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/Metadata.h"
30 #include "llvm/IR/ValueHandle.h"
31 #include "llvm/Pass.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetLibraryInfo.h"
36 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
37 #include "llvm/Transforms/Utils/Local.h"
38 #include "llvm/Transforms/Utils/SSAUpdater.h"
41 #define DEBUG_TYPE "jump-threading"
43 STATISTIC(NumThreads, "Number of jumps threaded");
44 STATISTIC(NumFolds, "Number of terminators folded");
45 STATISTIC(NumDupes, "Number of branch blocks duplicated to eliminate phi");
47 static cl::opt<unsigned>
48 Threshold("jump-threading-threshold",
49 cl::desc("Max block size to duplicate for jump threading"),
50 cl::init(6), cl::Hidden);
53 // These are at global scope so static functions can use them too.
54 typedef SmallVectorImpl<std::pair<Constant*, BasicBlock*> > PredValueInfo;
55 typedef SmallVector<std::pair<Constant*, BasicBlock*>, 8> PredValueInfoTy;
57 // This is used to keep track of what kind of constant we're currently hoping
59 enum ConstantPreference {
64 /// This pass performs 'jump threading', which looks at blocks that have
65 /// multiple predecessors and multiple successors. If one or more of the
66 /// predecessors of the block can be proven to always jump to one of the
67 /// successors, we forward the edge from the predecessor to the successor by
68 /// duplicating the contents of this block.
70 /// An example of when this can occur is code like this:
77 /// In this case, the unconditional branch at the end of the first if can be
78 /// revectored to the false side of the second if.
80 class JumpThreading : public FunctionPass {
82 TargetLibraryInfo *TLI;
85 SmallPtrSet<BasicBlock*, 16> LoopHeaders;
87 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders;
89 DenseSet<std::pair<Value*, BasicBlock*> > RecursionSet;
91 // RAII helper for updating the recursion stack.
92 struct RecursionSetRemover {
93 DenseSet<std::pair<Value*, BasicBlock*> > &TheSet;
94 std::pair<Value*, BasicBlock*> ThePair;
96 RecursionSetRemover(DenseSet<std::pair<Value*, BasicBlock*> > &S,
97 std::pair<Value*, BasicBlock*> P)
98 : TheSet(S), ThePair(P) { }
100 ~RecursionSetRemover() {
101 TheSet.erase(ThePair);
105 static char ID; // Pass identification
106 JumpThreading() : FunctionPass(ID) {
107 initializeJumpThreadingPass(*PassRegistry::getPassRegistry());
110 bool runOnFunction(Function &F) override;
112 void getAnalysisUsage(AnalysisUsage &AU) const override {
113 AU.addRequired<LazyValueInfo>();
114 AU.addPreserved<LazyValueInfo>();
115 AU.addRequired<TargetLibraryInfo>();
118 void FindLoopHeaders(Function &F);
119 bool ProcessBlock(BasicBlock *BB);
120 bool ThreadEdge(BasicBlock *BB, const SmallVectorImpl<BasicBlock*> &PredBBs,
122 bool DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
123 const SmallVectorImpl<BasicBlock *> &PredBBs);
125 bool ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB,
126 PredValueInfo &Result,
127 ConstantPreference Preference);
128 bool ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
129 ConstantPreference Preference);
131 bool ProcessBranchOnPHI(PHINode *PN);
132 bool ProcessBranchOnXOR(BinaryOperator *BO);
134 bool SimplifyPartiallyRedundantLoad(LoadInst *LI);
135 bool TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB);
139 char JumpThreading::ID = 0;
140 INITIALIZE_PASS_BEGIN(JumpThreading, "jump-threading",
141 "Jump Threading", false, false)
142 INITIALIZE_PASS_DEPENDENCY(LazyValueInfo)
143 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
144 INITIALIZE_PASS_END(JumpThreading, "jump-threading",
145 "Jump Threading", false, false)
147 // Public interface to the Jump Threading pass
148 FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); }
150 /// runOnFunction - Top level algorithm.
152 bool JumpThreading::runOnFunction(Function &F) {
153 if (skipOptnoneFunction(F))
156 DEBUG(dbgs() << "Jump threading on function '" << F.getName() << "'\n");
157 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
158 DL = DLP ? &DLP->getDataLayout() : nullptr;
159 TLI = &getAnalysis<TargetLibraryInfo>();
160 LVI = &getAnalysis<LazyValueInfo>();
162 // Remove unreachable blocks from function as they may result in infinite
163 // loop. We do threading if we found something profitable. Jump threading a
164 // branch can create other opportunities. If these opportunities form a cycle
165 // i.e. if any jump treading is undoing previous threading in the path, then
166 // we will loop forever. We take care of this issue by not jump threading for
167 // back edges. This works for normal cases but not for unreachable blocks as
168 // they may have cycle with no back edge.
169 removeUnreachableBlocks(F);
173 bool Changed, EverChanged = false;
176 for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
178 // Thread all of the branches we can over this block.
179 while (ProcessBlock(BB))
184 // If the block is trivially dead, zap it. This eliminates the successor
185 // edges which simplifies the CFG.
186 if (pred_begin(BB) == pred_end(BB) &&
187 BB != &BB->getParent()->getEntryBlock()) {
188 DEBUG(dbgs() << " JT: Deleting dead block '" << BB->getName()
189 << "' with terminator: " << *BB->getTerminator() << '\n');
190 LoopHeaders.erase(BB);
197 BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator());
199 // Can't thread an unconditional jump, but if the block is "almost
200 // empty", we can replace uses of it with uses of the successor and make
202 if (BI && BI->isUnconditional() &&
203 BB != &BB->getParent()->getEntryBlock() &&
204 // If the terminator is the only non-phi instruction, try to nuke it.
205 BB->getFirstNonPHIOrDbg()->isTerminator()) {
206 // Since TryToSimplifyUncondBranchFromEmptyBlock may delete the
207 // block, we have to make sure it isn't in the LoopHeaders set. We
208 // reinsert afterward if needed.
209 bool ErasedFromLoopHeaders = LoopHeaders.erase(BB);
210 BasicBlock *Succ = BI->getSuccessor(0);
212 // FIXME: It is always conservatively correct to drop the info
213 // for a block even if it doesn't get erased. This isn't totally
214 // awesome, but it allows us to use AssertingVH to prevent nasty
215 // dangling pointer issues within LazyValueInfo.
217 if (TryToSimplifyUncondBranchFromEmptyBlock(BB)) {
219 // If we deleted BB and BB was the header of a loop, then the
220 // successor is now the header of the loop.
224 if (ErasedFromLoopHeaders)
225 LoopHeaders.insert(BB);
228 EverChanged |= Changed;
235 /// getJumpThreadDuplicationCost - Return the cost of duplicating this block to
236 /// thread across it. Stop scanning the block when passing the threshold.
237 static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB,
238 unsigned Threshold) {
239 /// Ignore PHI nodes, these will be flattened when duplication happens.
240 BasicBlock::const_iterator I = BB->getFirstNonPHI();
242 // FIXME: THREADING will delete values that are just used to compute the
243 // branch, so they shouldn't count against the duplication cost.
245 // Sum up the cost of each instruction until we get to the terminator. Don't
246 // include the terminator because the copy won't include it.
248 for (; !isa<TerminatorInst>(I); ++I) {
250 // Stop scanning the block if we've reached the threshold.
251 if (Size > Threshold)
254 // Debugger intrinsics don't incur code size.
255 if (isa<DbgInfoIntrinsic>(I)) continue;
257 // If this is a pointer->pointer bitcast, it is free.
258 if (isa<BitCastInst>(I) && I->getType()->isPointerTy())
261 // All other instructions count for at least one unit.
264 // Calls are more expensive. If they are non-intrinsic calls, we model them
265 // as having cost of 4. If they are a non-vector intrinsic, we model them
266 // as having cost of 2 total, and if they are a vector intrinsic, we model
267 // them as having cost 1.
268 if (const CallInst *CI = dyn_cast<CallInst>(I)) {
269 if (CI->cannotDuplicate())
270 // Blocks with NoDuplicate are modelled as having infinite cost, so they
271 // are never duplicated.
273 else if (!isa<IntrinsicInst>(CI))
275 else if (!CI->getType()->isVectorTy())
280 // Threading through a switch statement is particularly profitable. If this
281 // block ends in a switch, decrease its cost to make it more likely to happen.
282 if (isa<SwitchInst>(I))
283 Size = Size > 6 ? Size-6 : 0;
285 // The same holds for indirect branches, but slightly more so.
286 if (isa<IndirectBrInst>(I))
287 Size = Size > 8 ? Size-8 : 0;
292 /// FindLoopHeaders - We do not want jump threading to turn proper loop
293 /// structures into irreducible loops. Doing this breaks up the loop nesting
294 /// hierarchy and pessimizes later transformations. To prevent this from
295 /// happening, we first have to find the loop headers. Here we approximate this
296 /// by finding targets of backedges in the CFG.
298 /// Note that there definitely are cases when we want to allow threading of
299 /// edges across a loop header. For example, threading a jump from outside the
300 /// loop (the preheader) to an exit block of the loop is definitely profitable.
301 /// It is also almost always profitable to thread backedges from within the loop
302 /// to exit blocks, and is often profitable to thread backedges to other blocks
303 /// within the loop (forming a nested loop). This simple analysis is not rich
304 /// enough to track all of these properties and keep it up-to-date as the CFG
305 /// mutates, so we don't allow any of these transformations.
307 void JumpThreading::FindLoopHeaders(Function &F) {
308 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges;
309 FindFunctionBackedges(F, Edges);
311 for (unsigned i = 0, e = Edges.size(); i != e; ++i)
312 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second));
315 /// getKnownConstant - Helper method to determine if we can thread over a
316 /// terminator with the given value as its condition, and if so what value to
317 /// use for that. What kind of value this is depends on whether we want an
318 /// integer or a block address, but an undef is always accepted.
319 /// Returns null if Val is null or not an appropriate constant.
320 static Constant *getKnownConstant(Value *Val, ConstantPreference Preference) {
324 // Undef is "known" enough.
325 if (UndefValue *U = dyn_cast<UndefValue>(Val))
328 if (Preference == WantBlockAddress)
329 return dyn_cast<BlockAddress>(Val->stripPointerCasts());
331 return dyn_cast<ConstantInt>(Val);
334 /// ComputeValueKnownInPredecessors - Given a basic block BB and a value V, see
335 /// if we can infer that the value is a known ConstantInt/BlockAddress or undef
336 /// in any of our predecessors. If so, return the known list of value and pred
337 /// BB in the result vector.
339 /// This returns true if there were any known values.
342 ComputeValueKnownInPredecessors(Value *V, BasicBlock *BB, PredValueInfo &Result,
343 ConstantPreference Preference) {
344 // This method walks up use-def chains recursively. Because of this, we could
345 // get into an infinite loop going around loops in the use-def chain. To
346 // prevent this, keep track of what (value, block) pairs we've already visited
347 // and terminate the search if we loop back to them
348 if (!RecursionSet.insert(std::make_pair(V, BB)).second)
351 // An RAII help to remove this pair from the recursion set once the recursion
352 // stack pops back out again.
353 RecursionSetRemover remover(RecursionSet, std::make_pair(V, BB));
355 // If V is a constant, then it is known in all predecessors.
356 if (Constant *KC = getKnownConstant(V, Preference)) {
357 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
358 Result.push_back(std::make_pair(KC, *PI));
363 // If V is a non-instruction value, or an instruction in a different block,
364 // then it can't be derived from a PHI.
365 Instruction *I = dyn_cast<Instruction>(V);
366 if (!I || I->getParent() != BB) {
368 // Okay, if this is a live-in value, see if it has a known value at the end
369 // of any of our predecessors.
371 // FIXME: This should be an edge property, not a block end property.
372 /// TODO: Per PR2563, we could infer value range information about a
373 /// predecessor based on its terminator.
375 // FIXME: change this to use the more-rich 'getPredicateOnEdge' method if
376 // "I" is a non-local compare-with-a-constant instruction. This would be
377 // able to handle value inequalities better, for example if the compare is
378 // "X < 4" and "X < 3" is known true but "X < 4" itself is not available.
379 // Perhaps getConstantOnEdge should be smart enough to do this?
381 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
383 // If the value is known by LazyValueInfo to be a constant in a
384 // predecessor, use that information to try to thread this block.
385 Constant *PredCst = LVI->getConstantOnEdge(V, P, BB);
386 if (Constant *KC = getKnownConstant(PredCst, Preference))
387 Result.push_back(std::make_pair(KC, P));
390 return !Result.empty();
393 /// If I is a PHI node, then we know the incoming values for any constants.
394 if (PHINode *PN = dyn_cast<PHINode>(I)) {
395 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
396 Value *InVal = PN->getIncomingValue(i);
397 if (Constant *KC = getKnownConstant(InVal, Preference)) {
398 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
400 Constant *CI = LVI->getConstantOnEdge(InVal,
401 PN->getIncomingBlock(i), BB);
402 if (Constant *KC = getKnownConstant(CI, Preference))
403 Result.push_back(std::make_pair(KC, PN->getIncomingBlock(i)));
407 return !Result.empty();
410 PredValueInfoTy LHSVals, RHSVals;
412 // Handle some boolean conditions.
413 if (I->getType()->getPrimitiveSizeInBits() == 1) {
414 assert(Preference == WantInteger && "One-bit non-integer type?");
416 // X & false -> false
417 if (I->getOpcode() == Instruction::Or ||
418 I->getOpcode() == Instruction::And) {
419 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
421 ComputeValueKnownInPredecessors(I->getOperand(1), BB, RHSVals,
424 if (LHSVals.empty() && RHSVals.empty())
427 ConstantInt *InterestingVal;
428 if (I->getOpcode() == Instruction::Or)
429 InterestingVal = ConstantInt::getTrue(I->getContext());
431 InterestingVal = ConstantInt::getFalse(I->getContext());
433 SmallPtrSet<BasicBlock*, 4> LHSKnownBBs;
435 // Scan for the sentinel. If we find an undef, force it to the
436 // interesting value: x|undef -> true and x&undef -> false.
437 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i)
438 if (LHSVals[i].first == InterestingVal ||
439 isa<UndefValue>(LHSVals[i].first)) {
440 Result.push_back(LHSVals[i]);
441 Result.back().first = InterestingVal;
442 LHSKnownBBs.insert(LHSVals[i].second);
444 for (unsigned i = 0, e = RHSVals.size(); i != e; ++i)
445 if (RHSVals[i].first == InterestingVal ||
446 isa<UndefValue>(RHSVals[i].first)) {
447 // If we already inferred a value for this block on the LHS, don't
449 if (!LHSKnownBBs.count(RHSVals[i].second)) {
450 Result.push_back(RHSVals[i]);
451 Result.back().first = InterestingVal;
455 return !Result.empty();
458 // Handle the NOT form of XOR.
459 if (I->getOpcode() == Instruction::Xor &&
460 isa<ConstantInt>(I->getOperand(1)) &&
461 cast<ConstantInt>(I->getOperand(1))->isOne()) {
462 ComputeValueKnownInPredecessors(I->getOperand(0), BB, Result,
467 // Invert the known values.
468 for (unsigned i = 0, e = Result.size(); i != e; ++i)
469 Result[i].first = ConstantExpr::getNot(Result[i].first);
474 // Try to simplify some other binary operator values.
475 } else if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
476 assert(Preference != WantBlockAddress
477 && "A binary operator creating a block address?");
478 if (ConstantInt *CI = dyn_cast<ConstantInt>(BO->getOperand(1))) {
479 PredValueInfoTy LHSVals;
480 ComputeValueKnownInPredecessors(BO->getOperand(0), BB, LHSVals,
483 // Try to use constant folding to simplify the binary operator.
484 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
485 Constant *V = LHSVals[i].first;
486 Constant *Folded = ConstantExpr::get(BO->getOpcode(), V, CI);
488 if (Constant *KC = getKnownConstant(Folded, WantInteger))
489 Result.push_back(std::make_pair(KC, LHSVals[i].second));
493 return !Result.empty();
496 // Handle compare with phi operand, where the PHI is defined in this block.
497 if (CmpInst *Cmp = dyn_cast<CmpInst>(I)) {
498 assert(Preference == WantInteger && "Compares only produce integers");
499 PHINode *PN = dyn_cast<PHINode>(Cmp->getOperand(0));
500 if (PN && PN->getParent() == BB) {
501 // We can do this simplification if any comparisons fold to true or false.
503 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
504 BasicBlock *PredBB = PN->getIncomingBlock(i);
505 Value *LHS = PN->getIncomingValue(i);
506 Value *RHS = Cmp->getOperand(1)->DoPHITranslation(BB, PredBB);
508 Value *Res = SimplifyCmpInst(Cmp->getPredicate(), LHS, RHS, DL);
510 if (!isa<Constant>(RHS))
513 LazyValueInfo::Tristate
514 ResT = LVI->getPredicateOnEdge(Cmp->getPredicate(), LHS,
515 cast<Constant>(RHS), PredBB, BB);
516 if (ResT == LazyValueInfo::Unknown)
518 Res = ConstantInt::get(Type::getInt1Ty(LHS->getContext()), ResT);
521 if (Constant *KC = getKnownConstant(Res, WantInteger))
522 Result.push_back(std::make_pair(KC, PredBB));
525 return !Result.empty();
529 // If comparing a live-in value against a constant, see if we know the
530 // live-in value on any predecessors.
531 if (isa<Constant>(Cmp->getOperand(1)) && Cmp->getType()->isIntegerTy()) {
532 if (!isa<Instruction>(Cmp->getOperand(0)) ||
533 cast<Instruction>(Cmp->getOperand(0))->getParent() != BB) {
534 Constant *RHSCst = cast<Constant>(Cmp->getOperand(1));
536 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB);PI != E; ++PI){
538 // If the value is known by LazyValueInfo to be a constant in a
539 // predecessor, use that information to try to thread this block.
540 LazyValueInfo::Tristate Res =
541 LVI->getPredicateOnEdge(Cmp->getPredicate(), Cmp->getOperand(0),
543 if (Res == LazyValueInfo::Unknown)
546 Constant *ResC = ConstantInt::get(Cmp->getType(), Res);
547 Result.push_back(std::make_pair(ResC, P));
550 return !Result.empty();
553 // Try to find a constant value for the LHS of a comparison,
554 // and evaluate it statically if we can.
555 if (Constant *CmpConst = dyn_cast<Constant>(Cmp->getOperand(1))) {
556 PredValueInfoTy LHSVals;
557 ComputeValueKnownInPredecessors(I->getOperand(0), BB, LHSVals,
560 for (unsigned i = 0, e = LHSVals.size(); i != e; ++i) {
561 Constant *V = LHSVals[i].first;
562 Constant *Folded = ConstantExpr::getCompare(Cmp->getPredicate(),
564 if (Constant *KC = getKnownConstant(Folded, WantInteger))
565 Result.push_back(std::make_pair(KC, LHSVals[i].second));
568 return !Result.empty();
573 if (SelectInst *SI = dyn_cast<SelectInst>(I)) {
574 // Handle select instructions where at least one operand is a known constant
575 // and we can figure out the condition value for any predecessor block.
576 Constant *TrueVal = getKnownConstant(SI->getTrueValue(), Preference);
577 Constant *FalseVal = getKnownConstant(SI->getFalseValue(), Preference);
578 PredValueInfoTy Conds;
579 if ((TrueVal || FalseVal) &&
580 ComputeValueKnownInPredecessors(SI->getCondition(), BB, Conds,
582 for (unsigned i = 0, e = Conds.size(); i != e; ++i) {
583 Constant *Cond = Conds[i].first;
585 // Figure out what value to use for the condition.
587 if (ConstantInt *CI = dyn_cast<ConstantInt>(Cond)) {
589 KnownCond = CI->isOne();
591 assert(isa<UndefValue>(Cond) && "Unexpected condition value");
592 // Either operand will do, so be sure to pick the one that's a known
594 // FIXME: Do this more cleverly if both values are known constants?
595 KnownCond = (TrueVal != nullptr);
598 // See if the select has a known constant value for this predecessor.
599 if (Constant *Val = KnownCond ? TrueVal : FalseVal)
600 Result.push_back(std::make_pair(Val, Conds[i].second));
603 return !Result.empty();
607 // If all else fails, see if LVI can figure out a constant value for us.
608 Constant *CI = LVI->getConstant(V, BB);
609 if (Constant *KC = getKnownConstant(CI, Preference)) {
610 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
611 Result.push_back(std::make_pair(KC, *PI));
614 return !Result.empty();
619 /// GetBestDestForBranchOnUndef - If we determine that the specified block ends
620 /// in an undefined jump, decide which block is best to revector to.
622 /// Since we can pick an arbitrary destination, we pick the successor with the
623 /// fewest predecessors. This should reduce the in-degree of the others.
625 static unsigned GetBestDestForJumpOnUndef(BasicBlock *BB) {
626 TerminatorInst *BBTerm = BB->getTerminator();
627 unsigned MinSucc = 0;
628 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc);
629 // Compute the successor with the minimum number of predecessors.
630 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
631 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) {
632 TestBB = BBTerm->getSuccessor(i);
633 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB));
634 if (NumPreds < MinNumPreds) {
636 MinNumPreds = NumPreds;
643 static bool hasAddressTakenAndUsed(BasicBlock *BB) {
644 if (!BB->hasAddressTaken()) return false;
646 // If the block has its address taken, it may be a tree of dead constants
647 // hanging off of it. These shouldn't keep the block alive.
648 BlockAddress *BA = BlockAddress::get(BB);
649 BA->removeDeadConstantUsers();
650 return !BA->use_empty();
653 /// ProcessBlock - If there are any predecessors whose control can be threaded
654 /// through to a successor, transform them now.
655 bool JumpThreading::ProcessBlock(BasicBlock *BB) {
656 // If the block is trivially dead, just return and let the caller nuke it.
657 // This simplifies other transformations.
658 if (pred_begin(BB) == pred_end(BB) &&
659 BB != &BB->getParent()->getEntryBlock())
662 // If this block has a single predecessor, and if that pred has a single
663 // successor, merge the blocks. This encourages recursive jump threading
664 // because now the condition in this block can be threaded through
665 // predecessors of our predecessor block.
666 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) {
667 if (SinglePred->getTerminator()->getNumSuccessors() == 1 &&
668 SinglePred != BB && !hasAddressTakenAndUsed(BB)) {
669 // If SinglePred was a loop header, BB becomes one.
670 if (LoopHeaders.erase(SinglePred))
671 LoopHeaders.insert(BB);
673 LVI->eraseBlock(SinglePred);
674 MergeBasicBlockIntoOnlyPred(BB);
680 // What kind of constant we're looking for.
681 ConstantPreference Preference = WantInteger;
683 // Look to see if the terminator is a conditional branch, switch or indirect
684 // branch, if not we can't thread it.
686 Instruction *Terminator = BB->getTerminator();
687 if (BranchInst *BI = dyn_cast<BranchInst>(Terminator)) {
688 // Can't thread an unconditional jump.
689 if (BI->isUnconditional()) return false;
690 Condition = BI->getCondition();
691 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(Terminator)) {
692 Condition = SI->getCondition();
693 } else if (IndirectBrInst *IB = dyn_cast<IndirectBrInst>(Terminator)) {
694 // Can't thread indirect branch with no successors.
695 if (IB->getNumSuccessors() == 0) return false;
696 Condition = IB->getAddress()->stripPointerCasts();
697 Preference = WantBlockAddress;
699 return false; // Must be an invoke.
702 // Run constant folding to see if we can reduce the condition to a simple
704 if (Instruction *I = dyn_cast<Instruction>(Condition)) {
705 Value *SimpleVal = ConstantFoldInstruction(I, DL, TLI);
707 I->replaceAllUsesWith(SimpleVal);
708 I->eraseFromParent();
709 Condition = SimpleVal;
713 // If the terminator is branching on an undef, we can pick any of the
714 // successors to branch to. Let GetBestDestForJumpOnUndef decide.
715 if (isa<UndefValue>(Condition)) {
716 unsigned BestSucc = GetBestDestForJumpOnUndef(BB);
718 // Fold the branch/switch.
719 TerminatorInst *BBTerm = BB->getTerminator();
720 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) {
721 if (i == BestSucc) continue;
722 BBTerm->getSuccessor(i)->removePredecessor(BB, true);
725 DEBUG(dbgs() << " In block '" << BB->getName()
726 << "' folding undef terminator: " << *BBTerm << '\n');
727 BranchInst::Create(BBTerm->getSuccessor(BestSucc), BBTerm);
728 BBTerm->eraseFromParent();
732 // If the terminator of this block is branching on a constant, simplify the
733 // terminator to an unconditional branch. This can occur due to threading in
735 if (getKnownConstant(Condition, Preference)) {
736 DEBUG(dbgs() << " In block '" << BB->getName()
737 << "' folding terminator: " << *BB->getTerminator() << '\n');
739 ConstantFoldTerminator(BB, true);
743 Instruction *CondInst = dyn_cast<Instruction>(Condition);
745 // All the rest of our checks depend on the condition being an instruction.
747 // FIXME: Unify this with code below.
748 if (ProcessThreadableEdges(Condition, BB, Preference))
754 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) {
755 // For a comparison where the LHS is outside this block, it's possible
756 // that we've branched on it before. Used LVI to see if we can simplify
757 // the branch based on that.
758 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
759 Constant *CondConst = dyn_cast<Constant>(CondCmp->getOperand(1));
760 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
761 if (CondBr && CondConst && CondBr->isConditional() && PI != PE &&
762 (!isa<Instruction>(CondCmp->getOperand(0)) ||
763 cast<Instruction>(CondCmp->getOperand(0))->getParent() != BB)) {
764 // For predecessor edge, determine if the comparison is true or false
765 // on that edge. If they're all true or all false, we can simplify the
767 // FIXME: We could handle mixed true/false by duplicating code.
768 LazyValueInfo::Tristate Baseline =
769 LVI->getPredicateOnEdge(CondCmp->getPredicate(), CondCmp->getOperand(0),
771 if (Baseline != LazyValueInfo::Unknown) {
772 // Check that all remaining incoming values match the first one.
774 LazyValueInfo::Tristate Ret =
775 LVI->getPredicateOnEdge(CondCmp->getPredicate(),
776 CondCmp->getOperand(0), CondConst, *PI, BB);
777 if (Ret != Baseline) break;
780 // If we terminated early, then one of the values didn't match.
782 unsigned ToRemove = Baseline == LazyValueInfo::True ? 1 : 0;
783 unsigned ToKeep = Baseline == LazyValueInfo::True ? 0 : 1;
784 CondBr->getSuccessor(ToRemove)->removePredecessor(BB, true);
785 BranchInst::Create(CondBr->getSuccessor(ToKeep), CondBr);
786 CondBr->eraseFromParent();
793 if (CondBr && CondConst && TryToUnfoldSelect(CondCmp, BB))
797 // Check for some cases that are worth simplifying. Right now we want to look
798 // for loads that are used by a switch or by the condition for the branch. If
799 // we see one, check to see if it's partially redundant. If so, insert a PHI
800 // which can then be used to thread the values.
802 Value *SimplifyValue = CondInst;
803 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue))
804 if (isa<Constant>(CondCmp->getOperand(1)))
805 SimplifyValue = CondCmp->getOperand(0);
807 // TODO: There are other places where load PRE would be profitable, such as
808 // more complex comparisons.
809 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue))
810 if (SimplifyPartiallyRedundantLoad(LI))
814 // Handle a variety of cases where we are branching on something derived from
815 // a PHI node in the current block. If we can prove that any predecessors
816 // compute a predictable value based on a PHI node, thread those predecessors.
818 if (ProcessThreadableEdges(CondInst, BB, Preference))
821 // If this is an otherwise-unfoldable branch on a phi node in the current
822 // block, see if we can simplify.
823 if (PHINode *PN = dyn_cast<PHINode>(CondInst))
824 if (PN->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
825 return ProcessBranchOnPHI(PN);
828 // If this is an otherwise-unfoldable branch on a XOR, see if we can simplify.
829 if (CondInst->getOpcode() == Instruction::Xor &&
830 CondInst->getParent() == BB && isa<BranchInst>(BB->getTerminator()))
831 return ProcessBranchOnXOR(cast<BinaryOperator>(CondInst));
834 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know
835 // "(X == 4)", thread through this block.
840 /// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant
841 /// load instruction, eliminate it by replacing it with a PHI node. This is an
842 /// important optimization that encourages jump threading, and needs to be run
843 /// interlaced with other jump threading tasks.
844 bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) {
845 // Don't hack volatile/atomic loads.
846 if (!LI->isSimple()) return false;
848 // If the load is defined in a block with exactly one predecessor, it can't be
849 // partially redundant.
850 BasicBlock *LoadBB = LI->getParent();
851 if (LoadBB->getSinglePredecessor())
854 // If the load is defined in a landing pad, it can't be partially redundant,
855 // because the edges between the invoke and the landing pad cannot have other
856 // instructions between them.
857 if (LoadBB->isLandingPad())
860 Value *LoadedPtr = LI->getOperand(0);
862 // If the loaded operand is defined in the LoadBB, it can't be available.
863 // TODO: Could do simple PHI translation, that would be fun :)
864 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr))
865 if (PtrOp->getParent() == LoadBB)
868 // Scan a few instructions up from the load, to see if it is obviously live at
869 // the entry to its block.
870 BasicBlock::iterator BBIt = LI;
872 if (Value *AvailableVal =
873 FindAvailableLoadedValue(LoadedPtr, LoadBB, BBIt, 6)) {
874 // If the value if the load is locally available within the block, just use
875 // it. This frequently occurs for reg2mem'd allocas.
876 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n";
878 // If the returned value is the load itself, replace with an undef. This can
879 // only happen in dead loops.
880 if (AvailableVal == LI) AvailableVal = UndefValue::get(LI->getType());
881 LI->replaceAllUsesWith(AvailableVal);
882 LI->eraseFromParent();
886 // Otherwise, if we scanned the whole block and got to the top of the block,
887 // we know the block is locally transparent to the load. If not, something
888 // might clobber its value.
889 if (BBIt != LoadBB->begin())
892 // If all of the loads and stores that feed the value have the same AA tags,
893 // then we can propagate them onto any newly inserted loads.
895 LI->getAAMetadata(AATags);
897 SmallPtrSet<BasicBlock*, 8> PredsScanned;
898 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy;
899 AvailablePredsTy AvailablePreds;
900 BasicBlock *OneUnavailablePred = nullptr;
902 // If we got here, the loaded value is transparent through to the start of the
903 // block. Check to see if it is available in any of the predecessor blocks.
904 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
906 BasicBlock *PredBB = *PI;
908 // If we already scanned this predecessor, skip it.
909 if (!PredsScanned.insert(PredBB))
912 // Scan the predecessor to see if the value is available in the pred.
913 BBIt = PredBB->end();
914 AAMDNodes ThisAATags;
915 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6,
916 nullptr, &ThisAATags);
917 if (!PredAvailable) {
918 OneUnavailablePred = PredBB;
922 // If AA tags disagree or are not present, forget about them.
923 if (AATags != ThisAATags) AATags = AAMDNodes();
925 // If so, this load is partially redundant. Remember this info so that we
926 // can create a PHI node.
927 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable));
930 // If the loaded value isn't available in any predecessor, it isn't partially
932 if (AvailablePreds.empty()) return false;
934 // Okay, the loaded value is available in at least one (and maybe all!)
935 // predecessors. If the value is unavailable in more than one unique
936 // predecessor, we want to insert a merge block for those common predecessors.
937 // This ensures that we only have to insert one reload, thus not increasing
939 BasicBlock *UnavailablePred = nullptr;
941 // If there is exactly one predecessor where the value is unavailable, the
942 // already computed 'OneUnavailablePred' block is it. If it ends in an
943 // unconditional branch, we know that it isn't a critical edge.
944 if (PredsScanned.size() == AvailablePreds.size()+1 &&
945 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) {
946 UnavailablePred = OneUnavailablePred;
947 } else if (PredsScanned.size() != AvailablePreds.size()) {
948 // Otherwise, we had multiple unavailable predecessors or we had a critical
949 // edge from the one.
950 SmallVector<BasicBlock*, 8> PredsToSplit;
951 SmallPtrSet<BasicBlock*, 8> AvailablePredSet;
953 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i)
954 AvailablePredSet.insert(AvailablePreds[i].first);
956 // Add all the unavailable predecessors to the PredsToSplit list.
957 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB);
960 // If the predecessor is an indirect goto, we can't split the edge.
961 if (isa<IndirectBrInst>(P->getTerminator()))
964 if (!AvailablePredSet.count(P))
965 PredsToSplit.push_back(P);
968 // Split them out to their own block.
970 SplitBlockPredecessors(LoadBB, PredsToSplit, "thread-pre-split", this);
973 // If the value isn't available in all predecessors, then there will be
974 // exactly one where it isn't available. Insert a load on that edge and add
975 // it to the AvailablePreds list.
976 if (UnavailablePred) {
977 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 &&
978 "Can't handle critical edge here!");
979 LoadInst *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", false,
981 UnavailablePred->getTerminator());
982 NewVal->setDebugLoc(LI->getDebugLoc());
984 NewVal->setAAMetadata(AATags);
986 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal));
989 // Now we know that each predecessor of this block has a value in
990 // AvailablePreds, sort them for efficient access as we're walking the preds.
991 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end());
993 // Create a PHI node at the start of the block for the PRE'd load value.
994 pred_iterator PB = pred_begin(LoadBB), PE = pred_end(LoadBB);
995 PHINode *PN = PHINode::Create(LI->getType(), std::distance(PB, PE), "",
998 PN->setDebugLoc(LI->getDebugLoc());
1000 // Insert new entries into the PHI for each predecessor. A single block may
1001 // have multiple entries here.
1002 for (pred_iterator PI = PB; PI != PE; ++PI) {
1003 BasicBlock *P = *PI;
1004 AvailablePredsTy::iterator I =
1005 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(),
1006 std::make_pair(P, (Value*)nullptr));
1008 assert(I != AvailablePreds.end() && I->first == P &&
1009 "Didn't find entry for predecessor!");
1011 PN->addIncoming(I->second, I->first);
1014 //cerr << "PRE: " << *LI << *PN << "\n";
1016 LI->replaceAllUsesWith(PN);
1017 LI->eraseFromParent();
1022 /// FindMostPopularDest - The specified list contains multiple possible
1023 /// threadable destinations. Pick the one that occurs the most frequently in
1026 FindMostPopularDest(BasicBlock *BB,
1027 const SmallVectorImpl<std::pair<BasicBlock*,
1028 BasicBlock*> > &PredToDestList) {
1029 assert(!PredToDestList.empty());
1031 // Determine popularity. If there are multiple possible destinations, we
1032 // explicitly choose to ignore 'undef' destinations. We prefer to thread
1033 // blocks with known and real destinations to threading undef. We'll handle
1034 // them later if interesting.
1035 DenseMap<BasicBlock*, unsigned> DestPopularity;
1036 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1037 if (PredToDestList[i].second)
1038 DestPopularity[PredToDestList[i].second]++;
1040 // Find the most popular dest.
1041 DenseMap<BasicBlock*, unsigned>::iterator DPI = DestPopularity.begin();
1042 BasicBlock *MostPopularDest = DPI->first;
1043 unsigned Popularity = DPI->second;
1044 SmallVector<BasicBlock*, 4> SamePopularity;
1046 for (++DPI; DPI != DestPopularity.end(); ++DPI) {
1047 // If the popularity of this entry isn't higher than the popularity we've
1048 // seen so far, ignore it.
1049 if (DPI->second < Popularity)
1051 else if (DPI->second == Popularity) {
1052 // If it is the same as what we've seen so far, keep track of it.
1053 SamePopularity.push_back(DPI->first);
1055 // If it is more popular, remember it.
1056 SamePopularity.clear();
1057 MostPopularDest = DPI->first;
1058 Popularity = DPI->second;
1062 // Okay, now we know the most popular destination. If there is more than one
1063 // destination, we need to determine one. This is arbitrary, but we need
1064 // to make a deterministic decision. Pick the first one that appears in the
1066 if (!SamePopularity.empty()) {
1067 SamePopularity.push_back(MostPopularDest);
1068 TerminatorInst *TI = BB->getTerminator();
1069 for (unsigned i = 0; ; ++i) {
1070 assert(i != TI->getNumSuccessors() && "Didn't find any successor!");
1072 if (std::find(SamePopularity.begin(), SamePopularity.end(),
1073 TI->getSuccessor(i)) == SamePopularity.end())
1076 MostPopularDest = TI->getSuccessor(i);
1081 // Okay, we have finally picked the most popular destination.
1082 return MostPopularDest;
1085 bool JumpThreading::ProcessThreadableEdges(Value *Cond, BasicBlock *BB,
1086 ConstantPreference Preference) {
1087 // If threading this would thread across a loop header, don't even try to
1089 if (LoopHeaders.count(BB))
1092 PredValueInfoTy PredValues;
1093 if (!ComputeValueKnownInPredecessors(Cond, BB, PredValues, Preference))
1096 assert(!PredValues.empty() &&
1097 "ComputeValueKnownInPredecessors returned true with no values");
1099 DEBUG(dbgs() << "IN BB: " << *BB;
1100 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1101 dbgs() << " BB '" << BB->getName() << "': FOUND condition = "
1102 << *PredValues[i].first
1103 << " for pred '" << PredValues[i].second->getName() << "'.\n";
1106 // Decide what we want to thread through. Convert our list of known values to
1107 // a list of known destinations for each pred. This also discards duplicate
1108 // predecessors and keeps track of the undefined inputs (which are represented
1109 // as a null dest in the PredToDestList).
1110 SmallPtrSet<BasicBlock*, 16> SeenPreds;
1111 SmallVector<std::pair<BasicBlock*, BasicBlock*>, 16> PredToDestList;
1113 BasicBlock *OnlyDest = nullptr;
1114 BasicBlock *MultipleDestSentinel = (BasicBlock*)(intptr_t)~0ULL;
1116 for (unsigned i = 0, e = PredValues.size(); i != e; ++i) {
1117 BasicBlock *Pred = PredValues[i].second;
1118 if (!SeenPreds.insert(Pred))
1119 continue; // Duplicate predecessor entry.
1121 // If the predecessor ends with an indirect goto, we can't change its
1123 if (isa<IndirectBrInst>(Pred->getTerminator()))
1126 Constant *Val = PredValues[i].first;
1129 if (isa<UndefValue>(Val))
1131 else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator()))
1132 DestBB = BI->getSuccessor(cast<ConstantInt>(Val)->isZero());
1133 else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
1134 DestBB = SI->findCaseValue(cast<ConstantInt>(Val)).getCaseSuccessor();
1136 assert(isa<IndirectBrInst>(BB->getTerminator())
1137 && "Unexpected terminator");
1138 DestBB = cast<BlockAddress>(Val)->getBasicBlock();
1141 // If we have exactly one destination, remember it for efficiency below.
1142 if (PredToDestList.empty())
1144 else if (OnlyDest != DestBB)
1145 OnlyDest = MultipleDestSentinel;
1147 PredToDestList.push_back(std::make_pair(Pred, DestBB));
1150 // If all edges were unthreadable, we fail.
1151 if (PredToDestList.empty())
1154 // Determine which is the most common successor. If we have many inputs and
1155 // this block is a switch, we want to start by threading the batch that goes
1156 // to the most popular destination first. If we only know about one
1157 // threadable destination (the common case) we can avoid this.
1158 BasicBlock *MostPopularDest = OnlyDest;
1160 if (MostPopularDest == MultipleDestSentinel)
1161 MostPopularDest = FindMostPopularDest(BB, PredToDestList);
1163 // Now that we know what the most popular destination is, factor all
1164 // predecessors that will jump to it into a single predecessor.
1165 SmallVector<BasicBlock*, 16> PredsToFactor;
1166 for (unsigned i = 0, e = PredToDestList.size(); i != e; ++i)
1167 if (PredToDestList[i].second == MostPopularDest) {
1168 BasicBlock *Pred = PredToDestList[i].first;
1170 // This predecessor may be a switch or something else that has multiple
1171 // edges to the block. Factor each of these edges by listing them
1172 // according to # occurrences in PredsToFactor.
1173 TerminatorInst *PredTI = Pred->getTerminator();
1174 for (unsigned i = 0, e = PredTI->getNumSuccessors(); i != e; ++i)
1175 if (PredTI->getSuccessor(i) == BB)
1176 PredsToFactor.push_back(Pred);
1179 // If the threadable edges are branching on an undefined value, we get to pick
1180 // the destination that these predecessors should get to.
1181 if (!MostPopularDest)
1182 MostPopularDest = BB->getTerminator()->
1183 getSuccessor(GetBestDestForJumpOnUndef(BB));
1185 // Ok, try to thread it!
1186 return ThreadEdge(BB, PredsToFactor, MostPopularDest);
1189 /// ProcessBranchOnPHI - We have an otherwise unthreadable conditional branch on
1190 /// a PHI node in the current block. See if there are any simplifications we
1191 /// can do based on inputs to the phi node.
1193 bool JumpThreading::ProcessBranchOnPHI(PHINode *PN) {
1194 BasicBlock *BB = PN->getParent();
1196 // TODO: We could make use of this to do it once for blocks with common PHI
1198 SmallVector<BasicBlock*, 1> PredBBs;
1201 // If any of the predecessor blocks end in an unconditional branch, we can
1202 // *duplicate* the conditional branch into that block in order to further
1203 // encourage jump threading and to eliminate cases where we have branch on a
1204 // phi of an icmp (branch on icmp is much better).
1205 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
1206 BasicBlock *PredBB = PN->getIncomingBlock(i);
1207 if (BranchInst *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator()))
1208 if (PredBr->isUnconditional()) {
1209 PredBBs[0] = PredBB;
1210 // Try to duplicate BB into PredBB.
1211 if (DuplicateCondBranchOnPHIIntoPred(BB, PredBBs))
1219 /// ProcessBranchOnXOR - We have an otherwise unthreadable conditional branch on
1220 /// a xor instruction in the current block. See if there are any
1221 /// simplifications we can do based on inputs to the xor.
1223 bool JumpThreading::ProcessBranchOnXOR(BinaryOperator *BO) {
1224 BasicBlock *BB = BO->getParent();
1226 // If either the LHS or RHS of the xor is a constant, don't do this
1228 if (isa<ConstantInt>(BO->getOperand(0)) ||
1229 isa<ConstantInt>(BO->getOperand(1)))
1232 // If the first instruction in BB isn't a phi, we won't be able to infer
1233 // anything special about any particular predecessor.
1234 if (!isa<PHINode>(BB->front()))
1237 // If we have a xor as the branch input to this block, and we know that the
1238 // LHS or RHS of the xor in any predecessor is true/false, then we can clone
1239 // the condition into the predecessor and fix that value to true, saving some
1240 // logical ops on that path and encouraging other paths to simplify.
1242 // This copies something like this:
1245 // %X = phi i1 [1], [%X']
1246 // %Y = icmp eq i32 %A, %B
1247 // %Z = xor i1 %X, %Y
1252 // %Y = icmp ne i32 %A, %B
1255 PredValueInfoTy XorOpValues;
1257 if (!ComputeValueKnownInPredecessors(BO->getOperand(0), BB, XorOpValues,
1259 assert(XorOpValues.empty());
1260 if (!ComputeValueKnownInPredecessors(BO->getOperand(1), BB, XorOpValues,
1266 assert(!XorOpValues.empty() &&
1267 "ComputeValueKnownInPredecessors returned true with no values");
1269 // Scan the information to see which is most popular: true or false. The
1270 // predecessors can be of the set true, false, or undef.
1271 unsigned NumTrue = 0, NumFalse = 0;
1272 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1273 if (isa<UndefValue>(XorOpValues[i].first))
1274 // Ignore undefs for the count.
1276 if (cast<ConstantInt>(XorOpValues[i].first)->isZero())
1282 // Determine which value to split on, true, false, or undef if neither.
1283 ConstantInt *SplitVal = nullptr;
1284 if (NumTrue > NumFalse)
1285 SplitVal = ConstantInt::getTrue(BB->getContext());
1286 else if (NumTrue != 0 || NumFalse != 0)
1287 SplitVal = ConstantInt::getFalse(BB->getContext());
1289 // Collect all of the blocks that this can be folded into so that we can
1290 // factor this once and clone it once.
1291 SmallVector<BasicBlock*, 8> BlocksToFoldInto;
1292 for (unsigned i = 0, e = XorOpValues.size(); i != e; ++i) {
1293 if (XorOpValues[i].first != SplitVal &&
1294 !isa<UndefValue>(XorOpValues[i].first))
1297 BlocksToFoldInto.push_back(XorOpValues[i].second);
1300 // If we inferred a value for all of the predecessors, then duplication won't
1301 // help us. However, we can just replace the LHS or RHS with the constant.
1302 if (BlocksToFoldInto.size() ==
1303 cast<PHINode>(BB->front()).getNumIncomingValues()) {
1305 // If all preds provide undef, just nuke the xor, because it is undef too.
1306 BO->replaceAllUsesWith(UndefValue::get(BO->getType()));
1307 BO->eraseFromParent();
1308 } else if (SplitVal->isZero()) {
1309 // If all preds provide 0, replace the xor with the other input.
1310 BO->replaceAllUsesWith(BO->getOperand(isLHS));
1311 BO->eraseFromParent();
1313 // If all preds provide 1, set the computed value to 1.
1314 BO->setOperand(!isLHS, SplitVal);
1320 // Try to duplicate BB into PredBB.
1321 return DuplicateCondBranchOnPHIIntoPred(BB, BlocksToFoldInto);
1325 /// AddPHINodeEntriesForMappedBlock - We're adding 'NewPred' as a new
1326 /// predecessor to the PHIBB block. If it has PHI nodes, add entries for
1327 /// NewPred using the entries from OldPred (suitably mapped).
1328 static void AddPHINodeEntriesForMappedBlock(BasicBlock *PHIBB,
1329 BasicBlock *OldPred,
1330 BasicBlock *NewPred,
1331 DenseMap<Instruction*, Value*> &ValueMap) {
1332 for (BasicBlock::iterator PNI = PHIBB->begin();
1333 PHINode *PN = dyn_cast<PHINode>(PNI); ++PNI) {
1334 // Ok, we have a PHI node. Figure out what the incoming value was for the
1336 Value *IV = PN->getIncomingValueForBlock(OldPred);
1338 // Remap the value if necessary.
1339 if (Instruction *Inst = dyn_cast<Instruction>(IV)) {
1340 DenseMap<Instruction*, Value*>::iterator I = ValueMap.find(Inst);
1341 if (I != ValueMap.end())
1345 PN->addIncoming(IV, NewPred);
1349 /// ThreadEdge - We have decided that it is safe and profitable to factor the
1350 /// blocks in PredBBs to one predecessor, then thread an edge from it to SuccBB
1351 /// across BB. Transform the IR to reflect this change.
1352 bool JumpThreading::ThreadEdge(BasicBlock *BB,
1353 const SmallVectorImpl<BasicBlock*> &PredBBs,
1354 BasicBlock *SuccBB) {
1355 // If threading to the same block as we come from, we would infinite loop.
1357 DEBUG(dbgs() << " Not threading across BB '" << BB->getName()
1358 << "' - would thread to self!\n");
1362 // If threading this would thread across a loop header, don't thread the edge.
1363 // See the comments above FindLoopHeaders for justifications and caveats.
1364 if (LoopHeaders.count(BB)) {
1365 DEBUG(dbgs() << " Not threading across loop header BB '" << BB->getName()
1366 << "' to dest BB '" << SuccBB->getName()
1367 << "' - it might create an irreducible loop!\n");
1371 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB, Threshold);
1372 if (JumpThreadCost > Threshold) {
1373 DEBUG(dbgs() << " Not threading BB '" << BB->getName()
1374 << "' - Cost is too high: " << JumpThreadCost << "\n");
1378 // And finally, do it! Start by factoring the predecessors is needed.
1380 if (PredBBs.size() == 1)
1381 PredBB = PredBBs[0];
1383 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1384 << " common predecessors.\n");
1385 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1388 // And finally, do it!
1389 DEBUG(dbgs() << " Threading edge from '" << PredBB->getName() << "' to '"
1390 << SuccBB->getName() << "' with cost: " << JumpThreadCost
1391 << ", across block:\n "
1394 LVI->threadEdge(PredBB, BB, SuccBB);
1396 // We are going to have to map operands from the original BB block to the new
1397 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to
1398 // account for entry from PredBB.
1399 DenseMap<Instruction*, Value*> ValueMapping;
1401 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(),
1402 BB->getName()+".thread",
1403 BB->getParent(), BB);
1404 NewBB->moveAfter(PredBB);
1406 BasicBlock::iterator BI = BB->begin();
1407 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1408 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1410 // Clone the non-phi instructions of BB into NewBB, keeping track of the
1411 // mapping and using it to remap operands in the cloned instructions.
1412 for (; !isa<TerminatorInst>(BI); ++BI) {
1413 Instruction *New = BI->clone();
1414 New->setName(BI->getName());
1415 NewBB->getInstList().push_back(New);
1416 ValueMapping[BI] = New;
1418 // Remap operands to patch up intra-block references.
1419 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1420 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1421 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1422 if (I != ValueMapping.end())
1423 New->setOperand(i, I->second);
1427 // We didn't copy the terminator from BB over to NewBB, because there is now
1428 // an unconditional jump to SuccBB. Insert the unconditional jump.
1429 BranchInst *NewBI =BranchInst::Create(SuccBB, NewBB);
1430 NewBI->setDebugLoc(BB->getTerminator()->getDebugLoc());
1432 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the
1433 // PHI nodes for NewBB now.
1434 AddPHINodeEntriesForMappedBlock(SuccBB, BB, NewBB, ValueMapping);
1436 // If there were values defined in BB that are used outside the block, then we
1437 // now have to update all uses of the value to use either the original value,
1438 // the cloned value, or some PHI derived value. This can require arbitrary
1439 // PHI insertion, of which we are prepared to do, clean these up now.
1440 SSAUpdater SSAUpdate;
1441 SmallVector<Use*, 16> UsesToRename;
1442 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1443 // Scan all uses of this instruction to see if it is used outside of its
1444 // block, and if so, record them in UsesToRename.
1445 for (Use &U : I->uses()) {
1446 Instruction *User = cast<Instruction>(U.getUser());
1447 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1448 if (UserPN->getIncomingBlock(U) == BB)
1450 } else if (User->getParent() == BB)
1453 UsesToRename.push_back(&U);
1456 // If there are no uses outside the block, we're done with this instruction.
1457 if (UsesToRename.empty())
1460 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1462 // We found a use of I outside of BB. Rename all uses of I that are outside
1463 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1464 // with the two values we know.
1465 SSAUpdate.Initialize(I->getType(), I->getName());
1466 SSAUpdate.AddAvailableValue(BB, I);
1467 SSAUpdate.AddAvailableValue(NewBB, ValueMapping[I]);
1469 while (!UsesToRename.empty())
1470 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1471 DEBUG(dbgs() << "\n");
1475 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to
1476 // NewBB instead of BB. This eliminates predecessors from BB, which requires
1477 // us to simplify any PHI nodes in BB.
1478 TerminatorInst *PredTerm = PredBB->getTerminator();
1479 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i)
1480 if (PredTerm->getSuccessor(i) == BB) {
1481 BB->removePredecessor(PredBB, true);
1482 PredTerm->setSuccessor(i, NewBB);
1485 // At this point, the IR is fully up to date and consistent. Do a quick scan
1486 // over the new instructions and zap any that are constants or dead. This
1487 // frequently happens because of phi translation.
1488 SimplifyInstructionsInBlock(NewBB, DL, TLI);
1490 // Threaded an edge!
1495 /// DuplicateCondBranchOnPHIIntoPred - PredBB contains an unconditional branch
1496 /// to BB which contains an i1 PHI node and a conditional branch on that PHI.
1497 /// If we can duplicate the contents of BB up into PredBB do so now, this
1498 /// improves the odds that the branch will be on an analyzable instruction like
1500 bool JumpThreading::DuplicateCondBranchOnPHIIntoPred(BasicBlock *BB,
1501 const SmallVectorImpl<BasicBlock *> &PredBBs) {
1502 assert(!PredBBs.empty() && "Can't handle an empty set");
1504 // If BB is a loop header, then duplicating this block outside the loop would
1505 // cause us to transform this into an irreducible loop, don't do this.
1506 // See the comments above FindLoopHeaders for justifications and caveats.
1507 if (LoopHeaders.count(BB)) {
1508 DEBUG(dbgs() << " Not duplicating loop header '" << BB->getName()
1509 << "' into predecessor block '" << PredBBs[0]->getName()
1510 << "' - it might create an irreducible loop!\n");
1514 unsigned DuplicationCost = getJumpThreadDuplicationCost(BB, Threshold);
1515 if (DuplicationCost > Threshold) {
1516 DEBUG(dbgs() << " Not duplicating BB '" << BB->getName()
1517 << "' - Cost is too high: " << DuplicationCost << "\n");
1521 // And finally, do it! Start by factoring the predecessors is needed.
1523 if (PredBBs.size() == 1)
1524 PredBB = PredBBs[0];
1526 DEBUG(dbgs() << " Factoring out " << PredBBs.size()
1527 << " common predecessors.\n");
1528 PredBB = SplitBlockPredecessors(BB, PredBBs, ".thr_comm", this);
1531 // Okay, we decided to do this! Clone all the instructions in BB onto the end
1533 DEBUG(dbgs() << " Duplicating block '" << BB->getName() << "' into end of '"
1534 << PredBB->getName() << "' to eliminate branch on phi. Cost: "
1535 << DuplicationCost << " block is:" << *BB << "\n");
1537 // Unless PredBB ends with an unconditional branch, split the edge so that we
1538 // can just clone the bits from BB into the end of the new PredBB.
1539 BranchInst *OldPredBranch = dyn_cast<BranchInst>(PredBB->getTerminator());
1541 if (!OldPredBranch || !OldPredBranch->isUnconditional()) {
1542 PredBB = SplitEdge(PredBB, BB, this);
1543 OldPredBranch = cast<BranchInst>(PredBB->getTerminator());
1546 // We are going to have to map operands from the original BB block into the
1547 // PredBB block. Evaluate PHI nodes in BB.
1548 DenseMap<Instruction*, Value*> ValueMapping;
1550 BasicBlock::iterator BI = BB->begin();
1551 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
1552 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB);
1554 // Clone the non-phi instructions of BB into PredBB, keeping track of the
1555 // mapping and using it to remap operands in the cloned instructions.
1556 for (; BI != BB->end(); ++BI) {
1557 Instruction *New = BI->clone();
1559 // Remap operands to patch up intra-block references.
1560 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i)
1561 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) {
1562 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst);
1563 if (I != ValueMapping.end())
1564 New->setOperand(i, I->second);
1567 // If this instruction can be simplified after the operands are updated,
1568 // just use the simplified value instead. This frequently happens due to
1570 if (Value *IV = SimplifyInstruction(New, DL)) {
1572 ValueMapping[BI] = IV;
1574 // Otherwise, insert the new instruction into the block.
1575 New->setName(BI->getName());
1576 PredBB->getInstList().insert(OldPredBranch, New);
1577 ValueMapping[BI] = New;
1581 // Check to see if the targets of the branch had PHI nodes. If so, we need to
1582 // add entries to the PHI nodes for branch from PredBB now.
1583 BranchInst *BBBranch = cast<BranchInst>(BB->getTerminator());
1584 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(0), BB, PredBB,
1586 AddPHINodeEntriesForMappedBlock(BBBranch->getSuccessor(1), BB, PredBB,
1589 // If there were values defined in BB that are used outside the block, then we
1590 // now have to update all uses of the value to use either the original value,
1591 // the cloned value, or some PHI derived value. This can require arbitrary
1592 // PHI insertion, of which we are prepared to do, clean these up now.
1593 SSAUpdater SSAUpdate;
1594 SmallVector<Use*, 16> UsesToRename;
1595 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
1596 // Scan all uses of this instruction to see if it is used outside of its
1597 // block, and if so, record them in UsesToRename.
1598 for (Use &U : I->uses()) {
1599 Instruction *User = cast<Instruction>(U.getUser());
1600 if (PHINode *UserPN = dyn_cast<PHINode>(User)) {
1601 if (UserPN->getIncomingBlock(U) == BB)
1603 } else if (User->getParent() == BB)
1606 UsesToRename.push_back(&U);
1609 // If there are no uses outside the block, we're done with this instruction.
1610 if (UsesToRename.empty())
1613 DEBUG(dbgs() << "JT: Renaming non-local uses of: " << *I << "\n");
1615 // We found a use of I outside of BB. Rename all uses of I that are outside
1616 // its block to be uses of the appropriate PHI node etc. See ValuesInBlocks
1617 // with the two values we know.
1618 SSAUpdate.Initialize(I->getType(), I->getName());
1619 SSAUpdate.AddAvailableValue(BB, I);
1620 SSAUpdate.AddAvailableValue(PredBB, ValueMapping[I]);
1622 while (!UsesToRename.empty())
1623 SSAUpdate.RewriteUse(*UsesToRename.pop_back_val());
1624 DEBUG(dbgs() << "\n");
1627 // PredBB no longer jumps to BB, remove entries in the PHI node for the edge
1629 BB->removePredecessor(PredBB, true);
1631 // Remove the unconditional branch at the end of the PredBB block.
1632 OldPredBranch->eraseFromParent();
1638 /// TryToUnfoldSelect - Look for blocks of the form
1644 /// %p = phi [%a, %bb] ...
1648 /// And expand the select into a branch structure if one of its arms allows %c
1649 /// to be folded. This later enables threading from bb1 over bb2.
1650 bool JumpThreading::TryToUnfoldSelect(CmpInst *CondCmp, BasicBlock *BB) {
1651 BranchInst *CondBr = dyn_cast<BranchInst>(BB->getTerminator());
1652 PHINode *CondLHS = dyn_cast<PHINode>(CondCmp->getOperand(0));
1653 Constant *CondRHS = cast<Constant>(CondCmp->getOperand(1));
1655 if (!CondBr || !CondBr->isConditional() || !CondLHS ||
1656 CondLHS->getParent() != BB)
1659 for (unsigned I = 0, E = CondLHS->getNumIncomingValues(); I != E; ++I) {
1660 BasicBlock *Pred = CondLHS->getIncomingBlock(I);
1661 SelectInst *SI = dyn_cast<SelectInst>(CondLHS->getIncomingValue(I));
1663 // Look if one of the incoming values is a select in the corresponding
1665 if (!SI || SI->getParent() != Pred || !SI->hasOneUse())
1668 BranchInst *PredTerm = dyn_cast<BranchInst>(Pred->getTerminator());
1669 if (!PredTerm || !PredTerm->isUnconditional())
1672 // Now check if one of the select values would allow us to constant fold the
1673 // terminator in BB. We don't do the transform if both sides fold, those
1674 // cases will be threaded in any case.
1675 LazyValueInfo::Tristate LHSFolds =
1676 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(1),
1678 LazyValueInfo::Tristate RHSFolds =
1679 LVI->getPredicateOnEdge(CondCmp->getPredicate(), SI->getOperand(2),
1681 if ((LHSFolds != LazyValueInfo::Unknown ||
1682 RHSFolds != LazyValueInfo::Unknown) &&
1683 LHSFolds != RHSFolds) {
1684 // Expand the select.
1693 BasicBlock *NewBB = BasicBlock::Create(BB->getContext(), "select.unfold",
1694 BB->getParent(), BB);
1695 // Move the unconditional branch to NewBB.
1696 PredTerm->removeFromParent();
1697 NewBB->getInstList().insert(NewBB->end(), PredTerm);
1698 // Create a conditional branch and update PHI nodes.
1699 BranchInst::Create(NewBB, BB, SI->getCondition(), Pred);
1700 CondLHS->setIncomingValue(I, SI->getFalseValue());
1701 CondLHS->addIncoming(SI->getTrueValue(), NewBB);
1702 // The select is now dead.
1703 SI->eraseFromParent();
1705 // Update any other PHI nodes in BB.
1706 for (BasicBlock::iterator BI = BB->begin();
1707 PHINode *Phi = dyn_cast<PHINode>(BI); ++BI)
1709 Phi->addIncoming(Phi->getIncomingValueForBlock(Pred), NewBB);