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.h"
15 #include "llvm/ADT/DenseMap.h"
16 #include "llvm/ADT/DenseSet.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/CFG.h"
22 #include "llvm/Analysis/ConstantFolding.h"
23 #include "llvm/Analysis/InstructionSimplify.h"
24 #include "llvm/Analysis/LazyValueInfo.h"
25 #include "llvm/Analysis/Loads.h"
26 #include "llvm/IR/DataLayout.h"
27 #include "llvm/IR/IntrinsicInst.h"
28 #include "llvm/IR/LLVMContext.h"
29 #include "llvm/IR/ValueHandle.h"
30 #include "llvm/Pass.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetLibraryInfo.h"
35 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
36 #include "llvm/Transforms/Utils/Local.h"
37 #include "llvm/Transforms/Utils/SSAUpdater.h"
40 #define DEBUG_TYPE "jump-threading"
42 STATISTIC(NumThreads, "Number of jumps threaded");
43 STATISTIC(NumFolds, "Number of terminators folded");
44 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
46 static cl::opt<unsigned>
47 Threshold("jump-threading-threshold",
48 cl::desc("Max block size to duplicate for jump threading"),
49 cl::init(6), cl::Hidden);
52 // These are at global scope so static functions can use them too.
53 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo;
54 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy;
56 // This is used to keep track of what kind of constant we're currently hoping
58 enum ConstantPreference {
63 /// This pass performs 'jump threading', which looks at blocks that have
64 /// multiple predecessors and multiple successors. If one or more of the
65 /// predecessors of the block can be proven to always jump to one of the
66 /// successors, we forward the edge from the predecessor to the successor by
67 /// duplicating the contents of this block.
69 /// An example of when this can occur is code like this:
76 /// In this case, the unconditional branch at the end of the first if can be
77 /// revectored to the false side of the second if.
79 class JumpThreading : public FunctionPass {
81 TargetLibraryInfo *TLI;
84 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
86 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
88 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
90 // RAII helper for updating the recursion stack.
91 struct RecursionSetRemover {
92 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
93 std::pair<Value*, BasicBlock*> ThePair;
95 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
96 std::pair<Value*, BasicBlock*> P)
97 : TheSet(S), ThePair(P) { }
99 ~RecursionSetRemover() {
100 TheSet.erase(ThePair);
104 static char ID; // Pass identification
105 JumpThreading() : FunctionPass(ID) {
106 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
109 bool runOnFunction(Function &F) override;
111 void getAnalysisUsage(AnalysisUsage &AU) const override {
112 AU.addRequired<LazyValueInfo>();
113 AU.addPreserved<LazyValueInfo>();
114 AU.addRequired<TargetLibraryInfo>();
117 void FindLoopHeaders(Function &F);
118 bool ProcessBlock(BasicBlock *BB);
119 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
121 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
122 const SmallVectorImpl<BasicBlock *> &PredBBs);
124 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
125 PredValueInfo &Result,
126 ConstantPreference Preference);
127 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
128 ConstantPreference Preference);
130 bool ProcessBranchOnPHI(PHINode *PN);
131 bool ProcessBranchOnXOR(BinaryOperator *BO);
133 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
134 bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB);
138 char JumpThreading::ID = 0;
139 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
140 "Jump Threading", false, false)
141 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
142 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
143 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
144 "Jump Threading", false, false)
146 // Public interface to the Jump Threading pass
147 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
149 /// runOnFunction - Top level algorithm.
151 bool JumpThreading::runOnFunction(Function &F) {
152 if (skipOptnoneFunction(F))
155 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
156 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
157 DL = DLP ? &DLP->getDataLayout() : nullptr;
158 TLI = &getAnalysis<TargetLibraryInfo>();
159 LVI = &getAnalysis<LazyValueInfo>();
161 // Remove unreachable blocks from function as they may result in infinite
162 // loop. We do threading if we found something profitable. Jump threading a
163 // branch can create other opportunities. If these opportunities form a cycle
164 // i.e. if any jump treading is undoing previous threading in the path, then
165 // we will loop forever. We take care of this issue by not jump threading for
166 // back edges. This works for normal cases but not for unreachable blocks as
167 // they may have cycle with no back edge.
168 removeUnreachableBlocks(F);
172 bool Changed, EverChanged = false;
175 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
177 // Thread all of the branches we can over this block.
178 while (ProcessBlock(BB))
183 // If the block is trivially dead, zap it. This eliminates the successor
184 // edges which simplifies the CFG.
185 if (pred_begin(BB) == pred_end(BB) &&
186 BB != &BB->getParent()->getEntryBlock()) {
187 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
188 << "' with terminator: " << *BB->getTerminator() << '\n');
189 LoopHeaders.erase(BB);
196 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
198 // Can't thread an unconditional jump, but if the block is "almost
199 // empty", we can replace uses of it with uses of the successor and make
201 if (BI && BI->isUnconditional() &&
202 BB != &BB->getParent()->getEntryBlock() &&
203 // If the terminator is the only non-phi instruction, try to nuke it.
204 BB->getFirstNonPHIOrDbg()->isTerminator()) {
205 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
206 // block, we have to make sure it isn't in the LoopHeaders set. We
207 // reinsert afterward if needed.
208 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
209 BasicBlock *Succ = BI->getSuccessor(0);
211 // FIXME: It is always conservatively correct to drop the info
212 // for a block even if it doesn't get erased. This isn't totally
213 // awesome, but it allows us to use AssertingVH to prevent nasty
214 // dangling pointer issues within LazyValueInfo.
216 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
218 // If we deleted BB and BB was the header of a loop, then the
219 // successor is now the header of the loop.
223 if (ErasedFromLoopHeaders)
224 LoopHeaders.insert(BB);
227 EverChanged |= Changed;
234 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
235 /// thread across it. Stop scanning the block when passing the threshold.
236 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
237 unsigned Threshold) {
238 /// Ignore PHI nodes, these will be flattened when duplication happens.
239 BasicBlock::const_iterator I = BB->getFirstNonPHI();
241 // FIXME: THREADING will delete values that are just used to compute the
242 // branch, so they shouldn't count against the duplication cost.
244 // Sum up the cost of each instruction until we get to the terminator. Don't
245 // include the terminator because the copy won't include it.
247 for (; !isa<TerminatorInst>(I); ++I) {
249 // Stop scanning the block if we've reached the threshold.
250 if (Size > Threshold)
253 // Debugger intrinsics don't incur code size.
254 if (isa<DbgInfoIntrinsic>(I)) continue;
256 // If this is a pointer->pointer bitcast, it is free.
257 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
260 // All other instructions count for at least one unit.
263 // Calls are more expensive. If they are non-intrinsic calls, we model them
264 // as having cost of 4. If they are a non-vector intrinsic, we model them
265 // as having cost of 2 total, and if they are a vector intrinsic, we model
266 // them as having cost 1.
267 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
268 if (CI->cannotDuplicate())
269 // Blocks with NoDuplicate are modelled as having infinite cost, so they
270 // are never duplicated.
272 else if (!isa<IntrinsicInst>(CI))
274 else if (!CI->getType()->isVectorTy())
279 // Threading through a switch statement is particularly profitable. If this
280 // block ends in a switch, decrease its cost to make it more likely to happen.
281 if (isa<SwitchInst>(I))
282 Size = Size > 6 ? Size-6 : 0;
284 // The same holds for indirect branches, but slightly more so.
285 if (isa<IndirectBrInst>(I))
286 Size = Size > 8 ? Size-8 : 0;
291 /// FindLoopHeaders - We do not want jump threading to turn proper loop
292 /// structures into irreducible loops. Doing this breaks up the loop nesting
293 /// hierarchy and pessimizes later transformations. To prevent this from
294 /// happening, we first have to find the loop headers. Here we approximate this
295 /// by finding targets of backedges in the CFG.
297 /// Note that there definitely are cases when we want to allow threading of
298 /// edges across a loop header. For example, threading a jump from outside the
299 /// loop (the preheader) to an exit block of the loop is definitely profitable.
300 /// It is also almost always profitable to thread backedges from within the loop
301 /// to exit blocks, and is often profitable to thread backedges to other blocks
302 /// within the loop (forming a nested loop). This simple analysis is not rich
303 /// enough to track all of these properties and keep it up-to-date as the CFG
304 /// mutates, so we don't allow any of these transformations.
306 void JumpThreading::FindLoopHeaders(Function &F) {
307 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
308 FindFunctionBackedges(F, Edges);
310 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
311 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
314 /// getKnownConstant - Helper method to determine if we can thread over a
315 /// terminator with the given value as its condition, and if so what value to
316 /// use for that. What kind of value this is depends on whether we want an
317 /// integer or a block address, but an undef is always accepted.
318 /// Returns null if Val is null or not an appropriate constant.
319 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
323 // Undef is "known" enough.
324 if (UndefValue *U = dyn_cast<UndefValue>(Val))
327 if (Preference == WantBlockAddress)
328 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
330 return dyn_cast<ConstantInt>(Val);
333 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
334 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
335 /// in any of our predecessors. If so, return the known list of value and pred
336 /// BB in the result vector.
338 /// This returns true if there were any known values.
341 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
342 ConstantPreference Preference) {
343 // This method walks up use-def chains recursively. Because of this, we could
344 // get into an infinite loop going around loops in the use-def chain. To
345 // prevent this, keep track of what (value, block) pairs we've already visited
346 // and terminate the search if we loop back to them
347 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
350 // An RAII help to remove this pair from the recursion set once the recursion
351 // stack pops back out again.
352 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
354 // If V is a constant, then it is known in all predecessors.
355 if (Constant *KC = getKnownConstant(V, Preference)) {
356 for (BasicBlock *Pred : predecessors(BB))
357 Result.push_back(std::make_pair(KC, Pred));
362 // If V is a non-instruction value, or an instruction in a different block,
363 // then it can't be derived from a PHI.
364 Instruction *I = dyn_cast<Instruction>(V);
365 if (!I || I->getParent() != BB) {
367 // Okay, if this is a live-in value, see if it has a known value at the end
368 // of any of our predecessors.
370 // FIXME: This should be an edge property, not a block end property.
371 /// TODO: Per PR2563, we could infer value range information about a
372 /// predecessor based on its terminator.
374 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
375 // "I" is a non-local compare-with-a-constant instruction. This would be
376 // able to handle value inequalities better, for example if the compare is
377 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
378 // Perhaps getConstantOnEdge should be smart enough to do this?
380 for (BasicBlock *P : predecessors(BB)) {
381 // If the value is known by LazyValueInfo to be a constant in a
382 // predecessor, use that information to try to thread this block.
383 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
384 if (Constant *KC = getKnownConstant(PredCst, Preference))
385 Result.push_back(std::make_pair(KC, P));
388 return !Result.empty();
391 /// If I is a PHI node, then we know the incoming values for any constants.
392 if (PHINode *PN = dyn_cast<PHINode>(I)) {
393 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
394 Value *InVal = PN->getIncomingValue(i);
395 if (Constant *KC = getKnownConstant(InVal, Preference)) {
396 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
398 Constant *CI = LVI->getConstantOnEdge(InVal,
399 PN->getIncomingBlock(i), BB);
400 if (Constant *KC = getKnownConstant(CI, Preference))
401 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
405 return !Result.empty();
408 PredValueInfoTy LHSVals, RHSVals;
410 // Handle some boolean conditions.
411 if (I->getType()->getPrimitiveSizeInBits() == 1) {
412 assert(Preference == WantInteger && "One-bit non-integer type?");
414 // X & false -> false
415 if (I->getOpcode() == Instruction::Or ||
416 I->getOpcode() == Instruction::And) {
417 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
419 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
422 if (LHSVals.empty() && RHSVals.empty())
425 ConstantInt *InterestingVal;
426 if (I->getOpcode() == Instruction::Or)
427 InterestingVal = ConstantInt::getTrue(I->getContext());
429 InterestingVal = ConstantInt::getFalse(I->getContext());
431 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
433 // Scan for the sentinel. If we find an undef, force it to the
434 // interesting value: x|undef -> true and x&undef -> false.
435 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
436 if (LHSVals[i].first == InterestingVal ||
437 isa<UndefValue>(LHSVals[i].first)) {
438 Result.push_back(LHSVals[i]);
439 Result.back().first = InterestingVal;
440 LHSKnownBBs.insert(LHSVals[i].second);
442 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
443 if (RHSVals[i].first == InterestingVal ||
444 isa<UndefValue>(RHSVals[i].first)) {
445 // If we already inferred a value for this block on the LHS, don't
447 if (!LHSKnownBBs.count(RHSVals[i].second)) {
448 Result.push_back(RHSVals[i]);
449 Result.back().first = InterestingVal;
453 return !Result.empty();
456 // Handle the NOT form of XOR.
457 if (I->getOpcode() == Instruction::Xor &&
458 isa<ConstantInt>(I->getOperand(1)) &&
459 cast<ConstantInt>(I->getOperand(1))->isOne()) {
460 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
465 // Invert the known values.
466 for (unsigned i = 0, e = Result.size(); i != e; ++i)
467 Result[i].first = ConstantExpr::getNot(Result[i].first);
472 // Try to simplify some other binary operator values.
473 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
474 assert(Preference != WantBlockAddress
475 && "A binary operator creating a block address?");
476 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
477 PredValueInfoTy LHSVals;
478 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
481 // Try to use constant folding to simplify the binary operator.
482 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
483 Constant *V = LHSVals[i].first;
484 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
486 if (Constant *KC = getKnownConstant(Folded, WantInteger))
487 Result.push_back(std::make_pair(KC, LHSVals[i].second));
491 return !Result.empty();
494 // Handle compare with phi operand, where the PHI is defined in this block.
495 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
496 assert(Preference == WantInteger && "Compares only produce integers");
497 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
498 if (PN && PN->getParent() == BB) {
499 // We can do this simplification if any comparisons fold to true or false.
501 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
502 BasicBlock *PredBB = PN->getIncomingBlock(i);
503 Value *LHS = PN->getIncomingValue(i);
504 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
506 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
508 if (!isa<Constant>(RHS))
511 LazyValueInfo::Tristate
512 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
513 cast<Constant>(RHS), PredBB, BB);
514 if (ResT == LazyValueInfo::Unknown)
516 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
519 if (Constant *KC = getKnownConstant(Res, WantInteger))
520 Result.push_back(std::make_pair(KC, PredBB));
523 return !Result.empty();
527 // If comparing a live-in value against a constant, see if we know the
528 // live-in value on any predecessors.
529 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
530 if (!isa<Instruction>(Cmp->getOperand(0)) ||
531 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
532 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
534 for (BasicBlock *P : predecessors(BB)) {
535 // If the value is known by LazyValueInfo to be a constant in a
536 // predecessor, use that information to try to thread this block.
537 LazyValueInfo::Tristate Res =
538 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
540 if (Res == LazyValueInfo::Unknown)
543 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
544 Result.push_back(std::make_pair(ResC, P));
547 return !Result.empty();
550 // Try to find a constant value for the LHS of a comparison,
551 // and evaluate it statically if we can.
552 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
553 PredValueInfoTy LHSVals;
554 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
557 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
558 Constant *V = LHSVals[i].first;
559 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
561 if (Constant *KC = getKnownConstant(Folded, WantInteger))
562 Result.push_back(std::make_pair(KC, LHSVals[i].second));
565 return !Result.empty();
570 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
571 // Handle select instructions where at least one operand is a known constant
572 // and we can figure out the condition value for any predecessor block.
573 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
574 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
575 PredValueInfoTy Conds;
576 if ((TrueVal || FalseVal) &&
577 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
579 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
580 Constant *Cond = Conds[i].first;
582 // Figure out what value to use for the condition.
584 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
586 KnownCond = CI->isOne();
588 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
589 // Either operand will do, so be sure to pick the one that's a known
591 // FIXME: Do this more cleverly if both values are known constants?
592 KnownCond = (TrueVal != nullptr);
595 // See if the select has a known constant value for this predecessor.
596 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
597 Result.push_back(std::make_pair(Val, Conds[i].second));
600 return !Result.empty();
604 // If all else fails, see if LVI can figure out a constant value for us.
605 Constant *CI = LVI->getConstant(V, BB);
606 if (Constant *KC = getKnownConstant(CI, Preference)) {
607 for (BasicBlock *Pred : predecessors(BB))
608 Result.push_back(std::make_pair(KC, Pred));
611 return !Result.empty();
616 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
617 /// in an undefined jump, decide which block is best to revector to.
619 /// Since we can pick an arbitrary destination, we pick the successor with the
620 /// fewest predecessors. This should reduce the in-degree of the others.
622 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
623 TerminatorInst *BBTerm = BB->getTerminator();
624 unsigned MinSucc = 0;
625 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
626 // Compute the successor with the minimum number of predecessors.
627 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
628 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
629 TestBB = BBTerm->getSuccessor(i);
630 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
631 if (NumPreds < MinNumPreds) {
633 MinNumPreds = NumPreds;
640 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
641 if (!BB->hasAddressTaken()) return false;
643 // If the block has its address taken, it may be a tree of dead constants
644 // hanging off of it. These shouldn't keep the block alive.
645 BlockAddress *BA = BlockAddress::get(BB);
646 BA->removeDeadConstantUsers();
647 return !BA->use_empty();
650 /// ProcessBlock - If there are any predecessors whose control can be threaded
651 /// through to a successor, transform them now.
652 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
653 // If the block is trivially dead, just return and let the caller nuke it.
654 // This simplifies other transformations.
655 if (pred_begin(BB) == pred_end(BB) &&
656 BB != &BB->getParent()->getEntryBlock())
659 // If this block has a single predecessor, and if that pred has a single
660 // successor, merge the blocks. This encourages recursive jump threading
661 // because now the condition in this block can be threaded through
662 // predecessors of our predecessor block.
663 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
664 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
665 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
666 // If SinglePred was a loop header, BB becomes one.
667 if (LoopHeaders.erase(SinglePred))
668 LoopHeaders.insert(BB);
670 LVI->eraseBlock(SinglePred);
671 MergeBasicBlockIntoOnlyPred(BB);
677 // What kind of constant we're looking for.
678 ConstantPreference Preference = WantInteger;
680 // Look to see if the terminator is a conditional branch, switch or indirect
681 // branch, if not we can't thread it.
683 Instruction *Terminator = BB->getTerminator();
684 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
685 // Can't thread an unconditional jump.
686 if (BI->isUnconditional()) return false;
687 Condition = BI->getCondition();
688 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
689 Condition = SI->getCondition();
690 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
691 // Can't thread indirect branch with no successors.
692 if (IB->getNumSuccessors() == 0) return false;
693 Condition = IB->getAddress()->stripPointerCasts();
694 Preference = WantBlockAddress;
696 return false; // Must be an invoke.
699 // Run constant folding to see if we can reduce the condition to a simple
701 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
702 Value *SimpleVal = ConstantFoldInstruction(I, DL, TLI);
704 I->replaceAllUsesWith(SimpleVal);
705 I->eraseFromParent();
706 Condition = SimpleVal;
710 // If the terminator is branching on an undef, we can pick any of the
711 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
712 if (isa<UndefValue>(Condition)) {
713 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
715 // Fold the branch/switch.
716 TerminatorInst *BBTerm = BB->getTerminator();
717 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
718 if (i == BestSucc) continue;
719 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
722 DEBUG(dbgs() << " In block '" << BB->getName()
723 << "' folding undef terminator: " << *BBTerm << '\n');
724 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
725 BBTerm->eraseFromParent();
729 // If the terminator of this block is branching on a constant, simplify the
730 // terminator to an unconditional branch. This can occur due to threading in
732 if (getKnownConstant(Condition, Preference)) {
733 DEBUG(dbgs() << " In block '" << BB->getName()
734 << "' folding terminator: " << *BB->getTerminator() << '\n');
736 ConstantFoldTerminator(BB, true);
740 Instruction *CondInst = dyn_cast<Instruction>(Condition);
742 // All the rest of our checks depend on the condition being an instruction.
744 // FIXME: Unify this with code below.
745 if (ProcessThreadableEdges(Condition, BB, Preference))
751 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
752 // For a comparison where the LHS is outside this block, it's possible
753 // that we've branched on it before. Used LVI to see if we can simplify
754 // the branch based on that.
755 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
756 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
757 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
758 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
759 (!isa<Instruction>(CondCmp->getOperand(0)) ||
760 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
761 // For predecessor edge, determine if the comparison is true or false
762 // on that edge. If they're all true or all false, we can simplify the
764 // FIXME: We could handle mixed true/false by duplicating code.
765 LazyValueInfo::Tristate Baseline =
766 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
768 if (Baseline != LazyValueInfo::Unknown) {
769 // Check that all remaining incoming values match the first one.
771 LazyValueInfo::Tristate Ret =
772 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
773 CondCmp->getOperand(0), CondConst, *PI, BB);
774 if (Ret != Baseline) break;
777 // If we terminated early, then one of the values didn't match.
779 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
780 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
781 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
782 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
783 CondBr->eraseFromParent();
790 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
794 // Check for some cases that are worth simplifying. Right now we want to look
795 // for loads that are used by a switch or by the condition for the branch. If
796 // we see one, check to see if it's partially redundant. If so, insert a PHI
797 // which can then be used to thread the values.
799 Value *SimplifyValue = CondInst;
800 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
801 if (isa<Constant>(CondCmp->getOperand(1)))
802 SimplifyValue = CondCmp->getOperand(0);
804 // TODO: There are other places where load PRE would be profitable, such as
805 // more complex comparisons.
806 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
807 if (SimplifyPartiallyRedundantLoad(LI))
811 // Handle a variety of cases where we are branching on something derived from
812 // a PHI node in the current block. If we can prove that any predecessors
813 // compute a predictable value based on a PHI node, thread those predecessors.
815 if (ProcessThreadableEdges(CondInst, BB, Preference))
818 // If this is an otherwise-unfoldable branch on a phi node in the current
819 // block, see if we can simplify.
820 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
821 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
822 return ProcessBranchOnPHI(PN);
825 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
826 if (CondInst->getOpcode() == Instruction::Xor &&
827 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
828 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
831 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
832 // "(X == 4)", thread through this block.
837 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
838 /// load instruction, eliminate it by replacing it with a PHI node. This is an
839 /// important optimization that encourages jump threading, and needs to be run
840 /// interlaced with other jump threading tasks.
841 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
842 // Don't hack volatile/atomic loads.
843 if (!LI->isSimple()) return false;
845 // If the load is defined in a block with exactly one predecessor, it can't be
846 // partially redundant.
847 BasicBlock *LoadBB = LI->getParent();
848 if (LoadBB->getSinglePredecessor())
851 // If the load is defined in a landing pad, it can't be partially redundant,
852 // because the edges between the invoke and the landing pad cannot have other
853 // instructions between them.
854 if (LoadBB->isLandingPad())
857 Value *LoadedPtr = LI->getOperand(0);
859 // If the loaded operand is defined in the LoadBB, it can't be available.
860 // TODO: Could do simple PHI translation, that would be fun :)
861 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
862 if (PtrOp->getParent() == LoadBB)
865 // Scan a few instructions up from the load, to see if it is obviously live at
866 // the entry to its block.
867 BasicBlock::iterator BBIt = LI;
869 if (Value *AvailableVal =
870 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
871 // If the value if the load is locally available within the block, just use
872 // it. This frequently occurs for reg2mem'd allocas.
873 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
875 // If the returned value is the load itself, replace with an undef. This can
876 // only happen in dead loops.
877 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
878 LI->replaceAllUsesWith(AvailableVal);
879 LI->eraseFromParent();
883 // Otherwise, if we scanned the whole block and got to the top of the block,
884 // we know the block is locally transparent to the load. If not, something
885 // might clobber its value.
886 if (BBIt != LoadBB->begin())
889 // If all of the loads and stores that feed the value have the same TBAA tag,
890 // then we can propagate it onto any newly inserted loads.
891 MDNode *TBAATag = LI->getMetadata(LLVMContext::MD_tbaa);
893 SmallPtrSet<BasicBlock*, 8> PredsScanned;
894 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
895 AvailablePredsTy AvailablePreds;
896 BasicBlock *OneUnavailablePred = nullptr;
898 // If we got here, the loaded value is transparent through to the start of the
899 // block. Check to see if it is available in any of the predecessor blocks.
900 for (BasicBlock *PredBB : predecessors(LoadBB)) {
901 // If we already scanned this predecessor, skip it.
902 if (!PredsScanned.insert(PredBB))
905 // Scan the predecessor to see if the value is available in the pred.
906 BBIt = PredBB->end();
907 MDNode *ThisTBAATag = nullptr;
908 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6,
909 nullptr, &ThisTBAATag);
910 if (!PredAvailable) {
911 OneUnavailablePred = PredBB;
915 // If tbaa tags disagree or are not present, forget about them.
916 if (TBAATag != ThisTBAATag) TBAATag = nullptr;
918 // If so, this load is partially redundant. Remember this info so that we
919 // can create a PHI node.
920 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
923 // If the loaded value isn't available in any predecessor, it isn't partially
925 if (AvailablePreds.empty()) return false;
927 // Okay, the loaded value is available in at least one (and maybe all!)
928 // predecessors. If the value is unavailable in more than one unique
929 // predecessor, we want to insert a merge block for those common predecessors.
930 // This ensures that we only have to insert one reload, thus not increasing
932 BasicBlock *UnavailablePred = nullptr;
934 // If there is exactly one predecessor where the value is unavailable, the
935 // already computed 'OneUnavailablePred' block is it. If it ends in an
936 // unconditional branch, we know that it isn't a critical edge.
937 if (PredsScanned.size() == AvailablePreds.size()+1 &&
938 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
939 UnavailablePred = OneUnavailablePred;
940 } else if (PredsScanned.size() != AvailablePreds.size()) {
941 // Otherwise, we had multiple unavailable predecessors or we had a critical
942 // edge from the one.
943 SmallVector<BasicBlock*, 8> PredsToSplit;
944 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
946 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
947 AvailablePredSet.insert(AvailablePreds[i].first);
949 // Add all the unavailable predecessors to the PredsToSplit list.
950 for (BasicBlock *P : predecessors(LoadBB)) {
951 // If the predecessor is an indirect goto, we can't split the edge.
952 if (isa<IndirectBrInst>(P->getTerminator()))
955 if (!AvailablePredSet.count(P))
956 PredsToSplit.push_back(P);
959 // Split them out to their own block.
961 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this);
964 // If the value isn't available in all predecessors, then there will be
965 // exactly one where it isn't available. Insert a load on that edge and add
966 // it to the AvailablePreds list.
967 if (UnavailablePred) {
968 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
969 "Can't handle critical edge here!");
970 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
972 UnavailablePred->getTerminator());
973 NewVal->setDebugLoc(LI->getDebugLoc());
975 NewVal->setMetadata(LLVMContext::MD_tbaa, TBAATag);
977 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
980 // Now we know that each predecessor of this block has a value in
981 // AvailablePreds, sort them for efficient access as we're walking the preds.
982 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
984 // Create a PHI node at the start of the block for the PRE'd load value.
985 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
986 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
989 PN->setDebugLoc(LI->getDebugLoc());
991 // Insert new entries into the PHI for each predecessor. A single block may
992 // have multiple entries here.
993 for (pred_iterator PI = PB; PI != PE; ++PI) {
995 AvailablePredsTy::iterator I =
996 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
997 std::make_pair(P, (Value*)nullptr));
999 assert(I != AvailablePreds.end() && I->first == P &&
1000 "Didn't find entry for predecessor!");
1002 PN->addIncoming(I->second, I->first);
1005 //cerr << "PRE: " << *LI << *PN << "\n";
1007 LI->replaceAllUsesWith(PN);
1008 LI->eraseFromParent();
1013 /// FindMostPopularDest - The specified list contains multiple possible
1014 /// threadable destinations. Pick the one that occurs the most frequently in
1017 FindMostPopularDest(BasicBlock *BB,
1018 const SmallVectorImpl<std::pair<BasicBlock*,
1019 BasicBlock*> > &PredToDestList) {
1020 assert(!PredToDestList.empty());
1022 // Determine popularity. If there are multiple possible destinations, we
1023 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1024 // blocks with known and real destinations to threading undef. We'll handle
1025 // them later if interesting.
1026 DenseMap<BasicBlock*, unsigned> DestPopularity;
1027 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1028 if (PredToDestList[i].second)
1029 DestPopularity[PredToDestList[i].second]++;
1031 // Find the most popular dest.
1032 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1033 BasicBlock *MostPopularDest = DPI->first;
1034 unsigned Popularity = DPI->second;
1035 SmallVector<BasicBlock*, 4> SamePopularity;
1037 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1038 // If the popularity of this entry isn't higher than the popularity we've
1039 // seen so far, ignore it.
1040 if (DPI->second < Popularity)
1042 else if (DPI->second == Popularity) {
1043 // If it is the same as what we've seen so far, keep track of it.
1044 SamePopularity.push_back(DPI->first);
1046 // If it is more popular, remember it.
1047 SamePopularity.clear();
1048 MostPopularDest = DPI->first;
1049 Popularity = DPI->second;
1053 // Okay, now we know the most popular destination. If there is more than one
1054 // destination, we need to determine one. This is arbitrary, but we need
1055 // to make a deterministic decision. Pick the first one that appears in the
1057 if (!SamePopularity.empty()) {
1058 SamePopularity.push_back(MostPopularDest);
1059 TerminatorInst *TI = BB->getTerminator();
1060 for (unsigned i = 0; ; ++i) {
1061 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1063 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1064 TI->getSuccessor(i)) == SamePopularity.end())
1067 MostPopularDest = TI->getSuccessor(i);
1072 // Okay, we have finally picked the most popular destination.
1073 return MostPopularDest;
1076 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1077 ConstantPreference Preference) {
1078 // If threading this would thread across a loop header, don't even try to
1080 if (LoopHeaders.count(BB))
1083 PredValueInfoTy PredValues;
1084 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference))
1087 assert(!PredValues.empty() &&
1088 "ComputeValueKnownInPredecessors returned true with no values");
1090 DEBUG(dbgs() << "IN BB: " << *BB;
1091 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1092 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1093 << *PredValues[i].first
1094 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1097 // Decide what we want to thread through. Convert our list of known values to
1098 // a list of known destinations for each pred. This also discards duplicate
1099 // predecessors and keeps track of the undefined inputs (which are represented
1100 // as a null dest in the PredToDestList).
1101 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1102 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1104 BasicBlock *OnlyDest = nullptr;
1105 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1107 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1108 BasicBlock *Pred = PredValues[i].second;
1109 if (!SeenPreds.insert(Pred))
1110 continue; // Duplicate predecessor entry.
1112 // If the predecessor ends with an indirect goto, we can't change its
1114 if (isa<IndirectBrInst>(Pred->getTerminator()))
1117 Constant *Val = PredValues[i].first;
1120 if (isa<UndefValue>(Val))
1122 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1123 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1124 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1125 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1127 assert(isa<IndirectBrInst>(BB->getTerminator())
1128 && "Unexpected terminator");
1129 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1132 // If we have exactly one destination, remember it for efficiency below.
1133 if (PredToDestList.empty())
1135 else if (OnlyDest != DestBB)
1136 OnlyDest = MultipleDestSentinel;
1138 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1141 // If all edges were unthreadable, we fail.
1142 if (PredToDestList.empty())
1145 // Determine which is the most common successor. If we have many inputs and
1146 // this block is a switch, we want to start by threading the batch that goes
1147 // to the most popular destination first. If we only know about one
1148 // threadable destination (the common case) we can avoid this.
1149 BasicBlock *MostPopularDest = OnlyDest;
1151 if (MostPopularDest == MultipleDestSentinel)
1152 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1154 // Now that we know what the most popular destination is, factor all
1155 // predecessors that will jump to it into a single predecessor.
1156 SmallVector<BasicBlock*, 16> PredsToFactor;
1157 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1158 if (PredToDestList[i].second == MostPopularDest) {
1159 BasicBlock *Pred = PredToDestList[i].first;
1161 // This predecessor may be a switch or something else that has multiple
1162 // edges to the block. Factor each of these edges by listing them
1163 // according to # occurrences in PredsToFactor.
1164 TerminatorInst *PredTI = Pred->getTerminator();
1165 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1166 if (PredTI->getSuccessor(i) == BB)
1167 PredsToFactor.push_back(Pred);
1170 // If the threadable edges are branching on an undefined value, we get to pick
1171 // the destination that these predecessors should get to.
1172 if (!MostPopularDest)
1173 MostPopularDest = BB->getTerminator()->
1174 getSuccessor(GetBestDestForJumpOnUndef(BB));
1176 // Ok, try to thread it!
1177 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1180 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1181 /// a PHI node in the current block. See if there are any simplifications we
1182 /// can do based on inputs to the phi node.
1184 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1185 BasicBlock *BB = PN->getParent();
1187 // TODO: We could make use of this to do it once for blocks with common PHI
1189 SmallVector<BasicBlock*, 1> PredBBs;
1192 // If any of the predecessor blocks end in an unconditional branch, we can
1193 // *duplicate* the conditional branch into that block in order to further
1194 // encourage jump threading and to eliminate cases where we have branch on a
1195 // phi of an icmp (branch on icmp is much better).
1196 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1197 BasicBlock *PredBB = PN->getIncomingBlock(i);
1198 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1199 if (PredBr->isUnconditional()) {
1200 PredBBs[0] = PredBB;
1201 // Try to duplicate BB into PredBB.
1202 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1210 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1211 /// a xor instruction in the current block. See if there are any
1212 /// simplifications we can do based on inputs to the xor.
1214 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1215 BasicBlock *BB = BO->getParent();
1217 // If either the LHS or RHS of the xor is a constant, don't do this
1219 if (isa<ConstantInt>(BO->getOperand(0)) ||
1220 isa<ConstantInt>(BO->getOperand(1)))
1223 // If the first instruction in BB isn't a phi, we won't be able to infer
1224 // anything special about any particular predecessor.
1225 if (!isa<PHINode>(BB->front()))
1228 // If we have a xor as the branch input to this block, and we know that the
1229 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1230 // the condition into the predecessor and fix that value to true, saving some
1231 // logical ops on that path and encouraging other paths to simplify.
1233 // This copies something like this:
1236 // %X = phi i1 [1], [%X']
1237 // %Y = icmp eq i32 %A, %B
1238 // %Z = xor i1 %X, %Y
1243 // %Y = icmp ne i32 %A, %B
1246 PredValueInfoTy XorOpValues;
1248 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1250 assert(XorOpValues.empty());
1251 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1257 assert(!XorOpValues.empty() &&
1258 "ComputeValueKnownInPredecessors returned true with no values");
1260 // Scan the information to see which is most popular: true or false. The
1261 // predecessors can be of the set true, false, or undef.
1262 unsigned NumTrue = 0, NumFalse = 0;
1263 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1264 if (isa<UndefValue>(XorOpValues[i].first))
1265 // Ignore undefs for the count.
1267 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1273 // Determine which value to split on, true, false, or undef if neither.
1274 ConstantInt *SplitVal = nullptr;
1275 if (NumTrue > NumFalse)
1276 SplitVal = ConstantInt::getTrue(BB->getContext());
1277 else if (NumTrue != 0 || NumFalse != 0)
1278 SplitVal = ConstantInt::getFalse(BB->getContext());
1280 // Collect all of the blocks that this can be folded into so that we can
1281 // factor this once and clone it once.
1282 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1283 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1284 if (XorOpValues[i].first != SplitVal &&
1285 !isa<UndefValue>(XorOpValues[i].first))
1288 BlocksToFoldInto.push_back(XorOpValues[i].second);
1291 // If we inferred a value for all of the predecessors, then duplication won't
1292 // help us. However, we can just replace the LHS or RHS with the constant.
1293 if (BlocksToFoldInto.size() ==
1294 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1296 // If all preds provide undef, just nuke the xor, because it is undef too.
1297 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1298 BO->eraseFromParent();
1299 } else if (SplitVal->isZero()) {
1300 // If all preds provide 0, replace the xor with the other input.
1301 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1302 BO->eraseFromParent();
1304 // If all preds provide 1, set the computed value to 1.
1305 BO->setOperand(!isLHS, SplitVal);
1311 // Try to duplicate BB into PredBB.
1312 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1316 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1317 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1318 /// NewPred using the entries from OldPred (suitably mapped).
1319 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1320 BasicBlock *OldPred,
1321 BasicBlock *NewPred,
1322 DenseMap<Instruction*, Value*> &ValueMap) {
1323 for (BasicBlock::iterator PNI = PHIBB->begin();
1324 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1325 // Ok, we have a PHI node. Figure out what the incoming value was for the
1327 Value *IV = PN->getIncomingValueForBlock(OldPred);
1329 // Remap the value if necessary.
1330 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1331 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1332 if (I != ValueMap.end())
1336 PN->addIncoming(IV, NewPred);
1340 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1341 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1342 /// across BB. Transform the IR to reflect this change.
1343 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1344 const SmallVectorImpl<BasicBlock*> &PredBBs,
1345 BasicBlock *SuccBB) {
1346 // If threading to the same block as we come from, we would infinite loop.
1348 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1349 << "' - would thread to self!\n");
1353 // If threading this would thread across a loop header, don't thread the edge.
1354 // See the comments above FindLoopHeaders for justifications and caveats.
1355 if (LoopHeaders.count(BB)) {
1356 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1357 << "' to dest BB '" << SuccBB->getName()
1358 << "' - it might create an irreducible loop!\n");
1362 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold);
1363 if (JumpThreadCost > Threshold) {
1364 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1365 << "' - Cost is too high: " << JumpThreadCost << "\n");
1369 // And finally, do it! Start by factoring the predecessors is needed.
1371 if (PredBBs.size() == 1)
1372 PredBB = PredBBs[0];
1374 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1375 << " common predecessors.\n");
1376 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1379 // And finally, do it!
1380 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1381 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1382 << ", across block:\n "
1385 LVI->threadEdge(PredBB, BB, SuccBB);
1387 // We are going to have to map operands from the original BB block to the new
1388 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1389 // account for entry from PredBB.
1390 DenseMap<Instruction*, Value*> ValueMapping;
1392 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1393 BB->getName()+".thread",
1394 BB->getParent(), BB);
1395 NewBB->moveAfter(PredBB);
1397 BasicBlock::iterator BI = BB->begin();
1398 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1399 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1401 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1402 // mapping and using it to remap operands in the cloned instructions.
1403 for (; !isa<TerminatorInst>(BI); ++BI) {
1404 Instruction *New = BI->clone();
1405 New->setName(BI->getName());
1406 NewBB->getInstList().push_back(New);
1407 ValueMapping[BI] = New;
1409 // Remap operands to patch up intra-block references.
1410 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1411 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1412 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1413 if (I != ValueMapping.end())
1414 New->setOperand(i, I->second);
1418 // We didn't copy the terminator from BB over to NewBB, because there is now
1419 // an unconditional jump to SuccBB. Insert the unconditional jump.
1420 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB);
1421 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1423 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1424 // PHI nodes for NewBB now.
1425 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1427 // If there were values defined in BB that are used outside the block, then we
1428 // now have to update all uses of the value to use either the original value,
1429 // the cloned value, or some PHI derived value. This can require arbitrary
1430 // PHI insertion, of which we are prepared to do, clean these up now.
1431 SSAUpdater SSAUpdate;
1432 SmallVector<Use*, 16> UsesToRename;
1433 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1434 // Scan all uses of this instruction to see if it is used outside of its
1435 // block, and if so, record them in UsesToRename.
1436 for (Use &U : I->uses()) {
1437 Instruction *User = cast<Instruction>(U.getUser());
1438 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1439 if (UserPN->getIncomingBlock(U) == BB)
1441 } else if (User->getParent() == BB)
1444 UsesToRename.push_back(&U);
1447 // If there are no uses outside the block, we're done with this instruction.
1448 if (UsesToRename.empty())
1451 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1453 // We found a use of I outside of BB. Rename all uses of I that are outside
1454 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1455 // with the two values we know.
1456 SSAUpdate.Initialize(I->getType(), I->getName());
1457 SSAUpdate.AddAvailableValue(BB, I);
1458 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1460 while (!UsesToRename.empty())
1461 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1462 DEBUG(dbgs() << "\n");
1466 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1467 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1468 // us to simplify any PHI nodes in BB.
1469 TerminatorInst *PredTerm = PredBB->getTerminator();
1470 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1471 if (PredTerm->getSuccessor(i) == BB) {
1472 BB->removePredecessor(PredBB, true);
1473 PredTerm->setSuccessor(i, NewBB);
1476 // At this point, the IR is fully up to date and consistent. Do a quick scan
1477 // over the new instructions and zap any that are constants or dead. This
1478 // frequently happens because of phi translation.
1479 SimplifyInstructionsInBlock(NewBB, DL, TLI);
1481 // Threaded an edge!
1486 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1487 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1488 /// If we can duplicate the contents of BB up into PredBB do so now, this
1489 /// improves the odds that the branch will be on an analyzable instruction like
1491 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1492 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1493 assert(!PredBBs.empty() && "Can't handle an empty set");
1495 // If BB is a loop header, then duplicating this block outside the loop would
1496 // cause us to transform this into an irreducible loop, don't do this.
1497 // See the comments above FindLoopHeaders for justifications and caveats.
1498 if (LoopHeaders.count(BB)) {
1499 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1500 << "' into predecessor block '" << PredBBs[0]->getName()
1501 << "' - it might create an irreducible loop!\n");
1505 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold);
1506 if (DuplicationCost > Threshold) {
1507 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1508 << "' - Cost is too high: " << DuplicationCost << "\n");
1512 // And finally, do it! Start by factoring the predecessors is needed.
1514 if (PredBBs.size() == 1)
1515 PredBB = PredBBs[0];
1517 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1518 << " common predecessors.\n");
1519 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1522 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1524 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1525 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1526 << DuplicationCost << " block is:" << *BB << "\n");
1528 // Unless PredBB ends with an unconditional branch, split the edge so that we
1529 // can just clone the bits from BB into the end of the new PredBB.
1530 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1532 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1533 PredBB = SplitEdge(PredBB, BB, this);
1534 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1537 // We are going to have to map operands from the original BB block into the
1538 // PredBB block. Evaluate PHI nodes in BB.
1539 DenseMap<Instruction*, Value*> ValueMapping;
1541 BasicBlock::iterator BI = BB->begin();
1542 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1543 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1545 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1546 // mapping and using it to remap operands in the cloned instructions.
1547 for (; BI != BB->end(); ++BI) {
1548 Instruction *New = BI->clone();
1550 // Remap operands to patch up intra-block references.
1551 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1552 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1553 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1554 if (I != ValueMapping.end())
1555 New->setOperand(i, I->second);
1558 // If this instruction can be simplified after the operands are updated,
1559 // just use the simplified value instead. This frequently happens due to
1561 if (Value *IV = SimplifyInstruction(New, DL)) {
1563 ValueMapping[BI] = IV;
1565 // Otherwise, insert the new instruction into the block.
1566 New->setName(BI->getName());
1567 PredBB->getInstList().insert(OldPredBranch, New);
1568 ValueMapping[BI] = New;
1572 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1573 // add entries to the PHI nodes for branch from PredBB now.
1574 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1575 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1577 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1580 // If there were values defined in BB that are used outside the block, then we
1581 // now have to update all uses of the value to use either the original value,
1582 // the cloned value, or some PHI derived value. This can require arbitrary
1583 // PHI insertion, of which we are prepared to do, clean these up now.
1584 SSAUpdater SSAUpdate;
1585 SmallVector<Use*, 16> UsesToRename;
1586 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1587 // Scan all uses of this instruction to see if it is used outside of its
1588 // block, and if so, record them in UsesToRename.
1589 for (Use &U : I->uses()) {
1590 Instruction *User = cast<Instruction>(U.getUser());
1591 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1592 if (UserPN->getIncomingBlock(U) == BB)
1594 } else if (User->getParent() == BB)
1597 UsesToRename.push_back(&U);
1600 // If there are no uses outside the block, we're done with this instruction.
1601 if (UsesToRename.empty())
1604 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1606 // We found a use of I outside of BB. Rename all uses of I that are outside
1607 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1608 // with the two values we know.
1609 SSAUpdate.Initialize(I->getType(), I->getName());
1610 SSAUpdate.AddAvailableValue(BB, I);
1611 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1613 while (!UsesToRename.empty())
1614 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1615 DEBUG(dbgs() << "\n");
1618 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1620 BB->removePredecessor(PredBB, true);
1622 // Remove the unconditional branch at the end of the PredBB block.
1623 OldPredBranch->eraseFromParent();
1629 /// TryToUnfoldSelect - Look for blocks of the form
1635 /// %p = phi [%a, %bb] ...
1639 /// And expand the select into a branch structure if one of its arms allows %c
1640 /// to be folded. This later enables threading from bb1 over bb2.
1641 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1642 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1643 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1644 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1646 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1647 CondLHS->getParent() != BB)
1650 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1651 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1652 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1654 // Look if one of the incoming values is a select in the corresponding
1656 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1659 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1660 if (!PredTerm || !PredTerm->isUnconditional())
1663 // Now check if one of the select values would allow us to constant fold the
1664 // terminator in BB. We don't do the transform if both sides fold, those
1665 // cases will be threaded in any case.
1666 LazyValueInfo::Tristate LHSFolds =
1667 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1669 LazyValueInfo::Tristate RHSFolds =
1670 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1672 if ((LHSFolds != LazyValueInfo::Unknown ||
1673 RHSFolds != LazyValueInfo::Unknown) &&
1674 LHSFolds != RHSFolds) {
1675 // Expand the select.
1684 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1685 BB->getParent(), BB);
1686 // Move the unconditional branch to NewBB.
1687 PredTerm->removeFromParent();
1688 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1689 // Create a conditional branch and update PHI nodes.
1690 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1691 CondLHS->setIncomingValue(I, SI->getFalseValue());
1692 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1693 // The select is now dead.
1694 SI->eraseFromParent();
1696 // Update any other PHI nodes in BB.
1697 for (BasicBlock::iterator BI = BB->begin();
1698 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1700 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);