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()->isVectorTy()) {
584 Constant *CmpConst = cast<Constant>(Cmp->getOperand(1));
586 if (!isa<Instruction>(Cmp->getOperand(0)) ||
587 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
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 CmpConst, 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 PredValueInfoTy LHSVals;
607 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
610 for (const auto &LHSVal : LHSVals) {
611 Constant *V = LHSVal.first;
612 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
614 if (Constant *KC = getKnownConstant(Folded, WantInteger))
615 Result.push_back(std::make_pair(KC, LHSVal.second));
618 return !Result.empty();
622 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
623 // Handle select instructions where at least one operand is a known constant
624 // and we can figure out the condition value for any predecessor block.
625 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
626 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
627 PredValueInfoTy Conds;
628 if ((TrueVal || FalseVal) &&
629 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
630 WantInteger, CxtI)) {
631 for (auto &C : Conds) {
632 Constant *Cond = C.first;
634 // Figure out what value to use for the condition.
636 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
638 KnownCond = CI->isOne();
640 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
641 // Either operand will do, so be sure to pick the one that's a known
643 // FIXME: Do this more cleverly if both values are known constants?
644 KnownCond = (TrueVal != nullptr);
647 // See if the select has a known constant value for this predecessor.
648 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
649 Result.push_back(std::make_pair(Val, C.second));
652 return !Result.empty();
656 // If all else fails, see if LVI can figure out a constant value for us.
657 Constant *CI = LVI->getConstant(V, BB, CxtI);
658 if (Constant *KC = getKnownConstant(CI, Preference)) {
659 for (BasicBlock *Pred : predecessors(BB))
660 Result.push_back(std::make_pair(KC, Pred));
663 return !Result.empty();
668 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
669 /// in an undefined jump, decide which block is best to revector to.
671 /// Since we can pick an arbitrary destination, we pick the successor with the
672 /// fewest predecessors. This should reduce the in-degree of the others.
674 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
675 TerminatorInst *BBTerm = BB->getTerminator();
676 unsigned MinSucc = 0;
677 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
678 // Compute the successor with the minimum number of predecessors.
679 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
680 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
681 TestBB = BBTerm->getSuccessor(i);
682 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
683 if (NumPreds < MinNumPreds) {
685 MinNumPreds = NumPreds;
692 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
693 if (!BB->hasAddressTaken()) return false;
695 // If the block has its address taken, it may be a tree of dead constants
696 // hanging off of it. These shouldn't keep the block alive.
697 BlockAddress *BA = BlockAddress::get(BB);
698 BA->removeDeadConstantUsers();
699 return !BA->use_empty();
702 /// ProcessBlock - If there are any predecessors whose control can be threaded
703 /// through to a successor, transform them now.
704 bool JumpThreadingPass::ProcessBlock(BasicBlock *BB) {
705 // If the block is trivially dead, just return and let the caller nuke it.
706 // This simplifies other transformations.
707 if (pred_empty(BB) &&
708 BB != &BB->getParent()->getEntryBlock())
711 // If this block has a single predecessor, and if that pred has a single
712 // successor, merge the blocks. This encourages recursive jump threading
713 // because now the condition in this block can be threaded through
714 // predecessors of our predecessor block.
715 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
716 const TerminatorInst *TI = SinglePred->getTerminator();
717 if (!TI->isExceptional() && TI->getNumSuccessors() == 1 &&
718 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
719 // If SinglePred was a loop header, BB becomes one.
720 if (LoopHeaders.erase(SinglePred))
721 LoopHeaders.insert(BB);
723 LVI->eraseBlock(SinglePred);
724 MergeBasicBlockIntoOnlyPred(BB);
730 if (TryToUnfoldSelectInCurrBB(BB))
733 // Look if we can propagate guards to predecessors.
734 if (HasGuards && ProcessGuards(BB))
737 // What kind of constant we're looking for.
738 ConstantPreference Preference = WantInteger;
740 // Look to see if the terminator is a conditional branch, switch or indirect
741 // branch, if not we can't thread it.
743 Instruction *Terminator = BB->getTerminator();
744 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
745 // Can't thread an unconditional jump.
746 if (BI->isUnconditional()) return false;
747 Condition = BI->getCondition();
748 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
749 Condition = SI->getCondition();
750 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
751 // Can't thread indirect branch with no successors.
752 if (IB->getNumSuccessors() == 0) return false;
753 Condition = IB->getAddress()->stripPointerCasts();
754 Preference = WantBlockAddress;
756 return false; // Must be an invoke.
759 // Run constant folding to see if we can reduce the condition to a simple
761 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
763 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
765 I->replaceAllUsesWith(SimpleVal);
766 if (isInstructionTriviallyDead(I, TLI))
767 I->eraseFromParent();
768 Condition = SimpleVal;
772 // If the terminator is branching on an undef, we can pick any of the
773 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
774 if (isa<UndefValue>(Condition)) {
775 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
777 // Fold the branch/switch.
778 TerminatorInst *BBTerm = BB->getTerminator();
779 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
780 if (i == BestSucc) continue;
781 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
784 DEBUG(dbgs() << " In block '" << BB->getName()
785 << "' folding undef terminator: " << *BBTerm << '\n');
786 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
787 BBTerm->eraseFromParent();
791 // If the terminator of this block is branching on a constant, simplify the
792 // terminator to an unconditional branch. This can occur due to threading in
794 if (getKnownConstant(Condition, Preference)) {
795 DEBUG(dbgs() << " In block '" << BB->getName()
796 << "' folding terminator: " << *BB->getTerminator() << '\n');
798 ConstantFoldTerminator(BB, true);
802 Instruction *CondInst = dyn_cast<Instruction>(Condition);
804 // All the rest of our checks depend on the condition being an instruction.
806 // FIXME: Unify this with code below.
807 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
812 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
813 // If we're branching on a conditional, LVI might be able to determine
814 // it's value at the branch instruction. We only handle comparisons
815 // against a constant at this time.
816 // TODO: This should be extended to handle switches as well.
817 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
818 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
819 if (CondBr && CondConst) {
820 // We should have returned as soon as we turn a conditional branch to
821 // unconditional. Because its no longer interesting as far as jump
822 // threading is concerned.
823 assert(CondBr->isConditional() && "Threading on unconditional terminator");
825 LazyValueInfo::Tristate Ret =
826 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
828 if (Ret != LazyValueInfo::Unknown) {
829 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
830 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
831 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
832 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
833 CondBr->eraseFromParent();
834 if (CondCmp->use_empty())
835 CondCmp->eraseFromParent();
836 else if (CondCmp->getParent() == BB) {
837 // If the fact we just learned is true for all uses of the
838 // condition, replace it with a constant value
839 auto *CI = Ret == LazyValueInfo::True ?
840 ConstantInt::getTrue(CondCmp->getType()) :
841 ConstantInt::getFalse(CondCmp->getType());
842 CondCmp->replaceAllUsesWith(CI);
843 CondCmp->eraseFromParent();
848 // We did not manage to simplify this branch, try to see whether
849 // CondCmp depends on a known phi-select pattern.
850 if (TryToUnfoldSelect(CondCmp, BB))
855 // Check for some cases that are worth simplifying. Right now we want to look
856 // for loads that are used by a switch or by the condition for the branch. If
857 // we see one, check to see if it's partially redundant. If so, insert a PHI
858 // which can then be used to thread the values.
860 Value *SimplifyValue = CondInst;
861 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
862 if (isa<Constant>(CondCmp->getOperand(1)))
863 SimplifyValue = CondCmp->getOperand(0);
865 // TODO: There are other places where load PRE would be profitable, such as
866 // more complex comparisons.
867 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
868 if (SimplifyPartiallyRedundantLoad(LI))
871 // Handle a variety of cases where we are branching on something derived from
872 // a PHI node in the current block. If we can prove that any predecessors
873 // compute a predictable value based on a PHI node, thread those predecessors.
875 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
878 // If this is an otherwise-unfoldable branch on a phi node in the current
879 // block, see if we can simplify.
880 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
881 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
882 return ProcessBranchOnPHI(PN);
884 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
885 if (CondInst->getOpcode() == Instruction::Xor &&
886 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
887 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
889 // Search for a stronger dominating condition that can be used to simplify a
890 // conditional branch leaving BB.
891 if (ProcessImpliedCondition(BB))
897 bool JumpThreadingPass::ProcessImpliedCondition(BasicBlock *BB) {
898 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
899 if (!BI || !BI->isConditional())
902 Value *Cond = BI->getCondition();
903 BasicBlock *CurrentBB = BB;
904 BasicBlock *CurrentPred = BB->getSinglePredecessor();
907 auto &DL = BB->getModule()->getDataLayout();
909 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
910 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
911 if (!PBI || !PBI->isConditional())
913 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
916 bool FalseDest = PBI->getSuccessor(1) == CurrentBB;
917 Optional<bool> Implication =
918 isImpliedCondition(PBI->getCondition(), Cond, DL, FalseDest);
920 BI->getSuccessor(*Implication ? 1 : 0)->removePredecessor(BB);
921 BranchInst::Create(BI->getSuccessor(*Implication ? 0 : 1), BI);
922 BI->eraseFromParent();
925 CurrentBB = CurrentPred;
926 CurrentPred = CurrentBB->getSinglePredecessor();
932 /// Return true if Op is an instruction defined in the given block.
933 static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
934 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
935 if (OpInst->getParent() == BB)
940 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
941 /// load instruction, eliminate it by replacing it with a PHI node. This is an
942 /// important optimization that encourages jump threading, and needs to be run
943 /// interlaced with other jump threading tasks.
944 bool JumpThreadingPass::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
945 // Don't hack volatile and ordered loads.
946 if (!LI->isUnordered()) return false;
948 // If the load is defined in a block with exactly one predecessor, it can't be
949 // partially redundant.
950 BasicBlock *LoadBB = LI->getParent();
951 if (LoadBB->getSinglePredecessor())
954 // If the load is defined in an EH pad, it can't be partially redundant,
955 // because the edges between the invoke and the EH pad cannot have other
956 // instructions between them.
957 if (LoadBB->isEHPad())
960 Value *LoadedPtr = LI->getOperand(0);
962 // If the loaded operand is defined in the LoadBB and its not a phi,
963 // it can't be available in predecessors.
964 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
967 // Scan a few instructions up from the load, to see if it is obviously live at
968 // the entry to its block.
969 BasicBlock::iterator BBIt(LI);
971 if (Value *AvailableVal = FindAvailableLoadedValue(
972 LI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
973 // If the value of the load is locally available within the block, just use
974 // it. This frequently occurs for reg2mem'd allocas.
977 LoadInst *NLI = cast<LoadInst>(AvailableVal);
978 combineMetadataForCSE(NLI, LI);
981 // If the returned value is the load itself, replace with an undef. This can
982 // only happen in dead loops.
983 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
984 if (AvailableVal->getType() != LI->getType())
986 CastInst::CreateBitOrPointerCast(AvailableVal, LI->getType(), "", LI);
987 LI->replaceAllUsesWith(AvailableVal);
988 LI->eraseFromParent();
992 // Otherwise, if we scanned the whole block and got to the top of the block,
993 // we know the block is locally transparent to the load. If not, something
994 // might clobber its value.
995 if (BBIt != LoadBB->begin())
998 // If all of the loads and stores that feed the value have the same AA tags,
999 // then we can propagate them onto any newly inserted loads.
1001 LI->getAAMetadata(AATags);
1003 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1004 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
1005 AvailablePredsTy AvailablePreds;
1006 BasicBlock *OneUnavailablePred = nullptr;
1007 SmallVector<LoadInst*, 8> CSELoads;
1009 // If we got here, the loaded value is transparent through to the start of the
1010 // block. Check to see if it is available in any of the predecessor blocks.
1011 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1012 // If we already scanned this predecessor, skip it.
1013 if (!PredsScanned.insert(PredBB).second)
1016 BBIt = PredBB->end();
1017 unsigned NumScanedInst = 0;
1018 Value *PredAvailable = nullptr;
1019 // NOTE: We don't CSE load that is volatile or anything stronger than
1020 // unordered, that should have been checked when we entered the function.
1021 assert(LI->isUnordered() && "Attempting to CSE volatile or atomic loads");
1022 // If this is a load on a phi pointer, phi-translate it and search
1023 // for available load/store to the pointer in predecessors.
1024 Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1025 PredAvailable = FindAvailablePtrLoadStore(
1026 Ptr, LI->getType(), LI->isAtomic(), PredBB, BBIt, DefMaxInstsToScan,
1027 AA, &IsLoadCSE, &NumScanedInst);
1029 // If PredBB has a single predecessor, continue scanning through the
1030 // single precessor.
1031 BasicBlock *SinglePredBB = PredBB;
1032 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1033 NumScanedInst < DefMaxInstsToScan) {
1034 SinglePredBB = SinglePredBB->getSinglePredecessor();
1036 BBIt = SinglePredBB->end();
1037 PredAvailable = FindAvailablePtrLoadStore(
1038 Ptr, LI->getType(), LI->isAtomic(), SinglePredBB, BBIt,
1039 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1044 if (!PredAvailable) {
1045 OneUnavailablePred = PredBB;
1050 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1052 // If so, this load is partially redundant. Remember this info so that we
1053 // can create a PHI node.
1054 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1057 // If the loaded value isn't available in any predecessor, it isn't partially
1059 if (AvailablePreds.empty()) return false;
1061 // Okay, the loaded value is available in at least one (and maybe all!)
1062 // predecessors. If the value is unavailable in more than one unique
1063 // predecessor, we want to insert a merge block for those common predecessors.
1064 // This ensures that we only have to insert one reload, thus not increasing
1066 BasicBlock *UnavailablePred = nullptr;
1068 // If there is exactly one predecessor where the value is unavailable, the
1069 // already computed 'OneUnavailablePred' block is it. If it ends in an
1070 // unconditional branch, we know that it isn't a critical edge.
1071 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1072 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1073 UnavailablePred = OneUnavailablePred;
1074 } else if (PredsScanned.size() != AvailablePreds.size()) {
1075 // Otherwise, we had multiple unavailable predecessors or we had a critical
1076 // edge from the one.
1077 SmallVector<BasicBlock*, 8> PredsToSplit;
1078 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1080 for (const auto &AvailablePred : AvailablePreds)
1081 AvailablePredSet.insert(AvailablePred.first);
1083 // Add all the unavailable predecessors to the PredsToSplit list.
1084 for (BasicBlock *P : predecessors(LoadBB)) {
1085 // If the predecessor is an indirect goto, we can't split the edge.
1086 if (isa<IndirectBrInst>(P->getTerminator()))
1089 if (!AvailablePredSet.count(P))
1090 PredsToSplit.push_back(P);
1093 // Split them out to their own block.
1094 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1097 // If the value isn't available in all predecessors, then there will be
1098 // exactly one where it isn't available. Insert a load on that edge and add
1099 // it to the AvailablePreds list.
1100 if (UnavailablePred) {
1101 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1102 "Can't handle critical edge here!");
1103 LoadInst *NewVal = new LoadInst(
1104 LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1105 LI->getName() + ".pr", false, LI->getAlignment(), LI->getOrdering(),
1106 LI->getSynchScope(), UnavailablePred->getTerminator());
1107 NewVal->setDebugLoc(LI->getDebugLoc());
1109 NewVal->setAAMetadata(AATags);
1111 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1114 // Now we know that each predecessor of this block has a value in
1115 // AvailablePreds, sort them for efficient access as we're walking the preds.
1116 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1118 // Create a PHI node at the start of the block for the PRE'd load value.
1119 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1120 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
1123 PN->setDebugLoc(LI->getDebugLoc());
1125 // Insert new entries into the PHI for each predecessor. A single block may
1126 // have multiple entries here.
1127 for (pred_iterator PI = PB; PI != PE; ++PI) {
1128 BasicBlock *P = *PI;
1129 AvailablePredsTy::iterator I =
1130 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1131 std::make_pair(P, (Value*)nullptr));
1133 assert(I != AvailablePreds.end() && I->first == P &&
1134 "Didn't find entry for predecessor!");
1136 // If we have an available predecessor but it requires casting, insert the
1137 // cast in the predecessor and use the cast. Note that we have to update the
1138 // AvailablePreds vector as we go so that all of the PHI entries for this
1139 // predecessor use the same bitcast.
1140 Value *&PredV = I->second;
1141 if (PredV->getType() != LI->getType())
1142 PredV = CastInst::CreateBitOrPointerCast(PredV, LI->getType(), "",
1143 P->getTerminator());
1145 PN->addIncoming(PredV, I->first);
1148 for (LoadInst *PredLI : CSELoads) {
1149 combineMetadataForCSE(PredLI, LI);
1152 LI->replaceAllUsesWith(PN);
1153 LI->eraseFromParent();
1158 /// FindMostPopularDest - The specified list contains multiple possible
1159 /// threadable destinations. Pick the one that occurs the most frequently in
1162 FindMostPopularDest(BasicBlock *BB,
1163 const SmallVectorImpl<std::pair<BasicBlock*,
1164 BasicBlock*> > &PredToDestList) {
1165 assert(!PredToDestList.empty());
1167 // Determine popularity. If there are multiple possible destinations, we
1168 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1169 // blocks with known and real destinations to threading undef. We'll handle
1170 // them later if interesting.
1171 DenseMap<BasicBlock*, unsigned> DestPopularity;
1172 for (const auto &PredToDest : PredToDestList)
1173 if (PredToDest.second)
1174 DestPopularity[PredToDest.second]++;
1176 // Find the most popular dest.
1177 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1178 BasicBlock *MostPopularDest = DPI->first;
1179 unsigned Popularity = DPI->second;
1180 SmallVector<BasicBlock*, 4> SamePopularity;
1182 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1183 // If the popularity of this entry isn't higher than the popularity we've
1184 // seen so far, ignore it.
1185 if (DPI->second < Popularity)
1187 else if (DPI->second == Popularity) {
1188 // If it is the same as what we've seen so far, keep track of it.
1189 SamePopularity.push_back(DPI->first);
1191 // If it is more popular, remember it.
1192 SamePopularity.clear();
1193 MostPopularDest = DPI->first;
1194 Popularity = DPI->second;
1198 // Okay, now we know the most popular destination. If there is more than one
1199 // destination, we need to determine one. This is arbitrary, but we need
1200 // to make a deterministic decision. Pick the first one that appears in the
1202 if (!SamePopularity.empty()) {
1203 SamePopularity.push_back(MostPopularDest);
1204 TerminatorInst *TI = BB->getTerminator();
1205 for (unsigned i = 0; ; ++i) {
1206 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1208 if (!is_contained(SamePopularity, TI->getSuccessor(i)))
1211 MostPopularDest = TI->getSuccessor(i);
1216 // Okay, we have finally picked the most popular destination.
1217 return MostPopularDest;
1220 bool JumpThreadingPass::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1221 ConstantPreference Preference,
1222 Instruction *CxtI) {
1223 // If threading this would thread across a loop header, don't even try to
1225 if (LoopHeaders.count(BB))
1228 PredValueInfoTy PredValues;
1229 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1232 assert(!PredValues.empty() &&
1233 "ComputeValueKnownInPredecessors returned true with no values");
1235 DEBUG(dbgs() << "IN BB: " << *BB;
1236 for (const auto &PredValue : PredValues) {
1237 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1239 << " for pred '" << PredValue.second->getName() << "'.\n";
1242 // Decide what we want to thread through. Convert our list of known values to
1243 // a list of known destinations for each pred. This also discards duplicate
1244 // predecessors and keeps track of the undefined inputs (which are represented
1245 // as a null dest in the PredToDestList).
1246 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1247 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1249 BasicBlock *OnlyDest = nullptr;
1250 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1251 Constant *OnlyVal = nullptr;
1252 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1254 unsigned PredWithKnownDest = 0;
1255 for (const auto &PredValue : PredValues) {
1256 BasicBlock *Pred = PredValue.second;
1257 if (!SeenPreds.insert(Pred).second)
1258 continue; // Duplicate predecessor entry.
1260 Constant *Val = PredValue.first;
1263 if (isa<UndefValue>(Val))
1265 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1266 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1267 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1268 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1269 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1270 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1272 assert(isa<IndirectBrInst>(BB->getTerminator())
1273 && "Unexpected terminator");
1274 assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress");
1275 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1278 // If we have exactly one destination, remember it for efficiency below.
1279 if (PredToDestList.empty()) {
1283 if (OnlyDest != DestBB)
1284 OnlyDest = MultipleDestSentinel;
1285 // It possible we have same destination, but different value, e.g. default
1286 // case in switchinst.
1288 OnlyVal = MultipleVal;
1291 // We know where this predecessor is going.
1292 ++PredWithKnownDest;
1294 // If the predecessor ends with an indirect goto, we can't change its
1296 if (isa<IndirectBrInst>(Pred->getTerminator()))
1299 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1302 // If all edges were unthreadable, we fail.
1303 if (PredToDestList.empty())
1306 // If all the predecessors go to a single known successor, we want to fold,
1307 // not thread. By doing so, we do not need to duplicate the current block and
1308 // also miss potential opportunities in case we dont/cant duplicate.
1309 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1310 if (PredWithKnownDest ==
1311 (size_t)std::distance(pred_begin(BB), pred_end(BB))) {
1312 bool SeenFirstBranchToOnlyDest = false;
1313 for (BasicBlock *SuccBB : successors(BB)) {
1314 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest)
1315 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1317 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1320 // Finally update the terminator.
1321 TerminatorInst *Term = BB->getTerminator();
1322 BranchInst::Create(OnlyDest, Term);
1323 Term->eraseFromParent();
1325 // If the condition is now dead due to the removal of the old terminator,
1327 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1328 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1329 CondInst->eraseFromParent();
1330 else if (OnlyVal && OnlyVal != MultipleVal &&
1331 CondInst->getParent() == BB) {
1332 // If we just learned Cond is the same value for all uses of the
1333 // condition, replace it with a constant value
1334 CondInst->replaceAllUsesWith(OnlyVal);
1335 if (!CondInst->mayHaveSideEffects())
1336 CondInst->eraseFromParent();
1343 // Determine which is the most common successor. If we have many inputs and
1344 // this block is a switch, we want to start by threading the batch that goes
1345 // to the most popular destination first. If we only know about one
1346 // threadable destination (the common case) we can avoid this.
1347 BasicBlock *MostPopularDest = OnlyDest;
1349 if (MostPopularDest == MultipleDestSentinel)
1350 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1352 // Now that we know what the most popular destination is, factor all
1353 // predecessors that will jump to it into a single predecessor.
1354 SmallVector<BasicBlock*, 16> PredsToFactor;
1355 for (const auto &PredToDest : PredToDestList)
1356 if (PredToDest.second == MostPopularDest) {
1357 BasicBlock *Pred = PredToDest.first;
1359 // This predecessor may be a switch or something else that has multiple
1360 // edges to the block. Factor each of these edges by listing them
1361 // according to # occurrences in PredsToFactor.
1362 for (BasicBlock *Succ : successors(Pred))
1364 PredsToFactor.push_back(Pred);
1367 // If the threadable edges are branching on an undefined value, we get to pick
1368 // the destination that these predecessors should get to.
1369 if (!MostPopularDest)
1370 MostPopularDest = BB->getTerminator()->
1371 getSuccessor(GetBestDestForJumpOnUndef(BB));
1373 // Ok, try to thread it!
1374 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1377 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1378 /// a PHI node in the current block. See if there are any simplifications we
1379 /// can do based on inputs to the phi node.
1381 bool JumpThreadingPass::ProcessBranchOnPHI(PHINode *PN) {
1382 BasicBlock *BB = PN->getParent();
1384 // TODO: We could make use of this to do it once for blocks with common PHI
1386 SmallVector<BasicBlock*, 1> PredBBs;
1389 // If any of the predecessor blocks end in an unconditional branch, we can
1390 // *duplicate* the conditional branch into that block in order to further
1391 // encourage jump threading and to eliminate cases where we have branch on a
1392 // phi of an icmp (branch on icmp is much better).
1393 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1394 BasicBlock *PredBB = PN->getIncomingBlock(i);
1395 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1396 if (PredBr->isUnconditional()) {
1397 PredBBs[0] = PredBB;
1398 // Try to duplicate BB into PredBB.
1399 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1407 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1408 /// a xor instruction in the current block. See if there are any
1409 /// simplifications we can do based on inputs to the xor.
1411 bool JumpThreadingPass::ProcessBranchOnXOR(BinaryOperator *BO) {
1412 BasicBlock *BB = BO->getParent();
1414 // If either the LHS or RHS of the xor is a constant, don't do this
1416 if (isa<ConstantInt>(BO->getOperand(0)) ||
1417 isa<ConstantInt>(BO->getOperand(1)))
1420 // If the first instruction in BB isn't a phi, we won't be able to infer
1421 // anything special about any particular predecessor.
1422 if (!isa<PHINode>(BB->front()))
1425 // If this BB is a landing pad, we won't be able to split the edge into it.
1429 // If we have a xor as the branch input to this block, and we know that the
1430 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1431 // the condition into the predecessor and fix that value to true, saving some
1432 // logical ops on that path and encouraging other paths to simplify.
1434 // This copies something like this:
1437 // %X = phi i1 [1], [%X']
1438 // %Y = icmp eq i32 %A, %B
1439 // %Z = xor i1 %X, %Y
1444 // %Y = icmp ne i32 %A, %B
1447 PredValueInfoTy XorOpValues;
1449 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1451 assert(XorOpValues.empty());
1452 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1458 assert(!XorOpValues.empty() &&
1459 "ComputeValueKnownInPredecessors returned true with no values");
1461 // Scan the information to see which is most popular: true or false. The
1462 // predecessors can be of the set true, false, or undef.
1463 unsigned NumTrue = 0, NumFalse = 0;
1464 for (const auto &XorOpValue : XorOpValues) {
1465 if (isa<UndefValue>(XorOpValue.first))
1466 // Ignore undefs for the count.
1468 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1474 // Determine which value to split on, true, false, or undef if neither.
1475 ConstantInt *SplitVal = nullptr;
1476 if (NumTrue > NumFalse)
1477 SplitVal = ConstantInt::getTrue(BB->getContext());
1478 else if (NumTrue != 0 || NumFalse != 0)
1479 SplitVal = ConstantInt::getFalse(BB->getContext());
1481 // Collect all of the blocks that this can be folded into so that we can
1482 // factor this once and clone it once.
1483 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1484 for (const auto &XorOpValue : XorOpValues) {
1485 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1488 BlocksToFoldInto.push_back(XorOpValue.second);
1491 // If we inferred a value for all of the predecessors, then duplication won't
1492 // help us. However, we can just replace the LHS or RHS with the constant.
1493 if (BlocksToFoldInto.size() ==
1494 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1496 // If all preds provide undef, just nuke the xor, because it is undef too.
1497 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1498 BO->eraseFromParent();
1499 } else if (SplitVal->isZero()) {
1500 // If all preds provide 0, replace the xor with the other input.
1501 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1502 BO->eraseFromParent();
1504 // If all preds provide 1, set the computed value to 1.
1505 BO->setOperand(!isLHS, SplitVal);
1511 // Try to duplicate BB into PredBB.
1512 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1516 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1517 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1518 /// NewPred using the entries from OldPred (suitably mapped).
1519 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1520 BasicBlock *OldPred,
1521 BasicBlock *NewPred,
1522 DenseMap<Instruction*, Value*> &ValueMap) {
1523 for (BasicBlock::iterator PNI = PHIBB->begin();
1524 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1525 // Ok, we have a PHI node. Figure out what the incoming value was for the
1527 Value *IV = PN->getIncomingValueForBlock(OldPred);
1529 // Remap the value if necessary.
1530 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1531 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1532 if (I != ValueMap.end())
1536 PN->addIncoming(IV, NewPred);
1540 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1541 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1542 /// across BB. Transform the IR to reflect this change.
1543 bool JumpThreadingPass::ThreadEdge(BasicBlock *BB,
1544 const SmallVectorImpl<BasicBlock *> &PredBBs,
1545 BasicBlock *SuccBB) {
1546 // If threading to the same block as we come from, we would infinite loop.
1548 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1549 << "' - would thread to self!\n");
1553 // If threading this would thread across a loop header, don't thread the edge.
1554 // See the comments above FindLoopHeaders for justifications and caveats.
1555 if (LoopHeaders.count(BB)) {
1556 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1557 << "' to dest BB '" << SuccBB->getName()
1558 << "' - it might create an irreducible loop!\n");
1562 unsigned JumpThreadCost =
1563 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1564 if (JumpThreadCost > BBDupThreshold) {
1565 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1566 << "' - Cost is too high: " << JumpThreadCost << "\n");
1570 // And finally, do it! Start by factoring the predecessors if needed.
1572 if (PredBBs.size() == 1)
1573 PredBB = PredBBs[0];
1575 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1576 << " common predecessors.\n");
1577 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1580 // And finally, do it!
1581 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1582 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1583 << ", across block:\n "
1586 LVI->threadEdge(PredBB, BB, SuccBB);
1588 // We are going to have to map operands from the original BB block to the new
1589 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1590 // account for entry from PredBB.
1591 DenseMap<Instruction*, Value*> ValueMapping;
1593 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1594 BB->getName()+".thread",
1595 BB->getParent(), BB);
1596 NewBB->moveAfter(PredBB);
1598 // Set the block frequency of NewBB.
1599 if (HasProfileData) {
1601 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1602 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1605 BasicBlock::iterator BI = BB->begin();
1606 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1607 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1609 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1610 // mapping and using it to remap operands in the cloned instructions.
1611 for (; !isa<TerminatorInst>(BI); ++BI) {
1612 Instruction *New = BI->clone();
1613 New->setName(BI->getName());
1614 NewBB->getInstList().push_back(New);
1615 ValueMapping[&*BI] = New;
1617 // Remap operands to patch up intra-block references.
1618 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1619 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1620 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1621 if (I != ValueMapping.end())
1622 New->setOperand(i, I->second);
1626 // We didn't copy the terminator from BB over to NewBB, because there is now
1627 // an unconditional jump to SuccBB. Insert the unconditional jump.
1628 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
1629 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1631 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1632 // PHI nodes for NewBB now.
1633 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1635 // If there were values defined in BB that are used outside the block, then we
1636 // now have to update all uses of the value to use either the original value,
1637 // the cloned value, or some PHI derived value. This can require arbitrary
1638 // PHI insertion, of which we are prepared to do, clean these up now.
1639 SSAUpdater SSAUpdate;
1640 SmallVector<Use*, 16> UsesToRename;
1641 for (Instruction &I : *BB) {
1642 // Scan all uses of this instruction to see if it is used outside of its
1643 // block, and if so, record them in UsesToRename.
1644 for (Use &U : I.uses()) {
1645 Instruction *User = cast<Instruction>(U.getUser());
1646 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1647 if (UserPN->getIncomingBlock(U) == BB)
1649 } else if (User->getParent() == BB)
1652 UsesToRename.push_back(&U);
1655 // If there are no uses outside the block, we're done with this instruction.
1656 if (UsesToRename.empty())
1659 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
1661 // We found a use of I outside of BB. Rename all uses of I that are outside
1662 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1663 // with the two values we know.
1664 SSAUpdate.Initialize(I.getType(), I.getName());
1665 SSAUpdate.AddAvailableValue(BB, &I);
1666 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
1668 while (!UsesToRename.empty())
1669 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1670 DEBUG(dbgs() << "\n");
1674 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1675 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1676 // us to simplify any PHI nodes in BB.
1677 TerminatorInst *PredTerm = PredBB->getTerminator();
1678 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1679 if (PredTerm->getSuccessor(i) == BB) {
1680 BB->removePredecessor(PredBB, true);
1681 PredTerm->setSuccessor(i, NewBB);
1684 // At this point, the IR is fully up to date and consistent. Do a quick scan
1685 // over the new instructions and zap any that are constants or dead. This
1686 // frequently happens because of phi translation.
1687 SimplifyInstructionsInBlock(NewBB, TLI);
1689 // Update the edge weight from BB to SuccBB, which should be less than before.
1690 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
1692 // Threaded an edge!
1697 /// Create a new basic block that will be the predecessor of BB and successor of
1698 /// all blocks in Preds. When profile data is available, update the frequency of
1700 BasicBlock *JumpThreadingPass::SplitBlockPreds(BasicBlock *BB,
1701 ArrayRef<BasicBlock *> Preds,
1702 const char *Suffix) {
1703 // Collect the frequencies of all predecessors of BB, which will be used to
1704 // update the edge weight on BB->SuccBB.
1705 BlockFrequency PredBBFreq(0);
1707 for (auto Pred : Preds)
1708 PredBBFreq += BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB);
1710 BasicBlock *PredBB = SplitBlockPredecessors(BB, Preds, Suffix);
1712 // Set the block frequency of the newly created PredBB, which is the sum of
1713 // frequencies of Preds.
1715 BFI->setBlockFreq(PredBB, PredBBFreq.getFrequency());
1719 bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
1720 const TerminatorInst *TI = BB->getTerminator();
1721 assert(TI->getNumSuccessors() > 1 && "not a split");
1723 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
1727 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
1728 if (MDName->getString() != "branch_weights")
1731 // Ensure there are weights for all of the successors. Note that the first
1732 // operand to the metadata node is a name, not a weight.
1733 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
1736 /// Update the block frequency of BB and branch weight and the metadata on the
1737 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
1738 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
1739 void JumpThreadingPass::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
1742 BasicBlock *SuccBB) {
1743 if (!HasProfileData)
1746 assert(BFI && BPI && "BFI & BPI should have been created here");
1748 // As the edge from PredBB to BB is deleted, we have to update the block
1750 auto BBOrigFreq = BFI->getBlockFreq(BB);
1751 auto NewBBFreq = BFI->getBlockFreq(NewBB);
1752 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
1753 auto BBNewFreq = BBOrigFreq - NewBBFreq;
1754 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
1756 // Collect updated outgoing edges' frequencies from BB and use them to update
1757 // edge probabilities.
1758 SmallVector<uint64_t, 4> BBSuccFreq;
1759 for (BasicBlock *Succ : successors(BB)) {
1760 auto SuccFreq = (Succ == SuccBB)
1761 ? BB2SuccBBFreq - NewBBFreq
1762 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
1763 BBSuccFreq.push_back(SuccFreq.getFrequency());
1766 uint64_t MaxBBSuccFreq =
1767 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
1769 SmallVector<BranchProbability, 4> BBSuccProbs;
1770 if (MaxBBSuccFreq == 0)
1771 BBSuccProbs.assign(BBSuccFreq.size(),
1772 {1, static_cast<unsigned>(BBSuccFreq.size())});
1774 for (uint64_t Freq : BBSuccFreq)
1775 BBSuccProbs.push_back(
1776 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
1777 // Normalize edge probabilities so that they sum up to one.
1778 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
1782 // Update edge probabilities in BPI.
1783 for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
1784 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
1786 // Update the profile metadata as well.
1788 // Don't do this if the profile of the transformed blocks was statically
1789 // estimated. (This could occur despite the function having an entry
1790 // frequency in completely cold parts of the CFG.)
1792 // In this case we don't want to suggest to subsequent passes that the
1793 // calculated weights are fully consistent. Consider this graph:
1808 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
1809 // the overall probabilities are inconsistent; the total probability that the
1810 // value is either 1, 2 or 3 is 150%.
1812 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
1813 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
1814 // the loop exit edge. Then based solely on static estimation we would assume
1815 // the loop was extremely hot.
1817 // FIXME this locally as well so that BPI and BFI are consistent as well. We
1818 // shouldn't make edges extremely likely or unlikely based solely on static
1820 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
1821 SmallVector<uint32_t, 4> Weights;
1822 for (auto Prob : BBSuccProbs)
1823 Weights.push_back(Prob.getNumerator());
1825 auto TI = BB->getTerminator();
1827 LLVMContext::MD_prof,
1828 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
1832 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1833 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1834 /// If we can duplicate the contents of BB up into PredBB do so now, this
1835 /// improves the odds that the branch will be on an analyzable instruction like
1837 bool JumpThreadingPass::DuplicateCondBranchOnPHIIntoPred(
1838 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
1839 assert(!PredBBs.empty() && "Can't handle an empty set");
1841 // If BB is a loop header, then duplicating this block outside the loop would
1842 // cause us to transform this into an irreducible loop, don't do this.
1843 // See the comments above FindLoopHeaders for justifications and caveats.
1844 if (LoopHeaders.count(BB)) {
1845 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1846 << "' into predecessor block '" << PredBBs[0]->getName()
1847 << "' - it might create an irreducible loop!\n");
1851 unsigned DuplicationCost =
1852 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1853 if (DuplicationCost > BBDupThreshold) {
1854 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1855 << "' - Cost is too high: " << DuplicationCost << "\n");
1859 // And finally, do it! Start by factoring the predecessors if needed.
1861 if (PredBBs.size() == 1)
1862 PredBB = PredBBs[0];
1864 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1865 << " common predecessors.\n");
1866 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1869 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1871 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1872 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1873 << DuplicationCost << " block is:" << *BB << "\n");
1875 // Unless PredBB ends with an unconditional branch, split the edge so that we
1876 // can just clone the bits from BB into the end of the new PredBB.
1877 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1879 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1880 PredBB = SplitEdge(PredBB, BB);
1881 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1884 // We are going to have to map operands from the original BB block into the
1885 // PredBB block. Evaluate PHI nodes in BB.
1886 DenseMap<Instruction*, Value*> ValueMapping;
1888 BasicBlock::iterator BI = BB->begin();
1889 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1890 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1891 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1892 // mapping and using it to remap operands in the cloned instructions.
1893 for (; BI != BB->end(); ++BI) {
1894 Instruction *New = BI->clone();
1896 // Remap operands to patch up intra-block references.
1897 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1898 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1899 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1900 if (I != ValueMapping.end())
1901 New->setOperand(i, I->second);
1904 // If this instruction can be simplified after the operands are updated,
1905 // just use the simplified value instead. This frequently happens due to
1907 if (Value *IV = SimplifyInstruction(
1909 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
1910 ValueMapping[&*BI] = IV;
1911 if (!New->mayHaveSideEffects()) {
1916 ValueMapping[&*BI] = New;
1919 // Otherwise, insert the new instruction into the block.
1920 New->setName(BI->getName());
1921 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
1925 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1926 // add entries to the PHI nodes for branch from PredBB now.
1927 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1928 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1930 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1933 // If there were values defined in BB that are used outside the block, then we
1934 // now have to update all uses of the value to use either the original value,
1935 // the cloned value, or some PHI derived value. This can require arbitrary
1936 // PHI insertion, of which we are prepared to do, clean these up now.
1937 SSAUpdater SSAUpdate;
1938 SmallVector<Use*, 16> UsesToRename;
1939 for (Instruction &I : *BB) {
1940 // Scan all uses of this instruction to see if it is used outside of its
1941 // block, and if so, record them in UsesToRename.
1942 for (Use &U : I.uses()) {
1943 Instruction *User = cast<Instruction>(U.getUser());
1944 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1945 if (UserPN->getIncomingBlock(U) == BB)
1947 } else if (User->getParent() == BB)
1950 UsesToRename.push_back(&U);
1953 // If there are no uses outside the block, we're done with this instruction.
1954 if (UsesToRename.empty())
1957 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
1959 // We found a use of I outside of BB. Rename all uses of I that are outside
1960 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1961 // with the two values we know.
1962 SSAUpdate.Initialize(I.getType(), I.getName());
1963 SSAUpdate.AddAvailableValue(BB, &I);
1964 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]);
1966 while (!UsesToRename.empty())
1967 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1968 DEBUG(dbgs() << "\n");
1971 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1973 BB->removePredecessor(PredBB, true);
1975 // Remove the unconditional branch at the end of the PredBB block.
1976 OldPredBranch->eraseFromParent();
1982 /// TryToUnfoldSelect - Look for blocks of the form
1988 /// %p = phi [%a, %bb1] ...
1992 /// And expand the select into a branch structure if one of its arms allows %c
1993 /// to be folded. This later enables threading from bb1 over bb2.
1994 bool JumpThreadingPass::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1995 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1996 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1997 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1999 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2000 CondLHS->getParent() != BB)
2003 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2004 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2005 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2007 // Look if one of the incoming values is a select in the corresponding
2009 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2012 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2013 if (!PredTerm || !PredTerm->isUnconditional())
2016 // Now check if one of the select values would allow us to constant fold the
2017 // terminator in BB. We don't do the transform if both sides fold, those
2018 // cases will be threaded in any case.
2019 LazyValueInfo::Tristate LHSFolds =
2020 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2021 CondRHS, Pred, BB, CondCmp);
2022 LazyValueInfo::Tristate RHSFolds =
2023 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2024 CondRHS, Pred, BB, CondCmp);
2025 if ((LHSFolds != LazyValueInfo::Unknown ||
2026 RHSFolds != LazyValueInfo::Unknown) &&
2027 LHSFolds != RHSFolds) {
2028 // Expand the select.
2037 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2038 BB->getParent(), BB);
2039 // Move the unconditional branch to NewBB.
2040 PredTerm->removeFromParent();
2041 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2042 // Create a conditional branch and update PHI nodes.
2043 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2044 CondLHS->setIncomingValue(I, SI->getFalseValue());
2045 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
2046 // The select is now dead.
2047 SI->eraseFromParent();
2049 // Update any other PHI nodes in BB.
2050 for (BasicBlock::iterator BI = BB->begin();
2051 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2053 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2060 /// TryToUnfoldSelectInCurrBB - Look for PHI/Select in the same BB of the form
2062 /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2063 /// %s = select p, trueval, falseval
2065 /// And expand the select into a branch structure. This later enables
2066 /// jump-threading over bb in this pass.
2068 /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2069 /// select if the associated PHI has at least one constant. If the unfolded
2070 /// select is not jump-threaded, it will be folded again in the later
2072 bool JumpThreadingPass::TryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2073 // If threading this would thread across a loop header, don't thread the edge.
2074 // See the comments above FindLoopHeaders for justifications and caveats.
2075 if (LoopHeaders.count(BB))
2078 // Look for a Phi/Select pair in the same basic block. The Phi feeds the
2079 // condition of the Select and at least one of the incoming values is a
2081 for (BasicBlock::iterator BI = BB->begin();
2082 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2083 unsigned NumPHIValues = PN->getNumIncomingValues();
2084 if (NumPHIValues == 0 || !PN->hasOneUse())
2087 SelectInst *SI = dyn_cast<SelectInst>(PN->user_back());
2088 if (!SI || SI->getParent() != BB)
2091 Value *Cond = SI->getCondition();
2092 if (!Cond || Cond != PN || !Cond->getType()->isIntegerTy(1))
2095 bool HasConst = false;
2096 for (unsigned i = 0; i != NumPHIValues; ++i) {
2097 if (PN->getIncomingBlock(i) == BB)
2099 if (isa<ConstantInt>(PN->getIncomingValue(i)))
2104 // Expand the select.
2105 TerminatorInst *Term =
2106 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
2107 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2108 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2109 NewPN->addIncoming(SI->getFalseValue(), BB);
2110 SI->replaceAllUsesWith(NewPN);
2111 SI->eraseFromParent();
2119 /// Try to propagate a guard from the current BB into one of its predecessors
2120 /// in case if another branch of execution implies that the condition of this
2121 /// guard is always true. Currently we only process the simplest case that
2126 /// br i1 %cond, label %T1, label %F1
2132 /// %condGuard = ...
2133 /// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2135 /// And cond either implies condGuard or !condGuard. In this case all the
2136 /// instructions before the guard can be duplicated in both branches, and the
2137 /// guard is then threaded to one of them.
2138 bool JumpThreadingPass::ProcessGuards(BasicBlock *BB) {
2139 using namespace PatternMatch;
2140 // We only want to deal with two predecessors.
2141 BasicBlock *Pred1, *Pred2;
2142 auto PI = pred_begin(BB), PE = pred_end(BB);
2154 // Try to thread one of the guards of the block.
2155 // TODO: Look up deeper than to immediate predecessor?
2156 auto *Parent = Pred1->getSinglePredecessor();
2157 if (!Parent || Parent != Pred2->getSinglePredecessor())
2160 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2162 if (match(&I, m_Intrinsic<Intrinsic::experimental_guard>()))
2163 if (ThreadGuard(BB, cast<IntrinsicInst>(&I), BI))
2169 /// Try to propagate the guard from BB which is the lower block of a diamond
2170 /// to one of its branches, in case if diamond's condition implies guard's
2172 bool JumpThreadingPass::ThreadGuard(BasicBlock *BB, IntrinsicInst *Guard,
2174 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?");
2175 assert(BI->isConditional() && "Unconditional branch has 2 successors?");
2176 Value *GuardCond = Guard->getArgOperand(0);
2177 Value *BranchCond = BI->getCondition();
2178 BasicBlock *TrueDest = BI->getSuccessor(0);
2179 BasicBlock *FalseDest = BI->getSuccessor(1);
2181 auto &DL = BB->getModule()->getDataLayout();
2182 bool TrueDestIsSafe = false;
2183 bool FalseDestIsSafe = false;
2185 // True dest is safe if BranchCond => GuardCond.
2186 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
2188 TrueDestIsSafe = true;
2190 // False dest is safe if !BranchCond => GuardCond.
2192 isImpliedCondition(BranchCond, GuardCond, DL, /* InvertAPred */ true);
2194 FalseDestIsSafe = true;
2197 if (!TrueDestIsSafe && !FalseDestIsSafe)
2200 BasicBlock *UnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
2201 BasicBlock *GuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
2203 ValueToValueMapTy UnguardedMapping, GuardedMapping;
2204 Instruction *AfterGuard = Guard->getNextNode();
2205 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
2206 if (Cost > BBDupThreshold)
2208 // Duplicate all instructions before the guard and the guard itself to the
2209 // branch where implication is not proved.
2210 GuardedBlock = DuplicateInstructionsInSplitBetween(
2211 BB, GuardedBlock, AfterGuard, GuardedMapping);
2212 assert(GuardedBlock && "Could not create the guarded block?");
2213 // Duplicate all instructions before the guard in the unguarded branch.
2214 // Since we have successfully duplicated the guarded block and this block
2215 // has fewer instructions, we expect it to succeed.
2216 UnguardedBlock = DuplicateInstructionsInSplitBetween(BB, UnguardedBlock,
2217 Guard, UnguardedMapping);
2218 assert(UnguardedBlock && "Could not create the unguarded block?");
2219 DEBUG(dbgs() << "Moved guard " << *Guard << " to block "
2220 << GuardedBlock->getName() << "\n");
2222 // Some instructions before the guard may still have uses. For them, we need
2223 // to create Phi nodes merging their copies in both guarded and unguarded
2224 // branches. Those instructions that have no uses can be just removed.
2225 SmallVector<Instruction *, 4> ToRemove;
2226 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
2227 if (!isa<PHINode>(&*BI))
2228 ToRemove.push_back(&*BI);
2230 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
2231 assert(InsertionPoint && "Empty block?");
2232 // Substitute with Phis & remove.
2233 for (auto *Inst : reverse(ToRemove)) {
2234 if (!Inst->use_empty()) {
2235 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
2236 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
2237 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
2238 NewPN->insertBefore(InsertionPoint);
2239 Inst->replaceAllUsesWith(NewPN);
2241 Inst->eraseFromParent();