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 // Replace uses of Cond with ToVal when safe to do so. If all uses are
257 // replaced, we can remove Cond. We cannot blindly replace all uses of Cond
258 // because we may incorrectly replace uses when guards/assumes are uses of
259 // of `Cond` and we used the guards/assume to reason about the `Cond` value
260 // at the end of block. RAUW unconditionally replaces all uses
261 // including the guards/assumes themselves and the uses before the
263 static void ReplaceFoldableUses(Instruction *Cond, Value *ToVal) {
264 assert(Cond->getType() == ToVal->getType());
265 auto *BB = Cond->getParent();
266 // We can unconditionally replace all uses in non-local blocks (i.e. uses
267 // strictly dominated by BB), since LVI information is true from the
269 replaceNonLocalUsesWith(Cond, ToVal);
270 for (Instruction &I : reverse(*BB)) {
271 // Reached the Cond whose uses we are trying to replace, so there are no
275 // We only replace uses in instructions that are guaranteed to reach the end
276 // of BB, where we know Cond is ToVal.
277 if (!isGuaranteedToTransferExecutionToSuccessor(&I))
279 I.replaceUsesOfWith(Cond, ToVal);
281 if (Cond->use_empty() && !Cond->mayHaveSideEffects())
282 Cond->eraseFromParent();
285 /// Return the cost of duplicating a piece of this block from first non-phi
286 /// and before StopAt instruction to thread across it. Stop scanning the block
287 /// when exceeding the threshold. If duplication is impossible, returns ~0U.
288 static unsigned getJumpThreadDuplicationCost(BasicBlock *BB,
290 unsigned Threshold) {
291 assert(StopAt->getParent() == BB && "Not an instruction from proper BB?");
292 /// Ignore PHI nodes, these will be flattened when duplication happens.
293 BasicBlock::const_iterator I(BB->getFirstNonPHI());
295 // FIXME: THREADING will delete values that are just used to compute the
296 // branch, so they shouldn't count against the duplication cost.
299 if (BB->getTerminator() == StopAt) {
300 // Threading through a switch statement is particularly profitable. If this
301 // block ends in a switch, decrease its cost to make it more likely to
303 if (isa<SwitchInst>(StopAt))
306 // The same holds for indirect branches, but slightly more so.
307 if (isa<IndirectBrInst>(StopAt))
311 // Bump the threshold up so the early exit from the loop doesn't skip the
312 // terminator-based Size adjustment at the end.
315 // Sum up the cost of each instruction until we get to the terminator. Don't
316 // include the terminator because the copy won't include it.
318 for (; &*I != StopAt; ++I) {
320 // Stop scanning the block if we've reached the threshold.
321 if (Size > Threshold)
324 // Debugger intrinsics don't incur code size.
325 if (isa<DbgInfoIntrinsic>(I)) continue;
327 // If this is a pointer->pointer bitcast, it is free.
328 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
331 // Bail out if this instruction gives back a token type, it is not possible
332 // to duplicate it if it is used outside this BB.
333 if (I->getType()->isTokenTy() && I->isUsedOutsideOfBlock(BB))
336 // All other instructions count for at least one unit.
339 // Calls are more expensive. If they are non-intrinsic calls, we model them
340 // as having cost of 4. If they are a non-vector intrinsic, we model them
341 // as having cost of 2 total, and if they are a vector intrinsic, we model
342 // them as having cost 1.
343 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
344 if (CI->cannotDuplicate() || CI->isConvergent())
345 // Blocks with NoDuplicate are modelled as having infinite cost, so they
346 // are never duplicated.
348 else if (!isa<IntrinsicInst>(CI))
350 else if (!CI->getType()->isVectorTy())
355 return Size > Bonus ? Size - Bonus : 0;
358 /// FindLoopHeaders - We do not want jump threading to turn proper loop
359 /// structures into irreducible loops. Doing this breaks up the loop nesting
360 /// hierarchy and pessimizes later transformations. To prevent this from
361 /// happening, we first have to find the loop headers. Here we approximate this
362 /// by finding targets of backedges in the CFG.
364 /// Note that there definitely are cases when we want to allow threading of
365 /// edges across a loop header. For example, threading a jump from outside the
366 /// loop (the preheader) to an exit block of the loop is definitely profitable.
367 /// It is also almost always profitable to thread backedges from within the loop
368 /// to exit blocks, and is often profitable to thread backedges to other blocks
369 /// within the loop (forming a nested loop). This simple analysis is not rich
370 /// enough to track all of these properties and keep it up-to-date as the CFG
371 /// mutates, so we don't allow any of these transformations.
373 void JumpThreadingPass::FindLoopHeaders(Function &F) {
374 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
375 FindFunctionBackedges(F, Edges);
377 for (const auto &Edge : Edges)
378 LoopHeaders.insert(Edge.second);
381 /// getKnownConstant - Helper method to determine if we can thread over a
382 /// terminator with the given value as its condition, and if so what value to
383 /// use for that. What kind of value this is depends on whether we want an
384 /// integer or a block address, but an undef is always accepted.
385 /// Returns null if Val is null or not an appropriate constant.
386 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
390 // Undef is "known" enough.
391 if (UndefValue *U = dyn_cast<UndefValue>(Val))
394 if (Preference == WantBlockAddress)
395 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
397 return dyn_cast<ConstantInt>(Val);
400 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
401 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
402 /// in any of our predecessors. If so, return the known list of value and pred
403 /// BB in the result vector.
405 /// This returns true if there were any known values.
407 bool JumpThreadingPass::ComputeValueKnownInPredecessors(
408 Value *V, BasicBlock *BB, PredValueInfo &Result,
409 ConstantPreference Preference, Instruction *CxtI) {
410 // This method walks up use-def chains recursively. Because of this, we could
411 // get into an infinite loop going around loops in the use-def chain. To
412 // prevent this, keep track of what (value, block) pairs we've already visited
413 // and terminate the search if we loop back to them
414 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
417 // An RAII help to remove this pair from the recursion set once the recursion
418 // stack pops back out again.
419 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
421 // If V is a constant, then it is known in all predecessors.
422 if (Constant *KC = getKnownConstant(V, Preference)) {
423 for (BasicBlock *Pred : predecessors(BB))
424 Result.push_back(std::make_pair(KC, Pred));
426 return !Result.empty();
429 // If V is a non-instruction value, or an instruction in a different block,
430 // then it can't be derived from a PHI.
431 Instruction *I = dyn_cast<Instruction>(V);
432 if (!I || I->getParent() != BB) {
434 // Okay, if this is a live-in value, see if it has a known value at the end
435 // of any of our predecessors.
437 // FIXME: This should be an edge property, not a block end property.
438 /// TODO: Per PR2563, we could infer value range information about a
439 /// predecessor based on its terminator.
441 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
442 // "I" is a non-local compare-with-a-constant instruction. This would be
443 // able to handle value inequalities better, for example if the compare is
444 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
445 // Perhaps getConstantOnEdge should be smart enough to do this?
447 for (BasicBlock *P : predecessors(BB)) {
448 // If the value is known by LazyValueInfo to be a constant in a
449 // predecessor, use that information to try to thread this block.
450 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB, CxtI);
451 if (Constant *KC = getKnownConstant(PredCst, Preference))
452 Result.push_back(std::make_pair(KC, P));
455 return !Result.empty();
458 /// If I is a PHI node, then we know the incoming values for any constants.
459 if (PHINode *PN = dyn_cast<PHINode>(I)) {
460 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
461 Value *InVal = PN->getIncomingValue(i);
462 if (Constant *KC = getKnownConstant(InVal, Preference)) {
463 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
465 Constant *CI = LVI->getConstantOnEdge(InVal,
466 PN->getIncomingBlock(i),
468 if (Constant *KC = getKnownConstant(CI, Preference))
469 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
473 return !Result.empty();
476 // Handle Cast instructions. Only see through Cast when the source operand is
477 // PHI or Cmp and the source type is i1 to save the compilation time.
478 if (CastInst *CI = dyn_cast<CastInst>(I)) {
479 Value *Source = CI->getOperand(0);
480 if (!Source->getType()->isIntegerTy(1))
482 if (!isa<PHINode>(Source) && !isa<CmpInst>(Source))
484 ComputeValueKnownInPredecessors(Source, BB, Result, Preference, CxtI);
488 // Convert the known values.
489 for (auto &R : Result)
490 R.first = ConstantExpr::getCast(CI->getOpcode(), R.first, CI->getType());
495 PredValueInfoTy LHSVals, RHSVals;
497 // Handle some boolean conditions.
498 if (I->getType()->getPrimitiveSizeInBits() == 1) {
499 assert(Preference == WantInteger && "One-bit non-integer type?");
501 // X & false -> false
502 if (I->getOpcode() == Instruction::Or ||
503 I->getOpcode() == Instruction::And) {
504 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
506 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
509 if (LHSVals.empty() && RHSVals.empty())
512 ConstantInt *InterestingVal;
513 if (I->getOpcode() == Instruction::Or)
514 InterestingVal = ConstantInt::getTrue(I->getContext());
516 InterestingVal = ConstantInt::getFalse(I->getContext());
518 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
520 // Scan for the sentinel. If we find an undef, force it to the
521 // interesting value: x|undef -> true and x&undef -> false.
522 for (const auto &LHSVal : LHSVals)
523 if (LHSVal.first == InterestingVal || isa<UndefValue>(LHSVal.first)) {
524 Result.emplace_back(InterestingVal, LHSVal.second);
525 LHSKnownBBs.insert(LHSVal.second);
527 for (const auto &RHSVal : RHSVals)
528 if (RHSVal.first == InterestingVal || isa<UndefValue>(RHSVal.first)) {
529 // If we already inferred a value for this block on the LHS, don't
531 if (!LHSKnownBBs.count(RHSVal.second))
532 Result.emplace_back(InterestingVal, RHSVal.second);
535 return !Result.empty();
538 // Handle the NOT form of XOR.
539 if (I->getOpcode() == Instruction::Xor &&
540 isa<ConstantInt>(I->getOperand(1)) &&
541 cast<ConstantInt>(I->getOperand(1))->isOne()) {
542 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
547 // Invert the known values.
548 for (auto &R : Result)
549 R.first = ConstantExpr::getNot(R.first);
554 // Try to simplify some other binary operator values.
555 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
556 assert(Preference != WantBlockAddress
557 && "A binary operator creating a block address?");
558 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
559 PredValueInfoTy LHSVals;
560 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
563 // Try to use constant folding to simplify the binary operator.
564 for (const auto &LHSVal : LHSVals) {
565 Constant *V = LHSVal.first;
566 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
568 if (Constant *KC = getKnownConstant(Folded, WantInteger))
569 Result.push_back(std::make_pair(KC, LHSVal.second));
573 return !Result.empty();
576 // Handle compare with phi operand, where the PHI is defined in this block.
577 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
578 assert(Preference == WantInteger && "Compares only produce integers");
579 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
580 if (PN && PN->getParent() == BB) {
581 const DataLayout &DL = PN->getModule()->getDataLayout();
582 // We can do this simplification if any comparisons fold to true or false.
584 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
585 BasicBlock *PredBB = PN->getIncomingBlock(i);
586 Value *LHS = PN->getIncomingValue(i);
587 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
589 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, {DL});
591 if (!isa<Constant>(RHS))
594 LazyValueInfo::Tristate
595 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
596 cast<Constant>(RHS), PredBB, BB,
598 if (ResT == LazyValueInfo::Unknown)
600 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
603 if (Constant *KC = getKnownConstant(Res, WantInteger))
604 Result.push_back(std::make_pair(KC, PredBB));
607 return !Result.empty();
610 // If comparing a live-in value against a constant, see if we know the
611 // live-in value on any predecessors.
612 if (isa<Constant>(Cmp->getOperand(1)) && !Cmp->getType()->isVectorTy()) {
613 Constant *CmpConst = cast<Constant>(Cmp->getOperand(1));
615 if (!isa<Instruction>(Cmp->getOperand(0)) ||
616 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
617 for (BasicBlock *P : predecessors(BB)) {
618 // If the value is known by LazyValueInfo to be a constant in a
619 // predecessor, use that information to try to thread this block.
620 LazyValueInfo::Tristate Res =
621 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
622 CmpConst, P, BB, CxtI ? CxtI : Cmp);
623 if (Res == LazyValueInfo::Unknown)
626 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
627 Result.push_back(std::make_pair(ResC, P));
630 return !Result.empty();
633 // Try to find a constant value for the LHS of a comparison,
634 // and evaluate it statically if we can.
635 PredValueInfoTy LHSVals;
636 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
639 for (const auto &LHSVal : LHSVals) {
640 Constant *V = LHSVal.first;
641 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
643 if (Constant *KC = getKnownConstant(Folded, WantInteger))
644 Result.push_back(std::make_pair(KC, LHSVal.second));
647 return !Result.empty();
651 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
652 // Handle select instructions where at least one operand is a known constant
653 // and we can figure out the condition value for any predecessor block.
654 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
655 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
656 PredValueInfoTy Conds;
657 if ((TrueVal || FalseVal) &&
658 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
659 WantInteger, CxtI)) {
660 for (auto &C : Conds) {
661 Constant *Cond = C.first;
663 // Figure out what value to use for the condition.
665 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
667 KnownCond = CI->isOne();
669 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
670 // Either operand will do, so be sure to pick the one that's a known
672 // FIXME: Do this more cleverly if both values are known constants?
673 KnownCond = (TrueVal != nullptr);
676 // See if the select has a known constant value for this predecessor.
677 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
678 Result.push_back(std::make_pair(Val, C.second));
681 return !Result.empty();
685 // If all else fails, see if LVI can figure out a constant value for us.
686 Constant *CI = LVI->getConstant(V, BB, CxtI);
687 if (Constant *KC = getKnownConstant(CI, Preference)) {
688 for (BasicBlock *Pred : predecessors(BB))
689 Result.push_back(std::make_pair(KC, Pred));
692 return !Result.empty();
697 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
698 /// in an undefined jump, decide which block is best to revector to.
700 /// Since we can pick an arbitrary destination, we pick the successor with the
701 /// fewest predecessors. This should reduce the in-degree of the others.
703 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
704 TerminatorInst *BBTerm = BB->getTerminator();
705 unsigned MinSucc = 0;
706 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
707 // Compute the successor with the minimum number of predecessors.
708 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
709 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
710 TestBB = BBTerm->getSuccessor(i);
711 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
712 if (NumPreds < MinNumPreds) {
714 MinNumPreds = NumPreds;
721 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
722 if (!BB->hasAddressTaken()) return false;
724 // If the block has its address taken, it may be a tree of dead constants
725 // hanging off of it. These shouldn't keep the block alive.
726 BlockAddress *BA = BlockAddress::get(BB);
727 BA->removeDeadConstantUsers();
728 return !BA->use_empty();
731 /// ProcessBlock - If there are any predecessors whose control can be threaded
732 /// through to a successor, transform them now.
733 bool JumpThreadingPass::ProcessBlock(BasicBlock *BB) {
734 // If the block is trivially dead, just return and let the caller nuke it.
735 // This simplifies other transformations.
736 if (pred_empty(BB) &&
737 BB != &BB->getParent()->getEntryBlock())
740 // If this block has a single predecessor, and if that pred has a single
741 // successor, merge the blocks. This encourages recursive jump threading
742 // because now the condition in this block can be threaded through
743 // predecessors of our predecessor block.
744 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
745 const TerminatorInst *TI = SinglePred->getTerminator();
746 if (!TI->isExceptional() && TI->getNumSuccessors() == 1 &&
747 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
748 // If SinglePred was a loop header, BB becomes one.
749 if (LoopHeaders.erase(SinglePred))
750 LoopHeaders.insert(BB);
752 LVI->eraseBlock(SinglePred);
753 MergeBasicBlockIntoOnlyPred(BB);
759 if (TryToUnfoldSelectInCurrBB(BB))
762 // Look if we can propagate guards to predecessors.
763 if (HasGuards && ProcessGuards(BB))
766 // What kind of constant we're looking for.
767 ConstantPreference Preference = WantInteger;
769 // Look to see if the terminator is a conditional branch, switch or indirect
770 // branch, if not we can't thread it.
772 Instruction *Terminator = BB->getTerminator();
773 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
774 // Can't thread an unconditional jump.
775 if (BI->isUnconditional()) return false;
776 Condition = BI->getCondition();
777 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
778 Condition = SI->getCondition();
779 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
780 // Can't thread indirect branch with no successors.
781 if (IB->getNumSuccessors() == 0) return false;
782 Condition = IB->getAddress()->stripPointerCasts();
783 Preference = WantBlockAddress;
785 return false; // Must be an invoke.
788 // Run constant folding to see if we can reduce the condition to a simple
790 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
792 ConstantFoldInstruction(I, BB->getModule()->getDataLayout(), TLI);
794 I->replaceAllUsesWith(SimpleVal);
795 if (isInstructionTriviallyDead(I, TLI))
796 I->eraseFromParent();
797 Condition = SimpleVal;
801 // If the terminator is branching on an undef, we can pick any of the
802 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
803 if (isa<UndefValue>(Condition)) {
804 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
806 // Fold the branch/switch.
807 TerminatorInst *BBTerm = BB->getTerminator();
808 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
809 if (i == BestSucc) continue;
810 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
813 DEBUG(dbgs() << " In block '" << BB->getName()
814 << "' folding undef terminator: " << *BBTerm << '\n');
815 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
816 BBTerm->eraseFromParent();
820 // If the terminator of this block is branching on a constant, simplify the
821 // terminator to an unconditional branch. This can occur due to threading in
823 if (getKnownConstant(Condition, Preference)) {
824 DEBUG(dbgs() << " In block '" << BB->getName()
825 << "' folding terminator: " << *BB->getTerminator() << '\n');
827 ConstantFoldTerminator(BB, true);
831 Instruction *CondInst = dyn_cast<Instruction>(Condition);
833 // All the rest of our checks depend on the condition being an instruction.
835 // FIXME: Unify this with code below.
836 if (ProcessThreadableEdges(Condition, BB, Preference, Terminator))
841 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
842 // If we're branching on a conditional, LVI might be able to determine
843 // it's value at the branch instruction. We only handle comparisons
844 // against a constant at this time.
845 // TODO: This should be extended to handle switches as well.
846 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
847 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
848 if (CondBr && CondConst) {
849 // We should have returned as soon as we turn a conditional branch to
850 // unconditional. Because its no longer interesting as far as jump
851 // threading is concerned.
852 assert(CondBr->isConditional() && "Threading on unconditional terminator");
854 LazyValueInfo::Tristate Ret =
855 LVI->getPredicateAt(CondCmp->getPredicate(), CondCmp->getOperand(0),
857 if (Ret != LazyValueInfo::Unknown) {
858 unsigned ToRemove = Ret == LazyValueInfo::True ? 1 : 0;
859 unsigned ToKeep = Ret == LazyValueInfo::True ? 0 : 1;
860 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
861 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
862 CondBr->eraseFromParent();
863 if (CondCmp->use_empty())
864 CondCmp->eraseFromParent();
865 // We can safely replace *some* uses of the CondInst if it has
866 // exactly one value as returned by LVI. RAUW is incorrect in the
867 // presence of guards and assumes, that have the `Cond` as the use. This
868 // is because we use the guards/assume to reason about the `Cond` value
869 // at the end of block, but RAUW unconditionally replaces all uses
870 // including the guards/assumes themselves and the uses before the
872 else if (CondCmp->getParent() == BB) {
873 auto *CI = Ret == LazyValueInfo::True ?
874 ConstantInt::getTrue(CondCmp->getType()) :
875 ConstantInt::getFalse(CondCmp->getType());
876 ReplaceFoldableUses(CondCmp, CI);
881 // We did not manage to simplify this branch, try to see whether
882 // CondCmp depends on a known phi-select pattern.
883 if (TryToUnfoldSelect(CondCmp, BB))
888 // Check for some cases that are worth simplifying. Right now we want to look
889 // for loads that are used by a switch or by the condition for the branch. If
890 // we see one, check to see if it's partially redundant. If so, insert a PHI
891 // which can then be used to thread the values.
893 Value *SimplifyValue = CondInst;
894 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
895 if (isa<Constant>(CondCmp->getOperand(1)))
896 SimplifyValue = CondCmp->getOperand(0);
898 // TODO: There are other places where load PRE would be profitable, such as
899 // more complex comparisons.
900 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
901 if (SimplifyPartiallyRedundantLoad(LI))
904 // Handle a variety of cases where we are branching on something derived from
905 // a PHI node in the current block. If we can prove that any predecessors
906 // compute a predictable value based on a PHI node, thread those predecessors.
908 if (ProcessThreadableEdges(CondInst, BB, Preference, Terminator))
911 // If this is an otherwise-unfoldable branch on a phi node in the current
912 // block, see if we can simplify.
913 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
914 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
915 return ProcessBranchOnPHI(PN);
917 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
918 if (CondInst->getOpcode() == Instruction::Xor &&
919 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
920 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
922 // Search for a stronger dominating condition that can be used to simplify a
923 // conditional branch leaving BB.
924 if (ProcessImpliedCondition(BB))
930 bool JumpThreadingPass::ProcessImpliedCondition(BasicBlock *BB) {
931 auto *BI = dyn_cast<BranchInst>(BB->getTerminator());
932 if (!BI || !BI->isConditional())
935 Value *Cond = BI->getCondition();
936 BasicBlock *CurrentBB = BB;
937 BasicBlock *CurrentPred = BB->getSinglePredecessor();
940 auto &DL = BB->getModule()->getDataLayout();
942 while (CurrentPred && Iter++ < ImplicationSearchThreshold) {
943 auto *PBI = dyn_cast<BranchInst>(CurrentPred->getTerminator());
944 if (!PBI || !PBI->isConditional())
946 if (PBI->getSuccessor(0) != CurrentBB && PBI->getSuccessor(1) != CurrentBB)
949 bool FalseDest = PBI->getSuccessor(1) == CurrentBB;
950 Optional<bool> Implication =
951 isImpliedCondition(PBI->getCondition(), Cond, DL, FalseDest);
953 BI->getSuccessor(*Implication ? 1 : 0)->removePredecessor(BB);
954 BranchInst::Create(BI->getSuccessor(*Implication ? 0 : 1), BI);
955 BI->eraseFromParent();
958 CurrentBB = CurrentPred;
959 CurrentPred = CurrentBB->getSinglePredecessor();
965 /// Return true if Op is an instruction defined in the given block.
966 static bool isOpDefinedInBlock(Value *Op, BasicBlock *BB) {
967 if (Instruction *OpInst = dyn_cast<Instruction>(Op))
968 if (OpInst->getParent() == BB)
973 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
974 /// load instruction, eliminate it by replacing it with a PHI node. This is an
975 /// important optimization that encourages jump threading, and needs to be run
976 /// interlaced with other jump threading tasks.
977 bool JumpThreadingPass::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
978 // Don't hack volatile and ordered loads.
979 if (!LI->isUnordered()) return false;
981 // If the load is defined in a block with exactly one predecessor, it can't be
982 // partially redundant.
983 BasicBlock *LoadBB = LI->getParent();
984 if (LoadBB->getSinglePredecessor())
987 // If the load is defined in an EH pad, it can't be partially redundant,
988 // because the edges between the invoke and the EH pad cannot have other
989 // instructions between them.
990 if (LoadBB->isEHPad())
993 Value *LoadedPtr = LI->getOperand(0);
995 // If the loaded operand is defined in the LoadBB and its not a phi,
996 // it can't be available in predecessors.
997 if (isOpDefinedInBlock(LoadedPtr, LoadBB) && !isa<PHINode>(LoadedPtr))
1000 // Scan a few instructions up from the load, to see if it is obviously live at
1001 // the entry to its block.
1002 BasicBlock::iterator BBIt(LI);
1004 if (Value *AvailableVal = FindAvailableLoadedValue(
1005 LI, LoadBB, BBIt, DefMaxInstsToScan, AA, &IsLoadCSE)) {
1006 // If the value of the load is locally available within the block, just use
1007 // it. This frequently occurs for reg2mem'd allocas.
1010 LoadInst *NLI = cast<LoadInst>(AvailableVal);
1011 combineMetadataForCSE(NLI, LI);
1014 // If the returned value is the load itself, replace with an undef. This can
1015 // only happen in dead loops.
1016 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
1017 if (AvailableVal->getType() != LI->getType())
1019 CastInst::CreateBitOrPointerCast(AvailableVal, LI->getType(), "", LI);
1020 LI->replaceAllUsesWith(AvailableVal);
1021 LI->eraseFromParent();
1025 // Otherwise, if we scanned the whole block and got to the top of the block,
1026 // we know the block is locally transparent to the load. If not, something
1027 // might clobber its value.
1028 if (BBIt != LoadBB->begin())
1031 // If all of the loads and stores that feed the value have the same AA tags,
1032 // then we can propagate them onto any newly inserted loads.
1034 LI->getAAMetadata(AATags);
1036 SmallPtrSet<BasicBlock*, 8> PredsScanned;
1037 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
1038 AvailablePredsTy AvailablePreds;
1039 BasicBlock *OneUnavailablePred = nullptr;
1040 SmallVector<LoadInst*, 8> CSELoads;
1042 // If we got here, the loaded value is transparent through to the start of the
1043 // block. Check to see if it is available in any of the predecessor blocks.
1044 for (BasicBlock *PredBB : predecessors(LoadBB)) {
1045 // If we already scanned this predecessor, skip it.
1046 if (!PredsScanned.insert(PredBB).second)
1049 BBIt = PredBB->end();
1050 unsigned NumScanedInst = 0;
1051 Value *PredAvailable = nullptr;
1052 // NOTE: We don't CSE load that is volatile or anything stronger than
1053 // unordered, that should have been checked when we entered the function.
1054 assert(LI->isUnordered() && "Attempting to CSE volatile or atomic loads");
1055 // If this is a load on a phi pointer, phi-translate it and search
1056 // for available load/store to the pointer in predecessors.
1057 Value *Ptr = LoadedPtr->DoPHITranslation(LoadBB, PredBB);
1058 PredAvailable = FindAvailablePtrLoadStore(
1059 Ptr, LI->getType(), LI->isAtomic(), PredBB, BBIt, DefMaxInstsToScan,
1060 AA, &IsLoadCSE, &NumScanedInst);
1062 // If PredBB has a single predecessor, continue scanning through the
1063 // single precessor.
1064 BasicBlock *SinglePredBB = PredBB;
1065 while (!PredAvailable && SinglePredBB && BBIt == SinglePredBB->begin() &&
1066 NumScanedInst < DefMaxInstsToScan) {
1067 SinglePredBB = SinglePredBB->getSinglePredecessor();
1069 BBIt = SinglePredBB->end();
1070 PredAvailable = FindAvailablePtrLoadStore(
1071 Ptr, LI->getType(), LI->isAtomic(), SinglePredBB, BBIt,
1072 (DefMaxInstsToScan - NumScanedInst), AA, &IsLoadCSE,
1077 if (!PredAvailable) {
1078 OneUnavailablePred = PredBB;
1083 CSELoads.push_back(cast<LoadInst>(PredAvailable));
1085 // If so, this load is partially redundant. Remember this info so that we
1086 // can create a PHI node.
1087 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
1090 // If the loaded value isn't available in any predecessor, it isn't partially
1092 if (AvailablePreds.empty()) return false;
1094 // Okay, the loaded value is available in at least one (and maybe all!)
1095 // predecessors. If the value is unavailable in more than one unique
1096 // predecessor, we want to insert a merge block for those common predecessors.
1097 // This ensures that we only have to insert one reload, thus not increasing
1099 BasicBlock *UnavailablePred = nullptr;
1101 // If there is exactly one predecessor where the value is unavailable, the
1102 // already computed 'OneUnavailablePred' block is it. If it ends in an
1103 // unconditional branch, we know that it isn't a critical edge.
1104 if (PredsScanned.size() == AvailablePreds.size()+1 &&
1105 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
1106 UnavailablePred = OneUnavailablePred;
1107 } else if (PredsScanned.size() != AvailablePreds.size()) {
1108 // Otherwise, we had multiple unavailable predecessors or we had a critical
1109 // edge from the one.
1110 SmallVector<BasicBlock*, 8> PredsToSplit;
1111 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
1113 for (const auto &AvailablePred : AvailablePreds)
1114 AvailablePredSet.insert(AvailablePred.first);
1116 // Add all the unavailable predecessors to the PredsToSplit list.
1117 for (BasicBlock *P : predecessors(LoadBB)) {
1118 // If the predecessor is an indirect goto, we can't split the edge.
1119 if (isa<IndirectBrInst>(P->getTerminator()))
1122 if (!AvailablePredSet.count(P))
1123 PredsToSplit.push_back(P);
1126 // Split them out to their own block.
1127 UnavailablePred = SplitBlockPreds(LoadBB, PredsToSplit, "thread-pre-split");
1130 // If the value isn't available in all predecessors, then there will be
1131 // exactly one where it isn't available. Insert a load on that edge and add
1132 // it to the AvailablePreds list.
1133 if (UnavailablePred) {
1134 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
1135 "Can't handle critical edge here!");
1136 LoadInst *NewVal = new LoadInst(
1137 LoadedPtr->DoPHITranslation(LoadBB, UnavailablePred),
1138 LI->getName() + ".pr", false, LI->getAlignment(), LI->getOrdering(),
1139 LI->getSynchScope(), UnavailablePred->getTerminator());
1140 NewVal->setDebugLoc(LI->getDebugLoc());
1142 NewVal->setAAMetadata(AATags);
1144 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
1147 // Now we know that each predecessor of this block has a value in
1148 // AvailablePreds, sort them for efficient access as we're walking the preds.
1149 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
1151 // Create a PHI node at the start of the block for the PRE'd load value.
1152 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
1153 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
1156 PN->setDebugLoc(LI->getDebugLoc());
1158 // Insert new entries into the PHI for each predecessor. A single block may
1159 // have multiple entries here.
1160 for (pred_iterator PI = PB; PI != PE; ++PI) {
1161 BasicBlock *P = *PI;
1162 AvailablePredsTy::iterator I =
1163 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1164 std::make_pair(P, (Value*)nullptr));
1166 assert(I != AvailablePreds.end() && I->first == P &&
1167 "Didn't find entry for predecessor!");
1169 // If we have an available predecessor but it requires casting, insert the
1170 // cast in the predecessor and use the cast. Note that we have to update the
1171 // AvailablePreds vector as we go so that all of the PHI entries for this
1172 // predecessor use the same bitcast.
1173 Value *&PredV = I->second;
1174 if (PredV->getType() != LI->getType())
1175 PredV = CastInst::CreateBitOrPointerCast(PredV, LI->getType(), "",
1176 P->getTerminator());
1178 PN->addIncoming(PredV, I->first);
1181 for (LoadInst *PredLI : CSELoads) {
1182 combineMetadataForCSE(PredLI, LI);
1185 LI->replaceAllUsesWith(PN);
1186 LI->eraseFromParent();
1191 /// FindMostPopularDest - The specified list contains multiple possible
1192 /// threadable destinations. Pick the one that occurs the most frequently in
1195 FindMostPopularDest(BasicBlock *BB,
1196 const SmallVectorImpl<std::pair<BasicBlock*,
1197 BasicBlock*> > &PredToDestList) {
1198 assert(!PredToDestList.empty());
1200 // Determine popularity. If there are multiple possible destinations, we
1201 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1202 // blocks with known and real destinations to threading undef. We'll handle
1203 // them later if interesting.
1204 DenseMap<BasicBlock*, unsigned> DestPopularity;
1205 for (const auto &PredToDest : PredToDestList)
1206 if (PredToDest.second)
1207 DestPopularity[PredToDest.second]++;
1209 // Find the most popular dest.
1210 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1211 BasicBlock *MostPopularDest = DPI->first;
1212 unsigned Popularity = DPI->second;
1213 SmallVector<BasicBlock*, 4> SamePopularity;
1215 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1216 // If the popularity of this entry isn't higher than the popularity we've
1217 // seen so far, ignore it.
1218 if (DPI->second < Popularity)
1220 else if (DPI->second == Popularity) {
1221 // If it is the same as what we've seen so far, keep track of it.
1222 SamePopularity.push_back(DPI->first);
1224 // If it is more popular, remember it.
1225 SamePopularity.clear();
1226 MostPopularDest = DPI->first;
1227 Popularity = DPI->second;
1231 // Okay, now we know the most popular destination. If there is more than one
1232 // destination, we need to determine one. This is arbitrary, but we need
1233 // to make a deterministic decision. Pick the first one that appears in the
1235 if (!SamePopularity.empty()) {
1236 SamePopularity.push_back(MostPopularDest);
1237 TerminatorInst *TI = BB->getTerminator();
1238 for (unsigned i = 0; ; ++i) {
1239 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1241 if (!is_contained(SamePopularity, TI->getSuccessor(i)))
1244 MostPopularDest = TI->getSuccessor(i);
1249 // Okay, we have finally picked the most popular destination.
1250 return MostPopularDest;
1253 bool JumpThreadingPass::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1254 ConstantPreference Preference,
1255 Instruction *CxtI) {
1256 // If threading this would thread across a loop header, don't even try to
1258 if (LoopHeaders.count(BB))
1261 PredValueInfoTy PredValues;
1262 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference, CxtI))
1265 assert(!PredValues.empty() &&
1266 "ComputeValueKnownInPredecessors returned true with no values");
1268 DEBUG(dbgs() << "IN BB: " << *BB;
1269 for (const auto &PredValue : PredValues) {
1270 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1272 << " for pred '" << PredValue.second->getName() << "'.\n";
1275 // Decide what we want to thread through. Convert our list of known values to
1276 // a list of known destinations for each pred. This also discards duplicate
1277 // predecessors and keeps track of the undefined inputs (which are represented
1278 // as a null dest in the PredToDestList).
1279 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1280 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1282 BasicBlock *OnlyDest = nullptr;
1283 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1284 Constant *OnlyVal = nullptr;
1285 Constant *MultipleVal = (Constant *)(intptr_t)~0ULL;
1287 unsigned PredWithKnownDest = 0;
1288 for (const auto &PredValue : PredValues) {
1289 BasicBlock *Pred = PredValue.second;
1290 if (!SeenPreds.insert(Pred).second)
1291 continue; // Duplicate predecessor entry.
1293 Constant *Val = PredValue.first;
1296 if (isa<UndefValue>(Val))
1298 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
1299 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1300 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1301 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1302 assert(isa<ConstantInt>(Val) && "Expecting a constant integer");
1303 DestBB = SI->findCaseValue(cast<ConstantInt>(Val))->getCaseSuccessor();
1305 assert(isa<IndirectBrInst>(BB->getTerminator())
1306 && "Unexpected terminator");
1307 assert(isa<BlockAddress>(Val) && "Expecting a constant blockaddress");
1308 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1311 // If we have exactly one destination, remember it for efficiency below.
1312 if (PredToDestList.empty()) {
1316 if (OnlyDest != DestBB)
1317 OnlyDest = MultipleDestSentinel;
1318 // It possible we have same destination, but different value, e.g. default
1319 // case in switchinst.
1321 OnlyVal = MultipleVal;
1324 // We know where this predecessor is going.
1325 ++PredWithKnownDest;
1327 // If the predecessor ends with an indirect goto, we can't change its
1329 if (isa<IndirectBrInst>(Pred->getTerminator()))
1332 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1335 // If all edges were unthreadable, we fail.
1336 if (PredToDestList.empty())
1339 // If all the predecessors go to a single known successor, we want to fold,
1340 // not thread. By doing so, we do not need to duplicate the current block and
1341 // also miss potential opportunities in case we dont/cant duplicate.
1342 if (OnlyDest && OnlyDest != MultipleDestSentinel) {
1343 if (PredWithKnownDest ==
1344 (size_t)std::distance(pred_begin(BB), pred_end(BB))) {
1345 bool SeenFirstBranchToOnlyDest = false;
1346 for (BasicBlock *SuccBB : successors(BB)) {
1347 if (SuccBB == OnlyDest && !SeenFirstBranchToOnlyDest)
1348 SeenFirstBranchToOnlyDest = true; // Don't modify the first branch.
1350 SuccBB->removePredecessor(BB, true); // This is unreachable successor.
1353 // Finally update the terminator.
1354 TerminatorInst *Term = BB->getTerminator();
1355 BranchInst::Create(OnlyDest, Term);
1356 Term->eraseFromParent();
1358 // If the condition is now dead due to the removal of the old terminator,
1360 if (auto *CondInst = dyn_cast<Instruction>(Cond)) {
1361 if (CondInst->use_empty() && !CondInst->mayHaveSideEffects())
1362 CondInst->eraseFromParent();
1363 // We can safely replace *some* uses of the CondInst if it has
1364 // exactly one value as returned by LVI. RAUW is incorrect in the
1365 // presence of guards and assumes, that have the `Cond` as the use. This
1366 // is because we use the guards/assume to reason about the `Cond` value
1367 // at the end of block, but RAUW unconditionally replaces all uses
1368 // including the guards/assumes themselves and the uses before the
1370 else if (OnlyVal && OnlyVal != MultipleVal &&
1371 CondInst->getParent() == BB)
1372 ReplaceFoldableUses(CondInst, OnlyVal);
1378 // Determine which is the most common successor. If we have many inputs and
1379 // this block is a switch, we want to start by threading the batch that goes
1380 // to the most popular destination first. If we only know about one
1381 // threadable destination (the common case) we can avoid this.
1382 BasicBlock *MostPopularDest = OnlyDest;
1384 if (MostPopularDest == MultipleDestSentinel)
1385 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1387 // Now that we know what the most popular destination is, factor all
1388 // predecessors that will jump to it into a single predecessor.
1389 SmallVector<BasicBlock*, 16> PredsToFactor;
1390 for (const auto &PredToDest : PredToDestList)
1391 if (PredToDest.second == MostPopularDest) {
1392 BasicBlock *Pred = PredToDest.first;
1394 // This predecessor may be a switch or something else that has multiple
1395 // edges to the block. Factor each of these edges by listing them
1396 // according to # occurrences in PredsToFactor.
1397 for (BasicBlock *Succ : successors(Pred))
1399 PredsToFactor.push_back(Pred);
1402 // If the threadable edges are branching on an undefined value, we get to pick
1403 // the destination that these predecessors should get to.
1404 if (!MostPopularDest)
1405 MostPopularDest = BB->getTerminator()->
1406 getSuccessor(GetBestDestForJumpOnUndef(BB));
1408 // Ok, try to thread it!
1409 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1412 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1413 /// a PHI node in the current block. See if there are any simplifications we
1414 /// can do based on inputs to the phi node.
1416 bool JumpThreadingPass::ProcessBranchOnPHI(PHINode *PN) {
1417 BasicBlock *BB = PN->getParent();
1419 // TODO: We could make use of this to do it once for blocks with common PHI
1421 SmallVector<BasicBlock*, 1> PredBBs;
1424 // If any of the predecessor blocks end in an unconditional branch, we can
1425 // *duplicate* the conditional branch into that block in order to further
1426 // encourage jump threading and to eliminate cases where we have branch on a
1427 // phi of an icmp (branch on icmp is much better).
1428 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1429 BasicBlock *PredBB = PN->getIncomingBlock(i);
1430 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1431 if (PredBr->isUnconditional()) {
1432 PredBBs[0] = PredBB;
1433 // Try to duplicate BB into PredBB.
1434 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1442 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1443 /// a xor instruction in the current block. See if there are any
1444 /// simplifications we can do based on inputs to the xor.
1446 bool JumpThreadingPass::ProcessBranchOnXOR(BinaryOperator *BO) {
1447 BasicBlock *BB = BO->getParent();
1449 // If either the LHS or RHS of the xor is a constant, don't do this
1451 if (isa<ConstantInt>(BO->getOperand(0)) ||
1452 isa<ConstantInt>(BO->getOperand(1)))
1455 // If the first instruction in BB isn't a phi, we won't be able to infer
1456 // anything special about any particular predecessor.
1457 if (!isa<PHINode>(BB->front()))
1460 // If this BB is a landing pad, we won't be able to split the edge into it.
1464 // If we have a xor as the branch input to this block, and we know that the
1465 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1466 // the condition into the predecessor and fix that value to true, saving some
1467 // logical ops on that path and encouraging other paths to simplify.
1469 // This copies something like this:
1472 // %X = phi i1 [1], [%X']
1473 // %Y = icmp eq i32 %A, %B
1474 // %Z = xor i1 %X, %Y
1479 // %Y = icmp ne i32 %A, %B
1482 PredValueInfoTy XorOpValues;
1484 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1486 assert(XorOpValues.empty());
1487 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1493 assert(!XorOpValues.empty() &&
1494 "ComputeValueKnownInPredecessors returned true with no values");
1496 // Scan the information to see which is most popular: true or false. The
1497 // predecessors can be of the set true, false, or undef.
1498 unsigned NumTrue = 0, NumFalse = 0;
1499 for (const auto &XorOpValue : XorOpValues) {
1500 if (isa<UndefValue>(XorOpValue.first))
1501 // Ignore undefs for the count.
1503 if (cast<ConstantInt>(XorOpValue.first)->isZero())
1509 // Determine which value to split on, true, false, or undef if neither.
1510 ConstantInt *SplitVal = nullptr;
1511 if (NumTrue > NumFalse)
1512 SplitVal = ConstantInt::getTrue(BB->getContext());
1513 else if (NumTrue != 0 || NumFalse != 0)
1514 SplitVal = ConstantInt::getFalse(BB->getContext());
1516 // Collect all of the blocks that this can be folded into so that we can
1517 // factor this once and clone it once.
1518 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1519 for (const auto &XorOpValue : XorOpValues) {
1520 if (XorOpValue.first != SplitVal && !isa<UndefValue>(XorOpValue.first))
1523 BlocksToFoldInto.push_back(XorOpValue.second);
1526 // If we inferred a value for all of the predecessors, then duplication won't
1527 // help us. However, we can just replace the LHS or RHS with the constant.
1528 if (BlocksToFoldInto.size() ==
1529 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1531 // If all preds provide undef, just nuke the xor, because it is undef too.
1532 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1533 BO->eraseFromParent();
1534 } else if (SplitVal->isZero()) {
1535 // If all preds provide 0, replace the xor with the other input.
1536 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1537 BO->eraseFromParent();
1539 // If all preds provide 1, set the computed value to 1.
1540 BO->setOperand(!isLHS, SplitVal);
1546 // Try to duplicate BB into PredBB.
1547 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1551 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1552 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1553 /// NewPred using the entries from OldPred (suitably mapped).
1554 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1555 BasicBlock *OldPred,
1556 BasicBlock *NewPred,
1557 DenseMap<Instruction*, Value*> &ValueMap) {
1558 for (BasicBlock::iterator PNI = PHIBB->begin();
1559 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1560 // Ok, we have a PHI node. Figure out what the incoming value was for the
1562 Value *IV = PN->getIncomingValueForBlock(OldPred);
1564 // Remap the value if necessary.
1565 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1566 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1567 if (I != ValueMap.end())
1571 PN->addIncoming(IV, NewPred);
1575 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1576 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1577 /// across BB. Transform the IR to reflect this change.
1578 bool JumpThreadingPass::ThreadEdge(BasicBlock *BB,
1579 const SmallVectorImpl<BasicBlock *> &PredBBs,
1580 BasicBlock *SuccBB) {
1581 // If threading to the same block as we come from, we would infinite loop.
1583 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1584 << "' - would thread to self!\n");
1588 // If threading this would thread across a loop header, don't thread the edge.
1589 // See the comments above FindLoopHeaders for justifications and caveats.
1590 if (LoopHeaders.count(BB)) {
1591 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1592 << "' to dest BB '" << SuccBB->getName()
1593 << "' - it might create an irreducible loop!\n");
1597 unsigned JumpThreadCost =
1598 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1599 if (JumpThreadCost > BBDupThreshold) {
1600 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1601 << "' - Cost is too high: " << JumpThreadCost << "\n");
1605 // And finally, do it! Start by factoring the predecessors if needed.
1607 if (PredBBs.size() == 1)
1608 PredBB = PredBBs[0];
1610 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1611 << " common predecessors.\n");
1612 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1615 // And finally, do it!
1616 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1617 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1618 << ", across block:\n "
1621 LVI->threadEdge(PredBB, BB, SuccBB);
1623 // We are going to have to map operands from the original BB block to the new
1624 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1625 // account for entry from PredBB.
1626 DenseMap<Instruction*, Value*> ValueMapping;
1628 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1629 BB->getName()+".thread",
1630 BB->getParent(), BB);
1631 NewBB->moveAfter(PredBB);
1633 // Set the block frequency of NewBB.
1634 if (HasProfileData) {
1636 BFI->getBlockFreq(PredBB) * BPI->getEdgeProbability(PredBB, BB);
1637 BFI->setBlockFreq(NewBB, NewBBFreq.getFrequency());
1640 BasicBlock::iterator BI = BB->begin();
1641 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1642 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1644 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1645 // mapping and using it to remap operands in the cloned instructions.
1646 for (; !isa<TerminatorInst>(BI); ++BI) {
1647 Instruction *New = BI->clone();
1648 New->setName(BI->getName());
1649 NewBB->getInstList().push_back(New);
1650 ValueMapping[&*BI] = New;
1652 // Remap operands to patch up intra-block references.
1653 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1654 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1655 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1656 if (I != ValueMapping.end())
1657 New->setOperand(i, I->second);
1661 // We didn't copy the terminator from BB over to NewBB, because there is now
1662 // an unconditional jump to SuccBB. Insert the unconditional jump.
1663 BranchInst *NewBI = BranchInst::Create(SuccBB, NewBB);
1664 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1666 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1667 // PHI nodes for NewBB now.
1668 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1670 // If there were values defined in BB that are used outside the block, then we
1671 // now have to update all uses of the value to use either the original value,
1672 // the cloned value, or some PHI derived value. This can require arbitrary
1673 // PHI insertion, of which we are prepared to do, clean these up now.
1674 SSAUpdater SSAUpdate;
1675 SmallVector<Use*, 16> UsesToRename;
1676 for (Instruction &I : *BB) {
1677 // Scan all uses of this instruction to see if it is used outside of its
1678 // block, and if so, record them in UsesToRename.
1679 for (Use &U : I.uses()) {
1680 Instruction *User = cast<Instruction>(U.getUser());
1681 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1682 if (UserPN->getIncomingBlock(U) == BB)
1684 } else if (User->getParent() == BB)
1687 UsesToRename.push_back(&U);
1690 // If there are no uses outside the block, we're done with this instruction.
1691 if (UsesToRename.empty())
1694 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
1696 // We found a use of I outside of BB. Rename all uses of I that are outside
1697 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1698 // with the two values we know.
1699 SSAUpdate.Initialize(I.getType(), I.getName());
1700 SSAUpdate.AddAvailableValue(BB, &I);
1701 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[&I]);
1703 while (!UsesToRename.empty())
1704 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1705 DEBUG(dbgs() << "\n");
1709 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1710 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1711 // us to simplify any PHI nodes in BB.
1712 TerminatorInst *PredTerm = PredBB->getTerminator();
1713 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1714 if (PredTerm->getSuccessor(i) == BB) {
1715 BB->removePredecessor(PredBB, true);
1716 PredTerm->setSuccessor(i, NewBB);
1719 // At this point, the IR is fully up to date and consistent. Do a quick scan
1720 // over the new instructions and zap any that are constants or dead. This
1721 // frequently happens because of phi translation.
1722 SimplifyInstructionsInBlock(NewBB, TLI);
1724 // Update the edge weight from BB to SuccBB, which should be less than before.
1725 UpdateBlockFreqAndEdgeWeight(PredBB, BB, NewBB, SuccBB);
1727 // Threaded an edge!
1732 /// Create a new basic block that will be the predecessor of BB and successor of
1733 /// all blocks in Preds. When profile data is available, update the frequency of
1735 BasicBlock *JumpThreadingPass::SplitBlockPreds(BasicBlock *BB,
1736 ArrayRef<BasicBlock *> Preds,
1737 const char *Suffix) {
1738 // Collect the frequencies of all predecessors of BB, which will be used to
1739 // update the edge weight on BB->SuccBB.
1740 BlockFrequency PredBBFreq(0);
1742 for (auto Pred : Preds)
1743 PredBBFreq += BFI->getBlockFreq(Pred) * BPI->getEdgeProbability(Pred, BB);
1745 BasicBlock *PredBB = SplitBlockPredecessors(BB, Preds, Suffix);
1747 // Set the block frequency of the newly created PredBB, which is the sum of
1748 // frequencies of Preds.
1750 BFI->setBlockFreq(PredBB, PredBBFreq.getFrequency());
1754 bool JumpThreadingPass::doesBlockHaveProfileData(BasicBlock *BB) {
1755 const TerminatorInst *TI = BB->getTerminator();
1756 assert(TI->getNumSuccessors() > 1 && "not a split");
1758 MDNode *WeightsNode = TI->getMetadata(LLVMContext::MD_prof);
1762 MDString *MDName = cast<MDString>(WeightsNode->getOperand(0));
1763 if (MDName->getString() != "branch_weights")
1766 // Ensure there are weights for all of the successors. Note that the first
1767 // operand to the metadata node is a name, not a weight.
1768 return WeightsNode->getNumOperands() == TI->getNumSuccessors() + 1;
1771 /// Update the block frequency of BB and branch weight and the metadata on the
1772 /// edge BB->SuccBB. This is done by scaling the weight of BB->SuccBB by 1 -
1773 /// Freq(PredBB->BB) / Freq(BB->SuccBB).
1774 void JumpThreadingPass::UpdateBlockFreqAndEdgeWeight(BasicBlock *PredBB,
1777 BasicBlock *SuccBB) {
1778 if (!HasProfileData)
1781 assert(BFI && BPI && "BFI & BPI should have been created here");
1783 // As the edge from PredBB to BB is deleted, we have to update the block
1785 auto BBOrigFreq = BFI->getBlockFreq(BB);
1786 auto NewBBFreq = BFI->getBlockFreq(NewBB);
1787 auto BB2SuccBBFreq = BBOrigFreq * BPI->getEdgeProbability(BB, SuccBB);
1788 auto BBNewFreq = BBOrigFreq - NewBBFreq;
1789 BFI->setBlockFreq(BB, BBNewFreq.getFrequency());
1791 // Collect updated outgoing edges' frequencies from BB and use them to update
1792 // edge probabilities.
1793 SmallVector<uint64_t, 4> BBSuccFreq;
1794 for (BasicBlock *Succ : successors(BB)) {
1795 auto SuccFreq = (Succ == SuccBB)
1796 ? BB2SuccBBFreq - NewBBFreq
1797 : BBOrigFreq * BPI->getEdgeProbability(BB, Succ);
1798 BBSuccFreq.push_back(SuccFreq.getFrequency());
1801 uint64_t MaxBBSuccFreq =
1802 *std::max_element(BBSuccFreq.begin(), BBSuccFreq.end());
1804 SmallVector<BranchProbability, 4> BBSuccProbs;
1805 if (MaxBBSuccFreq == 0)
1806 BBSuccProbs.assign(BBSuccFreq.size(),
1807 {1, static_cast<unsigned>(BBSuccFreq.size())});
1809 for (uint64_t Freq : BBSuccFreq)
1810 BBSuccProbs.push_back(
1811 BranchProbability::getBranchProbability(Freq, MaxBBSuccFreq));
1812 // Normalize edge probabilities so that they sum up to one.
1813 BranchProbability::normalizeProbabilities(BBSuccProbs.begin(),
1817 // Update edge probabilities in BPI.
1818 for (int I = 0, E = BBSuccProbs.size(); I < E; I++)
1819 BPI->setEdgeProbability(BB, I, BBSuccProbs[I]);
1821 // Update the profile metadata as well.
1823 // Don't do this if the profile of the transformed blocks was statically
1824 // estimated. (This could occur despite the function having an entry
1825 // frequency in completely cold parts of the CFG.)
1827 // In this case we don't want to suggest to subsequent passes that the
1828 // calculated weights are fully consistent. Consider this graph:
1843 // Assuming the blocks check_* all compare the same value against 1, 2 and 3,
1844 // the overall probabilities are inconsistent; the total probability that the
1845 // value is either 1, 2 or 3 is 150%.
1847 // As a consequence if we thread eq_1 -> check_2 to check_3, check_2->check_3
1848 // becomes 0%. This is even worse if the edge whose probability becomes 0% is
1849 // the loop exit edge. Then based solely on static estimation we would assume
1850 // the loop was extremely hot.
1852 // FIXME this locally as well so that BPI and BFI are consistent as well. We
1853 // shouldn't make edges extremely likely or unlikely based solely on static
1855 if (BBSuccProbs.size() >= 2 && doesBlockHaveProfileData(BB)) {
1856 SmallVector<uint32_t, 4> Weights;
1857 for (auto Prob : BBSuccProbs)
1858 Weights.push_back(Prob.getNumerator());
1860 auto TI = BB->getTerminator();
1862 LLVMContext::MD_prof,
1863 MDBuilder(TI->getParent()->getContext()).createBranchWeights(Weights));
1867 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1868 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1869 /// If we can duplicate the contents of BB up into PredBB do so now, this
1870 /// improves the odds that the branch will be on an analyzable instruction like
1872 bool JumpThreadingPass::DuplicateCondBranchOnPHIIntoPred(
1873 BasicBlock *BB, const SmallVectorImpl<BasicBlock *> &PredBBs) {
1874 assert(!PredBBs.empty() && "Can't handle an empty set");
1876 // If BB is a loop header, then duplicating this block outside the loop would
1877 // cause us to transform this into an irreducible loop, don't do this.
1878 // See the comments above FindLoopHeaders for justifications and caveats.
1879 if (LoopHeaders.count(BB)) {
1880 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1881 << "' into predecessor block '" << PredBBs[0]->getName()
1882 << "' - it might create an irreducible loop!\n");
1886 unsigned DuplicationCost =
1887 getJumpThreadDuplicationCost(BB, BB->getTerminator(), BBDupThreshold);
1888 if (DuplicationCost > BBDupThreshold) {
1889 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1890 << "' - Cost is too high: " << DuplicationCost << "\n");
1894 // And finally, do it! Start by factoring the predecessors if needed.
1896 if (PredBBs.size() == 1)
1897 PredBB = PredBBs[0];
1899 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1900 << " common predecessors.\n");
1901 PredBB = SplitBlockPreds(BB, PredBBs, ".thr_comm");
1904 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1906 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1907 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1908 << DuplicationCost << " block is:" << *BB << "\n");
1910 // Unless PredBB ends with an unconditional branch, split the edge so that we
1911 // can just clone the bits from BB into the end of the new PredBB.
1912 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1914 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1915 PredBB = SplitEdge(PredBB, BB);
1916 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1919 // We are going to have to map operands from the original BB block into the
1920 // PredBB block. Evaluate PHI nodes in BB.
1921 DenseMap<Instruction*, Value*> ValueMapping;
1923 BasicBlock::iterator BI = BB->begin();
1924 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1925 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1926 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1927 // mapping and using it to remap operands in the cloned instructions.
1928 for (; BI != BB->end(); ++BI) {
1929 Instruction *New = BI->clone();
1931 // Remap operands to patch up intra-block references.
1932 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1933 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1934 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1935 if (I != ValueMapping.end())
1936 New->setOperand(i, I->second);
1939 // If this instruction can be simplified after the operands are updated,
1940 // just use the simplified value instead. This frequently happens due to
1942 if (Value *IV = SimplifyInstruction(
1944 {BB->getModule()->getDataLayout(), TLI, nullptr, nullptr, New})) {
1945 ValueMapping[&*BI] = IV;
1946 if (!New->mayHaveSideEffects()) {
1951 ValueMapping[&*BI] = New;
1954 // Otherwise, insert the new instruction into the block.
1955 New->setName(BI->getName());
1956 PredBB->getInstList().insert(OldPredBranch->getIterator(), New);
1960 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1961 // add entries to the PHI nodes for branch from PredBB now.
1962 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1963 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1965 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1968 // If there were values defined in BB that are used outside the block, then we
1969 // now have to update all uses of the value to use either the original value,
1970 // the cloned value, or some PHI derived value. This can require arbitrary
1971 // PHI insertion, of which we are prepared to do, clean these up now.
1972 SSAUpdater SSAUpdate;
1973 SmallVector<Use*, 16> UsesToRename;
1974 for (Instruction &I : *BB) {
1975 // Scan all uses of this instruction to see if it is used outside of its
1976 // block, and if so, record them in UsesToRename.
1977 for (Use &U : I.uses()) {
1978 Instruction *User = cast<Instruction>(U.getUser());
1979 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1980 if (UserPN->getIncomingBlock(U) == BB)
1982 } else if (User->getParent() == BB)
1985 UsesToRename.push_back(&U);
1988 // If there are no uses outside the block, we're done with this instruction.
1989 if (UsesToRename.empty())
1992 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << I << "\n");
1994 // We found a use of I outside of BB. Rename all uses of I that are outside
1995 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1996 // with the two values we know.
1997 SSAUpdate.Initialize(I.getType(), I.getName());
1998 SSAUpdate.AddAvailableValue(BB, &I);
1999 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[&I]);
2001 while (!UsesToRename.empty())
2002 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
2003 DEBUG(dbgs() << "\n");
2006 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
2008 BB->removePredecessor(PredBB, true);
2010 // Remove the unconditional branch at the end of the PredBB block.
2011 OldPredBranch->eraseFromParent();
2017 /// TryToUnfoldSelect - Look for blocks of the form
2023 /// %p = phi [%a, %bb1] ...
2027 /// And expand the select into a branch structure if one of its arms allows %c
2028 /// to be folded. This later enables threading from bb1 over bb2.
2029 bool JumpThreadingPass::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
2030 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
2031 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
2032 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
2034 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
2035 CondLHS->getParent() != BB)
2038 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
2039 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
2040 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
2042 // Look if one of the incoming values is a select in the corresponding
2044 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
2047 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
2048 if (!PredTerm || !PredTerm->isUnconditional())
2051 // Now check if one of the select values would allow us to constant fold the
2052 // terminator in BB. We don't do the transform if both sides fold, those
2053 // cases will be threaded in any case.
2054 LazyValueInfo::Tristate LHSFolds =
2055 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
2056 CondRHS, Pred, BB, CondCmp);
2057 LazyValueInfo::Tristate RHSFolds =
2058 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
2059 CondRHS, Pred, BB, CondCmp);
2060 if ((LHSFolds != LazyValueInfo::Unknown ||
2061 RHSFolds != LazyValueInfo::Unknown) &&
2062 LHSFolds != RHSFolds) {
2063 // Expand the select.
2072 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
2073 BB->getParent(), BB);
2074 // Move the unconditional branch to NewBB.
2075 PredTerm->removeFromParent();
2076 NewBB->getInstList().insert(NewBB->end(), PredTerm);
2077 // Create a conditional branch and update PHI nodes.
2078 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
2079 CondLHS->setIncomingValue(I, SI->getFalseValue());
2080 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
2081 // The select is now dead.
2082 SI->eraseFromParent();
2084 // Update any other PHI nodes in BB.
2085 for (BasicBlock::iterator BI = BB->begin();
2086 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
2088 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);
2095 /// TryToUnfoldSelectInCurrBB - Look for PHI/Select in the same BB of the form
2097 /// %p = phi [false, %bb1], [true, %bb2], [false, %bb3], [true, %bb4], ...
2098 /// %s = select p, trueval, falseval
2100 /// And expand the select into a branch structure. This later enables
2101 /// jump-threading over bb in this pass.
2103 /// Using the similar approach of SimplifyCFG::FoldCondBranchOnPHI(), unfold
2104 /// select if the associated PHI has at least one constant. If the unfolded
2105 /// select is not jump-threaded, it will be folded again in the later
2107 bool JumpThreadingPass::TryToUnfoldSelectInCurrBB(BasicBlock *BB) {
2108 // If threading this would thread across a loop header, don't thread the edge.
2109 // See the comments above FindLoopHeaders for justifications and caveats.
2110 if (LoopHeaders.count(BB))
2113 // Look for a Phi/Select pair in the same basic block. The Phi feeds the
2114 // condition of the Select and at least one of the incoming values is a
2116 for (BasicBlock::iterator BI = BB->begin();
2117 PHINode *PN = dyn_cast<PHINode>(BI); ++BI) {
2118 unsigned NumPHIValues = PN->getNumIncomingValues();
2119 if (NumPHIValues == 0 || !PN->hasOneUse())
2122 SelectInst *SI = dyn_cast<SelectInst>(PN->user_back());
2123 if (!SI || SI->getParent() != BB)
2126 Value *Cond = SI->getCondition();
2127 if (!Cond || Cond != PN || !Cond->getType()->isIntegerTy(1))
2130 bool HasConst = false;
2131 for (unsigned i = 0; i != NumPHIValues; ++i) {
2132 if (PN->getIncomingBlock(i) == BB)
2134 if (isa<ConstantInt>(PN->getIncomingValue(i)))
2139 // Expand the select.
2140 TerminatorInst *Term =
2141 SplitBlockAndInsertIfThen(SI->getCondition(), SI, false);
2142 PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
2143 NewPN->addIncoming(SI->getTrueValue(), Term->getParent());
2144 NewPN->addIncoming(SI->getFalseValue(), BB);
2145 SI->replaceAllUsesWith(NewPN);
2146 SI->eraseFromParent();
2154 /// Try to propagate a guard from the current BB into one of its predecessors
2155 /// in case if another branch of execution implies that the condition of this
2156 /// guard is always true. Currently we only process the simplest case that
2161 /// br i1 %cond, label %T1, label %F1
2167 /// %condGuard = ...
2168 /// call void(i1, ...) @llvm.experimental.guard( i1 %condGuard )[ "deopt"() ]
2170 /// And cond either implies condGuard or !condGuard. In this case all the
2171 /// instructions before the guard can be duplicated in both branches, and the
2172 /// guard is then threaded to one of them.
2173 bool JumpThreadingPass::ProcessGuards(BasicBlock *BB) {
2174 using namespace PatternMatch;
2175 // We only want to deal with two predecessors.
2176 BasicBlock *Pred1, *Pred2;
2177 auto PI = pred_begin(BB), PE = pred_end(BB);
2189 // Try to thread one of the guards of the block.
2190 // TODO: Look up deeper than to immediate predecessor?
2191 auto *Parent = Pred1->getSinglePredecessor();
2192 if (!Parent || Parent != Pred2->getSinglePredecessor())
2195 if (auto *BI = dyn_cast<BranchInst>(Parent->getTerminator()))
2197 if (match(&I, m_Intrinsic<Intrinsic::experimental_guard>()))
2198 if (ThreadGuard(BB, cast<IntrinsicInst>(&I), BI))
2204 /// Try to propagate the guard from BB which is the lower block of a diamond
2205 /// to one of its branches, in case if diamond's condition implies guard's
2207 bool JumpThreadingPass::ThreadGuard(BasicBlock *BB, IntrinsicInst *Guard,
2209 assert(BI->getNumSuccessors() == 2 && "Wrong number of successors?");
2210 assert(BI->isConditional() && "Unconditional branch has 2 successors?");
2211 Value *GuardCond = Guard->getArgOperand(0);
2212 Value *BranchCond = BI->getCondition();
2213 BasicBlock *TrueDest = BI->getSuccessor(0);
2214 BasicBlock *FalseDest = BI->getSuccessor(1);
2216 auto &DL = BB->getModule()->getDataLayout();
2217 bool TrueDestIsSafe = false;
2218 bool FalseDestIsSafe = false;
2220 // True dest is safe if BranchCond => GuardCond.
2221 auto Impl = isImpliedCondition(BranchCond, GuardCond, DL);
2223 TrueDestIsSafe = true;
2225 // False dest is safe if !BranchCond => GuardCond.
2227 isImpliedCondition(BranchCond, GuardCond, DL, /* InvertAPred */ true);
2229 FalseDestIsSafe = true;
2232 if (!TrueDestIsSafe && !FalseDestIsSafe)
2235 BasicBlock *UnguardedBlock = TrueDestIsSafe ? TrueDest : FalseDest;
2236 BasicBlock *GuardedBlock = FalseDestIsSafe ? TrueDest : FalseDest;
2238 ValueToValueMapTy UnguardedMapping, GuardedMapping;
2239 Instruction *AfterGuard = Guard->getNextNode();
2240 unsigned Cost = getJumpThreadDuplicationCost(BB, AfterGuard, BBDupThreshold);
2241 if (Cost > BBDupThreshold)
2243 // Duplicate all instructions before the guard and the guard itself to the
2244 // branch where implication is not proved.
2245 GuardedBlock = DuplicateInstructionsInSplitBetween(
2246 BB, GuardedBlock, AfterGuard, GuardedMapping);
2247 assert(GuardedBlock && "Could not create the guarded block?");
2248 // Duplicate all instructions before the guard in the unguarded branch.
2249 // Since we have successfully duplicated the guarded block and this block
2250 // has fewer instructions, we expect it to succeed.
2251 UnguardedBlock = DuplicateInstructionsInSplitBetween(BB, UnguardedBlock,
2252 Guard, UnguardedMapping);
2253 assert(UnguardedBlock && "Could not create the unguarded block?");
2254 DEBUG(dbgs() << "Moved guard " << *Guard << " to block "
2255 << GuardedBlock->getName() << "\n");
2257 // Some instructions before the guard may still have uses. For them, we need
2258 // to create Phi nodes merging their copies in both guarded and unguarded
2259 // branches. Those instructions that have no uses can be just removed.
2260 SmallVector<Instruction *, 4> ToRemove;
2261 for (auto BI = BB->begin(); &*BI != AfterGuard; ++BI)
2262 if (!isa<PHINode>(&*BI))
2263 ToRemove.push_back(&*BI);
2265 Instruction *InsertionPoint = &*BB->getFirstInsertionPt();
2266 assert(InsertionPoint && "Empty block?");
2267 // Substitute with Phis & remove.
2268 for (auto *Inst : reverse(ToRemove)) {
2269 if (!Inst->use_empty()) {
2270 PHINode *NewPN = PHINode::Create(Inst->getType(), 2);
2271 NewPN->addIncoming(UnguardedMapping[Inst], UnguardedBlock);
2272 NewPN->addIncoming(GuardedMapping[Inst], GuardedBlock);
2273 NewPN->insertBefore(InsertionPoint);
2274 Inst->replaceAllUsesWith(NewPN);
2276 Inst->eraseFromParent();