1 //===- JumpThreading.cpp - Thread control through conditional blocks ------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the Jump Threading pass.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Scalar/JumpThreading.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/DenseSet.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
21 #include "llvm/Analysis/CFG.h"
22 #include "llvm/Analysis/ConstantFolding.h"
23 #include "llvm/Analysis/GlobalsModRef.h"
24 #include "llvm/Analysis/InstructionSimplify.h"
25 #include "llvm/Analysis/Loads.h"
26 #include "llvm/Analysis/LoopInfo.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/IR/ConstantRange.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/IntrinsicInst.h"
31 #include "llvm/IR/LLVMContext.h"
32 #include "llvm/IR/MDBuilder.h"
33 #include "llvm/IR/Metadata.h"
34 #include "llvm/IR/PatternMatch.h"
35 #include "llvm/Pass.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/raw_ostream.h"
39 #include "llvm/Transforms/Scalar.h"
40 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
41 #include "llvm/Transforms/Utils/Cloning.h"
42 #include "llvm/Transforms/Utils/Local.h"
43 #include "llvm/Transforms/Utils/SSAUpdater.h"
47 using namespace jumpthreading;
49 #define DEBUG_TYPE "jump-threading"
51 STATISTIC(NumThreads, "Number of jumps threaded");
52 STATISTIC(NumFolds, "Number of terminators folded");
53 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
55 static cl::opt<unsigned>
56 BBDuplicateThreshold("jump-threading-threshold",
57 cl::desc("Max block size to duplicate for jump threading"),
58 cl::init(6), cl::Hidden);
60 static cl::opt<unsigned>
61 ImplicationSearchThreshold(
62 "jump-threading-implication-search-threshold",
63 cl::desc("The number of predecessors to search for a stronger "
64 "condition to use to thread over a weaker condition"),
65 cl::init(3), cl::Hidden);
68 /// This pass performs 'jump threading', which looks at blocks that have
69 /// multiple predecessors and multiple successors. If one or more of the
70 /// predecessors of the block can be proven to always jump to one of the
71 /// successors, we forward the edge from the predecessor to the successor by
72 /// duplicating the contents of this block.
74 /// An example of when this can occur is code like this:
81 /// In this case, the unconditional branch at the end of the first if can be
82 /// revectored to the false side of the second if.
84 class JumpThreading : public FunctionPass {
85 JumpThreadingPass Impl;
88 static char ID; // Pass identification
89 JumpThreading(int T = -1) : FunctionPass(ID), Impl(T) {
90 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
93 bool runOnFunction(Function &F) override;
95 void getAnalysisUsage(AnalysisUsage &AU) const override {
96 AU.addRequired<AAResultsWrapperPass>();
97 AU.addRequired<LazyValueInfoWrapperPass>();
98 AU.addPreserved<LazyValueInfoWrapperPass>();
99 AU.addPreserved<GlobalsAAWrapperPass>();
100 AU.addRequired<TargetLibraryInfoWrapperPass>();
103 void releaseMemory() override { Impl.releaseMemory(); }
107 char JumpThreading::ID = 0;
108 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
109 "Jump Threading", false, false)
110 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)
111 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
112 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
113 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
114 "Jump Threading", false, false)
116 // Public interface to the Jump Threading pass
117 FunctionPass *llvm::createJumpThreadingPass(int Threshold) { return new JumpThreading(Threshold); }
119 JumpThreadingPass::JumpThreadingPass(int T) {
120 BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
123 /// runOnFunction - Top level algorithm.
125 bool JumpThreading::runOnFunction(Function &F) {
128 auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
129 auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
130 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
131 std::unique_ptr<BlockFrequencyInfo> BFI;
132 std::unique_ptr<BranchProbabilityInfo> BPI;
133 bool HasProfileData = F.getEntryCount().hasValue();
134 if (HasProfileData) {
135 LoopInfo LI{DominatorTree(F)};
136 BPI.reset(new BranchProbabilityInfo(F, LI, TLI));
137 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
140 return Impl.runImpl(F, TLI, LVI, AA, HasProfileData, std::move(BFI),
144 PreservedAnalyses JumpThreadingPass::run(Function &F,
145 FunctionAnalysisManager &AM) {
147 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
148 auto &LVI = AM.getResult<LazyValueAnalysis>(F);
149 auto &AA = AM.getResult<AAManager>(F);
151 std::unique_ptr<BlockFrequencyInfo> BFI;
152 std::unique_ptr<BranchProbabilityInfo> BPI;
153 bool HasProfileData = F.getEntryCount().hasValue();
154 if (HasProfileData) {
155 LoopInfo LI{DominatorTree(F)};
156 BPI.reset(new BranchProbabilityInfo(F, LI, &TLI));
157 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
160 bool Changed = runImpl(F, &TLI, &LVI, &AA, HasProfileData, std::move(BFI),
164 return PreservedAnalyses::all();
165 PreservedAnalyses PA;
166 PA.preserve<GlobalsAA>();
170 bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
171 LazyValueInfo *LVI_, AliasAnalysis *AA_,
172 bool HasProfileData_,
173 std::unique_ptr<BlockFrequencyInfo> BFI_,
174 std::unique_ptr<BranchProbabilityInfo> BPI_) {
176 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
182 // When profile data is available, we need to update edge weights after
183 // successful jump threading, which requires both BPI and BFI being available.
184 HasProfileData = HasProfileData_;
185 auto *GuardDecl = F.getParent()->getFunction(
186 Intrinsic::getName(Intrinsic::experimental_guard));
187 HasGuards = GuardDecl && !GuardDecl->use_empty();
188 if (HasProfileData) {
189 BPI = std::move(BPI_);
190 BFI = std::move(BFI_);
193 // Remove unreachable blocks from function as they may result in infinite
194 // loop. We do threading if we found something profitable. Jump threading a
195 // branch can create other opportunities. If these opportunities form a cycle
196 // i.e. if any jump threading is undoing previous threading in the path, then
197 // we will loop forever. We take care of this issue by not jump threading for
198 // back edges. This works for normal cases but not for unreachable blocks as
199 // they may have cycle with no back edge.
200 bool EverChanged = false;
201 EverChanged |= removeUnreachableBlocks(F, LVI);
208 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
209 BasicBlock *BB = &*I;
210 // Thread all of the branches we can over this block.
211 while (ProcessBlock(BB))
216 // If the block is trivially dead, zap it. This eliminates the successor
217 // edges which simplifies the CFG.
218 if (pred_empty(BB) &&
219 BB != &BB->getParent()->getEntryBlock()) {
220 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
221 << "' with terminator: " << *BB->getTerminator() << '\n');
222 LoopHeaders.erase(BB);
229 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
231 // Can't thread an unconditional jump, but if the block is "almost
232 // empty", we can replace uses of it with uses of the successor and make
234 // We should not eliminate the loop header either, because eliminating
235 // a loop header might later prevent LoopSimplify from transforming nested
236 // loops into simplified form.
237 if (BI && BI->isUnconditional() &&
238 BB != &BB->getParent()->getEntryBlock() &&
239 // If the terminator is the only non-phi instruction, try to nuke it.
240 BB->getFirstNonPHIOrDbg()->isTerminator() && !LoopHeaders.count(BB)) {
241 // FIXME: It is always conservatively correct to drop the info
242 // for a block even if it doesn't get erased. This isn't totally
243 // awesome, but it allows us to use AssertingVH to prevent nasty
244 // dangling pointer issues within LazyValueInfo.
246 if (TryToSimplifyUncondBranchFromEmptyBlock(BB))
250 EverChanged |= Changed;
257 // Replace uses of Cond with ToVal when safe to do so. If all uses are
258 // replaced, we can remove Cond. We cannot blindly replace all uses of Cond
259 // because we may incorrectly replace uses when guards/assumes are uses of
260 // of `Cond` and we used the guards/assume to reason about the `Cond` value
261 // at the end of block. RAUW unconditionally replaces all uses
262 // including the guards/assumes themselves and the uses before the
264 static void ReplaceFoldableUses(Instruction *Cond, Value *ToVal) {
265 assert(Cond->getType() == ToVal->getType());
266 auto *BB = Cond->getParent();
267 // We can unconditionally replace all uses in non-local blocks (i.e. uses
268 // strictly dominated by BB), since LVI information is true from the
270 replaceNonLocalUsesWith(Cond, ToVal);
271 for (Instruction &I : reverse(*BB)) {
272 // Reached the Cond whose uses we are trying to replace, so there are no
276 // We only replace uses in instructions that are guaranteed to reach the end
277 // of BB, where we know Cond is ToVal.
278 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
280 I.replaceUsesOfWith(Cond, ToVal);
282 if (Cond->use_empty() && !Cond->mayHaveSideEffects())
283 Cond->eraseFromParent();
286 /// Return the cost of duplicating a piece of this block from first non-phi
287 /// and before StopAt instruction to thread across it. Stop scanning the block
288 /// when exceeding the threshold. If duplication is impossible, returns ~0U.
289 static unsigned getJumpThreadDuplicationCost(BasicBlock *BB,
291 unsigned Threshold) {
292 assert(StopAt->getParent() == BB && "Not an instruction from proper BB?");
293 /// Ignore PHI nodes, these will be flattened when duplication happens.
294 BasicBlock::const_iterator I(BB->getFirstNonPHI());
296 // FIXME: THREADING will delete values that are just used to compute the
297 // branch, so they shouldn't count against the duplication cost.
300 if (BB->getTerminator() == StopAt) {
301 // Threading through a switch statement is particularly profitable. If this
302 // block ends in a switch, decrease its cost to make it more likely to
304 if (isa<SwitchInst>(StopAt))
307 // The same holds for indirect branches, but slightly more so.
308 if (isa<IndirectBrInst>(StopAt))
312 // Bump the threshold up so the early exit from the loop doesn't skip the
313 // terminator-based Size adjustment at the end.
316 // Sum up the cost of each instruction until we get to the terminator. Don't
317 // include the terminator because the copy won't include it.
319 for (; &*I != StopAt; ++I) {
321 // Stop scanning the block if we've reached the threshold.
322 if (Size > Threshold)
325 // Debugger intrinsics don't incur code size.
326 if (isa<DbgInfoIntrinsic>(I)) continue;
328 // If this is a pointer->pointer bitcast, it is free.
329 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
332 // Bail out if this instruction gives back a token type, it is not possible
333 // to duplicate it if it is used outside this BB.
334 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
337 // All other instructions count for at least one unit.
340 // Calls are more expensive. If they are non-intrinsic calls, we model them
341 // as having cost of 4. If they are a non-vector intrinsic, we model them
342 // as having cost of 2 total, and if they are a vector intrinsic, we model
343 // them as having cost 1.
344 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
345 if (CI->cannotDuplicate() || CI->isConvergent())
346 // Blocks with NoDuplicate are modelled as having infinite cost, so they
347 // are never duplicated.
349 else if (!isa<IntrinsicInst>(CI))
351 else if (!CI->getType()->isVectorTy())
356 return Size > Bonus ? Size - Bonus : 0;
359 /// FindLoopHeaders - We do not want jump threading to turn proper loop
360 /// structures into irreducible loops. Doing this breaks up the loop nesting
361 /// hierarchy and pessimizes later transformations. To prevent this from
362 /// happening, we first have to find the loop headers. Here we approximate this
363 /// by finding targets of backedges in the CFG.
365 /// Note that there definitely are cases when we want to allow threading of
366 /// edges across a loop header. For example, threading a jump from outside the
367 /// loop (the preheader) to an exit block of the loop is definitely profitable.
368 /// It is also almost always profitable to thread backedges from within the loop
369 /// to exit blocks, and is often profitable to thread backedges to other blocks
370 /// within the loop (forming a nested loop). This simple analysis is not rich
371 /// enough to track all of these properties and keep it up-to-date as the CFG
372 /// mutates, so we don't allow any of these transformations.
374 void JumpThreadingPass::FindLoopHeaders(Function &F) {
375 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
376 FindFunctionBackedges(F, Edges);
378 for (const auto &Edge : Edges)
379 LoopHeaders.insert(Edge.second);
382 /// getKnownConstant - Helper method to determine if we can thread over a
383 /// terminator with the given value as its condition, and if so what value to
384 /// use for that. What kind of value this is depends on whether we want an
385 /// integer or a block address, but an undef is always accepted.
386 /// Returns null if Val is null or not an appropriate constant.
387 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
391 // Undef is "known" enough.
392 if (UndefValue *U = dyn_cast<UndefValue>(Val))
395 if (Preference == WantBlockAddress)
396 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
398 return dyn_cast<ConstantInt>(Val);
401 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
402 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
403 /// in any of our predecessors. If so, return the known list of value and pred
404 /// BB in the result vector.
406 /// This returns true if there were any known values.
408 bool JumpThreadingPass::ComputeValueKnownInPredecessors(
409 Value *V, BasicBlock *BB, PredValueInfo &Result,
410 ConstantPreference Preference, Instruction *CxtI) {
411 // This method walks up use-def chains recursively. Because of this, we could
412 // get into an infinite loop going around loops in the use-def chain. To
413 // prevent this, keep track of what (value, block) pairs we've already visited
414 // and terminate the search if we loop back to them
415 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
418 // An RAII help to remove this pair from the recursion set once the recursion
419 // stack pops back out again.
420 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
422 // If V is a constant, then it is known in all predecessors.
423 if (Constant *KC = getKnownConstant(V, Preference)) {
424 for (BasicBlock *Pred : predecessors(BB))
425 Result.push_back(std::make_pair(KC, Pred));
427 return !Result.empty();
430 // If V is a non-instruction value, or an instruction in a different block,
431 // then it can't be derived from a PHI.
432 Instruction *I = dyn_cast<Instruction>(V);
433 if (!I || I->getParent() != BB) {
435 // Okay, if this is a live-in value, see if it has a known value at the end
436 // of any of our predecessors.
438 // FIXME: This should be an edge property, not a block end property.
439 /// TODO: Per PR2563, we could infer value range information about a
440 /// predecessor based on its terminator.
442 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
443 // "I" is a non-local compare-with-a-constant instruction. This would be
444 // able to handle value inequalities better, for example if the compare is
445 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
446 // Perhaps getConstantOnEdge should be smart enough to do this?
448 for (BasicBlock *P : predecessors(BB)) {
449 // If the value is known by LazyValueInfo to be a constant in a
450 // predecessor, use that information to try to thread this block.
451 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
452 if (Constant *KC = getKnownConstant(PredCst, Preference))
453 Result.push_back(std::make_pair(KC, P));
456 return !Result.empty();
459 /// If I is a PHI node, then we know the incoming values for any constants.
460 if (PHINode *PN = dyn_cast<PHINode>(I)) {
461 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
462 Value *InVal = PN->getIncomingValue(i);
463 if (Constant *KC = getKnownConstant(InVal, Preference)) {
464 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
466 Constant *CI = LVI->getConstantOnEdge(InVal,
467 PN->getIncomingBlock(i),
469 if (Constant *KC = getKnownConstant(CI, Preference))
470 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
474 return !Result.empty();
477 // Handle Cast instructions. Only see through Cast when the source operand is
478 // PHI or Cmp and the source type is i1 to save the compilation time.
479 if (CastInst *CI = dyn_cast<CastInst>(I)) {
480 Value *Source = CI->getOperand(0);
481 if (!Source->getType()->isIntegerTy(1))
483 if (!isa<PHINode>(Source) && !isa<CmpInst>(Source))
485 ComputeValueKnownInPredecessors(Source, BB, Result, Preference, CxtI);
489 // Convert the known values.
490 for (auto &R : Result)
491 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
496 PredValueInfoTy LHSVals, RHSVals;
498 // Handle some boolean conditions.
499 if (I->getType()->getPrimitiveSizeInBits() == 1) {
500 assert(Preference == WantInteger && "One-bit non-integer type?");
502 // X & false -> false
503 if (I->getOpcode() == Instruction::Or ||
504 I->getOpcode() == Instruction::And) {
505 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
507 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
510 if (LHSVals.empty() && RHSVals.empty())
513 ConstantInt *InterestingVal;
514 if (I->getOpcode() == Instruction::Or)
515 InterestingVal = ConstantInt::getTrue(I->getContext());
517 InterestingVal = ConstantInt::getFalse(I->getContext());
519 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
521 // Scan for the sentinel. If we find an undef, force it to the
522 // interesting value: x|undef -> true and x&undef -> false.
523 for (const auto &LHSVal : LHSVals)
524 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
525 Result.emplace_back(InterestingVal, LHSVal.second);
526 LHSKnownBBs.insert(LHSVal.second);
528 for (const auto &RHSVal : RHSVals)
529 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
530 // If we already inferred a value for this block on the LHS, don't
532 if (!LHSKnownBBs.count(RHSVal.second))
533 Result.emplace_back(InterestingVal, RHSVal.second);
536 return !Result.empty();
539 // Handle the NOT form of XOR.
540 if (I->getOpcode() == Instruction::Xor &&
541 isa<ConstantInt>(I->getOperand(1)) &&
542 cast<ConstantInt>(I->getOperand(1))->isOne()) {
543 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
548 // Invert the known values.
549 for (auto &R : Result)
550 R.first = ConstantExpr::getNot(R.first);
555 // Try to simplify some other binary operator values.
556 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
557 assert(Preference != WantBlockAddress
558 && "A binary operator creating a block address?");
559 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
560 PredValueInfoTy LHSVals;
561 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
564 // Try to use constant folding to simplify the binary operator.
565 for (const auto &LHSVal : LHSVals) {
566 Constant *V = LHSVal.first;
567 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
569 if (Constant *KC = getKnownConstant(Folded, WantInteger))
570 Result.push_back(std::make_pair(KC, LHSVal.second));
574 return !Result.empty();
577 // Handle compare with phi operand, where the PHI is defined in this block.
578 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
579 assert(Preference == WantInteger && "Compares only produce integers");
580 Type *CmpType = Cmp->getType();
581 Value *CmpLHS = Cmp->getOperand(0);
582 Value *CmpRHS = Cmp->getOperand(1);
583 CmpInst::Predicate Pred = Cmp->getPredicate();
585 PHINode *PN = dyn_cast<PHINode>(CmpLHS);
586 if (PN && PN->getParent() == BB) {
587 const DataLayout &DL = PN->getModule()->getDataLayout();
588 // We can do this simplification if any comparisons fold to true or false.
590 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
591 BasicBlock *PredBB = PN->getIncomingBlock(i);
592 Value *LHS = PN->getIncomingValue(i);
593 Value *RHS = CmpRHS->DoPHITranslation(BB, PredBB);
595 Value *Res = SimplifyCmpInst(Pred, LHS, RHS, {DL});
597 if (!isa<Constant>(RHS))
600 LazyValueInfo::Tristate
601 ResT = LVI->getPredicateOnEdge(Pred, LHS,
602 cast<Constant>(RHS), PredBB, BB,
604 if (ResT == LazyValueInfo::Unknown)
606 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
609 if (Constant *KC = getKnownConstant(Res, WantInteger))
610 Result.push_back(std::make_pair(KC, PredBB));
613 return !Result.empty();
616 // If comparing a live-in value against a constant, see if we know the
617 // live-in value on any predecessors.
618 if (isa<Constant>(CmpRHS) && !CmpType->isVectorTy()) {
619 Constant *CmpConst = cast<Constant>(CmpRHS);
621 if (!isa<Instruction>(CmpLHS) ||
622 cast<Instruction>(CmpLHS)->getParent() != BB) {
623 for (BasicBlock *P : predecessors(BB)) {
624 // If the value is known by LazyValueInfo to be a constant in a
625 // predecessor, use that information to try to thread this block.
626 LazyValueInfo::Tristate Res =
627 LVI->getPredicateOnEdge(Pred, CmpLHS,
628 CmpConst, P, BB, CxtI ? CxtI : Cmp);
629 if (Res == LazyValueInfo::Unknown)
632 Constant *ResC = ConstantInt::get(CmpType, Res);
633 Result.push_back(std::make_pair(ResC, P));
636 return !Result.empty();
639 // InstCombine can fold some forms of constant range checks into
640 // (icmp (add (x, C1)), C2). See if we have we have such a thing with
643 using namespace PatternMatch;
645 ConstantInt *AddConst;
646 if (isa<ConstantInt>(CmpConst) &&
647 match(CmpLHS, m_Add(m_Value(AddLHS), m_ConstantInt(AddConst)))) {
648 if (!isa<Instruction>(AddLHS) ||
649 cast<Instruction>(AddLHS)->getParent() != BB) {
650 for (BasicBlock *P : predecessors(BB)) {
651 // If the value is known by LazyValueInfo to be a ConstantRange in
652 // a predecessor, use that information to try to thread this
654 ConstantRange CR = LVI->getConstantRangeOnEdge(
655 AddLHS, P, BB, CxtI ? CxtI : cast<Instruction>(CmpLHS));
656 // Propagate the range through the addition.
657 CR = CR.add(AddConst->getValue());
659 // Get the range where the compare returns true.
660 ConstantRange CmpRange = ConstantRange::makeExactICmpRegion(
661 Pred, cast<ConstantInt>(CmpConst)->getValue());
664 if (CmpRange.contains(CR))
665 ResC = ConstantInt::getTrue(CmpType);
666 else if (CmpRange.inverse().contains(CR))
667 ResC = ConstantInt::getFalse(CmpType);
671 Result.push_back(std::make_pair(ResC, P));
674 return !Result.empty();
679 // Try to find a constant value for the LHS of a comparison,
680 // and evaluate it statically if we can.
681 PredValueInfoTy LHSVals;
682 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
685 for (const auto &LHSVal : LHSVals) {
686 Constant *V = LHSVal.first;
687 Constant *Folded = ConstantExpr::getCompare(Pred, V, CmpConst);
688 if (Constant *KC = getKnownConstant(Folded, WantInteger))
689 Result.push_back(std::make_pair(KC, LHSVal.second));
692 return !Result.empty();
696 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
697 // Handle select instructions where at least one operand is a known constant
698 // and we can figure out the condition value for any predecessor block.
699 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
700 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
701 PredValueInfoTy Conds;
702 if ((TrueVal || FalseVal) &&
703 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
704 WantInteger, CxtI)) {
705 for (auto &C : Conds) {
706 Constant *Cond = C.first;
708 // Figure out what value to use for the condition.
710 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
712 KnownCond = CI->isOne();
714 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
715 // Either operand will do, so be sure to pick the one that's a known
717 // FIXME: Do this more cleverly if both values are known constants?
718 KnownCond = (TrueVal != nullptr);
721 // See if the select has a known constant value for this predecessor.
722 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
723 Result.push_back(std::make_pair(Val, C.second));
726 return !Result.empty();
730 // If all else fails, see if LVI can figure out a constant value for us.
731 Constant *CI = LVI->getConstant(V, BB, CxtI);
732 if (Constant *KC = getKnownConstant(CI, Preference)) {
733 for (BasicBlock *Pred : predecessors(BB))
734 Result.push_back(std::make_pair(KC, Pred));
737 return !Result.empty();
742 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
743 /// in an undefined jump, decide which block is best to revector to.
745 /// Since we can pick an arbitrary destination, we pick the successor with the
746 /// fewest predecessors. This should reduce the in-degree of the others.
748 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
749 TerminatorInst *BBTerm = BB->getTerminator();
750 unsigned MinSucc = 0;
751 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
752 // Compute the successor with the minimum number of predecessors.
753 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
754 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
755 TestBB = BBTerm->getSuccessor(i);
756 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
757 if (NumPreds < MinNumPreds) {
759 MinNumPreds = NumPreds;
766 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
767 if (!BB->hasAddressTaken()) return false;
769 // If the block has its address taken, it may be a tree of dead constants
770 // hanging off of it. These shouldn't keep the block alive.
771 BlockAddress *BA = BlockAddress::get(BB);
772 BA->removeDeadConstantUsers();
773 return !BA->use_empty();
776 /// ProcessBlock - If there are any predecessors whose control can be threaded
777 /// through to a successor, transform them now.
778 bool JumpThreadingPass::ProcessBlock(BasicBlock *BB) {
779 // If the block is trivially dead, just return and let the caller nuke it.
780 // This simplifies other transformations.
781 if (pred_empty(BB) &&
782 BB != &BB->getParent()->getEntryBlock())
785 // If this block has a single predecessor, and if that pred has a single
786 // successor, merge the blocks. This encourages recursive jump threading
787 // because now the condition in this block can be threaded through
788 // predecessors of our predecessor block.
789 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
790 const TerminatorInst *TI = SinglePred->getTerminator();
791 if (!TI->isExceptional() && TI->getNumSuccessors() == 1 &&
792 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
793 // If SinglePred was a loop header, BB becomes one.
794 if (LoopHeaders.erase(SinglePred))
795 LoopHeaders.insert(BB);
797 LVI->eraseBlock(SinglePred);
798 MergeBasicBlockIntoOnlyPred(BB);
800 // Now that BB is merged into SinglePred (i.e. SinglePred Code followed by
801 // BB code within one basic block `BB`), we need to invalidate the LVI
802 // information associated with BB, because the LVI information need not be
803 // true for all of BB after the merge. For example,
804 // Before the merge, LVI info and code is as follows:
805 // SinglePred: <LVI info1 for %p val>
807 // call @exit() // need not transfer execution to successor.
808 // assume(%p) // from this point on %p is true
810 // BB: <LVI info2 for %p val, i.e. %p is true>
814 // Note that this LVI info for blocks BB and SinglPred is correct for %p
815 // (info2 and info1 respectively). After the merge and the deletion of the
816 // LVI info1 for SinglePred. We have the following code:
817 // BB: <LVI info2 for %p val>
821 // %x = use of %p <-- LVI info2 is correct from here onwards.
823 // LVI info2 for BB is incorrect at the beginning of BB.
825 // Invalidate LVI information for BB if the LVI is not provably true for
827 if (any_of(*BB, [](Instruction &I) {
828 return !isGuaranteedToTransferExecutionToSuccessor(&I);
835 if (TryToUnfoldSelectInCurrBB(BB))
838 // Look if we can propagate guards to predecessors.
839 if (HasGuards && ProcessGuards(BB))
842 // What kind of constant we're looking for.
843 ConstantPreference Preference = WantInteger;
845 // Look to see if the terminator is a conditional branch, switch or indirect
846 // branch, if not we can't thread it.
848 Instruction *Terminator = BB->getTerminator();
849 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
850 // Can't thread an unconditional jump.
851 if (BI->isUnconditional()) return false;
852 Condition = BI->getCondition();
853 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
854 Condition = SI->getCondition();
855 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
856 // Can't thread indirect branch with no successors.
857 if (IB->getNumSuccessors() == 0) return false;
858 Condition = IB->getAddress()->stripPointerCasts();
859 Preference = WantBlockAddress;
861 return false; // Must be an invoke.
864 // Run constant folding to see if we can reduce the condition to a simple
866 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
868 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
870 I->replaceAllUsesWith(SimpleVal);
871 if (isInstructionTriviallyDead(I, TLI))
872 I->eraseFromParent();
873 Condition = SimpleVal;
877 // If the terminator is branching on an undef, we can pick any of the
878 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
879 if (isa<UndefValue>(Condition)) {
880 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
882 // Fold the branch/switch.
883 TerminatorInst *BBTerm = BB->getTerminator();
884 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
885 if (i == BestSucc) continue;
886 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
889 DEBUG(dbgs() << " In block '" << BB->getName()
890 << "' folding undef terminator: " << *BBTerm << '\n');
891 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
892 BBTerm->eraseFromParent();
896 // If the terminator of this block is branching on a constant, simplify the
897 // terminator to an unconditional branch. This can occur due to threading in
899 if (getKnownConstant(Condition, Preference)) {
900 DEBUG(dbgs() << " In block '" << BB->getName()
901 << "' folding terminator: " << *BB->getTerminator() << '\n');
903 ConstantFoldTerminator(BB, true);
907 Instruction *CondInst = dyn_cast<Instruction>(Condition);
909 // All the rest of our checks depend on the condition being an instruction.
911 // FIXME: Unify this with code below.
912 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
917 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
918 // If we're branching on a conditional, LVI might be able to determine
919 // it's value at the branch instruction. We only handle comparisons
920 // against a constant at this time.
921 // TODO: This should be extended to handle switches as well.
922 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
923 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
924 if (CondBr && CondConst) {
925 // We should have returned as soon as we turn a conditional branch to
926 // unconditional. Because its no longer interesting as far as jump
927 // threading is concerned.
928 assert(CondBr->isConditional() && "Threading on unconditional terminator");
930 LazyValueInfo::Tristate Ret =
931 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
933 if (Ret != LazyValueInfo::Unknown) {
934 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
935 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
936 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
937 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
938 CondBr->eraseFromParent();
939 if (CondCmp->use_empty())
940 CondCmp->eraseFromParent();
941 // We can safely replace *some* uses of the CondInst if it has
942 // exactly one value as returned by LVI. RAUW is incorrect in the
943 // presence of guards and assumes, that have the `Cond` as the use. This
944 // is because we use the guards/assume to reason about the `Cond` value
945 // at the end of block, but RAUW unconditionally replaces all uses
946 // including the guards/assumes themselves and the uses before the
948 else if (CondCmp->getParent() == BB) {
949 auto *CI = Ret == LazyValueInfo::True ?
950 ConstantInt::getTrue(CondCmp->getType()) :
951 ConstantInt::getFalse(CondCmp->getType());
952 ReplaceFoldableUses(CondCmp, CI);
957 // We did not manage to simplify this branch, try to see whether
958 // CondCmp depends on a known phi-select pattern.
959 if (TryToUnfoldSelect(CondCmp, BB))
964 // Check for some cases that are worth simplifying. Right now we want to look
965 // for loads that are used by a switch or by the condition for the branch. If
966 // we see one, check to see if it's partially redundant. If so, insert a PHI
967 // which can then be used to thread the values.
969 Value *SimplifyValue = CondInst;
970 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
971 if (isa<Constant>(CondCmp->getOperand(1)))
972 SimplifyValue = CondCmp->getOperand(0);
974 // TODO: There are other places where load PRE would be profitable, such as
975 // more complex comparisons.
976 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
977 if (SimplifyPartiallyRedundantLoad(LI))
980 // Handle a variety of cases where we are branching on something derived from
981 // a PHI node in the current block. If we can prove that any predecessors
982 // compute a predictable value based on a PHI node, thread those predecessors.
984 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
987 // If this is an otherwise-unfoldable branch on a phi node in the current
988 // block, see if we can simplify.
989 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
990 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
991 return ProcessBranchOnPHI(PN);
993 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
994 if (CondInst->getOpcode() == Instruction::Xor &&
995 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
996 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
998 // Search for a stronger dominating condition that can be used to simplify a
999 // conditional branch leaving BB.
1000 if (ProcessImpliedCondition(BB))
1006 bool JumpThreadingPass::ProcessImpliedCondition(BasicBlock *BB) {
1007 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
1008 if (!BI || !BI->isConditional())
1011 Value *Cond = BI->getCondition();
1012 BasicBlock *CurrentBB = BB;
1013 BasicBlock *CurrentPred = BB->getSinglePredecessor();
1016 auto &DL = BB->getModule()->getDataLayout();
1018 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
1019 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
1020 if (!PBI || !PBI->isConditional())
1022 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
1025 bool FalseDest = PBI->getSuccessor(1) == CurrentBB;
1026 Optional<bool> Implication =
1027 isImpliedCondition(PBI->getCondition(), Cond, DL, FalseDest);
1029 BI->getSuccessor(*Implication ? 1 : 0)->removePredecessor(BB);
1030 BranchInst::Create(BI->getSuccessor(*Implication ? 0 : 1), BI);
1031 BI->eraseFromParent();
1034 CurrentBB = CurrentPred;
1035 CurrentPred = CurrentBB->getSinglePredecessor();
1041 /// Return true if Op is an instruction defined in the given block.
1042 static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
1043 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
1044 if (OpInst->getParent() == BB)
1049 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
1050 /// load instruction, eliminate it by replacing it with a PHI node. This is an
1051 /// important optimization that encourages jump threading, and needs to be run
1052 /// interlaced with other jump threading tasks.
1053 bool JumpThreadingPass::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
1054 // Don't hack volatile and ordered loads.
1055 if (!LI->isUnordered()) return false;
1057 // If the load is defined in a block with exactly one predecessor, it can't be
1058 // partially redundant.
1059 BasicBlock *LoadBB = LI->getParent();
1060 if (LoadBB->getSinglePredecessor())
1063 // If the load is defined in an EH pad, it can't be partially redundant,
1064 // because the edges between the invoke and the EH pad cannot have other
1065 // instructions between them.
1066 if (LoadBB->isEHPad())
1069 Value *LoadedPtr = LI->getOperand(0);
1071 // If the loaded operand is defined in the LoadBB and its not a phi,
1072 // it can't be available in predecessors.
1073 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
1076 // Scan a few instructions up from the load, to see if it is obviously live at
1077 // the entry to its block.
1078 BasicBlock::iterator BBIt(LI);
1080 if (Value *AvailableVal = FindAvailableLoadedValue(
1081 LI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1082 // If the value of the load is locally available within the block, just use
1083 // it. This frequently occurs for reg2mem'd allocas.
1086 LoadInst *NLI = cast<LoadInst>(AvailableVal);
1087 combineMetadataForCSE(NLI, LI);
1090 // If the returned value is the load itself, replace with an undef. This can
1091 // only happen in dead loops.
1092 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
1093 if (AvailableVal->getType() != LI->getType())
1095 CastInst::CreateBitOrPointerCast(AvailableVal, LI->getType(), "", LI);
1096 LI->replaceAllUsesWith(AvailableVal);
1097 LI->eraseFromParent();
1101 // Otherwise, if we scanned the whole block and got to the top of the block,
1102 // we know the block is locally transparent to the load. If not, something
1103 // might clobber its value.
1104 if (BBIt != LoadBB->begin())
1107 // If all of the loads and stores that feed the value have the same AA tags,
1108 // then we can propagate them onto any newly inserted loads.
1110 LI->getAAMetadata(AATags);
1112 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1113 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
1114 AvailablePredsTy AvailablePreds;
1115 BasicBlock *OneUnavailablePred = nullptr;
1116 SmallVector<LoadInst*, 8> CSELoads;
1118 // If we got here, the loaded value is transparent through to the start of the
1119 // block. Check to see if it is available in any of the predecessor blocks.
1120 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1121 // If we already scanned this predecessor, skip it.
1122 if (!PredsScanned.insert(PredBB).second)
1125 BBIt = PredBB->end();
1126 unsigned NumScanedInst = 0;
1127 Value *PredAvailable = nullptr;
1128 // NOTE: We don't CSE load that is volatile or anything stronger than
1129 // unordered, that should have been checked when we entered the function.
1130 assert(LI->isUnordered() && "Attempting to CSE volatile or atomic loads");
1131 // If this is a load on a phi pointer, phi-translate it and search
1132 // for available load/store to the pointer in predecessors.
1133 Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1134 PredAvailable = FindAvailablePtrLoadStore(
1135 Ptr, LI->getType(), LI->isAtomic(), PredBB, BBIt, DefMaxInstsToScan,
1136 AA, &IsLoadCSE, &NumScanedInst);
1138 // If PredBB has a single predecessor, continue scanning through the
1139 // single precessor.
1140 BasicBlock *SinglePredBB = PredBB;
1141 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1142 NumScanedInst < DefMaxInstsToScan) {
1143 SinglePredBB = SinglePredBB->getSinglePredecessor();
1145 BBIt = SinglePredBB->end();
1146 PredAvailable = FindAvailablePtrLoadStore(
1147 Ptr, LI->getType(), LI->isAtomic(), SinglePredBB, BBIt,
1148 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1153 if (!PredAvailable) {
1154 OneUnavailablePred = PredBB;
1159 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1161 // If so, this load is partially redundant. Remember this info so that we
1162 // can create a PHI node.
1163 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1166 // If the loaded value isn't available in any predecessor, it isn't partially
1168 if (AvailablePreds.empty()) return false;
1170 // Okay, the loaded value is available in at least one (and maybe all!)
1171 // predecessors. If the value is unavailable in more than one unique
1172 // predecessor, we want to insert a merge block for those common predecessors.
1173 // This ensures that we only have to insert one reload, thus not increasing
1175 BasicBlock *UnavailablePred = nullptr;
1177 // If there is exactly one predecessor where the value is unavailable, the
1178 // already computed 'OneUnavailablePred' block is it. If it ends in an
1179 // unconditional branch, we know that it isn't a critical edge.
1180 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1181 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1182 UnavailablePred = OneUnavailablePred;
1183 } else if (PredsScanned.size() != AvailablePreds.size()) {
1184 // Otherwise, we had multiple unavailable predecessors or we had a critical
1185 // edge from the one.
1186 SmallVector<BasicBlock*, 8> PredsToSplit;
1187 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1189 for (const auto &AvailablePred : AvailablePreds)
1190 AvailablePredSet.insert(AvailablePred.first);
1192 // Add all the unavailable predecessors to the PredsToSplit list.
1193 for (BasicBlock *P : predecessors(LoadBB)) {
1194 // If the predecessor is an indirect goto, we can't split the edge.
1195 if (isa<IndirectBrInst>(P->getTerminator()))
1198 if (!AvailablePredSet.count(P))
1199 PredsToSplit.push_back(P);
1202 // Split them out to their own block.
1203 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1206 // If the value isn't available in all predecessors, then there will be
1207 // exactly one where it isn't available. Insert a load on that edge and add
1208 // it to the AvailablePreds list.
1209 if (UnavailablePred) {
1210 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1211 "Can't handle critical edge here!");
1212 LoadInst *NewVal = new LoadInst(
1213 LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1214 LI->getName() + ".pr", false, LI->getAlignment(), LI->getOrdering(),
1215 LI->getSyncScopeID(), UnavailablePred->getTerminator());
1216 NewVal->setDebugLoc(LI->getDebugLoc());
1218 NewVal->setAAMetadata(AATags);
1220 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1223 // Now we know that each predecessor of this block has a value in
1224 // AvailablePreds, sort them for efficient access as we're walking the preds.
1225 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1227 // Create a PHI node at the start of the block for the PRE'd load value.
1228 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1229 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
1232 PN->setDebugLoc(LI->getDebugLoc());
1234 // Insert new entries into the PHI for each predecessor. A single block may
1235 // have multiple entries here.
1236 for (pred_iterator PI = PB; PI != PE; ++PI) {
1237 BasicBlock *P = *PI;
1238 AvailablePredsTy::iterator I =
1239 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1240 std::make_pair(P, (Value*)nullptr));
1242 assert(I != AvailablePreds.end() && I->first == P &&
1243 "Didn't find entry for predecessor!");
1245 // If we have an available predecessor but it requires casting, insert the
1246 // cast in the predecessor and use the cast. Note that we have to update the
1247 // AvailablePreds vector as we go so that all of the PHI entries for this
1248 // predecessor use the same bitcast.
1249 Value *&PredV = I->second;
1250 if (PredV->getType() != LI->getType())
1251 PredV = CastInst::CreateBitOrPointerCast(PredV, LI->getType(), "",
1252 P->getTerminator());
1254 PN->addIncoming(PredV, I->first);
1257 for (LoadInst *PredLI : CSELoads) {
1258 combineMetadataForCSE(PredLI, LI);
1261 LI->replaceAllUsesWith(PN);
1262 LI->eraseFromParent();
1267 /// FindMostPopularDest - The specified list contains multiple possible
1268 /// threadable destinations. Pick the one that occurs the most frequently in
1271 FindMostPopularDest(BasicBlock *BB,
1272 const SmallVectorImpl<std::pair<BasicBlock*,
1273 BasicBlock*> > &PredToDestList) {
1274 assert(!PredToDestList.empty());
1276 // Determine popularity. If there are multiple possible destinations, we
1277 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1278 // blocks with known and real destinations to threading undef. We'll handle
1279 // them later if interesting.
1280 DenseMap<BasicBlock*, unsigned> DestPopularity;
1281 for (const auto &PredToDest : PredToDestList)
1282 if (PredToDest.second)
1283 DestPopularity[PredToDest.second]++;
1285 // Find the most popular dest.
1286 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1287 BasicBlock *MostPopularDest = DPI->first;
1288 unsigned Popularity = DPI->second;
1289 SmallVector<BasicBlock*, 4> SamePopularity;
1291 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1292 // If the popularity of this entry isn't higher than the popularity we've
1293 // seen so far, ignore it.
1294 if (DPI->second < Popularity)
1296 else if (DPI->second == Popularity) {
1297 // If it is the same as what we've seen so far, keep track of it.
1298 SamePopularity.push_back(DPI->first);
1300 // If it is more popular, remember it.
1301 SamePopularity.clear();
1302 MostPopularDest = DPI->first;
1303 Popularity = DPI->second;
1307 // Okay, now we know the most popular destination. If there is more than one
1308 // destination, we need to determine one. This is arbitrary, but we need
1309 // to make a deterministic decision. Pick the first one that appears in the
1311 if (!SamePopularity.empty()) {
1312 SamePopularity.push_back(MostPopularDest);
1313 TerminatorInst *TI = BB->getTerminator();
1314 for (unsigned i = 0; ; ++i) {
1315 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1317 if (!is_contained(SamePopularity, TI->getSuccessor(i)))
1320 MostPopularDest = TI->getSuccessor(i);
1325 // Okay, we have finally picked the most popular destination.
1326 return MostPopularDest;
1329 bool JumpThreadingPass::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1330 ConstantPreference Preference,
1331 Instruction *CxtI) {
1332 // If threading this would thread across a loop header, don't even try to
1334 if (LoopHeaders.count(BB))
1337 PredValueInfoTy PredValues;
1338 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1341 assert(!PredValues.empty() &&
1342 "ComputeValueKnownInPredecessors returned true with no values");
1344 DEBUG(dbgs() << "IN BB: " << *BB;
1345 for (const auto &PredValue : PredValues) {
1346 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1348 << " for pred '" << PredValue.second->getName() << "'.\n";
1351 // Decide what we want to thread through. Convert our list of known values to
1352 // a list of known destinations for each pred. This also discards duplicate
1353 // predecessors and keeps track of the undefined inputs (which are represented
1354 // as a null dest in the PredToDestList).
1355 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1356 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1358 BasicBlock *OnlyDest = nullptr;
1359 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1360 Constant *OnlyVal = nullptr;
1361 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1363 unsigned PredWithKnownDest = 0;
1364 for (const auto &PredValue : PredValues) {
1365 BasicBlock *Pred = PredValue.second;
1366 if (!SeenPreds.insert(Pred).second)
1367 continue; // Duplicate predecessor entry.
1369 Constant *Val = PredValue.first;
1372 if (isa<UndefValue>(Val))
1374 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1375 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1376 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1377 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1378 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1379 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1381 assert(isa<IndirectBrInst>(BB->getTerminator())
1382 && "Unexpected terminator");
1383 assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress");
1384 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1387 // If we have exactly one destination, remember it for efficiency below.
1388 if (PredToDestList.empty()) {
1392 if (OnlyDest != DestBB)
1393 OnlyDest = MultipleDestSentinel;
1394 // It possible we have same destination, but different value, e.g. default
1395 // case in switchinst.
1397 OnlyVal = MultipleVal;
1400 // We know where this predecessor is going.
1401 ++PredWithKnownDest;
1403 // If the predecessor ends with an indirect goto, we can't change its
1405 if (isa<IndirectBrInst>(Pred->getTerminator()))
1408 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1411 // If all edges were unthreadable, we fail.
1412 if (PredToDestList.empty())
1415 // If all the predecessors go to a single known successor, we want to fold,
1416 // not thread. By doing so, we do not need to duplicate the current block and
1417 // also miss potential opportunities in case we dont/cant duplicate.
1418 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1419 if (PredWithKnownDest ==
1420 (size_t)std::distance(pred_begin(BB), pred_end(BB))) {
1421 bool SeenFirstBranchToOnlyDest = false;
1422 for (BasicBlock *SuccBB : successors(BB)) {
1423 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest)
1424 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1426 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1429 // Finally update the terminator.
1430 TerminatorInst *Term = BB->getTerminator();
1431 BranchInst::Create(OnlyDest, Term);
1432 Term->eraseFromParent();
1434 // If the condition is now dead due to the removal of the old terminator,
1436 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1437 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1438 CondInst->eraseFromParent();
1439 // We can safely replace *some* uses of the CondInst if it has
1440 // exactly one value as returned by LVI. RAUW is incorrect in the
1441 // presence of guards and assumes, that have the `Cond` as the use. This
1442 // is because we use the guards/assume to reason about the `Cond` value
1443 // at the end of block, but RAUW unconditionally replaces all uses
1444 // including the guards/assumes themselves and the uses before the
1446 else if (OnlyVal && OnlyVal != MultipleVal &&
1447 CondInst->getParent() == BB)
1448 ReplaceFoldableUses(CondInst, OnlyVal);
1454 // Determine which is the most common successor. If we have many inputs and
1455 // this block is a switch, we want to start by threading the batch that goes
1456 // to the most popular destination first. If we only know about one
1457 // threadable destination (the common case) we can avoid this.
1458 BasicBlock *MostPopularDest = OnlyDest;
1460 if (MostPopularDest == MultipleDestSentinel)
1461 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1463 // Now that we know what the most popular destination is, factor all
1464 // predecessors that will jump to it into a single predecessor.
1465 SmallVector<BasicBlock*, 16> PredsToFactor;
1466 for (const auto &PredToDest : PredToDestList)
1467 if (PredToDest.second == MostPopularDest) {
1468 BasicBlock *Pred = PredToDest.first;
1470 // This predecessor may be a switch or something else that has multiple
1471 // edges to the block. Factor each of these edges by listing them
1472 // according to # occurrences in PredsToFactor.
1473 for (BasicBlock *Succ : successors(Pred))
1475 PredsToFactor.push_back(Pred);
1478 // If the threadable edges are branching on an undefined value, we get to pick
1479 // the destination that these predecessors should get to.
1480 if (!MostPopularDest)
1481 MostPopularDest = BB->getTerminator()->
1482 getSuccessor(GetBestDestForJumpOnUndef(BB));
1484 // Ok, try to thread it!
1485 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1488 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1489 /// a PHI node in the current block. See if there are any simplifications we
1490 /// can do based on inputs to the phi node.
1492 bool JumpThreadingPass::ProcessBranchOnPHI(PHINode *PN) {
1493 BasicBlock *BB = PN->getParent();
1495 // TODO: We could make use of this to do it once for blocks with common PHI
1497 SmallVector<BasicBlock*, 1> PredBBs;
1500 // If any of the predecessor blocks end in an unconditional branch, we can
1501 // *duplicate* the conditional branch into that block in order to further
1502 // encourage jump threading and to eliminate cases where we have branch on a
1503 // phi of an icmp (branch on icmp is much better).
1504 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1505 BasicBlock *PredBB = PN->getIncomingBlock(i);
1506 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1507 if (PredBr->isUnconditional()) {
1508 PredBBs[0] = PredBB;
1509 // Try to duplicate BB into PredBB.
1510 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1518 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1519 /// a xor instruction in the current block. See if there are any
1520 /// simplifications we can do based on inputs to the xor.
1522 bool JumpThreadingPass::ProcessBranchOnXOR(BinaryOperator *BO) {
1523 BasicBlock *BB = BO->getParent();
1525 // If either the LHS or RHS of the xor is a constant, don't do this
1527 if (isa<ConstantInt>(BO->getOperand(0)) ||
1528 isa<ConstantInt>(BO->getOperand(1)))
1531 // If the first instruction in BB isn't a phi, we won't be able to infer
1532 // anything special about any particular predecessor.
1533 if (!isa<PHINode>(BB->front()))
1536 // If this BB is a landing pad, we won't be able to split the edge into it.
1540 // If we have a xor as the branch input to this block, and we know that the
1541 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1542 // the condition into the predecessor and fix that value to true, saving some
1543 // logical ops on that path and encouraging other paths to simplify.
1545 // This copies something like this:
1548 // %X = phi i1 [1], [%X']
1549 // %Y = icmp eq i32 %A, %B
1550 // %Z = xor i1 %X, %Y
1555 // %Y = icmp ne i32 %A, %B
1558 PredValueInfoTy XorOpValues;
1560 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1562 assert(XorOpValues.empty());
1563 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1569 assert(!XorOpValues.empty() &&
1570 "ComputeValueKnownInPredecessors returned true with no values");
1572 // Scan the information to see which is most popular: true or false. The
1573 // predecessors can be of the set true, false, or undef.
1574 unsigned NumTrue = 0, NumFalse = 0;
1575 for (const auto &XorOpValue : XorOpValues) {
1576 if (isa<UndefValue>(XorOpValue.first))
1577 // Ignore undefs for the count.
1579 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1585 // Determine which value to split on, true, false, or undef if neither.
1586 ConstantInt *SplitVal = nullptr;
1587 if (NumTrue > NumFalse)
1588 SplitVal = ConstantInt::getTrue(BB->getContext());
1589 else if (NumTrue != 0 || NumFalse != 0)
1590 SplitVal = ConstantInt::getFalse(BB->getContext());
1592 // Collect all of the blocks that this can be folded into so that we can
1593 // factor this once and clone it once.
1594 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1595 for (const auto &XorOpValue : XorOpValues) {
1596 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1599 BlocksToFoldInto.push_back(XorOpValue.second);
1602 // If we inferred a value for all of the predecessors, then duplication won't
1603 // help us. However, we can just replace the LHS or RHS with the constant.
1604 if (BlocksToFoldInto.size() ==
1605 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1607 // If all preds provide undef, just nuke the xor, because it is undef too.
1608 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1609 BO->eraseFromParent();
1610 } else if (SplitVal->isZero()) {
1611 // If all preds provide 0, replace the xor with the other input.
1612 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1613 BO->eraseFromParent();
1615 // If all preds provide 1, set the computed value to 1.
1616 BO->setOperand(!isLHS, SplitVal);
1622 // Try to duplicate BB into PredBB.
1623 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1627 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1628 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1629 /// NewPred using the entries from OldPred (suitably mapped).
1630 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1631 BasicBlock *OldPred,
1632 BasicBlock *NewPred,
1633 DenseMap<Instruction*, Value*> &ValueMap) {
1634 for (BasicBlock::iterator PNI = PHIBB->begin();
1635 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1636 // Ok, we have a PHI node. Figure out what the incoming value was for the
1638 Value *IV = PN->getIncomingValueForBlock(OldPred);
1640 // Remap the value if necessary.
1641 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1642 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1643 if (I != ValueMap.end())
1647 PN->addIncoming(IV, NewPred);
1651 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1652 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1653 /// across BB. Transform the IR to reflect this change.
1654 bool JumpThreadingPass::ThreadEdge(BasicBlock *BB,
1655 const SmallVectorImpl<BasicBlock *> &PredBBs,
1656 BasicBlock *SuccBB) {
1657 // If threading to the same block as we come from, we would infinite loop.
1659 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1660 << "' - would thread to self!\n");
1664 // If threading this would thread across a loop header, don't thread the edge.
1665 // See the comments above FindLoopHeaders for justifications and caveats.
1666 if (LoopHeaders.count(BB)) {
1667 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1668 << "' to dest BB '" << SuccBB->getName()
1669 << "' - it might create an irreducible loop!\n");
1673 unsigned JumpThreadCost =
1674 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1675 if (JumpThreadCost > BBDupThreshold) {
1676 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1677 << "' - Cost is too high: " << JumpThreadCost << "\n");
1681 // And finally, do it! Start by factoring the predecessors if needed.
1683 if (PredBBs.size() == 1)
1684 PredBB = PredBBs[0];
1686 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1687 << " common predecessors.\n");
1688 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1691 // And finally, do it!
1692 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1693 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1694 << ", across block:\n "
1697 LVI->threadEdge(PredBB, BB, SuccBB);
1699 // We are going to have to map operands from the original BB block to the new
1700 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1701 // account for entry from PredBB.
1702 DenseMap<Instruction*, Value*> ValueMapping;
1704 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1705 BB->getName()+".thread",
1706 BB->getParent(), BB);
1707 NewBB->moveAfter(PredBB);
1709 // Set the block frequency of NewBB.
1710 if (HasProfileData) {
1712 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1713 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1716 BasicBlock::iterator BI = BB->begin();
1717 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1718 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1720 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1721 // mapping and using it to remap operands in the cloned instructions.
1722 for (; !isa<TerminatorInst>(BI); ++BI) {
1723 Instruction *New = BI->clone();
1724 New->setName(BI->getName());
1725 NewBB->getInstList().push_back(New);
1726 ValueMapping[&*BI] = New;
1728 // Remap operands to patch up intra-block references.
1729 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1730 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1731 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1732 if (I != ValueMapping.end())
1733 New->setOperand(i, I->second);
1737 // We didn't copy the terminator from BB over to NewBB, because there is now
1738 // an unconditional jump to SuccBB. Insert the unconditional jump.
1739 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
1740 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1742 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1743 // PHI nodes for NewBB now.
1744 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1746 // If there were values defined in BB that are used outside the block, then we
1747 // now have to update all uses of the value to use either the original value,
1748 // the cloned value, or some PHI derived value. This can require arbitrary
1749 // PHI insertion, of which we are prepared to do, clean these up now.
1750 SSAUpdater SSAUpdate;
1751 SmallVector<Use*, 16> UsesToRename;
1752 for (Instruction &I : *BB) {
1753 // Scan all uses of this instruction to see if it is used outside of its
1754 // block, and if so, record them in UsesToRename.
1755 for (Use &U : I.uses()) {
1756 Instruction *User = cast<Instruction>(U.getUser());
1757 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1758 if (UserPN->getIncomingBlock(U) == BB)
1760 } else if (User->getParent() == BB)
1763 UsesToRename.push_back(&U);
1766 // If there are no uses outside the block, we're done with this instruction.
1767 if (UsesToRename.empty())
1770 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
1772 // We found a use of I outside of BB. Rename all uses of I that are outside
1773 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1774 // with the two values we know.
1775 SSAUpdate.Initialize(I.getType(), I.getName());
1776 SSAUpdate.AddAvailableValue(BB, &I);
1777 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
1779 while (!UsesToRename.empty())
1780 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1781 DEBUG(dbgs() << "\n");
1785 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1786 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1787 // us to simplify any PHI nodes in BB.
1788 TerminatorInst *PredTerm = PredBB->getTerminator();
1789 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1790 if (PredTerm->getSuccessor(i) == BB) {
1791 BB->removePredecessor(PredBB, true);
1792 PredTerm->setSuccessor(i, NewBB);
1795 // At this point, the IR is fully up to date and consistent. Do a quick scan
1796 // over the new instructions and zap any that are constants or dead. This
1797 // frequently happens because of phi translation.
1798 SimplifyInstructionsInBlock(NewBB, TLI);
1800 // Update the edge weight from BB to SuccBB, which should be less than before.
1801 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
1803 // Threaded an edge!
1808 /// Create a new basic block that will be the predecessor of BB and successor of
1809 /// all blocks in Preds. When profile data is available, update the frequency of
1811 BasicBlock *JumpThreadingPass::SplitBlockPreds(BasicBlock *BB,
1812 ArrayRef<BasicBlock *> Preds,
1813 const char *Suffix) {
1814 // Collect the frequencies of all predecessors of BB, which will be used to
1815 // update the edge weight on BB->SuccBB.
1816 BlockFrequency PredBBFreq(0);
1818 for (auto Pred : Preds)
1819 PredBBFreq += BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB);
1821 BasicBlock *PredBB = SplitBlockPredecessors(BB, Preds, Suffix);
1823 // Set the block frequency of the newly created PredBB, which is the sum of
1824 // frequencies of Preds.
1826 BFI->setBlockFreq(PredBB, PredBBFreq.getFrequency());
1830 bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
1831 const TerminatorInst *TI = BB->getTerminator();
1832 assert(TI->getNumSuccessors() > 1 && "not a split");
1834 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
1838 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
1839 if (MDName->getString() != "branch_weights")
1842 // Ensure there are weights for all of the successors. Note that the first
1843 // operand to the metadata node is a name, not a weight.
1844 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
1847 /// Update the block frequency of BB and branch weight and the metadata on the
1848 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
1849 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
1850 void JumpThreadingPass::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
1853 BasicBlock *SuccBB) {
1854 if (!HasProfileData)
1857 assert(BFI && BPI && "BFI & BPI should have been created here");
1859 // As the edge from PredBB to BB is deleted, we have to update the block
1861 auto BBOrigFreq = BFI->getBlockFreq(BB);
1862 auto NewBBFreq = BFI->getBlockFreq(NewBB);
1863 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
1864 auto BBNewFreq = BBOrigFreq - NewBBFreq;
1865 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
1867 // Collect updated outgoing edges' frequencies from BB and use them to update
1868 // edge probabilities.
1869 SmallVector<uint64_t, 4> BBSuccFreq;
1870 for (BasicBlock *Succ : successors(BB)) {
1871 auto SuccFreq = (Succ == SuccBB)
1872 ? BB2SuccBBFreq - NewBBFreq
1873 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
1874 BBSuccFreq.push_back(SuccFreq.getFrequency());
1877 uint64_t MaxBBSuccFreq =
1878 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
1880 SmallVector<BranchProbability, 4> BBSuccProbs;
1881 if (MaxBBSuccFreq == 0)
1882 BBSuccProbs.assign(BBSuccFreq.size(),
1883 {1, static_cast<unsigned>(BBSuccFreq.size())});
1885 for (uint64_t Freq : BBSuccFreq)
1886 BBSuccProbs.push_back(
1887 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
1888 // Normalize edge probabilities so that they sum up to one.
1889 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
1893 // Update edge probabilities in BPI.
1894 for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
1895 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
1897 // Update the profile metadata as well.
1899 // Don't do this if the profile of the transformed blocks was statically
1900 // estimated. (This could occur despite the function having an entry
1901 // frequency in completely cold parts of the CFG.)
1903 // In this case we don't want to suggest to subsequent passes that the
1904 // calculated weights are fully consistent. Consider this graph:
1919 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
1920 // the overall probabilities are inconsistent; the total probability that the
1921 // value is either 1, 2 or 3 is 150%.
1923 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
1924 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
1925 // the loop exit edge. Then based solely on static estimation we would assume
1926 // the loop was extremely hot.
1928 // FIXME this locally as well so that BPI and BFI are consistent as well. We
1929 // shouldn't make edges extremely likely or unlikely based solely on static
1931 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
1932 SmallVector<uint32_t, 4> Weights;
1933 for (auto Prob : BBSuccProbs)
1934 Weights.push_back(Prob.getNumerator());
1936 auto TI = BB->getTerminator();
1938 LLVMContext::MD_prof,
1939 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
1943 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1944 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1945 /// If we can duplicate the contents of BB up into PredBB do so now, this
1946 /// improves the odds that the branch will be on an analyzable instruction like
1948 bool JumpThreadingPass::DuplicateCondBranchOnPHIIntoPred(
1949 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
1950 assert(!PredBBs.empty() && "Can't handle an empty set");
1952 // If BB is a loop header, then duplicating this block outside the loop would
1953 // cause us to transform this into an irreducible loop, don't do this.
1954 // See the comments above FindLoopHeaders for justifications and caveats.
1955 if (LoopHeaders.count(BB)) {
1956 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1957 << "' into predecessor block '" << PredBBs[0]->getName()
1958 << "' - it might create an irreducible loop!\n");
1962 unsigned DuplicationCost =
1963 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1964 if (DuplicationCost > BBDupThreshold) {
1965 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1966 << "' - Cost is too high: " << DuplicationCost << "\n");
1970 // And finally, do it! Start by factoring the predecessors if needed.
1972 if (PredBBs.size() == 1)
1973 PredBB = PredBBs[0];
1975 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1976 << " common predecessors.\n");
1977 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1980 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1982 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1983 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1984 << DuplicationCost << " block is:" << *BB << "\n");
1986 // Unless PredBB ends with an unconditional branch, split the edge so that we
1987 // can just clone the bits from BB into the end of the new PredBB.
1988 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1990 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1991 PredBB = SplitEdge(PredBB, BB);
1992 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1995 // We are going to have to map operands from the original BB block into the
1996 // PredBB block. Evaluate PHI nodes in BB.
1997 DenseMap<Instruction*, Value*> ValueMapping;
1999 BasicBlock::iterator BI = BB->begin();
2000 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
2001 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
2002 // Clone the non-phi instructions of BB into PredBB, keeping track of the
2003 // mapping and using it to remap operands in the cloned instructions.
2004 for (; BI != BB->end(); ++BI) {
2005 Instruction *New = BI->clone();
2007 // Remap operands to patch up intra-block references.
2008 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
2009 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
2010 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
2011 if (I != ValueMapping.end())
2012 New->setOperand(i, I->second);
2015 // If this instruction can be simplified after the operands are updated,
2016 // just use the simplified value instead. This frequently happens due to
2018 if (Value *IV = SimplifyInstruction(
2020 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
2021 ValueMapping[&*BI] = IV;
2022 if (!New->mayHaveSideEffects()) {
2027 ValueMapping[&*BI] = New;
2030 // Otherwise, insert the new instruction into the block.
2031 New->setName(BI->getName());
2032 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
2036 // Check to see if the targets of the branch had PHI nodes. If so, we need to
2037 // add entries to the PHI nodes for branch from PredBB now.
2038 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
2039 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
2041 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
2044 // If there were values defined in BB that are used outside the block, then we
2045 // now have to update all uses of the value to use either the original value,
2046 // the cloned value, or some PHI derived value. This can require arbitrary
2047 // PHI insertion, of which we are prepared to do, clean these up now.
2048 SSAUpdater SSAUpdate;
2049 SmallVector<Use*, 16> UsesToRename;
2050 for (Instruction &I : *BB) {
2051 // Scan all uses of this instruction to see if it is used outside of its
2052 // block, and if so, record them in UsesToRename.
2053 for (Use &U : I.uses()) {
2054 Instruction *User = cast<Instruction>(U.getUser());
2055 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
2056 if (UserPN->getIncomingBlock(U) == BB)
2058 } else if (User->getParent() == BB)
2061 UsesToRename.push_back(&U);
2064 // If there are no uses outside the block, we're done with this instruction.
2065 if (UsesToRename.empty())
2068 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
2070 // We found a use of I outside of BB. Rename all uses of I that are outside
2071 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
2072 // with the two values we know.
2073 SSAUpdate.Initialize(I.getType(), I.getName());
2074 SSAUpdate.AddAvailableValue(BB, &I);
2075 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]);
2077 while (!UsesToRename.empty())
2078 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2079 DEBUG(dbgs() << "\n");
2082 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2084 BB->removePredecessor(PredBB, true);
2086 // Remove the unconditional branch at the end of the PredBB block.
2087 OldPredBranch->eraseFromParent();
2093 /// TryToUnfoldSelect - Look for blocks of the form
2099 /// %p = phi [%a, %bb1] ...
2103 /// And expand the select into a branch structure if one of its arms allows %c
2104 /// to be folded. This later enables threading from bb1 over bb2.
2105 bool JumpThreadingPass::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2106 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2107 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2108 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2110 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2111 CondLHS->getParent() != BB)
2114 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2115 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2116 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2118 // Look if one of the incoming values is a select in the corresponding
2120 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2123 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2124 if (!PredTerm || !PredTerm->isUnconditional())
2127 // Now check if one of the select values would allow us to constant fold the
2128 // terminator in BB. We don't do the transform if both sides fold, those
2129 // cases will be threaded in any case.
2130 LazyValueInfo::Tristate LHSFolds =
2131 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2132 CondRHS, Pred, BB, CondCmp);
2133 LazyValueInfo::Tristate RHSFolds =
2134 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2135 CondRHS, Pred, BB, CondCmp);
2136 if ((LHSFolds != LazyValueInfo::Unknown ||
2137 RHSFolds != LazyValueInfo::Unknown) &&
2138 LHSFolds != RHSFolds) {
2139 // Expand the select.
2148 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2149 BB->getParent(), BB);
2150 // Move the unconditional branch to NewBB.
2151 PredTerm->removeFromParent();
2152 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2153 // Create a conditional branch and update PHI nodes.
2154 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2155 CondLHS->setIncomingValue(I, SI->getFalseValue());
2156 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
2157 // The select is now dead.
2158 SI->eraseFromParent();
2160 // Update any other PHI nodes in BB.
2161 for (BasicBlock::iterator BI = BB->begin();
2162 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2164 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2171 /// TryToUnfoldSelectInCurrBB - Look for PHI/Select or PHI/CMP/Select in the
2172 /// same BB in the form
2174 /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2175 /// %s = select %p, trueval, falseval
2180 /// %p = phi [0, %bb1], [1, %bb2], [0, %bb3], [1, %bb4], ...
2182 /// %s = select %c, trueval, falseval
2184 /// And expand the select into a branch structure. This later enables
2185 /// jump-threading over bb in this pass.
2187 /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2188 /// select if the associated PHI has at least one constant. If the unfolded
2189 /// select is not jump-threaded, it will be folded again in the later
2191 bool JumpThreadingPass::TryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2192 // If threading this would thread across a loop header, don't thread the edge.
2193 // See the comments above FindLoopHeaders for justifications and caveats.
2194 if (LoopHeaders.count(BB))
2197 for (BasicBlock::iterator BI = BB->begin();
2198 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2199 // Look for a Phi having at least one constant incoming value.
2200 if (llvm::all_of(PN->incoming_values(),
2201 [](Value *V) { return !isa<ConstantInt>(V); }))
2204 auto isUnfoldCandidate = [BB](SelectInst *SI, Value *V) {
2205 // Check if SI is in BB and use V as condition.
2206 if (SI->getParent() != BB)
2208 Value *Cond = SI->getCondition();
2209 return (Cond && Cond == V && Cond->getType()->isIntegerTy(1));
2212 SelectInst *SI = nullptr;
2213 for (Use &U : PN->uses()) {
2214 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(U.getUser())) {
2215 // Look for a ICmp in BB that compares PN with a constant and is the
2216 // condition of a Select.
2217 if (Cmp->getParent() == BB && Cmp->hasOneUse() &&
2218 isa<ConstantInt>(Cmp->getOperand(1 - U.getOperandNo())))
2219 if (SelectInst *SelectI = dyn_cast<SelectInst>(Cmp->user_back()))
2220 if (isUnfoldCandidate(SelectI, Cmp->use_begin()->get())) {
2224 } else if (SelectInst *SelectI = dyn_cast<SelectInst>(U.getUser())) {
2225 // Look for a Select in BB that uses PN as condtion.
2226 if (isUnfoldCandidate(SelectI, U.get())) {
2235 // Expand the select.
2236 TerminatorInst *Term =
2237 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
2238 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2239 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2240 NewPN->addIncoming(SI->getFalseValue(), BB);
2241 SI->replaceAllUsesWith(NewPN);
2242 SI->eraseFromParent();
2248 /// Try to propagate a guard from the current BB into one of its predecessors
2249 /// in case if another branch of execution implies that the condition of this
2250 /// guard is always true. Currently we only process the simplest case that
2255 /// br i1 %cond, label %T1, label %F1
2261 /// %condGuard = ...
2262 /// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2264 /// And cond either implies condGuard or !condGuard. In this case all the
2265 /// instructions before the guard can be duplicated in both branches, and the
2266 /// guard is then threaded to one of them.
2267 bool JumpThreadingPass::ProcessGuards(BasicBlock *BB) {
2268 using namespace PatternMatch;
2269 // We only want to deal with two predecessors.
2270 BasicBlock *Pred1, *Pred2;
2271 auto PI = pred_begin(BB), PE = pred_end(BB);
2283 // Try to thread one of the guards of the block.
2284 // TODO: Look up deeper than to immediate predecessor?
2285 auto *Parent = Pred1->getSinglePredecessor();
2286 if (!Parent || Parent != Pred2->getSinglePredecessor())
2289 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2291 if (match(&I, m_Intrinsic<Intrinsic::experimental_guard>()))
2292 if (ThreadGuard(BB, cast<IntrinsicInst>(&I), BI))
2298 /// Try to propagate the guard from BB which is the lower block of a diamond
2299 /// to one of its branches, in case if diamond's condition implies guard's
2301 bool JumpThreadingPass::ThreadGuard(BasicBlock *BB, IntrinsicInst *Guard,
2303 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?");
2304 assert(BI->isConditional() && "Unconditional branch has 2 successors?");
2305 Value *GuardCond = Guard->getArgOperand(0);
2306 Value *BranchCond = BI->getCondition();
2307 BasicBlock *TrueDest = BI->getSuccessor(0);
2308 BasicBlock *FalseDest = BI->getSuccessor(1);
2310 auto &DL = BB->getModule()->getDataLayout();
2311 bool TrueDestIsSafe = false;
2312 bool FalseDestIsSafe = false;
2314 // True dest is safe if BranchCond => GuardCond.
2315 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
2317 TrueDestIsSafe = true;
2319 // False dest is safe if !BranchCond => GuardCond.
2321 isImpliedCondition(BranchCond, GuardCond, DL, /* InvertAPred */ true);
2323 FalseDestIsSafe = true;
2326 if (!TrueDestIsSafe && !FalseDestIsSafe)
2329 BasicBlock *UnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
2330 BasicBlock *GuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
2332 ValueToValueMapTy UnguardedMapping, GuardedMapping;
2333 Instruction *AfterGuard = Guard->getNextNode();
2334 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
2335 if (Cost > BBDupThreshold)
2337 // Duplicate all instructions before the guard and the guard itself to the
2338 // branch where implication is not proved.
2339 GuardedBlock = DuplicateInstructionsInSplitBetween(
2340 BB, GuardedBlock, AfterGuard, GuardedMapping);
2341 assert(GuardedBlock && "Could not create the guarded block?");
2342 // Duplicate all instructions before the guard in the unguarded branch.
2343 // Since we have successfully duplicated the guarded block and this block
2344 // has fewer instructions, we expect it to succeed.
2345 UnguardedBlock = DuplicateInstructionsInSplitBetween(BB, UnguardedBlock,
2346 Guard, UnguardedMapping);
2347 assert(UnguardedBlock && "Could not create the unguarded block?");
2348 DEBUG(dbgs() << "Moved guard " << *Guard << " to block "
2349 << GuardedBlock->getName() << "\n");
2351 // Some instructions before the guard may still have uses. For them, we need
2352 // to create Phi nodes merging their copies in both guarded and unguarded
2353 // branches. Those instructions that have no uses can be just removed.
2354 SmallVector<Instruction *, 4> ToRemove;
2355 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
2356 if (!isa<PHINode>(&*BI))
2357 ToRemove.push_back(&*BI);
2359 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
2360 assert(InsertionPoint && "Empty block?");
2361 // Substitute with Phis & remove.
2362 for (auto *Inst : reverse(ToRemove)) {
2363 if (!Inst->use_empty()) {
2364 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
2365 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
2366 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
2367 NewPN->insertBefore(InsertionPoint);
2368 Inst->replaceAllUsesWith(NewPN);
2370 Inst->eraseFromParent();