LLVM  7.0.0svn
JumpThreading.cpp
Go to the documentation of this file.
1 //===- JumpThreading.cpp - Thread control through conditional blocks ------===//
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 // This file implements the Jump Threading pass.
11 //
12 //===----------------------------------------------------------------------===//
13 
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/DenseSet.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Analysis/CFG.h"
30 #include "llvm/Analysis/Loads.h"
31 #include "llvm/Analysis/LoopInfo.h"
34 #include "llvm/IR/BasicBlock.h"
35 #include "llvm/IR/CFG.h"
36 #include "llvm/IR/Constant.h"
37 #include "llvm/IR/ConstantRange.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/Function.h"
42 #include "llvm/IR/InstrTypes.h"
43 #include "llvm/IR/Instruction.h"
44 #include "llvm/IR/Instructions.h"
45 #include "llvm/IR/IntrinsicInst.h"
46 #include "llvm/IR/Intrinsics.h"
47 #include "llvm/IR/LLVMContext.h"
48 #include "llvm/IR/MDBuilder.h"
49 #include "llvm/IR/Metadata.h"
50 #include "llvm/IR/Module.h"
51 #include "llvm/IR/PassManager.h"
52 #include "llvm/IR/PatternMatch.h"
53 #include "llvm/IR/Type.h"
54 #include "llvm/IR/Use.h"
55 #include "llvm/IR/User.h"
56 #include "llvm/IR/Value.h"
57 #include "llvm/Pass.h"
60 #include "llvm/Support/Casting.h"
62 #include "llvm/Support/Debug.h"
64 #include "llvm/Transforms/Scalar.h"
70 #include <algorithm>
71 #include <cassert>
72 #include <cstddef>
73 #include <cstdint>
74 #include <iterator>
75 #include <memory>
76 #include <utility>
77 
78 using namespace llvm;
79 using namespace jumpthreading;
80 
81 #define DEBUG_TYPE "jump-threading"
82 
83 STATISTIC(NumThreads, "Number of jumps threaded");
84 STATISTIC(NumFolds, "Number of terminators folded");
85 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
86 
87 static cl::opt<unsigned>
88 BBDuplicateThreshold("jump-threading-threshold",
89  cl::desc("Max block size to duplicate for jump threading"),
90  cl::init(6), cl::Hidden);
91 
92 static cl::opt<unsigned>
94  "jump-threading-implication-search-threshold",
95  cl::desc("The number of predecessors to search for a stronger "
96  "condition to use to thread over a weaker condition"),
97  cl::init(3), cl::Hidden);
98 
100  "print-lvi-after-jump-threading",
101  cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false),
102  cl::Hidden);
103 
104 namespace {
105 
106  /// This pass performs 'jump threading', which looks at blocks that have
107  /// multiple predecessors and multiple successors. If one or more of the
108  /// predecessors of the block can be proven to always jump to one of the
109  /// successors, we forward the edge from the predecessor to the successor by
110  /// duplicating the contents of this block.
111  ///
112  /// An example of when this can occur is code like this:
113  ///
114  /// if () { ...
115  /// X = 4;
116  /// }
117  /// if (X < 3) {
118  ///
119  /// In this case, the unconditional branch at the end of the first if can be
120  /// revectored to the false side of the second if.
121  class JumpThreading : public FunctionPass {
122  JumpThreadingPass Impl;
123 
124  public:
125  static char ID; // Pass identification
126 
127  JumpThreading(int T = -1) : FunctionPass(ID), Impl(T) {
129  }
130 
131  bool runOnFunction(Function &F) override;
132 
133  void getAnalysisUsage(AnalysisUsage &AU) const override {
141  }
142 
143  void releaseMemory() override { Impl.releaseMemory(); }
144  };
145 
146 } // end anonymous namespace
147 
148 char JumpThreading::ID = 0;
149 
150 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
151  "Jump Threading", false, false)
156 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
157  "Jump Threading", false, false)
158 
159 // Public interface to the Jump Threading pass
161  return new JumpThreading(Threshold);
162 }
163 
165  BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
166 }
167 
168 // Update branch probability information according to conditional
169 // branch probablity. This is usually made possible for cloned branches
170 // in inline instances by the context specific profile in the caller.
171 // For instance,
172 //
173 // [Block PredBB]
174 // [Branch PredBr]
175 // if (t) {
176 // Block A;
177 // } else {
178 // Block B;
179 // }
180 //
181 // [Block BB]
182 // cond = PN([true, %A], [..., %B]); // PHI node
183 // [Branch CondBr]
184 // if (cond) {
185 // ... // P(cond == true) = 1%
186 // }
187 //
188 // Here we know that when block A is taken, cond must be true, which means
189 // P(cond == true | A) = 1
190 //
191 // Given that P(cond == true) = P(cond == true | A) * P(A) +
192 // P(cond == true | B) * P(B)
193 // we get:
194 // P(cond == true ) = P(A) + P(cond == true | B) * P(B)
195 //
196 // which gives us:
197 // P(A) is less than P(cond == true), i.e.
198 // P(t == true) <= P(cond == true)
199 //
200 // In other words, if we know P(cond == true) is unlikely, we know
201 // that P(t == true) is also unlikely.
202 //
204  BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
205  if (!CondBr)
206  return;
207 
209  uint64_t TrueWeight, FalseWeight;
210  if (!CondBr->extractProfMetadata(TrueWeight, FalseWeight))
211  return;
212 
213  // Returns the outgoing edge of the dominating predecessor block
214  // that leads to the PhiNode's incoming block:
215  auto GetPredOutEdge =
216  [](BasicBlock *IncomingBB,
217  BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> {
218  auto *PredBB = IncomingBB;
219  auto *SuccBB = PhiBB;
220  while (true) {
221  BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
222  if (PredBr && PredBr->isConditional())
223  return {PredBB, SuccBB};
224  auto *SinglePredBB = PredBB->getSinglePredecessor();
225  if (!SinglePredBB)
226  return {nullptr, nullptr};
227  SuccBB = PredBB;
228  PredBB = SinglePredBB;
229  }
230  };
231 
232  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
233  Value *PhiOpnd = PN->getIncomingValue(i);
234  ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd);
235 
236  if (!CI || !CI->getType()->isIntegerTy(1))
237  continue;
238 
240  TrueWeight, TrueWeight + FalseWeight)
242  FalseWeight, TrueWeight + FalseWeight));
243 
244  auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB);
245  if (!PredOutEdge.first)
246  return;
247 
248  BasicBlock *PredBB = PredOutEdge.first;
249  BranchInst *PredBr = cast<BranchInst>(PredBB->getTerminator());
250 
251  uint64_t PredTrueWeight, PredFalseWeight;
252  // FIXME: We currently only set the profile data when it is missing.
253  // With PGO, this can be used to refine even existing profile data with
254  // context information. This needs to be done after more performance
255  // testing.
256  if (PredBr->extractProfMetadata(PredTrueWeight, PredFalseWeight))
257  continue;
258 
259  // We can not infer anything useful when BP >= 50%, because BP is the
260  // upper bound probability value.
261  if (BP >= BranchProbability(50, 100))
262  continue;
263 
264  SmallVector<uint32_t, 2> Weights;
265  if (PredBr->getSuccessor(0) == PredOutEdge.second) {
266  Weights.push_back(BP.getNumerator());
267  Weights.push_back(BP.getCompl().getNumerator());
268  } else {
269  Weights.push_back(BP.getCompl().getNumerator());
270  Weights.push_back(BP.getNumerator());
271  }
273  MDBuilder(PredBr->getParent()->getContext())
274  .createBranchWeights(Weights));
275  }
276 }
277 
278 /// runOnFunction - Toplevel algorithm.
280  if (skipFunction(F))
281  return false;
282  auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
283  // Get DT analysis before LVI. When LVI is initialized it conditionally adds
284  // DT if it's available.
285  auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
286  auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
287  auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
288  DeferredDominance DDT(*DT);
289  std::unique_ptr<BlockFrequencyInfo> BFI;
290  std::unique_ptr<BranchProbabilityInfo> BPI;
291  bool HasProfileData = F.hasProfileData();
292  if (HasProfileData) {
293  LoopInfo LI{DominatorTree(F)};
294  BPI.reset(new BranchProbabilityInfo(F, LI, TLI));
295  BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
296  }
297 
298  bool Changed = Impl.runImpl(F, TLI, LVI, AA, &DDT, HasProfileData,
299  std::move(BFI), std::move(BPI));
301  dbgs() << "LVI for function '" << F.getName() << "':\n";
302  LVI->printLVI(F, *DT, dbgs());
303  }
304  return Changed;
305 }
306 
309  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
310  // Get DT analysis before LVI. When LVI is initialized it conditionally adds
311  // DT if it's available.
312  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
313  auto &LVI = AM.getResult<LazyValueAnalysis>(F);
314  auto &AA = AM.getResult<AAManager>(F);
315  DeferredDominance DDT(DT);
316 
317  std::unique_ptr<BlockFrequencyInfo> BFI;
318  std::unique_ptr<BranchProbabilityInfo> BPI;
319  if (F.hasProfileData()) {
320  LoopInfo LI{DominatorTree(F)};
321  BPI.reset(new BranchProbabilityInfo(F, LI, &TLI));
322  BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
323  }
324 
325  bool Changed = runImpl(F, &TLI, &LVI, &AA, &DDT, HasProfileData,
326  std::move(BFI), std::move(BPI));
327 
328  if (!Changed)
329  return PreservedAnalyses::all();
331  PA.preserve<GlobalsAA>();
334  return PA;
335 }
336 
338  LazyValueInfo *LVI_, AliasAnalysis *AA_,
339  DeferredDominance *DDT_, bool HasProfileData_,
340  std::unique_ptr<BlockFrequencyInfo> BFI_,
341  std::unique_ptr<BranchProbabilityInfo> BPI_) {
342  DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
343  TLI = TLI_;
344  LVI = LVI_;
345  AA = AA_;
346  DDT = DDT_;
347  BFI.reset();
348  BPI.reset();
349  // When profile data is available, we need to update edge weights after
350  // successful jump threading, which requires both BPI and BFI being available.
351  HasProfileData = HasProfileData_;
352  auto *GuardDecl = F.getParent()->getFunction(
353  Intrinsic::getName(Intrinsic::experimental_guard));
354  HasGuards = GuardDecl && !GuardDecl->use_empty();
355  if (HasProfileData) {
356  BPI = std::move(BPI_);
357  BFI = std::move(BFI_);
358  }
359 
360  // Remove unreachable blocks from function as they may result in infinite
361  // loop. We do threading if we found something profitable. Jump threading a
362  // branch can create other opportunities. If these opportunities form a cycle
363  // i.e. if any jump threading is undoing previous threading in the path, then
364  // we will loop forever. We take care of this issue by not jump threading for
365  // back edges. This works for normal cases but not for unreachable blocks as
366  // they may have cycle with no back edge.
367  bool EverChanged = false;
368  EverChanged |= removeUnreachableBlocks(F, LVI, DDT);
369 
370  FindLoopHeaders(F);
371 
372  bool Changed;
373  do {
374  Changed = false;
375  for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
376  BasicBlock *BB = &*I;
377  // Thread all of the branches we can over this block.
378  while (ProcessBlock(BB))
379  Changed = true;
380 
381  ++I;
382 
383  // Don't thread branches over a block that's slated for deletion.
384  if (DDT->pendingDeletedBB(BB))
385  continue;
386 
387  // If the block is trivially dead, zap it. This eliminates the successor
388  // edges which simplifies the CFG.
389  if (pred_empty(BB) &&
390  BB != &BB->getParent()->getEntryBlock()) {
391  DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
392  << "' with terminator: " << *BB->getTerminator() << '\n');
393  LoopHeaders.erase(BB);
394  LVI->eraseBlock(BB);
395  DeleteDeadBlock(BB, DDT);
396  Changed = true;
397  continue;
398  }
399 
401 
402  // Can't thread an unconditional jump, but if the block is "almost
403  // empty", we can replace uses of it with uses of the successor and make
404  // this dead.
405  // We should not eliminate the loop header or latch either, because
406  // eliminating a loop header or latch might later prevent LoopSimplify
407  // from transforming nested loops into simplified form. We will rely on
408  // later passes in backend to clean up empty blocks.
409  if (BI && BI->isUnconditional() &&
410  BB != &BB->getParent()->getEntryBlock() &&
411  // If the terminator is the only non-phi instruction, try to nuke it.
412  BB->getFirstNonPHIOrDbg()->isTerminator() && !LoopHeaders.count(BB) &&
413  !LoopHeaders.count(BI->getSuccessor(0))) {
414  // FIXME: It is always conservatively correct to drop the info
415  // for a block even if it doesn't get erased. This isn't totally
416  // awesome, but it allows us to use AssertingVH to prevent nasty
417  // dangling pointer issues within LazyValueInfo.
418  LVI->eraseBlock(BB);
420  Changed = true;
421  }
422  }
423  EverChanged |= Changed;
424  } while (Changed);
425 
426  LoopHeaders.clear();
427  DDT->flush();
428  LVI->enableDT();
429  return EverChanged;
430 }
431 
432 // Replace uses of Cond with ToVal when safe to do so. If all uses are
433 // replaced, we can remove Cond. We cannot blindly replace all uses of Cond
434 // because we may incorrectly replace uses when guards/assumes are uses of
435 // of `Cond` and we used the guards/assume to reason about the `Cond` value
436 // at the end of block. RAUW unconditionally replaces all uses
437 // including the guards/assumes themselves and the uses before the
438 // guard/assume.
439 static void ReplaceFoldableUses(Instruction *Cond, Value *ToVal) {
440  assert(Cond->getType() == ToVal->getType());
441  auto *BB = Cond->getParent();
442  // We can unconditionally replace all uses in non-local blocks (i.e. uses
443  // strictly dominated by BB), since LVI information is true from the
444  // terminator of BB.
445  replaceNonLocalUsesWith(Cond, ToVal);
446  for (Instruction &I : reverse(*BB)) {
447  // Reached the Cond whose uses we are trying to replace, so there are no
448  // more uses.
449  if (&I == Cond)
450  break;
451  // We only replace uses in instructions that are guaranteed to reach the end
452  // of BB, where we know Cond is ToVal.
454  break;
455  I.replaceUsesOfWith(Cond, ToVal);
456  }
457  if (Cond->use_empty() && !Cond->mayHaveSideEffects())
458  Cond->eraseFromParent();
459 }
460 
461 /// Return the cost of duplicating a piece of this block from first non-phi
462 /// and before StopAt instruction to thread across it. Stop scanning the block
463 /// when exceeding the threshold. If duplication is impossible, returns ~0U.
465  Instruction *StopAt,
466  unsigned Threshold) {
467  assert(StopAt->getParent() == BB && "Not an instruction from proper BB?");
468  /// Ignore PHI nodes, these will be flattened when duplication happens.
470 
471  // FIXME: THREADING will delete values that are just used to compute the
472  // branch, so they shouldn't count against the duplication cost.
473 
474  unsigned Bonus = 0;
475  if (BB->getTerminator() == StopAt) {
476  // Threading through a switch statement is particularly profitable. If this
477  // block ends in a switch, decrease its cost to make it more likely to
478  // happen.
479  if (isa<SwitchInst>(StopAt))
480  Bonus = 6;
481 
482  // The same holds for indirect branches, but slightly more so.
483  if (isa<IndirectBrInst>(StopAt))
484  Bonus = 8;
485  }
486 
487  // Bump the threshold up so the early exit from the loop doesn't skip the
488  // terminator-based Size adjustment at the end.
489  Threshold += Bonus;
490 
491  // Sum up the cost of each instruction until we get to the terminator. Don't
492  // include the terminator because the copy won't include it.
493  unsigned Size = 0;
494  for (; &*I != StopAt; ++I) {
495 
496  // Stop scanning the block if we've reached the threshold.
497  if (Size > Threshold)
498  return Size;
499 
500  // Debugger intrinsics don't incur code size.
501  if (isa<DbgInfoIntrinsic>(I)) continue;
502 
503  // If this is a pointer->pointer bitcast, it is free.
504  if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
505  continue;
506 
507  // Bail out if this instruction gives back a token type, it is not possible
508  // to duplicate it if it is used outside this BB.
509  if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
510  return ~0U;
511 
512  // All other instructions count for at least one unit.
513  ++Size;
514 
515  // Calls are more expensive. If they are non-intrinsic calls, we model them
516  // as having cost of 4. If they are a non-vector intrinsic, we model them
517  // as having cost of 2 total, and if they are a vector intrinsic, we model
518  // them as having cost 1.
519  if (const CallInst *CI = dyn_cast<CallInst>(I)) {
520  if (CI->cannotDuplicate() || CI->isConvergent())
521  // Blocks with NoDuplicate are modelled as having infinite cost, so they
522  // are never duplicated.
523  return ~0U;
524  else if (!isa<IntrinsicInst>(CI))
525  Size += 3;
526  else if (!CI->getType()->isVectorTy())
527  Size += 1;
528  }
529  }
530 
531  return Size > Bonus ? Size - Bonus : 0;
532 }
533 
534 /// FindLoopHeaders - We do not want jump threading to turn proper loop
535 /// structures into irreducible loops. Doing this breaks up the loop nesting
536 /// hierarchy and pessimizes later transformations. To prevent this from
537 /// happening, we first have to find the loop headers. Here we approximate this
538 /// by finding targets of backedges in the CFG.
539 ///
540 /// Note that there definitely are cases when we want to allow threading of
541 /// edges across a loop header. For example, threading a jump from outside the
542 /// loop (the preheader) to an exit block of the loop is definitely profitable.
543 /// It is also almost always profitable to thread backedges from within the loop
544 /// to exit blocks, and is often profitable to thread backedges to other blocks
545 /// within the loop (forming a nested loop). This simple analysis is not rich
546 /// enough to track all of these properties and keep it up-to-date as the CFG
547 /// mutates, so we don't allow any of these transformations.
550  FindFunctionBackedges(F, Edges);
551 
552  for (const auto &Edge : Edges)
553  LoopHeaders.insert(Edge.second);
554 }
555 
556 /// getKnownConstant - Helper method to determine if we can thread over a
557 /// terminator with the given value as its condition, and if so what value to
558 /// use for that. What kind of value this is depends on whether we want an
559 /// integer or a block address, but an undef is always accepted.
560 /// Returns null if Val is null or not an appropriate constant.
562  if (!Val)
563  return nullptr;
564 
565  // Undef is "known" enough.
566  if (UndefValue *U = dyn_cast<UndefValue>(Val))
567  return U;
568 
569  if (Preference == WantBlockAddress)
570  return dyn_cast<BlockAddress>(Val->stripPointerCasts());
571 
572  return dyn_cast<ConstantInt>(Val);
573 }
574 
575 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
576 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
577 /// in any of our predecessors. If so, return the known list of value and pred
578 /// BB in the result vector.
579 ///
580 /// This returns true if there were any known values.
582  Value *V, BasicBlock *BB, PredValueInfo &Result,
584  // This method walks up use-def chains recursively. Because of this, we could
585  // get into an infinite loop going around loops in the use-def chain. To
586  // prevent this, keep track of what (value, block) pairs we've already visited
587  // and terminate the search if we loop back to them
588  if (!RecursionSet.insert(std::make_pair(V, BB)).second)
589  return false;
590 
591  // An RAII help to remove this pair from the recursion set once the recursion
592  // stack pops back out again.
593  RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
594 
595  // If V is a constant, then it is known in all predecessors.
596  if (Constant *KC = getKnownConstant(V, Preference)) {
597  for (BasicBlock *Pred : predecessors(BB))
598  Result.push_back(std::make_pair(KC, Pred));
599 
600  return !Result.empty();
601  }
602 
603  // If V is a non-instruction value, or an instruction in a different block,
604  // then it can't be derived from a PHI.
606  if (!I || I->getParent() != BB) {
607 
608  // Okay, if this is a live-in value, see if it has a known value at the end
609  // of any of our predecessors.
610  //
611  // FIXME: This should be an edge property, not a block end property.
612  /// TODO: Per PR2563, we could infer value range information about a
613  /// predecessor based on its terminator.
614  //
615  // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
616  // "I" is a non-local compare-with-a-constant instruction. This would be
617  // able to handle value inequalities better, for example if the compare is
618  // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
619  // Perhaps getConstantOnEdge should be smart enough to do this?
620 
621  if (DDT->pending())
622  LVI->disableDT();
623  else
624  LVI->enableDT();
625  for (BasicBlock *P : predecessors(BB)) {
626  // If the value is known by LazyValueInfo to be a constant in a
627  // predecessor, use that information to try to thread this block.
628  Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
629  if (Constant *KC = getKnownConstant(PredCst, Preference))
630  Result.push_back(std::make_pair(KC, P));
631  }
632 
633  return !Result.empty();
634  }
635 
636  /// If I is a PHI node, then we know the incoming values for any constants.
637  if (PHINode *PN = dyn_cast<PHINode>(I)) {
638  if (DDT->pending())
639  LVI->disableDT();
640  else
641  LVI->enableDT();
642  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
643  Value *InVal = PN->getIncomingValue(i);
644  if (Constant *KC = getKnownConstant(InVal, Preference)) {
645  Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
646  } else {
647  Constant *CI = LVI->getConstantOnEdge(InVal,
648  PN->getIncomingBlock(i),
649  BB, CxtI);
650  if (Constant *KC = getKnownConstant(CI, Preference))
651  Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
652  }
653  }
654 
655  return !Result.empty();
656  }
657 
658  // Handle Cast instructions. Only see through Cast when the source operand is
659  // PHI or Cmp and the source type is i1 to save the compilation time.
660  if (CastInst *CI = dyn_cast<CastInst>(I)) {
661  Value *Source = CI->getOperand(0);
662  if (!Source->getType()->isIntegerTy(1))
663  return false;
664  if (!isa<PHINode>(Source) && !isa<CmpInst>(Source))
665  return false;
666  ComputeValueKnownInPredecessors(Source, BB, Result, Preference, CxtI);
667  if (Result.empty())
668  return false;
669 
670  // Convert the known values.
671  for (auto &R : Result)
672  R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
673 
674  return true;
675  }
676 
677  // Handle some boolean conditions.
678  if (I->getType()->getPrimitiveSizeInBits() == 1) {
679  assert(Preference == WantInteger && "One-bit non-integer type?");
680  // X | true -> true
681  // X & false -> false
682  if (I->getOpcode() == Instruction::Or ||
683  I->getOpcode() == Instruction::And) {
684  PredValueInfoTy LHSVals, RHSVals;
685 
686  ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
687  WantInteger, CxtI);
688  ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
689  WantInteger, CxtI);
690 
691  if (LHSVals.empty() && RHSVals.empty())
692  return false;
693 
694  ConstantInt *InterestingVal;
695  if (I->getOpcode() == Instruction::Or)
696  InterestingVal = ConstantInt::getTrue(I->getContext());
697  else
698  InterestingVal = ConstantInt::getFalse(I->getContext());
699 
700  SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
701 
702  // Scan for the sentinel. If we find an undef, force it to the
703  // interesting value: x|undef -> true and x&undef -> false.
704  for (const auto &LHSVal : LHSVals)
705  if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
706  Result.emplace_back(InterestingVal, LHSVal.second);
707  LHSKnownBBs.insert(LHSVal.second);
708  }
709  for (const auto &RHSVal : RHSVals)
710  if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
711  // If we already inferred a value for this block on the LHS, don't
712  // re-add it.
713  if (!LHSKnownBBs.count(RHSVal.second))
714  Result.emplace_back(InterestingVal, RHSVal.second);
715  }
716 
717  return !Result.empty();
718  }
719 
720  // Handle the NOT form of XOR.
721  if (I->getOpcode() == Instruction::Xor &&
722  isa<ConstantInt>(I->getOperand(1)) &&
723  cast<ConstantInt>(I->getOperand(1))->isOne()) {
724  ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
725  WantInteger, CxtI);
726  if (Result.empty())
727  return false;
728 
729  // Invert the known values.
730  for (auto &R : Result)
731  R.first = ConstantExpr::getNot(R.first);
732 
733  return true;
734  }
735 
736  // Try to simplify some other binary operator values.
737  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
738  assert(Preference != WantBlockAddress
739  && "A binary operator creating a block address?");
740  if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
741  PredValueInfoTy LHSVals;
742  ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
743  WantInteger, CxtI);
744 
745  // Try to use constant folding to simplify the binary operator.
746  for (const auto &LHSVal : LHSVals) {
747  Constant *V = LHSVal.first;
748  Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
749 
750  if (Constant *KC = getKnownConstant(Folded, WantInteger))
751  Result.push_back(std::make_pair(KC, LHSVal.second));
752  }
753  }
754 
755  return !Result.empty();
756  }
757 
758  // Handle compare with phi operand, where the PHI is defined in this block.
759  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
760  assert(Preference == WantInteger && "Compares only produce integers");
761  Type *CmpType = Cmp->getType();
762  Value *CmpLHS = Cmp->getOperand(0);
763  Value *CmpRHS = Cmp->getOperand(1);
764  CmpInst::Predicate Pred = Cmp->getPredicate();
765 
766  PHINode *PN = dyn_cast<PHINode>(CmpLHS);
767  if (PN && PN->getParent() == BB) {
768  const DataLayout &DL = PN->getModule()->getDataLayout();
769  // We can do this simplification if any comparisons fold to true or false.
770  // See if any do.
771  if (DDT->pending())
772  LVI->disableDT();
773  else
774  LVI->enableDT();
775  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
776  BasicBlock *PredBB = PN->getIncomingBlock(i);
777  Value *LHS = PN->getIncomingValue(i);
778  Value *RHS = CmpRHS->DoPHITranslation(BB, PredBB);
779 
780  Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
781  if (!Res) {
782  if (!isa<Constant>(RHS))
783  continue;
784 
786  ResT = LVI->getPredicateOnEdge(Pred, LHS,
787  cast<Constant>(RHS), PredBB, BB,
788  CxtI ? CxtI : Cmp);
789  if (ResT == LazyValueInfo::Unknown)
790  continue;
791  Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
792  }
793 
794  if (Constant *KC = getKnownConstant(Res, WantInteger))
795  Result.push_back(std::make_pair(KC, PredBB));
796  }
797 
798  return !Result.empty();
799  }
800 
801  // If comparing a live-in value against a constant, see if we know the
802  // live-in value on any predecessors.
803  if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
804  Constant *CmpConst = cast<Constant>(CmpRHS);
805 
806  if (!isa<Instruction>(CmpLHS) ||
807  cast<Instruction>(CmpLHS)->getParent() != BB) {
808  if (DDT->pending())
809  LVI->disableDT();
810  else
811  LVI->enableDT();
812  for (BasicBlock *P : predecessors(BB)) {
813  // If the value is known by LazyValueInfo to be a constant in a
814  // predecessor, use that information to try to thread this block.
816  LVI->getPredicateOnEdge(Pred, CmpLHS,
817  CmpConst, P, BB, CxtI ? CxtI : Cmp);
818  if (Res == LazyValueInfo::Unknown)
819  continue;
820 
821  Constant *ResC = ConstantInt::get(CmpType, Res);
822  Result.push_back(std::make_pair(ResC, P));
823  }
824 
825  return !Result.empty();
826  }
827 
828  // InstCombine can fold some forms of constant range checks into
829  // (icmp (add (x, C1)), C2). See if we have we have such a thing with
830  // x as a live-in.
831  {
832  using namespace PatternMatch;
833 
834  Value *AddLHS;
835  ConstantInt *AddConst;
836  if (isa<ConstantInt>(CmpConst) &&
837  match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
838  if (!isa<Instruction>(AddLHS) ||
839  cast<Instruction>(AddLHS)->getParent() != BB) {
840  if (DDT->pending())
841  LVI->disableDT();
842  else
843  LVI->enableDT();
844  for (BasicBlock *P : predecessors(BB)) {
845  // If the value is known by LazyValueInfo to be a ConstantRange in
846  // a predecessor, use that information to try to thread this
847  // block.
848  ConstantRange CR = LVI->getConstantRangeOnEdge(
849  AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
850  // Propagate the range through the addition.
851  CR = CR.add(AddConst->getValue());
852 
853  // Get the range where the compare returns true.
855  Pred, cast<ConstantInt>(CmpConst)->getValue());
856 
857  Constant *ResC;
858  if (CmpRange.contains(CR))
859  ResC = ConstantInt::getTrue(CmpType);
860  else if (CmpRange.inverse().contains(CR))
861  ResC = ConstantInt::getFalse(CmpType);
862  else
863  continue;
864 
865  Result.push_back(std::make_pair(ResC, P));
866  }
867 
868  return !Result.empty();
869  }
870  }
871  }
872 
873  // Try to find a constant value for the LHS of a comparison,
874  // and evaluate it statically if we can.
875  PredValueInfoTy LHSVals;
876  ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
877  WantInteger, CxtI);
878 
879  for (const auto &LHSVal : LHSVals) {
880  Constant *V = LHSVal.first;
881  Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
882  if (Constant *KC = getKnownConstant(Folded, WantInteger))
883  Result.push_back(std::make_pair(KC, LHSVal.second));
884  }
885 
886  return !Result.empty();
887  }
888  }
889 
890  if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
891  // Handle select instructions where at least one operand is a known constant
892  // and we can figure out the condition value for any predecessor block.
893  Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
894  Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
895  PredValueInfoTy Conds;
896  if ((TrueVal || FalseVal) &&
897  ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
898  WantInteger, CxtI)) {
899  for (auto &C : Conds) {
900  Constant *Cond = C.first;
901 
902  // Figure out what value to use for the condition.
903  bool KnownCond;
904  if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
905  // A known boolean.
906  KnownCond = CI->isOne();
907  } else {
908  assert(isa<UndefValue>(Cond) && "Unexpected condition value");
909  // Either operand will do, so be sure to pick the one that's a known
910  // constant.
911  // FIXME: Do this more cleverly if both values are known constants?
912  KnownCond = (TrueVal != nullptr);
913  }
914 
915  // See if the select has a known constant value for this predecessor.
916  if (Constant *Val = KnownCond ? TrueVal : FalseVal)
917  Result.push_back(std::make_pair(Val, C.second));
918  }
919 
920  return !Result.empty();
921  }
922  }
923 
924  // If all else fails, see if LVI can figure out a constant value for us.
925  if (DDT->pending())
926  LVI->disableDT();
927  else
928  LVI->enableDT();
929  Constant *CI = LVI->getConstant(V, BB, CxtI);
930  if (Constant *KC = getKnownConstant(CI, Preference)) {
931  for (BasicBlock *Pred : predecessors(BB))
932  Result.push_back(std::make_pair(KC, Pred));
933  }
934 
935  return !Result.empty();
936 }
937 
938 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
939 /// in an undefined jump, decide which block is best to revector to.
940 ///
941 /// Since we can pick an arbitrary destination, we pick the successor with the
942 /// fewest predecessors. This should reduce the in-degree of the others.
943 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
944  TerminatorInst *BBTerm = BB->getTerminator();
945  unsigned MinSucc = 0;
946  BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
947  // Compute the successor with the minimum number of predecessors.
948  unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
949  for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
950  TestBB = BBTerm->getSuccessor(i);
951  unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
952  if (NumPreds < MinNumPreds) {
953  MinSucc = i;
954  MinNumPreds = NumPreds;
955  }
956  }
957 
958  return MinSucc;
959 }
960 
962  if (!BB->hasAddressTaken()) return false;
963 
964  // If the block has its address taken, it may be a tree of dead constants
965  // hanging off of it. These shouldn't keep the block alive.
968  return !BA->use_empty();
969 }
970 
971 /// ProcessBlock - If there are any predecessors whose control can be threaded
972 /// through to a successor, transform them now.
974  // If the block is trivially dead, just return and let the caller nuke it.
975  // This simplifies other transformations.
976  if (DDT->pendingDeletedBB(BB) ||
977  (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
978  return false;
979 
980  // If this block has a single predecessor, and if that pred has a single
981  // successor, merge the blocks. This encourages recursive jump threading
982  // because now the condition in this block can be threaded through
983  // predecessors of our predecessor block.
984  if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
985  const TerminatorInst *TI = SinglePred->getTerminator();
986  if (!TI->isExceptional() && TI->getNumSuccessors() == 1 &&
987  SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
988  // If SinglePred was a loop header, BB becomes one.
989  if (LoopHeaders.erase(SinglePred))
990  LoopHeaders.insert(BB);
991 
992  LVI->eraseBlock(SinglePred);
993  MergeBasicBlockIntoOnlyPred(BB, nullptr, DDT);
994 
995  // Now that BB is merged into SinglePred (i.e. SinglePred Code followed by
996  // BB code within one basic block `BB`), we need to invalidate the LVI
997  // information associated with BB, because the LVI information need not be
998  // true for all of BB after the merge. For example,
999  // Before the merge, LVI info and code is as follows:
1000  // SinglePred: <LVI info1 for %p val>
1001  // %y = use of %p
1002  // call @exit() // need not transfer execution to successor.
1003  // assume(%p) // from this point on %p is true
1004  // br label %BB
1005  // BB: <LVI info2 for %p val, i.e. %p is true>
1006  // %x = use of %p
1007  // br label exit
1008  //
1009  // Note that this LVI info for blocks BB and SinglPred is correct for %p
1010  // (info2 and info1 respectively). After the merge and the deletion of the
1011  // LVI info1 for SinglePred. We have the following code:
1012  // BB: <LVI info2 for %p val>
1013  // %y = use of %p
1014  // call @exit()
1015  // assume(%p)
1016  // %x = use of %p <-- LVI info2 is correct from here onwards.
1017  // br label exit
1018  // LVI info2 for BB is incorrect at the beginning of BB.
1019 
1020  // Invalidate LVI information for BB if the LVI is not provably true for
1021  // all of BB.
1022  if (any_of(*BB, [](Instruction &I) {
1024  }))
1025  LVI->eraseBlock(BB);
1026  return true;
1027  }
1028  }
1029 
1030  if (TryToUnfoldSelectInCurrBB(BB))
1031  return true;
1032 
1033  // Look if we can propagate guards to predecessors.
1034  if (HasGuards && ProcessGuards(BB))
1035  return true;
1036 
1037  // What kind of constant we're looking for.
1039 
1040  // Look to see if the terminator is a conditional branch, switch or indirect
1041  // branch, if not we can't thread it.
1042  Value *Condition;
1044  if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
1045  // Can't thread an unconditional jump.
1046  if (BI->isUnconditional()) return false;
1047  Condition = BI->getCondition();
1048  } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
1049  Condition = SI->getCondition();
1050  } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
1051  // Can't thread indirect branch with no successors.
1052  if (IB->getNumSuccessors() == 0) return false;
1053  Condition = IB->getAddress()->stripPointerCasts();
1054  Preference = WantBlockAddress;
1055  } else {
1056  return false; // Must be an invoke.
1057  }
1058 
1059  // Run constant folding to see if we can reduce the condition to a simple
1060  // constant.
1061  if (Instruction *I = dyn_cast<Instruction>(Condition)) {
1062  Value *SimpleVal =
1064  if (SimpleVal) {
1065  I->replaceAllUsesWith(SimpleVal);
1066  if (isInstructionTriviallyDead(I, TLI))
1067  I->eraseFromParent();
1068  Condition = SimpleVal;
1069  }
1070  }
1071 
1072  // If the terminator is branching on an undef, we can pick any of the
1073  // successors to branch to. Let GetBestDestForJumpOnUndef decide.
1074  if (isa<UndefValue>(Condition)) {
1075  unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
1076  std::vector<DominatorTree::UpdateType> Updates;
1077 
1078  // Fold the branch/switch.
1079  TerminatorInst *BBTerm = BB->getTerminator();
1080  Updates.reserve(BBTerm->getNumSuccessors());
1081  for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1082  if (i == BestSucc) continue;
1083  BasicBlock *Succ = BBTerm->getSuccessor(i);
1084  Succ->removePredecessor(BB, true);
1085  Updates.push_back({DominatorTree::Delete, BB, Succ});
1086  }
1087 
1088  DEBUG(dbgs() << " In block '" << BB->getName()
1089  << "' folding undef terminator: " << *BBTerm << '\n');
1090  BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
1091  BBTerm->eraseFromParent();
1092  DDT->applyUpdates(Updates);
1093  return true;
1094  }
1095 
1096  // If the terminator of this block is branching on a constant, simplify the
1097  // terminator to an unconditional branch. This can occur due to threading in
1098  // other blocks.
1099  if (getKnownConstant(Condition, Preference)) {
1100  DEBUG(dbgs() << " In block '" << BB->getName()
1101  << "' folding terminator: " << *BB->getTerminator() << '\n');
1102  ++NumFolds;
1103  ConstantFoldTerminator(BB, true, nullptr, DDT);
1104  return true;
1105  }
1106 
1107  Instruction *CondInst = dyn_cast<Instruction>(Condition);
1108 
1109  // All the rest of our checks depend on the condition being an instruction.
1110  if (!CondInst) {
1111  // FIXME: Unify this with code below.
1112  if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
1113  return true;
1114  return false;
1115  }
1116 
1117  if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
1118  // If we're branching on a conditional, LVI might be able to determine
1119  // it's value at the branch instruction. We only handle comparisons
1120  // against a constant at this time.
1121  // TODO: This should be extended to handle switches as well.
1122  BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1123  Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
1124  if (CondBr && CondConst) {
1125  // We should have returned as soon as we turn a conditional branch to
1126  // unconditional. Because its no longer interesting as far as jump
1127  // threading is concerned.
1128  assert(CondBr->isConditional() && "Threading on unconditional terminator");
1129 
1130  if (DDT->pending())
1131  LVI->disableDT();
1132  else
1133  LVI->enableDT();
1135  LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
1136  CondConst, CondBr);
1137  if (Ret != LazyValueInfo::Unknown) {
1138  unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
1139  unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
1140  BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove);
1141  ToRemoveSucc->removePredecessor(BB, true);
1142  BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
1143  CondBr->eraseFromParent();
1144  if (CondCmp->use_empty())
1145  CondCmp->eraseFromParent();
1146  // We can safely replace *some* uses of the CondInst if it has
1147  // exactly one value as returned by LVI. RAUW is incorrect in the
1148  // presence of guards and assumes, that have the `Cond` as the use. This
1149  // is because we use the guards/assume to reason about the `Cond` value
1150  // at the end of block, but RAUW unconditionally replaces all uses
1151  // including the guards/assumes themselves and the uses before the
1152  // guard/assume.
1153  else if (CondCmp->getParent() == BB) {
1154  auto *CI = Ret == LazyValueInfo::True ?
1155  ConstantInt::getTrue(CondCmp->getType()) :
1156  ConstantInt::getFalse(CondCmp->getType());
1157  ReplaceFoldableUses(CondCmp, CI);
1158  }
1159  DDT->deleteEdge(BB, ToRemoveSucc);
1160  return true;
1161  }
1162 
1163  // We did not manage to simplify this branch, try to see whether
1164  // CondCmp depends on a known phi-select pattern.
1165  if (TryToUnfoldSelect(CondCmp, BB))
1166  return true;
1167  }
1168  }
1169 
1170  // Check for some cases that are worth simplifying. Right now we want to look
1171  // for loads that are used by a switch or by the condition for the branch. If
1172  // we see one, check to see if it's partially redundant. If so, insert a PHI
1173  // which can then be used to thread the values.
1174  Value *SimplifyValue = CondInst;
1175  if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
1176  if (isa<Constant>(CondCmp->getOperand(1)))
1177  SimplifyValue = CondCmp->getOperand(0);
1178 
1179  // TODO: There are other places where load PRE would be profitable, such as
1180  // more complex comparisons.
1181  if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue))
1182  if (SimplifyPartiallyRedundantLoad(LoadI))
1183  return true;
1184 
1185  // Before threading, try to propagate profile data backwards:
1186  if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1187  if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1189 
1190  // Handle a variety of cases where we are branching on something derived from
1191  // a PHI node in the current block. If we can prove that any predecessors
1192  // compute a predictable value based on a PHI node, thread those predecessors.
1193  if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
1194  return true;
1195 
1196  // If this is an otherwise-unfoldable branch on a phi node in the current
1197  // block, see if we can simplify.
1198  if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1199  if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1200  return ProcessBranchOnPHI(PN);
1201 
1202  // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1203  if (CondInst->getOpcode() == Instruction::Xor &&
1204  CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1205  return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
1206 
1207  // Search for a stronger dominating condition that can be used to simplify a
1208  // conditional branch leaving BB.
1209  if (ProcessImpliedCondition(BB))
1210  return true;
1211 
1212  return false;
1213 }
1214 
1216  auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
1217  if (!BI || !BI->isConditional())
1218  return false;
1219 
1220  Value *Cond = BI->getCondition();
1221  BasicBlock *CurrentBB = BB;
1222  BasicBlock *CurrentPred = BB->getSinglePredecessor();
1223  unsigned Iter = 0;
1224 
1225  auto &DL = BB->getModule()->getDataLayout();
1226 
1227  while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1228  auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
1229  if (!PBI || !PBI->isConditional())
1230  return false;
1231  if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
1232  return false;
1233 
1234  bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB;
1235  Optional<bool> Implication =
1236  isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue);
1237  if (Implication) {
1238  BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1);
1239  BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0);
1240  RemoveSucc->removePredecessor(BB);
1241  BranchInst::Create(KeepSucc, BI);
1242  BI->eraseFromParent();
1243  DDT->deleteEdge(BB, RemoveSucc);
1244  return true;
1245  }
1246  CurrentBB = CurrentPred;
1247  CurrentPred = CurrentBB->getSinglePredecessor();
1248  }
1249 
1250  return false;
1251 }
1252 
1253 /// Return true if Op is an instruction defined in the given block.
1255  if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1256  if (OpInst->getParent() == BB)
1257  return true;
1258  return false;
1259 }
1260 
1261 /// SimplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1262 /// redundant load instruction, eliminate it by replacing it with a PHI node.
1263 /// This is an important optimization that encourages jump threading, and needs
1264 /// to be run interlaced with other jump threading tasks.
1266  // Don't hack volatile and ordered loads.
1267  if (!LoadI->isUnordered()) return false;
1268 
1269  // If the load is defined in a block with exactly one predecessor, it can't be
1270  // partially redundant.
1271  BasicBlock *LoadBB = LoadI->getParent();
1272  if (LoadBB->getSinglePredecessor())
1273  return false;
1274 
1275  // If the load is defined in an EH pad, it can't be partially redundant,
1276  // because the edges between the invoke and the EH pad cannot have other
1277  // instructions between them.
1278  if (LoadBB->isEHPad())
1279  return false;
1280 
1281  Value *LoadedPtr = LoadI->getOperand(0);
1282 
1283  // If the loaded operand is defined in the LoadBB and its not a phi,
1284  // it can't be available in predecessors.
1285  if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
1286  return false;
1287 
1288  // Scan a few instructions up from the load, to see if it is obviously live at
1289  // the entry to its block.
1290  BasicBlock::iterator BBIt(LoadI);
1291  bool IsLoadCSE;
1292  if (Value *AvailableVal = FindAvailableLoadedValue(
1293  LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1294  // If the value of the load is locally available within the block, just use
1295  // it. This frequently occurs for reg2mem'd allocas.
1296 
1297  if (IsLoadCSE) {
1298  LoadInst *NLoadI = cast<LoadInst>(AvailableVal);
1299  combineMetadataForCSE(NLoadI, LoadI);
1300  };
1301 
1302  // If the returned value is the load itself, replace with an undef. This can
1303  // only happen in dead loops.
1304  if (AvailableVal == LoadI)
1305  AvailableVal = UndefValue::get(LoadI->getType());
1306  if (AvailableVal->getType() != LoadI->getType())
1307  AvailableVal = CastInst::CreateBitOrPointerCast(
1308  AvailableVal, LoadI->getType(), "", LoadI);
1309  LoadI->replaceAllUsesWith(AvailableVal);
1310  LoadI->eraseFromParent();
1311  return true;
1312  }
1313 
1314  // Otherwise, if we scanned the whole block and got to the top of the block,
1315  // we know the block is locally transparent to the load. If not, something
1316  // might clobber its value.
1317  if (BBIt != LoadBB->begin())
1318  return false;
1319 
1320  // If all of the loads and stores that feed the value have the same AA tags,
1321  // then we can propagate them onto any newly inserted loads.
1322  AAMDNodes AATags;
1323  LoadI->getAAMetadata(AATags);
1324 
1325  SmallPtrSet<BasicBlock*, 8> PredsScanned;
1326 
1327  using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1328 
1329  AvailablePredsTy AvailablePreds;
1330  BasicBlock *OneUnavailablePred = nullptr;
1331  SmallVector<LoadInst*, 8> CSELoads;
1332 
1333  // If we got here, the loaded value is transparent through to the start of the
1334  // block. Check to see if it is available in any of the predecessor blocks.
1335  for (BasicBlock *PredBB : predecessors(LoadBB)) {
1336  // If we already scanned this predecessor, skip it.
1337  if (!PredsScanned.insert(PredBB).second)
1338  continue;
1339 
1340  BBIt = PredBB->end();
1341  unsigned NumScanedInst = 0;
1342  Value *PredAvailable = nullptr;
1343  // NOTE: We don't CSE load that is volatile or anything stronger than
1344  // unordered, that should have been checked when we entered the function.
1345  assert(LoadI->isUnordered() &&
1346  "Attempting to CSE volatile or atomic loads");
1347  // If this is a load on a phi pointer, phi-translate it and search
1348  // for available load/store to the pointer in predecessors.
1349  Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1350  PredAvailable = FindAvailablePtrLoadStore(
1351  Ptr, LoadI->getType(), LoadI->isAtomic(), PredBB, BBIt,
1352  DefMaxInstsToScan, AA, &IsLoadCSE, &NumScanedInst);
1353 
1354  // If PredBB has a single predecessor, continue scanning through the
1355  // single precessor.
1356  BasicBlock *SinglePredBB = PredBB;
1357  while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1358  NumScanedInst < DefMaxInstsToScan) {
1359  SinglePredBB = SinglePredBB->getSinglePredecessor();
1360  if (SinglePredBB) {
1361  BBIt = SinglePredBB->end();
1362  PredAvailable = FindAvailablePtrLoadStore(
1363  Ptr, LoadI->getType(), LoadI->isAtomic(), SinglePredBB, BBIt,
1364  (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1365  &NumScanedInst);
1366  }
1367  }
1368 
1369  if (!PredAvailable) {
1370  OneUnavailablePred = PredBB;
1371  continue;
1372  }
1373 
1374  if (IsLoadCSE)
1375  CSELoads.push_back(cast<LoadInst>(PredAvailable));
1376 
1377  // If so, this load is partially redundant. Remember this info so that we
1378  // can create a PHI node.
1379  AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1380  }
1381 
1382  // If the loaded value isn't available in any predecessor, it isn't partially
1383  // redundant.
1384  if (AvailablePreds.empty()) return false;
1385 
1386  // Okay, the loaded value is available in at least one (and maybe all!)
1387  // predecessors. If the value is unavailable in more than one unique
1388  // predecessor, we want to insert a merge block for those common predecessors.
1389  // This ensures that we only have to insert one reload, thus not increasing
1390  // code size.
1391  BasicBlock *UnavailablePred = nullptr;
1392 
1393  // If the value is unavailable in one of predecessors, we will end up
1394  // inserting a new instruction into them. It is only valid if all the
1395  // instructions before LoadI are guaranteed to pass execution to its
1396  // successor, or if LoadI is safe to speculate.
1397  // TODO: If this logic becomes more complex, and we will perform PRE insertion
1398  // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1399  // It requires domination tree analysis, so for this simple case it is an
1400  // overkill.
1401  if (PredsScanned.size() != AvailablePreds.size() &&
1403  for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1405  return false;
1406 
1407  // If there is exactly one predecessor where the value is unavailable, the
1408  // already computed 'OneUnavailablePred' block is it. If it ends in an
1409  // unconditional branch, we know that it isn't a critical edge.
1410  if (PredsScanned.size() == AvailablePreds.size()+1 &&
1411  OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1412  UnavailablePred = OneUnavailablePred;
1413  } else if (PredsScanned.size() != AvailablePreds.size()) {
1414  // Otherwise, we had multiple unavailable predecessors or we had a critical
1415  // edge from the one.
1416  SmallVector<BasicBlock*, 8> PredsToSplit;
1417  SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1418 
1419  for (const auto &AvailablePred : AvailablePreds)
1420  AvailablePredSet.insert(AvailablePred.first);
1421 
1422  // Add all the unavailable predecessors to the PredsToSplit list.
1423  for (BasicBlock *P : predecessors(LoadBB)) {
1424  // If the predecessor is an indirect goto, we can't split the edge.
1425  if (isa<IndirectBrInst>(P->getTerminator()))
1426  return false;
1427 
1428  if (!AvailablePredSet.count(P))
1429  PredsToSplit.push_back(P);
1430  }
1431 
1432  // Split them out to their own block.
1433  UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1434  }
1435 
1436  // If the value isn't available in all predecessors, then there will be
1437  // exactly one where it isn't available. Insert a load on that edge and add
1438  // it to the AvailablePreds list.
1439  if (UnavailablePred) {
1440  assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1441  "Can't handle critical edge here!");
1442  LoadInst *NewVal =
1443  new LoadInst(LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1444  LoadI->getName() + ".pr", false, LoadI->getAlignment(),
1445  LoadI->getOrdering(), LoadI->getSyncScopeID(),
1446  UnavailablePred->getTerminator());
1447  NewVal->setDebugLoc(LoadI->getDebugLoc());
1448  if (AATags)
1449  NewVal->setAAMetadata(AATags);
1450 
1451  AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1452  }
1453 
1454  // Now we know that each predecessor of this block has a value in
1455  // AvailablePreds, sort them for efficient access as we're walking the preds.
1456  array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1457 
1458  // Create a PHI node at the start of the block for the PRE'd load value.
1459  pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1460  PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "",
1461  &LoadBB->front());
1462  PN->takeName(LoadI);
1463  PN->setDebugLoc(LoadI->getDebugLoc());
1464 
1465  // Insert new entries into the PHI for each predecessor. A single block may
1466  // have multiple entries here.
1467  for (pred_iterator PI = PB; PI != PE; ++PI) {
1468  BasicBlock *P = *PI;
1469  AvailablePredsTy::iterator I =
1470  std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1471  std::make_pair(P, (Value*)nullptr));
1472 
1473  assert(I != AvailablePreds.end() && I->first == P &&
1474  "Didn't find entry for predecessor!");
1475 
1476  // If we have an available predecessor but it requires casting, insert the
1477  // cast in the predecessor and use the cast. Note that we have to update the
1478  // AvailablePreds vector as we go so that all of the PHI entries for this
1479  // predecessor use the same bitcast.
1480  Value *&PredV = I->second;
1481  if (PredV->getType() != LoadI->getType())
1482  PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "",
1483  P->getTerminator());
1484 
1485  PN->addIncoming(PredV, I->first);
1486  }
1487 
1488  for (LoadInst *PredLoadI : CSELoads) {
1489  combineMetadataForCSE(PredLoadI, LoadI);
1490  }
1491 
1492  LoadI->replaceAllUsesWith(PN);
1493  LoadI->eraseFromParent();
1494 
1495  return true;
1496 }
1497 
1498 /// FindMostPopularDest - The specified list contains multiple possible
1499 /// threadable destinations. Pick the one that occurs the most frequently in
1500 /// the list.
1501 static BasicBlock *
1503  const SmallVectorImpl<std::pair<BasicBlock *,
1504  BasicBlock *>> &PredToDestList) {
1505  assert(!PredToDestList.empty());
1506 
1507  // Determine popularity. If there are multiple possible destinations, we
1508  // explicitly choose to ignore 'undef' destinations. We prefer to thread
1509  // blocks with known and real destinations to threading undef. We'll handle
1510  // them later if interesting.
1511  DenseMap<BasicBlock*, unsigned> DestPopularity;
1512  for (const auto &PredToDest : PredToDestList)
1513  if (PredToDest.second)
1514  DestPopularity[PredToDest.second]++;
1515 
1516  // Find the most popular dest.
1517  DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1518  BasicBlock *MostPopularDest = DPI->first;
1519  unsigned Popularity = DPI->second;
1520  SmallVector<BasicBlock*, 4> SamePopularity;
1521 
1522  for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1523  // If the popularity of this entry isn't higher than the popularity we've
1524  // seen so far, ignore it.
1525  if (DPI->second < Popularity)
1526  ; // ignore.
1527  else if (DPI->second == Popularity) {
1528  // If it is the same as what we've seen so far, keep track of it.
1529  SamePopularity.push_back(DPI->first);
1530  } else {
1531  // If it is more popular, remember it.
1532  SamePopularity.clear();
1533  MostPopularDest = DPI->first;
1534  Popularity = DPI->second;
1535  }
1536  }
1537 
1538  // Okay, now we know the most popular destination. If there is more than one
1539  // destination, we need to determine one. This is arbitrary, but we need
1540  // to make a deterministic decision. Pick the first one that appears in the
1541  // successor list.
1542  if (!SamePopularity.empty()) {
1543  SamePopularity.push_back(MostPopularDest);
1544  TerminatorInst *TI = BB->getTerminator();
1545  for (unsigned i = 0; ; ++i) {
1546  assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1547 
1548  if (!is_contained(SamePopularity, TI->getSuccessor(i)))
1549  continue;
1550 
1551  MostPopularDest = TI->getSuccessor(i);
1552  break;
1553  }
1554  }
1555 
1556  // Okay, we have finally picked the most popular destination.
1557  return MostPopularDest;
1558 }
1559 
1562  Instruction *CxtI) {
1563  // If threading this would thread across a loop header, don't even try to
1564  // thread the edge.
1565  if (LoopHeaders.count(BB))
1566  return false;
1567 
1568  PredValueInfoTy PredValues;
1569  if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1570  return false;
1571 
1572  assert(!PredValues.empty() &&
1573  "ComputeValueKnownInPredecessors returned true with no values");
1574 
1575  DEBUG(dbgs() << "IN BB: " << *BB;
1576  for (const auto &PredValue : PredValues) {
1577  dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1578  << *PredValue.first
1579  << " for pred '" << PredValue.second->getName() << "'.\n";
1580  });
1581 
1582  // Decide what we want to thread through. Convert our list of known values to
1583  // a list of known destinations for each pred. This also discards duplicate
1584  // predecessors and keeps track of the undefined inputs (which are represented
1585  // as a null dest in the PredToDestList).
1586  SmallPtrSet<BasicBlock*, 16> SeenPreds;
1588 
1589  BasicBlock *OnlyDest = nullptr;
1590  BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1591  Constant *OnlyVal = nullptr;
1592  Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1593 
1594  unsigned PredWithKnownDest = 0;
1595  for (const auto &PredValue : PredValues) {
1596  BasicBlock *Pred = PredValue.second;
1597  if (!SeenPreds.insert(Pred).second)
1598  continue; // Duplicate predecessor entry.
1599 
1600  Constant *Val = PredValue.first;
1601 
1602  BasicBlock *DestBB;
1603  if (isa<UndefValue>(Val))
1604  DestBB = nullptr;
1605  else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1606  assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1607  DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1608  } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1609  assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1610  DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1611  } else {
1612  assert(isa<IndirectBrInst>(BB->getTerminator())
1613  && "Unexpected terminator");
1614  assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress");
1615  DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1616  }
1617 
1618  // If we have exactly one destination, remember it for efficiency below.
1619  if (PredToDestList.empty()) {
1620  OnlyDest = DestBB;
1621  OnlyVal = Val;
1622  } else {
1623  if (OnlyDest != DestBB)
1624  OnlyDest = MultipleDestSentinel;
1625  // It possible we have same destination, but different value, e.g. default
1626  // case in switchinst.
1627  if (Val != OnlyVal)
1628  OnlyVal = MultipleVal;
1629  }
1630 
1631  // We know where this predecessor is going.
1632  ++PredWithKnownDest;
1633 
1634  // If the predecessor ends with an indirect goto, we can't change its
1635  // destination.
1636  if (isa<IndirectBrInst>(Pred->getTerminator()))
1637  continue;
1638 
1639  PredToDestList.push_back(std::make_pair(Pred, DestBB));
1640  }
1641 
1642  // If all edges were unthreadable, we fail.
1643  if (PredToDestList.empty())
1644  return false;
1645 
1646  // If all the predecessors go to a single known successor, we want to fold,
1647  // not thread. By doing so, we do not need to duplicate the current block and
1648  // also miss potential opportunities in case we dont/cant duplicate.
1649  if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1650  if (PredWithKnownDest ==
1651  (size_t)std::distance(pred_begin(BB), pred_end(BB))) {
1652  bool SeenFirstBranchToOnlyDest = false;
1653  std::vector <DominatorTree::UpdateType> Updates;
1654  Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);
1655  for (BasicBlock *SuccBB : successors(BB)) {
1656  if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1657  SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1658  } else {
1659  SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1660  Updates.push_back({DominatorTree::Delete, BB, SuccBB});
1661  }
1662  }
1663 
1664  // Finally update the terminator.
1665  TerminatorInst *Term = BB->getTerminator();
1666  BranchInst::Create(OnlyDest, Term);
1667  Term->eraseFromParent();
1668  DDT->applyUpdates(Updates);
1669 
1670  // If the condition is now dead due to the removal of the old terminator,
1671  // erase it.
1672  if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1673  if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1674  CondInst->eraseFromParent();
1675  // We can safely replace *some* uses of the CondInst if it has
1676  // exactly one value as returned by LVI. RAUW is incorrect in the
1677  // presence of guards and assumes, that have the `Cond` as the use. This
1678  // is because we use the guards/assume to reason about the `Cond` value
1679  // at the end of block, but RAUW unconditionally replaces all uses
1680  // including the guards/assumes themselves and the uses before the
1681  // guard/assume.
1682  else if (OnlyVal && OnlyVal != MultipleVal &&
1683  CondInst->getParent() == BB)
1684  ReplaceFoldableUses(CondInst, OnlyVal);
1685  }
1686  return true;
1687  }
1688  }
1689 
1690  // Determine which is the most common successor. If we have many inputs and
1691  // this block is a switch, we want to start by threading the batch that goes
1692  // to the most popular destination first. If we only know about one
1693  // threadable destination (the common case) we can avoid this.
1694  BasicBlock *MostPopularDest = OnlyDest;
1695 
1696  if (MostPopularDest == MultipleDestSentinel)
1697  MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1698 
1699  // Now that we know what the most popular destination is, factor all
1700  // predecessors that will jump to it into a single predecessor.
1701  SmallVector<BasicBlock*, 16> PredsToFactor;
1702  for (const auto &PredToDest : PredToDestList)
1703  if (PredToDest.second == MostPopularDest) {
1704  BasicBlock *Pred = PredToDest.first;
1705 
1706  // This predecessor may be a switch or something else that has multiple
1707  // edges to the block. Factor each of these edges by listing them
1708  // according to # occurrences in PredsToFactor.
1709  for (BasicBlock *Succ : successors(Pred))
1710  if (Succ == BB)
1711  PredsToFactor.push_back(Pred);
1712  }
1713 
1714  // If the threadable edges are branching on an undefined value, we get to pick
1715  // the destination that these predecessors should get to.
1716  if (!MostPopularDest)
1717  MostPopularDest = BB->getTerminator()->
1718  getSuccessor(GetBestDestForJumpOnUndef(BB));
1719 
1720  // Ok, try to thread it!
1721  return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1722 }
1723 
1724 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1725 /// a PHI node in the current block. See if there are any simplifications we
1726 /// can do based on inputs to the phi node.
1728  BasicBlock *BB = PN->getParent();
1729 
1730  // TODO: We could make use of this to do it once for blocks with common PHI
1731  // values.
1733  PredBBs.resize(1);
1734 
1735  // If any of the predecessor blocks end in an unconditional branch, we can
1736  // *duplicate* the conditional branch into that block in order to further
1737  // encourage jump threading and to eliminate cases where we have branch on a
1738  // phi of an icmp (branch on icmp is much better).
1739  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1740  BasicBlock *PredBB = PN->getIncomingBlock(i);
1741  if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1742  if (PredBr->isUnconditional()) {
1743  PredBBs[0] = PredBB;
1744  // Try to duplicate BB into PredBB.
1745  if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1746  return true;
1747  }
1748  }
1749 
1750  return false;
1751 }
1752 
1753 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1754 /// a xor instruction in the current block. See if there are any
1755 /// simplifications we can do based on inputs to the xor.
1757  BasicBlock *BB = BO->getParent();
1758 
1759  // If either the LHS or RHS of the xor is a constant, don't do this
1760  // optimization.
1761  if (isa<ConstantInt>(BO->getOperand(0)) ||
1762  isa<ConstantInt>(BO->getOperand(1)))
1763  return false;
1764 
1765  // If the first instruction in BB isn't a phi, we won't be able to infer
1766  // anything special about any particular predecessor.
1767  if (!isa<PHINode>(BB->front()))
1768  return false;
1769 
1770  // If this BB is a landing pad, we won't be able to split the edge into it.
1771  if (BB->isEHPad())
1772  return false;
1773 
1774  // If we have a xor as the branch input to this block, and we know that the
1775  // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1776  // the condition into the predecessor and fix that value to true, saving some
1777  // logical ops on that path and encouraging other paths to simplify.
1778  //
1779  // This copies something like this:
1780  //
1781  // BB:
1782  // %X = phi i1 [1], [%X']
1783  // %Y = icmp eq i32 %A, %B
1784  // %Z = xor i1 %X, %Y
1785  // br i1 %Z, ...
1786  //
1787  // Into:
1788  // BB':
1789  // %Y = icmp ne i32 %A, %B
1790  // br i1 %Y, ...
1791 
1792  PredValueInfoTy XorOpValues;
1793  bool isLHS = true;
1794  if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1795  WantInteger, BO)) {
1796  assert(XorOpValues.empty());
1797  if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1798  WantInteger, BO))
1799  return false;
1800  isLHS = false;
1801  }
1802 
1803  assert(!XorOpValues.empty() &&
1804  "ComputeValueKnownInPredecessors returned true with no values");
1805 
1806  // Scan the information to see which is most popular: true or false. The
1807  // predecessors can be of the set true, false, or undef.
1808  unsigned NumTrue = 0, NumFalse = 0;
1809  for (const auto &XorOpValue : XorOpValues) {
1810  if (isa<UndefValue>(XorOpValue.first))
1811  // Ignore undefs for the count.
1812  continue;
1813  if (cast<ConstantInt>(XorOpValue.first)->isZero())
1814  ++NumFalse;
1815  else
1816  ++NumTrue;
1817  }
1818 
1819  // Determine which value to split on, true, false, or undef if neither.
1820  ConstantInt *SplitVal = nullptr;
1821  if (NumTrue > NumFalse)
1822  SplitVal = ConstantInt::getTrue(BB->getContext());
1823  else if (NumTrue != 0 || NumFalse != 0)
1824  SplitVal = ConstantInt::getFalse(BB->getContext());
1825 
1826  // Collect all of the blocks that this can be folded into so that we can
1827  // factor this once and clone it once.
1828  SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1829  for (const auto &XorOpValue : XorOpValues) {
1830  if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1831  continue;
1832 
1833  BlocksToFoldInto.push_back(XorOpValue.second);
1834  }
1835 
1836  // If we inferred a value for all of the predecessors, then duplication won't
1837  // help us. However, we can just replace the LHS or RHS with the constant.
1838  if (BlocksToFoldInto.size() ==
1839  cast<PHINode>(BB->front()).getNumIncomingValues()) {
1840  if (!SplitVal) {
1841  // If all preds provide undef, just nuke the xor, because it is undef too.
1843  BO->eraseFromParent();
1844  } else if (SplitVal->isZero()) {
1845  // If all preds provide 0, replace the xor with the other input.
1846  BO->replaceAllUsesWith(BO->getOperand(isLHS));
1847  BO->eraseFromParent();
1848  } else {
1849  // If all preds provide 1, set the computed value to 1.
1850  BO->setOperand(!isLHS, SplitVal);
1851  }
1852 
1853  return true;
1854  }
1855 
1856  // Try to duplicate BB into PredBB.
1857  return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1858 }
1859 
1860 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1861 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1862 /// NewPred using the entries from OldPred (suitably mapped).
1864  BasicBlock *OldPred,
1865  BasicBlock *NewPred,
1867  for (PHINode &PN : PHIBB->phis()) {
1868  // Ok, we have a PHI node. Figure out what the incoming value was for the
1869  // DestBlock.
1870  Value *IV = PN.getIncomingValueForBlock(OldPred);
1871 
1872  // Remap the value if necessary.
1873  if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1875  if (I != ValueMap.end())
1876  IV = I->second;
1877  }
1878 
1879  PN.addIncoming(IV, NewPred);
1880  }
1881 }
1882 
1883 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1884 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1885 /// across BB. Transform the IR to reflect this change.
1887  const SmallVectorImpl<BasicBlock *> &PredBBs,
1888  BasicBlock *SuccBB) {
1889  // If threading to the same block as we come from, we would infinite loop.
1890  if (SuccBB == BB) {
1891  DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1892  << "' - would thread to self!\n");
1893  return false;
1894  }
1895 
1896  // If threading this would thread across a loop header, don't thread the edge.
1897  // See the comments above FindLoopHeaders for justifications and caveats.
1898  if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
1899  DEBUG({
1900  bool BBIsHeader = LoopHeaders.count(BB);
1901  bool SuccIsHeader = LoopHeaders.count(SuccBB);
1902  dbgs() << " Not threading across "
1903  << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName()
1904  << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '")
1905  << SuccBB->getName() << "' - it might create an irreducible loop!\n";
1906  });
1907  return false;
1908  }
1909 
1910  unsigned JumpThreadCost =
1911  getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1912  if (JumpThreadCost > BBDupThreshold) {
1913  DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1914  << "' - Cost is too high: " << JumpThreadCost << "\n");
1915  return false;
1916  }
1917 
1918  // And finally, do it! Start by factoring the predecessors if needed.
1919  BasicBlock *PredBB;
1920  if (PredBBs.size() == 1)
1921  PredBB = PredBBs[0];
1922  else {
1923  DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1924  << " common predecessors.\n");
1925  PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1926  }
1927 
1928  // And finally, do it!
1929  DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1930  << SuccBB->getName() << "' with cost: " << JumpThreadCost
1931  << ", across block:\n "
1932  << *BB << "\n");
1933 
1934  if (DDT->pending())
1935  LVI->disableDT();
1936  else
1937  LVI->enableDT();
1938  LVI->threadEdge(PredBB, BB, SuccBB);
1939 
1940  // We are going to have to map operands from the original BB block to the new
1941  // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1942  // account for entry from PredBB.
1944 
1945  BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1946  BB->getName()+".thread",
1947  BB->getParent(), BB);
1948  NewBB->moveAfter(PredBB);
1949 
1950  // Set the block frequency of NewBB.
1951  if (HasProfileData) {
1952  auto NewBBFreq =
1953  BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1954  BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1955  }
1956 
1957  BasicBlock::iterator BI = BB->begin();
1958  for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1959  ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1960 
1961  // Clone the non-phi instructions of BB into NewBB, keeping track of the
1962  // mapping and using it to remap operands in the cloned instructions.
1963  for (; !isa<TerminatorInst>(BI); ++BI) {
1964  Instruction *New = BI->clone();
1965  New->setName(BI->getName());
1966  NewBB->getInstList().push_back(New);
1967  ValueMapping[&*BI] = New;
1968 
1969  // Remap operands to patch up intra-block references.
1970  for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1971  if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1972  DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1973  if (I != ValueMapping.end())
1974  New->setOperand(i, I->second);
1975  }
1976  }
1977 
1978  // We didn't copy the terminator from BB over to NewBB, because there is now
1979  // an unconditional jump to SuccBB. Insert the unconditional jump.
1980  BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
1981  NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1982 
1983  // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1984  // PHI nodes for NewBB now.
1985  AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1986 
1987  // If there were values defined in BB that are used outside the block, then we
1988  // now have to update all uses of the value to use either the original value,
1989  // the cloned value, or some PHI derived value. This can require arbitrary
1990  // PHI insertion, of which we are prepared to do, clean these up now.
1991  SSAUpdater SSAUpdate;
1992  SmallVector<Use*, 16> UsesToRename;
1993  for (Instruction &I : *BB) {
1994  // Scan all uses of this instruction to see if it is used outside of its
1995  // block, and if so, record them in UsesToRename.
1996  for (Use &U : I.uses()) {
1997  Instruction *User = cast<Instruction>(U.getUser());
1998  if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1999  if (UserPN->getIncomingBlock(U) == BB)
2000  continue;
2001  } else if (User->getParent() == BB)
2002  continue;
2003 
2004  UsesToRename.push_back(&U);
2005  }
2006 
2007  // If there are no uses outside the block, we're done with this instruction.
2008  if (UsesToRename.empty())
2009  continue;
2010 
2011  DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
2012 
2013  // We found a use of I outside of BB. Rename all uses of I that are outside
2014  // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2015  // with the two values we know.
2016  SSAUpdate.Initialize(I.getType(), I.getName());
2017  SSAUpdate.AddAvailableValue(BB, &I);
2018  SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
2019 
2020  while (!UsesToRename.empty())
2021  SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2022  DEBUG(dbgs() << "\n");
2023  }
2024 
2025  // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
2026  // NewBB instead of BB. This eliminates predecessors from BB, which requires
2027  // us to simplify any PHI nodes in BB.
2028  TerminatorInst *PredTerm = PredBB->getTerminator();
2029  for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2030  if (PredTerm->getSuccessor(i) == BB) {
2031  BB->removePredecessor(PredBB, true);
2032  PredTerm->setSuccessor(i, NewBB);
2033  }
2034 
2035  DDT->applyUpdates({{DominatorTree::Insert, NewBB, SuccBB},
2036  {DominatorTree::Insert, PredBB, NewBB},
2037  {DominatorTree::Delete, PredBB, BB}});
2038 
2039  // At this point, the IR is fully up to date and consistent. Do a quick scan
2040  // over the new instructions and zap any that are constants or dead. This
2041  // frequently happens because of phi translation.
2042  SimplifyInstructionsInBlock(NewBB, TLI);
2043 
2044  // Update the edge weight from BB to SuccBB, which should be less than before.
2045  UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
2046 
2047  // Threaded an edge!
2048  ++NumThreads;
2049  return true;
2050 }
2051 
2052 /// Create a new basic block that will be the predecessor of BB and successor of
2053 /// all blocks in Preds. When profile data is available, update the frequency of
2054 /// this new block.
2055 BasicBlock *JumpThreadingPass::SplitBlockPreds(BasicBlock *BB,
2056  ArrayRef<BasicBlock *> Preds,
2057  const char *Suffix) {
2059 
2060  // Collect the frequencies of all predecessors of BB, which will be used to
2061  // update the edge weight of the result of splitting predecessors.
2063  if (HasProfileData)
2064  for (auto Pred : Preds)
2065  FreqMap.insert(std::make_pair(
2066  Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
2067 
2068  // In the case when BB is a LandingPad block we create 2 new predecessors
2069  // instead of just one.
2070  if (BB->isLandingPad()) {
2071  std::string NewName = std::string(Suffix) + ".split-lp";
2072  SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);
2073  } else {
2074  NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));
2075  }
2076 
2077  std::vector<DominatorTree::UpdateType> Updates;
2078  Updates.reserve((2 * Preds.size()) + NewBBs.size());
2079  for (auto NewBB : NewBBs) {
2080  BlockFrequency NewBBFreq(0);
2081  Updates.push_back({DominatorTree::Insert, NewBB, BB});
2082  for (auto Pred : predecessors(NewBB)) {
2083  Updates.push_back({DominatorTree::Delete, Pred, BB});
2084  Updates.push_back({DominatorTree::Insert, Pred, NewBB});
2085  if (HasProfileData) // Update frequencies between Pred -> NewBB.
2086  NewBBFreq += FreqMap.lookup(Pred);
2087  }
2088  if (HasProfileData) // Apply the summed frequency to NewBB.
2089  BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2090  }
2091 
2092  DDT->applyUpdates(Updates);
2093  return NewBBs[0];
2094 }
2095 
2096 bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2097  const TerminatorInst *TI = BB->getTerminator();
2098  assert(TI->getNumSuccessors() > 1 && "not a split");
2099 
2100  MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
2101  if (!WeightsNode)
2102  return false;
2103 
2104  MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
2105  if (MDName->getString() != "branch_weights")
2106  return false;
2107 
2108  // Ensure there are weights for all of the successors. Note that the first
2109  // operand to the metadata node is a name, not a weight.
2110  return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
2111 }
2112 
2113 /// Update the block frequency of BB and branch weight and the metadata on the
2114 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2115 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
2116 void JumpThreadingPass::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2117  BasicBlock *BB,
2118  BasicBlock *NewBB,
2119  BasicBlock *SuccBB) {
2120  if (!HasProfileData)
2121  return;
2122 
2123  assert(BFI && BPI && "BFI & BPI should have been created here");
2124 
2125  // As the edge from PredBB to BB is deleted, we have to update the block
2126  // frequency of BB.
2127  auto BBOrigFreq = BFI->getBlockFreq(BB);
2128  auto NewBBFreq = BFI->getBlockFreq(NewBB);
2129  auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
2130  auto BBNewFreq = BBOrigFreq - NewBBFreq;
2131  BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
2132 
2133  // Collect updated outgoing edges' frequencies from BB and use them to update
2134  // edge probabilities.
2135  SmallVector<uint64_t, 4> BBSuccFreq;
2136  for (BasicBlock *Succ : successors(BB)) {
2137  auto SuccFreq = (Succ == SuccBB)
2138  ? BB2SuccBBFreq - NewBBFreq
2139  : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
2140  BBSuccFreq.push_back(SuccFreq.getFrequency());
2141  }
2142 
2143  uint64_t MaxBBSuccFreq =
2144  *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
2145 
2147  if (MaxBBSuccFreq == 0)
2148  BBSuccProbs.assign(BBSuccFreq.size(),
2149  {1, static_cast<unsigned>(BBSuccFreq.size())});
2150  else {
2151  for (uint64_t Freq : BBSuccFreq)
2152  BBSuccProbs.push_back(
2153  BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
2154  // Normalize edge probabilities so that they sum up to one.
2155  BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
2156  BBSuccProbs.end());
2157  }
2158 
2159  // Update edge probabilities in BPI.
2160  for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
2161  BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
2162 
2163  // Update the profile metadata as well.
2164  //
2165  // Don't do this if the profile of the transformed blocks was statically
2166  // estimated. (This could occur despite the function having an entry
2167  // frequency in completely cold parts of the CFG.)
2168  //
2169  // In this case we don't want to suggest to subsequent passes that the
2170  // calculated weights are fully consistent. Consider this graph:
2171  //
2172  // check_1
2173  // 50% / |
2174  // eq_1 | 50%
2175  // \ |
2176  // check_2
2177  // 50% / |
2178  // eq_2 | 50%
2179  // \ |
2180  // check_3
2181  // 50% / |
2182  // eq_3 | 50%
2183  // \ |
2184  //
2185  // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2186  // the overall probabilities are inconsistent; the total probability that the
2187  // value is either 1, 2 or 3 is 150%.
2188  //
2189  // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2190  // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2191  // the loop exit edge. Then based solely on static estimation we would assume
2192  // the loop was extremely hot.
2193  //
2194  // FIXME this locally as well so that BPI and BFI are consistent as well. We
2195  // shouldn't make edges extremely likely or unlikely based solely on static
2196  // estimation.
2197  if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
2198  SmallVector<uint32_t, 4> Weights;
2199  for (auto Prob : BBSuccProbs)
2200  Weights.push_back(Prob.getNumerator());
2201 
2202  auto TI = BB->getTerminator();
2203  TI->setMetadata(
2205  MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
2206  }
2207 }
2208 
2209 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2210 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2211 /// If we can duplicate the contents of BB up into PredBB do so now, this
2212 /// improves the odds that the branch will be on an analyzable instruction like
2213 /// a compare.
2215  BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2216  assert(!PredBBs.empty() && "Can't handle an empty set");
2217 
2218  // If BB is a loop header, then duplicating this block outside the loop would
2219  // cause us to transform this into an irreducible loop, don't do this.
2220  // See the comments above FindLoopHeaders for justifications and caveats.
2221  if (LoopHeaders.count(BB)) {
2222  DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
2223  << "' into predecessor block '" << PredBBs[0]->getName()
2224  << "' - it might create an irreducible loop!\n");
2225  return false;
2226  }
2227 
2228  unsigned DuplicationCost =
2229  getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2230  if (DuplicationCost > BBDupThreshold) {
2231  DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
2232  << "' - Cost is too high: " << DuplicationCost << "\n");
2233  return false;
2234  }
2235 
2236  // And finally, do it! Start by factoring the predecessors if needed.
2237  std::vector<DominatorTree::UpdateType> Updates;
2238  BasicBlock *PredBB;
2239  if (PredBBs.size() == 1)
2240  PredBB = PredBBs[0];
2241  else {
2242  DEBUG(dbgs() << " Factoring out " << PredBBs.size()
2243  << " common predecessors.\n");
2244  PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
2245  }
2246  Updates.push_back({DominatorTree::Delete, PredBB, BB});
2247 
2248  // Okay, we decided to do this! Clone all the instructions in BB onto the end
2249  // of PredBB.
2250  DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
2251  << PredBB->getName() << "' to eliminate branch on phi. Cost: "
2252  << DuplicationCost << " block is:" << *BB << "\n");
2253 
2254  // Unless PredBB ends with an unconditional branch, split the edge so that we
2255  // can just clone the bits from BB into the end of the new PredBB.
2256  BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2257 
2258  if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2259  BasicBlock *OldPredBB = PredBB;
2260  PredBB = SplitEdge(OldPredBB, BB);
2261  Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});
2262  Updates.push_back({DominatorTree::Insert, PredBB, BB});
2263  Updates.push_back({DominatorTree::Delete, OldPredBB, BB});
2264  OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
2265  }
2266 
2267  // We are going to have to map operands from the original BB block into the
2268  // PredBB block. Evaluate PHI nodes in BB.
2270 
2271  BasicBlock::iterator BI = BB->begin();
2272  for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2273  ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2274  // Clone the non-phi instructions of BB into PredBB, keeping track of the
2275  // mapping and using it to remap operands in the cloned instructions.
2276  for (; BI != BB->end(); ++BI) {
2277  Instruction *New = BI->clone();
2278 
2279  // Remap operands to patch up intra-block references.
2280  for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2281  if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2282  DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2283  if (I != ValueMapping.end())
2284  New->setOperand(i, I->second);
2285  }
2286 
2287  // If this instruction can be simplified after the operands are updated,
2288  // just use the simplified value instead. This frequently happens due to
2289  // phi translation.
2290  if (Value *IV = SimplifyInstruction(
2291  New,
2292  {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2293  ValueMapping[&*BI] = IV;
2294  if (!New->mayHaveSideEffects()) {
2295  New->deleteValue();
2296  New = nullptr;
2297  }
2298  } else {
2299  ValueMapping[&*BI] = New;
2300  }
2301  if (New) {
2302  // Otherwise, insert the new instruction into the block.
2303  New->setName(BI->getName());
2304  PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2305  // Update Dominance from simplified New instruction operands.
2306  for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2307  if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))
2308  Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});
2309  }
2310  }
2311 
2312  // Check to see if the targets of the branch had PHI nodes. If so, we need to
2313  // add entries to the PHI nodes for branch from PredBB now.
2314  BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2315  AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2316  ValueMapping);
2317  AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2318  ValueMapping);
2319 
2320  // If there were values defined in BB that are used outside the block, then we
2321  // now have to update all uses of the value to use either the original value,
2322  // the cloned value, or some PHI derived value. This can require arbitrary
2323  // PHI insertion, of which we are prepared to do, clean these up now.
2324  SSAUpdater SSAUpdate;
2325  SmallVector<Use*, 16> UsesToRename;
2326  for (Instruction &I : *BB) {
2327  // Scan all uses of this instruction to see if it is used outside of its
2328  // block, and if so, record them in UsesToRename.
2329  for (Use &U : I.uses()) {
2330  Instruction *User = cast<Instruction>(U.getUser());
2331  if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2332  if (UserPN->getIncomingBlock(U) == BB)
2333  continue;
2334  } else if (User->getParent() == BB)
2335  continue;
2336 
2337  UsesToRename.push_back(&U);
2338  }
2339 
2340  // If there are no uses outside the block, we're done with this instruction.
2341  if (UsesToRename.empty())
2342  continue;
2343 
2344  DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
2345 
2346  // We found a use of I outside of BB. Rename all uses of I that are outside
2347  // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2348  // with the two values we know.
2349  SSAUpdate.Initialize(I.getType(), I.getName());
2350  SSAUpdate.AddAvailableValue(BB, &I);
2351  SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]);
2352 
2353  while (!UsesToRename.empty())
2354  SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2355  DEBUG(dbgs() << "\n");
2356  }
2357 
2358  // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2359  // that we nuked.
2360  BB->removePredecessor(PredBB, true);
2361 
2362  // Remove the unconditional branch at the end of the PredBB block.
2363  OldPredBranch->eraseFromParent();
2364  DDT->applyUpdates(Updates);
2365 
2366  ++NumDupes;
2367  return true;
2368 }
2369 
2370 /// TryToUnfoldSelect - Look for blocks of the form
2371 /// bb1:
2372 /// %a = select
2373 /// br bb2
2374 ///
2375 /// bb2:
2376 /// %p = phi [%a, %bb1] ...
2377 /// %c = icmp %p
2378 /// br i1 %c
2379 ///
2380 /// And expand the select into a branch structure if one of its arms allows %c
2381 /// to be folded. This later enables threading from bb1 over bb2.
2383  BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2384  PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2385  Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2386 
2387  if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2388  CondLHS->getParent() != BB)
2389  return false;
2390 
2391  for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2392  BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2393  SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2394 
2395  // Look if one of the incoming values is a select in the corresponding
2396  // predecessor.
2397  if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2398  continue;
2399 
2400  BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2401  if (!PredTerm || !PredTerm->isUnconditional())
2402  continue;
2403 
2404  // Now check if one of the select values would allow us to constant fold the
2405  // terminator in BB. We don't do the transform if both sides fold, those
2406  // cases will be threaded in any case.
2407  if (DDT->pending())
2408  LVI->disableDT();
2409  else
2410  LVI->enableDT();
2411  LazyValueInfo::Tristate LHSFolds =
2412  LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2413  CondRHS, Pred, BB, CondCmp);
2414  LazyValueInfo::Tristate RHSFolds =
2415  LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2416  CondRHS, Pred, BB, CondCmp);
2417  if ((LHSFolds != LazyValueInfo::Unknown ||
2418  RHSFolds != LazyValueInfo::Unknown) &&
2419  LHSFolds != RHSFolds) {
2420  // Expand the select.
2421  //
2422  // Pred --
2423  // | v
2424  // | NewBB
2425  // | |
2426  // |-----
2427  // v
2428  // BB
2429  BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2430  BB->getParent(), BB);
2431  // Move the unconditional branch to NewBB.
2432  PredTerm->removeFromParent();
2433  NewBB->getInstList().insert(NewBB->end(), PredTerm);
2434  // Create a conditional branch and update PHI nodes.
2435  BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2436  CondLHS->setIncomingValue(I, SI->getFalseValue());
2437  CondLHS->addIncoming(SI->getTrueValue(), NewBB);
2438  // The select is now dead.
2439  SI->eraseFromParent();
2440 
2441  DDT->applyUpdates({{DominatorTree::Insert, NewBB, BB},
2442  {DominatorTree::Insert, Pred, NewBB}});
2443  // Update any other PHI nodes in BB.
2444  for (BasicBlock::iterator BI = BB->begin();
2445  PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2446  if (Phi != CondLHS)
2447  Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2448  return true;
2449  }
2450  }
2451  return false;
2452 }
2453 
2454 /// TryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2455 /// same BB in the form
2456 /// bb:
2457 /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2458 /// %s = select %p, trueval, falseval
2459 ///
2460 /// or
2461 ///
2462 /// bb:
2463 /// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2464 /// %c = cmp %p, 0
2465 /// %s = select %c, trueval, falseval
2466 ///
2467 /// And expand the select into a branch structure. This later enables
2468 /// jump-threading over bb in this pass.
2469 ///
2470 /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2471 /// select if the associated PHI has at least one constant. If the unfolded
2472 /// select is not jump-threaded, it will be folded again in the later
2473 /// optimizations.
2475  // If threading this would thread across a loop header, don't thread the edge.
2476  // See the comments above FindLoopHeaders for justifications and caveats.
2477  if (LoopHeaders.count(BB))
2478  return false;
2479 
2480  for (BasicBlock::iterator BI = BB->begin();
2481  PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2482  // Look for a Phi having at least one constant incoming value.
2483  if (llvm::all_of(PN->incoming_values(),
2484  [](Value *V) { return !isa<ConstantInt>(V); }))
2485  continue;
2486 
2487  auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2488  // Check if SI is in BB and use V as condition.
2489  if (SI->getParent() != BB)
2490  return false;
2491  Value *Cond = SI->getCondition();
2492  return (Cond && Cond == V && Cond->getType()->isIntegerTy(1));
2493  };
2494 
2495  SelectInst *SI = nullptr;
2496  for (Use &U : PN->uses()) {
2497  if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2498  // Look for a ICmp in BB that compares PN with a constant and is the
2499  // condition of a Select.
2500  if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2501  isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2502  if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2503  if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2504  SI = SelectI;
2505  break;
2506  }
2507  } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2508  // Look for a Select in BB that uses PN as condtion.
2509  if (isUnfoldCandidate(SelectI, U.get())) {
2510  SI = SelectI;
2511  break;
2512  }
2513  }
2514  }
2515 
2516  if (!SI)
2517  continue;
2518  // Expand the select.
2519  TerminatorInst *Term =
2520  SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
2521  BasicBlock *SplitBB = SI->getParent();
2522  BasicBlock *NewBB = Term->getParent();
2523  PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2524  NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2525  NewPN->addIncoming(SI->getFalseValue(), BB);
2526  SI->replaceAllUsesWith(NewPN);
2527  SI->eraseFromParent();
2528  // NewBB and SplitBB are newly created blocks which require insertion.
2529  std::vector<DominatorTree::UpdateType> Updates;
2530  Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2531  Updates.push_back({DominatorTree::Insert, BB, SplitBB});
2532  Updates.push_back({DominatorTree::Insert, BB, NewBB});
2533  Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});
2534  // BB's successors were moved to SplitBB, update DDT accordingly.
2535  for (auto *Succ : successors(SplitBB)) {
2536  Updates.push_back({DominatorTree::Delete, BB, Succ});
2537  Updates.push_back({DominatorTree::Insert, SplitBB, Succ});
2538  }
2539  DDT->applyUpdates(Updates);
2540  return true;
2541  }
2542  return false;
2543 }
2544 
2545 /// Try to propagate a guard from the current BB into one of its predecessors
2546 /// in case if another branch of execution implies that the condition of this
2547 /// guard is always true. Currently we only process the simplest case that
2548 /// looks like:
2549 ///
2550 /// Start:
2551 /// %cond = ...
2552 /// br i1 %cond, label %T1, label %F1
2553 /// T1:
2554 /// br label %Merge
2555 /// F1:
2556 /// br label %Merge
2557 /// Merge:
2558 /// %condGuard = ...
2559 /// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2560 ///
2561 /// And cond either implies condGuard or !condGuard. In this case all the
2562 /// instructions before the guard can be duplicated in both branches, and the
2563 /// guard is then threaded to one of them.
2565  using namespace PatternMatch;
2566 
2567  // We only want to deal with two predecessors.
2568  BasicBlock *Pred1, *Pred2;
2569  auto PI = pred_begin(BB), PE = pred_end(BB);
2570  if (PI == PE)
2571  return false;
2572  Pred1 = *PI++;
2573  if (PI == PE)
2574  return false;
2575  Pred2 = *PI++;
2576  if (PI != PE)
2577  return false;
2578  if (Pred1 == Pred2)
2579  return false;
2580 
2581  // Try to thread one of the guards of the block.
2582  // TODO: Look up deeper than to immediate predecessor?
2583  auto *Parent = Pred1->getSinglePredecessor();
2584  if (!Parent || Parent != Pred2->getSinglePredecessor())
2585  return false;
2586 
2587  if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2588  for (auto &I : *BB)
2589  if (match(&I, m_Intrinsic<Intrinsic::experimental_guard>()))
2590  if (ThreadGuard(BB, cast<IntrinsicInst>(&I), BI))
2591  return true;
2592 
2593  return false;
2594 }
2595 
2596 /// Try to propagate the guard from BB which is the lower block of a diamond
2597 /// to one of its branches, in case if diamond's condition implies guard's
2598 /// condition.
2600  BranchInst *BI) {
2601  assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?");
2602  assert(BI->isConditional() && "Unconditional branch has 2 successors?");
2603  Value *GuardCond = Guard->getArgOperand(0);
2604  Value *BranchCond = BI->getCondition();
2605  BasicBlock *TrueDest = BI->getSuccessor(0);
2606  BasicBlock *FalseDest = BI->getSuccessor(1);
2607 
2608  auto &DL = BB->getModule()->getDataLayout();
2609  bool TrueDestIsSafe = false;
2610  bool FalseDestIsSafe = false;
2611 
2612  // True dest is safe if BranchCond => GuardCond.
2613  auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
2614  if (Impl && *Impl)
2615  TrueDestIsSafe = true;
2616  else {
2617  // False dest is safe if !BranchCond => GuardCond.
2618  Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);
2619  if (Impl && *Impl)
2620  FalseDestIsSafe = true;
2621  }
2622 
2623  if (!TrueDestIsSafe && !FalseDestIsSafe)
2624  return false;
2625 
2626  BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
2627  BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
2628 
2629  ValueToValueMapTy UnguardedMapping, GuardedMapping;
2630  Instruction *AfterGuard = Guard->getNextNode();
2631  unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
2632  if (Cost > BBDupThreshold)
2633  return false;
2634  // Duplicate all instructions before the guard and the guard itself to the
2635  // branch where implication is not proved.
2637  BB, PredGuardedBlock, AfterGuard, GuardedMapping);
2638  assert(GuardedBlock && "Could not create the guarded block?");
2639  // Duplicate all instructions before the guard in the unguarded branch.
2640  // Since we have successfully duplicated the guarded block and this block
2641  // has fewer instructions, we expect it to succeed.
2643  BB, PredUnguardedBlock, Guard, UnguardedMapping);
2644  assert(UnguardedBlock && "Could not create the unguarded block?");
2645  DEBUG(dbgs() << "Moved guard " << *Guard << " to block "
2646  << GuardedBlock->getName() << "\n");
2647  // DuplicateInstructionsInSplitBetween inserts a new block "BB.split" between
2648  // PredBB and BB. We need to perform two inserts and one delete for each of
2649  // the above calls to update Dominators.
2650  DDT->applyUpdates(
2651  {// Guarded block split.
2652  {DominatorTree::Delete, PredGuardedBlock, BB},
2653  {DominatorTree::Insert, PredGuardedBlock, GuardedBlock},
2654  {DominatorTree::Insert, GuardedBlock, BB},
2655  // Unguarded block split.
2656  {DominatorTree::Delete, PredUnguardedBlock, BB},
2657  {DominatorTree::Insert, PredUnguardedBlock, UnguardedBlock},
2658  {DominatorTree::Insert, UnguardedBlock, BB}});
2659  // Some instructions before the guard may still have uses. For them, we need
2660  // to create Phi nodes merging their copies in both guarded and unguarded
2661  // branches. Those instructions that have no uses can be just removed.
2663  for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
2664  if (!isa<PHINode>(&*BI))
2665  ToRemove.push_back(&*BI);
2666 
2667  Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
2668  assert(InsertionPoint && "Empty block?");
2669  // Substitute with Phis & remove.
2670  for (auto *Inst : reverse(ToRemove)) {
2671  if (!Inst->use_empty()) {
2672  PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
2673  NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
2674  NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
2675  NewPN->insertBefore(InsertionPoint);
2676  Inst->replaceAllUsesWith(NewPN);
2677  }
2678  Inst->eraseFromParent();
2679  }
2680  return true;
2681 }
Legacy wrapper pass to provide the GlobalsAAResult object.
bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl< BasicBlock *> &PredBBs, BasicBlock *SuccBB)
ThreadEdge - We have decided that it is safe and profitable to factor the blocks in PredBBs to one pr...
uint64_t CallInst * C
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks &#39;this&#39; from the containing basic block and deletes it.
Definition: Instruction.cpp:67
IntegerType * getType() const
getType - Specialize the getType() method to always return an IntegerType, which reduces the amount o...
Definition: Constants.h:172
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:111
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:548
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:72
This class is the base class for the comparison instructions.
Definition: InstrTypes.h:843
iterator_range< use_iterator > uses()
Definition: Value.h:360
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
void removePredecessor(BasicBlock *Pred, bool DontDeleteUselessPHIs=false)
Notify the BasicBlock that the predecessor Pred is no longer able to reach it.
Definition: BasicBlock.cpp:277
static IntegerType * getInt1Ty(LLVMContext &C)
Definition: Type.cpp:173
Helper class for SSA formation on a set of values defined in multiple blocks.
Definition: SSAUpdater.h:39
void addIncoming(Value *V, BasicBlock *BB)
Add an incoming value to the end of the PHI list.
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
BranchProbability getCompl() const
NodeTy * getNextNode()
Get the next node, or nullptr for the list tail.
Definition: ilist_node.h:289
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:687
Compute iterated dominance frontiers using a linear time algorithm.
Definition: AllocatorList.h:24
bool isAtomic() const
Return true if this instruction has an AtomicOrdering of unordered or higher.
Wrapper around LazyValueInfo.
This is the interface for a simple mod/ref and alias analysis over globals.
void Initialize(Type *Ty, StringRef Name)
Reset this object to get ready for a new set of SSA updates with type &#39;Ty&#39;.
Definition: SSAUpdater.cpp:54
BasicBlock * getSuccessor(unsigned idx) const
Return the specified successor.
LLVM_ATTRIBUTE_ALWAYS_INLINE size_type size() const
Definition: SmallVector.h:136
iterator end()
Definition: Function.h:636
void initializeJumpThreadingPass(PassRegistry &)
void AddAvailableValue(BasicBlock *BB, Value *V)
Indicate that a rewritten value is available in the specified block with the specified value...
Definition: SSAUpdater.cpp:67
This class represents a function call, abstracting a target machine&#39;s calling convention.
cl::opt< unsigned > DefMaxInstsToScan
The default number of maximum instructions to scan in the block, used by FindAvailableLoadedValue().
This file contains the declarations for metadata subclasses.
const Value * getTrueValue() const
uint64_t getFrequency() const
Returns the frequency as a fixpoint number scaled by the entry frequency.
AtomicOrdering getOrdering() const
Returns the ordering constraint of this load instruction.
Definition: Instructions.h:233
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:738
bool isTerminator() const
Definition: Instruction.h:129
static CastInst * CreateBitOrPointerCast(Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Create a BitCast, a PtrToInt, or an IntToPTr cast instruction.
void deleteValue()
Delete a pointer to a generic Value.
Definition: Value.cpp:99
static cl::opt< unsigned > ImplicationSearchThreshold("jump-threading-implication-search-threshold", cl::desc("The number of predecessors to search for a stronger " "condition to use to thread over a weaker condition"), cl::init(3), cl::Hidden)
bool all_of(R &&range, UnaryPredicate P)
Provide wrappers to std::all_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:814
Value * FindAvailablePtrLoadStore(Value *Ptr, Type *AccessTy, bool AtLeastAtomic, BasicBlock *ScanBB, BasicBlock::iterator &ScanFrom, unsigned MaxInstsToScan, AliasAnalysis *AA, bool *IsLoad, unsigned *NumScanedInst)
Scan backwards to see if we have the value of the given pointer available locally within a small numb...
Definition: Loads.cpp:336
static BasicBlock * FindMostPopularDest(BasicBlock *BB, const SmallVectorImpl< std::pair< BasicBlock *, BasicBlock *>> &PredToDestList)
FindMostPopularDest - The specified list contains multiple possible threadable destinations.
BasicBlock * getSuccessor(unsigned i) const
STATISTIC(NumFunctions, "Total number of functions")
bool SimplifyInstructionsInBlock(BasicBlock *BB, const TargetLibraryInfo *TLI=nullptr)
Scan the specified basic block and try to simplify any instructions in it and recursively delete dead...
Definition: Local.cpp:557
Metadata node.
Definition: Metadata.h:862
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:231
F(f)
An instruction for reading from memory.
Definition: Instructions.h:164
static Constant * getCompare(unsigned short pred, Constant *C1, Constant *C2, bool OnlyIfReduced=false)
Return an ICmp or FCmp comparison operator constant expression.
Definition: Constants.cpp:1868
Value * getCondition() const
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:227
This defines the Use class.
const Value * DoPHITranslation(const BasicBlock *CurBB, const BasicBlock *PredBB) const
Translate PHI node to its predecessor from the given basic block.
Definition: Value.cpp:730
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:33
FunctionPass * createJumpThreadingPass(int Threshold=-1)
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:252
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:191
The address of a basic block.
Definition: Constants.h:818
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
AnalysisUsage & addRequired()
const Module * getModule() const
Return the module owning the function this basic block belongs to, or nullptr it the function does no...
Definition: BasicBlock.cpp:116
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:51
void MergeBasicBlockIntoOnlyPred(BasicBlock *BB, DominatorTree *DT=nullptr, DeferredDominance *DDT=nullptr)
BB is a block with one predecessor and its predecessor is known to have one successor (BB!)...
Definition: Local.cpp:640
StringRef getName(ID id)
Return the LLVM name for an intrinsic, such as "llvm.ppc.altivec.lvx".
Definition: Function.cpp:591
This class represents the LLVM &#39;select&#39; instruction.
const DataLayout & getDataLayout() const
Get the data layout for the module&#39;s target platform.
Definition: Module.cpp:361
This is the base class for all instructions that perform data casts.
Definition: InstrTypes.h:560
static cl::opt< unsigned > BBDuplicateThreshold("jump-threading-threshold", cl::desc("Max block size to duplicate for jump threading"), cl::init(6), cl::Hidden)
&#39;undef&#39; values are things that do not have specified contents.
Definition: Constants.h:1252
Value * FindAvailableLoadedValue(LoadInst *Load, BasicBlock *ScanBB, BasicBlock::iterator &ScanFrom, unsigned MaxInstsToScan=DefMaxInstsToScan, AliasAnalysis *AA=nullptr, bool *IsLoadCSE=nullptr, unsigned *NumScanedInst=nullptr)
Scan backwards to see if we have the value of the given load available locally within a small number ...
Definition: Loads.cpp:321
unsigned getNumSuccessors() const
A Use represents the edge between a Value definition and its users.
Definition: Use.h:56
static bool runImpl(CallGraphSCC &SCC, AARGetterT AARGetter)
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:42
jump threading
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:197
void setName(const Twine &Name)
Change the name of the value.
Definition: Value.cpp:295
bool contains(const APInt &Val) const
Return true if the specified value is in the set.
bool TryToUnfoldSelectInCurrBB(BasicBlock *BB)
TryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the same BB in the form bb: p = ...
BinaryOp_match< LHS, RHS, Instruction::Add > m_Add(const LHS &L, const RHS &R)
Definition: PatternMatch.h:539
void assign(size_type NumElts, const T &Elt)
Definition: SmallVector.h:425
auto reverse(ContainerTy &&C, typename std::enable_if< has_rbegin< ContainerTy >::value >::type *=nullptr) -> decltype(make_range(C.rbegin(), C.rend()))
Definition: STLExtras.h:233
bool isOne() const
This is just a convenience method to make client code smaller for a common case.
Definition: Constants.h:201
static Constant * get(unsigned Opcode, Constant *C1, Constant *C2, unsigned Flags=0, Type *OnlyIfReducedTy=nullptr)
get - Return a binary or shift operator constant expression, folding if possible. ...
Definition: Constants.cpp:1747
Value * SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, const SimplifyQuery &Q)
Given operands for a CmpInst, fold the result or return null.
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:245
static void normalizeProbabilities(ProbabilityIter Begin, ProbabilityIter End)
MDNode * getMetadata(unsigned KindID) const
Get the metadata of given kind attached to this Instruction.
Definition: Instruction.h:195
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:33
class_match< ConstantInt > m_ConstantInt()
Match an arbitrary ConstantInt and ignore it.
Definition: PatternMatch.h:83
const APInt & getValue() const
Return the constant as an APInt value reference.
Definition: Constants.h:138
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:126
auto lower_bound(R &&Range, ForwardIt I) -> decltype(adl_begin(Range))
Provide wrappers to std::lower_bound which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:910
bool removeUnreachableBlocks(Function &F, LazyValueInfo *LVI=nullptr, DeferredDominance *DDT=nullptr)
Remove all blocks that can not be reached from the function&#39;s entry.
Definition: Local.cpp:1922
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:439
void SplitLandingPadPredecessors(BasicBlock *OrigBB, ArrayRef< BasicBlock *> Preds, const char *Suffix, const char *Suffix2, SmallVectorImpl< BasicBlock *> &NewBBs, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr, bool PreserveLCSSA=false)
This method transforms the landing pad, OrigBB, by introducing two new basic blocks into the function...
static bool ProcessBlock(BasicBlock &BB, DominatorTree &DT, LoopInfo &LI, AAResults &AA)
Definition: Sink.cpp:201
bool runImpl(Function &F, TargetLibraryInfo *TLI_, LazyValueInfo *LVI_, AliasAnalysis *AA_, DeferredDominance *DDT_, bool HasProfileData_, std::unique_ptr< BlockFrequencyInfo > BFI_, std::unique_ptr< BranchProbabilityInfo > BPI_)
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:301
iterator begin()
Definition: Function.h:634
static Constant * getKnownConstant(Value *Val, ConstantPreference Preference)
getKnownConstant - Helper method to determine if we can thread over a terminator with the given value...
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree...
Definition: Dominators.h:142
Value * getOperand(unsigned i) const
Definition: User.h:154
void removeDeadConstantUsers() const
If there are any dead constant users dangling off of this constant, remove them.
Definition: Constants.cpp:500
StringRef getString() const
Definition: Metadata.cpp:464
const BasicBlock & getEntryBlock() const
Definition: Function.h:618
void getAAMetadata(AAMDNodes &N, bool Merge=false) const
Fills the AAMDNodes structure with AA metadata from this instruction.
BlockFrequencyInfo pass uses BlockFrequencyInfoImpl implementation to estimate IR basic block frequen...
static bool runOnFunction(Function &F, bool PostInlining)
#define P(N)
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:406
bool isGuaranteedToTransferExecutionToSuccessor(const Instruction *I)
Return true if this function can prove that the instruction I will always transfer execution to one o...
void array_pod_sort(IteratorTy Start, IteratorTy End)
array_pod_sort - This sorts an array with the specified start and end extent.
Definition: STLExtras.h:760
const Instruction * getFirstNonPHI() const
Returns a pointer to the first instruction in this block that is not a PHINode instruction.
Definition: BasicBlock.cpp:171
Subclasses of this class are all able to terminate a basic block.
Definition: InstrTypes.h:54
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:153
const_iterator getFirstInsertionPt() const
Returns an iterator to the first instruction in this block that is suitable for inserting a non-PHI i...
Definition: BasicBlock.cpp:200
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:282
void setSuccessor(unsigned idx, BasicBlock *B)
Update the specified successor to point at the provided block.
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:217
void insertBefore(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately before the specified instruction...
Definition: Instruction.cpp:73
LLVM Basic Block Representation.
Definition: BasicBlock.h:59
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:46
Conditional or Unconditional Branch instruction.
static BlockAddress * get(Function *F, BasicBlock *BB)
Return a BlockAddress for the specified function and basic block.
Definition: Constants.cpp:1375
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:42
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator begin()
Definition: SmallVector.h:116
static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB)
Return true if Op is an instruction defined in the given block.
Value * getIncomingValueForBlock(const BasicBlock *BB) const
This file contains the declarations for the subclasses of Constant, which represent the different fla...
const Instruction & front() const
Definition: BasicBlock.h:264
Indirect Branch Instruction.
A manager for alias analyses.
std::pair< iterator, bool > insert(PtrType Ptr)
Inserts Ptr if and only if there is no element in the container equal to Ptr.
Definition: SmallPtrSet.h:371
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:536
Constant * ConstantFoldInstruction(Instruction *I, const DataLayout &DL, const TargetLibraryInfo *TLI=nullptr)
ConstantFoldInstruction - Try to constant fold the specified instruction.
Interval::pred_iterator pred_begin(Interval *I)
pred_begin/pred_end - define methods so that Intervals may be used just like BasicBlocks can with the...
Definition: Interval.h:113
bool isUnordered() const
Definition: Instructions.h:264
Represent the analysis usage information of a pass.
bool any_of(R &&Range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:821
This instruction compares its operands according to the predicate given to the constructor.
Analysis pass providing a never-invalidated alias analysis result.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:853
jump Jump Threading
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:285
ConstantRange add(const ConstantRange &Other) const
Return a new range representing the possible values resulting from an addition of a value in this ran...
Interval::pred_iterator pred_end(Interval *I)
Definition: Interval.h:116
BasicBlock * SplitBlockPredecessors(BasicBlock *BB, ArrayRef< BasicBlock *> Preds, const char *Suffix, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr, bool PreserveLCSSA=false)
This method introduces at least one new basic block into the function and moves some of the predecess...
static BasicBlock * Create(LLVMContext &Context, const Twine &Name="", Function *Parent=nullptr, BasicBlock *InsertBefore=nullptr)
Creates a new BasicBlock.
Definition: BasicBlock.h:101
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
Definition: SmallPtrSet.h:382
bool pred_empty(const BasicBlock *BB)
Definition: CFG.h:107
static Constant * getNot(Constant *C)
Definition: Constants.cpp:2145
void FindLoopHeaders(Function &F)
FindLoopHeaders - We do not want jump threading to turn proper loop structures into irreducible loops...
bool ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions=false, const TargetLibraryInfo *TLI=nullptr, DeferredDominance *DDT=nullptr)
If a terminator instruction is predicated on a constant value, convert it into an unconditional branc...
Definition: Local.cpp:103
const Value * getCondition() const
static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB)
static UndefValue * get(Type *T)
Static factory methods - Return an &#39;undef&#39; object of the specified type.
Definition: Constants.cpp:1356
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs, and aliases.
Definition: Value.cpp:567
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
Value * getIncomingValue(unsigned i) const
Return incoming value number x.
INITIALIZE_PASS_END(RegBankSelect, DEBUG_TYPE, "Assign register bank of generic virtual registers", false, false) RegBankSelect
void setMetadata(unsigned KindID, MDNode *Node)
Set the metadata of the specified kind to the specified node.
Definition: Metadata.cpp:1222
static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB)
GetBestDestForBranchOnUndef - If we determine that the specified block ends in an undefined jump...
Tristate
This is used to return true/false/dunno results.
Definition: LazyValueInfo.h:63
bool TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB, DeferredDominance *DDT=nullptr)
BB is known to contain an unconditional branch, and contains no instructions other than PHI nodes...
Definition: Local.cpp:926
bool ProcessBranchOnPHI(PHINode *PN)
ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on a PHI node in the curren...
bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, jumpthreading::PredValueInfo &Result, jumpthreading::ConstantPreference Preference, Instruction *CxtI=nullptr)
ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see if we can infer that the ...
size_type size() const
Definition: SmallPtrSet.h:93
bool isLandingPad() const
Return true if this basic block is a landing pad.
Definition: BasicBlock.cpp:443
bool hasAddressTaken() const
Returns true if there are any uses of this basic block other than direct branches, switches, etc.
Definition: BasicBlock.h:376
const InstListType & getInstList() const
Return the underlying instruction list container.
Definition: BasicBlock.h:317
static cl::opt< bool > PrintLVIAfterJumpThreading("print-lvi-after-jump-threading", cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false), cl::Hidden)
Iterator for intrusive lists based on ilist_node.
unsigned getNumOperands() const
Definition: User.h:176
void moveAfter(BasicBlock *MovePos)
Unlink this basic block from its current function and insert it right after MovePos in the function M...
Definition: BasicBlock.cpp:110
static ConstantRange makeExactICmpRegion(CmpInst::Predicate Pred, const APInt &Other)
Produce the exact range such that all values in the returned range satisfy the given predicate with a...
See the file comment.
Definition: ValueMap.h:86
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:418
This is the shared class of boolean and integer constants.
Definition: Constants.h:84
iterator end()
Definition: BasicBlock.h:254
bool isExceptional() const
Definition: InstrTypes.h:84
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:862
Module.h This file contains the declarations for the Module class.
Provides information about what library functions are available for the current target.
This pass performs &#39;jump threading&#39;, which looks at blocks that have multiple predecessors and multip...
Definition: JumpThreading.h:77
This class represents a range of values.
Definition: ConstantRange.h:47
static BranchProbability getBranchProbability(uint64_t Numerator, uint64_t Denominator)
TerminatorInst * SplitBlockAndInsertIfThen(Value *Cond, Instruction *SplitBefore, bool Unreachable, MDNode *BranchWeights=nullptr, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr)
Split the containing block at the specified instruction - everything before SplitBefore stays in the ...
A collection of metadata nodes that might be associated with a memory access used by the alias-analys...
Definition: Metadata.h:642
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:383
static Constant * get(Type *Ty, uint64_t V, bool isSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:585
bool SimplifyPartiallyRedundantLoad(LoadInst *LI)
SimplifyPartiallyRedundantLoad - If LoadI is an obviously partially redundant load instruction...
static BranchInst * Create(BasicBlock *IfTrue, Instruction *InsertBefore=nullptr)
bool isConditional() const
static PHINode * Create(Type *Ty, unsigned NumReservedValues, const Twine &NameStr="", Instruction *InsertBefore=nullptr)
Constructors - NumReservedValues is a hint for the number of incoming edges that this phi node will h...
pred_range predecessors(BasicBlock *BB)
Definition: CFG.h:110
unsigned getNumIncomingValues() const
Return the number of incoming edges.
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:541
void setOperand(unsigned i, Value *Val)
Definition: User.h:159
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:132
static unsigned getJumpThreadDuplicationCost(BasicBlock *BB, Instruction *StopAt, unsigned Threshold)
Return the cost of duplicating a piece of this block from first non-phi and before StopAt instruction...
Function * getFunction(StringRef Name) const
Look up the specified function in the module symbol table.
Definition: Module.cpp:172
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:55
void push_back(pointer val)
Definition: ilist.h:326
Class to defer updates to a DominatorTree.
Definition: Dominators.h:313
BasicBlock * DuplicateInstructionsInSplitBetween(BasicBlock *BB, BasicBlock *PredBB, Instruction *StopAt, ValueToValueMapTy &ValueMapping)
Split edge between BB and PredBB and duplicate all non-Phi instructions from BB between its beginning...
static Constant * getCast(unsigned ops, Constant *C, Type *Ty, bool OnlyIfReduced=false)
Convenience function for getting a Cast operation.
Definition: Constants.cpp:1471
const Value * getFalseValue() const
void removeFromParent()
This method unlinks &#39;this&#39; from the containing basic block, but does not delete it.
Definition: Instruction.cpp:63
Predicate getPredicate() const
Return the predicate for this instruction.
Definition: InstrTypes.h:927
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:546
LLVM_ATTRIBUTE_ALWAYS_INLINE iterator end()
Definition: SmallVector.h:120
Analysis providing branch probability information.
iterator insert(iterator where, pointer New)
Definition: ilist.h:241
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:285
unsigned getAlignment() const
Return the alignment of the access that is being performed.
Definition: Instructions.h:226
void emplace_back(ArgTypes &&... Args)
Definition: SmallVector.h:654
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:61
iterator begin()
Definition: DenseMap.h:70
bool ProcessBranchOnXOR(BinaryOperator *BO)
ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on a xor instruction in the...
Value * getArgOperand(unsigned i) const
getArgOperand/setArgOperand - Return/set the i-th call argument.
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:224
BasicBlock * getIncomingBlock(unsigned i) const
Return incoming basic block number i.
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:108
SyncScope::ID getSyncScopeID() const
Returns the synchronization scope ID of this load instruction.
Definition: Instructions.h:245
bool ProcessImpliedCondition(BasicBlock *BB)
#define I(x, y, z)
Definition: MD5.cpp:58
bool isZero() const
This is just a convenience method to make client code smaller for a common code.
Definition: Constants.h:193
LLVM_NODISCARD std::enable_if<!is_simple_type< Y >::value, typename cast_retty< X, const Y >::ret_type >::type dyn_cast(const Y &Val)
Definition: Casting.h:323
static void ReplaceFoldableUses(Instruction *Cond, Value *ToVal)
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:174
iterator_range< const_phi_iterator > phis() const
Returns a range that iterates over the phis in the basic block.
Definition: BasicBlock.h:308
void combineMetadataForCSE(Instruction *K, const Instruction *J)
Combine the metadata of two instructions so that K can replace J.
Definition: Local.cpp:2036
bool isUnconditional() const
static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, BasicBlock *OldPred, BasicBlock *NewPred, DenseMap< Instruction *, Value *> &ValueMap)
AddPHINodeEntriesForMappedBlock - We&#39;re adding &#39;NewPred&#39; as a new predecessor to the PHIBB block...
ValueT lookup(const_arg_type_t< KeyT > Val) const
lookup - Return the entry for the specified key, or a default constructed value if no such entry exis...
Definition: DenseMap.h:181
bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB, jumpthreading::ConstantPreference Preference, Instruction *CxtI=nullptr)
Optional< bool > isImpliedCondition(const Value *LHS, const Value *RHS, const DataLayout &DL, bool LHSIsTrue=true, unsigned Depth=0)
Return true if RHS is known to be implied true by LHS.
void FindFunctionBackedges(const Function &F, SmallVectorImpl< std::pair< const BasicBlock *, const BasicBlock *> > &Result)
Analyze the specified function to find all of the loop backedges in the function and return them...
Definition: CFG.cpp:27
Analysis pass providing the TargetLibraryInfo.
static int const Threshold
TODO: Write a new FunctionPass AliasAnalysis so that it can keep a cache.
Multiway switch.
Helper struct that represents how a value is mapped through different register banks.
This pass computes, caches, and vends lazy value constraint information.
Definition: LazyValueInfo.h:32
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
unsigned getNumSuccessors() const
Return the number of successors that this terminator has.
See the comments on JumpThreadingPass.
void DeleteDeadBlock(BasicBlock *BB, DeferredDominance *DDT=nullptr)
Delete the specified block, which must have no predecessors.
bool isSafeToSpeculativelyExecute(const Value *V, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr)
Return true if the instruction does not have any effects besides calculating the result and does not ...
bool isEHPad() const
Return true if this basic block is an exception handling block.
Definition: BasicBlock.h:383
unsigned getPrimitiveSizeInBits() const LLVM_READONLY
Return the basic size of this type if it is a primitive type.
Definition: Type.cpp:115
bool ThreadGuard(BasicBlock *BB, IntrinsicInst *Guard, BranchInst *BI)
Try to propagate the guard from BB which is the lower block of a diamond to one of its branches...
bool ProcessGuards(BasicBlock *BB)
Try to propagate a guard from the current BB into one of its predecessors in case if another branch o...
Module * getParent()
Get the module that this global value is contained inside of...
Definition: GlobalValue.h:565
bool isInstructionTriviallyDead(Instruction *I, const TargetLibraryInfo *TLI=nullptr)
Return true if the result produced by the instruction is not used, and the instruction has no side ef...
Definition: Local.cpp:346
LLVM Value Representation.
Definition: Value.h:73
succ_range successors(BasicBlock *BB)
Definition: CFG.h:143
constexpr char Size[]
Key for Kernel::Arg::Metadata::mSize.
bool ProcessBlock(BasicBlock *BB)
ProcessBlock - If there are any predecessors whose control can be threaded through to a successor...
static const Function * getParent(const Value *V)
#define DEBUG(X)
Definition: Debug.h:118
BasicBlock * SplitEdge(BasicBlock *From, BasicBlock *To, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr)
Split the edge connecting specified block.
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:418
A single uniqued string.
Definition: Metadata.h:602
A container for analyses that lazily runs them and caches their results.
const Instruction * getFirstNonPHIOrDbg() const
Returns a pointer to the first instruction in this block that is not a PHINode or a debug intrinsic...
Definition: BasicBlock.cpp:178
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:260
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
const TerminatorInst * getTerminator() const LLVM_READONLY
Returns the terminator instruction if the block is well formed or null if the block is not well forme...
Definition: BasicBlock.cpp:120
This header defines various interfaces for pass management in LLVM.
INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading", "Jump Threading", false, false) INITIALIZE_PASS_END(JumpThreading
bool extractProfMetadata(uint64_t &TrueVal, uint64_t &FalseVal) const
Retrieve the raw weight values of a conditional branch or select.
Definition: Metadata.cpp:1311
op_range incoming_values()
ConstantRange inverse() const
Return a new range that is the logical not of the current set.
void RewriteUse(Use &U)
Rewrite a use of the symbolic value.
Definition: SSAUpdater.cpp:185
bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB, const SmallVectorImpl< BasicBlock *> &PredBBs)
DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch to BB which contains an i1...
static bool hasAddressTakenAndUsed(BasicBlock *BB)
Value * SimplifyInstruction(Instruction *I, const SimplifyQuery &Q, OptimizationRemarkEmitter *ORE=nullptr)
See if we can compute a simplified version of this instruction.
bool hasProfileData() const
Return true if the function is annotated with profile data.
Definition: Function.h:289
uint32_t getNumerator() const
bool use_empty() const
Definition: Value.h:328
unsigned replaceNonLocalUsesWith(Instruction *From, Value *To)
Definition: Local.cpp:2067
Analysis to compute lazy value information.
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:44
const BasicBlock * getParent() const
Definition: Instruction.h:67
bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB)
TryToUnfoldSelect - Look for blocks of the form bb1: a = select br bb2.
void resize(size_type N)
Definition: SmallVector.h:353
bool is_contained(R &&Range, const E &Element)
Wrapper function around std::find to detect if an element exists in a container.
Definition: STLExtras.h:873