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/Transforms/Scalar.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/DenseSet.h"
18 #include "llvm/ADT/STLExtras.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/Analysis/AliasAnalysis.h"
21 #include "llvm/Analysis/GlobalsModRef.h"
22 #include "llvm/Analysis/CFG.h"
23 #include "llvm/Analysis/BlockFrequencyInfoImpl.h"
24 #include "llvm/Analysis/ConstantFolding.h"
25 #include "llvm/Analysis/InstructionSimplify.h"
26 #include "llvm/Analysis/Loads.h"
27 #include "llvm/Analysis/LoopInfo.h"
28 #include "llvm/Analysis/ValueTracking.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/Utils/BasicBlockUtils.h"
40 #include "llvm/Transforms/Utils/Cloning.h"
41 #include "llvm/Transforms/Utils/Local.h"
42 #include "llvm/Transforms/Utils/SSAUpdater.h"
46 using namespace jumpthreading;
48 #define DEBUG_TYPE "jump-threading"
50 STATISTIC(NumThreads, "Number of jumps threaded");
51 STATISTIC(NumFolds, "Number of terminators folded");
52 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
54 static cl::opt<unsigned>
55 BBDuplicateThreshold("jump-threading-threshold",
56 cl::desc("Max block size to duplicate for jump threading"),
57 cl::init(6), cl::Hidden);
59 static cl::opt<unsigned>
60 ImplicationSearchThreshold(
61 "jump-threading-implication-search-threshold",
62 cl::desc("The number of predecessors to search for a stronger "
63 "condition to use to thread over a weaker condition"),
64 cl::init(3), cl::Hidden);
67 /// This pass performs 'jump threading', which looks at blocks that have
68 /// multiple predecessors and multiple successors. If one or more of the
69 /// predecessors of the block can be proven to always jump to one of the
70 /// successors, we forward the edge from the predecessor to the successor by
71 /// duplicating the contents of this block.
73 /// An example of when this can occur is code like this:
80 /// In this case, the unconditional branch at the end of the first if can be
81 /// revectored to the false side of the second if.
83 class JumpThreading : public FunctionPass {
84 JumpThreadingPass Impl;
87 static char ID; // Pass identification
88 JumpThreading(int T = -1) : FunctionPass(ID), Impl(T) {
89 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
92 bool runOnFunction(Function &F) override;
94 void getAnalysisUsage(AnalysisUsage &AU) const override {
95 AU.addRequired<AAResultsWrapperPass>();
96 AU.addRequired<LazyValueInfoWrapperPass>();
97 AU.addPreserved<LazyValueInfoWrapperPass>();
98 AU.addPreserved<GlobalsAAWrapperPass>();
99 AU.addRequired<TargetLibraryInfoWrapperPass>();
102 void releaseMemory() override { Impl.releaseMemory(); }
106 char JumpThreading::ID = 0;
107 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
108 "Jump Threading", false, false)
109 INITIALIZE_PASS_DEPENDENCY(LazyValueInfoWrapperPass)
110 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
111 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
112 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
113 "Jump Threading", false, false)
115 // Public interface to the Jump Threading pass
116 FunctionPass *llvm::createJumpThreadingPass(int Threshold) { return new JumpThreading(Threshold); }
118 JumpThreadingPass::JumpThreadingPass(int T) {
119 BBDupThreshold = (T == -1) ? BBDuplicateThreshold : unsigned(T);
122 /// runOnFunction - Top level algorithm.
124 bool JumpThreading::runOnFunction(Function &F) {
127 auto TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
128 auto LVI = &getAnalysis<LazyValueInfoWrapperPass>().getLVI();
129 auto AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
130 std::unique_ptr<BlockFrequencyInfo> BFI;
131 std::unique_ptr<BranchProbabilityInfo> BPI;
132 bool HasProfileData = F.getEntryCount().hasValue();
133 if (HasProfileData) {
134 LoopInfo LI{DominatorTree(F)};
135 BPI.reset(new BranchProbabilityInfo(F, LI));
136 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
139 return Impl.runImpl(F, TLI, LVI, AA, HasProfileData, std::move(BFI),
143 PreservedAnalyses JumpThreadingPass::run(Function &F,
144 FunctionAnalysisManager &AM) {
146 auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
147 auto &LVI = AM.getResult<LazyValueAnalysis>(F);
148 auto &AA = AM.getResult<AAManager>(F);
150 std::unique_ptr<BlockFrequencyInfo> BFI;
151 std::unique_ptr<BranchProbabilityInfo> BPI;
152 bool HasProfileData = F.getEntryCount().hasValue();
153 if (HasProfileData) {
154 LoopInfo LI{DominatorTree(F)};
155 BPI.reset(new BranchProbabilityInfo(F, LI));
156 BFI.reset(new BlockFrequencyInfo(F, *BPI, LI));
159 bool Changed = runImpl(F, &TLI, &LVI, &AA, HasProfileData, std::move(BFI),
163 return PreservedAnalyses::all();
164 PreservedAnalyses PA;
165 PA.preserve<GlobalsAA>();
169 bool JumpThreadingPass::runImpl(Function &F, TargetLibraryInfo *TLI_,
170 LazyValueInfo *LVI_, AliasAnalysis *AA_,
171 bool HasProfileData_,
172 std::unique_ptr<BlockFrequencyInfo> BFI_,
173 std::unique_ptr<BranchProbabilityInfo> BPI_) {
175 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
181 // When profile data is available, we need to update edge weights after
182 // successful jump threading, which requires both BPI and BFI being available.
183 HasProfileData = HasProfileData_;
184 auto *GuardDecl = F.getParent()->getFunction(
185 Intrinsic::getName(Intrinsic::experimental_guard));
186 HasGuards = GuardDecl && !GuardDecl->use_empty();
187 if (HasProfileData) {
188 BPI = std::move(BPI_);
189 BFI = std::move(BFI_);
192 // Remove unreachable blocks from function as they may result in infinite
193 // loop. We do threading if we found something profitable. Jump threading a
194 // branch can create other opportunities. If these opportunities form a cycle
195 // i.e. if any jump threading is undoing previous threading in the path, then
196 // we will loop forever. We take care of this issue by not jump threading for
197 // back edges. This works for normal cases but not for unreachable blocks as
198 // they may have cycle with no back edge.
199 bool EverChanged = false;
200 EverChanged |= removeUnreachableBlocks(F, LVI);
207 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
208 BasicBlock *BB = &*I;
209 // Thread all of the branches we can over this block.
210 while (ProcessBlock(BB))
215 // If the block is trivially dead, zap it. This eliminates the successor
216 // edges which simplifies the CFG.
217 if (pred_empty(BB) &&
218 BB != &BB->getParent()->getEntryBlock()) {
219 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
220 << "' with terminator: " << *BB->getTerminator() << '\n');
221 LoopHeaders.erase(BB);
228 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
230 // Can't thread an unconditional jump, but if the block is "almost
231 // empty", we can replace uses of it with uses of the successor and make
233 // We should not eliminate the loop header either, because eliminating
234 // a loop header might later prevent LoopSimplify from transforming nested
235 // loops into simplified form.
236 if (BI && BI->isUnconditional() &&
237 BB != &BB->getParent()->getEntryBlock() &&
238 // If the terminator is the only non-phi instruction, try to nuke it.
239 BB->getFirstNonPHIOrDbg()->isTerminator() && !LoopHeaders.count(BB)) {
240 // FIXME: It is always conservatively correct to drop the info
241 // for a block even if it doesn't get erased. This isn't totally
242 // awesome, but it allows us to use AssertingVH to prevent nasty
243 // dangling pointer issues within LazyValueInfo.
245 if (TryToSimplifyUncondBranchFromEmptyBlock(BB))
249 EverChanged |= Changed;
256 /// Return the cost of duplicating a piece of this block from first non-phi
257 /// and before StopAt instruction to thread across it. Stop scanning the block
258 /// when exceeding the threshold. If duplication is impossible, returns ~0U.
259 static unsigned getJumpThreadDuplicationCost(BasicBlock *BB,
261 unsigned Threshold) {
262 assert(StopAt->getParent() == BB && "Not an instruction from proper BB?");
263 /// Ignore PHI nodes, these will be flattened when duplication happens.
264 BasicBlock::const_iterator I(BB->getFirstNonPHI());
266 // FIXME: THREADING will delete values that are just used to compute the
267 // branch, so they shouldn't count against the duplication cost.
270 if (BB->getTerminator() == StopAt) {
271 // Threading through a switch statement is particularly profitable. If this
272 // block ends in a switch, decrease its cost to make it more likely to
274 if (isa<SwitchInst>(StopAt))
277 // The same holds for indirect branches, but slightly more so.
278 if (isa<IndirectBrInst>(StopAt))
282 // Bump the threshold up so the early exit from the loop doesn't skip the
283 // terminator-based Size adjustment at the end.
286 // Sum up the cost of each instruction until we get to the terminator. Don't
287 // include the terminator because the copy won't include it.
289 for (; &*I != StopAt; ++I) {
291 // Stop scanning the block if we've reached the threshold.
292 if (Size > Threshold)
295 // Debugger intrinsics don't incur code size.
296 if (isa<DbgInfoIntrinsic>(I)) continue;
298 // If this is a pointer->pointer bitcast, it is free.
299 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
302 // Bail out if this instruction gives back a token type, it is not possible
303 // to duplicate it if it is used outside this BB.
304 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
307 // All other instructions count for at least one unit.
310 // Calls are more expensive. If they are non-intrinsic calls, we model them
311 // as having cost of 4. If they are a non-vector intrinsic, we model them
312 // as having cost of 2 total, and if they are a vector intrinsic, we model
313 // them as having cost 1.
314 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
315 if (CI->cannotDuplicate() || CI->isConvergent())
316 // Blocks with NoDuplicate are modelled as having infinite cost, so they
317 // are never duplicated.
319 else if (!isa<IntrinsicInst>(CI))
321 else if (!CI->getType()->isVectorTy())
326 return Size > Bonus ? Size - Bonus : 0;
329 /// FindLoopHeaders - We do not want jump threading to turn proper loop
330 /// structures into irreducible loops. Doing this breaks up the loop nesting
331 /// hierarchy and pessimizes later transformations. To prevent this from
332 /// happening, we first have to find the loop headers. Here we approximate this
333 /// by finding targets of backedges in the CFG.
335 /// Note that there definitely are cases when we want to allow threading of
336 /// edges across a loop header. For example, threading a jump from outside the
337 /// loop (the preheader) to an exit block of the loop is definitely profitable.
338 /// It is also almost always profitable to thread backedges from within the loop
339 /// to exit blocks, and is often profitable to thread backedges to other blocks
340 /// within the loop (forming a nested loop). This simple analysis is not rich
341 /// enough to track all of these properties and keep it up-to-date as the CFG
342 /// mutates, so we don't allow any of these transformations.
344 void JumpThreadingPass::FindLoopHeaders(Function &F) {
345 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
346 FindFunctionBackedges(F, Edges);
348 for (const auto &Edge : Edges)
349 LoopHeaders.insert(Edge.second);
352 /// getKnownConstant - Helper method to determine if we can thread over a
353 /// terminator with the given value as its condition, and if so what value to
354 /// use for that. What kind of value this is depends on whether we want an
355 /// integer or a block address, but an undef is always accepted.
356 /// Returns null if Val is null or not an appropriate constant.
357 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
361 // Undef is "known" enough.
362 if (UndefValue *U = dyn_cast<UndefValue>(Val))
365 if (Preference == WantBlockAddress)
366 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
368 return dyn_cast<ConstantInt>(Val);
371 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
372 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
373 /// in any of our predecessors. If so, return the known list of value and pred
374 /// BB in the result vector.
376 /// This returns true if there were any known values.
378 bool JumpThreadingPass::ComputeValueKnownInPredecessors(
379 Value *V, BasicBlock *BB, PredValueInfo &Result,
380 ConstantPreference Preference, Instruction *CxtI) {
381 // This method walks up use-def chains recursively. Because of this, we could
382 // get into an infinite loop going around loops in the use-def chain. To
383 // prevent this, keep track of what (value, block) pairs we've already visited
384 // and terminate the search if we loop back to them
385 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
388 // An RAII help to remove this pair from the recursion set once the recursion
389 // stack pops back out again.
390 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
392 // If V is a constant, then it is known in all predecessors.
393 if (Constant *KC = getKnownConstant(V, Preference)) {
394 for (BasicBlock *Pred : predecessors(BB))
395 Result.push_back(std::make_pair(KC, Pred));
397 return !Result.empty();
400 // If V is a non-instruction value, or an instruction in a different block,
401 // then it can't be derived from a PHI.
402 Instruction *I = dyn_cast<Instruction>(V);
403 if (!I || I->getParent() != BB) {
405 // Okay, if this is a live-in value, see if it has a known value at the end
406 // of any of our predecessors.
408 // FIXME: This should be an edge property, not a block end property.
409 /// TODO: Per PR2563, we could infer value range information about a
410 /// predecessor based on its terminator.
412 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
413 // "I" is a non-local compare-with-a-constant instruction. This would be
414 // able to handle value inequalities better, for example if the compare is
415 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
416 // Perhaps getConstantOnEdge should be smart enough to do this?
418 for (BasicBlock *P : predecessors(BB)) {
419 // If the value is known by LazyValueInfo to be a constant in a
420 // predecessor, use that information to try to thread this block.
421 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
422 if (Constant *KC = getKnownConstant(PredCst, Preference))
423 Result.push_back(std::make_pair(KC, P));
426 return !Result.empty();
429 /// If I is a PHI node, then we know the incoming values for any constants.
430 if (PHINode *PN = dyn_cast<PHINode>(I)) {
431 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
432 Value *InVal = PN->getIncomingValue(i);
433 if (Constant *KC = getKnownConstant(InVal, Preference)) {
434 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
436 Constant *CI = LVI->getConstantOnEdge(InVal,
437 PN->getIncomingBlock(i),
439 if (Constant *KC = getKnownConstant(CI, Preference))
440 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
444 return !Result.empty();
447 // Handle Cast instructions. Only see through Cast when the source operand is
448 // PHI or Cmp and the source type is i1 to save the compilation time.
449 if (CastInst *CI = dyn_cast<CastInst>(I)) {
450 Value *Source = CI->getOperand(0);
451 if (!Source->getType()->isIntegerTy(1))
453 if (!isa<PHINode>(Source) && !isa<CmpInst>(Source))
455 ComputeValueKnownInPredecessors(Source, BB, Result, Preference, CxtI);
459 // Convert the known values.
460 for (auto &R : Result)
461 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
466 PredValueInfoTy LHSVals, RHSVals;
468 // Handle some boolean conditions.
469 if (I->getType()->getPrimitiveSizeInBits() == 1) {
470 assert(Preference == WantInteger && "One-bit non-integer type?");
472 // X & false -> false
473 if (I->getOpcode() == Instruction::Or ||
474 I->getOpcode() == Instruction::And) {
475 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
477 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
480 if (LHSVals.empty() && RHSVals.empty())
483 ConstantInt *InterestingVal;
484 if (I->getOpcode() == Instruction::Or)
485 InterestingVal = ConstantInt::getTrue(I->getContext());
487 InterestingVal = ConstantInt::getFalse(I->getContext());
489 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
491 // Scan for the sentinel. If we find an undef, force it to the
492 // interesting value: x|undef -> true and x&undef -> false.
493 for (const auto &LHSVal : LHSVals)
494 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
495 Result.emplace_back(InterestingVal, LHSVal.second);
496 LHSKnownBBs.insert(LHSVal.second);
498 for (const auto &RHSVal : RHSVals)
499 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
500 // If we already inferred a value for this block on the LHS, don't
502 if (!LHSKnownBBs.count(RHSVal.second))
503 Result.emplace_back(InterestingVal, RHSVal.second);
506 return !Result.empty();
509 // Handle the NOT form of XOR.
510 if (I->getOpcode() == Instruction::Xor &&
511 isa<ConstantInt>(I->getOperand(1)) &&
512 cast<ConstantInt>(I->getOperand(1))->isOne()) {
513 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
518 // Invert the known values.
519 for (auto &R : Result)
520 R.first = ConstantExpr::getNot(R.first);
525 // Try to simplify some other binary operator values.
526 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
527 assert(Preference != WantBlockAddress
528 && "A binary operator creating a block address?");
529 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
530 PredValueInfoTy LHSVals;
531 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
534 // Try to use constant folding to simplify the binary operator.
535 for (const auto &LHSVal : LHSVals) {
536 Constant *V = LHSVal.first;
537 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
539 if (Constant *KC = getKnownConstant(Folded, WantInteger))
540 Result.push_back(std::make_pair(KC, LHSVal.second));
544 return !Result.empty();
547 // Handle compare with phi operand, where the PHI is defined in this block.
548 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
549 assert(Preference == WantInteger && "Compares only produce integers");
550 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
551 if (PN && PN->getParent() == BB) {
552 const DataLayout &DL = PN->getModule()->getDataLayout();
553 // We can do this simplification if any comparisons fold to true or false.
555 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
556 BasicBlock *PredBB = PN->getIncomingBlock(i);
557 Value *LHS = PN->getIncomingValue(i);
558 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
560 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
562 if (!isa<Constant>(RHS))
565 LazyValueInfo::Tristate
566 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
567 cast<Constant>(RHS), PredBB, BB,
569 if (ResT == LazyValueInfo::Unknown)
571 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
574 if (Constant *KC = getKnownConstant(Res, WantInteger))
575 Result.push_back(std::make_pair(KC, PredBB));
578 return !Result.empty();
581 // If comparing a live-in value against a constant, see if we know the
582 // live-in value on any predecessors.
583 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
584 if (!isa<Instruction>(Cmp->getOperand(0)) ||
585 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
586 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
588 for (BasicBlock *P : predecessors(BB)) {
589 // If the value is known by LazyValueInfo to be a constant in a
590 // predecessor, use that information to try to thread this block.
591 LazyValueInfo::Tristate Res =
592 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
593 RHSCst, P, BB, CxtI ? CxtI : Cmp);
594 if (Res == LazyValueInfo::Unknown)
597 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
598 Result.push_back(std::make_pair(ResC, P));
601 return !Result.empty();
604 // Try to find a constant value for the LHS of a comparison,
605 // and evaluate it statically if we can.
606 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
607 PredValueInfoTy LHSVals;
608 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
611 for (const auto &LHSVal : LHSVals) {
612 Constant *V = LHSVal.first;
613 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
615 if (Constant *KC = getKnownConstant(Folded, WantInteger))
616 Result.push_back(std::make_pair(KC, LHSVal.second));
619 return !Result.empty();
624 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
625 // Handle select instructions where at least one operand is a known constant
626 // and we can figure out the condition value for any predecessor block.
627 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
628 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
629 PredValueInfoTy Conds;
630 if ((TrueVal || FalseVal) &&
631 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
632 WantInteger, CxtI)) {
633 for (auto &C : Conds) {
634 Constant *Cond = C.first;
636 // Figure out what value to use for the condition.
638 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
640 KnownCond = CI->isOne();
642 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
643 // Either operand will do, so be sure to pick the one that's a known
645 // FIXME: Do this more cleverly if both values are known constants?
646 KnownCond = (TrueVal != nullptr);
649 // See if the select has a known constant value for this predecessor.
650 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
651 Result.push_back(std::make_pair(Val, C.second));
654 return !Result.empty();
658 // If all else fails, see if LVI can figure out a constant value for us.
659 Constant *CI = LVI->getConstant(V, BB, CxtI);
660 if (Constant *KC = getKnownConstant(CI, Preference)) {
661 for (BasicBlock *Pred : predecessors(BB))
662 Result.push_back(std::make_pair(KC, Pred));
665 return !Result.empty();
670 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
671 /// in an undefined jump, decide which block is best to revector to.
673 /// Since we can pick an arbitrary destination, we pick the successor with the
674 /// fewest predecessors. This should reduce the in-degree of the others.
676 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
677 TerminatorInst *BBTerm = BB->getTerminator();
678 unsigned MinSucc = 0;
679 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
680 // Compute the successor with the minimum number of predecessors.
681 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
682 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
683 TestBB = BBTerm->getSuccessor(i);
684 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
685 if (NumPreds < MinNumPreds) {
687 MinNumPreds = NumPreds;
694 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
695 if (!BB->hasAddressTaken()) return false;
697 // If the block has its address taken, it may be a tree of dead constants
698 // hanging off of it. These shouldn't keep the block alive.
699 BlockAddress *BA = BlockAddress::get(BB);
700 BA->removeDeadConstantUsers();
701 return !BA->use_empty();
704 /// ProcessBlock - If there are any predecessors whose control can be threaded
705 /// through to a successor, transform them now.
706 bool JumpThreadingPass::ProcessBlock(BasicBlock *BB) {
707 // If the block is trivially dead, just return and let the caller nuke it.
708 // This simplifies other transformations.
709 if (pred_empty(BB) &&
710 BB != &BB->getParent()->getEntryBlock())
713 // If this block has a single predecessor, and if that pred has a single
714 // successor, merge the blocks. This encourages recursive jump threading
715 // because now the condition in this block can be threaded through
716 // predecessors of our predecessor block.
717 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
718 const TerminatorInst *TI = SinglePred->getTerminator();
719 if (!TI->isExceptional() && TI->getNumSuccessors() == 1 &&
720 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
721 // If SinglePred was a loop header, BB becomes one.
722 if (LoopHeaders.erase(SinglePred))
723 LoopHeaders.insert(BB);
725 LVI->eraseBlock(SinglePred);
726 MergeBasicBlockIntoOnlyPred(BB);
732 if (TryToUnfoldSelectInCurrBB(BB))
735 // Look if we can propagate guards to predecessors.
736 if (HasGuards && ProcessGuards(BB))
739 // What kind of constant we're looking for.
740 ConstantPreference Preference = WantInteger;
742 // Look to see if the terminator is a conditional branch, switch or indirect
743 // branch, if not we can't thread it.
745 Instruction *Terminator = BB->getTerminator();
746 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
747 // Can't thread an unconditional jump.
748 if (BI->isUnconditional()) return false;
749 Condition = BI->getCondition();
750 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
751 Condition = SI->getCondition();
752 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
753 // Can't thread indirect branch with no successors.
754 if (IB->getNumSuccessors() == 0) return false;
755 Condition = IB->getAddress()->stripPointerCasts();
756 Preference = WantBlockAddress;
758 return false; // Must be an invoke.
761 // Run constant folding to see if we can reduce the condition to a simple
763 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
765 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
767 I->replaceAllUsesWith(SimpleVal);
768 if (isInstructionTriviallyDead(I, TLI))
769 I->eraseFromParent();
770 Condition = SimpleVal;
774 // If the terminator is branching on an undef, we can pick any of the
775 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
776 if (isa<UndefValue>(Condition)) {
777 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
779 // Fold the branch/switch.
780 TerminatorInst *BBTerm = BB->getTerminator();
781 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
782 if (i == BestSucc) continue;
783 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
786 DEBUG(dbgs() << " In block '" << BB->getName()
787 << "' folding undef terminator: " << *BBTerm << '\n');
788 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
789 BBTerm->eraseFromParent();
793 // If the terminator of this block is branching on a constant, simplify the
794 // terminator to an unconditional branch. This can occur due to threading in
796 if (getKnownConstant(Condition, Preference)) {
797 DEBUG(dbgs() << " In block '" << BB->getName()
798 << "' folding terminator: " << *BB->getTerminator() << '\n');
800 ConstantFoldTerminator(BB, true);
804 Instruction *CondInst = dyn_cast<Instruction>(Condition);
806 // All the rest of our checks depend on the condition being an instruction.
808 // FIXME: Unify this with code below.
809 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
814 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
815 // If we're branching on a conditional, LVI might be able to determine
816 // it's value at the branch instruction. We only handle comparisons
817 // against a constant at this time.
818 // TODO: This should be extended to handle switches as well.
819 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
820 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
821 if (CondBr && CondConst) {
822 // We should have returned as soon as we turn a conditional branch to
823 // unconditional. Because its no longer interesting as far as jump
824 // threading is concerned.
825 assert(CondBr->isConditional() && "Threading on unconditional terminator");
827 LazyValueInfo::Tristate Ret =
828 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
830 if (Ret != LazyValueInfo::Unknown) {
831 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
832 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
833 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
834 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
835 CondBr->eraseFromParent();
836 if (CondCmp->use_empty())
837 CondCmp->eraseFromParent();
838 else if (CondCmp->getParent() == BB) {
839 // If the fact we just learned is true for all uses of the
840 // condition, replace it with a constant value
841 auto *CI = Ret == LazyValueInfo::True ?
842 ConstantInt::getTrue(CondCmp->getType()) :
843 ConstantInt::getFalse(CondCmp->getType());
844 CondCmp->replaceAllUsesWith(CI);
845 CondCmp->eraseFromParent();
850 // We did not manage to simplify this branch, try to see whether
851 // CondCmp depends on a known phi-select pattern.
852 if (TryToUnfoldSelect(CondCmp, BB))
857 // Check for some cases that are worth simplifying. Right now we want to look
858 // for loads that are used by a switch or by the condition for the branch. If
859 // we see one, check to see if it's partially redundant. If so, insert a PHI
860 // which can then be used to thread the values.
862 Value *SimplifyValue = CondInst;
863 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
864 if (isa<Constant>(CondCmp->getOperand(1)))
865 SimplifyValue = CondCmp->getOperand(0);
867 // TODO: There are other places where load PRE would be profitable, such as
868 // more complex comparisons.
869 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
870 if (SimplifyPartiallyRedundantLoad(LI))
873 // Handle a variety of cases where we are branching on something derived from
874 // a PHI node in the current block. If we can prove that any predecessors
875 // compute a predictable value based on a PHI node, thread those predecessors.
877 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
880 // If this is an otherwise-unfoldable branch on a phi node in the current
881 // block, see if we can simplify.
882 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
883 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
884 return ProcessBranchOnPHI(PN);
886 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
887 if (CondInst->getOpcode() == Instruction::Xor &&
888 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
889 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
891 // Search for a stronger dominating condition that can be used to simplify a
892 // conditional branch leaving BB.
893 if (ProcessImpliedCondition(BB))
899 bool JumpThreadingPass::ProcessImpliedCondition(BasicBlock *BB) {
900 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
901 if (!BI || !BI->isConditional())
904 Value *Cond = BI->getCondition();
905 BasicBlock *CurrentBB = BB;
906 BasicBlock *CurrentPred = BB->getSinglePredecessor();
909 auto &DL = BB->getModule()->getDataLayout();
911 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
912 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
913 if (!PBI || !PBI->isConditional())
915 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
918 bool FalseDest = PBI->getSuccessor(1) == CurrentBB;
919 Optional<bool> Implication =
920 isImpliedCondition(PBI->getCondition(), Cond, DL, FalseDest);
922 BI->getSuccessor(*Implication ? 1 : 0)->removePredecessor(BB);
923 BranchInst::Create(BI->getSuccessor(*Implication ? 0 : 1), BI);
924 BI->eraseFromParent();
927 CurrentBB = CurrentPred;
928 CurrentPred = CurrentBB->getSinglePredecessor();
934 /// Return true if Op is an instruction defined in the given block.
935 static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
936 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
937 if (OpInst->getParent() == BB)
942 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
943 /// load instruction, eliminate it by replacing it with a PHI node. This is an
944 /// important optimization that encourages jump threading, and needs to be run
945 /// interlaced with other jump threading tasks.
946 bool JumpThreadingPass::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
947 // Don't hack volatile and ordered loads.
948 if (!LI->isUnordered()) return false;
950 // If the load is defined in a block with exactly one predecessor, it can't be
951 // partially redundant.
952 BasicBlock *LoadBB = LI->getParent();
953 if (LoadBB->getSinglePredecessor())
956 // If the load is defined in an EH pad, it can't be partially redundant,
957 // because the edges between the invoke and the EH pad cannot have other
958 // instructions between them.
959 if (LoadBB->isEHPad())
962 Value *LoadedPtr = LI->getOperand(0);
964 // If the loaded operand is defined in the LoadBB and its not a phi,
965 // it can't be available in predecessors.
966 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
969 // Scan a few instructions up from the load, to see if it is obviously live at
970 // the entry to its block.
971 BasicBlock::iterator BBIt(LI);
973 if (Value *AvailableVal = FindAvailableLoadedValue(
974 LI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
975 // If the value of the load is locally available within the block, just use
976 // it. This frequently occurs for reg2mem'd allocas.
979 LoadInst *NLI = cast<LoadInst>(AvailableVal);
980 combineMetadataForCSE(NLI, LI);
983 // If the returned value is the load itself, replace with an undef. This can
984 // only happen in dead loops.
985 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
986 if (AvailableVal->getType() != LI->getType())
988 CastInst::CreateBitOrPointerCast(AvailableVal, LI->getType(), "", LI);
989 LI->replaceAllUsesWith(AvailableVal);
990 LI->eraseFromParent();
994 // Otherwise, if we scanned the whole block and got to the top of the block,
995 // we know the block is locally transparent to the load. If not, something
996 // might clobber its value.
997 if (BBIt != LoadBB->begin())
1000 // If all of the loads and stores that feed the value have the same AA tags,
1001 // then we can propagate them onto any newly inserted loads.
1003 LI->getAAMetadata(AATags);
1005 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1006 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
1007 AvailablePredsTy AvailablePreds;
1008 BasicBlock *OneUnavailablePred = nullptr;
1009 SmallVector<LoadInst*, 8> CSELoads;
1011 // If we got here, the loaded value is transparent through to the start of the
1012 // block. Check to see if it is available in any of the predecessor blocks.
1013 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1014 // If we already scanned this predecessor, skip it.
1015 if (!PredsScanned.insert(PredBB).second)
1018 BBIt = PredBB->end();
1019 unsigned NumScanedInst = 0;
1020 Value *PredAvailable = nullptr;
1021 // NOTE: We don't CSE load that is volatile or anything stronger than
1022 // unordered, that should have been checked when we entered the function.
1023 assert(LI->isUnordered() && "Attempting to CSE volatile or atomic loads");
1024 // If this is a load on a phi pointer, phi-translate it and search
1025 // for available load/store to the pointer in predecessors.
1026 Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1027 PredAvailable = FindAvailablePtrLoadStore(
1028 Ptr, LI->getType(), LI->isAtomic(), PredBB, BBIt, DefMaxInstsToScan,
1029 AA, &IsLoadCSE, &NumScanedInst);
1031 // If PredBB has a single predecessor, continue scanning through the
1032 // single precessor.
1033 BasicBlock *SinglePredBB = PredBB;
1034 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1035 NumScanedInst < DefMaxInstsToScan) {
1036 SinglePredBB = SinglePredBB->getSinglePredecessor();
1038 BBIt = SinglePredBB->end();
1039 PredAvailable = FindAvailablePtrLoadStore(
1040 Ptr, LI->getType(), LI->isAtomic(), SinglePredBB, BBIt,
1041 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1046 if (!PredAvailable) {
1047 OneUnavailablePred = PredBB;
1052 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1054 // If so, this load is partially redundant. Remember this info so that we
1055 // can create a PHI node.
1056 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1059 // If the loaded value isn't available in any predecessor, it isn't partially
1061 if (AvailablePreds.empty()) return false;
1063 // Okay, the loaded value is available in at least one (and maybe all!)
1064 // predecessors. If the value is unavailable in more than one unique
1065 // predecessor, we want to insert a merge block for those common predecessors.
1066 // This ensures that we only have to insert one reload, thus not increasing
1068 BasicBlock *UnavailablePred = nullptr;
1070 // If there is exactly one predecessor where the value is unavailable, the
1071 // already computed 'OneUnavailablePred' block is it. If it ends in an
1072 // unconditional branch, we know that it isn't a critical edge.
1073 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1074 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1075 UnavailablePred = OneUnavailablePred;
1076 } else if (PredsScanned.size() != AvailablePreds.size()) {
1077 // Otherwise, we had multiple unavailable predecessors or we had a critical
1078 // edge from the one.
1079 SmallVector<BasicBlock*, 8> PredsToSplit;
1080 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1082 for (const auto &AvailablePred : AvailablePreds)
1083 AvailablePredSet.insert(AvailablePred.first);
1085 // Add all the unavailable predecessors to the PredsToSplit list.
1086 for (BasicBlock *P : predecessors(LoadBB)) {
1087 // If the predecessor is an indirect goto, we can't split the edge.
1088 if (isa<IndirectBrInst>(P->getTerminator()))
1091 if (!AvailablePredSet.count(P))
1092 PredsToSplit.push_back(P);
1095 // Split them out to their own block.
1096 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1099 // If the value isn't available in all predecessors, then there will be
1100 // exactly one where it isn't available. Insert a load on that edge and add
1101 // it to the AvailablePreds list.
1102 if (UnavailablePred) {
1103 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1104 "Can't handle critical edge here!");
1105 LoadInst *NewVal = new LoadInst(
1106 LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1107 LI->getName() + ".pr", false, LI->getAlignment(), LI->getOrdering(),
1108 LI->getSynchScope(), UnavailablePred->getTerminator());
1109 NewVal->setDebugLoc(LI->getDebugLoc());
1111 NewVal->setAAMetadata(AATags);
1113 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1116 // Now we know that each predecessor of this block has a value in
1117 // AvailablePreds, sort them for efficient access as we're walking the preds.
1118 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1120 // Create a PHI node at the start of the block for the PRE'd load value.
1121 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1122 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
1125 PN->setDebugLoc(LI->getDebugLoc());
1127 // Insert new entries into the PHI for each predecessor. A single block may
1128 // have multiple entries here.
1129 for (pred_iterator PI = PB; PI != PE; ++PI) {
1130 BasicBlock *P = *PI;
1131 AvailablePredsTy::iterator I =
1132 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1133 std::make_pair(P, (Value*)nullptr));
1135 assert(I != AvailablePreds.end() && I->first == P &&
1136 "Didn't find entry for predecessor!");
1138 // If we have an available predecessor but it requires casting, insert the
1139 // cast in the predecessor and use the cast. Note that we have to update the
1140 // AvailablePreds vector as we go so that all of the PHI entries for this
1141 // predecessor use the same bitcast.
1142 Value *&PredV = I->second;
1143 if (PredV->getType() != LI->getType())
1144 PredV = CastInst::CreateBitOrPointerCast(PredV, LI->getType(), "",
1145 P->getTerminator());
1147 PN->addIncoming(PredV, I->first);
1150 for (LoadInst *PredLI : CSELoads) {
1151 combineMetadataForCSE(PredLI, LI);
1154 LI->replaceAllUsesWith(PN);
1155 LI->eraseFromParent();
1160 /// FindMostPopularDest - The specified list contains multiple possible
1161 /// threadable destinations. Pick the one that occurs the most frequently in
1164 FindMostPopularDest(BasicBlock *BB,
1165 const SmallVectorImpl<std::pair<BasicBlock*,
1166 BasicBlock*> > &PredToDestList) {
1167 assert(!PredToDestList.empty());
1169 // Determine popularity. If there are multiple possible destinations, we
1170 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1171 // blocks with known and real destinations to threading undef. We'll handle
1172 // them later if interesting.
1173 DenseMap<BasicBlock*, unsigned> DestPopularity;
1174 for (const auto &PredToDest : PredToDestList)
1175 if (PredToDest.second)
1176 DestPopularity[PredToDest.second]++;
1178 // Find the most popular dest.
1179 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1180 BasicBlock *MostPopularDest = DPI->first;
1181 unsigned Popularity = DPI->second;
1182 SmallVector<BasicBlock*, 4> SamePopularity;
1184 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1185 // If the popularity of this entry isn't higher than the popularity we've
1186 // seen so far, ignore it.
1187 if (DPI->second < Popularity)
1189 else if (DPI->second == Popularity) {
1190 // If it is the same as what we've seen so far, keep track of it.
1191 SamePopularity.push_back(DPI->first);
1193 // If it is more popular, remember it.
1194 SamePopularity.clear();
1195 MostPopularDest = DPI->first;
1196 Popularity = DPI->second;
1200 // Okay, now we know the most popular destination. If there is more than one
1201 // destination, we need to determine one. This is arbitrary, but we need
1202 // to make a deterministic decision. Pick the first one that appears in the
1204 if (!SamePopularity.empty()) {
1205 SamePopularity.push_back(MostPopularDest);
1206 TerminatorInst *TI = BB->getTerminator();
1207 for (unsigned i = 0; ; ++i) {
1208 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1210 if (!is_contained(SamePopularity, TI->getSuccessor(i)))
1213 MostPopularDest = TI->getSuccessor(i);
1218 // Okay, we have finally picked the most popular destination.
1219 return MostPopularDest;
1222 bool JumpThreadingPass::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1223 ConstantPreference Preference,
1224 Instruction *CxtI) {
1225 // If threading this would thread across a loop header, don't even try to
1227 if (LoopHeaders.count(BB))
1230 PredValueInfoTy PredValues;
1231 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1234 assert(!PredValues.empty() &&
1235 "ComputeValueKnownInPredecessors returned true with no values");
1237 DEBUG(dbgs() << "IN BB: " << *BB;
1238 for (const auto &PredValue : PredValues) {
1239 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1241 << " for pred '" << PredValue.second->getName() << "'.\n";
1244 // Decide what we want to thread through. Convert our list of known values to
1245 // a list of known destinations for each pred. This also discards duplicate
1246 // predecessors and keeps track of the undefined inputs (which are represented
1247 // as a null dest in the PredToDestList).
1248 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1249 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1251 BasicBlock *OnlyDest = nullptr;
1252 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1254 for (const auto &PredValue : PredValues) {
1255 BasicBlock *Pred = PredValue.second;
1256 if (!SeenPreds.insert(Pred).second)
1257 continue; // Duplicate predecessor entry.
1259 // If the predecessor ends with an indirect goto, we can't change its
1261 if (isa<IndirectBrInst>(Pred->getTerminator()))
1264 Constant *Val = PredValue.first;
1267 if (isa<UndefValue>(Val))
1269 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1270 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1271 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1272 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1274 assert(isa<IndirectBrInst>(BB->getTerminator())
1275 && "Unexpected terminator");
1276 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1279 // If we have exactly one destination, remember it for efficiency below.
1280 if (PredToDestList.empty())
1282 else if (OnlyDest != DestBB)
1283 OnlyDest = MultipleDestSentinel;
1285 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1288 // If all edges were unthreadable, we fail.
1289 if (PredToDestList.empty())
1292 // If all the predecessors go to a single known successor, we want to fold,
1293 // not thread. By doing so, we do not need to duplicate the current block and
1294 // also miss potential opportunities in case we dont/cant duplicate.
1295 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1296 if (PredToDestList.size() ==
1297 (size_t)std::distance(pred_begin(BB), pred_end(BB))) {
1298 bool SeenFirstBranchToOnlyDest = false;
1299 for (BasicBlock *SuccBB : successors(BB)) {
1300 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest)
1301 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1303 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1306 // Finally update the terminator.
1307 TerminatorInst *Term = BB->getTerminator();
1308 BranchInst::Create(OnlyDest, Term);
1309 Term->eraseFromParent();
1311 // If the condition is now dead due to the removal of the old terminator,
1313 auto *CondInst = dyn_cast<Instruction>(Cond);
1314 if (CondInst && CondInst->use_empty())
1315 CondInst->eraseFromParent();
1316 // FIXME: in case this instruction is defined in the current BB and it
1317 // resolves to a single value from all predecessors, we can do RAUW.
1322 // Determine which is the most common successor. If we have many inputs and
1323 // this block is a switch, we want to start by threading the batch that goes
1324 // to the most popular destination first. If we only know about one
1325 // threadable destination (the common case) we can avoid this.
1326 BasicBlock *MostPopularDest = OnlyDest;
1328 if (MostPopularDest == MultipleDestSentinel)
1329 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1331 // Now that we know what the most popular destination is, factor all
1332 // predecessors that will jump to it into a single predecessor.
1333 SmallVector<BasicBlock*, 16> PredsToFactor;
1334 for (const auto &PredToDest : PredToDestList)
1335 if (PredToDest.second == MostPopularDest) {
1336 BasicBlock *Pred = PredToDest.first;
1338 // This predecessor may be a switch or something else that has multiple
1339 // edges to the block. Factor each of these edges by listing them
1340 // according to # occurrences in PredsToFactor.
1341 for (BasicBlock *Succ : successors(Pred))
1343 PredsToFactor.push_back(Pred);
1346 // If the threadable edges are branching on an undefined value, we get to pick
1347 // the destination that these predecessors should get to.
1348 if (!MostPopularDest)
1349 MostPopularDest = BB->getTerminator()->
1350 getSuccessor(GetBestDestForJumpOnUndef(BB));
1352 // Ok, try to thread it!
1353 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1356 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1357 /// a PHI node in the current block. See if there are any simplifications we
1358 /// can do based on inputs to the phi node.
1360 bool JumpThreadingPass::ProcessBranchOnPHI(PHINode *PN) {
1361 BasicBlock *BB = PN->getParent();
1363 // TODO: We could make use of this to do it once for blocks with common PHI
1365 SmallVector<BasicBlock*, 1> PredBBs;
1368 // If any of the predecessor blocks end in an unconditional branch, we can
1369 // *duplicate* the conditional branch into that block in order to further
1370 // encourage jump threading and to eliminate cases where we have branch on a
1371 // phi of an icmp (branch on icmp is much better).
1372 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1373 BasicBlock *PredBB = PN->getIncomingBlock(i);
1374 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1375 if (PredBr->isUnconditional()) {
1376 PredBBs[0] = PredBB;
1377 // Try to duplicate BB into PredBB.
1378 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1386 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1387 /// a xor instruction in the current block. See if there are any
1388 /// simplifications we can do based on inputs to the xor.
1390 bool JumpThreadingPass::ProcessBranchOnXOR(BinaryOperator *BO) {
1391 BasicBlock *BB = BO->getParent();
1393 // If either the LHS or RHS of the xor is a constant, don't do this
1395 if (isa<ConstantInt>(BO->getOperand(0)) ||
1396 isa<ConstantInt>(BO->getOperand(1)))
1399 // If the first instruction in BB isn't a phi, we won't be able to infer
1400 // anything special about any particular predecessor.
1401 if (!isa<PHINode>(BB->front()))
1404 // If this BB is a landing pad, we won't be able to split the edge into it.
1408 // If we have a xor as the branch input to this block, and we know that the
1409 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1410 // the condition into the predecessor and fix that value to true, saving some
1411 // logical ops on that path and encouraging other paths to simplify.
1413 // This copies something like this:
1416 // %X = phi i1 [1], [%X']
1417 // %Y = icmp eq i32 %A, %B
1418 // %Z = xor i1 %X, %Y
1423 // %Y = icmp ne i32 %A, %B
1426 PredValueInfoTy XorOpValues;
1428 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1430 assert(XorOpValues.empty());
1431 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1437 assert(!XorOpValues.empty() &&
1438 "ComputeValueKnownInPredecessors returned true with no values");
1440 // Scan the information to see which is most popular: true or false. The
1441 // predecessors can be of the set true, false, or undef.
1442 unsigned NumTrue = 0, NumFalse = 0;
1443 for (const auto &XorOpValue : XorOpValues) {
1444 if (isa<UndefValue>(XorOpValue.first))
1445 // Ignore undefs for the count.
1447 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1453 // Determine which value to split on, true, false, or undef if neither.
1454 ConstantInt *SplitVal = nullptr;
1455 if (NumTrue > NumFalse)
1456 SplitVal = ConstantInt::getTrue(BB->getContext());
1457 else if (NumTrue != 0 || NumFalse != 0)
1458 SplitVal = ConstantInt::getFalse(BB->getContext());
1460 // Collect all of the blocks that this can be folded into so that we can
1461 // factor this once and clone it once.
1462 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1463 for (const auto &XorOpValue : XorOpValues) {
1464 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1467 BlocksToFoldInto.push_back(XorOpValue.second);
1470 // If we inferred a value for all of the predecessors, then duplication won't
1471 // help us. However, we can just replace the LHS or RHS with the constant.
1472 if (BlocksToFoldInto.size() ==
1473 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1475 // If all preds provide undef, just nuke the xor, because it is undef too.
1476 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1477 BO->eraseFromParent();
1478 } else if (SplitVal->isZero()) {
1479 // If all preds provide 0, replace the xor with the other input.
1480 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1481 BO->eraseFromParent();
1483 // If all preds provide 1, set the computed value to 1.
1484 BO->setOperand(!isLHS, SplitVal);
1490 // Try to duplicate BB into PredBB.
1491 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1495 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1496 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1497 /// NewPred using the entries from OldPred (suitably mapped).
1498 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1499 BasicBlock *OldPred,
1500 BasicBlock *NewPred,
1501 DenseMap<Instruction*, Value*> &ValueMap) {
1502 for (BasicBlock::iterator PNI = PHIBB->begin();
1503 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1504 // Ok, we have a PHI node. Figure out what the incoming value was for the
1506 Value *IV = PN->getIncomingValueForBlock(OldPred);
1508 // Remap the value if necessary.
1509 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1510 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1511 if (I != ValueMap.end())
1515 PN->addIncoming(IV, NewPred);
1519 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1520 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1521 /// across BB. Transform the IR to reflect this change.
1522 bool JumpThreadingPass::ThreadEdge(BasicBlock *BB,
1523 const SmallVectorImpl<BasicBlock *> &PredBBs,
1524 BasicBlock *SuccBB) {
1525 // If threading to the same block as we come from, we would infinite loop.
1527 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1528 << "' - would thread to self!\n");
1532 // If threading this would thread across a loop header, don't thread the edge.
1533 // See the comments above FindLoopHeaders for justifications and caveats.
1534 if (LoopHeaders.count(BB)) {
1535 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1536 << "' to dest BB '" << SuccBB->getName()
1537 << "' - it might create an irreducible loop!\n");
1541 unsigned JumpThreadCost =
1542 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1543 if (JumpThreadCost > BBDupThreshold) {
1544 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1545 << "' - Cost is too high: " << JumpThreadCost << "\n");
1549 // And finally, do it! Start by factoring the predecessors if needed.
1551 if (PredBBs.size() == 1)
1552 PredBB = PredBBs[0];
1554 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1555 << " common predecessors.\n");
1556 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1559 // And finally, do it!
1560 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1561 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1562 << ", across block:\n "
1565 LVI->threadEdge(PredBB, BB, SuccBB);
1567 // We are going to have to map operands from the original BB block to the new
1568 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1569 // account for entry from PredBB.
1570 DenseMap<Instruction*, Value*> ValueMapping;
1572 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1573 BB->getName()+".thread",
1574 BB->getParent(), BB);
1575 NewBB->moveAfter(PredBB);
1577 // Set the block frequency of NewBB.
1578 if (HasProfileData) {
1580 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1581 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1584 BasicBlock::iterator BI = BB->begin();
1585 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1586 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1588 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1589 // mapping and using it to remap operands in the cloned instructions.
1590 for (; !isa<TerminatorInst>(BI); ++BI) {
1591 Instruction *New = BI->clone();
1592 New->setName(BI->getName());
1593 NewBB->getInstList().push_back(New);
1594 ValueMapping[&*BI] = New;
1596 // Remap operands to patch up intra-block references.
1597 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1598 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1599 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1600 if (I != ValueMapping.end())
1601 New->setOperand(i, I->second);
1605 // We didn't copy the terminator from BB over to NewBB, because there is now
1606 // an unconditional jump to SuccBB. Insert the unconditional jump.
1607 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
1608 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1610 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1611 // PHI nodes for NewBB now.
1612 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1614 // If there were values defined in BB that are used outside the block, then we
1615 // now have to update all uses of the value to use either the original value,
1616 // the cloned value, or some PHI derived value. This can require arbitrary
1617 // PHI insertion, of which we are prepared to do, clean these up now.
1618 SSAUpdater SSAUpdate;
1619 SmallVector<Use*, 16> UsesToRename;
1620 for (Instruction &I : *BB) {
1621 // Scan all uses of this instruction to see if it is used outside of its
1622 // block, and if so, record them in UsesToRename.
1623 for (Use &U : I.uses()) {
1624 Instruction *User = cast<Instruction>(U.getUser());
1625 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1626 if (UserPN->getIncomingBlock(U) == BB)
1628 } else if (User->getParent() == BB)
1631 UsesToRename.push_back(&U);
1634 // If there are no uses outside the block, we're done with this instruction.
1635 if (UsesToRename.empty())
1638 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
1640 // We found a use of I outside of BB. Rename all uses of I that are outside
1641 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1642 // with the two values we know.
1643 SSAUpdate.Initialize(I.getType(), I.getName());
1644 SSAUpdate.AddAvailableValue(BB, &I);
1645 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
1647 while (!UsesToRename.empty())
1648 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1649 DEBUG(dbgs() << "\n");
1653 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1654 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1655 // us to simplify any PHI nodes in BB.
1656 TerminatorInst *PredTerm = PredBB->getTerminator();
1657 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1658 if (PredTerm->getSuccessor(i) == BB) {
1659 BB->removePredecessor(PredBB, true);
1660 PredTerm->setSuccessor(i, NewBB);
1663 // At this point, the IR is fully up to date and consistent. Do a quick scan
1664 // over the new instructions and zap any that are constants or dead. This
1665 // frequently happens because of phi translation.
1666 SimplifyInstructionsInBlock(NewBB, TLI);
1668 // Update the edge weight from BB to SuccBB, which should be less than before.
1669 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
1671 // Threaded an edge!
1676 /// Create a new basic block that will be the predecessor of BB and successor of
1677 /// all blocks in Preds. When profile data is available, update the frequency of
1679 BasicBlock *JumpThreadingPass::SplitBlockPreds(BasicBlock *BB,
1680 ArrayRef<BasicBlock *> Preds,
1681 const char *Suffix) {
1682 // Collect the frequencies of all predecessors of BB, which will be used to
1683 // update the edge weight on BB->SuccBB.
1684 BlockFrequency PredBBFreq(0);
1686 for (auto Pred : Preds)
1687 PredBBFreq += BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB);
1689 BasicBlock *PredBB = SplitBlockPredecessors(BB, Preds, Suffix);
1691 // Set the block frequency of the newly created PredBB, which is the sum of
1692 // frequencies of Preds.
1694 BFI->setBlockFreq(PredBB, PredBBFreq.getFrequency());
1698 bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
1699 const TerminatorInst *TI = BB->getTerminator();
1700 assert(TI->getNumSuccessors() > 1 && "not a split");
1702 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
1706 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
1707 if (MDName->getString() != "branch_weights")
1710 // Ensure there are weights for all of the successors. Note that the first
1711 // operand to the metadata node is a name, not a weight.
1712 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
1715 /// Update the block frequency of BB and branch weight and the metadata on the
1716 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
1717 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
1718 void JumpThreadingPass::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
1721 BasicBlock *SuccBB) {
1722 if (!HasProfileData)
1725 assert(BFI && BPI && "BFI & BPI should have been created here");
1727 // As the edge from PredBB to BB is deleted, we have to update the block
1729 auto BBOrigFreq = BFI->getBlockFreq(BB);
1730 auto NewBBFreq = BFI->getBlockFreq(NewBB);
1731 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
1732 auto BBNewFreq = BBOrigFreq - NewBBFreq;
1733 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
1735 // Collect updated outgoing edges' frequencies from BB and use them to update
1736 // edge probabilities.
1737 SmallVector<uint64_t, 4> BBSuccFreq;
1738 for (BasicBlock *Succ : successors(BB)) {
1739 auto SuccFreq = (Succ == SuccBB)
1740 ? BB2SuccBBFreq - NewBBFreq
1741 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
1742 BBSuccFreq.push_back(SuccFreq.getFrequency());
1745 uint64_t MaxBBSuccFreq =
1746 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
1748 SmallVector<BranchProbability, 4> BBSuccProbs;
1749 if (MaxBBSuccFreq == 0)
1750 BBSuccProbs.assign(BBSuccFreq.size(),
1751 {1, static_cast<unsigned>(BBSuccFreq.size())});
1753 for (uint64_t Freq : BBSuccFreq)
1754 BBSuccProbs.push_back(
1755 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
1756 // Normalize edge probabilities so that they sum up to one.
1757 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
1761 // Update edge probabilities in BPI.
1762 for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
1763 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
1765 // Update the profile metadata as well.
1767 // Don't do this if the profile of the transformed blocks was statically
1768 // estimated. (This could occur despite the function having an entry
1769 // frequency in completely cold parts of the CFG.)
1771 // In this case we don't want to suggest to subsequent passes that the
1772 // calculated weights are fully consistent. Consider this graph:
1787 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
1788 // the overall probabilities are inconsistent; the total probability that the
1789 // value is either 1, 2 or 3 is 150%.
1791 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
1792 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
1793 // the loop exit edge. Then based solely on static estimation we would assume
1794 // the loop was extremely hot.
1796 // FIXME this locally as well so that BPI and BFI are consistent as well. We
1797 // shouldn't make edges extremely likely or unlikely based solely on static
1799 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
1800 SmallVector<uint32_t, 4> Weights;
1801 for (auto Prob : BBSuccProbs)
1802 Weights.push_back(Prob.getNumerator());
1804 auto TI = BB->getTerminator();
1806 LLVMContext::MD_prof,
1807 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
1811 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1812 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1813 /// If we can duplicate the contents of BB up into PredBB do so now, this
1814 /// improves the odds that the branch will be on an analyzable instruction like
1816 bool JumpThreadingPass::DuplicateCondBranchOnPHIIntoPred(
1817 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
1818 assert(!PredBBs.empty() && "Can't handle an empty set");
1820 // If BB is a loop header, then duplicating this block outside the loop would
1821 // cause us to transform this into an irreducible loop, don't do this.
1822 // See the comments above FindLoopHeaders for justifications and caveats.
1823 if (LoopHeaders.count(BB)) {
1824 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1825 << "' into predecessor block '" << PredBBs[0]->getName()
1826 << "' - it might create an irreducible loop!\n");
1830 unsigned DuplicationCost =
1831 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1832 if (DuplicationCost > BBDupThreshold) {
1833 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1834 << "' - Cost is too high: " << DuplicationCost << "\n");
1838 // And finally, do it! Start by factoring the predecessors if needed.
1840 if (PredBBs.size() == 1)
1841 PredBB = PredBBs[0];
1843 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1844 << " common predecessors.\n");
1845 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1848 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1850 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1851 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1852 << DuplicationCost << " block is:" << *BB << "\n");
1854 // Unless PredBB ends with an unconditional branch, split the edge so that we
1855 // can just clone the bits from BB into the end of the new PredBB.
1856 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1858 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1859 PredBB = SplitEdge(PredBB, BB);
1860 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1863 // We are going to have to map operands from the original BB block into the
1864 // PredBB block. Evaluate PHI nodes in BB.
1865 DenseMap<Instruction*, Value*> ValueMapping;
1867 BasicBlock::iterator BI = BB->begin();
1868 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1869 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1870 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1871 // mapping and using it to remap operands in the cloned instructions.
1872 for (; BI != BB->end(); ++BI) {
1873 Instruction *New = BI->clone();
1875 // Remap operands to patch up intra-block references.
1876 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1877 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1878 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1879 if (I != ValueMapping.end())
1880 New->setOperand(i, I->second);
1883 // If this instruction can be simplified after the operands are updated,
1884 // just use the simplified value instead. This frequently happens due to
1887 SimplifyInstruction(New, BB->getModule()->getDataLayout())) {
1888 ValueMapping[&*BI] = IV;
1889 if (!New->mayHaveSideEffects()) {
1894 ValueMapping[&*BI] = New;
1897 // Otherwise, insert the new instruction into the block.
1898 New->setName(BI->getName());
1899 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
1903 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1904 // add entries to the PHI nodes for branch from PredBB now.
1905 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1906 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1908 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1911 // If there were values defined in BB that are used outside the block, then we
1912 // now have to update all uses of the value to use either the original value,
1913 // the cloned value, or some PHI derived value. This can require arbitrary
1914 // PHI insertion, of which we are prepared to do, clean these up now.
1915 SSAUpdater SSAUpdate;
1916 SmallVector<Use*, 16> UsesToRename;
1917 for (Instruction &I : *BB) {
1918 // Scan all uses of this instruction to see if it is used outside of its
1919 // block, and if so, record them in UsesToRename.
1920 for (Use &U : I.uses()) {
1921 Instruction *User = cast<Instruction>(U.getUser());
1922 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1923 if (UserPN->getIncomingBlock(U) == BB)
1925 } else if (User->getParent() == BB)
1928 UsesToRename.push_back(&U);
1931 // If there are no uses outside the block, we're done with this instruction.
1932 if (UsesToRename.empty())
1935 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
1937 // We found a use of I outside of BB. Rename all uses of I that are outside
1938 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1939 // with the two values we know.
1940 SSAUpdate.Initialize(I.getType(), I.getName());
1941 SSAUpdate.AddAvailableValue(BB, &I);
1942 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]);
1944 while (!UsesToRename.empty())
1945 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1946 DEBUG(dbgs() << "\n");
1949 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1951 BB->removePredecessor(PredBB, true);
1953 // Remove the unconditional branch at the end of the PredBB block.
1954 OldPredBranch->eraseFromParent();
1960 /// TryToUnfoldSelect - Look for blocks of the form
1966 /// %p = phi [%a, %bb1] ...
1970 /// And expand the select into a branch structure if one of its arms allows %c
1971 /// to be folded. This later enables threading from bb1 over bb2.
1972 bool JumpThreadingPass::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1973 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1974 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1975 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1977 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1978 CondLHS->getParent() != BB)
1981 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1982 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1983 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1985 // Look if one of the incoming values is a select in the corresponding
1987 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1990 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1991 if (!PredTerm || !PredTerm->isUnconditional())
1994 // Now check if one of the select values would allow us to constant fold the
1995 // terminator in BB. We don't do the transform if both sides fold, those
1996 // cases will be threaded in any case.
1997 LazyValueInfo::Tristate LHSFolds =
1998 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1999 CondRHS, Pred, BB, CondCmp);
2000 LazyValueInfo::Tristate RHSFolds =
2001 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2002 CondRHS, Pred, BB, CondCmp);
2003 if ((LHSFolds != LazyValueInfo::Unknown ||
2004 RHSFolds != LazyValueInfo::Unknown) &&
2005 LHSFolds != RHSFolds) {
2006 // Expand the select.
2015 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2016 BB->getParent(), BB);
2017 // Move the unconditional branch to NewBB.
2018 PredTerm->removeFromParent();
2019 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2020 // Create a conditional branch and update PHI nodes.
2021 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2022 CondLHS->setIncomingValue(I, SI->getFalseValue());
2023 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
2024 // The select is now dead.
2025 SI->eraseFromParent();
2027 // Update any other PHI nodes in BB.
2028 for (BasicBlock::iterator BI = BB->begin();
2029 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2031 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2038 /// TryToUnfoldSelectInCurrBB - Look for PHI/Select in the same BB of the form
2040 /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2041 /// %s = select p, trueval, falseval
2043 /// And expand the select into a branch structure. This later enables
2044 /// jump-threading over bb in this pass.
2046 /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2047 /// select if the associated PHI has at least one constant. If the unfolded
2048 /// select is not jump-threaded, it will be folded again in the later
2050 bool JumpThreadingPass::TryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2051 // If threading this would thread across a loop header, don't thread the edge.
2052 // See the comments above FindLoopHeaders for justifications and caveats.
2053 if (LoopHeaders.count(BB))
2056 // Look for a Phi/Select pair in the same basic block. The Phi feeds the
2057 // condition of the Select and at least one of the incoming values is a
2059 for (BasicBlock::iterator BI = BB->begin();
2060 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2061 unsigned NumPHIValues = PN->getNumIncomingValues();
2062 if (NumPHIValues == 0 || !PN->hasOneUse())
2065 SelectInst *SI = dyn_cast<SelectInst>(PN->user_back());
2066 if (!SI || SI->getParent() != BB)
2069 Value *Cond = SI->getCondition();
2070 if (!Cond || Cond != PN || !Cond->getType()->isIntegerTy(1))
2073 bool HasConst = false;
2074 for (unsigned i = 0; i != NumPHIValues; ++i) {
2075 if (PN->getIncomingBlock(i) == BB)
2077 if (isa<ConstantInt>(PN->getIncomingValue(i)))
2082 // Expand the select.
2083 TerminatorInst *Term =
2084 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
2085 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2086 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2087 NewPN->addIncoming(SI->getFalseValue(), BB);
2088 SI->replaceAllUsesWith(NewPN);
2089 SI->eraseFromParent();
2097 /// Try to propagate a guard from the current BB into one of its predecessors
2098 /// in case if another branch of execution implies that the condition of this
2099 /// guard is always true. Currently we only process the simplest case that
2104 /// br i1 %cond, label %T1, label %F1
2110 /// %condGuard = ...
2111 /// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2113 /// And cond either implies condGuard or !condGuard. In this case all the
2114 /// instructions before the guard can be duplicated in both branches, and the
2115 /// guard is then threaded to one of them.
2116 bool JumpThreadingPass::ProcessGuards(BasicBlock *BB) {
2117 using namespace PatternMatch;
2118 // We only want to deal with two predecessors.
2119 BasicBlock *Pred1, *Pred2;
2120 auto PI = pred_begin(BB), PE = pred_end(BB);
2132 // Try to thread one of the guards of the block.
2133 // TODO: Look up deeper than to immediate predecessor?
2134 auto *Parent = Pred1->getSinglePredecessor();
2135 if (!Parent || Parent != Pred2->getSinglePredecessor())
2138 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2140 if (match(&I, m_Intrinsic<Intrinsic::experimental_guard>()))
2141 if (ThreadGuard(BB, cast<IntrinsicInst>(&I), BI))
2147 /// Try to propagate the guard from BB which is the lower block of a diamond
2148 /// to one of its branches, in case if diamond's condition implies guard's
2150 bool JumpThreadingPass::ThreadGuard(BasicBlock *BB, IntrinsicInst *Guard,
2152 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?");
2153 assert(BI->isConditional() && "Unconditional branch has 2 successors?");
2154 Value *GuardCond = Guard->getArgOperand(0);
2155 Value *BranchCond = BI->getCondition();
2156 BasicBlock *TrueDest = BI->getSuccessor(0);
2157 BasicBlock *FalseDest = BI->getSuccessor(1);
2159 auto &DL = BB->getModule()->getDataLayout();
2160 bool TrueDestIsSafe = false;
2161 bool FalseDestIsSafe = false;
2163 // True dest is safe if BranchCond => GuardCond.
2164 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
2166 TrueDestIsSafe = true;
2168 // False dest is safe if !BranchCond => GuardCond.
2170 isImpliedCondition(BranchCond, GuardCond, DL, /* InvertAPred */ true);
2172 FalseDestIsSafe = true;
2175 if (!TrueDestIsSafe && !FalseDestIsSafe)
2178 BasicBlock *UnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
2179 BasicBlock *GuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
2181 ValueToValueMapTy UnguardedMapping, GuardedMapping;
2182 Instruction *AfterGuard = Guard->getNextNode();
2183 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
2184 if (Cost > BBDupThreshold)
2186 // Duplicate all instructions before the guard and the guard itself to the
2187 // branch where implication is not proved.
2188 GuardedBlock = DuplicateInstructionsInSplitBetween(
2189 BB, GuardedBlock, AfterGuard, GuardedMapping);
2190 assert(GuardedBlock && "Could not create the guarded block?");
2191 // Duplicate all instructions before the guard in the unguarded branch.
2192 // Since we have successfully duplicated the guarded block and this block
2193 // has fewer instructions, we expect it to succeed.
2194 UnguardedBlock = DuplicateInstructionsInSplitBetween(BB, UnguardedBlock,
2195 Guard, UnguardedMapping);
2196 assert(UnguardedBlock && "Could not create the unguarded block?");
2197 DEBUG(dbgs() << "Moved guard " << *Guard << " to block "
2198 << GuardedBlock->getName() << "\n");
2200 // Some instructions before the guard may still have uses. For them, we need
2201 // to create Phi nodes merging their copies in both guarded and unguarded
2202 // branches. Those instructions that have no uses can be just removed.
2203 SmallVector<Instruction *, 4> ToRemove;
2204 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
2205 if (!isa<PHINode>(&*BI))
2206 ToRemove.push_back(&*BI);
2208 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
2209 assert(InsertionPoint && "Empty block?");
2210 // Substitute with Phis & remove.
2211 for (auto *Inst : reverse(ToRemove)) {
2212 if (!Inst->use_empty()) {
2213 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
2214 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
2215 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
2216 NewPN->insertBefore(InsertionPoint);
2217 Inst->replaceAllUsesWith(NewPN);
2219 Inst->eraseFromParent();