LLVM  14.0.0git
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/MapVector.h"
17 #include "llvm/ADT/Optional.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Analysis/CFG.h"
32 #include "llvm/Analysis/Loads.h"
33 #include "llvm/Analysis/LoopInfo.h"
38 #include "llvm/IR/BasicBlock.h"
39 #include "llvm/IR/CFG.h"
40 #include "llvm/IR/Constant.h"
41 #include "llvm/IR/ConstantRange.h"
42 #include "llvm/IR/Constants.h"
43 #include "llvm/IR/DataLayout.h"
44 #include "llvm/IR/Dominators.h"
45 #include "llvm/IR/Function.h"
46 #include "llvm/IR/InstrTypes.h"
47 #include "llvm/IR/Instruction.h"
48 #include "llvm/IR/Instructions.h"
49 #include "llvm/IR/IntrinsicInst.h"
50 #include "llvm/IR/Intrinsics.h"
51 #include "llvm/IR/LLVMContext.h"
52 #include "llvm/IR/MDBuilder.h"
53 #include "llvm/IR/Metadata.h"
54 #include "llvm/IR/Module.h"
55 #include "llvm/IR/PassManager.h"
56 #include "llvm/IR/PatternMatch.h"
57 #include "llvm/IR/Type.h"
58 #include "llvm/IR/Use.h"
59 #include "llvm/IR/User.h"
60 #include "llvm/IR/Value.h"
61 #include "llvm/InitializePasses.h"
62 #include "llvm/Pass.h"
65 #include "llvm/Support/Casting.h"
67 #include "llvm/Support/Debug.h"
69 #include "llvm/Transforms/Scalar.h"
75 #include <algorithm>
76 #include <cassert>
77 #include <cstddef>
78 #include <cstdint>
79 #include <iterator>
80 #include <memory>
81 #include <utility>
82 
83 using namespace llvm;
84 using namespace jumpthreading;
85 
86 #define DEBUG_TYPE "jump-threading"
87 
88 STATISTIC(NumThreads, "Number of jumps threaded");
89 STATISTIC(NumFolds, "Number of terminators folded");
90 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
91 
92 static cl::opt<unsigned>
93 BBDuplicateThreshold("jump-threading-threshold",
94  cl::desc("Max block size to duplicate for jump threading"),
95  cl::init(6), cl::Hidden);
96 
97 static cl::opt<unsigned>
99  "jump-threading-implication-search-threshold",
100  cl::desc("The number of predecessors to search for a stronger "
101  "condition to use to thread over a weaker condition"),
102  cl::init(3), cl::Hidden);
103 
105  "print-lvi-after-jump-threading",
106  cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false),
107  cl::Hidden);
108 
110  "jump-threading-freeze-select-cond",
111  cl::desc("Freeze the condition when unfolding select"), cl::init(false),
112  cl::Hidden);
113 
115  "jump-threading-across-loop-headers",
116  cl::desc("Allow JumpThreading to thread across loop headers, for testing"),
117  cl::init(false), cl::Hidden);
118 
119 
120 namespace {
121 
122  /// This pass performs 'jump threading', which looks at blocks that have
123  /// multiple predecessors and multiple successors. If one or more of the
124  /// predecessors of the block can be proven to always jump to one of the
125  /// successors, we forward the edge from the predecessor to the successor by
126  /// duplicating the contents of this block.
127  ///
128  /// An example of when this can occur is code like this:
129  ///
130  /// if () { ...
131  /// X = 4;
132  /// }
133  /// if (X < 3) {
134  ///
135  /// In this case, the unconditional branch at the end of the first if can be
136  /// revectored to the false side of the second if.
137  class JumpThreading : public FunctionPass {
138  JumpThreadingPass Impl;
139 
140  public:
141  static char ID; // Pass identification
142 
143  JumpThreading(bool InsertFreezeWhenUnfoldingSelect = false, int T = -1)
144  : FunctionPass(ID), Impl(InsertFreezeWhenUnfoldingSelect, T) {
146  }
147 
148  bool runOnFunction(Function &F) override;
149 
150  void getAnalysisUsage(AnalysisUsage &AU) const override {
159  }
160 
161  void releaseMemory() override { Impl.releaseMemory(); }
162  };
163 
164 } // end anonymous namespace
165 
166 char JumpThreading::ID = 0;
167 
168 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
169  "Jump Threading", false, false)
176 
177 // Public interface to the Jump Threading pass
179  return new JumpThreading(InsertFr, Threshold);
180 }
181 
183  InsertFreezeWhenUnfoldingSelect = JumpThreadingFreezeSelectCond | InsertFr;
184  DefaultBBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
185 }
186 
187 // Update branch probability information according to conditional
188 // branch probability. This is usually made possible for cloned branches
189 // in inline instances by the context specific profile in the caller.
190 // For instance,
191 //
192 // [Block PredBB]
193 // [Branch PredBr]
194 // if (t) {
195 // Block A;
196 // } else {
197 // Block B;
198 // }
199 //
200 // [Block BB]
201 // cond = PN([true, %A], [..., %B]); // PHI node
202 // [Branch CondBr]
203 // if (cond) {
204 // ... // P(cond == true) = 1%
205 // }
206 //
207 // Here we know that when block A is taken, cond must be true, which means
208 // P(cond == true | A) = 1
209 //
210 // Given that P(cond == true) = P(cond == true | A) * P(A) +
211 // P(cond == true | B) * P(B)
212 // we get:
213 // P(cond == true ) = P(A) + P(cond == true | B) * P(B)
214 //
215 // which gives us:
216 // P(A) is less than P(cond == true), i.e.
217 // P(t == true) <= P(cond == true)
218 //
219 // In other words, if we know P(cond == true) is unlikely, we know
220 // that P(t == true) is also unlikely.
221 //
223  BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
224  if (!CondBr)
225  return;
226 
227  uint64_t TrueWeight, FalseWeight;
228  if (!CondBr->extractProfMetadata(TrueWeight, FalseWeight))
229  return;
230 
231  if (TrueWeight + FalseWeight == 0)
232  // Zero branch_weights do not give a hint for getting branch probabilities.
233  // Technically it would result in division by zero denominator, which is
234  // TrueWeight + FalseWeight.
235  return;
236 
237  // Returns the outgoing edge of the dominating predecessor block
238  // that leads to the PhiNode's incoming block:
239  auto GetPredOutEdge =
240  [](BasicBlock *IncomingBB,
241  BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> {
242  auto *PredBB = IncomingBB;
243  auto *SuccBB = PhiBB;
245  while (true) {
246  BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
247  if (PredBr && PredBr->isConditional())
248  return {PredBB, SuccBB};
249  Visited.insert(PredBB);
250  auto *SinglePredBB = PredBB->getSinglePredecessor();
251  if (!SinglePredBB)
252  return {nullptr, nullptr};
253 
254  // Stop searching when SinglePredBB has been visited. It means we see
255  // an unreachable loop.
256  if (Visited.count(SinglePredBB))
257  return {nullptr, nullptr};
258 
259  SuccBB = PredBB;
260  PredBB = SinglePredBB;
261  }
262  };
263 
264  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
265  Value *PhiOpnd = PN->getIncomingValue(i);
266  ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd);
267 
268  if (!CI || !CI->getType()->isIntegerTy(1))
269  continue;
270 
271  BranchProbability BP =
273  TrueWeight, TrueWeight + FalseWeight)
275  FalseWeight, TrueWeight + FalseWeight));
276 
277  auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB);
278  if (!PredOutEdge.first)
279  return;
280 
281  BasicBlock *PredBB = PredOutEdge.first;
282  BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
283  if (!PredBr)
284  return;
285 
286  uint64_t PredTrueWeight, PredFalseWeight;
287  // FIXME: We currently only set the profile data when it is missing.
288  // With PGO, this can be used to refine even existing profile data with
289  // context information. This needs to be done after more performance
290  // testing.
291  if (PredBr->extractProfMetadata(PredTrueWeight, PredFalseWeight))
292  continue;
293 
294  // We can not infer anything useful when BP >= 50%, because BP is the
295  // upper bound probability value.
296  if (BP >= BranchProbability(50, 100))
297  continue;
298 
299  SmallVector<uint32_t, 2> Weights;
300  if (PredBr->getSuccessor(0) == PredOutEdge.second) {
301  Weights.push_back(BP.getNumerator());
302  Weights.push_back(BP.getCompl().getNumerator());
303  } else {
304  Weights.push_back(BP.getCompl().getNumerator());
305  Weights.push_back(BP.getNumerator());
306  }
307  PredBr->setMetadata(LLVMContext::MD_prof,
308  MDBuilder(PredBr->getParent()->getContext())
309  .createBranchWeights(Weights));
310  }
311 }
312 
313 /// runOnFunction - Toplevel algorithm.
315  if (skipFunction(F))
316  return false;
317  auto TTI = &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
318  // Jump Threading has no sense for the targets with divergent CF
319  if (TTI->hasBranchDivergence())
320  return false;
321  auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
322  auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
323  auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
324  auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
326  std::unique_ptr<BlockFrequencyInfo> BFI;
327  std::unique_ptr<BranchProbabilityInfo> BPI;
328  if (F.hasProfileData()) {
329  LoopInfo LI{DominatorTree(F)};
330  BPI.reset(new BranchProbabilityInfo(F, LI, TLI));
331  BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
332  }
333 
334  bool Changed = Impl.runImpl(F, TLI, TTI, LVI, AA, &DTU, F.hasProfileData(),
335  std::move(BFI), std::move(BPI));
337  dbgs() << "LVI for function '" << F.getName() << "':\n";
338  LVI->printLVI(F, DTU.getDomTree(), dbgs());
339  }
340  return Changed;
341 }
342 
345  auto &TTI = AM.getResult<TargetIRAnalysis>(F);
346  // Jump Threading has no sense for the targets with divergent CF
347  if (TTI.hasBranchDivergence())
348  return PreservedAnalyses::all();
349  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
350  auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
351  auto &LVI = AM.getResult<LazyValueAnalysis>(F);
352  auto &AA = AM.getResult<AAManager>(F);
354 
355  std::unique_ptr<BlockFrequencyInfo> BFI;
356  std::unique_ptr<BranchProbabilityInfo> BPI;
357  if (F.hasProfileData()) {
358  LoopInfo LI{DominatorTree(F)};
359  BPI.reset(new BranchProbabilityInfo(F, LI, &TLI));
360  BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
361  }
362 
363  bool Changed = runImpl(F, &TLI, &TTI, &LVI, &AA, &DTU, F.hasProfileData(),
364  std::move(BFI), std::move(BPI));
365 
367  dbgs() << "LVI for function '" << F.getName() << "':\n";
368  LVI.printLVI(F, DTU.getDomTree(), dbgs());
369  }
370 
371  if (!Changed)
372  return PreservedAnalyses::all();
376  return PA;
377 }
378 
380  TargetTransformInfo *TTI_, LazyValueInfo *LVI_,
381  AliasAnalysis *AA_, DomTreeUpdater *DTU_,
382  bool HasProfileData_,
383  std::unique_ptr<BlockFrequencyInfo> BFI_,
384  std::unique_ptr<BranchProbabilityInfo> BPI_) {
385  LLVM_DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
386  TLI = TLI_;
387  TTI = TTI_;
388  LVI = LVI_;
389  AA = AA_;
390  DTU = DTU_;
391  BFI.reset();
392  BPI.reset();
393  // When profile data is available, we need to update edge weights after
394  // successful jump threading, which requires both BPI and BFI being available.
395  HasProfileData = HasProfileData_;
396  auto *GuardDecl = F.getParent()->getFunction(
397  Intrinsic::getName(Intrinsic::experimental_guard));
398  HasGuards = GuardDecl && !GuardDecl->use_empty();
399  if (HasProfileData) {
400  BPI = std::move(BPI_);
401  BFI = std::move(BFI_);
402  }
403 
404  // Reduce the number of instructions duplicated when optimizing strictly for
405  // size.
406  if (BBDuplicateThreshold.getNumOccurrences())
407  BBDupThreshold = BBDuplicateThreshold;
408  else if (F.hasFnAttribute(Attribute::MinSize))
409  BBDupThreshold = 3;
410  else
411  BBDupThreshold = DefaultBBDupThreshold;
412 
413  // JumpThreading must not processes blocks unreachable from entry. It's a
414  // waste of compute time and can potentially lead to hangs.
415  SmallPtrSet<BasicBlock *, 16> Unreachable;
416  assert(DTU && "DTU isn't passed into JumpThreading before using it.");
417  assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed.");
418  DominatorTree &DT = DTU->getDomTree();
419  for (auto &BB : F)
420  if (!DT.isReachableFromEntry(&BB))
421  Unreachable.insert(&BB);
422 
424  findLoopHeaders(F);
425 
426  bool EverChanged = false;
427  bool Changed;
428  do {
429  Changed = false;
430  for (auto &BB : F) {
431  if (Unreachable.count(&BB))
432  continue;
433  while (processBlock(&BB)) // Thread all of the branches we can over BB.
434  Changed = true;
435 
436  // Jump threading may have introduced redundant debug values into BB
437  // which should be removed.
438  if (Changed)
440 
441  // Stop processing BB if it's the entry or is now deleted. The following
442  // routines attempt to eliminate BB and locating a suitable replacement
443  // for the entry is non-trivial.
444  if (&BB == &F.getEntryBlock() || DTU->isBBPendingDeletion(&BB))
445  continue;
446 
447  if (pred_empty(&BB)) {
448  // When processBlock makes BB unreachable it doesn't bother to fix up
449  // the instructions in it. We must remove BB to prevent invalid IR.
450  LLVM_DEBUG(dbgs() << " JT: Deleting dead block '" << BB.getName()
451  << "' with terminator: " << *BB.getTerminator()
452  << '\n');
453  LoopHeaders.erase(&BB);
454  LVI->eraseBlock(&BB);
455  DeleteDeadBlock(&BB, DTU);
456  Changed = true;
457  continue;
458  }
459 
460  // processBlock doesn't thread BBs with unconditional TIs. However, if BB
461  // is "almost empty", we attempt to merge BB with its sole successor.
462  auto *BI = dyn_cast<BranchInst>(BB.getTerminator());
463  if (BI && BI->isUnconditional()) {
464  BasicBlock *Succ = BI->getSuccessor(0);
465  if (
466  // The terminator must be the only non-phi instruction in BB.
467  BB.getFirstNonPHIOrDbg(true)->isTerminator() &&
468  // Don't alter Loop headers and latches to ensure another pass can
469  // detect and transform nested loops later.
470  !LoopHeaders.count(&BB) && !LoopHeaders.count(Succ) &&
473  // BB is valid for cleanup here because we passed in DTU. F remains
474  // BB's parent until a DTU->getDomTree() event.
475  LVI->eraseBlock(&BB);
476  Changed = true;
477  }
478  }
479  }
480  EverChanged |= Changed;
481  } while (Changed);
482 
483  LoopHeaders.clear();
484  return EverChanged;
485 }
486 
487 // Replace uses of Cond with ToVal when safe to do so. If all uses are
488 // replaced, we can remove Cond. We cannot blindly replace all uses of Cond
489 // because we may incorrectly replace uses when guards/assumes are uses of
490 // of `Cond` and we used the guards/assume to reason about the `Cond` value
491 // at the end of block. RAUW unconditionally replaces all uses
492 // including the guards/assumes themselves and the uses before the
493 // guard/assume.
494 static void replaceFoldableUses(Instruction *Cond, Value *ToVal) {
495  assert(Cond->getType() == ToVal->getType());
496  auto *BB = Cond->getParent();
497  // We can unconditionally replace all uses in non-local blocks (i.e. uses
498  // strictly dominated by BB), since LVI information is true from the
499  // terminator of BB.
501  for (Instruction &I : reverse(*BB)) {
502  // Reached the Cond whose uses we are trying to replace, so there are no
503  // more uses.
504  if (&I == Cond)
505  break;
506  // We only replace uses in instructions that are guaranteed to reach the end
507  // of BB, where we know Cond is ToVal.
509  break;
510  I.replaceUsesOfWith(Cond, ToVal);
511  }
512  if (Cond->use_empty() && !Cond->mayHaveSideEffects())
513  Cond->eraseFromParent();
514 }
515 
516 /// Return the cost of duplicating a piece of this block from first non-phi
517 /// and before StopAt instruction to thread across it. Stop scanning the block
518 /// when exceeding the threshold. If duplication is impossible, returns ~0U.
520  BasicBlock *BB,
521  Instruction *StopAt,
522  unsigned Threshold) {
523  assert(StopAt->getParent() == BB && "Not an instruction from proper BB?");
524  /// Ignore PHI nodes, these will be flattened when duplication happens.
525  BasicBlock::const_iterator I(BB->getFirstNonPHI());
526 
527  // FIXME: THREADING will delete values that are just used to compute the
528  // branch, so they shouldn't count against the duplication cost.
529 
530  unsigned Bonus = 0;
531  if (BB->getTerminator() == StopAt) {
532  // Threading through a switch statement is particularly profitable. If this
533  // block ends in a switch, decrease its cost to make it more likely to
534  // happen.
535  if (isa<SwitchInst>(StopAt))
536  Bonus = 6;
537 
538  // The same holds for indirect branches, but slightly more so.
539  if (isa<IndirectBrInst>(StopAt))
540  Bonus = 8;
541  }
542 
543  // Bump the threshold up so the early exit from the loop doesn't skip the
544  // terminator-based Size adjustment at the end.
545  Threshold += Bonus;
546 
547  // Sum up the cost of each instruction until we get to the terminator. Don't
548  // include the terminator because the copy won't include it.
549  unsigned Size = 0;
550  for (; &*I != StopAt; ++I) {
551 
552  // Stop scanning the block if we've reached the threshold.
553  if (Size > Threshold)
554  return Size;
555 
556  // Bail out if this instruction gives back a token type, it is not possible
557  // to duplicate it if it is used outside this BB.
558  if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
559  return ~0U;
560 
561  // Blocks with NoDuplicate are modelled as having infinite cost, so they
562  // are never duplicated.
563  if (const CallInst *CI = dyn_cast<CallInst>(I))
564  if (CI->cannotDuplicate() || CI->isConvergent())
565  return ~0U;
566 
569  continue;
570 
571  // All other instructions count for at least one unit.
572  ++Size;
573 
574  // Calls are more expensive. If they are non-intrinsic calls, we model them
575  // as having cost of 4. If they are a non-vector intrinsic, we model them
576  // as having cost of 2 total, and if they are a vector intrinsic, we model
577  // them as having cost 1.
578  if (const CallInst *CI = dyn_cast<CallInst>(I)) {
579  if (!isa<IntrinsicInst>(CI))
580  Size += 3;
581  else if (!CI->getType()->isVectorTy())
582  Size += 1;
583  }
584  }
585 
586  return Size > Bonus ? Size - Bonus : 0;
587 }
588 
589 /// findLoopHeaders - We do not want jump threading to turn proper loop
590 /// structures into irreducible loops. Doing this breaks up the loop nesting
591 /// hierarchy and pessimizes later transformations. To prevent this from
592 /// happening, we first have to find the loop headers. Here we approximate this
593 /// by finding targets of backedges in the CFG.
594 ///
595 /// Note that there definitely are cases when we want to allow threading of
596 /// edges across a loop header. For example, threading a jump from outside the
597 /// loop (the preheader) to an exit block of the loop is definitely profitable.
598 /// It is also almost always profitable to thread backedges from within the loop
599 /// to exit blocks, and is often profitable to thread backedges to other blocks
600 /// within the loop (forming a nested loop). This simple analysis is not rich
601 /// enough to track all of these properties and keep it up-to-date as the CFG
602 /// mutates, so we don't allow any of these transformations.
605  FindFunctionBackedges(F, Edges);
606 
607  for (const auto &Edge : Edges)
608  LoopHeaders.insert(Edge.second);
609 }
610 
611 /// getKnownConstant - Helper method to determine if we can thread over a
612 /// terminator with the given value as its condition, and if so what value to
613 /// use for that. What kind of value this is depends on whether we want an
614 /// integer or a block address, but an undef is always accepted.
615 /// Returns null if Val is null or not an appropriate constant.
617  if (!Val)
618  return nullptr;
619 
620  // Undef is "known" enough.
621  if (UndefValue *U = dyn_cast<UndefValue>(Val))
622  return U;
623 
625  return dyn_cast<BlockAddress>(Val->stripPointerCasts());
626 
627  return dyn_cast<ConstantInt>(Val);
628 }
629 
630 /// computeValueKnownInPredecessors - Given a basic block BB and a value V, see
631 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
632 /// in any of our predecessors. If so, return the known list of value and pred
633 /// BB in the result vector.
634 ///
635 /// This returns true if there were any known values.
637  Value *V, BasicBlock *BB, PredValueInfo &Result,
639  Instruction *CxtI) {
640  // This method walks up use-def chains recursively. Because of this, we could
641  // get into an infinite loop going around loops in the use-def chain. To
642  // prevent this, keep track of what (value, block) pairs we've already visited
643  // and terminate the search if we loop back to them
644  if (!RecursionSet.insert(V).second)
645  return false;
646 
647  // If V is a constant, then it is known in all predecessors.
648  if (Constant *KC = getKnownConstant(V, Preference)) {
649  for (BasicBlock *Pred : predecessors(BB))
650  Result.emplace_back(KC, Pred);
651 
652  return !Result.empty();
653  }
654 
655  // If V is a non-instruction value, or an instruction in a different block,
656  // then it can't be derived from a PHI.
657  Instruction *I = dyn_cast<Instruction>(V);
658  if (!I || I->getParent() != BB) {
659 
660  // Okay, if this is a live-in value, see if it has a known value at the end
661  // of any of our predecessors.
662  //
663  // FIXME: This should be an edge property, not a block end property.
664  /// TODO: Per PR2563, we could infer value range information about a
665  /// predecessor based on its terminator.
666  //
667  // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
668  // "I" is a non-local compare-with-a-constant instruction. This would be
669  // able to handle value inequalities better, for example if the compare is
670  // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
671  // Perhaps getConstantOnEdge should be smart enough to do this?
672  for (BasicBlock *P : predecessors(BB)) {
673  // If the value is known by LazyValueInfo to be a constant in a
674  // predecessor, use that information to try to thread this block.
675  Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
676  if (Constant *KC = getKnownConstant(PredCst, Preference))
677  Result.emplace_back(KC, P);
678  }
679 
680  return !Result.empty();
681  }
682 
683  /// If I is a PHI node, then we know the incoming values for any constants.
684  if (PHINode *PN = dyn_cast<PHINode>(I)) {
685  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
686  Value *InVal = PN->getIncomingValue(i);
687  if (Constant *KC = getKnownConstant(InVal, Preference)) {
688  Result.emplace_back(KC, PN->getIncomingBlock(i));
689  } else {
690  Constant *CI = LVI->getConstantOnEdge(InVal,
691  PN->getIncomingBlock(i),
692  BB, CxtI);
693  if (Constant *KC = getKnownConstant(CI, Preference))
694  Result.emplace_back(KC, PN->getIncomingBlock(i));
695  }
696  }
697 
698  return !Result.empty();
699  }
700 
701  // Handle Cast instructions.
702  if (CastInst *CI = dyn_cast<CastInst>(I)) {
703  Value *Source = CI->getOperand(0);
704  computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
705  RecursionSet, CxtI);
706  if (Result.empty())
707  return false;
708 
709  // Convert the known values.
710  for (auto &R : Result)
711  R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
712 
713  return true;
714  }
715 
716  if (FreezeInst *FI = dyn_cast<FreezeInst>(I)) {
717  Value *Source = FI->getOperand(0);
718  computeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
719  RecursionSet, CxtI);
720 
721  erase_if(Result, [](auto &Pair) {
722  return !isGuaranteedNotToBeUndefOrPoison(Pair.first);
723  });
724 
725  return !Result.empty();
726  }
727 
728  // Handle some boolean conditions.
729  if (I->getType()->getPrimitiveSizeInBits() == 1) {
730  using namespace PatternMatch;
731 
732  assert(Preference == WantInteger && "One-bit non-integer type?");
733  // X | true -> true
734  // X & false -> false
735  Value *Op0, *Op1;
736  if (match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1))) ||
737  match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1)))) {
738  PredValueInfoTy LHSVals, RHSVals;
739 
740  computeValueKnownInPredecessorsImpl(Op0, BB, LHSVals, WantInteger,
741  RecursionSet, CxtI);
742  computeValueKnownInPredecessorsImpl(Op1, BB, RHSVals, WantInteger,
743  RecursionSet, CxtI);
744 
745  if (LHSVals.empty() && RHSVals.empty())
746  return false;
747 
748  ConstantInt *InterestingVal;
749  if (match(I, m_LogicalOr()))
750  InterestingVal = ConstantInt::getTrue(I->getContext());
751  else
752  InterestingVal = ConstantInt::getFalse(I->getContext());
753 
754  SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
755 
756  // Scan for the sentinel. If we find an undef, force it to the
757  // interesting value: x|undef -> true and x&undef -> false.
758  for (const auto &LHSVal : LHSVals)
759  if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
760  Result.emplace_back(InterestingVal, LHSVal.second);
761  LHSKnownBBs.insert(LHSVal.second);
762  }
763  for (const auto &RHSVal : RHSVals)
764  if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
765  // If we already inferred a value for this block on the LHS, don't
766  // re-add it.
767  if (!LHSKnownBBs.count(RHSVal.second))
768  Result.emplace_back(InterestingVal, RHSVal.second);
769  }
770 
771  return !Result.empty();
772  }
773 
774  // Handle the NOT form of XOR.
775  if (I->getOpcode() == Instruction::Xor &&
776  isa<ConstantInt>(I->getOperand(1)) &&
777  cast<ConstantInt>(I->getOperand(1))->isOne()) {
778  computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, Result,
779  WantInteger, RecursionSet, CxtI);
780  if (Result.empty())
781  return false;
782 
783  // Invert the known values.
784  for (auto &R : Result)
785  R.first = ConstantExpr::getNot(R.first);
786 
787  return true;
788  }
789 
790  // Try to simplify some other binary operator values.
791  } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
793  && "A binary operator creating a block address?");
794  if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
795  PredValueInfoTy LHSVals;
796  computeValueKnownInPredecessorsImpl(BO->getOperand(0), BB, LHSVals,
797  WantInteger, RecursionSet, CxtI);
798 
799  // Try to use constant folding to simplify the binary operator.
800  for (const auto &LHSVal : LHSVals) {
801  Constant *V = LHSVal.first;
802  Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
803 
804  if (Constant *KC = getKnownConstant(Folded, WantInteger))
805  Result.emplace_back(KC, LHSVal.second);
806  }
807  }
808 
809  return !Result.empty();
810  }
811 
812  // Handle compare with phi operand, where the PHI is defined in this block.
813  if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
814  assert(Preference == WantInteger && "Compares only produce integers");
815  Type *CmpType = Cmp->getType();
816  Value *CmpLHS = Cmp->getOperand(0);
817  Value *CmpRHS = Cmp->getOperand(1);
818  CmpInst::Predicate Pred = Cmp->getPredicate();
819 
820  PHINode *PN = dyn_cast<PHINode>(CmpLHS);
821  if (!PN)
822  PN = dyn_cast<PHINode>(CmpRHS);
823  if (PN && PN->getParent() == BB) {
824  const DataLayout &DL = PN->getModule()->getDataLayout();
825  // We can do this simplification if any comparisons fold to true or false.
826  // See if any do.
827  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
828  BasicBlock *PredBB = PN->getIncomingBlock(i);
829  Value *LHS, *RHS;
830  if (PN == CmpLHS) {
831  LHS = PN->getIncomingValue(i);
832  RHS = CmpRHS->DoPHITranslation(BB, PredBB);
833  } else {
834  LHS = CmpLHS->DoPHITranslation(BB, PredBB);
835  RHS = PN->getIncomingValue(i);
836  }
837  Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
838  if (!Res) {
839  if (!isa<Constant>(RHS))
840  continue;
841 
842  // getPredicateOnEdge call will make no sense if LHS is defined in BB.
843  auto LHSInst = dyn_cast<Instruction>(LHS);
844  if (LHSInst && LHSInst->getParent() == BB)
845  continue;
846 
848  ResT = LVI->getPredicateOnEdge(Pred, LHS,
849  cast<Constant>(RHS), PredBB, BB,
850  CxtI ? CxtI : Cmp);
851  if (ResT == LazyValueInfo::Unknown)
852  continue;
853  Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
854  }
855 
856  if (Constant *KC = getKnownConstant(Res, WantInteger))
857  Result.emplace_back(KC, PredBB);
858  }
859 
860  return !Result.empty();
861  }
862 
863  // If comparing a live-in value against a constant, see if we know the
864  // live-in value on any predecessors.
865  if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
866  Constant *CmpConst = cast<Constant>(CmpRHS);
867 
868  if (!isa<Instruction>(CmpLHS) ||
869  cast<Instruction>(CmpLHS)->getParent() != BB) {
870  for (BasicBlock *P : predecessors(BB)) {
871  // If the value is known by LazyValueInfo to be a constant in a
872  // predecessor, use that information to try to thread this block.
874  LVI->getPredicateOnEdge(Pred, CmpLHS,
875  CmpConst, P, BB, CxtI ? CxtI : Cmp);
876  if (Res == LazyValueInfo::Unknown)
877  continue;
878 
879  Constant *ResC = ConstantInt::get(CmpType, Res);
880  Result.emplace_back(ResC, P);
881  }
882 
883  return !Result.empty();
884  }
885 
886  // InstCombine can fold some forms of constant range checks into
887  // (icmp (add (x, C1)), C2). See if we have we have such a thing with
888  // x as a live-in.
889  {
890  using namespace PatternMatch;
891 
892  Value *AddLHS;
893  ConstantInt *AddConst;
894  if (isa<ConstantInt>(CmpConst) &&
895  match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
896  if (!isa<Instruction>(AddLHS) ||
897  cast<Instruction>(AddLHS)->getParent() != BB) {
898  for (BasicBlock *P : predecessors(BB)) {
899  // If the value is known by LazyValueInfo to be a ConstantRange in
900  // a predecessor, use that information to try to thread this
901  // block.
902  ConstantRange CR = LVI->getConstantRangeOnEdge(
903  AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
904  // Propagate the range through the addition.
905  CR = CR.add(AddConst->getValue());
906 
907  // Get the range where the compare returns true.
909  Pred, cast<ConstantInt>(CmpConst)->getValue());
910 
911  Constant *ResC;
912  if (CmpRange.contains(CR))
913  ResC = ConstantInt::getTrue(CmpType);
914  else if (CmpRange.inverse().contains(CR))
915  ResC = ConstantInt::getFalse(CmpType);
916  else
917  continue;
918 
919  Result.emplace_back(ResC, P);
920  }
921 
922  return !Result.empty();
923  }
924  }
925  }
926 
927  // Try to find a constant value for the LHS of a comparison,
928  // and evaluate it statically if we can.
929  PredValueInfoTy LHSVals;
930  computeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
931  WantInteger, RecursionSet, CxtI);
932 
933  for (const auto &LHSVal : LHSVals) {
934  Constant *V = LHSVal.first;
935  Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
936  if (Constant *KC = getKnownConstant(Folded, WantInteger))
937  Result.emplace_back(KC, LHSVal.second);
938  }
939 
940  return !Result.empty();
941  }
942  }
943 
944  if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
945  // Handle select instructions where at least one operand is a known constant
946  // and we can figure out the condition value for any predecessor block.
947  Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
948  Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
949  PredValueInfoTy Conds;
950  if ((TrueVal || FalseVal) &&
951  computeValueKnownInPredecessorsImpl(SI->getCondition(), BB, Conds,
952  WantInteger, RecursionSet, CxtI)) {
953  for (auto &C : Conds) {
954  Constant *Cond = C.first;
955 
956  // Figure out what value to use for the condition.
957  bool KnownCond;
958  if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
959  // A known boolean.
960  KnownCond = CI->isOne();
961  } else {
962  assert(isa<UndefValue>(Cond) && "Unexpected condition value");
963  // Either operand will do, so be sure to pick the one that's a known
964  // constant.
965  // FIXME: Do this more cleverly if both values are known constants?
966  KnownCond = (TrueVal != nullptr);
967  }
968 
969  // See if the select has a known constant value for this predecessor.
970  if (Constant *Val = KnownCond ? TrueVal : FalseVal)
971  Result.emplace_back(Val, C.second);
972  }
973 
974  return !Result.empty();
975  }
976  }
977 
978  // If all else fails, see if LVI can figure out a constant value for us.
979  assert(CxtI->getParent() == BB && "CxtI should be in BB");
980  Constant *CI = LVI->getConstant(V, CxtI);
981  if (Constant *KC = getKnownConstant(CI, Preference)) {
982  for (BasicBlock *Pred : predecessors(BB))
983  Result.emplace_back(KC, Pred);
984  }
985 
986  return !Result.empty();
987 }
988 
989 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
990 /// in an undefined jump, decide which block is best to revector to.
991 ///
992 /// Since we can pick an arbitrary destination, we pick the successor with the
993 /// fewest predecessors. This should reduce the in-degree of the others.
995  Instruction *BBTerm = BB->getTerminator();
996  unsigned MinSucc = 0;
997  BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
998  // Compute the successor with the minimum number of predecessors.
999  unsigned MinNumPreds = pred_size(TestBB);
1000  for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1001  TestBB = BBTerm->getSuccessor(i);
1002  unsigned NumPreds = pred_size(TestBB);
1003  if (NumPreds < MinNumPreds) {
1004  MinSucc = i;
1005  MinNumPreds = NumPreds;
1006  }
1007  }
1008 
1009  return MinSucc;
1010 }
1011 
1013  if (!BB->hasAddressTaken()) return false;
1014 
1015  // If the block has its address taken, it may be a tree of dead constants
1016  // hanging off of it. These shouldn't keep the block alive.
1019  return !BA->use_empty();
1020 }
1021 
1022 /// processBlock - If there are any predecessors whose control can be threaded
1023 /// through to a successor, transform them now.
1025  // If the block is trivially dead, just return and let the caller nuke it.
1026  // This simplifies other transformations.
1027  if (DTU->isBBPendingDeletion(BB) ||
1028  (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
1029  return false;
1030 
1031  // If this block has a single predecessor, and if that pred has a single
1032  // successor, merge the blocks. This encourages recursive jump threading
1033  // because now the condition in this block can be threaded through
1034  // predecessors of our predecessor block.
1035  if (maybeMergeBasicBlockIntoOnlyPred(BB))
1036  return true;
1037 
1038  if (tryToUnfoldSelectInCurrBB(BB))
1039  return true;
1040 
1041  // Look if we can propagate guards to predecessors.
1042  if (HasGuards && processGuards(BB))
1043  return true;
1044 
1045  // What kind of constant we're looking for.
1047 
1048  // Look to see if the terminator is a conditional branch, switch or indirect
1049  // branch, if not we can't thread it.
1050  Value *Condition;
1051  Instruction *Terminator = BB->getTerminator();
1052  if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
1053  // Can't thread an unconditional jump.
1054  if (BI->isUnconditional()) return false;
1055  Condition = BI->getCondition();
1056  } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
1057  Condition = SI->getCondition();
1058  } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
1059  // Can't thread indirect branch with no successors.
1060  if (IB->getNumSuccessors() == 0) return false;
1061  Condition = IB->getAddress()->stripPointerCasts();
1063  } else {
1064  return false; // Must be an invoke or callbr.
1065  }
1066 
1067  // Keep track if we constant folded the condition in this invocation.
1068  bool ConstantFolded = false;
1069 
1070  // Run constant folding to see if we can reduce the condition to a simple
1071  // constant.
1072  if (Instruction *I = dyn_cast<Instruction>(Condition)) {
1073  Value *SimpleVal =
1074  ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
1075  if (SimpleVal) {
1076  I->replaceAllUsesWith(SimpleVal);
1077  if (isInstructionTriviallyDead(I, TLI))
1078  I->eraseFromParent();
1079  Condition = SimpleVal;
1080  ConstantFolded = true;
1081  }
1082  }
1083 
1084  // If the terminator is branching on an undef or freeze undef, we can pick any
1085  // of the successors to branch to. Let getBestDestForJumpOnUndef decide.
1086  auto *FI = dyn_cast<FreezeInst>(Condition);
1087  if (isa<UndefValue>(Condition) ||
1088  (FI && isa<UndefValue>(FI->getOperand(0)) && FI->hasOneUse())) {
1089  unsigned BestSucc = getBestDestForJumpOnUndef(BB);
1090  std::vector<DominatorTree::UpdateType> Updates;
1091 
1092  // Fold the branch/switch.
1093  Instruction *BBTerm = BB->getTerminator();
1094  Updates.reserve(BBTerm->getNumSuccessors());
1095  for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1096  if (i == BestSucc) continue;
1097  BasicBlock *Succ = BBTerm->getSuccessor(i);
1098  Succ->removePredecessor(BB, true);
1099  Updates.push_back({DominatorTree::Delete, BB, Succ});
1100  }
1101 
1102  LLVM_DEBUG(dbgs() << " In block '" << BB->getName()
1103  << "' folding undef terminator: " << *BBTerm << '\n');
1104  BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
1105  ++NumFolds;
1106  BBTerm->eraseFromParent();
1107  DTU->applyUpdatesPermissive(Updates);
1108  if (FI)
1109  FI->eraseFromParent();
1110  return true;
1111  }
1112 
1113  // If the terminator of this block is branching on a constant, simplify the
1114  // terminator to an unconditional branch. This can occur due to threading in
1115  // other blocks.
1116  if (getKnownConstant(Condition, Preference)) {
1117  LLVM_DEBUG(dbgs() << " In block '" << BB->getName()
1118  << "' folding terminator: " << *BB->getTerminator()
1119  << '\n');
1120  ++NumFolds;
1121  ConstantFoldTerminator(BB, true, nullptr, DTU);
1122  if (HasProfileData)
1123  BPI->eraseBlock(BB);
1124  return true;
1125  }
1126 
1127  Instruction *CondInst = dyn_cast<Instruction>(Condition);
1128 
1129  // All the rest of our checks depend on the condition being an instruction.
1130  if (!CondInst) {
1131  // FIXME: Unify this with code below.
1132  if (processThreadableEdges(Condition, BB, Preference, Terminator))
1133  return true;
1134  return ConstantFolded;
1135  }
1136 
1137  if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
1138  // If we're branching on a conditional, LVI might be able to determine
1139  // it's value at the branch instruction. We only handle comparisons
1140  // against a constant at this time.
1141  // TODO: This should be extended to handle switches as well.
1142  BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1143  Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
1144  if (CondBr && CondConst) {
1145  // We should have returned as soon as we turn a conditional branch to
1146  // unconditional. Because its no longer interesting as far as jump
1147  // threading is concerned.
1148  assert(CondBr->isConditional() && "Threading on unconditional terminator");
1149 
1151  LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
1152  CondConst, CondBr, /*UseBlockValue=*/false);
1153  if (Ret != LazyValueInfo::Unknown) {
1154  unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
1155  unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
1156  BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove);
1157  ToRemoveSucc->removePredecessor(BB, true);
1158  BranchInst *UncondBr =
1159  BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
1160  UncondBr->setDebugLoc(CondBr->getDebugLoc());
1161  ++NumFolds;
1162  CondBr->eraseFromParent();
1163  if (CondCmp->use_empty())
1164  CondCmp->eraseFromParent();
1165  // We can safely replace *some* uses of the CondInst if it has
1166  // exactly one value as returned by LVI. RAUW is incorrect in the
1167  // presence of guards and assumes, that have the `Cond` as the use. This
1168  // is because we use the guards/assume to reason about the `Cond` value
1169  // at the end of block, but RAUW unconditionally replaces all uses
1170  // including the guards/assumes themselves and the uses before the
1171  // guard/assume.
1172  else if (CondCmp->getParent() == BB) {
1173  auto *CI = Ret == LazyValueInfo::True ?
1174  ConstantInt::getTrue(CondCmp->getType()) :
1175  ConstantInt::getFalse(CondCmp->getType());
1176  replaceFoldableUses(CondCmp, CI);
1177  }
1178  DTU->applyUpdatesPermissive(
1179  {{DominatorTree::Delete, BB, ToRemoveSucc}});
1180  if (HasProfileData)
1181  BPI->eraseBlock(BB);
1182  return true;
1183  }
1184 
1185  // We did not manage to simplify this branch, try to see whether
1186  // CondCmp depends on a known phi-select pattern.
1187  if (tryToUnfoldSelect(CondCmp, BB))
1188  return true;
1189  }
1190  }
1191 
1192  if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
1193  if (tryToUnfoldSelect(SI, BB))
1194  return true;
1195 
1196  // Check for some cases that are worth simplifying. Right now we want to look
1197  // for loads that are used by a switch or by the condition for the branch. If
1198  // we see one, check to see if it's partially redundant. If so, insert a PHI
1199  // which can then be used to thread the values.
1200  Value *SimplifyValue = CondInst;
1201 
1202  if (auto *FI = dyn_cast<FreezeInst>(SimplifyValue))
1203  // Look into freeze's operand
1204  SimplifyValue = FI->getOperand(0);
1205 
1206  if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
1207  if (isa<Constant>(CondCmp->getOperand(1)))
1208  SimplifyValue = CondCmp->getOperand(0);
1209 
1210  // TODO: There are other places where load PRE would be profitable, such as
1211  // more complex comparisons.
1212  if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue))
1213  if (simplifyPartiallyRedundantLoad(LoadI))
1214  return true;
1215 
1216  // Before threading, try to propagate profile data backwards:
1217  if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1218  if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1220 
1221  // Handle a variety of cases where we are branching on something derived from
1222  // a PHI node in the current block. If we can prove that any predecessors
1223  // compute a predictable value based on a PHI node, thread those predecessors.
1224  if (processThreadableEdges(CondInst, BB, Preference, Terminator))
1225  return true;
1226 
1227  // If this is an otherwise-unfoldable branch on a phi node or freeze(phi) in
1228  // the current block, see if we can simplify.
1229  PHINode *PN = dyn_cast<PHINode>(
1230  isa<FreezeInst>(CondInst) ? cast<FreezeInst>(CondInst)->getOperand(0)
1231  : CondInst);
1232 
1233  if (PN && PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1234  return processBranchOnPHI(PN);
1235 
1236  // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1237  if (CondInst->getOpcode() == Instruction::Xor &&
1238  CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1239  return processBranchOnXOR(cast<BinaryOperator>(CondInst));
1240 
1241  // Search for a stronger dominating condition that can be used to simplify a
1242  // conditional branch leaving BB.
1243  if (processImpliedCondition(BB))
1244  return true;
1245 
1246  return false;
1247 }
1248 
1250  auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
1251  if (!BI || !BI->isConditional())
1252  return false;
1253 
1254  Value *Cond = BI->getCondition();
1255  BasicBlock *CurrentBB = BB;
1256  BasicBlock *CurrentPred = BB->getSinglePredecessor();
1257  unsigned Iter = 0;
1258 
1259  auto &DL = BB->getModule()->getDataLayout();
1260 
1261  while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1262  auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
1263  if (!PBI || !PBI->isConditional())
1264  return false;
1265  if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
1266  return false;
1267 
1268  bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB;
1269  Optional<bool> Implication =
1270  isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue);
1271  if (Implication) {
1272  BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1);
1273  BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0);
1274  RemoveSucc->removePredecessor(BB);
1275  BranchInst *UncondBI = BranchInst::Create(KeepSucc, BI);
1276  UncondBI->setDebugLoc(BI->getDebugLoc());
1277  ++NumFolds;
1278  BI->eraseFromParent();
1279  DTU->applyUpdatesPermissive({{DominatorTree::Delete, BB, RemoveSucc}});
1280  if (HasProfileData)
1281  BPI->eraseBlock(BB);
1282  return true;
1283  }
1284  CurrentBB = CurrentPred;
1285  CurrentPred = CurrentBB->getSinglePredecessor();
1286  }
1287 
1288  return false;
1289 }
1290 
1291 /// Return true if Op is an instruction defined in the given block.
1293  if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1294  if (OpInst->getParent() == BB)
1295  return true;
1296  return false;
1297 }
1298 
1299 /// simplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1300 /// redundant load instruction, eliminate it by replacing it with a PHI node.
1301 /// This is an important optimization that encourages jump threading, and needs
1302 /// to be run interlaced with other jump threading tasks.
1304  // Don't hack volatile and ordered loads.
1305  if (!LoadI->isUnordered()) return false;
1306 
1307  // If the load is defined in a block with exactly one predecessor, it can't be
1308  // partially redundant.
1309  BasicBlock *LoadBB = LoadI->getParent();
1310  if (LoadBB->getSinglePredecessor())
1311  return false;
1312 
1313  // If the load is defined in an EH pad, it can't be partially redundant,
1314  // because the edges between the invoke and the EH pad cannot have other
1315  // instructions between them.
1316  if (LoadBB->isEHPad())
1317  return false;
1318 
1319  Value *LoadedPtr = LoadI->getOperand(0);
1320 
1321  // If the loaded operand is defined in the LoadBB and its not a phi,
1322  // it can't be available in predecessors.
1323  if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
1324  return false;
1325 
1326  // Scan a few instructions up from the load, to see if it is obviously live at
1327  // the entry to its block.
1328  BasicBlock::iterator BBIt(LoadI);
1329  bool IsLoadCSE;
1330  if (Value *AvailableVal = FindAvailableLoadedValue(
1331  LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1332  // If the value of the load is locally available within the block, just use
1333  // it. This frequently occurs for reg2mem'd allocas.
1334 
1335  if (IsLoadCSE) {
1336  LoadInst *NLoadI = cast<LoadInst>(AvailableVal);
1337  combineMetadataForCSE(NLoadI, LoadI, false);
1338  };
1339 
1340  // If the returned value is the load itself, replace with an undef. This can
1341  // only happen in dead loops.
1342  if (AvailableVal == LoadI)
1343  AvailableVal = UndefValue::get(LoadI->getType());
1344  if (AvailableVal->getType() != LoadI->getType())
1345  AvailableVal = CastInst::CreateBitOrPointerCast(
1346  AvailableVal, LoadI->getType(), "", LoadI);
1347  LoadI->replaceAllUsesWith(AvailableVal);
1348  LoadI->eraseFromParent();
1349  return true;
1350  }
1351 
1352  // Otherwise, if we scanned the whole block and got to the top of the block,
1353  // we know the block is locally transparent to the load. If not, something
1354  // might clobber its value.
1355  if (BBIt != LoadBB->begin())
1356  return false;
1357 
1358  // If all of the loads and stores that feed the value have the same AA tags,
1359  // then we can propagate them onto any newly inserted loads.
1360  AAMDNodes AATags = LoadI->getAAMetadata();
1361 
1362  SmallPtrSet<BasicBlock*, 8> PredsScanned;
1363 
1364  using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1365 
1366  AvailablePredsTy AvailablePreds;
1367  BasicBlock *OneUnavailablePred = nullptr;
1368  SmallVector<LoadInst*, 8> CSELoads;
1369 
1370  // If we got here, the loaded value is transparent through to the start of the
1371  // block. Check to see if it is available in any of the predecessor blocks.
1372  for (BasicBlock *PredBB : predecessors(LoadBB)) {
1373  // If we already scanned this predecessor, skip it.
1374  if (!PredsScanned.insert(PredBB).second)
1375  continue;
1376 
1377  BBIt = PredBB->end();
1378  unsigned NumScanedInst = 0;
1379  Value *PredAvailable = nullptr;
1380  // NOTE: We don't CSE load that is volatile or anything stronger than
1381  // unordered, that should have been checked when we entered the function.
1382  assert(LoadI->isUnordered() &&
1383  "Attempting to CSE volatile or atomic loads");
1384  // If this is a load on a phi pointer, phi-translate it and search
1385  // for available load/store to the pointer in predecessors.
1386  Type *AccessTy = LoadI->getType();
1387  const auto &DL = LoadI->getModule()->getDataLayout();
1388  MemoryLocation Loc(LoadedPtr->DoPHITranslation(LoadBB, PredBB),
1389  LocationSize::precise(DL.getTypeStoreSize(AccessTy)),
1390  AATags);
1391  PredAvailable = findAvailablePtrLoadStore(Loc, AccessTy, LoadI->isAtomic(),
1392  PredBB, BBIt, DefMaxInstsToScan,
1393  AA, &IsLoadCSE, &NumScanedInst);
1394 
1395  // If PredBB has a single predecessor, continue scanning through the
1396  // single predecessor.
1397  BasicBlock *SinglePredBB = PredBB;
1398  while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1399  NumScanedInst < DefMaxInstsToScan) {
1400  SinglePredBB = SinglePredBB->getSinglePredecessor();
1401  if (SinglePredBB) {
1402  BBIt = SinglePredBB->end();
1403  PredAvailable = findAvailablePtrLoadStore(
1404  Loc, AccessTy, LoadI->isAtomic(), SinglePredBB, BBIt,
1405  (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1406  &NumScanedInst);
1407  }
1408  }
1409 
1410  if (!PredAvailable) {
1411  OneUnavailablePred = PredBB;
1412  continue;
1413  }
1414 
1415  if (IsLoadCSE)
1416  CSELoads.push_back(cast<LoadInst>(PredAvailable));
1417 
1418  // If so, this load is partially redundant. Remember this info so that we
1419  // can create a PHI node.
1420  AvailablePreds.emplace_back(PredBB, PredAvailable);
1421  }
1422 
1423  // If the loaded value isn't available in any predecessor, it isn't partially
1424  // redundant.
1425  if (AvailablePreds.empty()) return false;
1426 
1427  // Okay, the loaded value is available in at least one (and maybe all!)
1428  // predecessors. If the value is unavailable in more than one unique
1429  // predecessor, we want to insert a merge block for those common predecessors.
1430  // This ensures that we only have to insert one reload, thus not increasing
1431  // code size.
1432  BasicBlock *UnavailablePred = nullptr;
1433 
1434  // If the value is unavailable in one of predecessors, we will end up
1435  // inserting a new instruction into them. It is only valid if all the
1436  // instructions before LoadI are guaranteed to pass execution to its
1437  // successor, or if LoadI is safe to speculate.
1438  // TODO: If this logic becomes more complex, and we will perform PRE insertion
1439  // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1440  // It requires domination tree analysis, so for this simple case it is an
1441  // overkill.
1442  if (PredsScanned.size() != AvailablePreds.size() &&
1444  for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1446  return false;
1447 
1448  // If there is exactly one predecessor where the value is unavailable, the
1449  // already computed 'OneUnavailablePred' block is it. If it ends in an
1450  // unconditional branch, we know that it isn't a critical edge.
1451  if (PredsScanned.size() == AvailablePreds.size()+1 &&
1452  OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1453  UnavailablePred = OneUnavailablePred;
1454  } else if (PredsScanned.size() != AvailablePreds.size()) {
1455  // Otherwise, we had multiple unavailable predecessors or we had a critical
1456  // edge from the one.
1457  SmallVector<BasicBlock*, 8> PredsToSplit;
1458  SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1459 
1460  for (const auto &AvailablePred : AvailablePreds)
1461  AvailablePredSet.insert(AvailablePred.first);
1462 
1463  // Add all the unavailable predecessors to the PredsToSplit list.
1464  for (BasicBlock *P : predecessors(LoadBB)) {
1465  // If the predecessor is an indirect goto, we can't split the edge.
1466  // Same for CallBr.
1467  if (isa<IndirectBrInst>(P->getTerminator()) ||
1468  isa<CallBrInst>(P->getTerminator()))
1469  return false;
1470 
1471  if (!AvailablePredSet.count(P))
1472  PredsToSplit.push_back(P);
1473  }
1474 
1475  // Split them out to their own block.
1476  UnavailablePred = splitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1477  }
1478 
1479  // If the value isn't available in all predecessors, then there will be
1480  // exactly one where it isn't available. Insert a load on that edge and add
1481  // it to the AvailablePreds list.
1482  if (UnavailablePred) {
1483  assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1484  "Can't handle critical edge here!");
1485  LoadInst *NewVal = new LoadInst(
1486  LoadI->getType(), LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1487  LoadI->getName() + ".pr", false, LoadI->getAlign(),
1488  LoadI->getOrdering(), LoadI->getSyncScopeID(),
1489  UnavailablePred->getTerminator());
1490  NewVal->setDebugLoc(LoadI->getDebugLoc());
1491  if (AATags)
1492  NewVal->setAAMetadata(AATags);
1493 
1494  AvailablePreds.emplace_back(UnavailablePred, NewVal);
1495  }
1496 
1497  // Now we know that each predecessor of this block has a value in
1498  // AvailablePreds, sort them for efficient access as we're walking the preds.
1499  array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1500 
1501  // Create a PHI node at the start of the block for the PRE'd load value.
1502  pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1503  PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "",
1504  &LoadBB->front());
1505  PN->takeName(LoadI);
1506  PN->setDebugLoc(LoadI->getDebugLoc());
1507 
1508  // Insert new entries into the PHI for each predecessor. A single block may
1509  // have multiple entries here.
1510  for (pred_iterator PI = PB; PI != PE; ++PI) {
1511  BasicBlock *P = *PI;
1512  AvailablePredsTy::iterator I =
1513  llvm::lower_bound(AvailablePreds, std::make_pair(P, (Value *)nullptr));
1514 
1515  assert(I != AvailablePreds.end() && I->first == P &&
1516  "Didn't find entry for predecessor!");
1517 
1518  // If we have an available predecessor but it requires casting, insert the
1519  // cast in the predecessor and use the cast. Note that we have to update the
1520  // AvailablePreds vector as we go so that all of the PHI entries for this
1521  // predecessor use the same bitcast.
1522  Value *&PredV = I->second;
1523  if (PredV->getType() != LoadI->getType())
1524  PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "",
1525  P->getTerminator());
1526 
1527  PN->addIncoming(PredV, I->first);
1528  }
1529 
1530  for (LoadInst *PredLoadI : CSELoads) {
1531  combineMetadataForCSE(PredLoadI, LoadI, true);
1532  }
1533 
1534  LoadI->replaceAllUsesWith(PN);
1535  LoadI->eraseFromParent();
1536 
1537  return true;
1538 }
1539 
1540 /// findMostPopularDest - The specified list contains multiple possible
1541 /// threadable destinations. Pick the one that occurs the most frequently in
1542 /// the list.
1543 static BasicBlock *
1545  const SmallVectorImpl<std::pair<BasicBlock *,
1546  BasicBlock *>> &PredToDestList) {
1547  assert(!PredToDestList.empty());
1548 
1549  // Determine popularity. If there are multiple possible destinations, we
1550  // explicitly choose to ignore 'undef' destinations. We prefer to thread
1551  // blocks with known and real destinations to threading undef. We'll handle
1552  // them later if interesting.
1553  MapVector<BasicBlock *, unsigned> DestPopularity;
1554 
1555  // Populate DestPopularity with the successors in the order they appear in the
1556  // successor list. This way, we ensure determinism by iterating it in the
1557  // same order in std::max_element below. We map nullptr to 0 so that we can
1558  // return nullptr when PredToDestList contains nullptr only.
1559  DestPopularity[nullptr] = 0;
1560  for (auto *SuccBB : successors(BB))
1561  DestPopularity[SuccBB] = 0;
1562 
1563  for (const auto &PredToDest : PredToDestList)
1564  if (PredToDest.second)
1565  DestPopularity[PredToDest.second]++;
1566 
1567  // Find the most popular dest.
1568  using VT = decltype(DestPopularity)::value_type;
1569  auto MostPopular = std::max_element(
1570  DestPopularity.begin(), DestPopularity.end(),
1571  [](const VT &L, const VT &R) { return L.second < R.second; });
1572 
1573  // Okay, we have finally picked the most popular destination.
1574  return MostPopular->first;
1575 }
1576 
1577 // Try to evaluate the value of V when the control flows from PredPredBB to
1578 // BB->getSinglePredecessor() and then on to BB.
1580  BasicBlock *PredPredBB,
1581  Value *V) {
1582  BasicBlock *PredBB = BB->getSinglePredecessor();
1583  assert(PredBB && "Expected a single predecessor");
1584 
1585  if (Constant *Cst = dyn_cast<Constant>(V)) {
1586  return Cst;
1587  }
1588 
1589  // Consult LVI if V is not an instruction in BB or PredBB.
1590  Instruction *I = dyn_cast<Instruction>(V);
1591  if (!I || (I->getParent() != BB && I->getParent() != PredBB)) {
1592  return LVI->getConstantOnEdge(V, PredPredBB, PredBB, nullptr);
1593  }
1594 
1595  // Look into a PHI argument.
1596  if (PHINode *PHI = dyn_cast<PHINode>(V)) {
1597  if (PHI->getParent() == PredBB)
1598  return dyn_cast<Constant>(PHI->getIncomingValueForBlock(PredPredBB));
1599  return nullptr;
1600  }
1601 
1602  // If we have a CmpInst, try to fold it for each incoming edge into PredBB.
1603  if (CmpInst *CondCmp = dyn_cast<CmpInst>(V)) {
1604  if (CondCmp->getParent() == BB) {
1605  Constant *Op0 =
1606  evaluateOnPredecessorEdge(BB, PredPredBB, CondCmp->getOperand(0));
1607  Constant *Op1 =
1608  evaluateOnPredecessorEdge(BB, PredPredBB, CondCmp->getOperand(1));
1609  if (Op0 && Op1) {
1610  return ConstantExpr::getCompare(CondCmp->getPredicate(), Op0, Op1);
1611  }
1612  }
1613  return nullptr;
1614  }
1615 
1616  return nullptr;
1617 }
1618 
1621  Instruction *CxtI) {
1622  // If threading this would thread across a loop header, don't even try to
1623  // thread the edge.
1624  if (LoopHeaders.count(BB))
1625  return false;
1626 
1627  PredValueInfoTy PredValues;
1628  if (!computeValueKnownInPredecessors(Cond, BB, PredValues, Preference,
1629  CxtI)) {
1630  // We don't have known values in predecessors. See if we can thread through
1631  // BB and its sole predecessor.
1632  return maybethreadThroughTwoBasicBlocks(BB, Cond);
1633  }
1634 
1635  assert(!PredValues.empty() &&
1636  "computeValueKnownInPredecessors returned true with no values");
1637 
1638  LLVM_DEBUG(dbgs() << "IN BB: " << *BB;
1639  for (const auto &PredValue : PredValues) {
1640  dbgs() << " BB '" << BB->getName()
1641  << "': FOUND condition = " << *PredValue.first
1642  << " for pred '" << PredValue.second->getName() << "'.\n";
1643  });
1644 
1645  // Decide what we want to thread through. Convert our list of known values to
1646  // a list of known destinations for each pred. This also discards duplicate
1647  // predecessors and keeps track of the undefined inputs (which are represented
1648  // as a null dest in the PredToDestList).
1649  SmallPtrSet<BasicBlock*, 16> SeenPreds;
1651 
1652  BasicBlock *OnlyDest = nullptr;
1653  BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1654  Constant *OnlyVal = nullptr;
1655  Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1656 
1657  for (const auto &PredValue : PredValues) {
1658  BasicBlock *Pred = PredValue.second;
1659  if (!SeenPreds.insert(Pred).second)
1660  continue; // Duplicate predecessor entry.
1661 
1662  Constant *Val = PredValue.first;
1663 
1664  BasicBlock *DestBB;
1665  if (isa<UndefValue>(Val))
1666  DestBB = nullptr;
1667  else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1668  assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1669  DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1670  } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1671  assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1672  DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1673  } else {
1674  assert(isa<IndirectBrInst>(BB->getTerminator())
1675  && "Unexpected terminator");
1676  assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress");
1677  DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1678  }
1679 
1680  // If we have exactly one destination, remember it for efficiency below.
1681  if (PredToDestList.empty()) {
1682  OnlyDest = DestBB;
1683  OnlyVal = Val;
1684  } else {
1685  if (OnlyDest != DestBB)
1686  OnlyDest = MultipleDestSentinel;
1687  // It possible we have same destination, but different value, e.g. default
1688  // case in switchinst.
1689  if (Val != OnlyVal)
1690  OnlyVal = MultipleVal;
1691  }
1692 
1693  // If the predecessor ends with an indirect goto, we can't change its
1694  // destination. Same for CallBr.
1695  if (isa<IndirectBrInst>(Pred->getTerminator()) ||
1696  isa<CallBrInst>(Pred->getTerminator()))
1697  continue;
1698 
1699  PredToDestList.emplace_back(Pred, DestBB);
1700  }
1701 
1702  // If all edges were unthreadable, we fail.
1703  if (PredToDestList.empty())
1704  return false;
1705 
1706  // If all the predecessors go to a single known successor, we want to fold,
1707  // not thread. By doing so, we do not need to duplicate the current block and
1708  // also miss potential opportunities in case we dont/cant duplicate.
1709  if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1710  if (BB->hasNPredecessors(PredToDestList.size())) {
1711  bool SeenFirstBranchToOnlyDest = false;
1712  std::vector <DominatorTree::UpdateType> Updates;
1713  Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);
1714  for (BasicBlock *SuccBB : successors(BB)) {
1715  if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1716  SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1717  } else {
1718  SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1719  Updates.push_back({DominatorTree::Delete, BB, SuccBB});
1720  }
1721  }
1722 
1723  // Finally update the terminator.
1724  Instruction *Term = BB->getTerminator();
1725  BranchInst::Create(OnlyDest, Term);
1726  ++NumFolds;
1727  Term->eraseFromParent();
1728  DTU->applyUpdatesPermissive(Updates);
1729  if (HasProfileData)
1730  BPI->eraseBlock(BB);
1731 
1732  // If the condition is now dead due to the removal of the old terminator,
1733  // erase it.
1734  if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1735  if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1736  CondInst->eraseFromParent();
1737  // We can safely replace *some* uses of the CondInst if it has
1738  // exactly one value as returned by LVI. RAUW is incorrect in the
1739  // presence of guards and assumes, that have the `Cond` as the use. This
1740  // is because we use the guards/assume to reason about the `Cond` value
1741  // at the end of block, but RAUW unconditionally replaces all uses
1742  // including the guards/assumes themselves and the uses before the
1743  // guard/assume.
1744  else if (OnlyVal && OnlyVal != MultipleVal &&
1745  CondInst->getParent() == BB)
1746  replaceFoldableUses(CondInst, OnlyVal);
1747  }
1748  return true;
1749  }
1750  }
1751 
1752  // Determine which is the most common successor. If we have many inputs and
1753  // this block is a switch, we want to start by threading the batch that goes
1754  // to the most popular destination first. If we only know about one
1755  // threadable destination (the common case) we can avoid this.
1756  BasicBlock *MostPopularDest = OnlyDest;
1757 
1758  if (MostPopularDest == MultipleDestSentinel) {
1759  // Remove any loop headers from the Dest list, threadEdge conservatively
1760  // won't process them, but we might have other destination that are eligible
1761  // and we still want to process.
1762  erase_if(PredToDestList,
1763  [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1764  return LoopHeaders.contains(PredToDest.second);
1765  });
1766 
1767  if (PredToDestList.empty())
1768  return false;
1769 
1770  MostPopularDest = findMostPopularDest(BB, PredToDestList);
1771  }
1772 
1773  // Now that we know what the most popular destination is, factor all
1774  // predecessors that will jump to it into a single predecessor.
1775  SmallVector<BasicBlock*, 16> PredsToFactor;
1776  for (const auto &PredToDest : PredToDestList)
1777  if (PredToDest.second == MostPopularDest) {
1778  BasicBlock *Pred = PredToDest.first;
1779 
1780  // This predecessor may be a switch or something else that has multiple
1781  // edges to the block. Factor each of these edges by listing them
1782  // according to # occurrences in PredsToFactor.
1783  for (BasicBlock *Succ : successors(Pred))
1784  if (Succ == BB)
1785  PredsToFactor.push_back(Pred);
1786  }
1787 
1788  // If the threadable edges are branching on an undefined value, we get to pick
1789  // the destination that these predecessors should get to.
1790  if (!MostPopularDest)
1791  MostPopularDest = BB->getTerminator()->
1792  getSuccessor(getBestDestForJumpOnUndef(BB));
1793 
1794  // Ok, try to thread it!
1795  return tryThreadEdge(BB, PredsToFactor, MostPopularDest);
1796 }
1797 
1798 /// processBranchOnPHI - We have an otherwise unthreadable conditional branch on
1799 /// a PHI node (or freeze PHI) in the current block. See if there are any
1800 /// simplifications we can do based on inputs to the phi node.
1802  BasicBlock *BB = PN->getParent();
1803 
1804  // TODO: We could make use of this to do it once for blocks with common PHI
1805  // values.
1807  PredBBs.resize(1);
1808 
1809  // If any of the predecessor blocks end in an unconditional branch, we can
1810  // *duplicate* the conditional branch into that block in order to further
1811  // encourage jump threading and to eliminate cases where we have branch on a
1812  // phi of an icmp (branch on icmp is much better).
1813  // This is still beneficial when a frozen phi is used as the branch condition
1814  // because it allows CodeGenPrepare to further canonicalize br(freeze(icmp))
1815  // to br(icmp(freeze ...)).
1816  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1817  BasicBlock *PredBB = PN->getIncomingBlock(i);
1818  if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1819  if (PredBr->isUnconditional()) {
1820  PredBBs[0] = PredBB;
1821  // Try to duplicate BB into PredBB.
1822  if (duplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1823  return true;
1824  }
1825  }
1826 
1827  return false;
1828 }
1829 
1830 /// processBranchOnXOR - We have an otherwise unthreadable conditional branch on
1831 /// a xor instruction in the current block. See if there are any
1832 /// simplifications we can do based on inputs to the xor.
1834  BasicBlock *BB = BO->getParent();
1835 
1836  // If either the LHS or RHS of the xor is a constant, don't do this
1837  // optimization.
1838  if (isa<ConstantInt>(BO->getOperand(0)) ||
1839  isa<ConstantInt>(BO->getOperand(1)))
1840  return false;
1841 
1842  // If the first instruction in BB isn't a phi, we won't be able to infer
1843  // anything special about any particular predecessor.
1844  if (!isa<PHINode>(BB->front()))
1845  return false;
1846 
1847  // If this BB is a landing pad, we won't be able to split the edge into it.
1848  if (BB->isEHPad())
1849  return false;
1850 
1851  // If we have a xor as the branch input to this block, and we know that the
1852  // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1853  // the condition into the predecessor and fix that value to true, saving some
1854  // logical ops on that path and encouraging other paths to simplify.
1855  //
1856  // This copies something like this:
1857  //
1858  // BB:
1859  // %X = phi i1 [1], [%X']
1860  // %Y = icmp eq i32 %A, %B
1861  // %Z = xor i1 %X, %Y
1862  // br i1 %Z, ...
1863  //
1864  // Into:
1865  // BB':
1866  // %Y = icmp ne i32 %A, %B
1867  // br i1 %Y, ...
1868 
1869  PredValueInfoTy XorOpValues;
1870  bool isLHS = true;
1871  if (!computeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1872  WantInteger, BO)) {
1873  assert(XorOpValues.empty());
1874  if (!computeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1875  WantInteger, BO))
1876  return false;
1877  isLHS = false;
1878  }
1879 
1880  assert(!XorOpValues.empty() &&
1881  "computeValueKnownInPredecessors returned true with no values");
1882 
1883  // Scan the information to see which is most popular: true or false. The
1884  // predecessors can be of the set true, false, or undef.
1885  unsigned NumTrue = 0, NumFalse = 0;
1886  for (const auto &XorOpValue : XorOpValues) {
1887  if (isa<UndefValue>(XorOpValue.first))
1888  // Ignore undefs for the count.
1889  continue;
1890  if (cast<ConstantInt>(XorOpValue.first)->isZero())
1891  ++NumFalse;
1892  else
1893  ++NumTrue;
1894  }
1895 
1896  // Determine which value to split on, true, false, or undef if neither.
1897  ConstantInt *SplitVal = nullptr;
1898  if (NumTrue > NumFalse)
1899  SplitVal = ConstantInt::getTrue(BB->getContext());
1900  else if (NumTrue != 0 || NumFalse != 0)
1901  SplitVal = ConstantInt::getFalse(BB->getContext());
1902 
1903  // Collect all of the blocks that this can be folded into so that we can
1904  // factor this once and clone it once.
1905  SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1906  for (const auto &XorOpValue : XorOpValues) {
1907  if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1908  continue;
1909 
1910  BlocksToFoldInto.push_back(XorOpValue.second);
1911  }
1912 
1913  // If we inferred a value for all of the predecessors, then duplication won't
1914  // help us. However, we can just replace the LHS or RHS with the constant.
1915  if (BlocksToFoldInto.size() ==
1916  cast<PHINode>(BB->front()).getNumIncomingValues()) {
1917  if (!SplitVal) {
1918  // If all preds provide undef, just nuke the xor, because it is undef too.
1920  BO->eraseFromParent();
1921  } else if (SplitVal->isZero()) {
1922  // If all preds provide 0, replace the xor with the other input.
1923  BO->replaceAllUsesWith(BO->getOperand(isLHS));
1924  BO->eraseFromParent();
1925  } else {
1926  // If all preds provide 1, set the computed value to 1.
1927  BO->setOperand(!isLHS, SplitVal);
1928  }
1929 
1930  return true;
1931  }
1932 
1933  // If any of predecessors end with an indirect goto, we can't change its
1934  // destination. Same for CallBr.
1935  if (any_of(BlocksToFoldInto, [](BasicBlock *Pred) {
1936  return isa<IndirectBrInst>(Pred->getTerminator()) ||
1937  isa<CallBrInst>(Pred->getTerminator());
1938  }))
1939  return false;
1940 
1941  // Try to duplicate BB into PredBB.
1942  return duplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1943 }
1944 
1945 /// addPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1946 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1947 /// NewPred using the entries from OldPred (suitably mapped).
1949  BasicBlock *OldPred,
1950  BasicBlock *NewPred,
1952  for (PHINode &PN : PHIBB->phis()) {
1953  // Ok, we have a PHI node. Figure out what the incoming value was for the
1954  // DestBlock.
1955  Value *IV = PN.getIncomingValueForBlock(OldPred);
1956 
1957  // Remap the value if necessary.
1958  if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1960  if (I != ValueMap.end())
1961  IV = I->second;
1962  }
1963 
1964  PN.addIncoming(IV, NewPred);
1965  }
1966 }
1967 
1968 /// Merge basic block BB into its sole predecessor if possible.
1970  BasicBlock *SinglePred = BB->getSinglePredecessor();
1971  if (!SinglePred)
1972  return false;
1973 
1974  const Instruction *TI = SinglePred->getTerminator();
1975  if (TI->isExceptionalTerminator() || TI->getNumSuccessors() != 1 ||
1976  SinglePred == BB || hasAddressTakenAndUsed(BB))
1977  return false;
1978 
1979  // If SinglePred was a loop header, BB becomes one.
1980  if (LoopHeaders.erase(SinglePred))
1981  LoopHeaders.insert(BB);
1982 
1983  LVI->eraseBlock(SinglePred);
1985 
1986  // Now that BB is merged into SinglePred (i.e. SinglePred code followed by
1987  // BB code within one basic block `BB`), we need to invalidate the LVI
1988  // information associated with BB, because the LVI information need not be
1989  // true for all of BB after the merge. For example,
1990  // Before the merge, LVI info and code is as follows:
1991  // SinglePred: <LVI info1 for %p val>
1992  // %y = use of %p
1993  // call @exit() // need not transfer execution to successor.
1994  // assume(%p) // from this point on %p is true
1995  // br label %BB
1996  // BB: <LVI info2 for %p val, i.e. %p is true>
1997  // %x = use of %p
1998  // br label exit
1999  //
2000  // Note that this LVI info for blocks BB and SinglPred is correct for %p
2001  // (info2 and info1 respectively). After the merge and the deletion of the
2002  // LVI info1 for SinglePred. We have the following code:
2003  // BB: <LVI info2 for %p val>
2004  // %y = use of %p
2005  // call @exit()
2006  // assume(%p)
2007  // %x = use of %p <-- LVI info2 is correct from here onwards.
2008  // br label exit
2009  // LVI info2 for BB is incorrect at the beginning of BB.
2010 
2011  // Invalidate LVI information for BB if the LVI is not provably true for
2012  // all of BB.
2014  LVI->eraseBlock(BB);
2015  return true;
2016 }
2017 
2018 /// Update the SSA form. NewBB contains instructions that are copied from BB.
2019 /// ValueMapping maps old values in BB to new ones in NewBB.
2021  BasicBlock *BB, BasicBlock *NewBB,
2022  DenseMap<Instruction *, Value *> &ValueMapping) {
2023  // If there were values defined in BB that are used outside the block, then we
2024  // now have to update all uses of the value to use either the original value,
2025  // the cloned value, or some PHI derived value. This can require arbitrary
2026  // PHI insertion, of which we are prepared to do, clean these up now.
2027  SSAUpdater SSAUpdate;
2028  SmallVector<Use *, 16> UsesToRename;
2029 
2030  for (Instruction &I : *BB) {
2031  // Scan all uses of this instruction to see if it is used outside of its
2032  // block, and if so, record them in UsesToRename.
2033  for (Use &U : I.uses()) {
2034  Instruction *User = cast<Instruction>(U.getUser());
2035  if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2036  if (UserPN->getIncomingBlock(U) == BB)
2037  continue;
2038  } else if (User->getParent() == BB)
2039  continue;
2040 
2041  UsesToRename.push_back(&U);
2042  }
2043 
2044  // If there are no uses outside the block, we're done with this instruction.
2045  if (UsesToRename.empty())
2046  continue;
2047  LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
2048 
2049  // We found a use of I outside of BB. Rename all uses of I that are outside
2050  // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2051  // with the two values we know.
2052  SSAUpdate.Initialize(I.getType(), I.getName());
2053  SSAUpdate.AddAvailableValue(BB, &I);
2054  SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
2055 
2056  while (!UsesToRename.empty())
2057  SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2058  LLVM_DEBUG(dbgs() << "\n");
2059  }
2060 }
2061 
2062 /// Clone instructions in range [BI, BE) to NewBB. For PHI nodes, we only clone
2063 /// arguments that come from PredBB. Return the map from the variables in the
2064 /// source basic block to the variables in the newly created basic block.
2067  BasicBlock::iterator BE, BasicBlock *NewBB,
2068  BasicBlock *PredBB) {
2069  // We are going to have to map operands from the source basic block to the new
2070  // copy of the block 'NewBB'. If there are PHI nodes in the source basic
2071  // block, evaluate them to account for entry from PredBB.
2072  DenseMap<Instruction *, Value *> ValueMapping;
2073 
2074  // Clone the phi nodes of the source basic block into NewBB. The resulting
2075  // phi nodes are trivial since NewBB only has one predecessor, but SSAUpdater
2076  // might need to rewrite the operand of the cloned phi.
2077  for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2078  PHINode *NewPN = PHINode::Create(PN->getType(), 1, PN->getName(), NewBB);
2079  NewPN->addIncoming(PN->getIncomingValueForBlock(PredBB), PredBB);
2080  ValueMapping[PN] = NewPN;
2081  }
2082 
2083  // Clone noalias scope declarations in the threaded block. When threading a
2084  // loop exit, we would otherwise end up with two idential scope declarations
2085  // visible at the same time.
2086  SmallVector<MDNode *> NoAliasScopes;
2087  DenseMap<MDNode *, MDNode *> ClonedScopes;
2088  LLVMContext &Context = PredBB->getContext();
2089  identifyNoAliasScopesToClone(BI, BE, NoAliasScopes);
2090  cloneNoAliasScopes(NoAliasScopes, ClonedScopes, "thread", Context);
2091 
2092  // Clone the non-phi instructions of the source basic block into NewBB,
2093  // keeping track of the mapping and using it to remap operands in the cloned
2094  // instructions.
2095  for (; BI != BE; ++BI) {
2096  Instruction *New = BI->clone();
2097  New->setName(BI->getName());
2098  NewBB->getInstList().push_back(New);
2099  ValueMapping[&*BI] = New;
2100  adaptNoAliasScopes(New, ClonedScopes, Context);
2101 
2102  // Remap operands to patch up intra-block references.
2103  for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2104  if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2105  DenseMap<Instruction *, Value *>::iterator I = ValueMapping.find(Inst);
2106  if (I != ValueMapping.end())
2107  New->setOperand(i, I->second);
2108  }
2109  }
2110 
2111  return ValueMapping;
2112 }
2113 
2114 /// Attempt to thread through two successive basic blocks.
2116  Value *Cond) {
2117  // Consider:
2118  //
2119  // PredBB:
2120  // %var = phi i32* [ null, %bb1 ], [ @a, %bb2 ]
2121  // %tobool = icmp eq i32 %cond, 0
2122  // br i1 %tobool, label %BB, label ...
2123  //
2124  // BB:
2125  // %cmp = icmp eq i32* %var, null
2126  // br i1 %cmp, label ..., label ...
2127  //
2128  // We don't know the value of %var at BB even if we know which incoming edge
2129  // we take to BB. However, once we duplicate PredBB for each of its incoming
2130  // edges (say, PredBB1 and PredBB2), we know the value of %var in each copy of
2131  // PredBB. Then we can thread edges PredBB1->BB and PredBB2->BB through BB.
2132 
2133  // Require that BB end with a Branch for simplicity.
2134  BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2135  if (!CondBr)
2136  return false;
2137 
2138  // BB must have exactly one predecessor.
2139  BasicBlock *PredBB = BB->getSinglePredecessor();
2140  if (!PredBB)
2141  return false;
2142 
2143  // Require that PredBB end with a conditional Branch. If PredBB ends with an
2144  // unconditional branch, we should be merging PredBB and BB instead. For
2145  // simplicity, we don't deal with a switch.
2146  BranchInst *PredBBBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2147  if (!PredBBBranch || PredBBBranch->isUnconditional())
2148  return false;
2149 
2150  // If PredBB has exactly one incoming edge, we don't gain anything by copying
2151  // PredBB.
2152  if (PredBB->getSinglePredecessor())
2153  return false;
2154 
2155  // Don't thread through PredBB if it contains a successor edge to itself, in
2156  // which case we would infinite loop. Suppose we are threading an edge from
2157  // PredPredBB through PredBB and BB to SuccBB with PredBB containing a
2158  // successor edge to itself. If we allowed jump threading in this case, we
2159  // could duplicate PredBB and BB as, say, PredBB.thread and BB.thread. Since
2160  // PredBB.thread has a successor edge to PredBB, we would immediately come up
2161  // with another jump threading opportunity from PredBB.thread through PredBB
2162  // and BB to SuccBB. This jump threading would repeatedly occur. That is, we
2163  // would keep peeling one iteration from PredBB.
2164  if (llvm::is_contained(successors(PredBB), PredBB))
2165  return false;
2166 
2167  // Don't thread across a loop header.
2168  if (LoopHeaders.count(PredBB))
2169  return false;
2170 
2171  // Avoid complication with duplicating EH pads.
2172  if (PredBB->isEHPad())
2173  return false;
2174 
2175  // Find a predecessor that we can thread. For simplicity, we only consider a
2176  // successor edge out of BB to which we thread exactly one incoming edge into
2177  // PredBB.
2178  unsigned ZeroCount = 0;
2179  unsigned OneCount = 0;
2180  BasicBlock *ZeroPred = nullptr;
2181  BasicBlock *OnePred = nullptr;
2182  for (BasicBlock *P : predecessors(PredBB)) {
2183  if (ConstantInt *CI = dyn_cast_or_null<ConstantInt>(
2184  evaluateOnPredecessorEdge(BB, P, Cond))) {
2185  if (CI->isZero()) {
2186  ZeroCount++;
2187  ZeroPred = P;
2188  } else if (CI->isOne()) {
2189  OneCount++;
2190  OnePred = P;
2191  }
2192  }
2193  }
2194 
2195  // Disregard complicated cases where we have to thread multiple edges.
2196  BasicBlock *PredPredBB;
2197  if (ZeroCount == 1) {
2198  PredPredBB = ZeroPred;
2199  } else if (OneCount == 1) {
2200  PredPredBB = OnePred;
2201  } else {
2202  return false;
2203  }
2204 
2205  BasicBlock *SuccBB = CondBr->getSuccessor(PredPredBB == ZeroPred);
2206 
2207  // If threading to the same block as we come from, we would infinite loop.
2208  if (SuccBB == BB) {
2209  LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
2210  << "' - would thread to self!\n");
2211  return false;
2212  }
2213 
2214  // If threading this would thread across a loop header, don't thread the edge.
2215  // See the comments above findLoopHeaders for justifications and caveats.
2216  if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
2217  LLVM_DEBUG({
2218  bool BBIsHeader = LoopHeaders.count(BB);
2219  bool SuccIsHeader = LoopHeaders.count(SuccBB);
2220  dbgs() << " Not threading across "
2221  << (BBIsHeader ? "loop header BB '" : "block BB '")
2222  << BB->getName() << "' to dest "
2223  << (SuccIsHeader ? "loop header BB '" : "block BB '")
2224  << SuccBB->getName()
2225  << "' - it might create an irreducible loop!\n";
2226  });
2227  return false;
2228  }
2229 
2230  // Compute the cost of duplicating BB and PredBB.
2231  unsigned BBCost = getJumpThreadDuplicationCost(
2232  TTI, BB, BB->getTerminator(), BBDupThreshold);
2233  unsigned PredBBCost = getJumpThreadDuplicationCost(
2234  TTI, PredBB, PredBB->getTerminator(), BBDupThreshold);
2235 
2236  // Give up if costs are too high. We need to check BBCost and PredBBCost
2237  // individually before checking their sum because getJumpThreadDuplicationCost
2238  // return (unsigned)~0 for those basic blocks that cannot be duplicated.
2239  if (BBCost > BBDupThreshold || PredBBCost > BBDupThreshold ||
2240  BBCost + PredBBCost > BBDupThreshold) {
2241  LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()
2242  << "' - Cost is too high: " << PredBBCost
2243  << " for PredBB, " << BBCost << "for BB\n");
2244  return false;
2245  }
2246 
2247  // Now we are ready to duplicate PredBB.
2248  threadThroughTwoBasicBlocks(PredPredBB, PredBB, BB, SuccBB);
2249  return true;
2250 }
2251 
2253  BasicBlock *PredBB,
2254  BasicBlock *BB,
2255  BasicBlock *SuccBB) {
2256  LLVM_DEBUG(dbgs() << " Threading through '" << PredBB->getName() << "' and '"
2257  << BB->getName() << "'\n");
2258 
2259  BranchInst *CondBr = cast<BranchInst>(BB->getTerminator());
2260  BranchInst *PredBBBranch = cast<BranchInst>(PredBB->getTerminator());
2261 
2262  BasicBlock *NewBB =
2263  BasicBlock::Create(PredBB->getContext(), PredBB->getName() + ".thread",
2264  PredBB->getParent(), PredBB);
2265  NewBB->moveAfter(PredBB);
2266 
2267  // Set the block frequency of NewBB.
2268  if (HasProfileData) {
2269  auto NewBBFreq = BFI->getBlockFreq(PredPredBB) *
2270  BPI->getEdgeProbability(PredPredBB, PredBB);
2271  BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2272  }
2273 
2274  // We are going to have to map operands from the original BB block to the new
2275  // copy of the block 'NewBB'. If there are PHI nodes in PredBB, evaluate them
2276  // to account for entry from PredPredBB.
2277  DenseMap<Instruction *, Value *> ValueMapping =
2278  cloneInstructions(PredBB->begin(), PredBB->end(), NewBB, PredPredBB);
2279 
2280  // Copy the edge probabilities from PredBB to NewBB.
2281  if (HasProfileData)
2282  BPI->copyEdgeProbabilities(PredBB, NewBB);
2283 
2284  // Update the terminator of PredPredBB to jump to NewBB instead of PredBB.
2285  // This eliminates predecessors from PredPredBB, which requires us to simplify
2286  // any PHI nodes in PredBB.
2287  Instruction *PredPredTerm = PredPredBB->getTerminator();
2288  for (unsigned i = 0, e = PredPredTerm->getNumSuccessors(); i != e; ++i)
2289  if (PredPredTerm->getSuccessor(i) == PredBB) {
2290  PredBB->removePredecessor(PredPredBB, true);
2291  PredPredTerm->setSuccessor(i, NewBB);
2292  }
2293 
2294  addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(0), PredBB, NewBB,
2295  ValueMapping);
2296  addPHINodeEntriesForMappedBlock(PredBBBranch->getSuccessor(1), PredBB, NewBB,
2297  ValueMapping);
2298 
2299  DTU->applyUpdatesPermissive(
2300  {{DominatorTree::Insert, NewBB, CondBr->getSuccessor(0)},
2301  {DominatorTree::Insert, NewBB, CondBr->getSuccessor(1)},
2302  {DominatorTree::Insert, PredPredBB, NewBB},
2303  {DominatorTree::Delete, PredPredBB, PredBB}});
2304 
2305  updateSSA(PredBB, NewBB, ValueMapping);
2306 
2307  // Clean up things like PHI nodes with single operands, dead instructions,
2308  // etc.
2309  SimplifyInstructionsInBlock(NewBB, TLI);
2310  SimplifyInstructionsInBlock(PredBB, TLI);
2311 
2312  SmallVector<BasicBlock *, 1> PredsToFactor;
2313  PredsToFactor.push_back(NewBB);
2314  threadEdge(BB, PredsToFactor, SuccBB);
2315 }
2316 
2317 /// tryThreadEdge - Thread an edge if it's safe and profitable to do so.
2319  BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs,
2320  BasicBlock *SuccBB) {
2321  // If threading to the same block as we come from, we would infinite loop.
2322  if (SuccBB == BB) {
2323  LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
2324  << "' - would thread to self!\n");
2325  return false;
2326  }
2327 
2328  // If threading this would thread across a loop header, don't thread the edge.
2329  // See the comments above findLoopHeaders for justifications and caveats.
2330  if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
2331  LLVM_DEBUG({
2332  bool BBIsHeader = LoopHeaders.count(BB);
2333  bool SuccIsHeader = LoopHeaders.count(SuccBB);
2334  dbgs() << " Not threading across "
2335  << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName()
2336  << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '")
2337  << SuccBB->getName() << "' - it might create an irreducible loop!\n";
2338  });
2339  return false;
2340  }
2341 
2342  unsigned JumpThreadCost = getJumpThreadDuplicationCost(
2343  TTI, BB, BB->getTerminator(), BBDupThreshold);
2344  if (JumpThreadCost > BBDupThreshold) {
2345  LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()
2346  << "' - Cost is too high: " << JumpThreadCost << "\n");
2347  return false;
2348  }
2349 
2350  threadEdge(BB, PredBBs, SuccBB);
2351  return true;
2352 }
2353 
2354 /// threadEdge - We have decided that it is safe and profitable to factor the
2355 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
2356 /// across BB. Transform the IR to reflect this change.
2358  const SmallVectorImpl<BasicBlock *> &PredBBs,
2359  BasicBlock *SuccBB) {
2360  assert(SuccBB != BB && "Don't create an infinite loop");
2361 
2362  assert(!LoopHeaders.count(BB) && !LoopHeaders.count(SuccBB) &&
2363  "Don't thread across loop headers");
2364 
2365  // And finally, do it! Start by factoring the predecessors if needed.
2366  BasicBlock *PredBB;
2367  if (PredBBs.size() == 1)
2368  PredBB = PredBBs[0];
2369  else {
2370  LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()
2371  << " common predecessors.\n");
2372  PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm");
2373  }
2374 
2375  // And finally, do it!
2376  LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName()
2377  << "' to '" << SuccBB->getName()
2378  << ", across block:\n " << *BB << "\n");
2379 
2380  LVI->threadEdge(PredBB, BB, SuccBB);
2381 
2382  BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
2383  BB->getName()+".thread",
2384  BB->getParent(), BB);
2385  NewBB->moveAfter(PredBB);
2386 
2387  // Set the block frequency of NewBB.
2388  if (HasProfileData) {
2389  auto NewBBFreq =
2390  BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
2391  BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2392  }
2393 
2394  // Copy all the instructions from BB to NewBB except the terminator.
2395  DenseMap<Instruction *, Value *> ValueMapping =
2396  cloneInstructions(BB->begin(), std::prev(BB->end()), NewBB, PredBB);
2397 
2398  // We didn't copy the terminator from BB over to NewBB, because there is now
2399  // an unconditional jump to SuccBB. Insert the unconditional jump.
2400  BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
2401  NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
2402 
2403  // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
2404  // PHI nodes for NewBB now.
2405  addPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
2406 
2407  // Update the terminator of PredBB to jump to NewBB instead of BB. This
2408  // eliminates predecessors from BB, which requires us to simplify any PHI
2409  // nodes in BB.
2410  Instruction *PredTerm = PredBB->getTerminator();
2411  for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2412  if (PredTerm->getSuccessor(i) == BB) {
2413  BB->removePredecessor(PredBB, true);
2414  PredTerm->setSuccessor(i, NewBB);
2415  }
2416 
2417  // Enqueue required DT updates.
2418  DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, SuccBB},
2419  {DominatorTree::Insert, PredBB, NewBB},
2420  {DominatorTree::Delete, PredBB, BB}});
2421 
2422  updateSSA(BB, NewBB, ValueMapping);
2423 
2424  // At this point, the IR is fully up to date and consistent. Do a quick scan
2425  // over the new instructions and zap any that are constants or dead. This
2426  // frequently happens because of phi translation.
2427  SimplifyInstructionsInBlock(NewBB, TLI);
2428 
2429  // Update the edge weight from BB to SuccBB, which should be less than before.
2430  updateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
2431 
2432  // Threaded an edge!
2433  ++NumThreads;
2434 }
2435 
2436 /// Create a new basic block that will be the predecessor of BB and successor of
2437 /// all blocks in Preds. When profile data is available, update the frequency of
2438 /// this new block.
2439 BasicBlock *JumpThreadingPass::splitBlockPreds(BasicBlock *BB,
2440  ArrayRef<BasicBlock *> Preds,
2441  const char *Suffix) {
2443 
2444  // Collect the frequencies of all predecessors of BB, which will be used to
2445  // update the edge weight of the result of splitting predecessors.
2447  if (HasProfileData)
2448  for (auto Pred : Preds)
2449  FreqMap.insert(std::make_pair(
2450  Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
2451 
2452  // In the case when BB is a LandingPad block we create 2 new predecessors
2453  // instead of just one.
2454  if (BB->isLandingPad()) {
2455  std::string NewName = std::string(Suffix) + ".split-lp";
2456  SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);
2457  } else {
2458  NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));
2459  }
2460 
2461  std::vector<DominatorTree::UpdateType> Updates;
2462  Updates.reserve((2 * Preds.size()) + NewBBs.size());
2463  for (auto NewBB : NewBBs) {
2464  BlockFrequency NewBBFreq(0);
2465  Updates.push_back({DominatorTree::Insert, NewBB, BB});
2466  for (auto Pred : predecessors(NewBB)) {
2467  Updates.push_back({DominatorTree::Delete, Pred, BB});
2468  Updates.push_back({DominatorTree::Insert, Pred, NewBB});
2469  if (HasProfileData) // Update frequencies between Pred -> NewBB.
2470  NewBBFreq += FreqMap.lookup(Pred);
2471  }
2472  if (HasProfileData) // Apply the summed frequency to NewBB.
2473  BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2474  }
2475 
2476  DTU->applyUpdatesPermissive(Updates);
2477  return NewBBs[0];
2478 }
2479 
2480 bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2481  const Instruction *TI = BB->getTerminator();
2482  assert(TI->getNumSuccessors() > 1 && "not a split");
2483 
2484  MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
2485  if (!WeightsNode)
2486  return false;
2487 
2488  MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
2489  if (MDName->getString() != "branch_weights")
2490  return false;
2491 
2492  // Ensure there are weights for all of the successors. Note that the first
2493  // operand to the metadata node is a name, not a weight.
2494  return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
2495 }
2496 
2497 /// Update the block frequency of BB and branch weight and the metadata on the
2498 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2499 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
2500 void JumpThreadingPass::updateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2501  BasicBlock *BB,
2502  BasicBlock *NewBB,
2503  BasicBlock *SuccBB) {
2504  if (!HasProfileData)
2505  return;
2506 
2507  assert(BFI && BPI && "BFI & BPI should have been created here");
2508 
2509  // As the edge from PredBB to BB is deleted, we have to update the block
2510  // frequency of BB.
2511  auto BBOrigFreq = BFI->getBlockFreq(BB);
2512  auto NewBBFreq = BFI->getBlockFreq(NewBB);
2513  auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
2514  auto BBNewFreq = BBOrigFreq - NewBBFreq;
2515  BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
2516 
2517  // Collect updated outgoing edges' frequencies from BB and use them to update
2518  // edge probabilities.
2519  SmallVector<uint64_t, 4> BBSuccFreq;
2520  for (BasicBlock *Succ : successors(BB)) {
2521  auto SuccFreq = (Succ == SuccBB)
2522  ? BB2SuccBBFreq - NewBBFreq
2523  : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
2524  BBSuccFreq.push_back(SuccFreq.getFrequency());
2525  }
2526 
2527  uint64_t MaxBBSuccFreq =
2528  *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
2529 
2531  if (MaxBBSuccFreq == 0)
2532  BBSuccProbs.assign(BBSuccFreq.size(),
2533  {1, static_cast<unsigned>(BBSuccFreq.size())});
2534  else {
2535  for (uint64_t Freq : BBSuccFreq)
2536  BBSuccProbs.push_back(
2537  BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
2538  // Normalize edge probabilities so that they sum up to one.
2539  BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
2540  BBSuccProbs.end());
2541  }
2542 
2543  // Update edge probabilities in BPI.
2544  BPI->setEdgeProbability(BB, BBSuccProbs);
2545 
2546  // Update the profile metadata as well.
2547  //
2548  // Don't do this if the profile of the transformed blocks was statically
2549  // estimated. (This could occur despite the function having an entry
2550  // frequency in completely cold parts of the CFG.)
2551  //
2552  // In this case we don't want to suggest to subsequent passes that the
2553  // calculated weights are fully consistent. Consider this graph:
2554  //
2555  // check_1
2556  // 50% / |
2557  // eq_1 | 50%
2558  // \ |
2559  // check_2
2560  // 50% / |
2561  // eq_2 | 50%
2562  // \ |
2563  // check_3
2564  // 50% / |
2565  // eq_3 | 50%
2566  // \ |
2567  //
2568  // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2569  // the overall probabilities are inconsistent; the total probability that the
2570  // value is either 1, 2 or 3 is 150%.
2571  //
2572  // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2573  // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2574  // the loop exit edge. Then based solely on static estimation we would assume
2575  // the loop was extremely hot.
2576  //
2577  // FIXME this locally as well so that BPI and BFI are consistent as well. We
2578  // shouldn't make edges extremely likely or unlikely based solely on static
2579  // estimation.
2580  if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
2581  SmallVector<uint32_t, 4> Weights;
2582  for (auto Prob : BBSuccProbs)
2583  Weights.push_back(Prob.getNumerator());
2584 
2585  auto TI = BB->getTerminator();
2586  TI->setMetadata(
2587  LLVMContext::MD_prof,
2588  MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
2589  }
2590 }
2591 
2592 /// duplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2593 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2594 /// If we can duplicate the contents of BB up into PredBB do so now, this
2595 /// improves the odds that the branch will be on an analyzable instruction like
2596 /// a compare.
2598  BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2599  assert(!PredBBs.empty() && "Can't handle an empty set");
2600 
2601  // If BB is a loop header, then duplicating this block outside the loop would
2602  // cause us to transform this into an irreducible loop, don't do this.
2603  // See the comments above findLoopHeaders for justifications and caveats.
2604  if (LoopHeaders.count(BB)) {
2605  LLVM_DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
2606  << "' into predecessor block '" << PredBBs[0]->getName()
2607  << "' - it might create an irreducible loop!\n");
2608  return false;
2609  }
2610 
2611  unsigned DuplicationCost = getJumpThreadDuplicationCost(
2612  TTI, BB, BB->getTerminator(), BBDupThreshold);
2613  if (DuplicationCost > BBDupThreshold) {
2614  LLVM_DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
2615  << "' - Cost is too high: " << DuplicationCost << "\n");
2616  return false;
2617  }
2618 
2619  // And finally, do it! Start by factoring the predecessors if needed.
2620  std::vector<DominatorTree::UpdateType> Updates;
2621  BasicBlock *PredBB;
2622  if (PredBBs.size() == 1)
2623  PredBB = PredBBs[0];
2624  else {
2625  LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()
2626  << " common predecessors.\n");
2627  PredBB = splitBlockPreds(BB, PredBBs, ".thr_comm");
2628  }
2629  Updates.push_back({DominatorTree::Delete, PredBB, BB});
2630 
2631  // Okay, we decided to do this! Clone all the instructions in BB onto the end
2632  // of PredBB.
2633  LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName()
2634  << "' into end of '" << PredBB->getName()
2635  << "' to eliminate branch on phi. Cost: "
2636  << DuplicationCost << " block is:" << *BB << "\n");
2637 
2638  // Unless PredBB ends with an unconditional branch, split the edge so that we
2639  // can just clone the bits from BB into the end of the new PredBB.
2640  BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2641 
2642  if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2643  BasicBlock *OldPredBB = PredBB;
2644  PredBB = SplitEdge(OldPredBB, BB);
2645  Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});
2646  Updates.push_back({DominatorTree::Insert, PredBB, BB});
2647  Updates.push_back({DominatorTree::Delete, OldPredBB, BB});
2648  OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
2649  }
2650 
2651  // We are going to have to map operands from the original BB block into the
2652  // PredBB block. Evaluate PHI nodes in BB.
2653  DenseMap<Instruction*, Value*> ValueMapping;
2654 
2655  BasicBlock::iterator BI = BB->begin();
2656  for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2657  ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2658  // Clone the non-phi instructions of BB into PredBB, keeping track of the
2659  // mapping and using it to remap operands in the cloned instructions.
2660  for (; BI != BB->end(); ++BI) {
2661  Instruction *New = BI->clone();
2662 
2663  // Remap operands to patch up intra-block references.
2664  for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2665  if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2666  DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2667  if (I != ValueMapping.end())
2668  New->setOperand(i, I->second);
2669  }
2670 
2671  // If this instruction can be simplified after the operands are updated,
2672  // just use the simplified value instead. This frequently happens due to
2673  // phi translation.
2674  if (Value *IV = SimplifyInstruction(
2675  New,
2676  {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2677  ValueMapping[&*BI] = IV;
2678  if (!New->mayHaveSideEffects()) {
2679  New->deleteValue();
2680  New = nullptr;
2681  }
2682  } else {
2683  ValueMapping[&*BI] = New;
2684  }
2685  if (New) {
2686  // Otherwise, insert the new instruction into the block.
2687  New->setName(BI->getName());
2688  PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2689  // Update Dominance from simplified New instruction operands.
2690  for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2691  if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))
2692  Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});
2693  }
2694  }
2695 
2696  // Check to see if the targets of the branch had PHI nodes. If so, we need to
2697  // add entries to the PHI nodes for branch from PredBB now.
2698  BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2699  addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2700  ValueMapping);
2701  addPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2702  ValueMapping);
2703 
2704  updateSSA(BB, PredBB, ValueMapping);
2705 
2706  // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2707  // that we nuked.
2708  BB->removePredecessor(PredBB, true);
2709 
2710  // Remove the unconditional branch at the end of the PredBB block.
2711  OldPredBranch->eraseFromParent();
2712  if (HasProfileData)
2713  BPI->copyEdgeProbabilities(BB, PredBB);
2714  DTU->applyUpdatesPermissive(Updates);
2715 
2716  ++NumDupes;
2717  return true;
2718 }
2719 
2720 // Pred is a predecessor of BB with an unconditional branch to BB. SI is
2721 // a Select instruction in Pred. BB has other predecessors and SI is used in
2722 // a PHI node in BB. SI has no other use.
2723 // A new basic block, NewBB, is created and SI is converted to compare and
2724 // conditional branch. SI is erased from parent.
2726  SelectInst *SI, PHINode *SIUse,
2727  unsigned Idx) {
2728  // Expand the select.
2729  //
2730  // Pred --
2731  // | v
2732  // | NewBB
2733  // | |
2734  // |-----
2735  // v
2736  // BB
2737  BranchInst *PredTerm = cast<BranchInst>(Pred->getTerminator());
2738  BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2739  BB->getParent(), BB);
2740  // Move the unconditional branch to NewBB.
2741  PredTerm->removeFromParent();
2742  NewBB->getInstList().insert(NewBB->end(), PredTerm);
2743  // Create a conditional branch and update PHI nodes.
2744  auto *BI = BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2745  BI->applyMergedLocation(PredTerm->getDebugLoc(), SI->getDebugLoc());
2746  SIUse->setIncomingValue(Idx, SI->getFalseValue());
2747  SIUse->addIncoming(SI->getTrueValue(), NewBB);
2748 
2749  // The select is now dead.
2750  SI->eraseFromParent();
2751  DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, BB},
2752  {DominatorTree::Insert, Pred, NewBB}});
2753 
2754  // Update any other PHI nodes in BB.
2755  for (BasicBlock::iterator BI = BB->begin();
2756  PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2757  if (Phi != SIUse)
2758  Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2759 }
2760 
2762  PHINode *CondPHI = dyn_cast<PHINode>(SI->getCondition());
2763 
2764  if (!CondPHI || CondPHI->getParent() != BB)
2765  return false;
2766 
2767  for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) {
2768  BasicBlock *Pred = CondPHI->getIncomingBlock(I);
2769  SelectInst *PredSI = dyn_cast<SelectInst>(CondPHI->getIncomingValue(I));
2770 
2771  // The second and third condition can be potentially relaxed. Currently
2772  // the conditions help to simplify the code and allow us to reuse existing
2773  // code, developed for tryToUnfoldSelect(CmpInst *, BasicBlock *)
2774  if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse())
2775  continue;
2776 
2777  BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2778  if (!PredTerm || !PredTerm->isUnconditional())
2779  continue;
2780 
2781  unfoldSelectInstr(Pred, BB, PredSI, CondPHI, I);
2782  return true;
2783  }
2784  return false;
2785 }
2786 
2787 /// tryToUnfoldSelect - Look for blocks of the form
2788 /// bb1:
2789 /// %a = select
2790 /// br bb2
2791 ///
2792 /// bb2:
2793 /// %p = phi [%a, %bb1] ...
2794 /// %c = icmp %p
2795 /// br i1 %c
2796 ///
2797 /// And expand the select into a branch structure if one of its arms allows %c
2798 /// to be folded. This later enables threading from bb1 over bb2.
2800  BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2801  PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2802  Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2803 
2804  if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2805  CondLHS->getParent() != BB)
2806  return false;
2807 
2808  for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2809  BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2810  SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2811 
2812  // Look if one of the incoming values is a select in the corresponding
2813  // predecessor.
2814  if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2815  continue;
2816 
2817  BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2818  if (!PredTerm || !PredTerm->isUnconditional())
2819  continue;
2820 
2821  // Now check if one of the select values would allow us to constant fold the
2822  // terminator in BB. We don't do the transform if both sides fold, those
2823  // cases will be threaded in any case.
2824  LazyValueInfo::Tristate LHSFolds =
2825  LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2826  CondRHS, Pred, BB, CondCmp);
2827  LazyValueInfo::Tristate RHSFolds =
2828  LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2829  CondRHS, Pred, BB, CondCmp);
2830  if ((LHSFolds != LazyValueInfo::Unknown ||
2831  RHSFolds != LazyValueInfo::Unknown) &&
2832  LHSFolds != RHSFolds) {
2833  unfoldSelectInstr(Pred, BB, SI, CondLHS, I);
2834  return true;
2835  }
2836  }
2837  return false;
2838 }
2839 
2840 /// tryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2841 /// same BB in the form
2842 /// bb:
2843 /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2844 /// %s = select %p, trueval, falseval
2845 ///
2846 /// or
2847 ///
2848 /// bb:
2849 /// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2850 /// %c = cmp %p, 0
2851 /// %s = select %c, trueval, falseval
2852 ///
2853 /// And expand the select into a branch structure. This later enables
2854 /// jump-threading over bb in this pass.
2855 ///
2856 /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2857 /// select if the associated PHI has at least one constant. If the unfolded
2858 /// select is not jump-threaded, it will be folded again in the later
2859 /// optimizations.
2861  // This transform would reduce the quality of msan diagnostics.
2862  // Disable this transform under MemorySanitizer.
2863  if (BB->getParent()->hasFnAttribute(Attribute::SanitizeMemory))
2864  return false;
2865 
2866  // If threading this would thread across a loop header, don't thread the edge.
2867  // See the comments above findLoopHeaders for justifications and caveats.
2868  if (LoopHeaders.count(BB))
2869  return false;
2870 
2871  for (BasicBlock::iterator BI = BB->begin();
2872  PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2873  // Look for a Phi having at least one constant incoming value.
2874  if (llvm::all_of(PN->incoming_values(),
2875  [](Value *V) { return !isa<ConstantInt>(V); }))
2876  continue;
2877 
2878  auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2879  using namespace PatternMatch;
2880 
2881  // Check if SI is in BB and use V as condition.
2882  if (SI->getParent() != BB)
2883  return false;
2884  Value *Cond = SI->getCondition();
2885  bool IsAndOr = match(SI, m_CombineOr(m_LogicalAnd(), m_LogicalOr()));
2886  return Cond && Cond == V && Cond->getType()->isIntegerTy(1) && !IsAndOr;
2887  };
2888 
2889  SelectInst *SI = nullptr;
2890  for (Use &U : PN->uses()) {
2891  if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2892  // Look for a ICmp in BB that compares PN with a constant and is the
2893  // condition of a Select.
2894  if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2895  isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2896  if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2897  if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2898  SI = SelectI;
2899  break;
2900  }
2901  } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2902  // Look for a Select in BB that uses PN as condition.
2903  if (isUnfoldCandidate(SelectI, U.get())) {
2904  SI = SelectI;
2905  break;
2906  }
2907  }
2908  }
2909 
2910  if (!SI)
2911  continue;
2912  // Expand the select.
2913  Value *Cond = SI->getCondition();
2914  if (InsertFreezeWhenUnfoldingSelect &&
2916  &DTU->getDomTree()))
2917  Cond = new FreezeInst(Cond, "cond.fr", SI);
2919  BasicBlock *SplitBB = SI->getParent();
2920  BasicBlock *NewBB = Term->getParent();
2921  PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2922  NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2923  NewPN->addIncoming(SI->getFalseValue(), BB);
2924  SI->replaceAllUsesWith(NewPN);
2925  SI->eraseFromParent();
2926  // NewBB and SplitBB are newly created blocks which require insertion.
2927  std::vector<DominatorTree::UpdateType> Updates;
2928  Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2929  Updates.push_back({DominatorTree::Insert, BB, SplitBB});
2930  Updates.push_back({DominatorTree::Insert, BB, NewBB});
2931  Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});
2932  // BB's successors were moved to SplitBB, update DTU accordingly.
2933  for (auto *Succ : successors(SplitBB)) {
2934  Updates.push_back({DominatorTree::Delete, BB, Succ});
2935  Updates.push_back({DominatorTree::Insert, SplitBB, Succ});
2936  }
2937  DTU->applyUpdatesPermissive(Updates);
2938  return true;
2939  }
2940  return false;
2941 }
2942 
2943 /// Try to propagate a guard from the current BB into one of its predecessors
2944 /// in case if another branch of execution implies that the condition of this
2945 /// guard is always true. Currently we only process the simplest case that
2946 /// looks like:
2947 ///
2948 /// Start:
2949 /// %cond = ...
2950 /// br i1 %cond, label %T1, label %F1
2951 /// T1:
2952 /// br label %Merge
2953 /// F1:
2954 /// br label %Merge
2955 /// Merge:
2956 /// %condGuard = ...
2957 /// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2958 ///
2959 /// And cond either implies condGuard or !condGuard. In this case all the
2960 /// instructions before the guard can be duplicated in both branches, and the
2961 /// guard is then threaded to one of them.
2963  using namespace PatternMatch;
2964 
2965  // We only want to deal with two predecessors.
2966  BasicBlock *Pred1, *Pred2;
2967  auto PI = pred_begin(BB), PE = pred_end(BB);
2968  if (PI == PE)
2969  return false;
2970  Pred1 = *PI++;
2971  if (PI == PE)
2972  return false;
2973  Pred2 = *PI++;
2974  if (PI != PE)
2975  return false;
2976  if (Pred1 == Pred2)
2977  return false;
2978 
2979  // Try to thread one of the guards of the block.
2980  // TODO: Look up deeper than to immediate predecessor?
2981  auto *Parent = Pred1->getSinglePredecessor();
2982  if (!Parent || Parent != Pred2->getSinglePredecessor())
2983  return false;
2984 
2985  if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2986  for (auto &I : *BB)
2987  if (isGuard(&I) && threadGuard(BB, cast<IntrinsicInst>(&I), BI))
2988  return true;
2989 
2990  return false;
2991 }
2992 
2993 /// Try to propagate the guard from BB which is the lower block of a diamond
2994 /// to one of its branches, in case if diamond's condition implies guard's
2995 /// condition.
2997  BranchInst *BI) {
2998  assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?");
2999  assert(BI->isConditional() && "Unconditional branch has 2 successors?");
3000  Value *GuardCond = Guard->getArgOperand(0);
3001  Value *BranchCond = BI->getCondition();
3002  BasicBlock *TrueDest = BI->getSuccessor(0);
3003  BasicBlock *FalseDest = BI->getSuccessor(1);
3004 
3005  auto &DL = BB->getModule()->getDataLayout();
3006  bool TrueDestIsSafe = false;
3007  bool FalseDestIsSafe = false;
3008 
3009  // True dest is safe if BranchCond => GuardCond.
3010  auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
3011  if (Impl && *Impl)
3012  TrueDestIsSafe = true;
3013  else {
3014  // False dest is safe if !BranchCond => GuardCond.
3015  Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);
3016  if (Impl && *Impl)
3017  FalseDestIsSafe = true;
3018  }
3019 
3020  if (!TrueDestIsSafe && !FalseDestIsSafe)
3021  return false;
3022 
3023  BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
3024  BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
3025 
3026  ValueToValueMapTy UnguardedMapping, GuardedMapping;
3027  Instruction *AfterGuard = Guard->getNextNode();
3028  unsigned Cost =
3029  getJumpThreadDuplicationCost(TTI, BB, AfterGuard, BBDupThreshold);
3030  if (Cost > BBDupThreshold)
3031  return false;
3032  // Duplicate all instructions before the guard and the guard itself to the
3033  // branch where implication is not proved.
3035  BB, PredGuardedBlock, AfterGuard, GuardedMapping, *DTU);
3036  assert(GuardedBlock && "Could not create the guarded block?");
3037  // Duplicate all instructions before the guard in the unguarded branch.
3038  // Since we have successfully duplicated the guarded block and this block
3039  // has fewer instructions, we expect it to succeed.
3041  BB, PredUnguardedBlock, Guard, UnguardedMapping, *DTU);
3042  assert(UnguardedBlock && "Could not create the unguarded block?");
3043  LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block "
3044  << GuardedBlock->getName() << "\n");
3045  // Some instructions before the guard may still have uses. For them, we need
3046  // to create Phi nodes merging their copies in both guarded and unguarded
3047  // branches. Those instructions that have no uses can be just removed.
3049  for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
3050  if (!isa<PHINode>(&*BI))
3051  ToRemove.push_back(&*BI);
3052 
3053  Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
3054  assert(InsertionPoint && "Empty block?");
3055  // Substitute with Phis & remove.
3056  for (auto *Inst : reverse(ToRemove)) {
3057  if (!Inst->use_empty()) {
3058  PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
3059  NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
3060  NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
3061  NewPN->insertBefore(InsertionPoint);
3062  Inst->replaceAllUsesWith(NewPN);
3063  }
3064  Inst->eraseFromParent();
3065  }
3066  return true;
3067 }
llvm::Check::Size
@ Size
Definition: FileCheck.h:73
i
i
Definition: README.txt:29
llvm::SSAUpdater::Initialize
void Initialize(Type *Ty, StringRef Name)
Reset this object to get ready for a new set of SSA updates with type 'Ty'.
Definition: SSAUpdater.cpp:53
llvm::PreservedAnalyses
A set of analyses that are preserved following a run of a transformation pass.
Definition: PassManager.h:155
llvm::array_pod_sort
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:1450
llvm::AAManager
A manager for alias analyses.
Definition: AliasAnalysis.h:1288
llvm::BasicBlock::end
iterator end()
Definition: BasicBlock.h:298
llvm::TargetIRAnalysis
Analysis pass providing the TargetTransformInfo.
Definition: TargetTransformInfo.h:2331
llvm::predecessors
pred_range predecessors(BasicBlock *BB)
Definition: CFG.h:127
llvm
This file implements support for optimizing divisions by a constant.
Definition: AllocatorList.h:23
getBestDestForJumpOnUndef
static unsigned getBestDestForJumpOnUndef(BasicBlock *BB)
GetBestDestForBranchOnUndef - If we determine that the specified block ends in an undefined jump,...
Definition: JumpThreading.cpp:994
llvm::SplitLandingPadPredecessors
void SplitLandingPadPredecessors(BasicBlock *OrigBB, ArrayRef< BasicBlock * > Preds, const char *Suffix, const char *Suffix2, SmallVectorImpl< BasicBlock * > &NewBBs, DominatorTree *DT, 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...
Definition: BasicBlockUtils.cpp:1268
llvm::Type::getInt1Ty
static IntegerType * getInt1Ty(LLVMContext &C)
Definition: Type.cpp:238
llvm::Instruction::getModule
const Module * getModule() const
Return the module owning the function this instruction belongs to or nullptr it the function does not...
Definition: Instruction.cpp:66
llvm::ConstantExpr::getNot
static Constant * getNot(Constant *C)
Definition: Constants.cpp:2684
Optional.h
ValueMapper.h
llvm::DataLayout
A parsed version of the target data layout string in and methods for querying it.
Definition: DataLayout.h:113
intptr_t
llvm::Value::hasOneUse
bool hasOneUse() const
Return true if there is exactly one use of this value.
Definition: Value.h:434
Metadata.h
replaceFoldableUses
static void replaceFoldableUses(Instruction *Cond, Value *ToVal)
Definition: JumpThreading.cpp:494
llvm::CmpInst::Predicate
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:720
llvm::JumpThreadingPass::unfoldSelectInstr
void unfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB, SelectInst *SI, PHINode *SIUse, unsigned Idx)
Definition: JumpThreading.cpp:2725
llvm::BasicBlock::iterator
InstListType::iterator iterator
Instruction iterators...
Definition: BasicBlock.h:90
llvm::isSafeToSpeculativelyExecute
bool isSafeToSpeculativelyExecute(const Value *V, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr, const TargetLibraryInfo *TLI=nullptr)
Return true if the instruction does not have any effects besides calculating the result and does not ...
Definition: ValueTracking.cpp:4611
llvm::FindFunctionBackedges
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:34
llvm::ConstantInt::getType
IntegerType * getType() const
getType - Specialize the getType() method to always return an IntegerType, which reduces the amount o...
Definition: Constants.h:173
llvm::BasicBlock::getParent
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:107
IntrinsicInst.h
llvm::AnalysisManager::getResult
PassT::Result & getResult(IRUnitT &IR, ExtraArgTs... ExtraArgs)
Get the result of an analysis pass for a given IR unit.
Definition: PassManager.h:783
llvm::ValueMap::end
iterator end()
Definition: ValueMap.h:136
Scalar.h
llvm::JumpThreadingPass::findLoopHeaders
void findLoopHeaders(Function &F)
findLoopHeaders - We do not want jump threading to turn proper loop structures into irreducible loops...
Definition: JumpThreading.cpp:603
Loads.h
T
llvm::Function
Definition: Function.h:62
llvm::DenseMapBase< DenseMap< KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >, KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >::lookup
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:197
llvm::lower_bound
auto lower_bound(R &&Range, T &&Value)
Provide wrappers to std::lower_bound which take ranges instead of having to pass begin/end explicitly...
Definition: STLExtras.h:1661
llvm::BranchProbability::getNumerator
uint32_t getNumerator() const
Definition: BranchProbability.h:65
P
This currently compiles esp xmm0 movsd esp eax eax esp ret We should use not the dag combiner This is because dagcombine2 needs to be able to see through the X86ISD::Wrapper which DAGCombine can t really do The code for turning x load into a single vector load is target independent and should be moved to the dag combiner The code for turning x load into a vector load can only handle a direct load from a global or a direct load from the stack It should be generalized to handle any load from P
Definition: README-SSE.txt:411
Pass.h
llvm::isImpliedCondition
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.
Definition: ValueTracking.cpp:6679
llvm::JumpThreadingPass::processBranchOnXOR
bool processBranchOnXOR(BinaryOperator *BO)
processBranchOnXOR - We have an otherwise unthreadable conditional branch on a xor instruction in the...
Definition: JumpThreading.cpp:1833
hasAddressTakenAndUsed
static bool hasAddressTakenAndUsed(BasicBlock *BB)
Definition: JumpThreading.cpp:1012
ImplicationSearchThreshold
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)
llvm::isGuaranteedNotToBeUndefOrPoison
bool isGuaranteedNotToBeUndefOrPoison(const Value *V, AssumptionCache *AC=nullptr, const Instruction *CtxI=nullptr, const DominatorTree *DT=nullptr, unsigned Depth=0)
Return true if this function can prove that V does not have undef bits and is never poison.
Definition: ValueTracking.cpp:5271
llvm::ConstantInt::getValue
const APInt & getValue() const
Return the constant as an APInt value reference.
Definition: Constants.h:133
llvm::LazyValueAnalysis
Analysis to compute lazy value information.
Definition: LazyValueInfo.h:131
llvm::ilist_node_with_parent::getNextNode
NodeTy * getNextNode()
Get the next node, or nullptr for the list tail.
Definition: ilist_node.h:288
llvm::SmallVector
This is a 'vector' (really, a variable-sized array), optimized for the case when the array is small.
Definition: SmallVector.h:1168
Statistic.h
llvm::jumpthreading::WantBlockAddress
@ WantBlockAddress
Definition: JumpThreading.h:61
llvm::replaceNonLocalUsesWith
unsigned replaceNonLocalUsesWith(Instruction *From, Value *To)
Definition: Local.cpp:2679
llvm::PatternMatch::m_Add
BinaryOp_match< LHS, RHS, Instruction::Add > m_Add(const LHS &L, const RHS &R)
Definition: PatternMatch.h:1008
llvm::PatternMatch::m_CombineOr
match_combine_or< LTy, RTy > m_CombineOr(const LTy &L, const RTy &R)
Combine two pattern matchers matching L || R.
Definition: PatternMatch.h:210
llvm::TargetTransformInfo
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
Definition: TargetTransformInfo.h:168
ToRemove
ReachingDefAnalysis InstSet & ToRemove
Definition: ARMLowOverheadLoops.cpp:546
llvm::Intrinsic::getName
StringRef getName(ID id)
Return the LLVM name for an intrinsic, such as "llvm.ppc.altivec.lvx".
Definition: Function.cpp:879
MapVector.h
DomTreeUpdater.h
llvm::erase_if
void erase_if(Container &C, UnaryPredicate P)
Provide a container algorithm similar to C++ Library Fundamentals v2's erase_if which is equivalent t...
Definition: STLExtras.h:1732
ValueTracking.h
Local.h
llvm::Instruction::insertBefore
void insertBefore(Instruction *InsertPos)
Insert an unlinked instruction into a basic block immediately before the specified instruction.
Definition: Instruction.cpp:84
llvm::DominatorTree
Concrete subclass of DominatorTreeBase that is used to compute a normal dominator tree.
Definition: Dominators.h:151
GlobalsModRef.h
llvm::cl::Hidden
@ Hidden
Definition: CommandLine.h:143
getJumpThreadDuplicationCost
static unsigned getJumpThreadDuplicationCost(const TargetTransformInfo *TTI, BasicBlock *BB, Instruction *StopAt, unsigned Threshold)
Return the cost of duplicating a piece of this block from first non-phi and before StopAt instruction...
Definition: JumpThreading.cpp:519
isOpDefinedInBlock
static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB)
Return true if Op is an instruction defined in the given block.
Definition: JumpThreading.cpp:1292
llvm::jumpthreading::ConstantPreference
ConstantPreference
Definition: JumpThreading.h:61
llvm::JumpThreadingPass::maybethreadThroughTwoBasicBlocks
bool maybethreadThroughTwoBasicBlocks(BasicBlock *BB, Value *Cond)
Attempt to thread through two successive basic blocks.
Definition: JumpThreading.cpp:2115
llvm::AAMDNodes
A collection of metadata nodes that might be associated with a memory access used by the alias-analys...
Definition: Metadata.h:651
llvm::DenseMapIterator
Definition: DenseMap.h:56
BlockFrequency.h
llvm::Type
The instances of the Type class are immutable: once they are created, they are never changed.
Definition: Type.h:45
DenseMap.h
llvm::JumpThreadingPass::JumpThreadingPass
JumpThreadingPass(bool InsertFreezeWhenUnfoldingSelect=false, int T=-1)
Definition: JumpThreading.cpp:182
Module.h
llvm::reverse
auto reverse(ContainerTy &&C, std::enable_if_t< has_rbegin< ContainerTy >::value > *=nullptr)
Definition: STLExtras.h:333
llvm::DominatorTreeBase< BasicBlock, false >::Insert
static constexpr UpdateKind Insert
Definition: GenericDomTree.h:242
llvm::JumpThreadingPass::evaluateOnPredecessorEdge
Constant * evaluateOnPredecessorEdge(BasicBlock *BB, BasicBlock *PredPredBB, Value *cond)
Definition: JumpThreading.cpp:1579
JumpThreading.h
llvm::Optional< bool >
llvm::BranchProbability::getBranchProbability
static BranchProbability getBranchProbability(uint64_t Numerator, uint64_t Denominator)
Definition: BranchProbability.cpp:53
llvm::BranchInst::getNumSuccessors
unsigned getNumSuccessors() const
Definition: Instructions.h:3159
llvm::SimplifyCmpInst
Value * SimplifyCmpInst(unsigned Predicate, Value *LHS, Value *RHS, const SimplifyQuery &Q)
Given operands for a CmpInst, fold the result or return null.
Definition: InstructionSimplify.cpp:5368
llvm::MapVector
This class implements a map that also provides access to all stored values in a deterministic order.
Definition: MapVector.h:37
llvm::SmallPtrSet
SmallPtrSet - This class implements a set which is optimized for holding SmallSize or less elements.
Definition: SmallPtrSet.h:449
llvm::tgtok::FalseVal
@ FalseVal
Definition: TGLexer.h:61
llvm::JumpThreadingPass::processImpliedCondition
bool processImpliedCondition(BasicBlock *BB)
Definition: JumpThreading.cpp:1249
LazyValueInfo.h
llvm::MipsISD::Ret
@ Ret
Definition: MipsISelLowering.h:116
llvm::BasicBlock::getSinglePredecessor
const BasicBlock * getSinglePredecessor() const
Return the predecessor of this block if it has a single predecessor block.
Definition: BasicBlock.cpp:268
STLExtras.h
llvm::SmallVectorImpl::pop_back_val
LLVM_NODISCARD T pop_back_val()
Definition: SmallVector.h:635
llvm::successors
succ_range successors(Instruction *I)
Definition: CFG.h:262
llvm::detail::DenseSetImpl::insert
std::pair< iterator, bool > insert(const ValueT &V)
Definition: DenseSet.h:206
llvm::LoadInst::getAlign
Align getAlign() const
Return the alignment of the access that is being performed.
Definition: Instructions.h:223
llvm::PHINode::setIncomingValue
void setIncomingValue(unsigned i, Value *V)
Definition: Instructions.h:2732
ConstantFolding.h
llvm::JumpThreadingPass::tryThreadEdge
bool tryThreadEdge(BasicBlock *BB, const SmallVectorImpl< BasicBlock * > &PredBBs, BasicBlock *SuccBB)
tryThreadEdge - Thread an edge if it's safe and profitable to do so.
Definition: JumpThreading.cpp:2318
llvm::Instruction::isExceptionalTerminator
bool isExceptionalTerminator() const
Definition: Instruction.h:170
Use.h
llvm::combineMetadataForCSE
void combineMetadataForCSE(Instruction *K, const Instruction *J, bool DoesKMove)
Combine the metadata of two instructions so that K can replace J.
Definition: Local.cpp:2565
LLVM_DEBUG
#define LLVM_DEBUG(X)
Definition: Debug.h:101
F
#define F(x, y, z)
Definition: MD5.cpp:56
llvm::Instruction::setMetadata
void setMetadata(unsigned KindID, MDNode *Node)
Set the metadata of the specified kind to the specified node.
Definition: Metadata.cpp:1336
llvm::JumpThreadingPass::duplicateCondBranchOnPHIIntoPred
bool duplicateCondBranchOnPHIIntoPred(BasicBlock *BB, const SmallVectorImpl< BasicBlock * > &PredBBs)
duplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch to BB which contains an i1...
Definition: JumpThreading.cpp:2597
llvm::DomTreeUpdater::UpdateStrategy::Lazy
@ Lazy
llvm::BasicBlock
LLVM Basic Block Representation.
Definition: BasicBlock.h:58
getKnownConstant
static Constant * getKnownConstant(Value *Val, ConstantPreference Preference)
getKnownConstant - Helper method to determine if we can thread over a terminator with the given value...
Definition: JumpThreading.cpp:616
llvm::MDNode::getNumOperands
unsigned getNumOperands() const
Return number of MDNode operands.
Definition: Metadata.h:1143
AliasAnalysis.h
llvm::isGuard
bool isGuard(const User *U)
Returns true iff U has semantics of a guard expressed in a form of call of llvm.experimental....
Definition: GuardUtils.cpp:18
Context
LLVMContext & Context
Definition: NVVMIntrRange.cpp:66
llvm::dbgs
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:163
Instruction.h
CommandLine.h
llvm::ConstantInt
This is the shared class of boolean and integer constants.
Definition: Constants.h:79
llvm::JumpThreadingPass::threadEdge
void 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...
Definition: JumpThreading.cpp:2357
llvm::Instruction::getNumSuccessors
unsigned getNumSuccessors() const
Return the number of successors that this instruction has.
Definition: Instruction.cpp:765
llvm::Instruction::getOpcode
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:160
llvm::all_of
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:1551
llvm::BlockFrequencyInfo
BlockFrequencyInfo pass uses BlockFrequencyInfoImpl implementation to estimate IR basic block frequen...
Definition: BlockFrequencyInfo.h:37
llvm::MapVector::begin
iterator begin()
Definition: MapVector.h:69
llvm::SimplifyInstructionsInBlock
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:701
BBDuplicateThreshold
static cl::opt< unsigned > BBDuplicateThreshold("jump-threading-threshold", cl::desc("Max block size to duplicate for jump threading"), cl::init(6), cl::Hidden)
llvm::PassRegistry::getPassRegistry
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
Definition: PassRegistry.cpp:31
Constants.h
llvm::PatternMatch::match
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
llvm::LazyValueInfo::Unknown
@ Unknown
Definition: LazyValueInfo.h:61
isZero
static bool isZero(Value *V, const DataLayout &DL, DominatorTree *DT, AssumptionCache *AC)
Definition: Lint.cpp:519
llvm::PHINode::getIncomingValue
Value * getIncomingValue(unsigned i) const
Return incoming value number x.
Definition: Instructions.h:2729
llvm::AAResults
Definition: AliasAnalysis.h:508
llvm::initializeJumpThreadingPass
void initializeJumpThreadingPass(PassRegistry &)
E
static GCRegistry::Add< CoreCLRGC > E("coreclr", "CoreCLR-compatible GC")
llvm::User
Definition: User.h:44
Intrinsics.h
C
(vector float) vec_cmpeq(*A, *B) C
Definition: README_ALTIVEC.txt:86
InstrTypes.h
llvm::JumpThreadingPass::releaseMemory
void releaseMemory()
Definition: JumpThreading.h:111
llvm::BasicBlock::begin
iterator begin()
Instruction iterator methods.
Definition: BasicBlock.h:296
llvm::TargetTransformInfo::hasBranchDivergence
bool hasBranchDivergence() const
Return true if branch divergence exists.
Definition: TargetTransformInfo.cpp:232
llvm::BranchProbabilityInfo
Analysis providing branch probability information.
Definition: BranchProbabilityInfo.h:115
llvm::MDBuilder::createBranchWeights
MDNode * createBranchWeights(uint32_t TrueWeight, uint32_t FalseWeight)
Return metadata containing two branch weights.
Definition: MDBuilder.cpp:37
llvm::jumpthreading::WantInteger
@ WantInteger
Definition: JumpThreading.h:61
llvm::AnalysisUsage
Represent the analysis usage information of a pass.
Definition: PassAnalysisSupport.h:47
runImpl
static bool runImpl(const TargetLibraryInfo &TLI, Function &F)
Definition: ReplaceWithVeclib.cpp:177
TargetLibraryInfo.h
llvm::Type::isVectorTy
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:226
llvm::Value::uses
iterator_range< use_iterator > uses()
Definition: Value.h:376
DenseSet.h
false
Definition: StackSlotColoring.cpp:142
llvm::LazyValueInfo::True
@ True
Definition: LazyValueInfo.h:61
llvm::PHINode::getIncomingValueForBlock
Value * getIncomingValueForBlock(const BasicBlock *BB) const
Definition: Instructions.h:2818
llvm::PatternMatch::m_ConstantInt
class_match< ConstantInt > m_ConstantInt()
Match an arbitrary ConstantInt and ignore it.
Definition: PatternMatch.h:145
llvm::Instruction
Definition: Instruction.h:45
llvm::SimplifyInstruction
Value * SimplifyInstruction(Instruction *I, const SimplifyQuery &Q, OptimizationRemarkEmitter *ORE=nullptr)
See if we can compute a simplified version of this instruction.
Definition: InstructionSimplify.cpp:6327
llvm::DominatorTreeWrapperPass
Legacy analysis pass which computes a DominatorTree.
Definition: Dominators.h:287
MDBuilder.h
llvm::BasicBlock::phis
iterator_range< const_phi_iterator > phis() const
Returns a range that iterates over the phis in the basic block.
Definition: BasicBlock.h:354
llvm::STATISTIC
STATISTIC(NumFunctions, "Total number of functions")
Threading
jump Jump Threading
Definition: JumpThreading.cpp:175
llvm::SmallVectorImpl::resize
void resize(size_type N)
Definition: SmallVector.h:606
llvm::UndefValue::get
static UndefValue * get(Type *T)
Static factory methods - Return an 'undef' object of the specified type.
Definition: Constants.cpp:1796
llvm::DomTreeUpdater
Definition: DomTreeUpdater.h:28
llvm::LocationSize::precise
static LocationSize precise(uint64_t Value)
Definition: MemoryLocation.h:100
llvm::ConstantInt::get
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:925
threading
jump threading
Definition: JumpThreading.cpp:174
SmallPtrSet.h
llvm::Instruction::getAAMetadata
AAMDNodes getAAMetadata() const
Returns the AA metadata for this instruction.
Definition: Metadata.cpp:1370
llvm::Instruction::getSuccessor
BasicBlock * getSuccessor(unsigned Idx) const
Return the specified successor. This instruction must be a terminator.
Definition: Instruction.cpp:777
llvm::SplitBlockPredecessors
BasicBlock * SplitBlockPredecessors(BasicBlock *BB, ArrayRef< BasicBlock * > Preds, const char *Suffix, DominatorTree *DT, 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...
Definition: BasicBlockUtils.cpp:1148
llvm::Instruction::removeFromParent
void removeFromParent()
This method unlinks 'this' from the containing basic block, but does not delete it.
Definition: Instruction.cpp:74
llvm::ConstantRange::add
ConstantRange add(const ConstantRange &Other) const
Return a new range representing the possible values resulting from an addition of a value in this ran...
Definition: ConstantRange.cpp:907
PatternMatch.h
llvm::RemoveRedundantDbgInstrs
bool RemoveRedundantDbgInstrs(BasicBlock *BB)
Try to remove redundant dbg.value instructions from given basic block.
Definition: BasicBlockUtils.cpp:432
llvm::JumpThreadingPass::tryToUnfoldSelect
bool tryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB)
tryToUnfoldSelect - Look for blocks of the form bb1: a = select br bb2
Definition: JumpThreading.cpp:2799
llvm::PHINode::getNumIncomingValues
unsigned getNumIncomingValues() const
Return the number of incoming edges.
Definition: Instructions.h:2725
llvm::Instruction::extractProfMetadata
bool extractProfMetadata(uint64_t &TrueVal, uint64_t &FalseVal) const
Retrieve the raw weight values of a conditional branch or select.
Definition: Metadata.cpp:1405
llvm::JumpThreadingPass::tryToUnfoldSelectInCurrBB
bool tryToUnfoldSelectInCurrBB(BasicBlock *BB)
tryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the same BB in the form bb: p = ...
Definition: JumpThreading.cpp:2860
llvm::LoadInst::getSyncScopeID
SyncScope::ID getSyncScopeID() const
Returns the synchronization scope ID of this load instruction.
Definition: Instructions.h:242
llvm::Value::use_empty
bool use_empty() const
Definition: Value.h:344
Type.h
BranchProbability.h
llvm::Instruction::getMetadata
MDNode * getMetadata(unsigned KindID) const
Get the metadata of given kind attached to this Instruction.
Definition: Instruction.h:282
INITIALIZE_PASS_END
#define INITIALIZE_PASS_END(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:58
CFG.h
LoopInfo.h
llvm::BranchInst::getCondition
Value * getCondition() const
Definition: Instructions.h:3149
jump
The object format emitted by the WebAssembly backed is documented see the home and packaging for producing WebAssembly applications that can run in browsers and other environments wasi sdk provides a more minimal C C SDK based on llvm and a libc based on for producing WebAssemmbly applictions that use the WASI ABI Rust provides WebAssembly support integrated into Cargo There are two main which provides a relatively minimal environment that has an emphasis on being native wasm32 unknown which uses Emscripten internally and provides standard C C filesystem GL and SDL bindings For more and br_table instructions can support having a value on the value stack across the jump(sometimes). We should(a) model this
llvm::findAvailablePtrLoadStore
Value * findAvailablePtrLoadStore(const MemoryLocation &Loc, Type *AccessTy, bool AtLeastAtomic, BasicBlock *ScanBB, BasicBlock::iterator &ScanFrom, unsigned MaxInstsToScan, AAResults *AA, bool *IsLoadCSE, unsigned *NumScanedInst)
Scan backwards to see if we have the value of the given pointer available locally within a small numb...
Definition: Loads.cpp:524
llvm::CmpInst
This class is the base class for the comparison instructions.
Definition: InstrTypes.h:710
PB
PassBuilder PB(Machine, PassOpts->PTO, None, &PIC)
llvm::Type::isIntegerTy
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:190
llvm::MDNode::getOperand
const MDOperand & getOperand(unsigned I) const
Definition: Metadata.h:1137
llvm::DenseSet< Value * >
BasicBlock.h
llvm::cl::opt
Definition: CommandLine.h:1432
llvm::DuplicateInstructionsInSplitBetween
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...
Definition: CloneFunction.cpp:887
llvm::TryToSimplifyUncondBranchFromEmptyBlock
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:1043
llvm::ConstantExpr::getCompare
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:2398
llvm::ConstantRange::inverse
ConstantRange inverse() const
Return a new range that is the logical not of the current set.
Definition: ConstantRange.cpp:1549
llvm::BlockFrequency
Definition: BlockFrequency.h:24
llvm::Constant
This is an important base class in LLVM.
Definition: Constant.h:41
llvm::Instruction::eraseFromParent
SymbolTableList< Instruction >::iterator eraseFromParent()
This method unlinks 'this' from the containing basic block and deletes it.
Definition: Instruction.cpp:78
BranchProbabilityInfo.h
llvm::ICmpInst
This instruction compares its operands according to the predicate given to the constructor.
Definition: Instructions.h:1203
llvm::TargetLibraryInfoWrapperPass
Definition: TargetLibraryInfo.h:465
uint64_t
llvm::TargetTransformInfoWrapperPass
Wrapper pass for TargetTransformInfo.
Definition: TargetTransformInfo.h:2387
llvm::PatternMatch::m_LogicalOr
LogicalOp_match< LHS, RHS, Instruction::Or > m_LogicalOr(const LHS &L, const RHS &R)
Matches L || R either in the form of L | R or L ? true : R.
Definition: PatternMatch.h:2522
llvm::LazyValueInfoWrapperPass
Wrapper around LazyValueInfo.
Definition: LazyValueInfo.h:142
llvm::JumpThreadingPass::simplifyPartiallyRedundantLoad
bool simplifyPartiallyRedundantLoad(LoadInst *LI)
simplifyPartiallyRedundantLoad - If LoadI is an obviously partially redundant load instruction,...
Definition: JumpThreading.cpp:1303
llvm::PreservedAnalyses::preserve
void preserve()
Mark an analysis as preserved.
Definition: PassManager.h:176
INITIALIZE_PASS_DEPENDENCY
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
llvm::PHINode::addIncoming
void addIncoming(Value *V, BasicBlock *BB)
Add an incoming value to the end of the PHI list.
Definition: Instructions.h:2783
findMostPopularDest
static BasicBlock * findMostPopularDest(BasicBlock *BB, const SmallVectorImpl< std::pair< BasicBlock *, BasicBlock * >> &PredToDestList)
findMostPopularDest - The specified list contains multiple possible threadable destinations.
Definition: JumpThreading.cpp:1544
move
compiles ldr LCPI1_0 ldr ldr mov lsr tst moveq r1 ldr LCPI1_1 and r0 bx lr It would be better to do something like to fold the shift into the conditional move
Definition: README.txt:546
llvm::LLVMContext
This is an important class for using LLVM in a threaded context.
Definition: LLVMContext.h:68
llvm::numbers::e
constexpr double e
Definition: MathExtras.h:57
llvm::BranchInst::Create
static BranchInst * Create(BasicBlock *IfTrue, Instruction *InsertBefore=nullptr)
Definition: Instructions.h:3124
MemoryLocation.h
llvm::DenseMap
Definition: DenseMap.h:714
llvm::ConstantExpr::get
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:2267
I
#define I(x, y, z)
Definition: MD5.cpp:59
Cloning.h
llvm::UndefValue
'undef' values are things that do not have specified contents.
Definition: Constants.h:1348
llvm::cl::init
initializer< Ty > init(const Ty &Val)
Definition: CommandLine.h:441
llvm::adaptNoAliasScopes
void adaptNoAliasScopes(llvm::Instruction *I, const DenseMap< MDNode *, MDNode * > &ClonedScopes, LLVMContext &Context)
Adapt the metadata for the specified instruction according to the provided mapping.
Definition: CloneFunction.cpp:957
llvm::is_contained
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:1616
llvm::Instruction::setDebugLoc
void setDebugLoc(DebugLoc Loc)
Set the debug location information for this instruction.
Definition: Instruction.h:367
llvm::DenseMapBase< DenseMap< KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >, KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >::find
iterator find(const_arg_type_t< KeyT > Val)
Definition: DenseMap.h:150
assert
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
SI
StandardInstrumentations SI(Debug, VerifyEach)
llvm::JumpThreadingPass
This pass performs 'jump threading', which looks at blocks that have multiple predecessors and multip...
Definition: JumpThreading.h:80
llvm::ConstantExpr::getCast
static Constant * getCast(unsigned ops, Constant *C, Type *Ty, bool OnlyIfReduced=false)
Convenience function for getting a Cast operation.
Definition: Constants.cpp:1988
llvm::DominatorTree::isReachableFromEntry
bool isReachableFromEntry(const Use &U) const
Provide an overload for a Use.
Definition: Dominators.cpp:328
llvm::SelectInst
This class represents the LLVM 'select' instruction.
Definition: Instructions.h:1738
llvm::cloneNoAliasScopes
void cloneNoAliasScopes(ArrayRef< MDNode * > NoAliasDeclScopes, DenseMap< MDNode *, MDNode * > &ClonedScopes, StringRef Ext, LLVMContext &Context)
Duplicate the specified list of noalias decl scopes.
Definition: CloneFunction.cpp:932
llvm::LazyValueInfo
This pass computes, caches, and vends lazy value constraint information.
Definition: LazyValueInfo.h:31
llvm::MDNode
Metadata node.
Definition: Metadata.h:906
llvm::SmallPtrSetImpl::count
size_type count(ConstPtrType Ptr) const
count - Return 1 if the specified pointer is in the set, 0 otherwise.
Definition: SmallPtrSet.h:382
llvm::BranchInst::isUnconditional
bool isUnconditional() const
Definition: Instructions.h:3146
llvm::PatternMatch::m_Value
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:76
llvm::User::setOperand
void setOperand(unsigned i, Value *Val)
Definition: User.h:174
llvm::JumpThreadingPass::processGuards
bool processGuards(BasicBlock *BB)
Try to propagate a guard from the current BB into one of its predecessors in case if another branch o...
Definition: JumpThreading.cpp:2962
llvm::DominatorTreeBase::reset
void reset()
Definition: GenericDomTree.h:806
CFG.h
llvm::TargetTransformInfo::TCC_Free
@ TCC_Free
Expected to fold away in lowering.
Definition: TargetTransformInfo.h:262
llvm::BlockAddress
The address of a basic block.
Definition: Constants.h:848
llvm::Sched::Source
@ Source
Definition: TargetLowering.h:100
llvm::ArrayRef
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory),...
Definition: APInt.h:32
llvm::isInstructionTriviallyDead
bool isInstructionTriviallyDead(Instruction *I, const TargetLibraryInfo *TLI=nullptr)
Return true if the result produced by the instruction is not used, and the instruction will return.
Definition: Local.cpp:398
llvm::LoopInfo
Definition: LoopInfo.h:1083
llvm::Constant::removeDeadConstantUsers
void removeDeadConstantUsers() const
If there are any dead constant users dangling off of this constant, remove them.
Definition: Constants.cpp:752
llvm::BinaryOperator
Definition: InstrTypes.h:189
llvm::any_of
bool any_of(R &&range, UnaryPredicate P)
Provide wrappers to std::any_of which take ranges instead of having to pass begin/end explicitly.
Definition: STLExtras.h:1558
llvm::BasicBlock::moveAfter
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:142
DataLayout.h
llvm::JumpThreadingPass::maybeMergeBasicBlockIntoOnlyPred
bool maybeMergeBasicBlockIntoOnlyPred(BasicBlock *BB)
Merge basic block BB into its sole predecessor if possible.
Definition: JumpThreading.cpp:1969
llvm::JumpThreadingPass::threadGuard
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,...
Definition: JumpThreading.cpp:2996
Cond
SmallVector< MachineOperand, 4 > Cond
Definition: BasicBlockSections.cpp:179
llvm::identifyNoAliasScopesToClone
void identifyNoAliasScopesToClone(ArrayRef< BasicBlock * > BBs, SmallVectorImpl< MDNode * > &NoAliasDeclScopes)
Find the 'llvm.experimental.noalias.scope.decl' intrinsics in the specified basic blocks and extract ...
Definition: CloneFunction.cpp:1028
llvm::Instruction::setSuccessor
void setSuccessor(unsigned Idx, BasicBlock *BB)
Update the specified successor to point at the provided block.
Definition: Instruction.cpp:789
llvm::TargetTransformInfo::TCK_SizeAndLatency
@ TCK_SizeAndLatency
The weighted sum of size and latency.
Definition: TargetTransformInfo.h:215
llvm::Value::getType
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:255
llvm::pred_size
unsigned pred_size(const BasicBlock *BB)
Get the number of predecessors of BB.
Definition: CFG.h:124
llvm::ConstantInt::isZero
bool isZero() const
This is just a convenience method to make client code smaller for a common code.
Definition: Constants.h:194
llvm::TargetTransformInfo::getUserCost
InstructionCost getUserCost(const User *U, ArrayRef< const Value * > Operands, TargetCostKind CostKind) const
Estimate the cost of a given IR user when lowered.
Definition: TargetTransformInfo.cpp:219
llvm::JumpThreadingPass::threadThroughTwoBasicBlocks
void threadThroughTwoBasicBlocks(BasicBlock *PredPredBB, BasicBlock *PredBB, BasicBlock *BB, BasicBlock *SuccBB)
Definition: JumpThreading.cpp:2252
llvm::AnalysisUsage::addPreserved
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
Definition: PassAnalysisSupport.h:98
llvm::Value::replaceAllUsesWith
void replaceAllUsesWith(Value *V)
Change all uses of this to point to a new Value.
Definition: Value.cpp:532
llvm::BasicBlock::Create
static BasicBlock * Create(LLVMContext &Context, const Twine &Name="", Function *Parent=nullptr, BasicBlock *InsertBefore=nullptr)
Creates a new BasicBlock.
Definition: BasicBlock.h:100
getParent
static const Function * getParent(const Value *V)
Definition: BasicAliasAnalysis.cpp:890
llvm::BranchProbability
Definition: BranchProbability.h:30
llvm::Value::getContext
LLVMContext & getContext() const
All values hold a context through their type.
Definition: Value.cpp:990
llvm::ilist_node_impl::getIterator
self_iterator getIterator()
Definition: ilist_node.h:81
DL
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
Definition: AArch64SLSHardening.cpp:76
ConstantRange.h
llvm::pred_empty
bool pred_empty(const BasicBlock *BB)
Definition: CFG.h:119
llvm::CastInst
This is the base class for all instructions that perform data casts.
Definition: InstrTypes.h:430
updatePredecessorProfileMetadata
static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB)
Definition: JumpThreading.cpp:222
SSAUpdater.h
BlockFrequencyInfo.h
llvm::AMDGPUISD::BFI
@ BFI
Definition: AMDGPUISelLowering.h:421
llvm::LoadInst::getOrdering
AtomicOrdering getOrdering() const
Returns the ordering constraint of this load instruction.
Definition: Instructions.h:232
llvm::PredIterator
Definition: CFG.h:43
llvm::Value::getName
StringRef getName() const
Return a constant reference to the value's name.
Definition: Value.cpp:309
llvm::ValueMap
See the file comment.
Definition: ValueMap.h:85
llvm::LoadInst
An instruction for reading from memory.
Definition: Instructions.h:175
llvm::DenseMapBase< DenseMap< KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >, KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >::insert
std::pair< iterator, bool > insert(const std::pair< KeyT, ValueT > &KV)
Definition: DenseMap.h:207
llvm::BasicBlock::getTerminator
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:152
llvm::Value::stripPointerCasts
const Value * stripPointerCasts() const
Strip off pointer casts, all-zero GEPs and address space casts.
Definition: Value.cpp:687
llvm::ConstantInt::getFalse
static ConstantInt * getFalse(LLVMContext &Context)
Definition: Constants.cpp:880
llvm::SmallPtrSetImplBase::size
size_type size() const
Definition: SmallPtrSet.h:92
llvm::MapVector::end
iterator end()
Definition: MapVector.h:71
llvm::BasicBlock::front
const Instruction & front() const
Definition: BasicBlock.h:308
llvm::BasicBlock::getContext
LLVMContext & getContext() const
Get the context in which this basic block lives.
Definition: BasicBlock.cpp:36
runOnFunction
static bool runOnFunction(Function &F, bool PostInlining)
Definition: EntryExitInstrumenter.cpp:69
addPHINodeEntriesForMappedBlock
static void addPHINodeEntriesForMappedBlock(BasicBlock *PHIBB, BasicBlock *OldPred, BasicBlock *NewPred, DenseMap< Instruction *, Value * > &ValueMap)
addPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new predecessor to the PHIBB block.
Definition: JumpThreading.cpp:1948
llvm::Instruction::isAtomic
bool isAtomic() const
Return true if this instruction has an AtomicOrdering of unordered or higher.
Definition: Instruction.cpp:604
Constant.h
llvm::SplitEdge
BasicBlock * SplitEdge(BasicBlock *From, BasicBlock *To, DominatorTree *DT=nullptr, LoopInfo *LI=nullptr, MemorySSAUpdater *MSSAU=nullptr, const Twine &BBName="")
Split the edge connecting the specified blocks, and return the newly created basic block between From...
Definition: BasicBlockUtils.cpp:493
llvm::ConstantFoldTerminator
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:128
llvm::ConstantInt::getTrue
static ConstantInt * getTrue(LLVMContext &Context)
Definition: Constants.cpp:873
PrintLVIAfterJumpThreading
static cl::opt< bool > PrintLVIAfterJumpThreading("print-lvi-after-jump-threading", cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false), cl::Hidden)
llvm::JumpThreadingPass::processThreadableEdges
bool processThreadableEdges(Value *Cond, BasicBlock *BB, jumpthreading::ConstantPreference Preference, Instruction *CxtI=nullptr)
Definition: JumpThreading.cpp:1619
llvm::CastInst::CreateBitOrPointerCast
static CastInst * CreateBitOrPointerCast(Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Create a BitCast, a PtrToInt, or an IntToPTr cast instruction.
Definition: Instructions.cpp:3231
llvm::DenseMapBase< DenseMap< KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >, KeyT, ValueT, DenseMapInfo< KeyT >, llvm::detail::DenseMapPair< KeyT, ValueT > >::end
iterator end()
Definition: DenseMap.h:83
llvm::SmallVectorImpl::assign
void assign(size_type NumElts, ValueParamT Elt)
Definition: SmallVector.h:669
llvm::AMDGPU::SendMsg::Op
Op
Definition: SIDefines.h:324
llvm::PreservedAnalyses::all
static PreservedAnalyses all()
Construct a special preserved set that preserves all passes.
Definition: PassManager.h:161
llvm::PHINode::Create
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...
Definition: Instructions.h:2675
llvm::pred_end
Interval::pred_iterator pred_end(Interval *I)
Definition: Interval.h:112
llvm::isGuaranteedToTransferExecutionToSuccessor
bool isGuaranteedToTransferExecutionToSuccessor(const Instruction *I)
Return true if this function can prove that the instruction I will always transfer execution to one o...
Definition: ValueTracking.cpp:5302
Casting.h
llvm::Instruction::setAAMetadata
void setAAMetadata(const AAMDNodes &N)
Sets the AA metadata on this instruction from the AAMDNodes structure.
Definition: Metadata.cpp:1379
Function.h
llvm::FindAvailableLoadedValue
Value * FindAvailableLoadedValue(LoadInst *Load, BasicBlock *ScanBB, BasicBlock::iterator &ScanFrom, unsigned MaxInstsToScan=DefMaxInstsToScan, AAResults *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:431
JumpThreadingFreezeSelectCond
static cl::opt< bool > JumpThreadingFreezeSelectCond("jump-threading-freeze-select-cond", cl::desc("Freeze the condition when unfolding select"), cl::init(false), cl::Hidden)
PassManager.h
llvm::TargetLibraryInfo
Provides information about what library functions are available for the current target.
Definition: TargetLibraryInfo.h:221
llvm::ValueMap::find
iterator find(const KeyT &Val)
Definition: ValueMap.h:156
llvm::JumpThreadingPass::runImpl
bool runImpl(Function &F, TargetLibraryInfo *TLI, TargetTransformInfo *TTI, LazyValueInfo *LVI, AAResults *AA, DomTreeUpdater *DTU, bool HasProfileData, std::unique_ptr< BlockFrequencyInfo > BFI, std::unique_ptr< BranchProbabilityInfo > BPI)
Definition: JumpThreading.cpp:379
llvm::ConstantRange::contains
bool contains(const APInt &Val) const
Return true if the specified value is in the set.
Definition: ConstantRange.cpp:393
INITIALIZE_PASS_BEGIN
INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading", "Jump Threading", false, false) INITIALIZE_PASS_END(JumpThreading
llvm::IndirectBrInst
Indirect Branch Instruction.
Definition: Instructions.h:3614
llvm::BranchProbability::getCompl
BranchProbability getCompl() const
Definition: BranchProbability.h:69
llvm::ConstantRange
This class represents a range of values.
Definition: ConstantRange.h:47
llvm::IntrinsicInst
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:45
GuardUtils.h
llvm::MDBuilder
Definition: MDBuilder.h:35
llvm::JumpThreadingPass::updateSSA
void updateSSA(BasicBlock *BB, BasicBlock *NewBB, DenseMap< Instruction *, Value * > &ValueMapping)
Update the SSA form.
Definition: JumpThreading.cpp:2020
llvm::pred_begin
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:109
llvm::DominatorTreeAnalysis
Analysis pass which computes a DominatorTree.
Definition: Dominators.h:252
llvm::BasicBlock::getInstList
const InstListType & getInstList() const
Return the underlying instruction list container.
Definition: BasicBlock.h:363
Instructions.h
llvm::JumpThreadingPass::computeValueKnownInPredecessorsImpl
bool computeValueKnownInPredecessorsImpl(Value *V, BasicBlock *BB, jumpthreading::PredValueInfo &Result, jumpthreading::ConstantPreference Preference, DenseSet< Value * > &RecursionSet, Instruction *CxtI=nullptr)
computeValueKnownInPredecessors - Given a basic block BB and a value V, see if we can infer that the ...
Definition: JumpThreading.cpp:636
llvm::M68kBeads::Term
@ Term
Definition: M68kBaseInfo.h:71
llvm::BranchProbability::normalizeProbabilities
static void normalizeProbabilities(ProbabilityIter Begin, ProbabilityIter End)
Definition: BranchProbability.h:205
SmallVector.h
llvm::LazyValueInfo::Tristate
Tristate
This is used to return true/false/dunno results.
Definition: LazyValueInfo.h:60
llvm::Instruction::getDebugLoc
const DebugLoc & getDebugLoc() const
Return the debug location for this node as a DebugLoc.
Definition: Instruction.h:370
User.h
llvm::Value::DoPHITranslation
const Value * DoPHITranslation(const BasicBlock *CurBB, const BasicBlock *PredBB) const
Translate PHI node to its predecessor from the given basic block.
Definition: Value.cpp:982
Dominators.h
llvm::CallBase::getArgOperand
Value * getArgOperand(unsigned i) const
Definition: InstrTypes.h:1328
llvm::FreezeInst
This class represents a freeze function that returns random concrete value if an operand is either a ...
Definition: Instructions.h:5358
llvm::AAResultsWrapperPass
A wrapper pass to provide the legacy pass manager access to a suitably prepared AAResults object.
Definition: AliasAnalysis.h:1336
llvm::SSAUpdater::RewriteUse
void RewriteUse(Use &U)
Rewrite a use of the symbolic value.
Definition: SSAUpdater.cpp:187
llvm::Instruction::getParent
const BasicBlock * getParent() const
Definition: Instruction.h:94
InstructionSimplify.h
llvm::CmpInst::getPredicate
Predicate getPredicate() const
Return the predicate for this instruction.
Definition: InstrTypes.h:796
llvm::BlockAddress::get
static BlockAddress * get(Function *F, BasicBlock *BB)
Return a BlockAddress for the specified function and basic block.
Definition: Constants.cpp:1834
llvm::GlobalsAAWrapperPass
Legacy wrapper pass to provide the GlobalsAAResult object.
Definition: GlobalsModRef.h:143
llvm::PHINode::getIncomingBlock
BasicBlock * getIncomingBlock(unsigned i) const
Return incoming basic block number i.
Definition: Instructions.h:2749
TargetTransformInfo.h
llvm::PHINode
Definition: Instructions.h:2633
Threshold
static cl::opt< unsigned > Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), cl::init(100), cl::Hidden)
llvm::BasicBlock::removePredecessor
void removePredecessor(BasicBlock *Pred, bool KeepOneInputPHIs=false)
Update PHI nodes in this BasicBlock before removal of predecessor Pred.
Definition: BasicBlock.cpp:325
llvm::SmallVectorImpl
This class consists of common code factored out of the SmallVector class to reduce code duplication b...
Definition: APFloat.h:43
llvm::ConstantRange::makeExactICmpRegion
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...
Definition: ConstantRange.cpp:138
llvm::Module::getDataLayout
const DataLayout & getDataLayout() const
Get the data layout for the module's target platform.
Definition: Module.cpp:401
llvm::DeleteDeadBlock
void DeleteDeadBlock(BasicBlock *BB, DomTreeUpdater *DTU=nullptr, bool KeepOneInputPHIs=false)
Delete the specified block, which must have no predecessors.
Definition: BasicBlockUtils.cpp:89
llvm::AnalysisManager
A container for analyses that lazily runs them and caches their results.
Definition: InstructionSimplify.h:44
ThreadAcrossLoopHeaders
static cl::opt< bool > ThreadAcrossLoopHeaders("jump-threading-across-loop-headers", cl::desc("Allow JumpThreading to thread across loop headers, for testing"), cl::init(false), cl::Hidden)
llvm::FunctionPass
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:298
llvm::CallInst
This class represents a function call, abstracting a target machine's calling convention.
Definition: Instructions.h:1475
llvm::ConstantInt::isOne
bool isOne() const
This is just a convenience method to make client code smaller for a common case.
Definition: Constants.h:200
BB
Common register allocation spilling lr str ldr sxth r3 ldr mla r4 can lr mov lr str ldr sxth r3 mla r4 and then merge mul and lr str ldr sxth r3 mla r4 It also increase the likelihood the store may become dead bb27 Successors according to LLVM BB
Definition: README.txt:39
llvm::LoadInst::isUnordered
bool isUnordered() const
Definition: Instructions.h:261
llvm::MergeBasicBlockIntoOnlyPred
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:741
llvm::AnalysisUsage::addRequired
AnalysisUsage & addRequired()
Definition: PassAnalysisSupport.h:75
LLVMContext.h
llvm::SwitchInst
Multiway switch.
Definition: Instructions.h:3212
llvm::Value::takeName
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:382
llvm::SplitBlockAndInsertIfThen
Instruction * SplitBlockAndInsertIfThen(Value *Cond, Instruction *SplitBefore, bool Unreachable, MDNode *BranchWeights, DominatorTree *DT, LoopInfo *LI=nullptr, BasicBlock *ThenBlock=nullptr)
Split the containing block at the specified instruction - everything before SplitBefore stays in the ...
Definition: BasicBlockUtils.cpp:1418
llvm::Sched::Preference
Preference
Definition: TargetLowering.h:98
llvm::SSAUpdater::AddAvailableValue
void AddAvailableValue(BasicBlock *BB, Value *V)
Indicate that a rewritten value is available in the specified block with the specified value.
Definition: SSAUpdater.cpp:70
llvm::User::getOperand
Value * getOperand(unsigned i) const
Definition: User.h:169
llvm::cl::desc
Definition: CommandLine.h:412
llvm::MDString::getString
StringRef getString() const
Definition: Metadata.cpp:483
llvm::BranchInst
Conditional or Unconditional Branch instruction.
Definition: Instructions.h:3068
llvm::ConstantFoldInstruction
Constant * ConstantFoldInstruction(Instruction *I, const DataLayout &DL, const TargetLibraryInfo *TLI=nullptr)
ConstantFoldInstruction - Try to constant fold the specified instruction.
Definition: ConstantFolding.cpp:1093
raw_ostream.h
llvm::SSAUpdater
Helper class for SSA formation on a set of values defined in multiple blocks.
Definition: SSAUpdater.h:38
llvm::JumpThreadingPass::cloneInstructions
DenseMap< Instruction *, Value * > cloneInstructions(BasicBlock::iterator BI, BasicBlock::iterator BE, BasicBlock *NewBB, BasicBlock *PredBB)
Clone instructions in range [BI, BE) to NewBB.
Definition: JumpThreading.cpp:2066
llvm::SmallVectorImpl::reserve
void reserve(size_type N)
Definition: SmallVector.h:624
BasicBlockUtils.h
llvm::JumpThreadingPass::run
PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM)
Definition: JumpThreading.cpp:343
llvm::MDString
A single uniqued string.
Definition: Metadata.h:611
llvm::tgtok::TrueVal
@ TrueVal
Definition: TGLexer.h:61
Value.h
InitializePasses.h
llvm::BasicBlock::isEHPad
bool isEHPad() const
Return true if this basic block is an exception handling block.
Definition: BasicBlock.h:465
llvm::JumpThreadingPass::processBlock
bool processBlock(BasicBlock *BB)
processBlock - If there are any predecessors whose control can be threaded through to a successor,...
Definition: JumpThreading.cpp:1024
llvm::MCID::Terminator
@ Terminator
Definition: MCInstrDesc.h:155
llvm::Value
LLVM Value Representation.
Definition: Value.h:74
Debug.h
llvm::TargetLibraryAnalysis
Analysis pass providing the TargetLibraryInfo.
Definition: TargetLibraryInfo.h:440
llvm::JumpThreadingPass::processBranchOnPHI
bool processBranchOnPHI(PHINode *PN)
processBranchOnPHI - We have an otherwise unthreadable conditional branch on a PHI node (or freeze PH...
Definition: JumpThreading.cpp:1801
llvm::BranchInst::isConditional
bool isConditional() const
Definition: Instructions.h:3147
llvm::PatternMatch::m_LogicalAnd
LogicalOp_match< LHS, RHS, Instruction::And > m_LogicalAnd(const LHS &L, const RHS &R)
Matches L && R either in the form of L & R or L ? R : false.
Definition: PatternMatch.h:2504
llvm::BranchInst::getSuccessor
BasicBlock * getSuccessor(unsigned i) const
Definition: Instructions.h:3161
llvm::MemoryLocation
Representation for a specific memory location.
Definition: MemoryLocation.h:209
llvm::DominatorTreeBase< BasicBlock, false >::Delete
static constexpr UpdateKind Delete
Definition: GenericDomTree.h:243
llvm::BasicBlock::const_iterator
InstListType::const_iterator const_iterator
Definition: BasicBlock.h:91
llvm::createJumpThreadingPass
FunctionPass * createJumpThreadingPass(bool FreezeSelectCond=false, int Threshold=-1)
Definition: JumpThreading.cpp:178
llvm::DefMaxInstsToScan
cl::opt< unsigned > DefMaxInstsToScan
The default number of maximum instructions to scan in the block, used by FindAvailableLoadedValue().
llvm::Use
A Use represents the edge between a Value definition and its users.
Definition: Use.h:44
llvm::SmallVectorImpl::emplace_back
reference emplace_back(ArgTypes &&... Args)
Definition: SmallVector.h:908
llvm::SmallPtrSetImpl::insert
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:364
llvm::Intrinsic::ID
unsigned ID
Definition: TargetTransformInfo.h:37