[InstCombine] Signed saturation patterns
[llvm-complete.git] / lib / Transforms / Scalar / JumpThreading.cpp
blob0cf00baaa24a01397c332106788a4658a62d0105
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.
11 //===----------------------------------------------------------------------===//
13 #include "llvm/Transforms/Scalar/JumpThreading.h"
14 #include "llvm/ADT/DenseMap.h"
15 #include "llvm/ADT/DenseSet.h"
16 #include "llvm/ADT/Optional.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallVector.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/Analysis/BlockFrequencyInfo.h"
23 #include "llvm/Analysis/BranchProbabilityInfo.h"
24 #include "llvm/Analysis/CFG.h"
25 #include "llvm/Analysis/ConstantFolding.h"
26 #include "llvm/Analysis/DomTreeUpdater.h"
27 #include "llvm/Analysis/GlobalsModRef.h"
28 #include "llvm/Analysis/GuardUtils.h"
29 #include "llvm/Analysis/InstructionSimplify.h"
30 #include "llvm/Analysis/LazyValueInfo.h"
31 #include "llvm/Analysis/Loads.h"
32 #include "llvm/Analysis/LoopInfo.h"
33 #include "llvm/Analysis/TargetLibraryInfo.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/IR/BasicBlock.h"
36 #include "llvm/IR/CFG.h"
37 #include "llvm/IR/Constant.h"
38 #include "llvm/IR/ConstantRange.h"
39 #include "llvm/IR/Constants.h"
40 #include "llvm/IR/DataLayout.h"
41 #include "llvm/IR/Dominators.h"
42 #include "llvm/IR/Function.h"
43 #include "llvm/IR/InstrTypes.h"
44 #include "llvm/IR/Instruction.h"
45 #include "llvm/IR/Instructions.h"
46 #include "llvm/IR/IntrinsicInst.h"
47 #include "llvm/IR/Intrinsics.h"
48 #include "llvm/IR/LLVMContext.h"
49 #include "llvm/IR/MDBuilder.h"
50 #include "llvm/IR/Metadata.h"
51 #include "llvm/IR/Module.h"
52 #include "llvm/IR/PassManager.h"
53 #include "llvm/IR/PatternMatch.h"
54 #include "llvm/IR/Type.h"
55 #include "llvm/IR/Use.h"
56 #include "llvm/IR/User.h"
57 #include "llvm/IR/Value.h"
58 #include "llvm/Pass.h"
59 #include "llvm/Support/BlockFrequency.h"
60 #include "llvm/Support/BranchProbability.h"
61 #include "llvm/Support/Casting.h"
62 #include "llvm/Support/CommandLine.h"
63 #include "llvm/Support/Debug.h"
64 #include "llvm/Support/raw_ostream.h"
65 #include "llvm/Transforms/Scalar.h"
66 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
67 #include "llvm/Transforms/Utils/Cloning.h"
68 #include "llvm/Transforms/Utils/Local.h"
69 #include "llvm/Transforms/Utils/SSAUpdater.h"
70 #include "llvm/Transforms/Utils/ValueMapper.h"
71 #include <algorithm>
72 #include <cassert>
73 #include <cstddef>
74 #include <cstdint>
75 #include <iterator>
76 #include <memory>
77 #include <utility>
79 using namespace llvm;
80 using namespace jumpthreading;
82 #define DEBUG_TYPE "jump-threading"
84 STATISTIC(NumThreads, "Number of jumps threaded");
85 STATISTIC(NumFolds, "Number of terminators folded");
86 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
88 static cl::opt<unsigned>
89 BBDuplicateThreshold("jump-threading-threshold",
90 cl::desc("Max block size to duplicate for jump threading"),
91 cl::init(6), cl::Hidden);
93 static cl::opt<unsigned>
94 ImplicationSearchThreshold(
95 "jump-threading-implication-search-threshold",
96 cl::desc("The number of predecessors to search for a stronger "
97 "condition to use to thread over a weaker condition"),
98 cl::init(3), cl::Hidden);
100 static cl::opt<bool> PrintLVIAfterJumpThreading(
101 "print-lvi-after-jump-threading",
102 cl::desc("Print the LazyValueInfo cache after JumpThreading"), cl::init(false),
103 cl::Hidden);
105 static cl::opt<bool> ThreadAcrossLoopHeaders(
106 "jump-threading-across-loop-headers",
107 cl::desc("Allow JumpThreading to thread across loop headers, for testing"),
108 cl::init(false), cl::Hidden);
111 namespace {
113 /// This pass performs 'jump threading', which looks at blocks that have
114 /// multiple predecessors and multiple successors. If one or more of the
115 /// predecessors of the block can be proven to always jump to one of the
116 /// successors, we forward the edge from the predecessor to the successor by
117 /// duplicating the contents of this block.
119 /// An example of when this can occur is code like this:
121 /// if () { ...
122 /// X = 4;
123 /// }
124 /// if (X < 3) {
126 /// In this case, the unconditional branch at the end of the first if can be
127 /// revectored to the false side of the second if.
128 class JumpThreading : public FunctionPass {
129 JumpThreadingPass Impl;
131 public:
132 static char ID; // Pass identification
134 JumpThreading(int T = -1) : FunctionPass(ID), Impl(T) {
135 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
138 bool runOnFunction(Function &F) override;
140 void getAnalysisUsage(AnalysisUsage &AU) const override {
141 AU.addRequired<DominatorTreeWrapperPass>();
142 AU.addPreserved<DominatorTreeWrapperPass>();
143 AU.addRequired<AAResultsWrapperPass>();
144 AU.addRequired<LazyValueInfoWrapperPass>();
145 AU.addPreserved<LazyValueInfoWrapperPass>();
146 AU.addPreserved<GlobalsAAWrapperPass>();
147 AU.addRequired<TargetLibraryInfoWrapperPass>();
150 void releaseMemory() override { Impl.releaseMemory(); }
153 } // end anonymous namespace
155 char JumpThreading::ID = 0;
157 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
158 "Jump Threading", false, false)
159 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
160 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)
161 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
162 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
163 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
164 "Jump Threading", false, false)
166 // Public interface to the Jump Threading pass
167 FunctionPass *llvm::createJumpThreadingPass(int Threshold) {
168 return new JumpThreading(Threshold);
171 JumpThreadingPass::JumpThreadingPass(int T) {
172 BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
175 // Update branch probability information according to conditional
176 // branch probability. This is usually made possible for cloned branches
177 // in inline instances by the context specific profile in the caller.
178 // For instance,
180 // [Block PredBB]
181 // [Branch PredBr]
182 // if (t) {
183 // Block A;
184 // } else {
185 // Block B;
186 // }
188 // [Block BB]
189 // cond = PN([true, %A], [..., %B]); // PHI node
190 // [Branch CondBr]
191 // if (cond) {
192 // ... // P(cond == true) = 1%
193 // }
195 // Here we know that when block A is taken, cond must be true, which means
196 // P(cond == true | A) = 1
198 // Given that P(cond == true) = P(cond == true | A) * P(A) +
199 // P(cond == true | B) * P(B)
200 // we get:
201 // P(cond == true ) = P(A) + P(cond == true | B) * P(B)
203 // which gives us:
204 // P(A) is less than P(cond == true), i.e.
205 // P(t == true) <= P(cond == true)
207 // In other words, if we know P(cond == true) is unlikely, we know
208 // that P(t == true) is also unlikely.
210 static void updatePredecessorProfileMetadata(PHINode *PN, BasicBlock *BB) {
211 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
212 if (!CondBr)
213 return;
215 BranchProbability BP;
216 uint64_t TrueWeight, FalseWeight;
217 if (!CondBr->extractProfMetadata(TrueWeight, FalseWeight))
218 return;
220 // Returns the outgoing edge of the dominating predecessor block
221 // that leads to the PhiNode's incoming block:
222 auto GetPredOutEdge =
223 [](BasicBlock *IncomingBB,
224 BasicBlock *PhiBB) -> std::pair<BasicBlock *, BasicBlock *> {
225 auto *PredBB = IncomingBB;
226 auto *SuccBB = PhiBB;
227 SmallPtrSet<BasicBlock *, 16> Visited;
228 while (true) {
229 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
230 if (PredBr && PredBr->isConditional())
231 return {PredBB, SuccBB};
232 Visited.insert(PredBB);
233 auto *SinglePredBB = PredBB->getSinglePredecessor();
234 if (!SinglePredBB)
235 return {nullptr, nullptr};
237 // Stop searching when SinglePredBB has been visited. It means we see
238 // an unreachable loop.
239 if (Visited.count(SinglePredBB))
240 return {nullptr, nullptr};
242 SuccBB = PredBB;
243 PredBB = SinglePredBB;
247 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
248 Value *PhiOpnd = PN->getIncomingValue(i);
249 ConstantInt *CI = dyn_cast<ConstantInt>(PhiOpnd);
251 if (!CI || !CI->getType()->isIntegerTy(1))
252 continue;
254 BP = (CI->isOne() ? BranchProbability::getBranchProbability(
255 TrueWeight, TrueWeight + FalseWeight)
256 : BranchProbability::getBranchProbability(
257 FalseWeight, TrueWeight + FalseWeight));
259 auto PredOutEdge = GetPredOutEdge(PN->getIncomingBlock(i), BB);
260 if (!PredOutEdge.first)
261 return;
263 BasicBlock *PredBB = PredOutEdge.first;
264 BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
265 if (!PredBr)
266 return;
268 uint64_t PredTrueWeight, PredFalseWeight;
269 // FIXME: We currently only set the profile data when it is missing.
270 // With PGO, this can be used to refine even existing profile data with
271 // context information. This needs to be done after more performance
272 // testing.
273 if (PredBr->extractProfMetadata(PredTrueWeight, PredFalseWeight))
274 continue;
276 // We can not infer anything useful when BP >= 50%, because BP is the
277 // upper bound probability value.
278 if (BP >= BranchProbability(50, 100))
279 continue;
281 SmallVector<uint32_t, 2> Weights;
282 if (PredBr->getSuccessor(0) == PredOutEdge.second) {
283 Weights.push_back(BP.getNumerator());
284 Weights.push_back(BP.getCompl().getNumerator());
285 } else {
286 Weights.push_back(BP.getCompl().getNumerator());
287 Weights.push_back(BP.getNumerator());
289 PredBr->setMetadata(LLVMContext::MD_prof,
290 MDBuilder(PredBr->getParent()->getContext())
291 .createBranchWeights(Weights));
295 /// runOnFunction - Toplevel algorithm.
296 bool JumpThreading::runOnFunction(Function &F) {
297 if (skipFunction(F))
298 return false;
299 auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
300 // Get DT analysis before LVI. When LVI is initialized it conditionally adds
301 // DT if it's available.
302 auto DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
303 auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
304 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
305 DomTreeUpdater DTU(*DT, DomTreeUpdater::UpdateStrategy::Lazy);
306 std::unique_ptr<BlockFrequencyInfo> BFI;
307 std::unique_ptr<BranchProbabilityInfo> BPI;
308 bool HasProfileData = F.hasProfileData();
309 if (HasProfileData) {
310 LoopInfo LI{DominatorTree(F)};
311 BPI.reset(new BranchProbabilityInfo(F, LI, TLI));
312 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
315 bool Changed = Impl.runImpl(F, TLI, LVI, AA, &DTU, HasProfileData,
316 std::move(BFI), std::move(BPI));
317 if (PrintLVIAfterJumpThreading) {
318 dbgs() << "LVI for function '" << F.getName() << "':\n";
319 LVI->printLVI(F, *DT, dbgs());
321 return Changed;
324 PreservedAnalyses JumpThreadingPass::run(Function &F,
325 FunctionAnalysisManager &AM) {
326 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
327 // Get DT analysis before LVI. When LVI is initialized it conditionally adds
328 // DT if it's available.
329 auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
330 auto &LVI = AM.getResult<LazyValueAnalysis>(F);
331 auto &AA = AM.getResult<AAManager>(F);
332 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
334 std::unique_ptr<BlockFrequencyInfo> BFI;
335 std::unique_ptr<BranchProbabilityInfo> BPI;
336 if (F.hasProfileData()) {
337 LoopInfo LI{DominatorTree(F)};
338 BPI.reset(new BranchProbabilityInfo(F, LI, &TLI));
339 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
342 bool Changed = runImpl(F, &TLI, &LVI, &AA, &DTU, HasProfileData,
343 std::move(BFI), std::move(BPI));
345 if (!Changed)
346 return PreservedAnalyses::all();
347 PreservedAnalyses PA;
348 PA.preserve<GlobalsAA>();
349 PA.preserve<DominatorTreeAnalysis>();
350 PA.preserve<LazyValueAnalysis>();
351 return PA;
354 bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
355 LazyValueInfo *LVI_, AliasAnalysis *AA_,
356 DomTreeUpdater *DTU_, bool HasProfileData_,
357 std::unique_ptr<BlockFrequencyInfo> BFI_,
358 std::unique_ptr<BranchProbabilityInfo> BPI_) {
359 LLVM_DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
360 TLI = TLI_;
361 LVI = LVI_;
362 AA = AA_;
363 DTU = DTU_;
364 BFI.reset();
365 BPI.reset();
366 // When profile data is available, we need to update edge weights after
367 // successful jump threading, which requires both BPI and BFI being available.
368 HasProfileData = HasProfileData_;
369 auto *GuardDecl = F.getParent()->getFunction(
370 Intrinsic::getName(Intrinsic::experimental_guard));
371 HasGuards = GuardDecl && !GuardDecl->use_empty();
372 if (HasProfileData) {
373 BPI = std::move(BPI_);
374 BFI = std::move(BFI_);
377 // JumpThreading must not processes blocks unreachable from entry. It's a
378 // waste of compute time and can potentially lead to hangs.
379 SmallPtrSet<BasicBlock *, 16> Unreachable;
380 assert(DTU && "DTU isn't passed into JumpThreading before using it.");
381 assert(DTU->hasDomTree() && "JumpThreading relies on DomTree to proceed.");
382 DominatorTree &DT = DTU->getDomTree();
383 for (auto &BB : F)
384 if (!DT.isReachableFromEntry(&BB))
385 Unreachable.insert(&BB);
387 if (!ThreadAcrossLoopHeaders)
388 FindLoopHeaders(F);
390 bool EverChanged = false;
391 bool Changed;
392 do {
393 Changed = false;
394 for (auto &BB : F) {
395 if (Unreachable.count(&BB))
396 continue;
397 while (ProcessBlock(&BB)) // Thread all of the branches we can over BB.
398 Changed = true;
399 // Stop processing BB if it's the entry or is now deleted. The following
400 // routines attempt to eliminate BB and locating a suitable replacement
401 // for the entry is non-trivial.
402 if (&BB == &F.getEntryBlock() || DTU->isBBPendingDeletion(&BB))
403 continue;
405 if (pred_empty(&BB)) {
406 // When ProcessBlock makes BB unreachable it doesn't bother to fix up
407 // the instructions in it. We must remove BB to prevent invalid IR.
408 LLVM_DEBUG(dbgs() << " JT: Deleting dead block '" << BB.getName()
409 << "' with terminator: " << *BB.getTerminator()
410 << '\n');
411 LoopHeaders.erase(&BB);
412 LVI->eraseBlock(&BB);
413 DeleteDeadBlock(&BB, DTU);
414 Changed = true;
415 continue;
418 // ProcessBlock doesn't thread BBs with unconditional TIs. However, if BB
419 // is "almost empty", we attempt to merge BB with its sole successor.
420 auto *BI = dyn_cast<BranchInst>(BB.getTerminator());
421 if (BI && BI->isUnconditional() &&
422 // The terminator must be the only non-phi instruction in BB.
423 BB.getFirstNonPHIOrDbg()->isTerminator() &&
424 // Don't alter Loop headers and latches to ensure another pass can
425 // detect and transform nested loops later.
426 !LoopHeaders.count(&BB) && !LoopHeaders.count(BI->getSuccessor(0)) &&
427 TryToSimplifyUncondBranchFromEmptyBlock(&BB, DTU)) {
428 // BB is valid for cleanup here because we passed in DTU. F remains
429 // BB's parent until a DTU->getDomTree() event.
430 LVI->eraseBlock(&BB);
431 Changed = true;
434 EverChanged |= Changed;
435 } while (Changed);
437 LoopHeaders.clear();
438 // Flush only the Dominator Tree.
439 DTU->getDomTree();
440 LVI->enableDT();
441 return EverChanged;
444 // Replace uses of Cond with ToVal when safe to do so. If all uses are
445 // replaced, we can remove Cond. We cannot blindly replace all uses of Cond
446 // because we may incorrectly replace uses when guards/assumes are uses of
447 // of `Cond` and we used the guards/assume to reason about the `Cond` value
448 // at the end of block. RAUW unconditionally replaces all uses
449 // including the guards/assumes themselves and the uses before the
450 // guard/assume.
451 static void ReplaceFoldableUses(Instruction *Cond, Value *ToVal) {
452 assert(Cond->getType() == ToVal->getType());
453 auto *BB = Cond->getParent();
454 // We can unconditionally replace all uses in non-local blocks (i.e. uses
455 // strictly dominated by BB), since LVI information is true from the
456 // terminator of BB.
457 replaceNonLocalUsesWith(Cond, ToVal);
458 for (Instruction &I : reverse(*BB)) {
459 // Reached the Cond whose uses we are trying to replace, so there are no
460 // more uses.
461 if (&I == Cond)
462 break;
463 // We only replace uses in instructions that are guaranteed to reach the end
464 // of BB, where we know Cond is ToVal.
465 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
466 break;
467 I.replaceUsesOfWith(Cond, ToVal);
469 if (Cond->use_empty() && !Cond->mayHaveSideEffects())
470 Cond->eraseFromParent();
473 /// Return the cost of duplicating a piece of this block from first non-phi
474 /// and before StopAt instruction to thread across it. Stop scanning the block
475 /// when exceeding the threshold. If duplication is impossible, returns ~0U.
476 static unsigned getJumpThreadDuplicationCost(BasicBlock *BB,
477 Instruction *StopAt,
478 unsigned Threshold) {
479 assert(StopAt->getParent() == BB && "Not an instruction from proper BB?");
480 /// Ignore PHI nodes, these will be flattened when duplication happens.
481 BasicBlock::const_iterator I(BB->getFirstNonPHI());
483 // FIXME: THREADING will delete values that are just used to compute the
484 // branch, so they shouldn't count against the duplication cost.
486 unsigned Bonus = 0;
487 if (BB->getTerminator() == StopAt) {
488 // Threading through a switch statement is particularly profitable. If this
489 // block ends in a switch, decrease its cost to make it more likely to
490 // happen.
491 if (isa<SwitchInst>(StopAt))
492 Bonus = 6;
494 // The same holds for indirect branches, but slightly more so.
495 if (isa<IndirectBrInst>(StopAt))
496 Bonus = 8;
499 // Bump the threshold up so the early exit from the loop doesn't skip the
500 // terminator-based Size adjustment at the end.
501 Threshold += Bonus;
503 // Sum up the cost of each instruction until we get to the terminator. Don't
504 // include the terminator because the copy won't include it.
505 unsigned Size = 0;
506 for (; &*I != StopAt; ++I) {
508 // Stop scanning the block if we've reached the threshold.
509 if (Size > Threshold)
510 return Size;
512 // Debugger intrinsics don't incur code size.
513 if (isa<DbgInfoIntrinsic>(I)) continue;
515 // If this is a pointer->pointer bitcast, it is free.
516 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
517 continue;
519 // Bail out if this instruction gives back a token type, it is not possible
520 // to duplicate it if it is used outside this BB.
521 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
522 return ~0U;
524 // All other instructions count for at least one unit.
525 ++Size;
527 // Calls are more expensive. If they are non-intrinsic calls, we model them
528 // as having cost of 4. If they are a non-vector intrinsic, we model them
529 // as having cost of 2 total, and if they are a vector intrinsic, we model
530 // them as having cost 1.
531 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
532 if (CI->cannotDuplicate() || CI->isConvergent())
533 // Blocks with NoDuplicate are modelled as having infinite cost, so they
534 // are never duplicated.
535 return ~0U;
536 else if (!isa<IntrinsicInst>(CI))
537 Size += 3;
538 else if (!CI->getType()->isVectorTy())
539 Size += 1;
543 return Size > Bonus ? Size - Bonus : 0;
546 /// FindLoopHeaders - We do not want jump threading to turn proper loop
547 /// structures into irreducible loops. Doing this breaks up the loop nesting
548 /// hierarchy and pessimizes later transformations. To prevent this from
549 /// happening, we first have to find the loop headers. Here we approximate this
550 /// by finding targets of backedges in the CFG.
552 /// Note that there definitely are cases when we want to allow threading of
553 /// edges across a loop header. For example, threading a jump from outside the
554 /// loop (the preheader) to an exit block of the loop is definitely profitable.
555 /// It is also almost always profitable to thread backedges from within the loop
556 /// to exit blocks, and is often profitable to thread backedges to other blocks
557 /// within the loop (forming a nested loop). This simple analysis is not rich
558 /// enough to track all of these properties and keep it up-to-date as the CFG
559 /// mutates, so we don't allow any of these transformations.
560 void JumpThreadingPass::FindLoopHeaders(Function &F) {
561 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
562 FindFunctionBackedges(F, Edges);
564 for (const auto &Edge : Edges)
565 LoopHeaders.insert(Edge.second);
568 /// getKnownConstant - Helper method to determine if we can thread over a
569 /// terminator with the given value as its condition, and if so what value to
570 /// use for that. What kind of value this is depends on whether we want an
571 /// integer or a block address, but an undef is always accepted.
572 /// Returns null if Val is null or not an appropriate constant.
573 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
574 if (!Val)
575 return nullptr;
577 // Undef is "known" enough.
578 if (UndefValue *U = dyn_cast<UndefValue>(Val))
579 return U;
581 if (Preference == WantBlockAddress)
582 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
584 return dyn_cast<ConstantInt>(Val);
587 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
588 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
589 /// in any of our predecessors. If so, return the known list of value and pred
590 /// BB in the result vector.
592 /// This returns true if there were any known values.
593 bool JumpThreadingPass::ComputeValueKnownInPredecessorsImpl(
594 Value *V, BasicBlock *BB, PredValueInfo &Result,
595 ConstantPreference Preference,
596 DenseSet<std::pair<Value *, BasicBlock *>> &RecursionSet,
597 Instruction *CxtI) {
598 // This method walks up use-def chains recursively. Because of this, we could
599 // get into an infinite loop going around loops in the use-def chain. To
600 // prevent this, keep track of what (value, block) pairs we've already visited
601 // and terminate the search if we loop back to them
602 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
603 return false;
605 // If V is a constant, then it is known in all predecessors.
606 if (Constant *KC = getKnownConstant(V, Preference)) {
607 for (BasicBlock *Pred : predecessors(BB))
608 Result.push_back(std::make_pair(KC, Pred));
610 return !Result.empty();
613 // If V is a non-instruction value, or an instruction in a different block,
614 // then it can't be derived from a PHI.
615 Instruction *I = dyn_cast<Instruction>(V);
616 if (!I || I->getParent() != BB) {
618 // Okay, if this is a live-in value, see if it has a known value at the end
619 // of any of our predecessors.
621 // FIXME: This should be an edge property, not a block end property.
622 /// TODO: Per PR2563, we could infer value range information about a
623 /// predecessor based on its terminator.
625 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
626 // "I" is a non-local compare-with-a-constant instruction. This would be
627 // able to handle value inequalities better, for example if the compare is
628 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
629 // Perhaps getConstantOnEdge should be smart enough to do this?
631 if (DTU->hasPendingDomTreeUpdates())
632 LVI->disableDT();
633 else
634 LVI->enableDT();
635 for (BasicBlock *P : predecessors(BB)) {
636 // If the value is known by LazyValueInfo to be a constant in a
637 // predecessor, use that information to try to thread this block.
638 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
639 if (Constant *KC = getKnownConstant(PredCst, Preference))
640 Result.push_back(std::make_pair(KC, P));
643 return !Result.empty();
646 /// If I is a PHI node, then we know the incoming values for any constants.
647 if (PHINode *PN = dyn_cast<PHINode>(I)) {
648 if (DTU->hasPendingDomTreeUpdates())
649 LVI->disableDT();
650 else
651 LVI->enableDT();
652 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
653 Value *InVal = PN->getIncomingValue(i);
654 if (Constant *KC = getKnownConstant(InVal, Preference)) {
655 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
656 } else {
657 Constant *CI = LVI->getConstantOnEdge(InVal,
658 PN->getIncomingBlock(i),
659 BB, CxtI);
660 if (Constant *KC = getKnownConstant(CI, Preference))
661 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
665 return !Result.empty();
668 // Handle Cast instructions. Only see through Cast when the source operand is
669 // PHI or Cmp to save the compilation time.
670 if (CastInst *CI = dyn_cast<CastInst>(I)) {
671 Value *Source = CI->getOperand(0);
672 if (!isa<PHINode>(Source) && !isa<CmpInst>(Source))
673 return false;
674 ComputeValueKnownInPredecessorsImpl(Source, BB, Result, Preference,
675 RecursionSet, CxtI);
676 if (Result.empty())
677 return false;
679 // Convert the known values.
680 for (auto &R : Result)
681 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
683 return true;
686 // Handle some boolean conditions.
687 if (I->getType()->getPrimitiveSizeInBits() == 1) {
688 assert(Preference == WantInteger && "One-bit non-integer type?");
689 // X | true -> true
690 // X & false -> false
691 if (I->getOpcode() == Instruction::Or ||
692 I->getOpcode() == Instruction::And) {
693 PredValueInfoTy LHSVals, RHSVals;
695 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
696 WantInteger, RecursionSet, CxtI);
697 ComputeValueKnownInPredecessorsImpl(I->getOperand(1), BB, RHSVals,
698 WantInteger, RecursionSet, CxtI);
700 if (LHSVals.empty() && RHSVals.empty())
701 return false;
703 ConstantInt *InterestingVal;
704 if (I->getOpcode() == Instruction::Or)
705 InterestingVal = ConstantInt::getTrue(I->getContext());
706 else
707 InterestingVal = ConstantInt::getFalse(I->getContext());
709 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
711 // Scan for the sentinel. If we find an undef, force it to the
712 // interesting value: x|undef -> true and x&undef -> false.
713 for (const auto &LHSVal : LHSVals)
714 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
715 Result.emplace_back(InterestingVal, LHSVal.second);
716 LHSKnownBBs.insert(LHSVal.second);
718 for (const auto &RHSVal : RHSVals)
719 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
720 // If we already inferred a value for this block on the LHS, don't
721 // re-add it.
722 if (!LHSKnownBBs.count(RHSVal.second))
723 Result.emplace_back(InterestingVal, RHSVal.second);
726 return !Result.empty();
729 // Handle the NOT form of XOR.
730 if (I->getOpcode() == Instruction::Xor &&
731 isa<ConstantInt>(I->getOperand(1)) &&
732 cast<ConstantInt>(I->getOperand(1))->isOne()) {
733 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, Result,
734 WantInteger, RecursionSet, CxtI);
735 if (Result.empty())
736 return false;
738 // Invert the known values.
739 for (auto &R : Result)
740 R.first = ConstantExpr::getNot(R.first);
742 return true;
745 // Try to simplify some other binary operator values.
746 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
747 assert(Preference != WantBlockAddress
748 && "A binary operator creating a block address?");
749 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
750 PredValueInfoTy LHSVals;
751 ComputeValueKnownInPredecessorsImpl(BO->getOperand(0), BB, LHSVals,
752 WantInteger, RecursionSet, CxtI);
754 // Try to use constant folding to simplify the binary operator.
755 for (const auto &LHSVal : LHSVals) {
756 Constant *V = LHSVal.first;
757 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
759 if (Constant *KC = getKnownConstant(Folded, WantInteger))
760 Result.push_back(std::make_pair(KC, LHSVal.second));
764 return !Result.empty();
767 // Handle compare with phi operand, where the PHI is defined in this block.
768 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
769 assert(Preference == WantInteger && "Compares only produce integers");
770 Type *CmpType = Cmp->getType();
771 Value *CmpLHS = Cmp->getOperand(0);
772 Value *CmpRHS = Cmp->getOperand(1);
773 CmpInst::Predicate Pred = Cmp->getPredicate();
775 PHINode *PN = dyn_cast<PHINode>(CmpLHS);
776 if (!PN)
777 PN = dyn_cast<PHINode>(CmpRHS);
778 if (PN && PN->getParent() == BB) {
779 const DataLayout &DL = PN->getModule()->getDataLayout();
780 // We can do this simplification if any comparisons fold to true or false.
781 // See if any do.
782 if (DTU->hasPendingDomTreeUpdates())
783 LVI->disableDT();
784 else
785 LVI->enableDT();
786 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
787 BasicBlock *PredBB = PN->getIncomingBlock(i);
788 Value *LHS, *RHS;
789 if (PN == CmpLHS) {
790 LHS = PN->getIncomingValue(i);
791 RHS = CmpRHS->DoPHITranslation(BB, PredBB);
792 } else {
793 LHS = CmpLHS->DoPHITranslation(BB, PredBB);
794 RHS = PN->getIncomingValue(i);
796 Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
797 if (!Res) {
798 if (!isa<Constant>(RHS))
799 continue;
801 // getPredicateOnEdge call will make no sense if LHS is defined in BB.
802 auto LHSInst = dyn_cast<Instruction>(LHS);
803 if (LHSInst && LHSInst->getParent() == BB)
804 continue;
806 LazyValueInfo::Tristate
807 ResT = LVI->getPredicateOnEdge(Pred, LHS,
808 cast<Constant>(RHS), PredBB, BB,
809 CxtI ? CxtI : Cmp);
810 if (ResT == LazyValueInfo::Unknown)
811 continue;
812 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
815 if (Constant *KC = getKnownConstant(Res, WantInteger))
816 Result.push_back(std::make_pair(KC, PredBB));
819 return !Result.empty();
822 // If comparing a live-in value against a constant, see if we know the
823 // live-in value on any predecessors.
824 if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
825 Constant *CmpConst = cast<Constant>(CmpRHS);
827 if (!isa<Instruction>(CmpLHS) ||
828 cast<Instruction>(CmpLHS)->getParent() != BB) {
829 if (DTU->hasPendingDomTreeUpdates())
830 LVI->disableDT();
831 else
832 LVI->enableDT();
833 for (BasicBlock *P : predecessors(BB)) {
834 // If the value is known by LazyValueInfo to be a constant in a
835 // predecessor, use that information to try to thread this block.
836 LazyValueInfo::Tristate Res =
837 LVI->getPredicateOnEdge(Pred, CmpLHS,
838 CmpConst, P, BB, CxtI ? CxtI : Cmp);
839 if (Res == LazyValueInfo::Unknown)
840 continue;
842 Constant *ResC = ConstantInt::get(CmpType, Res);
843 Result.push_back(std::make_pair(ResC, P));
846 return !Result.empty();
849 // InstCombine can fold some forms of constant range checks into
850 // (icmp (add (x, C1)), C2). See if we have we have such a thing with
851 // x as a live-in.
853 using namespace PatternMatch;
855 Value *AddLHS;
856 ConstantInt *AddConst;
857 if (isa<ConstantInt>(CmpConst) &&
858 match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
859 if (!isa<Instruction>(AddLHS) ||
860 cast<Instruction>(AddLHS)->getParent() != BB) {
861 if (DTU->hasPendingDomTreeUpdates())
862 LVI->disableDT();
863 else
864 LVI->enableDT();
865 for (BasicBlock *P : predecessors(BB)) {
866 // If the value is known by LazyValueInfo to be a ConstantRange in
867 // a predecessor, use that information to try to thread this
868 // block.
869 ConstantRange CR = LVI->getConstantRangeOnEdge(
870 AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
871 // Propagate the range through the addition.
872 CR = CR.add(AddConst->getValue());
874 // Get the range where the compare returns true.
875 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(
876 Pred, cast<ConstantInt>(CmpConst)->getValue());
878 Constant *ResC;
879 if (CmpRange.contains(CR))
880 ResC = ConstantInt::getTrue(CmpType);
881 else if (CmpRange.inverse().contains(CR))
882 ResC = ConstantInt::getFalse(CmpType);
883 else
884 continue;
886 Result.push_back(std::make_pair(ResC, P));
889 return !Result.empty();
894 // Try to find a constant value for the LHS of a comparison,
895 // and evaluate it statically if we can.
896 PredValueInfoTy LHSVals;
897 ComputeValueKnownInPredecessorsImpl(I->getOperand(0), BB, LHSVals,
898 WantInteger, RecursionSet, CxtI);
900 for (const auto &LHSVal : LHSVals) {
901 Constant *V = LHSVal.first;
902 Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
903 if (Constant *KC = getKnownConstant(Folded, WantInteger))
904 Result.push_back(std::make_pair(KC, LHSVal.second));
907 return !Result.empty();
911 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
912 // Handle select instructions where at least one operand is a known constant
913 // and we can figure out the condition value for any predecessor block.
914 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
915 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
916 PredValueInfoTy Conds;
917 if ((TrueVal || FalseVal) &&
918 ComputeValueKnownInPredecessorsImpl(SI->getCondition(), BB, Conds,
919 WantInteger, RecursionSet, CxtI)) {
920 for (auto &C : Conds) {
921 Constant *Cond = C.first;
923 // Figure out what value to use for the condition.
924 bool KnownCond;
925 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
926 // A known boolean.
927 KnownCond = CI->isOne();
928 } else {
929 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
930 // Either operand will do, so be sure to pick the one that's a known
931 // constant.
932 // FIXME: Do this more cleverly if both values are known constants?
933 KnownCond = (TrueVal != nullptr);
936 // See if the select has a known constant value for this predecessor.
937 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
938 Result.push_back(std::make_pair(Val, C.second));
941 return !Result.empty();
945 // If all else fails, see if LVI can figure out a constant value for us.
946 if (DTU->hasPendingDomTreeUpdates())
947 LVI->disableDT();
948 else
949 LVI->enableDT();
950 Constant *CI = LVI->getConstant(V, BB, CxtI);
951 if (Constant *KC = getKnownConstant(CI, Preference)) {
952 for (BasicBlock *Pred : predecessors(BB))
953 Result.push_back(std::make_pair(KC, Pred));
956 return !Result.empty();
959 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
960 /// in an undefined jump, decide which block is best to revector to.
962 /// Since we can pick an arbitrary destination, we pick the successor with the
963 /// fewest predecessors. This should reduce the in-degree of the others.
964 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
965 Instruction *BBTerm = BB->getTerminator();
966 unsigned MinSucc = 0;
967 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
968 // Compute the successor with the minimum number of predecessors.
969 unsigned MinNumPreds = pred_size(TestBB);
970 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
971 TestBB = BBTerm->getSuccessor(i);
972 unsigned NumPreds = pred_size(TestBB);
973 if (NumPreds < MinNumPreds) {
974 MinSucc = i;
975 MinNumPreds = NumPreds;
979 return MinSucc;
982 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
983 if (!BB->hasAddressTaken()) return false;
985 // If the block has its address taken, it may be a tree of dead constants
986 // hanging off of it. These shouldn't keep the block alive.
987 BlockAddress *BA = BlockAddress::get(BB);
988 BA->removeDeadConstantUsers();
989 return !BA->use_empty();
992 /// ProcessBlock - If there are any predecessors whose control can be threaded
993 /// through to a successor, transform them now.
994 bool JumpThreadingPass::ProcessBlock(BasicBlock *BB) {
995 // If the block is trivially dead, just return and let the caller nuke it.
996 // This simplifies other transformations.
997 if (DTU->isBBPendingDeletion(BB) ||
998 (pred_empty(BB) && BB != &BB->getParent()->getEntryBlock()))
999 return false;
1001 // If this block has a single predecessor, and if that pred has a single
1002 // successor, merge the blocks. This encourages recursive jump threading
1003 // because now the condition in this block can be threaded through
1004 // predecessors of our predecessor block.
1005 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
1006 const Instruction *TI = SinglePred->getTerminator();
1007 if (!TI->isExceptionalTerminator() && TI->getNumSuccessors() == 1 &&
1008 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
1009 // If SinglePred was a loop header, BB becomes one.
1010 if (LoopHeaders.erase(SinglePred))
1011 LoopHeaders.insert(BB);
1013 LVI->eraseBlock(SinglePred);
1014 MergeBasicBlockIntoOnlyPred(BB, DTU);
1016 // Now that BB is merged into SinglePred (i.e. SinglePred Code followed by
1017 // BB code within one basic block `BB`), we need to invalidate the LVI
1018 // information associated with BB, because the LVI information need not be
1019 // true for all of BB after the merge. For example,
1020 // Before the merge, LVI info and code is as follows:
1021 // SinglePred: <LVI info1 for %p val>
1022 // %y = use of %p
1023 // call @exit() // need not transfer execution to successor.
1024 // assume(%p) // from this point on %p is true
1025 // br label %BB
1026 // BB: <LVI info2 for %p val, i.e. %p is true>
1027 // %x = use of %p
1028 // br label exit
1030 // Note that this LVI info for blocks BB and SinglPred is correct for %p
1031 // (info2 and info1 respectively). After the merge and the deletion of the
1032 // LVI info1 for SinglePred. We have the following code:
1033 // BB: <LVI info2 for %p val>
1034 // %y = use of %p
1035 // call @exit()
1036 // assume(%p)
1037 // %x = use of %p <-- LVI info2 is correct from here onwards.
1038 // br label exit
1039 // LVI info2 for BB is incorrect at the beginning of BB.
1041 // Invalidate LVI information for BB if the LVI is not provably true for
1042 // all of BB.
1043 if (!isGuaranteedToTransferExecutionToSuccessor(BB))
1044 LVI->eraseBlock(BB);
1045 return true;
1049 if (TryToUnfoldSelectInCurrBB(BB))
1050 return true;
1052 // Look if we can propagate guards to predecessors.
1053 if (HasGuards && ProcessGuards(BB))
1054 return true;
1056 // What kind of constant we're looking for.
1057 ConstantPreference Preference = WantInteger;
1059 // Look to see if the terminator is a conditional branch, switch or indirect
1060 // branch, if not we can't thread it.
1061 Value *Condition;
1062 Instruction *Terminator = BB->getTerminator();
1063 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
1064 // Can't thread an unconditional jump.
1065 if (BI->isUnconditional()) return false;
1066 Condition = BI->getCondition();
1067 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
1068 Condition = SI->getCondition();
1069 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
1070 // Can't thread indirect branch with no successors.
1071 if (IB->getNumSuccessors() == 0) return false;
1072 Condition = IB->getAddress()->stripPointerCasts();
1073 Preference = WantBlockAddress;
1074 } else {
1075 return false; // Must be an invoke or callbr.
1078 // Run constant folding to see if we can reduce the condition to a simple
1079 // constant.
1080 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
1081 Value *SimpleVal =
1082 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
1083 if (SimpleVal) {
1084 I->replaceAllUsesWith(SimpleVal);
1085 if (isInstructionTriviallyDead(I, TLI))
1086 I->eraseFromParent();
1087 Condition = SimpleVal;
1091 // If the terminator is branching on an undef, we can pick any of the
1092 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
1093 if (isa<UndefValue>(Condition)) {
1094 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
1095 std::vector<DominatorTree::UpdateType> Updates;
1097 // Fold the branch/switch.
1098 Instruction *BBTerm = BB->getTerminator();
1099 Updates.reserve(BBTerm->getNumSuccessors());
1100 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
1101 if (i == BestSucc) continue;
1102 BasicBlock *Succ = BBTerm->getSuccessor(i);
1103 Succ->removePredecessor(BB, true);
1104 Updates.push_back({DominatorTree::Delete, BB, Succ});
1107 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()
1108 << "' folding undef terminator: " << *BBTerm << '\n');
1109 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
1110 BBTerm->eraseFromParent();
1111 DTU->applyUpdatesPermissive(Updates);
1112 return true;
1115 // If the terminator of this block is branching on a constant, simplify the
1116 // terminator to an unconditional branch. This can occur due to threading in
1117 // other blocks.
1118 if (getKnownConstant(Condition, Preference)) {
1119 LLVM_DEBUG(dbgs() << " In block '" << BB->getName()
1120 << "' folding terminator: " << *BB->getTerminator()
1121 << '\n');
1122 ++NumFolds;
1123 ConstantFoldTerminator(BB, true, nullptr, DTU);
1124 return true;
1127 Instruction *CondInst = dyn_cast<Instruction>(Condition);
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 false;
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");
1150 if (DTU->hasPendingDomTreeUpdates())
1151 LVI->disableDT();
1152 else
1153 LVI->enableDT();
1154 LazyValueInfo::Tristate Ret =
1155 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
1156 CondConst, CondBr);
1157 if (Ret != LazyValueInfo::Unknown) {
1158 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
1159 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
1160 BasicBlock *ToRemoveSucc = CondBr->getSuccessor(ToRemove);
1161 ToRemoveSucc->removePredecessor(BB, true);
1162 BranchInst *UncondBr =
1163 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
1164 UncondBr->setDebugLoc(CondBr->getDebugLoc());
1165 CondBr->eraseFromParent();
1166 if (CondCmp->use_empty())
1167 CondCmp->eraseFromParent();
1168 // We can safely replace *some* uses of the CondInst if it has
1169 // exactly one value as returned by LVI. RAUW is incorrect in the
1170 // presence of guards and assumes, that have the `Cond` as the use. This
1171 // is because we use the guards/assume to reason about the `Cond` value
1172 // at the end of block, but RAUW unconditionally replaces all uses
1173 // including the guards/assumes themselves and the uses before the
1174 // guard/assume.
1175 else if (CondCmp->getParent() == BB) {
1176 auto *CI = Ret == LazyValueInfo::True ?
1177 ConstantInt::getTrue(CondCmp->getType()) :
1178 ConstantInt::getFalse(CondCmp->getType());
1179 ReplaceFoldableUses(CondCmp, CI);
1181 DTU->applyUpdatesPermissive(
1182 {{DominatorTree::Delete, BB, ToRemoveSucc}});
1183 return true;
1186 // We did not manage to simplify this branch, try to see whether
1187 // CondCmp depends on a known phi-select pattern.
1188 if (TryToUnfoldSelect(CondCmp, BB))
1189 return true;
1193 if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator()))
1194 if (TryToUnfoldSelect(SI, BB))
1195 return true;
1197 // Check for some cases that are worth simplifying. Right now we want to look
1198 // for loads that are used by a switch or by the condition for the branch. If
1199 // we see one, check to see if it's partially redundant. If so, insert a PHI
1200 // which can then be used to thread the values.
1201 Value *SimplifyValue = CondInst;
1202 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
1203 if (isa<Constant>(CondCmp->getOperand(1)))
1204 SimplifyValue = CondCmp->getOperand(0);
1206 // TODO: There are other places where load PRE would be profitable, such as
1207 // more complex comparisons.
1208 if (LoadInst *LoadI = dyn_cast<LoadInst>(SimplifyValue))
1209 if (SimplifyPartiallyRedundantLoad(LoadI))
1210 return true;
1212 // Before threading, try to propagate profile data backwards:
1213 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1214 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1215 updatePredecessorProfileMetadata(PN, BB);
1217 // Handle a variety of cases where we are branching on something derived from
1218 // a PHI node in the current block. If we can prove that any predecessors
1219 // compute a predictable value based on a PHI node, thread those predecessors.
1220 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
1221 return true;
1223 // If this is an otherwise-unfoldable branch on a phi node in the current
1224 // block, see if we can simplify.
1225 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
1226 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1227 return ProcessBranchOnPHI(PN);
1229 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
1230 if (CondInst->getOpcode() == Instruction::Xor &&
1231 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
1232 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
1234 // Search for a stronger dominating condition that can be used to simplify a
1235 // conditional branch leaving BB.
1236 if (ProcessImpliedCondition(BB))
1237 return true;
1239 return false;
1242 bool JumpThreadingPass::ProcessImpliedCondition(BasicBlock *BB) {
1243 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
1244 if (!BI || !BI->isConditional())
1245 return false;
1247 Value *Cond = BI->getCondition();
1248 BasicBlock *CurrentBB = BB;
1249 BasicBlock *CurrentPred = BB->getSinglePredecessor();
1250 unsigned Iter = 0;
1252 auto &DL = BB->getModule()->getDataLayout();
1254 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1255 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
1256 if (!PBI || !PBI->isConditional())
1257 return false;
1258 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
1259 return false;
1261 bool CondIsTrue = PBI->getSuccessor(0) == CurrentBB;
1262 Optional<bool> Implication =
1263 isImpliedCondition(PBI->getCondition(), Cond, DL, CondIsTrue);
1264 if (Implication) {
1265 BasicBlock *KeepSucc = BI->getSuccessor(*Implication ? 0 : 1);
1266 BasicBlock *RemoveSucc = BI->getSuccessor(*Implication ? 1 : 0);
1267 RemoveSucc->removePredecessor(BB);
1268 BranchInst *UncondBI = BranchInst::Create(KeepSucc, BI);
1269 UncondBI->setDebugLoc(BI->getDebugLoc());
1270 BI->eraseFromParent();
1271 DTU->applyUpdatesPermissive({{DominatorTree::Delete, BB, RemoveSucc}});
1272 return true;
1274 CurrentBB = CurrentPred;
1275 CurrentPred = CurrentBB->getSinglePredecessor();
1278 return false;
1281 /// Return true if Op is an instruction defined in the given block.
1282 static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
1283 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1284 if (OpInst->getParent() == BB)
1285 return true;
1286 return false;
1289 /// SimplifyPartiallyRedundantLoad - If LoadI is an obviously partially
1290 /// redundant load instruction, eliminate it by replacing it with a PHI node.
1291 /// This is an important optimization that encourages jump threading, and needs
1292 /// to be run interlaced with other jump threading tasks.
1293 bool JumpThreadingPass::SimplifyPartiallyRedundantLoad(LoadInst *LoadI) {
1294 // Don't hack volatile and ordered loads.
1295 if (!LoadI->isUnordered()) return false;
1297 // If the load is defined in a block with exactly one predecessor, it can't be
1298 // partially redundant.
1299 BasicBlock *LoadBB = LoadI->getParent();
1300 if (LoadBB->getSinglePredecessor())
1301 return false;
1303 // If the load is defined in an EH pad, it can't be partially redundant,
1304 // because the edges between the invoke and the EH pad cannot have other
1305 // instructions between them.
1306 if (LoadBB->isEHPad())
1307 return false;
1309 Value *LoadedPtr = LoadI->getOperand(0);
1311 // If the loaded operand is defined in the LoadBB and its not a phi,
1312 // it can't be available in predecessors.
1313 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
1314 return false;
1316 // Scan a few instructions up from the load, to see if it is obviously live at
1317 // the entry to its block.
1318 BasicBlock::iterator BBIt(LoadI);
1319 bool IsLoadCSE;
1320 if (Value *AvailableVal = FindAvailableLoadedValue(
1321 LoadI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1322 // If the value of the load is locally available within the block, just use
1323 // it. This frequently occurs for reg2mem'd allocas.
1325 if (IsLoadCSE) {
1326 LoadInst *NLoadI = cast<LoadInst>(AvailableVal);
1327 combineMetadataForCSE(NLoadI, LoadI, false);
1330 // If the returned value is the load itself, replace with an undef. This can
1331 // only happen in dead loops.
1332 if (AvailableVal == LoadI)
1333 AvailableVal = UndefValue::get(LoadI->getType());
1334 if (AvailableVal->getType() != LoadI->getType())
1335 AvailableVal = CastInst::CreateBitOrPointerCast(
1336 AvailableVal, LoadI->getType(), "", LoadI);
1337 LoadI->replaceAllUsesWith(AvailableVal);
1338 LoadI->eraseFromParent();
1339 return true;
1342 // Otherwise, if we scanned the whole block and got to the top of the block,
1343 // we know the block is locally transparent to the load. If not, something
1344 // might clobber its value.
1345 if (BBIt != LoadBB->begin())
1346 return false;
1348 // If all of the loads and stores that feed the value have the same AA tags,
1349 // then we can propagate them onto any newly inserted loads.
1350 AAMDNodes AATags;
1351 LoadI->getAAMetadata(AATags);
1353 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1355 using AvailablePredsTy = SmallVector<std::pair<BasicBlock *, Value *>, 8>;
1357 AvailablePredsTy AvailablePreds;
1358 BasicBlock *OneUnavailablePred = nullptr;
1359 SmallVector<LoadInst*, 8> CSELoads;
1361 // If we got here, the loaded value is transparent through to the start of the
1362 // block. Check to see if it is available in any of the predecessor blocks.
1363 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1364 // If we already scanned this predecessor, skip it.
1365 if (!PredsScanned.insert(PredBB).second)
1366 continue;
1368 BBIt = PredBB->end();
1369 unsigned NumScanedInst = 0;
1370 Value *PredAvailable = nullptr;
1371 // NOTE: We don't CSE load that is volatile or anything stronger than
1372 // unordered, that should have been checked when we entered the function.
1373 assert(LoadI->isUnordered() &&
1374 "Attempting to CSE volatile or atomic loads");
1375 // If this is a load on a phi pointer, phi-translate it and search
1376 // for available load/store to the pointer in predecessors.
1377 Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1378 PredAvailable = FindAvailablePtrLoadStore(
1379 Ptr, LoadI->getType(), LoadI->isAtomic(), PredBB, BBIt,
1380 DefMaxInstsToScan, AA, &IsLoadCSE, &NumScanedInst);
1382 // If PredBB has a single predecessor, continue scanning through the
1383 // single predecessor.
1384 BasicBlock *SinglePredBB = PredBB;
1385 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1386 NumScanedInst < DefMaxInstsToScan) {
1387 SinglePredBB = SinglePredBB->getSinglePredecessor();
1388 if (SinglePredBB) {
1389 BBIt = SinglePredBB->end();
1390 PredAvailable = FindAvailablePtrLoadStore(
1391 Ptr, LoadI->getType(), LoadI->isAtomic(), SinglePredBB, BBIt,
1392 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1393 &NumScanedInst);
1397 if (!PredAvailable) {
1398 OneUnavailablePred = PredBB;
1399 continue;
1402 if (IsLoadCSE)
1403 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1405 // If so, this load is partially redundant. Remember this info so that we
1406 // can create a PHI node.
1407 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1410 // If the loaded value isn't available in any predecessor, it isn't partially
1411 // redundant.
1412 if (AvailablePreds.empty()) return false;
1414 // Okay, the loaded value is available in at least one (and maybe all!)
1415 // predecessors. If the value is unavailable in more than one unique
1416 // predecessor, we want to insert a merge block for those common predecessors.
1417 // This ensures that we only have to insert one reload, thus not increasing
1418 // code size.
1419 BasicBlock *UnavailablePred = nullptr;
1421 // If the value is unavailable in one of predecessors, we will end up
1422 // inserting a new instruction into them. It is only valid if all the
1423 // instructions before LoadI are guaranteed to pass execution to its
1424 // successor, or if LoadI is safe to speculate.
1425 // TODO: If this logic becomes more complex, and we will perform PRE insertion
1426 // farther than to a predecessor, we need to reuse the code from GVN's PRE.
1427 // It requires domination tree analysis, so for this simple case it is an
1428 // overkill.
1429 if (PredsScanned.size() != AvailablePreds.size() &&
1430 !isSafeToSpeculativelyExecute(LoadI))
1431 for (auto I = LoadBB->begin(); &*I != LoadI; ++I)
1432 if (!isGuaranteedToTransferExecutionToSuccessor(&*I))
1433 return false;
1435 // If there is exactly one predecessor where the value is unavailable, the
1436 // already computed 'OneUnavailablePred' block is it. If it ends in an
1437 // unconditional branch, we know that it isn't a critical edge.
1438 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1439 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1440 UnavailablePred = OneUnavailablePred;
1441 } else if (PredsScanned.size() != AvailablePreds.size()) {
1442 // Otherwise, we had multiple unavailable predecessors or we had a critical
1443 // edge from the one.
1444 SmallVector<BasicBlock*, 8> PredsToSplit;
1445 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1447 for (const auto &AvailablePred : AvailablePreds)
1448 AvailablePredSet.insert(AvailablePred.first);
1450 // Add all the unavailable predecessors to the PredsToSplit list.
1451 for (BasicBlock *P : predecessors(LoadBB)) {
1452 // If the predecessor is an indirect goto, we can't split the edge.
1453 // Same for CallBr.
1454 if (isa<IndirectBrInst>(P->getTerminator()) ||
1455 isa<CallBrInst>(P->getTerminator()))
1456 return false;
1458 if (!AvailablePredSet.count(P))
1459 PredsToSplit.push_back(P);
1462 // Split them out to their own block.
1463 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1466 // If the value isn't available in all predecessors, then there will be
1467 // exactly one where it isn't available. Insert a load on that edge and add
1468 // it to the AvailablePreds list.
1469 if (UnavailablePred) {
1470 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1471 "Can't handle critical edge here!");
1472 LoadInst *NewVal = new LoadInst(
1473 LoadI->getType(), LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1474 LoadI->getName() + ".pr", false, MaybeAlign(LoadI->getAlignment()),
1475 LoadI->getOrdering(), LoadI->getSyncScopeID(),
1476 UnavailablePred->getTerminator());
1477 NewVal->setDebugLoc(LoadI->getDebugLoc());
1478 if (AATags)
1479 NewVal->setAAMetadata(AATags);
1481 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1484 // Now we know that each predecessor of this block has a value in
1485 // AvailablePreds, sort them for efficient access as we're walking the preds.
1486 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1488 // Create a PHI node at the start of the block for the PRE'd load value.
1489 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1490 PHINode *PN = PHINode::Create(LoadI->getType(), std::distance(PB, PE), "",
1491 &LoadBB->front());
1492 PN->takeName(LoadI);
1493 PN->setDebugLoc(LoadI->getDebugLoc());
1495 // Insert new entries into the PHI for each predecessor. A single block may
1496 // have multiple entries here.
1497 for (pred_iterator PI = PB; PI != PE; ++PI) {
1498 BasicBlock *P = *PI;
1499 AvailablePredsTy::iterator I =
1500 llvm::lower_bound(AvailablePreds, std::make_pair(P, (Value *)nullptr));
1502 assert(I != AvailablePreds.end() && I->first == P &&
1503 "Didn't find entry for predecessor!");
1505 // If we have an available predecessor but it requires casting, insert the
1506 // cast in the predecessor and use the cast. Note that we have to update the
1507 // AvailablePreds vector as we go so that all of the PHI entries for this
1508 // predecessor use the same bitcast.
1509 Value *&PredV = I->second;
1510 if (PredV->getType() != LoadI->getType())
1511 PredV = CastInst::CreateBitOrPointerCast(PredV, LoadI->getType(), "",
1512 P->getTerminator());
1514 PN->addIncoming(PredV, I->first);
1517 for (LoadInst *PredLoadI : CSELoads) {
1518 combineMetadataForCSE(PredLoadI, LoadI, true);
1521 LoadI->replaceAllUsesWith(PN);
1522 LoadI->eraseFromParent();
1524 return true;
1527 /// FindMostPopularDest - The specified list contains multiple possible
1528 /// threadable destinations. Pick the one that occurs the most frequently in
1529 /// the list.
1530 static BasicBlock *
1531 FindMostPopularDest(BasicBlock *BB,
1532 const SmallVectorImpl<std::pair<BasicBlock *,
1533 BasicBlock *>> &PredToDestList) {
1534 assert(!PredToDestList.empty());
1536 // Determine popularity. If there are multiple possible destinations, we
1537 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1538 // blocks with known and real destinations to threading undef. We'll handle
1539 // them later if interesting.
1540 DenseMap<BasicBlock*, unsigned> DestPopularity;
1541 for (const auto &PredToDest : PredToDestList)
1542 if (PredToDest.second)
1543 DestPopularity[PredToDest.second]++;
1545 if (DestPopularity.empty())
1546 return nullptr;
1548 // Find the most popular dest.
1549 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1550 BasicBlock *MostPopularDest = DPI->first;
1551 unsigned Popularity = DPI->second;
1552 SmallVector<BasicBlock*, 4> SamePopularity;
1554 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1555 // If the popularity of this entry isn't higher than the popularity we've
1556 // seen so far, ignore it.
1557 if (DPI->second < Popularity)
1558 ; // ignore.
1559 else if (DPI->second == Popularity) {
1560 // If it is the same as what we've seen so far, keep track of it.
1561 SamePopularity.push_back(DPI->first);
1562 } else {
1563 // If it is more popular, remember it.
1564 SamePopularity.clear();
1565 MostPopularDest = DPI->first;
1566 Popularity = DPI->second;
1570 // Okay, now we know the most popular destination. If there is more than one
1571 // destination, we need to determine one. This is arbitrary, but we need
1572 // to make a deterministic decision. Pick the first one that appears in the
1573 // successor list.
1574 if (!SamePopularity.empty()) {
1575 SamePopularity.push_back(MostPopularDest);
1576 Instruction *TI = BB->getTerminator();
1577 for (unsigned i = 0; ; ++i) {
1578 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1580 if (!is_contained(SamePopularity, TI->getSuccessor(i)))
1581 continue;
1583 MostPopularDest = TI->getSuccessor(i);
1584 break;
1588 // Okay, we have finally picked the most popular destination.
1589 return MostPopularDest;
1592 bool JumpThreadingPass::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1593 ConstantPreference Preference,
1594 Instruction *CxtI) {
1595 // If threading this would thread across a loop header, don't even try to
1596 // thread the edge.
1597 if (LoopHeaders.count(BB))
1598 return false;
1600 PredValueInfoTy PredValues;
1601 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1602 return false;
1604 assert(!PredValues.empty() &&
1605 "ComputeValueKnownInPredecessors returned true with no values");
1607 LLVM_DEBUG(dbgs() << "IN BB: " << *BB;
1608 for (const auto &PredValue : PredValues) {
1609 dbgs() << " BB '" << BB->getName()
1610 << "': FOUND condition = " << *PredValue.first
1611 << " for pred '" << PredValue.second->getName() << "'.\n";
1614 // Decide what we want to thread through. Convert our list of known values to
1615 // a list of known destinations for each pred. This also discards duplicate
1616 // predecessors and keeps track of the undefined inputs (which are represented
1617 // as a null dest in the PredToDestList).
1618 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1619 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1621 BasicBlock *OnlyDest = nullptr;
1622 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1623 Constant *OnlyVal = nullptr;
1624 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1626 for (const auto &PredValue : PredValues) {
1627 BasicBlock *Pred = PredValue.second;
1628 if (!SeenPreds.insert(Pred).second)
1629 continue; // Duplicate predecessor entry.
1631 Constant *Val = PredValue.first;
1633 BasicBlock *DestBB;
1634 if (isa<UndefValue>(Val))
1635 DestBB = nullptr;
1636 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1637 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1638 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1639 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1640 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1641 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1642 } else {
1643 assert(isa<IndirectBrInst>(BB->getTerminator())
1644 && "Unexpected terminator");
1645 assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress");
1646 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1649 // If we have exactly one destination, remember it for efficiency below.
1650 if (PredToDestList.empty()) {
1651 OnlyDest = DestBB;
1652 OnlyVal = Val;
1653 } else {
1654 if (OnlyDest != DestBB)
1655 OnlyDest = MultipleDestSentinel;
1656 // It possible we have same destination, but different value, e.g. default
1657 // case in switchinst.
1658 if (Val != OnlyVal)
1659 OnlyVal = MultipleVal;
1662 // If the predecessor ends with an indirect goto, we can't change its
1663 // destination. Same for CallBr.
1664 if (isa<IndirectBrInst>(Pred->getTerminator()) ||
1665 isa<CallBrInst>(Pred->getTerminator()))
1666 continue;
1668 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1671 // If all edges were unthreadable, we fail.
1672 if (PredToDestList.empty())
1673 return false;
1675 // If all the predecessors go to a single known successor, we want to fold,
1676 // not thread. By doing so, we do not need to duplicate the current block and
1677 // also miss potential opportunities in case we dont/cant duplicate.
1678 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1679 if (BB->hasNPredecessors(PredToDestList.size())) {
1680 bool SeenFirstBranchToOnlyDest = false;
1681 std::vector <DominatorTree::UpdateType> Updates;
1682 Updates.reserve(BB->getTerminator()->getNumSuccessors() - 1);
1683 for (BasicBlock *SuccBB : successors(BB)) {
1684 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest) {
1685 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1686 } else {
1687 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1688 Updates.push_back({DominatorTree::Delete, BB, SuccBB});
1692 // Finally update the terminator.
1693 Instruction *Term = BB->getTerminator();
1694 BranchInst::Create(OnlyDest, Term);
1695 Term->eraseFromParent();
1696 DTU->applyUpdatesPermissive(Updates);
1698 // If the condition is now dead due to the removal of the old terminator,
1699 // erase it.
1700 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1701 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1702 CondInst->eraseFromParent();
1703 // We can safely replace *some* uses of the CondInst if it has
1704 // exactly one value as returned by LVI. RAUW is incorrect in the
1705 // presence of guards and assumes, that have the `Cond` as the use. This
1706 // is because we use the guards/assume to reason about the `Cond` value
1707 // at the end of block, but RAUW unconditionally replaces all uses
1708 // including the guards/assumes themselves and the uses before the
1709 // guard/assume.
1710 else if (OnlyVal && OnlyVal != MultipleVal &&
1711 CondInst->getParent() == BB)
1712 ReplaceFoldableUses(CondInst, OnlyVal);
1714 return true;
1718 // Determine which is the most common successor. If we have many inputs and
1719 // this block is a switch, we want to start by threading the batch that goes
1720 // to the most popular destination first. If we only know about one
1721 // threadable destination (the common case) we can avoid this.
1722 BasicBlock *MostPopularDest = OnlyDest;
1724 if (MostPopularDest == MultipleDestSentinel) {
1725 // Remove any loop headers from the Dest list, ThreadEdge conservatively
1726 // won't process them, but we might have other destination that are eligible
1727 // and we still want to process.
1728 erase_if(PredToDestList,
1729 [&](const std::pair<BasicBlock *, BasicBlock *> &PredToDest) {
1730 return LoopHeaders.count(PredToDest.second) != 0;
1733 if (PredToDestList.empty())
1734 return false;
1736 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1739 // Now that we know what the most popular destination is, factor all
1740 // predecessors that will jump to it into a single predecessor.
1741 SmallVector<BasicBlock*, 16> PredsToFactor;
1742 for (const auto &PredToDest : PredToDestList)
1743 if (PredToDest.second == MostPopularDest) {
1744 BasicBlock *Pred = PredToDest.first;
1746 // This predecessor may be a switch or something else that has multiple
1747 // edges to the block. Factor each of these edges by listing them
1748 // according to # occurrences in PredsToFactor.
1749 for (BasicBlock *Succ : successors(Pred))
1750 if (Succ == BB)
1751 PredsToFactor.push_back(Pred);
1754 // If the threadable edges are branching on an undefined value, we get to pick
1755 // the destination that these predecessors should get to.
1756 if (!MostPopularDest)
1757 MostPopularDest = BB->getTerminator()->
1758 getSuccessor(GetBestDestForJumpOnUndef(BB));
1760 // Ok, try to thread it!
1761 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1764 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1765 /// a PHI node in the current block. See if there are any simplifications we
1766 /// can do based on inputs to the phi node.
1767 bool JumpThreadingPass::ProcessBranchOnPHI(PHINode *PN) {
1768 BasicBlock *BB = PN->getParent();
1770 // TODO: We could make use of this to do it once for blocks with common PHI
1771 // values.
1772 SmallVector<BasicBlock*, 1> PredBBs;
1773 PredBBs.resize(1);
1775 // If any of the predecessor blocks end in an unconditional branch, we can
1776 // *duplicate* the conditional branch into that block in order to further
1777 // encourage jump threading and to eliminate cases where we have branch on a
1778 // phi of an icmp (branch on icmp is much better).
1779 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1780 BasicBlock *PredBB = PN->getIncomingBlock(i);
1781 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1782 if (PredBr->isUnconditional()) {
1783 PredBBs[0] = PredBB;
1784 // Try to duplicate BB into PredBB.
1785 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1786 return true;
1790 return false;
1793 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1794 /// a xor instruction in the current block. See if there are any
1795 /// simplifications we can do based on inputs to the xor.
1796 bool JumpThreadingPass::ProcessBranchOnXOR(BinaryOperator *BO) {
1797 BasicBlock *BB = BO->getParent();
1799 // If either the LHS or RHS of the xor is a constant, don't do this
1800 // optimization.
1801 if (isa<ConstantInt>(BO->getOperand(0)) ||
1802 isa<ConstantInt>(BO->getOperand(1)))
1803 return false;
1805 // If the first instruction in BB isn't a phi, we won't be able to infer
1806 // anything special about any particular predecessor.
1807 if (!isa<PHINode>(BB->front()))
1808 return false;
1810 // If this BB is a landing pad, we won't be able to split the edge into it.
1811 if (BB->isEHPad())
1812 return false;
1814 // If we have a xor as the branch input to this block, and we know that the
1815 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1816 // the condition into the predecessor and fix that value to true, saving some
1817 // logical ops on that path and encouraging other paths to simplify.
1819 // This copies something like this:
1821 // BB:
1822 // %X = phi i1 [1], [%X']
1823 // %Y = icmp eq i32 %A, %B
1824 // %Z = xor i1 %X, %Y
1825 // br i1 %Z, ...
1827 // Into:
1828 // BB':
1829 // %Y = icmp ne i32 %A, %B
1830 // br i1 %Y, ...
1832 PredValueInfoTy XorOpValues;
1833 bool isLHS = true;
1834 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1835 WantInteger, BO)) {
1836 assert(XorOpValues.empty());
1837 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1838 WantInteger, BO))
1839 return false;
1840 isLHS = false;
1843 assert(!XorOpValues.empty() &&
1844 "ComputeValueKnownInPredecessors returned true with no values");
1846 // Scan the information to see which is most popular: true or false. The
1847 // predecessors can be of the set true, false, or undef.
1848 unsigned NumTrue = 0, NumFalse = 0;
1849 for (const auto &XorOpValue : XorOpValues) {
1850 if (isa<UndefValue>(XorOpValue.first))
1851 // Ignore undefs for the count.
1852 continue;
1853 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1854 ++NumFalse;
1855 else
1856 ++NumTrue;
1859 // Determine which value to split on, true, false, or undef if neither.
1860 ConstantInt *SplitVal = nullptr;
1861 if (NumTrue > NumFalse)
1862 SplitVal = ConstantInt::getTrue(BB->getContext());
1863 else if (NumTrue != 0 || NumFalse != 0)
1864 SplitVal = ConstantInt::getFalse(BB->getContext());
1866 // Collect all of the blocks that this can be folded into so that we can
1867 // factor this once and clone it once.
1868 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1869 for (const auto &XorOpValue : XorOpValues) {
1870 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1871 continue;
1873 BlocksToFoldInto.push_back(XorOpValue.second);
1876 // If we inferred a value for all of the predecessors, then duplication won't
1877 // help us. However, we can just replace the LHS or RHS with the constant.
1878 if (BlocksToFoldInto.size() ==
1879 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1880 if (!SplitVal) {
1881 // If all preds provide undef, just nuke the xor, because it is undef too.
1882 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1883 BO->eraseFromParent();
1884 } else if (SplitVal->isZero()) {
1885 // If all preds provide 0, replace the xor with the other input.
1886 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1887 BO->eraseFromParent();
1888 } else {
1889 // If all preds provide 1, set the computed value to 1.
1890 BO->setOperand(!isLHS, SplitVal);
1893 return true;
1896 // Try to duplicate BB into PredBB.
1897 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1900 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1901 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1902 /// NewPred using the entries from OldPred (suitably mapped).
1903 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1904 BasicBlock *OldPred,
1905 BasicBlock *NewPred,
1906 DenseMap<Instruction*, Value*> &ValueMap) {
1907 for (PHINode &PN : PHIBB->phis()) {
1908 // Ok, we have a PHI node. Figure out what the incoming value was for the
1909 // DestBlock.
1910 Value *IV = PN.getIncomingValueForBlock(OldPred);
1912 // Remap the value if necessary.
1913 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1914 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1915 if (I != ValueMap.end())
1916 IV = I->second;
1919 PN.addIncoming(IV, NewPred);
1923 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1924 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1925 /// across BB. Transform the IR to reflect this change.
1926 bool JumpThreadingPass::ThreadEdge(BasicBlock *BB,
1927 const SmallVectorImpl<BasicBlock *> &PredBBs,
1928 BasicBlock *SuccBB) {
1929 // If threading to the same block as we come from, we would infinite loop.
1930 if (SuccBB == BB) {
1931 LLVM_DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1932 << "' - would thread to self!\n");
1933 return false;
1936 // If threading this would thread across a loop header, don't thread the edge.
1937 // See the comments above FindLoopHeaders for justifications and caveats.
1938 if (LoopHeaders.count(BB) || LoopHeaders.count(SuccBB)) {
1939 LLVM_DEBUG({
1940 bool BBIsHeader = LoopHeaders.count(BB);
1941 bool SuccIsHeader = LoopHeaders.count(SuccBB);
1942 dbgs() << " Not threading across "
1943 << (BBIsHeader ? "loop header BB '" : "block BB '") << BB->getName()
1944 << "' to dest " << (SuccIsHeader ? "loop header BB '" : "block BB '")
1945 << SuccBB->getName() << "' - it might create an irreducible loop!\n";
1947 return false;
1950 unsigned JumpThreadCost =
1951 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1952 if (JumpThreadCost > BBDupThreshold) {
1953 LLVM_DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1954 << "' - Cost is too high: " << JumpThreadCost << "\n");
1955 return false;
1958 // And finally, do it! Start by factoring the predecessors if needed.
1959 BasicBlock *PredBB;
1960 if (PredBBs.size() == 1)
1961 PredBB = PredBBs[0];
1962 else {
1963 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1964 << " common predecessors.\n");
1965 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1968 // And finally, do it!
1969 LLVM_DEBUG(dbgs() << " Threading edge from '" << PredBB->getName()
1970 << "' to '" << SuccBB->getName()
1971 << "' with cost: " << JumpThreadCost
1972 << ", across block:\n " << *BB << "\n");
1974 if (DTU->hasPendingDomTreeUpdates())
1975 LVI->disableDT();
1976 else
1977 LVI->enableDT();
1978 LVI->threadEdge(PredBB, BB, SuccBB);
1980 // We are going to have to map operands from the original BB block to the new
1981 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1982 // account for entry from PredBB.
1983 DenseMap<Instruction*, Value*> ValueMapping;
1985 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1986 BB->getName()+".thread",
1987 BB->getParent(), BB);
1988 NewBB->moveAfter(PredBB);
1990 // Set the block frequency of NewBB.
1991 if (HasProfileData) {
1992 auto NewBBFreq =
1993 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1994 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1997 BasicBlock::iterator BI = BB->begin();
1998 // Clone the phi nodes of BB into NewBB. The resulting phi nodes are trivial,
1999 // since NewBB only has one predecessor, but SSAUpdater might need to rewrite
2000 // the operand of the cloned phi.
2001 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2002 PHINode *NewPN = PHINode::Create(PN->getType(), 1, PN->getName(), NewBB);
2003 NewPN->addIncoming(PN->getIncomingValueForBlock(PredBB), PredBB);
2004 ValueMapping[PN] = NewPN;
2007 // Clone the non-phi instructions of BB into NewBB, keeping track of the
2008 // mapping and using it to remap operands in the cloned instructions.
2009 for (; !BI->isTerminator(); ++BI) {
2010 Instruction *New = BI->clone();
2011 New->setName(BI->getName());
2012 NewBB->getInstList().push_back(New);
2013 ValueMapping[&*BI] = New;
2015 // Remap operands to patch up intra-block references.
2016 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2017 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2018 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2019 if (I != ValueMapping.end())
2020 New->setOperand(i, I->second);
2024 // We didn't copy the terminator from BB over to NewBB, because there is now
2025 // an unconditional jump to SuccBB. Insert the unconditional jump.
2026 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
2027 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
2029 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
2030 // PHI nodes for NewBB now.
2031 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
2033 // Update the terminator of PredBB to jump to NewBB instead of BB. This
2034 // eliminates predecessors from BB, which requires us to simplify any PHI
2035 // nodes in BB.
2036 Instruction *PredTerm = PredBB->getTerminator();
2037 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
2038 if (PredTerm->getSuccessor(i) == BB) {
2039 BB->removePredecessor(PredBB, true);
2040 PredTerm->setSuccessor(i, NewBB);
2043 // Enqueue required DT updates.
2044 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, SuccBB},
2045 {DominatorTree::Insert, PredBB, NewBB},
2046 {DominatorTree::Delete, PredBB, BB}});
2048 // If there were values defined in BB that are used outside the block, then we
2049 // now have to update all uses of the value to use either the original value,
2050 // the cloned value, or some PHI derived value. This can require arbitrary
2051 // PHI insertion, of which we are prepared to do, clean these up now.
2052 SSAUpdater SSAUpdate;
2053 SmallVector<Use*, 16> UsesToRename;
2055 for (Instruction &I : *BB) {
2056 // Scan all uses of this instruction to see if their uses are no longer
2057 // dominated by the previous def and if so, record them in UsesToRename.
2058 // Also, skip phi operands from PredBB - we'll remove them anyway.
2059 for (Use &U : I.uses()) {
2060 Instruction *User = cast<Instruction>(U.getUser());
2061 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2062 if (UserPN->getIncomingBlock(U) == BB)
2063 continue;
2064 } else if (User->getParent() == BB)
2065 continue;
2067 UsesToRename.push_back(&U);
2070 // If there are no uses outside the block, we're done with this instruction.
2071 if (UsesToRename.empty())
2072 continue;
2073 LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
2075 // We found a use of I outside of BB. Rename all uses of I that are outside
2076 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2077 // with the two values we know.
2078 SSAUpdate.Initialize(I.getType(), I.getName());
2079 SSAUpdate.AddAvailableValue(BB, &I);
2080 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
2082 while (!UsesToRename.empty())
2083 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2084 LLVM_DEBUG(dbgs() << "\n");
2087 // At this point, the IR is fully up to date and consistent. Do a quick scan
2088 // over the new instructions and zap any that are constants or dead. This
2089 // frequently happens because of phi translation.
2090 SimplifyInstructionsInBlock(NewBB, TLI);
2092 // Update the edge weight from BB to SuccBB, which should be less than before.
2093 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
2095 // Threaded an edge!
2096 ++NumThreads;
2097 return true;
2100 /// Create a new basic block that will be the predecessor of BB and successor of
2101 /// all blocks in Preds. When profile data is available, update the frequency of
2102 /// this new block.
2103 BasicBlock *JumpThreadingPass::SplitBlockPreds(BasicBlock *BB,
2104 ArrayRef<BasicBlock *> Preds,
2105 const char *Suffix) {
2106 SmallVector<BasicBlock *, 2> NewBBs;
2108 // Collect the frequencies of all predecessors of BB, which will be used to
2109 // update the edge weight of the result of splitting predecessors.
2110 DenseMap<BasicBlock *, BlockFrequency> FreqMap;
2111 if (HasProfileData)
2112 for (auto Pred : Preds)
2113 FreqMap.insert(std::make_pair(
2114 Pred, BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB)));
2116 // In the case when BB is a LandingPad block we create 2 new predecessors
2117 // instead of just one.
2118 if (BB->isLandingPad()) {
2119 std::string NewName = std::string(Suffix) + ".split-lp";
2120 SplitLandingPadPredecessors(BB, Preds, Suffix, NewName.c_str(), NewBBs);
2121 } else {
2122 NewBBs.push_back(SplitBlockPredecessors(BB, Preds, Suffix));
2125 std::vector<DominatorTree::UpdateType> Updates;
2126 Updates.reserve((2 * Preds.size()) + NewBBs.size());
2127 for (auto NewBB : NewBBs) {
2128 BlockFrequency NewBBFreq(0);
2129 Updates.push_back({DominatorTree::Insert, NewBB, BB});
2130 for (auto Pred : predecessors(NewBB)) {
2131 Updates.push_back({DominatorTree::Delete, Pred, BB});
2132 Updates.push_back({DominatorTree::Insert, Pred, NewBB});
2133 if (HasProfileData) // Update frequencies between Pred -> NewBB.
2134 NewBBFreq += FreqMap.lookup(Pred);
2136 if (HasProfileData) // Apply the summed frequency to NewBB.
2137 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
2140 DTU->applyUpdatesPermissive(Updates);
2141 return NewBBs[0];
2144 bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
2145 const Instruction *TI = BB->getTerminator();
2146 assert(TI->getNumSuccessors() > 1 && "not a split");
2148 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
2149 if (!WeightsNode)
2150 return false;
2152 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
2153 if (MDName->getString() != "branch_weights")
2154 return false;
2156 // Ensure there are weights for all of the successors. Note that the first
2157 // operand to the metadata node is a name, not a weight.
2158 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
2161 /// Update the block frequency of BB and branch weight and the metadata on the
2162 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
2163 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
2164 void JumpThreadingPass::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
2165 BasicBlock *BB,
2166 BasicBlock *NewBB,
2167 BasicBlock *SuccBB) {
2168 if (!HasProfileData)
2169 return;
2171 assert(BFI && BPI && "BFI & BPI should have been created here");
2173 // As the edge from PredBB to BB is deleted, we have to update the block
2174 // frequency of BB.
2175 auto BBOrigFreq = BFI->getBlockFreq(BB);
2176 auto NewBBFreq = BFI->getBlockFreq(NewBB);
2177 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
2178 auto BBNewFreq = BBOrigFreq - NewBBFreq;
2179 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
2181 // Collect updated outgoing edges' frequencies from BB and use them to update
2182 // edge probabilities.
2183 SmallVector<uint64_t, 4> BBSuccFreq;
2184 for (BasicBlock *Succ : successors(BB)) {
2185 auto SuccFreq = (Succ == SuccBB)
2186 ? BB2SuccBBFreq - NewBBFreq
2187 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
2188 BBSuccFreq.push_back(SuccFreq.getFrequency());
2191 uint64_t MaxBBSuccFreq =
2192 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
2194 SmallVector<BranchProbability, 4> BBSuccProbs;
2195 if (MaxBBSuccFreq == 0)
2196 BBSuccProbs.assign(BBSuccFreq.size(),
2197 {1, static_cast<unsigned>(BBSuccFreq.size())});
2198 else {
2199 for (uint64_t Freq : BBSuccFreq)
2200 BBSuccProbs.push_back(
2201 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
2202 // Normalize edge probabilities so that they sum up to one.
2203 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
2204 BBSuccProbs.end());
2207 // Update edge probabilities in BPI.
2208 for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
2209 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
2211 // Update the profile metadata as well.
2213 // Don't do this if the profile of the transformed blocks was statically
2214 // estimated. (This could occur despite the function having an entry
2215 // frequency in completely cold parts of the CFG.)
2217 // In this case we don't want to suggest to subsequent passes that the
2218 // calculated weights are fully consistent. Consider this graph:
2220 // check_1
2221 // 50% / |
2222 // eq_1 | 50%
2223 // \ |
2224 // check_2
2225 // 50% / |
2226 // eq_2 | 50%
2227 // \ |
2228 // check_3
2229 // 50% / |
2230 // eq_3 | 50%
2231 // \ |
2233 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
2234 // the overall probabilities are inconsistent; the total probability that the
2235 // value is either 1, 2 or 3 is 150%.
2237 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
2238 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
2239 // the loop exit edge. Then based solely on static estimation we would assume
2240 // the loop was extremely hot.
2242 // FIXME this locally as well so that BPI and BFI are consistent as well. We
2243 // shouldn't make edges extremely likely or unlikely based solely on static
2244 // estimation.
2245 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
2246 SmallVector<uint32_t, 4> Weights;
2247 for (auto Prob : BBSuccProbs)
2248 Weights.push_back(Prob.getNumerator());
2250 auto TI = BB->getTerminator();
2251 TI->setMetadata(
2252 LLVMContext::MD_prof,
2253 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
2257 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
2258 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
2259 /// If we can duplicate the contents of BB up into PredBB do so now, this
2260 /// improves the odds that the branch will be on an analyzable instruction like
2261 /// a compare.
2262 bool JumpThreadingPass::DuplicateCondBranchOnPHIIntoPred(
2263 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
2264 assert(!PredBBs.empty() && "Can't handle an empty set");
2266 // If BB is a loop header, then duplicating this block outside the loop would
2267 // cause us to transform this into an irreducible loop, don't do this.
2268 // See the comments above FindLoopHeaders for justifications and caveats.
2269 if (LoopHeaders.count(BB)) {
2270 LLVM_DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
2271 << "' into predecessor block '" << PredBBs[0]->getName()
2272 << "' - it might create an irreducible loop!\n");
2273 return false;
2276 unsigned DuplicationCost =
2277 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
2278 if (DuplicationCost > BBDupThreshold) {
2279 LLVM_DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
2280 << "' - Cost is too high: " << DuplicationCost << "\n");
2281 return false;
2284 // And finally, do it! Start by factoring the predecessors if needed.
2285 std::vector<DominatorTree::UpdateType> Updates;
2286 BasicBlock *PredBB;
2287 if (PredBBs.size() == 1)
2288 PredBB = PredBBs[0];
2289 else {
2290 LLVM_DEBUG(dbgs() << " Factoring out " << PredBBs.size()
2291 << " common predecessors.\n");
2292 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
2294 Updates.push_back({DominatorTree::Delete, PredBB, BB});
2296 // Okay, we decided to do this! Clone all the instructions in BB onto the end
2297 // of PredBB.
2298 LLVM_DEBUG(dbgs() << " Duplicating block '" << BB->getName()
2299 << "' into end of '" << PredBB->getName()
2300 << "' to eliminate branch on phi. Cost: "
2301 << DuplicationCost << " block is:" << *BB << "\n");
2303 // Unless PredBB ends with an unconditional branch, split the edge so that we
2304 // can just clone the bits from BB into the end of the new PredBB.
2305 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
2307 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
2308 BasicBlock *OldPredBB = PredBB;
2309 PredBB = SplitEdge(OldPredBB, BB);
2310 Updates.push_back({DominatorTree::Insert, OldPredBB, PredBB});
2311 Updates.push_back({DominatorTree::Insert, PredBB, BB});
2312 Updates.push_back({DominatorTree::Delete, OldPredBB, BB});
2313 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
2316 // We are going to have to map operands from the original BB block into the
2317 // PredBB block. Evaluate PHI nodes in BB.
2318 DenseMap<Instruction*, Value*> ValueMapping;
2320 BasicBlock::iterator BI = BB->begin();
2321 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2322 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2323 // Clone the non-phi instructions of BB into PredBB, keeping track of the
2324 // mapping and using it to remap operands in the cloned instructions.
2325 for (; BI != BB->end(); ++BI) {
2326 Instruction *New = BI->clone();
2328 // Remap operands to patch up intra-block references.
2329 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2330 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2331 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2332 if (I != ValueMapping.end())
2333 New->setOperand(i, I->second);
2336 // If this instruction can be simplified after the operands are updated,
2337 // just use the simplified value instead. This frequently happens due to
2338 // phi translation.
2339 if (Value *IV = SimplifyInstruction(
2340 New,
2341 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2342 ValueMapping[&*BI] = IV;
2343 if (!New->mayHaveSideEffects()) {
2344 New->deleteValue();
2345 New = nullptr;
2347 } else {
2348 ValueMapping[&*BI] = New;
2350 if (New) {
2351 // Otherwise, insert the new instruction into the block.
2352 New->setName(BI->getName());
2353 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2354 // Update Dominance from simplified New instruction operands.
2355 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2356 if (BasicBlock *SuccBB = dyn_cast<BasicBlock>(New->getOperand(i)))
2357 Updates.push_back({DominatorTree::Insert, PredBB, SuccBB});
2361 // Check to see if the targets of the branch had PHI nodes. If so, we need to
2362 // add entries to the PHI nodes for branch from PredBB now.
2363 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2364 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2365 ValueMapping);
2366 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2367 ValueMapping);
2369 // If there were values defined in BB that are used outside the block, then we
2370 // now have to update all uses of the value to use either the original value,
2371 // the cloned value, or some PHI derived value. This can require arbitrary
2372 // PHI insertion, of which we are prepared to do, clean these up now.
2373 SSAUpdater SSAUpdate;
2374 SmallVector<Use*, 16> UsesToRename;
2375 for (Instruction &I : *BB) {
2376 // Scan all uses of this instruction to see if it is used outside of its
2377 // block, and if so, record them in UsesToRename.
2378 for (Use &U : I.uses()) {
2379 Instruction *User = cast<Instruction>(U.getUser());
2380 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2381 if (UserPN->getIncomingBlock(U) == BB)
2382 continue;
2383 } else if (User->getParent() == BB)
2384 continue;
2386 UsesToRename.push_back(&U);
2389 // If there are no uses outside the block, we're done with this instruction.
2390 if (UsesToRename.empty())
2391 continue;
2393 LLVM_DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
2395 // We found a use of I outside of BB. Rename all uses of I that are outside
2396 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2397 // with the two values we know.
2398 SSAUpdate.Initialize(I.getType(), I.getName());
2399 SSAUpdate.AddAvailableValue(BB, &I);
2400 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]);
2402 while (!UsesToRename.empty())
2403 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2404 LLVM_DEBUG(dbgs() << "\n");
2407 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2408 // that we nuked.
2409 BB->removePredecessor(PredBB, true);
2411 // Remove the unconditional branch at the end of the PredBB block.
2412 OldPredBranch->eraseFromParent();
2413 DTU->applyUpdatesPermissive(Updates);
2415 ++NumDupes;
2416 return true;
2419 // Pred is a predecessor of BB with an unconditional branch to BB. SI is
2420 // a Select instruction in Pred. BB has other predecessors and SI is used in
2421 // a PHI node in BB. SI has no other use.
2422 // A new basic block, NewBB, is created and SI is converted to compare and
2423 // conditional branch. SI is erased from parent.
2424 void JumpThreadingPass::UnfoldSelectInstr(BasicBlock *Pred, BasicBlock *BB,
2425 SelectInst *SI, PHINode *SIUse,
2426 unsigned Idx) {
2427 // Expand the select.
2429 // Pred --
2430 // | v
2431 // | NewBB
2432 // | |
2433 // |-----
2434 // v
2435 // BB
2436 BranchInst *PredTerm = cast<BranchInst>(Pred->getTerminator());
2437 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2438 BB->getParent(), BB);
2439 // Move the unconditional branch to NewBB.
2440 PredTerm->removeFromParent();
2441 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2442 // Create a conditional branch and update PHI nodes.
2443 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2444 SIUse->setIncomingValue(Idx, SI->getFalseValue());
2445 SIUse->addIncoming(SI->getTrueValue(), NewBB);
2447 // The select is now dead.
2448 SI->eraseFromParent();
2449 DTU->applyUpdatesPermissive({{DominatorTree::Insert, NewBB, BB},
2450 {DominatorTree::Insert, Pred, NewBB}});
2452 // Update any other PHI nodes in BB.
2453 for (BasicBlock::iterator BI = BB->begin();
2454 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2455 if (Phi != SIUse)
2456 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2459 bool JumpThreadingPass::TryToUnfoldSelect(SwitchInst *SI, BasicBlock *BB) {
2460 PHINode *CondPHI = dyn_cast<PHINode>(SI->getCondition());
2462 if (!CondPHI || CondPHI->getParent() != BB)
2463 return false;
2465 for (unsigned I = 0, E = CondPHI->getNumIncomingValues(); I != E; ++I) {
2466 BasicBlock *Pred = CondPHI->getIncomingBlock(I);
2467 SelectInst *PredSI = dyn_cast<SelectInst>(CondPHI->getIncomingValue(I));
2469 // The second and third condition can be potentially relaxed. Currently
2470 // the conditions help to simplify the code and allow us to reuse existing
2471 // code, developed for TryToUnfoldSelect(CmpInst *, BasicBlock *)
2472 if (!PredSI || PredSI->getParent() != Pred || !PredSI->hasOneUse())
2473 continue;
2475 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2476 if (!PredTerm || !PredTerm->isUnconditional())
2477 continue;
2479 UnfoldSelectInstr(Pred, BB, PredSI, CondPHI, I);
2480 return true;
2482 return false;
2485 /// TryToUnfoldSelect - Look for blocks of the form
2486 /// bb1:
2487 /// %a = select
2488 /// br bb2
2490 /// bb2:
2491 /// %p = phi [%a, %bb1] ...
2492 /// %c = icmp %p
2493 /// br i1 %c
2495 /// And expand the select into a branch structure if one of its arms allows %c
2496 /// to be folded. This later enables threading from bb1 over bb2.
2497 bool JumpThreadingPass::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2498 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2499 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2500 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2502 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2503 CondLHS->getParent() != BB)
2504 return false;
2506 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2507 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2508 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2510 // Look if one of the incoming values is a select in the corresponding
2511 // predecessor.
2512 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2513 continue;
2515 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2516 if (!PredTerm || !PredTerm->isUnconditional())
2517 continue;
2519 // Now check if one of the select values would allow us to constant fold the
2520 // terminator in BB. We don't do the transform if both sides fold, those
2521 // cases will be threaded in any case.
2522 if (DTU->hasPendingDomTreeUpdates())
2523 LVI->disableDT();
2524 else
2525 LVI->enableDT();
2526 LazyValueInfo::Tristate LHSFolds =
2527 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2528 CondRHS, Pred, BB, CondCmp);
2529 LazyValueInfo::Tristate RHSFolds =
2530 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2531 CondRHS, Pred, BB, CondCmp);
2532 if ((LHSFolds != LazyValueInfo::Unknown ||
2533 RHSFolds != LazyValueInfo::Unknown) &&
2534 LHSFolds != RHSFolds) {
2535 UnfoldSelectInstr(Pred, BB, SI, CondLHS, I);
2536 return true;
2539 return false;
2542 /// TryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2543 /// same BB in the form
2544 /// bb:
2545 /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2546 /// %s = select %p, trueval, falseval
2548 /// or
2550 /// bb:
2551 /// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2552 /// %c = cmp %p, 0
2553 /// %s = select %c, trueval, falseval
2555 /// And expand the select into a branch structure. This later enables
2556 /// jump-threading over bb in this pass.
2558 /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2559 /// select if the associated PHI has at least one constant. If the unfolded
2560 /// select is not jump-threaded, it will be folded again in the later
2561 /// optimizations.
2562 bool JumpThreadingPass::TryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2563 // If threading this would thread across a loop header, don't thread the edge.
2564 // See the comments above FindLoopHeaders for justifications and caveats.
2565 if (LoopHeaders.count(BB))
2566 return false;
2568 for (BasicBlock::iterator BI = BB->begin();
2569 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2570 // Look for a Phi having at least one constant incoming value.
2571 if (llvm::all_of(PN->incoming_values(),
2572 [](Value *V) { return !isa<ConstantInt>(V); }))
2573 continue;
2575 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2576 // Check if SI is in BB and use V as condition.
2577 if (SI->getParent() != BB)
2578 return false;
2579 Value *Cond = SI->getCondition();
2580 return (Cond && Cond == V && Cond->getType()->isIntegerTy(1));
2583 SelectInst *SI = nullptr;
2584 for (Use &U : PN->uses()) {
2585 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2586 // Look for a ICmp in BB that compares PN with a constant and is the
2587 // condition of a Select.
2588 if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2589 isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2590 if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2591 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2592 SI = SelectI;
2593 break;
2595 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2596 // Look for a Select in BB that uses PN as condition.
2597 if (isUnfoldCandidate(SelectI, U.get())) {
2598 SI = SelectI;
2599 break;
2604 if (!SI)
2605 continue;
2606 // Expand the select.
2607 Instruction *Term =
2608 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
2609 BasicBlock *SplitBB = SI->getParent();
2610 BasicBlock *NewBB = Term->getParent();
2611 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2612 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2613 NewPN->addIncoming(SI->getFalseValue(), BB);
2614 SI->replaceAllUsesWith(NewPN);
2615 SI->eraseFromParent();
2616 // NewBB and SplitBB are newly created blocks which require insertion.
2617 std::vector<DominatorTree::UpdateType> Updates;
2618 Updates.reserve((2 * SplitBB->getTerminator()->getNumSuccessors()) + 3);
2619 Updates.push_back({DominatorTree::Insert, BB, SplitBB});
2620 Updates.push_back({DominatorTree::Insert, BB, NewBB});
2621 Updates.push_back({DominatorTree::Insert, NewBB, SplitBB});
2622 // BB's successors were moved to SplitBB, update DTU accordingly.
2623 for (auto *Succ : successors(SplitBB)) {
2624 Updates.push_back({DominatorTree::Delete, BB, Succ});
2625 Updates.push_back({DominatorTree::Insert, SplitBB, Succ});
2627 DTU->applyUpdatesPermissive(Updates);
2628 return true;
2630 return false;
2633 /// Try to propagate a guard from the current BB into one of its predecessors
2634 /// in case if another branch of execution implies that the condition of this
2635 /// guard is always true. Currently we only process the simplest case that
2636 /// looks like:
2638 /// Start:
2639 /// %cond = ...
2640 /// br i1 %cond, label %T1, label %F1
2641 /// T1:
2642 /// br label %Merge
2643 /// F1:
2644 /// br label %Merge
2645 /// Merge:
2646 /// %condGuard = ...
2647 /// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2649 /// And cond either implies condGuard or !condGuard. In this case all the
2650 /// instructions before the guard can be duplicated in both branches, and the
2651 /// guard is then threaded to one of them.
2652 bool JumpThreadingPass::ProcessGuards(BasicBlock *BB) {
2653 using namespace PatternMatch;
2655 // We only want to deal with two predecessors.
2656 BasicBlock *Pred1, *Pred2;
2657 auto PI = pred_begin(BB), PE = pred_end(BB);
2658 if (PI == PE)
2659 return false;
2660 Pred1 = *PI++;
2661 if (PI == PE)
2662 return false;
2663 Pred2 = *PI++;
2664 if (PI != PE)
2665 return false;
2666 if (Pred1 == Pred2)
2667 return false;
2669 // Try to thread one of the guards of the block.
2670 // TODO: Look up deeper than to immediate predecessor?
2671 auto *Parent = Pred1->getSinglePredecessor();
2672 if (!Parent || Parent != Pred2->getSinglePredecessor())
2673 return false;
2675 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2676 for (auto &I : *BB)
2677 if (isGuard(&I) && ThreadGuard(BB, cast<IntrinsicInst>(&I), BI))
2678 return true;
2680 return false;
2683 /// Try to propagate the guard from BB which is the lower block of a diamond
2684 /// to one of its branches, in case if diamond's condition implies guard's
2685 /// condition.
2686 bool JumpThreadingPass::ThreadGuard(BasicBlock *BB, IntrinsicInst *Guard,
2687 BranchInst *BI) {
2688 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?");
2689 assert(BI->isConditional() && "Unconditional branch has 2 successors?");
2690 Value *GuardCond = Guard->getArgOperand(0);
2691 Value *BranchCond = BI->getCondition();
2692 BasicBlock *TrueDest = BI->getSuccessor(0);
2693 BasicBlock *FalseDest = BI->getSuccessor(1);
2695 auto &DL = BB->getModule()->getDataLayout();
2696 bool TrueDestIsSafe = false;
2697 bool FalseDestIsSafe = false;
2699 // True dest is safe if BranchCond => GuardCond.
2700 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
2701 if (Impl && *Impl)
2702 TrueDestIsSafe = true;
2703 else {
2704 // False dest is safe if !BranchCond => GuardCond.
2705 Impl = isImpliedCondition(BranchCond, GuardCond, DL, /* LHSIsTrue */ false);
2706 if (Impl && *Impl)
2707 FalseDestIsSafe = true;
2710 if (!TrueDestIsSafe && !FalseDestIsSafe)
2711 return false;
2713 BasicBlock *PredUnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
2714 BasicBlock *PredGuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
2716 ValueToValueMapTy UnguardedMapping, GuardedMapping;
2717 Instruction *AfterGuard = Guard->getNextNode();
2718 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
2719 if (Cost > BBDupThreshold)
2720 return false;
2721 // Duplicate all instructions before the guard and the guard itself to the
2722 // branch where implication is not proved.
2723 BasicBlock *GuardedBlock = DuplicateInstructionsInSplitBetween(
2724 BB, PredGuardedBlock, AfterGuard, GuardedMapping, *DTU);
2725 assert(GuardedBlock && "Could not create the guarded block?");
2726 // Duplicate all instructions before the guard in the unguarded branch.
2727 // Since we have successfully duplicated the guarded block and this block
2728 // has fewer instructions, we expect it to succeed.
2729 BasicBlock *UnguardedBlock = DuplicateInstructionsInSplitBetween(
2730 BB, PredUnguardedBlock, Guard, UnguardedMapping, *DTU);
2731 assert(UnguardedBlock && "Could not create the unguarded block?");
2732 LLVM_DEBUG(dbgs() << "Moved guard " << *Guard << " to block "
2733 << GuardedBlock->getName() << "\n");
2734 // Some instructions before the guard may still have uses. For them, we need
2735 // to create Phi nodes merging their copies in both guarded and unguarded
2736 // branches. Those instructions that have no uses can be just removed.
2737 SmallVector<Instruction *, 4> ToRemove;
2738 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
2739 if (!isa<PHINode>(&*BI))
2740 ToRemove.push_back(&*BI);
2742 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
2743 assert(InsertionPoint && "Empty block?");
2744 // Substitute with Phis & remove.
2745 for (auto *Inst : reverse(ToRemove)) {
2746 if (!Inst->use_empty()) {
2747 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
2748 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
2749 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
2750 NewPN->insertBefore(InsertionPoint);
2751 Inst->replaceAllUsesWith(NewPN);
2753 Inst->eraseFromParent();
2755 return true;