1//===-- BranchProbabilityInfo.cpp - Branch Probability Analysis -*- C++ -*-===//
2//
3//                     The LLVM Compiler Infrastructure
4//
5// This file is distributed under the University of Illinois Open Source
6// License. See LICENSE.TXT for details.
7//
8//===----------------------------------------------------------------------===//
9//
10// Loops should be simplified before this analysis.
11//
12//===----------------------------------------------------------------------===//
13
14#include "llvm/Constants.h"
15#include "llvm/Function.h"
16#include "llvm/Instructions.h"
17#include "llvm/LLVMContext.h"
18#include "llvm/Metadata.h"
19#include "llvm/Analysis/BranchProbabilityInfo.h"
20#include "llvm/Analysis/LoopInfo.h"
21#include "llvm/ADT/PostOrderIterator.h"
22#include "llvm/Support/CFG.h"
23#include "llvm/Support/Debug.h"
24
25using namespace llvm;
26
27INITIALIZE_PASS_BEGIN(BranchProbabilityInfo, "branch-prob",
28                      "Branch Probability Analysis", false, true)
29INITIALIZE_PASS_DEPENDENCY(LoopInfo)
30INITIALIZE_PASS_END(BranchProbabilityInfo, "branch-prob",
31                    "Branch Probability Analysis", false, true)
32
33char BranchProbabilityInfo::ID = 0;
34
35// Weights are for internal use only. They are used by heuristics to help to
36// estimate edges' probability. Example:
37//
38// Using "Loop Branch Heuristics" we predict weights of edges for the
39// block BB2.
40//         ...
41//          |
42//          V
43//         BB1<-+
44//          |   |
45//          |   | (Weight = 124)
46//          V   |
47//         BB2--+
48//          |
49//          | (Weight = 4)
50//          V
51//         BB3
52//
53// Probability of the edge BB2->BB1 = 124 / (124 + 4) = 0.96875
54// Probability of the edge BB2->BB3 = 4 / (124 + 4) = 0.03125
55static const uint32_t LBH_TAKEN_WEIGHT = 124;
56static const uint32_t LBH_NONTAKEN_WEIGHT = 4;
57
58/// \brief Unreachable-terminating branch taken weight.
59///
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.
62static const uint32_t UR_TAKEN_WEIGHT = 1;
63
64/// \brief Unreachable-terminating branch not-taken weight.
65///
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.
70static const uint32_t UR_NONTAKEN_WEIGHT = 1024*1024 - 1;
71
72static const uint32_t PH_TAKEN_WEIGHT = 20;
73static const uint32_t PH_NONTAKEN_WEIGHT = 12;
74
75static const uint32_t ZH_TAKEN_WEIGHT = 20;
76static const uint32_t ZH_NONTAKEN_WEIGHT = 12;
77
78static const uint32_t FPH_TAKEN_WEIGHT = 20;
79static const uint32_t FPH_NONTAKEN_WEIGHT = 12;
80
81// Standard weight value. Used when none of the heuristics set weight for
82// the edge.
83static const uint32_t NORMAL_WEIGHT = 16;
84
85// Minimum weight of an edge. Please note, that weight is NEVER 0.
86static const uint32_t MIN_WEIGHT = 1;
87
88static uint32_t getMaxWeightFor(BasicBlock *BB) {
89  return UINT32_MAX / BB->getTerminator()->getNumSuccessors();
90}
91
92
93/// \brief Calculate edge weights for successors lead to unreachable.
94///
95/// Predict that a successor which leads necessarily to an
96/// unreachable-terminated block as extremely unlikely.
97bool BranchProbabilityInfo::calcUnreachableHeuristics(BasicBlock *BB) {
98  TerminatorInst *TI = BB->getTerminator();
99  if (TI->getNumSuccessors() == 0) {
100    if (isa<UnreachableInst>(TI))
101      PostDominatedByUnreachable.insert(BB);
102    return false;
103  }
104
105  SmallPtrSet<BasicBlock *, 4> UnreachableEdges;
106  SmallPtrSet<BasicBlock *, 4> ReachableEdges;
107
108  for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
109    if (PostDominatedByUnreachable.count(*I))
110      UnreachableEdges.insert(*I);
111    else
112      ReachableEdges.insert(*I);
113  }
114
115  // If all successors are in the set of blocks post-dominated by unreachable,
116  // this block is too.
117  if (UnreachableEdges.size() == TI->getNumSuccessors())
118    PostDominatedByUnreachable.insert(BB);
119
120  // Skip probabilities if this block has a single successor or if all were
121  // reachable.
122  if (TI->getNumSuccessors() == 1 || UnreachableEdges.empty())
123    return false;
124
125  uint32_t UnreachableWeight =
126    std::max(UR_TAKEN_WEIGHT / UnreachableEdges.size(), MIN_WEIGHT);
127  for (SmallPtrSet<BasicBlock *, 4>::iterator I = UnreachableEdges.begin(),
128                                              E = UnreachableEdges.end();
129       I != E; ++I)
130    setEdgeWeight(BB, *I, UnreachableWeight);
131
132  if (ReachableEdges.empty())
133    return true;
134  uint32_t ReachableWeight =
135    std::max(UR_NONTAKEN_WEIGHT / ReachableEdges.size(), NORMAL_WEIGHT);
136  for (SmallPtrSet<BasicBlock *, 4>::iterator I = ReachableEdges.begin(),
137                                              E = ReachableEdges.end();
138       I != E; ++I)
139    setEdgeWeight(BB, *I, ReachableWeight);
140
141  return true;
142}
143
144// Propagate existing explicit probabilities from either profile data or
145// 'expect' intrinsic processing.
146bool BranchProbabilityInfo::calcMetadataWeights(BasicBlock *BB) {
147  TerminatorInst *TI = BB->getTerminator();
148  if (TI->getNumSuccessors() == 1)
149    return false;
150  if (!isa<BranchInst>(TI) && !isa<SwitchInst>(TI))
151    return false;
152
153  MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
154  if (!WeightsNode)
155    return false;
156
157  // Ensure there are weights for all of the successors. Note that the first
158  // operand to the metadata node is a name, not a weight.
159  if (WeightsNode->getNumOperands() != TI->getNumSuccessors() + 1)
160    return false;
161
162  // Build up the final weights that will be used in a temporary buffer, but
163  // don't add them until all weihts are present. Each weight value is clamped
164  // to [1, getMaxWeightFor(BB)].
165  uint32_t WeightLimit = getMaxWeightFor(BB);
166  SmallVector<uint32_t, 2> Weights;
167  Weights.reserve(TI->getNumSuccessors());
168  for (unsigned i = 1, e = WeightsNode->getNumOperands(); i != e; ++i) {
169    ConstantInt *Weight = dyn_cast<ConstantInt>(WeightsNode->getOperand(i));
170    if (!Weight)
171      return false;
172    Weights.push_back(
173      std::max<uint32_t>(1, Weight->getLimitedValue(WeightLimit)));
174  }
175  assert(Weights.size() == TI->getNumSuccessors() && "Checked above");
176  for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
177    setEdgeWeight(BB, TI->getSuccessor(i), Weights[i]);
178
179  return true;
180}
181
182// Calculate Edge Weights using "Pointer Heuristics". Predict a comparsion
183// between two pointer or pointer and NULL will fail.
184bool BranchProbabilityInfo::calcPointerHeuristics(BasicBlock *BB) {
185  BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
186  if (!BI || !BI->isConditional())
187    return false;
188
189  Value *Cond = BI->getCondition();
190  ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
191  if (!CI || !CI->isEquality())
192    return false;
193
194  Value *LHS = CI->getOperand(0);
195
196  if (!LHS->getType()->isPointerTy())
197    return false;
198
199  assert(CI->getOperand(1)->getType()->isPointerTy());
200
201  BasicBlock *Taken = BI->getSuccessor(0);
202  BasicBlock *NonTaken = BI->getSuccessor(1);
203
204  // p != 0   ->   isProb = true
205  // p == 0   ->   isProb = false
206  // p != q   ->   isProb = true
207  // p == q   ->   isProb = false;
208  bool isProb = CI->getPredicate() == ICmpInst::ICMP_NE;
209  if (!isProb)
210    std::swap(Taken, NonTaken);
211
212  setEdgeWeight(BB, Taken, PH_TAKEN_WEIGHT);
213  setEdgeWeight(BB, NonTaken, PH_NONTAKEN_WEIGHT);
214  return true;
215}
216
217// Calculate Edge Weights using "Loop Branch Heuristics". Predict backedges
218// as taken, exiting edges as not-taken.
219bool BranchProbabilityInfo::calcLoopBranchHeuristics(BasicBlock *BB) {
220  Loop *L = LI->getLoopFor(BB);
221  if (!L)
222    return false;
223
224  SmallPtrSet<BasicBlock *, 8> BackEdges;
225  SmallPtrSet<BasicBlock *, 8> ExitingEdges;
226  SmallPtrSet<BasicBlock *, 8> InEdges; // Edges from header to the loop.
227
228  for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
229    if (!L->contains(*I))
230      ExitingEdges.insert(*I);
231    else if (L->getHeader() == *I)
232      BackEdges.insert(*I);
233    else
234      InEdges.insert(*I);
235  }
236
237  if (uint32_t numBackEdges = BackEdges.size()) {
238    uint32_t backWeight = LBH_TAKEN_WEIGHT / numBackEdges;
239    if (backWeight < NORMAL_WEIGHT)
240      backWeight = NORMAL_WEIGHT;
241
242    for (SmallPtrSet<BasicBlock *, 8>::iterator EI = BackEdges.begin(),
243         EE = BackEdges.end(); EI != EE; ++EI) {
244      BasicBlock *Back = *EI;
245      setEdgeWeight(BB, Back, backWeight);
246    }
247  }
248
249  if (uint32_t numInEdges = InEdges.size()) {
250    uint32_t inWeight = LBH_TAKEN_WEIGHT / numInEdges;
251    if (inWeight < NORMAL_WEIGHT)
252      inWeight = NORMAL_WEIGHT;
253
254    for (SmallPtrSet<BasicBlock *, 8>::iterator EI = InEdges.begin(),
255         EE = InEdges.end(); EI != EE; ++EI) {
256      BasicBlock *Back = *EI;
257      setEdgeWeight(BB, Back, inWeight);
258    }
259  }
260
261  if (uint32_t numExitingEdges = ExitingEdges.size()) {
262    uint32_t exitWeight = LBH_NONTAKEN_WEIGHT / numExitingEdges;
263    if (exitWeight < MIN_WEIGHT)
264      exitWeight = MIN_WEIGHT;
265
266    for (SmallPtrSet<BasicBlock *, 8>::iterator EI = ExitingEdges.begin(),
267         EE = ExitingEdges.end(); EI != EE; ++EI) {
268      BasicBlock *Exiting = *EI;
269      setEdgeWeight(BB, Exiting, exitWeight);
270    }
271  }
272
273  return true;
274}
275
276bool BranchProbabilityInfo::calcZeroHeuristics(BasicBlock *BB) {
277  BranchInst * BI = dyn_cast<BranchInst>(BB->getTerminator());
278  if (!BI || !BI->isConditional())
279    return false;
280
281  Value *Cond = BI->getCondition();
282  ICmpInst *CI = dyn_cast<ICmpInst>(Cond);
283  if (!CI)
284    return false;
285
286  Value *RHS = CI->getOperand(1);
287  ConstantInt *CV = dyn_cast<ConstantInt>(RHS);
288  if (!CV)
289    return false;
290
291  bool isProb;
292  if (CV->isZero()) {
293    switch (CI->getPredicate()) {
294    case CmpInst::ICMP_EQ:
295      // X == 0   ->  Unlikely
296      isProb = false;
297      break;
298    case CmpInst::ICMP_NE:
299      // X != 0   ->  Likely
300      isProb = true;
301      break;
302    case CmpInst::ICMP_SLT:
303      // X < 0   ->  Unlikely
304      isProb = false;
305      break;
306    case CmpInst::ICMP_SGT:
307      // X > 0   ->  Likely
308      isProb = true;
309      break;
310    default:
311      return false;
312    }
313  } else if (CV->isOne() && CI->getPredicate() == CmpInst::ICMP_SLT) {
314    // InstCombine canonicalizes X <= 0 into X < 1.
315    // X <= 0   ->  Unlikely
316    isProb = false;
317  } else if (CV->isAllOnesValue() && CI->getPredicate() == CmpInst::ICMP_SGT) {
318    // InstCombine canonicalizes X >= 0 into X > -1.
319    // X >= 0   ->  Likely
320    isProb = true;
321  } else {
322    return false;
323  }
324
325  BasicBlock *Taken = BI->getSuccessor(0);
326  BasicBlock *NonTaken = BI->getSuccessor(1);
327
328  if (!isProb)
329    std::swap(Taken, NonTaken);
330
331  setEdgeWeight(BB, Taken, ZH_TAKEN_WEIGHT);
332  setEdgeWeight(BB, NonTaken, ZH_NONTAKEN_WEIGHT);
333
334  return true;
335}
336
337bool BranchProbabilityInfo::calcFloatingPointHeuristics(BasicBlock *BB) {
338  BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
339  if (!BI || !BI->isConditional())
340    return false;
341
342  Value *Cond = BI->getCondition();
343  FCmpInst *FCmp = dyn_cast<FCmpInst>(Cond);
344  if (!FCmp)
345    return false;
346
347  bool isProb;
348  if (FCmp->isEquality()) {
349    // f1 == f2 -> Unlikely
350    // f1 != f2 -> Likely
351    isProb = !FCmp->isTrueWhenEqual();
352  } else if (FCmp->getPredicate() == FCmpInst::FCMP_ORD) {
353    // !isnan -> Likely
354    isProb = true;
355  } else if (FCmp->getPredicate() == FCmpInst::FCMP_UNO) {
356    // isnan -> Unlikely
357    isProb = false;
358  } else {
359    return false;
360  }
361
362  BasicBlock *Taken = BI->getSuccessor(0);
363  BasicBlock *NonTaken = BI->getSuccessor(1);
364
365  if (!isProb)
366    std::swap(Taken, NonTaken);
367
368  setEdgeWeight(BB, Taken, FPH_TAKEN_WEIGHT);
369  setEdgeWeight(BB, NonTaken, FPH_NONTAKEN_WEIGHT);
370
371  return true;
372}
373
374void BranchProbabilityInfo::getAnalysisUsage(AnalysisUsage &AU) const {
375  AU.addRequired<LoopInfo>();
376  AU.setPreservesAll();
377}
378
379bool BranchProbabilityInfo::runOnFunction(Function &F) {
380  LastF = &F; // Store the last function we ran on for printing.
381  LI = &getAnalysis<LoopInfo>();
382  assert(PostDominatedByUnreachable.empty());
383
384  // Walk the basic blocks in post-order so that we can build up state about
385  // the successors of a block iteratively.
386  for (po_iterator<BasicBlock *> I = po_begin(&F.getEntryBlock()),
387                                 E = po_end(&F.getEntryBlock());
388       I != E; ++I) {
389    DEBUG(dbgs() << "Computing probabilities for " << I->getName() << "\n");
390    if (calcUnreachableHeuristics(*I))
391      continue;
392    if (calcMetadataWeights(*I))
393      continue;
394    if (calcLoopBranchHeuristics(*I))
395      continue;
396    if (calcPointerHeuristics(*I))
397      continue;
398    if (calcZeroHeuristics(*I))
399      continue;
400    calcFloatingPointHeuristics(*I);
401  }
402
403  PostDominatedByUnreachable.clear();
404  return false;
405}
406
407void BranchProbabilityInfo::print(raw_ostream &OS, const Module *) const {
408  OS << "---- Branch Probabilities ----\n";
409  // We print the probabilities from the last function the analysis ran over,
410  // or the function it is currently running over.
411  assert(LastF && "Cannot print prior to running over a function");
412  for (Function::const_iterator BI = LastF->begin(), BE = LastF->end();
413       BI != BE; ++BI) {
414    for (succ_const_iterator SI = succ_begin(BI), SE = succ_end(BI);
415         SI != SE; ++SI) {
416      printEdgeProbability(OS << "  ", BI, *SI);
417    }
418  }
419}
420
421uint32_t BranchProbabilityInfo::getSumForBlock(const BasicBlock *BB) const {
422  uint32_t Sum = 0;
423
424  for (succ_const_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
425    const BasicBlock *Succ = *I;
426    uint32_t Weight = getEdgeWeight(BB, Succ);
427    uint32_t PrevSum = Sum;
428
429    Sum += Weight;
430    assert(Sum > PrevSum); (void) PrevSum;
431  }
432
433  return Sum;
434}
435
436bool BranchProbabilityInfo::
437isEdgeHot(const BasicBlock *Src, const BasicBlock *Dst) const {
438  // Hot probability is at least 4/5 = 80%
439  // FIXME: Compare against a static "hot" BranchProbability.
440  return getEdgeProbability(Src, Dst) > BranchProbability(4, 5);
441}
442
443BasicBlock *BranchProbabilityInfo::getHotSucc(BasicBlock *BB) const {
444  uint32_t Sum = 0;
445  uint32_t MaxWeight = 0;
446  BasicBlock *MaxSucc = 0;
447
448  for (succ_iterator I = succ_begin(BB), E = succ_end(BB); I != E; ++I) {
449    BasicBlock *Succ = *I;
450    uint32_t Weight = getEdgeWeight(BB, Succ);
451    uint32_t PrevSum = Sum;
452
453    Sum += Weight;
454    assert(Sum > PrevSum); (void) PrevSum;
455
456    if (Weight > MaxWeight) {
457      MaxWeight = Weight;
458      MaxSucc = Succ;
459    }
460  }
461
462  // Hot probability is at least 4/5 = 80%
463  if (BranchProbability(MaxWeight, Sum) > BranchProbability(4, 5))
464    return MaxSucc;
465
466  return 0;
467}
468
469// Return edge's weight. If can't find it, return DEFAULT_WEIGHT value.
470uint32_t BranchProbabilityInfo::
471getEdgeWeight(const BasicBlock *Src, const BasicBlock *Dst) const {
472  Edge E(Src, Dst);
473  DenseMap<Edge, uint32_t>::const_iterator I = Weights.find(E);
474
475  if (I != Weights.end())
476    return I->second;
477
478  return DEFAULT_WEIGHT;
479}
480
481void BranchProbabilityInfo::
482setEdgeWeight(const BasicBlock *Src, const BasicBlock *Dst, uint32_t Weight) {
483  Weights[std::make_pair(Src, Dst)] = Weight;
484  DEBUG(dbgs() << "set edge " << Src->getName() << " -> "
485               << Dst->getName() << " weight to " << Weight
486               << (isEdgeHot(Src, Dst) ? " [is HOT now]\n" : "\n"));
487}
488
489
490BranchProbability BranchProbabilityInfo::
491getEdgeProbability(const BasicBlock *Src, const BasicBlock *Dst) const {
492
493  uint32_t N = getEdgeWeight(Src, Dst);
494  uint32_t D = getSumForBlock(Src);
495
496  return BranchProbability(N, D);
497}
498
499raw_ostream &
500BranchProbabilityInfo::printEdgeProbability(raw_ostream &OS,
501                                            const BasicBlock *Src,
502                                            const BasicBlock *Dst) const {
503
504  const BranchProbability Prob = getEdgeProbability(Src, Dst);
505  OS << "edge " << Src->getName() << " -> " << Dst->getName()
506     << " probability is " << Prob
507     << (isEdgeHot(Src, Dst) ? " [HOT edge]\n" : "\n");
508
509  return OS;
510}
511