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 (SmallVector<unsigned, 4>::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 (SmallVector<unsigned, 4>::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 ColdEdges.reserve(TI->getNumSuccessors());
226 SmallVector<unsigned, 4> NormalEdges;
227 NormalEdges.reserve(TI->getNumSuccessors());
228 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I)
229 if (PostDominatedByColdCall.count(*I))
230 ColdEdges.push_back(I.getSuccessorIndex());
232 NormalEdges.push_back(I.getSuccessorIndex());
234 // If all successors are in the set of blocks post-dominated by cold calls,
235 // this block is in the set post-dominated by cold calls.
236 if (ColdEdges.size() == TI->getNumSuccessors())
237 PostDominatedByColdCall.insert(BB);
239 // Otherwise, if the block itself contains a cold function, add it to the
240 // set of blocks postdominated by a cold call.
241 assert(!PostDominatedByColdCall.count(BB));
242 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
243 if (CallInst *CI = dyn_cast<CallInst>(I))
244 if (CI->hasFnAttr(Attribute::Cold)) {
245 PostDominatedByColdCall.insert(BB);
250 // Skip probabilities if this block has a single successor.
251 if (TI->getNumSuccessors() == 1 || ColdEdges.empty())
254 uint32_t ColdWeight =
255 std::max(CC_TAKEN_WEIGHT / (unsigned) ColdEdges.size(), MIN_WEIGHT);
256 for (SmallVector<unsigned, 4>::iterator I = ColdEdges.begin(),
259 setEdgeWeight(BB, *I, ColdWeight);
261 if (NormalEdges.empty())
263 uint32_t NormalWeight = std::max(
264 CC_NONTAKEN_WEIGHT / (unsigned) NormalEdges.size(), NORMAL_WEIGHT);
265 for (SmallVector<unsigned, 4>::iterator I = NormalEdges.begin(),
266 E = NormalEdges.end();
268 setEdgeWeight(BB, *I, NormalWeight);
273 // Calculate Edge Weights using "Pointer Heuristics". Predict a comparsion
274 // between two pointer or pointer and NULL will fail.
275 bool BranchProbabilityInfo::calcPointerHeuristics(BasicBlock *BB) {
276 BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
277 if (!BI || !BI->isConditional())
280 Value *Cond = BI->getCondition();
281 ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
282 if (!CI || !CI->isEquality())
285 Value *LHS = CI->getOperand(0);
287 if (!LHS->getType()->isPointerTy())
290 assert(CI->getOperand(1)->getType()->isPointerTy());
292 // p != 0 -> isProb = true
293 // p == 0 -> isProb = false
294 // p != q -> isProb = true
295 // p == q -> isProb = false;
296 unsigned TakenIdx = 0, NonTakenIdx = 1;
297 bool isProb = CI->getPredicate() == ICmpInst::ICMP_NE;
299 std::swap(TakenIdx, NonTakenIdx);
301 setEdgeWeight(BB, TakenIdx, PH_TAKEN_WEIGHT);
302 setEdgeWeight(BB, NonTakenIdx, PH_NONTAKEN_WEIGHT);
306 // Calculate Edge Weights using "Loop Branch Heuristics". Predict backedges
307 // as taken, exiting edges as not-taken.
308 bool BranchProbabilityInfo::calcLoopBranchHeuristics(BasicBlock *BB) {
309 Loop *L = LI->getLoopFor(BB);
313 SmallVector<unsigned, 8> BackEdges;
314 SmallVector<unsigned, 8> ExitingEdges;
315 SmallVector<unsigned, 8> InEdges; // Edges from header to the loop.
317 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
318 if (!L->contains(*I))
319 ExitingEdges.push_back(I.getSuccessorIndex());
320 else if (L->getHeader() == *I)
321 BackEdges.push_back(I.getSuccessorIndex());
323 InEdges.push_back(I.getSuccessorIndex());
326 if (uint32_t numBackEdges = BackEdges.size()) {
327 uint32_t backWeight = LBH_TAKEN_WEIGHT / numBackEdges;
328 if (backWeight < NORMAL_WEIGHT)
329 backWeight = NORMAL_WEIGHT;
331 for (SmallVector<unsigned, 8>::iterator EI = BackEdges.begin(),
332 EE = BackEdges.end(); EI != EE; ++EI) {
333 setEdgeWeight(BB, *EI, backWeight);
337 if (uint32_t numInEdges = InEdges.size()) {
338 uint32_t inWeight = LBH_TAKEN_WEIGHT / numInEdges;
339 if (inWeight < NORMAL_WEIGHT)
340 inWeight = NORMAL_WEIGHT;
342 for (SmallVector<unsigned, 8>::iterator EI = InEdges.begin(),
343 EE = InEdges.end(); EI != EE; ++EI) {
344 setEdgeWeight(BB, *EI, inWeight);
348 if (uint32_t numExitingEdges = ExitingEdges.size()) {
349 uint32_t exitWeight = LBH_NONTAKEN_WEIGHT / numExitingEdges;
350 if (exitWeight < MIN_WEIGHT)
351 exitWeight = MIN_WEIGHT;
353 for (SmallVector<unsigned, 8>::iterator EI = ExitingEdges.begin(),
354 EE = ExitingEdges.end(); EI != EE; ++EI) {
355 setEdgeWeight(BB, *EI, exitWeight);
362 bool BranchProbabilityInfo::calcZeroHeuristics(BasicBlock *BB) {
363 BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
364 if (!BI || !BI->isConditional())
367 Value *Cond = BI->getCondition();
368 ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
372 Value *RHS = CI->getOperand(1);
373 ConstantInt *CV = dyn_cast<ConstantInt>(RHS);
379 switch (CI->getPredicate()) {
380 case CmpInst::ICMP_EQ:
381 // X == 0 -> Unlikely
384 case CmpInst::ICMP_NE:
388 case CmpInst::ICMP_SLT:
392 case CmpInst::ICMP_SGT:
399 } else if (CV->isOne() && CI->getPredicate() == CmpInst::ICMP_SLT) {
400 // InstCombine canonicalizes X <= 0 into X < 1.
401 // X <= 0 -> Unlikely
403 } else if (CV->isAllOnesValue() && CI->getPredicate() == CmpInst::ICMP_SGT) {
404 // InstCombine canonicalizes X >= 0 into X > -1.
411 unsigned TakenIdx = 0, NonTakenIdx = 1;
414 std::swap(TakenIdx, NonTakenIdx);
416 setEdgeWeight(BB, TakenIdx, ZH_TAKEN_WEIGHT);
417 setEdgeWeight(BB, NonTakenIdx, ZH_NONTAKEN_WEIGHT);
422 bool BranchProbabilityInfo::calcFloatingPointHeuristics(BasicBlock *BB) {
423 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
424 if (!BI || !BI->isConditional())
427 Value *Cond = BI->getCondition();
428 FCmpInst *FCmp = dyn_cast<FCmpInst>(Cond);
433 if (FCmp->isEquality()) {
434 // f1 == f2 -> Unlikely
435 // f1 != f2 -> Likely
436 isProb = !FCmp->isTrueWhenEqual();
437 } else if (FCmp->getPredicate() == FCmpInst::FCMP_ORD) {
440 } else if (FCmp->getPredicate() == FCmpInst::FCMP_UNO) {
447 unsigned TakenIdx = 0, NonTakenIdx = 1;
450 std::swap(TakenIdx, NonTakenIdx);
452 setEdgeWeight(BB, TakenIdx, FPH_TAKEN_WEIGHT);
453 setEdgeWeight(BB, NonTakenIdx, FPH_NONTAKEN_WEIGHT);
458 bool BranchProbabilityInfo::calcInvokeHeuristics(BasicBlock *BB) {
459 InvokeInst *II = dyn_cast<InvokeInst>(BB->getTerminator());
463 setEdgeWeight(BB, 0/*Index for Normal*/, IH_TAKEN_WEIGHT);
464 setEdgeWeight(BB, 1/*Index for Unwind*/, IH_NONTAKEN_WEIGHT);
468 void BranchProbabilityInfo::getAnalysisUsage(AnalysisUsage &AU) const {
469 AU.addRequired<LoopInfo>();
470 AU.setPreservesAll();
473 bool BranchProbabilityInfo::runOnFunction(Function &F) {
474 LastF = &F; // Store the last function we ran on for printing.
475 LI = &getAnalysis<LoopInfo>();
476 assert(PostDominatedByUnreachable.empty());
477 assert(PostDominatedByColdCall.empty());
479 // Walk the basic blocks in post-order so that we can build up state about
480 // the successors of a block iteratively.
481 for (po_iterator<BasicBlock *> I = po_begin(&F.getEntryBlock()),
482 E = po_end(&F.getEntryBlock());
484 DEBUG(dbgs() << "Computing probabilities for " << I->getName() << "\n");
485 if (calcUnreachableHeuristics(*I))
487 if (calcMetadataWeights(*I))
489 if (calcColdCallHeuristics(*I))
491 if (calcLoopBranchHeuristics(*I))
493 if (calcPointerHeuristics(*I))
495 if (calcZeroHeuristics(*I))
497 if (calcFloatingPointHeuristics(*I))
499 calcInvokeHeuristics(*I);
502 PostDominatedByUnreachable.clear();
503 PostDominatedByColdCall.clear();
507 void BranchProbabilityInfo::print(raw_ostream &OS, const Module *) const {
508 OS << "---- Branch Probabilities ----\n";
509 // We print the probabilities from the last function the analysis ran over,
510 // or the function it is currently running over.
511 assert(LastF && "Cannot print prior to running over a function");
512 for (Function::const_iterator BI = LastF->begin(), BE = LastF->end();
514 for (succ_const_iterator SI = succ_begin(BI), SE = succ_end(BI);
516 printEdgeProbability(OS << " ", BI, *SI);
521 uint32_t BranchProbabilityInfo::getSumForBlock(const BasicBlock *BB) const {
524 for (succ_const_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
525 uint32_t Weight = getEdgeWeight(BB, I.getSuccessorIndex());
526 uint32_t PrevSum = Sum;
529 assert(Sum > PrevSum); (void) PrevSum;
535 bool BranchProbabilityInfo::
536 isEdgeHot(const BasicBlock *Src, const BasicBlock *Dst) const {
537 // Hot probability is at least 4/5 = 80%
538 // FIXME: Compare against a static "hot" BranchProbability.
539 return getEdgeProbability(Src, Dst) > BranchProbability(4, 5);
542 BasicBlock *BranchProbabilityInfo::getHotSucc(BasicBlock *BB) const {
544 uint32_t MaxWeight = 0;
545 BasicBlock *MaxSucc = 0;
547 for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
548 BasicBlock *Succ = *I;
549 uint32_t Weight = getEdgeWeight(BB, Succ);
550 uint32_t PrevSum = Sum;
553 assert(Sum > PrevSum); (void) PrevSum;
555 if (Weight > MaxWeight) {
561 // Hot probability is at least 4/5 = 80%
562 if (BranchProbability(MaxWeight, Sum) > BranchProbability(4, 5))
568 /// Get the raw edge weight for the edge. If can't find it, return
569 /// DEFAULT_WEIGHT value. Here an edge is specified using PredBlock and an index
570 /// to the successors.
571 uint32_t BranchProbabilityInfo::
572 getEdgeWeight(const BasicBlock *Src, unsigned IndexInSuccessors) const {
573 DenseMap<Edge, uint32_t>::const_iterator I =
574 Weights.find(std::make_pair(Src, IndexInSuccessors));
576 if (I != Weights.end())
579 return DEFAULT_WEIGHT;
582 /// Get the raw edge weight calculated for the block pair. This returns the sum
583 /// of all raw edge weights from Src to Dst.
584 uint32_t BranchProbabilityInfo::
585 getEdgeWeight(const BasicBlock *Src, const BasicBlock *Dst) const {
587 DenseMap<Edge, uint32_t>::const_iterator MapI;
588 for (succ_const_iterator I = succ_begin(Src), E = succ_end(Src); I != E; ++I)
590 MapI = Weights.find(std::make_pair(Src, I.getSuccessorIndex()));
591 if (MapI != Weights.end())
592 Weight += MapI->second;
594 return (Weight == 0) ? DEFAULT_WEIGHT : Weight;
597 /// Set the edge weight for a given edge specified by PredBlock and an index
598 /// to the successors.
599 void BranchProbabilityInfo::
600 setEdgeWeight(const BasicBlock *Src, unsigned IndexInSuccessors,
602 Weights[std::make_pair(Src, IndexInSuccessors)] = Weight;
603 DEBUG(dbgs() << "set edge " << Src->getName() << " -> "
604 << IndexInSuccessors << " successor weight to "
608 /// Get an edge's probability, relative to other out-edges from Src.
609 BranchProbability BranchProbabilityInfo::
610 getEdgeProbability(const BasicBlock *Src, unsigned IndexInSuccessors) const {
611 uint32_t N = getEdgeWeight(Src, IndexInSuccessors);
612 uint32_t D = getSumForBlock(Src);
614 return BranchProbability(N, D);
617 /// Get the probability of going from Src to Dst. It returns the sum of all
618 /// probabilities for edges from Src to Dst.
619 BranchProbability BranchProbabilityInfo::
620 getEdgeProbability(const BasicBlock *Src, const BasicBlock *Dst) const {
622 uint32_t N = getEdgeWeight(Src, Dst);
623 uint32_t D = getSumForBlock(Src);
625 return BranchProbability(N, D);
629 BranchProbabilityInfo::printEdgeProbability(raw_ostream &OS,
630 const BasicBlock *Src,
631 const BasicBlock *Dst) const {
633 const BranchProbability Prob = getEdgeProbability(Src, Dst);
634 OS << "edge " << Src->getName() << " -> " << Dst->getName()
635 << " probability is " << Prob
636 << (isEdgeHot(Src, Dst) ? " [HOT edge]\n" : "\n");