LLVM  9.0.0svn
JumpThreading.cpp
Go to the documentation of this file.
1 //===- JumpThreading.cpp - Thread control through conditional blocks ------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the Jump Threading pass.
10 //
11 //===----------------------------------------------------------------------===//
12 
14 #include "llvm/ADT/DenseMap.h"
15 #include "llvm/ADT/DenseSet.h"
16 #include "llvm/ADT/Optional.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
24 #include "llvm/Analysis/CFG.h"
31 #include "llvm/Analysis/Loads.h"
32 #include "llvm/Analysis/LoopInfo.h"
35 #include "llvm/IR/BasicBlock.h"
36 #include "llvm/IR/CFG.h"
37 #include "llvm/IR/Constant.h"
38 #include "llvm/IR/ConstantRange.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DataLayout.h"
41 #include "llvm/IR/Dominators.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/IR/InstrTypes.h"
44 #include "llvm/IR/Instruction.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/IR/Intrinsics.h"
48 #include "llvm/IR/LLVMContext.h"
49 #include "llvm/IR/MDBuilder.h"
50 #include "llvm/IR/Metadata.h"
51 #include "llvm/IR/Module.h"
52 #include "llvm/IR/PassManager.h"
53 #include "llvm/IR/PatternMatch.h"
54 #include "llvm/IR/Type.h"
55 #include "llvm/IR/Use.h"
56 #include "llvm/IR/User.h"
57 #include "llvm/IR/Value.h"
58 #include "llvm/Pass.h"
61 #include "llvm/Support/Casting.h"
63 #include "llvm/Support/Debug.h"
65 #include "llvm/Transforms/Scalar.h"
71 #include <algorithm>
72 #include <cassert>
73 #include <cstddef>
74 #include <cstdint>
75 #include <iterator>
76 #include <memory>
77 #include <utility>
78 
79 using namespace llvm;
80 using namespace jumpthreading;
81 
82 #define DEBUG_TYPE "jump-threading"
83 
84 STATISTIC(NumThreads, "Number of jumps threaded");
85 STATISTIC(NumFolds, "Number of terminators folded");
86 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
87 
88 static cl::opt<unsigned>
89 BBDuplicateThreshold("jump-threading-threshold",
90  cl::desc("Max block size to duplicate for jump threading"),
91  cl::init(6), cl::Hidden);
92 
93 static cl::opt<unsigned>
95  "jump-threading-implication-search-threshold",
96  cl::desc("The number of predecessors to search for a stronger "
97  "condition to use to thread over a weaker condition"),
98  cl::init(3), cl::Hidden);
99 
101  "print-lvi-after-jump-threading",
102  cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false),
103  cl::Hidden);
104 
105 namespace {
106 
107  /// This pass performs 'jump threading', which looks at blocks that have
108  /// multiple predecessors and multiple successors. If one or more of the
109  /// predecessors of the block can be proven to always jump to one of the
110  /// successors, we forward the edge from the predecessor to the successor by
111  /// duplicating the contents of this block.
112  ///
113  /// An example of when this can occur is code like this:
114  ///
115  /// if () { ...
116  /// X = 4;
117  /// }
118  /// if (X < 3) {
119  ///
120  /// In this case, the unconditional branch at the end of the first if can be
121  /// revectored to the false side of the second if.
122  class JumpThreading : public FunctionPass {
123  JumpThreadingPass Impl;
124 
125  public:
126  static char ID; // Pass identification
127 
128  JumpThreading(int T = -1) : FunctionPass(ID), Impl(T) {
130  }
131 
132  bool runOnFunction(Function &F) override;
133 
134  void getAnalysisUsage(AnalysisUsage &AU) const override {
142  }
143 
144  void releaseMemory() override { Impl.releaseMemory(); }
145  };
146 
147 } // end anonymous namespace
148 
149 char JumpThreading::ID = 0;
150 
151 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
152  "Jump Threading", false, false)
157 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
158  "Jump Threading", false, false)
159 
160 // Public interface to the Jump Threading pass
162  return new JumpThreading(Threshold);
163 }
164 
166  BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
167 }
168 
169 // Update branch probability information according to conditional
170 // branch probability. This is usually made possible for cloned branches
171 // in inline instances by the context specific profile in the caller.
172 // For instance,
173 //
174 // [Block PredBB]
175 // [Branch PredBr]
176 // if (t) {
177 // Block A;
178 // } else {
179 // Block B;
180 // }
181 //
182 // [Block BB]
183 // cond = PN([true, %A], [..., %B]); // PHI node
184 // [Branch CondBr]
185 // if (cond) {
186 // ... // P(cond == true) = 1%
187 // }
188 //
189 // Here we know that when block A is taken, cond must be true, which means
190 // P(cond == true | A) = 1
191 //
192 // Given that P(cond == true) = P(cond == true | A) * P(A) +
193 // P(cond == true | B) * P(B)
194 // we get:
195 // P(cond == true ) = P(A) + P(cond == true | B) * P(B)
196 //
197 // which gives us:
198 // P(A) is less than P(cond == true), i.e.
199 // P(t == true) <= P(cond == true)
200 //
201 // In other words, if we know P(cond == true) is unlikely, we know
202 // that P(t == true) is also unlikely.
203 //
205  BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
206  if (!CondBr)
207  return;
208 
210  uint64_t TrueWeight, FalseWeight;
211  if (!CondBr->extractProfMetadata(TrueWeight, FalseWeight))
212  return;
213 
214  // Returns the outgoing edge of the dominating predecessor block
215  // that leads to the PhiNode's incoming block:
216  auto GetPredOutEdge =
217  [](BasicBlock *IncomingBB,
218  BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> {
219  auto *PredBB = IncomingBB;
220  auto *SuccBB = PhiBB;
221  while (true) {
222  BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
223  if (PredBr && PredBr->isConditional())
224  return {PredBB, SuccBB};
225  auto *SinglePredBB = PredBB->getSinglePredecessor();
226  if (!SinglePredBB)
227  return {nullptr, nullptr};
228  SuccBB = PredBB;
229  PredBB = SinglePredBB;
230  }
231  };
232 
233  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
234  Value *PhiOpnd = PN->getIncomingValue(i);
235  ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd);
236 
237  if (!CI || !CI->getType()->isIntegerTy(1))
238  continue;
239 
241  TrueWeight, TrueWeight + FalseWeight)
243  FalseWeight, TrueWeight + FalseWeight));
244 
245  auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB);
246  if (!PredOutEdge.first)
247  return;
248 
249  BasicBlock *PredBB = PredOutEdge.first;
250  BranchInst *PredBr = cast<BranchInst>(PredBB->getTerminator());
251 
252  uint64_t PredTrueWeight, PredFalseWeight;
253  // FIXME: We currently only set the profile data when it is missing.
254  // With PGO, this can be used to refine even existing profile data with
255  // context information. This needs to be done after more performance
256  // testing.
257  if (PredBr->extractProfMetadata(PredTrueWeight, PredFalseWeight))
258  continue;
259 
260  // We can not infer anything useful when BP >= 50%, because BP is the
261  // upper bound probability value.
262  if (BP >= BranchProbability(50, 100))
263  continue;
264 
265  SmallVector<uint32_t, 2> Weights;
266  if (PredBr->getSuccessor(0) == PredOutEdge.second) {
267  Weights.push_back(BP.getNumerator());
268  Weights.push_back(BP.getCompl().getNumerator());
269  } else {
270  Weights.push_back(BP.getCompl().getNumerator());
271  Weights.push_back(BP.getNumerator());
272  }
274  MDBuilder(PredBr->getParent()->getContext())
275  .createBranchWeights(Weights));
276  }
277 }
278 
279 /// runOnFunction - Toplevel algorithm.
281  if (skipFunction(F))
282  return false;
283  auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
284  // Get DT analysis before LVI. When LVI is initialized it conditionally adds
285  // DT if it's available.
286  auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
287  auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
288  auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
290  std::unique_ptr<BlockFrequencyInfo> BFI;
291  std::unique_ptr<BranchProbabilityInfo> BPI;
292  bool HasProfileData = F.hasProfileData();
293  if (HasProfileData) {
294  LoopInfo LI{DominatorTree(F)};
295  BPI.reset(new BranchProbabilityInfo(F, LI, TLI));
296  BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
297  }
298 
299  bool Changed = Impl.runImpl(F, TLI, LVI, AA, &DTU, HasProfileData,
300  std::move(BFI), std::move(BPI));
302  dbgs() << "LVI for function '" << F.getName() << "':\n";
303  LVI->printLVI(F, *DT, dbgs());
304  }
305  return Changed;
306 }
307 
310  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
311  // Get DT analysis before LVI. When LVI is initialized it conditionally adds
312  // DT if it's available.
313  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
314  auto &LVI = AM.getResult<LazyValueAnalysis>(F);
315  auto &AA = AM.getResult<AAManager>(F);
317 
318  std::unique_ptr<BlockFrequencyInfo> BFI;
319  std::unique_ptr<BranchProbabilityInfo> BPI;
320  if (F.hasProfileData()) {
321  LoopInfo LI{DominatorTree(F)};
322  BPI.reset(new BranchProbabilityInfo(F, LI, &TLI));
323  BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
324  }
325 
326  bool Changed = runImpl(F, &TLI, &LVI, &AA, &DTU, HasProfileData,
327  std::move(BFI), std::move(BPI));
328 
329  if (!Changed)
330  return PreservedAnalyses::all();
332  PA.preserve<GlobalsAA>();
335  return PA;
336 }
337 
339  LazyValueInfo *LVI_, AliasAnalysis *AA_,
340  DomTreeUpdater *DTU_, bool HasProfileData_,
341  std::unique_ptr<BlockFrequencyInfo> BFI_,
342  std::unique_ptr<BranchProbabilityInfo> BPI_) {
343  LLVM_DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
344  TLI = TLI_;
345  LVI = LVI_;
346  AA = AA_;
347  DTU = DTU_;
348  BFI.reset();
349  BPI.reset();
350  // When profile data is available, we need to update edge weights after
351  // successful jump threading, which requires both BPI and BFI being available.
352  HasProfileData = HasProfileData_;
353  auto *GuardDecl = F.getParent()->getFunction(
354  Intrinsic::getName(Intrinsic::experimental_guard));
355  HasGuards = GuardDecl && !GuardDecl->use_empty();
356  if (HasProfileData) {
357  BPI = std::move(BPI_);
358  BFI = std::move(BFI_);
359  }
360 
361  // JumpThreading must not processes blocks unreachable from entry. It's a
362  // waste of compute time and can potentially lead to hangs.
363  SmallPtrSet<BasicBlock *, 16> Unreachable;
364  assert(DTU && "DTU isn't passed into JumpThreading before using it.");
365  assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed.");
366  DominatorTree &DT = DTU->getDomTree();
367  for (auto &BB : F)
368  if (!DT.isReachableFromEntry(&BB))
369  Unreachable.insert(&BB);
370 
371  FindLoopHeaders(F);
372 
373  bool EverChanged = false;
374  bool Changed;
375  do {
376  Changed = false;
377  for (auto &BB : F) {
378  if (Unreachable.count(&BB))
379  continue;
380  while (ProcessBlock(&BB)) // Thread all of the branches we can over BB.
381  Changed = true;
382  // Stop processing BB if it's the entry or is now deleted. The following
383  // routines attempt to eliminate BB and locating a suitable replacement
384  // for the entry is non-trivial.
385  if (&BB == &F.getEntryBlock() || DTU->isBBPendingDeletion(&BB))
386  continue;
387 
388  if (pred_empty(&BB)) {
389  // When ProcessBlock makes BB unreachable it doesn't bother to fix up
390  // the instructions in it. We must remove BB to prevent invalid IR.
391  LLVM_DEBUG(dbgs() << " JT: Deleting dead block '" << BB.getName()
392  << "' with terminator: " << *BB.getTerminator()
393  << '\n');
394  LoopHeaders.erase(&BB);
395  LVI->eraseBlock(&BB);
396  DeleteDeadBlock(&BB, DTU);
397  Changed = true;
398  continue;
399  }
400 
401  // ProcessBlock doesn't thread BBs with unconditional TIs. However, if BB
402  // is "almost empty", we attempt to merge BB with its sole successor.
403  auto *BI = dyn_cast<BranchInst>(BB.getTerminator());
404  if (BI && BI->isUnconditional() &&
405  // The terminator must be the only non-phi instruction in BB.
406  BB.getFirstNonPHIOrDbg()->isTerminator() &&
407  // Don't alter Loop headers and latches to ensure another pass can
408  // detect and transform nested loops later.
409  !LoopHeaders.count(&BB) && !LoopHeaders.count(BI->getSuccessor(0)) &&
411  // BB is valid for cleanup here because we passed in DTU. F remains
412  // BB's parent until a DTU->getDomTree() event.
413  LVI->eraseBlock(&BB);
414  Changed = true;
415  }
416  }
417  EverChanged |= Changed;
418  } while (Changed);
419 
420  LoopHeaders.clear();
421  // Flush only the Dominator Tree.
422  DTU->getDomTree();
423  LVI->enableDT();
424  return EverChanged;
425 }
426 
427 // Replace uses of Cond with ToVal when safe to do so. If all uses are
428 // replaced, we can remove Cond. We cannot blindly replace all uses of Cond
429 // because we may incorrectly replace uses when guards/assumes are uses of
430 // of `Cond` and we used the guards/assume to reason about the `Cond` value
431 // at the end of block. RAUW unconditionally replaces all uses
432 // including the guards/assumes themselves and the uses before the
433 // guard/assume.
434 static void ReplaceFoldableUses(Instruction *Cond, Value *ToVal) {
435  assert(Cond->getType() == ToVal->getType());
436  auto *BB = Cond->getParent();
437  // We can unconditionally replace all uses in non-local blocks (i.e. uses
438  // strictly dominated by BB), since LVI information is true from the
439  // terminator of BB.
440  replaceNonLocalUsesWith(Cond, ToVal);
441  for (Instruction &I : reverse(*BB)) {
442  // Reached the Cond whose uses we are trying to replace, so there are no
443  // more uses.
444  if (&I == Cond)
445  break;
446  // We only replace uses in instructions that are guaranteed to reach the end
447  // of BB, where we know Cond is ToVal.
449  break;
450  I.replaceUsesOfWith(Cond, ToVal);
451  }
452  if (Cond->use_empty() && !Cond->mayHaveSideEffects())
453  Cond->eraseFromParent();
454 }
455 
456 /// Return the cost of duplicating a piece of this block from first non-phi
457 /// and before StopAt instruction to thread across it. Stop scanning the block
458 /// when exceeding the threshold. If duplication is impossible, returns ~0U.
460  Instruction *StopAt,
461  unsigned Threshold) {
462  assert(StopAt->getParent() == BB && "Not an instruction from proper BB?");
463  /// Ignore PHI nodes, these will be flattened when duplication happens.
465 
466  // FIXME: THREADING will delete values that are just used to compute the
467  // branch, so they shouldn't count against the duplication cost.
468 
469  unsigned Bonus = 0;
470  if (BB->getTerminator() == StopAt) {
471  // Threading through a switch statement is particularly profitable. If this
472  // block ends in a switch, decrease its cost to make it more likely to
473  // happen.
474  if (isa<SwitchInst>(StopAt))
475  Bonus = 6;
476 
477  // The same holds for indirect branches, but slightly more so.
478  if (isa<IndirectBrInst>(StopAt))
479  Bonus = 8;
480  }
481 
482  // Bump the threshold up so the early exit from the loop doesn't skip the
483  // terminator-based Size adjustment at the end.
484  Threshold += Bonus;
485 
486  // Sum up the cost of each instruction until we get to the terminator. Don't
487  // include the terminator because the copy won't include it.
488  unsigned Size = 0;
489  for (; &*I != StopAt; ++I) {
490 
491  // Stop scanning the block if we've reached the threshold.
492  if (Size > Threshold)
493  return Size;
494 
495  // Debugger intrinsics don't incur code size.
496  if (isa<DbgInfoIntrinsic>(I)) continue;
497 
498  // If this is a pointer->pointer bitcast, it is free.
499  if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
500  continue;
501 
502  // Bail out if this instruction gives back a token type, it is not possible
503  // to duplicate it if it is used outside this BB.
504  if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
505  return ~0U;
506 
507  // All other instructions count for at least one unit.
508  ++Size;
509 
510  // Calls are more expensive. If they are non-intrinsic calls, we model them
511  // as having cost of 4. If they are a non-vector intrinsic, we model them
512  // as having cost of 2 total, and if they are a vector intrinsic, we model
513  // them as having cost 1.
514  if (const CallInst *CI = dyn_cast<CallInst>(I)) {
515  if (CI->cannotDuplicate() || CI->isConvergent())
516  // Blocks with NoDuplicate are modelled as having infinite cost, so they
517  // are never duplicated.
518  return ~0U;
519  else if (!isa<IntrinsicInst>(CI))
520  Size += 3;
521  else if (!CI->getType()->isVectorTy())
522  Size += 1;
523  }
524  }
525 
526  return Size > Bonus ? Size - Bonus : 0;
527 }
528 
529 /// FindLoopHeaders - We do not want jump threading to turn proper loop
530 /// structures into irreducible loops. Doing this breaks up the loop nesting
531 /// hierarchy and pessimizes later transformations. To prevent this from
532 /// happening, we first have to find the loop headers. Here we approximate this
533 /// by finding targets of backedges in the CFG.
534 ///
535 /// Note that there definitely are cases when we want to allow threading of
536 /// edges across a loop header. For example, threading a jump from outside the
537 /// loop (the preheader) to an exit block of the loop is definitely profitable.
538 /// It is also almost always profitable to thread backedges from within the loop
539 /// to exit blocks, and is often profitable to thread backedges to other blocks
540 /// within the loop (forming a nested loop). This simple analysis is not rich
541 /// enough to track all of these properties and keep it up-to-date as the CFG
542 /// mutates, so we don't allow any of these transformations.
545  FindFunctionBackedges(F, Edges);
546 
547  for (const auto &Edge : Edges)
548  LoopHeaders.insert(Edge.second);
549 }
550 
551 /// getKnownConstant - Helper method to determine if we can thread over a
552 /// terminator with the given value as its condition, and if so what value to
553 /// use for that. What kind of value this is depends on whether we want an
554 /// integer or a block address, but an undef is always accepted.
555 /// Returns null if Val is null or not an appropriate constant.
557  if (!Val)
558  return nullptr;
559 
560  // Undef is "known" enough.
561  if (UndefValue *U = dyn_cast<UndefValue>(Val))
562  return U;
563 
564  if (Preference == WantBlockAddress)
565  return dyn_cast<BlockAddress>(Val->stripPointerCasts());
566 
567  return dyn_cast<ConstantInt>(Val);
568 }
569 
570 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
571 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
572 /// in any of our predecessors. If so, return the known list of value and pred
573 /// BB in the result vector.
574 ///
575 /// This returns true if there were any known values.
577  Value *V, BasicBlock *BB, PredValueInfo &Result,
579  DenseSet<std::pair<Value *, BasicBlock *>> &RecursionSet,
580  Instruction *CxtI) {
581  // This method walks up use-def chains recursively. Because of this, we could
582  // get into an infinite loop going around loops in the use-def chain. To
583  // prevent this, keep track of what (value, block) pairs we've already visited
584  // and terminate the search if we loop back to them
585  if (!RecursionSet.insert(std::make_pair(V, BB)).second)
586  return false;
587 
588  // If V is a constant, then it is known in all predecessors.
589  if (Constant *KC = getKnownConstant(V, Preference)) {
590  for (BasicBlock *Pred : predecessors(BB))
591  Result.push_back(std::make_pair(KC, Pred));
592 
593  return !Result.empty();
594  }
595 
596  // If V is a non-instruction value, or an instruction in a different block,
597  // then it can't be derived from a PHI.
599  if (!I || I->getParent() != BB) {
600 
601  // Okay, if this is a live-in value, see if it has a known value at the end
602  // of any of our predecessors.
603  //
604  // FIXME: This should be an edge property, not a block end property.
605  /// TODO: Per PR2563, we could infer value range information about a
606  /// predecessor based on its terminator.
607  //
608  // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
609  // "I" is a non-local compare-with-a-constant instruction. This would be
610  // able to handle value inequalities better, for example if the compare is
611  // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
612  // Perhaps getConstantOnEdge should be smart enough to do this?
613 
614  if (DTU->hasPendingDomTreeUpdates())
615  LVI->disableDT();
616  else
617  LVI->enableDT();
618  for (BasicBlock *P : predecessors(BB)) {
619  // If the value is known by LazyValueInfo to be a constant in a
620  // predecessor, use that information to try to thread this block.
621  Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
622  if (Constant *KC = getKnownConstant(PredCst, Preference))
623  Result.push_back(std::make_pair(KC, P));
624  }
625 
626  return !Result.empty();
627  }
628 
629  /// If I is a PHI node, then we know the incoming values for any constants.
630  if (PHINode *PN = dyn_cast<PHINode>(I)) {
631  if (DTU->hasPendingDomTreeUpdates())
632  LVI->disableDT();
633  else
634  LVI->enableDT();
635  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
636  Value *InVal = PN->getIncomingValue(i);
637  if (Constant *KC = getKnownConstant(InVal, Preference)) {
638  Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
639  } else {
640  Constant *CI = LVI->getConstantOnEdge(InVal,
641  PN->getIncomingBlock(i),
642  BB, CxtI);
643  if (Constant *KC = getKnownConstant(CI, Preference))
644  Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
645  }
646  }
647 
648  return !Result.empty();
649  }
650 
651  // Handle Cast instructions. Only see through Cast when the source operand is
652  // PHI or Cmp to save the compilation time.
653  if (CastInst *CI = dyn_cast<CastInst>(I)) {
654  Value *Source = CI->getOperand(0);
655  if (!isa<PHINode>(Source) && !isa<CmpInst>(Source))
656  return false;
657  ComputeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
658  RecursionSet, CxtI);
659  if (Result.empty())
660  return false;
661 
662  // Convert the known values.
663  for (auto &R : Result)
664  R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
665 
666  return true;
667  }
668 
669  // Handle some boolean conditions.
670  if (I->getType()->getPrimitiveSizeInBits() == 1) {
671  assert(Preference == WantInteger && "One-bit non-integer type?");
672  // X | true -> true
673  // X & false -> false
674  if (I->getOpcode() == Instruction::Or ||
675  I->getOpcode() == Instruction::And) {
676  PredValueInfoTy LHSVals, RHSVals;
677 
678  ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
679  WantInteger, RecursionSet, CxtI);
680  ComputeValueKnownInPredecessorsImpl(I->getOperand(1), BB, RHSVals,
681  WantInteger, RecursionSet, CxtI);
682 
683  if (LHSVals.empty() && RHSVals.empty())
684  return false;
685 
686  ConstantInt *InterestingVal;
687  if (I->getOpcode() == Instruction::Or)
688  InterestingVal = ConstantInt::getTrue(I->getContext());
689  else
690  InterestingVal = ConstantInt::getFalse(I->getContext());
691 
692  SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
693 
694  // Scan for the sentinel. If we find an undef, force it to the
695  // interesting value: x|undef -> true and x&undef -> false.
696  for (const auto &LHSVal : LHSVals)
697  if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
698  Result.emplace_back(InterestingVal, LHSVal.second);
699  LHSKnownBBs.insert(LHSVal.second);
700  }
701  for (const auto &RHSVal : RHSVals)
702  if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
703  // If we already inferred a value for this block on the LHS, don't
704  // re-add it.
705  if (!LHSKnownBBs.count(RHSVal.second))
706  Result.emplace_back(InterestingVal, RHSVal.second);
707  }
708 
709  return !Result.empty();
710  }
711 
712  // Handle the NOT form of XOR.
713  if (I->getOpcode() == Instruction::Xor &&
714  isa<ConstantInt>(I->getOperand(1)) &&
715  cast<ConstantInt>(I->getOperand(1))->isOne()) {
716  ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, Result,
717  WantInteger, RecursionSet, CxtI);
718  if (Result.empty())
719  return false;
720 
721  // Invert the known values.
722  for (auto &R : Result)
723  R.first = ConstantExpr::getNot(R.first);
724 
725  return true;
726  }
727 
728  // Try to simplify some other binary operator values.
729  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
730  assert(Preference != WantBlockAddress
731  && "A binary operator creating a block address?");
732  if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
733  PredValueInfoTy LHSVals;
734  ComputeValueKnownInPredecessorsImpl(BO->getOperand(0), BB, LHSVals,
735  WantInteger, RecursionSet, CxtI);
736 
737  // Try to use constant folding to simplify the binary operator.
738  for (const auto &LHSVal : LHSVals) {
739  Constant *V = LHSVal.first;
740  Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
741 
742  if (Constant *KC = getKnownConstant(Folded, WantInteger))
743  Result.push_back(std::make_pair(KC, LHSVal.second));
744  }
745  }
746 
747  return !Result.empty();
748  }
749 
750  // Handle compare with phi operand, where the PHI is defined in this block.
751  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
752  assert(Preference == WantInteger && "Compares only produce integers");
753  Type *CmpType = Cmp->getType();
754  Value *CmpLHS = Cmp->getOperand(0);
755  Value *CmpRHS = Cmp->getOperand(1);
756  CmpInst::Predicate Pred = Cmp->getPredicate();
757 
758  PHINode *PN = dyn_cast<PHINode>(CmpLHS);
759  if (!PN)
760  PN = dyn_cast<PHINode>(CmpRHS);
761  if (PN && PN->getParent() == BB) {
762  const DataLayout &DL = PN->getModule()->getDataLayout();
763  // We can do this simplification if any comparisons fold to true or false.
764  // See if any do.
765  if (DTU->hasPendingDomTreeUpdates())
766  LVI->disableDT();
767  else
768  LVI->enableDT();
769  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
770  BasicBlock *PredBB = PN->getIncomingBlock(i);
771  Value *LHS, *RHS;
772  if (PN == CmpLHS) {
773  LHS = PN->getIncomingValue(i);
774  RHS = CmpRHS->DoPHITranslation(BB, PredBB);
775  } else {
776  LHS = CmpLHS->DoPHITranslation(BB, PredBB);
777  RHS = PN->getIncomingValue(i);
778  }
779  Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
780  if (!Res) {
781  if (!isa<Constant>(RHS))
782  continue;
783 
784  // getPredicateOnEdge call will make no sense if LHS is defined in BB.
785  auto LHSInst = dyn_cast<Instruction>(LHS);
786  if (LHSInst && LHSInst->getParent() == BB)
787  continue;
788 
790  ResT = LVI->getPredicateOnEdge(Pred, LHS,
791  cast<Constant>(RHS), PredBB, BB,
792  CxtI ? CxtI : Cmp);
793  if (ResT == LazyValueInfo::Unknown)
794  continue;
795  Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
796  }
797 
798  if (Constant *KC = getKnownConstant(Res, WantInteger))
799  Result.push_back(std::make_pair(KC, PredBB));
800  }
801 
802  return !Result.empty();
803  }
804 
805  // If comparing a live-in value against a constant, see if we know the
806  // live-in value on any predecessors.
807  if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
808  Constant *CmpConst = cast<Constant>(CmpRHS);
809 
810  if (!isa<Instruction>(CmpLHS) ||
811  cast<Instruction>(CmpLHS)->getParent() != BB) {
812  if (DTU->hasPendingDomTreeUpdates())
813  LVI->disableDT();
814  else
815  LVI->enableDT();
816  for (BasicBlock *P : predecessors(BB)) {
817  // If the value is known by LazyValueInfo to be a constant in a
818  // predecessor, use that information to try to thread this block.
820  LVI->getPredicateOnEdge(Pred, CmpLHS,
821  CmpConst, P, BB, CxtI ? CxtI : Cmp);
822  if (Res == LazyValueInfo::Unknown)
823  continue;
824 
825  Constant *ResC = ConstantInt::get(CmpType, Res);
826  Result.push_back(std::make_pair(ResC, P));
827  }
828 
829  return !Result.empty();
830  }
831 
832  // InstCombine can fold some forms of constant range checks into
833  // (icmp (add (x, C1)), C2). See if we have we have such a thing with
834  // x as a live-in.
835  {
836  using namespace PatternMatch;
837 
838  Value *AddLHS;
839  ConstantInt *AddConst;
840  if (isa<ConstantInt>(CmpConst) &&
841  match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
842  if (!isa<Instruction>(AddLHS) ||
843  cast<Instruction>(AddLHS)->getParent() != BB) {
844  if (DTU->hasPendingDomTreeUpdates())
845  LVI->disableDT();
846  else
847  LVI->enableDT();
848  for (BasicBlock *P : predecessors(BB)) {
849  // If the value is known by LazyValueInfo to be a ConstantRange in
850  // a predecessor, use that information to try to thread this
851  // block.
852  ConstantRange CR = LVI->getConstantRangeOnEdge(
853  AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
854  // Propagate the range through the addition.
855  CR = CR.add(AddConst->getValue());
856 
857  // Get the range where the compare returns true.
859  Pred, cast<ConstantInt>(CmpConst)->getValue());
860 
861  Constant *ResC;
862  if (CmpRange.contains(CR))
863  ResC = ConstantInt::getTrue(CmpType);
864  else if (CmpRange.inverse().contains(CR))
865  ResC = ConstantInt::getFalse(CmpType);
866  else
867  continue;
868 
869  Result.push_back(std::make_pair(ResC, P));
870  }
871 
872  return !Result.empty();
873  }
874  }
875  }
876 
877  // Try to find a constant value for the LHS of a comparison,
878  // and evaluate it statically if we can.
879  PredValueInfoTy LHSVals;
880  ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
881  WantInteger, RecursionSet, CxtI);
882 
883  for (const auto &LHSVal : LHSVals) {
884  Constant *V = LHSVal.first;
885  Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
886  if (Constant *KC = getKnownConstant(Folded, WantInteger))
887  Result.push_back(std::make_pair(KC, LHSVal.second));
888  }
889 
890  return !Result.empty();
891  }
892  }
893 
894  if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
895  // Handle select instructions where at least one operand is a known constant
896  // and we can figure out the condition value for any predecessor block.
897  Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
898  Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
899  PredValueInfoTy Conds;
900  if ((TrueVal || FalseVal) &&
901  ComputeValueKnownInPredecessorsImpl(SI->getCondition(), BB, Conds,
902  WantInteger, RecursionSet, CxtI)) {
903  for (auto &C : Conds) {
904  Constant *Cond = C.first;
905 
906  // Figure out what value to use for the condition.
907  bool KnownCond;
908  if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
909  // A known boolean.
910  KnownCond = CI->isOne();
911  } else {
912  assert(isa<UndefValue>(Cond) && "Unexpected condition value");
913  // Either operand will do, so be sure to pick the one that's a known
914  // constant.
915  // FIXME: Do this more cleverly if both values are known constants?
916  KnownCond = (TrueVal != nullptr);
917  }
918 
919  // See if the select has a known constant value for this predecessor.
920  if (Constant *Val = KnownCond ? TrueVal : FalseVal)
921  Result.push_back(std::make_pair(Val, C.second));
922  }
923 
924  return !Result.empty();
925  }
926  }
927 
928  // If all else fails, see if LVI can figure out a constant value for us.
929  if (DTU->hasPendingDomTreeUpdates())
930  LVI->disableDT();
931  else
932  LVI->enableDT();
933  Constant *CI = LVI->getConstant(V, BB, CxtI);
934  if (Constant *KC = getKnownConstant(CI, Preference)) {
935  for (BasicBlock *Pred : predecessors(BB))
936  Result.push_back(std::make_pair(KC, Pred));
937  }
938 
939  return !Result.empty();
940 }
941 
942 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
943 /// in an undefined jump, decide which block is best to revector to.
944 ///
945 /// Since we can pick an arbitrary destination, we pick the successor with the
946 /// fewest predecessors. This should reduce the in-degree of the others.
947 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
948  Instruction *BBTerm = BB->getTerminator();
949  unsigned MinSucc = 0;
950  BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
951  // Compute the successor with the minimum number of predecessors.
952  unsigned MinNumPreds = pred_size(TestBB);
953  for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
954  TestBB = BBTerm->getSuccessor(i);
955  unsigned NumPreds = pred_size(TestBB);
956  if (NumPreds < MinNumPreds) {
957  MinSucc = i;
958  MinNumPreds = NumPreds;
959  }
960  }
961 
962  return MinSucc;
963 }
964 
966  if (!BB->hasAddressTaken()) return false;
967 
968  // If the block has its address taken, it may be a tree of dead constants
969  // hanging off of it. These shouldn't keep the block alive.
972  return !BA->use_empty();
973 }
974 
975 /// ProcessBlock - If there are any predecessors whose control can be threaded
976 /// through to a successor, transform them now.
978  // If the block is trivially dead, just return and let the caller nuke it.
979  // This simplifies other transformations.
980  if (DTU->isBBPendingDeletion(BB) ||
981  (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
982  return false;
983 
984  // If this block has a single predecessor, and if that pred has a single
985  // successor, merge the blocks. This encourages recursive jump threading
986  // because now the condition in this block can be threaded through
987  // predecessors of our predecessor block.
988  if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
989  const Instruction *TI = SinglePred->getTerminator();
990  if (!TI->isExceptionalTerminator() && TI->getNumSuccessors() == 1 &&
991  SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
992  // If SinglePred was a loop header, BB becomes one.
993  if (LoopHeaders.erase(SinglePred))
994  LoopHeaders.insert(BB);
995 
996  LVI->eraseBlock(SinglePred);
998 
999  // Now that BB is merged into SinglePred (i.e. SinglePred Code followed by
1000  // BB code within one basic block `BB`), we need to invalidate the LVI
1001  // information associated with BB, because the LVI information need not be
1002  // true for all of BB after the merge. For example,
1003  // Before the merge, LVI info and code is as follows:
1004  // SinglePred: <LVI info1 for %p val>
1005  // %y = use of %p
1006  // call @exit() // need not transfer execution to successor.
1007  // assume(%p) // from this point on %p is true
1008  // br label %BB
1009  // BB: <LVI info2 for %p val, i.e. %p is true>
1010  // %x = use of %p
1011  // br label exit
1012  //
1013  // Note that this LVI info for blocks BB and SinglPred is correct for %p
1014  // (info2 and info1 respectively). After the merge and the deletion of the
1015  // LVI info1 for SinglePred. We have the following code:
1016  // BB: <LVI info2 for %p val>
1017  // %y = use of %p
1018  // call @exit()
1019  // assume(%p)
1020  // %x = use of %p <-- LVI info2 is correct from here onwards.
1021  // br label exit
1022  // LVI info2 for BB is incorrect at the beginning of BB.
1023 
1024  // Invalidate LVI information for BB if the LVI is not provably true for
1025  // all of BB.
1027  LVI->eraseBlock(BB);
1028  return true;
1029  }
1030  }
1031 
1032  if (TryToUnfoldSelectInCurrBB(BB))
1033  return true;
1034 
1035  // Look if we can propagate guards to predecessors.
1036  if (HasGuards && ProcessGuards(BB))
1037  return true;
1038 
1039  // What kind of constant we're looking for.
1041 
1042  // Look to see if the terminator is a conditional branch, switch or indirect
1043  // branch, if not we can't thread it.
1044  Value *Condition;
1046  if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
1047  // Can't thread an unconditional jump.
1048  if (BI->isUnconditional()) return false;
1049  Condition = BI->getCondition();
1050  } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
1051  Condition = SI->getCondition();
1052  } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
1053  // Can't thread indirect branch with no successors.
1054  if (IB->getNumSuccessors() == 0) return false;
1055  Condition = IB->getAddress()->stripPointerCasts();
1056  Preference = WantBlockAddress;
1057  } else {
1058  return false; // Must be an invoke or callbr.
1059  }
1060 
1061  // Run constant folding to see if we can reduce the condition to a simple
1062  // constant.
1063  if (Instruction *I = dyn_cast<Instruction>(Condition)) {
1064  Value *SimpleVal =
1066  if (SimpleVal) {
1067  I->replaceAllUsesWith(SimpleVal);
1068  if (isInstructionTriviallyDead(I, TLI))
1069  I->eraseFromParent();
1070  Condition = SimpleVal;
1071  }
1072  }
1073 
1074  // If the terminator is branching on an undef, we can pick any of the
1075  // successors to branch to. Let GetBestDestForJumpOnUndef decide.
1076  if (isa<UndefValue>(Condition)) {
1077  unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
1078  std::vector<DominatorTree::UpdateType> Updates;
1079 
1080  // Fold the branch/switch.
1081  Instruction *BBTerm = BB->getTerminator();
1082  Updates.reserve(BBTerm->getNumSuccessors());
1083  for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1084  if (i == BestSucc) continue;
1085  BasicBlock *Succ = BBTerm->getSuccessor(i);
1086  Succ->removePredecessor(BB, true);
1087  Updates.push_back({DominatorTree::Delete, BB, Succ});
1088  }
1089 
1090  LLVM_DEBUG(dbgs() << " In block '" << BB->getName()
1091  << "' folding undef terminator: " << *BBTerm << '\n');
1092  BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
1093  BBTerm->eraseFromParent();
1094  DTU->applyUpdatesPermissive(Updates);
1095  return true;
1096  }
1097 
1098  // If the terminator of this block is branching on a constant, simplify the
1099  // terminator to an unconditional branch. This can occur due to threading in
1100  // other blocks.
1101  if (getKnownConstant(Condition, Preference)) {
1102  LLVM_DEBUG(dbgs() << " In block '" << BB->getName()
1103  << "' folding terminator: " << *BB->getTerminator()
1104  << '\n');
1105  ++NumFolds;
1106  ConstantFoldTerminator(BB, true, nullptr, DTU);
1107  return true;
1108  }
1109 
1110  Instruction *CondInst = dyn_cast<Instruction>(Condition);
1111 
1112  // All the rest of our checks depend on the condition being an instruction.
1113  if (!CondInst) {
1114  // FIXME: Unify this with code below.
1115  if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
1116  return true;
1117  return false;
1118  }
1119 
1120  if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
1121  // If we're branching on a conditional, LVI might be able to determine
1122  // it's value at the branch instruction. We only handle comparisons
1123  // against a constant at this time.
1124  // TODO: This should be extended to handle switches as well.
1125  BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1126  Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
1127  if (CondBr && CondConst) {
1128  // We should have returned as soon as we turn a conditional branch to
1129  // unconditional. Because its no longer interesting as far as jump
1130  // threading is concerned.
1131  assert(CondBr->isConditional() && "Threading on unconditional terminator");
1132 
1133  if (DTU->hasPendingDomTreeUpdates())
1134  LVI->disableDT();
1135  else
1136  LVI->enableDT();
1138  LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
1139  CondConst, CondBr);
1140  if (Ret != LazyValueInfo::Unknown) {
1141  unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
1142  unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
1143  BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove);
1144  ToRemoveSucc->removePredecessor(BB, true);
1145  BranchInst *UncondBr =
1146  BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
1147  UncondBr->setDebugLoc(CondBr->getDebugLoc());
1148  CondBr->eraseFromParent();
1149  if (CondCmp->use_empty())
1150  CondCmp->eraseFromParent();
1151  // We can safely replace *some* uses of the CondInst if it has
1152  // exactly one value as returned by LVI. RAUW is incorrect in the
1153  // presence of guards and assumes, that have the `Cond` as the use. This
1154  // is because we use the guards/assume to reason about the `Cond` value
1155  // at the end of block, but RAUW unconditionally replaces all uses
1156  // including the guards/assumes themselves and the uses before the
1157  // guard/assume.
1158  else if (CondCmp->getParent() == BB) {
1159  auto *CI = Ret == LazyValueInfo::True ?
1160  ConstantInt::getTrue(CondCmp->getType()) :
1161  ConstantInt::getFalse(CondCmp->getType());
1162  ReplaceFoldableUses(CondCmp, CI);
1163  }
1164  DTU->applyUpdatesPermissive(
1165  {{DominatorTree::Delete, BB, ToRemoveSucc}});
1166  return true;
1167  }
1168 
1169  // We did not manage to simplify this branch, try to see whether
1170  // CondCmp depends on a known phi-select pattern.
1171  if (TryToUnfoldSelect(CondCmp, BB))
1172  return true;
1173  }
1174  }
1175 
1176  if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
1177  TryToUnfoldSelect(SI, BB);
1178 
1179  // Check for some cases that are worth simplifying. Right now we want to look
1180  // for loads that are used by a switch or by the condition for the branch. If
1181  // we see one, check to see if it's partially redundant. If so, insert a PHI
1182  // which can then be used to thread the values.
1183  Value *SimplifyValue = CondInst;
1184  if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
1185  if (isa<Constant>(CondCmp->getOperand(1)))
1186  SimplifyValue = CondCmp->getOperand(0);
1187 
1188  // TODO: There are other places where load PRE would be profitable, such as
1189  // more complex comparisons.
1190  if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue))
1191  if (SimplifyPartiallyRedundantLoad(LoadI))
1192  return true;
1193 
1194  // Before threading, try to propagate profile data backwards:
1195  if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1196  if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1198 
1199  // Handle a variety of cases where we are branching on something derived from
1200  // a PHI node in the current block. If we can prove that any predecessors
1201  // compute a predictable value based on a PHI node, thread those predecessors.
1202  if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
1203  return true;
1204 
1205  // If this is an otherwise-unfoldable branch on a phi node in the current
1206  // block, see if we can simplify.
1207  if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1208  if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1209  return ProcessBranchOnPHI(PN);
1210 
1211  // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1212  if (CondInst->getOpcode() == Instruction::Xor &&
1213  CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1214  return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
1215 
1216  // Search for a stronger dominating condition that can be used to simplify a
1217  // conditional branch leaving BB.
1218  if (ProcessImpliedCondition(BB))
1219  return true;
1220 
1221  return false;
1222 }
1223 
1225  auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
1226  if (!BI || !BI->isConditional())
1227  return false;
1228 
1229  Value *Cond = BI->getCondition();
1230  BasicBlock *CurrentBB = BB;
1231  BasicBlock *CurrentPred = BB->getSinglePredecessor();
1232  unsigned Iter = 0;
1233 
1234  auto &DL = BB->getModule()->getDataLayout();
1235 
1236  while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1237  auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
1238  if (!PBI || !PBI->isConditional())
1239  return false;
1240  if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
1241  return false;
1242 
1243  bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB;
1244  Optional<bool> Implication =
1245  isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue);
1246  if (Implication) {
1247  BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1);
1248  BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0);
1249  RemoveSucc->removePredecessor(BB);
1250  BranchInst *UncondBI = BranchInst::Create(KeepSucc, BI);
1251  UncondBI->setDebugLoc(BI->getDebugLoc());
1252  BI->eraseFromParent();
1253  DTU->applyUpdatesPermissive({{DominatorTree::Delete, BB, RemoveSucc}});
1254  return true;
1255  }
1256  CurrentBB = CurrentPred;
1257  CurrentPred = CurrentBB->getSinglePredecessor();
1258  }
1259 
1260  return false;
1261 }
1262 
1263 /// Return true if Op is an instruction defined in the given block.
1265  if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1266  if (OpInst->getParent() == BB)
1267  return true;
1268  return false;
1269 }
1270 
1271 /// SimplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1272 /// redundant load instruction, eliminate it by replacing it with a PHI node.
1273 /// This is an important optimization that encourages jump threading, and needs
1274 /// to be run interlaced with other jump threading tasks.
1276  // Don't hack volatile and ordered loads.
1277  if (!LoadI->isUnordered()) return false;
1278 
1279  // If the load is defined in a block with exactly one predecessor, it can't be
1280  // partially redundant.
1281  BasicBlock *LoadBB = LoadI->getParent();
1282  if (LoadBB->getSinglePredecessor())
1283  return false;
1284 
1285  // If the load is defined in an EH pad, it can't be partially redundant,
1286  // because the edges between the invoke and the EH pad cannot have other
1287  // instructions between them.
1288  if (LoadBB->isEHPad())
1289  return false;
1290 
1291  Value *LoadedPtr = LoadI->getOperand(0);
1292 
1293  // If the loaded operand is defined in the LoadBB and its not a phi,
1294  // it can't be available in predecessors.
1295  if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
1296  return false;
1297 
1298  // Scan a few instructions up from the load, to see if it is obviously live at
1299  // the entry to its block.
1300  BasicBlock::iterator BBIt(LoadI);
1301  bool IsLoadCSE;
1302  if (Value *AvailableVal = FindAvailableLoadedValue(
1303  LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1304  // If the value of the load is locally available within the block, just use
1305  // it. This frequently occurs for reg2mem'd allocas.
1306 
1307  if (IsLoadCSE) {
1308  LoadInst *NLoadI = cast<LoadInst>(AvailableVal);
1309  combineMetadataForCSE(NLoadI, LoadI, false);
1310  };
1311 
1312  // If the returned value is the load itself, replace with an undef. This can
1313  // only happen in dead loops.
1314  if (AvailableVal == LoadI)
1315  AvailableVal = UndefValue::get(LoadI->getType());
1316  if (AvailableVal->getType() != LoadI->getType())
1317  AvailableVal = CastInst::CreateBitOrPointerCast(
1318  AvailableVal, LoadI->getType(), "", LoadI);
1319  LoadI->replaceAllUsesWith(AvailableVal);
1320  LoadI->eraseFromParent();
1321  return true;
1322  }
1323 
1324  // Otherwise, if we scanned the whole block and got to the top of the block,
1325  // we know the block is locally transparent to the load. If not, something
1326  // might clobber its value.
1327  if (BBIt != LoadBB->begin())
1328  return false;
1329 
1330  // If all of the loads and stores that feed the value have the same AA tags,
1331  // then we can propagate them onto any newly inserted loads.
1332  AAMDNodes AATags;
1333  LoadI->getAAMetadata(AATags);
1334 
1335  SmallPtrSet<BasicBlock*, 8> PredsScanned;
1336 
1337  using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1338 
1339  AvailablePredsTy AvailablePreds;
1340  BasicBlock *OneUnavailablePred = nullptr;
1341  SmallVector<LoadInst*, 8> CSELoads;
1342 
1343  // If we got here, the loaded value is transparent through to the start of the
1344  // block. Check to see if it is available in any of the predecessor blocks.
1345  for (BasicBlock *PredBB : predecessors(LoadBB)) {
1346  // If we already scanned this predecessor, skip it.
1347  if (!PredsScanned.insert(PredBB).second)
1348  continue;
1349 
1350  BBIt = PredBB->end();
1351  unsigned NumScanedInst = 0;
1352  Value *PredAvailable = nullptr;
1353  // NOTE: We don't CSE load that is volatile or anything stronger than
1354  // unordered, that should have been checked when we entered the function.
1355  assert(LoadI->isUnordered() &&
1356  "Attempting to CSE volatile or atomic loads");
1357  // If this is a load on a phi pointer, phi-translate it and search
1358  // for available load/store to the pointer in predecessors.
1359  Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1360  PredAvailable = FindAvailablePtrLoadStore(
1361  Ptr, LoadI->getType(), LoadI->isAtomic(), PredBB, BBIt,
1362  DefMaxInstsToScan, AA, &IsLoadCSE, &NumScanedInst);
1363 
1364  // If PredBB has a single predecessor, continue scanning through the
1365  // single predecessor.
1366  BasicBlock *SinglePredBB = PredBB;
1367  while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1368  NumScanedInst < DefMaxInstsToScan) {
1369  SinglePredBB = SinglePredBB->getSinglePredecessor();
1370  if (SinglePredBB) {
1371  BBIt = SinglePredBB->end();
1372  PredAvailable = FindAvailablePtrLoadStore(
1373  Ptr, LoadI->getType(), LoadI->isAtomic(), SinglePredBB, BBIt,
1374  (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1375  &NumScanedInst);
1376  }
1377  }
1378 
1379  if (!PredAvailable) {
1380  OneUnavailablePred = PredBB;
1381  continue;
1382  }
1383 
1384  if (IsLoadCSE)
1385  CSELoads.push_back(cast<LoadInst>(PredAvailable));
1386 
1387  // If so, this load is partially redundant. Remember this info so that we
1388  // can create a PHI node.
1389  AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1390  }
1391 
1392  // If the loaded value isn't available in any predecessor, it isn't partially
1393  // redundant.
1394  if (AvailablePreds.empty()) return false;
1395 
1396  // Okay, the loaded value is available in at least one (and maybe all!)
1397  // predecessors. If the value is unavailable in more than one unique
1398  // predecessor, we want to insert a merge block for those common predecessors.
1399  // This ensures that we only have to insert one reload, thus not increasing
1400  // code size.
1401  BasicBlock *UnavailablePred = nullptr;
1402 
1403  // If the value is unavailable in one of predecessors, we will end up
1404  // inserting a new instruction into them. It is only valid if all the
1405  // instructions before LoadI are guaranteed to pass execution to its
1406  // successor, or if LoadI is safe to speculate.
1407  // TODO: If this logic becomes more complex, and we will perform PRE insertion
1408  // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1409  // It requires domination tree analysis, so for this simple case it is an
1410  // overkill.
1411  if (PredsScanned.size() != AvailablePreds.size() &&
1413  for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1415  return false;
1416 
1417  // If there is exactly one predecessor where the value is unavailable, the
1418  // already computed 'OneUnavailablePred' block is it. If it ends in an
1419  // unconditional branch, we know that it isn't a critical edge.
1420  if (PredsScanned.size() == AvailablePreds.size()+1 &&
1421  OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1422  UnavailablePred = OneUnavailablePred;
1423  } else if (PredsScanned.size() != AvailablePreds.size()) {
1424  // Otherwise, we had multiple unavailable predecessors or we had a critical
1425  // edge from the one.
1426  SmallVector<BasicBlock*, 8> PredsToSplit;
1427  SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1428 
1429  for (const auto &AvailablePred : AvailablePreds)
1430  AvailablePredSet.insert(AvailablePred.first);
1431 
1432  // Add all the unavailable predecessors to the PredsToSplit list.
1433  for (BasicBlock *P : predecessors(LoadBB)) {
1434  // If the predecessor is an indirect goto, we can't split the edge.
1435  // Same for CallBr.
1436  if (isa<IndirectBrInst>(P->getTerminator()) ||
1437  isa<CallBrInst>(P->getTerminator()))
1438  return false;
1439 
1440  if (!AvailablePredSet.count(P))
1441  PredsToSplit.push_back(P);
1442  }
1443 
1444  // Split them out to their own block.
1445  UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1446  }
1447 
1448  // If the value isn't available in all predecessors, then there will be
1449  // exactly one where it isn't available. Insert a load on that edge and add
1450  // it to the AvailablePreds list.
1451  if (UnavailablePred) {
1452  assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1453  "Can't handle critical edge here!");
1454  LoadInst *NewVal = new LoadInst(
1455  LoadI->getType(), LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1456  LoadI->getName() + ".pr", false, LoadI->getAlignment(),
1457  LoadI->getOrdering(), LoadI->getSyncScopeID(),
1458  UnavailablePred->getTerminator());
1459  NewVal->setDebugLoc(LoadI->getDebugLoc());
1460  if (AATags)
1461  NewVal->setAAMetadata(AATags);
1462 
1463  AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1464  }
1465 
1466  // Now we know that each predecessor of this block has a value in
1467  // AvailablePreds, sort them for efficient access as we're walking the preds.
1468  array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1469 
1470  // Create a PHI node at the start of the block for the PRE'd load value.
1471  pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1472  PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "",
1473  &LoadBB->front());
1474  PN->takeName(LoadI);
1475  PN->setDebugLoc(LoadI->getDebugLoc());
1476 
1477  // Insert new entries into the PHI for each predecessor. A single block may
1478  // have multiple entries here.
1479  for (pred_iterator PI = PB; PI != PE; ++PI) {
1480  BasicBlock *P = *PI;
1481  AvailablePredsTy::iterator I =
1482  std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1483  std::make_pair(P, (Value*)nullptr));
1484 
1485  assert(I != AvailablePreds.end() && I->first == P &&
1486  "Didn't find entry for predecessor!");
1487 
1488  // If we have an available predecessor but it requires casting, insert the
1489  // cast in the predecessor and use the cast. Note that we have to update the
1490  // AvailablePreds vector as we go so that all of the PHI entries for this
1491  // predecessor use the same bitcast.
1492  Value *&PredV = I->second;
1493  if (PredV->getType() != LoadI->getType())
1494  PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "",
1495  P->getTerminator());
1496 
1497  PN->addIncoming(PredV, I->first);
1498  }
1499 
1500  for (LoadInst *PredLoadI : CSELoads) {
1501  combineMetadataForCSE(PredLoadI, LoadI, true);
1502  }
1503 
1504  LoadI->replaceAllUsesWith(PN);
1505  LoadI->eraseFromParent();
1506 
1507  return true;
1508 }
1509 
1510 /// FindMostPopularDest - The specified list contains multiple possible
1511 /// threadable destinations. Pick the one that occurs the most frequently in
1512 /// the list.
1513 static BasicBlock *
1515  const SmallVectorImpl<std::pair<BasicBlock *,
1516  BasicBlock *>> &PredToDestList) {
1517  assert(!PredToDestList.empty());
1518 
1519  // Determine popularity. If there are multiple possible destinations, we
1520  // explicitly choose to ignore 'undef' destinations. We prefer to thread
1521  // blocks with known and real destinations to threading undef. We'll handle
1522  // them later if interesting.
1523  DenseMap<BasicBlock*, unsigned> DestPopularity;
1524  for (const auto &PredToDest : PredToDestList)
1525  if (PredToDest.second)
1526  DestPopularity[PredToDest.second]++;
1527 
1528  if (DestPopularity.empty())
1529  return nullptr;
1530 
1531  // Find the most popular dest.
1532  DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1533  BasicBlock *MostPopularDest = DPI->first;
1534  unsigned Popularity = DPI->second;
1535  SmallVector<BasicBlock*, 4> SamePopularity;
1536 
1537  for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1538  // If the popularity of this entry isn't higher than the popularity we've
1539  // seen so far, ignore it.
1540  if (DPI->second < Popularity)
1541  ; // ignore.
1542  else if (DPI->second == Popularity) {
1543  // If it is the same as what we've seen so far, keep track of it.
1544  SamePopularity.push_back(DPI->first);
1545  } else {
1546  // If it is more popular, remember it.
1547  SamePopularity.clear();
1548  MostPopularDest = DPI->first;
1549  Popularity = DPI->second;
1550  }
1551  }
1552 
1553  // Okay, now we know the most popular destination. If there is more than one
1554  // destination, we need to determine one. This is arbitrary, but we need
1555  // to make a deterministic decision. Pick the first one that appears in the
1556  // successor list.
1557  if (!SamePopularity.empty()) {
1558  SamePopularity.push_back(MostPopularDest);
1559  Instruction *TI = BB->getTerminator();
1560  for (unsigned i = 0; ; ++i) {
1561  assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1562 
1563  if (!is_contained(SamePopularity, TI->getSuccessor(i)))
1564  continue;
1565 
1566  MostPopularDest = TI->getSuccessor(i);
1567  break;
1568  }
1569  }
1570 
1571  // Okay, we have finally picked the most popular destination.
1572  return MostPopularDest;
1573 }
1574 
1577  Instruction *CxtI) {
1578  // If threading this would thread across a loop header, don't even try to
1579  // thread the edge.
1580  if (LoopHeaders.count(BB))
1581  return false;
1582 
1583  PredValueInfoTy PredValues;
1584  if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1585  return false;
1586 
1587  assert(!PredValues.empty() &&
1588  "ComputeValueKnownInPredecessors returned true with no values");
1589 
1590  LLVM_DEBUG(dbgs() << "IN BB: " << *BB;
1591  for (const auto &PredValue : PredValues) {
1592  dbgs() << " BB '" << BB->getName()
1593  << "': FOUND condition = " << *PredValue.first
1594  << " for pred '" << PredValue.second->getName() << "'.\n";
1595  });
1596 
1597  // Decide what we want to thread through. Convert our list of known values to
1598  // a list of known destinations for each pred. This also discards duplicate
1599  // predecessors and keeps track of the undefined inputs (which are represented
1600  // as a null dest in the PredToDestList).
1601  SmallPtrSet<BasicBlock*, 16> SeenPreds;
1603 
1604  BasicBlock *OnlyDest = nullptr;
1605  BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1606  Constant *OnlyVal = nullptr;
1607  Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1608 
1609  unsigned PredWithKnownDest = 0;
1610  for (const auto &PredValue : PredValues) {
1611  BasicBlock *Pred = PredValue.second;
1612  if (!SeenPreds.insert(Pred).second)
1613  continue; // Duplicate predecessor entry.
1614 
1615  Constant *Val = PredValue.first;
1616 
1617  BasicBlock *DestBB;
1618  if (isa<UndefValue>(Val))
1619  DestBB = nullptr;
1620  else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1621  assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1622  DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1623  } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1624  assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1625  DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1626  } else {
1627  assert(isa<IndirectBrInst>(BB->getTerminator())
1628  && "Unexpected terminator");
1629  assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress");
1630  DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1631  }
1632 
1633  // If we have exactly one destination, remember it for efficiency below.
1634  if (PredToDestList.empty()) {
1635  OnlyDest = DestBB;
1636  OnlyVal = Val;
1637  } else {
1638  if (OnlyDest != DestBB)
1639  OnlyDest = MultipleDestSentinel;
1640  // It possible we have same destination, but different value, e.g. default
1641  // case in switchinst.
1642  if (Val != OnlyVal)
1643  OnlyVal = MultipleVal;
1644  }
1645 
1646  // We know where this predecessor is going.
1647  ++PredWithKnownDest;
1648 
1649  // If the predecessor ends with an indirect goto, we can't change its
1650  // destination. Same for CallBr.
1651  if (isa<IndirectBrInst>(Pred->getTerminator()) ||
1652  isa<CallBrInst>(Pred->getTerminator()))
1653  continue;
1654 
1655  PredToDestList.push_back(std::make_pair(Pred, DestBB));
1656  }
1657 
1658  // If all edges were unthreadable, we fail.
1659  if (PredToDestList.empty())
1660  return false;
1661 
1662  // If all the predecessors go to a single known successor, we want to fold,
1663  // not thread. By doing so, we do not need to duplicate the current block and
1664  // also miss potential opportunities in case we dont/cant duplicate.
1665  if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1666  if (PredWithKnownDest == (size_t)pred_size(BB)) {
1667  bool SeenFirstBranchToOnlyDest = false;
1668  std::vector <DominatorTree::UpdateType> Updates;
1669  Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);
1670  for (BasicBlock *SuccBB : successors(BB)) {
1671  if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1672  SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1673  } else {
1674  SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1675  Updates.push_back({DominatorTree::Delete, BB, SuccBB});
1676  }
1677  }
1678 
1679  // Finally update the terminator.
1680  Instruction *Term = BB->getTerminator();
1681  BranchInst::Create(OnlyDest, Term);
1682  Term->eraseFromParent();
1683  DTU->applyUpdatesPermissive(Updates);
1684 
1685  // If the condition is now dead due to the removal of the old terminator,
1686  // erase it.
1687  if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1688  if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1689  CondInst->eraseFromParent();
1690  // We can safely replace *some* uses of the CondInst if it has
1691  // exactly one value as returned by LVI. RAUW is incorrect in the
1692  // presence of guards and assumes, that have the `Cond` as the use. This
1693  // is because we use the guards/assume to reason about the `Cond` value
1694  // at the end of block, but RAUW unconditionally replaces all uses
1695  // including the guards/assumes themselves and the uses before the
1696  // guard/assume.
1697  else if (OnlyVal && OnlyVal != MultipleVal &&
1698  CondInst->getParent() == BB)
1699  ReplaceFoldableUses(CondInst, OnlyVal);
1700  }
1701  return true;
1702  }
1703  }
1704 
1705  // Determine which is the most common successor. If we have many inputs and
1706  // this block is a switch, we want to start by threading the batch that goes
1707  // to the most popular destination first. If we only know about one
1708  // threadable destination (the common case) we can avoid this.
1709  BasicBlock *MostPopularDest = OnlyDest;
1710 
1711  if (MostPopularDest == MultipleDestSentinel) {
1712  // Remove any loop headers from the Dest list, ThreadEdge conservatively
1713  // won't process them, but we might have other destination that are eligible
1714  // and we still want to process.
1715  erase_if(PredToDestList,
1716  [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1717  return LoopHeaders.count(PredToDest.second) != 0;
1718  });
1719 
1720  if (PredToDestList.empty())
1721  return false;
1722 
1723  MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1724  }
1725 
1726  // Now that we know what the most popular destination is, factor all
1727  // predecessors that will jump to it into a single predecessor.
1728  SmallVector<BasicBlock*, 16> PredsToFactor;
1729  for (const auto &PredToDest : PredToDestList)
1730  if (PredToDest.second == MostPopularDest) {
1731  BasicBlock *Pred = PredToDest.first;
1732 
1733  // This predecessor may be a switch or something else that has multiple
1734  // edges to the block. Factor each of these edges by listing them
1735  // according to # occurrences in PredsToFactor.
1736  for (BasicBlock *Succ : successors(Pred))
1737  if (Succ == BB)
1738  PredsToFactor.push_back(Pred);
1739  }
1740 
1741  // If the threadable edges are branching on an undefined value, we get to pick
1742  // the destination that these predecessors should get to.
1743  if (!MostPopularDest)
1744  MostPopularDest = BB->getTerminator()->
1745  getSuccessor(GetBestDestForJumpOnUndef(BB));
1746 
1747  // Ok, try to thread it!
1748  return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1749 }
1750 
1751 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1752 /// a PHI node in the current block. See if there are any simplifications we
1753 /// can do based on inputs to the phi node.
1755  BasicBlock *BB = PN->getParent();
1756 
1757  // TODO: We could make use of this to do it once for blocks with common PHI
1758  // values.
1760  PredBBs.resize(1);
1761 
1762  // If any of the predecessor blocks end in an unconditional branch, we can
1763  // *duplicate* the conditional branch into that block in order to further
1764  // encourage jump threading and to eliminate cases where we have branch on a
1765  // phi of an icmp (branch on icmp is much better).
1766  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1767  BasicBlock *PredBB = PN->getIncomingBlock(i);
1768  if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1769  if (PredBr->isUnconditional()) {
1770  PredBBs[0] = PredBB;
1771  // Try to duplicate BB into PredBB.
1772  if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1773  return true;
1774  }
1775  }
1776 
1777  return false;
1778 }
1779 
1780 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1781 /// a xor instruction in the current block. See if there are any
1782 /// simplifications we can do based on inputs to the xor.
1784  BasicBlock *BB = BO->getParent();
1785 
1786  // If either the LHS or RHS of the xor is a constant, don't do this
1787  // optimization.
1788  if (isa<ConstantInt>(BO->getOperand(0)) ||
1789  isa<ConstantInt>(BO->getOperand(1)))
1790  return false;
1791 
1792  // If the first instruction in BB isn't a phi, we won't be able to infer
1793  // anything special about any particular predecessor.
1794  if (!isa<PHINode>(BB->front()))
1795  return false;
1796 
1797  // If this BB is a landing pad, we won't be able to split the edge into it.
1798  if (BB->isEHPad())
1799  return false;
1800 
1801  // If we have a xor as the branch input to this block, and we know that the
1802  // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1803  // the condition into the predecessor and fix that value to true, saving some
1804  // logical ops on that path and encouraging other paths to simplify.
1805  //
1806  // This copies something like this:
1807  //
1808  // BB:
1809  // %X = phi i1 [1], [%X']
1810  // %Y = icmp eq i32 %A, %B
1811  // %Z = xor i1 %X, %Y
1812  // br i1 %Z, ...
1813  //
1814  // Into:
1815  // BB':
1816  // %Y = icmp ne i32 %A, %B
1817  // br i1 %Y, ...
1818 
1819  PredValueInfoTy XorOpValues;
1820  bool isLHS = true;
1821  if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1822  WantInteger, BO)) {
1823  assert(XorOpValues.empty());
1824  if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1825  WantInteger, BO))
1826  return false;
1827  isLHS = false;
1828  }
1829 
1830  assert(!XorOpValues.empty() &&
1831  "ComputeValueKnownInPredecessors returned true with no values");
1832 
1833  // Scan the information to see which is most popular: true or false. The
1834  // predecessors can be of the set true, false, or undef.
1835  unsigned NumTrue = 0, NumFalse = 0;
1836  for (const auto &XorOpValue : XorOpValues) {
1837  if (isa<UndefValue>(XorOpValue.first))
1838  // Ignore undefs for the count.
1839  continue;
1840  if (cast<ConstantInt>(XorOpValue.first)->isZero())
1841  ++NumFalse;
1842  else
1843  ++NumTrue;
1844  }
1845 
1846  // Determine which value to split on, true, false, or undef if neither.
1847  ConstantInt *SplitVal = nullptr;
1848  if (NumTrue > NumFalse)
1849  SplitVal = ConstantInt::getTrue(BB->getContext());
1850  else if (NumTrue != 0 || NumFalse != 0)
1851  SplitVal = ConstantInt::getFalse(BB->getContext());
1852 
1853  // Collect all of the blocks that this can be folded into so that we can
1854  // factor this once and clone it once.
1855  SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1856  for (const auto &XorOpValue : XorOpValues) {
1857  if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1858  continue;
1859 
1860  BlocksToFoldInto.push_back(XorOpValue.second);
1861  }
1862 
1863  // If we inferred a value for all of the predecessors, then duplication won't
1864  // help us. However, we can just replace the LHS or RHS with the constant.
1865  if (BlocksToFoldInto.size() ==
1866  cast<PHINode>(BB->front()).getNumIncomingValues()) {
1867  if (!SplitVal) {
1868  // If all preds provide undef, just nuke the xor, because it is undef too.
1870  BO->eraseFromParent();
1871  } else if (SplitVal->isZero()) {
1872  // If all preds provide 0, replace the xor with the other input.
1873  BO->replaceAllUsesWith(BO->getOperand(isLHS));
1874  BO->eraseFromParent();
1875  } else {
1876  // If all preds provide 1, set the computed value to 1.
1877  BO->setOperand(!isLHS, SplitVal);
1878  }
1879 
1880  return true;
1881  }
1882 
1883  // Try to duplicate BB into PredBB.
1884  return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1885 }
1886 
1887 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1888 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1889 /// NewPred using the entries from OldPred (suitably mapped).
1891  BasicBlock *OldPred,
1892  BasicBlock *NewPred,
1894  for (PHINode &PN : PHIBB->phis()) {
1895  // Ok, we have a PHI node. Figure out what the incoming value was for the
1896  // DestBlock.
1897  Value *IV = PN.getIncomingValueForBlock(OldPred);
1898 
1899  // Remap the value if necessary.
1900  if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1902  if (I != ValueMap.end())
1903  IV = I->second;
1904  }
1905 
1906  PN.addIncoming(IV, NewPred);
1907  }
1908 }
1909 
1910 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1911 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1912 /// across BB. Transform the IR to reflect this change.
1914  const SmallVectorImpl<BasicBlock *> &PredBBs,
1915  BasicBlock *SuccBB) {
1916  // If threading to the same block as we come from, we would infinite loop.
1917  if (SuccBB == BB) {
1918  LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1919  << "' - would thread to self!\n");
1920  return false;
1921  }
1922 
1923  // If threading this would thread across a loop header, don't thread the edge.
1924  // See the comments above FindLoopHeaders for justifications and caveats.
1925  if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
1926  LLVM_DEBUG({
1927  bool BBIsHeader = LoopHeaders.count(BB);
1928  bool SuccIsHeader = LoopHeaders.count(SuccBB);
1929  dbgs() << " Not threading across "
1930  << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName()
1931  << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '")
1932  << SuccBB->getName() << "' - it might create an irreducible loop!\n";
1933  });
1934  return false;
1935  }
1936 
1937  unsigned JumpThreadCost =
1938  getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1939  if (JumpThreadCost > BBDupThreshold) {
1940  LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1941  << "' - Cost is too high: " << JumpThreadCost << "\n");
1942  return false;
1943  }
1944 
1945  // And finally, do it! Start by factoring the predecessors if needed.
1946  BasicBlock *PredBB;
1947  if (PredBBs.size() == 1)
1948  PredBB = PredBBs[0];
1949  else {
1950  LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1951  << " common predecessors.\n");
1952  PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1953  }
1954 
1955  // And finally, do it!
1956  LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName()
1957  << "' to '" << SuccBB->getName()
1958  << "' with cost: " << JumpThreadCost
1959  << ", across block:\n " << *BB << "\n");
1960 
1961  if (DTU->hasPendingDomTreeUpdates())
1962  LVI->disableDT();
1963  else
1964  LVI->enableDT();
1965  LVI->threadEdge(PredBB, BB, SuccBB);
1966 
1967  // We are going to have to map operands from the original BB block to the new
1968  // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1969  // account for entry from PredBB.
1971 
1972  BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1973  BB->getName()+".thread",
1974  BB->getParent(), BB);
1975  NewBB->moveAfter(PredBB);
1976 
1977  // Set the block frequency of NewBB.
1978  if (HasProfileData) {
1979  auto NewBBFreq =
1980  BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1981  BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1982  }
1983 
1984  BasicBlock::iterator BI = BB->begin();
1985  for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1986  ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1987 
1988  // Clone the non-phi instructions of BB into NewBB, keeping track of the
1989  // mapping and using it to remap operands in the cloned instructions.
1990  for (; !BI->isTerminator(); ++BI) {
1991  Instruction *New = BI->clone();
1992  New->setName(BI->getName());
1993  NewBB->getInstList().push_back(New);
1994  ValueMapping[&*BI] = New;
1995 
1996  // Remap operands to patch up intra-block references.
1997  for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1998  if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1999  DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2000  if (I != ValueMapping.end())
2001  New->setOperand(i, I->second);
2002  }
2003  }
2004 
2005  // We didn't copy the terminator from BB over to NewBB, because there is now
2006  // an unconditional jump to SuccBB. Insert the unconditional jump.
2007  BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
2008  NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
2009 
2010  // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
2011  // PHI nodes for NewBB now.
2012  AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
2013 
2014  // Update the terminator of PredBB to jump to NewBB instead of BB. This
2015  // eliminates predecessors from BB, which requires us to simplify any PHI
2016  // nodes in BB.
2017  Instruction *PredTerm = PredBB->getTerminator();
2018  for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2019  if (PredTerm->getSuccessor(i) == BB) {
2020  BB->removePredecessor(PredBB, true);
2021  PredTerm->setSuccessor(i, NewBB);
2022  }
2023 
2024  // Enqueue required DT updates.
2025  DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, SuccBB},
2026  {DominatorTree::Insert, PredBB, NewBB},
2027  {DominatorTree::Delete, PredBB, BB}});
2028 
2029  // If there were values defined in BB that are used outside the block, then we
2030  // now have to update all uses of the value to use either the original value,
2031  // the cloned value, or some PHI derived value. This can require arbitrary
2032  // PHI insertion, of which we are prepared to do, clean these up now.
2033  SSAUpdater SSAUpdate;
2034  SmallVector<Use*, 16> UsesToRename;
2035 
2036  for (Instruction &I : *BB) {
2037  // Scan all uses of this instruction to see if their uses are no longer
2038  // dominated by the previous def and if so, record them in UsesToRename.
2039  // Also, skip phi operands from PredBB - we'll remove them anyway.
2040  for (Use &U : I.uses()) {
2041  Instruction *User = cast<Instruction>(U.getUser());
2042  if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2043  if (UserPN->getIncomingBlock(U) == BB)
2044  continue;
2045  } else if (User->getParent() == BB)
2046  continue;
2047 
2048  UsesToRename.push_back(&U);
2049  }
2050 
2051  // If there are no uses outside the block, we're done with this instruction.
2052  if (UsesToRename.empty())
2053  continue;
2054  LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
2055 
2056  // We found a use of I outside of BB. Rename all uses of I that are outside
2057  // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2058  // with the two values we know.
2059  SSAUpdate.Initialize(I.getType(), I.getName());
2060  SSAUpdate.AddAvailableValue(BB, &I);
2061  SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
2062 
2063  while (!UsesToRename.empty())
2064  SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2065  LLVM_DEBUG(dbgs() << "\n");
2066  }
2067 
2068  // At this point, the IR is fully up to date and consistent. Do a quick scan
2069  // over the new instructions and zap any that are constants or dead. This
2070  // frequently happens because of phi translation.
2071  SimplifyInstructionsInBlock(NewBB, TLI);
2072 
2073  // Update the edge weight from BB to SuccBB, which should be less than before.
2074  UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
2075 
2076  // Threaded an edge!
2077  ++NumThreads;
2078  return true;
2079 }
2080 
2081 /// Create a new basic block that will be the predecessor of BB and successor of
2082 /// all blocks in Preds. When profile data is available, update the frequency of
2083 /// this new block.
2084 BasicBlock *JumpThreadingPass::SplitBlockPreds(BasicBlock *BB,
2085  ArrayRef<BasicBlock *> Preds,
2086  const char *Suffix) {
2088 
2089  // Collect the frequencies of all predecessors of BB, which will be used to
2090  // update the edge weight of the result of splitting predecessors.
2092  if (HasProfileData)
2093  for (auto Pred : Preds)
2094  FreqMap.insert(std::make_pair(
2095  Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
2096 
2097  // In the case when BB is a LandingPad block we create 2 new predecessors
2098  // instead of just one.
2099  if (BB->isLandingPad()) {
2100  std::string NewName = std::string(Suffix) + ".split-lp";
2101  SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);
2102  } else {
2103  NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));
2104  }
2105 
2106  std::vector<DominatorTree::UpdateType> Updates;
2107  Updates.reserve((2 * Preds.size()) + NewBBs.size());
2108  for (auto NewBB : NewBBs) {
2109  BlockFrequency NewBBFreq(0);
2110  Updates.push_back({DominatorTree::Insert, NewBB, BB});
2111  for (auto Pred : predecessors(NewBB)) {
2112  Updates.push_back({DominatorTree::Delete, Pred, BB});
2113  Updates.push_back({DominatorTree::Insert, Pred, NewBB});
2114  if (HasProfileData) // Update frequencies between Pred -> NewBB.
2115  NewBBFreq += FreqMap.lookup(Pred);
2116  }
2117  if (HasProfileData) // Apply the summed frequency to NewBB.
2118  BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2119  }
2120 
2121  DTU->applyUpdatesPermissive(Updates);
2122  return NewBBs[0];
2123 }
2124 
2125 bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2126  const Instruction *TI = BB->getTerminator();
2127  assert(TI->getNumSuccessors() > 1 && "not a split");
2128 
2129  MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
2130  if (!WeightsNode)
2131  return false;
2132 
2133  MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
2134  if (MDName->getString() != "branch_weights")
2135  return false;
2136 
2137  // Ensure there are weights for all of the successors. Note that the first
2138  // operand to the metadata node is a name, not a weight.
2139  return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
2140 }
2141 
2142 /// Update the block frequency of BB and branch weight and the metadata on the
2143 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2144 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
2145 void JumpThreadingPass::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2146  BasicBlock *BB,
2147  BasicBlock *NewBB,
2148  BasicBlock *SuccBB) {
2149  if (!HasProfileData)
2150  return;
2151 
2152  assert(BFI && BPI && "BFI & BPI should have been created here");
2153 
2154  // As the edge from PredBB to BB is deleted, we have to update the block
2155  // frequency of BB.
2156  auto BBOrigFreq = BFI->getBlockFreq(BB);
2157  auto NewBBFreq = BFI->getBlockFreq(NewBB);
2158  auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
2159  auto BBNewFreq = BBOrigFreq - NewBBFreq;
2160  BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
2161 
2162  // Collect updated outgoing edges' frequencies from BB and use them to update
2163  // edge probabilities.
2164  SmallVector<uint64_t, 4> BBSuccFreq;
2165  for (BasicBlock *Succ : successors(BB)) {
2166  auto SuccFreq = (Succ == SuccBB)
2167  ? BB2SuccBBFreq - NewBBFreq
2168  : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
2169  BBSuccFreq.push_back(SuccFreq.getFrequency());
2170  }
2171 
2172  uint64_t MaxBBSuccFreq =
2173  *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
2174 
2176  if (MaxBBSuccFreq == 0)
2177  BBSuccProbs.assign(BBSuccFreq.size(),
2178  {1, static_cast<unsigned>(BBSuccFreq.size())});
2179  else {
2180  for (uint64_t Freq : BBSuccFreq)
2181  BBSuccProbs.push_back(
2182  BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
2183  // Normalize edge probabilities so that they sum up to one.
2184  BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
2185  BBSuccProbs.end());
2186  }
2187 
2188  // Update edge probabilities in BPI.
2189  for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
2190  BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
2191 
2192  // Update the profile metadata as well.
2193  //
2194  // Don't do this if the profile of the transformed blocks was statically
2195  // estimated. (This could occur despite the function having an entry
2196  // frequency in completely cold parts of the CFG.)
2197  //
2198  // In this case we don't want to suggest to subsequent passes that the
2199  // calculated weights are fully consistent. Consider this graph:
2200  //
2201  // check_1
2202  // 50% / |
2203  // eq_1 | 50%
2204  // \ |
2205  // check_2
2206  // 50% / |
2207  // eq_2 | 50%
2208  // \ |
2209  // check_3
2210  // 50% / |
2211  // eq_3 | 50%
2212  // \ |
2213  //
2214  // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2215  // the overall probabilities are inconsistent; the total probability that the
2216  // value is either 1, 2 or 3 is 150%.
2217  //
2218  // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2219  // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2220  // the loop exit edge. Then based solely on static estimation we would assume
2221  // the loop was extremely hot.
2222  //
2223  // FIXME this locally as well so that BPI and BFI are consistent as well. We
2224  // shouldn't make edges extremely likely or unlikely based solely on static
2225  // estimation.
2226  if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
2227  SmallVector<uint32_t, 4> Weights;
2228  for (auto Prob : BBSuccProbs)
2229  Weights.push_back(Prob.getNumerator());
2230 
2231  auto TI = BB->getTerminator();
2232  TI->setMetadata(
2234  MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
2235  }
2236 }
2237 
2238 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2239 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2240 /// If we can duplicate the contents of BB up into PredBB do so now, this
2241 /// improves the odds that the branch will be on an analyzable instruction like
2242 /// a compare.
2244  BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2245  assert(!PredBBs.empty() && "Can't handle an empty set");
2246 
2247  // If BB is a loop header, then duplicating this block outside the loop would
2248  // cause us to transform this into an irreducible loop, don't do this.
2249  // See the comments above FindLoopHeaders for justifications and caveats.
2250  if (LoopHeaders.count(BB)) {
2251  LLVM_DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
2252  << "' into predecessor block '" << PredBBs[0]->getName()
2253  << "' - it might create an irreducible loop!\n");
2254  return false;
2255  }
2256 
2257  unsigned DuplicationCost =
2258  getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2259  if (DuplicationCost > BBDupThreshold) {
2260  LLVM_DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
2261  << "' - Cost is too high: " << DuplicationCost << "\n");
2262  return false;
2263  }
2264 
2265  // And finally, do it! Start by factoring the predecessors if needed.
2266  std::vector<DominatorTree::UpdateType> Updates;
2267  BasicBlock *PredBB;
2268  if (PredBBs.size() == 1)
2269  PredBB = PredBBs[0];
2270  else {
2271  LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()
2272  << " common predecessors.\n");
2273  PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
2274  }
2275  Updates.push_back({DominatorTree::Delete, PredBB, BB});
2276 
2277  // Okay, we decided to do this! Clone all the instructions in BB onto the end
2278  // of PredBB.
2279  LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName()
2280  << "' into end of '" << PredBB->getName()
2281  << "' to eliminate branch on phi. Cost: "
2282  << DuplicationCost << " block is:" << *BB << "\n");
2283 
2284  // Unless PredBB ends with an unconditional branch, split the edge so that we
2285  // can just clone the bits from BB into the end of the new PredBB.
2286  BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2287 
2288  if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2289  BasicBlock *OldPredBB = PredBB;
2290  PredBB = SplitEdge(OldPredBB, BB);
2291  Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});
2292  Updates.push_back({DominatorTree::Insert, PredBB, BB});
2293  Updates.push_back({DominatorTree::Delete, OldPredBB, BB});
2294  OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
2295  }
2296 
2297  // We are going to have to map operands from the original BB block into the
2298  // PredBB block. Evaluate PHI nodes in BB.
2300 
2301  BasicBlock::iterator BI = BB->begin();
2302  for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2303  ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2304  // Clone the non-phi instructions of BB into PredBB, keeping track of the
2305  // mapping and using it to remap operands in the cloned instructions.
2306  for (; BI != BB->end(); ++BI) {
2307  Instruction *New = BI->clone();
2308 
2309  // Remap operands to patch up intra-block references.
2310  for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2311  if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2312  DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2313  if (I != ValueMapping.end())
2314  New->setOperand(i, I->second);
2315  }
2316 
2317  // If this instruction can be simplified after the operands are updated,
2318  // just use the simplified value instead. This frequently happens due to
2319  // phi translation.
2320  if (Value *IV = SimplifyInstruction(
2321  New,
2322  {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2323  ValueMapping[&*BI] = IV;
2324  if (!New->mayHaveSideEffects()) {
2325  New->deleteValue();
2326  New = nullptr;
2327  }
2328  } else {
2329  ValueMapping[&*BI] = New;
2330  }
2331  if (New) {
2332  // Otherwise, insert the new instruction into the block.
2333  New->setName(BI->getName());
2334  PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2335  // Update Dominance from simplified New instruction operands.
2336  for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2337  if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))
2338  Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});
2339  }
2340  }
2341 
2342  // Check to see if the targets of the branch had PHI nodes. If so, we need to
2343  // add entries to the PHI nodes for branch from PredBB now.
2344  BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2345  AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2346  ValueMapping);
2347  AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2348  ValueMapping);
2349 
2350  // If there were values defined in BB that are used outside the block, then we
2351  // now have to update all uses of the value to use either the original value,
2352  // the cloned value, or some PHI derived value. This can require arbitrary
2353  // PHI insertion, of which we are prepared to do, clean these up now.
2354  SSAUpdater SSAUpdate;
2355  SmallVector<Use*, 16> UsesToRename;
2356  for (Instruction &I : *BB) {
2357  // Scan all uses of this instruction to see if it is used outside of its
2358  // block, and if so, record them in UsesToRename.
2359  for (Use &U : I.uses()) {
2360  Instruction *User = cast<Instruction>(U.getUser());
2361  if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2362  if (UserPN->getIncomingBlock(U) == BB)
2363  continue;
2364  } else if (User->getParent() == BB)
2365  continue;
2366 
2367  UsesToRename.push_back(&U);
2368  }
2369 
2370  // If there are no uses outside the block, we're done with this instruction.
2371  if (UsesToRename.empty())
2372  continue;
2373 
2374  LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
2375 
2376  // We found a use of I outside of BB. Rename all uses of I that are outside
2377  // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2378  // with the two values we know.
2379  SSAUpdate.Initialize(I.getType(), I.getName());
2380  SSAUpdate.AddAvailableValue(BB, &I);
2381  SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]);
2382 
2383  while (!UsesToRename.empty())
2384  SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2385  LLVM_DEBUG(dbgs() << "\n");
2386  }
2387 
2388  // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2389  // that we nuked.
2390  BB->removePredecessor(PredBB, true);
2391 
2392  // Remove the unconditional branch at the end of the PredBB block.
2393  OldPredBranch->eraseFromParent();
2394  DTU->applyUpdatesPermissive(Updates);
2395 
2396  ++NumDupes;
2397  return true;
2398 }
2399 
2400 // Pred is a predecessor of BB with an unconditional branch to BB. SI is
2401 // a Select instruction in Pred. BB has other predecessors and SI is used in
2402 // a PHI node in BB. SI has no other use.
2403 // A new basic block, NewBB, is created and SI is converted to compare and
2404 // conditional branch. SI is erased from parent.
2406  SelectInst *SI, PHINode *SIUse,
2407  unsigned Idx) {
2408  // Expand the select.
2409  //
2410  // Pred --
2411  // | v
2412  // | NewBB
2413  // | |
2414  // |-----
2415  // v
2416  // BB
2417  BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2418  BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2419  BB->getParent(), BB);
2420  // Move the unconditional branch to NewBB.
2421  PredTerm->removeFromParent();
2422  NewBB->getInstList().insert(NewBB->end(), PredTerm);
2423  // Create a conditional branch and update PHI nodes.
2424  BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2425  SIUse->setIncomingValue(Idx, SI->getFalseValue());
2426  SIUse->addIncoming(SI->getTrueValue(), NewBB);
2427 
2428  // The select is now dead.
2429  SI->eraseFromParent();
2430  DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, BB},
2431  {DominatorTree::Insert, Pred, NewBB}});
2432 
2433  // Update any other PHI nodes in BB.
2434  for (BasicBlock::iterator BI = BB->begin();
2435  PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2436  if (Phi != SIUse)
2437  Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2438 }
2439 
2441  PHINode *CondPHI = dyn_cast<PHINode>(SI->getCondition());
2442 
2443  if (!CondPHI || CondPHI->getParent() != BB)
2444  return false;
2445 
2446  for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) {
2447  BasicBlock *Pred = CondPHI->getIncomingBlock(I);
2448  SelectInst *PredSI = dyn_cast<SelectInst>(CondPHI->getIncomingValue(I));
2449 
2450  // The second and third condition can be potentially relaxed. Currently
2451  // the conditions help to simplify the code and allow us to reuse existing
2452  // code, developed for TryToUnfoldSelect(CmpInst *, BasicBlock *)
2453  if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse())
2454  continue;
2455 
2456  BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2457  if (!PredTerm || !PredTerm->isUnconditional())
2458  continue;
2459 
2460  UnfoldSelectInstr(Pred, BB, PredSI, CondPHI, I);
2461  return true;
2462  }
2463  return false;
2464 }
2465 
2466 /// TryToUnfoldSelect - Look for blocks of the form
2467 /// bb1:
2468 /// %a = select
2469 /// br bb2
2470 ///
2471 /// bb2:
2472 /// %p = phi [%a, %bb1] ...
2473 /// %c = icmp %p
2474 /// br i1 %c
2475 ///
2476 /// And expand the select into a branch structure if one of its arms allows %c
2477 /// to be folded. This later enables threading from bb1 over bb2.
2479  BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2480  PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2481  Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2482 
2483  if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2484  CondLHS->getParent() != BB)
2485  return false;
2486 
2487  for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2488  BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2489  SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2490 
2491  // Look if one of the incoming values is a select in the corresponding
2492  // predecessor.
2493  if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2494  continue;
2495 
2496  BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2497  if (!PredTerm || !PredTerm->isUnconditional())
2498  continue;
2499 
2500  // Now check if one of the select values would allow us to constant fold the
2501  // terminator in BB. We don't do the transform if both sides fold, those
2502  // cases will be threaded in any case.
2503  if (DTU->hasPendingDomTreeUpdates())
2504  LVI->disableDT();
2505  else
2506  LVI->enableDT();
2507  LazyValueInfo::Tristate LHSFolds =
2508  LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2509  CondRHS, Pred, BB, CondCmp);
2510  LazyValueInfo::Tristate RHSFolds =
2511  LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2512  CondRHS, Pred, BB, CondCmp);
2513  if ((LHSFolds != LazyValueInfo::Unknown ||
2514  RHSFolds != LazyValueInfo::Unknown) &&
2515  LHSFolds != RHSFolds) {
2516  UnfoldSelectInstr(Pred, BB, SI, CondLHS, I);
2517  return true;
2518  }
2519  }
2520  return false;
2521 }
2522 
2523 /// TryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2524 /// same BB in the form
2525 /// bb:
2526 /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2527 /// %s = select %p, trueval, falseval
2528 ///
2529 /// or
2530 ///
2531 /// bb:
2532 /// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2533 /// %c = cmp %p, 0
2534 /// %s = select %c, trueval, falseval
2535 ///
2536 /// And expand the select into a branch structure. This later enables
2537 /// jump-threading over bb in this pass.
2538 ///
2539 /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2540 /// select if the associated PHI has at least one constant. If the unfolded
2541 /// select is not jump-threaded, it will be folded again in the later
2542 /// optimizations.
2544  // If threading this would thread across a loop header, don't thread the edge.
2545  // See the comments above FindLoopHeaders for justifications and caveats.
2546  if (LoopHeaders.count(BB))
2547  return false;
2548 
2549  for (BasicBlock::iterator BI = BB->begin();
2550  PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2551  // Look for a Phi having at least one constant incoming value.
2552  if (llvm::all_of(PN->incoming_values(),
2553  [](Value *V) { return !isa<ConstantInt>(V); }))
2554  continue;
2555 
2556  auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2557  // Check if SI is in BB and use V as condition.
2558  if (SI->getParent() != BB)
2559  return false;
2560  Value *Cond = SI->getCondition();
2561  return (Cond && Cond == V && Cond->getType()->isIntegerTy(1));
2562  };
2563 
2564  SelectInst *SI = nullptr;
2565  for (Use &U : PN->uses()) {
2566  if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2567  // Look for a ICmp in BB that compares PN with a constant and is the
2568  // condition of a Select.
2569  if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2570  isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2571  if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2572  if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2573  SI = SelectI;
2574  break;
2575  }
2576  } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2577  // Look for a Select in BB that uses PN as condition.
2578  if (isUnfoldCandidate(SelectI, U.get())) {
2579  SI = SelectI;
2580  break;
2581  }
2582  }
2583  }
2584 
2585  if (!SI)
2586  continue;
2587  // Expand the select.
2588  Instruction *Term =
2589  SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
2590  BasicBlock *SplitBB = SI->getParent();
2591  BasicBlock *NewBB = Term->getParent();
2592  PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2593  NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2594  NewPN->addIncoming(SI->getFalseValue(), BB);
2595  SI->replaceAllUsesWith(NewPN);
2596  SI->eraseFromParent();
2597  // NewBB and SplitBB are newly created blocks which require insertion.
2598  std::vector<DominatorTree::UpdateType> Updates;
2599  Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2600  Updates.push_back({DominatorTree::Insert, BB, SplitBB});
2601  Updates.push_back({DominatorTree::Insert, BB, NewBB});
2602  Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});
2603  // BB's successors were moved to SplitBB, update DTU accordingly.
2604  for (auto *Succ : successors(SplitBB)) {
2605  Updates.push_back({DominatorTree::Delete, BB, Succ});
2606  Updates.push_back({DominatorTree::Insert, SplitBB, Succ});
2607  }
2608  DTU->applyUpdatesPermissive(Updates);
2609  return true;
2610  }
2611  return false;
2612 }
2613 
2614 /// Try to propagate a guard from the current BB into one of its predecessors
2615 /// in case if another branch of execution implies that the condition of this
2616 /// guard is always true. Currently we only process the simplest case that
2617 /// looks like:
2618 ///
2619 /// Start:
2620 /// %cond = ...
2621 /// br i1 %cond, label %T1, label %F1
2622 /// T1:
2623 /// br label %Merge
2624 /// F1:
2625 /// br label %Merge
2626 /// Merge:
2627 /// %condGuard = ...
2628 /// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2629 ///
2630 /// And cond either implies condGuard or !condGuard. In this case all the
2631 /// instructions before the guard can be duplicated in both branches, and the
2632 /// guard is then threaded to one of them.
2634  using namespace PatternMatch;
2635 
2636  // We only want to deal with two predecessors.
2637  BasicBlock *Pred1, *Pred2;
2638  auto PI = pred_begin(BB), PE = pred_end(BB);
2639  if (PI == PE)
2640  return false;
2641  Pred1 = *PI++;
2642  if (PI == PE)
2643  return false;
2644  Pred2 = *PI++;
2645  if (PI != PE)
2646  return false;
2647  if (Pred1 == Pred2)
2648  return false;
2649 
2650  // Try to thread one of the guards of the block.
2651  // TODO: Look up deeper than to immediate predecessor?
2652  auto *Parent = Pred1->getSinglePredecessor();
2653  if (!Parent || Parent != Pred2->getSinglePredecessor())
2654  return false;
2655 
2656  if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2657  for (auto &I : *BB)
2658  if (isGuard(&I) && ThreadGuard(BB, cast<IntrinsicInst>(&I), BI))
2659  return true;
2660 
2661  return false;
2662 }
2663 
2664 /// Try to propagate the guard from BB which is the lower block of a diamond
2665 /// to one of its branches, in case if diamond's condition implies guard's
2666 /// condition.
2668  BranchInst *BI) {
2669  assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?");
2670  assert(BI->isConditional() && "Unconditional branch has 2 successors?");
2671  Value *GuardCond = Guard->getArgOperand(0);
2672  Value *BranchCond = BI->getCondition();
2673  BasicBlock *TrueDest = BI->getSuccessor(0);
2674  BasicBlock *FalseDest = BI->getSuccessor(1);
2675 
2676  auto &DL = BB->getModule()->getDataLayout();
2677  bool TrueDestIsSafe = false;
2678  bool FalseDestIsSafe = false;
2679 
2680  // True dest is safe if BranchCond => GuardCond.
2681  auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
2682  if (Impl && *Impl)
2683  TrueDestIsSafe = true;
2684  else {
2685  // False dest is safe if !BranchCond => GuardCond.
2686  Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);
2687  if (Impl && *Impl)
2688  FalseDestIsSafe = true;
2689  }
2690 
2691  if (!TrueDestIsSafe && !FalseDestIsSafe)
2692  return false;
2693 
2694  BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
2695  BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
2696 
2697  ValueToValueMapTy UnguardedMapping, GuardedMapping;
2698  Instruction *AfterGuard = Guard->getNextNode();
2699  unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
2700  if (Cost > BBDupThreshold)
2701  return false;
2702  // Duplicate all instructions before the guard and the guard itself to the
2703  // branch where implication is not proved.
2705  BB, PredGuardedBlock, AfterGuard, GuardedMapping, *DTU);
2706  assert(GuardedBlock && "Could not create the guarded block?");
2707  // Duplicate all instructions before the guard in the unguarded branch.
2708  // Since we have successfully duplicated the guarded block and this block
2709  // has fewer instructions, we expect it to succeed.
2711  BB, PredUnguardedBlock, Guard, UnguardedMapping, *DTU);
2712  assert(UnguardedBlock && "Could not create the unguarded block?");
2713  LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block "
2714  << GuardedBlock->getName() << "\n");
2715  // Some instructions before the guard may still have uses. For them, we need
2716  // to create Phi nodes merging their copies in both guarded and unguarded
2717  // branches. Those instructions that have no uses can be just removed.
2719  for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
2720  if (!isa<PHINode>(&*BI))
2721  ToRemove.push_back(&*BI);
2722 
2723  Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
2724  assert(InsertionPoint && "Empty block?");
2725  // Substitute with Phis & remove.
2726  for (auto *Inst : reverse(ToRemove)) {
2727  if (!Inst->use_empty()) {
2728  PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
2729  NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
2730  NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
2731  NewPN->insertBefore(InsertionPoint);
2732  Inst->replaceAllUsesWith(NewPN);
2733  }
2734  Inst->eraseFromParent();
2735  }
2736  return true;
2737 }
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:171
A parsed version of the target data layout string in and methods for querying it. ...
Definition: DataLayout.h:110
reference emplace_back(ArgTypes &&... Args)
Definition: SmallVector.h:645
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:594
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:70
This class is the base class for the comparison instructions.
Definition: InstrTypes.h:636
static bool runImpl(Function &F, TargetLibraryInfo &TLI, DominatorTree &DT)
This is the entry point for all transforms.
iterator_range< use_iterator > uses()
Definition: Value.h:354
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
static IntegerType * getInt1Ty(LLVMContext &C)
Definition: Type.cpp:172
Helper class for SSA formation on a set of values defined in multiple blocks.
Definition: SSAUpdater.h:38
void DeleteDeadBlock(BasicBlock *BB, DomTreeUpdater *DTU=nullptr, bool KeepOneInputPHIs=false)
Delete the specified block, which must have no predecessors.
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
BasicBlock * DuplicateInstructionsInSplitBetween(BasicBlock *BB, BasicBlock *PredBB, Instruction *StopAt, ValueToValueMapTy &ValueMapping, DomTreeUpdater &DTU)
Split edge between BB and PredBB and duplicate all non-Phi instructions from BB between its beginning...
NodeTy * getNextNode()
Get the next node, or nullptr for the list tail.
Definition: ilist_node.h:288
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:769
This class represents lattice values for constants.
Definition: AllocatorList.h:23
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:53
bool ConstantFoldTerminator(BasicBlock *BB, bool DeleteDeadConditions=false, const TargetLibraryInfo *TLI=nullptr, DomTreeUpdater *DTU=nullptr)
If a terminator instruction is predicated on a constant value, convert it into an unconditional branc...
Definition: Local.cpp:105
BasicBlock * getSuccessor(unsigned Idx) const
Return the specified successor. This instruction must be a terminator.
void initializeJumpThreadingPass(PassRegistry &)
Implements a dense probed hash-table based set.
Definition: DenseSet.h:249
void removePredecessor(BasicBlock *Pred, bool KeepOneInputPHIs=false)
Notify the BasicBlock that the predecessor Pred is no longer able to reach it.
Definition: BasicBlock.cpp:301
void AddAvailableValue(BasicBlock *BB, Value *V)
Indicate that a rewritten value is available in the specified block with the specified value...
Definition: SSAUpdater.cpp:71
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.
Value * getCondition() const
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:247
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:709
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:98
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:1185
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:343
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:592
Metadata node.
Definition: Metadata.h:863
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:230
F(f)
An instruction for reading from memory.
Definition: Instructions.h:167
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:1965
Value * getCondition() const
const Instruction * 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:137
void setSuccessor(unsigned Idx, BasicBlock *BB)
Update the specified successor to point at the provided block.
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:229
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:701
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:32
FunctionPass * createJumpThreadingPass(int Threshold=-1)
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:268
BasicBlock * SplitEdge(BasicBlock *From, BasicBlock *To, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr, MemorySSAUpdater *MSSAU=nullptr)
Split the edge connecting specified block.
Value * getArgOperand(unsigned i) const
Definition: InstrTypes.h:1155
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:221
bool runImpl(Function &F, TargetLibraryInfo *TLI_, LazyValueInfo *LVI_, AliasAnalysis *AA_, DomTreeUpdater *DTU_, bool HasProfileData_, std::unique_ptr< BlockFrequencyInfo > BFI_, std::unique_ptr< BranchProbabilityInfo > BPI_)
The address of a basic block.
Definition: Constants.h:839
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:47
AnalysisUsage & addRequired()
const Module * getModule() const
Return the module owning the function this basic block belongs to, or nullptr if the function does no...
Definition: BasicBlock.cpp:133
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:50
void MergeBasicBlockIntoOnlyPred(BasicBlock *BB, DomTreeUpdater *DTU=nullptr)
BB is a block with one predecessor and its predecessor is known to have one successor (BB!)...
Definition: Local.cpp:675
StringRef getName(ID id)
Return the LLVM name for an intrinsic, such as "llvm.ppc.altivec.lvx".
Definition: Function.cpp:625
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:369
This is the base class for all instructions that perform data casts.
Definition: InstrTypes.h:353
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:1285
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:328
unsigned getNumSuccessors() const
A Use represents the edge between a Value definition and its users.
Definition: Use.h:55
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:41
jump threading
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:196
void setName(const Twine &Name)
Change the name of the value.
Definition: Value.cpp:285
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:641
void assign(size_type NumElts, const T &Elt)
Definition: SmallVector.h:416
auto reverse(ContainerTy &&C, typename std::enable_if< has_rbegin< ContainerTy >::value >::type *=nullptr) -> decltype(make_range(C.rbegin(), C.rend()))
Definition: STLExtras.h:266
bool isOne() const
This is just a convenience method to make client code smaller for a common case.
Definition: Constants.h:200
bool ComputeValueKnownInPredecessorsImpl(Value *V, BasicBlock *BB, jumpthreading::PredValueInfo &Result, jumpthreading::ConstantPreference Preference, DenseSet< std::pair< Value *, BasicBlock *>> &RecursionSet, Instruction *CxtI=nullptr)
ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see if we can infer that the ...
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:244
static void normalizeProbabilities(ProbabilityIter Begin, ProbabilityIter End)
bool TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB, DomTreeUpdater *DTU=nullptr)
BB is known to contain an unconditional branch, and contains no instructions other than PHI nodes...
Definition: Local.cpp:960
MDNode * getMetadata(unsigned KindID) const
Get the metadata of given kind attached to this Instruction.
Definition: Instruction.h:234
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:32
class_match< ConstantInt > m_ConstantInt()
Match an arbitrary ConstantInt and ignore it.
Definition: PatternMatch.h:81
const APInt & getValue() const
Return the constant as an APInt value reference.
Definition: Constants.h:137
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:125
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:1281
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:429
static bool ProcessBlock(BasicBlock &BB, DominatorTree &DT, LoopInfo &LI, AAResults &AA)
Definition: Sink.cpp:198
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:291
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:144
unsigned getNumSuccessors() const
Return the number of successors that this instruction has.
Value * getOperand(unsigned i) const
Definition: User.h:169
void removeDeadConstantUsers() const
If there are any dead constant users dangling off of this constant, remove them.
Definition: Constants.cpp:546
void SplitLandingPadPredecessors(BasicBlock *OrigBB, ArrayRef< BasicBlock *> Preds, const char *Suffix, const char *Suffix2, SmallVectorImpl< BasicBlock *> &NewBBs, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr, MemorySSAUpdater *MSSAU=nullptr, bool PreserveLCSSA=false)
This method transforms the landing pad, OrigBB, by introducing two new basic blocks into the function...
StringRef getString() const
Definition: Metadata.cpp:463
const BasicBlock & getEntryBlock() const
Definition: Function.h:639
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)
BasicBlock * SplitBlockPredecessors(BasicBlock *BB, ArrayRef< BasicBlock *> Preds, const char *Suffix, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr, MemorySSAUpdater *MSSAU=nullptr, bool PreserveLCSSA=false)
This method introduces at least one new basic block into the function and moves some of the predecess...
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:423
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:1082
const Instruction * getFirstNonPHI() const
Returns a pointer to the first instruction in this block that is not a PHINode instruction.
Definition: BasicBlock.cpp:189
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:216
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:321
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:233
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:57
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:45
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:1443
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
This is an important base class in LLVM.
Definition: Constant.h:41
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:280
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:370
bool mayHaveSideEffects() const
Return true if the instruction may have side effects.
Definition: Instruction.h:575
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:112
bool isUnordered() const
Definition: Instructions.h:278
Represent the analysis usage information of a pass.
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:646
jump Jump Threading
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:284
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:115
static BasicBlock * Create(LLVMContext &Context, const Twine &Name="", Function *Parent=nullptr, BasicBlock *InsertBefore=nullptr)
Creates a new BasicBlock.
Definition: BasicBlock.h:99
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
Definition: SmallPtrSet.h:381
bool pred_empty(const BasicBlock *BB)
Definition: CFG.h:116
static Constant * getNot(Constant *C)
Definition: Constants.cpp:2241
void FindLoopHeaders(Function &F)
FindLoopHeaders - We do not want jump threading to turn proper loop structures into irreducible loops...
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:1424
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs, and aliases.
Definition: Value.cpp:529
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:159
bool isExceptionalTerminator() const
Definition: Instruction.h:135
size_t size() const
Definition: SmallVector.h:52
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:1225
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:62
bool ProcessBranchOnPHI(PHINode *PN)
ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on a PHI node in the curren...
size_type size() const
Definition: SmallPtrSet.h:92
bool isLandingPad() const
Return true if this basic block is a landing pad.
Definition: BasicBlock.cpp:467
bool hasAddressTaken() const
Returns true if there are any uses of this basic block other than direct branches, switches, etc.
Definition: BasicBlock.h:391
const InstListType & getInstList() const
Return the underlying instruction list container.
Definition: BasicBlock.h:333
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:191
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:127
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:85
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements...
Definition: SmallPtrSet.h:417
This is the shared class of boolean and integer constants.
Definition: Constants.h:83
iterator end()
Definition: BasicBlock.h:270
This is a &#39;vector&#39; (really, a variable-sized array), optimized for the case when the array is small...
Definition: SmallVector.h:841
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)
A collection of metadata nodes that might be associated with a memory access used by the alias-analys...
Definition: Metadata.h:643
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:374
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:631
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:124
unsigned getNumIncomingValues() const
Return the number of incoming edges.
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:587
void erase_if(Container &C, UnaryPredicate P)
Provide a container algorithm similar to C++ Library Fundamentals v2&#39;s erase_if which is equivalent t...
Definition: STLExtras.h:1329
bool isGuard(const User *U)
Returns true iff U has semantics of a guard expressed in a form of call of llvm.experimental.guard intrinsic.
Definition: GuardUtils.cpp:17
void setOperand(unsigned i, Value *Val)
Definition: User.h:174
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:174
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:311
static cl::opt< unsigned > Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), cl::init(100), cl::Hidden)
static Constant * getCast(unsigned ops, Constant *C, Type *Ty, bool OnlyIfReduced=false)
Convenience function for getting a Cast operation.
Definition: Constants.cpp:1539
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:721
This file provides various utilities for inspecting and working with the control flow graph in LLVM I...
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:545
Analysis providing branch probability information.
iterator insert(iterator where, pointer New)
Definition: ilist.h:226
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:324
unsigned getAlignment() const
Return the alignment of the access that is being performed.
Definition: Instructions.h:240
LLVM_NODISCARD bool empty() const
Definition: SmallVector.h:55
void UnfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB, SelectInst *SI, PHINode *SIUse, unsigned Idx)
iterator begin()
Definition: DenseMap.h:99
bool ProcessBranchOnXOR(BinaryOperator *BO)
ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on a xor instruction in the...
StringRef getName() const
Return a constant reference to the value&#39;s name.
Definition: Value.cpp:214
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:106
SyncScope::ID getSyncScopeID() const
Returns the synchronization scope ID of this load instruction.
Definition: Instructions.h:259
bool ProcessImpliedCondition(BasicBlock *BB)
unsigned pred_size(const BasicBlock *BB)
Get the number of predecessors of BB.
Definition: CFG.h:121
#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:192
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:322
static void ReplaceFoldableUses(Instruction *Cond, Value *ToVal)
uint32_t Size
Definition: Profile.cpp:46
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:324
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:211
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:26
Analysis pass providing the TargetLibraryInfo.
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:31
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
See the comments on JumpThreadingPass.
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:398
unsigned getPrimitiveSizeInBits() const LLVM_READONLY
Return the basic size of this type if it is a primitive type.
Definition: Type.cpp:114
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:349
LLVM Value Representation.
Definition: Value.h:72
succ_range successors(Instruction *I)
Definition: CFG.h:259
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)
void combineMetadataForCSE(Instruction *K, const Instruction *J, bool DoesKMove)
Combine the metadata of two instructions so that K can replace J.
Definition: Local.cpp:2355
bool hasOneUse() const
Return true if there is exactly one user of this value.
Definition: Value.h:412
A single uniqued string.
Definition: Metadata.h:603
A container for analyses that lazily runs them and caches their results.
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:259
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object...
This header defines various interfaces for pass management in LLVM.
void setIncomingValue(unsigned i, Value *V)
INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading", "Jump Threading", false, false) INITIALIZE_PASS_END(JumpThreading
#define LLVM_DEBUG(X)
Definition: Debug.h:122
bool extractProfMetadata(uint64_t &TrueVal, uint64_t &FalseVal) const
Retrieve the raw weight values of a conditional branch or select.
Definition: Metadata.cpp:1314
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:189
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:307
uint32_t getNumerator() const
bool use_empty() const
Definition: Value.h:322
Instruction * SplitBlockAndInsertIfThen(Value *Cond, Instruction *SplitBefore, bool Unreachable, MDNode *BranchWeights=nullptr, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr, BasicBlock *ThenBlock=nullptr)
Split the containing block at the specified instruction - everything before SplitBefore stays in the ...
unsigned replaceNonLocalUsesWith(Instruction *From, Value *To)
Definition: Local.cpp:2420
Analysis to compute lazy value information.
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:43
const BasicBlock * getParent() const
Definition: Instruction.h:66
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:344
static Constant * get(unsigned Opcode, Constant *C1, unsigned Flags=0, Type *OnlyIfReducedTy=nullptr)
get - Return a unary operator constant expression, folding if possible.
Definition: Constants.cpp:1815
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:1244