PredBlock and an index
590 /// to the successors.
591 uint32_t BranchProbabilityInfo::
592 getEdgeWeight(const BasicBlock *Src, unsigned IndexInSuccessors) const {
593 DenseMap<Edge, uint32_t>::const_iterator I =
594 Weights.find(std::make_pair(Src, IndexInSuccessors));
596 if (I != Weights.end())
599 return DEFAULT_WEIGHT;
602 uint32_t BranchProbabilityInfo::getEdgeWeight(const BasicBlock *Src,
603 succ_const_iterator Dst) const {
604 return getEdgeWeight(Src, Dst.getSuccessorIndex());
607 /// Get the raw edge weight calculated for the block pair. This returns the sum
608 /// of all raw edge weights from Src to Dst.
609 uint32_t BranchProbabilityInfo::
610 getEdgeWeight(const BasicBlock *Src, const BasicBlock *Dst) const {
612 DenseMap<Edge, uint32_t>::const_iterator MapI;
613 for (succ_const_iterator I = succ_begin(Src), E = succ_end(Src); I != E; ++I)
615 MapI = Weights.find(std::make_pair(Src, I.getSuccessorIndex()));
616 if (MapI != Weights.end())
617 Weight += MapI->second;
619 return (Weight == 0) ? DEFAULT_WEIGHT : Weight;
622 /// Set the edge weight for a given edge specified by PredBlock and an index
623 /// to the successors.
624 void BranchProbabilityInfo::
625 setEdgeWeight(const BasicBlock *Src, unsigned IndexInSuccessors,
627 Weights[std::make_pair(Src, IndexInSuccessors)] = Weight;
628 DEBUG(dbgs() << "set edge " << Src->getName() << " -> "
629 << IndexInSuccessors << " successor weight to "
633 /// Get an edge's probability, relative to other out-edges from Src.
634 BranchProbability BranchProbabilityInfo::
635 getEdgeProbability(const BasicBlock *Src, unsigned IndexInSuccessors) const {
636 uint32_t N = getEdgeWeight(Src, IndexInSuccessors);
637 uint32_t D = getSumForBlock(Src);
639 return BranchProbability(N, D);
642 /// Get the probability of going from Src to Dst. It returns the sum of all
643 /// probabilities for edges from Src to Dst.
644 BranchProbability BranchProbabilityInfo::
645 getEdgeProbability(const BasicBlock *Src, const BasicBlock *Dst) const {
647 uint32_t N = getEdgeWeight(Src, Dst);
648 uint32_t D = getSumForBlock(Src);
650 return BranchProbability(N, D);
654 BranchProbabilityInfo::printEdgeProbability(raw_ostream &OS,
655 const BasicBlock *Src,
656 const BasicBlock *Dst) const {
658 const BranchProbability Prob = getEdgeProbability(Src, Dst);
659 OS << "edge " << Src->getName() << " -> " << Dst->getName()
660 << " probability is " << Prob
661 << (isEdgeHot(Src, Dst) ? " [HOT edge]\n" : "\n");